perf top tui: Wait till the first sample to refresh the screen.
[linux-2.6-block.git] / tools / perf / builtin-record.c
CommitLineData
abaff32a 1/*
bf9e1876
IM
2 * builtin-record.c
3 *
4 * Builtin record command: Record the profile of a workload
5 * (or a CPU, or a PID) into the perf.data output file - for
6 * later analysis via perf report.
abaff32a 7 */
b8f46c5a
XG
8#define _FILE_OFFSET_BITS 64
9
16f762a2 10#include "builtin.h"
bf9e1876
IM
11
12#include "perf.h"
13
6122e4e4 14#include "util/build-id.h"
6eda5838 15#include "util/util.h"
0e9b20b8 16#include "util/parse-options.h"
8ad8db37 17#include "util/parse-events.h"
6eda5838 18
7c6a1c65 19#include "util/header.h"
66e274f3 20#include "util/event.h"
361c99a6 21#include "util/evlist.h"
69aad6f1 22#include "util/evsel.h"
8f28827a 23#include "util/debug.h"
94c744b6 24#include "util/session.h"
8d06367f 25#include "util/symbol.h"
a12b51c4 26#include "util/cpumap.h"
fd78260b 27#include "util/thread_map.h"
7c6a1c65 28
97124d5e 29#include <unistd.h>
de9ac07b 30#include <sched.h>
a41794cd 31#include <sys/mman.h>
de9ac07b 32
69aad6f1 33#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
0a27d7f9 34#define SID(e, x, y) xyarray__entry(e->id, x, y)
69aad6f1 35
7865e817
FW
36enum write_mode_t {
37 WRITE_FORCE,
38 WRITE_APPEND
39};
40
3de29cab
SE
41static u64 user_interval = ULLONG_MAX;
42static u64 default_interval = 0;
640c03ce 43static u64 sample_type;
a21ca2ca 44
de9ac07b 45static unsigned int page_size;
42e59d7d 46static unsigned int mmap_pages = 128;
f9212819 47static unsigned int user_freq = UINT_MAX;
42e59d7d 48static int freq = 1000;
de9ac07b 49static int output;
529870e3 50static int pipe_output = 0;
d7065adb 51static const char *output_name = NULL;
42e59d7d 52static int group = 0;
1967936d 53static int realtime_prio = 0;
acac03fa 54static bool nodelay = false;
c0555642 55static bool raw_samples = false;
9c90a61c 56static bool sample_id_all_avail = true;
c0555642 57static bool system_wide = false;
42e59d7d 58static pid_t target_pid = -1;
d6d901c2 59static pid_t target_tid = -1;
42e59d7d 60static pid_t child_pid = -1;
2e6cdf99 61static bool no_inherit = false;
7865e817 62static enum write_mode_t write_mode = WRITE_FORCE;
c0555642
IM
63static bool call_graph = false;
64static bool inherit_stat = false;
65static bool no_samples = false;
66static bool sample_address = false;
9c90a61c 67static bool sample_time = false;
a1ac1d3c 68static bool no_buildid = false;
baa2f6ce 69static bool no_buildid_cache = false;
361c99a6 70static struct perf_evlist *evsel_list;
42e59d7d
IM
71
72static long samples = 0;
42e59d7d 73static u64 bytes_written = 0;
a21ca2ca 74
42e59d7d 75static int file_new = 1;
6122e4e4 76static off_t post_processing_offset;
7c6a1c65 77
94c744b6 78static struct perf_session *session;
c45c6ea2 79static const char *cpu_list;
f5970550 80
9215545e
TZ
81static void advance_output(size_t size)
82{
83 bytes_written += size;
84}
85
f5970550
PZ
86static void write_output(void *buf, size_t size)
87{
88 while (size) {
89 int ret = write(output, buf, size);
90
91 if (ret < 0)
92 die("failed to write");
93
94 size -= ret;
95 buf += ret;
96
97 bytes_written += ret;
98 }
99}
100
8115d60c 101static int process_synthesized_event(union perf_event *event,
8d50e5b4 102 struct perf_sample *sample __used,
d8f66248 103 struct perf_session *self __used)
234fbbf5 104{
6122e4e4 105 write_output(event, event->header.size);
234fbbf5
ACM
106 return 0;
107}
108
744bd8aa 109static void mmap_read(struct perf_mmap *md)
de9ac07b 110{
744bd8aa 111 unsigned int head = perf_mmap__read_head(md);
