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 #define FIO_RANDSEED            (0xb1899bedUL)
  22
  23 #define td_var_offset(var)      ((size_t) &((struct thread_data *)0)->var)
  24
  25 static int str_rw_cb(void *, const char *);
  26 static int str_ioengine_cb(void *, const char *);
  27 static int str_mem_cb(void *, const char *);
  28 static int str_verify_cb(void *, const char *);
  29 static int str_lockmem_cb(void *, unsigned long *);
  30 #ifdef FIO_HAVE_IOPRIO
  31 static int str_prio_cb(void *, unsigned int *);
  32 static int str_prioclass_cb(void *, unsigned int *);
  33 #endif
  34 static int str_exitall_cb(void);
  35 static int str_cpumask_cb(void *, unsigned int *);
  36
  37 #define __stringify_1(x)        #x
  38 #define __stringify(x)          __stringify_1(x)
  39
  40 /*
  41  * Map of job/command line options
  42  */
  43 static struct fio_option options[] = {
  44         {
  45                 .name   = "name",
  46                 .type   = FIO_OPT_STR_STORE,
  47                 .off1   = td_var_offset(name),
  48                 .help   = "Name of this job",
  49         },
  50         {
  51                 .name   = "directory",
  52                 .type   = FIO_OPT_STR_STORE,
  53                 .off1   = td_var_offset(directory),
  54                 .help   = "Directory to store files in",
  55         },
  56         {
  57                 .name   = "filename",
  58                 .type   = FIO_OPT_STR_STORE,
  59                 .off1   = td_var_offset(filename),
  60                 .help   = "Force the use of a specific file",
  61         },
  62         {
  63                 .name   = "rw",
  64                 .type   = FIO_OPT_STR,
  65                 .cb     = str_rw_cb,
  66                 .help   = "IO direction",
  67                 .def    = "read",
  68                 .posval = { "read", "write", "randwrite", "randread", "rw",
  69                                 "randrw", },
  70         },
  71         {
  72                 .name   = "ioengine",
  73                 .type   = FIO_OPT_STR,
  74                 .cb     = str_ioengine_cb,
  75                 .help   = "IO engine to use",
  76                 .def    = "sync",
  77                 .posval = { "sync", "libaio", "posixaio", "mmap", "splice",
  78                                 "sg", "null", },
  79         },
  80         {
  81                 .name   = "iodepth",
  82                 .type   = FIO_OPT_INT,
  83                 .off1   = td_var_offset(iodepth),
  84                 .help   = "Amount of IO buffers to keep in flight",
  85                 .def    = "1",
  86         },
  87         {
  88                 .name   = "size",
  89                 .type   = FIO_OPT_STR_VAL,
  90                 .off1   = td_var_offset(total_file_size),
  91                 .help   = "Size of device or file",
  92         },
  93         {
  94                 .name   = "bs",
  95                 .type   = FIO_OPT_STR_VAL_INT,
  96                 .off1   = td_var_offset(bs[DDIR_READ]),
  97                 .off2   = td_var_offset(bs[DDIR_WRITE]),
  98                 .help   = "Block size unit",
  99                 .def    = "4k",
 100         },
 101         {
 102                 .name   = "bsrange",
 103                 .type   = FIO_OPT_RANGE,
 104                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 105                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 106                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 107                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 108                 .help   = "Set block size range (in more detail than bs)",
 109         },
 110         {
 111                 .name   = "bs_unaligned",
 112                 .type   = FIO_OPT_STR_SET,
 113                 .off1   = td_var_offset(bs_unaligned),
 114                 .help   = "Don't sector align IO buffer sizes",
 115         },
 116         {
 117                 .name   = "offset",
 118                 .type   = FIO_OPT_STR_VAL,
 119                 .off1   = td_var_offset(start_offset),
 120                 .help   = "Start IO from this offset",
 121                 .def    = "0",
 122         },
 123         {
 124                 .name   = "randrepeat",
 125                 .type   = FIO_OPT_BOOL,
 126                 .off1   = td_var_offset(rand_repeatable),
 127                 .help   = "Use repeatable random IO pattern",
 128                 .def    = "1",
 129         },
 130         {
 131                 .name   = "norandommap",
 132                 .type   = FIO_OPT_STR_SET,
 133                 .off1   = td_var_offset(norandommap),
 134                 .help   = "Accept potential duplicate random blocks",
 135         },
 136         {
 137                 .name   = "nrfiles",
 138                 .type   = FIO_OPT_INT,
 139                 .off1   = td_var_offset(nr_files),
 140                 .help   = "Split job workload between this number of files",
 141                 .def    = "1",
 142         },
 143         {
 144                 .name   = "fsync",
 145                 .type   = FIO_OPT_INT,
 146                 .off1   = td_var_offset(fsync_blocks),
 147                 .help   = "Issue fsync for writes every given number of blocks",
 148                 .def    = "0",
 149         },
 150         {
 151                 .name   = "direct",
 152                 .type   = FIO_OPT_BOOL,
 153                 .off1   = td_var_offset(odirect),
 154                 .help   = "Use O_DIRECT IO",
 155                 .def    = "1",
 156         },
 157         {
 158                 .name   = "overwrite",
 159                 .type   = FIO_OPT_BOOL,
 160                 .off1   = td_var_offset(overwrite),
 161                 .help   = "When writing, set whether to overwrite current data",
 162                 .def    = "0",
 163         },
