git.oblomov.eu Git - linux-2.6/blob - 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 int                      profile_cpu                     = -1;
  39 static pid_t                    target_pid                      = -1;
  40 static int                      inherit                         = 1;
  41 static int                      force                           = 0;
  42 static int                      append_file                     = 0;
  43 static int                      call_graph                      = 0;
  44 static int                      verbose                         = 0;
  45 static int                      inherit_stat                    = 0;
  46 static int                      no_samples                      = 0;
  47 static int                      sample_address                  = 0;
  48
  49 static long                     samples;
  50 static struct timeval           last_read;
  51 static struct timeval           this_read;
  52
  53 static u64                      bytes_written;
  54
  55 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
  56
  57 static int                      nr_poll;
  58 static int                      nr_cpu;
  59
  60 static int                      file_new = 1;
  61
  62 struct perf_header              *header;
  63
  64 struct mmap_event {
  65         struct perf_event_header        header;
  66         u32                             pid;
  67         u32                             tid;
  68         u64                             start;
  69         u64                             len;
  70         u64                             pgoff;
  71         char                            filename[PATH_MAX];
  72 };
  73
  74 struct comm_event {
  75         struct perf_event_header        header;
  76         u32                             pid;
  77         u32                             tid;
  78         char                            comm[16];
  79 };
  80
  81
  82 struct mmap_data {
  83         int                     counter;
  84         void                    *base;
  85         unsigned int            mask;
  86         unsigned int            prev;
  87 };
  88
  89 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  90
  91 static unsigned long mmap_read_head(struct mmap_data *md)
  92 {
  93         struct perf_counter_mmap_page *pc = md->base;
  94         long head;
  95
  96         head = pc->data_head;
  97         rmb();
  98
  99         return head;
 100 }
 101
 102 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
 103 {
 104         struct perf_counter_mmap_page *pc = md->base;
 105
 106         /*
 107          * ensure all reads are done before we write the tail out.
 108          */
 109         /* mb(); */
 110         pc->data_tail = tail;
 111 }
 112
 113 static void write_output(void *buf, size_t size)
 114 {
 115         while (size) {
 116                 int ret = write(output, buf, size);
 117
 118                 if (ret < 0)
 119                         die("failed to write");
 120
 121                 size -= ret;
 122                 buf += ret;
 123
 124                 bytes_written += ret;
 125         }
 126 }
 127
 128 static void mmap_read(struct mmap_data *md)
 129 {
 130         unsigned int head = mmap_read_head(md);
 131         unsigned int old = md->prev;
 132         unsigned char *data = md->base + page_size;
 133         unsigned long size;
 134         void *buf;
 135         int diff;
 136
 137         gettimeofday(&this_read, NULL);
 138
 139         /*
 140          * If we're further behind than half the buffer, there's a chance
 141          * the writer will bite our tail and mess up the samples under us.
 142          *
 143          * If we somehow ended up ahead of the head, we got messed up.
 144          *
 145          * In either case, truncate and restart at head.
 146          */
 147         diff = head - old;
 148         if (diff < 0) {
 149                 struct timeval iv;
 150                 unsigned long msecs;
 151
 152                 timersub(&this_read, &last_read, &iv);
 153                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 154
 155                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 156                                 "  Last read %lu msecs ago.\n", msecs);
 157
 158                 /*
 159                  * head points to a known good entry, start there.
