SAFE public projects git trees. - safe/jmp/linux-2.6/blob - Documentation/perf_counter/builtin-record.c

   1
   2
   3 #include "perf.h"
   4 #include "builtin.h"
   5 #include "util/util.h"
   6 #include "util/parse-options.h"
   7 #include "util/parse-events.h"
   8 #include "util/string.h"
   9
  10 #include <unistd.h>
  11 #include <sched.h>
  12
  13 #define ALIGN(x, a)             __ALIGN_MASK(x, (typeof(x))(a)-1)
  14 #define __ALIGN_MASK(x, mask)   (((x)+(mask))&~(mask))
  15
  16 static int                      default_interval = 100000;
  17 static int                      event_count[MAX_COUNTERS];
  18
  19 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  20 static int                      nr_cpus                         =  0;
  21 static unsigned int             page_size;
  22 static unsigned int             mmap_pages                      = 16;
  23 static int                      output;
  24 static const char               *output_name                    = "perf.data";
  25 static int                      group                           = 0;
  26 static unsigned int             realtime_prio                   = 0;
  27 static int                      system_wide                     = 0;
  28 static pid_t                    target_pid                      = -1;
  29 static int                      inherit                         = 1;
  30 static int                      force                           = 0;
  31
  32 const unsigned int default_count[] = {
  33         1000000,
  34         1000000,
  35           10000,
  36           10000,
  37         1000000,
  38           10000,
  39 };
  40
  41 struct mmap_data {
  42         int counter;
  43         void *base;
  44         unsigned int mask;
  45         unsigned int prev;
  46 };
  47
  48 static unsigned int mmap_read_head(struct mmap_data *md)
  49 {
  50         struct perf_counter_mmap_page *pc = md->base;
  51         int head;
  52
  53         head = pc->data_head;
  54         rmb();
  55
  56         return head;
  57 }
  58
  59 static long events;
  60 static struct timeval last_read, this_read;
  61
  62 static void mmap_read(struct mmap_data *md)
  63 {
  64         unsigned int head = mmap_read_head(md);
  65         unsigned int old = md->prev;
  66         unsigned char *data = md->base + page_size;
  67         unsigned long size;
  68         void *buf;
  69         int diff;
  70
  71         gettimeofday(&this_read, NULL);
  72
  73         /*
  74          * If we're further behind than half the buffer, there's a chance
  75          * the writer will bite our tail and screw up the events under us.
  76          *
  77          * If we somehow ended up ahead of the head, we got messed up.
  78          *
  79          * In either case, truncate and restart at head.
  80          */
  81         diff = head - old;
  82         if (diff > md->mask / 2 || diff < 0) {
  83                 struct timeval iv;
  84                 unsigned long msecs;
  85
  86                 timersub(&this_read, &last_read, &iv);
  87                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
  88
  89                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
  90                                 "  Last read %lu msecs ago.\n", msecs);
  91
  92                 /*
  93                  * head points to a known good entry, start there.
