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
  34 #define __stringify_1(x)        #x
  35 #define __stringify(x)          __stringify_1(x)
  36
  37 /*
  38  * Map of job/command line options
  39  */
  40 static struct fio_option options[] = {
  41         {
  42                 .name   = "description",
  43                 .type   = FIO_OPT_STR_STORE,
  44                 .off1   = td_var_offset(description),
  45                 .help   = "Text job description",
  46         },
  47         {
  48                 .name   = "name",
  49                 .type   = FIO_OPT_STR_STORE,
  50                 .off1   = td_var_offset(name),
  51                 .help   = "Name of this job",
  52         },
  53         {
  54                 .name   = "directory",
  55                 .type   = FIO_OPT_STR_STORE,
  56                 .off1   = td_var_offset(directory),
  57                 .help   = "Directory to store files in",
  58         },
  59         {
  60                 .name   = "filename",
  61                 .type   = FIO_OPT_STR_STORE,
  62                 .off1   = td_var_offset(filename),
  63                 .help   = "Force the use of a specific file",
  64         },
  65         {
  66                 .name   = "rw",
  67                 .type   = FIO_OPT_STR,
  68                 .off1   = td_var_offset(td_ddir),
  69                 .help   = "IO direction",
  70                 .def    = "read",
  71                 .posval = {
  72                           { .ival = "read", .oval = TD_DDIR_READ },
  73                           { .ival = "write", .oval = TD_DDIR_WRITE },
  74                           { .ival = "randread", .oval = TD_DDIR_RANDREAD },
  75                           { .ival = "randwrite", .oval = TD_DDIR_RANDWRITE },
  76                           { .ival = "rw", .oval = TD_DDIR_RW },
  77                           { .ival = "randrw", .oval = TD_DDIR_RANDRW },
  78                           },
  79         },
  80         {
  81                 .name   = "ioengine",
  82                 .type   = FIO_OPT_STR_STORE,
  83                 .off1   = td_var_offset(ioengine),
  84                 .help   = "IO engine to use",
  85                 .def    = "sync",
  86                 .posval = {
  87                           { .ival = "sync", },
  88 #ifdef FIO_HAVE_LIBAIO
  89                           { .ival = "libaio", },
  90 #endif
  91 #ifdef FIO_HAVE_POSIXAIO
  92                           { .ival = "posixaio", },
  93 #endif
  94                           { .ival = "mmap", },
  95 #ifdef FIO_HAVE_SPLICE
  96                           { .ival = "splice", },
  97 #endif
  98 #ifdef FIO_HAVE_SGIO
  99                           { .ival = "sg", },
 100 #endif
 101                           { .ival = "null", }, { .ival = "net", },
 102 #ifdef FIO_HAVE_SYSLET
 103                           { .ival = "syslet-rw", },
 104 #endif
 105                           { .ival = "external", },
 106                           },
 107         },
 108         {
 109                 .name   = "iodepth",
 110                 .type   = FIO_OPT_INT,
 111                 .off1   = td_var_offset(iodepth),
 112                 .help   = "Amount of IO buffers to keep in flight",
 113                 .def    = "1",
 114         },
 115         {
 116                 .name   = "iodepth_batch",
 117                 .type   = FIO_OPT_INT,
 118                 .off1   = td_var_offset(iodepth_batch),
 119                 .help   = "Number of IO to submit in one go",
 120         },
 121         {
 122                 .name   = "iodepth_low",
 123                 .type   = FIO_OPT_INT,
 124                 .off1   = td_var_offset(iodepth_low),
 125                 .help   = "Low water mark for queuing depth",
 126         },
 127         {
 128                 .name   = "size",
 129                 .type   = FIO_OPT_STR_VAL,
 130                 .off1   = td_var_offset(total_file_size),
 131                 .help   = "Size of device or file",
 132         },
 133         {
 134                 .name   = "bs",
 135                 .type   = FIO_OPT_STR_VAL_INT,
 136                 .off1   = td_var_offset(bs[DDIR_READ]),
 137                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 138                 .help   = "Block size unit",
 139                 .def    = "4k",
 140         },
 141         {
 142                 .name   = "bsrange",
 143                 .type   = FIO_OPT_RANGE,
 144                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 145                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 146                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 147                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 148                 .help   = "Set block size range (in more detail than bs)",
 149         },
 150         {
 151                 .name   = "bs_unaligned",
 152                 .type   = FIO_OPT_STR_SET,
 153                 .off1   = td_var_offset(bs_unaligned),
 154                 .help   = "Don't sector align IO buffer sizes",
 155         },
 156         {
 157                 .name   = "offset",
 158                 .type   = FIO_OPT_STR_VAL,
 159                 .off1   = td_var_offset(start_offset),
 160                 .help   = "Start IO from this offset",
 161                 .def    = "0",
 162         },
 163         {
 164                 .name   = "randrepeat",
 165                 .type   = FIO_OPT_BOOL,
 166                 .off1   = td_var_offset(rand_repeatable),
 167                 .help   = "Use repeatable random IO pattern",
 168                 .def    = "1",
 169         },
 170         {
 171                 .name   = "norandommap",
 172                 .type   = FIO_OPT_STR_SET,
 173                 .off1   = td_var_offset(norandommap),
 174                 .help   = "Accept potential duplicate random blocks",