 164         {
 165                 .name   = "loops",
 166                 .type   = FIO_OPT_INT,
 167                 .off1   = td_var_offset(loops),
 168                 .help   = "Number of times to run the job",
 169                 .def    = "1",
 170         },
 171         {
 172                 .name   = "numjobs",
 173                 .type   = FIO_OPT_INT,
 174                 .off1   = td_var_offset(numjobs),
 175                 .help   = "Duplicate this job this many times",
 176                 .def    = "1",
 177         },
 178         {
 179                 .name   = "startdelay",
 180                 .type   = FIO_OPT_INT,
 181                 .off1   = td_var_offset(start_delay),
 182                 .help   = "Only start job when this period has passed",
 183                 .def    = "0",
 184         },
 185         {
 186                 .name   = "runtime",
 187                 .alias  = "timeout",
 188                 .type   = FIO_OPT_STR_VAL_TIME,
 189                 .off1   = td_var_offset(timeout),
 190                 .help   = "Stop workload when this amount of time has passed",
 191                 .def    = "0",
 192         },
 193         {
 194                 .name   = "mem",
 195                 .type   = FIO_OPT_STR,
 196                 .cb     = str_mem_cb,
 197                 .help   = "Backing type for IO buffers",
 198                 .def    = "malloc",
 199                 .posval =  { "malloc", "shm", "shmhuge", "mmap", "mmaphuge", },
 200         },
 201         {
 202                 .name   = "verify",
 203                 .type   = FIO_OPT_STR,
 204                 .cb     = str_verify_cb,
 205                 .help   = "Verify sum function",
 206                 .def    = "0",
 207                 .posval = { "crc32", "md5", },
 208         },
 209         {
 210                 .name   = "write_iolog",
 211                 .type   = FIO_OPT_STR_STORE,
 212                 .off1   = td_var_offset(write_iolog_file),
 213                 .help   = "Store IO pattern to file",
 214         },
 215         {
 216                 .name   = "read_iolog",
 217                 .type   = FIO_OPT_STR_STORE,
 218                 .off1   = td_var_offset(read_iolog_file),
 219                 .help   = "Playback IO pattern from file",
 220         },
 221         {
 222                 .name   = "exec_prerun",
 223                 .type   = FIO_OPT_STR_STORE,
 224                 .off1   = td_var_offset(exec_prerun),
 225                 .help   = "Execute this file prior to running job",
 226         },
 227         {
 228                 .name   = "exec_postrun",
 229                 .type   = FIO_OPT_STR_STORE,
 230                 .off1   = td_var_offset(exec_postrun),
 231                 .help   = "Execute this file after running job",
 232         },
 233 #ifdef FIO_HAVE_IOSCHED_SWITCH
 234         {
 235                 .name   = "ioscheduler",
 236                 .type   = FIO_OPT_STR_STORE,
 237                 .off1   = td_var_offset(ioscheduler),
 238                 .help   = "Use this IO scheduler on the backing device",
 239         },
 240 #endif
 241         {
 242                 .name   = "zonesize",
 243                 .type   = FIO_OPT_STR_VAL,
 244                 .off1   = td_var_offset(zone_size),
 245                 .help   = "Give size of an IO zone",
 246                 .def    = "0",
 247         },
 248         {
 249                 .name   = "zoneskip",
 250                 .type   = FIO_OPT_STR_VAL,
 251                 .off1   = td_var_offset(zone_skip),
 252                 .help   = "Space between IO zones",
 253                 .def    = "0",
 254         },
 255         {
 256                 .name   = "lockmem",
 257                 .type   = FIO_OPT_STR_VAL,
 258                 .cb     = str_lockmem_cb,
 259                 .help   = "Lock down this amount of memory",
 260                 .def    = "0",
 261         },
 262         {
 263                 .name   = "rwmixcycle",
 264                 .type   = FIO_OPT_INT,
 265                 .off1   = td_var_offset(rwmixcycle),
 266                 .help   = "Cycle period for mixed read/write workloads (msec)",
 267                 .def    = "500",
 268         },
 269         {
 270                 .name   = "rwmixread",
 271                 .type   = FIO_OPT_INT,
 272                 .off1   = td_var_offset(rwmixread),
 273                 .maxval = 100,
 274                 .help   = "Percentage of mixed workload that is reads",
 275                 .def    = "50",
 276         },
 277         {
 278                 .name   = "rwmixwrite",
 279                 .type   = FIO_OPT_INT,
 280                 .off1   = td_var_offset(rwmixwrite),
 281                 .maxval = 100,
 282                 .help   = "Percentage of mixed workload that is writes",
 283                 .def    = "50",
 284         },
 285         {
 286                 .name   = "nice",
 287                 .type   = FIO_OPT_INT,
 288                 .off1   = td_var_offset(nice),
 289                 .help   = "Set job CPU nice value",
 290                 .minval = -19,
 291                 .maxval = 20,
 292                 .def    = "0",
 293         },
 294 #ifdef FIO_HAVE_IOPRIO
 295         {
 296                 .name   = "prio",
 297                 .type   = FIO_OPT_INT,
 298                 .cb     = str_prio_cb,
 299                 .help   = "Set job IO priority value",
 300                 .minval = 0,
 301                 .maxval = 7,
 302         },
 303         {
 304                 .name   = "prioclass",
 305                 .type   = FIO_OPT_INT,
 306                 .cb     = str_prioclass_cb,
 307                 .help   = "Set job IO priority class",
 308                 .minval = 0,
 309                 .maxval = 3,
 310         },
 311 #endif
 312         {
 313                 .name   = "thinktime",
