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/session.h"
  21 #include "util/symbol.h"
  22
  23 #include <unistd.h>
  24 #include <sched.h>
  25
  26 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  27
  28 static long                     default_interval                =      0;
  29
  30 static int                      nr_cpus                         =      0;
  31 static unsigned int             page_size;
  32 static unsigned int             mmap_pages                      =    128;
  33 static int                      freq                            =   1000;
  34 static int                      output;
  35 static const char               *output_name                    = "perf.data";
  36 static int                      group                           =      0;
  37 static unsigned int             realtime_prio                   =      0;
  38 static int                      raw_samples                     =      0;
  39 static int                      system_wide                     =      0;
  40 static int                      profile_cpu                     =     -1;
  41 static pid_t                    target_pid                      =     -1;
  42 static pid_t                    child_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 static int                      multiplex                       =      0;
  51 static int                      multiplex_fd                    =     -1;
  52
  53 static long                     samples                         =      0;
  54 static struct timeval           last_read;
  55 static struct timeval           this_read;
  56
  57 static u64                      bytes_written                   =      0;
  58
  59 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
  60
  61 static int                      nr_poll                         =      0;
  62 static int                      nr_cpu                          =      0;
  63
  64 static int                      file_new                        =      1;
  65
  66 static struct perf_session      *session;
  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_event_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_event_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 write_event(event_t *buf, size_t size)
 115 {
 116         /*
 117         * Add it to the list of DSOs, so that when we finish this
 118          * record session we can pick the available build-ids.
 119          */
 120         if (buf->header.type == PERF_RECORD_MMAP)
 121                 dsos__findnew(buf->mmap.filename);
 122
 123         write_output(buf, size);
 124 }
 125
 126 static int process_synthesized_event(event_t *event)
 127 {
 128         write_event(event, event->header.size);
 129         return 0;
 130 }
 131
 132 static void mmap_read(struct mmap_data *md)
 133 {
 134         unsigned int head = mmap_read_head(md);
 135         unsigned int old = md->prev;
 136         unsigned char *data = md->base + page_size;
 137         unsigned long size;
 138         void *buf;
 139         int diff;
 140
 141         gettimeofday(&this_read, NULL);
 142
 143         /*
 144          * If we're further behind than half the buffer, there's a chance
 145          * the writer will bite our tail and mess up the samples under us.
 146          *
 147          * If we somehow ended up ahead of the head, we got messed up.
 148          *
 149          * In either case, truncate and restart at head.
 150          */
 151         diff = head - old;
 152         if (diff < 0) {
 153                 struct timeval iv;
 154                 unsigned long msecs;
 155
 156                 timersub(&this_read, &last_read, &iv);
 157                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 158
 159                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 160                                 "  Last read %lu msecs ago.\n", msecs);
 161
 162                 /*
 163                  * head points to a known good entry, start there.
 164                  */
 165                 old = head;
 166         }
 167
 168         last_read = this_read;
 169
 170         if (old != head)
 171                 samples++;
 172
 173         size = head - old;
 174
 175         if ((old & md->mask) + size != (head & md->mask)) {
 176                 buf = &data[old & md->mask];
 177                 size = md->mask + 1 - (old & md->mask);
 178                 old += size;
 179
 180                 write_event(buf, size);
 181         }
 182
 183         buf = &data[old & md->mask];
 184         size = head - old;
 185         old += size;
 186
 187         write_event(buf, size);
 188
 189         md->prev = old;
 190         mmap_write_tail(md, old);
 191 }
 192
 193 static volatile int done = 0;
 194 static volatile int signr = -1;
 195
 196 static void sig_handler(int sig)
 197 {
 198         done = 1;
 199         signr = sig;
 200 }
 201
 202 static void sig_atexit(void)
 203 {
 204         if (child_pid != -1)
 205                 kill(child_pid, SIGTERM);
 206
 207         if (signr == -1)
 208                 return;
 209
 210         signal(signr, SIG_DFL);
 211         kill(getpid(), signr);
 212 }
 213
 214 static int group_fd;
 215
 216 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 217 {
 218         struct perf_header_attr *h_attr;
 219
 220         if (nr < session->header.attrs) {
 221                 h_attr = session->header.attr[nr];
 222         } else {
 223                 h_attr = perf_header_attr__new(a);
 224                 if (h_attr != NULL)
 225                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 226                                 perf_header_attr__delete(h_attr);
 227                                 h_attr = NULL;
 228                         }
 229         }
 230
 231         return h_attr;
 232 }
 233
 234 static void create_counter(int counter, int cpu, pid_t pid)
 235 {
 236         char *filter = filters[counter];
 237         struct perf_event_attr *attr = attrs + counter;
 238         struct perf_header_attr *h_attr;
