init.c

   1 /*
   2  * This file contains job initialization and setup functions.
   3  */
   4 #include <stdio.h>
   5 #include <stdlib.h>
   6 #include <unistd.h>
   7 #include <fcntl.h>
   8 #include <ctype.h>
   9 #include <string.h>
  10 #include <errno.h>
  11 #include <getopt.h>
  12 #include <assert.h>
  13 #include <sys/ipc.h>
  14 #include <sys/shm.h>
  15 #include <sys/types.h>
  16 #include <sys/stat.h>
  17
  18 #include "fio.h"
  19 #include "parse.h"
  20
  21 /*
  22  * The default options
  23  */
  24 #define DEF_BS                  (4096)
  25 #define DEF_TIMEOUT             (0)
  26 #define DEF_RATE_CYCLE          (1000)
  27 #define DEF_ODIRECT             (1)
  28 #define DEF_IO_ENGINE           (FIO_SYNCIO)
  29 #define DEF_IO_ENGINE_NAME      "sync"
  30 #define DEF_SEQUENTIAL          (1)
  31 #define DEF_RAND_REPEAT         (1)
  32 #define DEF_OVERWRITE           (1)
  33 #define DEF_INVALIDATE          (1)
  34 #define DEF_SYNCIO              (0)
  35 #define DEF_RANDSEED            (0xb1899bedUL)
  36 #define DEF_BWAVGTIME           (500)
  37 #define DEF_CREATE_SER          (1)
  38 #define DEF_CREATE_FSYNC        (1)
  39 #define DEF_LOOPS               (1)
  40 #define DEF_VERIFY              (0)
  41 #define DEF_STONEWALL           (0)
  42 #define DEF_NUMJOBS             (1)
  43 #define DEF_USE_THREAD          (0)
  44 #define DEF_FILE_SIZE           (1024 * 1024 * 1024UL)
  45 #define DEF_ZONE_SIZE           (0)
  46 #define DEF_ZONE_SKIP           (0)
  47 #define DEF_RWMIX_CYCLE         (500)
  48 #define DEF_RWMIX_READ          (50)
  49 #define DEF_NICE                (0)
  50 #define DEF_NR_FILES            (1)
  51 #define DEF_UNLINK              (0)
  52 #define DEF_WRITE_BW_LOG        (0)
  53 #define DEF_WRITE_LAT_LOG       (0)
  54 #define DEF_NO_RAND_MAP         (0)
  55
  56 #define td_var_offset(var)      ((size_t) &((struct thread_data *)0)->var)
  57
  58 static int str_rw_cb(void *, const char *);
  59 static int str_ioengine_cb(void *, const char *);
  60 static int str_mem_cb(void *, const char *);
  61 static int str_verify_cb(void *, const char *);
  62 static int str_lockmem_cb(void *, unsigned long *);
  63 static int str_prio_cb(void *, unsigned int *);
  64 static int str_prioclass_cb(void *, unsigned int *);
  65 static int str_exitall_cb(void);
  66 static int str_cpumask_cb(void *, unsigned int *);
  67
  68 /*
  69  * Map of job/command line options
  70  */
  71 static struct fio_option options[] = {
  72         {
  73                 .name   = "name",
  74                 .type   = FIO_OPT_STR_STORE,
  75                 .off1   = td_var_offset(name),
  76         },
  77         {
  78                 .name   = "directory",
  79                 .type   = FIO_OPT_STR_STORE,
  80                 .off1   = td_var_offset(directory),
  81         },
  82         {
  83                 .name   = "filename",
  84                 .type   = FIO_OPT_STR_STORE,
  85                 .off1   = td_var_offset(filename),
  86         },
  87         {
  88                 .name   = "rw",
  89                 .type   = FIO_OPT_STR,
  90                 .cb     = str_rw_cb,
  91         },
  92         {
  93                 .name   = "ioengine",
  94                 .type   = FIO_OPT_STR,
  95                 .cb     = str_ioengine_cb,
  96         },
  97         {
  98                 .name   = "mem",
  99                 .type   = FIO_OPT_STR,
 100                 .cb     = str_mem_cb,
 101         },
 102         {
 103                 .name   = "verify",
 104                 .type   = FIO_OPT_STR,
 105                 .cb     = str_verify_cb,
 106         },
 107         {
 108                 .name   = "write_iolog",
 109                 .type   = FIO_OPT_STR_STORE,
 110                 .off1   = td_var_offset(write_iolog_file),
 111         },
 112         {
 113                 .name   = "read_iolog",
 114                 .type   = FIO_OPT_STR_STORE,
 115                 .off1   = td_var_offset(read_iolog_file),
 116         },
 117         {
 118                 .name   = "exec_prerun",
 119                 .type   = FIO_OPT_STR_STORE,
 120                 .off1   = td_var_offset(exec_prerun),
 121         },
 122         {
 123                 .name   = "exec_postrun",
 124                 .type   = FIO_OPT_STR_STORE,
 125                 .off1   = td_var_offset(exec_postrun),
 126         },
 127 #ifdef FIO_HAVE_IOSCHED_SWITCH
 128         {
 129                 .name   = "ioscheduler",
 130                 .type   = FIO_OPT_STR_STORE,
