Loading...
1#include <stdio.h>
2#include "evsel.h"
3#include "stat.h"
4#include "color.h"
5#include "pmu.h"
6
7enum {
8 CTX_BIT_USER = 1 << 0,
9 CTX_BIT_KERNEL = 1 << 1,
10 CTX_BIT_HV = 1 << 2,
11 CTX_BIT_HOST = 1 << 3,
12 CTX_BIT_IDLE = 1 << 4,
13 CTX_BIT_MAX = 1 << 5,
14};
15
16#define NUM_CTX CTX_BIT_MAX
17
18/*
19 * AGGR_GLOBAL: Use CPU 0
20 * AGGR_SOCKET: Use first CPU of socket
21 * AGGR_CORE: Use first CPU of core
22 * AGGR_NONE: Use matching CPU
23 * AGGR_THREAD: Not supported?
24 */
25static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
26static struct stats runtime_cycles_stats[NUM_CTX][MAX_NR_CPUS];
27static struct stats runtime_stalled_cycles_front_stats[NUM_CTX][MAX_NR_CPUS];
28static struct stats runtime_stalled_cycles_back_stats[NUM_CTX][MAX_NR_CPUS];
29static struct stats runtime_branches_stats[NUM_CTX][MAX_NR_CPUS];
30static struct stats runtime_cacherefs_stats[NUM_CTX][MAX_NR_CPUS];
31static struct stats runtime_l1_dcache_stats[NUM_CTX][MAX_NR_CPUS];
32static struct stats runtime_l1_icache_stats[NUM_CTX][MAX_NR_CPUS];
33static struct stats runtime_ll_cache_stats[NUM_CTX][MAX_NR_CPUS];
34static struct stats runtime_itlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
35static struct stats runtime_dtlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
36static struct stats runtime_cycles_in_tx_stats[NUM_CTX][MAX_NR_CPUS];
37static struct stats runtime_transaction_stats[NUM_CTX][MAX_NR_CPUS];
38static struct stats runtime_elision_stats[NUM_CTX][MAX_NR_CPUS];
39static bool have_frontend_stalled;
40
41struct stats walltime_nsecs_stats;
42
43void perf_stat__init_shadow_stats(void)
44{
45 have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
46}
47
48static int evsel_context(struct perf_evsel *evsel)
49{
50 int ctx = 0;
51
52 if (evsel->attr.exclude_kernel)
53 ctx |= CTX_BIT_KERNEL;
54 if (evsel->attr.exclude_user)
55 ctx |= CTX_BIT_USER;
56 if (evsel->attr.exclude_hv)
57 ctx |= CTX_BIT_HV;
58 if (evsel->attr.exclude_host)
59 ctx |= CTX_BIT_HOST;
60 if (evsel->attr.exclude_idle)
61 ctx |= CTX_BIT_IDLE;
62
63 return ctx;
64}
65
66void perf_stat__reset_shadow_stats(void)
67{
68 memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
69 memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
70 memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
71 memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
72 memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
73 memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
74 memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
75 memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
76 memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
77 memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
78 memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
79 memset(runtime_cycles_in_tx_stats, 0,
80 sizeof(runtime_cycles_in_tx_stats));
81 memset(runtime_transaction_stats, 0,
82 sizeof(runtime_transaction_stats));
83 memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
84 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
85}
86
87/*
88 * Update various tracking values we maintain to print
89 * more semantic information such as miss/hit ratios,
90 * instruction rates, etc:
91 */
92void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 *count,
93 int cpu)
94{
95 int ctx = evsel_context(counter);
96
97 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
98 update_stats(&runtime_nsecs_stats[cpu], count[0]);
99 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
100 update_stats(&runtime_cycles_stats[ctx][cpu], count[0]);
101 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
102 update_stats(&runtime_cycles_in_tx_stats[ctx][cpu], count[0]);
103 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
104 update_stats(&runtime_transaction_stats[ctx][cpu], count[0]);
105 else if (perf_stat_evsel__is(counter, ELISION_START))
106 update_stats(&runtime_elision_stats[ctx][cpu], count[0]);
107 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
108 update_stats(&runtime_stalled_cycles_front_stats[ctx][cpu], count[0]);
109 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
110 update_stats(&runtime_stalled_cycles_back_stats[ctx][cpu], count[0]);
111 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
112 update_stats(&runtime_branches_stats[ctx][cpu], count[0]);
113 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
114 update_stats(&runtime_cacherefs_stats[ctx][cpu], count[0]);
115 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
116 update_stats(&runtime_l1_dcache_stats[ctx][cpu], count[0]);
117 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
118 update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
119 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
120 update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
121 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
122 update_stats(&runtime_dtlb_cache_stats[ctx][cpu], count[0]);
123 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
124 update_stats(&runtime_itlb_cache_stats[ctx][cpu], count[0]);
125}
126
127/* used for get_ratio_color() */
128enum grc_type {
129 GRC_STALLED_CYCLES_FE,
130 GRC_STALLED_CYCLES_BE,
131 GRC_CACHE_MISSES,
132 GRC_MAX_NR
133};
134
135static const char *get_ratio_color(enum grc_type type, double ratio)
136{