 160                  */
 161                 old = head;
 162         }
 163
 164         last_read = this_read;
 165
 166         if (old != head)
 167                 samples++;
 168
 169         size = head - old;
 170
 171         if ((old & md->mask) + size != (head & md->mask)) {
 172                 buf = &data[old & md->mask];
 173                 size = md->mask + 1 - (old & md->mask);
 174                 old += size;
 175
 176                 write_output(buf, size);
 177         }
 178
 179         buf = &data[old & md->mask];
 180         size = head - old;
 181         old += size;
 182
 183         write_output(buf, size);
 184
 185         md->prev = old;
 186         mmap_write_tail(md, old);
 187 }
 188
 189 static volatile int done = 0;
 190 static volatile int signr = -1;
 191
 192 static void sig_handler(int sig)
 193 {
 194         done = 1;
 195         signr = sig;
 196 }
 197
 198 static void sig_atexit(void)
 199 {
 200         if (signr == -1)
 201                 return;
 202
 203         signal(signr, SIG_DFL);
 204         kill(getpid(), signr);
 205 }
 206
 207 static pid_t pid_synthesize_comm_event(pid_t pid, int full)
 208 {
 209         struct comm_event comm_ev;
 210         char filename[PATH_MAX];
 211         char bf[BUFSIZ];
 212         FILE *fp;
 213         size_t size = 0;
 214         DIR *tasks;
 215         struct dirent dirent, *next;
 216         pid_t tgid = 0;
 217
 218         snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
 219
 220         fp = fopen(filename, "r");
 221         if (fd == NULL) {
 222                 /*
 223                  * We raced with a task exiting - just return:
 224                  */
 225                 if (verbose)
 226                         fprintf(stderr, "couldn't open %s\n", filename);
 227                 return 0;
 228         }
 229
 230         memset(&comm_ev, 0, sizeof(comm_ev));
 231         while (!comm_ev.comm[0] || !comm_ev.pid) {
 232                 if (fgets(bf, sizeof(bf), fp) == NULL)
 233                         goto out_failure;
 234
 235                 if (memcmp(bf, "Name:", 5) == 0) {
 236                         char *name = bf + 5;
 237                         while (*name && isspace(*name))
 238                                 ++name;
 239                         size = strlen(name) - 1;
 240                         memcpy(comm_ev.comm, name, size++);
 241                 } else if (memcmp(bf, "Tgid:", 5) == 0) {
 242                         char *tgids = bf + 5;
 243                         while (*tgids && isspace(*tgids))
 244                                 ++tgids;
 245                         tgid = comm_ev.pid = atoi(tgids);
 246                 }
 247         }
 248
 249         comm_ev.header.type = PERF_EVENT_COMM;
 250         size = ALIGN(size, sizeof(u64));
 251         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 252
 253         if (!full) {
 254                 comm_ev.tid = pid;
 255
 256                 write_output(&comm_ev, comm_ev.header.size);
 257                 goto out_fclose;
 258         }
 259
 260         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 261
 262         tasks = opendir(filename);
 263         while (!readdir_r(tasks, &dirent, &next) && next) {
 264                 char *end;
 265                 pid = strtol(dirent.d_name, &end, 10);
 266                 if (*end)
 267                         continue;
 268
 269                 comm_ev.tid = pid;
 270
 271                 write_output(&comm_ev, comm_ev.header.size);
 272         }
 273         closedir(tasks);
 274
 275 out_fclose:
 276         fclose(fp);
 277         return tgid;
 278
 279 out_failure:
 280         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 281                 filename);
 282         exit(EXIT_FAILURE);
 283 }
 284
 285 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
 286 {
 287         char filename[PATH_MAX];
 288         FILE *fp;
 289
 290         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 291
 292         fp = fopen(filename, "r");
 293         if (fp == NULL) {
 294                 /*
 295                  * We raced with a task exiting - just return:
 296                  */
 297                 if (verbose)
 298                         fprintf(stderr, "couldn't open %s\n", filename);
 299                 return;
 300         }
 301         while (1) {
 302                 char bf[BUFSIZ], *pbf = bf;
 303                 struct mmap_event mmap_ev = {
 304                         .header = { .type = PERF_EVENT_MMAP },
 305                 };
 306                 int n;
 307                 size_t size;
 308                 if (fgets(bf, sizeof(bf), fp) == NULL)
 309                         break;
 310
 311                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 312                 n = hex2u64(pbf, &mmap_ev.start);
 313                 if (n < 0)
 314                         continue;
 315                 pbf += n + 1;
 316                 n = hex2u64(pbf, &mmap_ev.len);
 317                 if (n < 0)
 318                         continue;
 319                 pbf += n + 3;
 320                 if (*pbf == 'x') { /* vm_exec */
 321                         char *execname = strchr(bf, '/');
 322
 323                         /* Catch VDSO */
 324                         if (execname == NULL)
