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 #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 1.7";
 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 rate_quit = 0;
 421 int exitall_on_terminate = 0;
 422 int terse_output = 0;
 423 unsigned long long mlock_size = 0;
 424 FILE *f_out = NULL;
 425 FILE *f_err = NULL;
 426
 427 static int write_lat_log = DEF_WRITE_LAT_LOG;
 428 static int write_bw_log = DEF_WRITE_BW_LOG;
 429
 430 /*
 431  * Return a free job structure.
 432  */
 433 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 434 {
 435         struct thread_data *td;
 436
 437         if (global)
 438                 return &def_thread;
 439         if (thread_number >= max_jobs)
 440                 return NULL;
 441
 442         td = &threads[thread_number++];
 443         *td = *parent;
 444
 445         td->thread_number = thread_number;
 446         return td;
 447 }
 448
 449 static void put_job(struct thread_data *td)
 450 {
 451         if (td == &def_thread)
 452                 return;
 453
 454         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 455         thread_number--;
 456 }
 457
 458 /*
 459  * Lazy way of fixing up options that depend on each other. We could also
 460  * define option callback handlers, but this is easier.
 461  */
 462 static void fixup_options(struct thread_data *td)
 463 {
 464         if (!td->rwmixread && td->rwmixwrite)
 465                 td->rwmixread = 100 - td->rwmixwrite;
 466
 467         if (td->write_iolog_file && td->read_iolog_file) {
 468                 log_err("fio: read iolog overrides write_iolog\n");
 469                 free(td->write_iolog_file);
 470                 td->write_iolog_file = NULL;
 471         }
 472
 473         if (td->io_ops->flags & FIO_SYNCIO)
 474                 td->iodepth = 1;
 475         else {
 476                 if (!td->iodepth)
 477                         td->iodepth = td->nr_files;
 478         }
 479
 480         /*
 481          * only really works for sequential io for now, and with 1 file
 482          */
 483         if (td->zone_size && !td->sequential && td->nr_files == 1)
 484                 td->zone_size = 0;
 485
 486         /*
 487          * Reads can do overwrites, we always need to pre-create the file
 488          */
 489         if (td_read(td) || td_rw(td))
 490                 td->overwrite = 1;
 491
 492         if (!td->min_bs[DDIR_READ])
 493                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 494         if (!td->max_bs[DDIR_READ])
 495                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 496         if (!td->min_bs[DDIR_WRITE])
 497                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 498         if (!td->max_bs[DDIR_WRITE])
 499                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 500
 501         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 502
 503         if (td_read(td) && !td_rw(td))
 504                 td->verify = 0;
 505
 506         if (td->norandommap && td->verify != VERIFY_NONE) {
 507                 log_err("fio: norandommap given, verify disabled\n");
 508                 td->verify = VERIFY_NONE;
 509         }
 510         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 511                 log_err("fio: bs_unaligned may not work with raw io\n");
 512 }
 513
 514 /*
 515  * This function leaks the buffer
 516  */
 517 static char *to_kmg(unsigned int val)
 518 {
 519         char *buf = malloc(32);
 520         char post[] = { 0, 'K', 'G', 'P', -1 };
 521         char *p = post;
 522
 523         while (*p != -1) {
 524                 if (val & 1023)
 525                         break;
 526
 527                 val >>= 10;
 528                 p++;
 529         }
 530
 531         snprintf(buf, 31, "%u%c", val, *p);
 532         return buf;
 533 }
 534
 535 /*
 536  * Adds a job to the list of things todo. Sanitizes the various options
 537  * to make sure we don't have conflicts, and initializes various
 538  * members of td.