de9ac07b
PZ
112 unsigned int old = md->prev;
113 unsigned char *data = md->base + page_size;
114 unsigned long size;
115 void *buf;
de9ac07b 116
dc82009a
ACM
117 if (old == head)
118 return;
119
120 samples++;
de9ac07b
PZ
121
122 size = head - old;
123
124 if ((old & md->mask) + size != (head & md->mask)) {
125 buf = &data[old & md->mask];
126 size = md->mask + 1 - (old & md->mask);
127 old += size;
021e9f47 128
6122e4e4 129 write_output(buf, size);
de9ac07b
PZ
130 }
131
132 buf = &data[old & md->mask];
133 size = head - old;
134 old += size;
021e9f47 135
6122e4e4 136 write_output(buf, size);
de9ac07b
PZ
137
138 md->prev = old;
115d2d89 139 perf_mmap__write_tail(md, old);
de9ac07b
PZ
140}
141
142static volatile int done = 0;
f7b7c26e 143static volatile int signr = -1;
de9ac07b 144
16c8a109 145static void sig_handler(int sig)
de9ac07b 146{
16c8a109 147 done = 1;
f7b7c26e
PZ
148 signr = sig;
149}
150
151static void sig_atexit(void)
152{
5ffc8881 153 if (child_pid > 0)
933da83a
CW
154 kill(child_pid, SIGTERM);
155
18483b81 156 if (signr == -1 || signr == SIGUSR1)
f7b7c26e
PZ
157 return;
158
159 signal(signr, SIG_DFL);
160 kill(getpid(), signr);
de9ac07b
PZ
161}
162
cdd6c482 163static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
7c6a1c65
PZ
164{
165 struct perf_header_attr *h_attr;
166
94c744b6
ACM
167 if (nr < session->header.attrs) {
168 h_attr = session->header.attr[nr];
7c6a1c65
PZ
169 } else {
170 h_attr = perf_header_attr__new(a);
dc79c0fc 171 if (h_attr != NULL)
94c744b6 172 if (perf_header__add_attr(&session->header, h_attr) < 0) {
11deb1f9
ACM
173 perf_header_attr__delete(h_attr);
174 h_attr = NULL;
175 }
7c6a1c65
PZ
176 }
177
178 return h_attr;
179}
180
0a27d7f9 181static void create_counter(struct perf_evsel *evsel, int cpu)
de9ac07b 182{
69aad6f1
ACM
183 char *filter = evsel->filter;
184 struct perf_event_attr *attr = &evsel->attr;
7c6a1c65 185 struct perf_header_attr *h_attr;
0a27d7f9 186 struct perf_sample_id *sid;
d6d901c2 187 int thread_index;
c171b552 188 int ret;
dd7927f4 189
7e2ed097 190 for (thread_index = 0; thread_index < evsel_list->threads->nr; thread_index++) {
dd7927f4
ACM
191 h_attr = get_header_attr(attr, evsel->idx);
192 if (h_attr == NULL)
193 die("nomem\n");
194
195 if (!file_new) {
196 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
197 fprintf(stderr, "incompatible append\n");
198 exit(-1);
199 }
200 }
201
0a27d7f9
ACM
202 sid = SID(evsel, cpu, thread_index);
203 if (perf_header_attr__add_id(h_attr, sid->id) < 0) {
dd7927f4
ACM
204 pr_warning("Not enough memory to add id\n");
205 exit(-1);
206 }
207
dd7927f4
ACM
208 if (filter != NULL) {
209 ret = ioctl(FD(evsel, cpu, thread_index),
210 PERF_EVENT_IOC_SET_FILTER, filter);
211 if (ret) {
212 error("failed to set filter with %d (%s)\n", errno,
213 strerror(errno));
214 exit(-1);
215 }
216 }
217 }
218
219 if (!sample_type)
220 sample_type = attr->sample_type;
221}
222
223static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
224{
225 struct perf_event_attr *attr = &evsel->attr;
226 int track = !evsel->idx; /* only the first counter needs these */
7c6a1c65
PZ
227
228 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
229 PERF_FORMAT_TOTAL_TIME_RUNNING |
230 PERF_FORMAT_ID;
16c8a109 231
3a9f131f 232 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
3efa1cc9 233
361c99a6 234 if (evlist->nr_entries > 1)
8907fd60
EM
235 attr->sample_type |= PERF_SAMPLE_ID;
236
f9212819
FW
237 /*
238 * We default some events to a 1 default interval. But keep
239 * it a weak assumption overridable by the user.
