SAFE public projects git trees. - safe/jmp/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 #include "util/event.h"
  19
  20 #include <unistd.h>
  21 #include <sched.h>
  22
  23 #define ALIGN(x, a)             __ALIGN_MASK(x, (typeof(x))(a)-1)
  24 #define __ALIGN_MASK(x, mask)   (((x)+(mask))&~(mask))
  25
  26 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  27
  28 static long                     default_interval                = 100000;
  29
  30 static int                      nr_cpus                         = 0;
  31 static unsigned int             page_size;
  32 static unsigned int             mmap_pages                      = 128;
  33 static int                      freq                            = 0;
  34 static int                      output;
  35 static const char               *output_name                    = "perf.data";
  36 static int                      group                           = 0;
  37 static unsigned int             realtime_prio                   = 0;
  38 static int                      system_wide                     = 0;
  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                      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_data {
  64         int                     counter;
  65         void                    *base;
  66         unsigned int            mask;
  67         unsigned int            prev;
  68 };
  69
  70 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  71
  72 static unsigned long mmap_read_head(struct mmap_data *md)
  73 {
  74         struct perf_counter_mmap_page *pc = md->base;
  75         long head;
  76
  77         head = pc->data_head;
  78         rmb();
  79
  80         return head;
  81 }
  82
  83 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
  84 {
  85         struct perf_counter_mmap_page *pc = md->base;
  86
  87         /*
  88          * ensure all reads are done before we write the tail out.
  89          */
  90         /* mb(); */
  91         pc->data_tail = tail;
  92 }
  93
  94 static void write_output(void *buf, size_t size)
  95 {
  96         while (size) {
  97                 int ret = write(output, buf, size);
  98
  99                 if (ret < 0)
 100                         die("failed to write");
 101
 102                 size -= ret;
 103                 buf += ret;
 104
 105                 bytes_written += ret;
 106         }
 107 }
 108
 109 static void mmap_read(struct mmap_data *md)
 110 {
 111         unsigned int head = mmap_read_head(md);
 112         unsigned int old = md->prev;
 113         unsigned char *data = md->base + page_size;
 114         unsigned long size;
 115         void *buf;
 116         int diff;
 117
 118         gettimeofday(&this_read, NULL);
 119
 120         /*
 121          * If we're further behind than half the buffer, there's a chance
 122          * the writer will bite our tail and mess up the samples under us.
 123          *
 124          * If we somehow ended up ahead of the head, we got messed up.
 125          *
 126          * In either case, truncate and restart at head.
 127          */
 128         diff = head - old;
 129         if (diff < 0) {
 130                 struct timeval iv;
 131                 unsigned long msecs;
 132
 133                 timersub(&this_read, &last_read, &iv);
 134                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 135
 136                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 137                                 "  Last read %lu msecs ago.\n", msecs);
 138
 139                 /*
 140                  * head points to a known good entry, start there.
 141                  */
 142                 old = head;
 143         }
 144
 145         last_read = this_read;
 146
 147         if (old != head)
 148                 samples++;
 149
 150         size = head - old;
 151
 152         if ((old & md->mask) + size != (head & md->mask)) {
 153                 buf = &data[old & md->mask];
 154                 size = md->mask + 1 - (old & md->mask);
 155                 old += size;
 156
 157                 write_output(buf, size);
 158         }
 159
 160         buf = &data[old & md->mask];
 161         size = head - old;
 162         old += size;
 163
 164         write_output(buf, size);
 165
 166         md->prev = old;
 167         mmap_write_tail(md, old);
 168 }
 169
 170 static volatile int done = 0;
 171 static volatile int signr = -1;
 172
 173 static void sig_handler(int sig)
 174 {
 175         done = 1;
 176         signr = sig;
 177 }
 178
 179 static void sig_atexit(void)
 180 {
 181         if (signr == -1)
 182                 return;
 183
 184         signal(signr, SIG_DFL);
 185         kill(getpid(), signr);
 186 }
 187
 188 static void pid_synthesize_comm_event(pid_t pid, int full)
 189 {
 190         struct comm_event comm_ev;
 191         char filename[PATH_MAX];
 192         char bf[BUFSIZ];
 193         int fd;
 194         size_t size;
 195         char *field, *sep;
 196         DIR *tasks;
 197         struct dirent dirent, *next;
 198
 199         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 200
 201         fd = open(filename, O_RDONLY);
 202         if (fd < 0) {
 203                 /*
 204                  * We raced with a task exiting - just return:
 205                  */
 206                 if (verbose)
 207                         fprintf(stderr, "couldn't open %s\n", filename);
 208                 return;
 209         }
 210         if (read(fd, bf, sizeof(bf)) < 0) {
 211                 fprintf(stderr, "couldn't read %s\n", filename);
