SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 struct mmap_event {
  55         struct perf_event_header        header;
  56         __u32                           pid;
  57         __u32                           tid;
  58         __u64                           start;
  59         __u64                           len;
  60         __u64                           pgoff;
  61         char                            filename[PATH_MAX];
  62 };
  63
  64 struct comm_event {
  65         struct perf_event_header        header;
  66         __u32                           pid;
  67         __u32                           tid;
  68         char                            comm[16];
  69 };
  70
  71
  72 struct mmap_data {
  73         int                     counter;
  74         void                    *base;
  75         unsigned int            mask;
  76         unsigned int            prev;
  77 };
  78
  79 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  80
  81 static unsigned long mmap_read_head(struct mmap_data *md)
  82 {
  83         struct perf_counter_mmap_page *pc = md->base;
  84         long head;
  85
  86         head = pc->data_head;
  87         rmb();
  88
  89         return head;
  90 }
  91
  92 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
  93 {
  94         struct perf_counter_mmap_page *pc = md->base;
  95
  96         /*
  97          * ensure all reads are done before we write the tail out.
  98          */
  99         /* mb(); */
 100         pc->data_tail = tail;
 101 }
 102
 103 static void mmap_read(struct mmap_data *md)
 104 {
 105         unsigned int head = mmap_read_head(md);
 106         unsigned int old = md->prev;
 107         unsigned char *data = md->base + page_size;
 108         unsigned long size;
 109         void *buf;
 110         int diff;
 111
 112         gettimeofday(&this_read, NULL);
 113
 114         /*
 115          * If we're further behind than half the buffer, there's a chance
 116          * the writer will bite our tail and mess up the samples under us.
 117          *
 118          * If we somehow ended up ahead of the head, we got messed up.
 119          *
 120          * In either case, truncate and restart at head.
 121          */
 122         diff = head - old;
 123         if (diff < 0) {
 124                 struct timeval iv;
 125                 unsigned long msecs;
 126
 127                 timersub(&this_read, &last_read, &iv);
 128                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 129
 130                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 131                                 "  Last read %lu msecs ago.\n", msecs);
 132
 133                 /*
 134                  * head points to a known good entry, start there.
 135                  */
 136                 old = head;
 137         }
 138
 139         last_read = this_read;
 140
 141         if (old != head)
 142                 samples++;
 143
 144         size = head - old;
 145
 146         if ((old & md->mask) + size != (head & md->mask)) {
 147                 buf = &data[old & md->mask];
 148                 size = md->mask + 1 - (old & md->mask);
 149                 old += size;
 150
 151                 while (size) {
 152                         int ret = write(output, buf, size);
 153
 154                         if (ret < 0)
 155                                 die("failed to write");
 156
 157                         size -= ret;
 158                         buf += ret;
 159
 160                         bytes_written += ret;
 161                 }
 162         }
 163
 164         buf = &data[old & md->mask];
 165         size = head - old;
 166         old += size;
 167
 168         while (size) {
 169                 int ret = write(output, buf, size);
 170
 171                 if (ret < 0)
 172                         die("failed to write");
 173
 174                 size -= ret;
 175                 buf += ret;
 176
 177                 bytes_written += ret;
 178         }
 179
 180         md->prev = old;
 181         mmap_write_tail(md, old);
 182 }
 183
 184 static volatile int done = 0;
 185 static volatile int signr = -1;
 186
 187 static void sig_handler(int sig)
 188 {
 189         done = 1;
 190         signr = sig;
 191 }
 192
 193 static void sig_atexit(void)
 194 {
 195         if (signr == -1)
 196                 return;
 197
 198         signal(signr, SIG_DFL);
 199         kill(getpid(), signr);
 200 }
 201
 202 static void pid_synthesize_comm_event(pid_t pid, int full)
 203 {
 204         struct comm_event comm_ev;
 205         char filename[PATH_MAX];
 206         char bf[BUFSIZ];
 207         int fd, ret;
 208         size_t size;
