perf record: Stop intercepting events, use postprocessing to get build-ids
[linux-2.6.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 _LARGEFILE64_SOURCE
9 #define _FILE_OFFSET_BITS 64
10
11 #include "builtin.h"
12
13 #include "perf.h"
14
15 #include "util/build-id.h"
16 #include "util/util.h"
17 #include "util/parse-options.h"
18 #include "util/parse-events.h"
19 #include "util/string.h"
20
21 #include "util/header.h"
22 #include "util/event.h"
23 #include "util/debug.h"
24 #include "util/session.h"
25 #include "util/symbol.h"
26
27 #include <unistd.h>
28 #include <sched.h>
29
30 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
31
32 static long                     default_interval                =      0;
33
34 static int                      nr_cpus                         =      0;
35 static unsigned int             page_size;
36 static unsigned int             mmap_pages                      =    128;
37 static int                      freq                            =   1000;
38 static int                      output;
39 static const char               *output_name                    = "perf.data";
40 static int                      group                           =      0;
41 static unsigned int             realtime_prio                   =      0;
42 static int                      raw_samples                     =      0;
43 static int                      system_wide                     =      0;
44 static int                      profile_cpu                     =     -1;
45 static pid_t                    target_pid                      =     -1;
46 static pid_t                    child_pid                       =     -1;
47 static int                      inherit                         =      1;
48 static int                      force                           =      0;
49 static int                      append_file                     =      0;
50 static int                      call_graph                      =      0;
51 static int                      inherit_stat                    =      0;
52 static int                      no_samples                      =      0;
53 static int                      sample_address                  =      0;
54 static int                      multiplex                       =      0;
55 static int                      multiplex_fd                    =     -1;
56
57 static long                     samples                         =      0;
58 static struct timeval           last_read;
59 static struct timeval           this_read;
60
61 static u64                      bytes_written                   =      0;
62
63 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
64
65 static int                      nr_poll                         =      0;
66 static int                      nr_cpu                          =      0;
67
68 static int                      file_new                        =      1;
69 static off_t                    post_processing_offset;
70
71 static struct perf_session      *session;
72
73 struct mmap_data {
74         int                     counter;
75         void                    *base;
76         unsigned int            mask;
77         unsigned int            prev;
78 };
79
80 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
81
82 static unsigned long mmap_read_head(struct mmap_data *md)
83 {
84         struct perf_event_mmap_page *pc = md->base;
85         long head;
86
87         head = pc->data_head;
88         rmb();
89
90         return head;
91 }
92
93 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
94 {
95         struct perf_event_mmap_page *pc = md->base;
96
97         /*
98          * ensure all reads are done before we write the tail out.
99          */
100         /* mb(); */
101         pc->data_tail = tail;
102 }
103
104 static void write_output(void *buf, size_t size)
105 {
106         while (size) {
107                 int ret = write(output, buf, size);
108
109                 if (ret < 0)
110                         die("failed to write");
111
112                 size -= ret;
113                 buf += ret;
114
115                 bytes_written += ret;
116         }
117 }
118
119 static int process_synthesized_event(event_t *event,
120                                      struct perf_session *self __used)
121 {
122         write_output(event, event->header.size);
123         return 0;
124 }
125
126 static void mmap_read(struct mmap_data *md)
127 {
128         unsigned int head = mmap_read_head(md);
129         unsigned int old = md->prev;
130         unsigned char *data = md->base + page_size;
131         unsigned long size;
132         void *buf;
133         int diff;
134
135         gettimeofday(&this_read, NULL);
136
137         /*
138          * If we're further behind than half the buffer, there's a chance
139          * the writer will bite our tail and mess up the samples under us.
140          *
141          * If we somehow ended up ahead of the head, we got messed up.
142          *
143          * In either case, truncate and restart at head.
144          */
145         diff = head - old;
146         if (diff < 0) {
147                 struct timeval iv;
148                 unsigned long msecs;
149
150                 timersub(&this_read, &last_read, &iv);
151                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
152
153                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
154                                 "  Last read %lu msecs ago.\n", msecs);
155
156                 /*
157                  * head points to a known good entry, start there.
