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_mem_cb(void *, const char *);
  26 static int str_lockmem_cb(void *, unsigned long *);
  27 #ifdef FIO_HAVE_IOPRIO
  28 static int str_prio_cb(void *, unsigned int *);
  29 static int str_prioclass_cb(void *, unsigned int *);
  30 #endif
  31 static int str_exitall_cb(void);
  32 static int str_cpumask_cb(void *, unsigned int *);
  33 static int str_fst_cb(void *, const char *);
  34 static int str_filename_cb(void *, const char *);
  35 static int str_directory_cb(void *, const char *);
  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   = "description",
  46                 .type   = FIO_OPT_STR_STORE,
  47                 .off1   = td_var_offset(description),
  48                 .help   = "Text job description",
  49         },
  50         {
  51                 .name   = "name",
  52                 .type   = FIO_OPT_STR_STORE,
  53                 .off1   = td_var_offset(name),
  54                 .help   = "Name of this job",
  55         },
  56         {
  57                 .name   = "directory",
  58                 .type   = FIO_OPT_STR_STORE,
  59                 .off1   = td_var_offset(directory),
  60                 .cb     = str_directory_cb,
  61                 .help   = "Directory to store files in",
  62         },
  63         {
  64                 .name   = "filename",
  65                 .type   = FIO_OPT_STR_STORE,
  66                 .off1   = td_var_offset(filename),
  67                 .cb     = str_filename_cb,
  68                 .help   = "File(s) to use for the workload",
  69         },
  70         {
  71                 .name   = "rw",
  72                 .type   = FIO_OPT_STR,
  73                 .off1   = td_var_offset(td_ddir),
  74                 .help   = "IO direction",
  75                 .def    = "read",
  76                 .posval = {
  77                           { .ival = "read", .oval = TD_DDIR_READ },
  78                           { .ival = "write", .oval = TD_DDIR_WRITE },
  79                           { .ival = "randread", .oval = TD_DDIR_RANDREAD },
  80                           { .ival = "randwrite", .oval = TD_DDIR_RANDWRITE },
  81                           { .ival = "rw", .oval = TD_DDIR_RW },
  82                           { .ival = "randrw", .oval = TD_DDIR_RANDRW },
  83                           },
  84         },
  85         {
  86                 .name   = "ioengine",
  87                 .type   = FIO_OPT_STR_STORE,
  88                 .off1   = td_var_offset(ioengine),
  89                 .help   = "IO engine to use",
  90                 .def    = "sync",
  91                 .posval = {
  92                           { .ival = "sync", },
  93 #ifdef FIO_HAVE_LIBAIO
  94                           { .ival = "libaio", },
  95 #endif
  96 #ifdef FIO_HAVE_POSIXAIO
  97                           { .ival = "posixaio", },
  98 #endif
  99                           { .ival = "mmap", },
 100 #ifdef FIO_HAVE_SPLICE
 101                           { .ival = "splice", },
 102 #endif
 103 #ifdef FIO_HAVE_SGIO
 104                           { .ival = "sg", },
 105 #endif
 106                           { .ival = "null", }, { .ival = "net", },
 107 #ifdef FIO_HAVE_SYSLET
 108                           { .ival = "syslet-rw", },
 109 #endif
 110                           { .ival = "cpuio", },
 111                           { .ival = "external", },
 112                           },
 113         },
 114         {
 115                 .name   = "iodepth",
 116                 .type   = FIO_OPT_INT,
 117                 .off1   = td_var_offset(iodepth),
 118                 .help   = "Amount of IO buffers to keep in flight",
 119                 .def    = "1",
 120         },
 121         {
 122                 .name   = "iodepth_batch",
 123                 .type   = FIO_OPT_INT,
 124                 .off1   = td_var_offset(iodepth_batch),
 125                 .help   = "Number of IO to submit in one go",
 126         },
 127         {
 128                 .name   = "iodepth_low",
 129                 .type   = FIO_OPT_INT,
 130                 .off1   = td_var_offset(iodepth_low),
 131                 .help   = "Low water mark for queuing depth",
 132         },
 133         {
 134                 .name   = "size",
 135                 .type   = FIO_OPT_STR_VAL,
 136                 .off1   = td_var_offset(total_file_size),
 137                 .help   = "Size of device or file",
 138         },
 139         {
 140                 .name   = "bs",
 141                 .type   = FIO_OPT_STR_VAL_INT,
 142                 .off1   = td_var_offset(bs[DDIR_READ]),
 143                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 144                 .help   = "Block size unit",
 145                 .def    = "4k",
 146         },
 147         {
 148                 .name   = "bsrange",
 149                 .type   = FIO_OPT_RANGE,
 150                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 151                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 152                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 153                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 154                 .help   = "Set block size range (in more detail than bs)",
 155         },
 156         {
 157                 .name   = "bs_unaligned",
 158                 .type   = FIO_OPT_STR_SET,
 159                 .off1   = td_var_offset(bs_unaligned),
 160                 .help   = "Don't sector align IO buffer sizes",
 161         },
 162         {
 163                 .name   = "offset",
 164                 .type   = FIO_OPT_STR_VAL,
 165                 .off1   = td_var_offset(start_offset),
 166                 .help   = "Start IO from this offset",
 167                 .def    = "0",
 168         },
 169         {
 170                 .name   = "randrepeat",
 171                 .type   = FIO_OPT_BOOL,
 172                 .off1   = td_var_offset(rand_repeatable),
 173                 .help   = "Use repeatable random IO pattern",
 174                 .def    = "1",
