Root/tools/perf/builtin-stat.c

1/*
2 * builtin-stat.c
3 *
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
6 *
7 * Sample output:
8
9   $ perf stat ./hackbench 10
10
11  Time: 0.118
12
13  Performance counter stats for './hackbench 10':
14
15       1708.761321 task-clock # 11.037 CPUs utilized
16            41,190 context-switches # 0.024 M/sec
17             6,735 CPU-migrations # 0.004 M/sec
18            17,318 page-faults # 0.010 M/sec
19     5,205,202,243 cycles # 3.046 GHz
20     3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21     1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22     2,603,501,247 instructions # 0.50 insns per cycle
23                                             # 1.48 stalled cycles per insn
24       484,357,498 branches # 283.455 M/sec
25         6,388,934 branch-misses # 1.32% of all branches
26
27        0.154822978 seconds time elapsed
28
29 *
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
31 *
32 * Improvements and fixes by:
33 *
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
40 *
41 * Released under the GPL v2. (and only v2, not any later version)
42 */
43
44#include "perf.h"
45#include "builtin.h"
46#include "util/util.h"
47#include "util/parse-options.h"
48#include "util/parse-events.h"
49#include "util/event.h"
50#include "util/evlist.h"
51#include "util/evsel.h"
52#include "util/debug.h"
53#include "util/color.h"
54#include "util/header.h"
55#include "util/cpumap.h"
56#include "util/thread.h"
57#include "util/thread_map.h"
58
59#include <sys/prctl.h>
60#include <math.h>
61#include <locale.h>
62
63#define DEFAULT_SEPARATOR " "
64#define CNTR_NOT_SUPPORTED "<not supported>"
65#define CNTR_NOT_COUNTED "<not counted>"
66
67static struct perf_event_attr default_attrs[] = {
68
69  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
70  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
71  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
72  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
73
74  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
75  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
76  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
77  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
78  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
79  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
80
81};
82
83/*
84 * Detailed stats (-d), covering the L1 and last level data caches:
85 */
86static struct perf_event_attr detailed_attrs[] = {
87
88  { .type = PERF_TYPE_HW_CACHE,
89    .config =
90     PERF_COUNT_HW_CACHE_L1D << 0 |
91    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
92    (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
93
94  { .type = PERF_TYPE_HW_CACHE,
95    .config =
96     PERF_COUNT_HW_CACHE_L1D << 0 |
97    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
98    (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
99
100  { .type = PERF_TYPE_HW_CACHE,
101    .config =
102     PERF_COUNT_HW_CACHE_LL << 0 |
103    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
104    (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
105
106  { .type = PERF_TYPE_HW_CACHE,
107    .config =
108     PERF_COUNT_HW_CACHE_LL << 0 |
109    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
110    (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
111};
112
113/*
114 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
115 */
116static struct perf_event_attr very_detailed_attrs[] = {
117
118  { .type = PERF_TYPE_HW_CACHE,
119    .config =
120     PERF_COUNT_HW_CACHE_L1I << 0 |
121    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
122    (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
123
124  { .type = PERF_TYPE_HW_CACHE,
125    .config =
126     PERF_COUNT_HW_CACHE_L1I << 0 |
127    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
128    (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
129
130  { .type = PERF_TYPE_HW_CACHE,
131    .config =
132     PERF_COUNT_HW_CACHE_DTLB << 0 |
133    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
134    (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
135
136  { .type = PERF_TYPE_HW_CACHE,
137    .config =
138     PERF_COUNT_HW_CACHE_DTLB << 0 |
139    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
140    (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
141
142  { .type = PERF_TYPE_HW_CACHE,
143    .config =
144     PERF_COUNT_HW_CACHE_ITLB << 0 |
145    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
146    (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
147
148  { .type = PERF_TYPE_HW_CACHE,
149    .config =
150     PERF_COUNT_HW_CACHE_ITLB << 0 |
151    (PERF_COUNT_HW_CACHE_OP_READ << 8) |
152    (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
153
154};
155
156/*
157 * Very, very detailed stats (-d -d -d), adding prefetch events:
158 */
159static struct perf_event_attr very_very_detailed_attrs[] = {
160
