xref: /openbmc/linux/tools/perf/util/stat-shadow.c (revision 2c6467d2)
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 
12 /*
13  * AGGR_GLOBAL: Use CPU 0
14  * AGGR_SOCKET: Use first CPU of socket
15  * AGGR_CORE: Use first CPU of core
16  * AGGR_NONE: Use matching CPU
17  * AGGR_THREAD: Not supported?
18  */
19 static bool have_frontend_stalled;
20 
21 struct runtime_stat rt_stat;
22 struct stats walltime_nsecs_stats;
23 
24 struct saved_value {
25 	struct rb_node rb_node;
26 	struct perf_evsel *evsel;
27 	enum stat_type type;
28 	int ctx;
29 	int cpu;
30 	struct runtime_stat *stat;
31 	struct stats stats;
32 };
33 
34 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
35 {
36 	struct saved_value *a = container_of(rb_node,
37 					     struct saved_value,
38 					     rb_node);
39 	const struct saved_value *b = entry;
40 
41 	if (a->cpu != b->cpu)
42 		return a->cpu - b->cpu;
43 
44 	/*
45 	 * Previously the rbtree was used to link generic metrics.
46 	 * The keys were evsel/cpu. Now the rbtree is extended to support
47 	 * per-thread shadow stats. For shadow stats case, the keys
48 	 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
49 	 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
50 	 */
51 	if (a->type != b->type)
52 		return a->type - b->type;
53 
54 	if (a->ctx != b->ctx)
55 		return a->ctx - b->ctx;
56 
57 	if (a->evsel == NULL && b->evsel == NULL) {
58 		if (a->stat == b->stat)
59 			return 0;
60 
61 		if ((char *)a->stat < (char *)b->stat)
62 			return -1;
63 
64 		return 1;
65 	}
66 
67 	if (a->evsel == b->evsel)
68 		return 0;
69 	if ((char *)a->evsel < (char *)b->evsel)
70 		return -1;
71 	return +1;
72 }
73 
74 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
75 				     const void *entry)
76 {
77 	struct saved_value *nd = malloc(sizeof(struct saved_value));
78 
79 	if (!nd)
80 		return NULL;
81 	memcpy(nd, entry, sizeof(struct saved_value));
82 	return &nd->rb_node;
83 }
84 
85 static void saved_value_delete(struct rblist *rblist __maybe_unused,
86 			       struct rb_node *rb_node)
87 {
88 	struct saved_value *v;
89 
90 	BUG_ON(!rb_node);
91 	v = container_of(rb_node, struct saved_value, rb_node);
92 	free(v);
93 }
94 
95 static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
96 					      int cpu,
97 					      bool create,
98 					      enum stat_type type,
99 					      int ctx,
100 					      struct runtime_stat *st)
101 {
102 	struct rblist *rblist;
103 	struct rb_node *nd;
104 	struct saved_value dm = {
105 		.cpu = cpu,
106 		.evsel = evsel,
107 		.type = type,
108 		.ctx = ctx,
109 		.stat = st,
110 	};
111 
112 	rblist = &st->value_list;
113 
114 	nd = rblist__find(rblist, &dm);
115 	if (nd)
116 		return container_of(nd, struct saved_value, rb_node);
117 	if (create) {
118 		rblist__add_node(rblist, &dm);
119 		nd = rblist__find(rblist, &dm);
120 		if (nd)
121 			return container_of(nd, struct saved_value, rb_node);
122 	}
123 	return NULL;
124 }
125 
126 void runtime_stat__init(struct runtime_stat *st)
127 {
128 	struct rblist *rblist = &st->value_list;
129 
130 	rblist__init(rblist);
131 	rblist->node_cmp = saved_value_cmp;
132 	rblist->node_new = saved_value_new;
133 	rblist->node_delete = saved_value_delete;
134 }
135 
136 void runtime_stat__exit(struct runtime_stat *st)
137 {
138 	rblist__exit(&st->value_list);
139 }
140 
141 void perf_stat__init_shadow_stats(void)
142 {
143 	have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
144 	runtime_stat__init(&rt_stat);
145 }
146 
147 static int evsel_context(struct perf_evsel *evsel)
148 {
149 	int ctx = 0;
150 
151 	if (evsel->attr.exclude_kernel)
152 		ctx |= CTX_BIT_KERNEL;
153 	if (evsel->attr.exclude_user)
154 		ctx |= CTX_BIT_USER;
155 	if (evsel->attr.exclude_hv)
156 		ctx |= CTX_BIT_HV;
157 	if (evsel->attr.exclude_host)
158 		ctx |= CTX_BIT_HOST;
159 	if (evsel->attr.exclude_idle)
160 		ctx |= CTX_BIT_IDLE;
161 
162 	return ctx;
163 }
164 
165 static void reset_stat(struct runtime_stat *st)
166 {
167 	struct rblist *rblist;
168 	struct rb_node *pos, *next;
169 
170 	rblist = &st->value_list;
171 	next = rb_first(&rblist->entries);
172 	while (next) {
173 		pos = next;
174 		next = rb_next(pos);
175 		memset(&container_of(pos, struct saved_value, rb_node)->stats,
176 		       0,
177 		       sizeof(struct stats));
178 	}
179 }
180 
