perf: Convert perf header attrs into attr events
[deliverable/linux.git] / tools / perf / builtin-record.c
1 /*
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.
7 */
8 #define _FILE_OFFSET_BITS 64
9
10 #include "builtin.h"
11
12 #include "perf.h"
13
14 #include "util/build-id.h"
15 #include "util/util.h"
16 #include "util/parse-options.h"
17 #include "util/parse-events.h"
18
19 #include "util/header.h"
20 #include "util/event.h"
21 #include "util/debug.h"
22 #include "util/session.h"
23 #include "util/symbol.h"
24 #include "util/cpumap.h"
25
26 #include <unistd.h>
27 #include <sched.h>
28
29 static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
30
31 static long default_interval = 0;
32
33 static int nr_cpus = 0;
34 static unsigned int page_size;
35 static unsigned int mmap_pages = 128;
36 static int freq = 1000;
37 static int output;
38 static int pipe_output = 0;
39 static const char *output_name = "perf.data";
40 static int group = 0;
41 static unsigned int realtime_prio = 0;
42 static bool raw_samples = false;
43 static bool system_wide = false;
44 static int profile_cpu = -1;
45 static pid_t target_pid = -1;
46 static pid_t target_tid = -1;
47 static pid_t *all_tids = NULL;
48 static int thread_num = 0;
49 static pid_t child_pid = -1;
50 static bool inherit = true;
51 static bool force = false;
52 static bool append_file = false;
53 static bool call_graph = false;
54 static bool inherit_stat = false;
55 static bool no_samples = false;
56 static bool sample_address = false;
57 static bool multiplex = false;
58 static int multiplex_fd = -1;
59
60 static long samples = 0;
61 static struct timeval last_read;
62 static struct timeval this_read;
63
64 static u64 bytes_written = 0;
65
66 static struct pollfd *event_array;
67
68 static int nr_poll = 0;
69 static int nr_cpu = 0;
70
71 static int file_new = 1;
72 static off_t post_processing_offset;
73
74 static struct perf_session *session;
75
76 struct mmap_data {
77 int counter;
78 void *base;
79 unsigned int mask;
80 unsigned int prev;
81 };
82
83 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
84
85 static unsigned long mmap_read_head(struct mmap_data *md)
86 {
87 struct perf_event_mmap_page *pc = md->base;
88 long head;
89
90 head = pc->data_head;
91 rmb();
92
93 return head;
94 }
95
96 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
97 {
98 struct perf_event_mmap_page *pc = md->base;
99
100 /*
101 * ensure all reads are done before we write the tail out.
102 */
103 /* mb(); */
104 pc->data_tail = tail;
105 }
106
107 static void write_output(void *buf, size_t size)
108 {
109 while (size) {
110 int ret = write(output, buf, size);
111
112 if (ret < 0)
113 die("failed to write");
114
115 size -= ret;
116 buf += ret;
117
118 bytes_written += ret;
119 }
120 }
121
122 static int process_synthesized_event(event_t *event,
123 struct perf_session *self __used)
124 {
125 write_output(event, event->header.size);
126 return 0;
127 }
128
129 static void mmap_read(struct mmap_data *md)
130 {
131 unsigned int head = mmap_read_head(md);
132 unsigned int old = md->prev;
133 unsigned char *data = md->base + page_size;
134 unsigned long size;
135 void *buf;
136 int diff;
137
138 gettimeofday(&this_read, NULL);
139
140 /*
141 * If we're further behind than half the buffer, there's a chance
142 * the writer will bite our tail and mess up the samples under us.
143 *
144 * If we somehow ended up ahead of the head, we got messed up.
145 *
146 * In either case, truncate and restart at head.
147 */
148 diff = head - old;
149 if (diff < 0) {
150 struct timeval iv;
151 unsigned long msecs;
152
153 timersub(&this_read, &last_read, &iv);
154 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
155
156 fprintf(stderr, "WARNING: failed to keep up with mmap data."
157 " Last read %lu msecs ago.\n", msecs);
158
159 /*
160 * head points to a known good entry, start there.