161  { .type = PERF_TYPE_HW_CACHE,
162    .config =
163     PERF_COUNT_HW_CACHE_L1D << 0 |
164    (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
165    (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
166
167  { .type = PERF_TYPE_HW_CACHE,
168    .config =
169     PERF_COUNT_HW_CACHE_L1D << 0 |
170    (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
171    (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
172};
173
174
175
176static struct perf_evlist *evsel_list;
177
178static struct perf_target target = {
179    .uid = UINT_MAX,
180};
181
182static int run_idx = 0;
183static int run_count = 1;
184static bool no_inherit = false;
185static bool scale = true;
186static bool no_aggr = false;
187static pid_t child_pid = -1;
188static bool null_run = false;
189static int detailed_run = 0;
190static bool sync_run = false;
191static bool big_num = true;
192static int big_num_opt = -1;
193static const char *csv_sep = NULL;
194static bool csv_output = false;
195static bool group = false;
196static const char *output_name = NULL;
197static FILE *output = NULL;
198static int output_fd;
199
200static volatile int done = 0;
201
202struct stats
203{
204    double n, mean, M2;
205};
206
207struct perf_stat {
208    struct stats res_stats[3];
209};
210
211static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
212{
213    evsel->priv = zalloc(sizeof(struct perf_stat));
214    return evsel->priv == NULL ? -ENOMEM : 0;
215}
216
217static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
218{
219    free(evsel->priv);
220    evsel->priv = NULL;
221}
222
223static void update_stats(struct stats *stats, u64 val)
224{
225    double delta;
226
227    stats->n++;
228    delta = val - stats->mean;
229    stats->mean += delta / stats->n;
230    stats->M2 += delta*(val - stats->mean);
231}
232
233static double avg_stats(struct stats *stats)
234{
235    return stats->mean;
236}
237
238/*
239 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
240 *
241 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
242 * s^2 = -------------------------------
243 * n - 1
244 *
245 * http://en.wikipedia.org/wiki/Stddev
246 *
247 * The std dev of the mean is related to the std dev by:
248 *
249 * s
250 * s_mean = -------
251 * sqrt(n)
252 *
253 */
254static double stddev_stats(struct stats *stats)
255{
256    double variance, variance_mean;
257
258    if (!stats->n)
259        return 0.0;
260
261    variance = stats->M2 / (stats->n - 1);
262    variance_mean = variance / stats->n;
263
264    return sqrt(variance_mean);
265}
266
267static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
268static struct stats runtime_cycles_stats[MAX_NR_CPUS];
269static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
270static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
271static struct stats runtime_branches_stats[MAX_NR_CPUS];
272static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
273static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
274static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
275static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
276static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
277static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
278static struct stats walltime_nsecs_stats;
279
280static int create_perf_stat_counter(struct perf_evsel *evsel,
281                    struct perf_evsel *first)
282{
283    struct perf_event_attr *attr = &evsel->attr;
284    struct xyarray *group_fd = NULL;
285    bool exclude_guest_missing = false;
286    int ret;
287
288    if (group && evsel != first)
289        group_fd = first->fd;
290
291    if (scale)
292        attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
293                    PERF_FORMAT_TOTAL_TIME_RUNNING;
294
295    attr->inherit = !no_inherit;
296
297retry:
298    if (exclude_guest_missing)
299        evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
300
301    if (perf_target__has_cpu(&target)) {
302        ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
303                           group, group_fd);
304        if (ret)
305            goto check_ret;
306        return 0;
307    }
308
309    if (!perf_target__has_task(&target) && (!group || evsel == first)) {
310        attr->disabled = 1;
311        attr->enable_on_exec = 1;
312    }
313
314    ret = perf_evsel__open_per_thread(evsel, evsel_list->threads,
315                      group, group_fd);
316    if (!ret)
317        return 0;
318    /* fall through */
319check_ret:
320    if (ret && errno == EINVAL) {
321        if (!exclude_guest_missing &&
322            (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
323            pr_debug("Old kernel, cannot exclude "
324                 "guest or host samples.\n");
325            exclude_guest_missing = true;
326            goto retry;
327        }
328    }
329    return ret;
330}
331
332/*
333 * Does the counter have nsecs as a unit?