181 void perf_stat__reset_shadow_stats(void)
182 {
183 	reset_stat(&rt_stat);
184 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
185 }
186 
187 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
188 {
189 	reset_stat(st);
190 }
191 
192 static void update_runtime_stat(struct runtime_stat *st,
193 				enum stat_type type,
194 				int ctx, int cpu, u64 count)
195 {
196 	struct saved_value *v = saved_value_lookup(NULL, cpu, true,
197 						   type, ctx, st);
198 
199 	if (v)
200 		update_stats(&v->stats, count);
201 }
202 
203 /*
204  * Update various tracking values we maintain to print
205  * more semantic information such as miss/hit ratios,
206  * instruction rates, etc:
207  */
208 void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 count,
209 				    int cpu, struct runtime_stat *st)
210 {
211 	int ctx = evsel_context(counter);
212 	u64 count_ns = count;
213 
214 	count *= counter->scale;
215 
216 	if (perf_evsel__is_clock(counter))
217 		update_runtime_stat(st, STAT_NSECS, 0, cpu, count_ns);
218 	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
219 		update_runtime_stat(st, STAT_CYCLES, ctx, cpu, count);
220 	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
221 		update_runtime_stat(st, STAT_CYCLES_IN_TX, ctx, cpu, count);
222 	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
223 		update_runtime_stat(st, STAT_TRANSACTION, ctx, cpu, count);
224 	else if (perf_stat_evsel__is(counter, ELISION_START))
225 		update_runtime_stat(st, STAT_ELISION, ctx, cpu, count);
226 	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
227 		update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
228 				    ctx, cpu, count);
229 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
230 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
231 				    ctx, cpu, count);
232 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
233 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
234 				    ctx, cpu, count);
235 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
236 		update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
237 				    ctx, cpu, count);
238 	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
239 		update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
240 				    ctx, cpu, count);
241 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
242 		update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
243 				    ctx, cpu, count);
244 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
245 		update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
246 				    ctx, cpu, count);
247 	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
248 		update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count);
249 	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
250 		update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count);
251 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
252 		update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count);
253 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
254 		update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count);
255 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
256 		update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count);
257 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
258 		update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count);
259 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
260 		update_runtime_stat(st, STAT_ITLB_CACHE, ctx, cpu, count);
261 	else if (perf_stat_evsel__is(counter, SMI_NUM))
262 		update_runtime_stat(st, STAT_SMI_NUM, ctx, cpu, count);
263 	else if (perf_stat_evsel__is(counter, APERF))
264 		update_runtime_stat(st, STAT_APERF, ctx, cpu, count);
265 
266 	if (counter->collect_stat) {
267 		struct saved_value *v = saved_value_lookup(counter, cpu, true,
268 							   STAT_NONE, 0, st);
269 		update_stats(&v->stats, count);
270 	}
271 }
272 
273 /* used for get_ratio_color() */
274 enum grc_type {
275 	GRC_STALLED_CYCLES_FE,
276 	GRC_STALLED_CYCLES_BE,
277 	GRC_CACHE_MISSES,
278 	GRC_MAX_NR
279 };
280 
281 static const char *get_ratio_color(enum grc_type type, double ratio)
282 {
283 	static const double grc_table[GRC_MAX_NR][3] = {
284 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
285 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
