xref: /openbmc/linux/tools/perf/util/stat-shadow.c (revision 7d049cd1)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <math.h>
3 #include <stdio.h>
4 #include "evsel.h"
5 #include "stat.h"
6 #include "color.h"
7 #include "debug.h"
8 #include "pmu.h"
9 #include "rblist.h"
10 #include "evlist.h"
11 #include "expr.h"
12 #include "metricgroup.h"
13 #include "cgroup.h"
14 #include "units.h"
15 #include <linux/zalloc.h>
16 #include "iostat.h"
17 #include "util/hashmap.h"
18 
19 struct stats walltime_nsecs_stats;
20 struct rusage_stats ru_stats;
21 
22 enum {
23 	CTX_BIT_USER	= 1 << 0,
24 	CTX_BIT_KERNEL	= 1 << 1,
25 	CTX_BIT_HV	= 1 << 2,
26 	CTX_BIT_HOST	= 1 << 3,
27 	CTX_BIT_IDLE	= 1 << 4,
28 	CTX_BIT_MAX	= 1 << 5,
29 };
30 
31 enum stat_type {
32 	STAT_NONE = 0,
33 	STAT_NSECS,
34 	STAT_CYCLES,
35 	STAT_INSTRUCTIONS,
36 	STAT_STALLED_CYCLES_FRONT,
37 	STAT_STALLED_CYCLES_BACK,
38 	STAT_BRANCHES,
39 	STAT_BRANCH_MISS,
40 	STAT_CACHE_REFS,
41 	STAT_CACHE_MISSES,
42 	STAT_L1_DCACHE,
43 	STAT_L1_ICACHE,
44 	STAT_LL_CACHE,
45 	STAT_ITLB_CACHE,
46 	STAT_DTLB_CACHE,
47 	STAT_L1D_MISS,
48 	STAT_L1I_MISS,
49 	STAT_LL_MISS,
50 	STAT_DTLB_MISS,
51 	STAT_ITLB_MISS,
52 	STAT_MAX
53 };
54 
evsel_context(const struct evsel * evsel)55 static int evsel_context(const struct evsel *evsel)
56 {
57 	int ctx = 0;
58 
59 	if (evsel->core.attr.exclude_kernel)
60 		ctx |= CTX_BIT_KERNEL;
61 	if (evsel->core.attr.exclude_user)
62 		ctx |= CTX_BIT_USER;
63 	if (evsel->core.attr.exclude_hv)
64 		ctx |= CTX_BIT_HV;
65 	if (evsel->core.attr.exclude_host)
66 		ctx |= CTX_BIT_HOST;
67 	if (evsel->core.attr.exclude_idle)
68 		ctx |= CTX_BIT_IDLE;
69 
70 	return ctx;
71 }
72 
perf_stat__reset_shadow_stats(void)73 void perf_stat__reset_shadow_stats(void)
74 {
75 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
76 	memset(&ru_stats, 0, sizeof(ru_stats));
77 }
78 
evsel__stat_type(const struct evsel * evsel)79 static enum stat_type evsel__stat_type(const struct evsel *evsel)
80 {
81 	/* Fake perf_hw_cache_op_id values for use with evsel__match. */
82 	u64 PERF_COUNT_hw_cache_l1d_miss = PERF_COUNT_HW_CACHE_L1D |
83 		((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
84 		((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
85 	u64 PERF_COUNT_hw_cache_l1i_miss = PERF_COUNT_HW_CACHE_L1I |
86 		((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
87 		((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
88 	u64 PERF_COUNT_hw_cache_ll_miss = PERF_COUNT_HW_CACHE_LL |
89 		((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
90 		((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
91 	u64 PERF_COUNT_hw_cache_dtlb_miss = PERF_COUNT_HW_CACHE_DTLB |
92 		((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
93 		((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
94 	u64 PERF_COUNT_hw_cache_itlb_miss = PERF_COUNT_HW_CACHE_ITLB |
95 		((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
96 		((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
97 
98 	if (evsel__is_clock(evsel))
99 		return STAT_NSECS;
100 	else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES))
