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