 209         char *field, *sep;
 210         DIR *tasks;
 211         struct dirent dirent, *next;
 212
 213         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 214
 215         fd = open(filename, O_RDONLY);
 216         if (fd < 0) {
 217                 /*
 218                  * We raced with a task exiting - just return:
 219                  */
 220                 if (verbose)
 221                         fprintf(stderr, "couldn't open %s\n", filename);
 222                 return;
 223         }
 224         if (read(fd, bf, sizeof(bf)) < 0) {
 225                 fprintf(stderr, "couldn't read %s\n", filename);
 226                 exit(EXIT_FAILURE);
 227         }
 228         close(fd);
 229
 230         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 231         memset(&comm_ev, 0, sizeof(comm_ev));
 232         field = strchr(bf, '(');
 233         if (field == NULL)
 234                 goto out_failure;
 235         sep = strchr(++field, ')');
 236         if (sep == NULL)
 237                 goto out_failure;
 238         size = sep - field;
 239         memcpy(comm_ev.comm, field, size++);
 240
 241         comm_ev.pid = pid;
 242         comm_ev.header.type = PERF_EVENT_COMM;
 243         size = ALIGN(size, sizeof(__u64));
 244         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 245
 246         if (!full) {
 247                 comm_ev.tid = pid;
 248
 249                 ret = write(output, &comm_ev, comm_ev.header.size);
 250                 if (ret < 0) {
 251                         perror("failed to write");
 252                         exit(-1);
 253                 }
 254                 return;
 255         }
 256
 257         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 258
 259         tasks = opendir(filename);
 260         while (!readdir_r(tasks, &dirent, &next) && next) {
 261                 char *end;
 262                 pid = strtol(dirent.d_name, &end, 10);
 263                 if (*end)
 264                         continue;
 265
 266                 comm_ev.tid = pid;
 267
 268                 ret = write(output, &comm_ev, comm_ev.header.size);
 269                 if (ret < 0) {
 270                         perror("failed to write");
 271                         exit(-1);
 272                 }
 273         }
 274         closedir(tasks);
 275         return;
 276
 277 out_failure:
 278         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 279                 filename);
 280         exit(EXIT_FAILURE);
 281 }
 282
 283 static void pid_synthesize_mmap_samples(pid_t pid)
 284 {
 285         char filename[PATH_MAX];
 286         FILE *fp;
 287
 288         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 289
 290         fp = fopen(filename, "r");
 291         if (fp == NULL) {
 292                 /*
 293                  * We raced with a task exiting - just return:
 294                  */
 295                 if (verbose)
 296                         fprintf(stderr, "couldn't open %s\n", filename);
 297                 return;
 298         }
 299         while (1) {
 300                 char bf[BUFSIZ], *pbf = bf;
 301                 struct mmap_event mmap_ev = {
 302                         .header.type = PERF_EVENT_MMAP,
 303                 };
 304                 int n;
 305                 size_t size;
 306                 if (fgets(bf, sizeof(bf), fp) == NULL)
 307                         break;
 308
 309                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 310                 n = hex2u64(pbf, &mmap_ev.start);
 311                 if (n < 0)
 312                         continue;
 313                 pbf += n + 1;
 314                 n = hex2u64(pbf, &mmap_ev.len);
 315                 if (n < 0)
 316                         continue;
 317                 pbf += n + 3;
 318                 if (*pbf == 'x') { /* vm_exec */
 319                         char *execname = strrchr(bf, ' ');
 320
 321                         if (execname == NULL || execname[1] != '/')
 322                                 continue;
 323
 324                         execname += 1;
 325                         size = strlen(execname);
 326                         execname[size - 1] = '\0'; /* Remove \n */
 327                         memcpy(mmap_ev.filename, execname, size);
 328                         size = ALIGN(size, sizeof(__u64));
 329                         mmap_ev.len -= mmap_ev.start;
 330                         mmap_ev.header.size = (sizeof(mmap_ev) -
 331                                                (sizeof(mmap_ev.filename) - size));
 332                         mmap_ev.pid = pid;
 333                         mmap_ev.tid = pid;
 334