158                  */
159                 old = head;
160         }
161
162         last_read = this_read;
163
164         if (old != head)
165                 samples++;
166
167         size = head - old;
168
169         if ((old & md->mask) + size != (head & md->mask)) {
170                 buf = &data[old & md->mask];
171                 size = md->mask + 1 - (old & md->mask);
172                 old += size;
173
174                 write_output(buf, size);
175         }
176
177         buf = &data[old & md->mask];
178         size = head - old;
179         old += size;
180
181         write_output(buf, size);
182
183         md->prev = old;
184         mmap_write_tail(md, old);
185 }
186
187 static volatile int done = 0;
188 static volatile int signr = -1;
189
190 static void sig_handler(int sig)
191 {
192         done = 1;
193         signr = sig;
194 }
195
196 static void sig_atexit(void)
197 {
198         if (child_pid != -1)
199                 kill(child_pid, SIGTERM);
200
201         if (signr == -1)
202                 return;
203
204         signal(signr, SIG_DFL);
205         kill(getpid(), signr);
206 }
207
208 static int group_fd;
209
210 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
211 {
212         struct perf_header_attr *h_attr;
213
214         if (nr < session->header.attrs) {
215                 h_attr = session->header.attr[nr];
216         } else {
217                 h_attr = perf_header_attr__new(a);
218                 if (h_attr != NULL)
219                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
220                                 perf_header_attr__delete(h_attr);
221                                 h_attr = NULL;
222                         }
223         }
224
225         return h_attr;
226 }
227
228 static void create_counter(int counter, int cpu, pid_t pid)
229 {
230         char *filter = filters[counter];
231         struct perf_event_attr *attr = attrs + counter;
232         struct perf_header_attr *h_attr;
233         int track = !counter; /* only the first counter needs these */
234         int ret;
235         struct {
236                 u64 count;
237                 u64 time_enabled;
238                 u64 time_running;
239                 u64 id;
240         } read_data;
241
242         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
243                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
244                                   PERF_FORMAT_ID;
245
246         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
247
248         if (freq) {
249                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
250                 attr->freq              = 1;
251                 attr->sample_freq       = freq;
252         }
253
254         if (no_samples)
255                 attr->sample_freq = 0;
256
257         if (inherit_stat)
258                 attr->inherit_stat = 1;
259
260         if (sample_address)
261                 attr->sample_type       |= PERF_SAMPLE_ADDR;
262
263         if (call_graph)
264                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
265
266         if (raw_samples) {
267                 attr->sample_type       |= PERF_SAMPLE_TIME;
268                 attr->sample_type       |= PERF_SAMPLE_RAW;
269                 attr->sample_type       |= PERF_SAMPLE_CPU;
270         }
271
272         attr->mmap              = track;
273         attr->comm              = track;
274         attr->inherit           = inherit;
275         attr->disabled          = 1;
276
277 try_again:
278         fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
279
280         if (fd[nr_cpu][counter] < 0) {
281                 int err = errno;
282
283                 if (err == EPERM || err == EACCES)
284                         die("Permission error - are you root?\n");
285                 else if (err ==  ENODEV && profile_cpu != -1)
286                         die("No such device - did you specify an out-of-range profile CPU?\n");
287
288                 /*
289                  * If it's cycles then fall back to hrtimer
290                  * based cpu-clock-tick sw counter, which
291                  * is always available even if no PMU support:
292                  */
293                 if (attr->type == PERF_TYPE_HARDWARE
294                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
295
296                         if (verbose)
297                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
298                         attr->type = PERF_TYPE_SOFTWARE;
299                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
300                         goto try_again;
301                 }
302                 printf("\n");
303                 error("perfcounter syscall returned with %d (%s)\n",
304                         fd[nr_cpu][counter], strerror(err));
305
306 #if defined(__i386__) || defined(__x86_64__)
307                 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
308                         die("No hardware sampling interrupt available. No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.\n");
309 #endif
310
311                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
312                 exit(-1);
313         }
314
315         h_attr = get_header_attr(attr, counter);
316         if (h_attr == NULL)
317                 die("nomem\n");
318
319         if (!file_new) {
320                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
321                         fprintf(stderr, "incompatible append\n");