137 static const double grc_table[GRC_MAX_NR][3] = {
138 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
139 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
140 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
141 };
142 const char *color = PERF_COLOR_NORMAL;
143
144 if (ratio > grc_table[type][0])
145 color = PERF_COLOR_RED;
146 else if (ratio > grc_table[type][1])
147 color = PERF_COLOR_MAGENTA;
148 else if (ratio > grc_table[type][2])
149 color = PERF_COLOR_YELLOW;
150
151 return color;
152}
153
154static void print_stalled_cycles_frontend(int cpu,
155 struct perf_evsel *evsel, double avg,
156 struct perf_stat_output_ctx *out)
157{
158 double total, ratio = 0.0;
159 const char *color;
160 int ctx = evsel_context(evsel);
161
162 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
163
164 if (total)
165 ratio = avg / total * 100.0;
166
167 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
168
169 if (ratio)
170 out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
171 ratio);
172 else
173 out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
174}
175
176static void print_stalled_cycles_backend(int cpu,
177 struct perf_evsel *evsel, double avg,
178 struct perf_stat_output_ctx *out)
179{
180 double total, ratio = 0.0;
181 const char *color;
182 int ctx = evsel_context(evsel);
183
184 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
185
186 if (total)
187 ratio = avg / total * 100.0;
188
189 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
190
191 out->print_metric(out->ctx, color, "%6.2f%%", "backend cycles idle", ratio);
192}
193
194static void print_branch_misses(int cpu,
195 struct perf_evsel *evsel,
196 double avg,
197 struct perf_stat_output_ctx *out)
198{
199 double total, ratio = 0.0;
200 const char *color;
201 int ctx = evsel_context(evsel);
202
203 total = avg_stats(&runtime_branches_stats[ctx][cpu]);
204
205 if (total)
206 ratio = avg / total * 100.0;
207
208 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
209
210 out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
211}
212
213static void print_l1_dcache_misses(int cpu,
214 struct perf_evsel *evsel,
215 double avg,
216 struct perf_stat_output_ctx *out)
217{
218 double total, ratio = 0.0;
219 const char *color;
220 int ctx = evsel_context(evsel);
221
222 total = avg_stats(&runtime_l1_dcache_stats[ctx][cpu]);
223
224 if (total)
225 ratio = avg / total * 100.0;
226
227 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
228
229 out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
230}
231
232static void print_l1_icache_misses(int cpu,
233 struct perf_evsel *evsel,
234 double avg,
235 struct perf_stat_output_ctx *out)
236{
237 double total, ratio = 0.0;
238 const char *color;
239 int ctx = evsel_context(evsel);
240
241 total = avg_stats(&runtime_l1_icache_stats[ctx][cpu]);
242
243 if (total)
244 ratio = avg / total * 100.0;
245
246 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
247 out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
248}
249
250static void print_dtlb_cache_misses(int cpu,
251 struct perf_evsel *evsel,
252 double avg,
253 struct perf_stat_output_ctx *out)
254{
255 double total, ratio = 0.0;
256 const char *color;
257 int ctx = evsel_context(evsel);
258
259 total = avg_stats(&runtime_dtlb_cache_stats[ctx][cpu]);
260
261 if (total)
262 ratio = avg / total * 100.0;
263
264 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
265 out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
266}
267
268static void print_itlb_cache_misses(int cpu,
269 struct perf_evsel *evsel,
270 double avg,
271 struct perf_stat_output_ctx *out)
272{
273 double total, ratio = 0.0;
274 const char *color;
275 int ctx = evsel_context(evsel);
276
277 total = avg_stats(&runtime_itlb_cache_stats[ctx][cpu]);
278
279 if (total)
280 ratio = avg / total * 100.0;
281
282 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
283 out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
284}
285
286static void print_ll_cache_misses(int cpu,
287 struct perf_evsel *evsel,
288 double avg,
289 struct perf_stat_output_ctx *out)
290{
291 double total, ratio = 0.0;
292 const char *color;
293 int ctx = evsel_context(evsel);
294
295 total = avg_stats(&runtime_ll_cache_stats[ctx][cpu]);
296
297 if (total)
298 ratio = avg / total * 100.0;
299
300 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
301 out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
302}
303
304void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
305 double avg, int cpu,
306 struct perf_stat_output_ctx *out)
307{
308 void *ctxp = out->ctx;
309 print_metric_t print_metric = out->print_metric;
310 double total, ratio = 0.0, total2;
311 int ctx = evsel_context(evsel);
312
313 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
314 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
315 if (total) {
316 ratio = avg / total;
317 print_metric(ctxp, NULL, "%7.2f ",
318 "insn per cycle", ratio);
319 } else {
320 print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
321 }
322 total = avg_stats(&runtime_stalled_cycles_front_stats[ctx][cpu]);
323 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[ctx][cpu]));
324
325 if (total && avg) {
326 out->new_line(ctxp);
327 ratio = total / avg;
328 print_metric(ctxp, NULL, "%7.2f ",
329 "stalled cycles per insn",
330 ratio);
