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", }, { .ival = "libaio", },
  89                           { .ival = "posixaio", }, { .ival = "mmap", },
  90                           { .ival = "splice", }, { .ival = "sg", },
  91                           { .ival = "null", }, { .ival = "net", },
  92                           { .ival = "syslet-rw", },
  93                           },
  94         },
  95         {
  96                 .name   = "iodepth",
  97                 .type   = FIO_OPT_INT,
  98                 .off1   = td_var_offset(iodepth),
  99                 .help   = "Amount of IO buffers to keep in flight",
 100                 .def    = "1",
 101         },
 102         {
 103                 .name   = "iodepth_low",
 104                 .type   = FIO_OPT_INT,
 105                 .off1   = td_var_offset(iodepth_low),
 106                 .help   = "Low water mark for queuing depth",
 107         },
 108         {
 109                 .name   = "size",
 110                 .type   = FIO_OPT_STR_VAL,
 111                 .off1   = td_var_offset(total_file_size),
 112                 .help   = "Size of device or file",
 113         },
 114         {
 115                 .name   = "bs",
 116                 .type   = FIO_OPT_STR_VAL_INT,
 117                 .off1   = td_var_offset(bs[DDIR_READ]),
 118                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 119                 .help   = "Block size unit",
 120                 .def    = "4k",
 121         },
 122         {
 123                 .name   = "bsrange",
 124                 .type   = FIO_OPT_RANGE,
 125                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 126                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 127                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 128                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 129                 .help   = "Set block size range (in more detail than bs)",
 130         },
 131         {
 132                 .name   = "bs_unaligned",
 133                 .type   = FIO_OPT_STR_SET,
 134                 .off1   = td_var_offset(bs_unaligned),
 135                 .help   = "Don't sector align IO buffer sizes",
 136         },
 137         {
 138                 .name   = "offset",
 139                 .type   = FIO_OPT_STR_VAL,
 140                 .off1   = td_var_offset(start_offset),
 141                 .help   = "Start IO from this offset",
 142                 .def    = "0",
 143         },
 144         {
 145                 .name   = "randrepeat",
 146                 .type   = FIO_OPT_BOOL,
 147                 .off1   = td_var_offset(rand_repeatable),
 148                 .help   = "Use repeatable random IO pattern",
 149                 .def    = "1",
 150         },
 151         {
 152                 .name   = "norandommap",
 153                 .type   = FIO_OPT_STR_SET,
 154                 .off1   = td_var_offset(norandommap),
 155                 .help   = "Accept potential duplicate random blocks",
 156         },
 157         {
 158                 .name   = "nrfiles",
 159                 .type   = FIO_OPT_INT,
 160                 .off1   = td_var_offset(nr_files),
 161                 .help   = "Split job workload between this number of files",
 162                 .def    = "1",
 163         },
 164         {
 165                 .name   = "file_service_type",
 166                 .type   = FIO_OPT_STR,
 167                 .off1   = td_var_offset(file_service_type),
 168                 .help   = "How to select which file to service next",
 169                 .def    = "roundrobin",
 170                 .posval = {
 171                           { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
 172                           { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
 173                           },
 174         },
 175         {
 176                 .name   = "fsync",
 177                 .type   = FIO_OPT_INT,
 178                 .off1   = td_var_offset(fsync_blocks),
 179                 .help   = "Issue fsync for writes every given number of blocks",
 180                 .def    = "0",
 181         },
 182         {
 183                 .name   = "direct",
 184                 .type   = FIO_OPT_BOOL,
 185                 .off1   = td_var_offset(odirect),
 186                 .help   = "Use O_DIRECT IO (negates buffered)",
 187                 .def    = "0",
 188         },
 189         {
 190                 .name   = "buffered",
 191                 .type   = FIO_OPT_BOOL,
 192                 .off1   = td_var_offset(odirect),
 193                 .neg    = 1,
 194                 .help   = "Use buffered IO (negates direct)",
 195                 .def    = "1",
 196         },
 197         {
 198                 .name   = "overwrite",
 199                 .type   = FIO_OPT_BOOL,
 200                 .off1   = td_var_offset(overwrite),
 201                 .help   = "When writing, set whether to overwrite current data",
 202                 .def    = "0",
 203         },
 204         {
 205                 .name   = "loops",
 206                 .type   = FIO_OPT_INT,
 207                 .off1   = td_var_offset(loops),
 208                 .help   = "Number of times to run the job",
 209                 .def    = "1",
 210         },
 211         {
 212                 .name   = "numjobs",
 213                 .type   = FIO_OPT_INT,
 214                 .off1   = td_var_offset(numjobs),
 215                 .help   = "Duplicate this job this many times",
 216                 .def    = "1",
 217         },
 218         {
 219                 .name   = "startdelay",
 220                 .type   = FIO_OPT_INT,
 221                 .off1   = td_var_offset(start_delay),
 222                 .help   = "Only start job when this period has passed",
 223                 .def    = "0",
 224         },
 225         {
 226                 .name   = "runtime",
