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
19struct stats walltime_nsecs_stats;
20struct rusage_stats ru_stats;
21
22enum {
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
31enum 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
55static 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
73void 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
79static 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
139static 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
153static 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 aggr = &cur->stats->aggr[aggr_idx];
180 if (type == STAT_NSECS)
181 return aggr->counts.val;
182 return aggr->counts.val * cur->scale;
183 }
184 return 0.0;
185}
186
187static void print_ratio(struct perf_stat_config *config,
188 const struct evsel *evsel, int aggr_idx,
189 double numerator, struct perf_stat_output_ctx *out,
190 enum stat_type denominator_type,
191 const double color_ratios[3], const char *unit)
192{
193 double denominator = find_stat(evsel, aggr_idx, denominator_type);
194
195 if (numerator && denominator) {
196 double ratio = numerator / denominator * 100.0;
197 const char *color = get_ratio_color(color_ratios, ratio);
198
199 out->print_metric(config, out->ctx, color, "%7.2f%%", unit, ratio);
200 } else
201 out->print_metric(config, out->ctx, NULL, NULL, unit, 0);
202}
203
204static void print_stalled_cycles_front(struct perf_stat_config *config,
205 const struct evsel *evsel,
206 int aggr_idx, double stalled,
207 struct perf_stat_output_ctx *out)
208{
209 static const double color_ratios[3] = {50.0, 30.0, 10.0};
210
211 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, color_ratios,
212 "frontend cycles idle");
213}
214
215static void print_stalled_cycles_back(struct perf_stat_config *config,
216 const struct evsel *evsel,
217 int aggr_idx, double stalled,
218 struct perf_stat_output_ctx *out)
219{
220 static const double color_ratios[3] = {75.0, 50.0, 20.0};
221
222 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, color_ratios,
223 "backend cycles idle");
224}
225
226static void print_branch_miss(struct perf_stat_config *config,
227 const struct evsel *evsel,
228 int aggr_idx, double misses,
229 struct perf_stat_output_ctx *out)
230{
231 static const double color_ratios[3] = {20.0, 10.0, 5.0};
232
233 print_ratio(config, evsel, aggr_idx, misses, out, STAT_BRANCHES, color_ratios,
234 "of all branches");
235}
236
237static void print_l1d_miss(struct perf_stat_config *config,
238 const struct evsel *evsel,
239 int aggr_idx, double misses,
240 struct perf_stat_output_ctx *out)
241{
242 static const double color_ratios[3] = {20.0, 10.0, 5.0};
243
244 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_DCACHE, color_ratios,
245 "of all L1-dcache accesses");
246}
247
248static void print_l1i_miss(struct perf_stat_config *config,
249 const struct evsel *evsel,
250 int aggr_idx, double misses,
251 struct perf_stat_output_ctx *out)
252{
253 static const double color_ratios[3] = {20.0, 10.0, 5.0};
254
255 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_ICACHE, color_ratios,
256 "of all L1-icache accesses");
257}
258
259static void print_ll_miss(struct perf_stat_config *config,
260 const struct evsel *evsel,
261 int aggr_idx, double misses,
262 struct perf_stat_output_ctx *out)
263{
264 static const double color_ratios[3] = {20.0, 10.0, 5.0};
265
266 print_ratio(config, evsel, aggr_idx, misses, out, STAT_LL_CACHE, color_ratios,
267 "of all LL-cache accesses");
268}
269
270static void print_dtlb_miss(struct perf_stat_config *config,
271 const struct evsel *evsel,
272 int aggr_idx, double misses,
273 struct perf_stat_output_ctx *out)
274{
275 static const double color_ratios[3] = {20.0, 10.0, 5.0};
276
277 print_ratio(config, evsel, aggr_idx, misses, out, STAT_DTLB_CACHE, color_ratios,
278 "of all dTLB cache accesses");
279}
280
281static void print_itlb_miss(struct perf_stat_config *config,
282 const struct evsel *evsel,
283 int aggr_idx, double misses,
284 struct perf_stat_output_ctx *out)
285{
286 static const double color_ratios[3] = {20.0, 10.0, 5.0};
287
288 print_ratio(config, evsel, aggr_idx, misses, out, STAT_ITLB_CACHE, color_ratios,
289 "of all iTLB cache accesses");
290}
291