 325                                 execname = strstr(bf, "[vdso]");
 326
 327                         if (execname == NULL)
 328                                 continue;
 329
 330                         size = strlen(execname);
 331                         execname[size - 1] = '\0'; /* Remove \n */
 332                         memcpy(mmap_ev.filename, execname, size);
 333                         size = ALIGN(size, sizeof(u64));
 334                         mmap_ev.len -= mmap_ev.start;
 335                         mmap_ev.header.size = (sizeof(mmap_ev) -
 336                                                (sizeof(mmap_ev.filename) - size));
 337                         mmap_ev.pid = tgid;
 338                         mmap_ev.tid = pid;
 339
 340                         write_output(&mmap_ev, mmap_ev.header.size);
 341                 }
 342         }
 343
 344         fclose(fp);
 345 }
 346
 347 static void synthesize_all(void)
 348 {
 349         DIR *proc;
 350         struct dirent dirent, *next;
 351
 352         proc = opendir("/proc");
 353
 354         while (!readdir_r(proc, &dirent, &next) && next) {
 355                 char *end;
 356                 pid_t pid, tgid;
 357
 358                 pid = strtol(dirent.d_name, &end, 10);
 359                 if (*end) /* only interested in proper numerical dirents */
 360                         continue;
 361
 362                 tgid = pid_synthesize_comm_event(pid, 1);
 363                 pid_synthesize_mmap_samples(pid, tgid);
 364         }
 365
 366         closedir(proc);
 367 }
 368
 369 static int group_fd;
 370
 371 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
 372 {
 373         struct perf_header_attr *h_attr;
 374
 375         if (nr < header->attrs) {
 376                 h_attr = header->attr[nr];
 377         } else {
 378                 h_attr = perf_header_attr__new(a);
 379                 perf_header__add_attr(header, h_attr);
 380         }
 381
 382         return h_attr;
 383 }
 384
 385 static void create_counter(int counter, int cpu, pid_t pid)
 386 {
 387         struct perf_counter_attr *attr = attrs + counter;
 388         struct perf_header_attr *h_attr;
 389         int track = !counter; /* only the first counter needs these */
 390         struct {
 391                 u64 count;
 392                 u64 time_enabled;
 393                 u64 time_running;
 394                 u64 id;
 395         } read_data;
 396
 397         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 398                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 399                                   PERF_FORMAT_ID;
 400
 401         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 402
 403         if (freq) {
 404                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 405                 attr->freq              = 1;
 406                 attr->sample_freq       = freq;
 407         }
 408
 409         if (no_samples)
 410                 attr->sample_freq = 0;
 411
 412         if (inherit_stat)
 413                 attr->inherit_stat = 1;
 414
 415         if (sample_address)
 416                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 417
 418         if (call_graph)
 419                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 420
 421
 422         attr->mmap              = track;
 423         attr->comm              = track;
 424         attr->inherit           = (cpu < 0) && inherit;
 425         attr->disabled          = 1;
 426
 427 try_again:
 428         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 429
 430         if (fd[nr_cpu][counter] < 0) {
 431                 int err = errno;
 432
 433                 if (err == EPERM)
 434                         die("Permission error - are you root?\n");
 435                 else if (err ==  ENODEV && profile_cpu != -1)
 436                         die("No such device - did you specify an out-of-range profile CPU?\n");
 437
 438                 /*
 439                  * If it's cycles then fall back to hrtimer
 440                  * based cpu-clock-tick sw counter, which
 441                  * is always available even if no PMU support:
 442                  */
 443                 if (attr->type == PERF_TYPE_HARDWARE
 444                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 445
 446                         if (verbose)
 447                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 448                         attr->type = PERF_TYPE_SOFTWARE;
 449                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 450                         goto try_again;
 451                 }
 452                 printf("\n");
 453                 error("perfcounter syscall returned with %d (%s)\n",
 454                         fd[nr_cpu][counter], strerror(err));
 455                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 456                 exit(-1);
 457         }
 458
 459         h_attr = get_header_attr(attr, counter);
 460
 461         if (!file_new) {
 462                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 463                         fprintf(stderr, "incompatible append\n");
 464                         exit(-1);
 465                 }
 466         }
 467
 468         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
 469                 perror("Unable to read perf file descriptor\n");