161 */
162 old = head;
163 }
164
165 last_read = this_read;
166
167 if (old != head)
168 samples++;
169
170 size = head - old;
171
172 if ((old & md->mask) + size != (head & md->mask)) {
173 buf = &data[old & md->mask];
174 size = md->mask + 1 - (old & md->mask);
175 old += size;
176
177 write_output(buf, size);
178 }
179
180 buf = &data[old & md->mask];
181 size = head - old;
182 old += size;
183
184 write_output(buf, size);
185
186 md->prev = old;
187 mmap_write_tail(md, old);
188 }
189
190 static volatile int done = 0;
191 static volatile int signr = -1;
192
193 static void sig_handler(int sig)
194 {
195 done = 1;
196 signr = sig;
197 }
198
199 static void sig_atexit(void)
200 {
201 if (child_pid != -1)
202 kill(child_pid, SIGTERM);
203
204 if (signr == -1)
205 return;
206
207 signal(signr, SIG_DFL);
208 kill(getpid(), signr);
209 }
210
211 static int group_fd;
212
213 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
214 {
215 struct perf_header_attr *h_attr;
216
217 if (nr < session->header.attrs) {
218 h_attr = session->header.attr[nr];
219 } else {
220 h_attr = perf_header_attr__new(a);
221 if (h_attr != NULL)
222 if (perf_header__add_attr(&session->header, h_attr) < 0) {
223 perf_header_attr__delete(h_attr);
224 h_attr = NULL;
225 }
226 }
227
228 return h_attr;
229 }
230
231 static void create_counter(int counter, int cpu)
232 {
233 char *filter = filters[counter];
234 struct perf_event_attr *attr = attrs + counter;
235 struct perf_header_attr *h_attr;
236 int track = !counter; /* only the first counter needs these */
237 int thread_index;
238 int ret;
239 struct {
240 u64 count;
241 u64 time_enabled;
242 u64 time_running;
243 u64 id;
244 } read_data;
245
246 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
247 PERF_FORMAT_TOTAL_TIME_RUNNING |
248 PERF_FORMAT_ID;
249
250 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
251
252 if (nr_counters > 1)
253 attr->sample_type |= PERF_SAMPLE_ID;
254
255 if (freq) {
256 attr->sample_type |= PERF_SAMPLE_PERIOD;
257 attr->freq = 1;
258 attr->sample_freq = freq;
259 }
260
261 if (no_samples)
262 attr->sample_freq = 0;
263
264 if (inherit_stat)
265 attr->inherit_stat = 1;
266
267 if (sample_address)
268 attr->sample_type |= PERF_SAMPLE_ADDR;
269
270 if (call_graph)
271 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
272
273 if (raw_samples) {
274 attr->sample_type |= PERF_SAMPLE_TIME;
275 attr->sample_type |= PERF_SAMPLE_RAW;
276 attr->sample_type |= PERF_SAMPLE_CPU;
277 }
278
279 attr->mmap = track;
280 attr->comm = track;
281 attr->inherit = inherit;
282 if (target_pid == -1 && !system_wide) {
283 attr->disabled = 1;
284 attr->enable_on_exec = 1;
285 }
286
287 for (thread_index = 0; thread_index < thread_num; thread_index++) {
288 try_again:
289 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
290 all_tids[thread_index], cpu, group_fd, 0);
291
292 if (fd[nr_cpu][counter][thread_index] < 0) {
293 int err = errno;
294
295 if (err == EPERM || err == EACCES)
296 die("Permission error - are you root?\n"
297 "\t Consider tweaking"
298 " /proc/sys/kernel/perf_event_paranoid.\n");
299 else if (err == ENODEV && profile_cpu != -1) {
300 die("No such device - did you specify"
301 " an out-of-range profile CPU?\n");
302 }
303
304 /*
305 * If it's cycles then fall back to hrtimer
306 * based cpu-clock-tick sw counter, which
307 * is always available even if no PMU support:
308 */
309 if (attr->type == PERF_TYPE_HARDWARE
310 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
311
312 if (verbose)
313 warning(" ... trying to fall back to cpu-clock-ticks\n");
314 attr->type = PERF_TYPE_SOFTWARE;
315 attr->config = PERF_COUNT_SW_CPU_CLOCK;
316 goto try_again;
317 }
318 printf("\n");
319 error("perfcounter syscall returned with %d (%s)\n",
320 fd[nr_cpu][counter][thread_index], strerror(err));
321
322 #if defined(__i386__) || defined(__x86_64__)
323 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
324 die("No hardware sampling interrupt available."
