xref: /openbmc/linux/tools/perf/util/stat-shadow.c (revision 680ef72a)
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 enum {
13 	CTX_BIT_USER	= 1 << 0,
14 	CTX_BIT_KERNEL	= 1 << 1,
15 	CTX_BIT_HV	= 1 << 2,
16 	CTX_BIT_HOST	= 1 << 3,
17 	CTX_BIT_IDLE	= 1 << 4,
18 	CTX_BIT_MAX	= 1 << 5,
19 };
20 
21 #define NUM_CTX CTX_BIT_MAX
22 
23 /*
24  * AGGR_GLOBAL: Use CPU 0
25  * AGGR_SOCKET: Use first CPU of socket
26  * AGGR_CORE: Use first CPU of core
27  * AGGR_NONE: Use matching CPU
28  * AGGR_THREAD: Not supported?
29  */
30 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
31 static struct stats runtime_cycles_stats[NUM_CTX][MAX_NR_CPUS];
32 static struct stats runtime_stalled_cycles_front_stats[NUM_CTX][MAX_NR_CPUS];
33 static struct stats runtime_stalled_cycles_back_stats[NUM_CTX][MAX_NR_CPUS];
34 static struct stats runtime_branches_stats[NUM_CTX][MAX_NR_CPUS];
35 static struct stats runtime_cacherefs_stats[NUM_CTX][MAX_NR_CPUS];
36 static struct stats runtime_l1_dcache_stats[NUM_CTX][MAX_NR_CPUS];
37 static struct stats runtime_l1_icache_stats[NUM_CTX][MAX_NR_CPUS];
38 static struct stats runtime_ll_cache_stats[NUM_CTX][MAX_NR_CPUS];
39 static struct stats runtime_itlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
40 static struct stats runtime_dtlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
41 static struct stats runtime_cycles_in_tx_stats[NUM_CTX][MAX_NR_CPUS];
42 static struct stats runtime_transaction_stats[NUM_CTX][MAX_NR_CPUS];
43 static struct stats runtime_elision_stats[NUM_CTX][MAX_NR_CPUS];
44 static struct stats runtime_topdown_total_slots[NUM_CTX][MAX_NR_CPUS];
45 static struct stats runtime_topdown_slots_issued[NUM_CTX][MAX_NR_CPUS];
46 static struct stats runtime_topdown_slots_retired[NUM_CTX][MAX_NR_CPUS];
47 static struct stats runtime_topdown_fetch_bubbles[NUM_CTX][MAX_NR_CPUS];
48 static struct stats runtime_topdown_recovery_bubbles[NUM_CTX][MAX_NR_CPUS];
49 static struct stats runtime_smi_num_stats[NUM_CTX][MAX_NR_CPUS];
50 static struct stats runtime_aperf_stats[NUM_CTX][MAX_NR_CPUS];
51 static struct rblist runtime_saved_values;
52 static bool have_frontend_stalled;
53 
54 struct stats walltime_nsecs_stats;
55 
56 struct saved_value {
57 	struct rb_node rb_node;
58 	struct perf_evsel *evsel;
59 	int cpu;
60 	struct stats stats;
61 };
62 
63 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
64 {
65 	struct saved_value *a = container_of(rb_node,
66 					     struct saved_value,
67 					     rb_node);
68 	const struct saved_value *b = entry;
69 
70 	if (a->cpu != b->cpu)
71 		return a->cpu - b->cpu;
72 	if (a->evsel == b->evsel)
73 		return 0;
74 	if ((char *)a->evsel < (char *)b->evsel)
75 		return -1;
76 	return +1;
77 }
78 
79 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
80 				     const void *entry)
81 {
82 	struct saved_value *nd = malloc(sizeof(struct saved_value));
83 
84 	if (!nd)
85 		return NULL;
86 	memcpy(nd, entry, sizeof(struct saved_value));
87 	return &nd->rb_node;
88 }
89 
90 static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
91 					      int cpu,
92 					      bool create)
93 {
94 	struct rb_node *nd;
95 	struct saved_value dm = {
96 		.cpu = cpu,
97 		.evsel = evsel,
98 	};
99 	nd = rblist__find(&runtime_saved_values, &dm);
100 	if (nd)
101 		return container_of(nd, struct saved_value, rb_node);
102 	if (create) {
103 		rblist__add_node(&runtime_saved_values, &dm);
104 		nd = rblist__find(&runtime_saved_values, &dm);
105 		if (nd)
106 			return container_of(nd, struct saved_value, rb_node);
107 	}
108 	return NULL;
109 }
110 
111 void perf_stat__init_shadow_stats(void)
112 {
113 	have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
114 	rblist__init(&runtime_saved_values);
115 	runtime_saved_values.node_cmp = saved_value_cmp;
116 	runtime_saved_values.node_new = saved_value_new;
117 	/* No delete for now */
118 }
119 
120 static int evsel_context(struct perf_evsel *evsel)
121 {
122 	int ctx = 0;
123 
124 	if (evsel->attr.exclude_kernel)