101 		return STAT_CYCLES;
102 	else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS))
103 		return STAT_INSTRUCTIONS;
104 	else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
105 		return STAT_STALLED_CYCLES_FRONT;
106 	else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND))
107 		return STAT_STALLED_CYCLES_BACK;
108 	else if (evsel__match(evsel, HARDWARE, HW_BRANCH_INSTRUCTIONS))
109 		return STAT_BRANCHES;
110 	else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES))
111 		return STAT_BRANCH_MISS;
112 	else if (evsel__match(evsel, HARDWARE, HW_CACHE_REFERENCES))
113 		return STAT_CACHE_REFS;
114 	else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES))
115 		return STAT_CACHE_MISSES;
116 	else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1D))
117 		return STAT_L1_DCACHE;
118 	else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1I))
119 		return STAT_L1_ICACHE;
120 	else if (evsel__match(evsel, HW_CACHE, HW_CACHE_LL))
121 		return STAT_LL_CACHE;
122 	else if (evsel__match(evsel, HW_CACHE, HW_CACHE_DTLB))
123 		return STAT_DTLB_CACHE;
124 	else if (evsel__match(evsel, HW_CACHE, HW_CACHE_ITLB))
125 		return STAT_ITLB_CACHE;
126 	else if (evsel__match(evsel, HW_CACHE, hw_cache_l1d_miss))
127 		return STAT_L1D_MISS;
128 	else if (evsel__match(evsel, HW_CACHE, hw_cache_l1i_miss))
129 		return STAT_L1I_MISS;
130 	else if (evsel__match(evsel, HW_CACHE, hw_cache_ll_miss))
131 		return STAT_LL_MISS;
132 	else if (evsel__match(evsel, HW_CACHE, hw_cache_dtlb_miss))
133 		return STAT_DTLB_MISS;
134 	else if (evsel__match(evsel, HW_CACHE, hw_cache_itlb_miss))
135 		return STAT_ITLB_MISS;
136 	return STAT_NONE;
137 }
138 
get_ratio_color(const double ratios[3],double val)139 static const char *get_ratio_color(const double ratios[3], double val)
140 {
141 	const char *color = PERF_COLOR_NORMAL;
142 
143 	if (val > ratios[0])
144 		color = PERF_COLOR_RED;
145 	else if (val > ratios[1])
146 		color = PERF_COLOR_MAGENTA;
147 	else if (val > ratios[2])
148 		color = PERF_COLOR_YELLOW;
149 
150 	return color;
151 }
152 
find_stat(const struct evsel * evsel,int aggr_idx,enum stat_type type)153 static double find_stat(const struct evsel *evsel, int aggr_idx, enum stat_type type)
154 {
155 	const struct evsel *cur;
156 	int evsel_ctx = evsel_context(evsel);
157 
158 	evlist__for_each_entry(evsel->evlist, cur) {
159 		struct perf_stat_aggr *aggr;
160 
161 		/* Ignore the evsel that is being searched from. */
162 		if (evsel == cur)
163 			continue;
164 
165 		/* Ignore evsels that are part of different groups. */
166 		if (evsel->core.leader->nr_members > 1 &&
167 		    evsel->core.leader != cur->core.leader)
168 			continue;
169 		/* Ignore evsels with mismatched modifiers. */
170 		if (evsel_ctx != evsel_context(cur))
171 			continue;
172 		/* Ignore if not the cgroup we're looking for. */
173 		if (evsel->cgrp != cur->cgrp)
174 			continue;
175 		/* Ignore if not the stat we're looking for. */
176 		if (type != evsel__stat_type(cur))
177 			continue;
178 
179 		/*
180 		 * Except the SW CLOCK events,
181 		 * ignore if not the PMU we're looking for.
