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_rw_cb(void *, const char *);
  26 static int str_ioengine_cb(void *, const char *);
  27 static int str_mem_cb(void *, const char *);
  28 static int str_verify_cb(void *, const char *);
  29 static int str_lockmem_cb(void *, unsigned long *);
  30 #ifdef FIO_HAVE_IOPRIO
  31 static int str_prio_cb(void *, unsigned int *);
  32 static int str_prioclass_cb(void *, unsigned int *);
  33 #endif
  34 static int str_exitall_cb(void);
  35 static int str_cpumask_cb(void *, unsigned int *);
  36
  37 #define __stringify_1(x)        #x
  38 #define __stringify(x)          __stringify_1(x)
  39
  40 /*
  41  * Map of job/command line options
  42  */
  43 static struct fio_option options[] = {
  44         {
  45                 .name   = "description",
  46                 .type   = FIO_OPT_STR_STORE,
  47                 .off1   = td_var_offset(description),
  48                 .help   = "Text job description",
  49         },
  50         {
  51                 .name   = "name",
  52                 .type   = FIO_OPT_STR_STORE,
  53                 .off1   = td_var_offset(name),
  54                 .help   = "Name of this job",
  55         },
  56         {
  57                 .name   = "directory",
  58                 .type   = FIO_OPT_STR_STORE,
  59                 .off1   = td_var_offset(directory),
  60                 .help   = "Directory to store files in",
  61         },
  62         {
  63                 .name   = "filename",
  64                 .type   = FIO_OPT_STR_STORE,
  65                 .off1   = td_var_offset(filename),
  66                 .help   = "Force the use of a specific file",
  67         },
  68         {
  69                 .name   = "rw",
  70                 .type   = FIO_OPT_STR,
  71                 .cb     = str_rw_cb,
  72                 .help   = "IO direction",
  73                 .def    = "read",
  74                 .posval = { "read", "write", "randwrite", "randread", "rw",
  75                                 "randrw", },
  76         },
  77         {
  78                 .name   = "ioengine",
  79                 .type   = FIO_OPT_STR,
  80                 .cb     = str_ioengine_cb,
  81                 .help   = "IO engine to use",
  82                 .def    = "sync",
  83                 .posval = { "sync", "libaio", "posixaio", "mmap", "splice",
  84                                 "sg", "null", "net", "syslet-rw" },
  85         },
  86         {
  87                 .name   = "iodepth",
  88                 .type   = FIO_OPT_INT,
  89                 .off1   = td_var_offset(iodepth),
  90                 .help   = "Amount of IO buffers to keep in flight",
  91                 .def    = "1",
  92         },
  93         {
  94                 .name   = "size",
  95                 .type   = FIO_OPT_STR_VAL,
  96                 .off1   = td_var_offset(total_file_size),
  97                 .help   = "Size of device or file",
  98         },
  99         {
 100                 .name   = "bs",
 101                 .type   = FIO_OPT_STR_VAL_INT,
 102                 .off1   = td_var_offset(bs[DDIR_READ]),
 103                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 104                 .help   = "Block size unit",
 105                 .def    = "4k",
 106         },
 107         {
 108                 .name   = "bsrange",
 109                 .type   = FIO_OPT_RANGE,
 110                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 111                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 112                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 113                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 114                 .help   = "Set block size range (in more detail than bs)",
 115         },
 116         {
 117                 .name   = "bs_unaligned",
 118                 .type   = FIO_OPT_STR_SET,
 119                 .off1   = td_var_offset(bs_unaligned),
 120                 .help   = "Don't sector align IO buffer sizes",
 121         },
 122         {
 123                 .name   = "offset",
 124                 .type   = FIO_OPT_STR_VAL,
 125                 .off1   = td_var_offset(start_offset),
 126                 .help   = "Start IO from this offset",
 127                 .def    = "0",
 128         },
 129         {
 130                 .name   = "randrepeat",
 131                 .type   = FIO_OPT_BOOL,
 132                 .off1   = td_var_offset(rand_repeatable),
 133                 .help   = "Use repeatable random IO pattern",
