68a9be0d1513760a9dad9f938473566544ce7885
[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
19 #include <unistd.h>
20 #include <sched.h>
21
22 #define ALIGN(x, a)             __ALIGN_MASK(x, (typeof(x))(a)-1)
23 #define __ALIGN_MASK(x, mask)   (((x)+(mask))&~(mask))
24
25 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
26
27 static long                     default_interval                = 100000;
28
29 static int                      nr_cpus                         = 0;
30 static unsigned int             page_size;
31 static unsigned int             mmap_pages                      = 128;
32 static int                      freq                            = 0;
33 static int                      output;
34 static const char               *output_name                    = "perf.data";
35 static int                      group                           = 0;
36 static unsigned int             realtime_prio                   = 0;
37 static int                      system_wide                     = 0;
38 static pid_t                    target_pid                      = -1;
39 static int                      inherit                         = 1;
40 static int                      force                           = 0;
41 static int                      append_file                     = 0;
42 static int                      call_graph                      = 0;
43 static int                      verbose                         = 0;
44 static int                      inherit_stat                    = 0;
45 static int                      no_samples                      = 0;
46 static int                      sample_address                  = 0;
47
48 static long                     samples;
49 static struct timeval           last_read;
50 static struct timeval           this_read;
51
52 static u64                      bytes_written;
53
54 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
55
56 static int                      nr_poll;
57 static int                      nr_cpu;
58
59 static int                      file_new = 1;
60
61 struct perf_header              *header;
62
63 struct mmap_event {
64         struct perf_event_header        header;
65         u32                             pid;
66         u32                             tid;
67         u64                             start;
68         u64                             len;
69         u64                             pgoff;
70         char                            filename[PATH_MAX];
71 };
72
73 struct comm_event {
74         struct perf_event_header        header;
75         u32                             pid;
76         u32                             tid;
77         char                            comm[16];
78 };
79
80
81 struct mmap_data {
82         int                     counter;
83         void                    *base;
84         unsigned int            mask;
85         unsigned int            prev;
86 };
87
88 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
89
90 static unsigned long mmap_read_head(struct mmap_data *md)
91 {
92         struct perf_counter_mmap_page *pc = md->base;
93         long head;
94
95         head = pc->data_head;
96         rmb();
97
98         return head;
99 }
100
101 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
102 {
103         struct perf_counter_mmap_page *pc = md->base;
104
105         /*
106          * ensure all reads are done before we write the tail out.
107          */
108         /* mb(); */
109         pc->data_tail = tail;
110 }
111
112 static void write_output(void *buf, size_t size)
113 {
114         while (size) {
115                 int ret = write(output, buf, size);
116
117                 if (ret < 0)
118                         die("failed to write");
119
120                 size -= ret;
121                 buf += ret;
122
123                 bytes_written += ret;
124         }
125 }
126
127 static void mmap_read(struct mmap_data *md)
128 {
129         unsigned int head = mmap_read_head(md);
130         unsigned int old = md->prev;
131         unsigned char *data = md->base + page_size;
132         unsigned long size;
133         void *buf;
134         int diff;
135
136         gettimeofday(&this_read, NULL);
137
138         /*
139          * If we're further behind than half the buffer, there's a chance
140          * the writer will bite our tail and mess up the samples under us.
141          *
142          * If we somehow ended up ahead of the head, we got messed up.
143          *
144          * In either case, truncate and restart at head.
145          */
146         diff = head - old;
147         if (diff < 0) {
148                 struct timeval iv;
149                 unsigned long msecs;
150
151                 timersub(&this_read, &last_read, &iv);
152                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
153
154                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
155                                 "  Last read %lu msecs ago.\n", msecs);
156
157                 /*
158                  * head points to a known good entry, start there.
