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 int mmap_read_head(struct mmap_data *md)
  82 {
  83         struct perf_counter_mmap_page *pc = md->base;
  84         int head;
  85
  86         head = pc->data_head;
  87         rmb();
  88
  89         return head;
  90 }
  91
  92 static void mmap_read(struct mmap_data *md)
  93 {
  94         unsigned int head = mmap_read_head(md);
  95         unsigned int old = md->prev;
  96         unsigned char *data = md->base + page_size;
  97         unsigned long size;
  98         void *buf;
  99         int diff;
 100
 101         gettimeofday(&this_read, NULL);
 102
 103         /*
 104          * If we're further behind than half the buffer, there's a chance
 105          * the writer will bite our tail and mess up the samples under us.
 106          *
 107          * If we somehow ended up ahead of the head, we got messed up.
 108          *
 109          * In either case, truncate and restart at head.
 110          */
 111         diff = head - old;
 112         if (diff > md->mask / 2 || diff < 0) {
 113                 struct timeval iv;
 114                 unsigned long msecs;
 115
 116                 timersub(&this_read, &last_read, &iv);
 117                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 118
 119                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 120                                 "  Last read %lu msecs ago.\n", msecs);
 121
 122                 /*
 123                  * head points to a known good entry, start there.
 124                  */
 125                 old = head;
 126         }
 127
 128         last_read = this_read;
 129
 130         if (old != head)
 131                 samples++;
 132
 133         size = head - old;
 134
 135         if ((old & md->mask) + size != (head & md->mask)) {
 136                 buf = &data[old & md->mask];
 137                 size = md->mask + 1 - (old & md->mask);
 138                 old += size;
 139
 140                 while (size) {
 141                         int ret = write(output, buf, size);
 142
 143                         if (ret < 0)
 144                                 die("failed to write");
 145
 146                         size -= ret;
 147                         buf += ret;
 148
 149                         bytes_written += ret;
 150                 }
 151         }
 152
 153         buf = &data[old & md->mask];
 154         size = head - old;
 155         old += size;
 156
 157         while (size) {
 158                 int ret = write(output, buf, size);
 159
 160                 if (ret < 0)
 161                         die("failed to write");
 162
 163                 size -= ret;
 164                 buf += ret;
 165
 166                 bytes_written += ret;
 167         }
 168
 169         md->prev = old;
 170 }
 171
 172 static volatile int done = 0;
 173 static volatile int signr = -1;
 174
 175 static void sig_handler(int sig)
 176 {
 177         done = 1;
 178         signr = sig;
 179 }
 180
 181 static void sig_atexit(void)
 182 {
 183         if (signr == -1)
 184                 return;
 185
 186         signal(signr, SIG_DFL);
 187         kill(getpid(), signr);
 188 }
 189
 190 static void pid_synthesize_comm_event(pid_t pid, int full)
 191 {
 192         struct comm_event comm_ev;
 193         char filename[PATH_MAX];
 194         char bf[BUFSIZ];
 195         int fd, ret;
 196         size_t size;
 197         char *field, *sep;
 198         DIR *tasks;
 199         struct dirent dirent, *next;
 200
 201         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 202
 203         fd = open(filename, O_RDONLY);
 204         if (fd < 0) {
 205                 fprintf(stderr, "couldn't open %s\n", filename);
 206                 exit(EXIT_FAILURE);
 207         }
 208         if (read(fd, bf, sizeof(bf)) < 0) {
 209                 fprintf(stderr, "couldn't read %s\n", filename);
 210                 exit(EXIT_FAILURE);
 211         }
 212         close(fd);
 213
 214         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 215         memset(&comm_ev, 0, sizeof(comm_ev));
 216         field = strchr(bf, '(');
 217         if (field == NULL)
 218                 goto out_failure;
 219         sep = strchr(++field, ')');
 220         if (sep == NULL)
 221                 goto out_failure;
 222         size = sep - field;
 223         memcpy(comm_ev.comm, field, size++);
 224
 225         comm_ev.pid = pid;
 226         comm_ev.header.type = PERF_EVENT_COMM;
 227         size = ALIGN(size, sizeof(__u64));