  94                  */
  95                 old = head;
  96         }
  97
  98         last_read = this_read;
  99
 100         if (old != head)
 101                 events++;
 102
 103         size = head - old;
 104
 105         if ((old & md->mask) + size != (head & md->mask)) {
 106                 buf = &data[old & md->mask];
 107                 size = md->mask + 1 - (old & md->mask);
 108                 old += size;
 109                 while (size) {
 110                         int ret = write(output, buf, size);
 111                         if (ret < 0) {
 112                                 perror("failed to write");
 113                                 exit(-1);
 114                         }
 115                         size -= ret;
 116                         buf += ret;
 117                 }
 118         }
 119
 120         buf = &data[old & md->mask];
 121         size = head - old;
 122         old += size;
 123         while (size) {
 124                 int ret = write(output, buf, size);
 125                 if (ret < 0) {
 126                         perror("failed to write");
 127                         exit(-1);
 128                 }
 129                 size -= ret;
 130                 buf += ret;
 131         }
 132
 133         md->prev = old;
 134 }
 135
 136 static volatile int done = 0;
 137
 138 static void sig_handler(int sig)
 139 {
 140         done = 1;
 141 }
 142
 143 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
 144 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
 145
 146 static int nr_poll;
 147 static int nr_cpu;
 148
 149 struct mmap_event {
 150         struct perf_event_header        header;
 151         __u32                           pid;
 152         __u32                           tid;
 153         __u64                           start;
 154         __u64                           len;
 155         __u64                           pgoff;
 156         char                            filename[PATH_MAX];
 157 };
 158
 159 struct comm_event {
 160         struct perf_event_header        header;
 161         __u32                           pid;
 162         __u32                           tid;
 163         char                            comm[16];
 164 };
 165
 166 static void pid_synthesize_comm_event(pid_t pid, int full)
 167 {
 168         struct comm_event comm_ev;
 169         char filename[PATH_MAX];
 170         char bf[BUFSIZ];
 171         int fd, ret;
 172         size_t size;
 173         char *field, *sep;
 174         DIR *tasks;
 175         struct dirent dirent, *next;
 176
 177         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 178
 179         fd = open(filename, O_RDONLY);
 180         if (fd < 0) {
 181                 fprintf(stderr, "couldn't open %s\n", filename);
 182                 exit(EXIT_FAILURE);
 183         }
 184         if (read(fd, bf, sizeof(bf)) < 0) {
 185                 fprintf(stderr, "couldn't read %s\n", filename);
 186                 exit(EXIT_FAILURE);
 187         }
 188         close(fd);
 189
 190         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 191         memset(&comm_ev, 0, sizeof(comm_ev));
 192         field = strchr(bf, '(');
 193         if (field == NULL)
 194                 goto out_failure;
 195         sep = strchr(++field, ')');
 196         if (sep == NULL)
 197                 goto out_failure;
 198         size = sep - field;
 199         memcpy(comm_ev.comm, field, size++);
 200
 201         comm_ev.pid = pid;
 202         comm_ev.header.type = PERF_EVENT_COMM;
 203         size = ALIGN(size, sizeof(uint64_t));
 204         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 205
 206         if (!full) {
 207                 comm_ev.tid = pid;
 208
 209                 ret = write(output, &comm_ev, comm_ev.header.size);
 210                 if (ret < 0) {
 211                         perror("failed to write");
 212                         exit(-1);
 213                 }
 214                 return;
 215         }
 216
 217         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 218
 219         tasks = opendir(filename);
 220         while (!readdir_r(tasks, &dirent, &next) && next) {
 221                 char *end;
 222                 pid = strtol(dirent.d_name, &end, 10);
 223                 if (*end)
 224                         continue;
 225
 226                 comm_ev.tid = pid;
 227
 228                 ret = write(output, &comm_ev, comm_ev.header.size);
 229                 if (ret < 0) {
 230                         perror("failed to write");
 231                         exit(-1);
 232                 }
 233         }
 234         closedir(tasks);
 235         return;
 236
 237 out_failure:
 238         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 239                 filename);
 240         exit(EXIT_FAILURE);
 241 }
 242
 243 static void pid_synthesize_mmap_events(pid_t pid)
 244 {
 245         char filename[PATH_MAX];
 246         FILE *fp;
 247
 248         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 249
 250         fp = fopen(filename, "r");