 134                 .def    = "1",
 135         },
 136         {
 137                 .name   = "norandommap",
 138                 .type   = FIO_OPT_STR_SET,
 139                 .off1   = td_var_offset(norandommap),
 140                 .help   = "Accept potential duplicate random blocks",
 141         },
 142         {
 143                 .name   = "nrfiles",
 144                 .type   = FIO_OPT_INT,
 145                 .off1   = td_var_offset(nr_files),
 146                 .help   = "Split job workload between this number of files",
 147                 .def    = "1",
 148         },
 149         {
 150                 .name   = "fsync",
 151                 .type   = FIO_OPT_INT,
 152                 .off1   = td_var_offset(fsync_blocks),
 153                 .help   = "Issue fsync for writes every given number of blocks",
 154                 .def    = "0",
 155         },
 156         {
 157                 .name   = "direct",
 158                 .type   = FIO_OPT_BOOL,
 159                 .off1   = td_var_offset(odirect),
 160                 .help   = "Use O_DIRECT IO (negates buffered)",
 161                 .def    = "0",
 162         },
 163         {
 164                 .name   = "buffered",
 165                 .type   = FIO_OPT_BOOL,
 166                 .off1   = td_var_offset(odirect),
 167                 .neg    = 1,
 168                 .help   = "Use buffered IO (negates direct)",
 169                 .def    = "1",
 170         },
 171         {
 172                 .name   = "overwrite",
 173                 .type   = FIO_OPT_BOOL,
 174                 .off1   = td_var_offset(overwrite),
 175                 .help   = "When writing, set whether to overwrite current data",
 176                 .def    = "0",
 177         },
 178         {
 179                 .name   = "loops",
 180                 .type   = FIO_OPT_INT,
 181                 .off1   = td_var_offset(loops),
 182                 .help   = "Number of times to run the job",
 183                 .def    = "1",
 184         },
 185         {
 186                 .name   = "numjobs",
 187                 .type   = FIO_OPT_INT,
 188                 .off1   = td_var_offset(numjobs),
 189                 .help   = "Duplicate this job this many times",
 190                 .def    = "1",
 191         },
 192         {
 193                 .name   = "startdelay",
 194                 .type   = FIO_OPT_INT,
 195                 .off1   = td_var_offset(start_delay),
 196                 .help   = "Only start job when this period has passed",
 197                 .def    = "0",
 198         },
 199         {
 200                 .name   = "runtime",
 201                 .alias  = "timeout",
 202                 .type   = FIO_OPT_STR_VAL_TIME,
 203                 .off1   = td_var_offset(timeout),
 204                 .help   = "Stop workload when this amount of time has passed",
 205                 .def    = "0",
 206         },
 207         {
 208                 .name   = "mem",
 209                 .type   = FIO_OPT_STR,
 210                 .cb     = str_mem_cb,
 211                 .help   = "Backing type for IO buffers",
 212                 .def    = "malloc",
 213                 .posval =  { "malloc", "shm", "shmhuge", "mmap", "mmaphuge", },
 214         },
 215         {
 216                 .name   = "verify",
 217                 .type   = FIO_OPT_STR,
 218                 .cb     = str_verify_cb,
 219                 .help   = "Verify sum function",
 220                 .def    = "0",
 221                 .posval = { "crc32", "md5", },
 222         },
 223         {
 224                 .name   = "write_iolog",
 225                 .type   = FIO_OPT_STR_STORE,
 226                 .off1   = td_var_offset(write_iolog_file),
 227                 .help   = "Store IO pattern to file",
 228         },
 229         {
 230                 .name   = "read_iolog",
 231                 .type   = FIO_OPT_STR_STORE,
 232                 .off1   = td_var_offset(read_iolog_file),
 233                 .help   = "Playback IO pattern from file",
 234         },
 235         {
 236                 .name   = "exec_prerun",
 237                 .type   = FIO_OPT_STR_STORE,
 238                 .off1   = td_var_offset(exec_prerun),
 239                 .help   = "Execute this file prior to running job",
 240         },
 241         {
 242                 .name   = "exec_postrun",
 243                 .type   = FIO_OPT_STR_STORE,
 244                 .off1   = td_var_offset(exec_postrun),