334 */
335static inline int nsec_counter(struct perf_evsel *evsel)
336{
337    if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
338        perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
339        return 1;
340
341    return 0;
342}
343
344/*
345 * Update various tracking values we maintain to print
346 * more semantic information such as miss/hit ratios,
347 * instruction rates, etc:
348 */
349static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
350{
351    if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
352        update_stats(&runtime_nsecs_stats[0], count[0]);
353    else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
354        update_stats(&runtime_cycles_stats[0], count[0]);
355    else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
356        update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
357    else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
358        update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
359    else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
360        update_stats(&runtime_branches_stats[0], count[0]);
361    else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
362        update_stats(&runtime_cacherefs_stats[0], count[0]);
363    else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
364        update_stats(&runtime_l1_dcache_stats[0], count[0]);
365    else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
366        update_stats(&runtime_l1_icache_stats[0], count[0]);
367    else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
368        update_stats(&runtime_ll_cache_stats[0], count[0]);
369    else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
370        update_stats(&runtime_dtlb_cache_stats[0], count[0]);
371    else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
372        update_stats(&runtime_itlb_cache_stats[0], count[0]);
373}
374
375/*
376 * Read out the results of a single counter:
377 * aggregate counts across CPUs in system-wide mode
378 */
379static int read_counter_aggr(struct perf_evsel *counter)
380{
381    struct perf_stat *ps = counter->priv;
382    u64 *count = counter->counts->aggr.values;
383    int i;
384
385    if (__perf_evsel__read(counter, evsel_list->cpus->nr,
386                   evsel_list->threads->nr, scale) < 0)
387        return -1;
388
389    for (i = 0; i < 3; i++)
390        update_stats(&ps->res_stats[i], count[i]);
391
392    if (verbose) {
393        fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
394            event_name(counter), count[0], count[1], count[2]);
395    }
396
397    /*
398     * Save the full runtime - to allow normalization during printout:
399     */
400    update_shadow_stats(counter, count);
401
402    return 0;
403}
404
405/*
406 * Read out the results of a single counter:
407 * do not aggregate counts across CPUs in system-wide mode
408 */
409static int read_counter(struct perf_evsel *counter)
410{
411    u64 *count;
412    int cpu;
413
414    for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
415        if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
416            return -1;
417
418        count = counter->counts->cpu[cpu].values;
419
420        update_shadow_stats(counter, count);
421    }
422
423    return 0;
424}
425
426static int run_perf_stat(int argc __used, const char **argv)
427{
428    unsigned long long t0, t1;
429    struct perf_evsel *counter, *first;
430    int status = 0;
431    int child_ready_pipe[2], go_pipe[2];
432    const bool forks = (argc > 0);
433    char buf;
434
435    if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
436        perror("failed to create pipes");
437        exit(1);
438    }
439
440    if (forks) {
441        if ((child_pid = fork()) < 0)
442            perror("failed to fork");
443
444        if (!child_pid) {
445            close(child_ready_pipe[0]);
446            close(go_pipe[1]);
447            fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
448
449            /*
450             * Do a dummy execvp to get the PLT entry resolved,
451             * so we avoid the resolver overhead on the real
452             * execvp call.
453             */
454            execvp("", (char **)argv);
455
456            /*
457             * Tell the parent we're ready to go
458             */
459            close(child_ready_pipe[1]);
460
461            /*
462             * Wait until the parent tells us to go.
463             */
464            if (read(go_pipe[0], &buf, 1) == -1)
465                perror("unable to read pipe");
466
467            execvp(argv[0], (char **)argv);
468
469            perror(argv[0]);
470            exit(-1);
471        }
472
473        if (perf_target__none(&target))
474            evsel_list->threads->map[0] = child_pid;
475
476        /*
477         * Wait for the child to be ready to exec.
478         */
479        close(child_ready_pipe[1]);
480        close(go_pipe[0]);
481        if (read(child_ready_pipe[0], &buf, 1) == -1)
482            perror("unable to read pipe");
483        close(child_ready_pipe[0]);
484    }
485
486    first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
487
488    list_for_each_entry(counter, &evsel_list->entries, node) {
489        if (create_perf_stat_counter(counter, first) < 0) {
490            /*
491             * PPC returns ENXIO for HW counters until 2.6.37
492             * (behavior changed with commit b0a873e).
