tools/perf/builtin-record.c

   1 /*
   2  * builtin-record.c
   3  *
   4  * Builtin record command: Record the profile of a workload
   5  * (or a CPU, or a PID) into the perf.data output file - for
   6  * later analysis via perf report.
   7  */
   8 #include "builtin.h"
   9
  10 #include "perf.h"
  11
  12 #include "util/util.h"
  13 #include "util/parse-options.h"
  14 #include "util/parse-events.h"
  15 #include "util/string.h"
  16
  17 #include "util/header.h"
  18
  19 #include <unistd.h>
  20 #include <sched.h>
  21
  22 #define ALIGN(x, a)             __ALIGN_MASK(x, (typeof(x))(a)-1)
  23 #define __ALIGN_MASK(x, mask)   (((x)+(mask))&~(mask))
  24
  25 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  26
  27 static long                     default_interval                = 100000;
  28
  29 static int                      nr_cpus                         = 0;
  30 static unsigned int             page_size;
  31 static unsigned int             mmap_pages                      = 128;
  32 static int                      freq                            = 0;
  33 static int                      output;
  34 static const char               *output_name                    = "perf.data";
  35 static int                      group                           = 0;
  36 static unsigned int             realtime_prio                   = 0;
  37 static int                      system_wide                     = 0;
  38 static pid_t                    target_pid                      = -1;
  39 static int                      inherit                         = 1;
  40 static int                      force                           = 0;
  41 static int                      append_file                     = 0;
  42 static int                      call_graph                      = 0;
  43 static int                      verbose                         = 0;
  44 static int                      inherit_stat                    = 0;
  45 static int                      no_samples                      = 0;
  46
  47 static long                     samples;
  48 static struct timeval           last_read;
  49 static struct timeval           this_read;
  50
  51 static u64                      bytes_written;
  52
  53 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
  54
  55 static int                      nr_poll;
  56 static int                      nr_cpu;
  57
  58 static int                      file_new = 1;
  59
  60 struct perf_header              *header;
  61
  62 struct mmap_event {
  63         struct perf_event_header        header;
  64         u32                             pid;
  65         u32                             tid;
  66         u64                             start;
  67         u64                             len;
  68         u64                             pgoff;
  69         char                            filename[PATH_MAX];
  70 };
  71
  72 struct comm_event {
  73         struct perf_event_header        header;
  74         u32                             pid;
  75         u32                             tid;
  76         char                            comm[16];
  77 };
  78
  79
  80 struct mmap_data {
  81         int                     counter;
  82         void                    *base;
  83         unsigned int            mask;
  84         unsigned int            prev;
  85 };
  86
  87 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  88
  89 static unsigned long mmap_read_head(struct mmap_data *md)
  90 {
  91         struct perf_counter_mmap_page *pc = md->base;
  92         long head;
  93
  94         head = pc->data_head;
  95         rmb();
  96
  97         return head;
  98 }
  99
 100 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
 101 {
 102         struct perf_counter_mmap_page *pc = md->base;
 103
 104         /*
 105          * ensure all reads are done before we write the tail out.
 106          */
 107         /* mb(); */
 108         pc->data_tail = tail;
 109 }
 110
 111 static void write_output(void *buf, size_t size)
 112 {
 113         while (size) {
 114                 int ret = write(output, buf, size);
 115
 116                 if (ret < 0)
 117                         die("failed to write");
 118
 119                 size -= ret;
 120                 buf += ret;
 121
 122                 bytes_written += ret;
 123         }
 124 }
 125
 126 static void mmap_read(struct mmap_data *md)
 127 {
 128         unsigned int head = mmap_read_head(md);
 129         unsigned int old = md->prev;
 130         unsigned char *data = md->base + page_size;
 131         unsigned long size;
 132         void *buf;
 133         int diff;
 134
 135         gettimeofday(&this_read, NULL);
 136
 137         /*
 138          * If we're further behind than half the buffer, there's a chance
 139          * the writer will bite our tail and mess up the samples under us.
