8b2c860c49a23111fc5b677570ba158559d321e0
[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 #define ALIGN(x, a)             __ALIGN_MASK(x, (typeof(x))(a)-1)
25 #define __ALIGN_MASK(x, mask)   (((x)+(mask))&~(mask))
26
27 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
28
29 static long                     default_interval                =      0;
30
31 static int                      nr_cpus                         =      0;
32 static unsigned int             page_size;
33 static unsigned int             mmap_pages                      =    128;
34 static int                      freq                            =   1000;
35 static int                      output;
36 static const char               *output_name                    = "perf.data";
37 static int                      group                           =      0;
38 static unsigned int             realtime_prio                   =      0;
39 static int                      raw_samples                     =      0;
40 static int                      system_wide                     =      0;
41 static int                      profile_cpu                     =     -1;
42 static pid_t                    target_pid                      =     -1;
43 static pid_t                    child_pid                       =     -1;
44 static int                      inherit                         =      1;
45 static int                      force                           =      0;
46 static int                      append_file                     =      0;
47 static int                      call_graph                      =      0;
48 static int                      inherit_stat                    =      0;
49 static int                      no_samples                      =      0;
50 static int                      sample_address                  =      0;
51 static int                      multiplex                       =      0;
52 static int                      multiplex_fd                    =     -1;
53
54 static long                     samples                         =      0;
55 static struct timeval           last_read;
56 static struct timeval           this_read;
57
58 static u64                      bytes_written                   =      0;
59
60 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
61
62 static int                      nr_poll                         =      0;
63 static int                      nr_cpu                          =      0;
64
65 static int                      file_new                        =      1;
66
67 struct perf_header              *header                         =   NULL;
68
69 struct mmap_data {
70         int                     counter;
71         void                    *base;
72         unsigned int            mask;
73         unsigned int            prev;
74 };
75
76 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
77
78 static unsigned long mmap_read_head(struct mmap_data *md)
79 {
80         struct perf_event_mmap_page *pc = md->base;
81         long head;
82
83         head = pc->data_head;
84         rmb();
85
86         return head;
87 }
88
89 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
90 {
91         struct perf_event_mmap_page *pc = md->base;
92
93         /*
94          * ensure all reads are done before we write the tail out.
95          */
96         /* mb(); */
97         pc->data_tail = tail;
98 }
99
100 static void write_output(void *buf, size_t size)
101 {
102         while (size) {
103                 int ret = write(output, buf, size);
104
105                 if (ret < 0)
106                         die("failed to write");
107
108                 size -= ret;
109                 buf += ret;
110
111                 bytes_written += ret;
112         }
113 }
114
115 static void mmap_read(struct mmap_data *md)
116 {
117         unsigned int head = mmap_read_head(md);
118         unsigned int old = md->prev;
119         unsigned char *data = md->base + page_size;
120         unsigned long size;
121         void *buf;
122         int diff;
123
124         gettimeofday(&this_read, NULL);
125
126         /*
127          * If we're further behind than half the buffer, there's a chance
128          * the writer will bite our tail and mess up the samples under us.
129          *
130          * If we somehow ended up ahead of the head, we got messed up.
131          *
132          * In either case, truncate and restart at head.
133          */
134         diff = head - old;
135         if (diff < 0) {
136                 struct timeval iv;
137                 unsigned long msecs;
138
139                 timersub(&this_read, &last_read, &iv);
140                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
141
142                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
143                                 "  Last read %lu msecs ago.\n", msecs);
144
145                 /*
146                  * head points to a known good entry, start there.