 175         },
 176         {
 177                 .name   = "nrfiles",
 178                 .type   = FIO_OPT_INT,
 179                 .off1   = td_var_offset(nr_files),
 180                 .help   = "Split job workload between this number of files",
 181                 .def    = "1",
 182         },
 183         {
 184                 .name   = "openfiles",
 185                 .type   = FIO_OPT_INT,
 186                 .off1   = td_var_offset(open_files),
 187                 .help   = "Number of files to keep open at the same time",
 188         },
 189         {
 190                 .name   = "file_service_type",
 191                 .type   = FIO_OPT_STR,
 192                 .off1   = td_var_offset(file_service_type),
 193                 .help   = "How to select which file to service next",
 194                 .def    = "roundrobin",
 195                 .posval = {
 196                           { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
 197                           { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
 198                           },
 199         },
 200         {
 201                 .name   = "fsync",
 202                 .type   = FIO_OPT_INT,
 203                 .off1   = td_var_offset(fsync_blocks),
 204                 .help   = "Issue fsync for writes every given number of blocks",
 205                 .def    = "0",
 206         },
 207         {
 208                 .name   = "direct",
 209                 .type   = FIO_OPT_BOOL,
 210                 .off1   = td_var_offset(odirect),
 211                 .help   = "Use O_DIRECT IO (negates buffered)",
 212                 .def    = "0",
 213         },
 214         {
 215                 .name   = "buffered",
 216                 .type   = FIO_OPT_BOOL,
 217                 .off1   = td_var_offset(odirect),
 218                 .neg    = 1,
 219                 .help   = "Use buffered IO (negates direct)",
 220                 .def    = "1",
 221         },
 222         {
 223                 .name   = "overwrite",
 224                 .type   = FIO_OPT_BOOL,
 225                 .off1   = td_var_offset(overwrite),
 226                 .help   = "When writing, set whether to overwrite current data",
 227                 .def    = "0",
 228         },
 229         {
 230                 .name   = "loops",
 231                 .type   = FIO_OPT_INT,
 232                 .off1   = td_var_offset(loops),
 233                 .help   = "Number of times to run the job",
 234                 .def    = "1",
 235         },
 236         {
 237                 .name   = "numjobs",
 238                 .type   = FIO_OPT_INT,
 239                 .off1   = td_var_offset(numjobs),
 240                 .help   = "Duplicate this job this many times",
 241                 .def    = "1",
 242         },
 243         {
 244                 .name   = "startdelay",
 245                 .type   = FIO_OPT_INT,
 246                 .off1   = td_var_offset(start_delay),
 247                 .help   = "Only start job when this period has passed",
 248                 .def    = "0",
 249         },
 250         {
 251                 .name   = "runtime",
 252                 .alias  = "timeout",
 253                 .type   = FIO_OPT_STR_VAL_TIME,
 254                 .off1   = td_var_offset(timeout),
 255                 .help   = "Stop workload when this amount of time has passed",
 256                 .def    = "0",
 257         },
 258         {
 259                 .name   = "mem",
 260                 .type   = FIO_OPT_STR,
 261                 .cb     = str_mem_cb,
 262                 .off1   = td_var_offset(mem_type),
 263                 .help   = "Backing type for IO buffers",
 264                 .def    = "malloc",
 265                 .posval = {
 266                           { .ival = "malloc", .oval = MEM_MALLOC },
 267                           { .ival = "shm", .oval = MEM_SHM },
 268 #ifdef FIO_HAVE_HUGETLB
 269                           { .ival = "shmhuge", .oval = MEM_SHMHUGE },
 270 #endif
 271                           { .ival = "mmap", .oval = MEM_MMAP },
 272 #ifdef FIO_HAVE_HUGETLB
 273                           { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
 274 #endif
 275                           },
 276         },
 277         {
 278                 .name   = "verify",
 279                 .type   = FIO_OPT_STR,
 280                 .off1   = td_var_offset(verify),
 281                 .help   = "Verify sum function",
 282                 .def    = "0",
 283                 .posval = {
 284                           { .ival = "0", .oval = VERIFY_NONE },
 285                           { .ival = "crc32", .oval = VERIFY_CRC32 },
 286                           { .ival = "md5", .oval = VERIFY_MD5 },
 287                           },
 288         },
 289         {
 290                 .name   = "write_iolog",
 291                 .type   = FIO_OPT_STR_STORE,
 292                 .off1   = td_var_offset(write_iolog_file),
 293                 .help   = "Store IO pattern to file",
 294         },
 295         {
 296                 .name   = "read_iolog",
 297                 .type   = FIO_OPT_STR_STORE,
 298                 .off1   = td_var_offset(read_iolog_file),
 299                 .help   = "Playback IO pattern from file",
 300         },
 301         {
 302                 .name   = "exec_prerun",
 303                 .type   = FIO_OPT_STR_STORE,
 304                 .off1   = td_var_offset(exec_prerun),
 305                 .help   = "Execute this file prior to running job",
 306         },
 307         {
 308                 .name   = "exec_postrun",
 309                 .type   = FIO_OPT_STR_STORE,
 310                 .off1   = td_var_offset(exec_postrun),
 311                 .help   = "Execute this file after running job",
 312         },
 313 #ifdef FIO_HAVE_IOSCHED_SWITCH
