perf tools: Unify debug messages mechanisms
[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 #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
21 #include <unistd.h>
22 #include <sched.h>
23
24 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
25
26 static long                     default_interval                =      0;
27
28 static int                      nr_cpus                         =      0;
29 static unsigned int             page_size;
30 static unsigned int             mmap_pages                      =    128;
31 static int                      freq                            =   1000;
32 static int                      output;
33 static const char               *output_name                    = "perf.data";
34 static int                      group                           =      0;
35 static unsigned int             realtime_prio                   =      0;
36 static int                      raw_samples                     =      0;
37 static int                      system_wide                     =      0;
38 static int                      profile_cpu                     =     -1;
39 static pid_t                    target_pid                      =     -1;
40 static pid_t                    child_pid                       =     -1;
41 static int                      inherit                         =      1;
42 static int                      force                           =      0;
43 static int                      append_file                     =      0;
44 static int                      call_graph                      =      0;
45 static int                      inherit_stat                    =      0;
46 static int                      no_samples                      =      0;
47 static int                      sample_address                  =      0;
48 static int                      multiplex                       =      0;
49 static int                      multiplex_fd                    =     -1;
50
51 static long                     samples                         =      0;
52 static struct timeval           last_read;
53 static struct timeval           this_read;
54
55 static u64                      bytes_written                   =      0;
56
57 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
58
59 static int                      nr_poll                         =      0;
60 static int                      nr_cpu                          =      0;
61
62 static int                      file_new                        =      1;
63
64 struct perf_header              *header                         =   NULL;
65
66 struct mmap_data {
67         int                     counter;
68         void                    *base;
69         unsigned int            mask;
70         unsigned int            prev;
71 };
72
73 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
74
75 static unsigned long mmap_read_head(struct mmap_data *md)
76 {
77         struct perf_event_mmap_page *pc = md->base;
78         long head;
79
80         head = pc->data_head;
81         rmb();
82
83         return head;
84 }
85
86 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
87 {
88         struct perf_event_mmap_page *pc = md->base;
89
90         /*
91          * ensure all reads are done before we write the tail out.
92          */
93         /* mb(); */
94         pc->data_tail = tail;
95 }
96
97 static void write_output(void *buf, size_t size)
98 {
99         while (size) {
100                 int ret = write(output, buf, size);
101
102                 if (ret < 0)
103                         die("failed to write");
104
105                 size -= ret;
106                 buf += ret;
107
108                 bytes_written += ret;
109         }
110 }
111
112 static void mmap_read(struct mmap_data *md)
113 {
114         unsigned int head = mmap_read_head(md);
115         unsigned int old = md->prev;
116         unsigned char *data = md->base + page_size;
117         unsigned long size;
118         void *buf;
119         int diff;
120
121         gettimeofday(&this_read, NULL);
122
123         /*
124          * If we're further behind than half the buffer, there's a chance
125          * the writer will bite our tail and mess up the samples under us.
126          *
127          * If we somehow ended up ahead of the head, we got messed up.
128          *
129          * In either case, truncate and restart at head.
130          */
131         diff = head - old;
132         if (diff < 0) {
133                 struct timeval iv;
134                 unsigned long msecs;
135
136                 timersub(&this_read, &last_read, &iv);
137                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
138
139                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
140                                 "  Last read %lu msecs ago.\n", msecs);
141
142                 /*
143                  * head points to a known good entry, start there.
