4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
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
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
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>
41 * Released under the GPL v2. (and only v2, not any later version)
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"
59 #include <sys/prctl.h>
63 #define DEFAULT_SEPARATOR " "
64 #define CNTR_NOT_SUPPORTED "<not supported>"
65 #define CNTR_NOT_COUNTED "<not counted>"
67 static struct perf_event_attr default_attrs
[] = {
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
},
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
},
84 * Detailed stats (-d), covering the L1 and last level data caches:
86 static struct perf_event_attr detailed_attrs
[] = {
88 { .type
= PERF_TYPE_HW_CACHE
,
90 PERF_COUNT_HW_CACHE_L1D
<< 0 |
91 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
92 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
94 { .type
= PERF_TYPE_HW_CACHE
,
96 PERF_COUNT_HW_CACHE_L1D
<< 0 |
97 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
98 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
100 { .type
= PERF_TYPE_HW_CACHE
,
102 PERF_COUNT_HW_CACHE_LL
<< 0 |
103 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
104 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
106 { .type
= PERF_TYPE_HW_CACHE
,
108 PERF_COUNT_HW_CACHE_LL
<< 0 |
109 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
110 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
114 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
116 static struct perf_event_attr very_detailed_attrs
[] = {
118 { .type
= PERF_TYPE_HW_CACHE
,
120 PERF_COUNT_HW_CACHE_L1I
<< 0 |
121 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
122 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
124 { .type
= PERF_TYPE_HW_CACHE
,
126 PERF_COUNT_HW_CACHE_L1I
<< 0 |
127 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
128 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
130 { .type
= PERF_TYPE_HW_CACHE
,
132 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
133 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
134 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
136 { .type
= PERF_TYPE_HW_CACHE
,
138 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
139 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
140 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
142 { .type
= PERF_TYPE_HW_CACHE
,
144 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
145 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
146 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
148 { .type
= PERF_TYPE_HW_CACHE
,
150 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
151 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
152 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
157 * Very, very detailed stats (-d -d -d), adding prefetch events:
159 static struct perf_event_attr very_very_detailed_attrs
[] = {
161 { .type
= PERF_TYPE_HW_CACHE
,
163 PERF_COUNT_HW_CACHE_L1D
<< 0 |
164 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
165 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
167 { .type
= PERF_TYPE_HW_CACHE
,
169 PERF_COUNT_HW_CACHE_L1D
<< 0 |
170 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
171 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
176 struct perf_evlist
*evsel_list
;
178 static bool system_wide
= false;
179 static int run_idx
= 0;
181 static int run_count
= 1;
182 static bool no_inherit
= false;
183 static bool scale
= true;
184 static bool no_aggr
= false;
185 static pid_t target_pid
= -1;
186 static pid_t target_tid
= -1;
187 static pid_t child_pid
= -1;
188 static bool null_run
= false;
189 static int detailed_run
= 0;
190 static bool sync_run
= false;
191 static bool big_num
= true;
192 static int big_num_opt
= -1;
193 static const char *cpu_list
;
194 static const char *csv_sep
= NULL
;
195 static bool csv_output
= false;
196 static bool group
= false;
197 static const char *output_name
= NULL
;
198 static FILE *output
= NULL
;
199 static int output_fd
;
201 static volatile int done
= 0;
209 struct stats res_stats
[3];
212 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
214 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
215 return evsel
->priv
== NULL
? -ENOMEM
: 0;
218 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
224 static void update_stats(struct stats
*stats
, u64 val
)
229 delta
= val
- stats
->mean
;
230 stats
->mean
+= delta
/ stats
->n
;
231 stats
->M2
+= delta
*(val
- stats
->mean
);
234 static double avg_stats(struct stats
*stats
)
240 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
242 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
243 * s^2 = -------------------------------
246 * http://en.wikipedia.org/wiki/Stddev
248 * The std dev of the mean is related to the std dev by:
255 static double stddev_stats(struct stats
*stats
)
257 double variance
, variance_mean
;
262 variance
= stats
->M2
/ (stats