 314         {
 315                 .name   = "ioscheduler",
 316                 .type   = FIO_OPT_STR_STORE,
 317                 .off1   = td_var_offset(ioscheduler),
 318                 .help   = "Use this IO scheduler on the backing device",
 319         },
 320 #endif
 321         {
 322                 .name   = "zonesize",
 323                 .type   = FIO_OPT_STR_VAL,
 324                 .off1   = td_var_offset(zone_size),
 325                 .help   = "Give size of an IO zone",
 326                 .def    = "0",
 327         },
 328         {
 329                 .name   = "zoneskip",
 330                 .type   = FIO_OPT_STR_VAL,
 331                 .off1   = td_var_offset(zone_skip),
 332                 .help   = "Space between IO zones",
 333                 .def    = "0",
 334         },
 335         {
 336                 .name   = "lockmem",
 337                 .type   = FIO_OPT_STR_VAL,
 338                 .cb     = str_lockmem_cb,
 339                 .help   = "Lock down this amount of memory",
 340                 .def    = "0",
 341         },
 342         {
 343                 .name   = "rwmixcycle",
 344                 .type   = FIO_OPT_INT,
 345                 .off1   = td_var_offset(rwmixcycle),
 346                 .help   = "Cycle period for mixed read/write workloads (msec)",
 347                 .def    = "500",
 348         },
 349         {
 350                 .name   = "rwmixread",
 351                 .type   = FIO_OPT_INT,
 352                 .off1   = td_var_offset(rwmixread),
 353                 .maxval = 100,
 354                 .help   = "Percentage of mixed workload that is reads",
 355                 .def    = "50",
 356         },
 357         {
 358                 .name   = "rwmixwrite",
 359                 .type   = FIO_OPT_INT,
 360                 .off1   = td_var_offset(rwmixwrite),
 361                 .maxval = 100,
 362                 .help   = "Percentage of mixed workload that is writes",
 363                 .def    = "50",
 364         },
 365         {
 366                 .name   = "nice",
 367                 .type   = FIO_OPT_INT,
 368                 .off1   = td_var_offset(nice),
 369                 .help   = "Set job CPU nice value",
 370                 .minval = -19,
 371                 .maxval = 20,
 372                 .def    = "0",
 373         },
 374 #ifdef FIO_HAVE_IOPRIO
 375         {
 376                 .name   = "prio",
 377                 .type   = FIO_OPT_INT,
 378                 .cb     = str_prio_cb,
 379                 .help   = "Set job IO priority value",
 380                 .minval = 0,
 381                 .maxval = 7,
 382         },
 383         {
 384                 .name   = "prioclass",
 385                 .type   = FIO_OPT_INT,
 386                 .cb     = str_prioclass_cb,
 387                 .help   = "Set job IO priority class",
 388                 .minval = 0,
 389                 .maxval = 3,
 390         },
 391 #endif
 392         {
 393                 .name   = "thinktime",
 394                 .type   = FIO_OPT_INT,
 395                 .off1   = td_var_offset(thinktime),
 396                 .help   = "Idle time between IO buffers (usec)",
 397                 .def    = "0",
 398         },
 399         {
 400                 .name   = "thinktime_spin",
 401                 .type   = FIO_OPT_INT,
 402                 .off1   = td_var_offset(thinktime_spin),
 403                 .help   = "Start think time by spinning this amount (usec)",
 404                 .def    = "0",
 405         },
 406         {
 407                 .name   = "thinktime_blocks",
 408                 .type   = FIO_OPT_INT,
 409                 .off1   = td_var_offset(thinktime_blocks),
 410                 .help   = "IO buffer period between 'thinktime'",
 411                 .def    = "1",
 412         },
 413         {
 414                 .name   = "rate",
 415                 .type   = FIO_OPT_INT,
 416                 .off1   = td_var_offset(rate),
 417                 .help   = "Set bandwidth rate",
 418         },
 419         {
 420                 .name   = "ratemin",
 421                 .type   = FIO_OPT_INT,
 422                 .off1   = td_var_offset(ratemin),
 423                 .help   = "The bottom limit accepted",
 424         },
 425         {
 426                 .name   = "ratecycle",
 427                 .type   = FIO_OPT_INT,
 428                 .off1   = td_var_offset(ratecycle),
 429                 .help   = "Window average for rate limits (msec)",
 430                 .def    = "1000",
 431         },
 432         {
 433                 .name   = "invalidate",
 434                 .type   = FIO_OPT_BOOL,
 435                 .off1   = td_var_offset(invalidate_cache),
 436                 .help   = "Invalidate buffer/page cache prior to running job",
 437                 .def    = "1",
 438         },
 439         {
 440                 .name   = "sync",
 441                 .type   = FIO_OPT_BOOL,
 442                 .off1   = td_var_offset(sync_io),
 443                 .help   = "Use O_SYNC for buffered writes",
 444                 .def    = "0",
 445         },
 446         {
 447                 .name   = "bwavgtime",
 448                 .type   = FIO_OPT_INT,
 449                 .off1   = td_var_offset(bw_avg_time),
 450                 .help   = "Time window over which to calculate bandwidth (msec)",
 451                 .def    = "500",
 452         },
 453         {
 454                 .name   = "create_serialize",
 455                 .type   = FIO_OPT_BOOL,
 456                 .off1   = td_var_offset(create_serialize),
 457                 .help   = "Serialize creating of job files",
 458                 .def    = "1",
 459         },
 460         {
 461                 .name   = "create_fsync",
 462                 .type   = FIO_OPT_BOOL,
