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 #ifndef FIO_HAVE_LIBAIO
 547         if (td->io_engine == FIO_LIBAIO) {
 548                 log_err("Linux libaio not available\n");
 549                 return 1;
 550         }
 551 #endif
 552 #ifndef FIO_HAVE_POSIXAIO
 553         if (td->io_engine == FIO_POSIXAIO) {
 554                 log_err("posix aio not available\n");
 555                 return 1;
 556         }
 557 #endif
 558
 559         /*
 560          * the def_thread is just for options, it's not a real job
 561          */
 562         if (td == &def_thread)
 563                 return 0;
 564
 565         /*
 566          * Set default io engine, if none set
 567          */
 568         if (!td->io_ops) {
 569                 td->io_ops = load_ioengine(td, DEF_IO_ENGINE_NAME);
 570                 if (!td->io_ops) {
 571                         log_err("default engine %s not there?\n", DEF_IO_ENGINE_NAME);
 572                         return 1;
 573                 }
 574         }
 575
 576         if (td->odirect)
 577                 td->io_ops->flags |= FIO_RAWIO;
 578
 579         fixup_options(td);
 580
 581         td->filetype = FIO_TYPE_FILE;
 582         if (!stat(jobname, &sb)) {
 583                 if (S_ISBLK(sb.st_mode))
 584                         td->filetype = FIO_TYPE_BD;
 585                 else if (S_ISCHR(sb.st_mode))
 586                         td->filetype = FIO_TYPE_CHAR;
 587         }
 588
 589         if (td->filename)
 590                 td->nr_uniq_files = 1;
 591         else
 592                 td->nr_uniq_files = td->nr_files;
 593
 594         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 595                 char tmp[PATH_MAX];
 596                 int len = 0;
 597                 int i;
 598
 599                 if (td->directory && td->directory[0] != '\0')
 600                         sprintf(tmp, "%s/", td->directory);
 601
 602                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 603
 604                 for_each_file(td, f, i) {
 605                         memset(f, 0, sizeof(*f));
 606                         f->fd = -1;
 607
 608                         if (td->filename)
 609                                 sprintf(tmp + len, "%s", td->filename);
 610                         else
 611                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 612                         f->file_name = strdup(tmp);
 613                 }
 614         } else {
 615                 td->nr_files = 1;
 616                 td->files = malloc(sizeof(struct fio_file));
 617                 f = &td->files[0];
 618
 619                 memset(f, 0, sizeof(*f));
 620                 f->fd = -1;
 621                 f->file_name = strdup(jobname);
 622         }
 623
 624         for_each_file(td, f, i) {
 625                 f->file_size = td->total_file_size / td->nr_files;
 626                 f->file_offset = td->start_offset;
 627         }
 628
 629         fio_sem_init(&td->mutex, 0);
 630
 631         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
 632         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
 633         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
 634
 635         if (td->stonewall && td->thread_number > 1)
 636                 groupid++;
 637
 638         td->groupid = groupid;
 639
 640         if (setup_rate(td))
 641                 goto err;
 642
 643         if (td->write_lat_log) {
 644                 setup_log(&td->slat_log);
 645                 setup_log(&td->clat_log);
 646         }
 647         if (td->write_bw_log)
 648                 setup_log(&td->bw_log);
 649
 650         if (!td->name)
 651                 td->name = strdup(jobname);
 652
 653         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 654
 655         if (!terse_output) {
 656                 if (!job_add_num) {
 657                         if (td->io_ops->flags & FIO_CPUIO)
 658                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 659                         else
 660                                 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);
 661                 } else if (job_add_num == 1)
 662                         fprintf(f_out, "...\n");
 663         }
 664
 665         /*
 666          * recurse add identical jobs, clear numjobs and stonewall options
 667          * as they don't apply to sub-jobs
 668          */
 669         numjobs = td->numjobs;
 670         while (--numjobs) {
 671                 struct thread_data *td_new = get_new_job(0, td);
 672
 673                 if (!td_new)
 674                         goto err;
 675
 676                 td_new->numjobs = 1;
 677                 td_new->stonewall = 0;
 678                 job_add_num = numjobs - 1;
 679
 680                 if (add_job(td_new, jobname, job_add_num))
 681                         goto err;
 682         }
 683         return 0;
 684 err:
 685         put_job(td);
 686         return -1;
 687 }
 688
 689 /*
 690  * Initialize the various random states we need (random io, block size ranges,
 691  * read/write mix, etc).