 227                 .alias  = "timeout",
 228                 .type   = FIO_OPT_STR_VAL_TIME,
 229                 .off1   = td_var_offset(timeout),
 230                 .help   = "Stop workload when this amount of time has passed",
 231                 .def    = "0",
 232         },
 233         {
 234                 .name   = "mem",
 235                 .type   = FIO_OPT_STR,
 236                 .cb     = str_mem_cb,
 237                 .off1   = td_var_offset(mem_type),
 238                 .help   = "Backing type for IO buffers",
 239                 .def    = "malloc",
 240                 .posval = {
 241                           { .ival = "malloc", .oval = MEM_MALLOC },
 242                           { .ival = "shm", .oval = MEM_SHM },
 243 #ifdef FIO_HAVE_HUGETLB
 244                           { .ival = "shmhuge", .oval = MEM_SHMHUGE },
 245 #endif
 246                           { .ival = "mmap", .oval = MEM_MMAP },
 247 #ifdef FIO_HAVE_HUGETLB
 248                           { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
 249 #endif
 250                           },
 251         },
 252         {
 253                 .name   = "verify",
 254                 .type   = FIO_OPT_STR,
 255                 .off1   = td_var_offset(verify),
 256                 .help   = "Verify sum function",
 257                 .def    = "0",
 258                 .posval = {
 259                           { .ival = "0", .oval = VERIFY_NONE },
 260                           { .ival = "crc32", .oval = VERIFY_CRC32 },
 261                           { .ival = "md5", .oval = VERIFY_MD5 },
 262                           },
 263         },
 264         {
 265                 .name   = "write_iolog",
 266                 .type   = FIO_OPT_STR_STORE,
 267                 .off1   = td_var_offset(write_iolog_file),
 268                 .help   = "Store IO pattern to file",
 269         },
 270         {
 271                 .name   = "read_iolog",
 272                 .type   = FIO_OPT_STR_STORE,
 273                 .off1   = td_var_offset(read_iolog_file),
 274                 .help   = "Playback IO pattern from file",
 275         },
 276         {
 277                 .name   = "exec_prerun",
 278                 .type   = FIO_OPT_STR_STORE,
 279                 .off1   = td_var_offset(exec_prerun),
 280                 .help   = "Execute this file prior to running job",
 281         },
 282         {
 283                 .name   = "exec_postrun",
 284                 .type   = FIO_OPT_STR_STORE,
 285                 .off1   = td_var_offset(exec_postrun),
 286                 .help   = "Execute this file after running job",
 287         },
 288 #ifdef FIO_HAVE_IOSCHED_SWITCH
 289         {
 290                 .name   = "ioscheduler",
 291                 .type   = FIO_OPT_STR_STORE,
 292                 .off1   = td_var_offset(ioscheduler),
 293                 .help   = "Use this IO scheduler on the backing device",
 294         },
 295 #endif
 296         {
 297                 .name   = "zonesize",
 298                 .type   = FIO_OPT_STR_VAL,
 299                 .off1   = td_var_offset(zone_size),
 300                 .help   = "Give size of an IO zone",
 301                 .def    = "0",
 302         },
 303         {
 304                 .name   = "zoneskip",
 305                 .type   = FIO_OPT_STR_VAL,
 306                 .off1   = td_var_offset(zone_skip),
 307                 .help   = "Space between IO zones",
 308                 .def    = "0",
 309         },
 310         {
 311                 .name   = "lockmem",
 312                 .type   = FIO_OPT_STR_VAL,
 313                 .cb     = str_lockmem_cb,
 314                 .help   = "Lock down this amount of memory",
 315                 .def    = "0",
 316         },
 317         {
 318                 .name   = "rwmixcycle",
 319                 .type   = FIO_OPT_INT,
 320                 .off1   = td_var_offset(rwmixcycle),
 321                 .help   = "Cycle period for mixed read/write workloads (msec)",
 322                 .def    = "500",
 323         },
 324         {
 325                 .name   = "rwmixread",
 326                 .type   = FIO_OPT_INT,
 327                 .off1   = td_var_offset(rwmixread),
 328                 .maxval = 100,
 329                 .help   = "Percentage of mixed workload that is reads",
 330                 .def    = "50",
 331         },
 332         {
 333                 .name   = "rwmixwrite",
 334                 .type   = FIO_OPT_INT,
 335                 .off1   = td_var_offset(rwmixwrite),
 336                 .maxval = 100,
 337                 .help   = "Percentage of mixed workload that is writes",
 338                 .def    = "50",
 339         },
 340         {
 341                 .name   = "nice",
 342                 .type   = FIO_OPT_INT,
 343                 .off1   = td_var_offset(nice),
 344                 .help   = "Set job CPU nice value",
 345                 .minval = -19,
 346                 .maxval = 20,
 347                 .def    = "0",
 348         },
 349 #ifdef FIO_HAVE_IOPRIO
 350         {
 351                 .name   = "prio",
 352                 .type   = FIO_OPT_INT,
 353                 .cb     = str_prio_cb,
 354                 .help   = "Set job IO priority value",
 355                 .minval = 0,
 356                 .maxval = 7,
 357         },
 358         {
 359                 .name   = "prioclass",
 360                 .type   = FIO_OPT_INT,
 361                 .cb     = str_prioclass_cb,
 362                 .help   = "Set job IO priority class",
 363                 .minval = 0,
 364                 .maxval = 3,
 365         },
 366 #endif
 367         {
 368                 .name   = "thinktime",
 369                 .type   = FIO_OPT_INT,
 370                 .off1   = td_var_offset(thinktime),
 371                 .help   = "Idle time between IO buffers (usec)",
 372                 .def    = "0",
 373         },
 374         {
