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, "mmaphuge", 8)) {
 837 #ifdef FIO_HAVE_HUGETLB
 838                 char *hugefile;
 839
 840                 /*
 841                  * mmaphuge must be appended with the actual file
 842                  */
 843                 hugefile = strstr(mem, ":");
 844                 if (!hugefile) {
 845                         log_err("fio: mmaphuge:/path/to/file\n");
 846                         return 1;
 847                 }
 848
 849                 hugefile++;
 850                 strip_blank_front(&hugefile);
 851                 strip_blank_end(hugefile);
 852                 td->hugefile = strdup(hugefile);
 853                 td->mem_type = MEM_MMAPHUGE;
 854                 return 0;
 855 #else
 856                 log_err("fio: mmaphuge not available\n");
 857                 return 1;
 858 #endif
 859         } else if (!strncmp(mem, "mmap", 4)) {
 860                 td->mem_type = MEM_MMAP;
 861                 return 0;
 862         } else if (!strncmp(mem, "shmhuge", 7)) {
 863 #ifdef FIO_HAVE_HUGETLB
 864                 td->mem_type = MEM_SHMHUGE;
 865                 return 0;
 866 #else
 867                 log_err("fio: shmhuge not available\n");
 868                 return 1;
 869 #endif
 870         } else if (!strncmp(mem, "shm", 3)) {
 871                 td->mem_type = MEM_SHM;
 872                 return 0;
 873         }
 874
 875         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
 876         return 1;
 877 }
 878
 879 static int str_ioengine_cb(void *data, const char *str)
 880 {
 881         struct thread_data *td = data;
 882
 883         td->io_ops = load_ioengine(td, str);
 884         if (td->io_ops)
 885                 return 0;
 886
 887         log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
 888         log_err("fio: or specify path to dynamic ioengine module\n");
 889         return 1;
 890 }
 891
 892 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
 893 {
 894         mlock_size = *val;
 895         return 0;
 896 }
 897
 898 #ifdef FIO_HAVE_IOPRIO
 899 static int str_prioclass_cb(void *data, unsigned int *val)
 900 {
 901         struct thread_data *td = data;
 902
 903         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
 904         return 0;
 905 }
 906
 907 static int str_prio_cb(void *data, unsigned int *val)
 908 {
 909         struct thread_data *td = data;
 910
 911         td->ioprio |= *val;
 912         return 0;
 913 }
 914 #endif
 915
 916 static int str_exitall_cb(void)
 917 {
 918         exitall_on_terminate = 1;
 919         return 0;
 920 }
 921
 922 static int str_cpumask_cb(void *data, unsigned int *val)
 923 {
 924         struct thread_data *td = data;
 925
 926         fill_cpu_mask(td->cpumask, *val);
 927         return 0;
 928 }
 929
 930 /*
 931  * This is our [ini] type file parser.
 932  */
 933 static int parse_jobs_ini(char *file, int stonewall_flag)
 934 {
 935         unsigned int global;
 936         struct thread_data *td;
 937         char *string, *name;
 938         fpos_t off;
 939         FILE *f;
 940         char *p;
 941         int ret = 0, stonewall;
 942
 943         f = fopen(file, "r");
 944         if (!f) {
 945                 perror("fopen job file");
 946                 return 1;
 947         }
 948
 949         string = malloc(4096);
 950         name = malloc(256);
 951         memset(name, 0, 256);
 952
 953         stonewall = stonewall_flag;
 954         do {
 955                 p = fgets(string, 4095, f);
 956                 if (!p)
 957                         break;
 958                 if (is_empty_or_comment(p))
 959                         continue;
 960                 if (sscanf(p, "[%255s]", name) != 1)
 961                         continue;
 962
 963                 global = !strncmp(name, "global", 6);
 964
 965                 name[strlen(name) - 1] = '\0';
 966
 967                 td = get_new_job(global, &def_thread);
 968                 if (!td) {
 969                         ret = 1;
 970                         break;
 971                 }
 972
 973                 /*
 974                  * Seperate multiple job files by a stonewall
 975                  */
 976                 if (!global && stonewall) {
 977                         td->stonewall = stonewall;
 978                         stonewall = 0;
 979                 }
 980
 981                 fgetpos(f, &off);
 982                 while ((p = fgets(string, 4096, f)) != NULL) {
 983                         if (is_empty_or_comment(p))
 984                                 continue;
 985
 986                         strip_blank_front(&p);
 987
 988                         if (p[0] == '[')
 989                                 break;
 990
 991                         strip_blank_end(p);
 992
 993                         fgetpos(f, &off);
 994
 995                         /*
 996                          * Don't break here, continue parsing options so we
 997                          * dump all the bad ones. Makes trial/error fixups
 998                          * easier on the user.