 692  */
 693 int init_random_state(struct thread_data *td)
 694 {
 695         unsigned long seeds[4];
 696         int fd, num_maps, blocks, i;
 697         struct fio_file *f;
 698
 699         if (td->io_ops->flags & FIO_CPUIO)
 700                 return 0;
 701
 702         fd = open("/dev/urandom", O_RDONLY);
 703         if (fd == -1) {
 704                 td_verror(td, errno);
 705                 return 1;
 706         }
 707
 708         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 709                 td_verror(td, EIO);
 710                 close(fd);
 711                 return 1;
 712         }
 713
 714         close(fd);
 715
 716         os_random_seed(seeds[0], &td->bsrange_state);
 717         os_random_seed(seeds[1], &td->verify_state);
 718         os_random_seed(seeds[2], &td->rwmix_state);
 719
 720         if (td->sequential)
 721                 return 0;
 722
 723         if (td->rand_repeatable)
 724                 seeds[3] = DEF_RANDSEED;
 725
 726         if (!td->norandommap) {
 727                 for_each_file(td, f, i) {
 728                         blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 729                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 730                         f->file_map = malloc(num_maps * sizeof(long));
 731                         f->num_maps = num_maps;
 732                         memset(f->file_map, 0, num_maps * sizeof(long));
 733                 }
 734         }
 735
 736         os_random_seed(seeds[3], &td->random_state);
 737         return 0;
 738 }
 739
 740 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 741 {
 742 #ifdef FIO_HAVE_CPU_AFFINITY
 743         unsigned int i;
 744
 745         CPU_ZERO(&cpumask);
 746
 747         for (i = 0; i < sizeof(int) * 8; i++) {
 748                 if ((1 << i) & cpu)
 749                         CPU_SET(i, &cpumask);
 750         }
 751 #endif
 752 }
 753
 754 static int is_empty_or_comment(char *line)
 755 {
 756         unsigned int i;
 757
 758         for (i = 0; i < strlen(line); i++) {
 759                 if (line[i] == ';')
 760                         return 1;
 761                 if (!isspace(line[i]) && !iscntrl(line[i]))
 762                         return 0;
 763         }
 764
 765         return 1;
 766 }
 767
 768 static int str_rw_cb(void *data, const char *mem)
 769 {
 770         struct thread_data *td = data;
 771
 772         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 773                 td->ddir = DDIR_READ;
 774                 td->sequential = 1;
 775                 return 0;
 776         } else if (!strncmp(mem, "randread", 8)) {
 777                 td->ddir = DDIR_READ;
 778                 td->sequential = 0;
 779                 return 0;
 780         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 781                 td->ddir = DDIR_WRITE;
 782                 td->sequential = 1;
 783                 return 0;
 784         } else if (!strncmp(mem, "randwrite", 9)) {
 785                 td->ddir = DDIR_WRITE;
 786                 td->sequential = 0;
 787                 return 0;
 788         } else if (!strncmp(mem, "rw", 2)) {
 789                 td->ddir = 0;
 790                 td->iomix = 1;
 791                 td->sequential = 1;
 792                 return 0;
 793         } else if (!strncmp(mem, "randrw", 6)) {
 794                 td->ddir = 0;
 795                 td->iomix = 1;
 796                 td->sequential = 0;
 797                 return 0;
 798         }
 799
 800         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 801         return 1;
 802 }
 803
 804 static int str_verify_cb(void *data, const char *mem)
 805 {
 806         struct thread_data *td = data;
 807
 808         if (!strncmp(mem, "0", 1)) {
 809                 td->verify = VERIFY_NONE;
 810                 return 0;
 811         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 812                 td->verify = VERIFY_MD5;
 813                 return 0;
 814         } else if (!strncmp(mem, "crc32", 5)) {
 815                 td->verify = VERIFY_CRC32;
 816                 return 0;
 817         }
 818
