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                      verbose                         = 0;
  41
  42 static long                     samples;
  43 static struct timeval           last_read;
  44 static struct timeval           this_read;
  45
  46 static __u64                    bytes_written;
  47
  48 static struct pollfd            event_array[MAX_NR_CPUS * MAX_COUNTERS];
  49
  50 static int                      nr_poll;
  51 static int                      nr_cpu;
  52
  53 struct mmap_event {
  54         struct perf_event_header        header;
  55         __u32                           pid;
  56         __u32                           tid;
  57         __u64                           start;
  58         __u64                           len;
  59         __u64                           pgoff;
  60         char                            filename[PATH_MAX];
  61 };
  62
  63 struct comm_event {
  64         struct perf_event_header        header;
  65         __u32                           pid;
  66         __u32                           tid;
  67         char                            comm[16];
  68 };
  69
  70
  71 struct mmap_data {
  72         int                     counter;
  73         void                    *base;
  74         unsigned int            mask;
  75         unsigned int            prev;
  76 };
  77
  78 static struct mmap_data         mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  79
  80 static unsigned int mmap_read_head(struct mmap_data *md)
  81 {
  82         struct perf_counter_mmap_page *pc = md->base;
  83         int head;
  84
  85         head = pc->data_head;
  86         rmb();
  87
  88         return head;
  89 }
  90
  91 static void mmap_read(struct mmap_data *md)
  92 {
  93         unsigned int head = mmap_read_head(md);
  94         unsigned int old = md->prev;
  95         unsigned char *data = md->base + page_size;
  96         unsigned long size;
  97         void *buf;
  98         int diff;
  99
 100         gettimeofday(&this_read, NULL);
 101
 102         /*
 103          * If we're further behind than half the buffer, there's a chance
 104          * the writer will bite our tail and mess up the samples under us.
 105          *
 106          * If we somehow ended up ahead of the head, we got messed up.
 107          *
 108          * In either case, truncate and restart at head.
 109          */
 110         diff = head - old;
 111         if (diff > md->mask / 2 || diff < 0) {
 112                 struct timeval iv;
 113                 unsigned long msecs;
 114
 115                 timersub(&this_read, &last_read, &iv);
 116                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 117
 118                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 119                                 "  Last read %lu msecs ago.\n", msecs);
 120
 121                 /*
 122                  * head points to a known good entry, start there.
 123                  */
 124                 old = head;
 125         }
 126
 127         last_read = this_read;
 128
 129         if (old != head)
 130                 samples++;
 131
 132         size = head - old;
 133
 134         if ((old & md->mask) + size != (head & md->mask)) {
 135                 buf = &data[old & md->mask];
 136                 size = md->mask + 1 - (old & md->mask);
 137                 old += size;
 138
 139                 while (size) {
 140                         int ret = write(output, buf, size);
 141
 142                         if (ret < 0)
 143                                 die("failed to write");
 144
 145                         size -= ret;
 146                         buf += ret;
 147
 148                         bytes_written += ret;
 149                 }
 150         }
 151
 152         buf = &data[old & md->mask];
 153         size = head - old;
 154         old += size;
 155
 156         while (size) {
 157                 int ret = write(output, buf, size);
 158
 159                 if (ret < 0)
 160                         die("failed to write");
 161
 162                 size -= ret;
 163                 buf += ret;
 164
 165                 bytes_written += ret;
 166         }
 167
 168         md->prev = old;
 169 }
 170
 171 static volatile int done = 0;
 172
 173 static void sig_handler(int sig)
 174 {
 175         done = 1;
 176 }
 177
 178 static void pid_synthesize_comm_event(pid_t pid, int full)
 179 {
 180         struct comm_event comm_ev;
 181         char filename[PATH_MAX];
 182         char bf[BUFSIZ];
 183         int fd, ret;
 184         size_t size;
 185         char *field, *sep;
 186         DIR *tasks;
 187         struct dirent dirent, *next;