159                  */
160                 old = head;
161         }
162
163         last_read = this_read;
164
165         if (old != head)
166                 samples++;
167
168         size = head - old;
169
170         if ((old & md->mask) + size != (head & md->mask)) {
171                 buf = &data[old & md->mask];
172                 size = md->mask + 1 - (old & md->mask);
173                 old += size;
174
175                 write_output(buf, size);
176         }
177
178         buf = &data[old & md->mask];
179         size = head - old;
180         old += size;
181
182         write_output(buf, size);
183
184         md->prev = old;
185         mmap_write_tail(md, old);
186 }
187
188 static volatile int done = 0;
189 static volatile int signr = -1;
190
191 static void sig_handler(int sig)
192 {
193         done = 1;
194         signr = sig;
195 }
196
197 static void sig_atexit(void)
198 {
199         if (signr == -1)
200                 return;
201
202         signal(signr, SIG_DFL);
203         kill(getpid(), signr);
204 }
205
206 static void pid_synthesize_comm_event(pid_t pid, int full)
207 {
208         struct comm_event comm_ev;
209         char filename[PATH_MAX];
210         char bf[BUFSIZ];
211         int fd;
212         size_t size;
213         char *field, *sep;
214         DIR *tasks;
215         struct dirent dirent, *next;
216
217         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
218
219         fd = open(filename, O_RDONLY);
220         if (fd < 0) {
221                 /*
222                  * We raced with a task exiting - just return:
223                  */
224                 if (verbose)
225                         fprintf(stderr, "couldn't open %s\n", filename);
226                 return;
227         }
228         if (read(fd, bf, sizeof(bf)) < 0) {
229                 fprintf(stderr, "couldn't read %s\n", filename);
230                 exit(EXIT_FAILURE);
231         }
232         close(fd);
233
234         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
235         memset(&comm_ev, 0, sizeof(comm_ev));
236         field = strchr(bf, '(');
237         if (field == NULL)
238                 goto out_failure;
239         sep = strchr(++field, ')');
240         if (sep == NULL)
241                 goto out_failure;
242         size = sep - field;
243         memcpy(comm_ev.comm, field, size++);
244
245         comm_ev.pid = pid;
246         comm_ev.header.type = PERF_EVENT_COMM;
247         size = ALIGN(size, sizeof(u64));
248         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
249
250         if (!full) {
251                 comm_ev.tid = pid;
252
253                 write_output(&comm_ev, comm_ev.header.size);
254                 return;
255         }
256
257         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
258
259         tasks = opendir(filename);
260         while (!readdir_r(tasks, &dirent, &next) && next) {
261                 char *end;
262                 pid = strtol(dirent.d_name, &end, 10);
263                 if (*end)
264                         continue;
265
266                 comm_ev.tid = pid;
267
268                 write_output(&comm_ev, comm_ev.header.size);
269         }
270         closedir(tasks);
271         return;
272
273 out_failure:
274         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
275                 filename);
276         exit(EXIT_FAILURE);
277 }
278
279 static void pid_synthesize_mmap_samples(pid_t pid)
280 {
281         char filename[PATH_MAX];
282         FILE *fp;
283
284         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
285
286         fp = fopen(filename, "r");
287         if (fp == NULL) {
288                 /*
289                  * We raced with a task exiting - just return:
290                  */
291                 if (verbose)
292                         fprintf(stderr, "couldn't open %s\n", filename);
293                 return;
294         }
295         while (1) {
296                 char bf[BUFSIZ], *pbf = bf;
297                 struct mmap_event mmap_ev = {
298                         .header = { .type = PERF_EVENT_MMAP },
299                 };
300                 int n;
301                 size_t size;
302                 if (fgets(bf, sizeof(bf), fp) == NULL)
303                         break;
304
305                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
306                 n = hex2u64(pbf, &mmap_ev.start);
307                 if (n < 0)
308                         continue;
309                 pbf += n + 1;
310                 n = hex2u64(pbf, &mmap_ev.len);
311                 if (n < 0)
312                         continue;
313                 pbf += n + 3;
314                 if (*pbf == 'x') { /* vm_exec */
315                         char *execname = strchr(bf, '/');
316
317                         /* Catch VDSO */
318                         if (execname == NULL)
319                                 execname = strstr(bf, "[vdso]");
320
321                         if (execname == NULL)
322                                 continue;
323
324                         size = strlen(execname);
325                         execname[size - 1] = '\0'; /* Remove \n */
326                         memcpy(mmap_ev.filename, execname, size);
327                         size = ALIGN(size, sizeof(u64));
328                         mmap_ev.len -= mmap_ev.start;
329                         mmap_ev.header.size = (sizeof(mmap_ev) -
330                                                (sizeof(mmap_ev.filename) - size));
331                         mmap_ev.pid = pid;
332                         mmap_ev.tid = pid;
333
334                         write_output(&mmap_ev, mmap_ev.header.size);
335                 }
336         }
337
338         fclose(fp);
339 }
340
341 static void synthesize_all(void)
342 {
343         DIR *proc;
344         struct dirent dirent, *next;
345
346         proc = opendir("/proc");
347
348         while (!readdir_r(proc, &dirent, &next) && next) {
