git.oblomov.eu Git - linux-2.6/blob - Documentation/perf_counter/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 long                     default_interval = 100000;
  24 static long                     event_count[MAX_COUNTERS];
  25
  26 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  27 static int                      nr_cpus                         = 0;
  28 static unsigned int             page_size;
  29 static unsigned int             mmap_pages                      = 128;
  30 static int                      output;
  31 static const char               *output_name                    = "perf.data";
  32 static int                      group                           = 0;
  33 static unsigned int             realtime_prio                   = 0;
  34 static int                      system_wide                     = 0;
  35 static pid_t                    target_pid                      = -1;
  36 static int                      inherit                         = 1;
  37 static int                      force                           = 0;
  38 static int                      append_file                     = 0;
  39
  40 const unsigned int default_count[] = {
  41         1000000,
  42         1000000,
  43           10000,
  44           10000,
  45         1000000,
  46           10000,
  47 };
  48
  49 struct mmap_data {
  50         int counter;
  51         void *base;
  52         unsigned int mask;
  53         unsigned int prev;
  54 };
  55
  56 static unsigned int mmap_read_head(struct mmap_data *md)
  57 {
  58         struct perf_counter_mmap_page *pc = md->base;
  59         int head;
  60
  61         head = pc->data_head;
  62         rmb();
  63
  64         return head;
  65 }
  66
  67 static long events;
  68 static struct timeval last_read, this_read;
  69
  70 static __u64 bytes_written;
  71
  72 static void mmap_read(struct mmap_data *md)
  73 {
  74         unsigned int head = mmap_read_head(md);
  75         unsigned int old = md->prev;
  76         unsigned char *data = md->base + page_size;
  77         unsigned long size;
  78         void *buf;
  79         int diff;
  80
  81         gettimeofday(&this_read, NULL);
  82
  83         /*
  84          * If we're further behind than half the buffer, there's a chance
  85          * the writer will bite our tail and screw up the events under us.
  86          *
  87          * If we somehow ended up ahead of the head, we got messed up.
  88          *
  89          * In either case, truncate and restart at head.
  90          */
  91         diff = head - old;
  92         if (diff > md->mask / 2 || diff < 0) {
  93                 struct timeval iv;
  94                 unsigned long msecs;
  95
  96                 timersub(&this_read, &last_read, &iv);
  97                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
  98
  99                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 100                                 "  Last read %lu msecs ago.\n", msecs);
 101
 102                 /*
 103                  * head points to a known good entry, start there.
 104                  */
 105                 old = head;
 106         }
 107
 108         last_read = this_read;
 109
 110         if (old != head)
 111                 events++;
 112
 113         size = head - old;
 114
 115         if ((old & md->mask) + size != (head & md->mask)) {
 116                 buf = &data[old & md->mask];
 117                 size = md->mask + 1 - (old & md->mask);
 118                 old += size;
 119
 120                 while (size) {
 121                         int ret = write(output, buf, size);
 122
 123                         if (ret < 0)
 124                                 die("failed to write");
 125
 126                         size -= ret;
 127                         buf += ret;
 128
 129                         bytes_written += ret;
 130                 }
 131         }
 132
 133         buf = &data[old & md->mask];
 134         size = head - old;
 135         old += size;
 136
 137         while (size) {
 138                 int ret = write(output, buf, size);
 139
 140                 if (ret < 0)
 141                         die("failed to write");
 142
 143                 size -= ret;
 144                 buf += ret;
 145
 146                 bytes_written += ret;
 147         }
 148
 149         md->prev = old;
 150 }
 151
 152 static volatile int done = 0;
 153
 154 static void sig_handler(int sig)
 155 {
 156         done = 1;
 157 }
 158
 159 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
 160 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
 161
 162 static int nr_poll;
 163 static int nr_cpu;
 164
 165 struct mmap_event {