 228         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 229
 230         if (!full) {
 231                 comm_ev.tid = pid;
 232
 233                 ret = write(output, &comm_ev, comm_ev.header.size);
 234                 if (ret < 0) {
 235                         perror("failed to write");
 236                         exit(-1);
 237                 }
 238                 return;
 239         }
 240
 241         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 242
 243         tasks = opendir(filename);
 244         while (!readdir_r(tasks, &dirent, &next) && next) {
 245                 char *end;
 246                 pid = strtol(dirent.d_name, &end, 10);
 247                 if (*end)
 248                         continue;
 249
 250                 comm_ev.tid = pid;
 251
 252                 ret = write(output, &comm_ev, comm_ev.header.size);
 253                 if (ret < 0) {
 254                         perror("failed to write");
 255                         exit(-1);
 256                 }
 257         }
 258         closedir(tasks);
 259         return;
 260
 261 out_failure:
 262         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 263                 filename);
 264         exit(EXIT_FAILURE);
 265 }
 266
 267 static void pid_synthesize_mmap_samples(pid_t pid)
 268 {
 269         char filename[PATH_MAX];
 270         FILE *fp;
 271
 272         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 273
 274         fp = fopen(filename, "r");
 275         if (fp == NULL) {
 276                 fprintf(stderr, "couldn't open %s\n", filename);
 277                 exit(EXIT_FAILURE);
 278         }
 279         while (1) {
 280                 char bf[BUFSIZ], *pbf = bf;
 281                 struct mmap_event mmap_ev = {
 282                         .header.type = PERF_EVENT_MMAP,
 283                 };
 284                 int n;
 285                 size_t size;
 286                 if (fgets(bf, sizeof(bf), fp) == NULL)
 287                         break;
 288
 289                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 290                 n = hex2u64(pbf, &mmap_ev.start);
 291                 if (n < 0)
 292                         continue;
 293                 pbf += n + 1;
 294                 n = hex2u64(pbf, &mmap_ev.len);
 295                 if (n < 0)
 296                         continue;
 297                 pbf += n + 3;
 298                 if (*pbf == 'x') { /* vm_exec */
 299                         char *execname = strrchr(bf, ' ');
 300
 301                         if (execname == NULL || execname[1] != '/')
 302                                 continue;
 303
 304                         execname += 1;
 305                         size = strlen(execname);
 306                         execname[size - 1] = '\0'; /* Remove \n */
 307                         memcpy(mmap_ev.filename, execname, size);
 308                         size = ALIGN(size, sizeof(__u64));
 309                         mmap_ev.len -= mmap_ev.start;
 310                         mmap_ev.header.size = (sizeof(mmap_ev) -
 311                                                (sizeof(mmap_ev.filename) - size));
 312                         mmap_ev.pid = pid;
 313                         mmap_ev.tid = pid;
 314
 315                         if (write(output, &mmap_ev, mmap_ev.header.size) < 0) {
 316                                 perror("failed to write");
 317                                 exit(-1);
 318                         }
 319                 }
 320         }
 321
 322         fclose(fp);
 323 }
 324
 325 static void synthesize_samples(void)
 326 {
 327         DIR *proc;
 328         struct dirent dirent, *next;
 329
 330         proc = opendir("/proc");
 331
 332         while (!readdir_r(proc, &dirent, &next) && next) {
 333                 char *end;
 334                 pid_t pid;
 335
 336                 pid = strtol(dirent.d_name, &end, 10);
 337                 if (*end) /* only interested in proper numerical dirents */
 338                         continue;
 339
 340                 pid_synthesize_comm_event(pid, 1);
 341                 pid_synthesize_mmap_samples(pid);
 342         }
 343
 344         closedir(proc);
 345 }
 346
 347 static int group_fd;
 348
 349 static void create_counter(int counter, int cpu, pid_t pid)
 350 {
 351         struct perf_counter_attr *attr = attrs + counter;
 352         int track = 1;
 353
 354         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 355
 356         if (freq) {