 539  */
 540 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 541 {
 542         char *ddir_str[] = { "read", "write", "randread", "randwrite",
 543                              "rw", NULL, "randrw" };
 544         struct stat sb;
 545         int numjobs, ddir, i;
 546         struct fio_file *f;
 547
 548         /*
 549          * the def_thread is just for options, it's not a real job
 550          */
 551         if (td == &def_thread)
 552                 return 0;
 553
 554         /*
 555          * Set default io engine, if none set
 556          */
 557         if (!td->io_ops) {
 558                 td->io_ops = load_ioengine(td, DEF_IO_ENGINE_NAME);
 559                 if (!td->io_ops) {
 560                         log_err("default engine %s not there?\n", DEF_IO_ENGINE_NAME);
 561                         return 1;
 562                 }
 563         }
 564
 565         if (td->odirect)
 566                 td->io_ops->flags |= FIO_RAWIO;
 567
 568         fixup_options(td);
 569
 570         td->filetype = FIO_TYPE_FILE;
 571         if (!stat(jobname, &sb)) {
 572                 if (S_ISBLK(sb.st_mode))
 573                         td->filetype = FIO_TYPE_BD;
 574                 else if (S_ISCHR(sb.st_mode))
 575                         td->filetype = FIO_TYPE_CHAR;
 576         }
 577
 578         if (td->filename)
 579                 td->nr_uniq_files = 1;
 580         else
 581                 td->nr_uniq_files = td->nr_files;
 582
 583         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 584                 char tmp[PATH_MAX];
 585                 int len = 0;
 586                 int i;
 587
 588                 if (td->directory && td->directory[0] != '\0')
 589                         sprintf(tmp, "%s/", td->directory);
 590
 591                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 592
 593                 for_each_file(td, f, i) {
 594                         memset(f, 0, sizeof(*f));
 595                         f->fd = -1;
 596
 597                         if (td->filename)
 598                                 sprintf(tmp + len, "%s", td->filename);
 599                         else
 600                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 601                         f->file_name = strdup(tmp);
 602                 }
 603         } else {
 604                 td->nr_files = 1;
 605                 td->files = malloc(sizeof(struct fio_file));
 606                 f = &td->files[0];
 607
 608                 memset(f, 0, sizeof(*f));
 609                 f->fd = -1;
 610                 f->file_name = strdup(jobname);
 611         }
 612
 613         for_each_file(td, f, i) {
 614                 f->file_size = td->total_file_size / td->nr_files;
 615                 f->file_offset = td->start_offset;
 616         }
 617
 618         fio_sem_init(&td->mutex, 0);
 619
 620         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
 621         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
 622         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
 623
 624         if (td->stonewall && td->thread_number > 1)
 625                 groupid++;
 626
 627         td->groupid = groupid;
 628
 629         if (setup_rate(td))
 630                 goto err;
 631
 632         if (td->write_lat_log) {
 633                 setup_log(&td->slat_log);
 634                 setup_log(&td->clat_log);
 635         }
 636         if (td->write_bw_log)
 637                 setup_log(&td->bw_log);
 638
 639         if (!td->name)
 640                 td->name = strdup(jobname);
 641
 642         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 643
 644         if (!terse_output) {
 645                 if (!job_add_num) {
 646                         if (td->io_ops->flags & FIO_CPUIO)
 647                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 648                         else {
 649                                 char *c1, *c2, *c3, *c4;
 650
 651                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 652                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 653                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 654                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 655
 656                                 fprintf(f_out, "%s: (g=%d): rw=%s, odir=%d, bs=%s-%s/%s-%s, rate=%d, ioengine=%s, iodepth=%d\n", td->name, td->groupid, ddir_str[ddir], td->odirect, c1, c2, c3, c4, td->rate, td->io_ops->name, td->iodepth);
 657
 658                                 free(c1);
 659                                 free(c2);
 660                                 free(c3);
 661                                 free(c4);
 662                         }
 663                 } else if (job_add_num == 1)
 664                         fprintf(f_out, "...\n");
 665         }
 666
 667         /*
 668          * recurse add identical jobs, clear numjobs and stonewall options
 669          * as they don't apply to sub-jobs
 670          */
 671         numjobs = td->numjobs;
 672         while (--numjobs) {
 673                 struct thread_data *td_new = get_new_job(0, td);
 674
 675                 if (!td_new)
 676                         goto err;
 677
 678                 td_new->numjobs = 1;
 679                 td_new->stonewall = 0;
 680                 job_add_num = numjobs - 1;
 681
 682                 if (add_job(td_new, jobname, job_add_num))
 683                         goto err;
 684         }
 685         return 0;
 686 err:
 687         put_job(td);
 688         return -1;
 689 }
 690
 691 /*
 692  * Initialize the various random states we need (random io, block size ranges,
 693  * read/write mix, etc).