493             */
494            if (errno == EINVAL || errno == ENOSYS ||
495                errno == ENOENT || errno == EOPNOTSUPP ||
496                errno == ENXIO) {
497                if (verbose)
498                    ui__warning("%s event is not supported by the kernel.\n",
499                            event_name(counter));
500                counter->supported = false;
501                continue;
502            }
503
504            if (errno == EPERM || errno == EACCES) {
505                error("You may not have permission to collect %sstats.\n"
506                      "\t Consider tweaking"
507                      " /proc/sys/kernel/perf_event_paranoid or running as root.",
508                      target.system_wide ? "system-wide " : "");
509            } else {
510                error("open_counter returned with %d (%s). "
511                      "/bin/dmesg may provide additional information.\n",
512                       errno, strerror(errno));
513            }
514            if (child_pid != -1)
515                kill(child_pid, SIGTERM);
516            die("Not all events could be opened.\n");
517            return -1;
518        }
519        counter->supported = true;
520    }
521
522    if (perf_evlist__set_filters(evsel_list)) {
523        error("failed to set filter with %d (%s)\n", errno,
524            strerror(errno));
525        return -1;
526    }
527
528    /*
529     * Enable counters and exec the command:
530     */
531    t0 = rdclock();
532
533    if (forks) {
534        close(go_pipe[1]);
535        wait(&status);
536        if (WIFSIGNALED(status))
537            psignal(WTERMSIG(status), argv[0]);
538    } else {
539        while(!done) sleep(1);
540    }
541
542    t1 = rdclock();
543
544    update_stats(&walltime_nsecs_stats, t1 - t0);
545
546    if (no_aggr) {
547        list_for_each_entry(counter, &evsel_list->entries, node) {
548            read_counter(counter);
549            perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
550        }
551    } else {
552        list_for_each_entry(counter, &evsel_list->entries, node) {
553            read_counter_aggr(counter);
554            perf_evsel__close_fd(counter, evsel_list->cpus->nr,
555                         evsel_list->threads->nr);
556        }
557    }
558
559    return WEXITSTATUS(status);
560}
561
562static void print_noise_pct(double total, double avg)
563{
564    double pct = 0.0;
565
566    if (avg)
567        pct = 100.0*total/avg;
568
569    if (csv_output)
570        fprintf(output, "%s%.2f%%", csv_sep, pct);
571    else if (pct)
572        fprintf(output, " ( +-%6.2f%% )", pct);
573}
574
575static void print_noise(struct perf_evsel *evsel, double avg)
576{
577    struct perf_stat *ps;
578
579    if (run_count == 1)
580        return;
581
582    ps = evsel->priv;
583    print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
584}
585
586static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
587{
588    double msecs = avg / 1e6;
589    char cpustr[16] = { '\0', };
590    const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
591
592    if (no_aggr)
593        sprintf(cpustr, "CPU%*d%s",
594            csv_output ? 0 : -4,
595            evsel_list->cpus->map[cpu], csv_sep);
596
597    fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel));
598
599    if (evsel->cgrp)
600        fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
601
602    if (csv_output)
603        return;
604
605    if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
606        fprintf(output, " # %8.3f CPUs utilized ",
607            avg / avg_stats(&walltime_nsecs_stats));
608    else
609        fprintf(output, " ");
610}
611
612/* used for get_ratio_color() */
613enum grc_type {
614    GRC_STALLED_CYCLES_FE,
615    GRC_STALLED_CYCLES_BE,
616    GRC_CACHE_MISSES,
617    GRC_MAX_NR
618};
619
620static const char *get_ratio_color(enum grc_type type, double ratio)
621{
622    static const double grc_table[GRC_MAX_NR][3] = {
623        [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
624        [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
625        [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
626    };
627    const char *color = PERF_COLOR_NORMAL;
628
629    if (ratio > grc_table[type][0])
630        color = PERF_COLOR_RED;
631    else if (ratio > grc_table[type][1])
632        color = PERF_COLOR_MAGENTA;
633    else if (ratio > grc_table[type][2])
634        color = PERF_COLOR_YELLOW;
635
636    return color;
637}
638
639static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
640{
641    double total, ratio = 0.0;
642    const char *color;
643
644    total = avg_stats(&runtime_cycles_stats[cpu]);
645
646    if (total)
647        ratio = avg / total * 100.0;
648
649    color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
650
651    fprintf(output, " # ");
652    color_fprintf(output, color, "%6.2f%%", ratio);
653    fprintf(output, " frontend cycles idle ");
654}
655
656static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
657{
658    double total, ratio = 0.0;
659    const char *color;
660
661    total = avg_stats(&runtime_cycles_stats[cpu]);
662
663    if (total)
664        ratio = avg / total * 100.0;