 463                 .off1   = td_var_offset(create_fsync),
 464                 .help   = "Fsync file after creation",
 465                 .def    = "1",
 466         },
 467         {
 468                 .name   = "cpuload",
 469                 .type   = FIO_OPT_INT,
 470                 .off1   = td_var_offset(cpuload),
 471                 .help   = "Use this percentage of CPU",
 472         },
 473         {
 474                 .name   = "cpuchunks",
 475                 .type   = FIO_OPT_INT,
 476                 .off1   = td_var_offset(cpucycle),
 477                 .help   = "Length of the CPU burn cycles",
 478         },
 479 #ifdef FIO_HAVE_CPU_AFFINITY
 480         {
 481                 .name   = "cpumask",
 482                 .type   = FIO_OPT_INT,
 483                 .cb     = str_cpumask_cb,
 484                 .help   = "CPU affinity mask",
 485         },
 486 #endif
 487         {
 488                 .name   = "end_fsync",
 489                 .type   = FIO_OPT_BOOL,
 490                 .off1   = td_var_offset(end_fsync),
 491                 .help   = "Include fsync at the end of job",
 492                 .def    = "0",
 493         },
 494         {
 495                 .name   = "unlink",
 496                 .type   = FIO_OPT_BOOL,
 497                 .off1   = td_var_offset(unlink),
 498                 .help   = "Unlink created files after job has completed",
 499                 .def    = "0",
 500         },
 501         {
 502                 .name   = "exitall",
 503                 .type   = FIO_OPT_STR_SET,
 504                 .cb     = str_exitall_cb,
 505                 .help   = "Terminate all jobs when one exits",
 506         },
 507         {
 508                 .name   = "stonewall",
 509                 .type   = FIO_OPT_STR_SET,
 510                 .off1   = td_var_offset(stonewall),
 511                 .help   = "Insert a hard barrier between this job and previous",
 512         },
 513         {
 514                 .name   = "thread",
 515                 .type   = FIO_OPT_STR_SET,
 516                 .off1   = td_var_offset(use_thread),
 517                 .help   = "Use threads instead of forks",
 518         },
 519         {
 520                 .name   = "write_bw_log",
 521                 .type   = FIO_OPT_STR_SET,
 522                 .off1   = td_var_offset(write_bw_log),
 523                 .help   = "Write log of bandwidth during run",
 524         },
 525         {
 526                 .name   = "write_lat_log",
 527                 .type   = FIO_OPT_STR_SET,
 528                 .off1   = td_var_offset(write_lat_log),
 529                 .help   = "Write log of latency during run",
 530         },
 531         {
 532                 .name   = "hugepage-size",
 533                 .type   = FIO_OPT_STR_VAL,
 534                 .off1   = td_var_offset(hugepage_size),
 535                 .help   = "When using hugepages, specify size of each page",
 536                 .def    = __stringify(FIO_HUGE_PAGE),
 537         },
 538         {
 539                 .name   = "group_reporting",
 540                 .type   = FIO_OPT_STR_SET,
 541                 .off1   = td_var_offset(group_reporting),
 542                 .help   = "Do reporting on a per-group basis",
 543         },
 544         {
 545                 .name = NULL,
 546         },
 547 };
 548
 549 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 550 #define FIO_CMD_OPTS    (16)
 551 #define FIO_GETOPT_JOB  (0x89988998)
 552
 553 /*
 554  * Command line options. These will contain the above, plus a few
 555  * extra that only pertain to fio itself and not jobs.
 556  */
 557 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 558         {
 559                 .name           = "output",
 560                 .has_arg        = required_argument,
 561                 .val            = 'o',
 562         },
 563         {
 564                 .name           = "timeout",
 565                 .has_arg        = required_argument,
 566                 .val            = 't',
 567         },
 568         {
 569                 .name           = "latency-log",
 570                 .has_arg        = required_argument,
 571                 .val            = 'l',
 572         },
 573         {
 574                 .name           = "bandwidth-log",
 575                 .has_arg        = required_argument,
 576                 .val            = 'b',
 577         },
 578         {
 579                 .name           = "minimal",
 580                 .has_arg        = optional_argument,
 581                 .val            = 'm',
 582         },
 583         {
 584                 .name           = "version",
 585                 .has_arg        = no_argument,
 586                 .val            = 'v',
 587         },
 588         {
 589                 .name           = "help",
 590                 .has_arg        = no_argument,
 591                 .val            = 'h',
 592         },
 593         {
 594                 .name           = "cmdhelp",
 595                 .has_arg        = optional_argument,
 596                 .val            = 'c',
 597         },
 598         {
 599                 .name           = NULL,
 600         },
 601 };
 602
 603 static int def_timeout = 0;
 604
 605 static char fio_version_string[] = "fio 1.13";
 606
 607 static char **ini_file;
 608 static int max_jobs = MAX_JOBS;
 609
 610 struct thread_data def_thread;
 611 struct thread_data *threads = NULL;
 612
 613 int exitall_on_terminate = 0;
 614 int terse_output = 0;
 615 unsigned long long mlock_size = 0;
 616 FILE *f_out = NULL;
 617 FILE *f_err = NULL;
 618
 619 static int write_lat_log = 0;
 620 int write_bw_log = 0;
 621
 622 FILE *get_f_out()
 623 {
 624         return f_out;
 625 }
 626
 627 FILE *get_f_err()
 628 {
 629         return f_err;
 630 }
 631
 632 /*
 633  * Return a free job structure.