147                  */
148                 old = head;
149         }
150
151         last_read = this_read;
152
153         if (old != head)
154                 samples++;
155
156         size = head - old;
157
158         if ((old & md->mask) + size != (head & md->mask)) {
159                 buf = &data[old & md->mask];
160                 size = md->mask + 1 - (old & md->mask);
161                 old += size;
162
163                 write_output(buf, size);
164         }
165
166         buf = &data[old & md->mask];
167         size = head - old;
168         old += size;
169
170         write_output(buf, size);
171
172         md->prev = old;
173         mmap_write_tail(md, old);
174 }
175
176 static volatile int done = 0;
177 static volatile int signr = -1;
178
179 static void sig_handler(int sig)
180 {
181         done = 1;
182         signr = sig;
183 }
184
185 static void sig_atexit(void)
186 {
187         if (child_pid != -1)
188                 kill(child_pid, SIGTERM);
189
190         if (signr == -1)
191                 return;
192
193         signal(signr, SIG_DFL);
194         kill(getpid(), signr);
195 }
196
197 static pid_t pid_synthesize_comm_event(pid_t pid, int full)
198 {
199         struct comm_event comm_ev;
200         char filename[PATH_MAX];
201         char bf[BUFSIZ];
202         FILE *fp;
203         size_t size = 0;
204         DIR *tasks;
205         struct dirent dirent, *next;
206         pid_t tgid = 0;
207
208         snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
209
210         fp = fopen(filename, "r");
211         if (fp == NULL) {
212                 /*
213                  * We raced with a task exiting - just return:
214                  */
215                 if (verbose)
216                         fprintf(stderr, "couldn't open %s\n", filename);
217                 return 0;
218         }
219
220         memset(&comm_ev, 0, sizeof(comm_ev));
221         while (!comm_ev.comm[0] || !comm_ev.pid) {
222                 if (fgets(bf, sizeof(bf), fp) == NULL)
223                         goto out_failure;
224
225                 if (memcmp(bf, "Name:", 5) == 0) {
226                         char *name = bf + 5;
227                         while (*name && isspace(*name))
228                                 ++name;
229                         size = strlen(name) - 1;
230                         memcpy(comm_ev.comm, name, size++);
231                 } else if (memcmp(bf, "Tgid:", 5) == 0) {
232                         char *tgids = bf + 5;
233                         while (*tgids && isspace(*tgids))
234                                 ++tgids;
235                         tgid = comm_ev.pid = atoi(tgids);
236                 }
237         }
238
239         comm_ev.header.type = PERF_RECORD_COMM;
240         size = ALIGN(size, sizeof(u64));
241         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
242
243         if (!full) {
244                 comm_ev.tid = pid;
245
246                 write_output(&comm_ev, comm_ev.header.size);
247                 goto out_fclose;
248         }
249
250         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
251
252         tasks = opendir(filename);
253         while (!readdir_r(tasks, &dirent, &next) && next) {
254                 char *end;
255                 pid = strtol(dirent.d_name, &end, 10);
256                 if (*end)
257                         continue;
258
259                 comm_ev.tid = pid;
260
261                 write_output(&comm_ev, comm_ev.header.size);
262         }
263         closedir(tasks);
264
265 out_fclose:
266         fclose(fp);
267         return tgid;
268
269 out_failure:
270         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
271                 filename);
272         exit(EXIT_FAILURE);
273 }
274
275 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
276 {
277         char filename[PATH_MAX];
278         FILE *fp;
279
280         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
281
282         fp = fopen(filename, "r");
283         if (fp == NULL) {
284                 /*
285                  * We raced with a task exiting - just return:
286                  */
287                 if (verbose)
288                         fprintf(stderr, "couldn't open %s\n", filename);
289                 return;
290         }
291         while (1) {
292                 char bf[BUFSIZ], *pbf = bf;
293                 struct mmap_event mmap_ev = {
294                         .header = { .type = PERF_RECORD_MMAP },
295                 };
296                 int n;
297                 size_t size;
298                 if (fgets(bf, sizeof(bf), fp) == NULL)
299                         break;
300
301                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
302                 n = hex2u64(pbf, &mmap_ev.start);
303                 if (n < 0)
304                         continue;
305                 pbf += n + 1;
306                 n = hex2u64(pbf, &mmap_ev.len);
307                 if (n < 0)
308                         continue;
309                 pbf += n + 3;
310                 if (*pbf == 'x') { /* vm_exec */
311                         char *execname = strchr(bf, '/');
312
313                         /* Catch VDSO */
314                         if (execname == NULL)
315                                 execname = strstr(bf, "[vdso]");
316
317                         if (execname == NULL)
318                                 continue;
319
320                         size = strlen(execname);
321                         execname[size - 1] = '\0'; /* Remove \n */
322                         memcpy(mmap_ev.filename, execname, size);
323                         size = ALIGN(size, sizeof(u64));
324                         mmap_ev.len -= mmap_ev.start;
325                         mmap_ev.header.size = (sizeof(mmap_ev) -
326                                                (sizeof(mmap_ev.filename) - size));
327                         mmap_ev.pid = tgid;
