xref: /openbmc/linux/tools/perf/util/stat-shadow.c (revision 29c37341)
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 
22 struct runtime_stat rt_stat;
23 struct stats walltime_nsecs_stats;
24 
25 struct saved_value {
26 	struct rb_node rb_node;
27 	struct evsel *evsel;
28 	enum stat_type type;
29 	int ctx;
30 	int cpu;
31 	struct runtime_stat *stat;
32 	struct stats stats;
33 	u64 metric_total;
34 	int metric_other;
35 };
36 
37 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
38 {
39 	struct saved_value *a = container_of(rb_node,
40 					     struct saved_value,
41 					     rb_node);
42 	const struct saved_value *b = entry;
43 
44 	if (a->cpu != b->cpu)
45 		return a->cpu - b->cpu;
46 
47 	/*
48 	 * Previously the rbtree was used to link generic metrics.
49 	 * The keys were evsel/cpu. Now the rbtree is extended to support
50 	 * per-thread shadow stats. For shadow stats case, the keys
51 	 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
52 	 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
53 	 */
54 	if (a->type != b->type)
55 		return a->type - b->type;
56 
57 	if (a->ctx != b->ctx)
58 		return a->ctx - b->ctx;
59 
60 	if (a->evsel == NULL && b->evsel == NULL) {
61 		if (a->stat == b->stat)
62 			return 0;
63 
64 		if ((char *)a->stat < (char *)b->stat)
65 			return -1;
66 
67 		return 1;
68 	}
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 
77 static 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 
88 static 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 
98 static struct saved_value *saved_value_lookup(struct evsel *evsel,
99 					      int cpu,
100 					      bool create,
101 					      enum stat_type type,
102 					      int ctx,
103 					      struct runtime_stat *st)
104 {
105 	struct rblist *rblist;
106 	struct rb_node *nd;
107 	struct saved_value dm = {
108 		.cpu = cpu,
109 		.evsel = evsel,
110 		.type = type,
111 		.ctx = ctx,
112 		.stat = st,
113 	};
114 
115 	rblist = &st->value_list;
116 
117 	nd = rblist__find(rblist, &dm);
118 	if (nd)
119 		return container_of(nd, struct saved_value, rb_node);
120 	if (create) {
121 		rblist__add_node(rblist, &dm);
122 		nd = rblist__find(rblist, &dm);
123 		if (nd)
124 			return container_of(nd, struct saved_value, rb_node);
125 	}
126 	return NULL;
127 }
128 
129 void runtime_stat__init(struct runtime_stat *st)
130 {
131 	struct rblist *rblist = &st->value_list;
132 
133 	rblist__init(rblist);
134 	rblist->node_cmp = saved_value_cmp;
135 	rblist->node_new = saved_value_new;
136 	rblist->node_delete = saved_value_delete;
137 }
138 
139 void runtime_stat__exit(struct runtime_stat *st)
140 {
141 	rblist__exit(&st->value_list);
142 }
143 
144 void perf_stat__init_shadow_stats(void)
145 {
146 	runtime_stat__init(&rt_stat);
147 }
148 
149 static int evsel_context(struct evsel *evsel)
150 {
151 	int ctx = 0;
152 
153 	if (evsel->core.attr.exclude_kernel)
154 		ctx |= CTX_BIT_KERNEL;
155 	if (evsel->core.attr.exclude_user)
156 		ctx |= CTX_BIT_USER;
157 	if (evsel->core.attr.exclude_hv)
158 		ctx |= CTX_BIT_HV;
159 	if (evsel->core.attr.exclude_host)
160 		ctx |= CTX_BIT_HOST;
161 	if (evsel->core.attr.exclude_idle)
162 		ctx |= CTX_BIT_IDLE;
163 
164 	return ctx;
165 }
166 
167 static void reset_stat(struct runtime_stat *st)
168 {
169 	struct rblist *rblist;
170 	struct rb_node *pos, *next;
171 
172 	rblist = &st->value_list;
173 	next = rb_first_cached(&rblist->entries);
174 	while (next) {
175 		pos = next;
176 		next = rb_next(pos);
177 		memset(&container_of(pos, struct saved_value, rb_node)->stats,
178 		       0,
179 		       sizeof(struct stats));
180 	}
181 }
182 
183 void perf_stat__reset_shadow_stats(void)
184 {
185 	reset_stat(&rt_stat);
186 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
187 }
188 
189 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
190 {
191 	reset_stat(st);
192 }
193 
194 static void update_runtime_stat(struct runtime_stat *st,
195 				enum stat_type type,
196 				int ctx, int cpu, u64 count)
197 {
198 	struct saved_value *v = saved_value_lookup(NULL, cpu, true,
199 						   type, ctx, st);
200 
201 	if (v)
202 		update_stats(&v->stats, count);
203 }
204 
205 /*