286 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
287 	};
288 	const char *color = PERF_COLOR_NORMAL;
289 
290 	if (ratio > grc_table[type][0])
291 		color = PERF_COLOR_RED;
292 	else if (ratio > grc_table[type][1])
293 		color = PERF_COLOR_MAGENTA;
294 	else if (ratio > grc_table[type][2])
295 		color = PERF_COLOR_YELLOW;
296 
297 	return color;
298 }
299 
300 static struct perf_evsel *perf_stat__find_event(struct perf_evlist *evsel_list,
301 						const char *name)
302 {
303 	struct perf_evsel *c2;
304 
305 	evlist__for_each_entry (evsel_list, c2) {
306 		if (!strcasecmp(c2->name, name))
307 			return c2;
308 	}
309 	return NULL;
310 }
311 
312 /* Mark MetricExpr target events and link events using them to them. */
313 void perf_stat__collect_metric_expr(struct perf_evlist *evsel_list)
314 {
315 	struct perf_evsel *counter, *leader, **metric_events, *oc;
316 	bool found;
317 	const char **metric_names;
318 	int i;
319 	int num_metric_names;
320 
321 	evlist__for_each_entry(evsel_list, counter) {
322 		bool invalid = false;
323 
324 		leader = counter->leader;
325 		if (!counter->metric_expr)
326 			continue;
327 		metric_events = counter->metric_events;
328 		if (!metric_events) {
329 			if (expr__find_other(counter->metric_expr, counter->name,
330 						&metric_names, &num_metric_names) < 0)
331 				continue;
332 
333 			metric_events = calloc(sizeof(struct perf_evsel *),
334 					       num_metric_names + 1);
335 			if (!metric_events)
336 				return;
337 			counter->metric_events = metric_events;
338 		}
339 
340 		for (i = 0; i < num_metric_names; i++) {
341 			found = false;
342 			if (leader) {
343 				/* Search in group */
344 				for_each_group_member (oc, leader) {
345 					if (!strcasecmp(oc->name, metric_names[i])) {
346 						found = true;
347 						break;
348 					}
349 				}
350 			}
351 			if (!found) {
352 				/* Search ignoring groups */
353 				oc = perf_stat__find_event(evsel_list, metric_names[i]);
354 			}
355 			if (!oc) {
356 				/* Deduping one is good enough to handle duplicated PMUs. */
357 				static char *printed;
358 
359 				/*
360 				 * Adding events automatically would be difficult, because
361 				 * it would risk creating groups that are not schedulable.
362 				 * perf stat doesn't understand all the scheduling constraints
363 				 * of events. So we ask the user instead to add the missing
364 				 * events.
365 				 */
366 				if (!printed || strcasecmp(printed, metric_names[i])) {
367 					fprintf(stderr,
368 						"Add %s event to groups to get metric expression for %s\n",
369 						metric_names[i],
370 						counter->name);
371 					printed = strdup(metric_names[i]);
372 				}
373 				invalid = true;
374 				continue;
375 			}
376 			metric_events[i] = oc;
377 			oc->collect_stat = true;
378 		}
379 		metric_events[i] = NULL;
380 		free(metric_names);
381 		if (invalid) {
382 			free(metric_events);
383 			counter->metric_events = NULL;
384 			counter->metric_expr = NULL;
385 		}
386 	}
387 }
388 
389 static double runtime_stat_avg(struct runtime_stat *st,
390 			       enum stat_type type, int ctx, int cpu)
391 {
392 	struct saved_value *v;
393 
394 	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
395 	if (!v)
396 		return 0.0;
397 
398 	return avg_stats(&v->stats);
399 }
400 
401 static double runtime_stat_n(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 v->stats.n;
411 }
412 
413 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
414 					  int cpu,
415 					  struct perf_evsel *evsel, double avg,
416 					  struct perf_stat_output_ctx *out,
417 					  struct runtime_stat *st)
418 {
419 	double total, ratio = 0.0;
420 	const char *color;
421 	int ctx = evsel_context(evsel);
422 
423 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
424 
425 	if (total)
426 		ratio = avg / total * 100.0;
427 
428 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
429 
430 	if (ratio)
431 		out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
432 				  ratio);
433 	else
434 		out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
435 }
436 
437 static void print_stalled_cycles_backend(struct perf_stat_config *config,
438 					 int cpu,
439 					 struct perf_evsel *evsel, double avg,
440 					 struct perf_stat_output_ctx *out,
441 					 struct runtime_stat *st)
442 {
443 	double total, ratio = 0.0;
444 	const char *color;