125 		ctx |= CTX_BIT_KERNEL;
126 	if (evsel->attr.exclude_user)
127 		ctx |= CTX_BIT_USER;
128 	if (evsel->attr.exclude_hv)
129 		ctx |= CTX_BIT_HV;
130 	if (evsel->attr.exclude_host)
131 		ctx |= CTX_BIT_HOST;
132 	if (evsel->attr.exclude_idle)
133 		ctx |= CTX_BIT_IDLE;
134 
135 	return ctx;
136 }
137 
138 void perf_stat__reset_shadow_stats(void)
139 {
140 	struct rb_node *pos, *next;
141 
142 	memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
143 	memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
144 	memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
145 	memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
146 	memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
147 	memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
148 	memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
149 	memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
150 	memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
151 	memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
152 	memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
153 	memset(runtime_cycles_in_tx_stats, 0,
154 			sizeof(runtime_cycles_in_tx_stats));
155 	memset(runtime_transaction_stats, 0,
156 		sizeof(runtime_transaction_stats));
157 	memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
158 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
159 	memset(runtime_topdown_total_slots, 0, sizeof(runtime_topdown_total_slots));
160 	memset(runtime_topdown_slots_retired, 0, sizeof(runtime_topdown_slots_retired));
161 	memset(runtime_topdown_slots_issued, 0, sizeof(runtime_topdown_slots_issued));
162 	memset(runtime_topdown_fetch_bubbles, 0, sizeof(runtime_topdown_fetch_bubbles));
163 	memset(runtime_topdown_recovery_bubbles, 0, sizeof(runtime_topdown_recovery_bubbles));
164 	memset(runtime_smi_num_stats, 0, sizeof(runtime_smi_num_stats));
165 	memset(runtime_aperf_stats, 0, sizeof(runtime_aperf_stats));
166 
167 	next = rb_first(&runtime_saved_values.entries);
168 	while (next) {
169 		pos = next;
170 		next = rb_next(pos);
171 		memset(&container_of(pos, struct saved_value, rb_node)->stats,
172 		       0,
173 		       sizeof(struct stats));
174 	}
175 }
176 
177 /*
178  * Update various tracking values we maintain to print
179  * more semantic information such as miss/hit ratios,
180  * instruction rates, etc:
181  */
182 void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 count,
183 				    int cpu)
184 {
185 	int ctx = evsel_context(counter);
186 
187 	count *= counter->scale;
188 
189 	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK) ||
190 	    perf_evsel__match(counter, SOFTWARE, SW_CPU_CLOCK))
191 		update_stats(&runtime_nsecs_stats[cpu], count);
192 	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
193 		update_stats(&runtime_cycles_stats[ctx][cpu], count);
194 	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
195 		update_stats(&runtime_cycles_in_tx_stats[ctx][cpu], count);
196 	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
197 		update_stats(&runtime_transaction_stats[ctx][cpu], count);
198 	else if (perf_stat_evsel__is(counter, ELISION_START))
199 		update_stats(&runtime_elision_stats[ctx][cpu], count);
200 	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
201 		update_stats(&runtime_topdown_total_slots[ctx][cpu], count);
202 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
203 		update_stats(&runtime_topdown_slots_issued[ctx][cpu], count);
204 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
205 		update_stats(&runtime_topdown_slots_retired[ctx][cpu], count);
206 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
207 		update_stats(&runtime_topdown_fetch_bubbles[ctx][cpu], count);
208 	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
209 		update_stats(&runtime_topdown_recovery_bubbles[ctx][cpu], count);
210 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
211 		update_stats(&runtime_stalled_cycles_front_stats[ctx][cpu], count);
