tools/perf/builtin-record.c

   1 /*
   2  * builtin-record.c
   3  *
   4  * Builtin record command: Record the profile of a workload
   5  * (or a CPU, or a PID) into the perf.data output file - for
   6  * later analysis via perf report.
   7  */
   8 #include "builtin.h"
   9
  10 #include "perf.h"
  11
  12 #include "util/util.h"
  13 #include "util/parse-options.h"
  14 #include "util/parse-events.h"
  15 #include "util/string.h"
  16
  17 #include "util/header.h"
  18 #include "util/event.h"
  19 #include "util/debug.h"
  20
  21 #include <unistd.h>
  22 #include <sched.h>
  23
  24 #define ALIGN(x, a)             __ALIGN_MASK(x, (typeof(x))(a)-1)
  25 #define __ALIGN_MASK(x, mask)   (((x)+(mask))&~(mask))
  26
  27 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  28
  29 static long                     default_interval                = 100000;
  30
  31 static int                      nr_cpus                         = 0;
  32 static unsigned int             page_size;
  33 static unsigned int             mmap_pages                      = 128;
  34 static int                      freq                            = 0;
  35 static int                      output;
  36 static const char               *output_name                    = "perf.data";
  37 static int                      group                           = 0;
  38 static unsigned int             realtime_prio                   = 0;
  39 static int                      raw_samples                     = 0;
  40 static int                      system_wide                     = 0;
  41 static int                      profile_cpu                     = -1;
  42 static pid_t                    target_pid                      = -1;
  43 static int                      inherit                         = 1;
  44 static int                      force                           = 0;
  45 static int                      append_file                     = 0;
  46 static int                      call_graph                      = 0;
  47 static int                      inherit_stat                    = 0;
  48 static int                      no_samples                      = 0;
  49 static int                      sample_address                  = 0;
  50
  51 static long                     samples;
  52 static struct timeval           last_read;
  53 static struct timeval           this_read;
  54
  55 static u64                      bytes_written;
  56
  57 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
  58
  59 static int                      nr_poll;
  60 static int                      nr_cpu;
  61
  62 static int                      file_new = 1;
  63
  64 struct perf_header              *header;
  65
  66 struct mmap_data {
  67         int                     counter;
  68         void                    *base;
  69         unsigned int            mask;
  70         unsigned int            prev;
  71 };
  72
  73 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  74
  75 static unsigned long mmap_read_head(struct mmap_data *md)
  76 {
  77         struct perf_counter_mmap_page *pc = md->base;
  78         long head;
  79
  80         head = pc->data_head;
  81         rmb();
  82
  83         return head;
  84 }
  85
  86 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
  87 {
  88         struct perf_counter_mmap_page *pc = md->base;
  89
  90         /*
  91          * ensure all reads are done before we write the tail out.
  92          */
  93         /* mb(); */
  94         pc->data_tail = tail;
  95 }
  96
  97 static void write_output(void *buf, size_t size)
  98 {
  99         while (size) {
 100                 int ret = write(output, buf, size);
 101
 102                 if (ret < 0)
 103                         die("failed to write");
 104
 105                 size -= ret;
 106                 buf += ret;
 107
 108                 bytes_written += ret;
 109         }
 110 }
 111
 112 static void mmap_read(struct mmap_data *md)
 113 {
 114         unsigned int head = mmap_read_head(md);
 115         unsigned int old = md->prev;
 116         unsigned char *data = md->base + page_size;
 117         unsigned long size;
 118         void *buf;
 119         int diff;
 120
 121         gettimeofday(&this_read, NULL);
 122
 123         /*
 124          * If we're further behind than half the buffer, there's a chance
 125          * the writer will bite our tail and mess up the samples under us.
 126          *
 127          * If we somehow ended up ahead of the head, we got messed up.
 128          *
 129          * In either case, truncate and restart at head.
 130          */
 131         diff = head - old;
 132         if (diff < 0) {
 133                 struct timeval iv;
 134                 unsigned long msecs;
 135
 136                 timersub(&this_read, &last_read, &iv);
 137                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 138
 139                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 140                                 "  Last read %lu msecs ago.\n", msecs);
 141
 142                 /*
 143                  * head points to a known good entry, start there.