206  * Update various tracking values we maintain to print
207  * more semantic information such as miss/hit ratios,
208  * instruction rates, etc:
209  */
210 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
211 				    int cpu, struct runtime_stat *st)
212 {
213 	int ctx = evsel_context(counter);
214 	u64 count_ns = count;
215 	struct saved_value *v;
216 
217 	count *= counter->scale;
218 
219 	if (evsel__is_clock(counter))
220 		update_runtime_stat(st, STAT_NSECS, 0, cpu, count_ns);
221 	else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
222 		update_runtime_stat(st, STAT_CYCLES, ctx, cpu, count);
223 	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
224 		update_runtime_stat(st, STAT_CYCLES_IN_TX, ctx, cpu, count);
225 	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
226 		update_runtime_stat(st, STAT_TRANSACTION, ctx, cpu, count);
227 	else if (perf_stat_evsel__is(counter, ELISION_START))
228 		update_runtime_stat(st, STAT_ELISION, ctx, cpu, count);
229 	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
230 		update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
231 				    ctx, cpu, count);
232 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
233 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
234 				    ctx, cpu, count);
235 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
236 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
237 				    ctx, cpu, count);
238 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
239 		update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
240 				    ctx, cpu, count);
241 	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
242 		update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
243 				    ctx, cpu, count);
244 	else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
245 		update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
246 				    ctx, cpu, count);
247 	else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
248 		update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
249 				    ctx, cpu, count);
250 	else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
251 		update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count);
252 	else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
253 		update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count);
254 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
255 		update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count);
256 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
257 		update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count);
258 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
259 		update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count);
260 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
261 		update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count);
262 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
263 		update_runtime_stat(st, STAT_ITLB_CACHE, ctx, cpu, count);
264 	else if (perf_stat_evsel__is(counter, SMI_NUM))
265 		update_runtime_stat(st, STAT_SMI_NUM, ctx, cpu, count);
266 	else if (perf_stat_evsel__is(counter, APERF))
267 		update_runtime_stat(st, STAT_APERF, ctx, cpu, count);
268 
269 	if (counter->collect_stat) {
270 		v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st);
271 		update_stats(&v->stats, count);
272 		if (counter->metric_leader)
273 			v->metric_total += count;
274 	} else if (counter->metric_leader) {
275 		v = saved_value_lookup(counter->metric_leader,
276 				       cpu, true, STAT_NONE, 0, st);
277 		v->metric_total += count;
278 		v->metric_other++;
279 	}
280 }
281 
282 /* used for get_ratio_color() */
283 enum grc_type {
284 	GRC_STALLED_CYCLES_FE,
285 	GRC_STALLED_CYCLES_BE,
286 	GRC_CACHE_MISSES,
287 	GRC_MAX_NR
288 };
289 
290 static const char *get_ratio_color(enum grc_type type, double ratio)
291 {
292 	static const double grc_table[GRC_MAX_NR][3] = {
293 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
294 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
295 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
296 	};
297 	const char *color = PERF_COLOR_NORMAL;
298 
299 	if (ratio > grc_table[type][0])