 819         log_err("fio: verify types: md5, crc32\n");
 820         return 1;
 821 }
 822
 823 static int str_mem_cb(void *data, const char *mem)
 824 {
 825         struct thread_data *td = data;
 826
 827         if (!strncmp(mem, "malloc", 6)) {
 828                 td->mem_type = MEM_MALLOC;
 829                 return 0;
 830         } else if (!strncmp(mem, "shm", 3)) {
 831                 td->mem_type = MEM_SHM;
 832                 return 0;
 833         } else if (!strncmp(mem, "mmap", 4)) {
 834                 td->mem_type = MEM_MMAP;
 835                 return 0;
 836         }
 837
 838         log_err("fio: mem type: malloc, shm, mmap\n");
 839         return 1;
 840 }
 841
 842 static int str_ioengine_cb(void *data, const char *str)
 843 {
 844         struct thread_data *td = data;
 845
 846         td->io_ops = load_ioengine(td, str);
 847         if (td->io_ops)
 848                 return 0;
 849
 850         log_err("fio: ioengine: { linuxaio, aio, libaio }, posixaio, sync, mmap, sgio, splice, cpu\n");
 851         return 1;
 852 }
 853
 854 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
 855 {
 856         mlock_size = *val;
 857         return 0;
 858 }
 859
 860 static int str_prioclass_cb(void *data, unsigned int *val)
 861 {
 862         struct thread_data *td = data;
 863
 864         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
 865         return 0;
 866 }
 867
 868 static int str_prio_cb(void *data, unsigned int *val)
 869 {
 870         struct thread_data *td = data;
 871
 872         td->ioprio |= *val;
 873         return 0;
 874 }
 875
 876 static int str_exitall_cb(void)
 877 {
 878         exitall_on_terminate = 1;
 879         return 0;
 880 }
 881
 882 static int str_cpumask_cb(void *data, unsigned int *val)
 883 {
 884         struct thread_data *td = data;
 885
 886         fill_cpu_mask(td->cpumask, *val);
 887         return 0;
 888 }
 889
 890 /*
 891  * This is our [ini] type file parser.
 892  */
 893 int parse_jobs_ini(char *file, int stonewall_flag)
 894 {
 895         unsigned int global;
 896         struct thread_data *td;
 897         char *string, *name;
 898         fpos_t off;
 899         FILE *f;
 900         char *p;
 901         int ret = 0, stonewall;
 902
 903         f = fopen(file, "r");
 904         if (!f) {
 905                 perror("fopen job file");
 906                 return 1;
 907         }
 908
 909         string = malloc(4096);
 910         name = malloc(256);
 911         memset(name, 0, 256);
 912
 913         stonewall = stonewall_flag;
 914         do {
 915                 p = fgets(string, 4095, f);
 916                 if (!p)
 917                         break;
 918                 if (is_empty_or_comment(p))
 919                         continue;
 920                 if (sscanf(p, "[%255s]", name) != 1)
 921                         continue;
 922
 923                 global = !strncmp(name, "global", 6);
 924
 925                 name[strlen(name) - 1] = '\0';
 926
 927                 td = get_new_job(global, &def_thread);
 928                 if (!td) {
 929                         ret = 1;
 930                         break;
 931                 }
 932
 933                 /*
 934                  * Seperate multiple job files by a stonewall
 935                  */
 936                 if (!global && stonewall) {
 937                         td->stonewall = stonewall;
 938                         stonewall = 0;
 939                 }
 940
 941                 fgetpos(f, &off);
 942                 while ((p = fgets(string, 4096, f)) != NULL) {
 943                         if (is_empty_or_comment(p))
 944                                 continue;
 945
 946                         strip_blank_front(&p);
 947
 948                         if (p[0] == '[')
 949                                 break;
 950
 951                         strip_blank_end(p);
 952
 953                         fgetpos(f, &off);
 954
 955                         /*
 956                          * Don't break here, continue parsing options so we
 957                          * dump all the bad ones. Makes trial/error fixups
 958                          * easier on the user.
