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