445 	int ctx = evsel_context(evsel);
446 
447 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
448 
449 	if (total)
450 		ratio = avg / total * 100.0;
451 
452 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
453 
454 	out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
455 }
456 
457 static void print_branch_misses(struct perf_stat_config *config,
458 				int cpu,
459 				struct perf_evsel *evsel,
460 				double avg,
461 				struct perf_stat_output_ctx *out,
462 				struct runtime_stat *st)
463 {
464 	double total, ratio = 0.0;
465 	const char *color;
466 	int ctx = evsel_context(evsel);
467 
468 	total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
469 
470 	if (total)
471 		ratio = avg / total * 100.0;
472 
473 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
474 
475 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
476 }
477 
478 static void print_l1_dcache_misses(struct perf_stat_config *config,
479 				   int cpu,
480 				   struct perf_evsel *evsel,
481 				   double avg,
482 				   struct perf_stat_output_ctx *out,
483 				   struct runtime_stat *st)
484 
485 {
486 	double total, ratio = 0.0;
487 	const char *color;
488 	int ctx = evsel_context(evsel);
489 
490 	total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
491 
492 	if (total)
493 		ratio = avg / total * 100.0;
494 
495 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
496 
497 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
498 }
499 
500 static void print_l1_icache_misses(struct perf_stat_config *config,
501 				   int cpu,
502 				   struct perf_evsel *evsel,
503 				   double avg,
504 				   struct perf_stat_output_ctx *out,
505 				   struct runtime_stat *st)
506 
507 {
508 	double total, ratio = 0.0;
509 	const char *color;
510 	int ctx = evsel_context(evsel);
511 
512 	total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
513 
514 	if (total)
515 		ratio = avg / total * 100.0;
516 
517 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
518 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
519 }
520 
521 static void print_dtlb_cache_misses(struct perf_stat_config *config,
522 				    int cpu,
523 				    struct perf_evsel *evsel,
524 				    double avg,
525 				    struct perf_stat_output_ctx *out,
526 				    struct runtime_stat *st)
527 {
528 	double total, ratio = 0.0;
529 	const char *color;
530 	int ctx = evsel_context(evsel);
531 
532 	total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
533 
534 	if (total)
535 		ratio = avg / total * 100.0;
536 
537 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
538 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
539 }
540 
541 static void print_itlb_cache_misses(struct perf_stat_config *config,
542 				    int cpu,
543 				    struct perf_evsel *evsel,
544 				    double avg,
545 				    struct perf_stat_output_ctx *out,
546 				    struct runtime_stat *st)
547 {
548 	double total, ratio = 0.0;
549 	const char *color;
550 	int ctx = evsel_context(evsel);
551 
552 	total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
553 
554 	if (total)
555 		ratio = avg / total * 100.0;
556 
557 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
558 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
559 }
560 
561 static void print_ll_cache_misses(struct perf_stat_config *config,
562 				  int cpu,
563 				  struct perf_evsel *evsel,
564 				  double avg,
565 				  struct perf_stat_output_ctx *out,
566 				  struct runtime_stat *st)
567 {
568 	double total, ratio = 0.0;
569 	const char *color;
570 	int ctx = evsel_context(evsel);
571 
572 	total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
573 
574 	if (total)
575 		ratio = avg / total * 100.0;
576 
577 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
578 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
579 }
580 
581 /*
582  * High level "TopDown" CPU core pipe line bottleneck break down.
583  *
584  * Basic concept following
585  * Yasin, A Top Down Method for Performance analysis and Counter architecture
586  * ISPASS14
587  *
588  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
589  *
590  * Frontend -> Backend -> Retiring
591  * BadSpeculation in addition means out of order execution that is thrown away
592  * (for example branch mispredictions)
593  * Frontend is instruction decoding.
