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
  45 static long                     samples;
  46 static struct timeval           last_read;
  47 static struct timeval           this_read;
  48
  49 static u64                      bytes_written;
  50
  51 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
  52
  53 static int                      nr_poll;
  54 static int                      nr_cpu;
  55
  56 static int                      file_new = 1;
  57
  58 struct perf_header              *header;
  59
  60 struct mmap_event {
  61         struct perf_event_header        header;
  62         u32                             pid;
  63         u32                             tid;
  64         u64                             start;
  65         u64                             len;
  66         u64                             pgoff;
  67         char                            filename[PATH_MAX];
  68 };
  69
  70 struct comm_event {
  71         struct perf_event_header        header;
  72         u32                             pid;
  73         u32                             tid;
  74         char                            comm[16];
  75 };
  76
  77
  78 struct mmap_data {
  79         int                     counter;
  80         void                    *base;
  81         unsigned int            mask;
  82         unsigned int            prev;
  83 };
  84
  85 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  86
  87 static unsigned long mmap_read_head(struct mmap_data *md)
  88 {
  89         struct perf_counter_mmap_page *pc = md->base;
  90         long head;
  91
  92         head = pc->data_head;
  93         rmb();
  94
  95         return head;
  96 }
  97
  98 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
  99 {
 100         struct perf_counter_mmap_page *pc = md->base;
 101
 102         /*
 103          * ensure all reads are done before we write the tail out.
 104          */
 105         /* mb(); */
 106         pc->data_tail = tail;
 107 }
 108
 109 static void write_output(void *buf, size_t size)
 110 {
 111         while (size) {
 112                 int ret = write(output, buf, size);
 113
 114                 if (ret < 0)
 115                         die("failed to write");
 116
 117                 size -= ret;
 118                 buf += ret;
 119
 120                 bytes_written += ret;
 121         }
 122 }
 123
 124 static void mmap_read(struct mmap_data *md)
 125 {
 126         unsigned int head = mmap_read_head(md);
 127         unsigned int old = md->prev;
 128         unsigned char *data = md->base + page_size;
 129         unsigned long size;
 130         void *buf;
 131         int diff;
 132
 133         gettimeofday(&this_read, NULL);
 134
 135         /*
 136          * If we're further behind than half the buffer, there's a chance
 137          * the writer will bite our tail and mess up the samples under us.
 138          *
 139          * If we somehow ended up ahead of the head, we got messed up.
 140          *
 141          * In either case, truncate and restart at head.
 142          */
 143         diff = head - old;
 144         if (diff < 0) {
 145                 struct timeval iv;
 146                 unsigned long msecs;
 147
 148                 timersub(&this_read, &last_read, &iv);
 149                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 150
 151                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 152                                 "  Last read %lu msecs ago.\n", msecs);
 153
 154                 /*
 155                  * head points to a known good entry, start there.
 156                  */
 157                 old = head;
 158         }
 159
 160         last_read = this_read;
 161
 162         if (old != head)
 163                 samples++;
 164
 165         size = head - old;
 166
 167         if ((old & md->mask) + size != (head & md->mask)) {
 168                 buf = &data[old & md->mask];
 169                 size = md->mask + 1 - (old & md->mask);
 170                 old += size;
 171
 172                 write_output(buf, size);
 173         }
 174
 175         buf = &data[old & md->mask];
 176         size = head - old;
 177         old += size;
 178
 179         write_output(buf, size);
 180
 181         md->prev = old;
 182         mmap_write_tail(md, old);
 183 }
 184
 185 static volatile int done = 0;
 186 static volatile int signr = -1;
 187
 188 static void sig_handler(int sig)
 189 {
 190         done = 1;
 191         signr = sig;
 192 }
 193
 194 static void sig_atexit(void)
 195 {
 196         if (signr == -1)
 197                 return;
 198
 199         signal(signr, SIG_DFL);
 200         kill(getpid(), signr);
 201 }
 202
 203 static void pid_synthesize_comm_event(pid_t pid, int full)
 204 {
 205         struct comm_event comm_ev;
 206         char filename[PATH_MAX];
 207         char bf[BUFSIZ];
 208         int fd;
 209         size_t size;
 210         char *field, *sep;
 211         DIR *tasks;
 212         struct dirent dirent, *next;
 213
 214         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 215
 216         fd = open(filename, O_RDONLY);
 217         if (fd < 0) {
 218                 /*
 219                  * We raced with a task exiting - just return:
 220                  */