331 } else if (have_frontend_stalled) {
332 print_metric(ctxp, NULL, NULL,
333 "stalled cycles per insn", 0);
334 }
335 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
336 if (runtime_branches_stats[ctx][cpu].n != 0)
337 print_branch_misses(cpu, evsel, avg, out);
338 else
339 print_metric(ctxp, NULL, NULL, "of all branches", 0);
340 } else if (
341 evsel->attr.type == PERF_TYPE_HW_CACHE &&
342 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
343 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
344 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
345 if (runtime_l1_dcache_stats[ctx][cpu].n != 0)
346 print_l1_dcache_misses(cpu, evsel, avg, out);
347 else
348 print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
349 } else if (
350 evsel->attr.type == PERF_TYPE_HW_CACHE &&
351 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
352 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
353 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
354 if (runtime_l1_icache_stats[ctx][cpu].n != 0)
355 print_l1_icache_misses(cpu, evsel, avg, out);
356 else
357 print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
358 } else if (
359 evsel->attr.type == PERF_TYPE_HW_CACHE &&
360 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
361 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
362 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
363 if (runtime_dtlb_cache_stats[ctx][cpu].n != 0)
364 print_dtlb_cache_misses(cpu, evsel, avg, out);
365 else
366 print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
367 } else if (
368 evsel->attr.type == PERF_TYPE_HW_CACHE &&
369 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
370 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
371 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
372 if (runtime_itlb_cache_stats[ctx][cpu].n != 0)
373 print_itlb_cache_misses(cpu, evsel, avg, out);
374 else
375 print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
376 } else if (
377 evsel->attr.type == PERF_TYPE_HW_CACHE &&
378 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
379 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
380 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
381 if (runtime_ll_cache_stats[ctx][cpu].n != 0)
382 print_ll_cache_misses(cpu, evsel, avg, out);
383 else
384 print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
385 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
386 total = avg_stats(&runtime_cacherefs_stats[ctx][cpu]);
387
388 if (total)
389 ratio = avg * 100 / total;
390
391 if (runtime_cacherefs_stats[ctx][cpu].n != 0)
392 print_metric(ctxp, NULL, "%8.3f %%",
393 "of all cache refs", ratio);
394 else
395 print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
396 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
397 print_stalled_cycles_frontend(cpu, evsel, avg, out);
398 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
399 print_stalled_cycles_backend(cpu, evsel, avg, out);
400 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
401 total = avg_stats(&runtime_nsecs_stats[cpu]);
402
403 if (total) {
404 ratio = avg / total;
405 print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
406 } else {
407 print_metric(ctxp, NULL, NULL, "Ghz", 0);
408 }
409 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
410 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
411 if (total)
412 print_metric(ctxp, NULL,
413 "%7.2f%%", "transactional cycles",
414 100.0 * (avg / total));
415 else
416 print_metric(ctxp, NULL, NULL, "transactional cycles",
417 0);
418 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
419 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
420 total2 = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
421 if (total2 < avg)
422 total2 = avg;
423 if (total)
424 print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
425 100.0 * ((total2-avg) / total));
426 else
427 print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
428 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
429 total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
430
431 if (avg)
432 ratio = total / avg;
433
434 if (runtime_cycles_in_tx_stats[ctx][cpu].n != 0)
435 print_metric(ctxp, NULL, "%8.0f",
436 "cycles / transaction", ratio);
437 else
438 print_metric(ctxp, NULL, NULL, "cycles / transaction",
439 0);
440 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
441 total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
442
443 if (avg)
444 ratio = total / avg;
445
446 print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
447 } else if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) {
448 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
449 print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
450 avg / ratio);
451 else
452 print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
453 } else if (runtime_nsecs_stats[cpu].n != 0) {
454 char unit = 'M';
455 char unit_buf[10];
456
457 total = avg_stats(&runtime_nsecs_stats[cpu]);
458
459 if (total)
460 ratio = 1000.0 * avg / total;
461 if (ratio < 0.001) {
462 ratio *= 1000;
463 unit = 'K';
464 }
465 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
466 print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
467 } else {
468 print_metric(ctxp, NULL, NULL, NULL, 0);
469 }
470}
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#include "tool_pmu.h"