240 */
241 if (!attr->sample_period || (user_freq != UINT_MAX &&
3de29cab 242 user_interval != ULLONG_MAX)) {
f9212819
FW
243 if (freq) {
244 attr->sample_type |= PERF_SAMPLE_PERIOD;
245 attr->freq = 1;
246 attr->sample_freq = freq;
247 } else {
248 attr->sample_period = default_interval;
249 }
1dba15e7 250 }
3efa1cc9 251
649c48a9
PZ
252 if (no_samples)
253 attr->sample_freq = 0;
254
255 if (inherit_stat)
256 attr->inherit_stat = 1;
257
3af9e859 258 if (sample_address) {
4bba828d 259 attr->sample_type |= PERF_SAMPLE_ADDR;
3af9e859
EM
260 attr->mmap_data = track;
261 }
4bba828d 262
3efa1cc9
IM
263 if (call_graph)
264 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
265
f60f3593
AS
266 if (system_wide)
267 attr->sample_type |= PERF_SAMPLE_CPU;
268
a43d3f08
ACM
269 if (sample_id_all_avail &&
270 (sample_time || system_wide || !no_inherit || cpu_list))
9c90a61c
ACM
271 attr->sample_type |= PERF_SAMPLE_TIME;
272
cd6feeea 273 if (raw_samples) {
6ddf259d 274 attr->sample_type |= PERF_SAMPLE_TIME;
daac07b2 275 attr->sample_type |= PERF_SAMPLE_RAW;
cd6feeea
IM
276 attr->sample_type |= PERF_SAMPLE_CPU;
277 }
f413cdb8 278
acac03fa
KS
279 if (nodelay) {
280 attr->watermark = 0;
281 attr->wakeup_events = 1;
282 }
283
a21ca2ca
IM
284 attr->mmap = track;
285 attr->comm = track;
dd7927f4 286
2e6cdf99 287 if (target_pid == -1 && target_tid == -1 && !system_wide) {
46be604b 288 attr->disabled = 1;
bedbfdea 289 attr->enable_on_exec = 1;
46be604b 290 }
dd7927f4 291}
bedbfdea 292
dd7927f4
ACM
293static void open_counters(struct perf_evlist *evlist)
294{
295 struct perf_evsel *pos;
296 int cpu;
297
298 list_for_each_entry(pos, &evlist->entries, node) {
299 struct perf_event_attr *attr = &pos->attr;
300 /*
301 * Check if parse_single_tracepoint_event has already asked for
302 * PERF_SAMPLE_TIME.
303 *
304 * XXX this is kludgy but short term fix for problems introduced by
305 * eac23d1c that broke 'perf script' by having different sample_types
306 * when using multiple tracepoint events when we use a perf binary
307 * that tries to use sample_id_all on an older kernel.
308 *
309 * We need to move counter creation to perf_session, support
310 * different sample_types, etc.
311 */
312 bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
d6d901c2 313
dd7927f4
ACM
314 config_attr(pos, evlist);
315retry_sample_id:
316 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
317try_again:
7e2ed097
ACM
318 if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group,
319 !no_inherit) < 0) {
d6d901c2
ZY
320 int err = errno;
321
322 if (err == EPERM || err == EACCES)
323 die("Permission error - are you root?\n"
324 "\t Consider tweaking"
325 " /proc/sys/kernel/perf_event_paranoid.\n");
c45c6ea2 326 else if (err == ENODEV && cpu_list) {
d6d901c2
ZY
327 die("No such device - did you specify"
328 " an out-of-range profile CPU?\n");
9c90a61c
ACM
329 } else if (err == EINVAL && sample_id_all_avail) {
330 /*
331 * Old kernel, no attr->sample_id_type_all field
332 */
333 sample_id_all_avail = false;
a43d3f08 334 if (!sample_time && !raw_samples && !time_needed)
eac23d1c
IM
335 attr->sample_type &= ~PERF_SAMPLE_TIME;
336
9c90a61c 337 goto retry_sample_id;
d6d901c2 338 }
3da297a6 339
d6d901c2
ZY
340 /*
341 * If it's cycles then fall back to hrtimer
342 * based cpu-clock-tick sw counter, which
343 * is always available even if no PMU support:
344 */
345 if (attr->type == PERF_TYPE_HARDWARE
346 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
347
348 if (verbose)
349 warning(" ... trying to fall back to cpu-clock-ticks\n");
350 attr->type = PERF_TYPE_SOFTWARE;
351 attr->config = PERF_COUNT_SW_CPU_CLOCK;
352 goto try_again;
353 }
354 printf("\n");
d9cf837e 355 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
dd7927f4 356 err, strerror(err));
bfd45118
SK
357
358#if defined(__i386__) || defined(__x86_64__)
d6d901c2
ZY
359 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
360 die("No hardware sampling interrupt available."
