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           (0)
  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              (1)
  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 #ifdef FIO_HAVE_IOPRIO
  64 static int str_prio_cb(void *, unsigned int *);
  65 static int str_prioclass_cb(void *, unsigned int *);
  66 #endif
  67 static int str_exitall_cb(void);
  68 static int str_cpumask_cb(void *, unsigned int *);
  69
  70 /*
  71  * Map of job/command line options
  72  */
  73 static struct fio_option options[] = {
  74         {
  75                 .name   = "name",
  76                 .type   = FIO_OPT_STR_STORE,
  77                 .off1   = td_var_offset(name),
  78         },
  79         {
  80                 .name   = "directory",
  81                 .type   = FIO_OPT_STR_STORE,
  82                 .off1   = td_var_offset(directory),
  83         },
  84         {
  85                 .name   = "filename",
  86                 .type   = FIO_OPT_STR_STORE,
  87                 .off1   = td_var_offset(filename),
  88         },
  89         {
  90                 .name   = "rw",
  91                 .type   = FIO_OPT_STR,
  92                 .cb     = str_rw_cb,
  93         },
  94         {
  95                 .name   = "ioengine",
  96                 .type   = FIO_OPT_STR,
  97                 .cb     = str_ioengine_cb,
  98         },
  99         {
 100                 .name   = "mem",
 101                 .type   = FIO_OPT_STR,
 102                 .cb     = str_mem_cb,
 103         },
 104         {
 105                 .name   = "verify",
 106                 .type   = FIO_OPT_STR,
 107                 .cb     = str_verify_cb,
 108         },
 109         {
 110                 .name   = "write_iolog",
 111                 .type   = FIO_OPT_STR_STORE,
 112                 .off1   = td_var_offset(write_iolog_file),
 113         },
 114         {
 115                 .name   = "read_iolog",
 116                 .type   = FIO_OPT_STR_STORE,
 117                 .off1   = td_var_offset(read_iolog_file),
 118         },
 119         {
 120                 .name   = "exec_prerun",
 121                 .type   = FIO_OPT_STR_STORE,
 122                 .off1   = td_var_offset(exec_prerun),
 123         },
 124         {
 125                 .name   = "exec_postrun",
 126                 .type   = FIO_OPT_STR_STORE,
 127                 .off1   = td_var_offset(exec_postrun),
 128         },
 129 #ifdef FIO_HAVE_IOSCHED_SWITCH
 130         {
 131                 .name   = "ioscheduler",
 132                 .type   = FIO_OPT_STR_STORE,
 133                 .off1   = td_var_offset(ioscheduler),
 134         },
 135 #endif
 136         {
 137                 .name   = "size",
 138                 .type   = FIO_OPT_STR_VAL,
 139                 .off1   = td_var_offset(total_file_size),
 140         },
 141         {
 142                 .name   = "bs",
 143                 .type   = FIO_OPT_STR_VAL_INT,
 144                 .off1   = td_var_offset(bs[DDIR_READ]),
 145                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 146         },
 147         {
 148                 .name   = "offset",
 149                 .type   = FIO_OPT_STR_VAL,
 150                 .off1   = td_var_offset(start_offset),
 151         },
 152         {
 153                 .name   = "zonesize",
 154                 .type   = FIO_OPT_STR_VAL,
 155                 .off1   = td_var_offset(zone_size),
 156         },
 157         {
 158                 .name   = "zoneskip",
 159                 .type   = FIO_OPT_STR_VAL,
 160                 .off1   = td_var_offset(zone_skip),
 161         },
 162         {
 163                 .name   = "lockmem",
 164                 .type   = FIO_OPT_STR_VAL,
 165                 .cb     = str_lockmem_cb,
 166         },
 167         {
 168                 .name   = "bsrange",
 169                 .type   = FIO_OPT_RANGE,
 170                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 171                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 172                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 173                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 174         },
 175         {
 176                 .name   = "nrfiles",
 177                 .type   = FIO_OPT_INT,
 178                 .off1   = td_var_offset(nr_files),
 179         },
 180         {
 181                 .name   = "iodepth",
 182                 .type   = FIO_OPT_INT,
 183                 .off1   = td_var_offset(iodepth),
 184         },
 185         {
 186                 .name   = "fsync",
 187                 .type   = FIO_OPT_INT,
 188                 .off1   = td_var_offset(fsync_blocks),
 189         },
 190         {
 191                 .name   = "rwmixcycle",
 192                 .type   = FIO_OPT_INT,
 193                 .off1   = td_var_offset(rwmixcycle),
 194         },
 195         {
 196                 .name   = "rwmixread",
 197                 .type   = FIO_OPT_INT,
 198                 .off1   = td_var_offset(rwmixread),
 199                 .max_val= 100,
 200         },
 201         {
 202                 .name   = "rwmixwrite",
 203                 .type   = FIO_OPT_INT,
 204                 .off1   = td_var_offset(rwmixwrite),
 205                 .max_val= 100,
 206         },
 207         {
 208                 .name   = "nice",
 209                 .type   = FIO_OPT_INT,
 210                 .off1   = td_var_offset(nice),