 999                          */
1000                         ret |= parse_option(p, options, td);
1001                 }
1002
1003                 if (!ret) {
1004                         fsetpos(f, &off);
1005                         ret = add_job(td, name, 0);
1006                 } else {
1007                         log_err("fio: job %s dropped\n", name);
1008                         put_job(td);
1009                 }
1010         } while (!ret);
1011
1012         free(string);
1013         free(name);
1014         fclose(f);
1015         return ret;
1016 }
1017
1018 static int fill_def_thread(void)
1019 {
1020         memset(&def_thread, 0, sizeof(def_thread));
1021
1022         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1023                 perror("sched_getaffinity");
1024                 return 1;
1025         }
1026
1027         /*
1028          * fill globals
1029          */
1030         def_thread.ddir = DDIR_READ;
1031         def_thread.iomix = 0;
1032         def_thread.bs[DDIR_READ] = DEF_BS;
1033         def_thread.bs[DDIR_WRITE] = DEF_BS;
1034         def_thread.min_bs[DDIR_READ] = def_thread.min_bs[DDIR_WRITE] = 0;
1035         def_thread.max_bs[DDIR_READ] = def_thread.max_bs[DDIR_WRITE] = 0;
1036         def_thread.odirect = DEF_ODIRECT;
1037         def_thread.ratecycle = DEF_RATE_CYCLE;
1038         def_thread.sequential = DEF_SEQUENTIAL;
1039         def_thread.timeout = def_timeout;
1040         def_thread.overwrite = DEF_OVERWRITE;
1041         def_thread.invalidate_cache = DEF_INVALIDATE;
1042         def_thread.sync_io = DEF_SYNCIO;
1043         def_thread.mem_type = MEM_MALLOC;
1044         def_thread.bw_avg_time = DEF_BWAVGTIME;
1045         def_thread.create_serialize = DEF_CREATE_SER;
1046         def_thread.create_fsync = DEF_CREATE_FSYNC;
1047         def_thread.loops = DEF_LOOPS;
1048         def_thread.verify = DEF_VERIFY;
1049         def_thread.stonewall = DEF_STONEWALL;
1050         def_thread.numjobs = DEF_NUMJOBS;
1051         def_thread.use_thread = DEF_USE_THREAD;
1052         def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
1053         def_thread.rwmixread = DEF_RWMIX_READ;
1054         def_thread.nice = DEF_NICE;
1055         def_thread.rand_repeatable = DEF_RAND_REPEAT;
1056         def_thread.nr_files = DEF_NR_FILES;
1057         def_thread.unlink = DEF_UNLINK;
1058         def_thread.write_bw_log = write_bw_log;
1059         def_thread.write_lat_log = write_lat_log;
1060         def_thread.norandommap = DEF_NO_RAND_MAP;
1061 #ifdef FIO_HAVE_DISK_UTIL
1062         def_thread.do_disk_util = 1;
1063 #endif
1064
1065         return 0;
1066 }
1067
1068 static void usage(void)
1069 {
1070         printf("%s\n", fio_version_string);
1071         printf("\t--output\tWrite output to file\n");
1072         printf("\t--timeout\tRuntime in seconds\n");
1073         printf("\t--latency-log\tGenerate per-job latency logs\n");
1074         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1075         printf("\t--minimal\tMinimal (terse) output\n");
1076         printf("\t--version\tPrint version info and exit\n");
1077 }
1078
1079 static int parse_cmd_line(int argc, char *argv[])
1080 {
1081         struct thread_data *td = NULL;
1082         int c, ini_idx = 0, lidx, ret;
1083
1084         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1085                 switch (c) {
1086                 case 't':
1087                         def_timeout = atoi(optarg);
1088                         break;
1089                 case 'l':
1090                         write_lat_log = 1;
1091                         break;
1092                 case 'w':
1093                         write_bw_log = 1;
1094                         break;
1095                 case 'o':
1096                         f_out = fopen(optarg, "w+");
1097                         if (!f_out) {
1098                                 perror("fopen output");
1099                                 exit(1);
1100                         }
1101                         f_err = f_out;
1102                         break;
1103                 case 'm':
1104                         terse_output = 1;
1105                         break;
1106                 case 'h':
1107                         usage();
1108                         exit(0);
1109                 case 'v':
1110                         printf("%s\n", fio_version_string);
1111                         exit(0);
1112                 case FIO_GETOPT_JOB: {
1113                         const char *opt = long_options[lidx].name;
1114                         char *val = optarg;
1115
1116                         if (!strncmp(opt, "name", 4) && td) {
1117                                 ret = add_job(td, td->name ?: "fio", 0);
1118                                 if (ret) {