212 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
213 		update_stats(&runtime_stalled_cycles_back_stats[ctx][cpu], count);
214 	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
215 		update_stats(&runtime_branches_stats[ctx][cpu], count);
216 	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
217 		update_stats(&runtime_cacherefs_stats[ctx][cpu], count);
218 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
219 		update_stats(&runtime_l1_dcache_stats[ctx][cpu], count);
220 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
221 		update_stats(&runtime_ll_cache_stats[ctx][cpu], count);
222 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
223 		update_stats(&runtime_ll_cache_stats[ctx][cpu], count);
224 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
225 		update_stats(&runtime_dtlb_cache_stats[ctx][cpu], count);
226 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
227 		update_stats(&runtime_itlb_cache_stats[ctx][cpu], count);
228 	else if (perf_stat_evsel__is(counter, SMI_NUM))
229 		update_stats(&runtime_smi_num_stats[ctx][cpu], count);
230 	else if (perf_stat_evsel__is(counter, APERF))
231 		update_stats(&runtime_aperf_stats[ctx][cpu], count);
232 
233 	if (counter->collect_stat) {
234 		struct saved_value *v = saved_value_lookup(counter, cpu, true);
235 		update_stats(&v->stats, count);
236 	}
237 }
238 
239 /* used for get_ratio_color() */
240 enum grc_type {
241 	GRC_STALLED_CYCLES_FE,
242 	GRC_STALLED_CYCLES_BE,
243 	GRC_CACHE_MISSES,
244 	GRC_MAX_NR
245 };
246 
247 static const char *get_ratio_color(enum grc_type type, double ratio)
248 {
249 	static const double grc_table[GRC_MAX_NR][3] = {
250 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
251 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
252 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
253 	};
254 	const char *color = PERF_COLOR_NORMAL;
255 
256 	if (ratio > grc_table[type][0])
257 		color = PERF_COLOR_RED;
258 	else if (ratio > grc_table[type][1])
259 		color = PERF_COLOR_MAGENTA;
260 	else if (ratio > grc_table[type][2])
261 		color = PERF_COLOR_YELLOW;
262 
263 	return color;
264 }
265 
266 static struct perf_evsel *perf_stat__find_event(struct perf_evlist *evsel_list,
267 						const char *name)
268 {
269 	struct perf_evsel *c2;
270 
271 	evlist__for_each_entry (evsel_list, c2) {
272 		if (!strcasecmp(c2->name, name))
273 			return c2;
274 	}
275 	return NULL;
276 }
277 
278 /* Mark MetricExpr target events and link events using them to them. */
279 void perf_stat__collect_metric_expr(struct perf_evlist *evsel_list)
280 {
281 	struct perf_evsel *counter, *leader, **metric_events, *oc;
282 	bool found;
283 	const char **metric_names;
284 	int i;
285 	int num_metric_names;
286 
287 	evlist__for_each_entry(evsel_list, counter) {
288 		bool invalid = false;
289 
290 		leader = counter->leader;
291 		if (!counter->metric_expr)
292 			continue;
293 		metric_events = counter->metric_events;
294 		if (!metric_events) {
295 			if (expr__find_other(counter->metric_expr, counter->name,
296 						&metric_names, &num_metric_names) < 0)
297 				continue;
298 
299 			metric_events = calloc(sizeof(struct perf_evsel *),
300 					       num_metric_names + 1);
301 			if (!metric_events)
302 				return;
303 			counter->metric_events = metric_events;
304 		}
305 
306 		for (i = 0; i < num_metric_names; i++) {
307 			found = false;
308 			if (leader) {
309 				/* Search in group */
310 				for_each_group_member (oc, leader) {
311 					if (!strcasecmp(oc->name, metric_names[i])) {
312 						found = true;
313 						break;
314 					}
315 				}
316 			}
317 			if (!found) {
318 				/* Search ignoring groups */
319 				oc = perf_stat__find_event(evsel_list, metric_names[i]);
320 			}
321 			if (!oc) {
322 				/* Deduping one is good enough to handle duplicated PMUs. */
323 				static char *printed;
324 
325 				/*
326 				 * Adding events automatically would be difficult, because
327 				 * it would risk creating groups that are not schedulable.