 211         },
 212 #ifdef FIO_HAVE_IOPRIO
 213         {
 214                 .name   = "prio",
 215                 .type   = FIO_OPT_INT,
 216                 .cb     = str_prio_cb,
 217         },
 218         {
 219                 .name   = "prioclass",
 220                 .type   = FIO_OPT_INT,
 221                 .cb     = str_prioclass_cb,
 222         },
 223 #endif
 224         {
 225                 .name   = "thinktime",
 226                 .type   = FIO_OPT_INT,
 227                 .off1   = td_var_offset(thinktime)
 228         },
 229         {
 230                 .name   = "rate",
 231                 .type   = FIO_OPT_INT,
 232                 .off1   = td_var_offset(rate)
 233         },
 234         {
 235                 .name   = "ratemin",
 236                 .type   = FIO_OPT_INT,
 237                 .off1   = td_var_offset(ratemin)
 238         },
 239         {
 240                 .name   = "ratecycle",
 241                 .type   = FIO_OPT_INT,
 242                 .off1   = td_var_offset(ratecycle)
 243         },
 244         {
 245                 .name   = "startdelay",
 246                 .type   = FIO_OPT_INT,
 247                 .off1   = td_var_offset(start_delay)
 248         },
 249         {
 250                 .name   = "timeout",
 251                 .type   = FIO_OPT_STR_VAL_TIME,
 252                 .off1   = td_var_offset(timeout)
 253         },
 254         {
 255                 .name   = "invalidate",
 256                 .type   = FIO_OPT_INT,
 257                 .off1   = td_var_offset(invalidate_cache)
 258         },
 259         {
 260                 .name   = "sync",
 261                 .type   = FIO_OPT_INT,
 262                 .off1   = td_var_offset(sync_io)
 263         },
 264         {
 265                 .name   = "bwavgtime",
 266                 .type   = FIO_OPT_INT,
 267                 .off1   = td_var_offset(bw_avg_time)
 268         },
 269         {
 270                 .name   = "create_serialize",
 271                 .type   = FIO_OPT_INT,
 272                 .off1   = td_var_offset(create_serialize)
 273         },
 274         {
 275                 .name   = "create_fsync",
 276                 .type   = FIO_OPT_INT,
 277                 .off1   = td_var_offset(create_fsync)
 278         },
 279         {
 280                 .name   = "loops",
 281                 .type   = FIO_OPT_INT,
 282                 .off1   = td_var_offset(loops)
 283         },
 284         {
 285                 .name   = "numjobs",
 286                 .type   = FIO_OPT_INT,
 287                 .off1   = td_var_offset(numjobs)
 288         },
 289         {
 290                 .name   = "cpuload",
 291                 .type   = FIO_OPT_INT,
 292                 .off1   = td_var_offset(cpuload)
 293         },
 294         {
 295                 .name   = "cpuchunks",
 296                 .type   = FIO_OPT_INT,
 297                 .off1   = td_var_offset(cpucycle)
 298         },
 299         {
 300                 .name   = "direct",
 301                 .type   = FIO_OPT_INT,
 302                 .off1   = td_var_offset(odirect)
 303         },
 304         {
 305                 .name   = "overwrite",
 306                 .type   = FIO_OPT_INT,
 307                 .off1   = td_var_offset(overwrite)
 308         },
 309 #ifdef FIO_HAVE_CPU_AFFINITY
 310         {
 311                 .name   = "cpumask",
 312                 .type   = FIO_OPT_INT,
 313                 .cb     = str_cpumask_cb,
 314         },
 315 #endif
 316         {
 317                 .name   = "end_fsync",
 318                 .type   = FIO_OPT_INT,
 319                 .off1   = td_var_offset(end_fsync)
 320         },
 321         {
 322                 .name   = "unlink",
 323                 .type   = FIO_OPT_STR_SET,
 324                 .off1   = td_var_offset(unlink),
 325         },
 326         {
 327                 .name   = "exitall",
 328                 .type   = FIO_OPT_STR_SET,
 329                 .cb     = str_exitall_cb,
 330         },
 331         {
 332                 .name   = "stonewall",
 333                 .type   = FIO_OPT_STR_SET,
 334                 .off1   = td_var_offset(stonewall),
 335         },
 336         {
 337                 .name   = "thread",
 338                 .type   = FIO_OPT_STR_SET,
 339                 .off1   = td_var_offset(thread),
 340         },
 341         {
 342                 .name   = "write_bw_log",
 343                 .type   = FIO_OPT_STR_SET,
 344                 .off1   = td_var_offset(write_bw_log),
 345         },
 346         {
 347                 .name   = "write_lat_log",
 348                 .type   = FIO_OPT_STR_SET,
 349                 .off1   = td_var_offset(write_lat_log),
 350         },
 351         {
 352                 .name   = "norandommap",
 353                 .type   = FIO_OPT_STR_SET,
 354                 .off1   = td_var_offset(norandommap),
 355         },
 356         {
 357                 .name   = "bs_unaligned",
 358                 .type   = FIO_OPT_STR_SET,
 359                 .off1   = td_var_offset(bs_unaligned),
 360         },
 361         {
 362                 .name = NULL,
 363         },
 364 };
 365
 366 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 367 #define FIO_CMD_OPTS    (16)
 368 #define FIO_GETOPT_JOB  (0x89988998)
 369
 370 /*
 371  * Command line options. These will contain the above, plus a few
 372  * extra that only pertain to fio itself and not jobs.