19
20struct stats walltime_nsecs_stats;
21struct rusage_stats ru_stats;
22
23enum {
24 CTX_BIT_USER = 1 << 0,
25 CTX_BIT_KERNEL = 1 << 1,
26 CTX_BIT_HV = 1 << 2,
27 CTX_BIT_HOST = 1 << 3,
28 CTX_BIT_IDLE = 1 << 4,
29 CTX_BIT_MAX = 1 << 5,
30};
31
32enum stat_type {
33 STAT_NONE = 0,
34 STAT_NSECS,
35 STAT_CYCLES,
36 STAT_INSTRUCTIONS,
37 STAT_STALLED_CYCLES_FRONT,
38 STAT_STALLED_CYCLES_BACK,
39 STAT_BRANCHES,
40 STAT_BRANCH_MISS,
41 STAT_CACHE_REFS,
42 STAT_CACHE_MISSES,
43 STAT_L1_DCACHE,
44 STAT_L1_ICACHE,
45 STAT_LL_CACHE,
46 STAT_ITLB_CACHE,
47 STAT_DTLB_CACHE,
48 STAT_L1D_MISS,
49 STAT_L1I_MISS,
50 STAT_LL_MISS,
51 STAT_DTLB_MISS,
52 STAT_ITLB_MISS,
53 STAT_MAX
54};
55
56static int evsel_context(const struct evsel *evsel)
57{
58 int ctx = 0;
59
60 if (evsel->core.attr.exclude_kernel)
61 ctx |= CTX_BIT_KERNEL;
62 if (evsel->core.attr.exclude_user)
63 ctx |= CTX_BIT_USER;
64 if (evsel->core.attr.exclude_hv)
65 ctx |= CTX_BIT_HV;
66 if (evsel->core.attr.exclude_host)
67 ctx |= CTX_BIT_HOST;
68 if (evsel->core.attr.exclude_idle)
69 ctx |= CTX_BIT_IDLE;
70
71 return ctx;
72}
73
74void perf_stat__reset_shadow_stats(void)
75{
76 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
77 memset(&ru_stats, 0, sizeof(ru_stats));
78}
79
80static enum stat_type evsel__stat_type(struct evsel *evsel)
81{
82 /* Fake perf_hw_cache_op_id values for use with evsel__match. */
83 u64 PERF_COUNT_hw_cache_l1d_miss = PERF_COUNT_HW_CACHE_L1D |
84 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
85 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
86 u64 PERF_COUNT_hw_cache_l1i_miss = PERF_COUNT_HW_CACHE_L1I |
87 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
88 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
89 u64 PERF_COUNT_hw_cache_ll_miss = PERF_COUNT_HW_CACHE_LL |
90 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
91 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
92 u64 PERF_COUNT_hw_cache_dtlb_miss = PERF_COUNT_HW_CACHE_DTLB |
93 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
94 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
95 u64 PERF_COUNT_hw_cache_itlb_miss = PERF_COUNT_HW_CACHE_ITLB |
96 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
97 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
98
99 if (evsel__is_clock(evsel))
100 return STAT_NSECS;
101 else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES))
102 return STAT_CYCLES;
103 else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS))
104 return STAT_INSTRUCTIONS;
105 else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
106 return STAT_STALLED_CYCLES_FRONT;
107 else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND))
108 return STAT_STALLED_CYCLES_BACK;
109 else if (evsel__match(evsel, HARDWARE, HW_BRANCH_INSTRUCTIONS))
110 return STAT_BRANCHES;
111 else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES))
112 return STAT_BRANCH_MISS;
113 else if (evsel__match(evsel, HARDWARE, HW_CACHE_REFERENCES))
114 return STAT_CACHE_REFS;
115 else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES))
116 return STAT_CACHE_MISSES;
117 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1D))
118 return STAT_L1_DCACHE;
119 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1I))
120 return STAT_L1_ICACHE;
121 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_LL))
122 return STAT_LL_CACHE;
123 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_DTLB))
124 return STAT_DTLB_CACHE;
125 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_ITLB))
126 return STAT_ITLB_CACHE;
127 else if (evsel__match(evsel, HW_CACHE, hw_cache_l1d_miss))
128 return STAT_L1D_MISS;
129 else if (evsel__match(evsel, HW_CACHE, hw_cache_l1i_miss))
130 return STAT_L1I_MISS;
131 else if (evsel__match(evsel, HW_CACHE, hw_cache_ll_miss))
132 return STAT_LL_MISS;
133 else if (evsel__match(evsel, HW_CACHE, hw_cache_dtlb_miss))
134 return STAT_DTLB_MISS;
135 else if (evsel__match(evsel, HW_CACHE, hw_cache_itlb_miss))
136 return STAT_ITLB_MISS;
137 return STAT_NONE;
138}
139
140static enum metric_threshold_classify get_ratio_thresh(const double ratios[3], double val)
141{
142 assert(ratios[0] > ratios[1]);
143 assert(ratios[1] > ratios[2]);
144
145 return val > ratios[1]
146 ? (val > ratios[0] ? METRIC_THRESHOLD_BAD : METRIC_THRESHOLD_NEARLY_BAD)
147 : (val > ratios[2] ? METRIC_THRESHOLD_LESS_GOOD : METRIC_THRESHOLD_GOOD);
148}
149
150static double find_stat(const struct evsel *evsel, int aggr_idx, enum stat_type type)
151{
152 struct evsel *cur;
153 int evsel_ctx = evsel_context(evsel);
154
155 evlist__for_each_entry(evsel->evlist, cur) {
156 struct perf_stat_aggr *aggr;
157
158 /* Ignore the evsel that is being searched from. */
159 if (evsel == cur)
160 continue;
161
162 /* Ignore evsels that are part of different groups. */
163 if (evsel->core.leader->nr_members > 1 &&
164 evsel->core.leader != cur->core.leader)
165 continue;
166 /* Ignore evsels with mismatched modifiers. */
167 if (evsel_ctx != evsel_context(cur))
168 continue;
169 /* Ignore if not the cgroup we're looking for. */
170 if (evsel->cgrp != cur->cgrp)
171 continue;
172 /* Ignore if not the stat we're looking for. */
173 if (type != evsel__stat_type(cur))
174 continue;
175
176 /*
177 * Except the SW CLOCK events,
178 * ignore if not the PMU we're looking for.