 314                 .type   = FIO_OPT_INT,
 315                 .off1   = td_var_offset(thinktime),
 316                 .help   = "Idle time between IO buffers",
 317                 .def    = "0",
 318         },
 319         {
 320                 .name   = "thinktime_blocks",
 321                 .type   = FIO_OPT_INT,
 322                 .off1   = td_var_offset(thinktime_blocks),
 323                 .help   = "IO buffer period between 'thinktime'",
 324                 .def    = "1",
 325         },
 326         {
 327                 .name   = "rate",
 328                 .type   = FIO_OPT_INT,
 329                 .off1   = td_var_offset(rate),
 330                 .help   = "Set bandwidth rate",
 331         },
 332         {
 333                 .name   = "ratemin",
 334                 .type   = FIO_OPT_INT,
 335                 .off1   = td_var_offset(ratemin),
 336                 .help   = "The bottom limit accepted",
 337         },
 338         {
 339                 .name   = "ratecycle",
 340                 .type   = FIO_OPT_INT,
 341                 .off1   = td_var_offset(ratecycle),
 342                 .help   = "Window average for rate limits (msec)",
 343                 .def    = "1000",
 344         },
 345         {
 346                 .name   = "invalidate",
 347                 .type   = FIO_OPT_BOOL,
 348                 .off1   = td_var_offset(invalidate_cache),
 349                 .help   = "Invalidate buffer/page cache prior to running job",
 350                 .def    = "1",
 351         },
 352         {
 353                 .name   = "sync",
 354                 .type   = FIO_OPT_BOOL,
 355                 .off1   = td_var_offset(sync_io),
 356                 .help   = "Use O_SYNC for buffered writes",
 357                 .def    = "0",
 358         },
 359         {
 360                 .name   = "bwavgtime",
 361                 .type   = FIO_OPT_INT,
 362                 .off1   = td_var_offset(bw_avg_time),
 363                 .help   = "Time window over which to calculate bandwidth (msec)",
 364                 .def    = "500",
 365         },
 366         {
 367                 .name   = "create_serialize",
 368                 .type   = FIO_OPT_BOOL,
 369                 .off1   = td_var_offset(create_serialize),
 370                 .help   = "Serialize creating of job files",
 371                 .def    = "1",
 372         },
 373         {
 374                 .name   = "create_fsync",
 375                 .type   = FIO_OPT_BOOL,
 376                 .off1   = td_var_offset(create_fsync),
 377                 .help   = "Fsync file after creation",
 378                 .def    = "1",
 379         },
 380         {
 381                 .name   = "cpuload",
 382                 .type   = FIO_OPT_INT,
 383                 .off1   = td_var_offset(cpuload),
 384                 .help   = "Use this percentage of CPU",
 385         },
 386         {
 387                 .name   = "cpuchunks",
 388                 .type   = FIO_OPT_INT,
 389                 .off1   = td_var_offset(cpucycle),
 390                 .help   = "Length of the CPU burn cycles",
 391         },
 392 #ifdef FIO_HAVE_CPU_AFFINITY
 393         {
 394                 .name   = "cpumask",
 395                 .type   = FIO_OPT_INT,
 396                 .cb     = str_cpumask_cb,
 397                 .help   = "CPU affinity mask",
 398         },
 399 #endif
 400         {
 401                 .name   = "end_fsync",
 402                 .type   = FIO_OPT_BOOL,
 403                 .off1   = td_var_offset(end_fsync),
 404                 .help   = "Include fsync at the end of job",
 405                 .def    = "0",
 406         },
 407         {
 408                 .name   = "unlink",
 409                 .type   = FIO_OPT_BOOL,
 410                 .off1   = td_var_offset(unlink),
 411                 .help   = "Unlink created files after job has completed",
 412                 .def    = "1",
 413         },
 414         {
 415                 .name   = "exitall",
 416                 .type   = FIO_OPT_STR_SET,
 417                 .cb     = str_exitall_cb,
 418                 .help   = "Terminate all jobs when one exits",
 419         },
 420         {
 421                 .name   = "stonewall",
 422                 .type   = FIO_OPT_STR_SET,
 423                 .off1   = td_var_offset(stonewall),
 424                 .help   = "Insert a hard barrier between this job and previous",
 425         },
 426         {
 427                 .name   = "thread",
 428                 .type   = FIO_OPT_STR_SET,
 429                 .off1   = td_var_offset(use_thread),
 430                 .help   = "Use threads instead of forks",
 431         },
 432         {
 433                 .name   = "write_bw_log",
 434                 .type   = FIO_OPT_STR_SET,
 435                 .off1   = td_var_offset(write_bw_log),
 436                 .help   = "Write log of bandwidth during run",
 437         },
 438         {
 439                 .name   = "write_lat_log",
 440                 .type   = FIO_OPT_STR_SET,
 441                 .off1   = td_var_offset(write_lat_log),
 442                 .help   = "Write log of latency during run",
 443         },
 444         {
 445                 .name   = "hugepage-size",
 446                 .type   = FIO_OPT_STR_VAL,
 447                 .off1   = td_var_offset(hugepage_size),
 448                 .help   = "When using hugepages, specify size of each page",
 449                 .def    = __stringify(FIO_HUGE_PAGE),
 450         },
 451         {
 452                 .name = NULL,
 453         },
 454 };
 455
 456 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 457 #define FIO_CMD_OPTS    (16)
 458 #define FIO_GETOPT_JOB  (0x89988998)
 459
 460 /*
 461  * Command line options. These will contain the above, plus a few
 462  * extra that only pertain to fio itself and not jobs.