 634  */
 635 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 636 {
 637         struct thread_data *td;
 638
 639         if (global)
 640                 return &def_thread;
 641         if (thread_number >= max_jobs)
 642                 return NULL;
 643
 644         td = &threads[thread_number++];
 645         *td = *parent;
 646
 647         td->thread_number = thread_number;
 648         return td;
 649 }
 650
 651 static void put_job(struct thread_data *td)
 652 {
 653         if (td == &def_thread)
 654                 return;
 655
 656         if (td->error)
 657                 fprintf(f_out, "fio: %s\n", td->verror);
 658
 659         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 660         thread_number--;
 661 }
 662
 663 /*
 664  * Lazy way of fixing up options that depend on each other. We could also
 665  * define option callback handlers, but this is easier.
 666  */
 667 static void fixup_options(struct thread_data *td)
 668 {
 669         if (!td->rwmixread && td->rwmixwrite)
 670                 td->rwmixread = 100 - td->rwmixwrite;
 671
 672         if (td->write_iolog_file && td->read_iolog_file) {
 673                 log_err("fio: read iolog overrides write_iolog\n");
 674                 free(td->write_iolog_file);
 675                 td->write_iolog_file = NULL;
 676         }
 677
 678         if (td->io_ops->flags & FIO_SYNCIO)
 679                 td->iodepth = 1;
 680         else {
 681                 if (!td->iodepth)
 682                         td->iodepth = td->open_files;
 683         }
 684
 685         /*
 686          * only really works for sequential io for now, and with 1 file
 687          */
 688         if (td->zone_size && td_random(td) && td->open_files == 1)
 689                 td->zone_size = 0;
 690
 691         /*
 692          * Reads can do overwrites, we always need to pre-create the file
 693          */
 694         if (td_read(td) || td_rw(td))
 695                 td->overwrite = 1;
 696
 697         if (!td->min_bs[DDIR_READ])
 698                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 699         if (!td->max_bs[DDIR_READ])
 700                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 701         if (!td->min_bs[DDIR_WRITE])
 702                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 703         if (!td->max_bs[DDIR_WRITE])
 704                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 705
 706         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 707
 708         if (td_read(td) && !td_rw(td))
 709                 td->verify = 0;
 710
 711         if (td->norandommap && td->verify != VERIFY_NONE) {
 712                 log_err("fio: norandommap given, verify disabled\n");
 713                 td->verify = VERIFY_NONE;
 714         }
 715         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 716                 log_err("fio: bs_unaligned may not work with raw io\n");
 717
 718         /*
 719          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 720          */
 721         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 722                 td->odirect = 0;
 723
 724         /*
 725          * thinktime_spin must be less than thinktime
 726          */
 727         if (td->thinktime_spin > td->thinktime)
 728                 td->thinktime_spin = td->thinktime;
 729
 730         /*
 731          * The low water mark cannot be bigger than the iodepth
 732          */
 733         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
 734                 /*
 735                  * syslet work around - if the workload is sequential,
 736                  * we want to let the queue drain all the way down to
 737                  * avoid seeking between async threads
 738                  */
 739                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 740                         td->iodepth_low = 1;
 741                 else
 742                         td->iodepth_low = td->iodepth;
 743         }
 744
 745         /*
 746          * If batch number isn't set, default to the same as iodepth
 747          */
 748         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
 749                 td->iodepth_batch = td->iodepth;
 750
 751         if (td->open_files > td->nr_files || !td->open_files)
 752                 td->open_files = td->nr_files;
 753 }
 754
 755 /*
 756  * This function leaks the buffer
 757  */
 758 static char *to_kmg(unsigned int val)
 759 {
 760         char *buf = malloc(32);
 761         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 762         char *p = post;
 763
 764         do {
 765                 if (val & 1023)
 766                         break;
 767
 768                 val >>= 10;
 769                 p++;
 770         } while (*p);
 771
 772         snprintf(buf, 31, "%u%c", val, *p);
 773         return buf;
 774 }
 775
 776 /* External engines are specified by "external:name.o") */
 777 static const char *get_engine_name(const char *str)
 778 {
 779         char *p = strstr(str, ":");
 780
 781         if (!p)
 782                 return str;
 783
 784         p++;
 785         strip_blank_front(&p);
 786         strip_blank_end(p);
 787         return p;
 788 }
 789
 790 /*
 791  * Adds a job to the list of things todo. Sanitizes the various options
 792  * to make sure we don't have conflicts, and initializes various
 793  * members of td.