 144                  */
 145                 old = head;
 146         }
 147
 148         last_read = this_read;
 149
 150         if (old != head)
 151                 samples++;
 152
 153         size = head - old;
 154
 155         if ((old & md->mask) + size != (head & md->mask)) {
 156                 buf = &data[old & md->mask];
 157                 size = md->mask + 1 - (old & md->mask);
 158                 old += size;
 159
 160                 write_output(buf, size);
 161         }
 162
 163         buf = &data[old & md->mask];
 164         size = head - old;
 165         old += size;
 166
 167         write_output(buf, size);
 168
 169         md->prev = old;
 170         mmap_write_tail(md, old);
 171 }
 172
 173 static volatile int done = 0;
 174 static volatile int signr = -1;
 175
 176 static void sig_handler(int sig)
 177 {
 178         done = 1;
 179         signr = sig;
 180 }
 181
 182 static void sig_atexit(void)
 183 {
 184         if (signr == -1)
 185                 return;
 186
 187         signal(signr, SIG_DFL);
 188         kill(getpid(), signr);
 189 }
 190
 191 static pid_t pid_synthesize_comm_event(pid_t pid, int full)
 192 {
 193         struct comm_event comm_ev;
 194         char filename[PATH_MAX];
 195         char bf[BUFSIZ];
 196         FILE *fp;
 197         size_t size = 0;
 198         DIR *tasks;
 199         struct dirent dirent, *next;
 200         pid_t tgid = 0;
 201
 202         snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
 203
 204         fp = fopen(filename, "r");
 205         if (fp == NULL) {
 206                 /*
 207                  * We raced with a task exiting - just return:
 208                  */
 209                 if (verbose)
 210                         fprintf(stderr, "couldn't open %s\n", filename);
 211                 return 0;
 212         }
 213
 214         memset(&comm_ev, 0, sizeof(comm_ev));
 215         while (!comm_ev.comm[0] || !comm_ev.pid) {
 216                 if (fgets(bf, sizeof(bf), fp) == NULL)
 217                         goto out_failure;
 218
 219                 if (memcmp(bf, "Name:", 5) == 0) {
 220                         char *name = bf + 5;
 221                         while (*name && isspace(*name))
 222                                 ++name;
 223                         size = strlen(name) - 1;
 224                         memcpy(comm_ev.comm, name, size++);
 225                 } else if (memcmp(bf, "Tgid:", 5) == 0) {
 226                         char *tgids = bf + 5;
 227                         while (*tgids && isspace(*tgids))
 228                                 ++tgids;
 229                         tgid = comm_ev.pid = atoi(tgids);
 230                 }
 231         }
 232
 233         comm_ev.header.type = PERF_EVENT_COMM;
 234         size = ALIGN(size, sizeof(u64));
 235         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 236
 237         if (!full) {
 238                 comm_ev.tid = pid;
 239
 240                 write_output(&comm_ev, comm_ev.header.size);
 241                 goto out_fclose;
 242         }
 243
 244         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 245
 246         tasks = opendir(filename);
 247         while (!readdir_r(tasks, &dirent, &next) && next) {
 248                 char *end;
 249                 pid = strtol(dirent.d_name, &end, 10);
 250                 if (*end)
 251                         continue;
 252
 253                 comm_ev.tid = pid;
 254
 255                 write_output(&comm_ev, comm_ev.header.size);
 256         }
 257         closedir(tasks);
 258
 259 out_fclose:
 260         fclose(fp);
 261         return tgid;
 262
 263 out_failure:
 264         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 265                 filename);
 266         exit(EXIT_FAILURE);
 267 }
 268
 269 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
 270 {
 271         char filename[PATH_MAX];
 272         FILE *fp;
 273
 274         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 275
 276         fp = fopen(filename, "r");
 277         if (fp == NULL) {
 278                 /*
 279                  * We raced with a task exiting - just return:
 280                  */
 281                 if (verbose)
 282                         fprintf(stderr, "couldn't open %s\n", filename);
 283                 return;
 284         }
 285         while (1) {
 286                 char bf[BUFSIZ], *pbf = bf;
 287                 struct mmap_event mmap_ev = {
 288                         .header = { .type = PERF_EVENT_MMAP },
 289                 };
 290                 int n;
 291                 size_t size;
 292                 if (fgets(bf, sizeof(bf), fp) == NULL)
 293                         break;
 294
 295                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 296                 n = hex2u64(pbf, &mmap_ev.start);
 297                 if (n < 0)
 298                         continue;
 299                 pbf += n + 1;
 300                 n = hex2u64(pbf, &mmap_ev.len);
 301                 if (n < 0)
 302                         continue;
 303                 pbf += n + 3;
 304                 if (*pbf == 'x') { /* vm_exec */
 305                         char *execname = strchr(bf, '/');
 306
 307                         /* Catch VDSO */
 308                         if (execname == NULL)
