xref: /openbmc/linux/tools/perf/util/stat-shadow.c (revision 96de2506)
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 
213 	count *= counter->scale;
214 
215 	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK) ||
216 	    perf_evsel__match(counter, SOFTWARE, SW_CPU_CLOCK))
217 		update_runtime_stat(st, STAT_NSECS, 0, cpu, count);
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(int cpu,
414 					  struct perf_evsel *evsel, double avg,
415 					  struct perf_stat_output_ctx *out,
416 					  struct runtime_stat *st)
417 {
418 	double total, ratio = 0.0;
419 	const char *color;
420 	int ctx = evsel_context(evsel);
421 
422 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
423 
424 	if (total)
425 		ratio = avg / total * 100.0;
426 
427 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
428 
429 	if (ratio)
430 		out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
431 				  ratio);
432 	else
433 		out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
434 }
435 
436 static void print_stalled_cycles_backend(int cpu,
437 					 struct perf_evsel *evsel, double avg,
438 					 struct perf_stat_output_ctx *out,
439 					 struct runtime_stat *st)
440 {
441 	double total, ratio = 0.0;
442 	const char *color;
443 	int ctx = evsel_context(evsel);
444 
445 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
446 
447 	if (total)
448 		ratio = avg / total * 100.0;
449 
450 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
451 
452 	out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
453 }
454 
455 static void print_branch_misses(int cpu,
456 				struct perf_evsel *evsel,
457 				double avg,
458 				struct perf_stat_output_ctx *out,
459 				struct runtime_stat *st)
460 {
461 	double total, ratio = 0.0;
462 	const char *color;
463 	int ctx = evsel_context(evsel);
464 
465 	total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
466 
467 	if (total)
468 		ratio = avg / total * 100.0;
469 
470 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
471 
472 	out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
473 }
474 
475 static void print_l1_dcache_misses(int cpu,
476 				   struct perf_evsel *evsel,
477 				   double avg,
478 				   struct perf_stat_output_ctx *out,
479 				   struct runtime_stat *st)
480 
481 {
482 	double total, ratio = 0.0;
483 	const char *color;
484 	int ctx = evsel_context(evsel);
485 
486 	total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
487 
488 	if (total)
489 		ratio = avg / total * 100.0;
490 
491 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
492 
493 	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
494 }
495 
496 static void print_l1_icache_misses(int cpu,
497 				   struct perf_evsel *evsel,
498 				   double avg,
499 				   struct perf_stat_output_ctx *out,
500 				   struct runtime_stat *st)
501 
502 {
503 	double total, ratio = 0.0;
504 	const char *color;
505 	int ctx = evsel_context(evsel);
506 
507 	total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
508 
509 	if (total)
510 		ratio = avg / total * 100.0;
511 
512 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
513 	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
514 }
515 
516 static void print_dtlb_cache_misses(int cpu,
517 				    struct perf_evsel *evsel,
518 				    double avg,
519 				    struct perf_stat_output_ctx *out,
520 				    struct runtime_stat *st)
521 {
522 	double total, ratio = 0.0;
523 	const char *color;
524 	int ctx = evsel_context(evsel);
525 
526 	total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
527 
528 	if (total)
529 		ratio = avg / total * 100.0;
530 
531 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
532 	out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
533 }
534 
535 static void print_itlb_cache_misses(int cpu,
536 				    struct perf_evsel *evsel,
537 				    double avg,
538 				    struct perf_stat_output_ctx *out,
539 				    struct runtime_stat *st)
540 {
541 	double total, ratio = 0.0;
542 	const char *color;
543 	int ctx = evsel_context(evsel);
544 
545 	total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
546 
547 	if (total)
548 		ratio = avg / total * 100.0;
549 
550 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
551 	out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
552 }
553 
554 static void print_ll_cache_misses(int cpu,
555 				  struct perf_evsel *evsel,
556 				  double avg,
557 				  struct perf_stat_output_ctx *out,
558 				  struct runtime_stat *st)
559 {
560 	double total, ratio = 0.0;
561 	const char *color;
562 	int ctx = evsel_context(evsel);
563 
564 	total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
565 
566 	if (total)
567 		ratio = avg / total * 100.0;
568 
569 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
570 	out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
571 }
572 
573 /*
574  * High level "TopDown" CPU core pipe line bottleneck break down.