 140          *
 141          * If we somehow ended up ahead of the head, we got messed up.
 142          *
 143          * In either case, truncate and restart at head.
 144          */
 145         diff = head - old;
 146         if (diff < 0) {
 147                 struct timeval iv;
 148                 unsigned long msecs;
 149
 150                 timersub(&this_read, &last_read, &iv);
 151                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 152
 153                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 154                                 "  Last read %lu msecs ago.\n", msecs);
 155
 156                 /*
 157                  * head points to a known good entry, start there.
 158                  */
 159                 old = head;
 160         }
 161
 162         last_read = this_read;
 163
 164         if (old != head)
 165                 samples++;
 166
 167         size = head - old;
 168
 169         if ((old & md->mask) + size != (head & md->mask)) {
 170                 buf = &data[old & md->mask];
 171                 size = md->mask + 1 - (old & md->mask);
 172                 old += size;
 173
 174                 write_output(buf, size);
 175         }
 176
 177         buf = &data[old & md->mask];
 178         size = head - old;
 179         old += size;
 180
 181         write_output(buf, size);
 182
 183         md->prev = old;
 184         mmap_write_tail(md, old);
 185 }
 186
 187 static volatile int done = 0;
 188 static volatile int signr = -1;
 189
 190 static void sig_handler(int sig)
 191 {
 192         done = 1;
 193         signr = sig;
 194 }
 195
 196 static void sig_atexit(void)
 197 {
 198         if (signr == -1)
 199                 return;
 200
 201         signal(signr, SIG_DFL);
 202         kill(getpid(), signr);
 203 }
 204
 205 static void pid_synthesize_comm_event(pid_t pid, int full)
 206 {
 207         struct comm_event comm_ev;
 208         char filename[PATH_MAX];
 209         char bf[BUFSIZ];
 210         int fd;
 211         size_t size;
 212         char *field, *sep;
 213         DIR *tasks;
 214         struct dirent dirent, *next;
 215
 216         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 217
 218         fd = open(filename, O_RDONLY);
 219         if (fd < 0) {
 220                 /*
 221                  * We raced with a task exiting - just return:
 222                  */
 223                 if (verbose)
 224                         fprintf(stderr, "couldn't open %s\n", filename);
 225                 return;
 226         }
 227         if (read(fd, bf, sizeof(bf)) < 0) {
 228                 fprintf(stderr, "couldn't read %s\n", filename);
 229                 exit(EXIT_FAILURE);
 230         }
 231         close(fd);
 232
 233         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 234         memset(&comm_ev, 0, sizeof(comm_ev));
 235         field = strchr(bf, '(');
 236         if (field == NULL)
 237                 goto out_failure;
 238         sep = strchr(++field, ')');
 239         if (sep == NULL)
 240                 goto out_failure;
 241         size = sep - field;
 242         memcpy(comm_ev.comm, field, size++);
 243
 244         comm_ev.pid = pid;
 245         comm_ev.header.type = PERF_EVENT_COMM;
 246         size = ALIGN(size, sizeof(u64));
 247         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 248
 249         if (!full) {
 250                 comm_ev.tid = pid;
 251
 252                 write_output(&comm_ev, comm_ev.header.size);
 253                 return;
 254         }
 255
 256         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 257
 258         tasks = opendir(filename);
 259         while (!readdir_r(tasks, &dirent, &next) && next) {
 260                 char *end;
 261                 pid = strtol(dirent.d_name, &end, 10);
 262                 if (*end)
 263                         continue;
 264
 265                 comm_ev.tid = pid;
 266
 267                 write_output(&comm_ev, comm_ev.header.size);
 268         }
 269         closedir(tasks);
 270         return;
 271
 272 out_failure:
 273         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 274                 filename);
 275         exit(EXIT_FAILURE);
 276 }
 277
 278 static void pid_synthesize_mmap_samples(pid_t pid)
 279 {
 280         char filename[PATH_MAX];
 281         FILE *fp;
 282
 283         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 284
 285         fp = fopen(filename, "r");
 286         if (fp == NULL) {
 287                 /*
 288                  * We raced with a task exiting - just return:
 289                  */
 290                 if (verbose)
