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_ioengine_cb(void *, const char *);
  26 static int str_mem_cb(void *, const char *);
  27 static int str_lockmem_cb(void *, unsigned long *);
  28 #ifdef FIO_HAVE_IOPRIO
  29 static int str_prio_cb(void *, unsigned int *);
  30 static int str_prioclass_cb(void *, unsigned int *);
  31 #endif
  32 static int str_exitall_cb(void);
  33 static int str_cpumask_cb(void *, unsigned int *);
  34
  35 #define __stringify_1(x)        #x
  36 #define __stringify(x)          __stringify_1(x)
  37
  38 /*
  39  * Map of job/command line options
  40  */
  41 static struct fio_option options[] = {
  42         {
  43                 .name   = "description",
  44                 .type   = FIO_OPT_STR_STORE,
  45                 .off1   = td_var_offset(description),
  46                 .help   = "Text job description",
  47         },
  48         {
  49                 .name   = "name",
  50                 .type   = FIO_OPT_STR_STORE,
  51                 .off1   = td_var_offset(name),
  52                 .help   = "Name of this job",
  53         },
  54         {
  55                 .name   = "directory",
  56                 .type   = FIO_OPT_STR_STORE,
  57                 .off1   = td_var_offset(directory),
  58                 .help   = "Directory to store files in",
  59         },
  60         {
  61                 .name   = "filename",
  62                 .type   = FIO_OPT_STR_STORE,
  63                 .off1   = td_var_offset(filename),
  64                 .help   = "Force the use of a specific file",
  65         },
  66         {
  67                 .name   = "rw",
  68                 .type   = FIO_OPT_STR,
  69                 .off1   = td_var_offset(td_ddir),
  70                 .help   = "IO direction",
  71                 .def    = "read",
  72                 .posval = {
  73                           { .ival = "read", .oval = TD_DDIR_READ },
  74                           { .ival = "write", .oval = TD_DDIR_WRITE },
  75                           { .ival = "randread", .oval = TD_DDIR_RANDREAD },
  76                           { .ival = "randwrite", .oval = TD_DDIR_RANDWRITE },
  77                           { .ival = "rw", .oval = TD_DDIR_RW },
  78                           { .ival = "randrw", .oval = TD_DDIR_RANDRW },
  79                           },
  80         },
  81         {
  82                 .name   = "ioengine",
  83                 .type   = FIO_OPT_STR,
  84                 .cb     = str_ioengine_cb,
  85                 .help   = "IO engine to use",
  86                 .def    = "sync",
  87                 .posval = {
  88                           { .ival = "sync", },
  89 #ifdef FIO_HAVE_LIBAIO
  90                           { .ival = "libaio", },
  91 #endif
  92 #ifdef FIO_HAVE_POSIXAIO
  93                           { .ival = "posixaio", },
  94 #endif
  95                           { .ival = "mmap", },
  96 #ifdef FIO_HAVE_SPLICE
  97                           { .ival = "splice", },
  98 #endif
  99 #ifdef FIO_HAVE_SGIO
 100                           { .ival = "sg", },
 101 #endif
 102                           { .ival = "null", }, { .ival = "net", },
 103 #ifdef FIO_HAVE_SYSLET
 104                           { .ival = "syslet-rw", },
 105 #endif
 106                           { .ival = "external", },
 107                           },
 108         },
 109         {
 110                 .name   = "iodepth",
 111                 .type   = FIO_OPT_INT,
 112                 .off1   = td_var_offset(iodepth),
 113                 .help   = "Amount of IO buffers to keep in flight",
 114                 .def    = "1",
 115         },
 116         {
 117                 .name   = "iodepth_batch",
 118                 .type   = FIO_OPT_INT,
 119                 .off1   = td_var_offset(iodepth_batch),
 120                 .help   = "Number of IO to submit in one go",
 121         },
 122         {
 123                 .name   = "iodepth_low",
 124                 .type   = FIO_OPT_INT,
 125                 .off1   = td_var_offset(iodepth_low),
 126                 .help   = "Low water mark for queuing depth",
 127         },
 128         {
 129                 .name   = "size",
 130                 .type   = FIO_OPT_STR_VAL,
 131                 .off1   = td_var_offset(total_file_size),
 132                 .help   = "Size of device or file",
 133         },
 134         {
 135                 .name   = "bs",
 136                 .type   = FIO_OPT_STR_VAL_INT,
 137                 .off1   = td_var_offset(bs[DDIR_READ]),
 138                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 139                 .help   = "Block size unit",
 140                 .def    = "4k",
 141         },
 142         {
 143                 .name   = "bsrange",
 144                 .type   = FIO_OPT_RANGE,
 145                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 146                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 147                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 148                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 149                 .help   = "Set block size range (in more detail than bs)",
 150         },
 151         {
 152                 .name   = "bs_unaligned",
 153                 .type   = FIO_OPT_STR_SET,
 154                 .off1   = td_var_offset(bs_unaligned),
 155                 .help   = "Don't sector align IO buffer sizes",
 156         },
 157         {
 158                 .name   = "offset",
 159                 .type   = FIO_OPT_STR_VAL,
 160                 .off1   = td_var_offset(start_offset),
 161                 .help   = "Start IO from this offset",
 162                 .def    = "0",
 163         },
 164         {
 165                 .name   = "randrepeat",
 166                 .type   = FIO_OPT_BOOL,
 167                 .off1   = td_var_offset(rand_repeatable),
 168                 .help   = "Use repeatable random IO pattern",
