perf record: Add ID and to recorded event data when recording multiple events
[safe/jmp/linux-2.6] / 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 #include "util/string.h"
19
20 #include "util/header.h"
21 #include "util/event.h"
22 #include "util/debug.h"
23 #include "util/session.h"
24 #include "util/symbol.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 const char               *output_name                    = "perf.data";
39 static int                      group                           =      0;
40 static unsigned int             realtime_prio                   =      0;
41 static int                      raw_samples                     =      0;
42 static int                      system_wide                     =      0;
43 static int                      profile_cpu                     =     -1;
44 static pid_t                    target_pid                      =     -1;
45 static pid_t                    child_pid                       =     -1;
46 static int                      inherit                         =      1;
47 static int                      force                           =      0;
48 static int                      append_file                     =      0;
49 static int                      call_graph                      =      0;
50 static int                      inherit_stat                    =      0;
51 static int                      no_samples                      =      0;
52 static int                      sample_address                  =      0;
53 static int                      multiplex                       =      0;
54 static int                      multiplex_fd                    =     -1;
55
56 static long                     samples                         =      0;
57 static struct timeval           last_read;
58 static struct timeval           this_read;
59
60 static u64                      bytes_written                   =      0;
61
62 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
63
64 static int                      nr_poll                         =      0;
65 static int                      nr_cpu                          =      0;
66
67 static int                      file_new                        =      1;
68 static off_t                    post_processing_offset;
69
70 static struct perf_session      *session;
71
72 struct mmap_data {
73         int                     counter;
74         void                    *base;
75         unsigned int            mask;
76         unsigned int            prev;
77 };
78
79 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
80
81 static unsigned long mmap_read_head(struct mmap_data *md)
82 {
83         struct perf_event_mmap_page *pc = md->base;
84         long head;
85
86         head = pc->data_head;
87         rmb();
88
89         return head;
90 }
91
92 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
93 {
94         struct perf_event_mmap_page *pc = md->base;
95
96         /*
97          * ensure all reads are done before we write the tail out.
98          */
99         /* mb(); */
100         pc->data_tail = tail;
101 }
102
103 static void write_output(void *buf, size_t size)
104 {
105         while (size) {
106                 int ret = write(output, buf, size);
107
108                 if (ret < 0)
109                         die("failed to write");
110
111                 size -= ret;
112                 buf += ret;
113
114                 bytes_written += ret;
115         }
116 }
117
118 static int process_synthesized_event(event_t *event,
119                                      struct perf_session *self __used)
120 {
121         write_output(event, event->header.size);
122         return 0;
123 }
124
125 static void mmap_read(struct mmap_data *md)
126 {
127         unsigned int head = mmap_read_head(md);
128         unsigned int old = md->prev;
129         unsigned char *data = md->base + page_size;
130         unsigned long size;
131         void *buf;
132         int diff;
133
134         gettimeofday(&this_read, NULL);
135
136         /*
137          * If we're further behind than half the buffer, there's a chance
138          * the writer will bite our tail and mess up the samples under us.
139          *
140          * If we somehow ended up ahead of the head, we got messed up.
141          *
142          * In either case, truncate and restart at head.
143          */
144         diff = head - old;
145         if (diff < 0) {
146                 struct timeval iv;
147                 unsigned long msecs;
148
149                 timersub(&this_read, &last_read, &iv);
150                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
151
152                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
153                                 "  Last read %lu msecs ago.\n", msecs);
154
155                 /*
156                  * head points to a known good entry, start there.