179 */
180 if ((type != STAT_NSECS) && (evsel->pmu != cur->pmu))
181 continue;
182
183 aggr = &cur->stats->aggr[aggr_idx];
184 if (type == STAT_NSECS)
185 return aggr->counts.val;
186 return aggr->counts.val * cur->scale;
187 }
188 return 0.0;
189}
190
191static void print_ratio(struct perf_stat_config *config,
192 const struct evsel *evsel, int aggr_idx,
193 double numerator, struct perf_stat_output_ctx *out,
194 enum stat_type denominator_type,
195 const double thresh_ratios[3], const char *_unit)
196{
197 double denominator = find_stat(evsel, aggr_idx, denominator_type);
198 double ratio = 0;
199 enum metric_threshold_classify thresh = METRIC_THRESHOLD_UNKNOWN;
200 const char *fmt = NULL;
201 const char *unit = NULL;
202
203 if (numerator && denominator) {
204 ratio = numerator / denominator * 100.0;
205 thresh = get_ratio_thresh(thresh_ratios, ratio);
206 fmt = "%7.2f%%";
207 unit = _unit;
208 }
209 out->print_metric(config, out->ctx, thresh, fmt, unit, ratio);
210}
211
212static void print_stalled_cycles_front(struct perf_stat_config *config,
213 const struct evsel *evsel,
214 int aggr_idx, double stalled,
215 struct perf_stat_output_ctx *out)
216{
217 const double thresh_ratios[3] = {50.0, 30.0, 10.0};
218
219 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, thresh_ratios,
220 "frontend cycles idle");
221}
222
223static void print_stalled_cycles_back(struct perf_stat_config *config,
224 const struct evsel *evsel,
225 int aggr_idx, double stalled,
226 struct perf_stat_output_ctx *out)
227{
228 const double thresh_ratios[3] = {75.0, 50.0, 20.0};
229
230 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, thresh_ratios,
231 "backend cycles idle");
232}
233
234static void print_branch_miss(struct perf_stat_config *config,
235 const struct evsel *evsel,
236 int aggr_idx, double misses,
237 struct perf_stat_output_ctx *out)
238{
239 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
240
241 print_ratio(config, evsel, aggr_idx, misses, out, STAT_BRANCHES, thresh_ratios,
242 "of all branches");
243}
244
245static void print_l1d_miss(struct perf_stat_config *config,
246 const struct evsel *evsel,
247 int aggr_idx, double misses,
248 struct perf_stat_output_ctx *out)
249{
250 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
251
252 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_DCACHE, thresh_ratios,
253 "of all L1-dcache accesses");
254}
255
256static void print_l1i_miss(struct perf_stat_config *config,
257 const struct evsel *evsel,
258 int aggr_idx, double misses,
259 struct perf_stat_output_ctx *out)
260{
261 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
262
263 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_ICACHE, thresh_ratios,
264 "of all L1-icache accesses");
265}
266
267static void print_ll_miss(struct perf_stat_config *config,
268 const struct evsel *evsel,
269 int aggr_idx, double misses,
270 struct perf_stat_output_ctx *out)
271{
272 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
273
274 print_ratio(config, evsel, aggr_idx, misses, out, STAT_LL_CACHE, thresh_ratios,
275 "of all LL-cache accesses");
276}
277
278static void print_dtlb_miss(struct perf_stat_config *config,
279 const struct evsel *evsel,
280 int aggr_idx, double misses,
281 struct perf_stat_output_ctx *out)
282{
283 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
284
285 print_ratio(config, evsel, aggr_idx, misses, out, STAT_DTLB_CACHE, thresh_ratios,
286 "of all dTLB cache accesses");
287}
288
289static void print_itlb_miss(struct perf_stat_config *config,
290 const struct evsel *evsel,
291 int aggr_idx, double misses,
292 struct perf_stat_output_ctx *out)
293{
294 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
295
296 print_ratio(config, evsel, aggr_idx, misses, out, STAT_ITLB_CACHE, thresh_ratios,
297 "of all iTLB cache accesses");
298}
299
300static void print_cache_miss(struct perf_stat_config *config,
301 const struct evsel *evsel,
302 int aggr_idx, double misses,
303 struct perf_stat_output_ctx *out)
304{
305 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
306
307 print_ratio(config, evsel, aggr_idx, misses, out, STAT_CACHE_REFS, thresh_ratios,
308 "of all cache refs");
309}
310
311static void print_instructions(struct perf_stat_config *config,
312 const struct evsel *evsel,
313 int aggr_idx, double instructions,
314 struct perf_stat_output_ctx *out)