182 		 */
183 		if ((type != STAT_NSECS) && (evsel->pmu != cur->pmu))
184 			continue;
185 
186 		aggr = &cur->stats->aggr[aggr_idx];
187 		if (type == STAT_NSECS)
188 			return aggr->counts.val;
189 		return aggr->counts.val * cur->scale;
190 	}
191 	return 0.0;
192 }
193 
print_ratio(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double numerator,struct perf_stat_output_ctx * out,enum stat_type denominator_type,const double color_ratios[3],const char * unit)194 static void print_ratio(struct perf_stat_config *config,
195 			const struct evsel *evsel, int aggr_idx,
196 			double numerator, struct perf_stat_output_ctx *out,
197 			enum stat_type denominator_type,
198 			const double color_ratios[3], const char *unit)
199 {
200 	double denominator = find_stat(evsel, aggr_idx, denominator_type);
201 
202 	if (numerator && denominator) {
203 		double ratio = numerator / denominator * 100.0;
204 		const char *color = get_ratio_color(color_ratios, ratio);
205 
206 		out->print_metric(config, out->ctx, color, "%7.2f%%", unit, ratio);
207 	} else
208 		out->print_metric(config, out->ctx, NULL, NULL, unit, 0);
209 }
210 
print_stalled_cycles_front(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double stalled,struct perf_stat_output_ctx * out)211 static void print_stalled_cycles_front(struct perf_stat_config *config,
212 				const struct evsel *evsel,
213 				int aggr_idx, double stalled,
214 				struct perf_stat_output_ctx *out)
215 {
216 	static const double color_ratios[3] = {50.0, 30.0, 10.0};
217 
218 	print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, color_ratios,
219 		    "frontend cycles idle");
220 }
221 
print_stalled_cycles_back(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double stalled,struct perf_stat_output_ctx * out)222 static void print_stalled_cycles_back(struct perf_stat_config *config,
223 				const struct evsel *evsel,
224 				int aggr_idx, double stalled,
225 				struct perf_stat_output_ctx *out)
226 {
227 	static const double color_ratios[3] = {75.0, 50.0, 20.0};
228 
229 	print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, color_ratios,
230 		    "backend cycles idle");
231 }
232 
print_branch_miss(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double misses,struct perf_stat_output_ctx * out)233 static void print_branch_miss(struct perf_stat_config *config,
234 			const struct evsel *evsel,
235 			int aggr_idx, double misses,
236 			struct perf_stat_output_ctx *out)
237 {
238 	static const double color_ratios[3] = {20.0, 10.0, 5.0};
239 
240 	print_ratio(config, evsel, aggr_idx, misses, out, STAT_BRANCHES, color_ratios,
241 		    "of all branches");
242 }
243 
print_l1d_miss(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double misses,struct perf_stat_output_ctx * out)244 static void print_l1d_miss(struct perf_stat_config *config,
245 			const struct evsel *evsel,
246 			int aggr_idx, double misses,
247 			struct perf_stat_output_ctx *out)
248 {
249 	static const double color_ratios[3] = {20.0, 10.0, 5.0};
250 
251 	print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_DCACHE, color_ratios,
252 		    "of all L1-dcache accesses");
253 }
254 
print_l1i_miss(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double misses,struct perf_stat_output_ctx * out)255 static void print_l1i_miss(struct perf_stat_config *config,
256 			const struct evsel *evsel,
257 			int aggr_idx, double misses,
258 			struct perf_stat_output_ctx *out)
259 {
260 	static const double color_ratios[3] = {20.0, 10.0, 5.0};
261 
262 	print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_ICACHE, color_ratios,
263 		    "of all L1-icache accesses");
264 }
265 
print_ll_miss(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double misses,struct perf_stat_output_ctx * out)266 static void print_ll_miss(struct perf_stat_config *config,
267 			const struct evsel *evsel,
268 			int aggr_idx, double misses,
269 			struct perf_stat_output_ctx *out)
270 {
271 	static const double color_ratios[3] = {20.0, 10.0, 5.0};
272 
273 	print_ratio(config, evsel, aggr_idx, misses, out, STAT_LL_CACHE, color_ratios,
274 		    "of all LL-cache accesses");
275 }
276 
print_dtlb_miss(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double misses,struct perf_stat_output_ctx * out)277 static void print_dtlb_miss(struct perf_stat_config *config,
278 			const struct evsel *evsel,
279 			int aggr_idx, double misses,
280 			struct perf_stat_output_ctx *out)
281 {
282 	static const double color_ratios[3] = {20.0, 10.0, 5.0};
283 
284 	print_ratio(config, evsel, aggr_idx, misses, out, STAT_DTLB_CACHE, color_ratios,
285 		    "of all dTLB cache accesses");
286 }
287 
print_itlb_miss(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double misses,struct perf_stat_output_ctx * out)288 static void print_itlb_miss(struct perf_stat_config *config,
289 			const struct evsel *evsel,