 188
 189         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 190
 191         fd = open(filename, O_RDONLY);
 192         if (fd < 0) {
 193                 fprintf(stderr, "couldn't open %s\n", filename);
 194                 exit(EXIT_FAILURE);
 195         }
 196         if (read(fd, bf, sizeof(bf)) < 0) {
 197                 fprintf(stderr, "couldn't read %s\n", filename);
 198                 exit(EXIT_FAILURE);
 199         }
 200         close(fd);
 201
 202         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 203         memset(&comm_ev, 0, sizeof(comm_ev));
 204         field = strchr(bf, '(');
 205         if (field == NULL)
 206                 goto out_failure;
 207         sep = strchr(++field, ')');
 208         if (sep == NULL)
 209                 goto out_failure;
 210         size = sep - field;
 211         memcpy(comm_ev.comm, field, size++);
 212
 213         comm_ev.pid = pid;
 214         comm_ev.header.type = PERF_EVENT_COMM;
 215         size = ALIGN(size, sizeof(uint64_t));
 216         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 217
 218         if (!full) {
 219                 comm_ev.tid = pid;
 220
 221                 ret = write(output, &comm_ev, comm_ev.header.size);
 222                 if (ret < 0) {
 223                         perror("failed to write");
 224                         exit(-1);
 225                 }
 226                 return;
 227         }
 228
 229         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 230
 231         tasks = opendir(filename);
 232         while (!readdir_r(tasks, &dirent, &next) && next) {
 233                 char *end;
 234                 pid = strtol(dirent.d_name, &end, 10);
 235                 if (*end)
 236                         continue;
 237
 238                 comm_ev.tid = pid;
 239
 240                 ret = write(output, &comm_ev, comm_ev.header.size);
 241                 if (ret < 0) {
 242                         perror("failed to write");
 243                         exit(-1);
 244                 }
 245         }
 246         closedir(tasks);
 247         return;
 248
 249 out_failure:
 250         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 251                 filename);
 252         exit(EXIT_FAILURE);
 253 }
 254
 255 static void pid_synthesize_mmap_samples(pid_t pid)
 256 {
 257         char filename[PATH_MAX];
 258         FILE *fp;
 259
 260         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 261
 262         fp = fopen(filename, "r");
 263         if (fp == NULL) {
 264                 fprintf(stderr, "couldn't open %s\n", filename);
 265                 exit(EXIT_FAILURE);
 266         }
 267         while (1) {
 268                 char bf[BUFSIZ], *pbf = bf;
 269                 struct mmap_event mmap_ev = {
 270                         .header.type = PERF_EVENT_MMAP,
 271                 };
 272                 int n;
 273                 size_t size;
 274                 if (fgets(bf, sizeof(bf), fp) == NULL)
 275                         break;
 276
 277                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 278                 n = hex2u64(pbf, &mmap_ev.start);
 279                 if (n < 0)
 280                         continue;
 281                 pbf += n + 1;
 282                 n = hex2u64(pbf, &mmap_ev.len);
 283                 if (n < 0)
 284                         continue;
 285                 pbf += n + 3;
 286                 if (*pbf == 'x') { /* vm_exec */
 287                         char *execname = strrchr(bf, ' ');
 288
 289                         if (execname == NULL || execname[1] != '/')
 290                                 continue;
 291
 292                         execname += 1;
 293                         size = strlen(execname);
 294                         execname[size - 1] = '\0'; /* Remove \n */
 295                         memcpy(mmap_ev.filename, execname, size);
 296                         size = ALIGN(size, sizeof(uint64_t));
 297                         mmap_ev.len -= mmap_ev.start;
 298                         mmap_ev.header.size = (sizeof(mmap_ev) -
 299                                                (sizeof(mmap_ev.filename) - size));
 300                         mmap_ev.pid = pid;
 301                         mmap_ev.tid = pid;
 302
 303                         if (write(output, &mmap_ev, mmap_ev.header.size) < 0) {
 304                                 perror("failed to write");
 305                                 exit(-1);
 306                         }
 307                 }
 308         }
 309
 310         fclose(fp);
 311 }
 312
 313 static void synthesize_samples(void)
 314 {