328                         mmap_ev.tid = pid;
329
330                         write_output(&mmap_ev, mmap_ev.header.size);
331                 }
332         }
333
334         fclose(fp);
335 }
336
337 static void synthesize_all(void)
338 {
339         DIR *proc;
340         struct dirent dirent, *next;
341
342         proc = opendir("/proc");
343
344         while (!readdir_r(proc, &dirent, &next) && next) {
345                 char *end;
346                 pid_t pid, tgid;
347
348                 pid = strtol(dirent.d_name, &end, 10);
349                 if (*end) /* only interested in proper numerical dirents */
350                         continue;
351
352                 tgid = pid_synthesize_comm_event(pid, 1);
353                 pid_synthesize_mmap_samples(pid, tgid);
354         }
355
356         closedir(proc);
357 }
358
359 static int group_fd;
360
361 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
362 {
363         struct perf_header_attr *h_attr;
364
365         if (nr < header->attrs) {
366                 h_attr = header->attr[nr];
367         } else {
368                 h_attr = perf_header_attr__new(a);
369                 perf_header__add_attr(header, h_attr);
370         }
371
372         return h_attr;
373 }
374
375 static void create_counter(int counter, int cpu, pid_t pid)
376 {
377         char *filter = filters[counter];
378         struct perf_event_attr *attr = attrs + counter;
379         struct perf_header_attr *h_attr;
380         int track = !counter; /* only the first counter needs these */
381         int ret;
382         struct {
383                 u64 count;
384                 u64 time_enabled;
385                 u64 time_running;
386                 u64 id;
387         } read_data;
388
389         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
390                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
391                                   PERF_FORMAT_ID;
392
393         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
394
395         if (freq) {
396                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
397                 attr->freq              = 1;
398                 attr->sample_freq       = freq;
399         }
400
401         if (no_samples)
402                 attr->sample_freq = 0;
403
404         if (inherit_stat)
405                 attr->inherit_stat = 1;
406
407         if (sample_address)
408                 attr->sample_type       |= PERF_SAMPLE_ADDR;
409
410         if (call_graph)
411                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
412
413         if (raw_samples) {
414                 attr->sample_type       |= PERF_SAMPLE_TIME;
415                 attr->sample_type       |= PERF_SAMPLE_RAW;
416                 attr->sample_type       |= PERF_SAMPLE_CPU;
417         }
418
419         attr->mmap              = track;
420         attr->comm              = track;
421         attr->inherit           = (cpu < 0) && inherit;
422         attr->disabled          = 1;
423
424 try_again:
425         fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
426
427         if (fd[nr_cpu][counter] < 0) {
428                 int err = errno;
429
430                 if (err == EPERM)
431                         die("Permission error - are you root?\n");
432                 else if (err ==  ENODEV && profile_cpu != -1)
433                         die("No such device - did you specify an out-of-range profile CPU?\n");
434
435                 /*
436                  * If it's cycles then fall back to hrtimer
437                  * based cpu-clock-tick sw counter, which
438                  * is always available even if no PMU support:
439                  */
440                 if (attr->type == PERF_TYPE_HARDWARE
441                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
442
443                         if (verbose)
444                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
445                         attr->type = PERF_TYPE_SOFTWARE;
446                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
447                         goto try_again;
448                 }
449                 printf("\n");
450                 error("perfcounter syscall returned with %d (%s)\n",
451                         fd[nr_cpu][counter], strerror(err));
452                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
453                 exit(-1);
454         }
455
456         h_attr = get_header_attr(attr, counter);
457
458         if (!file_new) {
459                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
460                         fprintf(stderr, "incompatible append\n");
461                         exit(-1);
462                 }
463         }
464
465         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
466                 perror("Unable to read perf file descriptor\n");
467                 exit(-1);
468         }
469
470         perf_header_attr__add_id(h_attr, read_data.id);
471
472         assert(fd[nr_cpu][counter] >= 0);
473         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
474
475         /*
476          * First counter acts as the group leader:
477          */
478         if (group && group_fd == -1)
479                 group_fd = fd[nr_cpu][counter];
480         if (multiplex && multiplex_fd == -1)
481                 multiplex_fd = fd[nr_cpu][counter];
482
483         if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
484
485                 ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
486                 assert(ret != -1);
487         } else {
488                 event_array[nr_poll].fd = fd[nr_cpu][counter];
489                 event_array[nr_poll].events = POLLIN;
490                 nr_poll++;
491
492                 mmap_array[nr_cpu][counter].counter = counter;
493                 mmap_array[nr_cpu][counter].prev = 0;
494                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