665
666    color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
667
668    fprintf(output, " # ");
669    color_fprintf(output, color, "%6.2f%%", ratio);
670    fprintf(output, " backend cycles idle ");
671}
672
673static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
674{
675    double total, ratio = 0.0;
676    const char *color;
677
678    total = avg_stats(&runtime_branches_stats[cpu]);
679
680    if (total)
681        ratio = avg / total * 100.0;
682
683    color = get_ratio_color(GRC_CACHE_MISSES, ratio);
684
685    fprintf(output, " # ");
686    color_fprintf(output, color, "%6.2f%%", ratio);
687    fprintf(output, " of all branches ");
688}
689
690static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
691{
692    double total, ratio = 0.0;
693    const char *color;
694
695    total = avg_stats(&runtime_l1_dcache_stats[cpu]);
696
697    if (total)
698        ratio = avg / total * 100.0;
699
700    color = get_ratio_color(GRC_CACHE_MISSES, ratio);
701
702    fprintf(output, " # ");
703    color_fprintf(output, color, "%6.2f%%", ratio);
704    fprintf(output, " of all L1-dcache hits ");
705}
706
707static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
708{
709    double total, ratio = 0.0;
710    const char *color;
711
712    total = avg_stats(&runtime_l1_icache_stats[cpu]);
713
714    if (total)
715        ratio = avg / total * 100.0;
716
717    color = get_ratio_color(GRC_CACHE_MISSES, ratio);
718
719    fprintf(output, " # ");
720    color_fprintf(output, color, "%6.2f%%", ratio);
721    fprintf(output, " of all L1-icache hits ");
722}
723
724static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
725{
726    double total, ratio = 0.0;
727    const char *color;
728
729    total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
730
731    if (total)
732        ratio = avg / total * 100.0;
733
734    color = get_ratio_color(GRC_CACHE_MISSES, ratio);
735
736    fprintf(output, " # ");
737    color_fprintf(output, color, "%6.2f%%", ratio);
738    fprintf(output, " of all dTLB cache hits ");
739}
740
741static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
742{
743    double total, ratio = 0.0;
744    const char *color;
745
746    total = avg_stats(&runtime_itlb_cache_stats[cpu]);
747
748    if (total)
749        ratio = avg / total * 100.0;
750
751    color = get_ratio_color(GRC_CACHE_MISSES, ratio);
752
753    fprintf(output, " # ");
754    color_fprintf(output, color, "%6.2f%%", ratio);
755    fprintf(output, " of all iTLB cache hits ");
756}
757
758static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
759{
760    double total, ratio = 0.0;
761    const char *color;
762
763    total = avg_stats(&runtime_ll_cache_stats[cpu]);
764
765    if (total)
766        ratio = avg / total * 100.0;
767
768    color = get_ratio_color(GRC_CACHE_MISSES, ratio);
769
770    fprintf(output, " # ");
771    color_fprintf(output, color, "%6.2f%%", ratio);
772    fprintf(output, " of all LL-cache hits ");
773}
774
775static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
776{
777    double total, ratio = 0.0;
778    char cpustr[16] = { '\0', };
779    const char *fmt;
780
781    if (csv_output)
782        fmt = "%s%.0f%s%s";
783    else if (big_num)
784        fmt = "%s%'18.0f%s%-25s";
785    else
786        fmt = "%s%18.0f%s%-25s";
787
788    if (no_aggr)
789        sprintf(cpustr, "CPU%*d%s",
790            csv_output ? 0 : -4,
791            evsel_list->cpus->map[cpu], csv_sep);
792    else
793        cpu = 0;
794
795    fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel));
796
797    if (evsel->cgrp)
798        fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
799
800    if (csv_output)
801        return;
802
803    if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
804        total = avg_stats(&runtime_cycles_stats[cpu]);
805
806        if (total)
807            ratio = avg / total;
808
809        fprintf(output, " # %5.2f insns per cycle ", ratio);
810
811        total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
812        total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
813
814        if (total && avg) {
815            ratio = total / avg;
816            fprintf(output, "\n # %5.2f stalled cycles per insn", ratio);
817        }
818
819    } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
820            runtime_branches_stats[cpu].n != 0) {
821        print_branch_misses(cpu, evsel, avg);
822    } else if (
823        evsel->attr.type == PERF_TYPE_HW_CACHE &&
824        evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
825                    ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
826                    ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
827            runtime_l1_dcache_stats[cpu].n != 0) {
828        print_l1_dcache_misses(cpu, evsel, avg);
829    } else if (
830        evsel->attr.type == PERF_TYPE_HW_CACHE &&
831        evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
832                    ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
833                    ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
834            runtime_l1_icache_stats[cpu].n != 0) {
835        print_l1_icache_misses(cpu, evsel, avg);
