xref: /openbmc/linux/tools/perf/util/stat-shadow.c (revision 6c33a6f4)
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 (perf_evsel__is_clock(counter))
220 		update_runtime_stat(st, STAT_NSECS, 0, cpu, count_ns);
221 	else if (perf_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 (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
245 		update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
246 				    ctx, cpu, count);
247 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
248 		update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
249 				    ctx, cpu, count);
250 	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
251 		update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count);
252 	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
253 		update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count);
254 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
255 		update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count);
256 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
257 		update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count);
258 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
259 		update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count);
260 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
261 		update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count);
262 	else if (perf_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 	const char **metric_names;
327 	int i;
328 	int num_metric_names;
329 
330 	evlist__for_each_entry(evsel_list, counter) {
331 		bool invalid = false;
332 
333 		leader = counter->leader;
334 		if (!counter->metric_expr)
335 			continue;
336 		metric_events = counter->metric_events;
337 		if (!metric_events) {
338 			if (expr__find_other(counter->metric_expr, counter->name,
339 						&metric_names, &num_metric_names) < 0)
340 				continue;
341 
342 			metric_events = calloc(sizeof(struct evsel *),
343 					       num_metric_names + 1);
344 			if (!metric_events)
345 				return;
346 			counter->metric_events = metric_events;
347 		}
348 
349 		for (i = 0; i < num_metric_names; i++) {
350 			found = false;
351 			if (leader) {
352 				/* Search in group */
353 				for_each_group_member (oc, leader) {
354 					if (!strcasecmp(oc->name, metric_names[i]) &&
355 						!oc->collect_stat) {
356 						found = true;
357 						break;
358 					}
359 				}
360 			}
361 			if (!found) {
362 				/* Search ignoring groups */
363 				oc = perf_stat__find_event(evsel_list, metric_names[i]);
364 			}
365 			if (!oc) {
366 				/* Deduping one is good enough to handle duplicated PMUs. */
367 				static char *printed;
368 
369 				/*
370 				 * Adding events automatically would be difficult, because
371 				 * it would risk creating groups that are not schedulable.
372 				 * perf stat doesn't understand all the scheduling constraints
373 				 * of events. So we ask the user instead to add the missing
374 				 * events.
375 				 */
376 				if (!printed || strcasecmp(printed, metric_names[i])) {
377 					fprintf(stderr,
378 						"Add %s event to groups to get metric expression for %s\n",
379 						metric_names[i],
380 						counter->name);
381 					printed = strdup(metric_names[i]);
382 				}
383 				invalid = true;
384 				continue;
385 			}
386 			metric_events[i] = oc;
387 			oc->collect_stat = true;
388 		}
389 		metric_events[i] = NULL;
390 		free(metric_names);
391 		if (invalid) {
392 			free(metric_events);
393 			counter->metric_events = NULL;
394 			counter->metric_expr = NULL;
395 		}
396 	}
397 }
398 
399 static double runtime_stat_avg(struct runtime_stat *st,
400 			       enum stat_type type, int ctx, int cpu)
401 {
402 	struct saved_value *v;
403 
404 	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
405 	if (!v)
406 		return 0.0;
407 
408 	return avg_stats(&v->stats);
409 }
410 
411 static double runtime_stat_n(struct runtime_stat *st,
412 			     enum stat_type type, int ctx, int cpu)
413 {
414 	struct saved_value *v;
415 
416 	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
417 	if (!v)
418 		return 0.0;
419 
420 	return v->stats.n;
421 }
422 
423 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
424 					  int cpu,
425 					  struct evsel *evsel, double avg,
426 					  struct perf_stat_output_ctx *out,
427 					  struct runtime_stat *st)
428 {
429 	double total, ratio = 0.0;
430 	const char *color;
431 	int ctx = evsel_context(evsel);
432 
433 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
434 
435 	if (total)
436 		ratio = avg / total * 100.0;
437 
438 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
439 
440 	if (ratio)
441 		out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
442 				  ratio);
443 	else
444 		out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
445 }
446 
447 static void print_stalled_cycles_backend(struct perf_stat_config *config,
448 					 int cpu,
449 					 struct evsel *evsel, double avg,
450 					 struct perf_stat_output_ctx *out,
451 					 struct runtime_stat *st)
452 {
453 	double total, ratio = 0.0;
454 	const char *color;