328 				 * perf stat doesn't understand all the scheduling constraints
329 				 * of events. So we ask the user instead to add the missing
330 				 * events.
331 				 */
332 				if (!printed || strcasecmp(printed, metric_names[i])) {
333 					fprintf(stderr,
334 						"Add %s event to groups to get metric expression for %s\n",
335 						metric_names[i],
336 						counter->name);
337 					printed = strdup(metric_names[i]);
338 				}
339 				invalid = true;
340 				continue;
341 			}
342 			metric_events[i] = oc;
343 			oc->collect_stat = true;
344 		}
345 		metric_events[i] = NULL;
346 		free(metric_names);
347 		if (invalid) {
348 			free(metric_events);
349 			counter->metric_events = NULL;
350 			counter->metric_expr = NULL;
351 		}
352 	}
353 }
354 
355 static void print_stalled_cycles_frontend(int cpu,
356 					  struct perf_evsel *evsel, double avg,
357 					  struct perf_stat_output_ctx *out)
358 {
359 	double total, ratio = 0.0;
360 	const char *color;
361 	int ctx = evsel_context(evsel);
362 
363 	total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
364 
365 	if (total)
366 		ratio = avg / total * 100.0;
367 
368 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
369 
370 	if (ratio)
371 		out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
372 				  ratio);
373 	else
374 		out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
375 }
376 
377 static void print_stalled_cycles_backend(int cpu,
378 					 struct perf_evsel *evsel, double avg,
379 					 struct perf_stat_output_ctx *out)
380 {
381 	double total, ratio = 0.0;
382 	const char *color;
383 	int ctx = evsel_context(evsel);
384 
385 	total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
386 
387 	if (total)
388 		ratio = avg / total * 100.0;
389 
390 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
391 
392 	out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
393 }
394 
395 static void print_branch_misses(int cpu,
396 				struct perf_evsel *evsel,
397 				double avg,
398 				struct perf_stat_output_ctx *out)
399 {
400 	double total, ratio = 0.0;
401 	const char *color;
402 	int ctx = evsel_context(evsel);
403 
404 	total = avg_stats(&runtime_branches_stats[ctx][cpu]);
405 
406 	if (total)
407 		ratio = avg / total * 100.0;
408 
409 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
410 
411 	out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
412 }
413 
414 static void print_l1_dcache_misses(int cpu,
415 				   struct perf_evsel *evsel,
416 				   double avg,
417 				   struct perf_stat_output_ctx *out)
418 {
419 	double total, ratio = 0.0;
420 	const char *color;
421 	int ctx = evsel_context(evsel);
422 
423 	total = avg_stats(&runtime_l1_dcache_stats[ctx][cpu]);
424 
425 	if (total)
426 		ratio = avg / total * 100.0;
427 
428 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
429 
430 	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
431 }
432 
433 static void print_l1_icache_misses(int cpu,
434 				   struct perf_evsel *evsel,
435 				   double avg,
436 				   struct perf_stat_output_ctx *out)
437 {
438 	double total, ratio = 0.0;
439 	const char *color;
440 	int ctx = evsel_context(evsel);
441 
442 	total = avg_stats(&runtime_l1_icache_stats[ctx][cpu]);
443 
444 	if (total)
445 		ratio = avg / total * 100.0;
446 
447 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