325 " No APIC? If so then you can boot the kernel"
326 " with the \"lapic\" boot parameter to"
327 " force-enable it.\n");
328 #endif
329
330 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
331 exit(-1);
332 }
333
334 h_attr = get_header_attr(attr, counter);
335 if (h_attr == NULL)
336 die("nomem\n");
337
338 if (!file_new) {
339 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
340 fprintf(stderr, "incompatible append\n");
341 exit(-1);
342 }
343 }
344
345 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
346 perror("Unable to read perf file descriptor\n");
347 exit(-1);
348 }
349
350 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
351 pr_warning("Not enough memory to add id\n");
352 exit(-1);
353 }
354
355 assert(fd[nr_cpu][counter][thread_index] >= 0);
356 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
357
358 /*
359 * First counter acts as the group leader:
360 */
361 if (group && group_fd == -1)
362 group_fd = fd[nr_cpu][counter][thread_index];
363 if (multiplex && multiplex_fd == -1)
364 multiplex_fd = fd[nr_cpu][counter][thread_index];
365
366 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
367
368 ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
369 assert(ret != -1);
370 } else {
371 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
372 event_array[nr_poll].events = POLLIN;
373 nr_poll++;
374
375 mmap_array[nr_cpu][counter][thread_index].counter = counter;
376 mmap_array[nr_cpu][counter][thread_index].prev = 0;
377 mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
378 mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
379 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
380 if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
381 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
382 exit(-1);
383 }
384 }
385
386 if (filter != NULL) {
387 ret = ioctl(fd[nr_cpu][counter][thread_index],
388 PERF_EVENT_IOC_SET_FILTER, filter);
389 if (ret) {
390 error("failed to set filter with %d (%s)\n", errno,
391 strerror(errno));
392 exit(-1);
393 }
394 }
395 }
396 }
397
398 static void open_counters(int cpu)
399 {
400 int counter;
401
402 group_fd = -1;
403 for (counter = 0; counter < nr_counters; counter++)
404 create_counter(counter, cpu);
405
406 nr_cpu++;
407 }
408
409 static int process_buildids(void)
410 {
411 u64 size = lseek(output, 0, SEEK_CUR);
412
413 if (size == 0)
414 return 0;
415
416 session->fd = output;
417 return __perf_session__process_events(session, post_processing_offset,
418 size - post_processing_offset,
419 size, &build_id__mark_dso_hit_ops);
420 }
421
422 static void atexit_header(void)
423 {
424 session->header.data_size += bytes_written;
425
426 process_buildids();
427 perf_header__write(&session->header, output, true);
428 }
429
430 static int __cmd_record(int argc, const char **argv)
431 {
432 int i, counter;
433 struct stat st;
434 pid_t pid = 0;
435 int flags;
436 int err;
437 unsigned long waking = 0;
438 int child_ready_pipe[2], go_pipe[2];
439 const bool forks = argc > 0;
440 char buf;
441
442 page_size = sysconf(_SC_PAGE_SIZE);
443
444 atexit(sig_atexit);
445 signal(SIGCHLD, sig_handler);
446 signal(SIGINT, sig_handler);
447
448 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
449 perror("failed to create pipes");
450 exit(-1);
451 }
452
453 if (!strcmp(output_name, "-"))
454 pipe_output = 1;
455 else if (!stat(output_name, &st) && st.st_size) {
456 if (!force) {
457 if (!append_file) {
458 pr_err("Error, output file %s exists, use -A "
459 "to append or -f to overwrite.\n",
460 output_name);
461 exit(-1);
462 }
463 } else {
464 char oldname[PATH_MAX];
465 snprintf(oldname, sizeof(oldname), "%s.old",
466 output_name);
467 unlink(oldname);
468 rename(output_name, oldname);
469 }
470 } else {
471 append_file = false;
472 }
473
474 flags = O_CREAT|O_RDWR;
475 if (append_file)
476 file_new = 0;
477 else
478 flags |= O_TRUNC;
479
480 if (pipe_output)
481 output = STDOUT_FILENO;
482 else
483 output = open(output_name, flags, S_IRUSR | S_IWUSR);
484 if (output < 0) {
485 perror("failed to create output file");
486 exit(-1);
487 }
488
489 session = perf_session__new(output_name, O_WRONLY, force);
490 if (session == NULL) {
491 pr_err("Not enough memory for reading perf file header\n");
492 return -1;
493 }
494
495 if (!file_new) {
496 err = perf_header__read(session, output);
497 if (err < 0)
498 return err;
499 }
500
501 if (raw_samples) {
502 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
503 } else {
504 for (i = 0; i < nr_counters; i++) {
505 if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
506 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
507 break;
508 }
509 }
510 }
511
512 atexit(atexit_header);
513
514 if (forks) {
515 child_pid = fork();
516 if (pid < 0) {
517 perror("failed to fork");
518 exit(-1);
519 }
520
521 if (!child_pid) {
522 if (pipe_output)
523 dup2(2, 1);
524 close(child_ready_pipe[0]);
525 close(go_pipe[1]);
526 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
527
528 /*
529 * Do a dummy execvp to get the PLT entry resolved,
530 * so we avoid the resolver overhead on the real
531 * execvp call.
