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