349                 char *end;
350                 pid_t pid;
351
352                 pid = strtol(dirent.d_name, &end, 10);
353                 if (*end) /* only interested in proper numerical dirents */
354                         continue;
355
356                 pid_synthesize_comm_event(pid, 1);
357                 pid_synthesize_mmap_samples(pid);
358         }
359
360         closedir(proc);
361 }
362
363 static int group_fd;
364
365 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
366 {
367         struct perf_header_attr *h_attr;
368
369         if (nr < header->attrs) {
370                 h_attr = header->attr[nr];
371         } else {
372                 h_attr = perf_header_attr__new(a);
373                 perf_header__add_attr(header, h_attr);
374         }
375
376         return h_attr;
377 }
378
379 static void create_counter(int counter, int cpu, pid_t pid)
380 {
381         struct perf_counter_attr *attr = attrs + counter;
382         struct perf_header_attr *h_attr;
383         int track = !counter; /* only the first counter needs these */
384         struct {
385                 u64 count;
386                 u64 time_enabled;
387                 u64 time_running;
388                 u64 id;
389         } read_data;
390
391         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
392                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
393                                   PERF_FORMAT_ID;
394
395         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
396
397         if (freq) {
398                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
399                 attr->freq              = 1;
400                 attr->sample_freq       = freq;
401         }
402
403         if (no_samples)
404                 attr->sample_freq = 0;
405
406         if (inherit_stat)
407                 attr->inherit_stat = 1;
408
409         if (sample_address)
410                 attr->sample_type       |= PERF_SAMPLE_ADDR;
411
412         if (call_graph)
413                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
414
415         attr->mmap              = track;
416         attr->comm              = track;
417         attr->inherit           = (cpu < 0) && inherit;
418         attr->disabled          = 1;
419
420 try_again:
421         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
422
423         if (fd[nr_cpu][counter] < 0) {
424                 int err = errno;
425
426                 if (err == EPERM)
427                         die("Permission error - are you root?\n");
428
429                 /*
430                  * If it's cycles then fall back to hrtimer
431                  * based cpu-clock-tick sw counter, which
432                  * is always available even if no PMU support:
433                  */
434                 if (attr->type == PERF_TYPE_HARDWARE
435                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
436
437                         if (verbose)
438                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
439                         attr->type = PERF_TYPE_SOFTWARE;
440                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
441                         goto try_again;
442                 }
443                 printf("\n");
444                 error("perfcounter syscall returned with %d (%s)\n",
445                         fd[nr_cpu][counter], strerror(err));
446                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
447                 exit(-1);
448         }
449
450         h_attr = get_header_attr(attr, counter);
451
452         if (!file_new) {
453                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
454                         fprintf(stderr, "incompatible append\n");
455                         exit(-1);
456                 }
457         }
458
459         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
460                 perror("Unable to read perf file descriptor\n");
461                 exit(-1);
462         }
463
464         perf_header_attr__add_id(h_attr, read_data.id);
465
466         assert(fd[nr_cpu][counter] >= 0);
467         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
468
469         /*
470          * First counter acts as the group leader:
471          */
472         if (group && group_fd == -1)
473                 group_fd = fd[nr_cpu][counter];
474
475         event_array[nr_poll].fd = fd[nr_cpu][counter];
476         event_array[nr_poll].events = POLLIN;
477         nr_poll++;
478
479         mmap_array[nr_cpu][counter].counter = counter;
480         mmap_array[nr_cpu][counter].prev = 0;
481         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
482         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
483                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
484         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
485                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
486                 exit(-1);
487         }
488
489         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
490 }
491
492 static void open_counters(int cpu, pid_t pid)
493 {
494         int counter;
495
496         group_fd = -1;
497         for (counter = 0; counter < nr_counters; counter++)
498                 create_counter(counter, cpu, pid);
499
500         nr_cpu++;
501 }
502
503 static void atexit_header(void)
504 {
505         header->data_size += bytes_written;
506
507         perf_header__write(header, output);
508 }
509
510 static int __cmd_record(int argc, const char **argv)
511 {
512         int i, counter;
513         struct stat st;
514         pid_t pid = 0;
515         int flags;
516         int ret;
517
518         page_size = sysconf(_SC_PAGE_SIZE);
519         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
520         assert(nr_cpus <= MAX_NR_CPUS);
521         assert(nr_cpus >= 0);
522
523         atexit(sig_atexit);
524         signal(SIGCHLD, sig_handler);
525         signal(SIGINT, sig_handler);
526