 959                          */
 960                         ret |= parse_option(p, options, td);
 961                 }
 962
 963                 if (!ret) {
 964                         fsetpos(f, &off);
 965                         ret = add_job(td, name, 0);
 966                 } else {
 967                         log_err("fio: job %s dropped\n", name);
 968                         put_job(td);
 969                 }
 970         } while (!ret);
 971
 972         free(string);
 973         free(name);
 974         fclose(f);
 975         return ret;
 976 }
 977
 978 static int fill_def_thread(void)
 979 {
 980         memset(&def_thread, 0, sizeof(def_thread));
 981
 982         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
 983                 perror("sched_getaffinity");
 984                 return 1;
 985         }
 986
 987         /*
 988          * fill globals
 989          */
 990         def_thread.ddir = DDIR_READ;
 991         def_thread.iomix = 0;
 992         def_thread.bs[DDIR_READ] = DEF_BS;
 993         def_thread.bs[DDIR_WRITE] = DEF_BS;
 994         def_thread.min_bs[DDIR_READ] = def_thread.min_bs[DDIR_WRITE] = 0;
 995         def_thread.max_bs[DDIR_READ] = def_thread.max_bs[DDIR_WRITE] = 0;
 996         def_thread.odirect = DEF_ODIRECT;
 997         def_thread.ratecycle = DEF_RATE_CYCLE;
 998         def_thread.sequential = DEF_SEQUENTIAL;
 999         def_thread.timeout = def_timeout;
1000         def_thread.overwrite = DEF_OVERWRITE;
1001         def_thread.invalidate_cache = DEF_INVALIDATE;
1002         def_thread.sync_io = DEF_SYNCIO;
1003         def_thread.mem_type = MEM_MALLOC;
1004         def_thread.bw_avg_time = DEF_BWAVGTIME;
1005         def_thread.create_serialize = DEF_CREATE_SER;
1006         def_thread.create_fsync = DEF_CREATE_FSYNC;
1007         def_thread.loops = DEF_LOOPS;
1008         def_thread.verify = DEF_VERIFY;
1009         def_thread.stonewall = DEF_STONEWALL;
1010         def_thread.numjobs = DEF_NUMJOBS;
1011         def_thread.use_thread = DEF_USE_THREAD;
1012         def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
1013         def_thread.rwmixread = DEF_RWMIX_READ;
1014         def_thread.nice = DEF_NICE;
1015         def_thread.rand_repeatable = DEF_RAND_REPEAT;
1016         def_thread.nr_files = DEF_NR_FILES;
1017         def_thread.unlink = DEF_UNLINK;
1018         def_thread.write_bw_log = write_bw_log;
1019         def_thread.write_lat_log = write_lat_log;
1020         def_thread.norandommap = DEF_NO_RAND_MAP;
1021 #ifdef FIO_HAVE_DISK_UTIL
1022         def_thread.do_disk_util = 1;
1023 #endif
1024
1025         return 0;
1026 }
1027
1028 static void usage(void)
1029 {
1030         printf("%s\n", fio_version_string);
1031         printf("\t--output\tWrite output to file\n");
1032         printf("\t--timeout\tRuntime in seconds\n");
1033         printf("\t--latency-log\tGenerate per-job latency logs\n");
1034         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1035         printf("\t--minimal\tMinimal (terse) output\n");
1036         printf("\t--version\tPrint version info and exit\n");
1037 }
1038
1039 static int parse_cmd_line(int argc, char *argv[])
1040 {
1041         struct thread_data *td = NULL;
1042         int c, ini_idx = 0, lidx, ret;
1043
1044         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1045                 switch (c) {
1046                 case 't':
1047                         def_timeout = atoi(optarg);
1048                         break;
1049                 case 'l':
1050                         write_lat_log = 1;
1051                         break;
1052                 case 'w':
1053                         write_bw_log = 1;
1054                         break;
1055                 case 'o':
1056                         f_out = fopen(optarg, "w+");
1057                         if (!f_out) {
1058                                 perror("fopen output");
1059                                 exit(1);
1060                         }
1061                         f_err = f_out;
1062                         break;
1063                 case 'm':
1064                         terse_output = 1;
1065                         break;
1066                 case 'h':
1067                         usage();
1068                         exit(0);
1069                 case 'v':
1070                         printf("%s\n", fio_version_string);
1071                         exit(0);
1072                 case FIO_GETOPT_JOB: {
1073                         const char *opt = long_options[lidx].name;
1074                         char *val = optarg;
1075
1076                         if (!strncmp(opt, "name", 4) && td) {
1077                                 ret = add_job(td, td->name ?: "fio", 0);
1078                                 if (ret) {