144                  */
145                 old = head;
146         }
147
148         last_read = this_read;
149
150         if (old != head)
151                 samples++;
152
153         size = head - old;
154
155         if ((old & md->mask) + size != (head & md->mask)) {
156                 buf = &data[old & md->mask];
157                 size = md->mask + 1 - (old & md->mask);
158                 old += size;
159
160                 write_output(buf, size);
161         }
162
163         buf = &data[old & md->mask];
164         size = head - old;
165         old += size;
166
167         write_output(buf, size);
168
169         md->prev = old;
170         mmap_write_tail(md, old);
171 }
172
173 static volatile int done = 0;
174 static volatile int signr = -1;
175
176 static void sig_handler(int sig)
177 {
178         done = 1;
179         signr = sig;
180 }
181
182 static void sig_atexit(void)
183 {
184         if (child_pid != -1)
185                 kill(child_pid, SIGTERM);
186
187         if (signr == -1)
188                 return;
189
190         signal(signr, SIG_DFL);
191         kill(getpid(), signr);
192 }
193
194 static pid_t pid_synthesize_comm_event(pid_t pid, int full)
195 {
196         struct comm_event comm_ev;
197         char filename[PATH_MAX];
198         char bf[BUFSIZ];
199         FILE *fp;
200         size_t size = 0;
201         DIR *tasks;
202         struct dirent dirent, *next;
203         pid_t tgid = 0;
204
205         snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
206
207         fp = fopen(filename, "r");
208         if (fp == NULL) {
209                 /*
210                  * We raced with a task exiting - just return:
211                  */
212                 if (verbose)
213                         fprintf(stderr, "couldn't open %s\n", filename);
214                 return 0;
215         }
216
217         memset(&comm_ev, 0, sizeof(comm_ev));
218         while (!comm_ev.comm[0] || !comm_ev.pid) {
219                 if (fgets(bf, sizeof(bf), fp) == NULL)
220                         goto out_failure;
221
222                 if (memcmp(bf, "Name:", 5) == 0) {
223                         char *name = bf + 5;
224                         while (*name && isspace(*name))
225                                 ++name;
226                         size = strlen(name) - 1;
227                         memcpy(comm_ev.comm, name, size++);
228                 } else if (memcmp(bf, "Tgid:", 5) == 0) {
229                         char *tgids = bf + 5;
230                         while (*tgids && isspace(*tgids))
231                                 ++tgids;
232                         tgid = comm_ev.pid = atoi(tgids);
233                 }
234         }
235
236         comm_ev.header.type = PERF_RECORD_COMM;
237         size = ALIGN(size, sizeof(u64));
238         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
239
240         if (!full) {
241                 comm_ev.tid = pid;
242
243                 write_output(&comm_ev, comm_ev.header.size);
244                 goto out_fclose;
245         }
246
247         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
248
249         tasks = opendir(filename);
250         while (!readdir_r(tasks, &dirent, &next) && next) {
251                 char *end;
252                 pid = strtol(dirent.d_name, &end, 10);
253                 if (*end)
254                         continue;
255
256                 comm_ev.tid = pid;
257
258                 write_output(&comm_ev, comm_ev.header.size);
259         }
260         closedir(tasks);
261
262 out_fclose:
263         fclose(fp);
264         return tgid;
265
266 out_failure:
267         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
268                 filename);
269         exit(EXIT_FAILURE);
270 }
271
272 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
273 {
274         char filename[PATH_MAX];
275         FILE *fp;
276
277         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
278
279         fp = fopen(filename, "r");
280         if (fp == NULL) {
281                 /*
282                  * We raced with a task exiting - just return:
283                  */
284                 if (verbose)
285                         fprintf(stderr, "couldn't open %s\n", filename);
286                 return;
287         }
288         while (1) {
289                 char bf[BUFSIZ], *pbf = bf;
290                 struct mmap_event mmap_ev = {
291                         .header = { .type = PERF_RECORD_MMAP },
292                 };
293                 int n;
294                 size_t size;
295                 if (fgets(bf, sizeof(bf), fp) == NULL)
296                         break;
297
298                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
299                 n = hex2u64(pbf, &mmap_ev.start);
300                 if (n < 0)
301                         continue;
302                 pbf += n + 1;
303                 n = hex2u64(pbf, &mmap_ev.len);
304                 if (n < 0)
305                         continue;
306                 pbf += n + 3;
307                 if (*pbf == 'x') { /* vm_exec */
308                         char *execname = strchr(bf, '/');
309
310                         /* Catch VDSO */
311                         if (execname == NULL)
312                                 execname = strstr(bf, "[vdso]");
313
314                         if (execname == NULL)
315                                 continue;
316
317                         size = strlen(execname);
318                         execname[size - 1] = '\0'; /* Remove \n */
319                         memcpy(mmap_ev.filename, execname, size);
320                         size = ALIGN(size, sizeof(u64));
321                         mmap_ev.len -= mmap_ev.start;
322                         mmap_ev.header.size = (sizeof(mmap_ev) -
323                                                (sizeof(mmap_ev.filename) - size));
324                         mmap_ev.pid = tgid;