575  *
576  * Basic concept following
577  * Yasin, A Top Down Method for Performance analysis and Counter architecture
578  * ISPASS14
579  *
580  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
581  *
582  * Frontend -> Backend -> Retiring
583  * BadSpeculation in addition means out of order execution that is thrown away
584  * (for example branch mispredictions)
585  * Frontend is instruction decoding.
586  * Backend is execution, like computation and accessing data in memory
587  * Retiring is good execution that is not directly bottlenecked
588  *
589  * The formulas are computed in slots.
590  * A slot is an entry in the pipeline each for the pipeline width
591  * (for example a 4-wide pipeline has 4 slots for each cycle)
592  *
593  * Formulas:
594  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
595  *			TotalSlots
596  * Retiring = SlotsRetired / TotalSlots
597  * FrontendBound = FetchBubbles / TotalSlots
598  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
599  *
600  * The kernel provides the mapping to the low level CPU events and any scaling
601  * needed for the CPU pipeline width, for example:
602  *
603  * TotalSlots = Cycles * 4
604  *
605  * The scaling factor is communicated in the sysfs unit.
606  *
607  * In some cases the CPU may not be able to measure all the formulas due to
608  * missing events. In this case multiple formulas are combined, as possible.
609  *
610  * Full TopDown supports more levels to sub-divide each area: for example
611  * BackendBound into computing bound and memory bound. For now we only
612  * support Level 1 TopDown.
613  */
614 
615 static double sanitize_val(double x)
616 {
617 	if (x < 0 && x >= -0.02)
618 		return 0.0;
619 	return x;
620 }
621 
622 static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
623 {
624 	return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
625 }
626 
627 static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
628 {
629 	double bad_spec = 0;
630 	double total_slots;
631 	double total;
632 
633 	total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
634 		runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
635 		runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
636 
637 	total_slots = td_total_slots(ctx, cpu, st);
638 	if (total_slots)
639 		bad_spec = total / total_slots;
640 	return sanitize_val(bad_spec);
641 }
642 
643 static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
644 {
645 	double retiring = 0;
646 	double total_slots = td_total_slots(ctx, cpu, st);
647 	double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
648 					    ctx, cpu);
649 
650 	if (total_slots)
651 		retiring = ret_slots / total_slots;
652 	return retiring;
653 }
654 
655 static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
656 {
657 	double fe_bound = 0;
658 	double total_slots = td_total_slots(ctx, cpu, st);
659 	double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
660 					    ctx, cpu);
661 
662 	if (total_slots)
663 		fe_bound = fetch_bub / total_slots;
664 	return fe_bound;
665 }
666 
667 static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
668 {
669 	double sum = (td_fe_bound(ctx, cpu, st) +
670 		      td_bad_spec(ctx, cpu, st) +
671 		      td_retiring(ctx, cpu, st));
672 	if (sum == 0)
673 		return 0;
674 	return sanitize_val(1.0 - sum);
675 }
676 
677 static void print_smi_cost(int cpu, struct perf_evsel *evsel,
678 			   struct perf_stat_output_ctx *out,
679 			   struct runtime_stat *st)
680 {
681 	double smi_num, aperf, cycles, cost = 0.0;
682 	int ctx = evsel_context(evsel);
683 	const char *color = NULL;
684 
685 	smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
686 	aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
687 	cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
688 
689 	if ((cycles == 0) || (aperf == 0))
690 		return;
691 
692 	if (smi_num)
693 		cost = (aperf - cycles) / aperf * 100.00;
694 
695 	if (cost > 10)
696 		color = PERF_COLOR_RED;
697 	out->print_metric(out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
698 	out->print_metric(out->ctx, NULL, "%4.0f", "SMI#", smi_num);
699 }
700 
701 static void generic_metric(const char *metric_expr,
702 			   struct perf_evsel **metric_events,
703 			   char *name,
704 			   const char *metric_name,
705 			   double avg,
706 			   int cpu,
707 			   struct perf_stat_output_ctx *out,
708 			   struct runtime_stat *st)
709 {
710 	print_metric_t print_metric = out->print_metric;
711 	struct parse_ctx pctx;
712 	double ratio;
713 	int i;
714 	void *ctxp = out->ctx;
715 
716 	expr__ctx_init(&pctx);
717 	expr__add_id(&pctx, name, avg);
718 	for (i = 0; metric_events[i]; i++) {
719 		struct saved_value *v;
720 		struct stats *stats;
721 		double scale;
722 
723 		if (!strcmp(metric_events[i]->name, "duration_time")) {
724 			stats = &walltime_nsecs_stats;
725 			scale = 1e-9;
726 		} else {
727 			v = saved_value_lookup(metric_events[i], cpu, false,
728 					       STAT_NONE, 0, st);
729 			if (!v)
730 				break;
731 			stats = &v->stats;
732 			scale = 1.0;
733 		}
734 		expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
735 	}
736 	if (!metric_events[i]) {
737 		const char *p = metric_expr;
738 
739 		if (expr__parse(&ratio, &pctx, &p) == 0)
740 			print_metric(ctxp, NULL, "%8.1f",
741 				metric_name ?