1119                                         put_job(td);
1120                                         return 0;
1121                                 }
1122                                 td = NULL;
1123                         }
1124                         if (!td) {
1125                                 int global = !strncmp(val, "global", 6);
1126
1127                                 td = get_new_job(global, &def_thread);
1128                                 if (!td)
1129                                         return 0;
1130                         }
1131
1132                         ret = parse_cmd_option(opt, val, options, td);
1133                         if (ret) {
1134                                 log_err("fio: job dropped\n");
1135                                 put_job(td);
1136                                 td = NULL;
1137                         }
1138                         break;
1139                 }
1140                 default:
1141                         printf("optarg <<%s>>\n", argv[optind]);
1142                         break;
1143                 }
1144         }
1145
1146         if (td) {
1147                 ret = add_job(td, td->name ?: "fio", 0);
1148                 if (ret)
1149                         put_job(td);
1150         }
1151
1152         while (optind < argc) {
1153                 ini_idx++;
1154                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1155                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1156                 optind++;
1157         }
1158
1159         return ini_idx;
1160 }
1161
1162 static void free_shm(void)
1163 {
1164         struct shmid_ds sbuf;
1165
1166         if (threads) {
1167                 shmdt((void *) threads);
1168                 threads = NULL;
1169                 shmctl(shm_id, IPC_RMID, &sbuf);
1170         }
1171 }
1172
1173 /*
1174  * The thread area is shared between the main process and the job
1175  * threads/processes. So setup a shared memory segment that will hold
1176  * all the job info.
1177  */
1178 static int setup_thread_area(void)
1179 {
1180         /*
1181          * 1024 is too much on some machines, scale max_jobs if
1182          * we get a failure that looks like too large a shm segment
1183          */
1184         do {
1185                 size_t size = max_jobs * sizeof(struct thread_data);
1186
1187                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1188                 if (shm_id != -1)
1189                         break;
1190                 if (errno != EINVAL) {
1191                         perror("shmget");
1192                         break;
1193                 }
1194
1195                 max_jobs >>= 1;
1196         } while (max_jobs);
1197
1198         if (shm_id == -1)
1199                 return 1;
1200
1201         threads = shmat(shm_id, NULL, 0);
1202         if (threads == (void *) -1) {
1203                 perror("shmat");
1204                 return 1;
1205         }
1206
1207         atexit(free_shm);
1208         return 0;
1209 }
1210
1211 /*
1212  * Copy the fio options into the long options map, so we mirror
1213  * job and cmd line options.
1214  */
1215 static void dupe_job_options(void)
1216 {
1217         struct fio_option *o;
1218         unsigned int i;
1219
1220         i = 0;
1221         while (long_options[i].name)
1222                 i++;
1223
1224         o = &options[0];
1225         while (o->name) {
1226                 long_options[i].name = o->name;
1227                 long_options[i].val = FIO_GETOPT_JOB;
1228                 if (o->type == FIO_OPT_STR_SET)
1229                         long_options[i].has_arg = no_argument;
1230                 else
1231                         long_options[i].has_arg = required_argument;
1232
1233                 i++;
1234                 o++;
1235                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1236         }
1237 }
1238
1239 int parse_options(int argc, char *argv[])
1240 {
1241         int job_files, i;
1242
1243         f_out = stdout;
1244         f_err = stderr;
1245
1246         dupe_job_options();
1247
1248         if (setup_thread_area())
1249                 return 1;
1250         if (fill_def_thread())
1251                 return 1;
1252
1253         job_files = parse_cmd_line(argc, argv);
1254
1255         for (i = 0; i < job_files; i++) {
1256                 if (fill_def_thread())
1257                         return 1;
1258                 if (parse_jobs_ini(ini_file[i], i))
1259                         return 1;
1260                 free(ini_file[i]);
1261         }
1262
1263         free(ini_file);
1264
1265         if (!thread_number) {
1266                 log_err("No jobs defined(s)\n");
1267                 return 1;
1268         }
1269
1270         return 0;
1271 }