322                         exit(-1);
323                 }
324         }
325
326         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
327                 perror("Unable to read perf file descriptor\n");
328                 exit(-1);
329         }
330
331         if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
332                 pr_warning("Not enough memory to add id\n");
333                 exit(-1);
334         }
335
336         assert(fd[nr_cpu][counter] >= 0);
337         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
338
339         /*
340          * First counter acts as the group leader:
341          */
342         if (group && group_fd == -1)
343                 group_fd = fd[nr_cpu][counter];
344         if (multiplex && multiplex_fd == -1)
345                 multiplex_fd = fd[nr_cpu][counter];
346
347         if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
348
349                 ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
350                 assert(ret != -1);
351         } else {
352                 event_array[nr_poll].fd = fd[nr_cpu][counter];
353                 event_array[nr_poll].events = POLLIN;
354                 nr_poll++;
355
356                 mmap_array[nr_cpu][counter].counter = counter;
357                 mmap_array[nr_cpu][counter].prev = 0;
358                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
359                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
360                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
361                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
362                         error("failed to mmap with %d (%s)\n", errno, strerror(errno));
363                         exit(-1);
364                 }
365         }
366
367         if (filter != NULL) {
368                 ret = ioctl(fd[nr_cpu][counter],
369                             PERF_EVENT_IOC_SET_FILTER, filter);
370                 if (ret) {
371                         error("failed to set filter with %d (%s)\n", errno,
372                               strerror(errno));
373                         exit(-1);
374                 }
375         }
376
377         ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
378 }
379
380 static void open_counters(int cpu, pid_t pid)
381 {
382         int counter;
383
384         group_fd = -1;
385         for (counter = 0; counter < nr_counters; counter++)
386                 create_counter(counter, cpu, pid);
387
388         nr_cpu++;
389 }
390
391 static int process_buildids(void)
392 {
393         u64 size = lseek(output, 0, SEEK_CUR);
394
395         session->fd = output;
396         return __perf_session__process_events(session, post_processing_offset,
397                                               size - post_processing_offset,
398                                               size, &build_id__mark_dso_hit_ops);
399 }
400
401 static void atexit_header(void)
402 {
403         session->header.data_size += bytes_written;
404
405         process_buildids();
406         perf_header__write(&session->header, output, true);
407 }
408
409 static int __cmd_record(int argc, const char **argv)
410 {
411         int i, counter;
412         struct stat st;
413         pid_t pid = 0;
414         int flags;
415         int err;
416         unsigned long waking = 0;
417         int child_ready_pipe[2], go_pipe[2];
418         const bool forks = target_pid == -1 && argc > 0;
419         char buf;
420
421         page_size = sysconf(_SC_PAGE_SIZE);
422         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
423         assert(nr_cpus <= MAX_NR_CPUS);
424         assert(nr_cpus >= 0);
425
426         atexit(sig_atexit);
427         signal(SIGCHLD, sig_handler);
428         signal(SIGINT, sig_handler);
429
430         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
431                 perror("failed to create pipes");
432                 exit(-1);
433         }
434
435         if (!stat(output_name, &st) && st.st_size) {
436                 if (!force) {
437                         if (!append_file) {
438                                 pr_err("Error, output file %s exists, use -A "
439                                        "to append or -f to overwrite.\n",
440                                        output_name);
441                                 exit(-1);
442                         }
443                 } else {
444                         char oldname[PATH_MAX];
445                         snprintf(oldname, sizeof(oldname), "%s.old",
446                                  output_name);
447                         unlink(oldname);
448                         rename(output_name, oldname);
449                 }
450         } else {
451                 append_file = 0;
452         }
453
454         flags = O_CREAT|O_RDWR|O_LARGEFILE;
455         if (append_file)
456                 file_new = 0;
457         else
458                 flags |= O_TRUNC;
459
460         output = open(output_name, flags, S_IRUSR|S_IWUSR);
461         if (output < 0) {
462                 perror("failed to create output file");
463                 exit(-1);
464         }
465
466         session = perf_session__new(output_name, O_WRONLY, force);
467         if (session == NULL) {
468                 pr_err("Not enough memory for reading perf file header\n");
469                 return -1;
470         }
471
472         if (!file_new) {
473                 err = perf_header__read(&session->header, output);
