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   = "unlink",
 504                 .type   = FIO_OPT_BOOL,
 505                 .off1   = td_var_offset(unlink),
 506                 .help   = "Unlink created files after job has completed",
 507                 .def    = "0",
 508         },
 509         {
 510                 .name   = "exitall",
 511                 .type   = FIO_OPT_STR_SET,
 512                 .cb     = str_exitall_cb,
 513                 .help   = "Terminate all jobs when one exits",
 514         },
 515         {
 516                 .name   = "stonewall",
 517                 .type   = FIO_OPT_STR_SET,
 518                 .off1   = td_var_offset(stonewall),
 519                 .help   = "Insert a hard barrier between this job and previous",
 520         },
 521         {
 522                 .name   = "thread",
 523                 .type   = FIO_OPT_STR_SET,
 524                 .off1   = td_var_offset(use_thread),
 525                 .help   = "Use threads instead of forks",
 526         },
 527         {
 528                 .name   = "write_bw_log",
 529                 .type   = FIO_OPT_STR_SET,
 530                 .off1   = td_var_offset(write_bw_log),
 531                 .help   = "Write log of bandwidth during run",
 532         },
 533         {
 534                 .name   = "write_lat_log",
 535                 .type   = FIO_OPT_STR_SET,
 536                 .off1   = td_var_offset(write_lat_log),
 537                 .help   = "Write log of latency during run",
 538         },
 539         {
 540                 .name   = "hugepage-size",
 541                 .type   = FIO_OPT_STR_VAL,
 542                 .off1   = td_var_offset(hugepage_size),
 543                 .help   = "When using hugepages, specify size of each page",
 544                 .def    = __stringify(FIO_HUGE_PAGE),
 545         },
 546         {
 547                 .name   = "group_reporting",
 548                 .type   = FIO_OPT_STR_SET,
 549                 .off1   = td_var_offset(group_reporting),
 550                 .help   = "Do reporting on a per-group basis",
 551         },
 552         {
 553                 .name = NULL,
 554         },
 555 };
 556
 557 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 558 #define FIO_CMD_OPTS    (16)
 559 #define FIO_GETOPT_JOB  (0x89988998)
 560
 561 /*
 562  * Command line options. These will contain the above, plus a few
 563  * extra that only pertain to fio itself and not jobs.
 564  */
 565 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 566         {
 567                 .name           = "output",
 568                 .has_arg        = required_argument,
 569                 .val            = 'o',
 570         },
 571         {
 572                 .name           = "timeout",
 573                 .has_arg        = required_argument,
 574                 .val            = 't',
 575         },
 576         {
 577                 .name           = "latency-log",
 578                 .has_arg        = required_argument,
 579                 .val            = 'l',
 580         },
 581         {
 582                 .name           = "bandwidth-log",
 583                 .has_arg        = required_argument,
 584                 .val            = 'b',
 585         },
 586         {
 587                 .name           = "minimal",
 588                 .has_arg        = optional_argument,
 589                 .val            = 'm',
 590         },
 591         {
 592                 .name           = "version",
 593                 .has_arg        = no_argument,
 594                 .val            = 'v',
 595         },
 596         {
 597                 .name           = "help",
 598                 .has_arg        = no_argument,
 599                 .val            = 'h',
 600         },
 601         {
 602                 .name           = "cmdhelp",
 603                 .has_arg        = optional_argument,
 604                 .val            = 'c',
 605         },
 606         {
 607                 .name           = NULL,
 608         },
 609 };
 610
 611 static int def_timeout = 0;
 612
 613 static char fio_version_string[] = "fio 1.13";
 614
 615 static char **ini_file;
 616 static int max_jobs = MAX_JOBS;
 617
 618 struct thread_data def_thread;
 619 struct thread_data *threads = NULL;
 620
 621 int exitall_on_terminate = 0;
 622 int terse_output = 0;
 623 unsigned long long mlock_size = 0;
 624 FILE *f_out = NULL;
 625 FILE *f_err = NULL;
 626
 627 static int write_lat_log = 0;
 628 int write_bw_log = 0;
 629
 630 static int prev_group_jobs;
 631
 632 FILE *get_f_out()
 633 {
 634         return f_out;
 635 }
 636
 637 FILE *get_f_err()
 638 {
 639         return f_err;
 640 }
 641
 642 /*
 643  * Return a free job structure.