 463  */
 464 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 465         {
 466                 .name           = "output",
 467                 .has_arg        = required_argument,
 468                 .val            = 'o',
 469         },
 470         {
 471                 .name           = "timeout",
 472                 .has_arg        = required_argument,
 473                 .val            = 't',
 474         },
 475         {
 476                 .name           = "latency-log",
 477                 .has_arg        = required_argument,
 478                 .val            = 'l',
 479         },
 480         {
 481                 .name           = "bandwidth-log",
 482                 .has_arg        = required_argument,
 483                 .val            = 'b',
 484         },
 485         {
 486                 .name           = "minimal",
 487                 .has_arg        = optional_argument,
 488                 .val            = 'm',
 489         },
 490         {
 491                 .name           = "version",
 492                 .has_arg        = no_argument,
 493                 .val            = 'v',
 494         },
 495         {
 496                 .name           = "help",
 497                 .has_arg        = no_argument,
 498                 .val            = 'h',
 499         },
 500         {
 501                 .name           = "cmdhelp",
 502                 .has_arg        = required_argument,
 503                 .val            = 'c',
 504         },
 505         {
 506                 .name           = NULL,
 507         },
 508 };
 509
 510 static int def_timeout = 0;
 511
 512 static char fio_version_string[] = "fio 1.11";
 513
 514 static char **ini_file;
 515 static int max_jobs = MAX_JOBS;
 516
 517 struct thread_data def_thread;
 518 struct thread_data *threads = NULL;
 519
 520 int exitall_on_terminate = 0;
 521 int terse_output = 0;
 522 unsigned long long mlock_size = 0;
 523 FILE *f_out = NULL;
 524 FILE *f_err = NULL;
 525
 526 static int write_lat_log = 0;
 527 static int write_bw_log = 0;
 528
 529 /*
 530  * Return a free job structure.
 531  */
 532 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 533 {
 534         struct thread_data *td;
 535
 536         if (global)
 537                 return &def_thread;
 538         if (thread_number >= max_jobs)
 539                 return NULL;
 540
 541         td = &threads[thread_number++];
 542         *td = *parent;
 543
 544         td->thread_number = thread_number;
 545         return td;
 546 }
 547
 548 static void put_job(struct thread_data *td)
 549 {
 550         if (td == &def_thread)
 551                 return;
 552
 553         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 554         thread_number--;
 555 }
 556
 557 /*
 558  * Lazy way of fixing up options that depend on each other. We could also
 559  * define option callback handlers, but this is easier.
 560  */
 561 static void fixup_options(struct thread_data *td)
 562 {
 563         if (!td->rwmixread && td->rwmixwrite)
 564                 td->rwmixread = 100 - td->rwmixwrite;
 565
 566         if (td->write_iolog_file && td->read_iolog_file) {
 567                 log_err("fio: read iolog overrides write_iolog\n");
 568                 free(td->write_iolog_file);
 569                 td->write_iolog_file = NULL;
 570         }
 571
 572         if (td->io_ops->flags & FIO_SYNCIO)
 573                 td->iodepth = 1;
 574         else {
 575                 if (!td->iodepth)
 576                         td->iodepth = td->nr_files;
 577         }
 578
 579         /*
 580          * only really works for sequential io for now, and with 1 file
 581          */
 582         if (td->zone_size && !td->sequential && td->nr_files == 1)
 583                 td->zone_size = 0;
 584
 585         /*
 586          * Reads can do overwrites, we always need to pre-create the file
 587          */
 588         if (td_read(td) || td_rw(td))
 589                 td->overwrite = 1;
 590
 591         if (!td->min_bs[DDIR_READ])
 592                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 593         if (!td->max_bs[DDIR_READ])
 594                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 595         if (!td->min_bs[DDIR_WRITE])
 596                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 597         if (!td->max_bs[DDIR_WRITE])
 598                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 599
 600         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 601
 602         if (td_read(td) && !td_rw(td))
 603                 td->verify = 0;
 604
 605         if (td->norandommap && td->verify != VERIFY_NONE) {
 606                 log_err("fio: norandommap given, verify disabled\n");
 607                 td->verify = VERIFY_NONE;
 608         }
 609         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 610                 log_err("fio: bs_unaligned may not work with raw io\n");
 611
 612         /*
 613          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 614          */
 615         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 616                 td->odirect = 0;
 617 }
 618
 619 /*
 620  * This function leaks the buffer
 621  */
 622 static char *to_kmg(unsigned int val)
 623 {
 624         char *buf = malloc(32);
 625         char post[] = { 0, 'K', 'M', 'G', 'P', 0 };
 626         char *p = post;
 627
 628         do {
 629                 if (val & 1023)
 630                         break;
 631
 632                 val >>= 10;
 633                 p++;
 634         } while (*p);
 635
 636         snprintf(buf, 31, "%u%c", val, *p);
 637         return buf;
 638 }
 639
 640 /*
 641  * Adds a job to the list of things todo. Sanitizes the various options
 642  * to make sure we don't have conflicts, and initializes various
 643  * members of td.