292static void print_cache_miss(struct perf_stat_config *config,
293 const struct evsel *evsel,
294 int aggr_idx, double misses,
295 struct perf_stat_output_ctx *out)
296{
297 static const double color_ratios[3] = {20.0, 10.0, 5.0};
298
299 print_ratio(config, evsel, aggr_idx, misses, out, STAT_CACHE_REFS, color_ratios,
300 "of all cache refs");
301}
302
303static void print_instructions(struct perf_stat_config *config,
304 const struct evsel *evsel,
305 int aggr_idx, double instructions,
306 struct perf_stat_output_ctx *out)
307{
308 print_metric_t print_metric = out->print_metric;
309 void *ctxp = out->ctx;
310 double cycles = find_stat(evsel, aggr_idx, STAT_CYCLES);
311 double max_stalled = max(find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_FRONT),
312 find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_BACK));
313
314 if (cycles) {
315 print_metric(config, ctxp, NULL, "%7.2f ", "insn per cycle",
316 instructions / cycles);
317 } else
318 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
319
320 if (max_stalled && instructions) {
321 out->new_line(config, ctxp);
322 print_metric(config, ctxp, NULL, "%7.2f ", "stalled cycles per insn",
323 max_stalled / instructions);
324 }
325}
326
327static void print_cycles(struct perf_stat_config *config,
328 const struct evsel *evsel,
329 int aggr_idx, double cycles,
330 struct perf_stat_output_ctx *out)
331{
332 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
333
334 if (cycles && nsecs) {
335 double ratio = cycles / nsecs;
336
337 out->print_metric(config, out->ctx, NULL, "%8.3f", "GHz", ratio);
338 } else
339 out->print_metric(config, out->ctx, NULL, NULL, "GHz", 0);
340}
341
342static void print_nsecs(struct perf_stat_config *config,
343 const struct evsel *evsel,
344 int aggr_idx __maybe_unused, double nsecs,
345 struct perf_stat_output_ctx *out)
346{
347 print_metric_t print_metric = out->print_metric;
348 void *ctxp = out->ctx;
349 double wall_time = avg_stats(&walltime_nsecs_stats);
350
351 if (wall_time) {
352 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
353 nsecs / (wall_time * evsel->scale));
354 } else
355 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
356}
357
358static int prepare_metric(const struct metric_expr *mexp,
359 const struct evsel *evsel,
360 struct expr_parse_ctx *pctx,
361 int aggr_idx)
362{
363 struct evsel * const *metric_events = mexp->metric_events;
364 struct metric_ref *metric_refs = mexp->metric_refs;
365 int i;
366
367 for (i = 0; metric_events[i]; i++) {
368 char *n;
369 double val;
370 int source_count = 0;
371
372 if (evsel__is_tool(metric_events[i])) {
373 struct stats *stats;
374 double scale;
375
376 switch (metric_events[i]->tool_event) {
377 case PERF_TOOL_DURATION_TIME:
378 stats = &walltime_nsecs_stats;
379 scale = 1e-9;
380 break;
381 case PERF_TOOL_USER_TIME:
382 stats = &ru_stats.ru_utime_usec_stat;
383 scale = 1e-6;
384 break;
385 case PERF_TOOL_SYSTEM_TIME:
386 stats = &ru_stats.ru_stime_usec_stat;
387 scale = 1e-6;
388 break;
389 case PERF_TOOL_NONE:
390 pr_err("Invalid tool event 'none'");
391 abort();
392 case PERF_TOOL_MAX:
393 pr_err("Invalid tool event 'max'");
394 abort();
395 default:
396 pr_err("Unknown tool event '%s'", evsel__name(metric_events[i]));
397 abort();
398 }
399 val = avg_stats(stats) * scale;
400 source_count = 1;
401 } else {
402 struct perf_stat_evsel *ps = metric_events[i]->stats;
403 struct perf_stat_aggr *aggr;
404
405 /*
406 * If there are multiple uncore PMUs and we're not
407 * reading the leader's stats, determine the stats for
408 * the appropriate uncore PMU.
409 */
410 if (evsel && evsel->metric_leader &&
411 evsel->pmu != evsel->metric_leader->pmu &&
412 mexp->metric_events[i]->pmu == evsel->metric_leader->pmu) {
413 struct evsel *pos;
414
415 evlist__for_each_entry(evsel->evlist, pos) {
416 if (pos->pmu != evsel->pmu)
417 continue;
418 if (pos->metric_leader != mexp->metric_events[i])
419 continue;
420 ps = pos->stats;
421 source_count = 1;
422 break;
423 }
424 }
425 aggr = &ps->aggr[aggr_idx];
426 if (!aggr)
427 break;
428
429 if (!metric_events[i]->supported) {
430 /*
431 * Not supported events will have a count of 0,
432 * which can be confusing in a
433 * metric. Explicitly set the value to NAN. Not
434 * counted events (enable time of 0) are read as
435 * 0.