300 		color = PERF_COLOR_RED;
301 	else if (ratio > grc_table[type][1])
302 		color = PERF_COLOR_MAGENTA;
303 	else if (ratio > grc_table[type][2])
304 		color = PERF_COLOR_YELLOW;
305 
306 	return color;
307 }
308 
309 static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
310 						const char *name)
311 {
312 	struct evsel *c2;
313 
314 	evlist__for_each_entry (evsel_list, c2) {
315 		if (!strcasecmp(c2->name, name) && !c2->collect_stat)
316 			return c2;
317 	}
318 	return NULL;
319 }
320 
321 /* Mark MetricExpr target events and link events using them to them. */
322 void perf_stat__collect_metric_expr(struct evlist *evsel_list)
323 {
324 	struct evsel *counter, *leader, **metric_events, *oc;
325 	bool found;
326 	struct expr_parse_ctx ctx;
327 	struct hashmap_entry *cur;
328 	size_t bkt;
329 	int i;
330 
331 	expr__ctx_init(&ctx);
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 
339 		expr__ctx_clear(&ctx);
340 		metric_events = counter->metric_events;
341 		if (!metric_events) {
342 			if (expr__find_other(counter->metric_expr,
343 					     counter->name,
344 					     &ctx, 1) < 0)
345 				continue;
346 
347 			metric_events = calloc(sizeof(struct evsel *),
348 					       hashmap__size(&ctx.ids) + 1);
349 			if (!metric_events) {
350 				expr__ctx_clear(&ctx);
351 				return;
352 			}
353 			counter->metric_events = metric_events;
354 		}
355 
356 		i = 0;
357 		hashmap__for_each_entry((&ctx.ids), cur, bkt) {
358 			const char *metric_name = (const char *)cur->key;
359 
360 			found = false;
361 			if (leader) {
362 				/* Search in group */
363 				for_each_group_member (oc, leader) {
364 					if (!strcasecmp(oc->name,
365 							metric_name) &&
366 						!oc->collect_stat) {
367 						found = true;
368 						break;
369 					}
370 				}
371 			}
372 			if (!found) {
373 				/* Search ignoring groups */
374 				oc = perf_stat__find_event(evsel_list,
375 							   metric_name);
376 			}
377 			if (!oc) {
378 				/* Deduping one is good enough to handle duplicated PMUs. */
379 				static char *printed;
380 
381 				/*
382 				 * Adding events automatically would be difficult, because
383 				 * it would risk creating groups that are not schedulable.
384 				 * perf stat doesn't understand all the scheduling constraints
385 				 * of events. So we ask the user instead to add the missing
386 				 * events.
387 				 */
388 				if (!printed ||
389 				    strcasecmp(printed, metric_name)) {
390 					fprintf(stderr,
391 						"Add %s event to groups to get metric expression for %s\n",
392 						metric_name,
393 						counter->name);
394 					printed = strdup(metric_name);
395 				}
396 				invalid = true;
397 				continue;
398 			}
399 			metric_events[i++] = oc;
400 			oc->collect_stat = true;
401 		}
402 		metric_events[i] = NULL;
403 		if (invalid) {
404 			free(metric_events);
405 			counter->metric_events = NULL;
406 			counter->metric_expr = NULL;
407 		}
408 	}
409 	expr__ctx_clear(&ctx);
410 }
411 
412 static double runtime_stat_avg(struct runtime_stat *st,
413 			       enum stat_type type, int ctx, int cpu)
414 {
415 	struct saved_value *v;
416 
417 	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
418 	if (!v)
419 		return 0.0;
420 
421 	return avg_stats(&v->stats);
422 }
423 
424 static double runtime_stat_n(struct runtime_stat *st,
425 			     enum stat_type type, int ctx, int cpu)
426 {
427 	struct saved_value *v;
428 
429 	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
430 	if (!v)
431 		return 0.0;
432 
433 	return v->stats.n;
434 }
435 
436 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
437 					  int cpu,
438 					  struct evsel *evsel, double avg,
439 					  struct perf_stat_output_ctx *out,
440 					  struct runtime_stat *st)
441 {
442 	double total, ratio = 0.0;
443 	const char *color;
444 	int ctx = evsel_context(evsel);
445 
446 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
447 
448 	if (total)
449 		ratio = avg / total * 100.0;
450 
451 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
452 
453 	if (ratio)
454 		out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
455 				  ratio);
456 	else
457 		out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