 131                 .off1   = td_var_offset(ioscheduler),
 132         },
 133 #endif
 134         {
 135                 .name   = "size",
 136                 .type   = FIO_OPT_STR_VAL,
 137                 .off1   = td_var_offset(total_file_size),
 138         },
 139         {
 140                 .name   = "bs",
 141                 .type   = FIO_OPT_STR_VAL_INT,
 142                 .off1   = td_var_offset(bs[DDIR_READ]),
 143         },
 144         {
 145                 .name   = "read_bs",
 146                 .type   = FIO_OPT_STR_VAL_INT,
 147                 .off1   = td_var_offset(bs[DDIR_READ]),
 148         },
 149         {
 150                 .name   = "write_bs",
 151                 .type   = FIO_OPT_STR_VAL_INT,
 152                 .off1   = td_var_offset(bs[DDIR_WRITE]),
 153         },
 154         {
 155                 .name   = "offset",
 156                 .type   = FIO_OPT_STR_VAL,
 157                 .off1   = td_var_offset(start_offset),
 158         },
 159         {
 160                 .name   = "zonesize",
 161                 .type   = FIO_OPT_STR_VAL,
 162                 .off1   = td_var_offset(zone_size),
 163         },
 164         {
 165                 .name   = "zoneskip",
 166                 .type   = FIO_OPT_STR_VAL,
 167                 .off1   = td_var_offset(zone_skip),
 168         },
 169         {
 170                 .name   = "lockmem",
 171                 .type   = FIO_OPT_STR_VAL,
 172                 .cb     = str_lockmem_cb,
 173         },
 174         {
 175                 .name   = "bsrange",
 176                 .type   = FIO_OPT_RANGE,
 177                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 178                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 179         },
 180         {
 181                 .name   = "read_bsrange",
 182                 .type   = FIO_OPT_RANGE,
 183                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 184                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 185         },
 186         {
 187                 .name   = "write_bsrange",
 188                 .type   = FIO_OPT_RANGE,
 189                 .off1   = td_var_offset(min_bs[DDIR_WRITE]),
 190                 .off2   = td_var_offset(max_bs[DDIR_WRITE]),
 191         },
 192         {
 193                 .name   = "nrfiles",
 194                 .type   = FIO_OPT_INT,
 195                 .off1   = td_var_offset(nr_files),
 196         },
 197         {
 198                 .name   = "iodepth",
 199                 .type   = FIO_OPT_INT,
 200                 .off1   = td_var_offset(iodepth),
 201         },
 202         {
 203                 .name   = "fsync",
 204                 .type   = FIO_OPT_INT,
 205                 .off1   = td_var_offset(fsync_blocks),
 206         },
 207         {
 208                 .name   = "rwmixcycle",
 209                 .type   = FIO_OPT_INT,
 210                 .off1   = td_var_offset(rwmixcycle),
 211         },
 212         {
 213                 .name   = "rwmixread",
 214                 .type   = FIO_OPT_INT,
 215                 .off1   = td_var_offset(rwmixread),
 216                 .max_val= 100,
 217         },
 218         {
 219                 .name   = "rwmixwrite",
 220                 .type   = FIO_OPT_INT,
 221                 .off1   = td_var_offset(rwmixwrite),
 222                 .max_val= 100,
 223         },
 224         {
 225                 .name   = "nice",
 226                 .type   = FIO_OPT_INT,
 227                 .off1   = td_var_offset(nice),
 228         },
 229 #ifdef FIO_HAVE_IOPRIO
 230         {
 231                 .name   = "prio",
 232                 .type   = FIO_OPT_INT,
 233                 .cb     = str_prio_cb,
 234         },
 235         {
 236                 .name   = "prioclass",
 237                 .type   = FIO_OPT_INT,
 238                 .cb     = str_prioclass_cb,
 239         },
 240 #endif
 241         {
 242                 .name   = "thinktime",
 243                 .type   = FIO_OPT_INT,
 244                 .off1   = td_var_offset(thinktime)
 245         },
 246         {
 247                 .name   = "rate",
 248                 .type   = FIO_OPT_INT,
 249                 .off1   = td_var_offset(rate)
 250         },
 251         {
 252                 .name   = "ratemin",
 253                 .type   = FIO_OPT_INT,
 254                 .off1   = td_var_offset(ratemin)
 255         },
 256         {
 257                 .name   = "ratecycle",