 245                 .help   = "Execute this file after running job",
 246         },
 247 #ifdef FIO_HAVE_IOSCHED_SWITCH
 248         {
 249                 .name   = "ioscheduler",
 250                 .type   = FIO_OPT_STR_STORE,
 251                 .off1   = td_var_offset(ioscheduler),
 252                 .help   = "Use this IO scheduler on the backing device",
 253         },
 254 #endif
 255         {
 256                 .name   = "zonesize",
 257                 .type   = FIO_OPT_STR_VAL,
 258                 .off1   = td_var_offset(zone_size),
 259                 .help   = "Give size of an IO zone",
 260                 .def    = "0",
 261         },
 262         {
 263                 .name   = "zoneskip",
 264                 .type   = FIO_OPT_STR_VAL,
 265                 .off1   = td_var_offset(zone_skip),
 266                 .help   = "Space between IO zones",
 267                 .def    = "0",
 268         },
 269         {
 270                 .name   = "lockmem",
 271                 .type   = FIO_OPT_STR_VAL,
 272                 .cb     = str_lockmem_cb,
 273                 .help   = "Lock down this amount of memory",
 274                 .def    = "0",
 275         },
 276         {
 277                 .name   = "rwmixcycle",
 278                 .type   = FIO_OPT_INT,
 279                 .off1   = td_var_offset(rwmixcycle),
 280                 .help   = "Cycle period for mixed read/write workloads (msec)",
 281                 .def    = "500",
 282         },
 283         {
 284                 .name   = "rwmixread",
 285                 .type   = FIO_OPT_INT,
 286                 .off1   = td_var_offset(rwmixread),
 287                 .maxval = 100,
 288                 .help   = "Percentage of mixed workload that is reads",
 289                 .def    = "50",
 290         },
 291         {
 292                 .name   = "rwmixwrite",
 293                 .type   = FIO_OPT_INT,
 294                 .off1   = td_var_offset(rwmixwrite),
 295                 .maxval = 100,
 296                 .help   = "Percentage of mixed workload that is writes",
 297                 .def    = "50",
 298         },
 299         {
 300                 .name   = "nice",
 301                 .type   = FIO_OPT_INT,
 302                 .off1   = td_var_offset(nice),
 303                 .help   = "Set job CPU nice value",
 304                 .minval = -19,
 305                 .maxval = 20,
 306                 .def    = "0",
 307         },
 308 #ifdef FIO_HAVE_IOPRIO
 309         {
 310                 .name   = "prio",
 311                 .type   = FIO_OPT_INT,
 312                 .cb     = str_prio_cb,
 313                 .help   = "Set job IO priority value",
 314                 .minval = 0,
 315                 .maxval = 7,
 316         },
 317         {
 318                 .name   = "prioclass",
 319                 .type   = FIO_OPT_INT,
 320                 .cb     = str_prioclass_cb,
 321                 .help   = "Set job IO priority class",
 322                 .minval = 0,
 323                 .maxval = 3,
 324         },
 325 #endif
 326         {
 327                 .name   = "thinktime",
 328                 .type   = FIO_OPT_INT,
 329                 .off1   = td_var_offset(thinktime),
 330                 .help   = "Idle time between IO buffers (usec)",
 331                 .def    = "0",
 332         },
 333         {
 334                 .name   = "thinktime_blocks",
 335                 .type   = FIO_OPT_INT,
 336                 .off1   = td_var_offset(thinktime_blocks),
 337                 .help   = "IO buffer period between 'thinktime'",
 338                 .def    = "1",
 339         },
 340         {
 341                 .name   = "rate",
 342                 .type   = FIO_OPT_INT,
 343                 .off1   = td_var_offset(rate),
 344                 .help   = "Set bandwidth rate",
 345         },
 346         {
 347                 .name   = "ratemin",
 348                 .type   = FIO_OPT_INT,
 349                 .off1   = td_var_offset(ratemin),
 350                 .help   = "The bottom limit accepted",
 351         },
 352         {
 353                 .name   = "ratecycle",
 354                 .type   = FIO_OPT_INT,
 355                 .off1   = td_var_offset(ratecycle),
 356                 .help   = "Window average for rate limits (msec)",
 357                 .def    = "1000",
 358         },
 359         {
 360                 .name   = "invalidate",
 361                 .type   = FIO_OPT_BOOL,