 239         int track = !counter; /* only the first counter needs these */
 240         int ret;
 241         struct {
 242                 u64 count;
 243                 u64 time_enabled;
 244                 u64 time_running;
 245                 u64 id;
 246         } read_data;
 247
 248         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 249                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 250                                   PERF_FORMAT_ID;
 251
 252         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 253
 254         if (freq) {
 255                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 256                 attr->freq              = 1;
 257                 attr->sample_freq       = freq;
 258         }
 259
 260         if (no_samples)
 261                 attr->sample_freq = 0;
 262
 263         if (inherit_stat)
 264                 attr->inherit_stat = 1;
 265
 266         if (sample_address)
 267                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 268
 269         if (call_graph)
 270                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 271
 272         if (raw_samples) {
 273                 attr->sample_type       |= PERF_SAMPLE_TIME;
 274                 attr->sample_type       |= PERF_SAMPLE_RAW;
 275                 attr->sample_type       |= PERF_SAMPLE_CPU;
 276         }
 277
 278         attr->mmap              = track;
 279         attr->comm              = track;
 280         attr->inherit           = (cpu < 0) && inherit;
 281         attr->disabled          = 1;
 282
 283 try_again:
 284         fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
 285
 286         if (fd[nr_cpu][counter] < 0) {
 287                 int err = errno;
 288
 289                 if (err == EPERM || err == EACCES)
 290                         die("Permission error - are you root?\n");
 291                 else if (err ==  ENODEV && profile_cpu != -1)
 292                         die("No such device - did you specify an out-of-range profile CPU?\n");
 293
 294                 /*
 295                  * If it's cycles then fall back to hrtimer
 296                  * based cpu-clock-tick sw counter, which
 297                  * is always available even if no PMU support:
 298                  */
 299                 if (attr->type == PERF_TYPE_HARDWARE
 300                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 301
 302                         if (verbose)
 303                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 304                         attr->type = PERF_TYPE_SOFTWARE;
 305                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 306                         goto try_again;
 307                 }
 308                 printf("\n");
 309                 error("perfcounter syscall returned with %d (%s)\n",
 310                         fd[nr_cpu][counter], strerror(err));
 311
 312 #if defined(__i386__) || defined(__x86_64__)
 313                 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 314                         die("No hardware sampling interrupt available. No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.\n");
 315 #endif
 316
 317                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 318                 exit(-1);
 319         }
 320
 321         h_attr = get_header_attr(attr, counter);
 322         if (h_attr == NULL)
 323                 die("nomem\n");
 324
 325         if (!file_new) {
 326                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 327                         fprintf(stderr, "incompatible append\n");
 328                         exit(-1);
 329                 }
 330         }
 331
 332         if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
 333                 perror("Unable to read perf file descriptor\n");
 334                 exit(-1);
 335         }
 336
 337         if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 338                 pr_warning("Not enough memory to add id\n");
 339                 exit(-1);
 340         }
 341
 342         assert(fd[nr_cpu][counter] >= 0);
 343         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 344
 345         /*
 346          * First counter acts as the group leader:
 347          */
 348         if (group && group_fd == -1)
 349                 group_fd = fd[nr_cpu][counter];
 350         if (multiplex && multiplex_fd == -1)
 351                 multiplex_fd = fd[nr_cpu][counter];
 352
 353         if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
 354
 355                 ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
 356                 assert(ret != -1);
 357         } else {
 358                 event_array[nr_poll].fd = fd[nr_cpu][counter];
 359                 event_array[nr_poll].events = POLLIN;
 360                 nr_poll++;
 361
 362                 mmap_array[nr_cpu][counter].counter = counter;
 363                 mmap_array[nr_cpu][counter].prev = 0;
 364                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 365                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 366                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
 367                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 368                         error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 369                         exit(-1);
 370                 }
 371         }
 372
 373         if (filter != NULL) {
 374                 ret = ioctl(fd[nr_cpu][counter],
 375                             PERF_EVENT_IOC_SET_FILTER, filter);
 376                 if (ret) {
 377                         error("failed to set filter with %d (%s)\n", errno,
 378                               strerror(errno));
 379                         exit(-1);
 380                 }
 381         }
 382
 383         ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
 384 }
 385
 386 static void open_counters(int cpu, pid_t pid)
 387 {
 388         int counter;
 389
 390         group_fd = -1;