290 			int aggr_idx, double misses,
291 			struct perf_stat_output_ctx *out)
292 {
293 	static const double color_ratios[3] = {20.0, 10.0, 5.0};
294 
295 	print_ratio(config, evsel, aggr_idx, misses, out, STAT_ITLB_CACHE, color_ratios,
296 		    "of all iTLB cache accesses");
297 }
298 
print_cache_miss(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double misses,struct perf_stat_output_ctx * out)299 static void print_cache_miss(struct perf_stat_config *config,
300 			const struct evsel *evsel,
301 			int aggr_idx, double misses,
302 			struct perf_stat_output_ctx *out)
303 {
304 	static const double color_ratios[3] = {20.0, 10.0, 5.0};
305 
306 	print_ratio(config, evsel, aggr_idx, misses, out, STAT_CACHE_REFS, color_ratios,
307 		    "of all cache refs");
308 }
309 
print_instructions(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double instructions,struct perf_stat_output_ctx * out)310 static void print_instructions(struct perf_stat_config *config,
311 			const struct evsel *evsel,
312 			int aggr_idx, double instructions,
313 			struct perf_stat_output_ctx *out)
314 {
315 	print_metric_t print_metric = out->print_metric;
316 	void *ctxp = out->ctx;
317 	double cycles = find_stat(evsel, aggr_idx, STAT_CYCLES);
318 	double max_stalled = max(find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_FRONT),
319 				find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_BACK));
320 
321 	if (cycles) {
322 		print_metric(config, ctxp, NULL, "%7.2f ", "insn per cycle",
323 			instructions / cycles);
324 	} else
325 		print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
326 
327 	if (max_stalled && instructions) {
328 		out->new_line(config, ctxp);
329 		print_metric(config, ctxp, NULL, "%7.2f ", "stalled cycles per insn",
330 			max_stalled / instructions);
331 	}
332 }
333 
print_cycles(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx,double cycles,struct perf_stat_output_ctx * out)334 static void print_cycles(struct perf_stat_config *config,
335 			const struct evsel *evsel,
336 			int aggr_idx, double cycles,
337 			struct perf_stat_output_ctx *out)
338 {
339 	double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
340 
341 	if (cycles && nsecs) {
342 		double ratio = cycles / nsecs;
343 
344 		out->print_metric(config, out->ctx, NULL, "%8.3f", "GHz", ratio);
345 	} else
346 		out->print_metric(config, out->ctx, NULL, NULL, "GHz", 0);
347 }
348 
print_nsecs(struct perf_stat_config * config,const struct evsel * evsel,int aggr_idx __maybe_unused,double nsecs,struct perf_stat_output_ctx * out)349 static void print_nsecs(struct perf_stat_config *config,
350 			const struct evsel *evsel,
351 			int aggr_idx __maybe_unused, double nsecs,
352 			struct perf_stat_output_ctx *out)
353 {
354 	print_metric_t print_metric = out->print_metric;
355 	void *ctxp = out->ctx;
356 	double wall_time = avg_stats(&walltime_nsecs_stats);
357 
358 	if (wall_time) {
359 		print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
360 			nsecs / (wall_time * evsel->scale));
361 	} else
362 		print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
363 }
364 
prepare_metric(struct evsel ** metric_events,struct metric_ref * metric_refs,struct expr_parse_ctx * pctx,int aggr_idx)365 static int prepare_metric(struct evsel **metric_events,
366 			  struct metric_ref *metric_refs,
367 			  struct expr_parse_ctx *pctx,
368 			  int aggr_idx)
369 {
370 	int i;
371 
372 	for (i = 0; metric_events[i]; i++) {
373 		char *n;
374 		double val;
375 		int source_count = 0;
376 
377 		if (evsel__is_tool(metric_events[i])) {
378 			struct stats *stats;
379 			double scale;
380 
381 			switch (metric_events[i]->tool_event) {
382 			case PERF_TOOL_DURATION_TIME:
383 				stats = &walltime_nsecs_stats;
384 				scale = 1e-9;
385 				break;
386 			case PERF_TOOL_USER_TIME:
387 				stats = &ru_stats.ru_utime_usec_stat;
388 				scale = 1e-6;
389 				break;
390 			case PERF_TOOL_SYSTEM_TIME:
391 				stats = &ru_stats.ru_stime_usec_stat;
392 				scale = 1e-6;
393 				break;
394 			case PERF_TOOL_NONE:
395 				pr_err("Invalid tool event 'none'");
396 				abort();
397 			case PERF_TOOL_MAX:
398 				pr_err("Invalid tool event 'max'");
399 				abort();
400 			default:
401 				pr_err("Unknown tool event '%s'", evsel__name(metric_events[i]));
402 				abort();
403 			}
404 			val = avg_stats(stats) * scale;
405 			source_count = 1;
406 		} else {
407 			struct perf_stat_evsel *ps = metric_events[i]->stats;
408 			struct perf_stat_aggr *aggr = &ps->aggr[aggr_idx];
409 
410 			if (!aggr)
411 				break;
412 
413                         if (!metric_events[i]->supported) {
414 				/*
415 				 * Not supported events will have a count of 0,
416 				 * which can be confusing in a
417 				 * metric. Explicitly set the value to NAN. Not
418 				 * counted events (enable time of 0) are read as
419 				 * 0.