 251         if (fp == NULL) {
 252                 fprintf(stderr, "couldn't open %s\n", filename);
 253                 exit(EXIT_FAILURE);
 254         }
 255         while (1) {
 256                 char bf[BUFSIZ], *pbf = bf;
 257                 struct mmap_event mmap_ev = {
 258                         .header.type = PERF_EVENT_MMAP,
 259                 };
 260                 int n;
 261                 size_t size;
 262                 if (fgets(bf, sizeof(bf), fp) == NULL)
 263                         break;
 264
 265                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 266                 n = hex2u64(pbf, &mmap_ev.start);
 267                 if (n < 0)
 268                         continue;
 269                 pbf += n + 1;
 270                 n = hex2u64(pbf, &mmap_ev.len);
 271                 if (n < 0)
 272                         continue;
 273                 pbf += n + 3;
 274                 if (*pbf == 'x') { /* vm_exec */
 275                         char *execname = strrchr(bf, ' ');
 276
 277                         if (execname == NULL || execname[1] != '/')
 278                                 continue;
 279
 280                         execname += 1;
 281                         size = strlen(execname);
 282                         execname[size - 1] = '\0'; /* Remove \n */
 283                         memcpy(mmap_ev.filename, execname, size);
 284                         size = ALIGN(size, sizeof(uint64_t));
 285                         mmap_ev.len -= mmap_ev.start;
 286                         mmap_ev.header.size = (sizeof(mmap_ev) -
 287                                                (sizeof(mmap_ev.filename) - size));
 288                         mmap_ev.pid = pid;
 289                         mmap_ev.tid = pid;
 290
 291                         if (write(output, &mmap_ev, mmap_ev.header.size) < 0) {
 292                                 perror("failed to write");
 293                                 exit(-1);
 294                         }
 295                 }
 296         }
 297
 298         fclose(fp);
 299 }
 300
 301 static void synthesize_events(void)
 302 {
 303         DIR *proc;
 304         struct dirent dirent, *next;
 305
 306         proc = opendir("/proc");
 307
 308         while (!readdir_r(proc, &dirent, &next) && next) {
 309                 char *end;
 310                 pid_t pid;
 311
 312                 pid = strtol(dirent.d_name, &end, 10);
 313                 if (*end) /* only interested in proper numerical dirents */
 314                         continue;
 315
 316                 pid_synthesize_comm_event(pid, 1);
 317                 pid_synthesize_mmap_events(pid);
 318         }
 319
 320         closedir(proc);
 321 }
 322
 323 static void open_counters(int cpu, pid_t pid)
 324 {
 325         struct perf_counter_hw_event hw_event;
 326         int counter, group_fd;
 327         int track = 1;
 328
 329         if (pid > 0) {
 330                 pid_synthesize_comm_event(pid, 0);
 331                 pid_synthesize_mmap_events(pid);
 332         }
 333
 334         group_fd = -1;
 335         for (counter = 0; counter < nr_counters; counter++) {
 336
 337                 memset(&hw_event, 0, sizeof(hw_event));
 338                 hw_event.config         = event_id[counter];
 339                 hw_event.irq_period     = event_count[counter];
 340                 hw_event.record_type    = PERF_RECORD_IP | PERF_RECORD_TID;
 341                 hw_event.mmap           = track;
 342                 hw_event.comm           = track;
 343                 hw_event.inherit        = (cpu < 0) && inherit;
 344
 345                 track = 0; // only the first counter needs these
 346
 347                 fd[nr_cpu][counter] =
 348                         sys_perf_counter_open(&hw_event, pid, cpu, group_fd, 0);
 349
 350                 if (fd[nr_cpu][counter] < 0) {
 351                         int err = errno;
 352                         printf("kerneltop error: syscall returned with %d (%s)\n",
 353                                         fd[nr_cpu][counter], strerror(err));
 354                         if (err == EPERM)
 355                                 printf("Are you root?\n");
 356                         exit(-1);
 357                 }
 358                 assert(fd[nr_cpu][counter] >= 0);
 359                 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 360
 361                 /*
 362                  * First counter acts as the group leader:
 363                  */
 364                 if (group && group_fd == -1)
 365                         group_fd = fd[nr_cpu][counter];
 366
 367                 event_array[nr_poll].fd = fd[nr_cpu][counter];
 368                 event_array[nr_poll].events = POLLIN;
 369                 nr_poll++;
 370
 371                 mmap_array[nr_cpu][counter].counter = counter;
 372                 mmap_array[nr_cpu][counter].prev = 0;
 373                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 374                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 375                                 PROT_READ, MAP_SHARED, fd[nr_cpu][counter], 0);
