Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
tjh.dev
kernel
os
linux
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 L1-icache 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(struct evsel **metric_events,
359 struct metric_ref *metric_refs,
360 struct expr_parse_ctx *pctx,
361 int aggr_idx)
362{
363 int i;
364
365 for (i = 0; metric_events[i]; i++) {
366 char *n;
367 double val;
368 int source_count = 0;
369
370 if (evsel__is_tool(metric_events[i])) {
371 struct stats *stats;
372 double scale;
373
374 switch (metric_events[i]->tool_event) {
375 case PERF_TOOL_DURATION_TIME:
376 stats = &walltime_nsecs_stats;
377 scale = 1e-9;
378 break;
379 case PERF_TOOL_USER_TIME:
380 stats = &ru_stats.ru_utime_usec_stat;
381 scale = 1e-6;
382 break;
383 case PERF_TOOL_SYSTEM_TIME:
384 stats = &ru_stats.ru_stime_usec_stat;
385 scale = 1e-6;
386 break;
387 case PERF_TOOL_NONE:
388 pr_err("Invalid tool event 'none'");
389 abort();
390 case PERF_TOOL_MAX:
391 pr_err("Invalid tool event 'max'");
392 abort();
393 default:
394 pr_err("Unknown tool event '%s'", evsel__name(metric_events[i]));
395 abort();
396 }
397 val = avg_stats(stats) * scale;
398 source_count = 1;
399 } else {
400 struct perf_stat_evsel *ps = metric_events[i]->stats;
401 struct perf_stat_aggr *aggr = &ps->aggr[aggr_idx];
402
403 if (!aggr)
404 break;
405
406 /*
407 * If an event was scaled during stat gathering, reverse
408 * the scale before computing the metric.
409 */
410 val = aggr->counts.val * (1.0 / metric_events[i]->scale);
411 source_count = evsel__source_count(metric_events[i]);
412 }
413 n = strdup(evsel__metric_id(metric_events[i]));
414 if (!n)
415 return -ENOMEM;
416
417 expr__add_id_val_source_count(pctx, n, val, source_count);
418 }
419
420 for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
421 int ret = expr__add_ref(pctx, &metric_refs[j]);
422
423 if (ret)
424 return ret;
425 }
426
427 return i;
428}
429
430static void generic_metric(struct perf_stat_config *config,
431 const char *metric_expr,
432 const char *metric_threshold,
433 struct evsel **metric_events,
434 struct metric_ref *metric_refs,
435 char *name,
436 const char *metric_name,
437 const char *metric_unit,
438 int runtime,
439 int aggr_idx,
440 struct perf_stat_output_ctx *out)
441{
442 print_metric_t print_metric = out->print_metric;
443 struct expr_parse_ctx *pctx;
444 double ratio, scale, threshold;
445 int i;
446 void *ctxp = out->ctx;
447 const char *color = NULL;
448
449 pctx = expr__ctx_new();
450 if (!pctx)
451 return;
452
453 if (config->user_requested_cpu_list)
454 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
455 pctx->sctx.runtime = runtime;
456 pctx->sctx.system_wide = config->system_wide;
457 i = prepare_metric(metric_events, metric_refs, pctx, aggr_idx);
458 if (i < 0) {
459 expr__ctx_free(pctx);
460 return;
461 }
462 if (!metric_events[i]) {
463 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
464 char *unit;
465 char metric_bf[64];
466
467 if (metric_threshold &&
468 expr__parse(&threshold, pctx, metric_threshold) == 0 &&
469 !isnan(threshold)) {
470 color = fpclassify(threshold) == FP_ZERO
471 ? PERF_COLOR_GREEN : PERF_COLOR_RED;
472 }
473
474 if (metric_unit && metric_name) {
475 if (perf_pmu__convert_scale(metric_unit,
476 &unit, &scale) >= 0) {
477 ratio *= scale;
478 }
479 if (strstr(metric_expr, "?"))
