]> git.karo-electronics.de Git - mv-sheeva.git/blob - tools/perf/builtin-record.c
perf record: Fix race where process can disappear while reading its /proc/pid/tasks
[mv-sheeva.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
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 out_race:       
210                 /*
211                  * We raced with a task exiting - just return:
212                  */
213                 if (verbose)
214                         fprintf(stderr, "couldn't open %s\n", filename);
215                 return 0;
216         }
217
218         memset(&comm_ev, 0, sizeof(comm_ev));
219         while (!comm_ev.comm[0] || !comm_ev.pid) {
220                 if (fgets(bf, sizeof(bf), fp) == NULL)
221                         goto out_failure;
222
223                 if (memcmp(bf, "Name:", 5) == 0) {
224                         char *name = bf + 5;
225                         while (*name && isspace(*name))
226                                 ++name;
227                         size = strlen(name) - 1;
228                         memcpy(comm_ev.comm, name, size++);
229                 } else if (memcmp(bf, "Tgid:", 5) == 0) {
230                         char *tgids = bf + 5;
231                         while (*tgids && isspace(*tgids))
232                                 ++tgids;
233                         tgid = comm_ev.pid = atoi(tgids);
234                 }
235         }
236
237         comm_ev.header.type = PERF_RECORD_COMM;
238         size = ALIGN(size, sizeof(u64));
239         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
240
241         if (!full) {
242                 comm_ev.tid = pid;
243
244                 write_output(&comm_ev, comm_ev.header.size);
245                 goto out_fclose;
246         }
247
248         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
249
250         tasks = opendir(filename);
251         if (tasks == NULL)
252                 goto out_race;
253
254         while (!readdir_r(tasks, &dirent, &next) && next) {
255                 char *end;
256                 pid = strtol(dirent.d_name, &end, 10);
257                 if (*end)
258                         continue;
259
260                 comm_ev.tid = pid;
261
262                 write_output(&comm_ev, comm_ev.header.size);
263         }
264         closedir(tasks);
265
266 out_fclose:
267         fclose(fp);
268         return tgid;
269
270 out_failure:
271         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
272                 filename);
273         exit(EXIT_FAILURE);
274 }
275
276 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
277 {
278         char filename[PATH_MAX];
279         FILE *fp;
280
281         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
282
283         fp = fopen(filename, "r");
284         if (fp == NULL) {
285                 /*
286                  * We raced with a task exiting - just return:
287                  */
288                 if (verbose)
289                         fprintf(stderr, "couldn't open %s\n", filename);
290                 return;
291         }
292         while (1) {
293                 char bf[BUFSIZ], *pbf = bf;
294                 struct mmap_event mmap_ev = {
295                         .header = { .type = PERF_RECORD_MMAP },
296                 };
297                 int n;
298                 size_t size;
299                 if (fgets(bf, sizeof(bf), fp) == NULL)
300                         break;
301
302                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
303                 n = hex2u64(pbf, &mmap_ev.start);
304                 if (n < 0)
305                         continue;
306                 pbf += n + 1;
307                 n = hex2u64(pbf, &mmap_ev.len);
308                 if (n < 0)
309                         continue;
310                 pbf += n + 3;
311                 if (*pbf == 'x') { /* vm_exec */
312                         char *execname = strchr(bf, '/');
313
314                         /* Catch VDSO */
315                         if (execname == NULL)
316                                 execname = strstr(bf, "[vdso]");
317
318                         if (execname == NULL)
319                                 continue;
320
321                         size = strlen(execname);
322                         execname[size - 1] = '\0'; /* Remove \n */
323                         memcpy(mmap_ev.filename, execname, size);
324                         size = ALIGN(size, sizeof(u64));
325                         mmap_ev.len -= mmap_ev.start;
326                         mmap_ev.header.size = (sizeof(mmap_ev) -
327                                                (sizeof(mmap_ev.filename) - size));
328                         mmap_ev.pid = tgid;
329                         mmap_ev.tid = pid;
330
331                         write_output(&mmap_ev, mmap_ev.header.size);
332                 }
333         }
334
335         fclose(fp);
336 }
337
338 static void synthesize_all(void)
339 {
340         DIR *proc;
341         struct dirent dirent, *next;
342
343         proc = opendir("/proc");
344
345         while (!readdir_r(proc, &dirent, &next) && next) {
346                 char *end;
347                 pid_t pid, tgid;
348
349                 pid = strtol(dirent.d_name, &end, 10);
350                 if (*end) /* only interested in proper numerical dirents */
351                         continue;
352
353                 tgid = pid_synthesize_comm_event(pid, 1);
354                 pid_synthesize_mmap_samples(pid, tgid);
355         }
356
357         closedir(proc);