448 	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
449 }
450 
451 static void print_dtlb_cache_misses(int cpu,
452 				    struct perf_evsel *evsel,
453 				    double avg,
454 				    struct perf_stat_output_ctx *out)
455 {
456 	double total, ratio = 0.0;
457 	const char *color;
458 	int ctx = evsel_context(evsel);
459 
460 	total = avg_stats(&runtime_dtlb_cache_stats[ctx][cpu]);
461 
462 	if (total)
463 		ratio = avg / total * 100.0;
464 
465 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
466 	out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
467 }
468 
469 static void print_itlb_cache_misses(int cpu,
470 				    struct perf_evsel *evsel,
471 				    double avg,
472 				    struct perf_stat_output_ctx *out)
473 {
474 	double total, ratio = 0.0;
475 	const char *color;
476 	int ctx = evsel_context(evsel);
477 
478 	total = avg_stats(&runtime_itlb_cache_stats[ctx][cpu]);
479 
480 	if (total)
481 		ratio = avg / total * 100.0;
482 
483 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
484 	out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
485 }
486 
487 static void print_ll_cache_misses(int cpu,
488 				  struct perf_evsel *evsel,
489 				  double avg,
490 				  struct perf_stat_output_ctx *out)
491 {
492 	double total, ratio = 0.0;
493 	const char *color;
494 	int ctx = evsel_context(evsel);
495 
496 	total = avg_stats(&runtime_ll_cache_stats[ctx][cpu]);
497 
498 	if (total)
499 		ratio = avg / total * 100.0;
500 
501 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
502 	out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
503 }
504 
505 /*
506  * High level "TopDown" CPU core pipe line bottleneck break down.
507  *
508  * Basic concept following
509  * Yasin, A Top Down Method for Performance analysis and Counter architecture
510  * ISPASS14
511  *
512  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
513  *
514  * Frontend -> Backend -> Retiring
515  * BadSpeculation in addition means out of order execution that is thrown away
516  * (for example branch mispredictions)
517  * Frontend is instruction decoding.
518  * Backend is execution, like computation and accessing data in memory
519  * Retiring is good execution that is not directly bottlenecked
520  *
521  * The formulas are computed in slots.
522  * A slot is an entry in the pipeline each for the pipeline width
523  * (for example a 4-wide pipeline has 4 slots for each cycle)
524  *
525  * Formulas:
526  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
527  *			TotalSlots
528  * Retiring = SlotsRetired / TotalSlots
529  * FrontendBound = FetchBubbles / TotalSlots
530  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
531  *
532  * The kernel provides the mapping to the low level CPU events and any scaling
533  * needed for the CPU pipeline width, for example:
534  *
535  * TotalSlots = Cycles * 4
536  *
537  * The scaling factor is communicated in the sysfs unit.
538  *
539  * In some cases the CPU may not be able to measure all the formulas due to
540  * missing events. In this case multiple formulas are combined, as possible.
541  *
542  * Full TopDown supports more levels to sub-divide each area: for example
543  * BackendBound into computing bound and memory bound. For now we only
544  * support Level 1 TopDown.