->n
- 1);
263 variance_mean
= variance
/ stats
->n
;
265 return sqrt(variance_mean
);
268 struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
269 struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
270 struct stats runtime_stalled_cycles_front_stats
[MAX_NR_CPUS
];
271 struct stats runtime_stalled_cycles_back_stats
[MAX_NR_CPUS
];
272 struct stats runtime_branches_stats
[MAX_NR_CPUS
];
273 struct stats runtime_cacherefs_stats
[MAX_NR_CPUS
];
274 struct stats runtime_l1_dcache_stats
[MAX_NR_CPUS
];
275 struct stats runtime_l1_icache_stats
[MAX_NR_CPUS
];
276 struct stats runtime_ll_cache_stats
[MAX_NR_CPUS
];
277 struct stats runtime_itlb_cache_stats
[MAX_NR_CPUS
];
278 struct stats runtime_dtlb_cache_stats
[MAX_NR_CPUS
];
279 struct stats walltime_nsecs_stats
;
281 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
283 struct perf_event_attr
*attr
= &evsel
->attr
;
286 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
287 PERF_FORMAT_TOTAL_TIME_RUNNING
;
289 attr
->inherit
= !no_inherit
;
292 return perf_evsel__open_per_cpu(evsel
, evsel_list
->cpus
, group
);
294 if (target_pid
== -1 && target_tid
== -1) {
296 attr
->enable_on_exec
= 1;
299 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
, group
);
303 * Does the counter have nsecs as a unit?
305 static inline int nsec_counter(struct perf_evsel
*evsel
)
307 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
308 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
315 * Update various tracking values we maintain to print
316 * more semantic information such as miss/hit ratios,
317 * instruction rates, etc:
319 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
321 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
322 update_stats(&runtime_nsecs_stats
[0], count
[0]);
323 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
324 update_stats(&runtime_cycles_stats
[0], count
[0]);
325 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
326 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
327 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
328 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
329 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
330 update_stats(&runtime_branches_stats
[0], count
[0]);
331 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
332 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
333 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
334 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
335 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
336 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
337 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
338 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
339 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
340 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
341 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
342 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
346 * Read out the results of a single counter:
347 * aggregate counts across CPUs in system-wide mode
349 static int read_counter_aggr(struct perf_evsel
*counter
)
351 struct perf_stat
*ps
= counter
->priv
;
352 u64
*count
= counter
->counts
->aggr
.values
;
355 if (__perf_evsel__read(counter
, evsel_list
->cpus
->nr
,
356 evsel_list
->threads
->nr
, scale
) < 0)
359 for (i
= 0; i
< 3; i
++)
360 update_stats(&ps
->res_stats
[i
], count
[i
]);
363 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
364 event_name(counter
), count
[0], count
[1], count
[2]);
368 * Save the full runtime - to allow normalization during printout:
370 update_shadow_stats(counter
, count
);
376 * Read out the results of a single counter:
377 * do not aggregate counts across CPUs in system-wide mode
379 static int read_counter(struct perf_evsel
*counter
)
384 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; cpu
++) {
385 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
388 count
= counter
->counts
->cpu
[cpu
].values
;
390 update_shadow_stats(counter
, count
);
396 static int run_perf_stat(int argc __used
, const char **argv
)
398 unsigned long long t0
, t1
;
399 struct perf_evsel
*counter
;
401 int child_ready_pipe
[2], go_pipe
[2];
402 const bool forks
= (argc
> 0);
405 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
406 perror("failed to create pipes");
411 if ((child_pid
= fork()) < 0)
412 perror("failed to fork");
415 close(child_ready_pipe
[0]);
417 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
420 * Do a dummy execvp to get the PLT entry resolved,
421 * so we avoid the resolver overhead on the real
424 execvp("", (char **)argv
);
427 * Tell the parent we're ready to go
429 close(child_ready_pipe
[1]);
432 * Wait until the parent tells us to go.
434 if (read(go_pipe
[0], &buf
, 1) == -1)
435 perror("unable to read pipe");
437 execvp(argv
[0], (char **)argv
);
443 if (target_tid
== -1 && target_pid
== -1 && !system_wide
)
444 evsel_list
->threads
->map
[0] = child_pid
;
447 * Wait for the child to be ready to exec.