474                 if (err < 0)
475                         return err;
476         }
477
478         if (raw_samples) {
479                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
480         } else {
481                 for (i = 0; i < nr_counters; i++) {
482                         if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
483                                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
484                                 break;
485                         }
486                 }
487         }
488
489         atexit(atexit_header);
490
491         if (forks) {
492                 pid = fork();
493                 if (pid < 0) {
494                         perror("failed to fork");
495                         exit(-1);
496                 }
497
498                 if (!pid) {
499                         close(child_ready_pipe[0]);
500                         close(go_pipe[1]);
501                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
502
503                         /*
504                          * Do a dummy execvp to get the PLT entry resolved,
505                          * so we avoid the resolver overhead on the real
506                          * execvp call.
507                          */
508                         execvp("", (char **)argv);
509
510                         /*
511                          * Tell the parent we're ready to go
512                          */
513                         close(child_ready_pipe[1]);
514
515                         /*
516                          * Wait until the parent tells us to go.
517                          */
518                         if (read(go_pipe[0], &buf, 1) == -1)
519                                 perror("unable to read pipe");
520
521                         execvp(argv[0], (char **)argv);
522
523                         perror(argv[0]);
524                         exit(-1);
525                 }
526
527                 child_pid = pid;
528
529                 if (!system_wide)
530                         target_pid = pid;
531
532                 close(child_ready_pipe[1]);
533                 close(go_pipe[0]);
534                 /*
535                  * wait for child to settle
536                  */
537                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
538                         perror("unable to read pipe");
539                         exit(-1);
540                 }
541                 close(child_ready_pipe[0]);
542         }
543
544
545         if ((!system_wide && !inherit) || profile_cpu != -1) {
546                 open_counters(profile_cpu, target_pid);
547         } else {
548                 for (i = 0; i < nr_cpus; i++)
549                         open_counters(i, target_pid);
550         }
551
552         if (file_new) {
553                 err = perf_header__write(&session->header, output, false);
554                 if (err < 0)
555                         return err;
556         }
557
558         post_processing_offset = lseek(output, 0, SEEK_CUR);
559
560         err = event__synthesize_kernel_mmap(process_synthesized_event,
561                                             session, "_text");
562         if (err < 0) {
563                 pr_err("Couldn't record kernel reference relocation symbol.\n");
564                 return err;
565         }
566
567         err = event__synthesize_modules(process_synthesized_event, session);
568         if (err < 0) {
569                 pr_err("Couldn't record kernel reference relocation symbol.\n");
570                 return err;
571         }
572
573         if (!system_wide && profile_cpu == -1)
574                 event__synthesize_thread(pid, process_synthesized_event,
575                                          session);
576         else
577                 event__synthesize_threads(process_synthesized_event, session);
578
579         if (realtime_prio) {
580                 struct sched_param param;
581
582                 param.sched_priority = realtime_prio;
583                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
584                         pr_err("Could not set realtime priority.\n");
585                         exit(-1);
586                 }
587         }
588
589         /*
590          * Let the child rip
591          */
592         if (forks)
593                 close(go_pipe[1]);
594
595         for (;;) {
596                 int hits = samples;
597
598                 for (i = 0; i < nr_cpu; i++) {
599                         for (counter = 0; counter < nr_counters; counter++) {
600                                 if (mmap_array[i][counter].base)
601                                         mmap_read(&mmap_array[i][counter]);
602                         }
603                 }
604
605                 if (hits == samples) {
606                         if (done)
607                                 break;
608                         err = poll(event_array, nr_poll, -1);
609                         waking++;
610                 }
611
612                 if (done) {
613                         for (i = 0; i < nr_cpu; i++) {
614                                 for (counter = 0; counter < nr_counters; counter++)
615                                         ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
616                         }
617                 }
618         }
619
620         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
621
622         /*
623          * Approximate RIP event size: 24 bytes.