 470                 exit(-1);
 471         }
 472
 473         perf_header_attr__add_id(h_attr, read_data.id);
 474
 475         assert(fd[nr_cpu][counter] >= 0);
 476         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 477
 478         /*
 479          * First counter acts as the group leader:
 480          */
 481         if (group && group_fd == -1)
 482                 group_fd = fd[nr_cpu][counter];
 483
 484         event_array[nr_poll].fd = fd[nr_cpu][counter];
 485         event_array[nr_poll].events = POLLIN;
 486         nr_poll++;
 487
 488         mmap_array[nr_cpu][counter].counter = counter;
 489         mmap_array[nr_cpu][counter].prev = 0;
 490         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 491         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 492                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
 493         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 494                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 495                 exit(-1);
 496         }
 497
 498         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
 499 }
 500
 501 static void open_counters(int cpu, pid_t pid)
 502 {
 503         int counter;
 504
 505         group_fd = -1;
 506         for (counter = 0; counter < nr_counters; counter++)
 507                 create_counter(counter, cpu, pid);
 508
 509         nr_cpu++;
 510 }
 511
 512 static void atexit_header(void)
 513 {
 514         header->data_size += bytes_written;
 515
 516         perf_header__write(header, output);
 517 }
 518
 519 static int __cmd_record(int argc, const char **argv)
 520 {
 521         int i, counter;
 522         struct stat st;
 523         pid_t pid = 0;
 524         int flags;
 525         int ret;
 526
 527         page_size = sysconf(_SC_PAGE_SIZE);
 528         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 529         assert(nr_cpus <= MAX_NR_CPUS);
 530         assert(nr_cpus >= 0);
 531
 532         atexit(sig_atexit);
 533         signal(SIGCHLD, sig_handler);
 534         signal(SIGINT, sig_handler);
 535
 536         if (!stat(output_name, &st) && st.st_size) {
 537                 if (!force && !append_file) {
 538                         fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 539                                         output_name);
 540                         exit(-1);
 541                 }
 542         } else {
 543                 append_file = 0;
 544         }
 545
 546         flags = O_CREAT|O_RDWR;
 547         if (append_file)
 548                 file_new = 0;
 549         else
 550                 flags |= O_TRUNC;
 551
 552         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 553         if (output < 0) {
 554                 perror("failed to create output file");
 555                 exit(-1);
 556         }
 557
 558         if (!file_new)
 559                 header = perf_header__read(output);
 560         else
 561                 header = perf_header__new();
 562
 563         atexit(atexit_header);
 564
 565         if (!system_wide) {
 566                 pid = target_pid;
 567                 if (pid == -1)
 568                         pid = getpid();
 569
 570                 open_counters(profile_cpu, pid);
 571         } else {
 572                 if (profile_cpu != -1) {
 573                         open_counters(profile_cpu, target_pid);
 574                 } else {
 575                         for (i = 0; i < nr_cpus; i++)
 576                                 open_counters(i, target_pid);
 577                 }
 578         }
 579
 580         if (file_new)
 581                 perf_header__write(header, output);
 582
 583         if (!system_wide) {
 584                 pid_t tgid = pid_synthesize_comm_event(pid, 0);
 585                 pid_synthesize_mmap_samples(pid, tgid);
 586         } else
 587                 synthesize_all();
 588
 589         if (target_pid == -1 && argc) {
 590                 pid = fork();
 591                 if (pid < 0)
 592                         perror("failed to fork");
 593
 594                 if (!pid) {
 595                         if (execvp(argv[0], (char **)argv)) {
 596                                 perror(argv[0]);
 597                                 exit(-1);
 598                         }
 599                 }
 600         }
 601
 602         if (realtime_prio) {
 603                 struct sched_param param;
 604
 605                 param.sched_priority = realtime_prio;
 606                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 607                         printf("Could not set realtime priority.\n");
 608                         exit(-1);
 609                 }
 610         }
 611
 612         for (;;) {
 613                 int hits = samples;
 614
 615                 for (i = 0; i < nr_cpu; i++) {
 616                         for (counter = 0; counter < nr_counters; counter++)
 617                                 mmap_read(&mmap_array[i][counter]);
 618                 }
 619
 620                 if (hits == samples) {
 621                         if (done)
 622                                 break;
 623                         ret = poll(event_array, nr_poll, 100);
 624                 }
 625         }
 626
 627         /*
 628          * Approximate RIP event size: 24 bytes.