 644  */
 645 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 646 {
 647         struct thread_data *td;
 648
 649         if (global)
 650                 return &def_thread;
 651         if (thread_number >= max_jobs)
 652                 return NULL;
 653
 654         td = &threads[thread_number++];
 655         *td = *parent;
 656
 657         td->thread_number = thread_number;
 658         return td;
 659 }
 660
 661 static void put_job(struct thread_data *td)
 662 {
 663         if (td == &def_thread)
 664                 return;
 665
 666         if (td->error)
 667                 fprintf(f_out, "fio: %s\n", td->verror);
 668
 669         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 670         thread_number--;
 671 }
 672
 673 /*
 674  * Lazy way of fixing up options that depend on each other. We could also
 675  * define option callback handlers, but this is easier.
 676  */
 677 static void fixup_options(struct thread_data *td)
 678 {
 679         if (!td->rwmixread && td->rwmixwrite)
 680                 td->rwmixread = 100 - td->rwmixwrite;
 681
 682         if (td->write_iolog_file && td->read_iolog_file) {
 683                 log_err("fio: read iolog overrides write_iolog\n");
 684                 free(td->write_iolog_file);
 685                 td->write_iolog_file = NULL;
 686         }
 687
 688         if (td->io_ops->flags & FIO_SYNCIO)
 689                 td->iodepth = 1;
 690         else {
 691                 if (!td->iodepth)
 692                         td->iodepth = td->open_files;
 693         }
 694
 695         /*
 696          * only really works for sequential io for now, and with 1 file
 697          */
 698         if (td->zone_size && td_random(td) && td->open_files == 1)
 699                 td->zone_size = 0;
 700
 701         /*
 702          * Reads can do overwrites, we always need to pre-create the file
 703          */
 704         if (td_read(td) || td_rw(td))
 705                 td->overwrite = 1;
 706
 707         if (!td->min_bs[DDIR_READ])
 708                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 709         if (!td->max_bs[DDIR_READ])
 710                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 711         if (!td->min_bs[DDIR_WRITE])
 712                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 713         if (!td->max_bs[DDIR_WRITE])
 714                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 715
 716         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 717
 718         if (td_read(td) && !td_rw(td))
 719                 td->verify = 0;
 720
 721         if (td->norandommap && td->verify != VERIFY_NONE) {
 722                 log_err("fio: norandommap given, verify disabled\n");
 723                 td->verify = VERIFY_NONE;
 724         }
 725         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 726                 log_err("fio: bs_unaligned may not work with raw io\n");
 727
 728         /*
 729          * thinktime_spin must be less than thinktime
 730          */
 731         if (td->thinktime_spin > td->thinktime)
 732                 td->thinktime_spin = td->thinktime;
 733
 734         /*
 735          * The low water mark cannot be bigger than the iodepth
 736          */
 737         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
 738                 /*
 739                  * syslet work around - if the workload is sequential,
 740                  * we want to let the queue drain all the way down to
 741                  * avoid seeking between async threads
 742                  */
 743                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 744                         td->iodepth_low = 1;
 745                 else
 746                         td->iodepth_low = td->iodepth;
 747         }
 748
 749         /*
 750          * If batch number isn't set, default to the same as iodepth
 751          */
 752         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
 753                 td->iodepth_batch = td->iodepth;
 754
 755         if (td->open_files > td->nr_files || !td->open_files)
 756                 td->open_files = td->nr_files;
 757 }
 758
 759 /*
 760  * This function leaks the buffer
 761  */
 762 static char *to_kmg(unsigned int val)
 763 {
 764         char *buf = malloc(32);
 765         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 766         char *p = post;
 767
 768         do {
 769                 if (val & 1023)
 770                         break;
 771
 772                 val >>= 10;
 773                 p++;
 774         } while (*p);
 775
 776         snprintf(buf, 31, "%u%c", val, *p);
 777         return buf;
 778 }
 779
 780 /* External engines are specified by "external:name.o") */
 781 static const char *get_engine_name(const char *str)
 782 {
 783         char *p = strstr(str, ":");
 784
 785         if (!p)
 786                 return str;
 787
 788         p++;
 789         strip_blank_front(&p);
 790         strip_blank_end(p);
 791         return p;
 792 }
 793
 794 /*
 795  * Adds a job to the list of things todo. Sanitizes the various options
 796  * to make sure we don't have conflicts, and initializes various
 797  * members of td.