 335                         if (write(output, &mmap_ev, mmap_ev.header.size) < 0) {
 336                                 perror("failed to write");
 337                                 exit(-1);
 338                         }
 339                 }
 340         }
 341
 342         fclose(fp);
 343 }
 344
 345 static void synthesize_samples(void)
 346 {
 347         DIR *proc;
 348         struct dirent dirent, *next;
 349
 350         proc = opendir("/proc");
 351
 352         while (!readdir_r(proc, &dirent, &next) && next) {
 353                 char *end;
 354                 pid_t pid;
 355
 356                 pid = strtol(dirent.d_name, &end, 10);
 357                 if (*end) /* only interested in proper numerical dirents */
 358                         continue;
 359
 360                 pid_synthesize_comm_event(pid, 1);
 361                 pid_synthesize_mmap_samples(pid);
 362         }
 363
 364         closedir(proc);
 365 }
 366
 367 static int group_fd;
 368
 369 static void create_counter(int counter, int cpu, pid_t pid)
 370 {
 371         struct perf_counter_attr *attr = attrs + counter;
 372         int track = 1;
 373
 374         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 375
 376         if (freq) {
 377                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 378                 attr->freq              = 1;
 379                 attr->sample_freq       = freq;
 380         }
 381
 382         if (call_graph)
 383                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 384
 385         attr->mmap              = track;
 386         attr->comm              = track;
 387         attr->inherit           = (cpu < 0) && inherit;
 388         attr->disabled          = 1;
 389
 390         track = 0; /* only the first counter needs these */
 391
 392 try_again:
 393         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 394
 395         if (fd[nr_cpu][counter] < 0) {
 396                 int err = errno;
 397
 398                 if (err == EPERM)
 399                         die("Permission error - are you root?\n");
 400
 401                 /*
 402                  * If it's cycles then fall back to hrtimer
 403                  * based cpu-clock-tick sw counter, which
 404                  * is always available even if no PMU support:
 405                  */
 406                 if (attr->type == PERF_TYPE_HARDWARE
 407                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 408
 409                         if (verbose)
 410                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 411                         attr->type = PERF_TYPE_SOFTWARE;
 412                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 413                         goto try_again;
 414                 }
 415                 printf("\n");
 416                 error("perfcounter syscall returned with %d (%s)\n",
 417                         fd[nr_cpu][counter], strerror(err));
 418                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 419                 exit(-1);
 420         }
 421
 422         assert(fd[nr_cpu][counter] >= 0);
 423         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 424
 425         /*
 426          * First counter acts as the group leader:
 427          */
 428         if (group && group_fd == -1)
 429                 group_fd = fd[nr_cpu][counter];
 430
 431         event_array[nr_poll].fd = fd[nr_cpu][counter];
 432         event_array[nr_poll].events = POLLIN;
 433         nr_poll++;
 434
 435         mmap_array[nr_cpu][counter].counter = counter;
 436         mmap_array[nr_cpu][counter].prev = 0;
 437         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 438         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 439                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
 440         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 441                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 442                 exit(-1);
 443         }
 444
 445         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
 446 }
 447
 448 static void open_counters(int cpu, pid_t pid)
 449 {
 450         int counter;
 451
 452         if (pid > 0) {
 453                 pid_synthesize_comm_event(pid, 0);
 454                 pid_synthesize_mmap_samples(pid);
 455         }
 456
 457         group_fd = -1;
 458         for (counter = 0; counter < nr_counters; counter++)
 459                 create_counter(counter, cpu, pid);
 460
 461         nr_cpu++;
 462 }
 463
 464 static int __cmd_record(int argc, const char **argv)
 465 {
 466         int i, counter;
 467         struct stat st;
 468         pid_t pid;
 469         int flags;
 470         int ret;
 471