 375                 .name   = "thinktime_spin",
 376                 .type   = FIO_OPT_INT,
 377                 .off1   = td_var_offset(thinktime_spin),
 378                 .help   = "Start thinktime by spinning this amount (usec)",
 379                 .def    = "0",
 380         },
 381         {
 382                 .name   = "thinktime_blocks",
 383                 .type   = FIO_OPT_INT,
 384                 .off1   = td_var_offset(thinktime_blocks),
 385                 .help   = "IO buffer period between 'thinktime'",
 386                 .def    = "1",
 387         },
 388         {
 389                 .name   = "rate",
 390                 .type   = FIO_OPT_INT,
 391                 .off1   = td_var_offset(rate),
 392                 .help   = "Set bandwidth rate",
 393         },
 394         {
 395                 .name   = "ratemin",
 396                 .type   = FIO_OPT_INT,
 397                 .off1   = td_var_offset(ratemin),
 398                 .help   = "The bottom limit accepted",
 399         },
 400         {
 401                 .name   = "ratecycle",
 402                 .type   = FIO_OPT_INT,
 403                 .off1   = td_var_offset(ratecycle),
 404                 .help   = "Window average for rate limits (msec)",
 405                 .def    = "1000",
 406         },
 407         {
 408                 .name   = "invalidate",
 409                 .type   = FIO_OPT_BOOL,
 410                 .off1   = td_var_offset(invalidate_cache),
 411                 .help   = "Invalidate buffer/page cache prior to running job",
 412                 .def    = "1",
 413         },
 414         {
 415                 .name   = "sync",
 416                 .type   = FIO_OPT_BOOL,
 417                 .off1   = td_var_offset(sync_io),
 418                 .help   = "Use O_SYNC for buffered writes",
 419                 .def    = "0",
 420         },
 421         {
 422                 .name   = "bwavgtime",
 423                 .type   = FIO_OPT_INT,
 424                 .off1   = td_var_offset(bw_avg_time),
 425                 .help   = "Time window over which to calculate bandwidth (msec)",
 426                 .def    = "500",
 427         },
 428         {
 429                 .name   = "create_serialize",
 430                 .type   = FIO_OPT_BOOL,
 431                 .off1   = td_var_offset(create_serialize),
 432                 .help   = "Serialize creating of job files",
 433                 .def    = "1",
 434         },
 435         {
 436                 .name   = "create_fsync",
 437                 .type   = FIO_OPT_BOOL,
 438                 .off1   = td_var_offset(create_fsync),
 439                 .help   = "Fsync file after creation",
 440                 .def    = "1",
 441         },
 442         {
 443                 .name   = "cpuload",
 444                 .type   = FIO_OPT_INT,
 445                 .off1   = td_var_offset(cpuload),
 446                 .help   = "Use this percentage of CPU",
 447         },
 448         {
 449                 .name   = "cpuchunks",
 450                 .type   = FIO_OPT_INT,
 451                 .off1   = td_var_offset(cpucycle),
 452                 .help   = "Length of the CPU burn cycles",
 453         },
 454 #ifdef FIO_HAVE_CPU_AFFINITY
 455         {
 456                 .name   = "cpumask",
 457                 .type   = FIO_OPT_INT,
 458                 .cb     = str_cpumask_cb,
 459                 .help   = "CPU affinity mask",
 460         },
 461 #endif
 462         {
 463                 .name   = "end_fsync",
 464                 .type   = FIO_OPT_BOOL,
 465                 .off1   = td_var_offset(end_fsync),
 466                 .help   = "Include fsync at the end of job",
 467                 .def    = "0",
 468         },
 469         {
 470                 .name   = "unlink",
 471                 .type   = FIO_OPT_BOOL,
 472                 .off1   = td_var_offset(unlink),
 473                 .help   = "Unlink created files after job has completed",
 474                 .def    = "0",
 475         },
 476         {
 477                 .name   = "exitall",
 478                 .type   = FIO_OPT_STR_SET,
 479                 .cb     = str_exitall_cb,
 480                 .help   = "Terminate all jobs when one exits",
 481         },
 482         {
 483                 .name   = "stonewall",
 484                 .type   = FIO_OPT_STR_SET,
 485                 .off1   = td_var_offset(stonewall),
 486                 .help   = "Insert a hard barrier between this job and previous",
 487         },
 488         {
 489                 .name   = "thread",
 490                 .type   = FIO_OPT_STR_SET,
 491                 .off1   = td_var_offset(use_thread),
 492                 .help   = "Use threads instead of forks",
 493         },
 494         {
 495                 .name   = "write_bw_log",
 496                 .type   = FIO_OPT_STR_SET,
 497                 .off1   = td_var_offset(write_bw_log),
 498                 .help   = "Write log of bandwidth during run",
 499         },
 500         {
 501                 .name   = "write_lat_log",
 502                 .type   = FIO_OPT_STR_SET,
 503                 .off1   = td_var_offset(write_lat_log),
 504                 .help   = "Write log of latency during run",
 505         },
 506         {
 507                 .name   = "hugepage-size",
 508                 .type   = FIO_OPT_STR_VAL,
 509                 .off1   = td_var_offset(hugepage_size),
 510                 .help   = "When using hugepages, specify size of each page",
 511                 .def    = __stringify(FIO_HUGE_PAGE),
 512         },
 513         {
 514                 .name = NULL,
 515         },
 516 };
 517
 518 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 519 #define FIO_CMD_OPTS    (16)
 520 #define FIO_GETOPT_JOB  (0x89988998)
 521
 522 /*
 523  * Command line options. These will contain the above, plus a few
 524  * extra that only pertain to fio itself and not jobs.