 169                 .def    = "1",
 170         },
 171         {
 172                 .name   = "norandommap",
 173                 .type   = FIO_OPT_STR_SET,
 174                 .off1   = td_var_offset(norandommap),
 175                 .help   = "Accept potential duplicate random blocks",
 176         },
 177         {
 178                 .name   = "nrfiles",
 179                 .type   = FIO_OPT_INT,
 180                 .off1   = td_var_offset(nr_files),
 181                 .help   = "Split job workload between this number of files",
 182                 .def    = "1",
 183         },
 184         {
 185                 .name   = "openfiles",
 186                 .type   = FIO_OPT_INT,
 187                 .off1   = td_var_offset(open_files),
 188                 .help   = "Number of files to keep open at the same time",
 189         },
 190         {
 191                 .name   = "file_service_type",
 192                 .type   = FIO_OPT_STR,
 193                 .off1   = td_var_offset(file_service_type),
 194                 .help   = "How to select which file to service next",
 195                 .def    = "roundrobin",
 196                 .posval = {
 197                           { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
 198                           { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
 199                           },
 200         },
 201         {
 202                 .name   = "fsync",
 203                 .type   = FIO_OPT_INT,
 204                 .off1   = td_var_offset(fsync_blocks),
 205                 .help   = "Issue fsync for writes every given number of blocks",
 206                 .def    = "0",
 207         },
 208         {
 209                 .name   = "direct",
 210                 .type   = FIO_OPT_BOOL,
 211                 .off1   = td_var_offset(odirect),
 212                 .help   = "Use O_DIRECT IO (negates buffered)",
 213                 .def    = "0",
 214         },
 215         {
 216                 .name   = "buffered",
 217                 .type   = FIO_OPT_BOOL,
 218                 .off1   = td_var_offset(odirect),
 219                 .neg    = 1,
 220                 .help   = "Use buffered IO (negates direct)",
 221                 .def    = "1",
 222         },
 223         {
 224                 .name   = "overwrite",
 225                 .type   = FIO_OPT_BOOL,
 226                 .off1   = td_var_offset(overwrite),
 227                 .help   = "When writing, set whether to overwrite current data",
 228                 .def    = "0",
 229         },
 230         {
 231                 .name   = "loops",
 232                 .type   = FIO_OPT_INT,
 233                 .off1   = td_var_offset(loops),
 234                 .help   = "Number of times to run the job",
 235                 .def    = "1",
 236         },
 237         {
 238                 .name   = "numjobs",
 239                 .type   = FIO_OPT_INT,
 240                 .off1   = td_var_offset(numjobs),
 241                 .help   = "Duplicate this job this many times",
 242                 .def    = "1",
 243         },
 244         {
 245                 .name   = "startdelay",
 246                 .type   = FIO_OPT_INT,
 247                 .off1   = td_var_offset(start_delay),
 248                 .help   = "Only start job when this period has passed",
 249                 .def    = "0",
 250         },
 251         {
 252                 .name   = "runtime",
 253                 .alias  = "timeout",
 254                 .type   = FIO_OPT_STR_VAL_TIME,
 255                 .off1   = td_var_offset(timeout),
 256                 .help   = "Stop workload when this amount of time has passed",
 257                 .def    = "0",
 258         },
 259         {
 260                 .name   = "mem",
 261                 .type   = FIO_OPT_STR,
 262                 .cb     = str_mem_cb,
 263                 .off1   = td_var_offset(mem_type),
 264                 .help   = "Backing type for IO buffers",
 265                 .def    = "malloc",
 266                 .posval = {
 267                           { .ival = "malloc", .oval = MEM_MALLOC },
 268                           { .ival = "shm", .oval = MEM_SHM },
 269 #ifdef FIO_HAVE_HUGETLB
 270                           { .ival = "shmhuge", .oval = MEM_SHMHUGE },
 271 #endif
 272                           { .ival = "mmap", .oval = MEM_MMAP },
 273 #ifdef FIO_HAVE_HUGETLB
 274                           { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
 275 #endif
 276                           },
 277         },
 278         {
 279                 .name   = "verify",
 280                 .type   = FIO_OPT_STR,
 281                 .off1   = td_var_offset(verify),
 282                 .help   = "Verify sum function",
 283                 .def    = "0",
 284                 .posval = {
 285                           { .ival = "0", .oval = VERIFY_NONE },
 286                           { .ival = "crc32", .oval = VERIFY_CRC32 },
 287                           { .ival = "md5", .oval = VERIFY_MD5 },
 288                           },
 289         },
 290         {
 291                 .name   = "write_iolog",
 292                 .type   = FIO_OPT_STR_STORE,
 293                 .off1   = td_var_offset(write_iolog_file),
 294                 .help   = "Store IO pattern to file",
 295         },
 296         {
 297                 .name   = "read_iolog",
 298                 .type   = FIO_OPT_STR_STORE,
 299                 .off1   = td_var_offset(read_iolog_file),
 300                 .help   = "Playback IO pattern from file",
 301         },
 302         {
 303                 .name   = "exec_prerun",
 304                 .type   = FIO_OPT_STR_STORE,
 305                 .off1   = td_var_offset(exec_prerun),
 306                 .help   = "Execute this file prior to running job",
 307         },
 308         {
 309                 .name   = "exec_postrun",
 310                 .type   = FIO_OPT_STR_STORE,
