Documentation/perf_counter/builtin-record.c

   1 /*
   2  * perf record: Record the profile of a workload (or a CPU, or a PID) into
   3  * the perf.data output file - for later analysis via perf report.
   4  */
   5 #include "perf.h"
   6 #include "builtin.h"
   7 #include "util/util.h"
   8 #include "util/parse-options.h"
   9 #include "util/parse-events.h"
  10 #include "util/string.h"
  11
  12 #include <unistd.h>
  13 #include <sched.h>
  14
  15 #define ALIGN(x, a)             __ALIGN_MASK(x, (typeof(x))(a)-1)
  16 #define __ALIGN_MASK(x, mask)   (((x)+(mask))&~(mask))
  17
  18 static int                      default_interval = 100000;
  19 static int                      event_count[MAX_COUNTERS];
  20
  21 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  22 static int                      nr_cpus                         =  0;
  23 static unsigned int             page_size;
  24 static unsigned int             mmap_pages                      = 16;
  25 static int                      output;
  26 static const char               *output_name                    = "perf.data";
  27 static int                      group                           = 0;
  28 static unsigned int             realtime_prio                   = 0;
  29 static int                      system_wide                     = 0;
  30 static pid_t                    target_pid                      = -1;
  31 static int                      inherit                         = 1;
  32 static int                      force                           = 0;
  33 static int                      append_file                     = 0;
  34
  35 const unsigned int default_count[] = {
  36         1000000,
  37         1000000,
  38           10000,
  39           10000,
  40         1000000,
  41           10000,
  42 };
  43
  44 struct mmap_data {
  45         int counter;
  46         void *base;
  47         unsigned int mask;
  48         unsigned int prev;
  49 };
  50
  51 static unsigned int mmap_read_head(struct mmap_data *md)
  52 {
  53         struct perf_counter_mmap_page *pc = md->base;
  54         int head;
  55
  56         head = pc->data_head;
  57         rmb();
  58
  59         return head;
  60 }
  61
  62 static long events;
  63 static struct timeval last_read, this_read;
  64
  65 static void mmap_read(struct mmap_data *md)
  66 {
  67         unsigned int head = mmap_read_head(md);
  68         unsigned int old = md->prev;
  69         unsigned char *data = md->base + page_size;
  70         unsigned long size;
  71         void *buf;
  72         int diff;
  73
  74         gettimeofday(&this_read, NULL);
  75
  76         /*
  77          * If we're further behind than half the buffer, there's a chance
  78          * the writer will bite our tail and screw up the events under us.
  79          *
  80          * If we somehow ended up ahead of the head, we got messed up.
  81          *
  82          * In either case, truncate and restart at head.
  83          */
  84         diff = head - old;
  85         if (diff > md->mask / 2 || diff < 0) {
  86                 struct timeval iv;
  87                 unsigned long msecs;
  88
  89                 timersub(&this_read, &last_read, &iv);
  90                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
  91
  92                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
  93                                 "  Last read %lu msecs ago.\n", msecs);
  94
  95                 /*
  96                  * head points to a known good entry, start there.