358 }
359
360 static int group_fd;
361
362 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
363 {
364         struct perf_header_attr *h_attr;
365
366         if (nr < header->attrs) {
367                 h_attr = header->attr[nr];
368         } else {
369                 h_attr = perf_header_attr__new(a);
370                 perf_header__add_attr(header, h_attr);
371         }
372
373         return h_attr;
374 }
375
376 static void create_counter(int counter, int cpu, pid_t pid)
377 {
378         char *filter = filters[counter];
379         struct perf_event_attr *attr = attrs + counter;
380         struct perf_header_attr *h_attr;
381         int track = !counter; /* only the first counter needs these */
382         int ret;
383         struct {
384                 u64 count;
385                 u64 time_enabled;
386                 u64 time_running;
387                 u64 id;
388         } read_data;
389
390         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
391                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
392                                   PERF_FORMAT_ID;
393
394         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
395
396         if (freq) {
397                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
398                 attr->freq              = 1;
399                 attr->sample_freq       = freq;
400         }
401
402         if (no_samples)
403                 attr->sample_freq = 0;
404
405         if (inherit_stat)
406                 attr->inherit_stat = 1;
407
408         if (sample_address)
409                 attr->sample_type       |= PERF_SAMPLE_ADDR;
410
411         if (call_graph)
412                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
413
414         if (raw_samples) {
415                 attr->sample_type       |= PERF_SAMPLE_TIME;
416                 attr->sample_type       |= PERF_SAMPLE_RAW;
417                 attr->sample_type       |= PERF_SAMPLE_CPU;
418         }
419
420         attr->mmap              = track;
421         attr->comm              = track;
422         attr->inherit           = (cpu < 0) && inherit;
423         attr->disabled          = 1;
424
425 try_again:
426         fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
427
428         if (fd[nr_cpu][counter] < 0) {
429                 int err = errno;
430
431                 if (err == EPERM)
432                         die("Permission error - are you root?\n");
433                 else if (err ==  ENODEV && profile_cpu != -1)
434                         die("No such device - did you specify an out-of-range profile CPU?\n");
435
436                 /*
437                  * If it's cycles then fall back to hrtimer
438                  * based cpu-clock-tick sw counter, which
439                  * is always available even if no PMU support:
440                  */
441                 if (attr->type == PERF_TYPE_HARDWARE
442                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
443
444                         if (verbose)
445                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
446                         attr->type = PERF_TYPE_SOFTWARE;
447                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
448                         goto try_again;
449                 }
450                 printf("\n");
451                 error("perfcounter syscall returned with %d (%s)\n",
452                         fd[nr_cpu][counter], strerror(err));
453                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
454                 exit(-1);
455         }
456
457         h_attr = get_header_attr(attr, counter);
458
459         if (!file_new) {
460                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
461                         fprintf(stderr, "incompatible append\n");
462                         exit(-1);
463                 }
464         }
465
466         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
467                 perror("Unable to read perf file descriptor\n");
468                 exit(-1);
469         }
470
471         perf_header_attr__add_id(h_attr, read_data.id);
472
473         assert(fd[nr_cpu][counter] >= 0);
474         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
475
476         /*
477          * First counter acts as the group leader:
478          */
479         if (group && group_fd == -1)
480                 group_fd = fd[nr_cpu][counter];
481         if (multiplex && multiplex_fd == -1)
482                 multiplex_fd = fd[nr_cpu][counter];
483
484         if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
485
486                 ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
487                 assert(ret != -1);
488         } else {
489                 event_array[nr_poll].fd = fd[nr_cpu][counter];
490                 event_array[nr_poll].events = POLLIN;
491                 nr_poll++;
492
493                 mmap_array[nr_cpu][counter].counter = counter;
494                 mmap_array[nr_cpu][counter].prev = 0;
495                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
496                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
497                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
498                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
499                         error("failed to mmap with %d (%s)\n", errno, strerror(errno));
500                         exit(-1);
501                 }
502         }
503
504         if (filter != NULL) {