455 	int ctx = evsel_context(evsel);
456 
457 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
458 
459 	if (total)
460 		ratio = avg / total * 100.0;
461 
462 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
463 
464 	out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
465 }
466 
467 static void print_branch_misses(struct perf_stat_config *config,
468 				int cpu,
469 				struct evsel *evsel,
470 				double avg,
471 				struct perf_stat_output_ctx *out,
472 				struct runtime_stat *st)
473 {
474 	double total, ratio = 0.0;
475 	const char *color;
476 	int ctx = evsel_context(evsel);
477 
478 	total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
479 
480 	if (total)
481 		ratio = avg / total * 100.0;
482 
483 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
484 
485 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
486 }
487 
488 static void print_l1_dcache_misses(struct perf_stat_config *config,
489 				   int cpu,
490 				   struct evsel *evsel,
491 				   double avg,
492 				   struct perf_stat_output_ctx *out,
493 				   struct runtime_stat *st)
494 
495 {
496 	double total, ratio = 0.0;
497 	const char *color;
498 	int ctx = evsel_context(evsel);
499 
500 	total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
501 
502 	if (total)
503 		ratio = avg / total * 100.0;
504 
505 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
506 
507 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
508 }
509 
510 static void print_l1_icache_misses(struct perf_stat_config *config,
511 				   int cpu,
512 				   struct evsel *evsel,
513 				   double avg,
514 				   struct perf_stat_output_ctx *out,
515 				   struct runtime_stat *st)
516 
517 {
518 	double total, ratio = 0.0;
519 	const char *color;
520 	int ctx = evsel_context(evsel);
521 
522 	total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
523 
524 	if (total)
525 		ratio = avg / total * 100.0;
526 
527 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
528 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
529 }
530 
531 static void print_dtlb_cache_misses(struct perf_stat_config *config,
532 				    int cpu,
533 				    struct evsel *evsel,
534 				    double avg,
535 				    struct perf_stat_output_ctx *out,
536 				    struct runtime_stat *st)
537 {
538 	double total, ratio = 0.0;
539 	const char *color;
540 	int ctx = evsel_context(evsel);
541 
542 	total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
543 
544 	if (total)
545 		ratio = avg / total * 100.0;
546 
547 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
548 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
549 }
550 
551 static void print_itlb_cache_misses(struct perf_stat_config *config,
552 				    int cpu,
553 				    struct evsel *evsel,
554 				    double avg,
555 				    struct perf_stat_output_ctx *out,
556 				    struct runtime_stat *st)
557 {
558 	double total, ratio = 0.0;
559 	const char *color;
560 	int ctx = evsel_context(evsel);
561 
562 	total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
563 
564 	if (total)
565 		ratio = avg / total * 100.0;
566 
567 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
568 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
569 }
570 
571 static void print_ll_cache_misses(struct perf_stat_config *config,
572 				  int cpu,
573 				  struct evsel *evsel,
574 				  double avg,
575 				  struct perf_stat_output_ctx *out,
576 				  struct runtime_stat *st)
577 {
578 	double total, ratio = 0.0;
579 	const char *color;
580 	int ctx = evsel_context(evsel);
581 
582 	total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
583 
584 	if (total)
585 		ratio = avg / total * 100.0;
586 
587 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
588 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
589 }
590 
591 /*
592  * High level "TopDown" CPU core pipe line bottleneck break down.
593  *
594  * Basic concept following
595  * Yasin, A Top Down Method for Performance analysis and Counter architecture
596  * ISPASS14
597  *
598  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
599  *
600  * Frontend -> Backend -> Retiring
601  * BadSpeculation in addition means out of order execution that is thrown away
602  * (for example branch mispredictions)
603  * Frontend is instruction decoding.
604  * Backend is execution, like computation and accessing data in memory
605  * Retiring is good execution that is not directly bottlenecked
606  *
607  * The formulas are computed in slots.
608  * A slot is an entry in the pipeline each for the pipeline width
609  * (for example a 4-wide pipeline has 4 slots for each cycle)
610  *
611  * Formulas:
612  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
613  *			TotalSlots
614  * Retiring = SlotsRetired / TotalSlots
615  * FrontendBound = FetchBubbles / TotalSlots
616  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
617  *
618  * The kernel provides the mapping to the low level CPU events and any scaling
619  * needed for the CPU pipeline width, for example:
620  *
621  * TotalSlots = Cycles * 4
622  *
623  * The scaling factor is communicated in the sysfs unit.
