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