495                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
496                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
497                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
498                         error("failed to mmap with %d (%s)\n", errno, strerror(errno));
499                         exit(-1);
500                 }
501         }
502
503         if (filter != NULL) {
504                 ret = ioctl(fd[nr_cpu][counter],
505                             PERF_EVENT_IOC_SET_FILTER, filter);
506                 if (ret) {
507                         error("failed to set filter with %d (%s)\n", errno,
508                               strerror(errno));
509                         exit(-1);
510                 }
511         }
512
513         ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
514 }
515
516 static void open_counters(int cpu, pid_t pid)
517 {
518         int counter;
519
520         group_fd = -1;
521         for (counter = 0; counter < nr_counters; counter++)
522                 create_counter(counter, cpu, pid);
523
524         nr_cpu++;
525 }
526
527 static void atexit_header(void)
528 {
529         header->data_size += bytes_written;
530
531         perf_header__write(header, output);
532 }
533
534 static int __cmd_record(int argc, const char **argv)
535 {
536         int i, counter;
537         struct stat st;
538         pid_t pid = 0;
539         int flags;
540         int ret;
541         unsigned long waking = 0;
542
543         page_size = sysconf(_SC_PAGE_SIZE);
544         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
545         assert(nr_cpus <= MAX_NR_CPUS);
546         assert(nr_cpus >= 0);
547
548         atexit(sig_atexit);
549         signal(SIGCHLD, sig_handler);
550         signal(SIGINT, sig_handler);
551
552         if (!stat(output_name, &st) && st.st_size) {
553                 if (!force && !append_file) {
554                         fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
555                                         output_name);
556                         exit(-1);
557                 }
558         } else {
559                 append_file = 0;
560         }
561
562         flags = O_CREAT|O_RDWR;
563         if (append_file)
564                 file_new = 0;
565         else
566                 flags |= O_TRUNC;
567
568         output = open(output_name, flags, S_IRUSR|S_IWUSR);
569         if (output < 0) {
570                 perror("failed to create output file");
571                 exit(-1);
572         }
573
574         if (!file_new)
575                 header = perf_header__read(output);
576         else
577                 header = perf_header__new();
578
579         if (raw_samples) {
580                 perf_header__set_trace_info();
581         } else {
582                 for (i = 0; i < nr_counters; i++) {
583                         if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
584                                 perf_header__set_trace_info();
585                                 break;
586                         }
587                 }
588         }
589
590         atexit(atexit_header);
591
592         if (!system_wide) {
593                 pid = target_pid;
594                 if (pid == -1)
595                         pid = getpid();
596
597                 open_counters(profile_cpu, pid);
598         } else {
599                 if (profile_cpu != -1) {
600                         open_counters(profile_cpu, target_pid);
601                 } else {
602                         for (i = 0; i < nr_cpus; i++)
603                                 open_counters(i, target_pid);
604                 }
605         }
606
607         if (file_new)
608                 perf_header__write(header, output);
609
610         if (!system_wide) {
611                 pid_t tgid = pid_synthesize_comm_event(pid, 0);
612                 pid_synthesize_mmap_samples(pid, tgid);
613         } else
614                 synthesize_all();
615
616         if (target_pid == -1 && argc) {
617                 pid = fork();
618                 if (pid < 0)
619                         perror("failed to fork");
620
621                 if (!pid) {
622                         if (execvp(argv[0], (char **)argv)) {
623                                 perror(argv[0]);
624                                 exit(-1);
625                         }
626                 }
627
628                 child_pid = pid;
629         }
630
631         if (realtime_prio) {
632                 struct sched_param param;
633
634                 param.sched_priority = realtime_prio;
635                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
636                         printf("Could not set realtime priority.\n");
637                         exit(-1);
638                 }
639         }
640
641         for (;;) {
642                 int hits = samples;
643
644                 for (i = 0; i < nr_cpu; i++) {
645                         for (counter = 0; counter < nr_counters; counter++) {
646                                 if (mmap_array[i][counter].base)
647                                         mmap_read(&mmap_array[i][counter]);
648                         }
649                 }
650
651                 if (hits == samples) {
652                         if (done)
653                                 break;
654                         ret = poll(event_array, nr_poll, -1);
655                         waking++;
656                 }
657
658                 if (done) {
659                         for (i = 0; i < nr_cpu; i++) {
660                                 for (counter = 0; counter < nr_counters; counter++)
661                                         ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
662                         }
663                 }
664         }
665
666         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
667
668         /*
669          * Approximate RIP event size: 24 bytes.