624  *
625  * In some cases the CPU may not be able to measure all the formulas due to
626  * missing events. In this case multiple formulas are combined, as possible.
627  *
628  * Full TopDown supports more levels to sub-divide each area: for example
629  * BackendBound into computing bound and memory bound. For now we only
630  * support Level 1 TopDown.
631  */
632 
633 static double sanitize_val(double x)
634 {
635 	if (x < 0 && x >= -0.02)
636 		return 0.0;
637 	return x;
638 }
639 
640 static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
641 {
642 	return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
643 }
644 
645 static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
646 {
647 	double bad_spec = 0;
648 	double total_slots;
649 	double total;
650 
651 	total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
652 		runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
653 		runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
654 
655 	total_slots = td_total_slots(ctx, cpu, st);
656 	if (total_slots)
657 		bad_spec = total / total_slots;
658 	return sanitize_val(bad_spec);
659 }
660 
661 static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
662 {
663 	double retiring = 0;
664 	double total_slots = td_total_slots(ctx, cpu, st);
665 	double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
666 					    ctx, cpu);
667 
668 	if (total_slots)
669 		retiring = ret_slots / total_slots;
670 	return retiring;
671 }
672 
673 static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
674 {
675 	double fe_bound = 0;
676 	double total_slots = td_total_slots(ctx, cpu, st);
677 	double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
678 					    ctx, cpu);
679 
680 	if (total_slots)
681 		fe_bound = fetch_bub / total_slots;
682 	return fe_bound;
683 }
684 
685 static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
686 {
687 	double sum = (td_fe_bound(ctx, cpu, st) +
688 		      td_bad_spec(ctx, cpu, st) +
689 		      td_retiring(ctx, cpu, st));
690 	if (sum == 0)
691 		return 0;
692 	return sanitize_val(1.0 - sum);
693 }
694 
695 static void print_smi_cost(struct perf_stat_config *config,
696 			   int cpu, struct evsel *evsel,
697 			   struct perf_stat_output_ctx *out,
698 			   struct runtime_stat *st)
699 {
700 	double smi_num, aperf, cycles, cost = 0.0;
701 	int ctx = evsel_context(evsel);
702 	const char *color = NULL;
703 
704 	smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
705 	aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
706 	cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
707 
708 	if ((cycles == 0) || (aperf == 0))
709 		return;
710 
711 	if (smi_num)
712 		cost = (aperf - cycles) / aperf * 100.00;
713 
714 	if (cost > 10)
715 		color = PERF_COLOR_RED;
716 	out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
717 	out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
718 }
719 
720 static void generic_metric(struct perf_stat_config *config,
721 			   const char *metric_expr,
722 			   struct evsel **metric_events,
723 			   char *name,
724 			   const char *metric_name,
725 			   const char *metric_unit,
726 			   double avg,
727 			   int cpu,
728 			   struct perf_stat_output_ctx *out,
729 			   struct runtime_stat *st)
730 {
731 	print_metric_t print_metric = out->print_metric;
732 	struct parse_ctx pctx;
733 	double ratio, scale;
734 	int i;
735 	void *ctxp = out->ctx;
736 	char *n, *pn;
737 
738 	expr__ctx_init(&pctx);
739 	/* Must be first id entry */
740 	expr__add_id(&pctx, name, avg);
741 	for (i = 0; metric_events[i]; i++) {
742 		struct saved_value *v;
743 		struct stats *stats;
744 		u64 metric_total = 0;
745 
746 		if (!strcmp(metric_events[i]->name, "duration_time")) {
747 			stats = &walltime_nsecs_stats;
748 			scale = 1e-9;
749 		} else {
750 			v = saved_value_lookup(metric_events[i], cpu, false,
751 					       STAT_NONE, 0, st);
752 			if (!v)
753 				break;
754 			stats = &v->stats;
755 			scale = 1.0;
756 
757 			if (v->metric_other)
758 				metric_total = v->metric_total;
759 		}
760 
761 		n = strdup(metric_events[i]->name);
762 		if (!n)
763 			return;
764 		/*
765 		 * This display code with --no-merge adds [cpu] postfixes.
766 		 * These are not supported by the parser. Remove everything
767 		 * after the space.