505                 ret = ioctl(fd[nr_cpu][counter],
506                             PERF_EVENT_IOC_SET_FILTER, filter);
507                 if (ret) {
508                         error("failed to set filter with %d (%s)\n", errno,
509                               strerror(errno));
510                         exit(-1);
511                 }
512         }
513
514         ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
515 }
516
517 static void open_counters(int cpu, pid_t pid)
518 {
519         int counter;
520
521         group_fd = -1;
522         for (counter = 0; counter < nr_counters; counter++)
523                 create_counter(counter, cpu, pid);
524
525         nr_cpu++;
526 }
527
528 static void atexit_header(void)
529 {
530         header->data_size += bytes_written;
531
532         perf_header__write(header, output);
533 }
534
535 static int __cmd_record(int argc, const char **argv)
536 {
537         int i, counter;
538         struct stat st;
539         pid_t pid = 0;
540         int flags;
541         int ret;
542         unsigned long waking = 0;
543
544         page_size = sysconf(_SC_PAGE_SIZE);
545         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
546         assert(nr_cpus <= MAX_NR_CPUS);
547         assert(nr_cpus >= 0);
548
549         atexit(sig_atexit);
550         signal(SIGCHLD, sig_handler);
551         signal(SIGINT, sig_handler);
552
553         if (!stat(output_name, &st) && st.st_size) {
554                 if (!force && !append_file) {
555                         fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
556                                         output_name);
557                         exit(-1);
558                 }
559         } else {
560                 append_file = 0;
561         }
562
563         flags = O_CREAT|O_RDWR;
564         if (append_file)
565                 file_new = 0;
566         else
567                 flags |= O_TRUNC;
568
569         output = open(output_name, flags, S_IRUSR|S_IWUSR);
570         if (output < 0) {
571                 perror("failed to create output file");
572                 exit(-1);
573         }
574
575         if (!file_new)
576                 header = perf_header__read(output);
577         else
578                 header = perf_header__new();
579
580         if (raw_samples) {
581                 perf_header__feat_trace_info(header);
582         } else {
583                 for (i = 0; i < nr_counters; i++) {
584                         if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
585                                 perf_header__feat_trace_info(header);
586                                 break;
587                         }
588                 }
589         }
590
591         atexit(atexit_header);
592
593         if (!system_wide) {
594                 pid = target_pid;
595                 if (pid == -1)
596                         pid = getpid();
597
598                 open_counters(profile_cpu, pid);
599         } else {
600                 if (profile_cpu != -1) {
601                         open_counters(profile_cpu, target_pid);
602                 } else {
603                         for (i = 0; i < nr_cpus; i++)
604                                 open_counters(i, target_pid);
605                 }
606         }
607
608         if (file_new)
609                 perf_header__write(header, output);
610
611         if (!system_wide) {
612                 pid_t tgid = pid_synthesize_comm_event(pid, 0);
613                 pid_synthesize_mmap_samples(pid, tgid);
614         } else
615                 synthesize_all();
616
617         if (target_pid == -1 && argc) {
618                 pid = fork();
619                 if (pid < 0)
620                         perror("failed to fork");
621
622                 if (!pid) {
623                         if (execvp(argv[0], (char **)argv)) {
624                                 perror(argv[0]);
625                                 exit(-1);
626                         }
627                 }
628
629                 child_pid = pid;
630         }
631
632         if (realtime_prio) {
633                 struct sched_param param;
634
635                 param.sched_priority = realtime_prio;
636                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
637                         pr_err("Could not set realtime priority.\n");
638                         exit(-1);
639                 }
640         }
641
642         for (;;) {
643                 int hits = samples;
644
645                 for (i = 0; i < nr_cpu; i++) {
646                         for (counter = 0; counter < nr_counters; counter++) {
647                                 if (mmap_array[i][counter].base)
648                                         mmap_read(&mmap_array[i][counter]);
649                         }
650                 }
651
652                 if (hits == samples) {
653                         if (done)
654                                 break;
655                         ret = poll(event_array, nr_poll, -1);
656                         waking++;
657                 }
658
659                 if (done) {
660                         for (i = 0; i < nr_cpu; i++) {
661                                 for (counter = 0; counter < nr_counters; counter++)
662                                         ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
663                         }
664                 }
665         }
666
667         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
668
669         /*
670          * Approximate RIP event size: 24 bytes.