836    } else if (
837        evsel->attr.type == PERF_TYPE_HW_CACHE &&
838        evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
839                    ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
840                    ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
841            runtime_dtlb_cache_stats[cpu].n != 0) {
842        print_dtlb_cache_misses(cpu, evsel, avg);
843    } else if (
844        evsel->attr.type == PERF_TYPE_HW_CACHE &&
845        evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
846                    ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
847                    ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
848            runtime_itlb_cache_stats[cpu].n != 0) {
849        print_itlb_cache_misses(cpu, evsel, avg);
850    } else if (
851        evsel->attr.type == PERF_TYPE_HW_CACHE &&
852        evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
853                    ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
854                    ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
855            runtime_ll_cache_stats[cpu].n != 0) {
856        print_ll_cache_misses(cpu, evsel, avg);
857    } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
858            runtime_cacherefs_stats[cpu].n != 0) {
859        total = avg_stats(&runtime_cacherefs_stats[cpu]);
860
861        if (total)
862            ratio = avg * 100 / total;
863
864        fprintf(output, " # %8.3f %% of all cache refs ", ratio);
865
866    } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
867        print_stalled_cycles_frontend(cpu, evsel, avg);
868    } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
869        print_stalled_cycles_backend(cpu, evsel, avg);
870    } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
871        total = avg_stats(&runtime_nsecs_stats[cpu]);
872
873        if (total)
874            ratio = 1.0 * avg / total;
875
876        fprintf(output, " # %8.3f GHz ", ratio);
877    } else if (runtime_nsecs_stats[cpu].n != 0) {
878        char unit = 'M';
879
880        total = avg_stats(&runtime_nsecs_stats[cpu]);
881
882        if (total)
883            ratio = 1000.0 * avg / total;
884        if (ratio < 0.001) {
885            ratio *= 1000;
886            unit = 'K';
887        }
888
889        fprintf(output, " # %8.3f %c/sec ", ratio, unit);
890    } else {
891        fprintf(output, " ");
892    }
893}
894
895/*
896 * Print out the results of a single counter:
897 * aggregated counts in system-wide mode
898 */
899static void print_counter_aggr(struct perf_evsel *counter)
900{
901    struct perf_stat *ps = counter->priv;
902    double avg = avg_stats(&ps->res_stats[0]);
903    int scaled = counter->counts->scaled;
904
905    if (scaled == -1) {
906        fprintf(output, "%*s%s%*s",
907            csv_output ? 0 : 18,
908            counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
909            csv_sep,
910            csv_output ? 0 : -24,
911            event_name(counter));
912
913        if (counter->cgrp)
914            fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
915
916        fputc('\n', output);
917        return;
918    }
919
920    if (nsec_counter(counter))
921        nsec_printout(-1, counter, avg);
922    else
923        abs_printout(-1, counter, avg);
924
925    print_noise(counter, avg);
926
927    if (csv_output) {
928        fputc('\n', output);
929        return;
930    }
931
932    if (scaled) {
933        double avg_enabled, avg_running;
934
935        avg_enabled = avg_stats(&ps->res_stats[1]);
936        avg_running = avg_stats(&ps->res_stats[2]);
937
938        fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
939    }
940    fprintf(output, "\n");
941}
942
943/*
944 * Print out the results of a single counter:
945 * does not use aggregated count in system-wide
946 */
947static void print_counter(struct perf_evsel *counter)
948{
949    u64 ena, run, val;
950    int cpu;
951
952    for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
953        val = counter->counts->cpu[cpu].val;
954        ena = counter->counts->cpu[cpu].ena;
955        run = counter->counts->cpu[cpu].run;
956        if (run == 0 || ena == 0) {
957            fprintf(output, "CPU%*d%s%*s%s%*s",
958                csv_output ? 0 : -4,
959                evsel_list->cpus->map[cpu], csv_sep,
960                csv_output ? 0 : 18,
961                counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
962                csv_sep,
963                csv_output ? 0 : -24,
964                event_name(counter));
965
966            if (counter->cgrp)
967                fprintf(output, "%s%s",
968                    csv_sep, counter->cgrp->name);
969
970            fputc('\n', output);
971            continue;
972        }
973
974        if (nsec_counter(counter))
975            nsec_printout(cpu, counter, val);
976        else
977            abs_printout(cpu, counter, val);
978
979        if (!csv_output) {
980            print_noise(counter, 1.0);
981
982            if (run != ena)
983                fprintf(output, " (%.2f%%)",
984                    100.0 * run / ena);
985        }
986        fputc('\n', output);
987    }
988}
989