768 		 */
769 		pn = strchr(n, ' ');
770 		if (pn)
771 			*pn = 0;
772 
773 		if (metric_total)
774 			expr__add_id(&pctx, n, metric_total);
775 		else
776 			expr__add_id(&pctx, n, avg_stats(stats)*scale);
777 	}
778 
779 	if (!metric_events[i]) {
780 		const char *p = metric_expr;
781 
782 		if (expr__parse(&ratio, &pctx, &p) == 0) {
783 			char *unit;
784 			char metric_bf[64];
785 
786 			if (metric_unit && metric_name) {
787 				if (perf_pmu__convert_scale(metric_unit,
788 					&unit, &scale) >= 0) {
789 					ratio *= scale;
790 				}
791 
792 				scnprintf(metric_bf, sizeof(metric_bf),
793 					  "%s  %s", unit, metric_name);
794 				print_metric(config, ctxp, NULL, "%8.1f",
795 					     metric_bf, ratio);
796 			} else {
797 				print_metric(config, ctxp, NULL, "%8.1f",
798 					metric_name ?
799 					metric_name :
800 					out->force_header ?  name : "",
801 					ratio);
802 			}
803 		} else {
804 			print_metric(config, ctxp, NULL, NULL,
805 				     out->force_header ?
806 				     (metric_name ? metric_name : name) : "", 0);
807 		}
808 	} else
809 		print_metric(config, ctxp, NULL, NULL, "", 0);
810 
811 	for (i = 1; i < pctx.num_ids; i++)
812 		zfree(&pctx.ids[i].name);
813 }
814 
815 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
816 				   struct evsel *evsel,
817 				   double avg, int cpu,
818 				   struct perf_stat_output_ctx *out,
819 				   struct rblist *metric_events,
820 				   struct runtime_stat *st)
821 {
822 	void *ctxp = out->ctx;
823 	print_metric_t print_metric = out->print_metric;
824 	double total, ratio = 0.0, total2;
825 	const char *color = NULL;
826 	int ctx = evsel_context(evsel);
827 	struct metric_event *me;
828 	int num = 1;
829 
830 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
831 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
832 
833 		if (total) {
834 			ratio = avg / total;
835 			print_metric(config, ctxp, NULL, "%7.2f ",
836 					"insn per cycle", ratio);
837 		} else {
838 			print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
839 		}
840 
841 		total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
842 					 ctx, cpu);
843 
844 		total = max(total, runtime_stat_avg(st,
845 						    STAT_STALLED_CYCLES_BACK,
846 						    ctx, cpu));
847 
848 		if (total && avg) {
849 			out->new_line(config, ctxp);
850 			ratio = total / avg;
851 			print_metric(config, ctxp, NULL, "%7.2f ",
852 					"stalled cycles per insn",
853 					ratio);
854 		}
855 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
856 		if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
857 			print_branch_misses(config, cpu, evsel, avg, out, st);
858 		else
859 			print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
860 	} else if (
861 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
862 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
863 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
864 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
865 
866 		if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
867 			print_l1_dcache_misses(config, cpu, evsel, avg, out, st);
868 		else
869 			print_metric(config, ctxp, NULL, NULL, "of all L1-dcache hits", 0);
870 	} else if (
871 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
872 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
873 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
874 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
875 
876 		if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
877 			print_l1_icache_misses(config, cpu, evsel, avg, out, st);
878 		else
879 			print_metric(config, ctxp, NULL, NULL, "of all L1-icache hits", 0);
880 	} else if (
881 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
882 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
883 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
884 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
885 
886 		if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
887 			print_dtlb_cache_misses(config, cpu, evsel, avg, out, st);
888 		else
889 			print_metric(config, ctxp, NULL, NULL, "of all dTLB cache hits", 0);
890 	} else if (
891 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
892 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
893 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
894 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
895 
896 		if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
897 			print_itlb_cache_misses(config, cpu, evsel, avg, out, st);
898 		else
899 			print_metric(config, ctxp, NULL, NULL, "of all iTLB cache hits", 0);
900 	} else if (
901 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
902 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
903 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
904 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
905 
906 		if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
907 			print_ll_cache_misses(config, cpu, evsel, avg, out, st);
908 		else
909 			print_metric(config, ctxp, NULL, NULL, "of all LL-cache hits", 0);
910 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
911 		total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
912 
913 		if (total)
914 			ratio = avg * 100 / total;
915 
916 		if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