 472         page_size = sysconf(_SC_PAGE_SIZE);
 473         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 474         assert(nr_cpus <= MAX_NR_CPUS);
 475         assert(nr_cpus >= 0);
 476
 477         if (!stat(output_name, &st) && !force && !append_file) {
 478                 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 479                                 output_name);
 480                 exit(-1);
 481         }
 482
 483         flags = O_CREAT|O_RDWR;
 484         if (append_file)
 485                 flags |= O_APPEND;
 486         else
 487                 flags |= O_TRUNC;
 488
 489         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 490         if (output < 0) {
 491                 perror("failed to create output file");
 492                 exit(-1);
 493         }
 494
 495         if (!system_wide) {
 496                 open_counters(-1, target_pid != -1 ? target_pid : getpid());
 497         } else for (i = 0; i < nr_cpus; i++)
 498                 open_counters(i, target_pid);
 499
 500         atexit(sig_atexit);
 501         signal(SIGCHLD, sig_handler);
 502         signal(SIGINT, sig_handler);
 503
 504         if (target_pid == -1 && argc) {
 505                 pid = fork();
 506                 if (pid < 0)
 507                         perror("failed to fork");
 508
 509                 if (!pid) {
 510                         if (execvp(argv[0], (char **)argv)) {
 511                                 perror(argv[0]);
 512                                 exit(-1);
 513                         }
 514                 }
 515         }
 516
 517         if (realtime_prio) {
 518                 struct sched_param param;
 519
 520                 param.sched_priority = realtime_prio;
 521                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 522                         printf("Could not set realtime priority.\n");
 523                         exit(-1);
 524                 }
 525         }
 526
 527         if (system_wide)
 528                 synthesize_samples();
 529
 530         while (!done) {
 531                 int hits = samples;
 532
 533                 for (i = 0; i < nr_cpu; i++) {
 534                         for (counter = 0; counter < nr_counters; counter++)
 535                                 mmap_read(&mmap_array[i][counter]);
 536                 }
 537
 538                 if (hits == samples)
 539                         ret = poll(event_array, nr_poll, 100);
 540         }
 541
 542         /*
 543          * Approximate RIP event size: 24 bytes.
 544          */
 545         fprintf(stderr,
 546                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 547                 (double)bytes_written / 1024.0 / 1024.0,
 548                 output_name,
 549                 bytes_written / 24);
 550
 551         return 0;
 552 }
 553
 554 static const char * const record_usage[] = {
 555         "perf record [<options>] [<command>]",
 556         "perf record [<options>] -- <command> [<options>]",
 557         NULL
 558 };
 559
 560 static const struct option options[] = {
 561         OPT_CALLBACK('e', "event", NULL, "event",
 562                      "event selector. use 'perf list' to list available events",
 563                      parse_events),
 564         OPT_INTEGER('p', "pid", &target_pid,
 565                     "record events on existing pid"),
 566         OPT_INTEGER('r', "realtime", &realtime_prio,
 567                     "collect data with this RT SCHED_FIFO priority"),
 568         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 569                             "system-wide collection from all CPUs"),
 570         OPT_BOOLEAN('A', "append", &append_file,
 571                             "append to the output file to do incremental profiling"),
 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_END()
 589 };
 590
 591 int cmd_record(int argc, const char **argv, const char *prefix)
 592 {
 593         int counter;
 594
 595         argc = parse_options(argc, argv, options, record_usage, 0);
 596         if (!argc && target_pid == -1 && !system_wide)
 597                 usage_with_options(record_usage, options);
 598
 599         if (!nr_counters) {
 600                 nr_counters     = 1;
 601                 attrs[0].type   = PERF_TYPE_HARDWARE;
 602                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 603         }
 604
 605         for (counter = 0; counter < nr_counters; counter++) {
 606                 if (attrs[counter].sample_period)
 607                         continue;
 608
 609                 attrs[counter].sample_period = default_interval;
 610         }
 611
 612         return __cmd_record(argc, argv);
 613 }