594  * Backend is execution, like computation and accessing data in memory
595  * Retiring is good execution that is not directly bottlenecked
596  *
597  * The formulas are computed in slots.
598  * A slot is an entry in the pipeline each for the pipeline width
599  * (for example a 4-wide pipeline has 4 slots for each cycle)
600  *
601  * Formulas:
602  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
603  *			TotalSlots
604  * Retiring = SlotsRetired / TotalSlots
605  * FrontendBound = FetchBubbles / TotalSlots
606  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
607  *
608  * The kernel provides the mapping to the low level CPU events and any scaling
609  * needed for the CPU pipeline width, for example:
610  *
611  * TotalSlots = Cycles * 4
612  *
613  * The scaling factor is communicated in the sysfs unit.
614  *
615  * In some cases the CPU may not be able to measure all the formulas due to
616  * missing events. In this case multiple formulas are combined, as possible.
617  *
618  * Full TopDown supports more levels to sub-divide each area: for example
619  * BackendBound into computing bound and memory bound. For now we only
620  * support Level 1 TopDown.
621  */
622 
623 static double sanitize_val(double x)
624 {
625 	if (x < 0 && x >= -0.02)
626 		return 0.0;
627 	return x;
628 }
629 
630 static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
631 {
632 	return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
633 }
634 
635 static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
636 {
637 	double bad_spec = 0;
638 	double total_slots;
639 	double total;
640 
641 	total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
642 		runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
643 		runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
644 
645 	total_slots = td_total_slots(ctx, cpu, st);
646 	if (total_slots)
647 		bad_spec = total / total_slots;
648 	return sanitize_val(bad_spec);
649 }
650 
651 static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
652 {
653 	double retiring = 0;
654 	double total_slots = td_total_slots(ctx, cpu, st);
655 	double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
656 					    ctx, cpu);
657 
658 	if (total_slots)
659 		retiring = ret_slots / total_slots;
660 	return retiring;
661 }
662 
663 static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
664 {
665 	double fe_bound = 0;
666 	double total_slots = td_total_slots(ctx, cpu, st);
667 	double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
668 					    ctx, cpu);
669 
670 	if (total_slots)
671 		fe_bound = fetch_bub / total_slots;
672 	return fe_bound;
673 }
674 
675 static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
676 {
677 	double sum = (td_fe_bound(ctx, cpu, st) +
678 		      td_bad_spec(ctx, cpu, st) +
679 		      td_retiring(ctx, cpu, st));
680 	if (sum == 0)
681 		return 0;
682 	return sanitize_val(1.0 - sum);
683 }
684 
685 static void print_smi_cost(struct perf_stat_config *config,
686 			   int cpu, struct perf_evsel *evsel,
687 			   struct perf_stat_output_ctx *out,
688 			   struct runtime_stat *st)
689 {
690 	double smi_num, aperf, cycles, cost = 0.0;
691 	int ctx = evsel_context(evsel);
692 	const char *color = NULL;
693 
694 	smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
695 	aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
696 	cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
697 
698 	if ((cycles == 0) || (aperf == 0))
699 		return;
700 
701 	if (smi_num)
702 		cost = (aperf - cycles) / aperf * 100.00;
703 
704 	if (cost > 10)
705 		color = PERF_COLOR_RED;
706 	out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
707 	out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
708 }
709 
710 static void generic_metric(struct perf_stat_config *config,
711 			   const char *metric_expr,
712 			   struct perf_evsel **metric_events,
713 			   char *name,
714 			   const char *metric_name,
715 			   double avg,
716 			   int cpu,
717 			   struct perf_stat_output_ctx *out,
718 			   struct runtime_stat *st)
719 {
720 	print_metric_t print_metric = out->print_metric;
721 	struct parse_ctx pctx;
722 	double ratio;
723 	int i;
724 	void *ctxp = out->ctx;
725 
726 	expr__ctx_init(&pctx);
727 	expr__add_id(&pctx, name, avg);
728 	for (i = 0; metric_events[i]; i++) {
729 		struct saved_value *v;
730 		struct stats *stats;
731 		double scale;
732 
733 		if (!strcmp(metric_events[i]->name, "duration_time")) {
734 			stats = &walltime_nsecs_stats;
735 			scale = 1e-9;
736 		} else {
737 			v = saved_value_lookup(metric_events[i], cpu, false,
738 					       STAT_NONE, 0, st);
739 			if (!v)
740 				break;
741 			stats = &v->stats;
742 			scale = 1.0;
743 		}
744 		expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
745 	}
746 	if (!metric_events[i]) {
747 		const char *p = metric_expr;
748 
749 		if (expr__parse(&ratio, &pctx, &p) == 0)
750 			print_metric(config, ctxp, NULL, "%8.1f",
751 				metric_name ?