 644  */
 645 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 646 {
 647         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
 648                                    "rw", NULL, "randrw" };
 649         struct stat sb;
 650         int numjobs, ddir, i;
 651         struct fio_file *f;
 652
 653         /*
 654          * the def_thread is just for options, it's not a real job
 655          */
 656         if (td == &def_thread)
 657                 return 0;
 658
 659         assert(td->io_ops);
 660
 661         if (td->odirect)
 662                 td->io_ops->flags |= FIO_RAWIO;
 663
 664         td->filetype = FIO_TYPE_FILE;
 665         if (!stat(jobname, &sb)) {
 666                 if (S_ISBLK(sb.st_mode))
 667                         td->filetype = FIO_TYPE_BD;
 668                 else if (S_ISCHR(sb.st_mode))
 669                         td->filetype = FIO_TYPE_CHAR;
 670         }
 671
 672         fixup_options(td);
 673
 674         if (td->filename)
 675                 td->nr_uniq_files = 1;
 676         else
 677                 td->nr_uniq_files = td->nr_files;
 678
 679         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 680                 char tmp[PATH_MAX];
 681                 int len = 0;
 682
 683                 if (td->directory && td->directory[0] != '\0')
 684                         len = sprintf(tmp, "%s/", td->directory);
 685
 686                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 687
 688                 for_each_file(td, f, i) {
 689                         memset(f, 0, sizeof(*f));
 690                         f->fd = -1;
 691
 692                         if (td->filename)
 693                                 sprintf(tmp + len, "%s", td->filename);
 694                         else
 695                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 696                         f->file_name = strdup(tmp);
 697                 }
 698         } else {
 699                 td->nr_files = 1;
 700                 td->files = malloc(sizeof(struct fio_file));
 701                 f = &td->files[0];
 702
 703                 memset(f, 0, sizeof(*f));
 704                 f->fd = -1;
 705                 f->file_name = strdup(jobname);
 706         }
 707
 708         for_each_file(td, f, i) {
 709                 f->file_size = td->total_file_size / td->nr_files;
 710                 f->file_offset = td->start_offset;
 711         }
 712
 713         fio_sem_init(&td->mutex, 0);
 714
 715         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
 716         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
 717         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
 718
 719         if (td->stonewall && td->thread_number > 1)
 720                 groupid++;
 721
 722         td->groupid = groupid;
 723
 724         if (setup_rate(td))
 725                 goto err;
 726
 727         if (td->write_lat_log) {
 728                 setup_log(&td->slat_log);
 729                 setup_log(&td->clat_log);
 730         }
 731         if (td->write_bw_log)
 732                 setup_log(&td->bw_log);
 733
 734         if (!td->name)
 735                 td->name = strdup(jobname);
 736
 737         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 738
 739         if (!terse_output) {
 740                 if (!job_add_num) {
 741                         if (td->io_ops->flags & FIO_CPUIO)
 742                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 743                         else {
 744                                 char *c1, *c2, *c3, *c4;
 745
 746                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 747                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 748                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 749                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 750
 751                                 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);
 752
 753                                 free(c1);
 754                                 free(c2);
 755                                 free(c3);
 756                                 free(c4);
 757                         }
 758                 } else if (job_add_num == 1)
 759                         fprintf(f_out, "...\n");
 760         }
 761
 762         /*
 763          * recurse add identical jobs, clear numjobs and stonewall options
 764          * as they don't apply to sub-jobs
 765          */
 766         numjobs = td->numjobs;
 767         while (--numjobs) {
 768                 struct thread_data *td_new = get_new_job(0, td);
 769
 770                 if (!td_new)
 771                         goto err;
 772
 773                 td_new->numjobs = 1;
 774                 td_new->stonewall = 0;
 775                 job_add_num = numjobs - 1;
 776
 777                 if (add_job(td_new, jobname, job_add_num))
 778                         goto err;
 779         }
 780         return 0;
 781 err:
 782         put_job(td);
 783         return -1;
 784 }
 785
 786 /*
 787  * Initialize the various random states we need (random io, block size ranges,
 788  * read/write mix, etc).