671          */
672         fprintf(stderr,
673                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
674                 (double)bytes_written / 1024.0 / 1024.0,
675                 output_name,
676                 bytes_written / 24);
677
678         return 0;
679 }
680
681 static const char * const record_usage[] = {
682         "perf record [<options>] [<command>]",
683         "perf record [<options>] -- <command> [<options>]",
684         NULL
685 };
686
687 static const struct option options[] = {
688         OPT_CALLBACK('e', "event", NULL, "event",
689                      "event selector. use 'perf list' to list available events",
690                      parse_events),
691         OPT_CALLBACK(0, "filter", NULL, "filter",
692                      "event filter", parse_filter),
693         OPT_INTEGER('p', "pid", &target_pid,
694                     "record events on existing pid"),
695         OPT_INTEGER('r', "realtime", &realtime_prio,
696                     "collect data with this RT SCHED_FIFO priority"),
697         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
698                     "collect raw sample records from all opened counters"),
699         OPT_BOOLEAN('a', "all-cpus", &system_wide,
700                             "system-wide collection from all CPUs"),
701         OPT_BOOLEAN('A', "append", &append_file,
702                             "append to the output file to do incremental profiling"),
703         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
704                             "CPU to profile on"),
705         OPT_BOOLEAN('f', "force", &force,
706                         "overwrite existing data file"),
707         OPT_LONG('c', "count", &default_interval,
708                     "event period to sample"),
709         OPT_STRING('o', "output", &output_name, "file",
710                     "output file name"),
711         OPT_BOOLEAN('i', "inherit", &inherit,
712                     "child tasks inherit counters"),
713         OPT_INTEGER('F', "freq", &freq,
714                     "profile at this frequency"),
715         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
716                     "number of mmap data pages"),
717         OPT_BOOLEAN('g', "call-graph", &call_graph,
718                     "do call-graph (stack chain/backtrace) recording"),
719         OPT_BOOLEAN('v', "verbose", &verbose,
720                     "be more verbose (show counter open errors, etc)"),
721         OPT_BOOLEAN('s', "stat", &inherit_stat,
722                     "per thread counts"),
723         OPT_BOOLEAN('d', "data", &sample_address,
724                     "Sample addresses"),
725         OPT_BOOLEAN('n', "no-samples", &no_samples,
726                     "don't sample"),
727         OPT_BOOLEAN('M', "multiplex", &multiplex,
728                     "multiplex counter output in a single channel"),
729         OPT_END()
730 };
731
732 int cmd_record(int argc, const char **argv, const char *prefix __used)
733 {
734         int counter;
735
736         argc = parse_options(argc, argv, options, record_usage,
737                 PARSE_OPT_STOP_AT_NON_OPTION);
738         if (!argc && target_pid == -1 && !system_wide)
739                 usage_with_options(record_usage, options);
740
741         if (!nr_counters) {
742                 nr_counters     = 1;
743                 attrs[0].type   = PERF_TYPE_HARDWARE;
744                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
745         }
746
747         /*
748          * User specified count overrides default frequency.
749          */
750         if (default_interval)
751                 freq = 0;
752         else if (freq) {
753                 default_interval = freq;
754         } else {
755                 fprintf(stderr, "frequency and count are zero, aborting\n");
756                 exit(EXIT_FAILURE);
757         }
758
759         for (counter = 0; counter < nr_counters; counter++) {
760                 if (attrs[counter].sample_period)
761                         continue;
762
763                 attrs[counter].sample_period = default_interval;
764         }
765
766         return __cmd_record(argc, argv);
767 }