458 }
459 
460 static void print_stalled_cycles_backend(struct perf_stat_config *config,
461 					 int cpu,
462 					 struct evsel *evsel, double avg,
463 					 struct perf_stat_output_ctx *out,
464 					 struct runtime_stat *st)
465 {
466 	double total, ratio = 0.0;
467 	const char *color;
468 	int ctx = evsel_context(evsel);
469 
470 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
471 
472 	if (total)
473 		ratio = avg / total * 100.0;
474 
475 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
476 
477 	out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
478 }
479 
480 static void print_branch_misses(struct perf_stat_config *config,
481 				int cpu,
482 				struct evsel *evsel,
483 				double avg,
484 				struct perf_stat_output_ctx *out,
485 				struct runtime_stat *st)
486 {
487 	double total, ratio = 0.0;
488 	const char *color;
489 	int ctx = evsel_context(evsel);
490 
491 	total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
492 
493 	if (total)
494 		ratio = avg / total * 100.0;
495 
496 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
497 
498 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
499 }
500 
501 static void print_l1_dcache_misses(struct perf_stat_config *config,
502 				   int cpu,
503 				   struct evsel *evsel,
504 				   double avg,
505 				   struct perf_stat_output_ctx *out,
506 				   struct runtime_stat *st)
507 
508 {
509 	double total, ratio = 0.0;
510 	const char *color;
511 	int ctx = evsel_context(evsel);
512 
513 	total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
514 
515 	if (total)
516 		ratio = avg / total * 100.0;
517 
518 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
519 
520 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
521 }
522 
523 static void print_l1_icache_misses(struct perf_stat_config *config,
524 				   int cpu,
525 				   struct evsel *evsel,
526 				   double avg,
527 				   struct perf_stat_output_ctx *out,
528 				   struct runtime_stat *st)
529 
530 {
531 	double total, ratio = 0.0;
532 	const char *color;
533 	int ctx = evsel_context(evsel);
534 
535 	total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
536 
537 	if (total)
538 		ratio = avg / total * 100.0;
539 
540 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
541 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
542 }
543 
544 static void print_dtlb_cache_misses(struct perf_stat_config *config,
545 				    int cpu,
546 				    struct evsel *evsel,
547 				    double avg,
548 				    struct perf_stat_output_ctx *out,
549 				    struct runtime_stat *st)
550 {
551 	double total, ratio = 0.0;
552 	const char *color;
553 	int ctx = evsel_context(evsel);
554 
555 	total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
556 
557 	if (total)
558 		ratio = avg / total * 100.0;
559 
560 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
561 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
562 }
563 
564 static void print_itlb_cache_misses(struct perf_stat_config *config,
565 				    int cpu,
566 				    struct evsel *evsel,
567 				    double avg,
568 				    struct perf_stat_output_ctx *out,
569 				    struct runtime_stat *st)
570 {
571 	double total, ratio = 0.0;
572 	const char *color;
573 	int ctx = evsel_context(evsel);
574 
575 	total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
576 
577 	if (total)
578 		ratio = avg / total * 100.0;
579 
580 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
581 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
582 }
583 
584 static void print_ll_cache_misses(struct perf_stat_config *config,
585 				  int cpu,
586 				  struct evsel *evsel,
587 				  double avg,
588 				  struct perf_stat_output_ctx *out,
589 				  struct runtime_stat *st)
590 {
591 	double total, ratio = 0.0;
592 	const char *color;
593 	int ctx = evsel_context(evsel);
594 
595 	total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
596 
597 	if (total)
598 		ratio = avg / total * 100.0;
599 
600 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
601 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
602 }
603 
604 /*
605  * High level "TopDown" CPU core pipe line bottleneck break down.