 166         struct perf_event_header        header;
 167         __u32                           pid;
 168         __u32                           tid;
 169         __u64                           start;
 170         __u64                           len;
 171         __u64                           pgoff;
 172         char                            filename[PATH_MAX];
 173 };
 174
 175 struct comm_event {
 176         struct perf_event_header        header;
 177         __u32                           pid;
 178         __u32                           tid;
 179         char                            comm[16];
 180 };
 181
 182 static void pid_synthesize_comm_event(pid_t pid, int full)
 183 {
 184         struct comm_event comm_ev;
 185         char filename[PATH_MAX];
 186         char bf[BUFSIZ];
 187         int fd, ret;
 188         size_t size;
 189         char *field, *sep;
 190         DIR *tasks;
 191         struct dirent dirent, *next;
 192
 193         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 194
 195         fd = open(filename, O_RDONLY);
 196         if (fd < 0) {
 197                 fprintf(stderr, "couldn't open %s\n", filename);
 198                 exit(EXIT_FAILURE);
 199         }
 200         if (read(fd, bf, sizeof(bf)) < 0) {
 201                 fprintf(stderr, "couldn't read %s\n", filename);
 202                 exit(EXIT_FAILURE);
 203         }
 204         close(fd);
 205
 206         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 207         memset(&comm_ev, 0, sizeof(comm_ev));
 208         field = strchr(bf, '(');
 209         if (field == NULL)
 210                 goto out_failure;
 211         sep = strchr(++field, ')');
 212         if (sep == NULL)
 213                 goto out_failure;
 214         size = sep - field;
 215         memcpy(comm_ev.comm, field, size++);
 216
 217         comm_ev.pid = pid;
 218         comm_ev.header.type = PERF_EVENT_COMM;
 219         size = ALIGN(size, sizeof(uint64_t));
 220         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 221
 222         if (!full) {
 223                 comm_ev.tid = pid;
 224
 225                 ret = write(output, &comm_ev, comm_ev.header.size);
 226                 if (ret < 0) {
 227                         perror("failed to write");
 228                         exit(-1);
 229                 }
 230                 return;
 231         }
 232
 233         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 234
 235         tasks = opendir(filename);
 236         while (!readdir_r(tasks, &dirent, &next) && next) {
 237                 char *end;
 238                 pid = strtol(dirent.d_name, &end, 10);
 239                 if (*end)
 240                         continue;
 241
 242                 comm_ev.tid = pid;
 243
 244                 ret = write(output, &comm_ev, comm_ev.header.size);
 245                 if (ret < 0) {
 246                         perror("failed to write");
 247                         exit(-1);
 248                 }
 249         }
 250         closedir(tasks);
 251         return;
 252
 253 out_failure:
 254         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 255                 filename);
 256         exit(EXIT_FAILURE);
 257 }
 258
 259 static void pid_synthesize_mmap_events(pid_t pid)
 260 {
 261         char filename[PATH_MAX];
 262         FILE *fp;
 263
 264         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 265
 266         fp = fopen(filename, "r");
 267         if (fp == NULL) {
 268                 fprintf(stderr, "couldn't open %s\n", filename);
 269                 exit(EXIT_FAILURE);
 270         }
 271         while (1) {
 272                 char bf[BUFSIZ], *pbf = bf;
 273                 struct mmap_event mmap_ev = {
 274                         .header.type = PERF_EVENT_MMAP,
 275                 };
 276                 int n;
 277                 size_t size;
 278                 if (fgets(bf, sizeof(bf), fp) == NULL)
 279                         break;
 280
 281                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 282                 n = hex2u64(pbf, &mmap_ev.start);
 283                 if (n < 0)
 284                         continue;
 285                 pbf += n + 1;
 286                 n = hex2u64(pbf, &mmap_ev.len);
 287                 if (n < 0)
 288                         continue;
 289                 pbf += n + 3;
 290                 if (*pbf == 'x') { /* vm_exec */
 291                         char *execname = strrchr(bf, ' ');
 292
 293                         if (execname == NULL || execname[1] != '/')
 294                                 continue;
 295
 296                         execname += 1;
 297                         size = strlen(execname);