 798  */
 799 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 800 {
 801         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 802                                    "randread", "randwrite", "randrw" };
 803         unsigned int i;
 804         struct fio_file *f;
 805         const char *engine;
 806         char fname[PATH_MAX];
 807         int numjobs;
 808
 809         /*
 810          * the def_thread is just for options, it's not a real job
 811          */
 812         if (td == &def_thread)
 813                 return 0;
 814
 815         engine = get_engine_name(td->ioengine);
 816         td->io_ops = load_ioengine(td, engine);
 817         if (!td->io_ops) {
 818                 log_err("fio: failed to load engine %s\n", engine);
 819                 return 1;
 820         }
 821
 822         if (td->use_thread)
 823                 nr_thread++;
 824         else
 825                 nr_process++;
 826
 827         if (td->odirect)
 828                 td->io_ops->flags |= FIO_RAWIO;
 829
 830         if (!td->filename) {
 831                 td->filename = strdup(jobname);
 832
 833                 for (i = 0; i < td->nr_files; i++) {
 834                         sprintf(fname, "%s.%d.%d", td->filename, td->thread_number, i);
 835                         add_file(td, fname);
 836                 }
 837         }
 838
 839         fixup_options(td);
 840
 841         for_each_file(td, f, i) {
 842                 if (td->directory && f->filetype == FIO_TYPE_FILE) {
 843                         sprintf(fname, "%s/%s", td->directory, f->file_name);
 844                         f->file_name = strdup(fname);
 845                 }
 846
 847                 f->file_size = td->total_file_size / td->nr_files;
 848                 f->file_offset = td->start_offset;
 849         }
 850
 851         td->mutex = fio_sem_init(0);
 852
 853         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 854         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 855         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 856
 857         if ((td->stonewall || td->numjobs > 1) && prev_group_jobs) {
 858                 prev_group_jobs = 0;
 859                 groupid++;
 860         }
 861
 862         td->groupid = groupid;
 863         prev_group_jobs++;
 864
 865         if (setup_rate(td))
 866                 goto err;
 867
 868         if (td->write_lat_log) {
 869                 setup_log(&td->ts.slat_log);
 870                 setup_log(&td->ts.clat_log);
 871         }
 872         if (td->write_bw_log)
 873                 setup_log(&td->ts.bw_log);
 874
 875         if (!td->name)
 876                 td->name = strdup(jobname);
 877
 878         if (!terse_output) {
 879                 if (!job_add_num) {
 880                         if (!strcmp(td->io_ops->name, "cpuio"))
 881                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 882                         else {
 883                                 char *c1, *c2, *c3, *c4;
 884
 885                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 886                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 887                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 888                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 889
 890                                 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);
 891
 892                                 free(c1);
 893                                 free(c2);
 894                                 free(c3);
 895                                 free(c4);
 896                         }
 897                 } else if (job_add_num == 1)
 898                         fprintf(f_out, "...\n");
 899         }
 900
 901         /*
 902          * recurse add identical jobs, clear numjobs and stonewall options
 903          * as they don't apply to sub-jobs
 904          */
 905         numjobs = td->numjobs;
 906         while (--numjobs) {
 907                 struct thread_data *td_new = get_new_job(0, td);
 908
 909                 if (!td_new)
 910                         goto err;
 911
 912                 td_new->numjobs = 1;
 913                 td_new->stonewall = 0;
 914                 job_add_num = numjobs - 1;
 915
 916                 if (add_job(td_new, jobname, job_add_num))
 917                         goto err;
 918         }
 919
 920         if (td->numjobs > 1) {
 921                 groupid++;
 922                 prev_group_jobs = 0;
 923         }
 924
 925         return 0;
 926 err:
 927         put_job(td);
 928         return -1;
 929 }
 930
 931 /*
 932  * Initialize the various random states we need (random io, block size ranges,
 933  * read/write mix, etc).