606  *
607  * Basic concept following
608  * Yasin, A Top Down Method for Performance analysis and Counter architecture
609  * ISPASS14
610  *
611  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
612  *
613  * Frontend -> Backend -> Retiring
614  * BadSpeculation in addition means out of order execution that is thrown away
615  * (for example branch mispredictions)
616  * Frontend is instruction decoding.
617  * Backend is execution, like computation and accessing data in memory
618  * Retiring is good execution that is not directly bottlenecked
619  *
620  * The formulas are computed in slots.
621  * A slot is an entry in the pipeline each for the pipeline width
622  * (for example a 4-wide pipeline has 4 slots for each cycle)
623  *
624  * Formulas:
625  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
626  *			TotalSlots
627  * Retiring = SlotsRetired / TotalSlots
628  * FrontendBound = FetchBubbles / TotalSlots
629  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
630  *
631  * The kernel provides the mapping to the low level CPU events and any scaling
632  * needed for the CPU pipeline width, for example:
633  *
634  * TotalSlots = Cycles * 4
635  *
636  * The scaling factor is communicated in the sysfs unit.
637  *
638  * In some cases the CPU may not be able to measure all the formulas due to
639  * missing events. In this case multiple formulas are combined, as possible.
640  *
641  * Full TopDown supports more levels to sub-divide each area: for example
642  * BackendBound into computing bound and memory bound. For now we only
643  * support Level 1 TopDown.
644  */
645 
646 static double sanitize_val(double x)
647 {
648 	if (x < 0 && x >= -0.02)
649 		return 0.0;
650 	return x;
651 }
652 
653 static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
654 {
655 	return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
656 }
657 
658 static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
659 {
660 	double bad_spec = 0;
661 	double total_slots;
662 	double total;
663 
664 	total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
665 		runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
666 		runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
667 
668 	total_slots = td_total_slots(ctx, cpu, st);
669 	if (total_slots)
670 		bad_spec = total / total_slots;
671 	return sanitize_val(bad_spec);
672 }
673 
674 static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
675 {
676 	double retiring = 0;
677 	double total_slots = td_total_slots(ctx, cpu, st);
678 	double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
679 					    ctx, cpu);
680 
681 	if (total_slots)
682 		retiring = ret_slots / total_slots;
683 	return retiring;
684 }
685 
686 static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
687 {
688 	double fe_bound = 0;
689 	double total_slots = td_total_slots(ctx, cpu, st);
690 	double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
691 					    ctx, cpu);
692 
693 	if (total_slots)
694 		fe_bound = fetch_bub / total_slots;
695 	return fe_bound;
696 }
697 
698 static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
699 {
700 	double sum = (td_fe_bound(ctx, cpu, st) +
701 		      td_bad_spec(ctx, cpu, st) +
702 		      td_retiring(ctx, cpu, st));
703 	if (sum == 0)
704 		return 0;
705 	return sanitize_val(1.0 - sum);
706 }
707 
708 static void print_smi_cost(struct perf_stat_config *config,
709 			   int cpu, struct evsel *evsel,
710 			   struct perf_stat_output_ctx *out,
711 			   struct runtime_stat *st)
712 {
713 	double smi_num, aperf, cycles, cost = 0.0;
714 	int ctx = evsel_context(evsel);
715 	const char *color = NULL;
716 
717 	smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
718 	aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
719 	cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
720 
721 	if ((cycles == 0) || (aperf == 0))
722 		return;
723 
724 	if (smi_num)
725 		cost = (aperf - cycles) / aperf * 100.00;
726 
727 	if (cost > 10)
728 		color = PERF_COLOR_RED;
729 	out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
730 	out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
731 }
732 
733 static int prepare_metric(struct evsel **metric_events,
734 			  struct metric_ref *metric_refs,
735 			  struct expr_parse_ctx *pctx,
736 			  int cpu,
737 			  struct runtime_stat *st)
738 {
739 	double scale;
740 	char *n, *pn;
741 	int i, j, ret;
742 
743 	expr__ctx_init(pctx);
744 	for (i = 0; metric_events[i]; i++) {
745 		struct saved_value *v;
746 		struct stats *stats;
747 		u64 metric_total = 0;
748 
749 		if (!strcmp(metric_events[i]->name, "duration_time")) {
750 			stats = &walltime_nsecs_stats;
751 			scale = 1e-9;
752 		} else {
753 			v = saved_value_lookup(metric_events[i], cpu, false,
754 					       STAT_NONE, 0, st);
755 			if (!v)
756 				break;
757 			stats = &v->stats;
758 			scale = 1.0;
759 
760 			if (v->metric_other)
761 				metric_total = v->metric_total;
762 		}
763 
764 		n = strdup(metric_events[i]->name);
765 		if (!n)
766 			return -ENOMEM;
767 		/*
768 		 * This display code with --no-merge adds [cpu] postfixes.