624          */
625         fprintf(stderr,
626                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
627                 (double)bytes_written / 1024.0 / 1024.0,
628                 output_name,
629                 bytes_written / 24);
630
631         return 0;
632 }
633
634 static const char * const record_usage[] = {
635         "perf record [<options>] [<command>]",
636         "perf record [<options>] -- <command> [<options>]",
637         NULL
638 };
639
640 static const struct option options[] = {
641         OPT_CALLBACK('e', "event", NULL, "event",
642                      "event selector. use 'perf list' to list available events",
643                      parse_events),
644         OPT_CALLBACK(0, "filter", NULL, "filter",
645                      "event filter", parse_filter),
646         OPT_INTEGER('p', "pid", &target_pid,
647                     "record events on existing pid"),
648         OPT_INTEGER('r', "realtime", &realtime_prio,
649                     "collect data with this RT SCHED_FIFO priority"),
650         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
651                     "collect raw sample records from all opened counters"),
652         OPT_BOOLEAN('a', "all-cpus", &system_wide,
653                             "system-wide collection from all CPUs"),
654         OPT_BOOLEAN('A', "append", &append_file,
655                             "append to the output file to do incremental profiling"),
656         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
657                             "CPU to profile on"),
658         OPT_BOOLEAN('f', "force", &force,
659                         "overwrite existing data file"),
660         OPT_LONG('c', "count", &default_interval,
661                     "event period to sample"),
662         OPT_STRING('o', "output", &output_name, "file",
663                     "output file name"),
664         OPT_BOOLEAN('i', "inherit", &inherit,
665                     "child tasks inherit counters"),
666         OPT_INTEGER('F', "freq", &freq,
667                     "profile at this frequency"),
668         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
669                     "number of mmap data pages"),
670         OPT_BOOLEAN('g', "call-graph", &call_graph,
671                     "do call-graph (stack chain/backtrace) recording"),
672         OPT_BOOLEAN('v', "verbose", &verbose,
673                     "be more verbose (show counter open errors, etc)"),
674         OPT_BOOLEAN('s', "stat", &inherit_stat,
675                     "per thread counts"),
676         OPT_BOOLEAN('d', "data", &sample_address,
677                     "Sample addresses"),
678         OPT_BOOLEAN('n', "no-samples", &no_samples,
679                     "don't sample"),
680         OPT_BOOLEAN('M', "multiplex", &multiplex,
681                     "multiplex counter output in a single channel"),
682         OPT_END()
683 };
684
685 int cmd_record(int argc, const char **argv, const char *prefix __used)
686 {
687         int counter;
688
689         argc = parse_options(argc, argv, options, record_usage,
690                             PARSE_OPT_STOP_AT_NON_OPTION);
691         if (!argc && target_pid == -1 && !system_wide && profile_cpu == -1)
692                 usage_with_options(record_usage, options);
693
694         symbol__init();
695
696         if (!nr_counters) {
697                 nr_counters     = 1;
698                 attrs[0].type   = PERF_TYPE_HARDWARE;
699                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
700         }
701
702         /*
703          * User specified count overrides default frequency.
704          */
705         if (default_interval)
706                 freq = 0;
707         else if (freq) {
708                 default_interval = freq;
709         } else {
710                 fprintf(stderr, "frequency and count are zero, aborting\n");
711                 exit(EXIT_FAILURE);
712         }
713
714         for (counter = 0; counter < nr_counters; counter++) {
715                 if (attrs[counter].sample_period)
716                         continue;
717
718                 attrs[counter].sample_period = default_interval;
719         }
720
721         return __cmd_record(argc, argv);
722 }