 258                 .type   = FIO_OPT_INT,
 259                 .off1   = td_var_offset(ratecycle)
 260         },
 261         {
 262                 .name   = "startdelay",
 263                 .type   = FIO_OPT_INT,
 264                 .off1   = td_var_offset(start_delay)
 265         },
 266         {
 267                 .name   = "timeout",
 268                 .type   = FIO_OPT_STR_VAL_TIME,
 269                 .off1   = td_var_offset(timeout)
 270         },
 271         {
 272                 .name   = "invalidate",
 273                 .type   = FIO_OPT_INT,
 274                 .off1   = td_var_offset(invalidate_cache)
 275         },
 276         {
 277                 .name   = "sync",
 278                 .type   = FIO_OPT_INT,
 279                 .off1   = td_var_offset(sync_io)
 280         },
 281         {
 282                 .name   = "bwavgtime",
 283                 .type   = FIO_OPT_INT,
 284                 .off1   = td_var_offset(bw_avg_time)
 285         },
 286         {
 287                 .name   = "create_serialize",
 288                 .type   = FIO_OPT_INT,
 289                 .off1   = td_var_offset(create_serialize)
 290         },
 291         {
 292                 .name   = "create_fsync",
 293                 .type   = FIO_OPT_INT,
 294                 .off1   = td_var_offset(create_fsync)
 295         },
 296         {
 297                 .name   = "loops",
 298                 .type   = FIO_OPT_INT,
 299                 .off1   = td_var_offset(loops)
 300         },
 301         {
 302                 .name   = "numjobs",
 303                 .type   = FIO_OPT_INT,
 304                 .off1   = td_var_offset(numjobs)
 305         },
 306         {
 307                 .name   = "cpuload",
 308                 .type   = FIO_OPT_INT,
 309                 .off1   = td_var_offset(cpuload)
 310         },
 311         {
 312                 .name   = "cpuchunks",
 313                 .type   = FIO_OPT_INT,
 314                 .off1   = td_var_offset(cpucycle)
 315         },
 316         {
 317                 .name   = "direct",
 318                 .type   = FIO_OPT_INT,
 319                 .off1   = td_var_offset(odirect)
 320         },
 321         {
 322                 .name   = "overwrite",
 323                 .type   = FIO_OPT_INT,
 324                 .off1   = td_var_offset(overwrite)
 325         },
 326 #ifdef FIO_HAVE_CPU_AFFINITY
 327         {
 328                 .name   = "cpumask",
 329                 .type   = FIO_OPT_INT,
 330                 .cb     = str_cpumask_cb,
 331         },
 332 #endif
 333         {
 334                 .name   = "end_fsync",
 335                 .type   = FIO_OPT_INT,
 336                 .off1   = td_var_offset(end_fsync)
 337         },
 338         {
 339                 .name   = "unlink",
 340                 .type   = FIO_OPT_STR_SET,
 341                 .off1   = td_var_offset(unlink),
 342         },
 343         {
 344                 .name   = "exitall",
 345                 .type   = FIO_OPT_STR_SET,
 346                 .cb     = str_exitall_cb,
 347         },
 348         {
 349                 .name   = "stonewall",
 350                 .type   = FIO_OPT_STR_SET,
 351                 .off1   = td_var_offset(stonewall),
 352         },
 353         {
 354                 .name   = "thread",
 355                 .type   = FIO_OPT_STR_SET,
 356                 .off1   = td_var_offset(thread),
 357         },
 358         {
 359                 .name   = "write_bw_log",
 360                 .type   = FIO_OPT_STR_SET,
 361                 .off1   = td_var_offset(write_bw_log),
 362         },
 363         {
 364                 .name   = "write_lat_log",
 365                 .type   = FIO_OPT_STR_SET,
 366                 .off1   = td_var_offset(write_lat_log),
 367         },
 368         {
 369                 .name   = "norandommap",
 370                 .type   = FIO_OPT_STR_SET,
 371                 .off1   = td_var_offset(norandommap),
 372         },
 373         {
 374                 .name   = "bs_unaligned",
 375                 .type   = FIO_OPT_STR_SET,
 376                 .off1   = td_var_offset(bs_unaligned),
 377         },
 378         {
 379                 .name = NULL,
 380         },
 381 };
 382
 383 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 384 #define FIO_CMD_OPTS    (16)
 385 #define FIO_GETOPT_JOB  (0x89988998)
 386
 387 /*
 388  * Command line options. These will contain the above, plus a few
 389  * extra that only pertain to fio itself and not jobs.
