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 #include "util/trace-event.h"
  21
  22 #include <unistd.h>
  23 #include <sched.h>
  24
  25 #define ALIGN(x, a)             __ALIGN_MASK(x, (typeof(x))(a)-1)
  26 #define __ALIGN_MASK(x, mask)   (((x)+(mask))&~(mask))
  27
  28 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  29
  30 static long                     default_interval                = 100000;
  31
  32 static int                      nr_cpus                         = 0;
  33 static unsigned int             page_size;
  34 static unsigned int             mmap_pages                      = 128;
  35 static int                      freq                            = 0;
  36 static int                      output;
  37 static const char               *output_name                    = "perf.data";
  38 static int                      group                           = 0;
  39 static unsigned int             realtime_prio                   = 0;
  40 static int                      raw_samples                     = 0;
  41 static int                      system_wide                     = 0;
  42 static int                      profile_cpu                     = -1;
  43 static pid_t                    target_pid                      = -1;
  44 static int                      inherit                         = 1;
  45 static int                      force                           = 0;
  46 static int                      append_file                     = 0;
  47 static int                      call_graph                      = 0;
  48 static int                      inherit_stat                    = 0;
  49 static int                      no_samples                      = 0;
  50 static int                      sample_address                  = 0;
  51 static int                      multiplex                       = 0;
  52
  53 static long                     samples;
  54 static struct timeval           last_read;
  55 static struct timeval           this_read;
  56
  57 static u64                      bytes_written;
  58
  59 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
  60
  61 static int                      nr_poll;
  62 static int                      nr_cpu;
  63
  64 static int                      file_new = 1;
  65
  66 struct perf_header              *header;
  67
  68 struct mmap_data {
  69         int                     counter;
  70         void                    *base;
  71         unsigned int            mask;
  72         unsigned int            prev;
  73 };
  74
  75 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  76
  77 static unsigned long mmap_read_head(struct mmap_data *md)
  78 {
  79         struct perf_counter_mmap_page *pc = md->base;
  80         long head;
  81
  82         head = pc->data_head;
  83         rmb();
  84
  85         return head;
  86 }
  87
  88 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
  89 {
  90         struct perf_counter_mmap_page *pc = md->base;
  91
  92         /*
  93          * ensure all reads are done before we write the tail out.
  94          */
  95         /* mb(); */
  96         pc->data_tail = tail;
  97 }
  98
  99 static void write_output(void *buf, size_t size)
 100 {
 101         while (size) {
 102                 int ret = write(output, buf, size);
 103
 104                 if (ret < 0)
 105                         die("failed to write");
 106
 107                 size -= ret;
 108                 buf += ret;
 109
 110                 bytes_written += ret;
 111         }
 112 }
 113
 114 static void mmap_read(struct mmap_data *md)
 115 {
 116         unsigned int head = mmap_read_head(md);
 117         unsigned int old = md->prev;
 118         unsigned char *data = md->base + page_size;
 119         unsigned long size;
 120         void *buf;
 121         int diff;
 122
 123         gettimeofday(&this_read, NULL);
 124
 125         /*
 126          * If we're further behind than half the buffer, there's a chance
 127          * the writer will bite our tail and mess up the samples under us.
 128          *
 129          * If we somehow ended up ahead of the head, we got messed up.
 130          *
 131          * In either case, truncate and restart at head.
 132          */
 133         diff = head - old;
 134         if (diff < 0) {
 135                 struct timeval iv;
 136                 unsigned long msecs;
 137
 138                 timersub(&this_read, &last_read, &iv);
 139                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 140
 141                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 142                                 "  Last read %lu msecs ago.\n", msecs);
 143
 144                 /*
 145                  * head points to a known good entry, start there.