 525  */
 526 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 527         {
 528                 .name           = "output",
 529                 .has_arg        = required_argument,
 530                 .val            = 'o',
 531         },
 532         {
 533                 .name           = "timeout",
 534                 .has_arg        = required_argument,
 535                 .val            = 't',
 536         },
 537         {
 538                 .name           = "latency-log",
 539                 .has_arg        = required_argument,
 540                 .val            = 'l',
 541         },
 542         {
 543                 .name           = "bandwidth-log",
 544                 .has_arg        = required_argument,
 545                 .val            = 'b',
 546         },
 547         {
 548                 .name           = "minimal",
 549                 .has_arg        = optional_argument,
 550                 .val            = 'm',
 551         },
 552         {
 553                 .name           = "version",
 554                 .has_arg        = no_argument,
 555                 .val            = 'v',
 556         },
 557         {
 558                 .name           = "help",
 559                 .has_arg        = no_argument,
 560                 .val            = 'h',
 561         },
 562         {
 563                 .name           = "cmdhelp",
 564                 .has_arg        = required_argument,
 565                 .val            = 'c',
 566         },
 567         {
 568                 .name           = NULL,
 569         },
 570 };
 571
 572 static int def_timeout = 0;
 573
 574 static char fio_version_string[] = "fio 1.11";
 575
 576 static char **ini_file;
 577 static int max_jobs = MAX_JOBS;
 578
 579 struct thread_data def_thread;
 580 struct thread_data *threads = NULL;
 581
 582 int exitall_on_terminate = 0;
 583 int terse_output = 0;
 584 unsigned long long mlock_size = 0;
 585 FILE *f_out = NULL;
 586 FILE *f_err = NULL;
 587
 588 static int write_lat_log = 0;
 589 int write_bw_log = 0;
 590
 591 /*
 592  * Return a free job structure.
 593  */
 594 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 595 {
 596         struct thread_data *td;
 597
 598         if (global)
 599                 return &def_thread;
 600         if (thread_number >= max_jobs)
 601                 return NULL;
 602
 603         td = &threads[thread_number++];
 604         *td = *parent;
 605
 606         td->thread_number = thread_number;
 607         return td;
 608 }
 609
 610 static void put_job(struct thread_data *td)
 611 {
 612         if (td == &def_thread)
 613                 return;
 614
 615         if (td->error)
 616                 fprintf(f_out, "fio: %s\n", td->verror);
 617
 618         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 619         thread_number--;
 620 }
 621
 622 /*
 623  * Lazy way of fixing up options that depend on each other. We could also
 624  * define option callback handlers, but this is easier.
