2 * perf stat: /usr/bin/time -alike performance counter statistics utility
4 It summarizes the counter events of all tasks (and child tasks),
5 covering all CPUs that the command (or workload) executes on.
6 It only counts the per-task events of the workload started,
7 independent of how many other tasks run on those CPUs.
11 $ perf stat -e 1 -e 3 -e 5 ls -lR /usr/include/ >/dev/null
13 Performance counter stats for 'ls':
15 163516953 instructions
19 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
21 * Improvements and fixes by:
23 * Arjan van de Ven <arjan@linux.intel.com>
24 * Yanmin Zhang <yanmin.zhang@intel.com>
25 * Wu Fengguang <fengguang.wu@intel.com>
26 * Mike Galbraith <efault@gmx.de>
27 * Paul Mackerras <paulus@samba.org>
29 * Released under the GPL v2. (and only v2, not any later version)
34 #include "util/util.h"
35 #include "util/parse-options.h"
36 #include "util/parse-events.h"
38 #include <sys/prctl.h>
40 static int system_wide
= 0;
41 static int inherit
= 1;
43 static __u64 default_event_id
[MAX_COUNTERS
] = {
44 EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_TASK_CLOCK
),
45 EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_CONTEXT_SWITCHES
),
46 EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_CPU_MIGRATIONS
),
47 EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_PAGE_FAULTS
),
49 EID(PERF_TYPE_HARDWARE
, PERF_COUNT_CPU_CYCLES
),
50 EID(PERF_TYPE_HARDWARE
, PERF_COUNT_INSTRUCTIONS
),
51 EID(PERF_TYPE_HARDWARE
, PERF_COUNT_CACHE_REFERENCES
),
52 EID(PERF_TYPE_HARDWARE
, PERF_COUNT_CACHE_MISSES
),
55 static int default_interval
= 100000;
56 static int event_count
[MAX_COUNTERS
];
57 static int fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
59 static int target_pid
= -1;
60 static int nr_cpus
= 0;
61 static unsigned int page_size
;
65 static const unsigned int default_count
[] = {
74 static void create_perfstat_counter(int counter
)
76 struct perf_counter_hw_event hw_event
;
78 memset(&hw_event
, 0, sizeof(hw_event
));
79 hw_event
.config
= event_id
[counter
];
80 hw_event
.record_type
= 0;
82 hw_event
.exclude_kernel
= event_mask
[counter
] & EVENT_MASK_KERNEL
;
83 hw_event
.exclude_user
= event_mask
[counter
] & EVENT_MASK_USER
;
86 hw_event
.read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
87 PERF_FORMAT_TOTAL_TIME_RUNNING
;
91 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
92 fd
[cpu
][counter
] = sys_perf_counter_open(&hw_event
, -1, cpu
, -1, 0);
93 if (fd
[cpu
][counter
] < 0) {
94 printf("perfstat error: syscall returned with %d (%s)\n",
95 fd
[cpu
][counter
], strerror(errno
));
100 hw_event
.inherit
= inherit
;
101 hw_event
.disabled
= 1;
103 fd
[0][counter
] = sys_perf_counter_open(&hw_event
, 0, -1, -1, 0);
104 if (fd
[0][counter
] < 0) {
105 printf("perfstat error: syscall returned with %d (%s)\n",
106 fd
[0][counter
], strerror(errno
));
113 * Does the counter have nsecs as a unit?