 309                                 execname = strstr(bf, "[vdso]");
 310
 311                         if (execname == NULL)
 312                                 continue;
 313
 314                         size = strlen(execname);
 315                         execname[size - 1] = '\0'; /* Remove \n */
 316                         memcpy(mmap_ev.filename, execname, size);
 317                         size = ALIGN(size, sizeof(u64));
 318                         mmap_ev.len -= mmap_ev.start;
 319                         mmap_ev.header.size = (sizeof(mmap_ev) -
 320                                                (sizeof(mmap_ev.filename) - size));
 321                         mmap_ev.pid = tgid;
 322                         mmap_ev.tid = pid;
 323
 324                         write_output(&mmap_ev, mmap_ev.header.size);
 325                 }
 326         }
 327
 328         fclose(fp);
 329 }
 330
 331 static void synthesize_all(void)
 332 {
 333         DIR *proc;
 334         struct dirent dirent, *next;
 335
 336         proc = opendir("/proc");
 337
 338         while (!readdir_r(proc, &dirent, &next) && next) {
 339                 char *end;
 340                 pid_t pid, tgid;
 341
 342                 pid = strtol(dirent.d_name, &end, 10);
 343                 if (*end) /* only interested in proper numerical dirents */
 344                         continue;
 345
 346                 tgid = pid_synthesize_comm_event(pid, 1);
 347                 pid_synthesize_mmap_samples(pid, tgid);
 348         }
 349
 350         closedir(proc);
 351 }
 352
 353 static int group_fd;
 354
 355 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
 356 {
 357         struct perf_header_attr *h_attr;
 358
 359         if (nr < header->attrs) {
 360                 h_attr = header->attr[nr];
 361         } else {
 362                 h_attr = perf_header_attr__new(a);
 363                 perf_header__add_attr(header, h_attr);
 364         }
 365
 366         return h_attr;
 367 }
 368
 369 static void create_counter(int counter, int cpu, pid_t pid)
 370 {
 371         struct perf_counter_attr *attr = attrs + counter;
 372         struct perf_header_attr *h_attr;
 373         int track = !counter; /* only the first counter needs these */
 374         struct {
 375                 u64 count;
 376                 u64 time_enabled;
 377                 u64 time_running;
 378                 u64 id;
 379         } read_data;
 380
 381         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 382                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 383                                   PERF_FORMAT_ID;
 384
 385         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 386
 387         if (freq) {
 388                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 389                 attr->freq              = 1;
 390                 attr->sample_freq       = freq;
 391         }
 392
 393         if (no_samples)
 394                 attr->sample_freq = 0;
 395
 396         if (inherit_stat)
 397                 attr->inherit_stat = 1;
 398
 399         if (sample_address)
 400                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 401
 402         if (call_graph)
 403                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 404
 405         if (raw_samples)
 406                 attr->sample_type       |= PERF_SAMPLE_RAW;
 407
 408         attr->mmap              = track;
 409         attr->comm              = track;
 410         attr->inherit           = (cpu < 0) && inherit;
 411         attr->disabled          = 1;
 412
 413 try_again:
 414         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 415
 416         if (fd[nr_cpu][counter] < 0) {
 417                 int err = errno;
 418
 419                 if (err == EPERM)
 420                         die("Permission error - are you root?\n");
 421                 else if (err ==  ENODEV && profile_cpu != -1)
 422                         die("No such device - did you specify an out-of-range profile CPU?\n");
 423
 424                 /*
 425                  * If it's cycles then fall back to hrtimer
 426                  * based cpu-clock-tick sw counter, which
 427                  * is always available even if no PMU support:
 428                  */
 429                 if (attr->type == PERF_TYPE_HARDWARE
 430                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 431
 432                         if (verbose)
 433                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 434                         attr->type = PERF_TYPE_SOFTWARE;
 435                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 436                         goto try_again;
 437                 }
 438                 printf("\n");
 439                 error("perfcounter syscall returned with %d (%s)\n",
 440                         fd[nr_cpu][counter], strerror(err));
 441                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 442                 exit(-1);
 443         }
 444
 445         h_attr = get_header_attr(attr, counter);
 446
 447         if (!file_new) {
 448                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 449                         fprintf(stderr, "incompatible append\n");
 450                         exit(-1);