 175         },
 176         {
 177                 .name   = "norandommap",
 178                 .type   = FIO_OPT_STR_SET,
 179                 .off1   = td_var_offset(norandommap),
 180                 .help   = "Accept potential duplicate random blocks",
 181         },
 182         {
 183                 .name   = "nrfiles",
 184                 .type   = FIO_OPT_INT,
 185                 .off1   = td_var_offset(nr_files),
 186                 .help   = "Split job workload between this number of files",
 187                 .def    = "1",
 188         },
 189         {
 190                 .name   = "openfiles",
 191                 .type   = FIO_OPT_INT,
 192                 .off1   = td_var_offset(open_files),
 193                 .help   = "Number of files to keep open at the same time",
 194         },
 195         {
 196                 .name   = "file_service_type",
 197                 .type   = FIO_OPT_STR,
 198                 .cb     = str_fst_cb,
 199                 .off1   = td_var_offset(file_service_type),
 200                 .help   = "How to select which file to service next",
 201                 .def    = "roundrobin",
 202                 .posval = {
 203                           { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
 204                           { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
 205                           },
 206         },
 207         {
 208                 .name   = "fsync",
 209                 .type   = FIO_OPT_INT,
 210                 .off1   = td_var_offset(fsync_blocks),
 211                 .help   = "Issue fsync for writes every given number of blocks",
 212                 .def    = "0",
 213         },
 214         {
 215                 .name   = "direct",
 216                 .type   = FIO_OPT_BOOL,
 217                 .off1   = td_var_offset(odirect),
 218                 .help   = "Use O_DIRECT IO (negates buffered)",
 219                 .def    = "0",
 220         },
 221         {
 222                 .name   = "buffered",
 223                 .type   = FIO_OPT_BOOL,
 224                 .off1   = td_var_offset(odirect),
 225                 .neg    = 1,
 226                 .help   = "Use buffered IO (negates direct)",
 227                 .def    = "1",
 228         },
 229         {
 230                 .name   = "overwrite",
 231                 .type   = FIO_OPT_BOOL,
 232                 .off1   = td_var_offset(overwrite),
 233                 .help   = "When writing, set whether to overwrite current data",
 234                 .def    = "0",
 235         },
 236         {
 237                 .name   = "loops",
 238                 .type   = FIO_OPT_INT,
 239                 .off1   = td_var_offset(loops),
 240                 .help   = "Number of times to run the job",
 241                 .def    = "1",
 242         },
 243         {
 244                 .name   = "numjobs",
 245                 .type   = FIO_OPT_INT,
 246                 .off1   = td_var_offset(numjobs),
 247                 .help   = "Duplicate this job this many times",
 248                 .def    = "1",
 249         },
 250         {
 251                 .name   = "startdelay",
 252                 .type   = FIO_OPT_INT,
 253                 .off1   = td_var_offset(start_delay),
 254                 .help   = "Only start job when this period has passed",
 255                 .def    = "0",
 256         },
 257         {
 258                 .name   = "runtime",
 259                 .alias  = "timeout",
 260                 .type   = FIO_OPT_STR_VAL_TIME,
 261                 .off1   = td_var_offset(timeout),
 262                 .help   = "Stop workload when this amount of time has passed",
 263                 .def    = "0",
 264         },
 265         {
 266                 .name   = "mem",
 267                 .type   = FIO_OPT_STR,
 268                 .cb     = str_mem_cb,
 269                 .off1   = td_var_offset(mem_type),
 270                 .help   = "Backing type for IO buffers",
 271                 .def    = "malloc",
 272                 .posval = {
 273                           { .ival = "malloc", .oval = MEM_MALLOC },
 274                           { .ival = "shm", .oval = MEM_SHM },
 275 #ifdef FIO_HAVE_HUGETLB
 276                           { .ival = "shmhuge", .oval = MEM_SHMHUGE },
 277 #endif
 278                           { .ival = "mmap", .oval = MEM_MMAP },
 279 #ifdef FIO_HAVE_HUGETLB
 280                           { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
 281 #endif
 282                           },
 283         },
 284         {
 285                 .name   = "verify",
 286                 .type   = FIO_OPT_STR,
 287                 .off1   = td_var_offset(verify),
 288                 .help   = "Verify sum function",
 289                 .def    = "0",
 290                 .posval = {
 291                           { .ival = "0", .oval = VERIFY_NONE },
 292                           { .ival = "crc32", .oval = VERIFY_CRC32 },
 293                           { .ival = "md5", .oval = VERIFY_MD5 },
 294                           },
 295         },
 296         {
 297                 .name   = "write_iolog",
 298                 .type   = FIO_OPT_STR_STORE,
 299                 .off1   = td_var_offset(write_iolog_file),
 300                 .help   = "Store IO pattern to file",
 301         },
 302         {
 303                 .name   = "read_iolog",
 304                 .type   = FIO_OPT_STR_STORE,
 305                 .off1   = td_var_offset(read_iolog_file),
 306                 .help   = "Playback IO pattern from file",
 307         },
 308         {
 309                 .name   = "exec_prerun",
 310                 .type   = FIO_OPT_STR_STORE,
 311                 .off1   = td_var_offset(exec_prerun),
 312                 .help   = "Execute this file prior to running job",
 313         },
 314         {
 315                 .name   = "exec_postrun",
 316                 .type   = FIO_OPT_STR_STORE,
 317                 .off1   = td_var_offset(exec_postrun),
 318                 .help   = "Execute this file after running job",