917 			print_metric(config, ctxp, NULL, "%8.3f %%",
918 				     "of all cache refs", ratio);
919 		else
920 			print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
921 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
922 		print_stalled_cycles_frontend(config, cpu, evsel, avg, out, st);
923 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
924 		print_stalled_cycles_backend(config, cpu, evsel, avg, out, st);
925 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
926 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
927 
928 		if (total) {
929 			ratio = avg / total;
930 			print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
931 		} else {
932 			print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
933 		}
934 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
935 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
936 
937 		if (total)
938 			print_metric(config, ctxp, NULL,
939 					"%7.2f%%", "transactional cycles",
940 					100.0 * (avg / total));
941 		else
942 			print_metric(config, ctxp, NULL, NULL, "transactional cycles",
943 				     0);
944 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
945 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
946 		total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
947 
948 		if (total2 < avg)
949 			total2 = avg;
950 		if (total)
951 			print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
952 				100.0 * ((total2-avg) / total));
953 		else
954 			print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
955 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
956 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
957 					 ctx, cpu);
958 
959 		if (avg)
960 			ratio = total / avg;
961 
962 		if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
963 			print_metric(config, ctxp, NULL, "%8.0f",
964 				     "cycles / transaction", ratio);
965 		else
966 			print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
967 				      0);
968 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
969 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
970 					 ctx, cpu);
971 
972 		if (avg)
973 			ratio = total / avg;
974 
975 		print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
976 	} else if (perf_evsel__is_clock(evsel)) {
977 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
978 			print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
979 				     avg / (ratio * evsel->scale));
980 		else
981 			print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
982 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
983 		double fe_bound = td_fe_bound(ctx, cpu, st);
984 
985 		if (fe_bound > 0.2)
986 			color = PERF_COLOR_RED;
987 		print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
988 				fe_bound * 100.);
989 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
990 		double retiring = td_retiring(ctx, cpu, st);
991 
992 		if (retiring > 0.7)
993 			color = PERF_COLOR_GREEN;
994 		print_metric(config, ctxp, color, "%8.1f%%", "retiring",
995 				retiring * 100.);
996 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
997 		double bad_spec = td_bad_spec(ctx, cpu, st);
998 
999 		if (bad_spec > 0.1)
1000 			color = PERF_COLOR_RED;
1001 		print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1002 				bad_spec * 100.);
1003 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1004 		double be_bound = td_be_bound(ctx, cpu, st);
1005 		const char *name = "backend bound";
1006 		static int have_recovery_bubbles = -1;
1007 
1008 		/* In case the CPU does not support topdown-recovery-bubbles */
1009 		if (have_recovery_bubbles < 0)
1010 			have_recovery_bubbles = pmu_have_event("cpu",
1011 					"topdown-recovery-bubbles");
1012 		if (!have_recovery_bubbles)
1013 			name = "backend bound/bad spec";
1014 
1015 		if (be_bound > 0.2)
1016 			color = PERF_COLOR_RED;
1017 		if (td_total_slots(ctx, cpu, st) > 0)
1018 			print_metric(config, ctxp, color, "%8.1f%%", name,
1019 					be_bound * 100.);
1020 		else
1021 			print_metric(config, ctxp, NULL, NULL, name, 0);
1022 	} else if (evsel->metric_expr) {
1023 		generic_metric(config, evsel->metric_expr, evsel->metric_events, evsel->name,
1024 				evsel->metric_name, NULL, avg, cpu, out, st);
1025 	} else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
1026 		char unit = 'M';
1027 		char unit_buf[10];
1028 
1029 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
1030 
1031 		if (total)
1032 			ratio = 1000.0 * avg / total;
1033 		if (ratio < 0.001) {
1034 			ratio *= 1000;
1035 			unit = 'K';
1036 		}
1037 		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1038 		print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1039 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1040 		print_smi_cost(config, cpu, evsel, out, st);
1041 	} else {
1042 		num = 0;
1043 	}
1044 
1045 	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1046 		struct metric_expr *mexp;
1047 
1048 		list_for_each_entry (mexp, &me->head, nd) {
1049 			if (num++ > 0)
1050 				out->new_line(config, ctxp);
1051 			generic_metric(config, mexp->metric_expr, mexp->metric_events,
1052 					evsel->name, mexp->metric_name,
1053 					mexp->metric_unit, avg, cpu, out, st);
1054 		}
1055 	}
1056 	if (num == 0)
1057 		print_metric(config, ctxp, NULL, NULL, NULL, 0);
1058 }
1059