752 				metric_name :
753 				out->force_header ?  name : "",
754 				ratio);
755 		else
756 			print_metric(config, ctxp, NULL, NULL,
757 				     out->force_header ?
758 				     (metric_name ? metric_name : name) : "", 0);
759 	} else
760 		print_metric(config, ctxp, NULL, NULL, "", 0);
761 }
762 
763 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
764 				   struct perf_evsel *evsel,
765 				   double avg, int cpu,
766 				   struct perf_stat_output_ctx *out,
767 				   struct rblist *metric_events,
768 				   struct runtime_stat *st)
769 {
770 	void *ctxp = out->ctx;
771 	print_metric_t print_metric = out->print_metric;
772 	double total, ratio = 0.0, total2;
773 	const char *color = NULL;
774 	int ctx = evsel_context(evsel);
775 	struct metric_event *me;
776 	int num = 1;
777 
778 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
779 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
780 
781 		if (total) {
782 			ratio = avg / total;
783 			print_metric(config, ctxp, NULL, "%7.2f ",
784 					"insn per cycle", ratio);
785 		} else {
786 			print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
787 		}
788 
789 		total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
790 					 ctx, cpu);
791 
792 		total = max(total, runtime_stat_avg(st,
793 						    STAT_STALLED_CYCLES_BACK,
794 						    ctx, cpu));
795 
796 		if (total && avg) {
797 			out->new_line(config, ctxp);
798 			ratio = total / avg;
799 			print_metric(config, ctxp, NULL, "%7.2f ",
800 					"stalled cycles per insn",
801 					ratio);
802 		} else if (have_frontend_stalled) {
803 			print_metric(config, ctxp, NULL, NULL,
804 				     "stalled cycles per insn", 0);
805 		}
806 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
807 		if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
808 			print_branch_misses(config, cpu, evsel, avg, out, st);
809 		else
810 			print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
811 	} else if (
812 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
813 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
814 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
815 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
816 
817 		if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
818 			print_l1_dcache_misses(config, cpu, evsel, avg, out, st);
819 		else
820 			print_metric(config, ctxp, NULL, NULL, "of all L1-dcache hits", 0);
821 	} else if (
822 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
823 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
824 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
825 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
826 
827 		if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
828 			print_l1_icache_misses(config, cpu, evsel, avg, out, st);
829 		else
830 			print_metric(config, ctxp, NULL, NULL, "of all L1-icache hits", 0);
831 	} else if (
832 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
833 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
834 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
835 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
836 
837 		if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
838 			print_dtlb_cache_misses(config, cpu, evsel, avg, out, st);
839 		else
840 			print_metric(config, ctxp, NULL, NULL, "of all dTLB cache hits", 0);
841 	} else if (
842 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
843 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
844 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
845 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
846 
847 		if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
848 			print_itlb_cache_misses(config, cpu, evsel, avg, out, st);
849 		else
850 			print_metric(config, ctxp, NULL, NULL, "of all iTLB cache hits", 0);
851 	} else if (
852 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
853 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
854 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
855 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
856 
857 		if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
858 			print_ll_cache_misses(config, cpu, evsel, avg, out, st);
859 		else
860 			print_metric(config, ctxp, NULL, NULL, "of all LL-cache hits", 0);
861 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
862 		total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
863 
864 		if (total)
865 			ratio = avg * 100 / total;
866 
867 		if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
868 			print_metric(config, ctxp, NULL, "%8.3f %%",
869 				     "of all cache refs", ratio);
870 		else
871 			print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
872 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
873 		print_stalled_cycles_frontend(config, cpu, evsel, avg, out, st);
874 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