 311                 .off1   = td_var_offset(exec_postrun),
 312                 .help   = "Execute this file after running job",
 313         },
 314 #ifdef FIO_HAVE_IOSCHED_SWITCH
 315         {
 316                 .name   = "ioscheduler",
 317                 .type   = FIO_OPT_STR_STORE,
 318                 .off1   = td_var_offset(ioscheduler),
 319                 .help   = "Use this IO scheduler on the backing device",
 320         },
 321 #endif
 322         {
 323                 .name   = "zonesize",
 324                 .type   = FIO_OPT_STR_VAL,
 325                 .off1   = td_var_offset(zone_size),
 326                 .help   = "Give size of an IO zone",
 327                 .def    = "0",
 328         },
 329         {
 330                 .name   = "zoneskip",
 331                 .type   = FIO_OPT_STR_VAL,
 332                 .off1   = td_var_offset(zone_skip),
 333                 .help   = "Space between IO zones",
 334                 .def    = "0",
 335         },
 336         {
 337                 .name   = "lockmem",
 338                 .type   = FIO_OPT_STR_VAL,
 339                 .cb     = str_lockmem_cb,
 340                 .help   = "Lock down this amount of memory",
 341                 .def    = "0",
 342         },
 343         {
 344                 .name   = "rwmixcycle",
 345                 .type   = FIO_OPT_INT,
 346                 .off1   = td_var_offset(rwmixcycle),
 347                 .help   = "Cycle period for mixed read/write workloads (msec)",
 348                 .def    = "500",
 349         },
 350         {
 351                 .name   = "rwmixread",
 352                 .type   = FIO_OPT_INT,
 353                 .off1   = td_var_offset(rwmixread),
 354                 .maxval = 100,
 355                 .help   = "Percentage of mixed workload that is reads",
 356                 .def    = "50",
 357         },
 358         {
 359                 .name   = "rwmixwrite",
 360                 .type   = FIO_OPT_INT,
 361                 .off1   = td_var_offset(rwmixwrite),
 362                 .maxval = 100,
 363                 .help   = "Percentage of mixed workload that is writes",
 364                 .def    = "50",
 365         },
 366         {
 367                 .name   = "nice",
 368                 .type   = FIO_OPT_INT,
 369                 .off1   = td_var_offset(nice),
 370                 .help   = "Set job CPU nice value",
 371                 .minval = -19,
 372                 .maxval = 20,
 373                 .def    = "0",
 374         },
 375 #ifdef FIO_HAVE_IOPRIO
 376         {
 377                 .name   = "prio",
 378                 .type   = FIO_OPT_INT,
 379                 .cb     = str_prio_cb,
 380                 .help   = "Set job IO priority value",
 381                 .minval = 0,
 382                 .maxval = 7,
 383         },
 384         {
 385                 .name   = "prioclass",
 386                 .type   = FIO_OPT_INT,
 387                 .cb     = str_prioclass_cb,
 388                 .help   = "Set job IO priority class",
 389                 .minval = 0,
 390                 .maxval = 3,
 391         },
 392 #endif
 393         {
 394                 .name   = "thinktime",
 395                 .type   = FIO_OPT_INT,
 396                 .off1   = td_var_offset(thinktime),
 397                 .help   = "Idle time between IO buffers (usec)",
 398                 .def    = "0",
 399         },
 400         {
 401                 .name   = "thinktime_spin",
 402                 .type   = FIO_OPT_INT,
 403                 .off1   = td_var_offset(thinktime_spin),
 404                 .help   = "Start think time by spinning this amount (usec)",
 405                 .def    = "0",
 406         },
 407         {
 408                 .name   = "thinktime_blocks",
 409                 .type   = FIO_OPT_INT,
 410                 .off1   = td_var_offset(thinktime_blocks),
 411                 .help   = "IO buffer period between 'thinktime'",
 412                 .def    = "1",
 413         },
 414         {
 415                 .name   = "rate",
 416                 .type   = FIO_OPT_INT,
 417                 .off1   = td_var_offset(rate),
 418                 .help   = "Set bandwidth rate",
 419         },
 420         {
 421                 .name   = "ratemin",
 422                 .type   = FIO_OPT_INT,
 423                 .off1   = td_var_offset(ratemin),
 424                 .help   = "The bottom limit accepted",
 425         },
 426         {
 427                 .name   = "ratecycle",
 428                 .type   = FIO_OPT_INT,
 429                 .off1   = td_var_offset(ratecycle),
 430                 .help   = "Window average for rate limits (msec)",
 431                 .def    = "1000",
 432         },
 433         {
 434                 .name   = "invalidate",
 435                 .type   = FIO_OPT_BOOL,
 436                 .off1   = td_var_offset(invalidate_cache),
 437                 .help   = "Invalidate buffer/page cache prior to running job",
 438                 .def    = "1",
 439         },
 440         {
 441                 .name   = "sync",
 442                 .type   = FIO_OPT_BOOL,
 443                 .off1   = td_var_offset(sync_io),
 444                 .help   = "Use O_SYNC for buffered writes",
 445                 .def    = "0",
 446         },
 447         {
 448                 .name   = "bwavgtime",
 449                 .type   = FIO_OPT_INT,
 450                 .off1   = td_var_offset(bw_avg_time),
 451                 .help   = "Time window over which to calculate bandwidth (msec)",
 452                 .def    = "500",
 453         },
 454         {
 455                 .name   = "create_serialize",
 456                 .type   = FIO_OPT_BOOL,
 457                 .off1   = td_var_offset(create_serialize),
 458                 .help   = "Serialize creating of job files",
 459                 .def    = "1",
 460         },
 461         {