 362                 .off1   = td_var_offset(invalidate_cache),
 363                 .help   = "Invalidate buffer/page cache prior to running job",
 364                 .def    = "1",
 365         },
 366         {
 367                 .name   = "sync",
 368                 .type   = FIO_OPT_BOOL,
 369                 .off1   = td_var_offset(sync_io),
 370                 .help   = "Use O_SYNC for buffered writes",
 371                 .def    = "0",
 372         },
 373         {
 374                 .name   = "bwavgtime",
 375                 .type   = FIO_OPT_INT,
 376                 .off1   = td_var_offset(bw_avg_time),
 377                 .help   = "Time window over which to calculate bandwidth (msec)",
 378                 .def    = "500",
 379         },
 380         {
 381                 .name   = "create_serialize",
 382                 .type   = FIO_OPT_BOOL,
 383                 .off1   = td_var_offset(create_serialize),
 384                 .help   = "Serialize creating of job files",
 385                 .def    = "1",
 386         },
 387         {
 388                 .name   = "create_fsync",
 389                 .type   = FIO_OPT_BOOL,
 390                 .off1   = td_var_offset(create_fsync),
 391                 .help   = "Fsync file after creation",
 392                 .def    = "1",
 393         },
 394         {
 395                 .name   = "cpuload",
 396                 .type   = FIO_OPT_INT,
 397                 .off1   = td_var_offset(cpuload),
 398                 .help   = "Use this percentage of CPU",
 399         },
 400         {
 401                 .name   = "cpuchunks",
 402                 .type   = FIO_OPT_INT,
 403                 .off1   = td_var_offset(cpucycle),
 404                 .help   = "Length of the CPU burn cycles",
 405         },
 406 #ifdef FIO_HAVE_CPU_AFFINITY
 407         {
 408                 .name   = "cpumask",
 409                 .type   = FIO_OPT_INT,
 410                 .cb     = str_cpumask_cb,
 411                 .help   = "CPU affinity mask",
 412         },
 413 #endif
 414         {
 415                 .name   = "end_fsync",
 416                 .type   = FIO_OPT_BOOL,
 417                 .off1   = td_var_offset(end_fsync),
 418                 .help   = "Include fsync at the end of job",
 419                 .def    = "0",
 420         },
 421         {
 422                 .name   = "unlink",
 423                 .type   = FIO_OPT_BOOL,
 424                 .off1   = td_var_offset(unlink),
 425                 .help   = "Unlink created files after job has completed",
 426                 .def    = "0",
 427         },
 428         {
 429                 .name   = "exitall",
 430                 .type   = FIO_OPT_STR_SET,
 431                 .cb     = str_exitall_cb,
 432                 .help   = "Terminate all jobs when one exits",
 433         },
 434         {
 435                 .name   = "stonewall",
 436                 .type   = FIO_OPT_STR_SET,
 437                 .off1   = td_var_offset(stonewall),
 438                 .help   = "Insert a hard barrier between this job and previous",
 439         },
 440         {
 441                 .name   = "thread",
 442                 .type   = FIO_OPT_STR_SET,
 443                 .off1   = td_var_offset(use_thread),
 444                 .help   = "Use threads instead of forks",
 445         },
 446         {
 447                 .name   = "write_bw_log",
 448                 .type   = FIO_OPT_STR_SET,
 449                 .off1   = td_var_offset(write_bw_log),
 450                 .help   = "Write log of bandwidth during run",
 451         },
 452         {
 453                 .name   = "write_lat_log",
 454                 .type   = FIO_OPT_STR_SET,
 455                 .off1   = td_var_offset(write_lat_log),
 456                 .help   = "Write log of latency during run",
 457         },
 458         {
 459                 .name   = "hugepage-size",
 460                 .type   = FIO_OPT_STR_VAL,
 461                 .off1   = td_var_offset(hugepage_size),
 462                 .help   = "When using hugepages, specify size of each page",
 463                 .def    = __stringify(FIO_HUGE_PAGE),
 464         },
 465         {
 466                 .name = NULL,
 467         },
 468 };
 469
 470 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 471 #define FIO_CMD_OPTS    (16)
 472 #define FIO_GETOPT_JOB  (0x89988998)
 473
 474 /*
 475  * Command line options. These will contain the above, plus a few
 476  * extra that only pertain to fio itself and not jobs.