 794  */
 795 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 796 {
 797         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 798                                    "randread", "randwrite", "randrw" };
 799         struct stat sb;
 800         int numjobs, i;
 801         struct fio_file *f;
 802         const char *engine;
 803
 804         /*
 805          * the def_thread is just for options, it's not a real job
 806          */
 807         if (td == &def_thread)
 808                 return 0;
 809
 810         engine = get_engine_name(td->ioengine);
 811         td->io_ops = load_ioengine(td, engine);
 812         if (!td->io_ops) {
 813                 log_err("fio: failed to load engine %s\n", engine);
 814                 return 1;
 815         }
 816
 817         if (td->odirect)
 818                 td->io_ops->flags |= FIO_RAWIO;
 819
 820         td->filetype = FIO_TYPE_FILE;
 821         if (td->filename && !lstat(td->filename, &sb)) {
 822                 if (S_ISBLK(sb.st_mode))
 823                         td->filetype = FIO_TYPE_BD;
 824                 else if (S_ISCHR(sb.st_mode))
 825                         td->filetype = FIO_TYPE_CHAR;
 826         }
 827
 828         fixup_options(td);
 829
 830         if (td->filename)
 831                 td->nr_uniq_files = 1;
 832         else
 833                 td->nr_uniq_files = td->open_files;
 834
 835         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 836                 char tmp[PATH_MAX];
 837                 int len = 0;
 838
 839                 if (td->directory && td->directory[0] != '\0') {
 840                         if (lstat(td->directory, &sb) < 0) {
 841                                 log_err("fio: %s is not a directory\n", td->directory);
 842                                 td_verror(td, errno, "lstat");
 843                                 return 1;
 844                         }
 845                         if (!S_ISDIR(sb.st_mode)) {
 846                                 log_err("fio: %s is not a directory\n", td->directory);
 847                                 return 1;
 848                         }
 849                         len = sprintf(tmp, "%s/", td->directory);
 850                 }
 851
 852                 td->files = malloc(sizeof(struct fio_file) * td->open_files);
 853
 854                 for_each_file(td, f, i) {
 855                         memset(f, 0, sizeof(*f));
 856                         f->fd = -1;
 857
 858                         if (td->filename)
 859                                 sprintf(tmp + len, "%s", td->filename);
 860                         else
 861                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 862                         f->file_name = strdup(tmp);
 863                 }
 864         } else {
 865                 td->open_files = td->nr_files = 1;
 866                 td->files = malloc(sizeof(struct fio_file));
 867                 f = &td->files[0];
 868
 869                 memset(f, 0, sizeof(*f));
 870                 f->fd = -1;
 871                 f->file_name = strdup(jobname);
 872         }
 873
 874         for_each_file(td, f, i) {
 875                 f->file_size = td->total_file_size / td->nr_files;
 876                 f->file_offset = td->start_offset;
 877         }
 878
 879         td->mutex = fio_sem_init(0);
 880
 881         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 882         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 883         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 884
 885         if (td->stonewall && td->thread_number > 1)
 886                 groupid++;
 887
 888         td->groupid = groupid;
 889
 890         if (setup_rate(td))
 891                 goto err;
 892
 893         if (td->write_lat_log) {
 894                 setup_log(&td->ts.slat_log);
 895                 setup_log(&td->ts.clat_log);
 896         }
 897         if (td->write_bw_log)
 898                 setup_log(&td->ts.bw_log);
 899
 900         if (!td->name)
 901                 td->name = strdup(jobname);
 902
 903         if (!terse_output) {
 904                 if (!job_add_num) {
 905                         if (td->io_ops->flags & FIO_CPUIO)
 906                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 907                         else {
 908                                 char *c1, *c2, *c3, *c4;
 909
 910                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 911                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 912                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 913                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 914
 915                                 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);
 916
 917                                 free(c1);
 918                                 free(c2);
 919                                 free(c3);
 920                                 free(c4);
 921                         }
 922                 } else if (job_add_num == 1)
 923                         fprintf(f_out, "...\n");
 924         }
 925
 926         /*
 927          * recurse add identical jobs, clear numjobs and stonewall options
 928          * as they don't apply to sub-jobs
 929          */
 930         numjobs = td->numjobs;
 931         while (--numjobs) {
 932                 struct thread_data *td_new = get_new_job(0, td);
 933
 934                 if (!td_new)
 935                         goto err;
 936
 937                 td_new->numjobs = 1;
 938                 td_new->stonewall = 0;
 939                 job_add_num = numjobs - 1;
 940
 941                 if (add_job(td_new, jobname, job_add_num))
 942                         goto err;
 943         }
 944         return 0;
 945 err:
 946         put_job(td);
 947         return -1;
 948 }
 949
 950 /*
 951  * Initialize the various random states we need (random io, block size ranges,
 952  * read/write mix, etc).
 953  */
 954 int init_random_state(struct thread_data *td)
 955 {
 956         unsigned long seeds[5];
 957         int fd, num_maps, blocks, i;
 958         struct fio_file *f;
 959
 960         if (td->io_ops->flags & FIO_CPUIO)
 961                 return 0;
 962
 963         fd = open("/dev/urandom", O_RDONLY);