 315         DIR *proc;
 316         struct dirent dirent, *next;
 317
 318         proc = opendir("/proc");
 319
 320         while (!readdir_r(proc, &dirent, &next) && next) {
 321                 char *end;
 322                 pid_t pid;
 323
 324                 pid = strtol(dirent.d_name, &end, 10);
 325                 if (*end) /* only interested in proper numerical dirents */
 326                         continue;
 327
 328                 pid_synthesize_comm_event(pid, 1);
 329                 pid_synthesize_mmap_samples(pid);
 330         }
 331
 332         closedir(proc);
 333 }
 334
 335 static int group_fd;
 336
 337 static void create_counter(int counter, int cpu, pid_t pid)
 338 {
 339         struct perf_counter_attr *attr = attrs + counter;
 340         int track = 1;
 341
 342         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_PERIOD;
 343         if (freq) {
 344                 attr->freq              = 1;
 345                 attr->sample_freq       = freq;
 346         }
 347         attr->mmap              = track;
 348         attr->comm              = track;
 349         attr->inherit           = (cpu < 0) && inherit;
 350
 351         track = 0; /* only the first counter needs these */
 352
 353 try_again:
 354         fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
 355
 356         if (fd[nr_cpu][counter] < 0) {
 357                 int err = errno;
 358
 359                 if (err == EPERM)
 360                         die("Permission error - are you root?\n");
 361
 362                 /*
 363                  * If it's cycles then fall back to hrtimer
 364                  * based cpu-clock-tick sw counter, which
 365                  * is always available even if no PMU support:
 366                  */
 367                 if (attr->type == PERF_TYPE_HARDWARE
 368                         && attr->config == PERF_COUNT_CPU_CYCLES) {
 369
 370                         if (verbose)
 371                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
 372                         attr->type = PERF_TYPE_SOFTWARE;
 373                         attr->config = PERF_COUNT_CPU_CLOCK;
 374                         goto try_again;
 375                 }
 376                 printf("\n");
 377                 error("perfcounter syscall returned with %d (%s)\n",
 378                         fd[nr_cpu][counter], strerror(err));
 379                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
 380                 exit(-1);
 381         }
 382
 383         assert(fd[nr_cpu][counter] >= 0);
 384         fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 385
 386         /*
 387          * First counter acts as the group leader:
 388          */
 389         if (group && group_fd == -1)
 390                 group_fd = fd[nr_cpu][counter];
 391
 392         event_array[nr_poll].fd = fd[nr_cpu][counter];
 393         event_array[nr_poll].events = POLLIN;
 394         nr_poll++;
 395
 396         mmap_array[nr_cpu][counter].counter = counter;
 397         mmap_array[nr_cpu][counter].prev = 0;
 398         mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 399         mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 400                         PROT_READ, MAP_SHARED, fd[nr_cpu][counter], 0);
 401         if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 402                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 403                 exit(-1);
 404         }
 405 }
 406
 407 static void open_counters(int cpu, pid_t pid)
 408 {
 409         int counter;
 410
 411         if (pid > 0) {
 412                 pid_synthesize_comm_event(pid, 0);
 413                 pid_synthesize_mmap_samples(pid);
 414         }
 415
 416         group_fd = -1;
 417         for (counter = 0; counter < nr_counters; counter++)
 418                 create_counter(counter, cpu, pid);
 419
 420         nr_cpu++;
 421 }
 422
 423 static int __cmd_record(int argc, const char **argv)
 424 {
 425         int i, counter;
 426         struct stat st;
 427         pid_t pid;
 428         int flags;
 429         int ret;
 430
 431         page_size = sysconf(_SC_PAGE_SIZE);
 432         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 433         assert(nr_cpus <= MAX_NR_CPUS);
 434         assert(nr_cpus >= 0);
 435
 436         if (!stat(output_name, &st) && !force && !append_file) {
 437                 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 438                                 output_name);
 439                 exit(-1);
 440         }
 441
 442         flags = O_CREAT|O_RDWR;
 443         if (append_file)
 444                 flags |= O_APPEND;
 445         else