742 				metric_name :
743 				out->force_header ?  name : "",
744 				ratio);
745 		else
746 			print_metric(ctxp, NULL, NULL,
747 				     out->force_header ?
748 				     (metric_name ? metric_name : name) : "", 0);
749 	} else
750 		print_metric(ctxp, NULL, NULL, "", 0);
751 }
752 
753 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
754 				   double avg, int cpu,
755 				   struct perf_stat_output_ctx *out,
756 				   struct rblist *metric_events,
757 				   struct runtime_stat *st)
758 {
759 	void *ctxp = out->ctx;
760 	print_metric_t print_metric = out->print_metric;
761 	double total, ratio = 0.0, total2;
762 	const char *color = NULL;
763 	int ctx = evsel_context(evsel);
764 	struct metric_event *me;
765 	int num = 1;
766 
767 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
768 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
769 
770 		if (total) {
771 			ratio = avg / total;
772 			print_metric(ctxp, NULL, "%7.2f ",
773 					"insn per cycle", ratio);
774 		} else {
775 			print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
776 		}
777 
778 		total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
779 					 ctx, cpu);
780 
781 		total = max(total, runtime_stat_avg(st,
782 						    STAT_STALLED_CYCLES_BACK,
783 						    ctx, cpu));
784 
785 		if (total && avg) {
786 			out->new_line(ctxp);
787 			ratio = total / avg;
788 			print_metric(ctxp, NULL, "%7.2f ",
789 					"stalled cycles per insn",
790 					ratio);
791 		} else if (have_frontend_stalled) {
792 			print_metric(ctxp, NULL, NULL,
793 				     "stalled cycles per insn", 0);
794 		}
795 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
796 		if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
797 			print_branch_misses(cpu, evsel, avg, out, st);
798 		else
799 			print_metric(ctxp, NULL, NULL, "of all branches", 0);
800 	} else if (
801 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
802 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
803 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
804 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
805 
806 		if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
807 			print_l1_dcache_misses(cpu, evsel, avg, out, st);
808 		else
809 			print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
810 	} else if (
811 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
812 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
813 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
814 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
815 
816 		if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
817 			print_l1_icache_misses(cpu, evsel, avg, out, st);
818 		else
819 			print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
820 	} else if (
821 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
822 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
823 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
824 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
825 
826 		if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
827 			print_dtlb_cache_misses(cpu, evsel, avg, out, st);
828 		else
829 			print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
830 	} else if (
831 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
832 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
833 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
834 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
835 
836 		if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
837 			print_itlb_cache_misses(cpu, evsel, avg, out, st);
838 		else
839 			print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
840 	} else if (
841 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
842 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
843 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
844 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
845 
846 		if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
847 			print_ll_cache_misses(cpu, evsel, avg, out, st);
848 		else
849 			print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
850 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
851 		total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
852 
853 		if (total)
854 			ratio = avg * 100 / total;
855 
856 		if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
857 			print_metric(ctxp, NULL, "%8.3f %%",
858 				     "of all cache refs", ratio);
859 		else
860 			print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
861 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
862 		print_stalled_cycles_frontend(cpu, evsel, avg, out, st);
863 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
864 		print_stalled_cycles_backend(cpu, evsel, avg, out, st);
