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