 694  */
 695 int init_random_state(struct thread_data *td)
 696 {
 697         unsigned long seeds[4];
 698         int fd, num_maps, blocks, i;
 699         struct fio_file *f;
 700
 701         if (td->io_ops->flags & FIO_CPUIO)
 702                 return 0;
 703
 704         fd = open("/dev/urandom", O_RDONLY);
 705         if (fd == -1) {
 706                 td_verror(td, errno);
 707                 return 1;
 708         }
 709
 710         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 711                 td_verror(td, EIO);
 712                 close(fd);
 713                 return 1;
 714         }
 715
 716         close(fd);
 717
 718         os_random_seed(seeds[0], &td->bsrange_state);
 719         os_random_seed(seeds[1], &td->verify_state);
 720         os_random_seed(seeds[2], &td->rwmix_state);
 721
 722         if (td->sequential)
 723                 return 0;
 724
 725         if (td->rand_repeatable)
 726                 seeds[3] = DEF_RANDSEED;
 727
 728         if (!td->norandommap) {
 729                 for_each_file(td, f, i) {
 730                         blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 731                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 732                         f->file_map = malloc(num_maps * sizeof(long));
 733                         f->num_maps = num_maps;
 734                         memset(f->file_map, 0, num_maps * sizeof(long));
 735                 }
 736         }
 737
 738         os_random_seed(seeds[3], &td->random_state);
 739         return 0;
 740 }
 741
 742 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 743 {
 744 #ifdef FIO_HAVE_CPU_AFFINITY
 745         unsigned int i;
 746
 747         CPU_ZERO(&cpumask);
 748
 749         for (i = 0; i < sizeof(int) * 8; i++) {
 750                 if ((1 << i) & cpu)
 751                         CPU_SET(i, &cpumask);
 752         }
 753 #endif
 754 }
 755
 756 static int is_empty_or_comment(char *line)
 757 {
 758         unsigned int i;
 759
 760         for (i = 0; i < strlen(line); i++) {
 761                 if (line[i] == ';')
 762                         return 1;
 763                 if (!isspace(line[i]) && !iscntrl(line[i]))
 764                         return 0;
 765         }
 766
 767         return 1;
 768 }
 769
 770 static int str_rw_cb(void *data, const char *mem)
 771 {
 772         struct thread_data *td = data;
 773
 774         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 775                 td->ddir = DDIR_READ;
 776                 td->sequential = 1;
 777                 return 0;
 778         } else if (!strncmp(mem, "randread", 8)) {
 779                 td->ddir = DDIR_READ;
 780                 td->sequential = 0;
 781                 return 0;
 782         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 783                 td->ddir = DDIR_WRITE;
 784                 td->sequential = 1;
 785                 return 0;
 786         } else if (!strncmp(mem, "randwrite", 9)) {
 787                 td->ddir = DDIR_WRITE;
 788                 td->sequential = 0;
 789                 return 0;
 790         } else if (!strncmp(mem, "rw", 2)) {
 791                 td->ddir = 0;
 792                 td->iomix = 1;
 793                 td->sequential = 1;
 794                 return 0;
 795         } else if (!strncmp(mem, "randrw", 6)) {
 796                 td->ddir = 0;
 797                 td->iomix = 1;
 798                 td->sequential = 0;
 799                 return 0;
 800         }
 801
 802         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 803         return 1;
 804 }
 805
 806 static int str_verify_cb(void *data, const char *mem)
 807 {
 808         struct thread_data *td = data;
 809
 810         if (!strncmp(mem, "0", 1)) {
 811                 td->verify = VERIFY_NONE;
 812                 return 0;
 813         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 814                 td->verify = VERIFY_MD5;
 815                 return 0;
 816         } else if (!strncmp(mem, "crc32", 5)) {
 817                 td->verify = VERIFY_CRC32;
 818                 return 0;
 819         }
 820
 821         log_err("fio: verify types: md5, crc32\n");