361 " No APIC? If so then you can boot the kernel"
362 " with the \"lapic\" boot parameter to"
363 " force-enable it.\n");
bfd45118
SK
364#endif
365
d6d901c2 366 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
c171b552
LZ
367 }
368 }
a43d3f08 369
7e2ed097 370 if (perf_evlist__mmap(evlist, mmap_pages, false) < 0)
0a27d7f9
ACM
371 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
372
7e2ed097 373 for (cpu = 0; cpu < evsel_list->cpus->nr; ++cpu) {
dd7927f4 374 list_for_each_entry(pos, &evlist->entries, node)
0a27d7f9 375 create_counter(pos, cpu);
dd7927f4 376 }
16c8a109
PZ
377}
378
6122e4e4
ACM
379static int process_buildids(void)
380{
381 u64 size = lseek(output, 0, SEEK_CUR);
382
9f591fd7
ACM
383 if (size == 0)
384 return 0;
385
6122e4e4
ACM
386 session->fd = output;
387 return __perf_session__process_events(session, post_processing_offset,
388 size - post_processing_offset,
389 size, &build_id__mark_dso_hit_ops);
390}
391
f5970550
PZ
392static void atexit_header(void)
393{
c7929e47
TZ
394 if (!pipe_output) {
395 session->header.data_size += bytes_written;
f5970550 396
baa2f6ce
ACM
397 if (!no_buildid)
398 process_buildids();
361c99a6 399 perf_header__write(&session->header, evsel_list, output, true);
39d17dac 400 perf_session__delete(session);
361c99a6 401 perf_evlist__delete(evsel_list);
d65a458b 402 symbol__exit();
c7929e47 403 }
f5970550
PZ
404}
405
8115d60c 406static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
a1645ce1
ZY
407{
408 int err;
23346f21 409 struct perf_session *psession = data;
a1645ce1 410
23346f21 411 if (machine__is_host(machine))
a1645ce1
ZY
412 return;
413
414 /*
415 *As for guest kernel when processing subcommand record&report,
416 *we arrange module mmap prior to guest kernel mmap and trigger
417 *a preload dso because default guest module symbols are loaded
418 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
419 *method is used to avoid symbol missing when the first addr is
420 *in module instead of in guest kernel.
421 */
8115d60c
ACM
422 err = perf_event__synthesize_modules(process_synthesized_event,
423 psession, machine);
a1645ce1
ZY
424 if (err < 0)
425 pr_err("Couldn't record guest kernel [%d]'s reference"
23346f21 426 " relocation symbol.\n", machine->pid);
a1645ce1 427
a1645ce1
ZY
428 /*
429 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
430 * have no _text sometimes.