 146                  */
 147                 old = head;
 148         }
 149
 150         last_read = this_read;
 151
 152         if (old != head)
 153                 samples++;
 154
 155         size = head - old;
 156
 157         if ((old & md->mask) + size != (head & md->mask)) {
 158                 buf = &data[old & md->mask];
 159                 size = md->mask + 1 - (old & md->mask);
 160                 old += size;
 161
 162                 write_output(buf, size);
 163         }
 164
 165         buf = &data[old & md->mask];
 166         size = head - old;
 167         old += size;
 168
 169         write_output(buf, size);
 170
 171         md->prev = old;
 172         mmap_write_tail(md, old);
 173 }
 174
 175 static volatile int done = 0;
 176 static volatile int signr = -1;
 177
 178 static void sig_handler(int sig)
 179 {
 180         done = 1;
 181         signr = sig;
 182 }
 183
 184 static void sig_atexit(void)
 185 {
 186         if (signr == -1)
 187                 return;
 188
 189         signal(signr, SIG_DFL);
 190         kill(getpid(), signr);
 191 }
 192
 193 static pid_t pid_synthesize_comm_event(pid_t pid, int full)
 194 {
 195         struct comm_event comm_ev;
 196         char filename[PATH_MAX];
 197         char bf[BUFSIZ];
 198         FILE *fp;
 199         size_t size = 0;
 200         DIR *tasks;
 201         struct dirent dirent, *next;
 202         pid_t tgid = 0;
 203
 204         snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
 205
 206         fp = fopen(filename, "r");
 207         if (fp == NULL) {
 208                 /*
 209                  * We raced with a task exiting - just return:
 210                  */
 211                 if (verbose)
 212                         fprintf(stderr, "couldn't open %s\n", filename);
 213                 return 0;
 214         }
 215
 216         memset(&comm_ev, 0, sizeof(comm_ev));
 217         while (!comm_ev.comm[0] || !comm_ev.pid) {
 218                 if (fgets(bf, sizeof(bf), fp) == NULL)
 219                         goto out_failure;
 220
 221                 if (memcmp(bf, "Name:", 5) == 0) {
 222                         char *name = bf + 5;
 223                         while (*name && isspace(*name))
 224                                 ++name;
 225                         size = strlen(name) - 1;
 226                         memcpy(comm_ev.comm, name, size++);
 227                 } else if (memcmp(bf, "Tgid:", 5) == 0) {
 228                         char *tgids = bf + 5;
 229                         while (*tgids && isspace(*tgids))
 230                                 ++tgids;
 231                         tgid = comm_ev.pid = atoi(tgids);
 232                 }
 233         }
 234
 235         comm_ev.header.type = PERF_EVENT_COMM;
 236         size = ALIGN(size, sizeof(u64));
 237         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 238
 239         if (!full) {
 240                 comm_ev.tid = pid;
 241
 242                 write_output(&comm_ev, comm_ev.header.size);
 243                 goto out_fclose;
 244         }
 245
 246         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 247
 248         tasks = opendir(filename);
 249         while (!readdir_r(tasks, &dirent, &next) && next) {
 250                 char *end;
 251                 pid = strtol(dirent.d_name, &end, 10);
 252                 if (*end)
 253                         continue;
 254
 255                 comm_ev.tid = pid;
 256
 257                 write_output(&comm_ev, comm_ev.header.size);
 258         }
 259         closedir(tasks);
 260
 261 out_fclose:
 262         fclose(fp);
 263         return tgid;
 264
 265 out_failure:
 266         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 267                 filename);
 268         exit(EXIT_FAILURE);
 269 }
 270
 271 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
 272 {
 273         char filename[PATH_MAX];
 274         FILE *fp;
 275
 276         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 277
 278         fp = fopen(filename, "r");
 279         if (fp == NULL) {
 280                 /*
 281                  * We raced with a task exiting - just return:
 282                  */
 283                 if (verbose)
 284                         fprintf(stderr, "couldn't open %s\n", filename);
 285                 return;
 286         }
 287         while (1) {
 288                 char bf[BUFSIZ], *pbf = bf;
 289                 struct mmap_event mmap_ev = {
 290                         .header = { .type = PERF_EVENT_MMAP },
 291                 };
 292                 int n;
 293                 size_t size;
 294                 if (fgets(bf, sizeof(bf), fp) == NULL)
 295                         break;
 296
 297                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 298                 n = hex2u64(pbf, &mmap_ev.start);
 299                 if (n < 0)
 300                         continue;
 301                 pbf += n + 1;
 302                 n = hex2u64(pbf, &mmap_ev.len);
 303                 if (n < 0)
 304                         continue;
 305                 pbf += n + 3;
 306                 if (*pbf == 'x') { /* vm_exec */
 307                         char *execname = strchr(bf, '/');
 308
 309                         /* Catch VDSO */
 310                         if (execname == NULL)
 311                                 execname = strstr(bf, "[vdso]");
 312
 313                         if (execname == NULL)
 314                                 continue;
 315
 316                         size = strlen(execname);
 317                         execname[size - 1] = '\0'; /* Remove \n */