 319         },
 320 #ifdef FIO_HAVE_IOSCHED_SWITCH
 321         {
 322                 .name   = "ioscheduler",
 323                 .type   = FIO_OPT_STR_STORE,
 324                 .off1   = td_var_offset(ioscheduler),
 325                 .help   = "Use this IO scheduler on the backing device",
 326         },
 327 #endif
 328         {
 329                 .name   = "zonesize",
 330                 .type   = FIO_OPT_STR_VAL,
 331                 .off1   = td_var_offset(zone_size),
 332                 .help   = "Give size of an IO zone",
 333                 .def    = "0",
 334         },
 335         {
 336                 .name   = "zoneskip",
 337                 .type   = FIO_OPT_STR_VAL,
 338                 .off1   = td_var_offset(zone_skip),
 339                 .help   = "Space between IO zones",
 340                 .def    = "0",
 341         },
 342         {
 343                 .name   = "lockmem",
 344                 .type   = FIO_OPT_STR_VAL,
 345                 .cb     = str_lockmem_cb,
 346                 .help   = "Lock down this amount of memory",
 347                 .def    = "0",
 348         },
 349         {
 350                 .name   = "rwmixcycle",
 351                 .type   = FIO_OPT_INT,
 352                 .off1   = td_var_offset(rwmixcycle),
 353                 .help   = "Cycle period for mixed read/write workloads (msec)",
 354                 .def    = "500",
 355         },
 356         {
 357                 .name   = "rwmixread",
 358                 .type   = FIO_OPT_INT,
 359                 .off1   = td_var_offset(rwmixread),
 360                 .maxval = 100,
 361                 .help   = "Percentage of mixed workload that is reads",
 362                 .def    = "50",
 363         },
 364         {
 365                 .name   = "rwmixwrite",
 366                 .type   = FIO_OPT_INT,
 367                 .off1   = td_var_offset(rwmixwrite),
 368                 .maxval = 100,
 369                 .help   = "Percentage of mixed workload that is writes",
 370                 .def    = "50",
 371         },
 372         {
 373                 .name   = "nice",
 374                 .type   = FIO_OPT_INT,
 375                 .off1   = td_var_offset(nice),
 376                 .help   = "Set job CPU nice value",
 377                 .minval = -19,
 378                 .maxval = 20,
 379                 .def    = "0",
 380         },
 381 #ifdef FIO_HAVE_IOPRIO
 382         {
 383                 .name   = "prio",
 384                 .type   = FIO_OPT_INT,
 385                 .cb     = str_prio_cb,
 386                 .help   = "Set job IO priority value",
 387                 .minval = 0,
 388                 .maxval = 7,
 389         },
 390         {
 391                 .name   = "prioclass",
 392                 .type   = FIO_OPT_INT,
 393                 .cb     = str_prioclass_cb,
 394                 .help   = "Set job IO priority class",
 395                 .minval = 0,
 396                 .maxval = 3,
 397         },
 398 #endif
 399         {
 400                 .name   = "thinktime",
 401                 .type   = FIO_OPT_INT,
 402                 .off1   = td_var_offset(thinktime),
 403                 .help   = "Idle time between IO buffers (usec)",
 404                 .def    = "0",
 405         },
 406         {
 407                 .name   = "thinktime_spin",
 408                 .type   = FIO_OPT_INT,
 409                 .off1   = td_var_offset(thinktime_spin),
 410                 .help   = "Start think time by spinning this amount (usec)",
 411                 .def    = "0",
 412         },
 413         {
 414                 .name   = "thinktime_blocks",
 415                 .type   = FIO_OPT_INT,
 416                 .off1   = td_var_offset(thinktime_blocks),
 417                 .help   = "IO buffer period between 'thinktime'",
 418                 .def    = "1",
 419         },
 420         {
 421                 .name   = "rate",
 422                 .type   = FIO_OPT_INT,
 423                 .off1   = td_var_offset(rate),
 424                 .help   = "Set bandwidth rate",
 425         },
 426         {
 427                 .name   = "ratemin",
 428                 .type   = FIO_OPT_INT,
 429                 .off1   = td_var_offset(ratemin),
 430                 .help   = "The bottom limit accepted",
 431         },
 432         {
 433                 .name   = "ratecycle",
 434                 .type   = FIO_OPT_INT,
 435                 .off1   = td_var_offset(ratecycle),
 436                 .help   = "Window average for rate limits (msec)",
 437                 .def    = "1000",
 438         },
 439         {
 440                 .name   = "invalidate",
 441                 .type   = FIO_OPT_BOOL,
 442                 .off1   = td_var_offset(invalidate_cache),
 443                 .help   = "Invalidate buffer/page cache prior to running job",
 444                 .def    = "1",
 445         },
 446         {
 447                 .name   = "sync",
 448                 .type   = FIO_OPT_BOOL,
 449                 .off1   = td_var_offset(sync_io),
 450                 .help   = "Use O_SYNC for buffered writes",
 451                 .def    = "0",
 452         },
 453         {
 454                 .name   = "bwavgtime",
 455                 .type   = FIO_OPT_INT,
 456                 .off1   = td_var_offset(bw_avg_time),
 457                 .help   = "Time window over which to calculate bandwidth (msec)",
 458                 .def    = "500",
 459         },
 460         {
 461                 .name   = "create_serialize",
 462                 .type   = FIO_OPT_BOOL,
 463                 .off1   = td_var_offset(create_serialize),
 464                 .help   = "Serialize creating of job files",
 465                 .def    = "1",
 466         },
 467         {
 468                 .name   = "create_fsync",
 469                 .type   = FIO_OPT_BOOL,
 470                 .off1   = td_var_offset(create_fsync),
 471                 .help   = "Fsync file after creation",
 472                 .def    = "1",
 473         },