 212                 exit(EXIT_FAILURE);
 213         }
 214         close(fd);
 215
 216         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 217         memset(&comm_ev, 0, sizeof(comm_ev));
 218         field = strchr(bf, '(');
 219         if (field == NULL)
 220                 goto out_failure;
 221         sep = strchr(++field, ')');
 222         if (sep == NULL)
 223                 goto out_failure;
 224         size = sep - field;
 225         memcpy(comm_ev.comm, field, size++);
 226
 227         comm_ev.pid = pid;
 228         comm_ev.header.type = PERF_EVENT_COMM;
 229         size = ALIGN(size, sizeof(u64));
 230         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 231
 232         if (!full) {
 233                 comm_ev.tid = pid;
 234
 235                 write_output(&comm_ev, comm_ev.header.size);
 236                 return;
 237         }
 238
 239         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 240
 241         tasks = opendir(filename);
 242         while (!readdir_r(tasks, &dirent, &next) && next) {
 243                 char *end;
 244                 pid = strtol(dirent.d_name, &end, 10);
 245                 if (*end)
 246                         continue;
 247
 248                 comm_ev.tid = pid;
 249
 250                 write_output(&comm_ev, comm_ev.header.size);
 251         }
 252         closedir(tasks);
 253         return;
 254
 255 out_failure:
 256         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 257                 filename);
 258         exit(EXIT_FAILURE);
 259 }
 260
 261 static void pid_synthesize_mmap_samples(pid_t pid)
 262 {
 263         char filename[PATH_MAX];
 264         FILE *fp;
 265
 266         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 267
 268         fp = fopen(filename, "r");
 269         if (fp == NULL) {
 270                 /*
 271                  * We raced with a task exiting - just return:
 272                  */
 273                 if (verbose)
 274                         fprintf(stderr, "couldn't open %s\n", filename);
 275                 return;
 276         }
 277         while (1) {
 278                 char bf[BUFSIZ], *pbf = bf;
 279                 struct mmap_event mmap_ev = {
 280                         .header = { .type = PERF_EVENT_MMAP },
 281                 };
 282                 int n;
 283                 size_t size;
 284                 if (fgets(bf, sizeof(bf), fp) == NULL)
 285                         break;
 286
 287                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 288                 n = hex2u64(pbf, &mmap_ev.start);
 289                 if (n < 0)
 290                         continue;
 291                 pbf += n + 1;
 292                 n = hex2u64(pbf, &mmap_ev.len);
 293                 if (n < 0)
 294                         continue;
 295                 pbf += n + 3;
 296                 if (*pbf == 'x') { /* vm_exec */
 297                         char *execname = strchr(bf, '/');
 298
 299                         /* Catch VDSO */
 300                         if (execname == NULL)
 301                                 execname = strstr(bf, "[vdso]");
 302
 303                         if (execname == NULL)
 304                                 continue;
 305
 306                         size = strlen(execname);
 307                         execname[size - 1] = '\0'; /* Remove \n */
 308                         memcpy(mmap_ev.filename, execname, size);
 309                         size = ALIGN(size, sizeof(u64));
 310                         mmap_ev.len -= mmap_ev.start;
 311                         mmap_ev.header.size = (sizeof(mmap_ev) -
 312                                                (sizeof(mmap_ev.filename) - size));
 313                         mmap_ev.pid = pid;
 314                         mmap_ev.tid = pid;
 315
 316                         write_output(&mmap_ev, mmap_ev.header.size);
 317                 }
 318         }
 319
 320         fclose(fp);
 321 }
 322
 323 static void synthesize_all(void)
 324 {
 325         DIR *proc;
 326         struct dirent dirent, *next;
 327
 328         proc = opendir("/proc");
 329
 330         while (!readdir_r(proc, &dirent, &next) && next) {
 331                 char *end;
 332                 pid_t pid;
 333
 334                 pid = strtol(dirent.d_name, &end, 10);
 335                 if (*end) /* only interested in proper numerical dirents */
 336                         continue;
 337
 338                 pid_synthesize_comm_event(pid, 1);
 339                 pid_synthesize_mmap_samples(pid);
 340         }
 341
 342         closedir(proc);
 343 }
 344
 345 static int group_fd;
 346
 347 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
 348 {
 349         struct perf_header_attr *h_attr;
 350
 351         if (nr < header->attrs) {
 352                 h_attr = header->attr[nr];
 353         } else {
 354                 h_attr = perf_header_attr__new(a);
 355                 perf_header__add_attr(header, h_attr);
 356         }
 357
 358         return h_attr;
 359 }
 360
 361 static void create_counter(int counter, int cpu, pid_t pid)
 362 {
 363         struct perf_counter_attr *attr = attrs + counter;
 364         struct perf_header_attr *h_attr;
 365         int track = !counter; /* only the first counter needs these */
 366         struct {
 367                 u64 count;
 368                 u64 time_enabled;
 369                 u64 time_running;
 370                 u64 id;
 371         } read_data;
 372