 357                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
 358                 attr->freq              = 1;
 359                 attr->sample_freq       = freq;
 360         }
 361
 362         if (call_graph)
 363                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 364
 365         attr->mmap              = track;
 366         attr->comm              = track;
 367         attr->inherit           = (cpu < 0) && inherit;
 368         attr->disabled          = 1;
 369
 370         track = 0; /* only the first counter needs these */
 371
 372 try_again:
 373         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 374
 375         if (fd[nr_cpu][counter] < 0) {
 376                 int err = errno;
 377
 378                 if (err == EPERM)
 379                         die("Permission error - are you root?\n");
 380
 381                 /*
 382                  * If it's cycles then fall back to hrtimer
 383                  * based cpu-clock-tick sw counter, which
 384                  * is always available even if no PMU support:
 385                  */
 386                 if (attr->type == PERF_TYPE_HARDWARE
 387                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 388
 389                         if (verbose)
 390                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 391                         attr->type = PERF_TYPE_SOFTWARE;
 392                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
 393                         goto try_again;
 394                 }
 395                 printf("\n");
 396                 error("perfcounter syscall returned with %d (%s)\n",
 397                         fd[nr_cpu][counter], strerror(err));
 398                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 399                 exit(-1);
 400         }
 401
 402         assert(fd[nr_cpu][counter] >= 0);
 403         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 404
 405         /*
 406          * First counter acts as the group leader:
 407          */
 408         if (group && group_fd == -1)
 409                 group_fd = fd[nr_cpu][counter];
 410
 411         event_array[nr_poll].fd = fd[nr_cpu][counter];
 412         event_array[nr_poll].events = POLLIN;
 413         nr_poll++;
 414
 415         mmap_array[nr_cpu][counter].counter = counter;
 416         mmap_array[nr_cpu][counter].prev = 0;
 417         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 418         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 419                         PROT_READ, MAP_SHARED, fd[nr_cpu][counter], 0);
 420         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 421                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 422                 exit(-1);
 423         }
 424
 425         ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
 426 }
 427
 428 static void open_counters(int cpu, pid_t pid)
 429 {
 430         int counter;
 431
 432         if (pid > 0) {
 433                 pid_synthesize_comm_event(pid, 0);
 434                 pid_synthesize_mmap_samples(pid);
 435         }
 436
 437         group_fd = -1;
 438         for (counter = 0; counter < nr_counters; counter++)
 439                 create_counter(counter, cpu, pid);
 440
 441         nr_cpu++;
 442 }
 443
 444 static int __cmd_record(int argc, const char **argv)
 445 {
 446         int i, counter;
 447         struct stat st;
 448         pid_t pid;
 449         int flags;
 450         int ret;
 451
 452         page_size = sysconf(_SC_PAGE_SIZE);
 453         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 454         assert(nr_cpus <= MAX_NR_CPUS);
 455         assert(nr_cpus >= 0);
 456
 457         if (!stat(output_name, &st) && !force && !append_file) {
 458                 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 459                                 output_name);
 460                 exit(-1);
 461         }
 462
 463         flags = O_CREAT|O_RDWR;
 464         if (append_file)
 465                 flags |= O_APPEND;
 466         else
 467                 flags |= O_TRUNC;
 468
 469         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 470         if (output < 0) {
 471                 perror("failed to create output file");
 472                 exit(-1);
 473         }
 474
 475         if (!system_wide) {
 476                 open_counters(-1, target_pid != -1 ? target_pid : getpid());
 477         } else for (i = 0; i < nr_cpus; i++)
 478                 open_counters(i, target_pid);
 479