  97                  */
  98                 old = head;
  99         }
 100
 101         last_read = this_read;
 102
 103         if (old != head)
 104                 events++;
 105
 106         size = head - old;
 107
 108         if ((old & md->mask) + size != (head & md->mask)) {
 109                 buf = &data[old & md->mask];
 110                 size = md->mask + 1 - (old & md->mask);
 111                 old += size;
 112                 while (size) {
 113                         int ret = write(output, buf, size);
 114                         if (ret < 0) {
 115                                 perror("failed to write");
 116                                 exit(-1);
 117                         }
 118                         size -= ret;
 119                         buf += ret;
 120                 }
 121         }
 122
 123         buf = &data[old & md->mask];
 124         size = head - old;
 125         old += size;
 126         while (size) {
 127                 int ret = write(output, buf, size);
 128                 if (ret < 0) {
 129                         perror("failed to write");
 130                         exit(-1);
 131                 }
 132                 size -= ret;
 133                 buf += ret;
 134         }
 135
 136         md->prev = old;
 137 }
 138
 139 static volatile int done = 0;
 140
 141 static void sig_handler(int sig)
 142 {
 143         done = 1;
 144 }
 145
 146 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
 147 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
 148
 149 static int nr_poll;
 150 static int nr_cpu;
 151
 152 struct mmap_event {
 153         struct perf_event_header        header;
 154         __u32                           pid;
 155         __u32                           tid;
 156         __u64                           start;
 157         __u64                           len;
 158         __u64                           pgoff;
 159         char                            filename[PATH_MAX];
 160 };
 161
 162 struct comm_event {
 163         struct perf_event_header        header;
 164         __u32                           pid;
 165         __u32                           tid;
 166         char                            comm[16];
 167 };
 168
 169 static void pid_synthesize_comm_event(pid_t pid, int full)
 170 {
 171         struct comm_event comm_ev;
 172         char filename[PATH_MAX];
 173         char bf[BUFSIZ];
 174         int fd, ret;
 175         size_t size;
 176         char *field, *sep;
 177         DIR *tasks;
 178         struct dirent dirent, *next;
 179
 180         snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
 181
 182         fd = open(filename, O_RDONLY);
 183         if (fd < 0) {
 184                 fprintf(stderr, "couldn't open %s\n", filename);
 185                 exit(EXIT_FAILURE);
 186         }
 187         if (read(fd, bf, sizeof(bf)) < 0) {
 188                 fprintf(stderr, "couldn't read %s\n", filename);
 189                 exit(EXIT_FAILURE);
 190         }
 191         close(fd);
 192
 193         /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
 194         memset(&comm_ev, 0, sizeof(comm_ev));
 195         field = strchr(bf, '(');
 196         if (field == NULL)
 197                 goto out_failure;
 198         sep = strchr(++field, ')');
 199         if (sep == NULL)
 200                 goto out_failure;
 201         size = sep - field;
 202         memcpy(comm_ev.comm, field, size++);
 203
 204         comm_ev.pid = pid;
 205         comm_ev.header.type = PERF_EVENT_COMM;
 206         size = ALIGN(size, sizeof(uint64_t));
 207         comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
 208
 209         if (!full) {
 210                 comm_ev.tid = pid;
 211
 212                 ret = write(output, &comm_ev, comm_ev.header.size);
 213                 if (ret < 0) {
 214                         perror("failed to write");
 215                         exit(-1);
 216                 }
 217                 return;
 218         }
 219
 220         snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
 221
 222         tasks = opendir(filename);
 223         while (!readdir_r(tasks, &dirent, &next) && next) {
 224                 char *end;
 225                 pid = strtol(dirent.d_name, &end, 10);
 226                 if (*end)
 227                         continue;
 228
 229                 comm_ev.tid = pid;
 230
 231                 ret = write(output, &comm_ev, comm_ev.header.size);
 232                 if (ret < 0) {
 233                         perror("failed to write");
 234                         exit(-1);
 235                 }
 236         }
 237         closedir(tasks);
 238         return;
 239
 240 out_failure:
 241         fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
 242                 filename);
 243         exit(EXIT_FAILURE);
 244 }
 245
 246 static void pid_synthesize_mmap_events(pid_t pid)
 247 {
 248         char filename[PATH_MAX];
 249         FILE *fp;
 250
 251         snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
 252
 253         fp = fopen(filename, "r");
 254         if (fp == NULL) {
 255                 fprintf(stderr, "couldn't open %s\n", filename);