 474         {
 475                 .name   = "cpuload",
 476                 .type   = FIO_OPT_INT,
 477                 .off1   = td_var_offset(cpuload),
 478                 .help   = "Use this percentage of CPU",
 479         },
 480         {
 481                 .name   = "cpuchunks",
 482                 .type   = FIO_OPT_INT,
 483                 .off1   = td_var_offset(cpucycle),
 484                 .help   = "Length of the CPU burn cycles (usecs)",
 485                 .def    = "50000",
 486         },
 487 #ifdef FIO_HAVE_CPU_AFFINITY
 488         {
 489                 .name   = "cpumask",
 490                 .type   = FIO_OPT_INT,
 491                 .cb     = str_cpumask_cb,
 492                 .help   = "CPU affinity mask",
 493         },
 494 #endif
 495         {
 496                 .name   = "end_fsync",
 497                 .type   = FIO_OPT_BOOL,
 498                 .off1   = td_var_offset(end_fsync),
 499                 .help   = "Include fsync at the end of job",
 500                 .def    = "0",
 501         },
 502         {
 503                 .name   = "fsync_on_close",
 504                 .type   = FIO_OPT_BOOL,
 505                 .off1   = td_var_offset(fsync_on_close),
 506                 .help   = "fsync files on close",
 507                 .def    = "0",
 508         },
 509         {
 510                 .name   = "unlink",
 511                 .type   = FIO_OPT_BOOL,
 512                 .off1   = td_var_offset(unlink),
 513                 .help   = "Unlink created files after job has completed",
 514                 .def    = "0",
 515         },
 516         {
 517                 .name   = "exitall",
 518                 .type   = FIO_OPT_STR_SET,
 519                 .cb     = str_exitall_cb,
 520                 .help   = "Terminate all jobs when one exits",
 521         },
 522         {
 523                 .name   = "stonewall",
 524                 .type   = FIO_OPT_STR_SET,
 525                 .off1   = td_var_offset(stonewall),
 526                 .help   = "Insert a hard barrier between this job and previous",
 527         },
 528         {
 529                 .name   = "thread",
 530                 .type   = FIO_OPT_STR_SET,
 531                 .off1   = td_var_offset(use_thread),
 532                 .help   = "Use threads instead of forks",
 533         },
 534         {
 535                 .name   = "write_bw_log",
 536                 .type   = FIO_OPT_STR_SET,
 537                 .off1   = td_var_offset(write_bw_log),
 538                 .help   = "Write log of bandwidth during run",
 539         },
 540         {
 541                 .name   = "write_lat_log",
 542                 .type   = FIO_OPT_STR_SET,
 543                 .off1   = td_var_offset(write_lat_log),
 544                 .help   = "Write log of latency during run",
 545         },
 546         {
 547                 .name   = "hugepage-size",
 548                 .type   = FIO_OPT_STR_VAL,
 549                 .off1   = td_var_offset(hugepage_size),
 550                 .help   = "When using hugepages, specify size of each page",
 551                 .def    = __stringify(FIO_HUGE_PAGE),
 552         },
 553         {
 554                 .name   = "group_reporting",
 555                 .type   = FIO_OPT_STR_SET,
 556                 .off1   = td_var_offset(group_reporting),
 557                 .help   = "Do reporting on a per-group basis",
 558         },
 559         {
 560                 .name = NULL,
 561         },
 562 };
 563
 564 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 565 #define FIO_CMD_OPTS    (16)
 566 #define FIO_GETOPT_JOB  (0x89988998)
 567
 568 /*
 569  * Command line options. These will contain the above, plus a few
 570  * extra that only pertain to fio itself and not jobs.
 571  */
 572 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 573         {
 574                 .name           = "output",
 575                 .has_arg        = required_argument,
 576                 .val            = 'o',
 577         },
 578         {
 579                 .name           = "timeout",
 580                 .has_arg        = required_argument,
 581                 .val            = 't',
 582         },
 583         {
 584                 .name           = "latency-log",
 585                 .has_arg        = required_argument,
 586                 .val            = 'l',
 587         },
 588         {
 589                 .name           = "bandwidth-log",
 590                 .has_arg        = required_argument,
 591                 .val            = 'b',
 592         },
 593         {
 594                 .name           = "minimal",
 595                 .has_arg        = optional_argument,
 596                 .val            = 'm',
 597         },
 598         {
 599                 .name           = "version",
 600                 .has_arg        = no_argument,
 601                 .val            = 'v',
 602         },
 603         {
 604                 .name           = "help",
 605                 .has_arg        = no_argument,
 606                 .val            = 'h',
 607         },
 608         {
 609                 .name           = "cmdhelp",
 610                 .has_arg        = optional_argument,
 611                 .val            = 'c',
 612         },
 613         {
 614                 .name           = NULL,
 615         },
 616 };
 617
 618 static int def_timeout = 0;
 619
 620 static char fio_version_string[] = "fio 1.13";
 621
 622 static char **ini_file;
 623 static int max_jobs = MAX_JOBS;
 624
 625 struct thread_data def_thread;
 626 struct thread_data *threads = NULL;
 627
 628 int exitall_on_terminate = 0;
 629 int terse_output = 0;
 630 unsigned long long mlock_size = 0;
 631 FILE *f_out = NULL;
 632 FILE *f_err = NULL;
 633
 634 static int write_lat_log = 0;
 635 int write_bw_log = 0;
 636
 637 static int prev_group_jobs;
 638
 639 FILE *get_f_out()
 640 {
 641         return f_out;
 642 }
 643
 644 FILE *get_f_err()
 645 {
 646         return f_err;
 647 }
 648
 649 /*
 650  * Return a free job structure.