 822         return 1;
 823 }
 824
 825 static int str_mem_cb(void *data, const char *mem)
 826 {
 827         struct thread_data *td = data;
 828
 829         if (!strncmp(mem, "malloc", 6)) {
 830                 td->mem_type = MEM_MALLOC;
 831                 return 0;
 832         } else if (!strncmp(mem, "shm", 3)) {
 833                 td->mem_type = MEM_SHM;
 834                 return 0;
 835         } else if (!strncmp(mem, "mmap", 4)) {
 836                 td->mem_type = MEM_MMAP;
 837                 return 0;
 838         }
 839
 840         log_err("fio: mem type: malloc, shm, mmap\n");
 841         return 1;
 842 }
 843
 844 static int str_ioengine_cb(void *data, const char *str)
 845 {
 846         struct thread_data *td = data;
 847
 848         td->io_ops = load_ioengine(td, str);
 849         if (td->io_ops)
 850                 return 0;
 851
 852         log_err("fio: ioengine: { linuxaio, aio, libaio }, posixaio, sync, mmap, sgio, splice, cpu\n");
 853         return 1;
 854 }
 855
 856 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
 857 {
 858         mlock_size = *val;
 859         return 0;
 860 }
 861
 862 #ifdef FIO_HAVE_IOPRIO
 863 static int str_prioclass_cb(void *data, unsigned int *val)
 864 {
 865         struct thread_data *td = data;
 866
 867         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
 868         return 0;
 869 }
 870
 871 static int str_prio_cb(void *data, unsigned int *val)
 872 {
 873         struct thread_data *td = data;
 874
 875         td->ioprio |= *val;
 876         return 0;
 877 }
 878 #endif
 879
 880 static int str_exitall_cb(void)
 881 {
 882         exitall_on_terminate = 1;
 883         return 0;
 884 }
 885
 886 static int str_cpumask_cb(void *data, unsigned int *val)
 887 {
 888         struct thread_data *td = data;
 889
 890         fill_cpu_mask(td->cpumask, *val);
 891         return 0;
 892 }
 893
 894 /*
 895  * This is our [ini] type file parser.
 896  */
 897 int parse_jobs_ini(char *file, int stonewall_flag)
 898 {
 899         unsigned int global;
 900         struct thread_data *td;
 901         char *string, *name;
 902         fpos_t off;
 903         FILE *f;
 904         char *p;
 905         int ret = 0, stonewall;
 906
 907         f = fopen(file, "r");
 908         if (!f) {
 909                 perror("fopen job file");
 910                 return 1;
 911         }
 912
 913         string = malloc(4096);
 914         name = malloc(256);
 915         memset(name, 0, 256);
 916
 917         stonewall = stonewall_flag;
 918         do {
 919                 p = fgets(string, 4095, f);
 920                 if (!p)
 921                         break;
 922                 if (is_empty_or_comment(p))
 923                         continue;
 924                 if (sscanf(p, "[%255s]", name) != 1)
 925                         continue;
 926
 927                 global = !strncmp(name, "global", 6);
 928
 929                 name[strlen(name) - 1] = '\0';
 930
 931                 td = get_new_job(global, &def_thread);
 932                 if (!td) {
 933                         ret = 1;
 934                         break;
 935                 }
 936
 937                 /*
 938                  * Seperate multiple job files by a stonewall
 939                  */
 940                 if (!global && stonewall) {
 941                         td->stonewall = stonewall;
 942                         stonewall = 0;
 943                 }
 944
 945                 fgetpos(f, &off);
 946                 while ((p = fgets(string, 4096, f)) != NULL) {
 947                         if (is_empty_or_comment(p))
 948                                 continue;
 949
 950                         strip_blank_front(&p);
 951
 952                         if (p[0] == '[')
 953                                 break;
 954
 955                         strip_blank_end(p);
 956
 957                         fgetpos(f, &off);
 958
 959                         /*
 960                          * Don't break here, continue parsing options so we
 961                          * dump all the bad ones. Makes trial/error fixups
 962                          * easier on the user.