 451                 }
 452         }
 453
 454         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
 455                 perror("Unable to read perf file descriptor\n");
 456                 exit(-1);
 457         }
 458
 459         perf_header_attr__add_id(h_attr, read_data.id);
 460
 461         assert(fd[nr_cpu][counter] >= 0);
 462         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 463
 464         /*
 465          * First counter acts as the group leader:
 466          */
 467         if (group && group_fd == -1)
 468                 group_fd = fd[nr_cpu][counter];
 469
 470         event_array[nr_poll].fd = fd[nr_cpu][counter];
 471         event_array[nr_poll].events = POLLIN;
 472         nr_poll++;
 473
 474         mmap_array[nr_cpu][counter].counter = counter;
 475         mmap_array[nr_cpu][counter].prev = 0;
 476         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 477         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 478                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
 479         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 480                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 481                 exit(-1);
 482         }
 483
 484         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
 485 }
 486
 487 static void open_counters(int cpu, pid_t pid)
 488 {
 489         int counter;
 490
 491         group_fd = -1;
 492         for (counter = 0; counter < nr_counters; counter++)
 493                 create_counter(counter, cpu, pid);
 494
 495         nr_cpu++;
 496 }
 497
 498 static void atexit_header(void)
 499 {
 500         header->data_size += bytes_written;
 501
 502         perf_header__write(header, output);
 503 }
 504
 505 static int __cmd_record(int argc, const char **argv)
 506 {
 507         int i, counter;
 508         struct stat st;
 509         pid_t pid = 0;
 510         int flags;
 511         int ret;
 512
 513         page_size = sysconf(_SC_PAGE_SIZE);
 514         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 515         assert(nr_cpus <= MAX_NR_CPUS);
 516         assert(nr_cpus >= 0);
 517
 518         atexit(sig_atexit);
 519         signal(SIGCHLD, sig_handler);
 520         signal(SIGINT, sig_handler);
 521
 522         if (!stat(output_name, &st) && st.st_size) {
 523                 if (!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         } else {
 529                 append_file = 0;
 530         }
 531
 532         flags = O_CREAT|O_RDWR;
 533         if (append_file)
 534                 file_new = 0;
 535         else
 536                 flags |= O_TRUNC;
 537
 538         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 539         if (output < 0) {
 540                 perror("failed to create output file");
 541                 exit(-1);
 542         }
 543
 544         if (!file_new)
 545                 header = perf_header__read(output);
 546         else
 547                 header = perf_header__new();
 548
 549         atexit(atexit_header);
 550
 551         if (!system_wide) {
 552                 pid = target_pid;
 553                 if (pid == -1)
 554                         pid = getpid();
 555
 556                 open_counters(profile_cpu, pid);
 557         } else {
 558                 if (profile_cpu != -1) {
 559                         open_counters(profile_cpu, target_pid);
 560                 } else {
 561                         for (i = 0; i < nr_cpus; i++)
 562                                 open_counters(i, target_pid);
 563                 }
 564         }
 565
 566         if (file_new)
 567                 perf_header__write(header, output);
 568
 569         if (!system_wide) {
 570                 pid_t tgid = pid_synthesize_comm_event(pid, 0);
 571                 pid_synthesize_mmap_samples(pid, tgid);
 572         } else
 573                 synthesize_all();
 574
 575         if (target_pid == -1 && argc) {
 576                 pid = fork();
 577                 if (pid < 0)
 578                         perror("failed to fork");
 579
 580                 if (!pid) {
 581                         if (execvp(argv[0], (char **)argv)) {
 582                                 perror(argv[0]);
 583                                 exit(-1);
 584                         }
 585                 }
 586         }
 587
 588         if (realtime_prio) {
 589                 struct sched_param param;
 590
 591                 param.sched_priority = realtime_prio;
 592                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 593                         printf("Could not set realtime priority.\n");
 594                         exit(-1);
 595                 }
 596         }
 597
 598         for (;;) {
 599                 int hits = samples;
 600
 601                 for (i = 0; i < nr_cpu; i++) {
 602                         for (counter = 0; counter < nr_counters; counter++)
 603                                 mmap_read(&mmap_array[i][counter]);
 604                 }
 605
 606                 if (hits == samples) {
 607                         if (done)
 608                                 break;
 609                         ret = poll(event_array, nr_poll, 100);
 610                 }
 611         }
 612
 613         /*
 614          * Approximate RIP event size: 24 bytes.