420 				 */
421 				val = NAN;
422 				source_count = 0;
423 			} else {
424 				val = aggr->counts.val;
425 				source_count = evsel__source_count(metric_events[i]);
426 			}
427 		}
428 		n = strdup(evsel__metric_id(metric_events[i]));
429 		if (!n)
430 			return -ENOMEM;
431 
432 		expr__add_id_val_source_count(pctx, n, val, source_count);
433 	}
434 
435 	for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
436 		int ret = expr__add_ref(pctx, &metric_refs[j]);
437 
438 		if (ret)
439 			return ret;
440 	}
441 
442 	return i;
443 }
444 
generic_metric(struct perf_stat_config * config,const char * metric_expr,const char * metric_threshold,struct evsel ** metric_events,struct metric_ref * metric_refs,char * name,const char * metric_name,const char * metric_unit,int runtime,int aggr_idx,struct perf_stat_output_ctx * out)445 static void generic_metric(struct perf_stat_config *config,
446 			   const char *metric_expr,
447 			   const char *metric_threshold,
448 			   struct evsel **metric_events,
449 			   struct metric_ref *metric_refs,
450 			   char *name,
451 			   const char *metric_name,
452 			   const char *metric_unit,
453 			   int runtime,
454 			   int aggr_idx,
455 			   struct perf_stat_output_ctx *out)
456 {
457 	print_metric_t print_metric = out->print_metric;
458 	struct expr_parse_ctx *pctx;
459 	double ratio, scale, threshold;
460 	int i;
461 	void *ctxp = out->ctx;
462 	const char *color = NULL;
463 
464 	pctx = expr__ctx_new();
465 	if (!pctx)
466 		return;
467 
468 	if (config->user_requested_cpu_list)
469 		pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
470 	pctx->sctx.runtime = runtime;
471 	pctx->sctx.system_wide = config->system_wide;
472 	i = prepare_metric(metric_events, metric_refs, pctx, aggr_idx);
473 	if (i < 0) {
474 		expr__ctx_free(pctx);
475 		return;
476 	}
477 	if (!metric_events[i]) {
478 		if (expr__parse(&ratio, pctx, metric_expr) == 0) {
479 			char *unit;
480 			char metric_bf[64];
481 
482 			if (metric_threshold &&
483 			    expr__parse(&threshold, pctx, metric_threshold) == 0 &&
484 			    !isnan(threshold)) {
485 				color = fpclassify(threshold) == FP_ZERO
486 					? PERF_COLOR_GREEN : PERF_COLOR_RED;
487 			}
488 
489 			if (metric_unit && metric_name) {
490 				if (perf_pmu__convert_scale(metric_unit,
491 					&unit, &scale) >= 0) {
492 					ratio *= scale;
493 				}
494 				if (strstr(metric_expr, "?"))
495 					scnprintf(metric_bf, sizeof(metric_bf),
496 					  "%s  %s_%d", unit, metric_name, runtime);
497 				else
498 					scnprintf(metric_bf, sizeof(metric_bf),
499 					  "%s  %s", unit, metric_name);
500 
501 				print_metric(config, ctxp, color, "%8.1f",
502 					     metric_bf, ratio);
503 			} else {
504 				print_metric(config, ctxp, color, "%8.2f",
505 					metric_name ?