527         if (!stat(output_name, &st) && !force && !append_file) {
528                 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
529                                 output_name);
530                 exit(-1);
531         }
532
533         flags = O_CREAT|O_RDWR;
534         if (append_file)
535                 file_new = 0;
536         else
537                 flags |= O_TRUNC;
538
539         output = open(output_name, flags, S_IRUSR|S_IWUSR);
540         if (output < 0) {
541                 perror("failed to create output file");
542                 exit(-1);
543         }
544
545         if (!file_new)
546                 header = perf_header__read(output);
547         else
548                 header = perf_header__new();
549
550         atexit(atexit_header);
551
552         if (!system_wide) {
553                 pid = target_pid;
554                 if (pid == -1)
555                         pid = getpid();
556
557                 open_counters(-1, pid);
558         } else for (i = 0; i < nr_cpus; i++)
559                 open_counters(i, target_pid);
560
561         if (file_new)
562                 perf_header__write(header, output);
563
564         if (!system_wide) {
565                 pid_synthesize_comm_event(pid, 0);
566                 pid_synthesize_mmap_samples(pid);
567         } else
568                 synthesize_all();
569
570         if (target_pid == -1 && argc) {
571                 pid = fork();
572                 if (pid < 0)
573                         perror("failed to fork");
574
575                 if (!pid) {
576                         if (execvp(argv[0], (char **)argv)) {
577                                 perror(argv[0]);
578                                 exit(-1);
579                         }
580                 }
581         }
582
583         if (realtime_prio) {
584                 struct sched_param param;
585
586                 param.sched_priority = realtime_prio;
587                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
588                         printf("Could not set realtime priority.\n");
589                         exit(-1);
590                 }
591         }
592
593         for (;;) {
594                 int hits = samples;
595
596                 for (i = 0; i < nr_cpu; i++) {
597                         for (counter = 0; counter < nr_counters; counter++)
598                                 mmap_read(&mmap_array[i][counter]);
599                 }
600
601                 if (hits == samples) {
602                         if (done)
603                                 break;
604                         ret = poll(event_array, nr_poll, 100);
605                 }
606         }
607
608         /*
609          * Approximate RIP event size: 24 bytes.
610          */
611         fprintf(stderr,
612                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
613                 (double)bytes_written / 1024.0 / 1024.0,
614                 output_name,
615                 bytes_written / 24);
616
617         return 0;
618 }
619
620 static const char * const record_usage[] = {
621         "perf record [<options>] [<command>]",
622         "perf record [<options>] -- <command> [<options>]",
623         NULL
624 };
625
626 static const struct option options[] = {
627         OPT_CALLBACK('e', "event", NULL, "event",
628                      "event selector. use 'perf list' to list available events",
629                      parse_events),
630         OPT_INTEGER('p', "pid", &target_pid,
631                     "record events on existing pid"),
632         OPT_INTEGER('r', "realtime", &realtime_prio,
633                     "collect data with this RT SCHED_FIFO priority"),
634         OPT_BOOLEAN('a', "all-cpus", &system_wide,
635                             "system-wide collection from all CPUs"),
636         OPT_BOOLEAN('A', "append", &append_file,
637                             "append to the output file to do incremental profiling"),
638         OPT_BOOLEAN('f', "force", &force,
639                         "overwrite existing data file"),
640         OPT_LONG('c', "count", &default_interval,
641                     "event period to sample"),
642         OPT_STRING('o', "output", &output_name, "file",
643                     "output file name"),
644         OPT_BOOLEAN('i', "inherit", &inherit,
645                     "child tasks inherit counters"),
646         OPT_INTEGER('F', "freq", &freq,
647                     "profile at this frequency"),
648         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
649                     "number of mmap data pages"),
650         OPT_BOOLEAN('g', "call-graph", &call_graph,
651                     "do call-graph (stack chain/backtrace) recording"),
652         OPT_BOOLEAN('v', "verbose", &verbose,
653                     "be more verbose (show counter open errors, etc)"),
654         OPT_BOOLEAN('s', "stat", &inherit_stat,
655                     "per thread counts"),
656         OPT_BOOLEAN('d', "data", &sample_address,
657                     "Sample addresses"),
658         OPT_BOOLEAN('n', "no-samples", &no_samples,
659                     "don't sample"),
660         OPT_END()
661 };
662
663 int cmd_record(int argc, const char **argv, const char *prefix __used)
664 {
665         int counter;
666
667         argc = parse_options(argc, argv, options, record_usage, 0);
668         if (!argc && target_pid == -1 && !system_wide)
669                 usage_with_options(record_usage, options);
670
671         if (!nr_counters) {
672                 nr_counters     = 1;
673                 attrs[0].type   = PERF_TYPE_HARDWARE;
674                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
675         }
676
677         for (counter = 0; counter < nr_counters; counter++) {
678                 if (attrs[counter].sample_period)
679                         continue;
680
681                 attrs[counter].sample_period = default_interval;
682         }
683
684         return __cmd_record(argc, argv);
685 }