449 close(child_ready_pipe
[1]);
451 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
452 perror("unable to read pipe");
453 close(child_ready_pipe
[0]);
456 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
457 if (create_perf_stat_counter(counter
) < 0) {
458 if (errno
== EINVAL
|| errno
== ENOSYS
|| errno
== ENOENT
) {
460 ui__warning("%s event is not supported by the kernel.\n",
461 event_name(counter
));
462 counter
->supported
= false;
466 if (errno
== EPERM
|| errno
== EACCES
) {
467 error("You may not have permission to collect %sstats.\n"
468 "\t Consider tweaking"
469 " /proc/sys/kernel/perf_event_paranoid or running as root.",
470 system_wide
? "system-wide " : "");
472 error("open_counter returned with %d (%s). "
473 "/bin/dmesg may provide additional information.\n",
474 errno
, strerror(errno
));
477 kill(child_pid
, SIGTERM
);
478 die("Not all events could be opened.\n");
481 counter
->supported
= true;
484 if (perf_evlist__set_filters(evsel_list
)) {
485 error("failed to set filter with %d (%s)\n", errno
,
491 * Enable counters and exec the command:
498 if (WIFSIGNALED(status
))
499 psignal(WTERMSIG(status
), argv
[0]);
501 while(!done
) sleep(1);
506 update_stats(&walltime_nsecs_stats
, t1
- t0
);
509 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
510 read_counter(counter
);
511 perf_evsel__close_fd(counter
, evsel_list
->cpus
->nr
, 1);
514 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
515 read_counter_aggr(counter
);
516 perf_evsel__close_fd(counter
, evsel_list
->cpus
->nr
,
517 evsel_list
->threads
->nr
);
521 return WEXITSTATUS(status
);
524 static void print_noise_pct(double total
, double avg
)
529 pct
= 100.0*total
/avg
;
532 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
534 fprintf(output
, " ( +-%6.2f%% )", pct
);
537 static void print_noise(struct perf_evsel
*evsel
, double avg
)
539 struct perf_stat
*ps
;
545 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
548 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
550 double msecs
= avg
/ 1e6
;
551 char cpustr
[16] = { '\0', };
552 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
555 sprintf(cpustr
, "CPU%*d%s",
557 evsel_list
->cpus
->map
[cpu
], csv_sep
);
559 fprintf(output
, fmt
, cpustr
, msecs
, csv_sep
, event_name(evsel
));
562 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
567 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
568 fprintf(output
, " # %8.3f CPUs utilized ",
569 avg
/ avg_stats(&walltime_nsecs_stats
));
572 static void print_stalled_cycles_frontend(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
574 double total
, ratio
= 0.0;
577 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
580 ratio
= avg
/ total
* 100.0;
582 color
= PERF_COLOR_NORMAL
;
584 color
= PERF_COLOR_RED
;
585 else if (ratio
> 30.0)
586 color
= PERF_COLOR_MAGENTA
;
587 else if (ratio
> 10.0)
588 color
= PERF_COLOR_YELLOW
;
590 fprintf(output
, " # ");
591 color_fprintf(output
, color
, "%6.2f%%", ratio
);
592 fprintf(output
, " frontend cycles idle ");
595 static void print_stalled_cycles_backend(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
597 double total
, ratio
= 0.0;
600 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
603 ratio
= avg
/ total
* 100.0;
605 color
= PERF_COLOR_NORMAL
;
607 color
= PERF_COLOR_RED
;
608 else if (ratio
> 50.0)
609 color
= PERF_COLOR_MAGENTA
;
610 else if (ratio
> 20.0)
611 color
= PERF_COLOR_YELLOW
;
613 fprintf(output
, " # ");
614 color_fprintf(output
, color
, "%6.2f%%", ratio
);
615 fprintf(output
, " backend cycles idle ");
618 static void print_branch_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
620 double total
, ratio
= 0.0;
623 total
= avg_stats(&runtime_branches_stats
[cpu
]);
626 ratio
= avg
/ total
* 100.0;
628 color
= PERF_COLOR_NORMAL
;
630 color
= PERF_COLOR_RED
;
631 else if (ratio
> 10.0)
632 color
= PERF_COLOR_MAGENTA
;
633 else if (ratio
> 5.0)
634 color
= PERF_COLOR_YELLOW
;
636 fprintf(output
, " # ");
637 color_fprintf(output