 789  */
 790 int init_random_state(struct thread_data *td)
 791 {
 792         unsigned long seeds[4];
 793         int fd, num_maps, blocks, i;
 794         struct fio_file *f;
 795
 796         if (td->io_ops->flags & FIO_CPUIO)
 797                 return 0;
 798
 799         fd = open("/dev/urandom", O_RDONLY);
 800         if (fd == -1) {
 801                 td_verror(td, errno);
 802                 return 1;
 803         }
 804
 805         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 806                 td_verror(td, EIO);
 807                 close(fd);
 808                 return 1;
 809         }
 810
 811         close(fd);
 812
 813         os_random_seed(seeds[0], &td->bsrange_state);
 814         os_random_seed(seeds[1], &td->verify_state);
 815         os_random_seed(seeds[2], &td->rwmix_state);
 816
 817         if (td->sequential)
 818                 return 0;
 819
 820         if (td->rand_repeatable)
 821                 seeds[3] = FIO_RANDSEED;
 822
 823         if (!td->norandommap) {
 824                 for_each_file(td, f, i) {
 825                         blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 826                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 827                         f->file_map = malloc(num_maps * sizeof(long));
 828                         f->num_maps = num_maps;
 829                         memset(f->file_map, 0, num_maps * sizeof(long));
 830                 }
 831         }
 832
 833         os_random_seed(seeds[3], &td->random_state);
 834         return 0;
 835 }
 836
 837 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 838 {
 839 #ifdef FIO_HAVE_CPU_AFFINITY
 840         unsigned int i;
 841
 842         CPU_ZERO(&cpumask);
 843
 844         for (i = 0; i < sizeof(int) * 8; i++) {
 845                 if ((1 << i) & cpu)
 846                         CPU_SET(i, &cpumask);
 847         }
 848 #endif
 849 }
 850
 851 static int is_empty_or_comment(char *line)
 852 {
 853         unsigned int i;
 854
 855         for (i = 0; i < strlen(line); i++) {
 856                 if (line[i] == ';')
 857                         return 1;
 858                 if (!isspace(line[i]) && !iscntrl(line[i]))
 859                         return 0;
 860         }
 861
 862         return 1;
 863 }
 864
 865 static int str_rw_cb(void *data, const char *mem)
 866 {
 867         struct thread_data *td = data;
 868
 869         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 870                 td->ddir = DDIR_READ;
 871                 td->sequential = 1;
 872                 return 0;
 873         } else if (!strncmp(mem, "randread", 8)) {
 874                 td->ddir = DDIR_READ;
 875                 td->sequential = 0;
 876                 return 0;
 877         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 878                 td->ddir = DDIR_WRITE;
 879                 td->sequential = 1;
 880                 return 0;
 881         } else if (!strncmp(mem, "randwrite", 9)) {
 882                 td->ddir = DDIR_WRITE;
 883                 td->sequential = 0;
 884                 return 0;
 885         } else if (!strncmp(mem, "rw", 2)) {
 886                 td->ddir = DDIR_READ;
 887                 td->iomix = 1;
 888                 td->sequential = 1;
 889                 return 0;
 890         } else if (!strncmp(mem, "randrw", 6)) {
 891                 td->ddir = DDIR_READ;
 892                 td->iomix = 1;
 893                 td->sequential = 0;
 894                 return 0;
 895         }
 896
 897         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 898         return 1;
 899 }
 900
 901 static int str_verify_cb(void *data, const char *mem)
 902 {
 903         struct thread_data *td = data;
 904
 905         if (!strncmp(mem, "0", 1)) {
 906                 td->verify = VERIFY_NONE;
 907                 return 0;
 908         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 909                 td->verify = VERIFY_MD5;
 910                 return 0;
 911         } else if (!strncmp(mem, "crc32", 5)) {
 912                 td->verify = VERIFY_CRC32;
 913                 return 0;
 914         }
 915
 916         log_err("fio: verify types: md5, crc32\n");
 917         return 1;
 918 }
 919
 920 /*
 921  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
 922  */
 923 static char *get_mmap_file(const char *str)
 924 {
 925         char *p = strstr(str, ":");
 926
 927         if (!p)
 928                 return NULL;
 929
 930         p++;
 931         strip_blank_front(&p);
 932         strip_blank_end(p);
 933         return strdup(p);
 934 }
 935
 936 static int str_mem_cb(void *data, const char *mem)
 937 {
 938         struct thread_data *td = data;
 939
 940         if (!strncmp(mem, "malloc", 6)) {
 941                 td->mem_type = MEM_MALLOC;
 942                 return 0;
 943         } else if (!strncmp(mem, "mmaphuge", 8)) {
 944 #ifdef FIO_HAVE_HUGETLB
 945                 /*
 946                  * mmaphuge must be appended with the actual file
 947                  */
 948                 td->mmapfile = get_mmap_file(mem);
 949                 if (!td->mmapfile) {
 950                         log_err("fio: mmaphuge:/path/to/file\n");
 951                         return 1;
 952                 }
 953
 954                 td->mem_type = MEM_MMAPHUGE;
 955                 return 0;
 956 #else
 957                 log_err("fio: mmaphuge not available\n");
 958                 return 1;
 959 #endif
 960         } else if (!strncmp(mem, "mmap", 4)) {
 961                 /*
 962                  * Check if the user wants file backed memory. It's ok
 963                  * if there's no file given, we'll just use anon mamp then.