 934  */
 935 int init_random_state(struct thread_data *td)
 936 {
 937         unsigned long seeds[5];
 938         int fd, num_maps, blocks;
 939         struct fio_file *f;
 940         unsigned int i;
 941
 942         if (td->io_ops->flags & FIO_DISKLESSIO)
 943                 return 0;
 944
 945         fd = open("/dev/urandom", O_RDONLY);
 946         if (fd == -1) {
 947                 td_verror(td, errno, "open");
 948                 return 1;
 949         }
 950
 951         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 952                 td_verror(td, EIO, "read");
 953                 close(fd);
 954                 return 1;
 955         }
 956
 957         close(fd);
 958
 959         os_random_seed(seeds[0], &td->bsrange_state);
 960         os_random_seed(seeds[1], &td->verify_state);
 961         os_random_seed(seeds[2], &td->rwmix_state);
 962
 963         if (td->file_service_type == FIO_FSERVICE_RANDOM)
 964                 os_random_seed(seeds[3], &td->next_file_state);
 965
 966         if (!td_random(td))
 967                 return 0;
 968
 969         if (td->rand_repeatable)
 970                 seeds[4] = FIO_RANDSEED * td->thread_number;
 971
 972         if (!td->norandommap) {
 973                 for_each_file(td, f, i) {
 974                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 975                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 976                         f->file_map = malloc(num_maps * sizeof(long));
 977                         if (!f->file_map) {
 978                                 log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
 979                                 return 1;
 980                         }
 981                         f->num_maps = num_maps;
 982                         memset(f->file_map, 0, num_maps * sizeof(long));
 983                 }
 984         }
 985
 986         os_random_seed(seeds[4], &td->random_state);
 987         return 0;
 988 }
 989
 990 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 991 {
 992 #ifdef FIO_HAVE_CPU_AFFINITY
 993         unsigned int i;
 994
 995         CPU_ZERO(&cpumask);
 996
 997         for (i = 0; i < sizeof(int) * 8; i++) {
 998                 if ((1 << i) & cpu)
 999                         CPU_SET(i, &cpumask);
1000         }
1001 #endif
1002 }
1003
1004 static int is_empty_or_comment(char *line)
1005 {
1006         unsigned int i;
1007
1008         for (i = 0; i < strlen(line); i++) {
1009                 if (line[i] == ';')
1010                         return 1;
1011                 if (line[i] == '#')
1012                         return 1;
1013                 if (!isspace(line[i]) && !iscntrl(line[i]))
1014                         return 0;
1015         }
1016
1017         return 1;
1018 }
1019
1020 /*
1021  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
1022  */
1023 static char *get_opt_postfix(const char *str)
1024 {
1025         char *p = strstr(str, ":");
1026
1027         if (!p)
1028                 return NULL;
1029
1030         p++;
1031         strip_blank_front(&p);
1032         strip_blank_end(p);
1033         return strdup(p);
1034 }
1035
1036 static int str_mem_cb(void *data, const char *mem)
1037 {
1038         struct thread_data *td = data;
1039
1040         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1041                 td->mmapfile = get_opt_postfix(mem);
1042                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1043                         log_err("fio: mmaphuge:/path/to/file\n");
1044                         return 1;
1045                 }
1046         }
1047
1048         return 0;
1049 }
1050
1051 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1052 {
1053         mlock_size = *val;
1054         return 0;
1055 }
1056
1057 #ifdef FIO_HAVE_IOPRIO
1058 static int str_prioclass_cb(void *data, unsigned int *val)
1059 {
1060         struct thread_data *td = data;
1061
1062         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1063         return 0;
1064 }
1065
1066 static int str_prio_cb(void *data, unsigned int *val)
1067 {
1068         struct thread_data *td = data;
1069
1070         td->ioprio |= *val;
1071         return 0;
1072 }
1073 #endif
1074
1075 static int str_exitall_cb(void)
1076 {
1077         exitall_on_terminate = 1;
1078         return 0;
1079 }
1080
1081 static int str_cpumask_cb(void *data, unsigned int *val)
1082 {
1083         struct thread_data *td = data;
1084
1085         fill_cpu_mask(td->cpumask, *val);
1086         return 0;
1087 }
1088
1089 static int str_fst_cb(void *data, const char *str)
1090 {
1091         struct thread_data *td = data;
1092         char *nr = get_opt_postfix(str);
1093
1094         td->file_service_nr = 1;
1095         if (nr)
1096                 td->file_service_nr = atoi(nr);
1097
1098         return 0;
1099 }
1100
1101 static int str_filename_cb(void *data, const char *input)
1102 {
1103         struct thread_data *td = data;
1104         char *fname, *str;
1105
1106         str = strdup(input);
1107         while ((fname = strsep(&str, ":")) != NULL)
1108                 add_file(td, fname);
1109
1110         return 0;
1111 }
1112
1113 static int str_directory_cb(void *data, const char fio_unused *str)
1114 {
1115         struct thread_data *td = data;
1116         struct stat sb;
1117
1118         if (lstat(td->directory, &sb) < 0) {
1119                 log_err("fio: %s is not a directory\n", td->directory);
1120                 td_verror(td, errno, "lstat");
1121                 return 1;
1122         }
1123         if (!S_ISDIR(sb.st_mode)) {
1124                 log_err("fio: %s is not a directory\n", td->directory);
1125                 return 1;
1126         }
1127
1128         return 0;
1129 }
1130
1131 /*
1132  * This is our [ini] type file parser.