545  */
546 
547 static double sanitize_val(double x)
548 {
549 	if (x < 0 && x >= -0.02)
550 		return 0.0;
551 	return x;
552 }
553 
554 static double td_total_slots(int ctx, int cpu)
555 {
556 	return avg_stats(&runtime_topdown_total_slots[ctx][cpu]);
557 }
558 
559 static double td_bad_spec(int ctx, int cpu)
560 {
561 	double bad_spec = 0;
562 	double total_slots;
563 	double total;
564 
565 	total = avg_stats(&runtime_topdown_slots_issued[ctx][cpu]) -
566 		avg_stats(&runtime_topdown_slots_retired[ctx][cpu]) +
567 		avg_stats(&runtime_topdown_recovery_bubbles[ctx][cpu]);
568 	total_slots = td_total_slots(ctx, cpu);
569 	if (total_slots)
570 		bad_spec = total / total_slots;
571 	return sanitize_val(bad_spec);
572 }
573 
574 static double td_retiring(int ctx, int cpu)
575 {
576 	double retiring = 0;
577 	double total_slots = td_total_slots(ctx, cpu);
578 	double ret_slots = avg_stats(&runtime_topdown_slots_retired[ctx][cpu]);
579 
580 	if (total_slots)
581 		retiring = ret_slots / total_slots;
582 	return retiring;
583 }
584 
585 static double td_fe_bound(int ctx, int cpu)
586 {
587 	double fe_bound = 0;
588 	double total_slots = td_total_slots(ctx, cpu);
589 	double fetch_bub = avg_stats(&runtime_topdown_fetch_bubbles[ctx][cpu]);
590 
591 	if (total_slots)
592 		fe_bound = fetch_bub / total_slots;
593 	return fe_bound;
594 }
595 
596 static double td_be_bound(int ctx, int cpu)
597 {
598 	double sum = (td_fe_bound(ctx, cpu) +
599 		      td_bad_spec(ctx, cpu) +
600 		      td_retiring(ctx, cpu));
601 	if (sum == 0)
602 		return 0;
603 	return sanitize_val(1.0 - sum);
604 }
605 
606 static void print_smi_cost(int cpu, struct perf_evsel *evsel,
607 			   struct perf_stat_output_ctx *out)
608 {
609 	double smi_num, aperf, cycles, cost = 0.0;
610 	int ctx = evsel_context(evsel);
611 	const char *color = NULL;
612 
613 	smi_num = avg_stats(&runtime_smi_num_stats[ctx][cpu]);
614 	aperf = avg_stats(&runtime_aperf_stats[ctx][cpu]);
615 	cycles = avg_stats(&runtime_cycles_stats[ctx][cpu]);
616 
617 	if ((cycles == 0) || (aperf == 0))
618 		return;
619 
620 	if (smi_num)
621 		cost = (aperf - cycles) / aperf * 100.00;
622 
623 	if (cost > 10)
624 		color = PERF_COLOR_RED;
625 	out->print_metric(out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
626 	out->print_metric(out->ctx, NULL, "%4.0f", "SMI#", smi_num);
627 }
628 
629 static void generic_metric(const char *metric_expr,
630 			   struct perf_evsel **metric_events,
631 			   char *name,
632 			   const char *metric_name,
633 			   double avg,
634 			   int cpu,
635 			   struct perf_stat_output_ctx *out)
636 {
637 	print_metric_t print_metric = out->print_metric;
638 	struct parse_ctx pctx;
639 	double ratio;
640 	int i;
641 	void *ctxp = out->ctx;
642 
643 	expr__ctx_init(&pctx);
644 	expr__add_id(&pctx, name, avg);
645 	for (i = 0; metric_events[i]; i++) {
646 		struct saved_value *v;
647 		struct stats *stats;
648 		double scale;
649 
650 		if (!strcmp(metric_events[i]->name, "duration_time")) {
651 			stats = &walltime_nsecs_stats;
652 			scale = 1e-9;
653 		} else {
654 			v = saved_value_lookup(metric_events[i], cpu, false);
655 			if (!v)
656 				break;
657 			stats = &v->stats;
658 			scale = 1.0;
659 		}
660 		expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
661 	}
662 	if (!metric_events[i]) {
663 		const char *p = metric_expr;
664 
665 		if (expr__parse(&ratio, &pctx, &p) == 0)
666 			print_metric(ctxp, NULL, "%8.1f",
667 				metric_name ?