506 					metric_name :
507 					out->force_header ?  name : "",
508 					ratio);
509 			}
510 		} else {
511 			print_metric(config, ctxp, color, /*unit=*/NULL,
512 				     out->force_header ?
513 				     (metric_name ? metric_name : name) : "", 0);
514 		}
515 	} else {
516 		print_metric(config, ctxp, color, /*unit=*/NULL,
517 			     out->force_header ?
518 			     (metric_name ? metric_name : name) : "", 0);
519 	}
520 
521 	expr__ctx_free(pctx);
522 }
523 
test_generic_metric(struct metric_expr * mexp,int aggr_idx)524 double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
525 {
526 	struct expr_parse_ctx *pctx;
527 	double ratio = 0.0;
528 
529 	pctx = expr__ctx_new();
530 	if (!pctx)
531 		return NAN;
532 
533 	if (prepare_metric(mexp->metric_events, mexp->metric_refs, pctx, aggr_idx) < 0)
534 		goto out;
535 
536 	if (expr__parse(&ratio, pctx, mexp->metric_expr))
537 		ratio = 0.0;
538 
539 out:
540 	expr__ctx_free(pctx);
541 	return ratio;
542 }
543 
perf_stat__print_metricgroup_header(struct perf_stat_config * config,struct evsel * evsel,void * ctxp,const char * name,struct perf_stat_output_ctx * out)544 static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,
545 						struct evsel *evsel,
546 						void *ctxp,
547 						const char *name,
548 						struct perf_stat_output_ctx *out)
549 {
550 	bool need_full_name = perf_pmus__num_core_pmus() > 1;
551 	static const char *last_name;
552 	static const char *last_pmu;
553 	char full_name[64];
554 
555 	/*
556 	 * A metricgroup may have several metric events,
557 	 * e.g.,TopdownL1 on e-core of ADL.
558 	 * The name has been output by the first metric
559 	 * event. Only align with other metics from
560 	 * different metric events.
561 	 */
562 	if (last_name && !strcmp(last_name, name)) {
563 		if (!need_full_name || !strcmp(last_pmu, evsel->pmu_name)) {
564 			out->print_metricgroup_header(config, ctxp, NULL);
565 			return;
566 		}
567 	}
568 
569 	if (need_full_name)
570 		scnprintf(full_name, sizeof(full_name), "%s (%s)", name, evsel->pmu_name);
571 	else
572 		scnprintf(full_name, sizeof(full_name), "%s", name);
573 
574 	out->print_metricgroup_header(config, ctxp, full_name);
575 
576 	last_name = name;
577 	last_pmu = evsel->pmu_name;
578 }
579 
580 /**
581  * perf_stat__print_shadow_stats_metricgroup - Print out metrics associated with the evsel
582  *					       For the non-default, all metrics associated
583  *					       with the evsel are printed.
584  *					       For the default mode, only the metrics from
585  *					       the same metricgroup and the name of the
586  *					       metricgroup are printed. To print the metrics
587  *					       from the next metricgroup (if available),
588  *					       invoke the function with correspoinding
589  *					       metric_expr.
590  */
perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config * config,struct evsel * evsel,int aggr_idx,int * num,void * from,struct perf_stat_output_ctx * out,struct rblist * metric_events)591 void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
592 						struct evsel *evsel,
593 						int aggr_idx,
594 						int *num,
595 						void *from,
596 						struct perf_stat_output_ctx *out,
597 						struct rblist *metric_events)
598 {
599 	struct metric_event *me;
600 	struct metric_expr *mexp = from;
601 	void *ctxp = out->ctx;
602 	bool header_printed = false;
603 	const char *name = NULL;
604 
605 	me = metricgroup__lookup(metric_events, evsel, false);
606 	if (me == NULL)
607 		return NULL;
608 
609 	if (!mexp)
610 		mexp = list_first_entry(&me->head, typeof(*mexp), nd);
611 
612 	list_for_each_entry_from(mexp, &me->head, nd) {
613 		/* Print the display name of the Default metricgroup */
614 		if (!config->metric_only && me->is_default) {
615 			if (!name)
616 				name = mexp->default_metricgroup_name;
617 			/*
618 			 * Two or more metricgroup may share the same metric
619 			 * event, e.g., TopdownL1 and TopdownL2 on SPR.