431 */
8115d60c
ACM
432 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
433 psession, machine, "_text");
a1645ce1 434 if (err < 0)
8115d60c
ACM
435 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
436 psession, machine,
437 "_stext");
a1645ce1
ZY
438 if (err < 0)
439 pr_err("Couldn't record guest kernel [%d]'s reference"
23346f21 440 " relocation symbol.\n", machine->pid);
a1645ce1
ZY
441}
442
98402807
FW
443static struct perf_event_header finished_round_event = {
444 .size = sizeof(struct perf_event_header),
445 .type = PERF_RECORD_FINISHED_ROUND,
446};
447
448static void mmap_read_all(void)
449{
0e2e63dd 450 int i;
98402807 451
7e2ed097 452 for (i = 0; i < evsel_list->cpus->nr; i++) {
0a27d7f9
ACM
453 if (evsel_list->mmap[i].base)
454 mmap_read(&evsel_list->mmap[i]);
98402807
FW
455 }
456
457 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
458 write_output(&finished_round_event, sizeof(finished_round_event));
459}
460
d4db3f16 461static int __cmd_record(int argc, const char **argv)
16c8a109 462{
69aad6f1 463 int i;
abaff32a 464 struct stat st;
abaff32a 465 int flags;
4dc0a04b 466 int err;
8b412664 467 unsigned long waking = 0;
856e9660 468 int child_ready_pipe[2], go_pipe[2];
46be604b 469 const bool forks = argc > 0;
856e9660 470 char buf;
23346f21 471 struct machine *machine;
de9ac07b
PZ
472
473 page_size = sysconf(_SC_PAGE_SIZE);
de9ac07b 474
f5970550
PZ
475 atexit(sig_atexit);
476 signal(SIGCHLD, sig_handler);
477 signal(SIGINT, sig_handler);
18483b81 478 signal(SIGUSR1, sig_handler);
f5970550 479
d4db3f16 480 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
856e9660
PZ
481 perror("failed to create pipes");
482 exit(-1);
483 }
484
d7065adb
FBH
485 if (!output_name) {
486 if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
487 pipe_output = 1;
488 else
489 output_name = "perf.data";
490 }
491 if (output_name) {
492 if (!strcmp(output_name, "-"))
493 pipe_output = 1;
494 else if (!stat(output_name, &st) && st.st_size) {
495 if (write_mode == WRITE_FORCE) {
496 char oldname[PATH_MAX];
497 snprintf(oldname, sizeof(oldname), "%s.old",
498 output_name);
499 unlink(oldname);
500 rename(output_name, oldname);
501 }
502 } else if (write_mode == WRITE_APPEND) {
503 write_mode = WRITE_FORCE;
266e0e21 504 }
97124d5e
PZ
505 }
506
f887f301 507 flags = O_CREAT|O_RDWR;
7865e817 508 if (write_mode == WRITE_APPEND)
f5970550 509 file_new = 0;
abaff32a
IM
510 else
511 flags |= O_TRUNC;
512
529870e3
TZ
513 if (pipe_output)
514 output = STDOUT_FILENO;
515 else
516 output = open(output_name, flags, S_IRUSR | S_IWUSR);
de9ac07b
PZ
517 if (output < 0) {
518 perror("failed to create output file");
519 exit(-1);
520 }
521
7865e817 522 session = perf_session__new(output_name, O_WRONLY,
21ef97f0 523 write_mode == WRITE_FORCE, false, NULL);
94c744b6 524 if (session == NULL) {
a9a70bbc
ACM
525 pr_err("Not enough memory for reading perf file header\n");
526 return -1;
527 }
528
baa2f6ce
ACM
529 if (!no_buildid)
530 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
531
4dc0a04b 532 if (!file_new) {
8dc58101 533 err = perf_header__read(session, output);
4dc0a04b 534 if (err < 0)
39d17dac 535 goto out_delete_session;
4dc0a04b
ACM
536 }
537
361c99a6 538 if (have_tracepoints(&evsel_list->entries))
94c744b6 539 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
03456a15 540
d4db3f16 541 if (forks) {
46be604b 542 child_pid = fork();
2fb750e8 543 if (child_pid < 0) {
856e9660
PZ
544 perror("failed to fork");
545 exit(-1);
546 }
7c6a1c65 547
46be604b 548 if (!child_pid) {
529870e3
TZ
549 if (pipe_output)
550 dup2(2, 1);
856e9660
PZ
551 close(child_ready_pipe[0]);
552 close(go_pipe[1]);
553 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
554
555 /*
556 * Do a dummy execvp to get the PLT entry resolved,
557 * so we avoid the resolver overhead on the real
558 * execvp call.
559 */
560 execvp("", (char **)argv);
561
562 /*
563 * Tell the parent we're ready to go
564 */
565 close(child_ready_pipe[1]);
566
567 /*
568 * Wait until the parent tells us to go.