 462                 .name   = "create_fsync",
 463                 .type   = FIO_OPT_BOOL,
 464                 .off1   = td_var_offset(create_fsync),
 465                 .help   = "Fsync file after creation",
 466                 .def    = "1",
 467         },
 468         {
 469                 .name   = "cpuload",
 470                 .type   = FIO_OPT_INT,
 471                 .off1   = td_var_offset(cpuload),
 472                 .help   = "Use this percentage of CPU",
 473         },
 474         {
 475                 .name   = "cpuchunks",
 476                 .type   = FIO_OPT_INT,
 477                 .off1   = td_var_offset(cpucycle),
 478                 .help   = "Length of the CPU burn cycles",
 479         },
 480 #ifdef FIO_HAVE_CPU_AFFINITY
 481         {
 482                 .name   = "cpumask",
 483                 .type   = FIO_OPT_INT,
 484                 .cb     = str_cpumask_cb,
 485                 .help   = "CPU affinity mask",
 486         },
 487 #endif
 488         {
 489                 .name   = "end_fsync",
 490                 .type   = FIO_OPT_BOOL,
 491                 .off1   = td_var_offset(end_fsync),
 492                 .help   = "Include fsync at the end of job",
 493                 .def    = "0",
 494         },
 495         {
 496                 .name   = "unlink",
 497                 .type   = FIO_OPT_BOOL,
 498                 .off1   = td_var_offset(unlink),
 499                 .help   = "Unlink created files after job has completed",
 500                 .def    = "0",
 501         },
 502         {
 503                 .name   = "exitall",
 504                 .type   = FIO_OPT_STR_SET,
 505                 .cb     = str_exitall_cb,
 506                 .help   = "Terminate all jobs when one exits",
 507         },
 508         {
 509                 .name   = "stonewall",
 510                 .type   = FIO_OPT_STR_SET,
 511                 .off1   = td_var_offset(stonewall),
 512                 .help   = "Insert a hard barrier between this job and previous",
 513         },
 514         {
 515                 .name   = "thread",
 516                 .type   = FIO_OPT_STR_SET,
 517                 .off1   = td_var_offset(use_thread),
 518                 .help   = "Use threads instead of forks",
 519         },
 520         {
 521                 .name   = "write_bw_log",
 522                 .type   = FIO_OPT_STR_SET,
 523                 .off1   = td_var_offset(write_bw_log),
 524                 .help   = "Write log of bandwidth during run",
 525         },
 526         {
 527                 .name   = "write_lat_log",
 528                 .type   = FIO_OPT_STR_SET,
 529                 .off1   = td_var_offset(write_lat_log),
 530                 .help   = "Write log of latency during run",
 531         },
 532         {
 533                 .name   = "hugepage-size",
 534                 .type   = FIO_OPT_STR_VAL,
 535                 .off1   = td_var_offset(hugepage_size),
 536                 .help   = "When using hugepages, specify size of each page",
 537                 .def    = __stringify(FIO_HUGE_PAGE),
 538         },
 539         {
 540                 .name   = "group_reporting",
 541                 .type   = FIO_OPT_STR_SET,
 542                 .off1   = td_var_offset(group_reporting),
 543                 .help   = "Do reporting on a per-group basis",
 544         },
 545         {
 546                 .name = NULL,
 547         },
 548 };
 549
 550 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 551 #define FIO_CMD_OPTS    (16)
 552 #define FIO_GETOPT_JOB  (0x89988998)
 553
 554 /*
 555  * Command line options. These will contain the above, plus a few
 556  * extra that only pertain to fio itself and not jobs.
 557  */
 558 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 559         {
 560                 .name           = "output",
 561                 .has_arg        = required_argument,
 562                 .val            = 'o',
 563         },
 564         {
 565                 .name           = "timeout",
 566                 .has_arg        = required_argument,
 567                 .val            = 't',
 568         },
 569         {
 570                 .name           = "latency-log",
 571                 .has_arg        = required_argument,
 572                 .val            = 'l',
 573         },
 574         {
 575                 .name           = "bandwidth-log",
 576                 .has_arg        = required_argument,
 577                 .val            = 'b',
 578         },
 579         {
 580                 .name           = "minimal",
 581                 .has_arg        = optional_argument,
 582                 .val            = 'm',
 583         },
 584         {
 585                 .name           = "version",
 586                 .has_arg        = no_argument,
 587                 .val            = 'v',
 588         },
 589         {
 590                 .name           = "help",
 591                 .has_arg        = no_argument,
 592                 .val            = 'h',
 593         },
 594         {
 595                 .name           = "cmdhelp",
 596                 .has_arg        = optional_argument,
 597                 .val            = 'c',
 598         },
 599         {
 600                 .name           = NULL,
 601         },
 602 };
 603
 604 static int def_timeout = 0;
 605
 606 static char fio_version_string[] = "fio 1.13";
 607
 608 static char **ini_file;
 609 static int max_jobs = MAX_JOBS;
 610
 611 struct thread_data def_thread;
 612 struct thread_data *threads = NULL;
 613
 614 int exitall_on_terminate = 0;
 615 int terse_output = 0;
 616 unsigned long long mlock_size = 0;
 617 FILE *f_out = NULL;
 618 FILE *f_err = NULL;
 619
 620 static int write_lat_log = 0;
 621 int write_bw_log = 0;
 622
 623 FILE *get_f_out()
 624 {
 625         return f_out;
 626 }
 627
 628 FILE *get_f_err()
 629 {
 630         return f_err;
 631 }
 632
 633 /*
 634  * Return a free job structure.