 391         for (counter = 0; counter < nr_counters; counter++)
 392                 create_counter(counter, cpu, pid);
 393
 394         nr_cpu++;
 395 }
 396
 397 static void atexit_header(void)
 398 {
 399         session->header.data_size += bytes_written;
 400
 401         perf_header__write(&session->header, output, true);
 402 }
 403
 404 static int __cmd_record(int argc, const char **argv)
 405 {
 406         int i, counter;
 407         struct stat st;
 408         pid_t pid = 0;
 409         int flags;
 410         int err;
 411         unsigned long waking = 0;
 412
 413         page_size = sysconf(_SC_PAGE_SIZE);
 414         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 415         assert(nr_cpus <= MAX_NR_CPUS);
 416         assert(nr_cpus >= 0);
 417
 418         atexit(sig_atexit);
 419         signal(SIGCHLD, sig_handler);
 420         signal(SIGINT, sig_handler);
 421
 422         if (!stat(output_name, &st) && st.st_size) {
 423                 if (!force && !append_file) {
 424                         fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 425                                         output_name);
 426                         exit(-1);
 427                 }
 428         } else {
 429                 append_file = 0;
 430         }
 431
 432         flags = O_CREAT|O_RDWR;
 433         if (append_file)
 434                 file_new = 0;
 435         else
 436                 flags |= O_TRUNC;
 437
 438         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 439         if (output < 0) {
 440                 perror("failed to create output file");
 441                 exit(-1);
 442         }
 443
 444         session = perf_session__new(output_name, O_WRONLY, force);
 445         if (session == NULL) {
 446                 pr_err("Not enough memory for reading perf file header\n");
 447                 return -1;
 448         }
 449
 450         if (!file_new) {
 451                 err = perf_header__read(&session->header, output);
 452                 if (err < 0)
 453                         return err;
 454         }
 455
 456         if (raw_samples) {
 457                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 458         } else {
 459                 for (i = 0; i < nr_counters; i++) {
 460                         if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
 461                                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 462                                 break;
 463                         }
 464                 }
 465         }
 466
 467         atexit(atexit_header);
 468
 469         if (!system_wide) {
 470                 pid = target_pid;
 471                 if (pid == -1)
 472                         pid = getpid();
 473
 474                 open_counters(profile_cpu, pid);
 475         } else {
 476                 if (profile_cpu != -1) {
 477                         open_counters(profile_cpu, target_pid);
 478                 } else {
 479                         for (i = 0; i < nr_cpus; i++)
 480                                 open_counters(i, target_pid);
 481                 }
 482         }
 483
 484         if (file_new) {
 485                 err = perf_header__write(&session->header, output, false);
 486                 if (err < 0)
 487                         return err;
 488         }
 489
 490         if (!system_wide)
 491                 event__synthesize_thread(pid, process_synthesized_event);
 492         else
 493                 event__synthesize_threads(process_synthesized_event);
 494
 495         if (target_pid == -1 && argc) {
 496                 pid = fork();
 497                 if (pid < 0)
 498                         die("failed to fork");
 499
 500                 if (!pid) {
 501                         if (execvp(argv[0], (char **)argv)) {
 502                                 perror(argv[0]);
 503                                 exit(-1);
 504                         }
 505                 } else {
 506                         /*
 507                          * Wait a bit for the execv'ed child to appear
 508                          * and be updated in /proc
 509                          * FIXME: Do you know a less heuristical solution?
 510                          */
 511                         usleep(1000);
 512                         event__synthesize_thread(pid,
 513                                                  process_synthesized_event);
 514                 }
 515
 516                 child_pid = pid;
 517         }
 518
 519         if (realtime_prio) {
 520                 struct sched_param param;
 521
 522                 param.sched_priority = realtime_prio;
 523                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 524                         pr_err("Could not set realtime priority.\n");
 525                         exit(-1);
 526                 }
 527         }
 528
 529         for (;;) {
 530                 int hits = samples;
 531
 532                 for (i = 0; i < nr_cpu; i++) {
 533                         for (counter = 0; counter < nr_counters; counter++) {
 534                                 if (mmap_array[i][counter].base)
 535                                         mmap_read(&mmap_array[i][counter]);
 536                         }
 537                 }
 538
 539                 if (hits == samples) {
 540                         if (done)
 541                                 break;
 542                         err = poll(event_array, nr_poll, -1);
 543                         waking++;
 544                 }
 545
 546                 if (done) {
 547                         for (i = 0; i < nr_cpu; i++) {
 548                                 for (counter = 0; counter < nr_counters; counter++)
 549                                         ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
 550                         }
 551                 }
 552         }
 553
 554         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 555
 556         /*
 557          * Approximate RIP event size: 24 bytes.