480 scnprintf(metric_bf, sizeof(metric_bf),
481 "%s %s_%d", unit, metric_name, runtime);
482 else
483 scnprintf(metric_bf, sizeof(metric_bf),
484 "%s %s", unit, metric_name);
485
486 print_metric(config, ctxp, color, "%8.1f",
487 metric_bf, ratio);
488 } else {
489 print_metric(config, ctxp, color, "%8.2f",
490 metric_name ?
491 metric_name :
492 out->force_header ? name : "",
493 ratio);
494 }
495 } else {
496 print_metric(config, ctxp, color, /*unit=*/NULL,
497 out->force_header ?
498 (metric_name ? metric_name : name) : "", 0);
499 }
500 } else {
501 print_metric(config, ctxp, color, /*unit=*/NULL,
502 out->force_header ?
503 (metric_name ? metric_name : name) : "", 0);
504 }
505
506 expr__ctx_free(pctx);
507}
508
509double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
510{
511 struct expr_parse_ctx *pctx;
512 double ratio = 0.0;
513
514 pctx = expr__ctx_new();
515 if (!pctx)
516 return NAN;
517
518 if (prepare_metric(mexp->metric_events, mexp->metric_refs, pctx, aggr_idx) < 0)
519 goto out;
520
521 if (expr__parse(&ratio, pctx, mexp->metric_expr))
522 ratio = 0.0;
523
524out:
525 expr__ctx_free(pctx);
526 return ratio;
527}
528
529void perf_stat__print_shadow_stats(struct perf_stat_config *config,
530 struct evsel *evsel,
531 double avg, int aggr_idx,
532 struct perf_stat_output_ctx *out,
533 struct rblist *metric_events)
534{
535 typedef void (*stat_print_function_t)(struct perf_stat_config *config,
536 const struct evsel *evsel,
537 int aggr_idx, double misses,
538 struct perf_stat_output_ctx *out);
539 static const stat_print_function_t stat_print_function[STAT_MAX] = {
540 [STAT_INSTRUCTIONS] = print_instructions,
541 [STAT_BRANCH_MISS] = print_branch_miss,
542 [STAT_L1D_MISS] = print_l1d_miss,
543 [STAT_L1I_MISS] = print_l1i_miss,
544 [STAT_DTLB_MISS] = print_dtlb_miss,
545 [STAT_ITLB_MISS] = print_itlb_miss,
546 [STAT_LL_MISS] = print_ll_miss,
547 [STAT_CACHE_MISSES] = print_cache_miss,
548 [STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
549 [STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
550 [STAT_CYCLES] = print_cycles,
551 [STAT_NSECS] = print_nsecs,
552 };
553 print_metric_t print_metric = out->print_metric;
554 void *ctxp = out->ctx;
555 struct metric_event *me;
556 int num = 1;
557
558 if (config->iostat_run) {
559 iostat_print_metric(config, evsel, out);
560 } else {
561 stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
562
563 if (fn)
564 fn(config, evsel, aggr_idx, avg, out);
565 else {
566 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
567
568 if (nsecs) {
569 char unit = ' ';
570 char unit_buf[10] = "/sec";
571 double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
572 &unit);
573
574 if (unit != ' ')
575 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
576 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
577 } else
578 num = 0;
579 }
580 }
581
582 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
583 struct metric_expr *mexp;
584
585 list_for_each_entry (mexp, &me->head, nd) {
586 if (num++ > 0)
587 out->new_line(config, ctxp);
588 generic_metric(config, mexp->metric_expr, mexp->metric_threshold,
589 mexp->metric_events, mexp->metric_refs, evsel->name,
590 mexp->metric_name, mexp->metric_unit, mexp->runtime,
591 aggr_idx, out);
592 }
593 }
594 if (num == 0)
595 print_metric(config, ctxp, NULL, NULL, NULL, 0);
596}