 651  */
 652 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 653 {
 654         struct thread_data *td;
 655
 656         if (global)
 657                 return &def_thread;
 658         if (thread_number >= max_jobs)
 659                 return NULL;
 660
 661         td = &threads[thread_number++];
 662         *td = *parent;
 663
 664         td->thread_number = thread_number;
 665         return td;
 666 }
 667
 668 static void put_job(struct thread_data *td)
 669 {
 670         if (td == &def_thread)
 671                 return;
 672
 673         if (td->error)
 674                 fprintf(f_out, "fio: %s\n", td->verror);
 675
 676         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 677         thread_number--;
 678 }
 679
 680 /*
 681  * Lazy way of fixing up options that depend on each other. We could also
 682  * define option callback handlers, but this is easier.
 683  */
 684 static void fixup_options(struct thread_data *td)
 685 {
 686         if (!td->rwmixread && td->rwmixwrite)
 687                 td->rwmixread = 100 - td->rwmixwrite;
 688
 689         if (td->write_iolog_file && td->read_iolog_file) {
 690                 log_err("fio: read iolog overrides write_iolog\n");
 691                 free(td->write_iolog_file);
 692                 td->write_iolog_file = NULL;
 693         }
 694
 695         if (td->io_ops->flags & FIO_SYNCIO)
 696                 td->iodepth = 1;
 697         else {
 698                 if (!td->iodepth)
 699                         td->iodepth = td->open_files;
 700         }
 701
 702         /*
 703          * only really works for sequential io for now, and with 1 file
 704          */
 705         if (td->zone_size && td_random(td) && td->open_files == 1)
 706                 td->zone_size = 0;
 707
 708         /*
 709          * Reads can do overwrites, we always need to pre-create the file
 710          */
 711         if (td_read(td) || td_rw(td))
 712                 td->overwrite = 1;
 713
 714         if (!td->min_bs[DDIR_READ])
 715                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 716         if (!td->max_bs[DDIR_READ])
 717                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 718         if (!td->min_bs[DDIR_WRITE])
 719                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 720         if (!td->max_bs[DDIR_WRITE])
 721                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 722
 723         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 724
 725         if (td_read(td) && !td_rw(td))
 726                 td->verify = 0;
 727
 728         if (td->norandommap && td->verify != VERIFY_NONE) {
 729                 log_err("fio: norandommap given, verify disabled\n");
 730                 td->verify = VERIFY_NONE;
 731         }
 732         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 733                 log_err("fio: bs_unaligned may not work with raw io\n");
 734
 735         /*
 736          * thinktime_spin must be less than thinktime
 737          */
 738         if (td->thinktime_spin > td->thinktime)
 739                 td->thinktime_spin = td->thinktime;
 740
 741         /*
 742          * The low water mark cannot be bigger than the iodepth
 743          */
 744         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
 745                 /*
 746                  * syslet work around - if the workload is sequential,
 747                  * we want to let the queue drain all the way down to
 748                  * avoid seeking between async threads
 749                  */
 750                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 751                         td->iodepth_low = 1;
 752                 else
 753                         td->iodepth_low = td->iodepth;
 754         }
 755
 756         /*
 757          * If batch number isn't set, default to the same as iodepth
 758          */
 759         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
 760                 td->iodepth_batch = td->iodepth;
 761
 762         if (!td->nr_files)
 763                 td->nr_files = td->files_index;
 764         else if (td->nr_files > td->files_index)
 765                 td->nr_files = td->files_index;
 766
 767         if (td->open_files > td->nr_files || !td->open_files)
 768                 td->open_files = td->nr_files;
 769 }
 770
 771 /*
 772  * This function leaks the buffer
 773  */
 774 static char *to_kmg(unsigned int val)
 775 {
 776         char *buf = malloc(32);
 777         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 778         char *p = post;
 779
 780         do {
 781                 if (val & 1023)
 782                         break;
 783
 784                 val >>= 10;
 785                 p++;
 786         } while (*p);
 787
 788         snprintf(buf, 31, "%u%c", val, *p);
 789         return buf;
 790 }
 791
 792 /* External engines are specified by "external:name.o") */
 793 static const char *get_engine_name(const char *str)
 794 {
 795         char *p = strstr(str, ":");
 796
 797         if (!p)
 798                 return str;
 799
 800         p++;
 801         strip_blank_front(&p);
 802         strip_blank_end(p);
 803         return p;
 804 }
 805
 806 /*
 807  * Adds a job to the list of things todo. Sanitizes the various options
 808  * to make sure we don't have conflicts, and initializes various
 809  * members of td.