 964         if (fd == -1) {
 965                 td_verror(td, errno, "open");
 966                 return 1;
 967         }
 968
 969         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 970                 td_verror(td, EIO, "read");
 971                 close(fd);
 972                 return 1;
 973         }
 974
 975         close(fd);
 976
 977         os_random_seed(seeds[0], &td->bsrange_state);
 978         os_random_seed(seeds[1], &td->verify_state);
 979         os_random_seed(seeds[2], &td->rwmix_state);
 980
 981         if (td->file_service_type == FIO_FSERVICE_RANDOM)
 982                 os_random_seed(seeds[3], &td->next_file_state);
 983
 984         if (!td_random(td))
 985                 return 0;
 986
 987         if (td->rand_repeatable)
 988                 seeds[4] = FIO_RANDSEED * td->thread_number;
 989
 990         if (!td->norandommap) {
 991                 for_each_file(td, f, i) {
 992                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 993                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 994                         f->file_map = malloc(num_maps * sizeof(long));
 995                         f->num_maps = num_maps;
 996                         memset(f->file_map, 0, num_maps * sizeof(long));
 997                 }
 998         }
 999
1000         os_random_seed(seeds[4], &td->random_state);
1001         return 0;
1002 }
1003
1004 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
1005 {
1006 #ifdef FIO_HAVE_CPU_AFFINITY
1007         unsigned int i;
1008
1009         CPU_ZERO(&cpumask);
1010
1011         for (i = 0; i < sizeof(int) * 8; i++) {
1012                 if ((1 << i) & cpu)
1013                         CPU_SET(i, &cpumask);
1014         }
1015 #endif
1016 }
1017
1018 static int is_empty_or_comment(char *line)
1019 {
1020         unsigned int i;
1021
1022         for (i = 0; i < strlen(line); i++) {
1023                 if (line[i] == ';')
1024                         return 1;
1025                 if (line[i] == '#')
1026                         return 1;
1027                 if (!isspace(line[i]) && !iscntrl(line[i]))
1028                         return 0;
1029         }
1030
1031         return 1;
1032 }
1033
1034 /*
1035  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
1036  */
1037 static char *get_mmap_file(const char *str)
1038 {
1039         char *p = strstr(str, ":");
1040
1041         if (!p)
1042                 return NULL;
1043
1044         p++;
1045         strip_blank_front(&p);
1046         strip_blank_end(p);
1047         return strdup(p);
1048 }
1049
1050 static int str_mem_cb(void *data, const char *mem)
1051 {
1052         struct thread_data *td = data;
1053
1054         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1055                 td->mmapfile = get_mmap_file(mem);
1056                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1057                         log_err("fio: mmaphuge:/path/to/file\n");
1058                         return 1;
1059                 }
1060         }
1061
1062         return 0;
1063 }
1064
1065 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1066 {
1067         mlock_size = *val;
1068         return 0;
1069 }
1070
1071 #ifdef FIO_HAVE_IOPRIO
1072 static int str_prioclass_cb(void *data, unsigned int *val)
1073 {
1074         struct thread_data *td = data;
1075
1076         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1077         return 0;
1078 }
1079
1080 static int str_prio_cb(void *data, unsigned int *val)
1081 {
1082         struct thread_data *td = data;
1083
1084         td->ioprio |= *val;
1085         return 0;
1086 }
1087 #endif
1088
1089 static int str_exitall_cb(void)
1090 {
1091         exitall_on_terminate = 1;
1092         return 0;
1093 }
1094
1095 static int str_cpumask_cb(void *data, unsigned int *val)
1096 {
1097         struct thread_data *td = data;
1098
1099         fill_cpu_mask(td->cpumask, *val);
1100         return 0;
1101 }
1102
1103 /*
1104  * This is our [ini] type file parser.
1105  */
1106 static int parse_jobs_ini(char *file, int stonewall_flag)
1107 {
1108         unsigned int global;
1109         struct thread_data *td;
1110         char *string, *name;
1111         fpos_t off;
1112         FILE *f;
1113         char *p;
1114         int ret = 0, stonewall;
1115
1116         f = fopen(file, "r");
1117         if (!f) {
1118                 perror("fopen job file");
1119                 return 1;
1120         }
1121
1122         string = malloc(4096);
1123         name = malloc(256);
1124         memset(name, 0, 256);
1125
1126         stonewall = stonewall_flag;
1127         do {
1128                 p = fgets(string, 4095, f);
1129                 if (!p)
1130                         break;
1131                 if (is_empty_or_comment(p))
1132                         continue;
1133                 if (sscanf(p, "[%255s]", name) != 1)
1134                         continue;
1135
1136                 global = !strncmp(name, "global", 6);
1137
1138                 name[strlen(name) - 1] = '\0';
1139
1140                 td = get_new_job(global, &def_thread);
1141                 if (!td) {
1142                         ret = 1;
1143                         break;
1144                 }
1145
1146                 /*
1147                  * Seperate multiple job files by a stonewall
1148                  */
1149                 if (!global && stonewall) {
1150                         td->stonewall = stonewall;
1151                         stonewall = 0;
1152                 }
1153
1154                 fgetpos(f, &off);
1155                 while ((p = fgets(string, 4096, f)) != NULL) {
1156                         if (is_empty_or_comment(p))
1157                                 continue;
1158
1159                         strip_blank_front(&p);
1160
1161                         if (p[0] == '[')
1162                                 break;
1163
1164                         strip_blank_end(p);
1165
1166                         fgetpos(f, &off);
1167
1168                         /*
1169                          * Don't break here, continue parsing options so we
1170                          * dump all the bad ones. Makes trial/error fixups
1171                          * easier on the user.