 373         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 374                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 375                                   PERF_FORMAT_ID;
 376
 377         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 378
 379         if (freq) {
 380                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 381                 attr->freq              = 1;
 382                 attr->sample_freq       = freq;
 383         }
 384
 385         if (no_samples)
 386                 attr->sample_freq = 0;
 387
 388         if (inherit_stat)
 389                 attr->inherit_stat = 1;
 390
 391         if (sample_address)
 392                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 393
 394         if (call_graph)
 395                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 396
 397
 398         attr->mmap              = track;
 399         attr->comm              = track;
 400         attr->inherit           = (cpu < 0) && inherit;
 401         attr->disabled          = 1;
 402
 403 try_again:
 404         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 405
 406         if (fd[nr_cpu][counter] < 0) {
 407                 int err = errno;
 408
 409                 if (err == EPERM)
 410                         die("Permission error - are you root?\n");
 411
 412                 /*
 413                  * If it's cycles then fall back to hrtimer
 414                  * based cpu-clock-tick sw counter, which
 415                  * is always available even if no PMU support:
 416                  */
 417                 if (attr->type == PERF_TYPE_HARDWARE
 418                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 419
 420                         if (verbose)
 421                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 422                         attr->type = PERF_TYPE_SOFTWARE;
 423                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 424                         goto try_again;
 425                 }
 426                 printf("\n");
 427                 error("perfcounter syscall returned with %d (%s)\n",
 428                         fd[nr_cpu][counter], strerror(err));
 429                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 430                 exit(-1);
 431         }
 432
 433         h_attr = get_header_attr(attr, counter);
 434
 435         if (!file_new) {
 436                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 437                         fprintf(stderr, "incompatible append\n");
 438                         exit(-1);
 439                 }
 440         }
 441
 442         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
 443                 perror("Unable to read perf file descriptor\n");
 444                 exit(-1);
 445         }
 446
 447         perf_header_attr__add_id(h_attr, read_data.id);
 448
 449         assert(fd[nr_cpu][counter] >= 0);
 450         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 451
 452         /*
 453          * First counter acts as the group leader:
 454          */
 455         if (group && group_fd == -1)
 456                 group_fd = fd[nr_cpu][counter];
 457
 458         event_array[nr_poll].fd = fd[nr_cpu][counter];
 459         event_array[nr_poll].events = POLLIN;
 460         nr_poll++;
 461
 462         mmap_array[nr_cpu][counter].counter = counter;
 463         mmap_array[nr_cpu][counter].prev = 0;
 464         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 465         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 466                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
 467         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 468                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 469                 exit(-1);
 470         }
 471
 472         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
 473 }
 474
 475 static void open_counters(int cpu, pid_t pid)
 476 {
 477         int counter;
 478
 479         group_fd = -1;
 480         for (counter = 0; counter < nr_counters; counter++)
 481                 create_counter(counter, cpu, pid);
 482
 483         nr_cpu++;
 484 }
 485
 486 static void atexit_header(void)
 487 {
 488         header->data_size += bytes_written;
 489
 490         perf_header__write(header, output);
 491 }
 492
 493 static int __cmd_record(int argc, const char **argv)
 494 {
 495         int i, counter;
 496         struct stat st;
 497         pid_t pid = 0;
 498         int flags;
 499         int ret;
 500
 501         page_size = sysconf(_SC_PAGE_SIZE);
 502         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 503         assert(nr_cpus <= MAX_NR_CPUS);
 504         assert(nr_cpus >= 0);
 505
 506         atexit(sig_atexit);
 507         signal(SIGCHLD, sig_handler);
 508         signal(SIGINT, sig_handler);
 509
 510         if (!stat(output_name, &st) && st.st_size) {
 511                 if (!force && !append_file) {
 512                         fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 513                                         output_name);
 514                         exit(-1);
 515                 }
 516         } else {
 517                 append_file = 0;
 518         }
 519
 520         flags = O_CREAT|O_RDWR;
 521         if (append_file)
 522                 file_new = 0;
 523         else
 524                 flags |= O_TRUNC;
 525
 526         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 527         if (output < 0) {
 528                 perror("failed to create output file");
 529                 exit(-1);
 530         }
 531