 477  */
 478 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 479         {
 480                 .name           = "output",
 481                 .has_arg        = required_argument,
 482                 .val            = 'o',
 483         },
 484         {
 485                 .name           = "timeout",
 486                 .has_arg        = required_argument,
 487                 .val            = 't',
 488         },
 489         {
 490                 .name           = "latency-log",
 491                 .has_arg        = required_argument,
 492                 .val            = 'l',
 493         },
 494         {
 495                 .name           = "bandwidth-log",
 496                 .has_arg        = required_argument,
 497                 .val            = 'b',
 498         },
 499         {
 500                 .name           = "minimal",
 501                 .has_arg        = optional_argument,
 502                 .val            = 'm',
 503         },
 504         {
 505                 .name           = "version",
 506                 .has_arg        = no_argument,
 507                 .val            = 'v',
 508         },
 509         {
 510                 .name           = "help",
 511                 .has_arg        = no_argument,
 512                 .val            = 'h',
 513         },
 514         {
 515                 .name           = "cmdhelp",
 516                 .has_arg        = required_argument,
 517                 .val            = 'c',
 518         },
 519         {
 520                 .name           = NULL,
 521         },
 522 };
 523
 524 static int def_timeout = 0;
 525
 526 static char fio_version_string[] = "fio 1.11";
 527
 528 static char **ini_file;
 529 static int max_jobs = MAX_JOBS;
 530
 531 struct thread_data def_thread;
 532 struct thread_data *threads = NULL;
 533
 534 int exitall_on_terminate = 0;
 535 int terse_output = 0;
 536 unsigned long long mlock_size = 0;
 537 FILE *f_out = NULL;
 538 FILE *f_err = NULL;
 539
 540 static int write_lat_log = 0;
 541 int write_bw_log = 0;
 542
 543 /*
 544  * Return a free job structure.
 545  */
 546 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 547 {
 548         struct thread_data *td;
 549
 550         if (global)
 551                 return &def_thread;
 552         if (thread_number >= max_jobs)
 553                 return NULL;
 554
 555         td = &threads[thread_number++];
 556         *td = *parent;
 557
 558         td->thread_number = thread_number;
 559         return td;
 560 }
 561
 562 static void put_job(struct thread_data *td)
 563 {
 564         if (td == &def_thread)
 565                 return;
 566
 567         if (td->error)
 568                 fprintf(f_out, "fio: %s\n", td->verror);
 569
 570         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 571         thread_number--;
 572 }
 573
 574 /*
 575  * Lazy way of fixing up options that depend on each other. We could also
 576  * define option callback handlers, but this is easier.
 577  */
 578 static void fixup_options(struct thread_data *td)
 579 {
 580         if (!td->rwmixread && td->rwmixwrite)
 581                 td->rwmixread = 100 - td->rwmixwrite;
 582
 583         if (td->write_iolog_file && td->read_iolog_file) {
 584                 log_err("fio: read iolog overrides write_iolog\n");
 585                 free(td->write_iolog_file);
 586                 td->write_iolog_file = NULL;
 587         }
 588
 589         if (td->io_ops->flags & FIO_SYNCIO)
 590                 td->iodepth = 1;
 591         else {
 592                 if (!td->iodepth)
 593                         td->iodepth = td->nr_files;
 594         }
 595
 596         /*
 597          * only really works for sequential io for now, and with 1 file
 598          */
 599         if (td->zone_size && !td->sequential && td->nr_files == 1)
 600                 td->zone_size = 0;
 601
 602         /*
 603          * Reads can do overwrites, we always need to pre-create the file
 604          */
 605         if (td_read(td) || td_rw(td))
 606                 td->overwrite = 1;
 607
 608         if (!td->min_bs[DDIR_READ])
 609                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 610         if (!td->max_bs[DDIR_READ])
 611                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 612         if (!td->min_bs[DDIR_WRITE])
 613                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 614         if (!td->max_bs[DDIR_WRITE])
 615                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 616
 617         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 618
 619         if (td_read(td) && !td_rw(td))
 620                 td->verify = 0;
 621
 622         if (td->norandommap && td->verify != VERIFY_NONE) {
 623                 log_err("fio: norandommap given, verify disabled\n");
 624                 td->verify = VERIFY_NONE;
 625         }
 626         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 627                 log_err("fio: bs_unaligned may not work with raw io\n");
 628
 629         /*
 630          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 631          */
 632         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 633                 td->odirect = 0;
 634 }
 635
 636 /*
 637  * This function leaks the buffer
 638  */
 639 static char *to_kmg(unsigned int val)
 640 {
 641         char *buf = malloc(32);
 642         char post[] = { 0, 'K', 'M', 'G', 'P', 0 };
 643         char *p = post;
 644
 645         do {
 646                 if (val & 1023)
 647                         break;
 648
 649                 val >>= 10;
 650                 p++;
 651         } while (*p);
 652
 653         snprintf(buf, 31, "%u%c", val, *p);
 654         return buf;
 655 }
 656
 657 /*
 658  * Adds a job to the list of things todo. Sanitizes the various options
 659  * to make sure we don't have conflicts, and initializes various
 660  * members of td.