 963                          */
 964                         ret |= parse_option(p, options, td);
 965                 }
 966
 967                 if (!ret) {
 968                         fsetpos(f, &off);
 969                         ret = add_job(td, name, 0);
 970                 } else {
 971                         log_err("fio: job %s dropped\n", name);
 972                         put_job(td);
 973                 }
 974         } while (!ret);
 975
 976         free(string);
 977         free(name);
 978         fclose(f);
 979         return ret;
 980 }
 981
 982 static int fill_def_thread(void)
 983 {
 984         memset(&def_thread, 0, sizeof(def_thread));
 985
 986         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
 987                 perror("sched_getaffinity");
 988                 return 1;
 989         }
 990
 991         /*
 992          * fill globals
 993          */
 994         def_thread.ddir = DDIR_READ;
 995         def_thread.iomix = 0;
 996         def_thread.bs[DDIR_READ] = DEF_BS;
 997         def_thread.bs[DDIR_WRITE] = DEF_BS;
 998         def_thread.min_bs[DDIR_READ] = def_thread.min_bs[DDIR_WRITE] = 0;
 999         def_thread.max_bs[DDIR_READ] = def_thread.max_bs[DDIR_WRITE] = 0;
1000         def_thread.odirect = DEF_ODIRECT;
1001         def_thread.ratecycle = DEF_RATE_CYCLE;
1002         def_thread.sequential = DEF_SEQUENTIAL;
1003         def_thread.timeout = def_timeout;
1004         def_thread.overwrite = DEF_OVERWRITE;
1005         def_thread.invalidate_cache = DEF_INVALIDATE;
1006         def_thread.sync_io = DEF_SYNCIO;
1007         def_thread.mem_type = MEM_MALLOC;
1008         def_thread.bw_avg_time = DEF_BWAVGTIME;
1009         def_thread.create_serialize = DEF_CREATE_SER;
1010         def_thread.create_fsync = DEF_CREATE_FSYNC;
1011         def_thread.loops = DEF_LOOPS;
1012         def_thread.verify = DEF_VERIFY;
1013         def_thread.stonewall = DEF_STONEWALL;
1014         def_thread.numjobs = DEF_NUMJOBS;
1015         def_thread.use_thread = DEF_USE_THREAD;
1016         def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
1017         def_thread.rwmixread = DEF_RWMIX_READ;
1018         def_thread.nice = DEF_NICE;
1019         def_thread.rand_repeatable = DEF_RAND_REPEAT;
1020         def_thread.nr_files = DEF_NR_FILES;
1021         def_thread.unlink = DEF_UNLINK;
1022         def_thread.write_bw_log = write_bw_log;
1023         def_thread.write_lat_log = write_lat_log;
1024         def_thread.norandommap = DEF_NO_RAND_MAP;
1025 #ifdef FIO_HAVE_DISK_UTIL
1026         def_thread.do_disk_util = 1;
1027 #endif
1028
1029         return 0;
1030 }
1031
1032 static void usage(void)
1033 {
1034         printf("%s\n", fio_version_string);
1035         printf("\t--output\tWrite output to file\n");
1036         printf("\t--timeout\tRuntime in seconds\n");
1037         printf("\t--latency-log\tGenerate per-job latency logs\n");
1038         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1039         printf("\t--minimal\tMinimal (terse) output\n");
1040         printf("\t--version\tPrint version info and exit\n");
1041 }
1042
1043 static int parse_cmd_line(int argc, char *argv[])
1044 {
1045         struct thread_data *td = NULL;
1046         int c, ini_idx = 0, lidx, ret;
1047
1048         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1049                 switch (c) {
1050                 case 't':
1051                         def_timeout = atoi(optarg);
1052                         break;
1053                 case 'l':
1054                         write_lat_log = 1;
1055                         break;
1056                 case 'w':
1057                         write_bw_log = 1;
1058                         break;
1059                 case 'o':
1060                         f_out = fopen(optarg, "w+");
1061                         if (!f_out) {
1062                                 perror("fopen output");
1063                                 exit(1);
1064                         }
1065                         f_err = f_out;
1066                         break;
1067                 case 'm':
1068                         terse_output = 1;
1069                         break;
1070                 case 'h':
1071                         usage();
1072                         exit(0);
1073                 case 'v':
1074                         printf("%s\n", fio_version_string);
1075                         exit(0);
1076                 case FIO_GETOPT_JOB: {
1077                         const char *opt = long_options[lidx].name;
1078                         char *val = optarg;
1079
1080                         if (!strncmp(opt, "name", 4) && td) {
1081                                 ret = add_job(td, td->name ?: "fio", 0);
1082                                 if (ret) {