620 			 * Return and print the prefix, e.g., noise, running
621 			 * for the next metricgroup.
622 			 */
623 			if (strcmp(name, mexp->default_metricgroup_name))
624 				return (void *)mexp;
625 			/* Only print the name of the metricgroup once */
626 			if (!header_printed) {
627 				header_printed = true;
628 				perf_stat__print_metricgroup_header(config, evsel, ctxp,
629 								    name, out);
630 			}
631 		}
632 
633 		if ((*num)++ > 0)
634 			out->new_line(config, ctxp);
635 		generic_metric(config, mexp->metric_expr, mexp->metric_threshold,
636 			       mexp->metric_events, mexp->metric_refs, evsel->name,
637 			       mexp->metric_name, mexp->metric_unit, mexp->runtime,
638 			       aggr_idx, out);
639 	}
640 
641 	return NULL;
642 }
643 
perf_stat__print_shadow_stats(struct perf_stat_config * config,struct evsel * evsel,double avg,int aggr_idx,struct perf_stat_output_ctx * out,struct rblist * metric_events)644 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
645 				   struct evsel *evsel,
646 				   double avg, int aggr_idx,
647 				   struct perf_stat_output_ctx *out,
648 				   struct rblist *metric_events)
649 {
650 	typedef void (*stat_print_function_t)(struct perf_stat_config *config,
651 					const struct evsel *evsel,
652 					int aggr_idx, double misses,
653 					struct perf_stat_output_ctx *out);
654 	static const stat_print_function_t stat_print_function[STAT_MAX] = {
655 		[STAT_INSTRUCTIONS] = print_instructions,
656 		[STAT_BRANCH_MISS] = print_branch_miss,
657 		[STAT_L1D_MISS] = print_l1d_miss,
658 		[STAT_L1I_MISS] = print_l1i_miss,
659 		[STAT_DTLB_MISS] = print_dtlb_miss,
660 		[STAT_ITLB_MISS] = print_itlb_miss,
661 		[STAT_LL_MISS] = print_ll_miss,
662 		[STAT_CACHE_MISSES] = print_cache_miss,
663 		[STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
664 		[STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
665 		[STAT_CYCLES] = print_cycles,
666 		[STAT_NSECS] = print_nsecs,
667 	};
668 	print_metric_t print_metric = out->print_metric;
669 	void *ctxp = out->ctx;
670 	int num = 1;
671 
672 	if (config->iostat_run) {
673 		iostat_print_metric(config, evsel, out);
674 	} else {
675 		stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
676 
677 		if (fn)
678 			fn(config, evsel, aggr_idx, avg, out);
679 		else {
680 			double nsecs =	find_stat(evsel, aggr_idx, STAT_NSECS);
681 
682 			if (nsecs) {
683 				char unit = ' ';
684 				char unit_buf[10] = "/sec";
685 				double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
686 								   &unit);
687 
688 				if (unit != ' ')
689 					snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
690 				print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
691 			} else
692 				num = 0;
693 		}
694 	}
695 
696 	perf_stat__print_shadow_stats_metricgroup(config, evsel, aggr_idx,
697 						  &num, NULL, out, metric_events);
698 
699 	if (num == 0)
700 		print_metric(config, ctxp, NULL, NULL, NULL, 0);
701 }
702 
703 /**
704  * perf_stat__skip_metric_event - Skip the evsel in the Default metricgroup,
705  *				  if it's not running or not the metric event.
706  */
perf_stat__skip_metric_event(struct evsel * evsel,struct rblist * metric_events,u64 ena,u64 run)707 bool perf_stat__skip_metric_event(struct evsel *evsel,
708 				  struct rblist *metric_events,
709 				  u64 ena, u64 run)
710 {
711 	if (!evsel->default_metricgroup)
712 		return false;
713 
714 	if (!ena || !run)
715 		return true;
716 
717 	return !metricgroup__lookup(metric_events, evsel, false);
718 }
719