 625  */
 626 static void fixup_options(struct thread_data *td)
 627 {
 628         if (!td->rwmixread && td->rwmixwrite)
 629                 td->rwmixread = 100 - td->rwmixwrite;
 630
 631         if (td->write_iolog_file && td->read_iolog_file) {
 632                 log_err("fio: read iolog overrides write_iolog\n");
 633                 free(td->write_iolog_file);
 634                 td->write_iolog_file = NULL;
 635         }
 636
 637         if (td->io_ops->flags & FIO_SYNCIO)
 638                 td->iodepth = 1;
 639         else {
 640                 if (!td->iodepth)
 641                         td->iodepth = td->nr_files;
 642         }
 643
 644         /*
 645          * only really works for sequential io for now, and with 1 file
 646          */
 647         if (td->zone_size && td_random(td) && td->nr_files == 1)
 648                 td->zone_size = 0;
 649
 650         /*
 651          * Reads can do overwrites, we always need to pre-create the file
 652          */
 653         if (td_read(td) || td_rw(td))
 654                 td->overwrite = 1;
 655
 656         if (!td->min_bs[DDIR_READ])
 657                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 658         if (!td->max_bs[DDIR_READ])
 659                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 660         if (!td->min_bs[DDIR_WRITE])
 661                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 662         if (!td->max_bs[DDIR_WRITE])
 663                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 664
 665         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 666
 667         if (td_read(td) && !td_rw(td))
 668                 td->verify = 0;
 669
 670         if (td->norandommap && td->verify != VERIFY_NONE) {
 671                 log_err("fio: norandommap given, verify disabled\n");
 672                 td->verify = VERIFY_NONE;
 673         }
 674         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 675                 log_err("fio: bs_unaligned may not work with raw io\n");
 676
 677         /*
 678          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 679          */
 680         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 681                 td->odirect = 0;
 682
 683         /*
 684          * thinktime_spin must be less than thinktime
 685          */
 686         if (td->thinktime_spin > td->thinktime)
 687                 td->thinktime_spin = td->thinktime;
 688
 689         /*
 690          * The low water mark cannot be bigger than the iodepth
 691          */
 692         if (td->iodepth_low > td->iodepth || !td->iodepth_low)
 693                 td->iodepth_low = td->iodepth;
 694 }
 695
 696 /*
 697  * This function leaks the buffer
 698  */
 699 static char *to_kmg(unsigned int val)
 700 {
 701         char *buf = malloc(32);
 702         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 703         char *p = post;
 704
 705         do {
 706                 if (val & 1023)
 707                         break;
 708
 709                 val >>= 10;
 710                 p++;
 711         } while (*p);
 712
 713         snprintf(buf, 31, "%u%c", val, *p);
 714         return buf;
 715 }
 716
 717 /*
 718  * Adds a job to the list of things todo. Sanitizes the various options
 719  * to make sure we don't have conflicts, and initializes various
 720  * members of td.
 721  */
 722 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 723 {
 724         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 725                                    "randread", "randwrite", "randrw" };
 726         struct stat sb;
 727         int numjobs, i;
 728         struct fio_file *f;
 729
 730         /*
 731          * the def_thread is just for options, it's not a real job
 732          */
 733         if (td == &def_thread)
 734                 return 0;
 735
 736         assert(td->io_ops);
 737
 738         if (td->odirect)
 739                 td->io_ops->flags |= FIO_RAWIO;
 740
 741         td->filetype = FIO_TYPE_FILE;
 742         if (td->filename && !lstat(td->filename, &sb)) {
 743                 if (S_ISBLK(sb.st_mode))
 744                         td->filetype = FIO_TYPE_BD;
 745                 else if (S_ISCHR(sb.st_mode))
 746                         td->filetype = FIO_TYPE_CHAR;
 747         }
 748
 749         fixup_options(td);
 750
 751         if (td->filename)
 752                 td->nr_uniq_files = 1;
 753         else
 754                 td->nr_uniq_files = td->nr_files;
 755
 756         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 757                 char tmp[PATH_MAX];
 758                 int len = 0;
 759
 760                 if (td->directory && td->directory[0] != '\0') {
 761                         if (lstat(td->directory, &sb) < 0) {
 762                                 log_err("fio: %s is not a directory\n", td->directory);
 763                                 td_verror(td, errno, "lstat");
 764                                 return 1;
 765                         }
 766                         if (!S_ISDIR(sb.st_mode)) {
 767                                 log_err("fio: %s is not a directory\n", td->directory);
 768                                 return 1;
 769                         }
 770                         len = sprintf(tmp, "%s/", td->directory);
 771                 }
 772
 773                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 774
 775                 for_each_file(td, f, i) {
 776                         memset(f, 0, sizeof(*f));
 777                         f->fd = -1;
 778
 779                         if (td->filename)
 780                                 sprintf(tmp + len, "%s", td->filename);
 781                         else
 782                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 783                         f->file_name = strdup(tmp);
 784                 }
 785         } else {
 786                 td->nr_files = 1;
 787                 td->files = malloc(sizeof(struct fio_file));
 788                 f = &td->files[0];
 789
 790                 memset(f, 0, sizeof(*f));
 791                 f->fd = -1;
 792                 f->file_name = strdup(jobname);
 793         }
 794
 795         for_each_file(td, f, i) {
 796                 f->file_size = td->total_file_size / td->nr_files;
 797                 f->file_offset = td->start_offset;
 798         }
 799
 800         fio_sem_init(&td->mutex, 0);
 801
 802         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 803         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 804         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 805
 806         if (td->stonewall && td->thread_number > 1)
 807                 groupid++;
 808
 809         td->groupid = groupid;
 810
 811         if (setup_rate(td))
 812                 goto err;
 813
 814         if (td->write_lat_log) {
 815                 setup_log(&td->ts.slat_log);
 816                 setup_log(&td->ts.clat_log);
 817         }
 818         if (td->write_bw_log)
 819                 setup_log(&td->ts.bw_log);
 820
 821         if (!td->name)
 822                 td->name = strdup(jobname);
 823
 824         if (!terse_output) {
 825                 if (!job_add_num) {
 826                         if (td->io_ops->flags & FIO_CPUIO)
 827                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 828                         else {
 829                                 char *c1, *c2, *c3, *c4;
 830
 831                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 832                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 833                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 834                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 835
 836                                 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);
 837
 838                                 free(c1);
 839                                 free(c2);
 840                                 free(c3);
 841                                 free(c4);
 842                         }
 843                 } else if (job_add_num == 1)
 844                         fprintf(f_out, "...\n");
 845         }
 846
 847         /*
 848          * recurse add identical jobs, clear numjobs and stonewall options
 849          * as they don't apply to sub-jobs
 850          */
 851         numjobs = td->numjobs;
 852         while (--numjobs) {
 853                 struct thread_data *td_new = get_new_job(0, td);
 854
 855                 if (!td_new)
 856                         goto err;
 857
 858                 td_new->numjobs = 1;
 859                 td_new->stonewall = 0;
 860                 job_add_num = numjobs - 1;
 861
 862                 if (add_job(td_new, jobname, job_add_num))
 863                         goto err;
 864         }
 865         return 0;
 866 err:
 867         put_job(td);
 868         return -1;
 869 }
 870
 871 /*
 872  * Initialize the various random states we need (random io, block size ranges,
 873  * read/write mix, etc).