 532         if (!file_new)
 533                 header = perf_header__read(output);
 534         else
 535                 header = perf_header__new();
 536
 537         atexit(atexit_header);
 538
 539         if (!system_wide) {
 540                 pid = target_pid;
 541                 if (pid == -1)
 542                         pid = getpid();
 543
 544                 open_counters(-1, pid);
 545         } else for (i = 0; i < nr_cpus; i++)
 546                 open_counters(i, target_pid);
 547
 548         if (file_new)
 549                 perf_header__write(header, output);
 550
 551         if (!system_wide) {
 552                 pid_synthesize_comm_event(pid, 0);
 553                 pid_synthesize_mmap_samples(pid);
 554         } else
 555                 synthesize_all();
 556
 557         if (target_pid == -1 && argc) {
 558                 pid = fork();
 559                 if (pid < 0)
 560                         perror("failed to fork");
 561
 562                 if (!pid) {
 563                         if (execvp(argv[0], (char **)argv)) {
 564                                 perror(argv[0]);
 565                                 exit(-1);
 566                         }
 567                 }
 568         }
 569
 570         if (realtime_prio) {
 571                 struct sched_param param;
 572
 573                 param.sched_priority = realtime_prio;
 574                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 575                         printf("Could not set realtime priority.\n");
 576                         exit(-1);
 577                 }
 578         }
 579
 580         for (;;) {
 581                 int hits = samples;
 582
 583                 for (i = 0; i < nr_cpu; i++) {
 584                         for (counter = 0; counter < nr_counters; counter++)
 585                                 mmap_read(&mmap_array[i][counter]);
 586                 }
 587
 588                 if (hits == samples) {
 589                         if (done)
 590                                 break;
 591                         ret = poll(event_array, nr_poll, 100);
 592                 }
 593         }
 594
 595         /*
 596          * Approximate RIP event size: 24 bytes.
 597          */
 598         fprintf(stderr,
 599                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 600                 (double)bytes_written / 1024.0 / 1024.0,
 601                 output_name,
 602                 bytes_written / 24);
 603
 604         return 0;
 605 }
 606
 607 static const char * const record_usage[] = {
 608         "perf record [<options>] [<command>]",
 609         "perf record [<options>] -- <command> [<options>]",
 610         NULL
 611 };
 612
 613 static const struct option options[] = {
 614         OPT_CALLBACK('e', "event", NULL, "event",
 615                      "event selector. use 'perf list' to list available events",
 616                      parse_events),
 617         OPT_INTEGER('p', "pid", &target_pid,
 618                     "record events on existing pid"),
 619         OPT_INTEGER('r', "realtime", &realtime_prio,
 620                     "collect data with this RT SCHED_FIFO priority"),
 621         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 622                             "system-wide collection from all CPUs"),
 623         OPT_BOOLEAN('A', "append", &append_file,
 624                             "append to the output file to do incremental profiling"),
 625         OPT_BOOLEAN('f', "force", &force,
 626                         "overwrite existing data file"),
 627         OPT_LONG('c', "count", &default_interval,
 628                     "event period to sample"),
 629         OPT_STRING('o', "output", &output_name, "file",
 630                     "output file name"),
 631         OPT_BOOLEAN('i', "inherit", &inherit,
 632                     "child tasks inherit counters"),
 633         OPT_INTEGER('F', "freq", &freq,
 634                     "profile at this frequency"),
 635         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 636                     "number of mmap data pages"),
 637         OPT_BOOLEAN('g', "call-graph", &call_graph,
 638                     "do call-graph (stack chain/backtrace) recording"),
 639         OPT_BOOLEAN('v', "verbose", &verbose,
 640                     "be more verbose (show counter open errors, etc)"),
 641         OPT_BOOLEAN('s', "stat", &inherit_stat,
 642                     "per thread counts"),
 643         OPT_BOOLEAN('d', "data", &sample_address,
 644                     "Sample addresses"),
 645         OPT_BOOLEAN('n', "no-samples", &no_samples,
 646                     "don't sample"),
 647         OPT_END()
 648 };
 649
 650 int cmd_record(int argc, const char **argv, const char *prefix __used)
 651 {
 652         int counter;
 653
 654         argc = parse_options(argc, argv, options, record_usage,
 655                 PARSE_OPT_STOP_AT_NON_OPTION);
 656         if (!argc && target_pid == -1 && !system_wide)
 657                 usage_with_options(record_usage, options);
 658
 659         if (!nr_counters) {
 660                 nr_counters     = 1;
 661                 attrs[0].type   = PERF_TYPE_HARDWARE;
 662                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 663         }
 664
 665         for (counter = 0; counter < nr_counters; counter++) {
 666                 if (attrs[counter].sample_period)
 667                         continue;
 668
 669                 attrs[counter].sample_period = default_interval;
 670         }
 671
 672         return __cmd_record(argc, argv);
 673 }