157                  */
158                 old = head;
159         }
160
161         last_read = this_read;
162
163         if (old != head)
164                 samples++;
165
166         size = head - old;
167
168         if ((old & md->mask) + size != (head & md->mask)) {
169                 buf = &data[old & md->mask];
170                 size = md->mask + 1 - (old & md->mask);
171                 old += size;
172
173                 write_output(buf, size);
174         }
175
176         buf = &data[old & md->mask];
177         size = head - old;
178         old += size;
179
180         write_output(buf, size);
181
182         md->prev = old;
183         mmap_write_tail(md, old);
184 }
185
186 static volatile int done = 0;
187 static volatile int signr = -1;
188
189 static void sig_handler(int sig)
190 {
191         done = 1;
192         signr = sig;
193 }
194
195 static void sig_atexit(void)
196 {
197         if (child_pid != -1)
198                 kill(child_pid, SIGTERM);
199
200         if (signr == -1)
201                 return;
202
203         signal(signr, SIG_DFL);
204         kill(getpid(), signr);
205 }
206
207 static int group_fd;
208
209 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
210 {
211         struct perf_header_attr *h_attr;
212
213         if (nr < session->header.attrs) {
214                 h_attr = session->header.attr[nr];
215         } else {
216                 h_attr = perf_header_attr__new(a);
217                 if (h_attr != NULL)
218                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
219                                 perf_header_attr__delete(h_attr);
220                                 h_attr = NULL;
221                         }
222         }
223
224         return h_attr;
225 }
226
227 static void create_counter(int counter, int cpu, pid_t pid)
228 {
229         char *filter = filters[counter];
230         struct perf_event_attr *attr = attrs + counter;
231         struct perf_header_attr *h_attr;
232         int track = !counter; /* only the first counter needs these */
233         int ret;
234         struct {
235                 u64 count;
236                 u64 time_enabled;
237                 u64 time_running;
238                 u64 id;
239         } read_data;
240
241         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
242                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
243                                   PERF_FORMAT_ID;
244
245         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
246
247         if (nr_counters > 1)
248                 attr->sample_type |= PERF_SAMPLE_ID;
249
250         if (freq) {
251                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
252                 attr->freq              = 1;
253                 attr->sample_freq       = freq;
254         }
255
256         if (no_samples)
257                 attr->sample_freq = 0;
258
259         if (inherit_stat)
260                 attr->inherit_stat = 1;
261
262         if (sample_address)
263                 attr->sample_type       |= PERF_SAMPLE_ADDR;
264
265         if (call_graph)
266                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
267
268         if (raw_samples) {
269                 attr->sample_type       |= PERF_SAMPLE_TIME;
270                 attr->sample_type       |= PERF_SAMPLE_RAW;
271                 attr->sample_type       |= PERF_SAMPLE_CPU;
272         }
273
274         attr->mmap              = track;
275         attr->comm              = track;
276         attr->inherit           = inherit;
277         attr->disabled          = 1;
278
279 try_again:
280         fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
281
282         if (fd[nr_cpu][counter] < 0) {
283                 int err = errno;
284
285                 if (err == EPERM || err == EACCES)
286                         die("Permission error - are you root?\n");
287                 else if (err ==  ENODEV && profile_cpu != -1)
288                         die("No such device - did you specify an out-of-range profile CPU?\n");
289
290                 /*
291                  * If it's cycles then fall back to hrtimer
292                  * based cpu-clock-tick sw counter, which
293                  * is always available even if no PMU support:
294                  */
295                 if (attr->type == PERF_TYPE_HARDWARE
296                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
297
298                         if (verbose)
299                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
300                         attr->type = PERF_TYPE_SOFTWARE;
301                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
302                         goto try_again;
303                 }
304                 printf("\n");
305                 error("perfcounter syscall returned with %d (%s)\n",
306                         fd[nr_cpu][counter], strerror(err));
307
308 #if defined(__i386__) || defined(__x86_64__)
309                 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
310                         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");
311 #endif
312
313                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
314                 exit(-1);
315         }
316
317         h_attr = get_header_attr(attr, counter);
318         if (h_attr == NULL)
319                 die("nomem\n");
320
321         if (!file_new) {
322                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
323                         fprintf(stderr, "incompatible append\n");
324                         exit(-1);
325                 }
326         }
327
328         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
329                 perror("Unable to read perf file descriptor\n");
330                 exit(-1);
331         }
332
333         if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
334                 pr_warning("Not enough memory to add id\n");
335                 exit(-1);
336         }
337
338         assert(fd[nr_cpu][counter] >= 0);
339         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
340
341         /*