 874  */
 875 int init_random_state(struct thread_data *td)
 876 {
 877         unsigned long seeds[5];
 878         int fd, num_maps, blocks, i;
 879         struct fio_file *f;
 880
 881         if (td->io_ops->flags & FIO_CPUIO)
 882                 return 0;
 883
 884         fd = open("/dev/urandom", O_RDONLY);
 885         if (fd == -1) {
 886                 td_verror(td, errno, "open");
 887                 return 1;
 888         }
 889
 890         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 891                 td_verror(td, EIO, "read");
 892                 close(fd);
 893                 return 1;
 894         }
 895
 896         close(fd);
 897
 898         os_random_seed(seeds[0], &td->bsrange_state);
 899         os_random_seed(seeds[1], &td->verify_state);
 900         os_random_seed(seeds[2], &td->rwmix_state);
 901
 902         if (td->file_service_type == FIO_FSERVICE_RANDOM)
 903                 os_random_seed(seeds[3], &td->next_file_state);
 904
 905         if (!td_random(td))
 906                 return 0;
 907
 908         if (td->rand_repeatable)
 909                 seeds[4] = FIO_RANDSEED * td->thread_number;
 910
 911         if (!td->norandommap) {
 912                 for_each_file(td, f, i) {
 913                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 914                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 915                         f->file_map = malloc(num_maps * sizeof(long));
 916                         f->num_maps = num_maps;
 917                         memset(f->file_map, 0, num_maps * sizeof(long));
 918                 }
 919         }
 920
 921         os_random_seed(seeds[4], &td->random_state);
 922         return 0;
 923 }
 924
 925 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 926 {
 927 #ifdef FIO_HAVE_CPU_AFFINITY
 928         unsigned int i;
 929
 930         CPU_ZERO(&cpumask);
 931
 932         for (i = 0; i < sizeof(int) * 8; i++) {
 933                 if ((1 << i) & cpu)
 934                         CPU_SET(i, &cpumask);
 935         }
 936 #endif
 937 }
 938
 939 static int is_empty_or_comment(char *line)
 940 {
 941         unsigned int i;
 942
 943         for (i = 0; i < strlen(line); i++) {
 944                 if (line[i] == ';')
 945                         return 1;
 946                 if (line[i] == '#')
 947                         return 1;
 948                 if (!isspace(line[i]) && !iscntrl(line[i]))
 949                         return 0;
 950         }
 951
 952         return 1;
 953 }
 954
 955 /*
 956  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
 957  */
 958 static char *get_mmap_file(const char *str)
 959 {
 960         char *p = strstr(str, ":");
 961
 962         if (!p)
 963                 return NULL;
 964
 965         p++;
 966         strip_blank_front(&p);
 967         strip_blank_end(p);
 968         return strdup(p);
 969 }
 970
 971 static int str_mem_cb(void *data, const char *mem)
 972 {
 973         struct thread_data *td = data;
 974
 975         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
 976                 td->mmapfile = get_mmap_file(mem);
 977                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
 978                         log_err("fio: mmaphuge:/path/to/file\n");
 979                         return 1;
 980                 }
 981         }
 982
 983         return 0;
 984 }
 985
 986 static int str_ioengine_cb(void *data, const char *str)
 987 {
 988         struct thread_data *td = data;
 989
 990         td->io_ops = load_ioengine(td, str);
 991         if (td->io_ops)
 992                 return 0;
 993
 994         log_err("fio: ioengine= libaio, posixaio, sync, syslet-rw, mmap, sgio, splice, cpu, null\n");
 995         log_err("fio: or specify path to dynamic ioengine module\n");
 996         return 1;
 997 }
 998
 999 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1000 {
1001         mlock_size = *val;
1002         return 0;
1003 }
1004
1005 #ifdef FIO_HAVE_IOPRIO
1006 static int str_prioclass_cb(void *data, unsigned int *val)
1007 {
1008         struct thread_data *td = data;
1009
1010         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1011         return 0;
1012 }
1013
1014 static int str_prio_cb(void *data, unsigned int *val)
1015 {
1016         struct thread_data *td = data;
1017
1018         td->ioprio |= *val;
1019         return 0;
1020 }
1021 #endif
1022
1023 static int str_exitall_cb(void)
1024 {
1025         exitall_on_terminate = 1;
1026         return 0;
1027 }
1028
1029 static int str_cpumask_cb(void *data, unsigned int *val)
1030 {
1031         struct thread_data *td = data;
1032
1033         fill_cpu_mask(td->cpumask, *val);
1034         return 0;
1035 }
1036
1037 /*
1038  * This is our [ini] type file parser.