 373  */
 374 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 375         {
 376                 .name           = "output",
 377                 .has_arg        = required_argument,
 378                 .val            = 'o',
 379         },
 380         {
 381                 .name           = "timeout",
 382                 .has_arg        = required_argument,
 383                 .val            = 't',
 384         },
 385         {
 386                 .name           = "latency-log",
 387                 .has_arg        = required_argument,
 388                 .val            = 'l',
 389         },
 390         {
 391                 .name           = "bandwidth-log",
 392                 .has_arg        = required_argument,
 393                 .val            = 'b',
 394         },
 395         {
 396                 .name           = "minimal",
 397                 .has_arg        = optional_argument,
 398                 .val            = 'm',
 399         },
 400         {
 401                 .name           = "version",
 402                 .has_arg        = no_argument,
 403                 .val            = 'v',
 404         },
 405         {
 406                 .name           = NULL,
 407         },
 408 };
 409
 410 static int def_timeout = DEF_TIMEOUT;
 411
 412 static char fio_version_string[] = "fio 1.9";
 413
 414 static char **ini_file;
 415 static int max_jobs = MAX_JOBS;
 416
 417 struct thread_data def_thread;
 418 struct thread_data *threads = NULL;
 419
 420 int exitall_on_terminate = 0;
 421 int terse_output = 0;
 422 unsigned long long mlock_size = 0;
 423 FILE *f_out = NULL;
 424 FILE *f_err = NULL;
 425
 426 static int write_lat_log = DEF_WRITE_LAT_LOG;
 427 static int write_bw_log = DEF_WRITE_BW_LOG;
 428
 429 /*
 430  * Return a free job structure.
 431  */
 432 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 433 {
 434         struct thread_data *td;
 435
 436         if (global)
 437                 return &def_thread;
 438         if (thread_number >= max_jobs)
 439                 return NULL;
 440
 441         td = &threads[thread_number++];
 442         *td = *parent;
 443
 444         td->thread_number = thread_number;
 445         return td;
 446 }
 447
 448 static void put_job(struct thread_data *td)
 449 {
 450         if (td == &def_thread)
 451                 return;
 452
 453         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 454         thread_number--;
 455 }
 456
 457 /*
 458  * Lazy way of fixing up options that depend on each other. We could also
 459  * define option callback handlers, but this is easier.