990static void print_stat(int argc, const char **argv)
991{
992    struct perf_evsel *counter;
993    int i;
994
995    fflush(stdout);
996
997    if (!csv_output) {
998        fprintf(output, "\n");
999        fprintf(output, " Performance counter stats for ");
1000        if (!perf_target__has_task(&target)) {
1001            fprintf(output, "\'%s", argv[0]);
1002            for (i = 1; i < argc; i++)
1003                fprintf(output, " %s", argv[i]);
1004        } else if (target.pid)
1005            fprintf(output, "process id \'%s", target.pid);
1006        else
1007            fprintf(output, "thread id \'%s", target.tid);
1008
1009        fprintf(output, "\'");
1010        if (run_count > 1)
1011            fprintf(output, " (%d runs)", run_count);
1012        fprintf(output, ":\n\n");
1013    }
1014
1015    if (no_aggr) {
1016        list_for_each_entry(counter, &evsel_list->entries, node)
1017            print_counter(counter);
1018    } else {
1019        list_for_each_entry(counter, &evsel_list->entries, node)
1020            print_counter_aggr(counter);
1021    }
1022
1023    if (!csv_output) {
1024        if (!null_run)
1025            fprintf(output, "\n");
1026        fprintf(output, " %17.9f seconds time elapsed",
1027                avg_stats(&walltime_nsecs_stats)/1e9);
1028        if (run_count > 1) {
1029            fprintf(output, " ");
1030            print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1031                    avg_stats(&walltime_nsecs_stats));
1032        }
1033        fprintf(output, "\n\n");
1034    }
1035}
1036
1037static volatile int signr = -1;
1038
1039static void skip_signal(int signo)
1040{
1041    if(child_pid == -1)
1042        done = 1;
1043
1044    signr = signo;
1045}
1046
1047static void sig_atexit(void)
1048{
1049    if (child_pid != -1)
1050        kill(child_pid, SIGTERM);
1051
1052    if (signr == -1)
1053        return;
1054
1055    signal(signr, SIG_DFL);
1056    kill(getpid(), signr);
1057}
1058
1059static const char * const stat_usage[] = {
1060    "perf stat [<options>] [<command>]",
1061    NULL
1062};
1063
1064static int stat__set_big_num(const struct option *opt __used,
1065                 const char *s __used, int unset)
1066{
1067    big_num_opt = unset ? 0 : 1;
1068    return 0;
1069}
1070
1071static bool append_file;
1072
1073static const struct option options[] = {
1074    OPT_CALLBACK('e', "event", &evsel_list, "event",
1075             "event selector. use 'perf list' to list available events",
1076             parse_events_option),
1077    OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1078             "event filter", parse_filter),
1079    OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1080            "child tasks do not inherit counters"),
1081    OPT_STRING('p', "pid", &target.pid, "pid",
1082           "stat events on existing process id"),
1083    OPT_STRING('t', "tid", &target.tid, "tid",
1084           "stat events on existing thread id"),
1085    OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1086            "system-wide collection from all CPUs"),
1087    OPT_BOOLEAN('g', "group", &group,
1088            "put the counters into a counter group"),
1089    OPT_BOOLEAN('c', "scale", &scale,
1090            "scale/normalize counters"),
1091    OPT_INCR('v', "verbose", &verbose,
1092            "be more verbose (show counter open errors, etc)"),
1093    OPT_INTEGER('r', "repeat", &run_count,
1094            "repeat command and print average + stddev (max: 100)"),
1095    OPT_BOOLEAN('n', "null", &null_run,
1096            "null run - dont start any counters"),
1097    OPT_INCR('d', "detailed", &detailed_run,
1098            "detailed run - start a lot of events"),
1099    OPT_BOOLEAN('S', "sync", &sync_run,
1100            "call sync() before starting a run"),
1101    OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1102               "print large numbers with thousands\' separators",
1103               stat__set_big_num),
1104    OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1105            "list of cpus to monitor in system-wide"),
1106    OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1107            "disable CPU count aggregation"),
1108    OPT_STRING('x', "field-separator", &csv_sep, "separator",
1109           "print counts with custom separator"),
1110    OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1111             "monitor event in cgroup name only",
1112             parse_cgroups),
1113    OPT_STRING('o', "output", &output_name, "file",
1114            "output file name"),
1115    OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1116    OPT_INTEGER(0, "log-fd", &output_fd,
1117            "log output to fd, instead of stderr"),
1118    OPT_END()
1119};
1120
1121/*
1122 * Add default attributes, if there were no attributes specified or
1123 * if -d/--detailed, -d -d or -d -d -d is used:
1124 */
1125static int add_default_attributes(void)
1126{
1127    /* Set attrs if no event is selected and !null_run: */
1128    if (null_run)
1129        return 0;
1130
1131    if (!evsel_list->nr_entries) {
1132        if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1133            return -1;
1134    }
1135