 390  */
 391 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 392         {
 393                 .name           = "output",
 394                 .has_arg        = required_argument,
 395                 .val            = 'o',
 396         },
 397         {
 398                 .name           = "timeout",
 399                 .has_arg        = required_argument,
 400                 .val            = 't',
 401         },
 402         {
 403                 .name           = "latency-log",
 404                 .has_arg        = required_argument,
 405                 .val            = 'l',
 406         },
 407         {
 408                 .name           = "bandwidth-log",
 409                 .has_arg        = required_argument,
 410                 .val            = 'b',
 411         },
 412         {
 413                 .name           = "minimal",
 414                 .has_arg        = optional_argument,
 415                 .val            = 'm',
 416         },
 417         {
 418                 .name           = "version",
 419                 .has_arg        = no_argument,
 420                 .val            = 'v',
 421         },
 422         {
 423                 .name           = NULL,
 424         },
 425 };
 426
 427 static int def_timeout = DEF_TIMEOUT;
 428
 429 static char fio_version_string[] = "fio 1.7";
 430
 431 static char **ini_file;
 432 static int max_jobs = MAX_JOBS;
 433
 434 struct thread_data def_thread;
 435 struct thread_data *threads = NULL;
 436
 437 int rate_quit = 0;
 438 int exitall_on_terminate = 0;
 439 int terse_output = 0;
 440 unsigned long long mlock_size = 0;
 441 FILE *f_out = NULL;
 442 FILE *f_err = NULL;
 443
 444 static int write_lat_log = DEF_WRITE_LAT_LOG;
 445 static int write_bw_log = DEF_WRITE_BW_LOG;
 446
 447 /*
 448  * Return a free job structure.
 449  */
 450 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 451 {
 452         struct thread_data *td;
 453
 454         if (global)
 455                 return &def_thread;
 456         if (thread_number >= max_jobs)
 457                 return NULL;
 458
 459         td = &threads[thread_number++];
 460         *td = *parent;
 461
 462         td->thread_number = thread_number;
 463         return td;
 464 }
 465
 466 static void put_job(struct thread_data *td)
 467 {
 468         if (td == &def_thread)
 469                 return;
 470
 471         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 472         thread_number--;
 473 }
 474
 475 /*
 476  * Lazy way of fixing up options that depend on each other. We could also
 477  * define option callback handlers, but this is easier.
 478  */
 479 static void fixup_options(struct thread_data *td)
 480 {
 481         if (!td->rwmixread && td->rwmixwrite)
 482                 td->rwmixread = 100 - td->rwmixwrite;
 483
 484         if (td->write_iolog_file && td->read_iolog_file) {
 485                 log_err("fio: read iolog overrides write_iolog\n");
 486                 free(td->write_iolog_file);
 487                 td->write_iolog_file = NULL;
 488         }
 489
 490         if (td->io_ops->flags & FIO_SYNCIO)
 491                 td->iodepth = 1;
 492         else {
 493                 if (!td->iodepth)
 494                         td->iodepth = td->nr_files;
 495         }
 496
 497         /*
 498          * only really works for sequential io for now, and with 1 file
 499          */
 500         if (td->zone_size && !td->sequential && td->nr_files == 1)
 501                 td->zone_size = 0;
 502
 503         /*
 504          * Reads can do overwrites, we always need to pre-create the file
 505          */
 506         if (td_read(td) || td_rw(td))
 507                 td->overwrite = 1;
 508
 509         if (td->bs[DDIR_READ] != DEF_BS)
 510                 td->bs[DDIR_WRITE] = td->bs[DDIR_READ];
 511         if (!td->min_bs[DDIR_READ])
 512                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 513         if (!td->max_bs[DDIR_READ])
 514                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 515         if (!td->min_bs[DDIR_WRITE])
 516                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 517         if (!td->max_bs[DDIR_WRITE])
 518                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 519
 520         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 521
 522         if (td_read(td) && !td_rw(td))
 523                 td->verify = 0;
 524
 525         if (td->norandommap && td->verify != VERIFY_NONE) {
 526                 log_err("fio: norandommap given, verify disabled\n");
 527                 td->verify = VERIFY_NONE;
 528         }
 529         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 530                 log_err("fio: bs_unaligned may not work with raw io\n");
 531 }
 532
 533 /*
 534  * Adds a job to the list of things todo. Sanitizes the various options
 535  * to make sure we don't have conflicts, and initializes various
 536  * members of td.