 221                 if (verbose)
 222                         fprintf(stderr, "couldn't open %s\n", filename);
 223                 return;
 224         }
 225         if (read(fd, bf, sizeof(bf)) < 0) {
 226                 fprintf(stderr, "couldn't read %s\n", filename);
 227                 exit(EXIT_FAILURE);
 228         }
 229         close(fd);
 230
 231         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 232         memset(&comm_ev, 0, sizeof(comm_ev));
 233         field = strchr(bf, '(');
 234         if (field == NULL)
 235                 goto out_failure;
 236         sep = strchr(++field, ')');
 237         if (sep == NULL)
 238                 goto out_failure;
 239         size = sep - field;
 240         memcpy(comm_ev.comm, field, size++);
 241
 242         comm_ev.pid = pid;
 243         comm_ev.header.type = PERF_EVENT_COMM;
 244         size = ALIGN(size, sizeof(u64));
 245         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 246
 247         if (!full) {
 248                 comm_ev.tid = pid;
 249
 250                 write_output(&comm_ev, comm_ev.header.size);
 251                 return;
 252         }
 253
 254         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 255
 256         tasks = opendir(filename);
 257         while (!readdir_r(tasks, &dirent, &next) && next) {
 258                 char *end;
 259                 pid = strtol(dirent.d_name, &end, 10);
 260                 if (*end)
 261                         continue;
 262
 263                 comm_ev.tid = pid;
 264
 265                 write_output(&comm_ev, comm_ev.header.size);
 266         }
 267         closedir(tasks);
 268         return;
 269
 270 out_failure:
 271         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 272                 filename);
 273         exit(EXIT_FAILURE);
 274 }
 275
 276 static void pid_synthesize_mmap_samples(pid_t pid)
 277 {
 278         char filename[PATH_MAX];
 279         FILE *fp;
 280
 281         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 282
 283         fp = fopen(filename, "r");
 284         if (fp == NULL) {
 285                 /*
 286                  * We raced with a task exiting - just return:
 287                  */
 288                 if (verbose)
 289                         fprintf(stderr, "couldn't open %s\n", filename);
 290                 return;
 291         }
 292         while (1) {
 293                 char bf[BUFSIZ], *pbf = bf;
 294                 struct mmap_event mmap_ev = {
 295                         .header.type = PERF_EVENT_MMAP,
 296                 };
 297                 int n;
 298                 size_t size;
 299                 if (fgets(bf, sizeof(bf), fp) == NULL)
 300                         break;
 301
 302                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 303                 n = hex2u64(pbf, &mmap_ev.start);
 304                 if (n < 0)
 305                         continue;
 306                 pbf += n + 1;
 307                 n = hex2u64(pbf, &mmap_ev.len);
 308                 if (n < 0)
 309                         continue;
 310                 pbf += n + 3;
 311                 if (*pbf == 'x') { /* vm_exec */
 312                         char *execname = strchr(bf, '/');
 313
 314                         if (execname == NULL)
 315                                 continue;
 316
 317                         size = strlen(execname);
 318                         execname[size - 1] = '\0'; /* Remove \n */
 319                         memcpy(mmap_ev.filename, execname, size);
 320                         size = ALIGN(size, sizeof(u64));
 321                         mmap_ev.len -= mmap_ev.start;
 322                         mmap_ev.header.size = (sizeof(mmap_ev) -
 323                                                (sizeof(mmap_ev.filename) - size));
 324                         mmap_ev.pid = pid;
 325                         mmap_ev.tid = pid;
 326
 327                         write_output(&mmap_ev, mmap_ev.header.size);
 328                 }
 329         }
 330
 331         fclose(fp);
 332 }
 333
 334 static void synthesize_all(void)
 335 {
 336         DIR *proc;
 337         struct dirent dirent, *next;
 338
 339         proc = opendir("/proc");
 340
 341         while (!readdir_r(proc, &dirent, &next) && next) {
 342                 char *end;
 343                 pid_t pid;
 344
 345                 pid = strtol(dirent.d_name, &end, 10);
 346                 if (*end) /* only interested in proper numerical dirents */
 347                         continue;
 348
 349                 pid_synthesize_comm_event(pid, 1);
 350                 pid_synthesize_mmap_samples(pid);
 351         }
 352
 353         closedir(proc);
 354 }
 355
 356 static int group_fd;
 357
 358 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
 359 {
 360         struct perf_header_attr *h_attr;
 361
 362         if (nr < header->attrs) {
 363                 h_attr = header->attr[nr];
 364         } else {
 365                 h_attr = perf_header_attr__new(a);
 366                 perf_header__add_attr(header, h_attr);
 367         }
 368
 369         return h_attr;
 370 }
 371
 372 static void create_counter(int counter, int cpu, pid_t pid)
 373 {
 374         struct perf_counter_attr *attr = attrs + counter;