 661  */
 662 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 663 {
 664         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
 665                                    "rw", NULL, "randrw" };
 666         struct stat sb;
 667         int numjobs, ddir, i;
 668         struct fio_file *f;
 669
 670         /*
 671          * the def_thread is just for options, it's not a real job
 672          */
 673         if (td == &def_thread)
 674                 return 0;
 675
 676         assert(td->io_ops);
 677
 678         if (td->odirect)
 679                 td->io_ops->flags |= FIO_RAWIO;
 680
 681         td->filetype = FIO_TYPE_FILE;
 682         if (!stat(jobname, &sb)) {
 683                 if (S_ISBLK(sb.st_mode))
 684                         td->filetype = FIO_TYPE_BD;
 685                 else if (S_ISCHR(sb.st_mode))
 686                         td->filetype = FIO_TYPE_CHAR;
 687         }
 688
 689         fixup_options(td);
 690
 691         if (td->filename)
 692                 td->nr_uniq_files = 1;
 693         else
 694                 td->nr_uniq_files = td->nr_files;
 695
 696         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 697                 char tmp[PATH_MAX];
 698                 int len = 0;
 699
 700                 if (td->directory && td->directory[0] != '\0') {
 701                         if (lstat(td->directory, &sb) < 0) {
 702                                 log_err("fio: %s is not a directory\n", td->directory);
 703                                 td_verror(td, errno);
 704                                 return 1;
 705                         }
 706                         if (!S_ISDIR(sb.st_mode)) {
 707                                 log_err("fio: %s is not a directory\n", td->directory);
 708                                 return 1;
 709                         }
 710                         len = sprintf(tmp, "%s/", td->directory);
 711                 }
 712
 713                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 714
 715                 for_each_file(td, f, i) {
 716                         memset(f, 0, sizeof(*f));
 717                         f->fd = -1;
 718
 719                         if (td->filename)
 720                                 sprintf(tmp + len, "%s", td->filename);
 721                         else
 722                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 723                         f->file_name = strdup(tmp);
 724                 }
 725         } else {
 726                 td->nr_files = 1;
 727                 td->files = malloc(sizeof(struct fio_file));
 728                 f = &td->files[0];
 729
 730                 memset(f, 0, sizeof(*f));
 731                 f->fd = -1;
 732                 f->file_name = strdup(jobname);
 733         }
 734
 735         for_each_file(td, f, i) {
 736                 f->file_size = td->total_file_size / td->nr_files;
 737                 f->file_offset = td->start_offset;
 738         }
 739
 740         fio_sem_init(&td->mutex, 0);
 741
 742         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
 743         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
 744         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
 745
 746         if (td->stonewall && td->thread_number > 1)
 747                 groupid++;
 748
 749         td->groupid = groupid;
 750
 751         if (setup_rate(td))
 752                 goto err;
 753
 754         if (td->write_lat_log) {
 755                 setup_log(&td->slat_log);
 756                 setup_log(&td->clat_log);
 757         }
 758         if (td->write_bw_log)
 759                 setup_log(&td->bw_log);
 760
 761         if (!td->name)
 762                 td->name = strdup(jobname);
 763
 764         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 765
 766         if (!terse_output) {
 767                 if (!job_add_num) {
 768                         if (td->io_ops->flags & FIO_CPUIO)
 769                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 770                         else {
 771                                 char *c1, *c2, *c3, *c4;
 772
 773                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 774                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 775                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 776                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 777