 810  */
 811 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 812 {
 813         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 814                                    "randread", "randwrite", "randrw" };
 815         unsigned int i;
 816         struct fio_file *f;
 817         const char *engine;
 818         char fname[PATH_MAX];
 819         int numjobs;
 820
 821         /*
 822          * the def_thread is just for options, it's not a real job
 823          */
 824         if (td == &def_thread)
 825                 return 0;
 826
 827         engine = get_engine_name(td->ioengine);
 828         td->io_ops = load_ioengine(td, engine);
 829         if (!td->io_ops) {
 830                 log_err("fio: failed to load engine %s\n", engine);
 831                 return 1;
 832         }
 833
 834         if (td->use_thread)
 835                 nr_thread++;
 836         else
 837                 nr_process++;
 838
 839         if (td->odirect)
 840                 td->io_ops->flags |= FIO_RAWIO;
 841
 842         if (!td->filename) {
 843                 td->filename = strdup(jobname);
 844
 845                 if (td->nr_files == 1)
 846                         add_file(td, td->filename);
 847                 else {
 848                         for (i = 0; i < td->nr_files; i++) {
 849                                 sprintf(fname, "%s.%d.%d", td->filename, td->thread_number, i);
 850                                 add_file(td, fname);
 851                         }
 852                 }
 853         }
 854
 855         fixup_options(td);
 856
 857         for_each_file(td, f, i) {
 858                 if (td->directory && f->filetype == FIO_TYPE_FILE) {
 859                         sprintf(fname, "%s/%s", td->directory, f->file_name);
 860                         f->file_name = strdup(fname);
 861                 }
 862         }
 863
 864         td->mutex = fio_sem_init(0);
 865
 866         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 867         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 868         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 869
 870         if ((td->stonewall || td->numjobs > 1) && prev_group_jobs) {
 871                 prev_group_jobs = 0;
 872                 groupid++;
 873         }
 874
 875         td->groupid = groupid;
 876         prev_group_jobs++;
 877
 878         if (setup_rate(td))
 879                 goto err;
 880
 881         if (td->write_lat_log) {
 882                 setup_log(&td->ts.slat_log);
 883                 setup_log(&td->ts.clat_log);
 884         }
 885         if (td->write_bw_log)
 886                 setup_log(&td->ts.bw_log);
 887
 888         if (!td->name)
 889                 td->name = strdup(jobname);
 890
 891         if (!terse_output) {
 892                 if (!job_add_num) {
 893                         if (!strcmp(td->io_ops->name, "cpuio"))
 894                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 895                         else {
 896                                 char *c1, *c2, *c3, *c4;
 897
 898                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 899                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 900                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 901                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 902
 903                                 fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[td->td_ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
 904
 905                                 free(c1);
 906                                 free(c2);
 907                                 free(c3);
 908                                 free(c4);
 909                         }
 910                 } else if (job_add_num == 1)
 911                         fprintf(f_out, "...\n");
 912         }
 913
 914         /*
 915          * recurse add identical jobs, clear numjobs and stonewall options
 916          * as they don't apply to sub-jobs
 917          */
 918         numjobs = td->numjobs;
 919         while (--numjobs) {
 920                 struct thread_data *td_new = get_new_job(0, td);
 921
 922                 if (!td_new)
 923                         goto err;
 924
 925                 td_new->numjobs = 1;
 926                 td_new->stonewall = 0;
 927                 job_add_num = numjobs - 1;
 928
 929                 if (add_job(td_new, jobname, job_add_num))
 930                         goto err;
 931         }
 932
 933         if (td->numjobs > 1) {
 934                 groupid++;
 935                 prev_group_jobs = 0;
 936         }
 937
 938         return 0;
 939 err:
 940         put_job(td);
 941         return -1;
 942 }
 943
 944 /*
 945  * Initialize the various random states we need (random io, block size ranges,
 946  * read/write mix, etc).
 947  */
 948 int init_random_state(struct thread_data *td)
 949 {
 950         unsigned long seeds[5];
 951         int fd, num_maps, blocks;
 952         struct fio_file *f;
 953         unsigned int i;
 954
 955         if (td->io_ops->flags & FIO_DISKLESSIO)
 956                 return 0;
 957
 958         fd = open("/dev/urandom", O_RDONLY);
 959         if (fd == -1) {
 960                 td_verror(td, errno, "open");
 961                 return 1;
 962         }
 963
 964         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 965                 td_verror(td, EIO, "read");
 966                 close(fd);
 967                 return 1;
 968         }
 969
 970         close(fd);
 971
 972         os_random_seed(seeds[0], &td->bsrange_state);
 973         os_random_seed(seeds[1], &td->verify_state);
 974         os_random_seed(seeds[2], &td->rwmix_state);
 975
 976         if (td->file_service_type == FIO_FSERVICE_RANDOM)
 977                 os_random_seed(seeds[3], &td->next_file_state);
 978
 979         if (!td_random(td))
 980                 return 0;
 981
 982         if (td->rand_repeatable)
 983                 seeds[4] = FIO_RANDSEED * td->thread_number;
 984
 985         if (!td->norandommap) {
 986                 for_each_file(td, f, i) {
 987                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 988                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 989                         f->file_map = malloc(num_maps * sizeof(long));
 990                         if (!f->file_map) {
 991                                 log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
 992                                 return 1;