 460  */
 461 static void fixup_options(struct thread_data *td)
 462 {
 463         if (!td->rwmixread && td->rwmixwrite)
 464                 td->rwmixread = 100 - td->rwmixwrite;
 465
 466         if (td->write_iolog_file && td->read_iolog_file) {
 467                 log_err("fio: read iolog overrides write_iolog\n");
 468                 free(td->write_iolog_file);
 469                 td->write_iolog_file = NULL;
 470         }
 471
 472         if (td->io_ops->flags & FIO_SYNCIO)
 473                 td->iodepth = 1;
 474         else {
 475                 if (!td->iodepth)
 476                         td->iodepth = td->nr_files;
 477         }
 478
 479         /*
 480          * only really works for sequential io for now, and with 1 file
 481          */
 482         if (td->zone_size && !td->sequential && td->nr_files == 1)
 483                 td->zone_size = 0;
 484
 485         /*
 486          * Reads can do overwrites, we always need to pre-create the file
 487          */
 488         if (td_read(td) || td_rw(td))
 489                 td->overwrite = 1;
 490
 491         if (!td->min_bs[DDIR_READ])
 492                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 493         if (!td->max_bs[DDIR_READ])
 494                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 495         if (!td->min_bs[DDIR_WRITE])
 496                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 497         if (!td->max_bs[DDIR_WRITE])
 498                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 499
 500         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 501
 502         if (td_read(td) && !td_rw(td))
 503                 td->verify = 0;
 504
 505         if (td->norandommap && td->verify != VERIFY_NONE) {
 506                 log_err("fio: norandommap given, verify disabled\n");
 507                 td->verify = VERIFY_NONE;
 508         }
 509         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 510                 log_err("fio: bs_unaligned may not work with raw io\n");
 511
 512         /*
 513          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 514          */
 515         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 516                 td->odirect = 0;
 517 }
 518
 519 /*
 520  * This function leaks the buffer
 521  */
 522 static char *to_kmg(unsigned int val)
 523 {
 524         char *buf = malloc(32);
 525         char post[] = { 0, 'K', 'M', 'G', 'P', 0 };
 526         char *p = post;
 527
 528         do {
 529                 if (val & 1023)
 530                         break;
 531
 532                 val >>= 10;
 533                 p++;
 534         } while (*p);
 535
 536         snprintf(buf, 31, "%u%c", val, *p);
 537         return buf;
 538 }
 539
 540 /*
 541  * Adds a job to the list of things todo. Sanitizes the various options
 542  * to make sure we don't have conflicts, and initializes various
 543  * members of td.
 544  */
 545 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 546 {
 547         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
 548                                    "rw", NULL, "randrw" };
 549         struct stat sb;
 550         int numjobs, ddir, i;
 551         struct fio_file *f;
 552
 553         /*
 554          * the def_thread is just for options, it's not a real job
 555          */
 556         if (td == &def_thread)
 557                 return 0;
 558
 559         /*
 560          * Set default io engine, if none set
 561          */
 562         if (!td->io_ops) {
 563                 td->io_ops = load_ioengine(td, DEF_IO_ENGINE_NAME);
 564                 if (!td->io_ops) {
 565                         log_err("default engine %s not there?\n", DEF_IO_ENGINE_NAME);
 566                         return 1;
 567                 }
 568         }
 569
 570         if (td->odirect)
 571                 td->io_ops->flags |= FIO_RAWIO;
 572
 573         td->filetype = FIO_TYPE_FILE;
 574         if (!stat(jobname, &sb)) {
 575                 if (S_ISBLK(sb.st_mode))
 576                         td->filetype = FIO_TYPE_BD;
 577                 else if (S_ISCHR(sb.st_mode))
 578                         td->filetype = FIO_TYPE_CHAR;
 579         }
 580
 581         fixup_options(td);
 582
 583         if (td->filename)
 584                 td->nr_uniq_files = 1;
 585         else
 586                 td->nr_uniq_files = td->nr_files;
 587
 588         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 589                 char tmp[PATH_MAX];
 590                 int len = 0;
 591
 592                 if (td->directory && td->directory[0] != '\0')
 593                         sprintf(tmp, "%s/", td->directory);
 594
 595                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 596
 597                 for_each_file(td, f, i) {
 598                         memset(f, 0, sizeof(*f));
 599                         f->fd = -1;
 600
 601                         if (td->filename)
 602                                 sprintf(tmp + len, "%s", td->filename);
 603                         else
 604                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 605                         f->file_name = strdup(tmp);
 606                 }
 607         } else {
 608                 td->nr_files = 1;
 609                 td->files = malloc(sizeof(struct fio_file));
 610                 f = &td->files[0];
 611
 612                 memset(f, 0, sizeof(*f));
 613                 f->fd = -1;
 614                 f->file_name = strdup(jobname);
 615         }
 616
 617         for_each_file(td, f, i) {
 618                 f->file_size = td->total_file_size / td->nr_files;
 619                 f->file_offset = td->start_offset;
 620         }
 621
 622         fio_sem_init(&td->mutex, 0);