 558          */
 559         fprintf(stderr,
 560                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 561                 (double)bytes_written / 1024.0 / 1024.0,
 562                 output_name,
 563                 bytes_written / 24);
 564
 565         return 0;
 566 }
 567
 568 static const char * const record_usage[] = {
 569         "perf record [<options>] [<command>]",
 570         "perf record [<options>] -- <command> [<options>]",
 571         NULL
 572 };
 573
 574 static const struct option options[] = {
 575         OPT_CALLBACK('e', "event", NULL, "event",
 576                      "event selector. use 'perf list' to list available events",
 577                      parse_events),
 578         OPT_CALLBACK(0, "filter", NULL, "filter",
 579                      "event filter", parse_filter),
 580         OPT_INTEGER('p', "pid", &target_pid,
 581                     "record events on existing pid"),
 582         OPT_INTEGER('r', "realtime", &realtime_prio,
 583                     "collect data with this RT SCHED_FIFO priority"),
 584         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 585                     "collect raw sample records from all opened counters"),
 586         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 587                             "system-wide collection from all CPUs"),
 588         OPT_BOOLEAN('A', "append", &append_file,
 589                             "append to the output file to do incremental profiling"),
 590         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
 591                             "CPU to profile on"),
 592         OPT_BOOLEAN('f', "force", &force,
 593                         "overwrite existing data file"),
 594         OPT_LONG('c', "count", &default_interval,
 595                     "event period to sample"),
 596         OPT_STRING('o', "output", &output_name, "file",
 597                     "output file name"),
 598         OPT_BOOLEAN('i', "inherit", &inherit,
 599                     "child tasks inherit counters"),
 600         OPT_INTEGER('F', "freq", &freq,
 601                     "profile at this frequency"),
 602         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 603                     "number of mmap data pages"),
 604         OPT_BOOLEAN('g', "call-graph", &call_graph,
 605                     "do call-graph (stack chain/backtrace) recording"),
 606         OPT_BOOLEAN('v', "verbose", &verbose,
 607                     "be more verbose (show counter open errors, etc)"),
 608         OPT_BOOLEAN('s', "stat", &inherit_stat,
 609                     "per thread counts"),
 610         OPT_BOOLEAN('d', "data", &sample_address,
 611                     "Sample addresses"),
 612         OPT_BOOLEAN('n', "no-samples", &no_samples,
 613                     "don't sample"),
 614         OPT_BOOLEAN('M', "multiplex", &multiplex,
 615                     "multiplex counter output in a single channel"),
 616         OPT_END()
 617 };
 618
 619 int cmd_record(int argc, const char **argv, const char *prefix __used)
 620 {
 621         int counter;
 622
 623         symbol__init(0);
 624
 625         argc = parse_options(argc, argv, options, record_usage,
 626                 PARSE_OPT_STOP_AT_NON_OPTION);
 627         if (!argc && target_pid == -1 && !system_wide)
 628                 usage_with_options(record_usage, options);
 629
 630         if (!nr_counters) {
 631                 nr_counters     = 1;
 632                 attrs[0].type   = PERF_TYPE_HARDWARE;
 633                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 634         }
 635
 636         /*
 637          * User specified count overrides default frequency.
 638          */
 639         if (default_interval)
 640                 freq = 0;
 641         else if (freq) {
 642                 default_interval = freq;
 643         } else {
 644                 fprintf(stderr, "frequency and count are zero, aborting\n");
 645                 exit(EXIT_FAILURE);
 646         }
 647
 648         for (counter = 0; counter < nr_counters; counter++) {
 649                 if (attrs[counter].sample_period)
 650                         continue;
 651
 652                 attrs[counter].sample_period = default_interval;
 653         }
 654
 655         return __cmd_record(argc, argv);
 656 }