 629          */
 630         fprintf(stderr,
 631                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 632                 (double)bytes_written / 1024.0 / 1024.0,
 633                 output_name,
 634                 bytes_written / 24);
 635
 636         return 0;
 637 }
 638
 639 static const char * const record_usage[] = {
 640         "perf record [<options>] [<command>]",
 641         "perf record [<options>] -- <command> [<options>]",
 642         NULL
 643 };
 644
 645 static const struct option options[] = {
 646         OPT_CALLBACK('e', "event", NULL, "event",
 647                      "event selector. use 'perf list' to list available events",
 648                      parse_events),
 649         OPT_INTEGER('p', "pid", &target_pid,
 650                     "record events on existing pid"),
 651         OPT_INTEGER('r', "realtime", &realtime_prio,
 652                     "collect data with this RT SCHED_FIFO priority"),
 653         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 654                             "system-wide collection from all CPUs"),
 655         OPT_BOOLEAN('A', "append", &append_file,
 656                             "append to the output file to do incremental profiling"),
 657         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
 658                             "CPU to profile on"),
 659         OPT_BOOLEAN('f', "force", &force,
 660                         "overwrite existing data file"),
 661         OPT_LONG('c', "count", &default_interval,
 662                     "event period to sample"),
 663         OPT_STRING('o', "output", &output_name, "file",
 664                     "output file name"),
 665         OPT_BOOLEAN('i', "inherit", &inherit,
 666                     "child tasks inherit counters"),
 667         OPT_INTEGER('F', "freq", &freq,
 668                     "profile at this frequency"),
 669         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 670                     "number of mmap data pages"),
 671         OPT_BOOLEAN('g', "call-graph", &call_graph,
 672                     "do call-graph (stack chain/backtrace) recording"),
 673         OPT_BOOLEAN('v', "verbose", &verbose,
 674                     "be more verbose (show counter open errors, etc)"),
 675         OPT_BOOLEAN('s', "stat", &inherit_stat,
 676                     "per thread counts"),
 677         OPT_BOOLEAN('d', "data", &sample_address,
 678                     "Sample addresses"),
 679         OPT_BOOLEAN('n', "no-samples", &no_samples,
 680                     "don't sample"),
 681         OPT_END()
 682 };
 683
 684 int cmd_record(int argc, const char **argv, const char *prefix __used)
 685 {
 686         int counter;
 687
 688         argc = parse_options(argc, argv, options, record_usage,
 689                 PARSE_OPT_STOP_AT_NON_OPTION);
 690         if (!argc && target_pid == -1 && !system_wide)
 691                 usage_with_options(record_usage, options);
 692
 693         if (!nr_counters) {
 694                 nr_counters     = 1;
 695                 attrs[0].type   = PERF_TYPE_HARDWARE;
 696                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 697         }
 698
 699         for (counter = 0; counter < nr_counters; counter++) {
 700                 if (attrs[counter].sample_period)
 701                         continue;
 702
 703                 attrs[counter].sample_period = default_interval;
 704         }
 705
 706         return __cmd_record(argc, argv);
 707 }