1136    /* Detailed events get appended to the event list: */
1137
1138    if (detailed_run < 1)
1139        return 0;
1140
1141    /* Append detailed run extra attributes: */
1142    if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1143        return -1;
1144
1145    if (detailed_run < 2)
1146        return 0;
1147
1148    /* Append very detailed run extra attributes: */
1149    if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1150        return -1;
1151
1152    if (detailed_run < 3)
1153        return 0;
1154
1155    /* Append very, very detailed run extra attributes: */
1156    return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1157}
1158
1159int cmd_stat(int argc, const char **argv, const char *prefix __used)
1160{
1161    struct perf_evsel *pos;
1162    int status = -ENOMEM;
1163    const char *mode;
1164
1165    setlocale(LC_ALL, "");
1166
1167    evsel_list = perf_evlist__new(NULL, NULL);
1168    if (evsel_list == NULL)
1169        return -ENOMEM;
1170
1171    argc = parse_options(argc, argv, options, stat_usage,
1172        PARSE_OPT_STOP_AT_NON_OPTION);
1173
1174    output = stderr;
1175    if (output_name && strcmp(output_name, "-"))
1176        output = NULL;
1177
1178    if (output_name && output_fd) {
1179        fprintf(stderr, "cannot use both --output and --log-fd\n");
1180        usage_with_options(stat_usage, options);
1181    }
1182
1183    if (output_fd < 0) {
1184        fprintf(stderr, "argument to --log-fd must be a > 0\n");
1185        usage_with_options(stat_usage, options);
1186    }
1187
1188    if (!output) {
1189        struct timespec tm;
1190        mode = append_file ? "a" : "w";
1191
1192        output = fopen(output_name, mode);
1193        if (!output) {
1194            perror("failed to create output file");
1195            exit(-1);
1196        }
1197        clock_gettime(CLOCK_REALTIME, &tm);
1198        fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1199    } else if (output_fd > 0) {
1200        mode = append_file ? "a" : "w";
1201        output = fdopen(output_fd, mode);
1202        if (!output) {
1203            perror("Failed opening logfd");
1204            return -errno;
1205        }
1206    }
1207
1208    if (csv_sep) {
1209        csv_output = true;
1210        if (!strcmp(csv_sep, "\\t"))
1211            csv_sep = "\t";
1212    } else
1213        csv_sep = DEFAULT_SEPARATOR;
1214
1215    /*
1216     * let the spreadsheet do the pretty-printing
1217     */
1218    if (csv_output) {
1219        /* User explicitly passed -B? */
1220        if (big_num_opt == 1) {
1221            fprintf(stderr, "-B option not supported with -x\n");
1222            usage_with_options(stat_usage, options);
1223        } else /* Nope, so disable big number formatting */
1224            big_num = false;
1225    } else if (big_num_opt == 0) /* User passed --no-big-num */
1226        big_num = false;
1227
1228    if (!argc && !perf_target__has_task(&target))
1229        usage_with_options(stat_usage, options);
1230    if (run_count <= 0)
1231        usage_with_options(stat_usage, options);
1232
1233    /* no_aggr, cgroup are for system-wide only */
1234    if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) {
1235        fprintf(stderr, "both cgroup and no-aggregation "
1236            "modes only available in system-wide mode\n");
1237
1238        usage_with_options(stat_usage, options);
1239    }
1240
1241    if (add_default_attributes())
1242        goto out;
1243
1244    perf_target__validate(&target);
1245
1246    if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1247        if (perf_target__has_task(&target))
1248            pr_err("Problems finding threads of monitor\n");
1249        if (perf_target__has_cpu(&target))
1250            perror("failed to parse CPUs map");
1251
1252        usage_with_options(stat_usage, options);
1253        return -1;
1254    }
1255
1256    list_for_each_entry(pos, &evsel_list->entries, node) {
1257        if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1258            perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
1259            goto out_free_fd;
1260    }
1261
1262    /*
1263     * We dont want to block the signals - that would cause
1264     * child tasks to inherit that and Ctrl-C would not work.
1265     * What we want is for Ctrl-C to work in the exec()-ed
1266     * task, but being ignored by perf stat itself:
1267     */
1268    atexit(sig_atexit);
1269    signal(SIGINT, skip_signal);
1270    signal(SIGALRM, skip_signal);
1271    signal(SIGABRT, skip_signal);
1272
1273    status = 0;
1274    for (run_idx = 0; run_idx < run_count; run_idx++) {
1275        if (run_count != 1 && verbose)
1276            fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1277                run_idx + 1);
1278
1279        if (sync_run)
1280            sync();
1281
1282        status = run_perf_stat(argc, argv);
1283    }
1284
1285    if (status != -1)
1286        print_stat(argc, argv);
1287out_free_fd:
1288    list_for_each_entry(pos, &evsel_list->entries, node)
1289        perf_evsel__free_stat_priv(pos);
1290    perf_evlist__delete_maps(evsel_list);
1291out:
1292    perf_evlist__delete(evsel_list);
1293    return status;
1294}
1295

Archive Download this file



interactive