315{
316 print_metric_t print_metric = out->print_metric;
317 void *ctxp = out->ctx;
318 double cycles = find_stat(evsel, aggr_idx, STAT_CYCLES);
319 double max_stalled = max(find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_FRONT),
320 find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_BACK));
321
322 if (cycles) {
323 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%7.2f ",
324 "insn per cycle", instructions / cycles);
325 } else {
326 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
327 "insn per cycle", 0);
328 }
329 if (max_stalled && instructions) {
330 out->new_line(config, ctxp);
331 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%7.2f ",
332 "stalled cycles per insn", max_stalled / instructions);
333 }
334}
335
336static void print_cycles(struct perf_stat_config *config,
337 const struct evsel *evsel,
338 int aggr_idx, double cycles,
339 struct perf_stat_output_ctx *out)
340{
341 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
342
343 if (cycles && nsecs) {
344 double ratio = cycles / nsecs;
345
346 out->print_metric(config, out->ctx, METRIC_THRESHOLD_UNKNOWN, "%8.3f",
347 "GHz", ratio);
348 } else {
349 out->print_metric(config, out->ctx, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
350 "GHz", 0);
351 }
352}
353
354static void print_nsecs(struct perf_stat_config *config,
355 const struct evsel *evsel,
356 int aggr_idx __maybe_unused, double nsecs,
357 struct perf_stat_output_ctx *out)
358{
359 print_metric_t print_metric = out->print_metric;
360 void *ctxp = out->ctx;
361 double wall_time = avg_stats(&walltime_nsecs_stats);
362
363 if (wall_time) {
364 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%8.3f", "CPUs utilized",
365 nsecs / (wall_time * evsel->scale));
366 } else {
367 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
368 "CPUs utilized", 0);
369 }
370}
371
372static int prepare_metric(const struct metric_expr *mexp,
373 const struct evsel *evsel,
374 struct expr_parse_ctx *pctx,
375 int aggr_idx)
376{
377 struct evsel * const *metric_events = mexp->metric_events;
378 struct metric_ref *metric_refs = mexp->metric_refs;
379 int i;
380
381 for (i = 0; metric_events[i]; i++) {
382 char *n;
383 double val;
384 int source_count = 0;
385
386 if (evsel__is_tool(metric_events[i])) {
387 struct stats *stats;
388 double scale;
389
390 switch (evsel__tool_event(metric_events[i])) {
391 case TOOL_PMU__EVENT_DURATION_TIME:
392 stats = &walltime_nsecs_stats;
393 scale = 1e-9;
394 break;
395 case TOOL_PMU__EVENT_USER_TIME:
396 stats = &ru_stats.ru_utime_usec_stat;
397 scale = 1e-6;
398 break;
399 case TOOL_PMU__EVENT_SYSTEM_TIME:
400 stats = &ru_stats.ru_stime_usec_stat;
401 scale = 1e-6;
402 break;
403 case TOOL_PMU__EVENT_NONE:
404 pr_err("Invalid tool event 'none'");
405 abort();
406 case TOOL_PMU__EVENT_MAX:
407 pr_err("Invalid tool event 'max'");
408 abort();
409 case TOOL_PMU__EVENT_HAS_PMEM:
410 case TOOL_PMU__EVENT_NUM_CORES:
411 case TOOL_PMU__EVENT_NUM_CPUS:
412 case TOOL_PMU__EVENT_NUM_CPUS_ONLINE:
413 case TOOL_PMU__EVENT_NUM_DIES:
414 case TOOL_PMU__EVENT_NUM_PACKAGES:
415 case TOOL_PMU__EVENT_SLOTS:
416 case TOOL_PMU__EVENT_SMT_ON:
417 case TOOL_PMU__EVENT_SYSTEM_TSC_FREQ:
418 default:
419 pr_err("Unexpected tool event '%s'", evsel__name(metric_events[i]));
420 abort();
421 }
422 val = avg_stats(stats) * scale;
423 source_count = 1;
424 } else {
425 struct perf_stat_evsel *ps = metric_events[i]->stats;
426 struct perf_stat_aggr *aggr;
427
428 /*
429 * If there are multiple uncore PMUs and we're not
430 * reading the leader's stats, determine the stats for
431 * the appropriate uncore PMU.
432 */
433 if (evsel && evsel->metric_leader &&
434 evsel->pmu != evsel->metric_leader->pmu &&
435 mexp->metric_events[i]->pmu == evsel->metric_leader->pmu) {
436 struct evsel *pos;
437
438 evlist__for_each_entry(evsel->evlist, pos) {
439 if (pos->pmu != evsel->pmu)
440 continue;
441 if (pos->metric_leader != mexp->metric_events[i])
442 continue;
443 ps = pos->stats;
444 source_count = 1;
445 break;
446 }
447 }
448 aggr = &ps->aggr[aggr_idx];
449 if (!aggr)
450 break;
451
452 if (!metric_events[i]->supported) {
453 /*
454 * Not supported events will have a count of 0,
455 * which can be confusing in a
456 * metric. Explicitly set the value to NAN. Not
457 * counted events (enable time of 0) are read as
458 * 0.