1083                                         put_job(td);
1084                                         return 0;
1085                                 }
1086                                 td = NULL;
1087                         }
1088                         if (!td) {
1089                                 int global = !strncmp(val, "global", 6);
1090
1091                                 td = get_new_job(global, &def_thread);
1092                                 if (!td)
1093                                         return 0;
1094                         }
1095
1096                         ret = parse_cmd_option(opt, val, options, td);
1097                         if (ret) {
1098                                 log_err("fio: job dropped\n");
1099                                 put_job(td);
1100                                 td = NULL;
1101                         }
1102                         break;
1103                 }
1104                 default:
1105                         printf("optarg <<%s>>\n", argv[optind]);
1106                         break;
1107                 }
1108         }
1109
1110         if (td) {
1111                 ret = add_job(td, td->name ?: "fio", 0);
1112                 if (ret)
1113                         put_job(td);
1114         }
1115
1116         while (optind < argc) {
1117                 ini_idx++;
1118                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1119                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1120                 optind++;
1121         }
1122
1123         return ini_idx;
1124 }
1125
1126 static void free_shm(void)
1127 {
1128         struct shmid_ds sbuf;
1129
1130         if (threads) {
1131                 shmdt((void *) threads);
1132                 threads = NULL;
1133                 shmctl(shm_id, IPC_RMID, &sbuf);
1134         }
1135 }
1136
1137 /*
1138  * The thread area is shared between the main process and the job
1139  * threads/processes. So setup a shared memory segment that will hold
1140  * all the job info.
1141  */
1142 static int setup_thread_area(void)
1143 {
1144         /*
1145          * 1024 is too much on some machines, scale max_jobs if
1146          * we get a failure that looks like too large a shm segment
1147          */
1148         do {
1149                 size_t size = max_jobs * sizeof(struct thread_data);
1150
1151                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1152                 if (shm_id != -1)
1153                         break;
1154                 if (errno != EINVAL) {
1155                         perror("shmget");
1156                         break;
1157                 }
1158
1159                 max_jobs >>= 1;
1160         } while (max_jobs);
1161
1162         if (shm_id == -1)
1163                 return 1;
1164
1165         threads = shmat(shm_id, NULL, 0);
1166         if (threads == (void *) -1) {
1167                 perror("shmat");
1168                 return 1;
1169         }
1170
1171         atexit(free_shm);
1172         return 0;
1173 }
1174
1175 /*
1176  * Copy the fio options into the long options map, so we mirror
1177  * job and cmd line options.
1178  */
1179 static void dupe_job_options(void)
1180 {
1181         struct fio_option *o;
1182         unsigned int i;
1183
1184         i = 0;
1185         while (long_options[i].name)
1186                 i++;
1187
1188         o = &options[0];
1189         while (o->name) {
1190                 long_options[i].name = o->name;
1191                 long_options[i].val = FIO_GETOPT_JOB;
1192                 if (o->type == FIO_OPT_STR_SET)
1193                         long_options[i].has_arg = no_argument;
1194                 else
1195                         long_options[i].has_arg = required_argument;
1196
1197                 i++;
1198                 o++;
1199                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1200         }
1201 }
1202
1203 int parse_options(int argc, char *argv[])
1204 {
1205         int job_files, i;
1206
1207         f_out = stdout;
1208         f_err = stderr;
1209
1210         dupe_job_options();
1211
1212         if (setup_thread_area())
1213                 return 1;
1214         if (fill_def_thread())
1215                 return 1;
1216
1217         job_files = parse_cmd_line(argc, argv);
1218
1219         for (i = 0; i < job_files; i++) {
1220                 if (fill_def_thread())
1221                         return 1;
1222                 if (parse_jobs_ini(ini_file[i], i))
1223                         return 1;
1224                 free(ini_file[i]);
1225         }
1226
1227         free(ini_file);
1228
1229         if (!thread_number) {
1230                 log_err("No jobs defined(s)\n");
1231                 return 1;
1232         }
1233
1234         return 0;
1235 }