1172                          */
1173                         ret |= parse_option(p, options, td);
1174                 }
1175
1176                 if (!ret) {
1177                         fsetpos(f, &off);
1178                         ret = add_job(td, name, 0);
1179                 } else {
1180                         log_err("fio: job %s dropped\n", name);
1181                         put_job(td);
1182                 }
1183         } while (!ret);
1184
1185         free(string);
1186         free(name);
1187         fclose(f);
1188         return ret;
1189 }
1190
1191 static int fill_def_thread(void)
1192 {
1193         memset(&def_thread, 0, sizeof(def_thread));
1194
1195         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1196                 perror("sched_getaffinity");
1197                 return 1;
1198         }
1199
1200         /*
1201          * fill default options
1202          */
1203         fill_default_options(&def_thread, options);
1204
1205         def_thread.timeout = def_timeout;
1206         def_thread.write_bw_log = write_bw_log;
1207         def_thread.write_lat_log = write_lat_log;
1208
1209 #ifdef FIO_HAVE_DISK_UTIL
1210         def_thread.do_disk_util = 1;
1211 #endif
1212
1213         return 0;
1214 }
1215
1216 static void usage(void)
1217 {
1218         printf("%s\n", fio_version_string);
1219         printf("\t--output\tWrite output to file\n");
1220         printf("\t--timeout\tRuntime in seconds\n");
1221         printf("\t--latency-log\tGenerate per-job latency logs\n");
1222         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1223         printf("\t--minimal\tMinimal (terse) output\n");
1224         printf("\t--version\tPrint version info and exit\n");
1225         printf("\t--help\t\tPrint this page\n");
1226         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1227 }
1228
1229 static int parse_cmd_line(int argc, char *argv[])
1230 {
1231         struct thread_data *td = NULL;
1232         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1233
1234         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
1235                 switch (c) {
1236                 case 't':
1237                         def_timeout = atoi(optarg);
1238                         break;
1239                 case 'l':
1240                         write_lat_log = 1;
1241                         break;
1242                 case 'w':
1243                         write_bw_log = 1;
1244                         break;
1245                 case 'o':
1246                         f_out = fopen(optarg, "w+");
1247                         if (!f_out) {
1248                                 perror("fopen output");
1249                                 exit(1);
1250                         }
1251                         f_err = f_out;
1252                         break;
1253                 case 'm':
1254                         terse_output = 1;
1255                         break;
1256                 case 'h':
1257                         usage();
1258                         exit(0);
1259                 case 'c':
1260                         ret = show_cmd_help(options, optarg);
1261                         exit(ret);
1262                 case 'v':
1263                         printf("%s\n", fio_version_string);
1264                         exit(0);
1265                 case FIO_GETOPT_JOB: {
1266                         const char *opt = long_options[lidx].name;
1267                         char *val = optarg;
1268
1269                         if (!strncmp(opt, "name", 4) && td) {
1270                                 ret = add_job(td, td->name ?: "fio", 0);
1271                                 if (ret) {
1272                                         put_job(td);
1273                                         return 0;
1274                                 }
1275                                 td = NULL;
1276                         }
1277                         if (!td) {
1278                                 int global = !strncmp(val, "global", 6);
1279
1280                                 td = get_new_job(global, &def_thread);
1281                                 if (!td)
1282                                         return 0;
1283                         }
1284
1285                         ret = parse_cmd_option(opt, val, options, td);
1286                         if (ret)
1287                                 dont_add_job = 1;
1288                         break;
1289                 }
1290                 default:
1291                         break;
1292                 }
1293         }
1294
1295         if (td) {
1296                 if (dont_add_job)
1297                         put_job(td);
1298                 else {
1299                         ret = add_job(td, td->name ?: "fio", 0);
1300                         if (ret)
1301                                 put_job(td);
1302                 }
1303         }
1304
1305         while (optind < argc) {
1306                 ini_idx++;
1307                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1308                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1309                 optind++;
1310         }
1311
1312         return ini_idx;
1313 }
1314
1315 static void free_shm(void)
1316 {
1317         struct shmid_ds sbuf;
1318
1319         if (threads) {
1320                 shmdt((void *) threads);
1321                 threads = NULL;
1322                 shmctl(shm_id, IPC_RMID, &sbuf);
1323         }
1324 }
1325
1326 /*
1327  * The thread area is shared between the main process and the job
1328  * threads/processes. So setup a shared memory segment that will hold
1329  * all the job info.
1330  */
1331 static int setup_thread_area(void)
1332 {
1333         /*
1334          * 1024 is too much on some machines, scale max_jobs if
1335          * we get a failure that looks like too large a shm segment
1336          */
1337         do {
1338                 size_t size = max_jobs * sizeof(struct thread_data);
1339
1340                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1341                 if (shm_id != -1)
1342                         break;
1343                 if (errno != EINVAL) {
1344                         perror("shmget");
1345                         break;
1346                 }
1347
1348                 max_jobs >>= 1;
1349         } while (max_jobs);
1350
1351         if (shm_id == -1)
1352                 return 1;
1353
1354         threads = shmat(shm_id, NULL, 0);
1355         if (threads == (void *) -1) {
1356                 perror("shmat");
1357                 return 1;
1358         }
1359
1360         atexit(free_shm);
1361         return 0;
1362 }
1363
1364 /*
1365  * Copy the fio options into the long options map, so we mirror
1366  * job and cmd line options.
1367  */
1368 static void dupe_job_options(void)
1369 {
1370         struct fio_option *o;
1371         unsigned int i;
1372
1373         i = 0;
1374         while (long_options[i].name)
1375                 i++;
1376
1377         o = &options[0];
1378         while (o->name) {
1379                 long_options[i].name = o->name;
1380                 long_options[i].val = FIO_GETOPT_JOB;
1381                 if (o->type == FIO_OPT_STR_SET)
1382                         long_options[i].has_arg = no_argument;
1383                 else
1384                         long_options[i].has_arg = required_argument;
1385
1386                 i++;
1387                 o++;
1388                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1389         }
1390 }
1391
1392 int parse_options(int argc, char *argv[])
1393 {
1394         int job_files, i;
1395
1396         f_out = stdout;
1397         f_err = stderr;
1398
1399         options_init(options);
1400
1401         dupe_job_options();
1402
1403         if (setup_thread_area())
1404                 return 1;
1405         if (fill_def_thread())
1406                 return 1;
1407
1408         job_files = parse_cmd_line(argc, argv);
1409
1410         for (i = 0; i < job_files; i++) {
1411                 if (fill_def_thread())
1412                         return 1;
1413                 if (parse_jobs_ini(ini_file[i], i))
1414                         return 1;
1415                 free(ini_file[i]);
1416         }
1417
1418         free(ini_file);
1419
1420         if (!thread_number) {
1421                 log_err("No jobs defined(s)\n");
1422                 return 1;
1423         }
1424
1425         return 0;
1426 }