436 */
437 val = NAN;
438 source_count = 0;
439 } else {
440 val = aggr->counts.val;
441 if (!source_count)
442 source_count = evsel__source_count(metric_events[i]);
443 }
444 }
445 n = strdup(evsel__metric_id(metric_events[i]));
446 if (!n)
447 return -ENOMEM;
448
449 expr__add_id_val_source_count(pctx, n, val, source_count);
450 }
451
452 for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
453 int ret = expr__add_ref(pctx, &metric_refs[j]);
454
455 if (ret)
456 return ret;
457 }
458
459 return i;
460}
461
462static void generic_metric(struct perf_stat_config *config,
463 struct metric_expr *mexp,
464 struct evsel *evsel,
465 int aggr_idx,
466 struct perf_stat_output_ctx *out)
467{
468 print_metric_t print_metric = out->print_metric;
469 const char *metric_name = mexp->metric_name;
470 const char *metric_expr = mexp->metric_expr;
471 const char *metric_threshold = mexp->metric_threshold;
472 const char *metric_unit = mexp->metric_unit;
473 struct evsel * const *metric_events = mexp->metric_events;
474 int runtime = mexp->runtime;
475 struct expr_parse_ctx *pctx;
476 double ratio, scale, threshold;
477 int i;
478 void *ctxp = out->ctx;
479 const char *color = NULL;
480
481 pctx = expr__ctx_new();
482 if (!pctx)
483 return;
484
485 if (config->user_requested_cpu_list)
486 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
487 pctx->sctx.runtime = runtime;
488 pctx->sctx.system_wide = config->system_wide;
489 i = prepare_metric(mexp, evsel, pctx, aggr_idx);
490 if (i < 0) {
491 expr__ctx_free(pctx);
492 return;
493 }
494 if (!metric_events[i]) {
495 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
496 char *unit;
497 char metric_bf[64];
498
499 if (metric_threshold &&
500 expr__parse(&threshold, pctx, metric_threshold) == 0 &&
501 !isnan(threshold)) {
502 color = fpclassify(threshold) == FP_ZERO
503 ? PERF_COLOR_GREEN : PERF_COLOR_RED;
504 }
505
506 if (metric_unit && metric_name) {
507 if (perf_pmu__convert_scale(metric_unit,
508 &unit, &scale) >= 0) {
509 ratio *= scale;
510 }
511 if (strstr(metric_expr, "?"))
512 scnprintf(metric_bf, sizeof(metric_bf),
513 "%s %s_%d", unit, metric_name, runtime);
514 else
515 scnprintf(metric_bf, sizeof(metric_bf),
516 "%s %s", unit, metric_name);
517
518 print_metric(config, ctxp, color, "%8.1f",
519 metric_bf, ratio);
520 } else {
521 print_metric(config, ctxp, color, "%8.2f",
522 metric_name ?
523 metric_name :
524 out->force_header ? evsel->name : "",
525 ratio);
526 }
527 } else {
528 print_metric(config, ctxp, color, /*unit=*/NULL,
529 out->force_header ?
530 (metric_name ?: evsel->name) : "", 0);
531 }
532 } else {
533 print_metric(config, ctxp, color, /*unit=*/NULL,
534 out->force_header ?
535 (metric_name ?: evsel->name) : "", 0);
536 }
537
538 expr__ctx_free(pctx);
539}
540
541double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
542{
543 struct expr_parse_ctx *pctx;
544 double ratio = 0.0;
545
546 pctx = expr__ctx_new();
547 if (!pctx)
548 return NAN;
549
550 if (prepare_metric(mexp, /*evsel=*/NULL, pctx, aggr_idx) < 0)
551 goto out;
552
553 if (expr__parse(&ratio, pctx, mexp->metric_expr))
554 ratio = 0.0;
555
556out:
557 expr__ctx_free(pctx);
558 return ratio;
559}
560
561static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,
562 struct evsel *evsel,
563 void *ctxp,
564 const char *name,
565 struct perf_stat_output_ctx *out)
566{
567 bool need_full_name = perf_pmus__num_core_pmus() > 1;
568 static const char *last_name;
569 static const char *last_pmu;
570 char full_name[64];
571
572 /*
573 * A metricgroup may have several metric events,
574 * e.g.,TopdownL1 on e-core of ADL.
575 * The name has been output by the first metric
576 * event. Only align with other metics from
577 * different metric events.
578 */
579 if (last_name && !strcmp(last_name, name)) {
580 if (!need_full_name || !strcmp(last_pmu, evsel->pmu_name)) {
581 out->print_metricgroup_header(config, ctxp, NULL);
582 return;
583 }
584 }
585
586 if (need_full_name)
587 scnprintf(full_name, sizeof(full_name), "%s (%s)", name, evsel->pmu_name);
588 else
589 scnprintf(full_name, sizeof(full_name), "%s", name);
590
591 out->print_metricgroup_header(config, ctxp, full_name);
592
593 last_name = name;
594 last_pmu = evsel->pmu_name;
595}
596
597/**
598 * perf_stat__print_shadow_stats_metricgroup - Print out metrics associated with the evsel
599 * For the non-default, all metrics associated
600 * with the evsel are printed.