 623
 624         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
 625         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
 626         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
 627
 628         if (td->stonewall && td->thread_number > 1)
 629                 groupid++;
 630
 631         td->groupid = groupid;
 632
 633         if (setup_rate(td))
 634                 goto err;
 635
 636         if (td->write_lat_log) {
 637                 setup_log(&td->slat_log);
 638                 setup_log(&td->clat_log);
 639         }
 640         if (td->write_bw_log)
 641                 setup_log(&td->bw_log);
 642
 643         if (!td->name)
 644                 td->name = strdup(jobname);
 645
 646         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 647
 648         if (!terse_output) {
 649                 if (!job_add_num) {
 650                         if (td->io_ops->flags & FIO_CPUIO)
 651                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 652                         else {
 653                                 char *c1, *c2, *c3, *c4;
 654
 655                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 656                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 657                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 658                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 659
 660                                 fprintf(f_out, "%s: (g=%d): rw=%s, odir=%u, bs=%s-%s/%s-%s, rate=%u, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[ddir], td->odirect, c1, c2, c3, c4, td->rate, td->io_ops->name, td->iodepth);
 661
 662                                 free(c1);
 663                                 free(c2);
 664                                 free(c3);
 665                                 free(c4);
 666                         }
 667                 } else if (job_add_num == 1)
 668                         fprintf(f_out, "...\n");
 669         }
 670
 671         /*
 672          * recurse add identical jobs, clear numjobs and stonewall options
 673          * as they don't apply to sub-jobs
 674          */
 675         numjobs = td->numjobs;
 676         while (--numjobs) {
 677                 struct thread_data *td_new = get_new_job(0, td);
 678
 679                 if (!td_new)
 680                         goto err;
 681
 682                 td_new->numjobs = 1;
 683                 td_new->stonewall = 0;
 684                 job_add_num = numjobs - 1;
 685
 686                 if (add_job(td_new, jobname, job_add_num))
 687                         goto err;
 688         }
 689         return 0;
 690 err:
 691         put_job(td);
 692         return -1;
 693 }
 694
 695 /*
 696  * Initialize the various random states we need (random io, block size ranges,
 697  * read/write mix, etc).
 698  */
 699 int init_random_state(struct thread_data *td)
 700 {
 701         unsigned long seeds[4];
 702         int fd, num_maps, blocks, i;
 703         struct fio_file *f;
 704
 705         if (td->io_ops->flags & FIO_CPUIO)
 706                 return 0;
 707
 708         fd = open("/dev/urandom", O_RDONLY);
 709         if (fd == -1) {
 710                 td_verror(td, errno);
 711                 return 1;
 712         }
 713
 714         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 715                 td_verror(td, EIO);
 716                 close(fd);
 717                 return 1;
 718         }
 719
 720         close(fd);
 721
 722         os_random_seed(seeds[0], &td->bsrange_state);
 723         os_random_seed(seeds[1], &td->verify_state);
 724         os_random_seed(seeds[2], &td->rwmix_state);
 725
 726         if (td->sequential)
 727                 return 0;
 728
 729         if (td->rand_repeatable)
 730                 seeds[3] = DEF_RANDSEED;
 731
 732         if (!td->norandommap) {
 733                 for_each_file(td, f, i) {
 734                         blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 735                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 736                         f->file_map = malloc(num_maps * sizeof(long));
 737                         f->num_maps = num_maps;
 738                         memset(f->file_map, 0, num_maps * sizeof(long));
 739                 }
 740         }
 741
 742         os_random_seed(seeds[3], &td->random_state);
 743         return 0;
 744 }
 745
 746 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 747 {
 748 #ifdef FIO_HAVE_CPU_AFFINITY
 749         unsigned int i;
 750
 751         CPU_ZERO(&cpumask);
 752
 753         for (i = 0; i < sizeof(int) * 8; i++) {
 754                 if ((1 << i) & cpu)
 755                         CPU_SET(i, &cpumask);
 756         }
 757 #endif
 758 }
 759
 760 static int is_empty_or_comment(char *line)
 761 {
 762         unsigned int i;
 763
 764         for (i = 0; i < strlen(line); i++) {
 765                 if (line[i] == ';')
 766                         return 1;
 767                 if (!isspace(line[i]) && !iscntrl(line[i]))
 768                         return 0;
 769         }
 770
 771         return 1;
 772 }
 773
 774 static int str_rw_cb(void *data, const char *mem)
 775 {
 776         struct thread_data *td = data;
 777
 778         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 779                 td->ddir = DDIR_READ;
 780                 td->sequential = 1;
 781                 return 0;
 782         } else if (!strncmp(mem, "randread", 8)) {
 783                 td->ddir = DDIR_READ;
 784                 td->sequential = 0;
 785                 return 0;
 786         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 787                 td->ddir = DDIR_WRITE;
 788                 td->sequential = 1;
 789                 return 0;