459 */
460 val = NAN;
461 source_count = 0;
462 } else {
463 val = aggr->counts.val;
464 if (!source_count)
465 source_count = evsel__source_count(metric_events[i]);
466 }
467 }
468 n = strdup(evsel__metric_id(metric_events[i]));
469 if (!n)
470 return -ENOMEM;
471
472 expr__add_id_val_source_count(pctx, n, val, source_count);
473 }
474
475 for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
476 int ret = expr__add_ref(pctx, &metric_refs[j]);
477
478 if (ret)
479 return ret;
480 }
481
482 return i;
483}
484
485static void generic_metric(struct perf_stat_config *config,
486 struct metric_expr *mexp,
487 struct evsel *evsel,
488 int aggr_idx,
489 struct perf_stat_output_ctx *out)
490{
491 print_metric_t print_metric = out->print_metric;
492 const char *metric_name = mexp->metric_name;
493 const char *metric_expr = mexp->metric_expr;
494 const char *metric_threshold = mexp->metric_threshold;
495 const char *metric_unit = mexp->metric_unit;
496 struct evsel * const *metric_events = mexp->metric_events;
497 int runtime = mexp->runtime;
498 struct expr_parse_ctx *pctx;
499 double ratio, scale, threshold;
500 int i;
501 void *ctxp = out->ctx;
502 enum metric_threshold_classify thresh = METRIC_THRESHOLD_UNKNOWN;
503
504 pctx = expr__ctx_new();
505 if (!pctx)
506 return;
507
508 if (config->user_requested_cpu_list)
509 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
510 pctx->sctx.runtime = runtime;
511 pctx->sctx.system_wide = config->system_wide;
512 i = prepare_metric(mexp, evsel, pctx, aggr_idx);
513 if (i < 0) {
514 expr__ctx_free(pctx);
515 return;
516 }
517 if (!metric_events[i]) {
518 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
519 char *unit;
520 char metric_bf[128];
521
522 if (metric_threshold &&
523 expr__parse(&threshold, pctx, metric_threshold) == 0 &&
524 !isnan(threshold)) {
525 thresh = fpclassify(threshold) == FP_ZERO
526 ? METRIC_THRESHOLD_GOOD : METRIC_THRESHOLD_BAD;
527 }
528
529 if (metric_unit && metric_name) {
530 if (perf_pmu__convert_scale(metric_unit,
531 &unit, &scale) >= 0) {
532 ratio *= scale;
533 }
534 if (strstr(metric_expr, "?"))
535 scnprintf(metric_bf, sizeof(metric_bf),
536 "%s %s_%d", unit, metric_name, runtime);
537 else
538 scnprintf(metric_bf, sizeof(metric_bf),
539 "%s %s", unit, metric_name);
540
541 print_metric(config, ctxp, thresh, "%8.1f",
542 metric_bf, ratio);
543 } else {
544 print_metric(config, ctxp, thresh, "%8.2f",
545 metric_name ?
546 metric_name :
547 out->force_header ? evsel->name : "",
548 ratio);
549 }
550 } else {
551 print_metric(config, ctxp, thresh, /*fmt=*/NULL,
552 out->force_header ?
553 (metric_name ?: evsel->name) : "", 0);
554 }
555 } else {
556 print_metric(config, ctxp, thresh, /*fmt=*/NULL,
557 out->force_header ?
558 (metric_name ?: evsel->name) : "", 0);
559 }
560
561 expr__ctx_free(pctx);
562}
563
564double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
565{
566 struct expr_parse_ctx *pctx;
567 double ratio = 0.0;
568
569 pctx = expr__ctx_new();
570 if (!pctx)
571 return NAN;
572
573 if (prepare_metric(mexp, /*evsel=*/NULL, pctx, aggr_idx) < 0)
574 goto out;
575
576 if (expr__parse(&ratio, pctx, mexp->metric_expr))
577 ratio = 0.0;
578
579out:
580 expr__ctx_free(pctx);
581 return ratio;
582}
583
584static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,
585 struct evsel *evsel,
586 void *ctxp,
587 const char *name,
588 struct perf_stat_output_ctx *out)
589{
590 bool need_full_name = perf_pmus__num_core_pmus() > 1;
591 static const char *last_name;
592 static const struct perf_pmu *last_pmu;
593 char full_name[64];
594
595 /*
596 * A metricgroup may have several metric events,
597 * e.g.,TopdownL1 on e-core of ADL.
598 * The name has been output by the first metric
599 * event. Only align with other metics from
600 * different metric events.