 291                         fprintf(stderr, "couldn't open %s\n", filename);
 292                 return;
 293         }
 294         while (1) {
 295                 char bf[BUFSIZ], *pbf = bf;
 296                 struct mmap_event mmap_ev = {
 297                         .header.type = PERF_EVENT_MMAP,
 298                 };
 299                 int n;
 300                 size_t size;
 301                 if (fgets(bf, sizeof(bf), fp) == NULL)
 302                         break;
 303
 304                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 305                 n = hex2u64(pbf, &mmap_ev.start);
 306                 if (n < 0)
 307                         continue;
 308                 pbf += n + 1;
 309                 n = hex2u64(pbf, &mmap_ev.len);
 310                 if (n < 0)
 311                         continue;
 312                 pbf += n + 3;
 313                 if (*pbf == 'x') { /* vm_exec */
 314                         char *execname = strchr(bf, '/');
 315
 316                         if (execname == NULL)
 317                                 continue;
 318
 319                         size = strlen(execname);
 320                         execname[size - 1] = '\0'; /* Remove \n */
 321                         memcpy(mmap_ev.filename, execname, size);
 322                         size = ALIGN(size, sizeof(u64));
 323                         mmap_ev.len -= mmap_ev.start;
 324                         mmap_ev.header.size = (sizeof(mmap_ev) -
 325                                                (sizeof(mmap_ev.filename) - size));
 326                         mmap_ev.pid = pid;
 327                         mmap_ev.tid = pid;
 328
 329                         write_output(&mmap_ev, mmap_ev.header.size);
 330                 }
 331         }
 332
 333         fclose(fp);
 334 }
 335
 336 static void synthesize_all(void)
 337 {
 338         DIR *proc;
 339         struct dirent dirent, *next;
 340
 341         proc = opendir("/proc");
 342
 343         while (!readdir_r(proc, &dirent, &next) && next) {
 344                 char *end;
 345                 pid_t pid;
 346
 347                 pid = strtol(dirent.d_name, &end, 10);
 348                 if (*end) /* only interested in proper numerical dirents */
 349                         continue;
 350
 351                 pid_synthesize_comm_event(pid, 1);
 352                 pid_synthesize_mmap_samples(pid);
 353         }
 354
 355         closedir(proc);
 356 }
 357
 358 static int group_fd;
 359
 360 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
 361 {
 362         struct perf_header_attr *h_attr;
 363
 364         if (nr < header->attrs) {
 365                 h_attr = header->attr[nr];
 366         } else {
 367                 h_attr = perf_header_attr__new(a);
 368                 perf_header__add_attr(header, h_attr);
 369         }
 370
 371         return h_attr;
 372 }
 373
 374 static void create_counter(int counter, int cpu, pid_t pid)
 375 {
 376         struct perf_counter_attr *attr = attrs + counter;
 377         struct perf_header_attr *h_attr;
 378         int track = !counter; /* only the first counter needs these */
 379         struct {
 380                 u64 count;
 381                 u64 time_enabled;
 382                 u64 time_running;
 383                 u64 id;
 384         } read_data;
 385
 386         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 387                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 388                                   PERF_FORMAT_ID;
 389
 390         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 391
 392         if (freq) {
 393                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 394                 attr->freq              = 1;
 395                 attr->sample_freq       = freq;
 396         }
 397
 398         if (no_samples)
 399                 attr->sample_freq = 0;
 400
 401         if (inherit_stat)
 402                 attr->inherit_stat = 1;
 403
 404         if (call_graph)
 405                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 406
 407         attr->mmap              = track;
 408         attr->comm              = track;
 409         attr->inherit           = (cpu < 0) && inherit;
 410         attr->disabled          = 1;
 411
 412 try_again:
 413         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 414
 415         if (fd[nr_cpu][counter] < 0) {
 416                 int err = errno;
 417
 418                 if (err == EPERM)
 419                         die("Permission error - are you root?\n");
 420
 421                 /*