 790         } else if (!strncmp(mem, "randwrite", 9)) {
 791                 td->ddir = DDIR_WRITE;
 792                 td->sequential = 0;
 793                 return 0;
 794         } else if (!strncmp(mem, "rw", 2)) {
 795                 td->ddir = DDIR_READ;
 796                 td->iomix = 1;
 797                 td->sequential = 1;
 798                 return 0;
 799         } else if (!strncmp(mem, "randrw", 6)) {
 800                 td->ddir = DDIR_READ;
 801                 td->iomix = 1;
 802                 td->sequential = 0;
 803                 return 0;
 804         }
 805
 806         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 807         return 1;
 808 }
 809
 810 static int str_verify_cb(void *data, const char *mem)
 811 {
 812         struct thread_data *td = data;
 813
 814         if (!strncmp(mem, "0", 1)) {
 815                 td->verify = VERIFY_NONE;
 816                 return 0;
 817         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 818                 td->verify = VERIFY_MD5;
 819                 return 0;
 820         } else if (!strncmp(mem, "crc32", 5)) {
 821                 td->verify = VERIFY_CRC32;
 822                 return 0;
 823         }
 824
 825         log_err("fio: verify types: md5, crc32\n");
 826         return 1;
 827 }
 828
 829 static int str_mem_cb(void *data, const char *mem)
 830 {
 831         struct thread_data *td = data;
 832
 833         if (!strncmp(mem, "malloc", 6)) {
 834                 td->mem_type = MEM_MALLOC;
 835                 return 0;
 836         } else if (!strncmp(mem, "mmap", 4)) {
 837                 td->mem_type = MEM_MMAP;
 838                 return 0;
 839         } else if (!strncmp(mem, "shmhuge", 7)) {
 840 #ifdef FIO_HAVE_HUGETLB
 841                 td->mem_type = MEM_SHMHUGE;
 842                 return 0;
 843 #else
 844                 log_err("fio: shmhuge not available\n");
 845                 return 1;
 846 #endif
 847         } else if (!strncmp(mem, "shm", 3)) {
 848                 td->mem_type = MEM_SHM;
 849                 return 0;
 850         }
 851
 852         log_err("fio: mem type: malloc, shm, mmap, shmhuge\n");
 853         return 1;
 854 }
 855
 856 static int str_ioengine_cb(void *data, const char *str)
 857 {
 858         struct thread_data *td = data;
 859
 860         td->io_ops = load_ioengine(td, str);
 861         if (td->io_ops)
 862                 return 0;
 863
 864         log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
 865         log_err("fio: or specify path to dynamic ioengine module\n");
 866         return 1;
 867 }
 868
 869 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
 870 {
 871         mlock_size = *val;
 872         return 0;
 873 }
 874
 875 #ifdef FIO_HAVE_IOPRIO
 876 static int str_prioclass_cb(void *data, unsigned int *val)
 877 {
 878         struct thread_data *td = data;
 879
 880         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
 881         return 0;
 882 }
 883
 884 static int str_prio_cb(void *data, unsigned int *val)
 885 {
 886         struct thread_data *td = data;
 887
 888         td->ioprio |= *val;
 889         return 0;
 890 }
 891 #endif
 892
 893 static int str_exitall_cb(void)
 894 {
 895         exitall_on_terminate = 1;
 896         return 0;
 897 }
 898
 899 static int str_cpumask_cb(void *data, unsigned int *val)
 900 {
 901         struct thread_data *td = data;
 902
 903         fill_cpu_mask(td->cpumask, *val);
 904         return 0;
 905 }
 906
 907 /*
 908  * This is our [ini] type file parser.
 909  */
 910 static int parse_jobs_ini(char *file, int stonewall_flag)
 911 {
 912         unsigned int global;
 913         struct thread_data *td;
 914         char *string, *name;
 915         fpos_t off;
 916         FILE *f;
 917         char *p;
 918         int ret = 0, stonewall;
 919
 920         f = fopen(file, "r");
 921         if (!f) {
 922                 perror("fopen job file");
 923                 return 1;
 924         }
 925
 926         string = malloc(4096);
 927         name = malloc(256);
 928         memset(name, 0, 256);
 929
 930         stonewall = stonewall_flag;
 931         do {
 932                 p = fgets(string, 4095, f);
 933                 if (!p)
 934                         break;
 935                 if (is_empty_or_comment(p))
 936                         continue;
 937                 if (sscanf(p, "[%255s]", name) != 1)
 938                         continue;
 939
 940                 global = !strncmp(name, "global", 6);
 941
 942                 name[strlen(name) - 1] = '\0';
 943
 944                 td = get_new_job(global, &def_thread);
 945                 if (!td) {
 946                         ret = 1;
 947                         break;
 948                 }
 949
 950                 /*
 951                  * Seperate multiple job files by a stonewall
 952                  */
 953                 if (!global && stonewall) {
 954                         td->stonewall = stonewall;
 955                         stonewall = 0;
 956                 }
 957
 958                 fgetpos(f, &off);
 959                 while ((p = fgets(string, 4096, f)) != NULL) {
 960                         if (is_empty_or_comment(p))
 961                                 continue;
 962
 963                         strip_blank_front(&p);
 964
 965                         if (p[0] == '[')
 966                                 break;
 967
 968                         strip_blank_end(p);
 969
 970                         fgetpos(f, &off);
 971
 972                         /*
 973                          * Don't break here, continue parsing options so we
 974                          * dump all the bad ones. Makes trial/error fixups
 975                          * easier on the user.