 993                         }
 994                         f->num_maps = num_maps;
 995                         memset(f->file_map, 0, num_maps * sizeof(long));
 996                 }
 997         }
 998
 999         os_random_seed(seeds[4], &td->random_state);
1000         return 0;
1001 }
1002
1003 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
1004 {
1005 #ifdef FIO_HAVE_CPU_AFFINITY
1006         unsigned int i;
1007
1008         CPU_ZERO(&cpumask);
1009
1010         for (i = 0; i < sizeof(int) * 8; i++) {
1011                 if ((1 << i) & cpu)
1012                         CPU_SET(i, &cpumask);
1013         }
1014 #endif
1015 }
1016
1017 static int is_empty_or_comment(char *line)
1018 {
1019         unsigned int i;
1020
1021         for (i = 0; i < strlen(line); i++) {
1022                 if (line[i] == ';')
1023                         return 1;
1024                 if (line[i] == '#')
1025                         return 1;
1026                 if (!isspace(line[i]) && !iscntrl(line[i]))
1027                         return 0;
1028         }
1029
1030         return 1;
1031 }
1032
1033 /*
1034  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
1035  */
1036 static char *get_opt_postfix(const char *str)
1037 {
1038         char *p = strstr(str, ":");
1039
1040         if (!p)
1041                 return NULL;
1042
1043         p++;
1044         strip_blank_front(&p);
1045         strip_blank_end(p);
1046         return strdup(p);
1047 }
1048
1049 static int str_mem_cb(void *data, const char *mem)
1050 {
1051         struct thread_data *td = data;
1052
1053         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1054                 td->mmapfile = get_opt_postfix(mem);
1055                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1056                         log_err("fio: mmaphuge:/path/to/file\n");
1057                         return 1;
1058                 }
1059         }
1060
1061         return 0;
1062 }
1063
1064 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1065 {
1066         mlock_size = *val;
1067         return 0;
1068 }
1069
1070 #ifdef FIO_HAVE_IOPRIO
1071 static int str_prioclass_cb(void *data, unsigned int *val)
1072 {
1073         struct thread_data *td = data;
1074
1075         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1076         return 0;
1077 }
1078
1079 static int str_prio_cb(void *data, unsigned int *val)
1080 {
1081         struct thread_data *td = data;
1082
1083         td->ioprio |= *val;
1084         return 0;
1085 }
1086 #endif
1087
1088 static int str_exitall_cb(void)
1089 {
1090         exitall_on_terminate = 1;
1091         return 0;
1092 }
1093
1094 static int str_cpumask_cb(void *data, unsigned int *val)
1095 {
1096         struct thread_data *td = data;
1097
1098         fill_cpu_mask(td->cpumask, *val);
1099         return 0;
1100 }
1101
1102 static int str_fst_cb(void *data, const char *str)
1103 {
1104         struct thread_data *td = data;
1105         char *nr = get_opt_postfix(str);
1106
1107         td->file_service_nr = 1;
1108         if (nr)
1109                 td->file_service_nr = atoi(nr);
1110
1111         return 0;
1112 }
1113
1114 static int str_filename_cb(void *data, const char *input)
1115 {
1116         struct thread_data *td = data;
1117         char *fname, *str, *p;
1118
1119         p = str = strdup(input);
1120
1121         strip_blank_front(&str);
1122         strip_blank_end(str);
1123
1124         while ((fname = strsep(&str, ":")) != NULL) {
1125                 if (!strlen(fname))
1126                         break;
1127                 add_file(td, fname);
1128         }
1129
1130         free(p);
1131         return 0;
1132 }
1133
1134 static int str_directory_cb(void *data, const char fio_unused *str)
1135 {
1136         struct thread_data *td = data;
1137         struct stat sb;
1138
1139         if (lstat(td->directory, &sb) < 0) {
1140                 log_err("fio: %s is not a directory\n", td->directory);
1141                 td_verror(td, errno, "lstat");
1142                 return 1;
1143         }
1144         if (!S_ISDIR(sb.st_mode)) {
1145                 log_err("fio: %s is not a directory\n", td->directory);
1146                 return 1;
1147         }
1148
1149         return 0;
1150 }
1151
1152 /*
1153  * This is our [ini] type file parser.
1154  */
1155 static int parse_jobs_ini(char *file, int stonewall_flag)
1156 {
1157         unsigned int global;
1158         struct thread_data *td;
1159         char *string, *name;
1160         fpos_t off;
1161         FILE *f;
1162         char *p;
1163         int ret = 0, stonewall;
1164
1165         f = fopen(file, "r");
1166         if (!f) {
1167                 perror("fopen job file");
1168                 return 1;
1169         }
1170
1171         string = malloc(4096);
1172         name = malloc(256);
1173         memset(name, 0, 256);
1174
1175         stonewall = stonewall_flag;
1176         do {
1177                 p = fgets(string, 4095, f);
1178                 if (!p)
1179                         break;
1180                 if (is_empty_or_comment(p))
1181                         continue;
1182                 if (sscanf(p, "[%255s]", name) != 1)
1183                         continue;
1184
1185                 global = !strncmp(name, "global", 6);
1186
1187                 name[strlen(name) - 1] = '\0';
1188
1189                 td = get_new_job(global, &def_thread);
1190                 if (!td) {
1191                         ret = 1;
1192                         break;
1193                 }
1194
1195                 /*
1196                  * Seperate multiple job files by a stonewall
1197                  */
1198                 if (!global && stonewall) {
1199                         td->stonewall = stonewall;
1200                         stonewall = 0;
1201                 }
1202
1203                 fgetpos(f, &off);
1204                 while ((p = fgets(string, 4096, f)) != NULL) {
1205                         if (is_empty_or_comment(p))
1206                                 continue;
1207
1208                         strip_blank_front(&p);
1209
1210                         if (p[0] == '[')
1211                                 break;
1212
1213                         strip_blank_end(p);
1214
1215                         fgetpos(f, &off);
1216
1217                         /*
1218                          * Don't break here, continue parsing options so we
1219                          * dump all the bad ones. Makes trial/error fixups
1220                          * easier on the user.