 964                  */
 965                 td->mmapfile = get_mmap_file(mem);
 966                 td->mem_type = MEM_MMAP;
 967                 return 0;
 968         } else if (!strncmp(mem, "shmhuge", 7)) {
 969 #ifdef FIO_HAVE_HUGETLB
 970                 td->mem_type = MEM_SHMHUGE;
 971                 return 0;
 972 #else
 973                 log_err("fio: shmhuge not available\n");
 974                 return 1;
 975 #endif
 976         } else if (!strncmp(mem, "shm", 3)) {
 977                 td->mem_type = MEM_SHM;
 978                 return 0;
 979         }
 980
 981         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
 982         return 1;
 983 }
 984
 985 static int str_ioengine_cb(void *data, const char *str)
 986 {
 987         struct thread_data *td = data;
 988
 989         td->io_ops = load_ioengine(td, str);
 990         if (td->io_ops)
 991                 return 0;
 992
 993         log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
 994         log_err("fio: or specify path to dynamic ioengine module\n");
 995         return 1;
 996 }
 997
 998 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
 999 {
1000         mlock_size = *val;
1001         return 0;
1002 }
1003
1004 #ifdef FIO_HAVE_IOPRIO
1005 static int str_prioclass_cb(void *data, unsigned int *val)
1006 {
1007         struct thread_data *td = data;
1008
1009         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1010         return 0;
1011 }
1012
1013 static int str_prio_cb(void *data, unsigned int *val)
1014 {
1015         struct thread_data *td = data;
1016
1017         td->ioprio |= *val;
1018         return 0;
1019 }
1020 #endif
1021
1022 static int str_exitall_cb(void)
1023 {
1024         exitall_on_terminate = 1;
1025         return 0;
1026 }
1027
1028 static int str_cpumask_cb(void *data, unsigned int *val)
1029 {
1030         struct thread_data *td = data;
1031
1032         fill_cpu_mask(td->cpumask, *val);
1033         return 0;
1034 }
1035
1036 /*
1037  * This is our [ini] type file parser.
1038  */
1039 static int parse_jobs_ini(char *file, int stonewall_flag)
1040 {
1041         unsigned int global;
1042         struct thread_data *td;
1043         char *string, *name;
1044         fpos_t off;
1045         FILE *f;
1046         char *p;
1047         int ret = 0, stonewall;
1048
1049         f = fopen(file, "r");
1050         if (!f) {
1051                 perror("fopen job file");
1052                 return 1;
1053         }
1054
1055         string = malloc(4096);
1056         name = malloc(256);
1057         memset(name, 0, 256);
1058
1059         stonewall = stonewall_flag;
1060         do {
1061                 p = fgets(string, 4095, f);
1062                 if (!p)
1063                         break;
1064                 if (is_empty_or_comment(p))
1065                         continue;
1066                 if (sscanf(p, "[%255s]", name) != 1)
1067                         continue;
1068
1069                 global = !strncmp(name, "global", 6);
1070
1071                 name[strlen(name) - 1] = '\0';
1072
1073                 td = get_new_job(global, &def_thread);
1074                 if (!td) {
1075                         ret = 1;
1076                         break;
1077                 }
1078
1079                 /*
1080                  * Seperate multiple job files by a stonewall
1081                  */
1082                 if (!global && stonewall) {
1083                         td->stonewall = stonewall;
1084                         stonewall = 0;
1085                 }
1086
1087                 fgetpos(f, &off);
1088                 while ((p = fgets(string, 4096, f)) != NULL) {
1089                         if (is_empty_or_comment(p))
1090                                 continue;
1091
1092                         strip_blank_front(&p);
1093
1094                         if (p[0] == '[')
1095                                 break;
1096
1097                         strip_blank_end(p);
1098
1099                         fgetpos(f, &off);
1100
1101                         /*
1102                          * Don't break here, continue parsing options so we
1103                          * dump all the bad ones. Makes trial/error fixups
1104                          * easier on the user.
1105                          */
1106                         ret |= parse_option(p, options, td);
1107                 }
1108
1109                 if (!ret) {
1110                         fsetpos(f, &off);
1111                         ret = add_job(td, name, 0);
1112                 } else {
1113                         log_err("fio: job %s dropped\n", name);
1114                         put_job(td);
1115                 }
1116         } while (!ret);
1117
1118         free(string);
1119         free(name);
1120         fclose(f);
1121         return ret;
1122 }
1123
1124 static int fill_def_thread(void)
1125 {
1126         memset(&def_thread, 0, sizeof(def_thread));
1127
1128         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1129                 perror("sched_getaffinity");
1130                 return 1;
1131         }
1132
1133         /*
1134          * fill default options
1135          */
1136         fill_default_options(&def_thread, options);
1137
1138         def_thread.timeout = def_timeout;
1139         def_thread.write_bw_log = write_bw_log;
1140         def_thread.write_lat_log = write_lat_log;
1141
1142 #ifdef FIO_HAVE_DISK_UTIL
1143         def_thread.do_disk_util = 1;
1144 #endif
1145
1146         return 0;
1147 }
1148
1149 static void usage(void)
1150 {
1151         printf("%s\n", fio_version_string);
1152         printf("\t--output\tWrite output to file\n");
1153         printf("\t--timeout\tRuntime in seconds\n");
1154         printf("\t--latency-log\tGenerate per-job latency logs\n");
1155         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1156         printf("\t--minimal\tMinimal (terse) output\n");
1157         printf("\t--version\tPrint version info and exit\n");
1158         printf("\t--help\t\tPrint this page\n");
1159         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1160 }
1161
1162 static int parse_cmd_line(int argc, char *argv[])
1163 {
1164         struct thread_data *td = NULL;
1165         int c, ini_idx = 0, lidx, ret;
1166
1167         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1168                 switch (c) {