670          */
671         fprintf(stderr,
672                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
673                 (double)bytes_written / 1024.0 / 1024.0,
674                 output_name,
675                 bytes_written / 24);
676
677         return 0;
678 }
679
680 static const char * const record_usage[] = {
681         "perf record [<options>] [<command>]",
682         "perf record [<options>] -- <command> [<options>]",
683         NULL
684 };
685
686 static const struct option options[] = {
687         OPT_CALLBACK('e', "event", NULL, "event",
688                      "event selector. use 'perf list' to list available events",
689                      parse_events),
690         OPT_CALLBACK(0, "filter", NULL, "filter",
691                      "event filter", parse_filter),
692         OPT_INTEGER('p', "pid", &target_pid,
693                     "record events on existing pid"),
694         OPT_INTEGER('r', "realtime", &realtime_prio,
695                     "collect data with this RT SCHED_FIFO priority"),
696         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
697                     "collect raw sample records from all opened counters"),
698         OPT_BOOLEAN('a', "all-cpus", &system_wide,
699                             "system-wide collection from all CPUs"),
700         OPT_BOOLEAN('A', "append", &append_file,
701                             "append to the output file to do incremental profiling"),
702         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
703                             "CPU to profile on"),
704         OPT_BOOLEAN('f', "force", &force,
705                         "overwrite existing data file"),
706         OPT_LONG('c', "count", &default_interval,
707                     "event period to sample"),
708         OPT_STRING('o', "output", &output_name, "file",
709                     "output file name"),
710         OPT_BOOLEAN('i', "inherit", &inherit,
711                     "child tasks inherit counters"),
712         OPT_INTEGER('F', "freq", &freq,
713                     "profile at this frequency"),
714         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
715                     "number of mmap data pages"),
716         OPT_BOOLEAN('g', "call-graph", &call_graph,
717                     "do call-graph (stack chain/backtrace) recording"),
718         OPT_BOOLEAN('v', "verbose", &verbose,
719                     "be more verbose (show counter open errors, etc)"),
720         OPT_BOOLEAN('s', "stat", &inherit_stat,
721                     "per thread counts"),
722         OPT_BOOLEAN('d', "data", &sample_address,
723                     "Sample addresses"),
724         OPT_BOOLEAN('n', "no-samples", &no_samples,
725                     "don't sample"),
726         OPT_BOOLEAN('M', "multiplex", &multiplex,
727                     "multiplex counter output in a single channel"),
728         OPT_END()
729 };
730
731 int cmd_record(int argc, const char **argv, const char *prefix __used)
732 {
733         int counter;
734
735         argc = parse_options(argc, argv, options, record_usage,
736                 PARSE_OPT_STOP_AT_NON_OPTION);
737         if (!argc && target_pid == -1 && !system_wide)
738                 usage_with_options(record_usage, options);
739
740         if (!nr_counters) {
741                 nr_counters     = 1;
742                 attrs[0].type   = PERF_TYPE_HARDWARE;
743                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
744         }
745
746         /*
747          * User specified count overrides default frequency.
748          */
749         if (default_interval)
750                 freq = 0;
751         else if (freq) {
752                 default_interval = freq;
753         } else {
754                 fprintf(stderr, "frequency and count are zero, aborting\n");
755                 exit(EXIT_FAILURE);
756         }
757
758         for (counter = 0; counter < nr_counters; counter++) {
759                 if (attrs[counter].sample_period)
760                         continue;
761
762                 attrs[counter].sample_period = default_interval;
763         }
764
765         return __cmd_record(argc, argv);
766 }