569 */
570 if (read(go_pipe[0], &buf, 1) == -1)
571 perror("unable to read pipe");
572
573 execvp(argv[0], (char **)argv);
574
575 perror(argv[0]);
18483b81 576 kill(getppid(), SIGUSR1);
856e9660 577 exit(-1);
0a5ac846 578 }
856e9660 579
d6d901c2 580 if (!system_wide && target_tid == -1 && target_pid == -1)
7e2ed097 581 evsel_list->threads->map[0] = child_pid;
d6d901c2 582
856e9660
PZ
583 close(child_ready_pipe[1]);
584 close(go_pipe[0]);
585 /*
586 * wait for child to settle
587 */
588 if (read(child_ready_pipe[0], &buf, 1) == -1) {
589 perror("unable to read pipe");
590 exit(-1);
591 }
592 close(child_ready_pipe[0]);
593 }
594
dd7927f4 595 open_counters(evsel_list);
de9ac07b 596
640c03ce
ACM
597 perf_session__set_sample_type(session, sample_type);
598
712a4b60
ACM
599 /*
600 * perf_session__delete(session) will be called at atexit_header()
601 */
602 atexit(atexit_header);
603
529870e3
TZ
604 if (pipe_output) {
605 err = perf_header__write_pipe(output);
606 if (err < 0)
607 return err;
608 } else if (file_new) {
361c99a6
ACM
609 err = perf_header__write(&session->header, evsel_list,
610 output, false);
d5eed904
ACM
611 if (err < 0)
612 return err;
56b03f3c
ACM
613 }
614
6122e4e4
ACM
615 post_processing_offset = lseek(output, 0, SEEK_CUR);
616
9c90a61c
ACM
617 perf_session__set_sample_id_all(session, sample_id_all_avail);
618
2c46dbb5 619 if (pipe_output) {
8115d60c
ACM
620 err = perf_event__synthesize_attrs(&session->header,
621 process_synthesized_event,
622 session);
2c46dbb5
TZ
623 if (err < 0) {
624 pr_err("Couldn't synthesize attrs.\n");
625 return err;
626 }
cd19a035 627
8115d60c
ACM
628 err = perf_event__synthesize_event_types(process_synthesized_event,
629 session);
cd19a035
TZ
630 if (err < 0) {
631 pr_err("Couldn't synthesize event_types.\n");
632 return err;
633 }
9215545e 634
361c99a6 635 if (have_tracepoints(&evsel_list->entries)) {
63e0c771
TZ
636 /*
637 * FIXME err <= 0 here actually means that
638 * there were no tracepoints so its not really
639 * an error, just that we don't need to
640 * synthesize anything. We really have to
641 * return this more properly and also
642 * propagate errors that now are calling die()
643 */
8115d60c
ACM
644 err = perf_event__synthesize_tracing_data(output, evsel_list,
645 process_synthesized_event,
646 session);
63e0c771
TZ
647 if (err <= 0) {
648 pr_err("Couldn't record tracing data.\n");
649 return err;
650 }
2c9faa06 651 advance_output(err);
63e0c771 652 }
2c46dbb5
TZ
653 }
654
23346f21
ACM
655 machine = perf_session__find_host_machine(session);
656 if (!machine) {
a1645ce1
ZY
657 pr_err("Couldn't find native kernel information.\n");
658 return -1;
659 }
660
8115d60c
ACM
661 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
662 session, machine, "_text");
70162138 663 if (err < 0)
8115d60c
ACM
664 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
665 session, machine, "_stext");
c1a3a4b9
ACM
666 if (err < 0)
667 pr_err("Couldn't record kernel reference relocation symbol\n"
668 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
669 "Check /proc/kallsyms permission or run as root.\n");
b7cece76 670
8115d60c
ACM
671 err = perf_event__synthesize_modules(process_synthesized_event,
672 session, machine);
c1a3a4b9
ACM
673 if (err < 0)
674 pr_err("Couldn't record kernel module information.\n"
675 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
676 "Check /proc/modules permission or run as root.\n");
677
a1645ce1 678 if (perf_guest)
8115d60c
ACM
679 perf_session__process_machines(session,
680 perf_event__synthesize_guest_os);
7c6a1c65 681
cf103a14 682 if (!system_wide)
7c940c18
ACM
683 perf_event__synthesize_thread_map(evsel_list->threads,
684 process_synthesized_event,
685 session);
234fbbf5 686 else
8115d60c
ACM
687 perf_event__synthesize_threads(process_synthesized_event,
688 session);
7c6a1c65 689
de9ac07b
PZ
690 if (realtime_prio) {
691 struct sched_param param;
692
693 param.sched_priority = realtime_prio;
694 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
6beba7ad 695 pr_err("Could not set realtime priority.\n");
de9ac07b
PZ
696 exit(-1);
697 }
698 }
699
856e9660
PZ
700 /*
701 * Let the child rip
702 */
d4db3f16
ACM
703 if (forks)
704 close(go_pipe[1]);
856e9660 705
649c48a9 706 for (;;) {
2debbc83 707 int hits = samples;
d6d901c2 708 int thread;
de9ac07b 709
98402807 710 mmap_read_all();
de9ac07b 711
649c48a9
PZ
712 if (hits == samples) {
713 if (done)
714 break;
5c581041 715 err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
8b412664
PZ
716 waking++;
717 }
718
719 if (done) {
7e2ed097 720 for (i = 0; i < evsel_list->cpus->nr; i++) {
69aad6f1
ACM
721 struct perf_evsel *pos;
722
361c99a6 723 list_for_each_entry(pos, &evsel_list->entries, node) {
d6d901c2 724 for (thread = 0;
7e2ed097 725 thread < evsel_list->threads->nr;
d6d901c2 726 thread++)
69aad6f1 727 ioctl(FD(pos, i, thread),
d6d901c2
ZY
728 PERF_EVENT_IOC_DISABLE);
729 }
8b412664 730 }
649c48a9 731 }
de9ac07b
PZ
732 }
733
18483b81 734 if (quiet || signr == SIGUSR1)
b44308f5
ACM
735 return 0;
736
8b412664
PZ
737 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
738
021e9f47
IM
739 /*
740 * Approximate RIP event size: 24 bytes.