 615          */
 616         fprintf(stderr,
 617                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 618                 (double)bytes_written / 1024.0 / 1024.0,
 619                 output_name,
 620                 bytes_written / 24);
 621
 622         return 0;
 623 }
 624
 625 static const char * const record_usage[] = {
 626         "perf record [<options>] [<command>]",
 627         "perf record [<options>] -- <command> [<options>]",
 628         NULL
 629 };
 630
 631 static const struct option options[] = {
 632         OPT_CALLBACK('e', "event", NULL, "event",
 633                      "event selector. use 'perf list' to list available events",
 634                      parse_events),
 635         OPT_INTEGER('p', "pid", &target_pid,
 636                     "record events on existing pid"),
 637         OPT_INTEGER('r', "realtime", &realtime_prio,
 638                     "collect data with this RT SCHED_FIFO priority"),
 639         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 640                     "collect raw sample records from all opened counters"),
 641         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 642                             "system-wide collection from all CPUs"),
 643         OPT_BOOLEAN('A', "append", &append_file,
 644                             "append to the output file to do incremental profiling"),
 645         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
 646                             "CPU to profile on"),
 647         OPT_BOOLEAN('f', "force", &force,
 648                         "overwrite existing data file"),
 649         OPT_LONG('c', "count", &default_interval,
 650                     "event period to sample"),
 651         OPT_STRING('o', "output", &output_name, "file",
 652                     "output file name"),
 653         OPT_BOOLEAN('i', "inherit", &inherit,
 654                     "child tasks inherit counters"),
 655         OPT_INTEGER('F', "freq", &freq,
 656                     "profile at this frequency"),
 657         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 658                     "number of mmap data pages"),
 659         OPT_BOOLEAN('g', "call-graph", &call_graph,
 660                     "do call-graph (stack chain/backtrace) recording"),
 661         OPT_BOOLEAN('v', "verbose", &verbose,
 662                     "be more verbose (show counter open errors, etc)"),
 663         OPT_BOOLEAN('s', "stat", &inherit_stat,
 664                     "per thread counts"),
 665         OPT_BOOLEAN('d', "data", &sample_address,
 666                     "Sample addresses"),
 667         OPT_BOOLEAN('n', "no-samples", &no_samples,
 668                     "don't sample"),
 669         OPT_END()
 670 };
 671
 672 int cmd_record(int argc, const char **argv, const char *prefix __used)
 673 {
 674         int counter;
 675
 676         argc = parse_options(argc, argv, options, record_usage,
 677                 PARSE_OPT_STOP_AT_NON_OPTION);
 678         if (!argc && target_pid == -1 && !system_wide)
 679                 usage_with_options(record_usage, options);
 680
 681         if (!nr_counters) {
 682                 nr_counters     = 1;
 683                 attrs[0].type   = PERF_TYPE_HARDWARE;
 684                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 685         }
 686
 687         for (counter = 0; counter < nr_counters; counter++) {
 688                 if (attrs[counter].sample_period)
 689                         continue;
 690
 691                 attrs[counter].sample_period = default_interval;
 692         }
 693
 694         return __cmd_record(argc, argv);
 695 }