1169                 case 't':
1170                         def_timeout = atoi(optarg);
1171                         break;
1172                 case 'l':
1173                         write_lat_log = 1;
1174                         break;
1175                 case 'w':
1176                         write_bw_log = 1;
1177                         break;
1178                 case 'o':
1179                         f_out = fopen(optarg, "w+");
1180                         if (!f_out) {
1181                                 perror("fopen output");
1182                                 exit(1);
1183                         }
1184                         f_err = f_out;
1185                         break;
1186                 case 'm':
1187                         terse_output = 1;
1188                         break;
1189                 case 'h':
1190                         usage();
1191                         exit(0);
1192                 case 'c':
1193                         ret = show_cmd_help(options, optarg);
1194                         exit(ret);
1195                 case 'v':
1196                         printf("%s\n", fio_version_string);
1197                         exit(0);
1198                 case FIO_GETOPT_JOB: {
1199                         const char *opt = long_options[lidx].name;
1200                         char *val = optarg;
1201
1202                         if (!strncmp(opt, "name", 4) && td) {
1203                                 ret = add_job(td, td->name ?: "fio", 0);
1204                                 if (ret) {
1205                                         put_job(td);
1206                                         return 0;
1207                                 }
1208                                 td = NULL;
1209                         }
1210                         if (!td) {
1211                                 int global = !strncmp(val, "global", 6);
1212
1213                                 td = get_new_job(global, &def_thread);
1214                                 if (!td)
1215                                         return 0;
1216                         }
1217
1218                         ret = parse_cmd_option(opt, val, options, td);
1219                         if (ret) {
1220                                 log_err("fio: job dropped\n");
1221                                 put_job(td);
1222                                 td = NULL;
1223                         }
1224                         break;
1225                 }
1226                 default:
1227                         break;
1228                 }
1229         }
1230
1231         if (td) {
1232                 ret = add_job(td, td->name ?: "fio", 0);
1233                 if (ret)
1234                         put_job(td);
1235         }
1236
1237         while (optind < argc) {
1238                 ini_idx++;
1239                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1240                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1241                 optind++;
1242         }
1243
1244         return ini_idx;
1245 }
1246
1247 static void free_shm(void)
1248 {
1249         struct shmid_ds sbuf;
1250
1251         if (threads) {
1252                 shmdt((void *) threads);
1253                 threads = NULL;
1254                 shmctl(shm_id, IPC_RMID, &sbuf);
1255         }
1256 }
1257
1258 /*
1259  * The thread area is shared between the main process and the job
1260  * threads/processes. So setup a shared memory segment that will hold
1261  * all the job info.
1262  */
1263 static int setup_thread_area(void)
1264 {
1265         /*
1266          * 1024 is too much on some machines, scale max_jobs if
1267          * we get a failure that looks like too large a shm segment
1268          */
1269         do {
1270                 size_t size = max_jobs * sizeof(struct thread_data);
1271
1272                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1273                 if (shm_id != -1)
1274                         break;
1275                 if (errno != EINVAL) {
1276                         perror("shmget");
1277                         break;
1278                 }
1279
1280                 max_jobs >>= 1;
1281         } while (max_jobs);
1282
1283         if (shm_id == -1)
1284                 return 1;
1285
1286         threads = shmat(shm_id, NULL, 0);
1287         if (threads == (void *) -1) {
1288                 perror("shmat");
1289                 return 1;
1290         }
1291
1292         atexit(free_shm);
1293         return 0;
1294 }
1295
1296 /*
1297  * Copy the fio options into the long options map, so we mirror
1298  * job and cmd line options.
1299  */
1300 static void dupe_job_options(void)
1301 {
1302         struct fio_option *o;
1303         unsigned int i;
1304
1305         i = 0;
1306         while (long_options[i].name)
1307                 i++;
1308
1309         o = &options[0];
1310         while (o->name) {
1311                 long_options[i].name = o->name;
1312                 long_options[i].val = FIO_GETOPT_JOB;
1313                 if (o->type == FIO_OPT_STR_SET)
1314                         long_options[i].has_arg = no_argument;
1315                 else
1316                         long_options[i].has_arg = required_argument;
1317
1318                 i++;
1319                 o++;
1320                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1321         }
1322 }
1323
1324 int parse_options(int argc, char *argv[])
1325 {
1326         int job_files, i;
1327
1328         f_out = stdout;
1329         f_err = stderr;
1330
1331         options_init(options);
1332
1333         dupe_job_options();
1334
1335         if (setup_thread_area())
1336                 return 1;
1337         if (fill_def_thread())
1338                 return 1;
1339
1340         job_files = parse_cmd_line(argc, argv);
1341
1342         for (i = 0; i < job_files; i++) {
1343                 if (fill_def_thread())
1344                         return 1;
1345                 if (parse_jobs_ini(ini_file[i], i))
1346                         return 1;
1347                 free(ini_file[i]);
1348         }
1349
1350         free(ini_file);
1351
1352         if (!thread_number) {
1353                 log_err("No jobs defined(s)\n");
1354                 return 1;
1355         }
1356
1357         return 0;
1358 }