769 		 * These are not supported by the parser. Remove everything
770 		 * after the space.
771 		 */
772 		pn = strchr(n, ' ');
773 		if (pn)
774 			*pn = 0;
775 
776 		if (metric_total)
777 			expr__add_id_val(pctx, n, metric_total);
778 		else
779 			expr__add_id_val(pctx, n, avg_stats(stats)*scale);
780 	}
781 
782 	for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
783 		ret = expr__add_ref(pctx, &metric_refs[j]);
784 		if (ret)
785 			return ret;
786 	}
787 
788 	return i;
789 }
790 
791 static void generic_metric(struct perf_stat_config *config,
792 			   const char *metric_expr,
793 			   struct evsel **metric_events,
794 			   struct metric_ref *metric_refs,
795 			   char *name,
796 			   const char *metric_name,
797 			   const char *metric_unit,
798 			   int runtime,
799 			   int cpu,
800 			   struct perf_stat_output_ctx *out,
801 			   struct runtime_stat *st)
802 {
803 	print_metric_t print_metric = out->print_metric;
804 	struct expr_parse_ctx pctx;
805 	double ratio, scale;
806 	int i;
807 	void *ctxp = out->ctx;
808 
809 	i = prepare_metric(metric_events, metric_refs, &pctx, cpu, st);
810 	if (i < 0)
811 		return;
812 
813 	if (!metric_events[i]) {
814 		if (expr__parse(&ratio, &pctx, metric_expr, runtime) == 0) {
815 			char *unit;
816 			char metric_bf[64];
817 
818 			if (metric_unit && metric_name) {
819 				if (perf_pmu__convert_scale(metric_unit,
820 					&unit, &scale) >= 0) {
821 					ratio *= scale;
822 				}
823 				if (strstr(metric_expr, "?"))
824 					scnprintf(metric_bf, sizeof(metric_bf),
825 					  "%s  %s_%d", unit, metric_name, runtime);
826 				else
827 					scnprintf(metric_bf, sizeof(metric_bf),
828 					  "%s  %s", unit, metric_name);
829 
830 				print_metric(config, ctxp, NULL, "%8.1f",
831 					     metric_bf, ratio);
832 			} else {
833 				print_metric(config, ctxp, NULL, "%8.2f",
834 					metric_name ?
835 					metric_name :
836 					out->force_header ?  name : "",
837 					ratio);
838 			}
839 		} else {
840 			print_metric(config, ctxp, NULL, NULL,
841 				     out->force_header ?
842 				     (metric_name ? metric_name : name) : "", 0);
843 		}
844 	} else {
845 		print_metric(config, ctxp, NULL, NULL,
846 			     out->force_header ?