 480         atexit(sig_atexit);
 481         signal(SIGCHLD, sig_handler);
 482         signal(SIGINT, sig_handler);
 483
 484         if (target_pid == -1 && argc) {
 485                 pid = fork();
 486                 if (pid < 0)
 487                         perror("failed to fork");
 488
 489                 if (!pid) {
 490                         if (execvp(argv[0], (char **)argv)) {
 491                                 perror(argv[0]);
 492                                 exit(-1);
 493                         }
 494                 }
 495         }
 496
 497         if (realtime_prio) {
 498                 struct sched_param param;
 499
 500                 param.sched_priority = realtime_prio;
 501                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 502                         printf("Could not set realtime priority.\n");
 503                         exit(-1);
 504                 }
 505         }
 506
 507         if (system_wide)
 508                 synthesize_samples();
 509
 510         while (!done) {
 511                 int hits = samples;
 512
 513                 for (i = 0; i < nr_cpu; i++) {
 514                         for (counter = 0; counter < nr_counters; counter++)
 515                                 mmap_read(&mmap_array[i][counter]);
 516                 }
 517
 518                 if (hits == samples)
 519                         ret = poll(event_array, nr_poll, 100);
 520         }
 521
 522         /*
 523          * Approximate RIP event size: 24 bytes.
 524          */
 525         fprintf(stderr,
 526                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 527                 (double)bytes_written / 1024.0 / 1024.0,
 528                 output_name,
 529                 bytes_written / 24);
 530
 531         return 0;
 532 }
 533
 534 static const char * const record_usage[] = {
 535         "perf record [<options>] [<command>]",
 536         "perf record [<options>] -- <command> [<options>]",
 537         NULL
 538 };
 539
 540 static const struct option options[] = {
 541         OPT_CALLBACK('e', "event", NULL, "event",
 542                      "event selector. use 'perf list' to list available events",
 543                      parse_events),
 544         OPT_INTEGER('p', "pid", &target_pid,
 545                     "record events on existing pid"),
 546         OPT_INTEGER('r', "realtime", &realtime_prio,
 547                     "collect data with this RT SCHED_FIFO priority"),
 548         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 549                             "system-wide collection from all CPUs"),
 550         OPT_BOOLEAN('A', "append", &append_file,
 551                             "append to the output file to do incremental profiling"),
 552         OPT_BOOLEAN('f', "force", &force,
 553                         "overwrite existing data file"),
 554         OPT_LONG('c', "count", &default_interval,
 555                     "event period to sample"),
 556         OPT_STRING('o', "output", &output_name, "file",
 557                     "output file name"),
 558         OPT_BOOLEAN('i', "inherit", &inherit,
 559                     "child tasks inherit counters"),
 560         OPT_INTEGER('F', "freq", &freq,
 561                     "profile at this frequency"),
 562         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 563                     "number of mmap data pages"),
 564         OPT_BOOLEAN('g', "call-graph", &call_graph,
 565                     "do call-graph (stack chain/backtrace) recording"),
 566         OPT_BOOLEAN('v', "verbose", &verbose,
 567                     "be more verbose (show counter open errors, etc)"),
 568         OPT_END()
 569 };
 570
 571 int cmd_record(int argc, const char **argv, const char *prefix)
 572 {
 573         int counter;
 574
 575         argc = parse_options(argc, argv, options, record_usage, 0);
 576         if (!argc && target_pid == -1 && !system_wide)
 577                 usage_with_options(record_usage, options);
 578
 579         if (!nr_counters) {
 580                 nr_counters     = 1;
 581                 attrs[0].type   = PERF_TYPE_HARDWARE;
 582                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 583         }
 584
 585         for (counter = 0; counter < nr_counters; counter++) {
 586                 if (attrs[counter].sample_period)
 587                         continue;
 588
 589                 attrs[counter].sample_period = default_interval;
 590         }
 591
 592         return __cmd_record(argc, argv);
 593 }