865 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
866 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
867 
868 		if (total) {
869 			ratio = avg / total;
870 			print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
871 		} else {
872 			print_metric(ctxp, NULL, NULL, "Ghz", 0);
873 		}
874 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
875 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
876 
877 		if (total)
878 			print_metric(ctxp, NULL,
879 					"%7.2f%%", "transactional cycles",
880 					100.0 * (avg / total));
881 		else
882 			print_metric(ctxp, NULL, NULL, "transactional cycles",
883 				     0);
884 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
885 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
886 		total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
887 
888 		if (total2 < avg)
889 			total2 = avg;
890 		if (total)
891 			print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
892 				100.0 * ((total2-avg) / total));
893 		else
894 			print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
895 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
896 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
897 					 ctx, cpu);
898 
899 		if (avg)
900 			ratio = total / avg;
901 
902 		if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
903 			print_metric(ctxp, NULL, "%8.0f",
904 				     "cycles / transaction", ratio);
905 		else
906 			print_metric(ctxp, NULL, NULL, "cycles / transaction",
907 				      0);
908 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
909 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
910 					 ctx, cpu);
911 
912 		if (avg)
913 			ratio = total / avg;
914 
915 		print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
916 	} else if (perf_evsel__is_clock(evsel)) {
917 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
918 			print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
919 				     avg / (ratio * evsel->scale));
920 		else
921 			print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
922 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
923 		double fe_bound = td_fe_bound(ctx, cpu, st);
924 
925 		if (fe_bound > 0.2)
926 			color = PERF_COLOR_RED;
927 		print_metric(ctxp, color, "%8.1f%%", "frontend bound",
928 				fe_bound * 100.);
929 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
930 		double retiring = td_retiring(ctx, cpu, st);
931 
932 		if (retiring > 0.7)
933 			color = PERF_COLOR_GREEN;
934 		print_metric(ctxp, color, "%8.1f%%", "retiring",
935 				retiring * 100.);
936 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
937 		double bad_spec = td_bad_spec(ctx, cpu, st);
938 
939 		if (bad_spec > 0.1)
940 			color = PERF_COLOR_RED;
941 		print_metric(ctxp, color, "%8.1f%%", "bad speculation",
942 				bad_spec * 100.);
943 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
944 		double be_bound = td_be_bound(ctx, cpu, st);
945 		const char *name = "backend bound";
946 		static int have_recovery_bubbles = -1;
947 
948 		/* In case the CPU does not support topdown-recovery-bubbles */
949 		if (have_recovery_bubbles < 0)
950 			have_recovery_bubbles = pmu_have_event("cpu",
951 					"topdown-recovery-bubbles");
952 		if (!have_recovery_bubbles)
953 			name = "backend bound/bad spec";
954 
955 		if (be_bound > 0.2)
956 			color = PERF_COLOR_RED;
957 		if (td_total_slots(ctx, cpu, st) > 0)
958 			print_metric(ctxp, color, "%8.1f%%", name,
959 					be_bound * 100.);
960 		else
961 			print_metric(ctxp, NULL, NULL, name, 0);
962 	} else if (evsel->metric_expr) {
963 		generic_metric(evsel->metric_expr, evsel->metric_events, evsel->name,
964 				evsel->metric_name, avg, cpu, out, st);
965 	} else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
966 		char unit = 'M';
967 		char unit_buf[10];
968 
969 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
970 
971 		if (total)
972 			ratio = 1000.0 * avg / total;
973 		if (ratio < 0.001) {
974 			ratio *= 1000;
975 			unit = 'K';
976 		}
977 		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
978 		print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
979 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
980 		print_smi_cost(cpu, evsel, out, st);
981 	} else {
982 		num = 0;
983 	}
984 
985 	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
986 		struct metric_expr *mexp;
987 
988 		list_for_each_entry (mexp, &me->head, nd) {
989 			if (num++ > 0)
990 				out->new_line(ctxp);
991 			generic_metric(mexp->metric_expr, mexp->metric_events,
992 					evsel->name, mexp->metric_name,
993 					avg, cpu, out, st);
994 		}
995 	}
996 	if (num == 0)
997 		print_metric(ctxp, NULL, NULL, NULL, 0);
998 }
999