 256                 exit(EXIT_FAILURE);
 257         }
 258         while (1) {
 259                 char bf[BUFSIZ], *pbf = bf;
 260                 struct mmap_event mmap_ev = {
 261                         .header.type = PERF_EVENT_MMAP,
 262                 };
 263                 int n;
 264                 size_t size;
 265                 if (fgets(bf, sizeof(bf), fp) == NULL)
 266                         break;
 267
 268                 /* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
 269                 n = hex2u64(pbf, &mmap_ev.start);
 270                 if (n < 0)
 271                         continue;
 272                 pbf += n + 1;
 273                 n = hex2u64(pbf, &mmap_ev.len);
 274                 if (n < 0)
 275                         continue;
 276                 pbf += n + 3;
 277                 if (*pbf == 'x') { /* vm_exec */
 278                         char *execname = strrchr(bf, ' ');
 279
 280                         if (execname == NULL || execname[1] != '/')
 281                                 continue;
 282
 283                         execname += 1;
 284                         size = strlen(execname);
 285                         execname[size - 1] = '\0'; /* Remove \n */
 286                         memcpy(mmap_ev.filename, execname, size);
 287                         size = ALIGN(size, sizeof(uint64_t));
 288                         mmap_ev.len -= mmap_ev.start;
 289                         mmap_ev.header.size = (sizeof(mmap_ev) -
 290                                                (sizeof(mmap_ev.filename) - size));
 291                         mmap_ev.pid = pid;
 292                         mmap_ev.tid = pid;
 293
 294                         if (write(output, &mmap_ev, mmap_ev.header.size) < 0) {
 295                                 perror("failed to write");
 296                                 exit(-1);
 297                         }
 298                 }
 299         }
 300
 301         fclose(fp);
 302 }
 303
 304 static void synthesize_events(void)
 305 {
 306         DIR *proc;
 307         struct dirent dirent, *next;
 308
 309         proc = opendir("/proc");
 310
 311         while (!readdir_r(proc, &dirent, &next) && next) {
 312                 char *end;
 313                 pid_t pid;
 314
 315                 pid = strtol(dirent.d_name, &end, 10);
 316                 if (*end) /* only interested in proper numerical dirents */
 317                         continue;
 318
 319                 pid_synthesize_comm_event(pid, 1);
 320                 pid_synthesize_mmap_events(pid);
 321         }
 322
 323         closedir(proc);
 324 }
 325
 326 static void open_counters(int cpu, pid_t pid)
 327 {
 328         struct perf_counter_attr attr;
 329         int counter, group_fd;
 330         int track = 1;
 331
 332         if (pid > 0) {
 333                 pid_synthesize_comm_event(pid, 0);
 334                 pid_synthesize_mmap_events(pid);
 335         }
 336
 337         group_fd = -1;
 338         for (counter = 0; counter < nr_counters; counter++) {
 339
 340                 memset(&attr, 0, sizeof(attr));
 341                 attr.config             = event_id[counter];
 342                 attr.sample_period      = event_count[counter];
 343                 attr.sample_type        = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 344                 attr.mmap               = track;
 345                 attr.comm               = track;
 346                 attr.inherit    = (cpu < 0) && inherit;
 347
 348                 track = 0; // only the first counter needs these
 349
 350                 fd[nr_cpu][counter] =
 351                         sys_perf_counter_open(&attr, pid, cpu, group_fd, 0);
 352
 353                 if (fd[nr_cpu][counter] < 0) {
 354                         int err = errno;
 355                         printf("kerneltop error: syscall returned with %d (%s)\n",
 356                                         fd[nr_cpu][counter], strerror(err));
 357                         if (err == EPERM)
 358                                 printf("Are you root?\n");
 359                         exit(-1);
 360                 }
 361                 assert(fd[nr_cpu][counter] >= 0);
 362                 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
 363
 364                 /*
 365                  * First counter acts as the group leader:
 366                  */
 367                 if (group && group_fd == -1)
 368                         group_fd = fd[nr_cpu][counter];
 369
 370                 event_array[nr_poll].fd = fd[nr_cpu][counter];
 371                 event_array[nr_poll].events = POLLIN;
 372                 nr_poll++;
 373
 374                 mmap_array[nr_cpu][counter].counter = counter;
 375                 mmap_array[nr_cpu][counter].prev = 0;
 376                 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
 377                 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
 378                                 PROT_READ, MAP_SHARED, fd[nr_cpu][counter], 0);
 379                 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
 380                         printf("kerneltop error: failed to mmap with %d (%s)\n",