, color
, "%6.2f%%", ratio
);
638 fprintf(output
, " of all branches ");
641 static void print_l1_dcache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
643 double total
, ratio
= 0.0;
646 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
649 ratio
= avg
/ total
* 100.0;
651 color
= PERF_COLOR_NORMAL
;
653 color
= PERF_COLOR_RED
;
654 else if (ratio
> 10.0)
655 color
= PERF_COLOR_MAGENTA
;
656 else if (ratio
> 5.0)
657 color
= PERF_COLOR_YELLOW
;
659 fprintf(output
, " # ");
660 color_fprintf(output
, color
, "%6.2f%%", ratio
);
661 fprintf(output
, " of all L1-dcache hits ");
664 static void print_l1_icache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
666 double total
, ratio
= 0.0;
669 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
672 ratio
= avg
/ total
* 100.0;
674 color
= PERF_COLOR_NORMAL
;
676 color
= PERF_COLOR_RED
;
677 else if (ratio
> 10.0)
678 color
= PERF_COLOR_MAGENTA
;
679 else if (ratio
> 5.0)
680 color
= PERF_COLOR_YELLOW
;
682 fprintf(output
, " # ");
683 color_fprintf(output
, color
, "%6.2f%%", ratio
);
684 fprintf(output
, " of all L1-icache hits ");
687 static void print_dtlb_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
689 double total
, ratio
= 0.0;
692 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
695 ratio
= avg
/ total
* 100.0;
697 color
= PERF_COLOR_NORMAL
;
699 color
= PERF_COLOR_RED
;
700 else if (ratio
> 10.0)
701 color
= PERF_COLOR_MAGENTA
;
702 else if (ratio
> 5.0)
703 color
= PERF_COLOR_YELLOW
;
705 fprintf(output
, " # ");
706 color_fprintf(output
, color
, "%6.2f%%", ratio
);
707 fprintf(output
, " of all dTLB cache hits ");
710 static void print_itlb_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
712 double total
, ratio
= 0.0;
715 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
718 ratio
= avg
/ total
* 100.0;
720 color
= PERF_COLOR_NORMAL
;
722 color
= PERF_COLOR_RED
;
723 else if (ratio
> 10.0)
724 color
= PERF_COLOR_MAGENTA
;
725 else if (ratio
> 5.0)
726 color
= PERF_COLOR_YELLOW
;
728 fprintf(output
, " # ");
729 color_fprintf(output
, color
, "%6.2f%%", ratio
);
730 fprintf(output
, " of all iTLB cache hits ");
733 static void print_ll_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
735 double total
, ratio
= 0.0;
738 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
741 ratio
= avg
/ total
* 100.0;
743 color
= PERF_COLOR_NORMAL
;
745 color
= PERF_COLOR_RED
;
746 else if (ratio
> 10.0)
747 color
= PERF_COLOR_MAGENTA
;
748 else if (ratio
> 5.0)
749 color
= PERF_COLOR_YELLOW
;
751 fprintf(output
, " # ");
752 color_fprintf(output
, color
, "%6.2f%%", ratio
);
753 fprintf(output
, " of all LL-cache hits ");
756 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
758 double total
, ratio
= 0.0;
759 char cpustr
[16] = { '\0', };
765 fmt
= "%s%'18.0f%s%-25s";
767 fmt
= "%s%18.0f%s%-25s";
770 sprintf(cpustr
, "CPU%*d%s",
772 evsel_list
->cpus
->map
[cpu
], csv_sep
);
776 fprintf(output
, fmt
, cpustr
, avg
, csv_sep
, event_name(evsel
));
779 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
784 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
785 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
790 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
792 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
793 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
797 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
800 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
801 runtime_branches_stats
[cpu
].n
!= 0) {
802 print_branch_misses(cpu
, evsel
, avg
);
804 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
805 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
806 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
807 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
808 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
809 print_l1_dcache_misses(cpu
, evsel
, avg
);
811 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
812 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