 375         struct perf_header_attr *h_attr;
 376         int track = !counter; /* only the first counter needs these */
 377         struct {
 378                 u64 count;
 379                 u64 time_enabled;
 380                 u64 time_running;
 381                 u64 id;
 382         } read_data;
 383
 384         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 385                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 386                                   PERF_FORMAT_ID;
 387
 388         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 389
 390         if (freq) {
 391                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 392                 attr->freq              = 1;
 393                 attr->sample_freq       = freq;
 394         }
 395
 396         if (call_graph)
 397                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 398
 399         attr->mmap              = track;
 400         attr->comm              = track;
 401         attr->inherit           = (cpu < 0) && inherit;
 402         attr->disabled          = 1;
 403
 404 try_again:
 405         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 406
 407         if (fd[nr_cpu][counter] < 0) {
 408                 int err = errno;
 409
 410                 if (err == EPERM)
 411                         die("Permission error - are you root?\n");
 412
 413                 /*
 414                  * If it's cycles then fall back to hrtimer
 415                  * based cpu-clock-tick sw counter, which
 416                  * is always available even if no PMU support:
 417                  */
 418                 if (attr->type == PERF_TYPE_HARDWARE
 419                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 420
 421                         if (verbose)
 422                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 423                         attr->type = PERF_TYPE_SOFTWARE;
 424                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 425                         goto try_again;
 426                 }
 427                 printf("\n");
 428                 error("perfcounter syscall returned with %d (%s)\n",
 429                         fd[nr_cpu][counter], strerror(err));
 430                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 431                 exit(-1);
 432         }
 433
 434         h_attr = get_header_attr(attr, counter);
 435
 436         if (!file_new) {
 437                 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 438                         fprintf(stderr, "incompatible append\n");
 439                         exit(-1);
 440                 }
 441         }
 442
 443         read(fd[nr_cpu][counter], &read_data, sizeof(read_data));
 444
 445         perf_header_attr__add_id(h_attr, read_data.id);
 446
 447         assert(fd[nr_cpu][counter] >= 0);
 448         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 449
 450         /*
 451          * First counter acts as the group leader:
 452          */
 453         if (group && group_fd == -1)
 454                 group_fd = fd[nr_cpu][counter];
 455
 456         event_array[nr_poll].fd = fd[nr_cpu][counter];
 457         event_array[nr_poll].events = POLLIN;
 458         nr_poll++;
 459
 460         mmap_array[nr_cpu][counter].counter = counter;
 461         mmap_array[nr_cpu][counter].prev = 0;
 462         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 463         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 464                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
 465         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 466                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 467                 exit(-1);
 468         }
 469
 470         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
 471 }
 472
 473 static void open_counters(int cpu, pid_t pid)
 474 {
 475         int counter;
 476
 477         group_fd = -1;
 478         for (counter = 0; counter < nr_counters; counter++)
 479                 create_counter(counter, cpu, pid);
 480
 481         nr_cpu++;
 482 }
 483
 484 static void atexit_header(void)
 485 {
 486         header->data_size += bytes_written;
 487
 488         perf_header__write(header, output);
 489 }
 490
 491 static int __cmd_record(int argc, const char **argv)
 492 {
 493         int i, counter;
 494         struct stat st;
 495         pid_t pid = 0;
 496         int flags;
 497         int ret;
 498
 499         page_size = sysconf(_SC_PAGE_SIZE);
 500         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 501         assert(nr_cpus <= MAX_NR_CPUS);
 502         assert(nr_cpus >= 0);
 503
 504         atexit(sig_atexit);
 505         signal(SIGCHLD, sig_handler);
 506         signal(SIGINT, sig_handler);
 507
 508         if (!stat(output_name, &st) && !force && !append_file) {
 509                 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 510                                 output_name);
 511                 exit(-1);
 512         }
 513
 514         flags = O_CREAT|O_RDWR;
 515         if (append_file)
 516                 file_new = 0;
 517         else
 518                 flags |= O_TRUNC;
 519
 520         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 521         if (output < 0) {