 318                         memcpy(mmap_ev.filename, execname, size);
 319                         size = ALIGN(size, sizeof(u64));
 320                         mmap_ev.len -= mmap_ev.start;
 321                         mmap_ev.header.size = (sizeof(mmap_ev) -
 322                                                (sizeof(mmap_ev.filename) - size));
 323                         mmap_ev.pid = tgid;
 324                         mmap_ev.tid = pid;
 325
 326                         write_output(&mmap_ev, mmap_ev.header.size);
 327                 }
 328         }
 329
 330         fclose(fp);
 331 }
 332
 333 static void synthesize_all(void)
 334 {
 335         DIR *proc;
 336         struct dirent dirent, *next;
 337
 338         proc = opendir("/proc");
 339
 340         while (!readdir_r(proc, &dirent, &next) && next) {
 341                 char *end;
 342                 pid_t pid, tgid;
 343
 344                 pid = strtol(dirent.d_name, &end, 10);
 345                 if (*end) /* only interested in proper numerical dirents */
 346                         continue;
 347
 348                 tgid = pid_synthesize_comm_event(pid, 1);
 349                 pid_synthesize_mmap_samples(pid, tgid);
 350         }
 351
 352         closedir(proc);
 353 }
 354
 355 static int group_fd;
 356
 357 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
 358 {
 359         struct perf_header_attr *h_attr;
 360
 361         if (nr < header->attrs) {
 362                 h_attr = header->attr[nr];
 363         } else {
 364                 h_attr = perf_header_attr__new(a);
 365                 perf_header__add_attr(header, h_attr);
 366         }
 367
 368         return h_attr;
 369 }
 370
 371 static void create_counter(int counter, int cpu, pid_t pid)
 372 {
 373         struct perf_counter_attr *attr = attrs + counter;
 374         struct perf_header_attr *h_attr;
 375         int track = !counter; /* only the first counter needs these */
 376         struct {
 377                 u64 count;
 378                 u64 time_enabled;
 379                 u64 time_running;
 380                 u64 id;
 381         } read_data;
 382
 383         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 384                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 385                                   PERF_FORMAT_ID;
 386
 387         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 388
 389         if (freq) {
 390                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 391                 attr->freq              = 1;
 392                 attr->sample_freq       = freq;
 393         }
 394
 395         if (no_samples)
 396                 attr->sample_freq = 0;
 397
 398         if (inherit_stat)
 399                 attr->inherit_stat = 1;
 400
 401         if (sample_address)
 402                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 403
 404         if (call_graph)
 405                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 406
 407         if (raw_samples) {
 408                 attr->sample_type       |= PERF_SAMPLE_TIME;
 409                 attr->sample_type       |= PERF_SAMPLE_RAW;
 410                 attr->sample_type       |= PERF_SAMPLE_CPU;
 411         }
 412
 413         attr->mmap              = track;
 414         attr->comm              = track;
 415         attr->inherit           = (cpu < 0) && inherit;
 416         attr->disabled          = 1;
 417
 418 try_again:
 419         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 420
 421         if (fd[nr_cpu][counter] < 0) {
 422                 int err = errno;
 423
 424                 if (err == EPERM)
 425                         die("Permission error - are you root?\n");
 426                 else if (err ==  ENODEV && profile_cpu != -1)
 427                         die("No such device - did you specify an out-of-range profile CPU?\n");
 428
 429                 /*
 430                  * If it's cycles then fall back to hrtimer
 431                  * based cpu-clock-tick sw counter, which
 432                  * is always available even if no PMU support:
 433                  */
 434                 if (attr->type == PERF_TYPE_HARDWARE
 435                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 436
 437                         if (verbose)
 438                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 439                         attr->type = PERF_TYPE_SOFTWARE;
 440                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 441                         goto try_again;
 442                 }
 443                 printf("\n");
 444                 error("perfcounter syscall returned with %d (%s)\n",
 445                         fd[nr_cpu][counter], strerror(err));
 446                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 447                 exit(-1);
 448         }
 449
 450         h_attr = get_header_attr(attr, counter);
 451
 452         if (!file_new) {
 453                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 454                         fprintf(stderr, "incompatible append\n");
 455                         exit(-1);
 456                 }
 457         }
 458
 459         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
 460                 perror("Unable to read perf file descriptor\n");
 461                 exit(-1);
 462         }
 463
 464         perf_header_attr__add_id(h_attr, read_data.id);
 465
 466         assert(fd[nr_cpu][counter] >= 0);
 467         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 468
 469         /*
 470          * First counter acts as the group leader:
 471          */
 472         if (group && group_fd == -1)
 473                 group_fd = fd[nr_cpu][counter];
 474
 475         event_array[nr_poll].fd = fd[nr_cpu][counter];
 476         event_array[nr_poll].events = POLLIN;
 477         nr_poll++;
 478
 479         mmap_array[nr_cpu][counter].counter = counter;
 480         mmap_array[nr_cpu][counter].prev = 0;
 481         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 482         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 483                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
 484         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 485                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 486                 exit(-1);
 487         }
 488
 489         if (multiplex && fd[nr_cpu][counter] != group_fd)
 490                 ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_SET_OUTPUT, group_fd);
 491
 492         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
 493 }
 494
 495 static void open_counters(int cpu, pid_t pid)
 496 {
 497         int counter;
 498
 499         group_fd = -1;
 500         for (counter = 0; counter < nr_counters; counter++)
 501                 create_counter(counter, cpu, pid);
 502
 503         nr_cpu++;
 504 }
 505
 506 static void atexit_header(void)
 507 {
 508         header->data_size += bytes_written;
 509
 510         perf_header__write(header, output);
 511 }
 512
 513 static int __cmd_record(int argc, const char **argv)
 514 {
 515         int i, counter;
 516         struct stat st;
 517         pid_t pid = 0;
 518         int flags;
 519         int ret;
 520
 521         page_size = sysconf(_SC_PAGE_SIZE);
 522         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 523         assert(nr_cpus <= MAX_NR_CPUS);
 524         assert(nr_cpus >= 0);
 525
 526         atexit(sig_atexit);
 527         signal(SIGCHLD, sig_handler);
 528         signal(SIGINT, sig_handler);
 529
 530         if (!stat(output_name, &st) && st.st_size) {
 531                 if (!force && !append_file) {
 532                         fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 533                                         output_name);
 534                         exit(-1);
 535                 }
 536         } else {
 537                 append_file = 0;
 538         }
 539
 540         flags = O_CREAT|O_RDWR;
 541         if (append_file)
 542                 file_new = 0;
 543         else
 544                 flags |= O_TRUNC;
 545
 546         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 547         if (output < 0) {
 548                 perror("failed to create output file");
 549                 exit(-1);
 550         }
 551
 552         if (!file_new)
 553                 header = perf_header__read(output);
 554         else
 555                 header = perf_header__new();
 556
 557
 558         if (raw_samples) {
 559                 read_tracing_data(attrs, nr_counters);
 560         } else {
 561                 for (i = 0; i < nr_counters; i++) {
 562                         if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
 563                                 read_tracing_data(attrs, nr_counters);
 564                                 break;
 565                         }
 566                 }
 567         }
 568         atexit(atexit_header);
 569
 570         if (!system_wide) {
 571                 pid = target_pid;
 572                 if (pid == -1)
 573                         pid = getpid();
 574
 575                 open_counters(profile_cpu, pid);
 576         } else {
 577                 if (profile_cpu != -1) {
 578                         open_counters(profile_cpu, target_pid);
 579                 } else {
 580                         for (i = 0; i < nr_cpus; i++)
 581                                 open_counters(i, target_pid);
 582                 }
 583         }
 584
 585         if (file_new)
 586                 perf_header__write(header, output);
 587
 588         if (!system_wide) {
 589                 pid_t tgid = pid_synthesize_comm_event(pid, 0);
 590                 pid_synthesize_mmap_samples(pid, tgid);
 591         } else
 592                 synthesize_all();
 593
 594         if (target_pid == -1 && argc) {
 595                 pid = fork();
 596                 if (pid < 0)
 597                         perror("failed to fork");
 598
 599                 if (!pid) {
 600                         if (execvp(argv[0], (char **)argv)) {
 601                                 perror(argv[0]);
 602                                 exit(-1);
 603                         }
 604                 }
 605         }
 606
 607         if (realtime_prio) {
 608                 struct sched_param param;
 609
 610                 param.sched_priority = realtime_prio;
 611                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 612                         printf("Could not set realtime priority.\n");
 613                         exit(-1);
 614                 }
 615         }
 616
 617         for (;;) {
 618                 int hits = samples;
 619
 620                 for (i = 0; i < nr_cpu; i++) {
 621                         for (counter = 0; counter < nr_counters; counter++)
 622                                 mmap_read(&mmap_array[i][counter]);
 623                 }
 624
 625                 if (hits == samples) {
 626                         if (done)
 627                                 break;
 628                         ret = poll(event_array, nr_poll, 100);
 629                 }
 630         }
 631
 632         /*
 633          * Approximate RIP event size: 24 bytes.