342          * First counter acts as the group leader:
343          */
344         if (group && group_fd == -1)
345                 group_fd = fd[nr_cpu][counter];
346         if (multiplex && multiplex_fd == -1)
347                 multiplex_fd = fd[nr_cpu][counter];
348
349         if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
350
351                 ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
352                 assert(ret != -1);
353         } else {
354                 event_array[nr_poll].fd = fd[nr_cpu][counter];
355                 event_array[nr_poll].events = POLLIN;
356                 nr_poll++;
357
358                 mmap_array[nr_cpu][counter].counter = counter;
359                 mmap_array[nr_cpu][counter].prev = 0;
360                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
361                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
362                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
363                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
364                         error("failed to mmap with %d (%s)\n", errno, strerror(errno));
365                         exit(-1);
366                 }
367         }
368
369         if (filter != NULL) {
370                 ret = ioctl(fd[nr_cpu][counter],
371                             PERF_EVENT_IOC_SET_FILTER, filter);
372                 if (ret) {
373                         error("failed to set filter with %d (%s)\n", errno,
374                               strerror(errno));
375                         exit(-1);
376                 }
377         }
378
379         ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
380 }
381
382 static void open_counters(int cpu, pid_t pid)
383 {
384         int counter;
385
386         group_fd = -1;
387         for (counter = 0; counter < nr_counters; counter++)
388                 create_counter(counter, cpu, pid);
389
390         nr_cpu++;
391 }
392
393 static int process_buildids(void)
394 {
395         u64 size = lseek(output, 0, SEEK_CUR);
396
397         session->fd = output;
398         return __perf_session__process_events(session, post_processing_offset,
399                                               size - post_processing_offset,
400                                               size, &build_id__mark_dso_hit_ops);
401 }
402
403 static void atexit_header(void)
404 {
405         session->header.data_size += bytes_written;
406
407         process_buildids();
408         perf_header__write(&session->header, output, true);
409 }
410
411 static int __cmd_record(int argc, const char **argv)
412 {
413         int i, counter;
414         struct stat st;
415         pid_t pid = 0;
416         int flags;
417         int err;
418         unsigned long waking = 0;
419         int child_ready_pipe[2], go_pipe[2];
420         const bool forks = target_pid == -1 && argc > 0;
421         char buf;
422
423         page_size = sysconf(_SC_PAGE_SIZE);
424         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
425         assert(nr_cpus <= MAX_NR_CPUS);
426         assert(nr_cpus >= 0);
427
428         atexit(sig_atexit);
429         signal(SIGCHLD, sig_handler);
430         signal(SIGINT, sig_handler);
431
432         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
433                 perror("failed to create pipes");
434                 exit(-1);
435         }
436
437         if (!stat(output_name, &st) && st.st_size) {
438                 if (!force) {
439                         if (!append_file) {
440                                 pr_err("Error, output file %s exists, use -A "
441                                        "to append or -f to overwrite.\n",
442                                        output_name);
443                                 exit(-1);
444                         }
445                 } else {
446                         char oldname[PATH_MAX];
447                         snprintf(oldname, sizeof(oldname), "%s.old",
448                                  output_name);
449                         unlink(oldname);
450                         rename(output_name, oldname);
451                 }
452         } else {
453                 append_file = 0;
454         }
455
456         flags = O_CREAT|O_RDWR;
457         if (append_file)
458                 file_new = 0;
459         else
460                 flags |= O_TRUNC;
461
462         output = open(output_name, flags, S_IRUSR|S_IWUSR);
463         if (output < 0) {
464                 perror("failed to create output file");
465                 exit(-1);
466         }
467
468         session = perf_session__new(output_name, O_WRONLY, force);
469         if (session == NULL) {
470                 pr_err("Not enough memory for reading perf file header\n");
471                 return -1;
472         }
473
474         if (!file_new) {
475                 err = perf_header__read(&session->header, output);
476                 if (err < 0)
477                         return err;
478         }
479
480         if (raw_samples) {
481                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
482         } else {
483                 for (i = 0; i < nr_counters; i++) {
484                         if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
485                                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
486                                 break;
487                         }
488                 }
489         }
490
491         atexit(atexit_header);
492
493         if (forks) {
494                 pid = fork();
495                 if (pid < 0) {
496                         perror("failed to fork");
497                         exit(-1);
498                 }
499
500                 if (!pid) {
501                         close(child_ready_pipe[0]);
502                         close(go_pipe[1]);
503                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
504
505                         /*
506                          * Do a dummy execvp to get the PLT entry resolved,
507                          * so we avoid the resolver overhead on the real
508                          * execvp call.