 537  */
 538 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 539 {
 540         char *ddir_str[] = { "read", "write", "randread", "randwrite",
 541                              "rw", NULL, "randrw" };
 542         struct stat sb;
 543         int numjobs, ddir, i;
 544         struct fio_file *f;
 545
 546         /*
 547          * the def_thread is just for options, it's not a real job
 548          */
 549         if (td == &def_thread)
 550                 return 0;
 551
 552         /*
 553          * Set default io engine, if none set
 554          */
 555         if (!td->io_ops) {
 556                 td->io_ops = load_ioengine(td, DEF_IO_ENGINE_NAME);
 557                 if (!td->io_ops) {
 558                         log_err("default engine %s not there?\n", DEF_IO_ENGINE_NAME);
 559                         return 1;
 560                 }
 561         }
 562
 563         if (td->odirect)
 564                 td->io_ops->flags |= FIO_RAWIO;
 565
 566         fixup_options(td);
 567
 568         td->filetype = FIO_TYPE_FILE;
 569         if (!stat(jobname, &sb)) {
 570                 if (S_ISBLK(sb.st_mode))
 571                         td->filetype = FIO_TYPE_BD;
 572                 else if (S_ISCHR(sb.st_mode))
 573                         td->filetype = FIO_TYPE_CHAR;
 574         }
 575
 576         if (td->filename)
 577                 td->nr_uniq_files = 1;
 578         else
 579                 td->nr_uniq_files = td->nr_files;
 580
 581         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 582                 char tmp[PATH_MAX];
 583                 int len = 0;
 584                 int i;
 585
 586                 if (td->directory && td->directory[0] != '\0')
 587                         sprintf(tmp, "%s/", td->directory);
 588
 589                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 590
 591                 for_each_file(td, f, i) {
 592                         memset(f, 0, sizeof(*f));
 593                         f->fd = -1;
 594
 595                         if (td->filename)
 596                                 sprintf(tmp + len, "%s", td->filename);
 597                         else
 598                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 599                         f->file_name = strdup(tmp);
 600                 }
 601         } else {
 602                 td->nr_files = 1;
 603                 td->files = malloc(sizeof(struct fio_file));
 604                 f = &td->files[0];
 605
 606                 memset(f, 0, sizeof(*f));
 607                 f->fd = -1;
 608                 f->file_name = strdup(jobname);
 609         }
 610
 611         for_each_file(td, f, i) {
 612                 f->file_size = td->total_file_size / td->nr_files;
 613                 f->file_offset = td->start_offset;
 614         }
 615
 616         fio_sem_init(&td->mutex, 0);
 617
 618         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
 619         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
 620         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
 621
 622         if (td->stonewall && td->thread_number > 1)
 623                 groupid++;
 624
 625         td->groupid = groupid;
 626
 627         if (setup_rate(td))
 628                 goto err;
 629
 630         if (td->write_lat_log) {
 631                 setup_log(&td->slat_log);
 632                 setup_log(&td->clat_log);
 633         }
 634         if (td->write_bw_log)
 635                 setup_log(&td->bw_log);
 636
 637         if (!td->name)
 638                 td->name = strdup(jobname);
 639
 640         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 641
 642         if (!terse_output) {
 643                 if (!job_add_num) {
 644                         if (td->io_ops->flags & FIO_CPUIO)
 645                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 646                         else
 647                                 fprintf(f_out, "%s: (g=%d): rw=%s, odir=%d, bs=%d-%d/%d-%d, rate=%d, ioengine=%s, iodepth=%d\n", td->name, td->groupid, ddir_str[ddir], td->odirect, td->min_bs[DDIR_READ], td->max_bs[DDIR_READ], td->min_bs[DDIR_WRITE], td->max_bs[DDIR_WRITE], td->rate, td->io_ops->name, td->iodepth);
 648                 } else if (job_add_num == 1)
 649                         fprintf(f_out, "...\n");
 650         }
 651
 652         /*
 653          * recurse add identical jobs, clear numjobs and stonewall options
 654          * as they don't apply to sub-jobs
 655          */
 656         numjobs = td->numjobs;
 657         while (--numjobs) {
 658                 struct thread_data *td_new = get_new_job(0, td);
 659
 660                 if (!td_new)
 661                         goto err;
 662
 663                 td_new->numjobs = 1;
 664                 td_new->stonewall = 0;
 665                 job_add_num = numjobs - 1;
 666
 667                 if (add_job(td_new, jobname, job_add_num))
 668                         goto err;
 669         }
 670         return 0;
 671 err:
 672         put_job(td);
 673         return -1;
 674 }
 675
 676 /*
 677  * Initialize the various random states we need (random io, block size ranges,
 678  * read/write mix, etc).
 679  */
 680 int init_random_state(struct thread_data *td)
 681 {
 682         unsigned long seeds[4];
 683         int fd, num_maps, blocks, i;
 684         struct fio_file *f;
 685
 686         if (td->io_ops->flags & FIO_CPUIO)
 687                 return 0;
 688
 689         fd = open("/dev/urandom", O_RDONLY);
 690         if (fd == -1) {
 691                 td_verror(td, errno);
 692                 return 1;
 693         }
 694