847 			     (metric_name ? metric_name : name) : "", 0);
848 	}
849 
850 	expr__ctx_clear(&pctx);
851 }
852 
853 double test_generic_metric(struct metric_expr *mexp, int cpu, struct runtime_stat *st)
854 {
855 	struct expr_parse_ctx pctx;
856 	double ratio;
857 
858 	if (prepare_metric(mexp->metric_events, mexp->metric_refs, &pctx, cpu, st) < 0)
859 		return 0.;
860 
861 	if (expr__parse(&ratio, &pctx, mexp->metric_expr, 1))
862 		return 0.;
863 
864 	return ratio;
865 }
866 
867 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
868 				   struct evsel *evsel,
869 				   double avg, int cpu,
870 				   struct perf_stat_output_ctx *out,
871 				   struct rblist *metric_events,
872 				   struct runtime_stat *st)
873 {
874 	void *ctxp = out->ctx;
875 	print_metric_t print_metric = out->print_metric;
876 	double total, ratio = 0.0, total2;
877 	const char *color = NULL;
878 	int ctx = evsel_context(evsel);
879 	struct metric_event *me;
880 	int num = 1;
881 
882 	if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
883 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
884 
885 		if (total) {
886 			ratio = avg / total;
887 			print_metric(config, ctxp, NULL, "%7.2f ",
888 					"insn per cycle", ratio);
889 		} else {
890 			print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
891 		}
892 
893 		total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
894 					 ctx, cpu);
895 
896 		total = max(total, runtime_stat_avg(st,
897 						    STAT_STALLED_CYCLES_BACK,
898 						    ctx, cpu));
899 
900 		if (total && avg) {
901 			out->new_line(config, ctxp);
902 			ratio = total / avg;
903 			print_metric(config, ctxp, NULL, "%7.2f ",
904 					"stalled cycles per insn",
905 					ratio);
906 		}
907 	} else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
908 		if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
909 			print_branch_misses(config, cpu, evsel, avg, out, st);
910 		else
911 			print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
912 	} else if (
913 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
914 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
915 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
916 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
917 
918 		if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
919 			print_l1_dcache_misses(config, cpu, evsel, avg, out, st);
920 		else
921 			print_metric(config, ctxp, NULL, NULL, "of all L1-dcache hits", 0);
922 	} else if (
923 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
924 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
925 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
926 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
927 
928 		if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
929 			print_l1_icache_misses(config, cpu, evsel, avg, out, st);
930 		else
931 			print_metric(config, ctxp, NULL, NULL, "of all L1-icache hits", 0);
932 	} else if (
933 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
934 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
935 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
936 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
937 
938 		if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
939 			print_dtlb_cache_misses(config, cpu, evsel, avg, out, st);
940 		else
941 			print_metric(config, ctxp, NULL, NULL, "of all dTLB cache hits", 0);
942 	} else if (
943 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
944 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
945 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
946 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
947 
948 		if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
949 			print_itlb_cache_misses(config, cpu, evsel, avg, out, st);
950 		else
951 			print_metric(config, ctxp, NULL, NULL, "of all iTLB cache hits", 0);
952 	} else if (
953 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
954 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
955 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
956 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
957 
958 		if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
959 			print_ll_cache_misses(config, cpu, evsel, avg, out, st);
960 		else
961 			print_metric(config, ctxp, NULL, NULL, "of all LL-cache hits", 0);
962 	} else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
963 		total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
964 
965 		if (total)
966 			ratio = avg * 100 / total;
967 
968 		if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
969 			print_metric(config, ctxp, NULL, "%8.3f %%",
970 				     "of all cache refs", ratio);
971 		else
972 			print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
973 	} else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
974 		print_stalled_cycles_frontend(config, cpu, evsel, avg, out, st);
975 	} else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