532 */
533 execvp("", (char **)argv);
534
535 /*
536 * Tell the parent we're ready to go
537 */
538 close(child_ready_pipe[1]);
539
540 /*
541 * Wait until the parent tells us to go.
542 */
543 if (read(go_pipe[0], &buf, 1) == -1)
544 perror("unable to read pipe");
545
546 execvp(argv[0], (char **)argv);
547
548 perror(argv[0]);
549 exit(-1);
550 }
551
552 if (!system_wide && target_tid == -1 && target_pid == -1)
553 all_tids[0] = child_pid;
554
555 close(child_ready_pipe[1]);
556 close(go_pipe[0]);
557 /*
558 * wait for child to settle
559 */
560 if (read(child_ready_pipe[0], &buf, 1) == -1) {
561 perror("unable to read pipe");
562 exit(-1);
563 }
564 close(child_ready_pipe[0]);
565 }
566
567 if ((!system_wide && !inherit) || profile_cpu != -1) {
568 open_counters(profile_cpu);
569 } else {
570 nr_cpus = read_cpu_map();
571 for (i = 0; i < nr_cpus; i++)
572 open_counters(cpumap[i]);
573 }
574
575 if (pipe_output) {
576 err = perf_header__write_pipe(output);
577 if (err < 0)
578 return err;
579 } else if (file_new) {
580 err = perf_header__write(&session->header, output, false);
581 if (err < 0)
582 return err;
583 }
584
585 post_processing_offset = lseek(output, 0, SEEK_CUR);
586
587 if (pipe_output) {
588 err = event__synthesize_attrs(&session->header,
589 process_synthesized_event,
590 session);
591 if (err < 0) {
592 pr_err("Couldn't synthesize attrs.\n");
593 return err;
594 }
595 }
596
597 err = event__synthesize_kernel_mmap(process_synthesized_event,
598 session, "_text");
599 if (err < 0)
600 err = event__synthesize_kernel_mmap(process_synthesized_event,
601 session, "_stext");
602 if (err < 0) {
603 pr_err("Couldn't record kernel reference relocation symbol.\n");
604 return err;
605 }
606
607 err = event__synthesize_modules(process_synthesized_event, session);
608 if (err < 0) {
609 pr_err("Couldn't record kernel reference relocation symbol.\n");
610 return err;
611 }
612
613 if (!system_wide && profile_cpu == -1)
614 event__synthesize_thread(target_tid, process_synthesized_event,
615 session);
616 else
617 event__synthesize_threads(process_synthesized_event, session);
618
619 if (realtime_prio) {
620 struct sched_param param;
621
622 param.sched_priority = realtime_prio;
623 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
624 pr_err("Could not set realtime priority.\n");
625 exit(-1);
626 }
627 }
628
629 /*
630 * Let the child rip
631 */
632 if (forks)
633 close(go_pipe[1]);
634
635 for (;;) {
636 int hits = samples;
637 int thread;
638
639 for (i = 0; i < nr_cpu; i++) {
640 for (counter = 0; counter < nr_counters; counter++) {
641 for (thread = 0;
642 thread < thread_num; thread++) {
643 if (mmap_array[i][counter][thread].base)
644 mmap_read(&mmap_array[i][counter][thread]);
645 }
646
647 }
648 }
649
650 if (hits == samples) {
651 if (done)
652 break;
653 err = poll(event_array, nr_poll, -1);
654 waking++;
655 }
656
657 if (done) {
658 for (i = 0; i < nr_cpu; i++) {
659 for (counter = 0;
660 counter < nr_counters;
661 counter++) {
662 for (thread = 0;
663 thread < thread_num;
664 thread++)
665 ioctl(fd[i][counter][thread],
666 PERF_EVENT_IOC_DISABLE);
667 }
668 }
669 }
670 }
671
672 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
673
674 /*
675 * Approximate RIP event size: 24 bytes.