 695         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 696                 td_verror(td, EIO);
 697                 close(fd);
 698                 return 1;
 699         }
 700
 701         close(fd);
 702
 703         os_random_seed(seeds[0], &td->bsrange_state);
 704         os_random_seed(seeds[1], &td->verify_state);
 705         os_random_seed(seeds[2], &td->rwmix_state);
 706
 707         if (td->sequential)
 708                 return 0;
 709
 710         if (td->rand_repeatable)
 711                 seeds[3] = DEF_RANDSEED;
 712
 713         if (!td->norandommap) {
 714                 for_each_file(td, f, i) {
 715                         blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 716                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 717                         f->file_map = malloc(num_maps * sizeof(long));
 718                         f->num_maps = num_maps;
 719                         memset(f->file_map, 0, num_maps * sizeof(long));
 720                 }
 721         }
 722
 723         os_random_seed(seeds[3], &td->random_state);
 724         return 0;
 725 }
 726
 727 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 728 {
 729 #ifdef FIO_HAVE_CPU_AFFINITY
 730         unsigned int i;
 731
 732         CPU_ZERO(&cpumask);
 733
 734         for (i = 0; i < sizeof(int) * 8; i++) {
 735                 if ((1 << i) & cpu)
 736                         CPU_SET(i, &cpumask);
 737         }
 738 #endif
 739 }
 740
 741 static int is_empty_or_comment(char *line)
 742 {
 743         unsigned int i;
 744
 745         for (i = 0; i < strlen(line); i++) {
 746                 if (line[i] == ';')
 747                         return 1;
 748                 if (!isspace(line[i]) && !iscntrl(line[i]))
 749                         return 0;
 750         }
 751
 752         return 1;
 753 }
 754
 755 static int str_rw_cb(void *data, const char *mem)
 756 {
 757         struct thread_data *td = data;
 758
 759         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 760                 td->ddir = DDIR_READ;
 761                 td->sequential = 1;
 762                 return 0;
 763         } else if (!strncmp(mem, "randread", 8)) {
 764                 td->ddir = DDIR_READ;
 765                 td->sequential = 0;
 766                 return 0;
 767         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 768                 td->ddir = DDIR_WRITE;
 769                 td->sequential = 1;
 770                 return 0;
 771         } else if (!strncmp(mem, "randwrite", 9)) {
 772                 td->ddir = DDIR_WRITE;
 773                 td->sequential = 0;
 774                 return 0;
 775         } else if (!strncmp(mem, "rw", 2)) {
 776                 td->ddir = 0;
 777                 td->iomix = 1;
 778                 td->sequential = 1;
 779                 return 0;
 780         } else if (!strncmp(mem, "randrw", 6)) {
 781                 td->ddir = 0;
 782                 td->iomix = 1;
 783                 td->sequential = 0;
 784                 return 0;
 785         }
 786
 787         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 788         return 1;
 789 }
 790
 791 static int str_verify_cb(void *data, const char *mem)
 792 {
 793         struct thread_data *td = data;
 794
 795         if (!strncmp(mem, "0", 1)) {
 796                 td->verify = VERIFY_NONE;
 797                 return 0;
 798         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 799                 td->verify = VERIFY_MD5;
 800                 return 0;
 801         } else if (!strncmp(mem, "crc32", 5)) {
 802                 td->verify = VERIFY_CRC32;
 803                 return 0;
 804         }
 805
 806         log_err("fio: verify types: md5, crc32\n");
 807         return 1;
 808 }
 809
 810 static int str_mem_cb(void *data, const char *mem)
 811 {
 812         struct thread_data *td = data;
 813
 814         if (!strncmp(mem, "malloc", 6)) {
 815                 td->mem_type = MEM_MALLOC;
 816                 return 0;
 817         } else if (!strncmp(mem, "shm", 3)) {
 818                 td->mem_type = MEM_SHM;
 819                 return 0;
 820         } else if (!strncmp(mem, "mmap", 4)) {
 821                 td->mem_type = MEM_MMAP;
 822                 return 0;
 823         }
 824
 825         log_err("fio: mem type: malloc, shm, mmap\n");
 826         return 1;
 827 }
 828
 829 static int str_ioengine_cb(void *data, const char *str)
 830 {
 831         struct thread_data *td = data;
 832
 833         td->io_ops = load_ioengine(td, str);
 834         if (td->io_ops)
 835                 return 0;
 836
 837         log_err("fio: ioengine: { linuxaio, aio, libaio }, posixaio, sync, mmap, sgio, splice, cpu\n");
 838         return 1;
 839 }
 840
 841 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
 842 {
 843         mlock_size = *val;
 844         return 0;
 845 }
 846
 847 static int str_prioclass_cb(void *data, unsigned int *val)
 848 {
 849         struct thread_data *td = data;
 850
 851         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
 852         return 0;
 853 }
 854
 855 static int str_prio_cb(void *data, unsigned int *val)
 856 {
 857         struct thread_data *td = data;
 858
 859         td->ioprio |= *val;
 860         return 0;
 861 }
 862
 863 static int str_exitall_cb(void)
 864 {
 865         exitall_on_terminate = 1;
 866         return 0;
 867 }
 868
 869 static int str_cpumask_cb(void *data, unsigned int *val)
 870 {
 871         struct thread_data *td = data;
 872
 873         fill_cpu_mask(td->cpumask, *val);
 874         return 0;
 875 }
 876
 877 /*
 878  * This is our [ini] type file parser.