509                          */
510                         execvp("", (char **)argv);
511
512                         /*
513                          * Tell the parent we're ready to go
514                          */
515                         close(child_ready_pipe[1]);
516
517                         /*
518                          * Wait until the parent tells us to go.
519                          */
520                         if (read(go_pipe[0], &buf, 1) == -1)
521                                 perror("unable to read pipe");
522
523                         execvp(argv[0], (char **)argv);
524
525                         perror(argv[0]);
526                         exit(-1);
527                 }
528
529                 child_pid = pid;
530
531                 if (!system_wide)
532                         target_pid = pid;
533
534                 close(child_ready_pipe[1]);
535                 close(go_pipe[0]);
536                 /*
537                  * wait for child to settle
538                  */
539                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
540                         perror("unable to read pipe");
541                         exit(-1);
542                 }
543                 close(child_ready_pipe[0]);
544         }
545
546
547         if ((!system_wide && !inherit) || profile_cpu != -1) {
548                 open_counters(profile_cpu, target_pid);
549         } else {
550                 for (i = 0; i < nr_cpus; i++)
551                         open_counters(i, target_pid);
552         }
553
554         if (file_new) {
555                 err = perf_header__write(&session->header, output, false);
556                 if (err < 0)
557                         return err;
558         }
559
560         post_processing_offset = lseek(output, 0, SEEK_CUR);
561
562         err = event__synthesize_kernel_mmap(process_synthesized_event,
563                                             session, "_text");
564         if (err < 0) {
565                 pr_err("Couldn't record kernel reference relocation symbol.\n");
566                 return err;
567         }
568
569         err = event__synthesize_modules(process_synthesized_event, session);
570         if (err < 0) {
571                 pr_err("Couldn't record kernel reference relocation symbol.\n");
572                 return err;
573         }
574
575         if (!system_wide && profile_cpu == -1)
576                 event__synthesize_thread(target_pid, process_synthesized_event,
577                                          session);
578         else
579                 event__synthesize_threads(process_synthesized_event, session);
580
581         if (realtime_prio) {
582                 struct sched_param param;
583
584                 param.sched_priority = realtime_prio;
585                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
586                         pr_err("Could not set realtime priority.\n");
587                         exit(-1);
588                 }
589         }
590
591         /*
592          * Let the child rip
593          */
594         if (forks)
595                 close(go_pipe[1]);
596
597         for (;;) {
598                 int hits = samples;
599
600                 for (i = 0; i < nr_cpu; i++) {
601                         for (counter = 0; counter < nr_counters; counter++) {
602                                 if (mmap_array[i][counter].base)
603                                         mmap_read(&mmap_array[i][counter]);
604                         }
605                 }
606
607                 if (hits == samples) {
608                         if (done)
609                                 break;
610                         err = poll(event_array, nr_poll, -1);
611                         waking++;
612                 }
613
614                 if (done) {
615                         for (i = 0; i < nr_cpu; i++) {
616                                 for (counter = 0; counter < nr_counters; counter++)
617                                         ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
618                         }
619                 }
620         }
621
622         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
623
624         /*
625          * Approximate RIP event size: 24 bytes.