 635  */
 636 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 637 {
 638         struct thread_data *td;
 639
 640         if (global)
 641                 return &def_thread;
 642         if (thread_number >= max_jobs)
 643                 return NULL;
 644
 645         td = &threads[thread_number++];
 646         *td = *parent;
 647
 648         td->thread_number = thread_number;
 649         return td;
 650 }
 651
 652 static void put_job(struct thread_data *td)
 653 {
 654         if (td == &def_thread)
 655                 return;
 656
 657         if (td->error)
 658                 fprintf(f_out, "fio: %s\n", td->verror);
 659
 660         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 661         thread_number--;
 662 }
 663
 664 /*
 665  * Lazy way of fixing up options that depend on each other. We could also
 666  * define option callback handlers, but this is easier.
 667  */
 668 static void fixup_options(struct thread_data *td)
 669 {
 670         if (!td->rwmixread && td->rwmixwrite)
 671                 td->rwmixread = 100 - td->rwmixwrite;
 672
 673         if (td->write_iolog_file && td->read_iolog_file) {
 674                 log_err("fio: read iolog overrides write_iolog\n");
 675                 free(td->write_iolog_file);
 676                 td->write_iolog_file = NULL;
 677         }
 678
 679         if (td->io_ops->flags & FIO_SYNCIO)
 680                 td->iodepth = 1;
 681         else {
 682                 if (!td->iodepth)
 683                         td->iodepth = td->open_files;
 684         }
 685
 686         /*
 687          * only really works for sequential io for now, and with 1 file
 688          */
 689         if (td->zone_size && td_random(td) && td->open_files == 1)
 690                 td->zone_size = 0;
 691
 692         /*
 693          * Reads can do overwrites, we always need to pre-create the file
 694          */
 695         if (td_read(td) || td_rw(td))
 696                 td->overwrite = 1;
 697
 698         if (!td->min_bs[DDIR_READ])
 699                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 700         if (!td->max_bs[DDIR_READ])
 701                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 702         if (!td->min_bs[DDIR_WRITE])
 703                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 704         if (!td->max_bs[DDIR_WRITE])
 705                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 706
 707         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 708
 709         if (td_read(td) && !td_rw(td))
 710                 td->verify = 0;
 711
 712         if (td->norandommap && td->verify != VERIFY_NONE) {
 713                 log_err("fio: norandommap given, verify disabled\n");
 714                 td->verify = VERIFY_NONE;
 715         }
 716         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 717                 log_err("fio: bs_unaligned may not work with raw io\n");
 718
 719         /*
 720          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 721          */
 722         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 723                 td->odirect = 0;
 724
 725         /*
 726          * thinktime_spin must be less than thinktime
 727          */
 728         if (td->thinktime_spin > td->thinktime)
 729                 td->thinktime_spin = td->thinktime;
 730
 731         /*
 732          * The low water mark cannot be bigger than the iodepth
 733          */
 734         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
 735                 /*
 736                  * syslet work around - if the workload is sequential,
 737                  * we want to let the queue drain all the way down to
 738                  * avoid seeking between async threads
 739                  */
 740                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 741                         td->iodepth_low = 1;
 742                 else
 743                         td->iodepth_low = td->iodepth;
 744         }
 745
 746         /*
 747          * If batch number isn't set, default to the same as iodepth
 748          */
 749         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
 750                 td->iodepth_batch = td->iodepth;
 751
 752         if (td->open_files > td->nr_files || !td->open_files)
 753                 td->open_files = td->nr_files;
 754 }
 755
 756 /*
 757  * This function leaks the buffer
 758  */
 759 static char *to_kmg(unsigned int val)
 760 {
 761         char *buf = malloc(32);
 762         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 763         char *p = post;
 764
 765         do {
 766                 if (val & 1023)
 767                         break;
 768
 769                 val >>= 10;
 770                 p++;
 771         } while (*p);
 772
 773         snprintf(buf, 31, "%u%c", val, *p);
 774         return buf;
 775 }
 776
 777 /*
 778  * Adds a job to the list of things todo. Sanitizes the various options
 779  * to make sure we don't have conflicts, and initializes various
 780  * members of td.