325                         mmap_ev.tid = pid;
326
327                         write_output(&mmap_ev, mmap_ev.header.size);
328                 }
329         }
330
331         fclose(fp);
332 }
333
334 static void synthesize_all(void)
335 {
336         DIR *proc;
337         struct dirent dirent, *next;
338
339         proc = opendir("/proc");
340
341         while (!readdir_r(proc, &dirent, &next) && next) {
342                 char *end;
343                 pid_t pid, tgid;
344
345                 pid = strtol(dirent.d_name, &end, 10);
346                 if (*end) /* only interested in proper numerical dirents */
347                         continue;
348
349                 tgid = pid_synthesize_comm_event(pid, 1);
350                 pid_synthesize_mmap_samples(pid, tgid);
351         }
352
353         closedir(proc);
354 }
355
356 static int group_fd;
357
358 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
359 {
360         struct perf_header_attr *h_attr;
361
362         if (nr < header->attrs) {
363                 h_attr = header->attr[nr];
364         } else {
365                 h_attr = perf_header_attr__new(a);
366                 perf_header__add_attr(header, h_attr);
367         }
368
369         return h_attr;
370 }
371
372 static void create_counter(int counter, int cpu, pid_t pid)
373 {
374         char *filter = filters[counter];
375         struct perf_event_attr *attr = attrs + counter;
376         struct perf_header_attr *h_attr;
377         int track = !counter; /* only the first counter needs these */
378         int ret;
379         struct {
380                 u64 count;
381                 u64 time_enabled;
382                 u64 time_running;
383                 u64 id;
384         } read_data;
385
386         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
387                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
388                                   PERF_FORMAT_ID;
389
390         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
391
392         if (freq) {
393                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
394                 attr->freq              = 1;
395                 attr->sample_freq       = freq;
396         }
397
398         if (no_samples)
399                 attr->sample_freq = 0;
400
401         if (inherit_stat)
402                 attr->inherit_stat = 1;
403
404         if (sample_address)
405                 attr->sample_type       |= PERF_SAMPLE_ADDR;
406
407         if (call_graph)
408                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
409
410         if (raw_samples) {
411                 attr->sample_type       |= PERF_SAMPLE_TIME;
412                 attr->sample_type       |= PERF_SAMPLE_RAW;
413                 attr->sample_type       |= PERF_SAMPLE_CPU;
414         }
415
416         attr->mmap              = track;
417         attr->comm              = track;
418         attr->inherit           = (cpu < 0) && inherit;
419         attr->disabled          = 1;
420
421 try_again:
422         fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
423
424         if (fd[nr_cpu][counter] < 0) {
425                 int err = errno;
426
427                 if (err == EPERM)
428                         die("Permission error - are you root?\n");
429                 else if (err ==  ENODEV && profile_cpu != -1)
430                         die("No such device - did you specify an out-of-range profile CPU?\n");
431
432                 /*
433                  * If it's cycles then fall back to hrtimer
434                  * based cpu-clock-tick sw counter, which
435                  * is always available even if no PMU support:
436                  */
437                 if (attr->type == PERF_TYPE_HARDWARE
438                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
439
440                         if (verbose)
441                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
442                         attr->type = PERF_TYPE_SOFTWARE;
443                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
444                         goto try_again;
445                 }
446                 printf("\n");
447                 error("perfcounter syscall returned with %d (%s)\n",
448                         fd[nr_cpu][counter], strerror(err));
449                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
450                 exit(-1);
451         }
452
453         h_attr = get_header_attr(attr, counter);
454
455         if (!file_new) {
456                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
457                         fprintf(stderr, "incompatible append\n");
458                         exit(-1);
459                 }
460         }
461
462         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
463                 perror("Unable to read perf file descriptor\n");
464                 exit(-1);
465         }
466
467         perf_header_attr__add_id(h_attr, read_data.id);
468
469         assert(fd[nr_cpu][counter] >= 0);
470         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
471
472         /*
473          * First counter acts as the group leader:
474          */
475         if (group && group_fd == -1)
476                 group_fd = fd[nr_cpu][counter];
477         if (multiplex && multiplex_fd == -1)
478                 multiplex_fd = fd[nr_cpu][counter];
479
480         if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
481
482                 ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
483                 assert(ret != -1);
484         } else {
485                 event_array[nr_poll].fd = fd[nr_cpu][counter];
486                 event_array[nr_poll].events = POLLIN;
487                 nr_poll++;
488
489                 mmap_array[nr_cpu][counter].counter = counter;
490                 mmap_array[nr_cpu][counter].prev = 0;
491                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