 381                                         errno, strerror(errno));
 382                         exit(-1);
 383                 }
 384         }
 385         nr_cpu++;
 386 }
 387
 388 static int __cmd_record(int argc, const char **argv)
 389 {
 390         int i, counter;
 391         struct stat st;
 392         pid_t pid;
 393         int flags;
 394         int ret;
 395
 396         page_size = sysconf(_SC_PAGE_SIZE);
 397         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 398         assert(nr_cpus <= MAX_NR_CPUS);
 399         assert(nr_cpus >= 0);
 400
 401         if (!stat(output_name, &st) && !force && !append_file) {
 402                 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
 403                                 output_name);
 404                 exit(-1);
 405         }
 406
 407         flags = O_CREAT|O_RDWR;
 408         if (append_file)
 409                 flags |= O_APPEND;
 410         else
 411                 flags |= O_TRUNC;
 412
 413         output = open(output_name, flags, S_IRUSR|S_IWUSR);
 414         if (output < 0) {
 415                 perror("failed to create output file");
 416                 exit(-1);
 417         }
 418
 419         if (!system_wide) {
 420                 open_counters(-1, target_pid != -1 ? target_pid : 0);
 421         } else for (i = 0; i < nr_cpus; i++)
 422                 open_counters(i, target_pid);
 423
 424         signal(SIGCHLD, sig_handler);
 425         signal(SIGINT, sig_handler);
 426
 427         if (target_pid == -1 && argc) {
 428                 pid = fork();
 429                 if (pid < 0)
 430                         perror("failed to fork");
 431
 432                 if (!pid) {
 433                         if (execvp(argv[0], (char **)argv)) {
 434                                 perror(argv[0]);
 435                                 exit(-1);
 436                         }
 437                 }
 438         }
 439
 440         if (realtime_prio) {
 441                 struct sched_param param;
 442
 443                 param.sched_priority = realtime_prio;
 444                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 445                         printf("Could not set realtime priority.\n");
 446                         exit(-1);
 447                 }
 448         }
 449
 450         if (system_wide)
 451                 synthesize_events();
 452
 453         while (!done) {
 454                 int hits = events;
 455
 456                 for (i = 0; i < nr_cpu; i++) {
 457                         for (counter = 0; counter < nr_counters; counter++)
 458                                 mmap_read(&mmap_array[i][counter]);
 459                 }
 460
 461                 if (hits == events)
 462                         ret = poll(event_array, nr_poll, 100);
 463         }
 464
 465         return 0;
 466 }
 467
 468 static const char * const record_usage[] = {
 469         "perf record [<options>] [<command>]",
 470         "perf record [<options>] -- <command> [<options>]",
 471         NULL
 472 };
 473
 474 static char events_help_msg[EVENTS_HELP_MAX];
 475
 476 static const struct option options[] = {
 477         OPT_CALLBACK('e', "event", NULL, "event",
 478                      events_help_msg, parse_events),
 479         OPT_INTEGER('p', "pid", &target_pid,
 480                     "record events on existing pid"),
 481         OPT_INTEGER('r', "realtime", &realtime_prio,
 482                     "collect data with this RT SCHED_FIFO priority"),
 483         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 484                             "system-wide collection from all CPUs"),
 485         OPT_BOOLEAN('A', "append", &append_file,
 486                             "append to the output file to do incremental profiling"),
 487         OPT_BOOLEAN('f', "force", &force,
 488                         "overwrite existing data file"),
 489         OPT_INTEGER('c', "count", &default_interval,
 490                     "event period to sample"),
 491         OPT_STRING('o', "output", &output_name, "file",
 492                     "output file name"),
 493         OPT_BOOLEAN('i', "inherit", &inherit,
 494                     "child tasks inherit counters"),
 495         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 496                     "number of mmap data pages"),
 497         OPT_END()
 498 };
 499
 500 int cmd_record(int argc, const char **argv, const char *prefix)
 501 {
 502         int counter;
 503
 504         create_events_help(events_help_msg);
 505
 506         argc = parse_options(argc, argv, options, record_usage, 0);
 507         if (!argc && target_pid == -1 && !system_wide)
 508                 usage_with_options(record_usage, options);
 509
 510         if (!nr_counters) {
 511                 nr_counters = 1;
 512                 event_id[0] = 0;
 513         }
 514
 515         for (counter = 0; counter < nr_counters; counter++) {
 516                 if (event_count[counter])
 517                         continue;
 518
 519                 event_count[counter] = default_interval;
 520         }
 521
 522         return __cmd_record(argc, argv);
 523 }