676 */
677 fprintf(stderr,
678 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
679 (double)bytes_written / 1024.0 / 1024.0,
680 output_name,
681 bytes_written / 24);
682
683 return 0;
684 }
685
686 static const char * const record_usage[] = {
687 "perf record [<options>] [<command>]",
688 "perf record [<options>] -- <command> [<options>]",
689 NULL
690 };
691
692 static const struct option options[] = {
693 OPT_CALLBACK('e', "event", NULL, "event",
694 "event selector. use 'perf list' to list available events",
695 parse_events),
696 OPT_CALLBACK(0, "filter", NULL, "filter",
697 "event filter", parse_filter),
698 OPT_INTEGER('p', "pid", &target_pid,
699 "record events on existing process id"),
700 OPT_INTEGER('t', "tid", &target_tid,
701 "record events on existing thread id"),
702 OPT_INTEGER('r', "realtime", &realtime_prio,
703 "collect data with this RT SCHED_FIFO priority"),
704 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
705 "collect raw sample records from all opened counters"),
706 OPT_BOOLEAN('a', "all-cpus", &system_wide,
707 "system-wide collection from all CPUs"),
708 OPT_BOOLEAN('A', "append", &append_file,
709 "append to the output file to do incremental profiling"),
710 OPT_INTEGER('C', "profile_cpu", &profile_cpu,
711 "CPU to profile on"),
712 OPT_BOOLEAN('f', "force", &force,
713 "overwrite existing data file"),
714 OPT_LONG('c', "count", &default_interval,
715 "event period to sample"),
716 OPT_STRING('o', "output", &output_name, "file",
717 "output file name"),
718 OPT_BOOLEAN('i', "inherit", &inherit,
719 "child tasks inherit counters"),
720 OPT_INTEGER('F', "freq", &freq,
721 "profile at this frequency"),
722 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
723 "number of mmap data pages"),
724 OPT_BOOLEAN('g', "call-graph", &call_graph,
725 "do call-graph (stack chain/backtrace) recording"),
726 OPT_INCR('v', "verbose", &verbose,
727 "be more verbose (show counter open errors, etc)"),
728 OPT_BOOLEAN('s', "stat", &inherit_stat,
729 "per thread counts"),
730 OPT_BOOLEAN('d', "data", &sample_address,
731 "Sample addresses"),
732 OPT_BOOLEAN('n', "no-samples", &no_samples,
733 "don't sample"),
734 OPT_BOOLEAN('M', "multiplex", &multiplex,
735 "multiplex counter output in a single channel"),
736 OPT_END()
737 };
738
739 int cmd_record(int argc, const char **argv, const char *prefix __used)
740 {
741 int counter;
742 int i,j;
743
744 argc = parse_options(argc, argv, options, record_usage,
745 PARSE_OPT_STOP_AT_NON_OPTION);
746 if (!argc && target_pid == -1 && target_tid == -1 &&
747 !system_wide && profile_cpu == -1)
748 usage_with_options(record_usage, options);
749
750 symbol__init();
751
752 if (!nr_counters) {
753 nr_counters = 1;
754 attrs[0].type = PERF_TYPE_HARDWARE;
755 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
756 }
757
758 if (target_pid != -1) {
759 target_tid = target_pid;
760 thread_num = find_all_tid(target_pid, &all_tids);
761 if (thread_num <= 0) {
762 fprintf(stderr, "Can't find all threads of pid %d\n",
763 target_pid);
764 usage_with_options(record_usage, options);
765 }
766 } else {
767 all_tids=malloc(sizeof(pid_t));
768 if (!all_tids)
769 return -ENOMEM;
770
771 all_tids[0] = target_tid;
772 thread_num = 1;
773 }
774
775 for (i = 0; i < MAX_NR_CPUS; i++) {
776 for (j = 0; j < MAX_COUNTERS; j++) {
777 fd[i][j] = malloc(sizeof(int)*thread_num);
778 mmap_array[i][j] = zalloc(
779 sizeof(struct mmap_data)*thread_num);
780 if (!fd[i][j] || !mmap_array[i][j])
781 return -ENOMEM;
782 }
783 }
784 event_array = malloc(
785 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
786 if (!event_array)
787 return -ENOMEM;
788
789 /*
790 * User specified count overrides default frequency.
791 */
792 if (default_interval)
793 freq = 0;
794 else if (freq) {
795 default_interval = freq;
796 } else {
797 fprintf(stderr, "frequency and count are zero, aborting\n");
798 exit(EXIT_FAILURE);
799 }
800
801 for (counter = 0; counter < nr_counters; counter++) {
802 if (attrs[counter].sample_period)
803 continue;
804
805 attrs[counter].sample_period = default_interval;
806 }
807
808 return __cmd_record(argc, argv);
809 }
This page took 0.064599 seconds and 5 git commands to generate.