 879  */
 880 int parse_jobs_ini(char *file, int stonewall_flag)
 881 {
 882         unsigned int global;
 883         struct thread_data *td;
 884         char *string, *name;
 885         fpos_t off;
 886         FILE *f;
 887         char *p;
 888         int ret = 0, stonewall;
 889
 890         f = fopen(file, "r");
 891         if (!f) {
 892                 perror("fopen job file");
 893                 return 1;
 894         }
 895
 896         string = malloc(4096);
 897         name = malloc(256);
 898         memset(name, 0, 256);
 899
 900         stonewall = stonewall_flag;
 901         do {
 902                 p = fgets(string, 4095, f);
 903                 if (!p)
 904                         break;
 905                 if (is_empty_or_comment(p))
 906                         continue;
 907                 if (sscanf(p, "[%255s]", name) != 1)
 908                         continue;
 909
 910                 global = !strncmp(name, "global", 6);
 911
 912                 name[strlen(name) - 1] = '\0';
 913
 914                 td = get_new_job(global, &def_thread);
 915                 if (!td) {
 916                         ret = 1;
 917                         break;
 918                 }
 919
 920                 /*
 921                  * Seperate multiple job files by a stonewall
 922                  */
 923                 if (!global && stonewall) {
 924                         td->stonewall = stonewall;
 925                         stonewall = 0;
 926                 }
 927
 928                 fgetpos(f, &off);
 929                 while ((p = fgets(string, 4096, f)) != NULL) {
 930                         if (is_empty_or_comment(p))
 931                                 continue;
 932
 933                         strip_blank_front(&p);
 934
 935                         if (p[0] == '[')
 936                                 break;
 937
 938                         strip_blank_end(p);
 939
 940                         fgetpos(f, &off);
 941
 942                         /*
 943                          * Don't break here, continue parsing options so we
 944                          * dump all the bad ones. Makes trial/error fixups
 945                          * easier on the user.
 946                          */
 947                         ret |= parse_option(p, options, td);
 948                 }
 949
 950                 if (!ret) {
 951                         fsetpos(f, &off);
 952                         ret = add_job(td, name, 0);
 953                 } else {
 954                         log_err("fio: job %s dropped\n", name);
 955                         put_job(td);
 956                 }
 957         } while (!ret);
 958
 959         free(string);
 960         free(name);
 961         fclose(f);
 962         return ret;
 963 }
 964
 965 static int fill_def_thread(void)
 966 {
 967         memset(&def_thread, 0, sizeof(def_thread));
 968
 969         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
 970                 perror("sched_getaffinity");
 971                 return 1;
 972         }
 973
 974         /*
 975          * fill globals
 976          */
 977         def_thread.ddir = DDIR_READ;
 978         def_thread.iomix = 0;
 979         def_thread.bs[DDIR_READ] = DEF_BS;
 980         def_thread.bs[DDIR_WRITE] = DEF_BS;
 981         def_thread.min_bs[DDIR_READ] = def_thread.min_bs[DDIR_WRITE] = 0;
 982         def_thread.max_bs[DDIR_READ] = def_thread.max_bs[DDIR_WRITE] = 0;
 983         def_thread.odirect = DEF_ODIRECT;
 984         def_thread.ratecycle = DEF_RATE_CYCLE;
 985         def_thread.sequential = DEF_SEQUENTIAL;
 986         def_thread.timeout = def_timeout;
 987         def_thread.overwrite = DEF_OVERWRITE;
 988         def_thread.invalidate_cache = DEF_INVALIDATE;
 989         def_thread.sync_io = DEF_SYNCIO;
 990         def_thread.mem_type = MEM_MALLOC;
 991         def_thread.bw_avg_time = DEF_BWAVGTIME;
 992         def_thread.create_serialize = DEF_CREATE_SER;
 993         def_thread.create_fsync = DEF_CREATE_FSYNC;
 994         def_thread.loops = DEF_LOOPS;
 995         def_thread.verify = DEF_VERIFY;
 996         def_thread.stonewall = DEF_STONEWALL;
 997         def_thread.numjobs = DEF_NUMJOBS;
 998         def_thread.use_thread = DEF_USE_THREAD;
 999         def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
1000         def_thread.rwmixread = DEF_RWMIX_READ;
1001         def_thread.nice = DEF_NICE;
1002         def_thread.rand_repeatable = DEF_RAND_REPEAT;
1003         def_thread.nr_files = DEF_NR_FILES;
1004         def_thread.unlink = DEF_UNLINK;
1005         def_thread.write_bw_log = write_bw_log;
1006         def_thread.write_lat_log = write_lat_log;
1007         def_thread.norandommap = DEF_NO_RAND_MAP;
1008 #ifdef FIO_HAVE_DISK_UTIL
1009         def_thread.do_disk_util = 1;
1010 #endif
1011
1012         return 0;
1013 }
1014
1015 static void usage(void)
1016 {
1017         printf("%s\n", fio_version_string);
1018         printf("\t--output\tWrite output to file\n");
1019         printf("\t--timeout\tRuntime in seconds\n");
1020         printf("\t--latency-log\tGenerate per-job latency logs\n");
1021         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1022         printf("\t--minimal\tMinimal (terse) output\n");
1023         printf("\t--version\tPrint version info and exit\n");
1024 }
1025
1026 static int parse_cmd_line(int argc, char *argv[])
1027 {
1028         struct thread_data *td = NULL;
1029         int c, ini_idx = 0, lidx, ret;
1030
1031         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1032                 switch (c) {
1033                 case 't':
1034                         def_timeout = atoi(optarg);
1035                         break;
1036                 case 'l':
1037                         write_lat_log = 1;
1038                         break;
1039                 case 'w':
1040                         write_bw_log = 1;