 781  */
 782 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 783 {
 784         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 785                                    "randread", "randwrite", "randrw" };
 786         struct stat sb;
 787         int numjobs, i;
 788         struct fio_file *f;
 789
 790         /*
 791          * the def_thread is just for options, it's not a real job
 792          */
 793         if (td == &def_thread)
 794                 return 0;
 795
 796         assert(td->io_ops);
 797
 798         if (td->odirect)
 799                 td->io_ops->flags |= FIO_RAWIO;
 800
 801         td->filetype = FIO_TYPE_FILE;
 802         if (td->filename && !lstat(td->filename, &sb)) {
 803                 if (S_ISBLK(sb.st_mode))
 804                         td->filetype = FIO_TYPE_BD;
 805                 else if (S_ISCHR(sb.st_mode))
 806                         td->filetype = FIO_TYPE_CHAR;
 807         }
 808
 809         fixup_options(td);
 810
 811         if (td->filename)
 812                 td->nr_uniq_files = 1;
 813         else
 814                 td->nr_uniq_files = td->open_files;
 815
 816         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 817                 char tmp[PATH_MAX];
 818                 int len = 0;
 819
 820                 if (td->directory && td->directory[0] != '\0') {
 821                         if (lstat(td->directory, &sb) < 0) {
 822                                 log_err("fio: %s is not a directory\n", td->directory);
 823                                 td_verror(td, errno, "lstat");
 824                                 return 1;
 825                         }
 826                         if (!S_ISDIR(sb.st_mode)) {
 827                                 log_err("fio: %s is not a directory\n", td->directory);
 828                                 return 1;
 829                         }
 830                         len = sprintf(tmp, "%s/", td->directory);
 831                 }
 832
 833                 td->files = malloc(sizeof(struct fio_file) * td->open_files);
 834
 835                 for_each_file(td, f, i) {
 836                         memset(f, 0, sizeof(*f));
 837                         f->fd = -1;
 838
 839                         if (td->filename)
 840                                 sprintf(tmp + len, "%s", td->filename);
 841                         else
 842                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 843                         f->file_name = strdup(tmp);
 844                 }
 845         } else {
 846                 td->open_files = td->nr_files = 1;
 847                 td->files = malloc(sizeof(struct fio_file));
 848                 f = &td->files[0];
 849
 850                 memset(f, 0, sizeof(*f));
 851                 f->fd = -1;
 852                 f->file_name = strdup(jobname);
 853         }
 854
 855         for_each_file(td, f, i) {
 856                 f->file_size = td->total_file_size / td->nr_files;
 857                 f->file_offset = td->start_offset;
 858         }
 859
 860         fio_sem_init(&td->mutex, 0);
 861
 862         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 863         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 864         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 865
 866         if (td->stonewall && td->thread_number > 1)
 867                 groupid++;
 868
 869         td->groupid = groupid;
 870
 871         if (setup_rate(td))
 872                 goto err;
 873
 874         if (td->write_lat_log) {
 875                 setup_log(&td->ts.slat_log);
 876                 setup_log(&td->ts.clat_log);
 877         }
 878         if (td->write_bw_log)
 879                 setup_log(&td->ts.bw_log);
 880
 881         if (!td->name)
 882                 td->name = strdup(jobname);
 883
 884         if (!terse_output) {
 885                 if (!job_add_num) {
 886                         if (td->io_ops->flags & FIO_CPUIO)
 887                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 888                         else {
 889                                 char *c1, *c2, *c3, *c4;
 890
 891                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 892                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 893                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 894                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 895
 896                                 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);
 897
 898                                 free(c1);
 899                                 free(c2);
 900                                 free(c3);
 901                                 free(c4);
 902                         }
 903                 } else if (job_add_num == 1)
 904                         fprintf(f_out, "...\n");
 905         }
 906
 907         /*
 908          * recurse add identical jobs, clear numjobs and stonewall options
 909          * as they don't apply to sub-jobs
 910          */
 911         numjobs = td->numjobs;
 912         while (--numjobs) {
 913                 struct thread_data *td_new = get_new_job(0, td);
 914
 915                 if (!td_new)
 916                         goto err;
 917
 918                 td_new->numjobs = 1;
 919                 td_new->stonewall = 0;
 920                 job_add_num = numjobs - 1;
 921
 922                 if (add_job(td_new, jobname, job_add_num))
 923                         goto err;
 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, i;
 939         struct fio_file *f;
 940
 941         if (td->io_ops->flags & FIO_CPUIO)
 942                 return 0;
 943
 944         fd = open("/dev/urandom", O_RDONLY);
 945         if (fd == -1) {
 946                 td_verror(td, errno, "open");
 947                 return 1;
 948         }
 949
 950         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 951                 td_verror(td, EIO, "read");
 952                 close(fd);
 953                 return 1;
 954         }
 955
 956         close(fd);
 957
 958         os_random_seed(seeds[0], &td->bsrange_state);
 959         os_random_seed(seeds[1], &td->verify_state);
 960         os_random_seed(seeds[2], &td->rwmix_state);
 961
 962         if (td->file_service_type == FIO_FSERVICE_RANDOM)