 376                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 377                         printf("kerneltop error: failed to mmap with %d (%s)\n",
 378                                         errno, strerror(errno));
 379                         exit(-1);
 380                 }
 381         }
 382         nr_cpu++;
 383 }
 384
 385 static int __cmd_record(int argc, const char **argv)
 386 {
 387         int i, counter;
 388         pid_t pid;
 389         int ret;
 390         struct stat st;
 391
 392         page_size = sysconf(_SC_PAGE_SIZE);
 393         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 394         assert(nr_cpus <= MAX_NR_CPUS);
 395         assert(nr_cpus >= 0);
 396
 397         if (!stat(output_name, &st) && !force) {
 398                 fprintf(stderr, "Error, output file: %s exists, use -f to overwrite.\n",
 399                                 output_name);
 400                 exit(-1);
 401         }
 402
 403         output = open(output_name, O_CREAT|O_TRUNC|O_RDWR, S_IRUSR|S_IWUSR);
 404         if (output < 0) {
 405                 perror("failed to create output file");
 406                 exit(-1);
 407         }
 408
 409         if (!system_wide) {
 410                 open_counters(-1, target_pid != -1 ? target_pid : 0);
 411         } else for (i = 0; i < nr_cpus; i++)
 412                 open_counters(i, target_pid);
 413
 414         signal(SIGCHLD, sig_handler);
 415         signal(SIGINT, sig_handler);
 416
 417         if (target_pid == -1 && argc) {
 418                 pid = fork();
 419                 if (pid < 0)
 420                         perror("failed to fork");
 421
 422                 if (!pid) {
 423                         if (execvp(argv[0], (char **)argv)) {
 424                                 perror(argv[0]);
 425                                 exit(-1);
 426                         }
 427                 }
 428         }
 429
 430         if (realtime_prio) {
 431                 struct sched_param param;
 432
 433                 param.sched_priority = realtime_prio;
 434                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 435                         printf("Could not set realtime priority.\n");
 436                         exit(-1);
 437                 }
 438         }
 439
 440         if (system_wide)
 441                 synthesize_events();
 442
 443         while (!done) {
 444                 int hits = events;
 445
 446                 for (i = 0; i < nr_cpu; i++) {
 447                         for (counter = 0; counter < nr_counters; counter++)
 448                                 mmap_read(&mmap_array[i][counter]);
 449                 }
 450
 451                 if (hits == events)
 452                         ret = poll(event_array, nr_poll, 100);
 453         }
 454
 455         return 0;
 456 }
 457
 458 static const char * const record_usage[] = {
 459         "perf record [<options>] [<command>]",
 460         "perf record [<options>] -- <command> [<options>]",
 461         NULL
 462 };
 463
 464 static char events_help_msg[EVENTS_HELP_MAX];
 465
 466 static const struct option options[] = {
 467         OPT_CALLBACK('e', "event", NULL, "event",
 468                      events_help_msg, parse_events),
 469         OPT_INTEGER('c', "count", &default_interval,
 470                     "event period to sample"),
 471         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 472                     "number of mmap data pages"),
 473         OPT_STRING('o', "output", &output_name, "file",
 474                     "output file name"),
 475         OPT_BOOLEAN('i', "inherit", &inherit,
 476                     "child tasks inherit counters"),
 477         OPT_INTEGER('p', "pid", &target_pid,
 478                     "record events on existing pid"),
 479         OPT_INTEGER('r', "realtime", &realtime_prio,
 480                     "collect data with this RT SCHED_FIFO priority"),
 481         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 482                             "system-wide collection from all CPUs"),
 483         OPT_BOOLEAN('f', "force", &force,
 484                         "overwrite existing data file"),
 485         OPT_END()
 486 };
 487
 488 int cmd_record(int argc, const char **argv, const char *prefix)
 489 {
 490         int counter;
 491
 492         create_events_help(events_help_msg);
 493
 494         argc = parse_options(argc, argv, options, record_usage, 0);
 495         if (!argc && target_pid == -1 && !system_wide)
 496                 usage_with_options(record_usage, options);
 497
 498         if (!nr_counters) {
 499                 nr_counters = 1;
 500                 event_id[0] = 0;
 501         }
 502
 503         for (counter = 0; counter < nr_counters; counter++) {
 504                 if (event_count[counter])
 505                         continue;
 506
 507                 event_count[counter] = default_interval;
 508         }
 509
 510         return __cmd_record(argc, argv);
 511 }