741 */
742 fprintf(stderr,
9486aa38 743 "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
021e9f47
IM
744 (double)bytes_written / 1024.0 / 1024.0,
745 output_name,
746 bytes_written / 24);
addc2785 747
de9ac07b 748 return 0;
39d17dac
ACM
749
750out_delete_session:
751 perf_session__delete(session);
752 return err;
de9ac07b 753}
0e9b20b8 754
0e9b20b8 755static const char * const record_usage[] = {
9e096753
MG
756 "perf record [<options>] [<command>]",
757 "perf record [<options>] -- <command> [<options>]",
0e9b20b8
IM
758 NULL
759};
760
7865e817
FW
761static bool force, append_file;
762
bca647aa 763const struct option record_options[] = {
361c99a6 764 OPT_CALLBACK('e', "event", &evsel_list, "event",
86847b62
TG
765 "event selector. use 'perf list' to list available events",
766 parse_events),
361c99a6 767 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
c171b552 768 "event filter", parse_filter),
0e9b20b8 769 OPT_INTEGER('p', "pid", &target_pid,
d6d901c2
ZY
770 "record events on existing process id"),
771 OPT_INTEGER('t', "tid", &target_tid,
772 "record events on existing thread id"),
0e9b20b8
IM
773 OPT_INTEGER('r', "realtime", &realtime_prio,
774 "collect data with this RT SCHED_FIFO priority"),
acac03fa
KS
775 OPT_BOOLEAN('D', "no-delay", &nodelay,
776 "collect data without buffering"),
daac07b2
FW
777 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
778 "collect raw sample records from all opened counters"),
0e9b20b8
IM
779 OPT_BOOLEAN('a', "all-cpus", &system_wide,
780 "system-wide collection from all CPUs"),
abaff32a
IM
781 OPT_BOOLEAN('A', "append", &append_file,
782 "append to the output file to do incremental profiling"),
c45c6ea2
SE
783 OPT_STRING('C', "cpu", &cpu_list, "cpu",
784 "list of cpus to monitor"),
97124d5e 785 OPT_BOOLEAN('f', "force", &force,
7865e817 786 "overwrite existing data file (deprecated)"),
3de29cab 787 OPT_U64('c', "count", &user_interval, "event period to sample"),
abaff32a
IM
788 OPT_STRING('o', "output", &output_name, "file",
789 "output file name"),
2e6cdf99
SE
790 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
791 "child tasks do not inherit counters"),
1967936d
ACM
792 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
793 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
3efa1cc9
IM
794 OPT_BOOLEAN('g', "call-graph", &call_graph,
795 "do call-graph (stack chain/backtrace) recording"),
c0555642 796 OPT_INCR('v', "verbose", &verbose,
3da297a6 797 "be more verbose (show counter open errors, etc)"),
b44308f5 798 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
649c48a9
PZ
799 OPT_BOOLEAN('s', "stat", &inherit_stat,
800 "per thread counts"),
4bba828d
AB
801 OPT_BOOLEAN('d', "data", &sample_address,
802 "Sample addresses"),
9c90a61c 803 OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
649c48a9
PZ
804 OPT_BOOLEAN('n', "no-samples", &no_samples,
805 "don't sample"),
baa2f6ce 806 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
a1ac1d3c 807 "do not update the buildid cache"),
baa2f6ce
ACM
808 OPT_BOOLEAN('B', "no-buildid", &no_buildid,
809 "do not collect buildids in perf.data"),