 446                 flags |= O_TRUNC;
 447
 448         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 449         if (output < 0) {
 450                 perror("failed to create output file");
 451                 exit(-1);
 452         }
 453
 454         if (!system_wide) {
 455                 open_counters(-1, target_pid != -1 ? target_pid : getpid());
 456         } else for (i = 0; i < nr_cpus; i++)
 457                 open_counters(i, target_pid);
 458
 459         signal(SIGCHLD, sig_handler);
 460         signal(SIGINT, sig_handler);
 461
 462         if (target_pid == -1 && argc) {
 463                 pid = fork();
 464                 if (pid < 0)
 465                         perror("failed to fork");
 466
 467                 if (!pid) {
 468                         if (execvp(argv[0], (char **)argv)) {
 469                                 perror(argv[0]);
 470                                 exit(-1);
 471                         }
 472                 }
 473         }
 474
 475         if (realtime_prio) {
 476                 struct sched_param param;
 477
 478                 param.sched_priority = realtime_prio;
 479                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 480                         printf("Could not set realtime priority.\n");
 481                         exit(-1);
 482                 }
 483         }
 484
 485         if (system_wide)
 486                 synthesize_samples();
 487
 488         while (!done) {
 489                 int hits = samples;
 490
 491                 for (i = 0; i < nr_cpu; i++) {
 492                         for (counter = 0; counter < nr_counters; counter++)
 493                                 mmap_read(&mmap_array[i][counter]);
 494                 }
 495
 496                 if (hits == samples)
 497                         ret = poll(event_array, nr_poll, 100);
 498         }
 499
 500         /*
 501          * Approximate RIP event size: 24 bytes.
 502          */
 503         fprintf(stderr,
 504                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 505                 (double)bytes_written / 1024.0 / 1024.0,
 506                 output_name,
 507                 bytes_written / 24);
 508
 509         return 0;
 510 }
 511
 512 static const char * const record_usage[] = {
 513         "perf record [<options>] [<command>]",
 514         "perf record [<options>] -- <command> [<options>]",
 515         NULL
 516 };
 517
 518 static const struct option options[] = {
 519         OPT_CALLBACK('e', "event", NULL, "event",
 520                      "event selector. use 'perf list' to list available events",
 521                      parse_events),
 522         OPT_INTEGER('p', "pid", &target_pid,
 523                     "record events on existing pid"),
 524         OPT_INTEGER('r', "realtime", &realtime_prio,
 525                     "collect data with this RT SCHED_FIFO priority"),
 526         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 527                             "system-wide collection from all CPUs"),
 528         OPT_BOOLEAN('A', "append", &append_file,
 529                             "append to the output file to do incremental profiling"),
 530         OPT_BOOLEAN('f', "force", &force,
 531                         "overwrite existing data file"),
 532         OPT_LONG('c', "count", &default_interval,
 533                     "event period to sample"),
 534         OPT_STRING('o', "output", &output_name, "file",
 535                     "output file name"),
 536         OPT_BOOLEAN('i', "inherit", &inherit,
 537                     "child tasks inherit counters"),
 538         OPT_INTEGER('F', "freq", &freq,
 539                     "profile at this frequency"),
 540         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 541                     "number of mmap data pages"),
 542         OPT_BOOLEAN('v', "verbose", &verbose,
 543                     "be more verbose (show counter open errors, etc)"),
 544         OPT_END()
 545 };
 546
 547 int cmd_record(int argc, const char **argv, const char *prefix)
 548 {
 549         int counter;
 550
 551         argc = parse_options(argc, argv, options, record_usage, 0);
 552         if (!argc && target_pid == -1 && !system_wide)
 553                 usage_with_options(record_usage, options);
 554
 555         if (!nr_counters)
 556                 nr_counters = 1;
 557
 558         for (counter = 0; counter < nr_counters; counter++) {
 559                 if (attrs[counter].sample_period)
 560                         continue;
 561
 562                 attrs[counter].sample_period = default_interval;
 563         }
 564
 565         return __cmd_record(argc, argv);
 566 }