813 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
814 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
815 runtime_l1_icache_stats
[cpu
].n
!= 0) {
816 print_l1_icache_misses(cpu
, evsel
, avg
);
818 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
819 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
820 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
821 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
822 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
823 print_dtlb_cache_misses(cpu
, evsel
, avg
);
825 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
826 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
827 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
828 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
829 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
830 print_itlb_cache_misses(cpu
, evsel
, avg
);
832 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
833 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
834 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
835 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
836 runtime_ll_cache_stats
[cpu
].n
!= 0) {
837 print_ll_cache_misses(cpu
, evsel
, avg
);
838 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
839 runtime_cacherefs_stats
[cpu
].n
!= 0) {
840 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
843 ratio
= avg
* 100 / total
;
845 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
847 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
848 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
849 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
850 print_stalled_cycles_backend(cpu
, evsel
, avg
);
851 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
852 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
855 ratio
= 1.0 * avg
/ total
;
857 fprintf(output
, " # %8.3f GHz ", ratio
);
858 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
859 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
862 ratio
= 1000.0 * avg
/ total
;
864 fprintf(output
, " # %8.3f M/sec ", ratio
);
866 fprintf(output
, " ");
871 * Print out the results of a single counter:
872 * aggregated counts in system-wide mode
874 static void print_counter_aggr(struct perf_evsel
*counter
)
876 struct perf_stat
*ps
= counter
->priv
;
877 double avg
= avg_stats(&ps
->res_stats
[0]);
878 int scaled
= counter
->counts
->scaled
;
881 fprintf(output
, "%*s%s%*s",
883 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
885 csv_output
? 0 : -24,
886 event_name(counter
));
889 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
895 if (nsec_counter(counter
))
896 nsec_printout(-1, counter
, avg
);
898 abs_printout(-1, counter
, avg
);
900 print_noise(counter
, avg
);
908 double avg_enabled
, avg_running
;
910 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
911 avg_running
= avg_stats(&ps
->res_stats
[2]);
913 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
915 fprintf(output
, "\n");
919 * Print out the results of a single counter:
920 * does not use aggregated count in system-wide
922 static void print_counter(struct perf_evsel
*counter
)
927 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; cpu
++) {
928 val
= counter
->counts
->cpu
[cpu
].val
;
929 ena
= counter
->counts
->cpu
[cpu
].ena
;
930 run
= counter
->counts
->cpu
[cpu
].run
;
931 if (run
== 0 || ena
== 0) {
932 fprintf(output
, "CPU%*d%s%*s%s%*s",
934 evsel_list
->cpus
->map
[cpu
], csv_sep
,
936 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
938 csv_output
? 0 : -24,
939 event_name(counter
));
942 fprintf(output
, "%s%s",
943 csv_sep
, counter
->cgrp
->name
);
949 if (nsec_counter(counter
))
950 nsec_printout(cpu
, counter
, val
);
952 abs_printout(cpu
, counter
, val
);
955 print_noise(counter
, 1.0);
958 fprintf(output
, " (%.2f%%)",
965 static void print_stat(int argc
, const char **argv
)
967 struct perf_evsel
*counter
;
973 fprintf(output
, "\n");
974 fprintf(output
, " Performance counter stats for ");