 976                          */
 977                         ret |= parse_option(p, options, td);
 978                 }
 979
 980                 if (!ret) {
 981                         fsetpos(f, &off);
 982                         ret = add_job(td, name, 0);
 983                 } else {
 984                         log_err("fio: job %s dropped\n", name);
 985                         put_job(td);
 986                 }
 987         } while (!ret);
 988
 989         free(string);
 990         free(name);
 991         fclose(f);
 992         return ret;
 993 }
 994
 995 static int fill_def_thread(void)
 996 {
 997         memset(&def_thread, 0, sizeof(def_thread));
 998
 999         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1000                 perror("sched_getaffinity");
1001                 return 1;
1002         }
1003
1004         /*
1005          * fill globals
1006          */
1007         def_thread.ddir = DDIR_READ;
1008         def_thread.iomix = 0;
1009         def_thread.bs[DDIR_READ] = DEF_BS;
1010         def_thread.bs[DDIR_WRITE] = DEF_BS;
1011         def_thread.min_bs[DDIR_READ] = def_thread.min_bs[DDIR_WRITE] = 0;
1012         def_thread.max_bs[DDIR_READ] = def_thread.max_bs[DDIR_WRITE] = 0;
1013         def_thread.odirect = DEF_ODIRECT;
1014         def_thread.ratecycle = DEF_RATE_CYCLE;
1015         def_thread.sequential = DEF_SEQUENTIAL;
1016         def_thread.timeout = def_timeout;
1017         def_thread.overwrite = DEF_OVERWRITE;
1018         def_thread.invalidate_cache = DEF_INVALIDATE;
1019         def_thread.sync_io = DEF_SYNCIO;
1020         def_thread.mem_type = MEM_MALLOC;
1021         def_thread.bw_avg_time = DEF_BWAVGTIME;
1022         def_thread.create_serialize = DEF_CREATE_SER;
1023         def_thread.create_fsync = DEF_CREATE_FSYNC;
1024         def_thread.loops = DEF_LOOPS;
1025         def_thread.verify = DEF_VERIFY;
1026         def_thread.stonewall = DEF_STONEWALL;
1027         def_thread.numjobs = DEF_NUMJOBS;
1028         def_thread.use_thread = DEF_USE_THREAD;
1029         def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
1030         def_thread.rwmixread = DEF_RWMIX_READ;
1031         def_thread.nice = DEF_NICE;
1032         def_thread.rand_repeatable = DEF_RAND_REPEAT;
1033         def_thread.nr_files = DEF_NR_FILES;
1034         def_thread.unlink = DEF_UNLINK;
1035         def_thread.write_bw_log = write_bw_log;
1036         def_thread.write_lat_log = write_lat_log;
1037         def_thread.norandommap = DEF_NO_RAND_MAP;
1038 #ifdef FIO_HAVE_DISK_UTIL
1039         def_thread.do_disk_util = 1;
1040 #endif
1041
1042         return 0;
1043 }
1044
1045 static void usage(void)
1046 {
1047         printf("%s\n", fio_version_string);
1048         printf("\t--output\tWrite output to file\n");
1049         printf("\t--timeout\tRuntime in seconds\n");
1050         printf("\t--latency-log\tGenerate per-job latency logs\n");
1051         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1052         printf("\t--minimal\tMinimal (terse) output\n");
1053         printf("\t--version\tPrint version info and exit\n");
1054 }
1055
1056 static int parse_cmd_line(int argc, char *argv[])
1057 {
1058         struct thread_data *td = NULL;
1059         int c, ini_idx = 0, lidx, ret;
1060
1061         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1062                 switch (c) {
1063                 case 't':
1064                         def_timeout = atoi(optarg);
1065                         break;
1066                 case 'l':
1067                         write_lat_log = 1;
1068                         break;
1069                 case 'w':
1070                         write_bw_log = 1;
1071                         break;
1072                 case 'o':
1073                         f_out = fopen(optarg, "w+");
1074                         if (!f_out) {
1075                                 perror("fopen output");
1076                                 exit(1);
1077                         }
1078                         f_err = f_out;
1079                         break;
1080                 case 'm':
1081                         terse_output = 1;
1082                         break;
1083                 case 'h':
1084                         usage();
1085                         exit(0);
1086                 case 'v':
1087                         printf("%s\n", fio_version_string);
1088                         exit(0);
1089                 case FIO_GETOPT_JOB: {
1090                         const char *opt = long_options[lidx].name;
1091                         char *val = optarg;
1092
1093                         if (!strncmp(opt, "name", 4) && td) {
1094                                 ret = add_job(td, td->name ?: "fio", 0);
1095                                 if (ret) {