1039  */
1040 static int parse_jobs_ini(char *file, int stonewall_flag)
1041 {
1042         unsigned int global;
1043         struct thread_data *td;
1044         char *string, *name;
1045         fpos_t off;
1046         FILE *f;
1047         char *p;
1048         int ret = 0, stonewall;
1049
1050         f = fopen(file, "r");
1051         if (!f) {
1052                 perror("fopen job file");
1053                 return 1;
1054         }
1055
1056         string = malloc(4096);
1057         name = malloc(256);
1058         memset(name, 0, 256);
1059
1060         stonewall = stonewall_flag;
1061         do {
1062                 p = fgets(string, 4095, f);
1063                 if (!p)
1064                         break;
1065                 if (is_empty_or_comment(p))
1066                         continue;
1067                 if (sscanf(p, "[%255s]", name) != 1)
1068                         continue;
1069
1070                 global = !strncmp(name, "global", 6);
1071
1072                 name[strlen(name) - 1] = '\0';
1073
1074                 td = get_new_job(global, &def_thread);
1075                 if (!td) {
1076                         ret = 1;
1077                         break;
1078                 }
1079
1080                 /*
1081                  * Seperate multiple job files by a stonewall
1082                  */
1083                 if (!global && stonewall) {
1084                         td->stonewall = stonewall;
1085                         stonewall = 0;
1086                 }
1087
1088                 fgetpos(f, &off);
1089                 while ((p = fgets(string, 4096, f)) != NULL) {
1090                         if (is_empty_or_comment(p))
1091                                 continue;
1092
1093                         strip_blank_front(&p);
1094
1095                         if (p[0] == '[')
1096                                 break;
1097
1098                         strip_blank_end(p);
1099
1100                         fgetpos(f, &off);
1101
1102                         /*
1103                          * Don't break here, continue parsing options so we
1104                          * dump all the bad ones. Makes trial/error fixups
1105                          * easier on the user.
1106                          */
1107                         ret |= parse_option(p, options, td);
1108                 }
1109
1110                 if (!ret) {
1111                         fsetpos(f, &off);
1112                         ret = add_job(td, name, 0);
1113                 } else {
1114                         log_err("fio: job %s dropped\n", name);
1115                         put_job(td);
1116                 }
1117         } while (!ret);
1118
1119         free(string);
1120         free(name);
1121         fclose(f);
1122         return ret;
1123 }
1124
1125 static int fill_def_thread(void)
1126 {
1127         memset(&def_thread, 0, sizeof(def_thread));
1128
1129         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1130                 perror("sched_getaffinity");
1131                 return 1;
1132         }
1133
1134         /*
1135          * fill default options
1136          */
1137         fill_default_options(&def_thread, options);
1138
1139         def_thread.timeout = def_timeout;
1140         def_thread.write_bw_log = write_bw_log;
1141         def_thread.write_lat_log = write_lat_log;
1142
1143 #ifdef FIO_HAVE_DISK_UTIL
1144         def_thread.do_disk_util = 1;
1145 #endif
1146
1147         return 0;
1148 }
1149
1150 static void usage(void)
1151 {
1152         printf("%s\n", fio_version_string);
1153         printf("\t--output\tWrite output to file\n");
1154         printf("\t--timeout\tRuntime in seconds\n");
1155         printf("\t--latency-log\tGenerate per-job latency logs\n");
1156         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1157         printf("\t--minimal\tMinimal (terse) output\n");
1158         printf("\t--version\tPrint version info and exit\n");
1159         printf("\t--help\t\tPrint this page\n");
1160         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1161 }
1162
1163 static int parse_cmd_line(int argc, char *argv[])
1164 {
1165         struct thread_data *td = NULL;
1166         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1167
1168         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1169                 switch (c) {
1170                 case 't':
1171                         def_timeout = atoi(optarg);
1172                         break;
1173                 case 'l':
1174                         write_lat_log = 1;
1175                         break;
1176                 case 'w':
1177                         write_bw_log = 1;
1178                         break;
1179                 case 'o':
1180                         f_out = fopen(optarg, "w+");
1181                         if (!f_out) {
1182                                 perror("fopen output");
1183                                 exit(1);
1184                         }
1185                         f_err = f_out;
1186                         break;
1187                 case 'm':
1188                         terse_output = 1;
1189                         break;
1190                 case 'h':
1191                         usage();
1192                         exit(0);
1193                 case 'c':
1194                         ret = show_cmd_help(options, optarg);