975 if(target_pid
== -1 && target_tid
== -1) {
976 fprintf(output
, "\'%s", argv
[0]);
977 for (i
= 1; i
< argc
; i
++)
978 fprintf(output
, " %s", argv
[i
]);
979 } else if (target_pid
!= -1)
980 fprintf(output
, "process id \'%d", target_pid
);
982 fprintf(output
, "thread id \'%d", target_tid
);
984 fprintf(output
, "\'");
986 fprintf(output
, " (%d runs)", run_count
);
987 fprintf(output
, ":\n\n");
991 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
992 print_counter(counter
);
994 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
995 print_counter_aggr(counter
);
1000 fprintf(output
, "\n");
1001 fprintf(output
, " %17.9f seconds time elapsed",
1002 avg_stats(&walltime_nsecs_stats
)/1e9
);
1003 if (run_count
> 1) {
1004 fprintf(output
, " ");
1005 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1006 avg_stats(&walltime_nsecs_stats
));
1008 fprintf(output
, "\n\n");
1012 static volatile int signr
= -1;
1014 static void skip_signal(int signo
)
1022 static void sig_atexit(void)
1024 if (child_pid
!= -1)
1025 kill(child_pid
, SIGTERM
);
1030 signal(signr
, SIG_DFL
);
1031 kill(getpid(), signr
);
1034 static const char * const stat_usage
[] = {
1035 "perf stat [<options>] [<command>]",
1039 static int stat__set_big_num(const struct option
*opt __used
,
1040 const char *s __used
, int unset
)
1042 big_num_opt
= unset
? 0 : 1;
1046 static bool append_file
;
1048 static const struct option options
[] = {
1049 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1050 "event selector. use 'perf list' to list available events",
1051 parse_events_option
),
1052 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1053 "event filter", parse_filter
),
1054 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1055 "child tasks do not inherit counters"),
1056 OPT_INTEGER('p', "pid", &target_pid
,
1057 "stat events on existing process id"),
1058 OPT_INTEGER('t', "tid", &target_tid
,
1059 "stat events on existing thread id"),
1060 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
1061 "system-wide collection from all CPUs"),
1062 OPT_BOOLEAN('g', "group", &group
,
1063 "put the counters into a counter group"),
1064 OPT_BOOLEAN('c', "scale", &scale
,
1065 "scale/normalize counters"),
1066 OPT_INCR('v', "verbose", &verbose
,
1067 "be more verbose (show counter open errors, etc)"),
1068 OPT_INTEGER('r', "repeat", &run_count
,
1069 "repeat command and print average + stddev (max: 100)"),
1070 OPT_BOOLEAN('n', "null", &null_run
,
1071 "null run - dont start any counters"),
1072 OPT_INCR('d', "detailed", &detailed_run
,
1073 "detailed run - start a lot of events"),
1074 OPT_BOOLEAN('S', "sync", &sync_run
,
1075 "call sync() before starting a run"),
1076 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1077 "print large numbers with thousands\' separators",
1079 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
1080 "list of cpus to monitor in system-wide"),
1081 OPT_BOOLEAN('A', "no-aggr", &no_aggr
,
1082 "disable CPU count aggregation"),
1083 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1084 "print counts with custom separator"),
1085 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1086 "monitor event in cgroup name only",
1088 OPT_STRING('o', "output", &output_name
, "file",
1089 "output file name"),
1090 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1091 OPT_INTEGER(0, "log-fd", &output_fd
,
1092 "log output to fd, instead of stderr"),
1097 * Add default attributes, if there were no attributes specified or
1098 * if -d/--detailed, -d -d or -d -d -d is used:
1100 static int add_default_attributes(void)
1102 struct perf_evsel
*pos
;
1106 /* Set attrs if no event is selected and !null_run: */
1110 if (!evsel_list
->nr_entries
) {
1111 for (c
= 0; c
< ARRAY_SIZE(default_attrs
); c
++) {
1112 pos
= perf_evsel__new(default_attrs
+ c
, c
+ attr_nr
);
1115 perf_evlist__add(evsel_list
, pos
);
1120 /* Detailed events get appended to the event list: */
1122 if (detailed_run
< 1)
1125 /* Append detailed run extra attributes: */