1079                                         put_job(td);
1080                                         return 0;
1081                                 }
1082                                 td = NULL;
1083                         }
1084                         if (!td) {
1085                                 int global = !strncmp(val, "global", 6);
1086
1087                                 td = get_new_job(global, &def_thread);
1088                                 if (!td)
1089                                         return 0;
1090                         }
1091
1092                         ret = parse_cmd_option(opt, val, options, td);
1093                         if (ret) {
1094                                 log_err("fio: job dropped\n");
1095                                 put_job(td);
1096                                 td = NULL;
1097                         }
1098                         break;
1099                 }
1100                 default:
1101                         printf("optarg <<%s>>\n", argv[optind]);
1102                         break;
1103                 }
1104         }
1105
1106         if (td) {
1107                 ret = add_job(td, td->name ?: "fio", 0);
1108                 if (ret)
1109                         put_job(td);
1110         }
1111
1112         while (optind < argc) {
1113                 ini_idx++;
1114                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1115                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1116                 optind++;
1117         }
1118
1119         return ini_idx;
1120 }
1121
1122 static void free_shm(void)
1123 {
1124         struct shmid_ds sbuf;
1125
1126         if (threads) {
1127                 shmdt((void *) threads);
1128                 threads = NULL;
1129                 shmctl(shm_id, IPC_RMID, &sbuf);
1130         }
1131 }
1132
1133 /*
1134  * The thread area is shared between the main process and the job
1135  * threads/processes. So setup a shared memory segment that will hold
1136  * all the job info.
1137  */
1138 static int setup_thread_area(void)
1139 {
1140         /*
1141          * 1024 is too much on some machines, scale max_jobs if
1142          * we get a failure that looks like too large a shm segment
1143          */
1144         do {
1145                 size_t size = max_jobs * sizeof(struct thread_data);
1146
1147                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1148                 if (shm_id != -1)
1149                         break;
1150                 if (errno != EINVAL) {
1151                         perror("shmget");
1152                         break;
1153                 }
1154
1155                 max_jobs >>= 1;
1156         } while (max_jobs);
1157
1158         if (shm_id == -1)
1159                 return 1;
1160
1161         threads = shmat(shm_id, NULL, 0);
1162         if (threads == (void *) -1) {
1163                 perror("shmat");
1164                 return 1;
1165         }
1166
1167         atexit(free_shm);
1168         return 0;
1169 }
1170
1171 /*
1172  * Copy the fio options into the long options map, so we mirror
1173  * job and cmd line options.
1174  */
1175 static void dupe_job_options(void)
1176 {
1177         struct fio_option *o;
1178         unsigned int i;
1179
1180         i = 0;
1181         while (long_options[i].name)
1182                 i++;
1183
1184         o = &options[0];
1185         while (o->name) {
1186                 long_options[i].name = o->name;
1187                 long_options[i].val = FIO_GETOPT_JOB;
1188                 if (o->type == FIO_OPT_STR_SET)
1189                         long_options[i].has_arg = no_argument;
1190                 else
1191                         long_options[i].has_arg = required_argument;
1192
1193                 i++;
1194                 o++;
1195                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1196         }
1197 }
1198
1199 int parse_options(int argc, char *argv[])
1200 {
1201         int job_files, i;
1202
1203         f_out = stdout;
1204         f_err = stderr;
1205
1206         dupe_job_options();
1207
1208         if (setup_thread_area())
1209                 return 1;
1210         if (fill_def_thread())
1211                 return 1;
1212
1213         job_files = parse_cmd_line(argc, argv);
1214
1215         for (i = 0; i < job_files; i++) {
1216                 if (fill_def_thread())
1217                         return 1;
1218                 if (parse_jobs_ini(ini_file[i], i))
1219                         return 1;
1220                 free(ini_file[i]);
1221         }
1222
1223         free(ini_file);
1224
1225         if (!thread_number) {
1226                 log_err("No jobs defined(s)\n");
1227                 return 1;
1228         }
1229
1230         return 0;
1231 }