976 		print_stalled_cycles_backend(config, cpu, evsel, avg, out, st);
977 	} else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
978 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
979 
980 		if (total) {
981 			ratio = avg / total;
982 			print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
983 		} else {
984 			print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
985 		}
986 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
987 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
988 
989 		if (total)
990 			print_metric(config, ctxp, NULL,
991 					"%7.2f%%", "transactional cycles",
992 					100.0 * (avg / total));
993 		else
994 			print_metric(config, ctxp, NULL, NULL, "transactional cycles",
995 				     0);
996 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
997 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
998 		total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
999 
1000 		if (total2 < avg)
1001 			total2 = avg;
1002 		if (total)
1003 			print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1004 				100.0 * ((total2-avg) / total));
1005 		else
1006 			print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1007 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1008 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
1009 					 ctx, cpu);
1010 
1011 		if (avg)
1012 			ratio = total / avg;
1013 
1014 		if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
1015 			print_metric(config, ctxp, NULL, "%8.0f",
1016 				     "cycles / transaction", ratio);
1017 		else
1018 			print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1019 				      0);
1020 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1021 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
1022 					 ctx, cpu);
1023 
1024 		if (avg)
1025 			ratio = total / avg;
1026 
1027 		print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1028 	} else if (evsel__is_clock(evsel)) {
1029 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1030 			print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1031 				     avg / (ratio * evsel->scale));
1032 		else
1033 			print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1034 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1035 		double fe_bound = td_fe_bound(ctx, cpu, st);
1036 
1037 		if (fe_bound > 0.2)
1038 			color = PERF_COLOR_RED;
1039 		print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1040 				fe_bound * 100.);
1041 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1042 		double retiring = td_retiring(ctx, cpu, st);
1043 
1044 		if (retiring > 0.7)
1045 			color = PERF_COLOR_GREEN;
1046 		print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1047 				retiring * 100.);
1048 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1049 		double bad_spec = td_bad_spec(ctx, cpu, st);
1050 
1051 		if (bad_spec > 0.1)
1052 			color = PERF_COLOR_RED;
1053 		print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1054 				bad_spec * 100.);
1055 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1056 		double be_bound = td_be_bound(ctx, cpu, st);
1057 		const char *name = "backend bound";
1058 		static int have_recovery_bubbles = -1;
1059 
1060 		/* In case the CPU does not support topdown-recovery-bubbles */
1061 		if (have_recovery_bubbles < 0)
1062 			have_recovery_bubbles = pmu_have_event("cpu",
1063 					"topdown-recovery-bubbles");
1064 		if (!have_recovery_bubbles)
1065 			name = "backend bound/bad spec";
1066 
1067 		if (be_bound > 0.2)
1068 			color = PERF_COLOR_RED;
1069 		if (td_total_slots(ctx, cpu, st) > 0)
1070 			print_metric(config, ctxp, color, "%8.1f%%", name,
1071 					be_bound * 100.);
1072 		else
1073 			print_metric(config, ctxp, NULL, NULL, name, 0);
1074 	} else if (evsel->metric_expr) {
1075 		generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1076 				evsel->name, evsel->metric_name, NULL, 1, cpu, out, st);
1077 	} else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
1078 		char unit = 'M';
1079 		char unit_buf[10];
1080 
1081 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
1082 
1083 		if (total)
1084 			ratio = 1000.0 * avg / total;
1085 		if (ratio < 0.001) {
1086 			ratio *= 1000;
1087 			unit = 'K';
1088 		}
1089 		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1090 		print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1091 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1092 		print_smi_cost(config, cpu, evsel, out, st);
1093 	} else {
1094 		num = 0;
1095 	}
1096 
1097 	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1098 		struct metric_expr *mexp;
1099 
1100 		list_for_each_entry (mexp, &me->head, nd) {
1101 			if (num++ > 0)
1102 				out->new_line(config, ctxp);
1103 			generic_metric(config, mexp->metric_expr, mexp->metric_events,
1104 					mexp->metric_refs, evsel->name, mexp->metric_name,
1105 					mexp->metric_unit, mexp->runtime, cpu, out, st);
1106 		}
1107 	}
1108 	if (num == 0)
1109 		print_metric(config, ctxp, NULL, NULL, NULL, 0);
1110 }
1111