 522                 perror("failed to create output file");
 523                 exit(-1);
 524         }
 525
 526         if (!file_new)
 527                 header = perf_header__read(output);
 528         else
 529                 header = perf_header__new();
 530
 531         atexit(atexit_header);
 532
 533         if (!system_wide) {
 534                 pid = target_pid;
 535                 if (pid == -1)
 536                         pid = getpid();
 537
 538                 open_counters(-1, pid);
 539         } else for (i = 0; i < nr_cpus; i++)
 540                 open_counters(i, target_pid);
 541
 542         if (file_new)
 543                 perf_header__write(header, output);
 544
 545         if (!system_wide) {
 546                 pid_synthesize_comm_event(pid, 0);
 547                 pid_synthesize_mmap_samples(pid);
 548         } else
 549                 synthesize_all();
 550
 551         if (target_pid == -1 && argc) {
 552                 pid = fork();
 553                 if (pid < 0)
 554                         perror("failed to fork");
 555
 556                 if (!pid) {
 557                         if (execvp(argv[0], (char **)argv)) {
 558                                 perror(argv[0]);
 559                                 exit(-1);
 560                         }
 561                 }
 562         }
 563
 564         if (realtime_prio) {
 565                 struct sched_param param;
 566
 567                 param.sched_priority = realtime_prio;
 568                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 569                         printf("Could not set realtime priority.\n");
 570                         exit(-1);
 571                 }
 572         }
 573
 574         while (!done) {
 575                 int hits = samples;
 576
 577                 for (i = 0; i < nr_cpu; i++) {
 578                         for (counter = 0; counter < nr_counters; counter++)
 579                                 mmap_read(&mmap_array[i][counter]);
 580                 }
 581
 582                 if (hits == samples)
 583                         ret = poll(event_array, nr_poll, 100);
 584         }
 585
 586         /*
 587          * Approximate RIP event size: 24 bytes.
 588          */
 589         fprintf(stderr,
 590                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 591                 (double)bytes_written / 1024.0 / 1024.0,
 592                 output_name,
 593                 bytes_written / 24);
 594
 595         return 0;
 596 }
 597
 598 static const char * const record_usage[] = {
 599         "perf record [<options>] [<command>]",
 600         "perf record [<options>] -- <command> [<options>]",
 601         NULL
 602 };
 603
 604 static const struct option options[] = {
 605         OPT_CALLBACK('e', "event", NULL, "event",
 606                      "event selector. use 'perf list' to list available events",
 607                      parse_events),
 608         OPT_INTEGER('p', "pid", &target_pid,
 609                     "record events on existing pid"),
 610         OPT_INTEGER('r', "realtime", &realtime_prio,
 611                     "collect data with this RT SCHED_FIFO priority"),
 612         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 613                             "system-wide collection from all CPUs"),
 614         OPT_BOOLEAN('A', "append", &append_file,
 615                             "append to the output file to do incremental profiling"),
 616         OPT_BOOLEAN('f', "force", &force,
 617                         "overwrite existing data file"),
 618         OPT_LONG('c', "count", &default_interval,
 619                     "event period to sample"),
 620         OPT_STRING('o', "output", &output_name, "file",
 621                     "output file name"),
 622         OPT_BOOLEAN('i', "inherit", &inherit,
 623                     "child tasks inherit counters"),
 624         OPT_INTEGER('F', "freq", &freq,
 625                     "profile at this frequency"),
 626         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 627                     "number of mmap data pages"),
 628         OPT_BOOLEAN('g', "call-graph", &call_graph,
 629                     "do call-graph (stack chain/backtrace) recording"),
 630         OPT_BOOLEAN('v', "verbose", &verbose,
 631                     "be more verbose (show counter open errors, etc)"),
 632         OPT_END()
 633 };
 634
 635 int cmd_record(int argc, const char **argv, const char *prefix)
 636 {
 637         int counter;
 638
 639         argc = parse_options(argc, argv, options, record_usage, 0);
 640         if (!argc && target_pid == -1 && !system_wide)
 641                 usage_with_options(record_usage, options);
 642
 643         if (!nr_counters) {
 644                 nr_counters     = 1;
 645                 attrs[0].type   = PERF_TYPE_HARDWARE;
 646                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 647         }
 648
 649         for (counter = 0; counter < nr_counters; counter++) {
 650                 if (attrs[counter].sample_period)
 651                         continue;
 652
 653                 attrs[counter].sample_period = default_interval;
 654         }
 655
 656         return __cmd_record(argc, argv);
 657 }