 634          */
 635         fprintf(stderr,
 636                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 637                 (double)bytes_written / 1024.0 / 1024.0,
 638                 output_name,
 639                 bytes_written / 24);
 640
 641         return 0;
 642 }
 643
 644 static const char * const record_usage[] = {
 645         "perf record [<options>] [<command>]",
 646         "perf record [<options>] -- <command> [<options>]",
 647         NULL
 648 };
 649
 650 static const struct option options[] = {
 651         OPT_CALLBACK('e', "event", NULL, "event",
 652                      "event selector. use 'perf list' to list available events",
 653                      parse_events),
 654         OPT_INTEGER('p', "pid", &target_pid,
 655                     "record events on existing pid"),
 656         OPT_INTEGER('r', "realtime", &realtime_prio,
 657                     "collect data with this RT SCHED_FIFO priority"),
 658         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 659                     "collect raw sample records from all opened counters"),
 660         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 661                             "system-wide collection from all CPUs"),
 662         OPT_BOOLEAN('A', "append", &append_file,
 663                             "append to the output file to do incremental profiling"),
 664         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
 665                             "CPU to profile on"),
 666         OPT_BOOLEAN('f', "force", &force,
 667                         "overwrite existing data file"),
 668         OPT_LONG('c', "count", &default_interval,
 669                     "event period to sample"),
 670         OPT_STRING('o', "output", &output_name, "file",
 671                     "output file name"),
 672         OPT_BOOLEAN('i', "inherit", &inherit,
 673                     "child tasks inherit counters"),
 674         OPT_INTEGER('F', "freq", &freq,
 675                     "profile at this frequency"),
 676         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 677                     "number of mmap data pages"),
 678         OPT_BOOLEAN('g', "call-graph", &call_graph,
 679                     "do call-graph (stack chain/backtrace) recording"),
 680         OPT_BOOLEAN('v', "verbose", &verbose,
 681                     "be more verbose (show counter open errors, etc)"),
 682         OPT_BOOLEAN('s', "stat", &inherit_stat,
 683                     "per thread counts"),
 684         OPT_BOOLEAN('d', "data", &sample_address,
 685                     "Sample addresses"),
 686         OPT_BOOLEAN('n', "no-samples", &no_samples,
 687                     "don't sample"),
 688         OPT_BOOLEAN('M', "multiplex", &multiplex,
 689                     "multiplex counter output in a single channel"),
 690         OPT_END()
 691 };
 692
 693 int cmd_record(int argc, const char **argv, const char *prefix __used)
 694 {
 695         int counter;
 696
 697         argc = parse_options(argc, argv, options, record_usage,
 698                 PARSE_OPT_STOP_AT_NON_OPTION);
 699         if (!argc && target_pid == -1 && !system_wide)
 700                 usage_with_options(record_usage, options);
 701
 702         if (!nr_counters) {
 703                 nr_counters     = 1;
 704                 attrs[0].type   = PERF_TYPE_HARDWARE;
 705                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 706         }
 707
 708         for (counter = 0; counter < nr_counters; counter++) {
 709                 if (attrs[counter].sample_period)
 710                         continue;
 711
 712                 attrs[counter].sample_period = default_interval;
 713         }
 714
 715         return __cmd_record(argc, argv);
 716 }