492                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
493                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
494                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
495                         error("failed to mmap with %d (%s)\n", errno, strerror(errno));
496                         exit(-1);
497                 }
498         }
499
500         if (filter != NULL) {
501                 ret = ioctl(fd[nr_cpu][counter],
502                             PERF_EVENT_IOC_SET_FILTER, filter);
503                 if (ret) {
504                         error("failed to set filter with %d (%s)\n", errno,
505                               strerror(errno));
506                         exit(-1);
507                 }
508         }
509
510         ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
511 }
512
513 static void open_counters(int cpu, pid_t pid)
514 {
515         int counter;
516
517         group_fd = -1;
518         for (counter = 0; counter < nr_counters; counter++)
519                 create_counter(counter, cpu, pid);
520
521         nr_cpu++;
522 }
523
524 static void atexit_header(void)
525 {
526         header->data_size += bytes_written;
527
528         perf_header__write(header, output);
529 }
530
531 static int __cmd_record(int argc, const char **argv)
532 {
533         int i, counter;
534         struct stat st;
535         pid_t pid = 0;
536         int flags;
537         int ret;
538         unsigned long waking = 0;
539
540         page_size = sysconf(_SC_PAGE_SIZE);
541         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
542         assert(nr_cpus <= MAX_NR_CPUS);
543         assert(nr_cpus >= 0);
544
545         atexit(sig_atexit);
546         signal(SIGCHLD, sig_handler);
547         signal(SIGINT, sig_handler);
548
549         if (!stat(output_name, &st) && st.st_size) {
550                 if (!force && !append_file) {
551                         fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
552                                         output_name);
553                         exit(-1);
554                 }
555         } else {
556                 append_file = 0;
557         }
558
559         flags = O_CREAT|O_RDWR;
560         if (append_file)
561                 file_new = 0;
562         else
563                 flags |= O_TRUNC;
564
565         output = open(output_name, flags, S_IRUSR|S_IWUSR);
566         if (output < 0) {
567                 perror("failed to create output file");
568                 exit(-1);
569         }
570
571         if (!file_new)
572                 header = perf_header__read(output);
573         else
574                 header = perf_header__new();
575
576         if (raw_samples) {
577                 perf_header__feat_trace_info(header);
578         } else {
579                 for (i = 0; i < nr_counters; i++) {
580                         if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
581                                 perf_header__feat_trace_info(header);
582                                 break;
583                         }
584                 }
585         }
586
587         atexit(atexit_header);
588
589         if (!system_wide) {
590                 pid = target_pid;
591                 if (pid == -1)
592                         pid = getpid();
593
594                 open_counters(profile_cpu, pid);
595         } else {
596                 if (profile_cpu != -1) {
597                         open_counters(profile_cpu, target_pid);
598                 } else {
599                         for (i = 0; i < nr_cpus; i++)
600                                 open_counters(i, target_pid);
601                 }
602         }
603
604         if (file_new)
605                 perf_header__write(header, output);
606
607         if (!system_wide) {
608                 pid_t tgid = pid_synthesize_comm_event(pid, 0);
609                 pid_synthesize_mmap_samples(pid, tgid);
610         } else
611                 synthesize_all();
612
613         if (target_pid == -1 && argc) {
614                 pid = fork();
615                 if (pid < 0)
616                         perror("failed to fork");
617
618                 if (!pid) {
619                         if (execvp(argv[0], (char **)argv)) {
620                                 perror(argv[0]);
621                                 exit(-1);
622                         }
623                 }
624
625                 child_pid = pid;
626         }
627
628         if (realtime_prio) {
629                 struct sched_param param;
630
631                 param.sched_priority = realtime_prio;
632                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
633                         pr_err("Could not set realtime priority.\n");
634                         exit(-1);
635                 }
636         }
637
638         for (;;) {
639                 int hits = samples;
640
641                 for (i = 0; i < nr_cpu; i++) {
642                         for (counter = 0; counter < nr_counters; counter++) {
643                                 if (mmap_array[i][counter].base)
644                                         mmap_read(&mmap_array[i][counter]);
645                         }
646                 }
647
648                 if (hits == samples) {
649                         if (done)
650                                 break;
651                         ret = poll(event_array, nr_poll, -1);
652                         waking++;
653                 }
654
655                 if (done) {
656                         for (i = 0; i < nr_cpu; i++) {
657                                 for (counter = 0; counter < nr_counters; counter++)
658                                         ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
659                         }
660                 }
661         }
662
663         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
664
665         /*
666          * Approximate RIP event size: 24 bytes.