668 				metric_name :
669 				out->force_header ?  name : "",
670 				ratio);
671 		else
672 			print_metric(ctxp, NULL, NULL,
673 				     out->force_header ?
674 				     (metric_name ? metric_name : name) : "", 0);
675 	} else
676 		print_metric(ctxp, NULL, NULL, "", 0);
677 }
678 
679 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
680 				   double avg, int cpu,
681 				   struct perf_stat_output_ctx *out,
682 				   struct rblist *metric_events)
683 {
684 	void *ctxp = out->ctx;
685 	print_metric_t print_metric = out->print_metric;
686 	double total, ratio = 0.0, total2;
687 	const char *color = NULL;
688 	int ctx = evsel_context(evsel);
689 	struct metric_event *me;
690 	int num = 1;
691 
692 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
693 		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
694 		if (total) {
695 			ratio = avg / total;
696 			print_metric(ctxp, NULL, "%7.2f ",
697 					"insn per cycle", ratio);
698 		} else {
699 			print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
700 		}
701 		total = avg_stats(&runtime_stalled_cycles_front_stats[ctx][cpu]);
702 		total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[ctx][cpu]));
703 
704 		if (total && avg) {
705 			out->new_line(ctxp);
706 			ratio = total / avg;
707 			print_metric(ctxp, NULL, "%7.2f ",
708 					"stalled cycles per insn",
709 					ratio);
710 		} else if (have_frontend_stalled) {
711 			print_metric(ctxp, NULL, NULL,
712 				     "stalled cycles per insn", 0);
713 		}
714 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
715 		if (runtime_branches_stats[ctx][cpu].n != 0)
716 			print_branch_misses(cpu, evsel, avg, out);
717 		else
718 			print_metric(ctxp, NULL, NULL, "of all branches", 0);
719 	} else if (
720 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
721 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
722 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
723 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
724 		if (runtime_l1_dcache_stats[ctx][cpu].n != 0)
725 			print_l1_dcache_misses(cpu, evsel, avg, out);
726 		else
727 			print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
728 	} else if (
729 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
730 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
731 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
732 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
733 		if (runtime_l1_icache_stats[ctx][cpu].n != 0)
734 			print_l1_icache_misses(cpu, evsel, avg, out);
735 		else
736 			print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
737 	} else if (
738 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
739 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
740 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
741 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
742 		if (runtime_dtlb_cache_stats[ctx][cpu].n != 0)
743 			print_dtlb_cache_misses(cpu, evsel, avg, out);
744 		else
745 			print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
746 	} else if (
747 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
748 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
749 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
750 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
751 		if (runtime_itlb_cache_stats[ctx][cpu].n != 0)
752 			print_itlb_cache_misses(cpu, evsel, avg, out);
753 		else
754 			print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
755 	} else if (
756 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
757 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
758 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
759 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
760 		if (runtime_ll_cache_stats[ctx][cpu].n != 0)
761 			print_ll_cache_misses(cpu, evsel, avg, out);
762 		else
763 			print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
764 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
765 		total = avg_stats(&runtime_cacherefs_stats[ctx][cpu]);
766 
767 		if (total)
768 			ratio = avg * 100 / total;
769 
770 		if (runtime_cacherefs_stats[ctx][cpu].n != 0)
771 			print_metric(ctxp, NULL, "%8.3f %%",
772 				     "of all cache refs", ratio);
773 		else
774 			print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
775 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
776 		print_stalled_cycles_frontend(cpu, evsel, avg, out);
777 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
778 		print_stalled_cycles_backend(cpu, evsel, avg, out);
779 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