023695d9
SE
810 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
811 "monitor event in cgroup name only",
812 parse_cgroups),
0e9b20b8
IM
813 OPT_END()
814};
815
f37a291c 816int cmd_record(int argc, const char **argv, const char *prefix __used)
0e9b20b8 817{
69aad6f1
ACM
818 int err = -ENOMEM;
819 struct perf_evsel *pos;
0e9b20b8 820
7e2ed097 821 evsel_list = perf_evlist__new(NULL, NULL);
361c99a6
ACM
822 if (evsel_list == NULL)
823 return -ENOMEM;
824
bca647aa 825 argc = parse_options(argc, argv, record_options, record_usage,
655000e7 826 PARSE_OPT_STOP_AT_NON_OPTION);
d6d901c2 827 if (!argc && target_pid == -1 && target_tid == -1 &&
c45c6ea2 828 !system_wide && !cpu_list)
bca647aa 829 usage_with_options(record_usage, record_options);
0e9b20b8 830
7865e817
FW
831 if (force && append_file) {
832 fprintf(stderr, "Can't overwrite and append at the same time."
833 " You need to choose between -f and -A");
bca647aa 834 usage_with_options(record_usage, record_options);
7865e817
FW
835 } else if (append_file) {
836 write_mode = WRITE_APPEND;
837 } else {
838 write_mode = WRITE_FORCE;
839 }
840
023695d9
SE
841 if (nr_cgroups && !system_wide) {
842 fprintf(stderr, "cgroup monitoring only available in"
843 " system-wide mode\n");
844 usage_with_options(record_usage, record_options);
845 }
846
655000e7 847 symbol__init();
baa2f6ce
ACM
848
849 if (no_buildid_cache || no_buildid)
a1ac1d3c 850 disable_buildid_cache();
655000e7 851
361c99a6
ACM
852 if (evsel_list->nr_entries == 0 &&
853 perf_evlist__add_default(evsel_list) < 0) {
69aad6f1
ACM
854 pr_err("Not enough memory for event selector list\n");
855 goto out_symbol_exit;
bbd36e5e 856 }
0e9b20b8 857
5c98d466 858 if (target_pid != -1)
d6d901c2 859 target_tid = target_pid;
d6d901c2 860
7e2ed097
ACM
861 if (perf_evlist__create_maps(evsel_list, target_pid,
862 target_tid, cpu_list) < 0)
dd7927f4 863 usage_with_options(record_usage, record_options);
69aad6f1 864
361c99a6 865 list_for_each_entry(pos, &evsel_list->entries, node) {
7e2ed097
ACM
866 if (perf_evsel__alloc_fd(pos, evsel_list->cpus->nr,
867 evsel_list->threads->nr) < 0)
69aad6f1 868 goto out_free_fd;
ad7f4e3f
ACM
869 if (perf_header__push_event(pos->attr.config, event_name(pos)))
870 goto out_free_fd;
d6d901c2 871 }
5c581041 872
7e2ed097 873 if (perf_evlist__alloc_pollfd(evsel_list) < 0)
39d17dac 874 goto out_free_fd;
d6d901c2 875
3de29cab 876 if (user_interval != ULLONG_MAX)
f9212819
FW
877 default_interval = user_interval;
878 if (user_freq != UINT_MAX)
879 freq = user_freq;
880
7e4ff9e3
MG
881 /*
882 * User specified count overrides default frequency.
883 */
884 if (default_interval)
885 freq = 0;
886 else if (freq) {
887 default_interval = freq;
888 } else {
889 fprintf(stderr, "frequency and count are zero, aborting\n");
39d17dac 890 err = -EINVAL;
5c581041 891 goto out_free_fd;
7e4ff9e3
MG
892 }
893
39d17dac 894 err = __cmd_record(argc, argv);
39d17dac 895out_free_fd:
7e2ed097 896 perf_evlist__delete_maps(evsel_list);
d65a458b
ACM
897out_symbol_exit:
898 symbol__exit();
39d17dac 899 return err;
0e9b20b8 900}