 422                  * If it's cycles then fall back to hrtimer
 423                  * based cpu-clock-tick sw counter, which
 424                  * is always available even if no PMU support:
 425                  */
 426                 if (attr->type == PERF_TYPE_HARDWARE
 427                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 428
 429                         if (verbose)
 430                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 431                         attr->type = PERF_TYPE_SOFTWARE;
 432                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 433                         goto try_again;
 434                 }
 435                 printf("\n");
 436                 error("perfcounter syscall returned with %d (%s)\n",
 437                         fd[nr_cpu][counter], strerror(err));
 438                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 439                 exit(-1);
 440         }
 441
 442         h_attr = get_header_attr(attr, counter);
 443
 444         if (!file_new) {
 445                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 446                         fprintf(stderr, "incompatible append\n");
 447                         exit(-1);
 448                 }
 449         }
 450
 451         read(fd[nr_cpu][counter], &read_data, sizeof(read_data));
 452
 453         perf_header_attr__add_id(h_attr, read_data.id);
 454
 455         assert(fd[nr_cpu][counter] >= 0);
 456         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 457
 458         /*
 459          * First counter acts as the group leader:
 460          */
 461         if (group && group_fd == -1)
 462                 group_fd = fd[nr_cpu][counter];
 463
 464         event_array[nr_poll].fd = fd[nr_cpu][counter];
 465         event_array[nr_poll].events = POLLIN;
 466         nr_poll++;
 467
 468         mmap_array[nr_cpu][counter].counter = counter;
 469         mmap_array[nr_cpu][counter].prev = 0;
 470         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 471         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 472                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
 473         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 474                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 475                 exit(-1);
 476         }
 477
 478         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
 479 }
 480
 481 static void open_counters(int cpu, pid_t pid)
 482 {
 483         int counter;
 484
 485         group_fd = -1;
 486         for (counter = 0; counter < nr_counters; counter++)
 487                 create_counter(counter, cpu, pid);
 488
 489         nr_cpu++;
 490 }
 491
 492 static void atexit_header(void)
 493 {
 494         header->data_size += bytes_written;
 495
 496         perf_header__write(header, output);
 497 }
 498
 499 static int __cmd_record(int argc, const char **argv)
 500 {
 501         int i, counter;
 502         struct stat st;
 503         pid_t pid = 0;
 504         int flags;
 505         int ret;
 506
 507         page_size = sysconf(_SC_PAGE_SIZE);
 508         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 509         assert(nr_cpus <= MAX_NR_CPUS);
 510         assert(nr_cpus >= 0);
 511
 512         atexit(sig_atexit);
 513         signal(SIGCHLD, sig_handler);
 514         signal(SIGINT, sig_handler);
 515
 516         if (!stat(output_name, &st) && !force && !append_file) {
 517                 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 518                                 output_name);
 519                 exit(-1);
 520         }
 521
 522         flags = O_CREAT|O_RDWR;
 523         if (append_file)
 524                 file_new = 0;
 525         else
 526                 flags |= O_TRUNC;
 527
 528         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 529         if (output < 0) {
 530                 perror("failed to create output file");
 531                 exit(-1);
 532         }
 533
 534         if (!file_new)
 535                 header = perf_header__read(output);
 536         else
 537                 header = perf_header__new();
 538
 539         atexit(atexit_header);
 540
 541         if (!system_wide) {
 542                 pid = target_pid;
 543                 if (pid == -1)
 544                         pid = getpid();
 545
 546                 open_counters(-1, pid);
 547         } else for (i = 0; i < nr_cpus; i++)
 548                 open_counters(i, target_pid);
 549
 550         if (file_new)
 551                 perf_header__write(header, output);
 552
 553         if (!system_wide) {