1133  */
1134 static int parse_jobs_ini(char *file, int stonewall_flag)
1135 {
1136         unsigned int global;
1137         struct thread_data *td;
1138         char *string, *name;
1139         fpos_t off;
1140         FILE *f;
1141         char *p;
1142         int ret = 0, stonewall;
1143
1144         f = fopen(file, "r");
1145         if (!f) {
1146                 perror("fopen job file");
1147                 return 1;
1148         }
1149
1150         string = malloc(4096);
1151         name = malloc(256);
1152         memset(name, 0, 256);
1153
1154         stonewall = stonewall_flag;
1155         do {
1156                 p = fgets(string, 4095, f);
1157                 if (!p)
1158                         break;
1159                 if (is_empty_or_comment(p))
1160                         continue;
1161                 if (sscanf(p, "[%255s]", name) != 1)
1162                         continue;
1163
1164                 global = !strncmp(name, "global", 6);
1165
1166                 name[strlen(name) - 1] = '\0';
1167
1168                 td = get_new_job(global, &def_thread);
1169                 if (!td) {
1170                         ret = 1;
1171                         break;
1172                 }
1173
1174                 /*
1175                  * Seperate multiple job files by a stonewall
1176                  */
1177                 if (!global && stonewall) {
1178                         td->stonewall = stonewall;
1179                         stonewall = 0;
1180                 }
1181
1182                 fgetpos(f, &off);
1183                 while ((p = fgets(string, 4096, f)) != NULL) {
1184                         if (is_empty_or_comment(p))
1185                                 continue;
1186
1187                         strip_blank_front(&p);
1188
1189                         if (p[0] == '[')
1190                                 break;
1191
1192                         strip_blank_end(p);
1193
1194                         fgetpos(f, &off);
1195
1196                         /*
1197                          * Don't break here, continue parsing options so we
1198                          * dump all the bad ones. Makes trial/error fixups
1199                          * easier on the user.
1200                          */
1201                         ret |= parse_option(p, options, td);
1202                 }
1203
1204                 if (!ret) {
1205                         fsetpos(f, &off);
1206                         ret = add_job(td, name, 0);
1207                 } else {
1208                         log_err("fio: job %s dropped\n", name);
1209                         put_job(td);
1210                 }
1211         } while (!ret);
1212
1213         free(string);
1214         free(name);
1215         fclose(f);
1216         return ret;
1217 }
1218
1219 static int fill_def_thread(void)
1220 {
1221         memset(&def_thread, 0, sizeof(def_thread));
1222
1223         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1224                 perror("sched_getaffinity");
1225                 return 1;
1226         }
1227
1228         /*
1229          * fill default options
1230          */
1231         fill_default_options(&def_thread, options);
1232
1233         def_thread.timeout = def_timeout;
1234         def_thread.write_bw_log = write_bw_log;
1235         def_thread.write_lat_log = write_lat_log;
1236
1237 #ifdef FIO_HAVE_DISK_UTIL
1238         def_thread.do_disk_util = 1;
1239 #endif
1240
1241         return 0;
1242 }
1243
1244 static void usage(void)
1245 {
1246         printf("%s\n", fio_version_string);
1247         printf("\t--output\tWrite output to file\n");
1248         printf("\t--timeout\tRuntime in seconds\n");
1249         printf("\t--latency-log\tGenerate per-job latency logs\n");
1250         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1251         printf("\t--minimal\tMinimal (terse) output\n");
1252         printf("\t--version\tPrint version info and exit\n");
1253         printf("\t--help\t\tPrint this page\n");
1254         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1255 }
1256
1257 static int parse_cmd_line(int argc, char *argv[])
1258 {
1259         struct thread_data *td = NULL;
1260         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1261
1262         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
1263                 switch (c) {
1264                 case 't':
1265                         def_timeout = atoi(optarg);
1266                         break;
1267                 case 'l':
1268                         write_lat_log = 1;
1269                         break;
1270                 case 'w':
1271                         write_bw_log = 1;
1272                         break;
1273                 case 'o':
1274                         f_out = fopen(optarg, "w+");
1275                         if (!f_out) {
1276                                 perror("fopen output");
1277                                 exit(1);
1278                         }
1279                         f_err = f_out;
1280                         break;
1281                 case 'm':
1282                         terse_output = 1;
1283                         break;
1284                 case 'h':
1285                         usage();
1286                         exit(0);
1287                 case 'c':