 298                         execname[size - 1] = '\0'; /* Remove \n */
 299                         memcpy(mmap_ev.filename, execname, size);
 300                         size = ALIGN(size, sizeof(uint64_t));
 301                         mmap_ev.len -= mmap_ev.start;
 302                         mmap_ev.header.size = (sizeof(mmap_ev) -
 303                                                (sizeof(mmap_ev.filename) - size));
 304                         mmap_ev.pid = pid;
 305                         mmap_ev.tid = pid;
 306
 307                         if (write(output, &mmap_ev, mmap_ev.header.size) < 0) {
 308                                 perror("failed to write");
 309                                 exit(-1);
 310                         }
 311                 }
 312         }
 313
 314         fclose(fp);
 315 }
 316
 317 static void synthesize_events(void)
 318 {
 319         DIR *proc;
 320         struct dirent dirent, *next;
 321
 322         proc = opendir("/proc");
 323
 324         while (!readdir_r(proc, &dirent, &next) && next) {
 325                 char *end;
 326                 pid_t pid;
 327
 328                 pid = strtol(dirent.d_name, &end, 10);
 329                 if (*end) /* only interested in proper numerical dirents */
 330                         continue;
 331
 332                 pid_synthesize_comm_event(pid, 1);
 333                 pid_synthesize_mmap_events(pid);
 334         }
 335
 336         closedir(proc);
 337 }
 338
 339 static void open_counters(int cpu, pid_t pid)
 340 {
 341         struct perf_counter_attr attr;
 342         int counter, group_fd;
 343         int track = 1;
 344
 345         if (pid > 0) {
 346                 pid_synthesize_comm_event(pid, 0);
 347                 pid_synthesize_mmap_events(pid);
 348         }
 349
 350         group_fd = -1;
 351         for (counter = 0; counter < nr_counters; counter++) {
 352
 353                 memset(&attr, 0, sizeof(attr));
 354                 attr.config             = event_id[counter];
 355                 attr.sample_period      = event_count[counter];
 356                 attr.sample_type        = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 357                 attr.mmap               = track;
 358                 attr.comm               = track;
 359                 attr.inherit    = (cpu < 0) && inherit;
 360
 361                 track = 0; // only the first counter needs these
 362
 363                 fd[nr_cpu][counter] =
 364                         sys_perf_counter_open(&attr, pid, cpu, group_fd, 0);
 365
 366                 if (fd[nr_cpu][counter] < 0) {
 367                         int err = errno;
 368
 369                         error("syscall returned with %d (%s)\n",
 370                                         fd[nr_cpu][counter], strerror(err));
 371                         if (err == EPERM)
 372                                 printf("Are you root?\n");
 373                         exit(-1);
 374                 }
 375                 assert(fd[nr_cpu][counter] >= 0);
 376                 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 377
 378                 /*
 379                  * First counter acts as the group leader:
 380                  */
 381                 if (group && group_fd == -1)
 382                         group_fd = fd[nr_cpu][counter];
 383
 384                 event_array[nr_poll].fd = fd[nr_cpu][counter];
 385                 event_array[nr_poll].events = POLLIN;
 386                 nr_poll++;
 387
 388                 mmap_array[nr_cpu][counter].counter = counter;
 389                 mmap_array[nr_cpu][counter].prev = 0;
 390                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 391                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 392                                 PROT_READ, MAP_SHARED, fd[nr_cpu][counter], 0);
 393                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 394                         error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 395                         exit(-1);
 396                 }
 397         }
 398         nr_cpu++;
 399 }
 400
 401 static int __cmd_record(int argc, const char **argv)
 402 {
 403         int i, counter;
 404         struct stat st;
 405         pid_t pid;
 406         int flags;
 407         int ret;
 408
 409         page_size = sysconf(_SC_PAGE_SIZE);
 410         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 411         assert(nr_cpus <= MAX_NR_CPUS);
 412         assert(nr_cpus >= 0);
 413
 414         if (!stat(output_name, &st) && !force && !append_file) {
 415                 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 416                                 output_name);