 963                 os_random_seed(seeds[3], &td->next_file_state);
 964
 965         if (!td_random(td))
 966                 return 0;
 967
 968         if (td->rand_repeatable)
 969                 seeds[4] = FIO_RANDSEED * td->thread_number;
 970
 971         if (!td->norandommap) {
 972                 for_each_file(td, f, i) {
 973                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 974                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 975                         f->file_map = malloc(num_maps * sizeof(long));
 976                         f->num_maps = num_maps;
 977                         memset(f->file_map, 0, num_maps * sizeof(long));
 978                 }
 979         }
 980
 981         os_random_seed(seeds[4], &td->random_state);
 982         return 0;
 983 }
 984
 985 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 986 {
 987 #ifdef FIO_HAVE_CPU_AFFINITY
 988         unsigned int i;
 989
 990         CPU_ZERO(&cpumask);
 991
 992         for (i = 0; i < sizeof(int) * 8; i++) {
 993                 if ((1 << i) & cpu)
 994                         CPU_SET(i, &cpumask);
 995         }
 996 #endif
 997 }
 998
 999 static int is_empty_or_comment(char *line)
1000 {
1001         unsigned int i;
1002
1003         for (i = 0; i < strlen(line); i++) {
1004                 if (line[i] == ';')
1005                         return 1;
1006                 if (line[i] == '#')
1007                         return 1;
1008                 if (!isspace(line[i]) && !iscntrl(line[i]))
1009                         return 0;
1010         }
1011
1012         return 1;
1013 }
1014
1015 /*
1016  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
1017  */
1018 static char *get_mmap_file(const char *str)
1019 {
1020         char *p = strstr(str, ":");
1021
1022         if (!p)
1023                 return NULL;
1024
1025         p++;
1026         strip_blank_front(&p);
1027         strip_blank_end(p);
1028         return strdup(p);
1029 }
1030
1031 static int str_mem_cb(void *data, const char *mem)
1032 {
1033         struct thread_data *td = data;
1034
1035         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1036                 td->mmapfile = get_mmap_file(mem);
1037                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1038                         log_err("fio: mmaphuge:/path/to/file\n");
1039                         return 1;
1040                 }
1041         }
1042
1043         return 0;
1044 }
1045
1046 /* External engines are specified by "external:name.o") */
1047 static const char *get_engine_name(const char *str)
1048 {
1049         char *p = strstr(str, ":");
1050
1051         if (!p)
1052                 return str;
1053
1054         p++;
1055         strip_blank_front(&p);
1056         strip_blank_end(p);
1057         return p;
1058 }
1059
1060 static int str_ioengine_cb(void *data, const char *str)
1061 {
1062         struct thread_data *td = data;
1063         const char *name = get_engine_name(str);
1064
1065         td->io_ops = load_ioengine(td, name);
1066         if (td->io_ops)
1067                 return 0;
1068
1069         return 1;
1070 }
1071
1072 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1073 {
1074         mlock_size = *val;
1075         return 0;
1076 }
1077
1078 #ifdef FIO_HAVE_IOPRIO
1079 static int str_prioclass_cb(void *data, unsigned int *val)
1080 {
1081         struct thread_data *td = data;
1082
1083         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1084         return 0;
1085 }
1086
1087 static int str_prio_cb(void *data, unsigned int *val)
1088 {
1089         struct thread_data *td = data;
1090
1091         td->ioprio |= *val;
1092         return 0;
1093 }
1094 #endif
1095
1096 static int str_exitall_cb(void)
1097 {
1098         exitall_on_terminate = 1;
1099         return 0;
1100 }
1101
1102 static int str_cpumask_cb(void *data, unsigned int *val)
1103 {
1104         struct thread_data *td = data;
1105
1106         fill_cpu_mask(td->cpumask, *val);
1107         return 0;
1108 }
1109
1110 /*
1111  * This is our [ini] type file parser.
1112  */
1113 static int parse_jobs_ini(char *file, int stonewall_flag)
1114 {
1115         unsigned int global;
1116         struct thread_data *td;
1117         char *string, *name;
1118         fpos_t off;
1119         FILE *f;
1120         char *p;
1121         int ret = 0, stonewall;
1122
1123         f = fopen(file, "r");
1124         if (!f) {
1125                 perror("fopen job file");
1126                 return 1;
1127         }
1128
1129         string = malloc(4096);
1130         name = malloc(256);
1131         memset(name, 0, 256);
1132
1133         stonewall = stonewall_flag;
1134         do {
1135                 p = fgets(string, 4095, f);
1136                 if (!p)
1137                         break;
1138                 if (is_empty_or_comment(p))
1139                         continue;
1140                 if (sscanf(p, "[%255s]", name) != 1)
1141                         continue;
1142
1143                 global = !strncmp(name, "global", 6);
1144
1145                 name[strlen(name) - 1] = '\0';
1146
1147                 td = get_new_job(global, &def_thread);
1148                 if (!td) {
1149                         ret = 1;
1150                         break;
1151                 }
1152
1153                 /*
1154                  * Seperate multiple job files by a stonewall
1155                  */
1156                 if (!global && stonewall) {
1157                         td->stonewall = stonewall;
1158                         stonewall = 0;
1159                 }
1160
1161                 fgetpos(f, &off);
1162                 while ((p = fgets(string, 4096, f)) != NULL) {
1163                         if (is_empty_or_comment(p))
1164                                 continue;
1165
1166                         strip_blank_front(&p);
1167
1168                         if (p[0] == '[')
1169                                 break;
1170
1171                         strip_blank_end(p);
1172
1173                         fgetpos(f, &off);
1174
1175                         /*
1176                          * Don't break here, continue parsing options so we
1177                          * dump all the bad ones. Makes trial/error fixups
1178                          * easier on the user.