 554                 pid_synthesize_comm_event(pid, 0);
 555                 pid_synthesize_mmap_samples(pid);
 556         } else
 557                 synthesize_all();
 558
 559         if (target_pid == -1 && argc) {
 560                 pid = fork();
 561                 if (pid < 0)
 562                         perror("failed to fork");
 563
 564                 if (!pid) {
 565                         if (execvp(argv[0], (char **)argv)) {
 566                                 perror(argv[0]);
 567                                 exit(-1);
 568                         }
 569                 }
 570         }
 571
 572         if (realtime_prio) {
 573                 struct sched_param param;
 574
 575                 param.sched_priority = realtime_prio;
 576                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 577                         printf("Could not set realtime priority.\n");
 578                         exit(-1);
 579                 }
 580         }
 581
 582         for (;;) {
 583                 int hits = samples;
 584
 585                 for (i = 0; i < nr_cpu; i++) {
 586                         for (counter = 0; counter < nr_counters; counter++)
 587                                 mmap_read(&mmap_array[i][counter]);
 588                 }
 589
 590                 if (hits == samples) {
 591                         if (done)
 592                                 break;
 593                         ret = poll(event_array, nr_poll, 100);
 594                 }
 595         }
 596
 597         /*
 598          * Approximate RIP event size: 24 bytes.
 599          */
 600         fprintf(stderr,
 601                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 602                 (double)bytes_written / 1024.0 / 1024.0,
 603                 output_name,
 604                 bytes_written / 24);
 605
 606         return 0;
 607 }
 608
 609 static const char * const record_usage[] = {
 610         "perf record [<options>] [<command>]",
 611         "perf record [<options>] -- <command> [<options>]",
 612         NULL
 613 };
 614
 615 static const struct option options[] = {
 616         OPT_CALLBACK('e', "event", NULL, "event",
 617                      "event selector. use 'perf list' to list available events",
 618                      parse_events),
 619         OPT_INTEGER('p', "pid", &target_pid,
 620                     "record events on existing pid"),
 621         OPT_INTEGER('r', "realtime", &realtime_prio,
 622                     "collect data with this RT SCHED_FIFO priority"),
 623         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 624                             "system-wide collection from all CPUs"),
 625         OPT_BOOLEAN('A', "append", &append_file,
 626                             "append to the output file to do incremental profiling"),
 627         OPT_BOOLEAN('f', "force", &force,
 628                         "overwrite existing data file"),
 629         OPT_LONG('c', "count", &default_interval,
 630                     "event period to sample"),
 631         OPT_STRING('o', "output", &output_name, "file",
 632                     "output file name"),
 633         OPT_BOOLEAN('i', "inherit", &inherit,
 634                     "child tasks inherit counters"),
 635         OPT_INTEGER('F', "freq", &freq,
 636                     "profile at this frequency"),
 637         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 638                     "number of mmap data pages"),
 639         OPT_BOOLEAN('g', "call-graph", &call_graph,
 640                     "do call-graph (stack chain/backtrace) recording"),
 641         OPT_BOOLEAN('v', "verbose", &verbose,
 642                     "be more verbose (show counter open errors, etc)"),
 643         OPT_BOOLEAN('s', "stat", &inherit_stat,
 644                     "per thread counts"),
 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)
 651 {
 652         int counter;
 653
 654         argc = parse_options(argc, argv, options, record_usage, 0);
 655         if (!argc && target_pid == -1 && !system_wide)
 656                 usage_with_options(record_usage, options);
 657
 658         if (!nr_counters) {
 659                 nr_counters     = 1;
 660                 attrs[0].type   = PERF_TYPE_HARDWARE;
 661                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 662         }
 663
 664         for (counter = 0; counter < nr_counters; counter++) {
 665                 if (attrs[counter].sample_period)
 666                         continue;
 667
 668                 attrs[counter].sample_period = default_interval;
 669         }
 670
 671         return __cmd_record(argc, argv);
 672 }