1096                                         put_job(td);
1097                                         return 0;
1098                                 }
1099                                 td = NULL;
1100                         }
1101                         if (!td) {
1102                                 int global = !strncmp(val, "global", 6);
1103
1104                                 td = get_new_job(global, &def_thread);
1105                                 if (!td)
1106                                         return 0;
1107                         }
1108
1109                         ret = parse_cmd_option(opt, val, options, td);
1110                         if (ret) {
1111                                 log_err("fio: job dropped\n");
1112                                 put_job(td);
1113                                 td = NULL;
1114                         }
1115                         break;
1116                 }
1117                 default:
1118                         printf("optarg <<%s>>\n", argv[optind]);
1119                         break;
1120                 }
1121         }
1122
1123         if (td) {
1124                 ret = add_job(td, td->name ?: "fio", 0);
1125                 if (ret)
1126                         put_job(td);
1127         }
1128
1129         while (optind < argc) {
1130                 ini_idx++;
1131                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1132                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1133                 optind++;
1134         }
1135
1136         return ini_idx;
1137 }
1138
1139 static void free_shm(void)
1140 {
1141         struct shmid_ds sbuf;
1142
1143         if (threads) {
1144                 shmdt((void *) threads);
1145                 threads = NULL;
1146                 shmctl(shm_id, IPC_RMID, &sbuf);
1147         }
1148 }
1149
1150 /*
1151  * The thread area is shared between the main process and the job
1152  * threads/processes. So setup a shared memory segment that will hold
1153  * all the job info.
1154  */
1155 static int setup_thread_area(void)
1156 {
1157         /*
1158          * 1024 is too much on some machines, scale max_jobs if
1159          * we get a failure that looks like too large a shm segment
1160          */
1161         do {
1162                 size_t size = max_jobs * sizeof(struct thread_data);
1163
1164                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1165                 if (shm_id != -1)
1166                         break;
1167                 if (errno != EINVAL) {
1168                         perror("shmget");
1169                         break;
1170                 }
1171
1172                 max_jobs >>= 1;
1173         } while (max_jobs);
1174
1175         if (shm_id == -1)
1176                 return 1;
1177
1178         threads = shmat(shm_id, NULL, 0);
1179         if (threads == (void *) -1) {
1180                 perror("shmat");
1181                 return 1;
1182         }
1183
1184         atexit(free_shm);
1185         return 0;
1186 }
1187
1188 /*
1189  * Copy the fio options into the long options map, so we mirror
1190  * job and cmd line options.
1191  */
1192 static void dupe_job_options(void)
1193 {
1194         struct fio_option *o;
1195         unsigned int i;
1196
1197         i = 0;
1198         while (long_options[i].name)
1199                 i++;
1200
1201         o = &options[0];
1202         while (o->name) {
1203                 long_options[i].name = o->name;
1204                 long_options[i].val = FIO_GETOPT_JOB;
1205                 if (o->type == FIO_OPT_STR_SET)
1206                         long_options[i].has_arg = no_argument;
1207                 else
1208                         long_options[i].has_arg = required_argument;
1209
1210                 i++;
1211                 o++;
1212                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1213         }
1214 }
1215
1216 int parse_options(int argc, char *argv[])
1217 {
1218         int job_files, i;
1219
1220         f_out = stdout;
1221         f_err = stderr;
1222
1223         dupe_job_options();
1224
1225         if (setup_thread_area())
1226                 return 1;
1227         if (fill_def_thread())
1228                 return 1;
1229
1230         job_files = parse_cmd_line(argc, argv);
1231
1232         for (i = 0; i < job_files; i++) {
1233                 if (fill_def_thread())
1234                         return 1;
1235                 if (parse_jobs_ini(ini_file[i], i))
1236                         return 1;
1237                 free(ini_file[i]);
1238         }
1239
1240         free(ini_file);
1241
1242         if (!thread_number) {
1243                 log_err("No jobs defined(s)\n");
1244                 return 1;
1245         }
1246
1247         return 0;
1248 }