1195                         exit(ret);
1196                 case 'v':
1197                         printf("%s\n", fio_version_string);
1198                         exit(0);
1199                 case FIO_GETOPT_JOB: {
1200                         const char *opt = long_options[lidx].name;
1201                         char *val = optarg;
1202
1203                         if (!strncmp(opt, "name", 4) && td) {
1204                                 ret = add_job(td, td->name ?: "fio", 0);
1205                                 if (ret) {
1206                                         put_job(td);
1207                                         return 0;
1208                                 }
1209                                 td = NULL;
1210                         }
1211                         if (!td) {
1212                                 int global = !strncmp(val, "global", 6);
1213
1214                                 td = get_new_job(global, &def_thread);
1215                                 if (!td)
1216                                         return 0;
1217                         }
1218
1219                         ret = parse_cmd_option(opt, val, options, td);
1220                         if (ret) {
1221                                 dont_add_job = 1;
1222                                 log_err("fio: job dropped\n");
1223                         }
1224                         break;
1225                 }
1226                 default:
1227                         break;
1228                 }
1229         }
1230
1231         if (td) {
1232                 if (dont_add_job)
1233                         put_job(td);
1234                 else {
1235                         ret = add_job(td, td->name ?: "fio", 0);
1236                         if (ret)
1237                                 put_job(td);
1238                 }
1239         }
1240
1241         while (optind < argc) {
1242                 ini_idx++;
1243                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1244                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1245                 optind++;
1246         }
1247
1248         return ini_idx;
1249 }
1250
1251 static void free_shm(void)
1252 {
1253         struct shmid_ds sbuf;
1254
1255         if (threads) {
1256                 shmdt((void *) threads);
1257                 threads = NULL;
1258                 shmctl(shm_id, IPC_RMID, &sbuf);
1259         }
1260 }
1261
1262 /*
1263  * The thread area is shared between the main process and the job
1264  * threads/processes. So setup a shared memory segment that will hold
1265  * all the job info.
1266  */
1267 static int setup_thread_area(void)
1268 {
1269         /*
1270          * 1024 is too much on some machines, scale max_jobs if
1271          * we get a failure that looks like too large a shm segment
1272          */
1273         do {
1274                 size_t size = max_jobs * sizeof(struct thread_data);
1275
1276                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1277                 if (shm_id != -1)
1278                         break;
1279                 if (errno != EINVAL) {
1280                         perror("shmget");
1281                         break;
1282                 }
1283
1284                 max_jobs >>= 1;
1285         } while (max_jobs);
1286
1287         if (shm_id == -1)
1288                 return 1;
1289
1290         threads = shmat(shm_id, NULL, 0);
1291         if (threads == (void *) -1) {
1292                 perror("shmat");
1293                 return 1;
1294         }
1295
1296         atexit(free_shm);
1297         return 0;
1298 }
1299
1300 /*
1301  * Copy the fio options into the long options map, so we mirror
1302  * job and cmd line options.
1303  */
1304 static void dupe_job_options(void)
1305 {
1306         struct fio_option *o;
1307         unsigned int i;
1308
1309         i = 0;
1310         while (long_options[i].name)
1311                 i++;
1312
1313         o = &options[0];
1314         while (o->name) {
1315                 long_options[i].name = o->name;
1316                 long_options[i].val = FIO_GETOPT_JOB;
1317                 if (o->type == FIO_OPT_STR_SET)
1318                         long_options[i].has_arg = no_argument;
1319                 else
1320                         long_options[i].has_arg = required_argument;
1321
1322                 i++;
1323                 o++;
1324                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1325         }
1326 }
1327
1328 int parse_options(int argc, char *argv[])
1329 {
1330         int job_files, i;
1331
1332         f_out = stdout;
1333         f_err = stderr;
1334
1335         options_init(options);
1336
1337         dupe_job_options();
1338
1339         if (setup_thread_area())
1340                 return 1;
1341         if (fill_def_thread())
1342                 return 1;
1343
1344         job_files = parse_cmd_line(argc, argv);
1345
1346         for (i = 0; i < job_files; i++) {
1347                 if (fill_def_thread())
1348                         return 1;
1349                 if (parse_jobs_ini(ini_file[i], i))
1350                         return 1;
1351                 free(ini_file[i]);
1352         }
1353
1354         free(ini_file);
1355
1356         if (!thread_number) {
1357                 log_err("No jobs defined(s)\n");
1358                 return 1;
1359         }
1360
1361         return 0;
1362 }