601 */
602 if (last_name && !strcmp(last_name, name)) {
603 if (!need_full_name || last_pmu != evsel->pmu) {
604 out->print_metricgroup_header(config, ctxp, NULL);
605 return;
606 }
607 }
608
609 if (need_full_name && evsel->pmu)
610 scnprintf(full_name, sizeof(full_name), "%s (%s)", name, evsel->pmu->name);
611 else
612 scnprintf(full_name, sizeof(full_name), "%s", name);
613
614 out->print_metricgroup_header(config, ctxp, full_name);
615
616 last_name = name;
617 last_pmu = evsel->pmu;
618}
619
620/**
621 * perf_stat__print_shadow_stats_metricgroup - Print out metrics associated with the evsel
622 * For the non-default, all metrics associated
623 * with the evsel are printed.
624 * For the default mode, only the metrics from
625 * the same metricgroup and the name of the
626 * metricgroup are printed. To print the metrics
627 * from the next metricgroup (if available),
628 * invoke the function with correspoinding
629 * metric_expr.
630 */
631void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
632 struct evsel *evsel,
633 int aggr_idx,
634 int *num,
635 void *from,
636 struct perf_stat_output_ctx *out,
637 struct rblist *metric_events)
638{
639 struct metric_event *me;
640 struct metric_expr *mexp = from;
641 void *ctxp = out->ctx;
642 bool header_printed = false;
643 const char *name = NULL;
644
645 me = metricgroup__lookup(metric_events, evsel, false);
646 if (me == NULL)
647 return NULL;
648
649 if (!mexp)
650 mexp = list_first_entry(&me->head, typeof(*mexp), nd);
651
652 list_for_each_entry_from(mexp, &me->head, nd) {
653 /* Print the display name of the Default metricgroup */
654 if (!config->metric_only && me->is_default) {
655 if (!name)
656 name = mexp->default_metricgroup_name;
657 /*
658 * Two or more metricgroup may share the same metric
659 * event, e.g., TopdownL1 and TopdownL2 on SPR.
660 * Return and print the prefix, e.g., noise, running
661 * for the next metricgroup.
662 */
663 if (strcmp(name, mexp->default_metricgroup_name))
664 return (void *)mexp;
665 /* Only print the name of the metricgroup once */
666 if (!header_printed) {
667 header_printed = true;
668 perf_stat__print_metricgroup_header(config, evsel, ctxp,
669 name, out);
670 }
671 }
672
673 if ((*num)++ > 0)
674 out->new_line(config, ctxp);
675 generic_metric(config, mexp, evsel, aggr_idx, out);
676 }
677
678 return NULL;
679}
680
681void perf_stat__print_shadow_stats(struct perf_stat_config *config,
682 struct evsel *evsel,
683 double avg, int aggr_idx,
684 struct perf_stat_output_ctx *out,
685 struct rblist *metric_events)
686{
687 typedef void (*stat_print_function_t)(struct perf_stat_config *config,
688 const struct evsel *evsel,
689 int aggr_idx, double misses,
690 struct perf_stat_output_ctx *out);
691 static const stat_print_function_t stat_print_function[STAT_MAX] = {
692 [STAT_INSTRUCTIONS] = print_instructions,
693 [STAT_BRANCH_MISS] = print_branch_miss,
694 [STAT_L1D_MISS] = print_l1d_miss,
695 [STAT_L1I_MISS] = print_l1i_miss,
696 [STAT_DTLB_MISS] = print_dtlb_miss,
697 [STAT_ITLB_MISS] = print_itlb_miss,
698 [STAT_LL_MISS] = print_ll_miss,
699 [STAT_CACHE_MISSES] = print_cache_miss,
700 [STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
701 [STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
702 [STAT_CYCLES] = print_cycles,
703 [STAT_NSECS] = print_nsecs,
704 };
705 print_metric_t print_metric = out->print_metric;
706 void *ctxp = out->ctx;
707 int num = 1;
708
709 if (config->iostat_run) {
710 iostat_print_metric(config, evsel, out);
711 } else {
712 stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
713
714 if (fn)
715 fn(config, evsel, aggr_idx, avg, out);
716 else {
717 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
718
719 if (nsecs) {
720 char unit = ' ';
721 char unit_buf[10] = "/sec";
722 double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
723 &unit);
724
725 if (unit != ' ')
726 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
727 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%8.3f",
728 unit_buf, ratio);
729 } else {
730 num = 0;
731 }
732 }
733 }
734
735 perf_stat__print_shadow_stats_metricgroup(config, evsel, aggr_idx,
736 &num, NULL, out, metric_events);
737
738 if (num == 0) {
739 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN,
740 /*fmt=*/NULL, /*unit=*/NULL, 0);
741 }
742}
743
744/**
745 * perf_stat__skip_metric_event - Skip the evsel in the Default metricgroup,
746 * if it's not running or not the metric event.
747 */
748bool perf_stat__skip_metric_event(struct evsel *evsel,
749 struct rblist *metric_events,
750 u64 ena, u64 run)
751{
752 if (!evsel->default_metricgroup)
753 return false;
754
755 if (!ena || !run)
756 return true;
757
758 return !metricgroup__lookup(metric_events, evsel, false);
759}