1041                         break;
1042                 case 'o':
1043                         f_out = fopen(optarg, "w+");
1044                         if (!f_out) {
1045                                 perror("fopen output");
1046                                 exit(1);
1047                         }
1048                         f_err = f_out;
1049                         break;
1050                 case 'm':
1051                         terse_output = 1;
1052                         break;
1053                 case 'h':
1054                         usage();
1055                         exit(0);
1056                 case 'v':
1057                         printf("%s\n", fio_version_string);
1058                         exit(0);
1059                 case FIO_GETOPT_JOB: {
1060                         const char *opt = long_options[lidx].name;
1061                         char *val = optarg;
1062
1063                         if (!strncmp(opt, "name", 4) && td) {
1064                                 ret = add_job(td, td->name ?: "fio", 0);
1065                                 if (ret) {
1066                                         put_job(td);
1067                                         return 0;
1068                                 }
1069                                 td = NULL;
1070                         }
1071                         if (!td) {
1072                                 int global = !strncmp(val, "global", 6);
1073
1074                                 td = get_new_job(global, &def_thread);
1075                                 if (!td)
1076                                         return 0;
1077                         }
1078
1079                         ret = parse_cmd_option(opt, val, options, td);
1080                         if (ret) {
1081                                 log_err("fio: job dropped\n");
1082                                 put_job(td);
1083                                 td = NULL;
1084                         }
1085                         break;
1086                 }
1087                 default:
1088                         printf("optarg <<%s>>\n", argv[optind]);
1089                         break;
1090                 }
1091         }
1092
1093         if (td) {
1094                 ret = add_job(td, td->name ?: "fio", 0);
1095                 if (ret)
1096                         put_job(td);
1097         }
1098
1099         while (optind < argc) {
1100                 ini_idx++;
1101                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1102                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1103                 optind++;
1104         }
1105
1106         return ini_idx;
1107 }
1108
1109 static void free_shm(void)
1110 {
1111         struct shmid_ds sbuf;
1112
1113         if (threads) {
1114                 shmdt((void *) threads);
1115                 threads = NULL;
1116                 shmctl(shm_id, IPC_RMID, &sbuf);
1117         }
1118 }
1119
1120 /*
1121  * The thread area is shared between the main process and the job
1122  * threads/processes. So setup a shared memory segment that will hold
1123  * all the job info.
1124  */
1125 static int setup_thread_area(void)
1126 {
1127         /*
1128          * 1024 is too much on some machines, scale max_jobs if
1129          * we get a failure that looks like too large a shm segment
1130          */
1131         do {
1132                 size_t size = max_jobs * sizeof(struct thread_data);
1133
1134                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1135                 if (shm_id != -1)
1136                         break;
1137                 if (errno != EINVAL) {
1138                         perror("shmget");
1139                         break;
1140                 }
1141
1142                 max_jobs >>= 1;
1143         } while (max_jobs);
1144
1145         if (shm_id == -1)
1146                 return 1;
1147
1148         threads = shmat(shm_id, NULL, 0);
1149         if (threads == (void *) -1) {
1150                 perror("shmat");
1151                 return 1;
1152         }
1153
1154         atexit(free_shm);
1155         return 0;
1156 }
1157
1158 /*
1159  * Copy the fio options into the long options map, so we mirror
1160  * job and cmd line options.
1161  */
1162 static void dupe_job_options(void)
1163 {
1164         struct fio_option *o;
1165         unsigned int i;
1166
1167         i = 0;
1168         while (long_options[i].name)
1169                 i++;
1170
1171         o = &options[0];
1172         while (o->name) {
1173                 long_options[i].name = o->name;
1174                 long_options[i].val = FIO_GETOPT_JOB;
1175                 if (o->type == FIO_OPT_STR_SET)
1176                         long_options[i].has_arg = no_argument;
1177                 else
1178                         long_options[i].has_arg = required_argument;
1179
1180                 i++;
1181                 o++;
1182                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1183         }
1184 }
1185
1186 int parse_options(int argc, char *argv[])
1187 {
1188         int job_files, i;
1189
1190         f_out = stdout;
1191         f_err = stderr;
1192
1193         dupe_job_options();
1194
1195         if (setup_thread_area())
1196                 return 1;
1197         if (fill_def_thread())
1198                 return 1;
1199
1200         job_files = parse_cmd_line(argc, argv);
1201
1202         for (i = 0; i < job_files; i++) {
1203                 if (fill_def_thread())
1204                         return 1;
1205                 if (parse_jobs_ini(ini_file[i], i))
1206                         return 1;
1207                 free(ini_file[i]);
1208         }
1209
1210         free(ini_file);
1211
1212         if (!thread_number) {
1213                 log_err("No jobs defined(s)\n");
1214                 return 1;
1215         }
1216
1217         return 0;
1218 }