1221                          */
1222                         ret |= parse_option(p, options, td);
1223                 }
1224
1225                 if (!ret) {
1226                         fsetpos(f, &off);
1227                         ret = add_job(td, name, 0);
1228                 } else {
1229                         log_err("fio: job %s dropped\n", name);
1230                         put_job(td);
1231                 }
1232         } while (!ret);
1233
1234         free(string);
1235         free(name);
1236         fclose(f);
1237         return ret;
1238 }
1239
1240 static int fill_def_thread(void)
1241 {
1242         memset(&def_thread, 0, sizeof(def_thread));
1243
1244         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1245                 perror("sched_getaffinity");
1246                 return 1;
1247         }
1248
1249         /*
1250          * fill default options
1251          */
1252         fill_default_options(&def_thread, options);
1253
1254         def_thread.timeout = def_timeout;
1255         def_thread.write_bw_log = write_bw_log;
1256         def_thread.write_lat_log = write_lat_log;
1257
1258 #ifdef FIO_HAVE_DISK_UTIL
1259         def_thread.do_disk_util = 1;
1260 #endif
1261
1262         return 0;
1263 }
1264
1265 static void usage(void)
1266 {
1267         printf("%s\n", fio_version_string);
1268         printf("\t--output\tWrite output to file\n");
1269         printf("\t--timeout\tRuntime in seconds\n");
1270         printf("\t--latency-log\tGenerate per-job latency logs\n");
1271         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1272         printf("\t--minimal\tMinimal (terse) output\n");
1273         printf("\t--version\tPrint version info and exit\n");
1274         printf("\t--help\t\tPrint this page\n");
1275         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1276 }
1277
1278 static int parse_cmd_line(int argc, char *argv[])
1279 {
1280         struct thread_data *td = NULL;
1281         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1282
1283         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
1284                 switch (c) {
1285                 case 't':
1286                         def_timeout = atoi(optarg);
1287                         break;
1288                 case 'l':
1289                         write_lat_log = 1;
1290                         break;
1291                 case 'w':
1292                         write_bw_log = 1;
1293                         break;
1294                 case 'o':
1295                         f_out = fopen(optarg, "w+");
1296                         if (!f_out) {
1297                                 perror("fopen output");
1298                                 exit(1);
1299                         }
1300                         f_err = f_out;
1301                         break;
1302                 case 'm':
1303                         terse_output = 1;
1304                         break;
1305                 case 'h':
1306                         usage();
1307                         exit(0);
1308                 case 'c':
1309                         ret = show_cmd_help(options, optarg);
1310                         exit(ret);
1311                 case 'v':
1312                         printf("%s\n", fio_version_string);
1313                         exit(0);
1314                 case FIO_GETOPT_JOB: {
1315                         const char *opt = long_options[lidx].name;
1316                         char *val = optarg;
1317
1318                         if (!strncmp(opt, "name", 4) && td) {
1319                                 ret = add_job(td, td->name ?: "fio", 0);
1320                                 if (ret) {
1321                                         put_job(td);
1322                                         return 0;
1323                                 }
1324                                 td = NULL;
1325                         }
1326                         if (!td) {
1327                                 int global = !strncmp(val, "global", 6);
1328
1329                                 td = get_new_job(global, &def_thread);
1330                                 if (!td)
1331                                         return 0;
1332                         }
1333
1334                         ret = parse_cmd_option(opt, val, options, td);
1335                         if (ret)
1336                                 dont_add_job = 1;
1337                         break;
1338                 }
1339                 default:
1340                         break;
1341                 }
1342         }
1343
1344         if (td) {
1345                 if (dont_add_job)
1346                         put_job(td);
1347                 else {
1348                         ret = add_job(td, td->name ?: "fio", 0);
1349                         if (ret)
1350                                 put_job(td);
1351                 }
1352         }
1353
1354         while (optind < argc) {
1355                 ini_idx++;
1356                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1357                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1358                 optind++;
1359         }
1360
1361         return ini_idx;
1362 }
1363
1364 static void free_shm(void)
1365 {
1366         struct shmid_ds sbuf;
1367
1368         if (threads) {
1369                 shmdt((void *) threads);
1370                 threads = NULL;
1371                 shmctl(shm_id, IPC_RMID, &sbuf);
1372         }
1373 }
1374
1375 /*
1376  * The thread area is shared between the main process and the job
1377  * threads/processes. So setup a shared memory segment that will hold
1378  * all the job info.
1379  */
1380 static int setup_thread_area(void)
1381 {
1382         /*
1383          * 1024 is too much on some machines, scale max_jobs if
1384          * we get a failure that looks like too large a shm segment
1385          */
1386         do {
1387                 size_t size = max_jobs * sizeof(struct thread_data);
1388
1389                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1390                 if (shm_id != -1)
1391                         break;
1392                 if (errno != EINVAL) {
1393                         perror("shmget");
1394                         break;
1395                 }
1396
1397                 max_jobs >>= 1;
1398         } while (max_jobs);
1399
1400         if (shm_id == -1)
1401                 return 1;
1402
1403         threads = shmat(shm_id, NULL, 0);
1404         if (threads == (void *) -1) {
1405                 perror("shmat");
1406                 return 1;
1407         }
1408
1409         atexit(free_shm);
1410         return 0;
1411 }
1412
1413 /*
1414  * Copy the fio options into the long options map, so we mirror
1415  * job and cmd line options.
1416  */
1417 static void dupe_job_options(void)
1418 {
1419         struct fio_option *o;
1420         unsigned int i;
1421
1422         i = 0;
1423         while (long_options[i].name)
1424                 i++;
1425
1426         o = &options[0];
1427         while (o->name) {
1428                 long_options[i].name = o->name;
1429                 long_options[i].val = FIO_GETOPT_JOB;
1430                 if (o->type == FIO_OPT_STR_SET)
1431                         long_options[i].has_arg = no_argument;
1432                 else
1433                         long_options[i].has_arg = required_argument;
1434
1435                 i++;
1436                 o++;
1437                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1438         }
1439 }
1440
1441 int parse_options(int argc, char *argv[])
1442 {
1443         int job_files, i;
1444
1445         f_out = stdout;
1446         f_err = stderr;
1447
1448         options_init(options);
1449
1450         dupe_job_options();
1451
1452         if (setup_thread_area())
1453                 return 1;
1454         if (fill_def_thread())
1455                 return 1;
1456
1457         job_files = parse_cmd_line(argc, argv);
1458
1459         for (i = 0; i < job_files; i++) {
1460                 if (fill_def_thread())
1461                         return 1;
1462                 if (parse_jobs_ini(ini_file[i], i))
1463                         return 1;
1464                 free(ini_file[i]);
1465         }
1466
1467         free(ini_file);
1468
1469         if (!thread_number) {
1470                 log_err("No jobs defined(s)\n");
1471                 return 1;
1472         }
1473
1474         return 0;
1475 }