626          */
627         fprintf(stderr,
628                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
629                 (double)bytes_written / 1024.0 / 1024.0,
630                 output_name,
631                 bytes_written / 24);
632
633         return 0;
634 }
635
636 static const char * const record_usage[] = {
637         "perf record [<options>] [<command>]",
638         "perf record [<options>] -- <command> [<options>]",
639         NULL
640 };
641
642 static const struct option options[] = {
643         OPT_CALLBACK('e', "event", NULL, "event",
644                      "event selector. use 'perf list' to list available events",
645                      parse_events),
646         OPT_CALLBACK(0, "filter", NULL, "filter",
647                      "event filter", parse_filter),
648         OPT_INTEGER('p', "pid", &target_pid,
649                     "record events on existing pid"),
650         OPT_INTEGER('r', "realtime", &realtime_prio,
651                     "collect data with this RT SCHED_FIFO priority"),
652         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
653                     "collect raw sample records from all opened counters"),
654         OPT_BOOLEAN('a', "all-cpus", &system_wide,
655                             "system-wide collection from all CPUs"),
656         OPT_BOOLEAN('A', "append", &append_file,
657                             "append to the output file to do incremental profiling"),
658         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
659                             "CPU to profile on"),
660         OPT_BOOLEAN('f', "force", &force,
661                         "overwrite existing data file"),
662         OPT_LONG('c', "count", &default_interval,
663                     "event period to sample"),
664         OPT_STRING('o', "output", &output_name, "file",
665                     "output file name"),
666         OPT_BOOLEAN('i', "inherit", &inherit,
667                     "child tasks inherit counters"),
668         OPT_INTEGER('F', "freq", &freq,
669                     "profile at this frequency"),
670         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
671                     "number of mmap data pages"),
672         OPT_BOOLEAN('g', "call-graph", &call_graph,
673                     "do call-graph (stack chain/backtrace) recording"),
674         OPT_BOOLEAN('v', "verbose", &verbose,
675                     "be more verbose (show counter open errors, etc)"),
676         OPT_BOOLEAN('s', "stat", &inherit_stat,
677                     "per thread counts"),
678         OPT_BOOLEAN('d', "data", &sample_address,
679                     "Sample addresses"),
680         OPT_BOOLEAN('n', "no-samples", &no_samples,
681                     "don't sample"),
682         OPT_BOOLEAN('M', "multiplex", &multiplex,
683                     "multiplex counter output in a single channel"),
684         OPT_END()
685 };
686
687 int cmd_record(int argc, const char **argv, const char *prefix __used)
688 {
689         int counter;
690
691         argc = parse_options(argc, argv, options, record_usage,
692                             PARSE_OPT_STOP_AT_NON_OPTION);
693         if (!argc && target_pid == -1 && !system_wide && profile_cpu == -1)
694                 usage_with_options(record_usage, options);
695
696         symbol__init();
697
698         if (!nr_counters) {
699                 nr_counters     = 1;
700                 attrs[0].type   = PERF_TYPE_HARDWARE;
701                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
702         }
703
704         /*
705          * User specified count overrides default frequency.
706          */
707         if (default_interval)
708                 freq = 0;
709         else if (freq) {
710                 default_interval = freq;
711         } else {
712                 fprintf(stderr, "frequency and count are zero, aborting\n");
713                 exit(EXIT_FAILURE);
714         }
715
716         for (counter = 0; counter < nr_counters; counter++) {
717                 if (attrs[counter].sample_period)
718                         continue;
719
720                 attrs[counter].sample_period = default_interval;
721         }
722
723         return __cmd_record(argc, argv);
724 }