1179                          */
1180                         ret |= parse_option(p, options, td);
1181                 }
1182
1183                 if (!ret) {
1184                         fsetpos(f, &off);
1185                         ret = add_job(td, name, 0);
1186                 } else {
1187                         log_err("fio: job %s dropped\n", name);
1188                         put_job(td);
1189                 }
1190         } while (!ret);
1191
1192         free(string);
1193         free(name);
1194         fclose(f);
1195         return ret;
1196 }
1197
1198 static int fill_def_thread(void)
1199 {
1200         memset(&def_thread, 0, sizeof(def_thread));
1201
1202         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1203                 perror("sched_getaffinity");
1204                 return 1;
1205         }
1206
1207         /*
1208          * fill default options
1209          */
1210         fill_default_options(&def_thread, options);
1211
1212         def_thread.timeout = def_timeout;
1213         def_thread.write_bw_log = write_bw_log;
1214         def_thread.write_lat_log = write_lat_log;
1215
1216 #ifdef FIO_HAVE_DISK_UTIL
1217         def_thread.do_disk_util = 1;
1218 #endif
1219
1220         return 0;
1221 }
1222
1223 static void usage(void)
1224 {
1225         printf("%s\n", fio_version_string);
1226         printf("\t--output\tWrite output to file\n");
1227         printf("\t--timeout\tRuntime in seconds\n");
1228         printf("\t--latency-log\tGenerate per-job latency logs\n");
1229         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1230         printf("\t--minimal\tMinimal (terse) output\n");
1231         printf("\t--version\tPrint version info and exit\n");
1232         printf("\t--help\t\tPrint this page\n");
1233         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1234 }
1235
1236 static int parse_cmd_line(int argc, char *argv[])
1237 {
1238         struct thread_data *td = NULL;
1239         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1240
1241         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
1242                 switch (c) {
1243                 case 't':
1244                         def_timeout = atoi(optarg);
1245                         break;
1246                 case 'l':
1247                         write_lat_log = 1;
1248                         break;
1249                 case 'w':
1250                         write_bw_log = 1;
1251                         break;
1252                 case 'o':
1253                         f_out = fopen(optarg, "w+");
1254                         if (!f_out) {
1255                                 perror("fopen output");
1256                                 exit(1);
1257                         }
1258                         f_err = f_out;
1259                         break;
1260                 case 'm':
1261                         terse_output = 1;
1262                         break;
1263                 case 'h':
1264                         usage();
1265                         exit(0);
1266                 case 'c':
1267                         ret = show_cmd_help(options, optarg);
1268                         exit(ret);
1269                 case 'v':
1270                         printf("%s\n", fio_version_string);
1271                         exit(0);
1272                 case FIO_GETOPT_JOB: {
1273                         const char *opt = long_options[lidx].name;
1274                         char *val = optarg;
1275
1276                         if (!strncmp(opt, "name", 4) && td) {
1277                                 ret = add_job(td, td->name ?: "fio", 0);
1278                                 if (ret) {
1279                                         put_job(td);
1280                                         return 0;
1281                                 }
1282                                 td = NULL;
1283                         }
1284                         if (!td) {
1285                                 int global = !strncmp(val, "global", 6);
1286
1287                                 td = get_new_job(global, &def_thread);
1288                                 if (!td)
1289                                         return 0;
1290                         }
1291
1292                         ret = parse_cmd_option(opt, val, options, td);
1293                         if (ret)
1294                                 dont_add_job = 1;
1295                         break;
1296                 }
1297                 default:
1298                         break;
1299                 }
1300         }
1301
1302         if (td) {
1303                 if (dont_add_job)
1304                         put_job(td);
1305                 else {
1306                         ret = add_job(td, td->name ?: "fio", 0);
1307                         if (ret)
1308                                 put_job(td);
1309                 }
1310         }
1311
1312         while (optind < argc) {
1313                 ini_idx++;
1314                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1315                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1316                 optind++;
1317         }
1318
1319         return ini_idx;
1320 }
1321
1322 static void free_shm(void)
1323 {
1324         struct shmid_ds sbuf;
1325
1326         if (threads) {
1327                 shmdt((void *) threads);
1328                 threads = NULL;
1329                 shmctl(shm_id, IPC_RMID, &sbuf);
1330         }
1331 }
1332
1333 /*
1334  * The thread area is shared between the main process and the job
1335  * threads/processes. So setup a shared memory segment that will hold
1336  * all the job info.
1337  */
1338 static int setup_thread_area(void)
1339 {
1340         /*
1341          * 1024 is too much on some machines, scale max_jobs if
1342          * we get a failure that looks like too large a shm segment
1343          */
1344         do {
1345                 size_t size = max_jobs * sizeof(struct thread_data);
1346
1347                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1348                 if (shm_id != -1)
1349                         break;
1350                 if (errno != EINVAL) {
1351                         perror("shmget");
1352                         break;
1353                 }
1354
1355                 max_jobs >>= 1;
1356         } while (max_jobs);
1357
1358         if (shm_id == -1)
1359                 return 1;
1360
1361         threads = shmat(shm_id, NULL, 0);
1362         if (threads == (void *) -1) {
1363                 perror("shmat");
1364                 return 1;
1365         }
1366
1367         atexit(free_shm);
1368         return 0;
1369 }
1370
1371 /*
1372  * Copy the fio options into the long options map, so we mirror
1373  * job and cmd line options.
1374  */
1375 static void dupe_job_options(void)
1376 {
1377         struct fio_option *o;
1378         unsigned int i;
1379
1380         i = 0;
1381         while (long_options[i].name)
1382                 i++;
1383
1384         o = &options[0];
1385         while (o->name) {
1386                 long_options[i].name = o->name;
1387                 long_options[i].val = FIO_GETOPT_JOB;
1388                 if (o->type == FIO_OPT_STR_SET)
1389                         long_options[i].has_arg = no_argument;
1390                 else
1391                         long_options[i].has_arg = required_argument;
1392
1393                 i++;
1394                 o++;
1395                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1396         }
1397 }
1398
1399 int parse_options(int argc, char *argv[])
1400 {
1401         int job_files, i;
1402
1403         f_out = stdout;
1404         f_err = stderr;
1405
1406         options_init(options);
1407
1408         dupe_job_options();
1409
1410         if (setup_thread_area())
1411                 return 1;
1412         if (fill_def_thread())
1413                 return 1;
1414
1415         job_files = parse_cmd_line(argc, argv);
1416
1417         for (i = 0; i < job_files; i++) {
1418                 if (fill_def_thread())
1419                         return 1;
1420                 if (parse_jobs_ini(ini_file[i], i))
1421                         return 1;
1422                 free(ini_file[i]);
1423         }
1424
1425         free(ini_file);
1426
1427         if (!thread_number) {
1428                 log_err("No jobs defined(s)\n");
1429                 return 1;
1430         }
1431
1432         return 0;
1433 }