601 * For the default mode, only the metrics from
602 * the same metricgroup and the name of the
603 * metricgroup are printed. To print the metrics
604 * from the next metricgroup (if available),
605 * invoke the function with correspoinding
606 * metric_expr.
607 */
608void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
609 struct evsel *evsel,
610 int aggr_idx,
611 int *num,
612 void *from,
613 struct perf_stat_output_ctx *out,
614 struct rblist *metric_events)
615{
616 struct metric_event *me;
617 struct metric_expr *mexp = from;
618 void *ctxp = out->ctx;
619 bool header_printed = false;
620 const char *name = NULL;
621
622 me = metricgroup__lookup(metric_events, evsel, false);
623 if (me == NULL)
624 return NULL;
625
626 if (!mexp)
627 mexp = list_first_entry(&me->head, typeof(*mexp), nd);
628
629 list_for_each_entry_from(mexp, &me->head, nd) {
630 /* Print the display name of the Default metricgroup */
631 if (!config->metric_only && me->is_default) {
632 if (!name)
633 name = mexp->default_metricgroup_name;
634 /*
635 * Two or more metricgroup may share the same metric
636 * event, e.g., TopdownL1 and TopdownL2 on SPR.
637 * Return and print the prefix, e.g., noise, running
638 * for the next metricgroup.
639 */
640 if (strcmp(name, mexp->default_metricgroup_name))
641 return (void *)mexp;
642 /* Only print the name of the metricgroup once */
643 if (!header_printed) {
644 header_printed = true;
645 perf_stat__print_metricgroup_header(config, evsel, ctxp,
646 name, out);
647 }
648 }
649
650 if ((*num)++ > 0)
651 out->new_line(config, ctxp);
652 generic_metric(config, mexp, evsel, aggr_idx, out);
653 }
654
655 return NULL;
656}
657
658void perf_stat__print_shadow_stats(struct perf_stat_config *config,
659 struct evsel *evsel,
660 double avg, int aggr_idx,
661 struct perf_stat_output_ctx *out,
662 struct rblist *metric_events)
663{
664 typedef void (*stat_print_function_t)(struct perf_stat_config *config,
665 const struct evsel *evsel,
666 int aggr_idx, double misses,
667 struct perf_stat_output_ctx *out);
668 static const stat_print_function_t stat_print_function[STAT_MAX] = {
669 [STAT_INSTRUCTIONS] = print_instructions,
670 [STAT_BRANCH_MISS] = print_branch_miss,
671 [STAT_L1D_MISS] = print_l1d_miss,
672 [STAT_L1I_MISS] = print_l1i_miss,
673 [STAT_DTLB_MISS] = print_dtlb_miss,
674 [STAT_ITLB_MISS] = print_itlb_miss,
675 [STAT_LL_MISS] = print_ll_miss,
676 [STAT_CACHE_MISSES] = print_cache_miss,
677 [STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
678 [STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
679 [STAT_CYCLES] = print_cycles,
680 [STAT_NSECS] = print_nsecs,
681 };
682 print_metric_t print_metric = out->print_metric;
683 void *ctxp = out->ctx;
684 int num = 1;
685
686 if (config->iostat_run) {
687 iostat_print_metric(config, evsel, out);
688 } else {
689 stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
690
691 if (fn)
692 fn(config, evsel, aggr_idx, avg, out);
693 else {
694 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
695
696 if (nsecs) {
697 char unit = ' ';
698 char unit_buf[10] = "/sec";
699 double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
700 &unit);
701
702 if (unit != ' ')
703 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
704 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
705 } else
706 num = 0;
707 }
708 }
709
710 perf_stat__print_shadow_stats_metricgroup(config, evsel, aggr_idx,
711 &num, NULL, out, metric_events);
712
713 if (num == 0)
714 print_metric(config, ctxp, NULL, NULL, NULL, 0);
715}
716
717/**
718 * perf_stat__skip_metric_event - Skip the evsel in the Default metricgroup,
719 * if it's not running or not the metric event.
720 */
721bool perf_stat__skip_metric_event(struct evsel *evsel,
722 struct rblist *metric_events,
723 u64 ena, u64 run)
724{
725 if (!evsel->default_metricgroup)
726 return false;
727
728 if (!ena || !run)
729 return true;
730
731 return !metricgroup__lookup(metric_events, evsel, false);
732}