115 static inline int nsec_counter(int counter
)
117 if (event_id
[counter
] == EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_CPU_CLOCK
))
119 if (event_id
[counter
] == EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_TASK_CLOCK
))
126 * Print out the results of a single counter:
128 static void print_counter(int counter
)
130 __u64 count
[3], single_count
[3];
135 count
[0] = count
[1] = count
[2] = 0;
137 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
138 res
= read(fd
[cpu
][counter
], single_count
, nv
* sizeof(__u64
));
139 assert(res
== nv
* sizeof(__u64
));
141 count
[0] += single_count
[0];
143 count
[1] += single_count
[1];
144 count
[2] += single_count
[2];
151 fprintf(stderr
, " %14s %-20s\n",
152 "<not counted>", event_name(counter
));
155 if (count
[2] < count
[1]) {
157 count
[0] = (unsigned long long)
158 ((double)count
[0] * count
[1] / count
[2] + 0.5);
162 if (nsec_counter(counter
)) {
163 double msecs
= (double)count
[0] / 1000000;
165 fprintf(stderr
, " %14.6f %-20s (msecs)",
166 msecs
, event_name(counter
));
168 fprintf(stderr
, " %14Ld %-20s (events)",
169 count
[0], event_name(counter
));
172 fprintf(stderr
, " (scaled from %.2f%%)",
173 (double) count
[2] / count
[1] * 100);
174 fprintf(stderr
, "\n");
177 static int do_perfstat(int argc
, const char **argv
)
179 unsigned long long t0
, t1
;
187 for (counter
= 0; counter
< nr_counters
; counter
++)
188 create_perfstat_counter(counter
);
191 * Enable counters and exec the command:
194 prctl(PR_TASK_PERF_COUNTERS_ENABLE
);
196 if ((pid
= fork()) < 0)
197 perror("failed to fork");
199 if (execvp(argv
[0], (char **)argv
)) {
204 while (wait(&status
) >= 0)
206 prctl(PR_TASK_PERF_COUNTERS_DISABLE
);
211 fprintf(stderr
, "\n");
212 fprintf(stderr
, " Performance counter stats for \'%s\':\n",
214 fprintf(stderr
, "\n");
216 for (counter
= 0; counter
< nr_counters
; counter
++)
217 print_counter(counter
);
220 fprintf(stderr
, "\n");
221 fprintf(stderr
, " Wall-clock time elapsed: %12.6f msecs\n",
222 (double)(t1
-t0
)/1e6
);
223 fprintf(stderr
, "\n");
228 static void skip_signal(int signo
)
232 static const char * const stat_usage
[] = {
233 "perf stat [<options>] <command>",
237 static char events_help_msg
[EVENTS_HELP_MAX
];
239 static const struct option options
[] = {
240 OPT_CALLBACK('e', "event", NULL
, "event",
241 events_help_msg
, parse_events
),
242 OPT_INTEGER('c', "count", &default_interval
,
243 "event period to sample"),
244 OPT_BOOLEAN('i', "inherit", &inherit
,
245 "child tasks inherit counters"),
246 OPT_INTEGER('p', "pid", &target_pid
,
247 "stat events on existing pid"),
248 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
249 "system-wide collection from all CPUs"),
250 OPT_BOOLEAN('l', "scale", &scale
,
251 "scale/normalize counters"),
255 int cmd_stat(int argc
, const char **argv
, const char *prefix
)
259 page_size
= sysconf(_SC_PAGE_SIZE
);
261 create_events_help(events_help_msg
);
262 memcpy(event_id
, default_event_id
, sizeof(default_event_id
));
264 argc
= parse_options(argc
, argv
, options
, stat_usage
, 0);
266 usage_with_options(stat_usage
, options
);
272 for (counter
= 0; counter
< nr_counters
; counter
++) {
273 if (event_count
[counter
])
276 event_count
[counter
] = default_interval
;
278 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
279 assert(nr_cpus
<= MAX_NR_CPUS
);
280 assert(nr_cpus
>= 0);
283 * We dont want to block the signals - that would cause
284 * child tasks to inherit that and Ctrl-C would not work.
285 * What we want is for Ctrl-C to work in the exec()-ed
286 * task, but being ignored by perf stat itself:
288 signal(SIGINT
, skip_signal
);
289 signal(SIGALRM
, skip_signal
);
290 signal(SIGABRT
, skip_signal
);
292 return do_perfstat(argc
, argv
);