667          */
668         fprintf(stderr,
669                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
670                 (double)bytes_written / 1024.0 / 1024.0,
671                 output_name,
672                 bytes_written / 24);
673
674         return 0;
675 }
676
677 static const char * const record_usage[] = {
678         "perf record [<options>] [<command>]",
679         "perf record [<options>] -- <command> [<options>]",
680         NULL
681 };
682
683 static const struct option options[] = {
684         OPT_CALLBACK('e', "event", NULL, "event",
685                      "event selector. use 'perf list' to list available events",
686                      parse_events),
687         OPT_CALLBACK(0, "filter", NULL, "filter",
688                      "event filter", parse_filter),
689         OPT_INTEGER('p', "pid", &target_pid,
690                     "record events on existing pid"),
691         OPT_INTEGER('r', "realtime", &realtime_prio,
692                     "collect data with this RT SCHED_FIFO priority"),
693         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
694                     "collect raw sample records from all opened counters"),
695         OPT_BOOLEAN('a', "all-cpus", &system_wide,
696                             "system-wide collection from all CPUs"),
697         OPT_BOOLEAN('A', "append", &append_file,
698                             "append to the output file to do incremental profiling"),
699         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
700                             "CPU to profile on"),
701         OPT_BOOLEAN('f', "force", &force,
702                         "overwrite existing data file"),
703         OPT_LONG('c', "count", &default_interval,
704                     "event period to sample"),
705         OPT_STRING('o', "output", &output_name, "file",
706                     "output file name"),
707         OPT_BOOLEAN('i', "inherit", &inherit,
708                     "child tasks inherit counters"),
709         OPT_INTEGER('F', "freq", &freq,
710                     "profile at this frequency"),
711         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
712                     "number of mmap data pages"),
713         OPT_BOOLEAN('g', "call-graph", &call_graph,
714                     "do call-graph (stack chain/backtrace) recording"),
715         OPT_BOOLEAN('v', "verbose", &verbose,
716                     "be more verbose (show counter open errors, etc)"),
717         OPT_BOOLEAN('s', "stat", &inherit_stat,
718                     "per thread counts"),
719         OPT_BOOLEAN('d', "data", &sample_address,
720                     "Sample addresses"),
721         OPT_BOOLEAN('n', "no-samples", &no_samples,
722                     "don't sample"),
723         OPT_BOOLEAN('M', "multiplex", &multiplex,
724                     "multiplex counter output in a single channel"),
725         OPT_END()
726 };
727
728 int cmd_record(int argc, const char **argv, const char *prefix __used)
729 {
730         int counter;
731
732         argc = parse_options(argc, argv, options, record_usage,
733                 PARSE_OPT_STOP_AT_NON_OPTION);
734         if (!argc && target_pid == -1 && !system_wide)
735                 usage_with_options(record_usage, options);
736
737         if (!nr_counters) {
738                 nr_counters     = 1;
739                 attrs[0].type   = PERF_TYPE_HARDWARE;
740                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
741         }
742
743         /*
744          * User specified count overrides default frequency.
745          */
746         if (default_interval)
747                 freq = 0;
748         else if (freq) {
749                 default_interval = freq;
750         } else {
751                 fprintf(stderr, "frequency and count are zero, aborting\n");
752                 exit(EXIT_FAILURE);
753         }
754
755         for (counter = 0; counter < nr_counters; counter++) {
756                 if (attrs[counter].sample_period)
757                         continue;
758
759                 attrs[counter].sample_period = default_interval;
760         }
761
762         return __cmd_record(argc, argv);
763 }