1288                         ret = show_cmd_help(options, optarg);
1289                         exit(ret);
1290                 case 'v':
1291                         printf("%s\n", fio_version_string);
1292                         exit(0);
1293                 case FIO_GETOPT_JOB: {
1294                         const char *opt = long_options[lidx].name;
1295                         char *val = optarg;
1296
1297                         if (!strncmp(opt, "name", 4) && td) {
1298                                 ret = add_job(td, td->name ?: "fio", 0);
1299                                 if (ret) {
1300                                         put_job(td);
1301                                         return 0;
1302                                 }
1303                                 td = NULL;
1304                         }
1305                         if (!td) {
1306                                 int global = !strncmp(val, "global", 6);
1307
1308                                 td = get_new_job(global, &def_thread);
1309                                 if (!td)
1310                                         return 0;
1311                         }
1312
1313                         ret = parse_cmd_option(opt, val, options, td);
1314                         if (ret)
1315                                 dont_add_job = 1;
1316                         break;
1317                 }
1318                 default:
1319                         break;
1320                 }
1321         }
1322
1323         if (td) {
1324                 if (dont_add_job)
1325                         put_job(td);
1326                 else {
1327                         ret = add_job(td, td->name ?: "fio", 0);
1328                         if (ret)
1329                                 put_job(td);
1330                 }
1331         }
1332
1333         while (optind < argc) {
1334                 ini_idx++;
1335                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1336                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1337                 optind++;
1338         }
1339
1340         return ini_idx;
1341 }
1342
1343 static void free_shm(void)
1344 {
1345         struct shmid_ds sbuf;
1346
1347         if (threads) {
1348                 shmdt((void *) threads);
1349                 threads = NULL;
1350                 shmctl(shm_id, IPC_RMID, &sbuf);
1351         }
1352 }
1353
1354 /*
1355  * The thread area is shared between the main process and the job
1356  * threads/processes. So setup a shared memory segment that will hold
1357  * all the job info.
1358  */
1359 static int setup_thread_area(void)
1360 {
1361         /*
1362          * 1024 is too much on some machines, scale max_jobs if
1363          * we get a failure that looks like too large a shm segment
1364          */
1365         do {
1366                 size_t size = max_jobs * sizeof(struct thread_data);
1367
1368                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1369                 if (shm_id != -1)
1370                         break;
1371                 if (errno != EINVAL) {
1372                         perror("shmget");
1373                         break;
1374                 }
1375
1376                 max_jobs >>= 1;
1377         } while (max_jobs);
1378
1379         if (shm_id == -1)
1380                 return 1;
1381
1382         threads = shmat(shm_id, NULL, 0);
1383         if (threads == (void *) -1) {
1384                 perror("shmat");
1385                 return 1;
1386         }
1387
1388         atexit(free_shm);
1389         return 0;
1390 }
1391
1392 /*
1393  * Copy the fio options into the long options map, so we mirror
1394  * job and cmd line options.
1395  */
1396 static void dupe_job_options(void)
1397 {
1398         struct fio_option *o;
1399         unsigned int i;
1400
1401         i = 0;
1402         while (long_options[i].name)
1403                 i++;
1404
1405         o = &options[0];
1406         while (o->name) {
1407                 long_options[i].name = o->name;
1408                 long_options[i].val = FIO_GETOPT_JOB;
1409                 if (o->type == FIO_OPT_STR_SET)
1410                         long_options[i].has_arg = no_argument;
1411                 else
1412                         long_options[i].has_arg = required_argument;
1413
1414                 i++;
1415                 o++;
1416                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1417         }
1418 }
1419
1420 int parse_options(int argc, char *argv[])
1421 {
1422         int job_files, i;
1423
1424         f_out = stdout;
1425         f_err = stderr;
1426
1427         options_init(options);
1428
1429         dupe_job_options();
1430
1431         if (setup_thread_area())
1432                 return 1;
1433         if (fill_def_thread())
1434                 return 1;
1435
1436         job_files = parse_cmd_line(argc, argv);
1437
1438         for (i = 0; i < job_files; i++) {
1439                 if (fill_def_thread())
1440                         return 1;
1441                 if (parse_jobs_ini(ini_file[i], i))
1442                         return 1;
1443                 free(ini_file[i]);
1444         }
1445
1446         free(ini_file);
1447
1448         if (!thread_number) {
1449                 log_err("No jobs defined(s)\n");
1450                 return 1;
1451         }
1452
1453         return 0;
1454 }