1126 for (c
= 0; c
< ARRAY_SIZE(detailed_attrs
); c
++) {
1127 pos
= perf_evsel__new(detailed_attrs
+ c
, c
+ attr_nr
);
1130 perf_evlist__add(evsel_list
, pos
);
1134 if (detailed_run
< 2)
1137 /* Append very detailed run extra attributes: */
1138 for (c
= 0; c
< ARRAY_SIZE(very_detailed_attrs
); c
++) {
1139 pos
= perf_evsel__new(very_detailed_attrs
+ c
, c
+ attr_nr
);
1142 perf_evlist__add(evsel_list
, pos
);
1145 if (detailed_run
< 3)
1148 /* Append very, very detailed run extra attributes: */
1149 for (c
= 0; c
< ARRAY_SIZE(very_very_detailed_attrs
); c
++) {
1150 pos
= perf_evsel__new(very_very_detailed_attrs
+ c
, c
+ attr_nr
);
1153 perf_evlist__add(evsel_list
, pos
);
1160 int cmd_stat(int argc
, const char **argv
, const char *prefix __used
)
1162 struct perf_evsel
*pos
;
1163 int status
= -ENOMEM
;
1166 setlocale(LC_ALL
, "");
1168 evsel_list
= perf_evlist__new(NULL
, NULL
);
1169 if (evsel_list
== NULL
)
1172 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1173 PARSE_OPT_STOP_AT_NON_OPTION
);
1176 if (output_name
&& strcmp(output_name
, "-"))
1179 if (output_name
&& output_fd
) {
1180 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1181 usage_with_options(stat_usage
, options
);
1185 mode
= append_file
? "a" : "w";
1187 output
= fopen(output_name
, mode
);
1189 perror("failed to create output file");
1192 clock_gettime(CLOCK_REALTIME
, &tm
);
1193 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1194 } else if (output_fd
!= 2) {
1195 mode
= append_file
? "a" : "w";
1196 output
= fdopen(output_fd
, mode
);
1198 perror("Failed opening logfd");
1205 if (!strcmp(csv_sep
, "\\t"))
1208 csv_sep
= DEFAULT_SEPARATOR
;
1211 * let the spreadsheet do the pretty-printing
1214 /* User explicitly passed -B? */
1215 if (big_num_opt
== 1) {
1216 fprintf(stderr
, "-B option not supported with -x\n");
1217 usage_with_options(stat_usage
, options
);
1218 } else /* Nope, so disable big number formatting */
1220 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1223 if (!argc
&& target_pid
== -1 && target_tid
== -1)
1224 usage_with_options(stat_usage
, options
);
1226 usage_with_options(stat_usage
, options
);
1228 /* no_aggr, cgroup are for system-wide only */
1229 if ((no_aggr
|| nr_cgroups
) && !system_wide
) {
1230 fprintf(stderr
, "both cgroup and no-aggregation "
1231 "modes only available in system-wide mode\n");
1233 usage_with_options(stat_usage
, options
);
1236 if (add_default_attributes())
1239 if (target_pid
!= -1)
1240 target_tid
= target_pid
;
1242 evsel_list
->threads
= thread_map__new(target_pid
, target_tid
);
1243 if (evsel_list
->threads
== NULL
) {
1244 pr_err("Problems finding threads of monitor\n");
1245 usage_with_options(stat_usage
, options
);
1249 evsel_list
->cpus
= cpu_map__new(cpu_list
);
1251 evsel_list
->cpus
= cpu_map__dummy_new();
1253 if (evsel_list
->cpus
== NULL
) {
1254 perror("failed to parse CPUs map");
1255 usage_with_options(stat_usage
, options
);
1259 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1260 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
1261 perf_evsel__alloc_counts(pos
, evsel_list
->cpus
->nr
) < 0 ||
1262 perf_evsel__alloc_fd(pos
, evsel_list
->cpus
->nr
, evsel_list
->threads
->nr
) < 0)
1267 * We dont want to block the signals - that would cause
1268 * child tasks to inherit that and Ctrl-C would not work.
1269 * What we want is for Ctrl-C to work in the exec()-ed
1270 * task, but being ignored by perf stat itself:
1273 signal(SIGINT
, skip_signal
);
1274 signal(SIGALRM
, skip_signal
);
1275 signal(SIGABRT
, skip_signal
);
1278 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
1279 if (run_count
!= 1 && verbose
)
1280 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1286 status
= run_perf_stat(argc
, argv
);
1290 print_stat(argc
, argv
);
1292 list_for_each_entry(pos
, &evsel_list
->entries
, node
)
1293 perf_evsel__free_stat_priv(pos
);
1294 perf_evlist__delete_maps(evsel_list
);
1296 perf_evlist__delete(evsel_list
);