780 		total = avg_stats(&runtime_nsecs_stats[cpu]);
781 
782 		if (total) {
783 			ratio = avg / total;
784 			print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
785 		} else {
786 			print_metric(ctxp, NULL, NULL, "Ghz", 0);
787 		}
788 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
789 		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
790 		if (total)
791 			print_metric(ctxp, NULL,
792 					"%7.2f%%", "transactional cycles",
793 					100.0 * (avg / total));
794 		else
795 			print_metric(ctxp, NULL, NULL, "transactional cycles",
796 				     0);
797 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
798 		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
799 		total2 = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
800 		if (total2 < avg)
801 			total2 = avg;
802 		if (total)
803 			print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
804 				100.0 * ((total2-avg) / total));
805 		else
806 			print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
807 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
808 		total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
809 
810 		if (avg)
811 			ratio = total / avg;
812 
813 		if (runtime_cycles_in_tx_stats[ctx][cpu].n != 0)
814 			print_metric(ctxp, NULL, "%8.0f",
815 				     "cycles / transaction", ratio);
816 		else
817 			print_metric(ctxp, NULL, NULL, "cycles / transaction",
818 				     0);
819 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
820 		total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
821 
822 		if (avg)
823 			ratio = total / avg;
824 
825 		print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
826 	} else if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK) ||
827 		   perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK)) {
828 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
829 			print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
830 				     avg / ratio);
831 		else
832 			print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
833 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
834 		double fe_bound = td_fe_bound(ctx, cpu);
835 
836 		if (fe_bound > 0.2)
837 			color = PERF_COLOR_RED;
838 		print_metric(ctxp, color, "%8.1f%%", "frontend bound",
839 				fe_bound * 100.);
840 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
841 		double retiring = td_retiring(ctx, cpu);
842 
843 		if (retiring > 0.7)
844 			color = PERF_COLOR_GREEN;
845 		print_metric(ctxp, color, "%8.1f%%", "retiring",
846 				retiring * 100.);
847 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
848 		double bad_spec = td_bad_spec(ctx, cpu);
849 
850 		if (bad_spec > 0.1)
851 			color = PERF_COLOR_RED;
852 		print_metric(ctxp, color, "%8.1f%%", "bad speculation",
853 				bad_spec * 100.);
854 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
855 		double be_bound = td_be_bound(ctx, cpu);
856 		const char *name = "backend bound";
857 		static int have_recovery_bubbles = -1;
858 
859 		/* In case the CPU does not support topdown-recovery-bubbles */
860 		if (have_recovery_bubbles < 0)
861 			have_recovery_bubbles = pmu_have_event("cpu",
862 					"topdown-recovery-bubbles");
863 		if (!have_recovery_bubbles)
864 			name = "backend bound/bad spec";
865 
866 		if (be_bound > 0.2)
867 			color = PERF_COLOR_RED;
868 		if (td_total_slots(ctx, cpu) > 0)
869 			print_metric(ctxp, color, "%8.1f%%", name,
870 					be_bound * 100.);
871 		else
872 			print_metric(ctxp, NULL, NULL, name, 0);
873 	} else if (evsel->metric_expr) {
874 		generic_metric(evsel->metric_expr, evsel->metric_events, evsel->name,
875 				evsel->metric_name, avg, cpu, out);
876 	} else if (runtime_nsecs_stats[cpu].n != 0) {
877 		char unit = 'M';
878 		char unit_buf[10];
879 
880 		total = avg_stats(&runtime_nsecs_stats[cpu]);
881 
882 		if (total)
883 			ratio = 1000.0 * avg / total;
884 		if (ratio < 0.001) {
885 			ratio *= 1000;
886 			unit = 'K';
887 		}
888 		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
889 		print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
890 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
891 		print_smi_cost(cpu, evsel, out);
892 	} else {
893 		num = 0;
894 	}
895 
896 	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
897 		struct metric_expr *mexp;
898 
899 		list_for_each_entry (mexp, &me->head, nd) {
900 			if (num++ > 0)
901 				out->new_line(ctxp);
902 			generic_metric(mexp->metric_expr, mexp->metric_events,
903 					evsel->name, mexp->metric_name,
904 					avg, cpu, out);
905 		}
906 	}
907 	if (num == 0)
908 		print_metric(ctxp, NULL, NULL, NULL, 0);
909 }
910