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