875 		print_stalled_cycles_backend(config, cpu, evsel, avg, out, st);
876 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
877 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
878 
879 		if (total) {
880 			ratio = avg / total;
881 			print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
882 		} else {
883 			print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
884 		}
885 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
886 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
887 
888 		if (total)
889 			print_metric(config, ctxp, NULL,
890 					"%7.2f%%", "transactional cycles",
891 					100.0 * (avg / total));
892 		else
893 			print_metric(config, ctxp, NULL, NULL, "transactional cycles",
894 				     0);
895 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
896 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
897 		total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
898 
899 		if (total2 < avg)
900 			total2 = avg;
901 		if (total)
902 			print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
903 				100.0 * ((total2-avg) / total));
904 		else
905 			print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
906 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
907 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
908 					 ctx, cpu);
909 
910 		if (avg)
911 			ratio = total / avg;
912 
913 		if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
914 			print_metric(config, ctxp, NULL, "%8.0f",
915 				     "cycles / transaction", ratio);
916 		else
917 			print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
918 				      0);
919 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
920 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
921 					 ctx, cpu);
922 
923 		if (avg)
924 			ratio = total / avg;
925 
926 		print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
927 	} else if (perf_evsel__is_clock(evsel)) {
928 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
929 			print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
930 				     avg / (ratio * evsel->scale));
931 		else
932 			print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
933 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
934 		double fe_bound = td_fe_bound(ctx, cpu, st);
935 
936 		if (fe_bound > 0.2)
937 			color = PERF_COLOR_RED;
938 		print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
939 				fe_bound * 100.);
940 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
941 		double retiring = td_retiring(ctx, cpu, st);
942 
943 		if (retiring > 0.7)
944 			color = PERF_COLOR_GREEN;
945 		print_metric(config, ctxp, color, "%8.1f%%", "retiring",
946 				retiring * 100.);
947 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
948 		double bad_spec = td_bad_spec(ctx, cpu, st);
949 
950 		if (bad_spec > 0.1)
951 			color = PERF_COLOR_RED;
952 		print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
953 				bad_spec * 100.);
954 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
955 		double be_bound = td_be_bound(ctx, cpu, st);
956 		const char *name = "backend bound";
957 		static int have_recovery_bubbles = -1;
958 
959 		/* In case the CPU does not support topdown-recovery-bubbles */
960 		if (have_recovery_bubbles < 0)
961 			have_recovery_bubbles = pmu_have_event("cpu",
962 					"topdown-recovery-bubbles");
963 		if (!have_recovery_bubbles)
964 			name = "backend bound/bad spec";
965 
966 		if (be_bound > 0.2)
967 			color = PERF_COLOR_RED;
968 		if (td_total_slots(ctx, cpu, st) > 0)
969 			print_metric(config, ctxp, color, "%8.1f%%", name,
970 					be_bound * 100.);
971 		else
972 			print_metric(config, ctxp, NULL, NULL, name, 0);
973 	} else if (evsel->metric_expr) {
974 		generic_metric(config, evsel->metric_expr, evsel->metric_events, evsel->name,
975 				evsel->metric_name, avg, cpu, out, st);
976 	} else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
977 		char unit = 'M';
978 		char unit_buf[10];
979 
980 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
981 
982 		if (total)
983 			ratio = 1000.0 * avg / total;
984 		if (ratio < 0.001) {
985 			ratio *= 1000;
986 			unit = 'K';
987 		}
988 		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
989 		print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
990 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
991 		print_smi_cost(config, cpu, evsel, out, st);
992 	} else {
993 		num = 0;
994 	}
995 
996 	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
997 		struct metric_expr *mexp;
998 
999 		list_for_each_entry (mexp, &me->head, nd) {
1000 			if (num++ > 0)
1001 				out->new_line(config, ctxp);
1002 			generic_metric(config, mexp->metric_expr, mexp->metric_events,
1003 					evsel->name, mexp->metric_name,
1004 					avg, cpu, out, st);
1005 		}
1006 	}
1007 	if (num == 0)
1008 		print_metric(config, ctxp, NULL, NULL, NULL, 0);
1009 }
1010