 417                 exit(-1);
 418         }
 419
 420         flags = O_CREAT|O_RDWR;
 421         if (append_file)
 422                 flags |= O_APPEND;
 423         else
 424                 flags |= O_TRUNC;
 425
 426         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 427         if (output < 0) {
 428                 perror("failed to create output file");
 429                 exit(-1);
 430         }
 431
 432         if (!system_wide) {
 433                 open_counters(-1, target_pid != -1 ? target_pid : getpid());
 434         } else for (i = 0; i < nr_cpus; i++)
 435                 open_counters(i, target_pid);
 436
 437         signal(SIGCHLD, sig_handler);
 438         signal(SIGINT, sig_handler);
 439
 440         if (target_pid == -1 && argc) {
 441                 pid = fork();
 442                 if (pid < 0)
 443                         perror("failed to fork");
 444
 445                 if (!pid) {
 446                         if (execvp(argv[0], (char **)argv)) {
 447                                 perror(argv[0]);
 448                                 exit(-1);
 449                         }
 450                 }
 451         }
 452
 453         if (realtime_prio) {
 454                 struct sched_param param;
 455
 456                 param.sched_priority = realtime_prio;
 457                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 458                         printf("Could not set realtime priority.\n");
 459                         exit(-1);
 460                 }
 461         }
 462
 463         if (system_wide)
 464                 synthesize_events();
 465
 466         while (!done) {
 467                 int hits = events;
 468
 469                 for (i = 0; i < nr_cpu; i++) {
 470                         for (counter = 0; counter < nr_counters; counter++)
 471                                 mmap_read(&mmap_array[i][counter]);
 472                 }
 473
 474                 if (hits == events)
 475                         ret = poll(event_array, nr_poll, 100);
 476         }
 477
 478         /*
 479          * Approximate RIP event size: 24 bytes.
 480          */
 481         fprintf(stderr,
 482                 "[ perf record: Captured and wrote %.3f MB %s (~%lld events) ]\n",
 483                 (double)bytes_written / 1024.0 / 1024.0,
 484                 output_name,
 485                 bytes_written / 24);
 486
 487         return 0;
 488 }
 489
 490 static const char * const record_usage[] = {
 491         "perf record [<options>] [<command>]",
 492         "perf record [<options>] -- <command> [<options>]",
 493         NULL
 494 };
 495
 496 static char events_help_msg[EVENTS_HELP_MAX];
 497
 498 static const struct option options[] = {
 499         OPT_CALLBACK('e', "event", NULL, "event",
 500                      events_help_msg, parse_events),
 501         OPT_INTEGER('p', "pid", &target_pid,
 502                     "record events on existing pid"),
 503         OPT_INTEGER('r', "realtime", &realtime_prio,
 504                     "collect data with this RT SCHED_FIFO priority"),
 505         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 506                             "system-wide collection from all CPUs"),
 507         OPT_BOOLEAN('A', "append", &append_file,
 508                             "append to the output file to do incremental profiling"),
 509         OPT_BOOLEAN('f', "force", &force,
 510                         "overwrite existing data file"),
 511         OPT_LONG('c', "count", &default_interval,
 512                     "event period to sample"),
 513         OPT_STRING('o', "output", &output_name, "file",
 514                     "output file name"),
 515         OPT_BOOLEAN('i', "inherit", &inherit,
 516                     "child tasks inherit counters"),
 517         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 518                     "number of mmap data pages"),
 519         OPT_END()
 520 };
 521
 522 int cmd_record(int argc, const char **argv, const char *prefix)
 523 {
 524         int counter;
 525
 526         create_events_help(events_help_msg);
 527
 528         argc = parse_options(argc, argv, options, record_usage, 0);
 529         if (!argc && target_pid == -1 && !system_wide)
 530                 usage_with_options(record_usage, options);
 531
 532         if (!nr_counters) {
 533                 nr_counters = 1;
 534                 event_id[0] = 0;
 535         }
 536
 537         for (counter = 0; counter < nr_counters; counter++) {
 538                 if (event_count[counter])
 539                         continue;
 540
 541                 event_count[counter] = default_interval;
 542         }
 543
 544         return __cmd_record(argc, argv);
 545 }