 778                                 fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
 779
 780                                 free(c1);
 781                                 free(c2);
 782                                 free(c3);
 783                                 free(c4);
 784                         }
 785                 } else if (job_add_num == 1)
 786                         fprintf(f_out, "...\n");
 787         }
 788
 789         /*
 790          * recurse add identical jobs, clear numjobs and stonewall options
 791          * as they don't apply to sub-jobs
 792          */
 793         numjobs = td->numjobs;
 794         while (--numjobs) {
 795                 struct thread_data *td_new = get_new_job(0, td);
 796
 797                 if (!td_new)
 798                         goto err;
 799
 800                 td_new->numjobs = 1;
 801                 td_new->stonewall = 0;
 802                 job_add_num = numjobs - 1;
 803
 804                 if (add_job(td_new, jobname, job_add_num))
 805                         goto err;
 806         }
 807         return 0;
 808 err:
 809         put_job(td);
 810         return -1;
 811 }
 812
 813 /*
 814  * Initialize the various random states we need (random io, block size ranges,
 815  * read/write mix, etc).
 816  */
 817 int init_random_state(struct thread_data *td)
 818 {
 819         unsigned long seeds[4];
 820         int fd, num_maps, blocks, i;
 821         struct fio_file *f;
 822
 823         if (td->io_ops->flags & FIO_CPUIO)
 824                 return 0;
 825
 826         fd = open("/dev/urandom", O_RDONLY);
 827         if (fd == -1) {
 828                 td_verror(td, errno);
 829                 return 1;
 830         }
 831
 832         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 833                 td_verror(td, EIO);
 834                 close(fd);
 835                 return 1;
 836         }
 837
 838         close(fd);
 839
 840         os_random_seed(seeds[0], &td->bsrange_state);
 841         os_random_seed(seeds[1], &td->verify_state);
 842         os_random_seed(seeds[2], &td->rwmix_state);
 843
 844         if (td->sequential)
 845                 return 0;
 846
 847         if (td->rand_repeatable)
 848                 seeds[3] = FIO_RANDSEED;
 849
 850         if (!td->norandommap) {
 851                 for_each_file(td, f, i) {
 852                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 853                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 854                         f->file_map = malloc(num_maps * sizeof(long));
 855                         f->num_maps = num_maps;
 856                         memset(f->file_map, 0, num_maps * sizeof(long));
 857                 }
 858         }
 859
 860         os_random_seed(seeds[3], &td->random_state);
 861         return 0;
 862 }
 863
 864 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 865 {
 866 #ifdef FIO_HAVE_CPU_AFFINITY
 867         unsigned int i;
 868
 869         CPU_ZERO(&cpumask);
 870
 871         for (i = 0; i < sizeof(int) * 8; i++) {
 872                 if ((1 << i) & cpu)
 873                         CPU_SET(i, &cpumask);
 874         }
 875 #endif
 876 }
 877
 878 static int is_empty_or_comment(char *line)
 879 {
 880         unsigned int i;
 881
 882         for (i = 0; i < strlen(line); i++) {
 883                 if (line[i] == ';')
 884                         return 1;
 885                 if (!isspace(line[i]) && !iscntrl(line[i]))
 886                         return 0;
 887         }
 888
 889         return 1;
 890 }
 891
 892 static int str_rw_cb(void *data, const char *mem)
 893 {
 894         struct thread_data *td = data;
 895
 896         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 897                 td->ddir = DDIR_READ;
 898                 td->sequential = 1;
 899                 return 0;
 900         } else if (!strncmp(mem, "randread", 8)) {
 901                 td->ddir = DDIR_READ;
 902                 td->sequential = 0;
 903                 return 0;
 904         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 905                 td->ddir = DDIR_WRITE;
 906                 td->sequential = 1;
 907                 return 0;
 908         } else if (!strncmp(mem, "randwrite", 9)) {
 909                 td->ddir = DDIR_WRITE;
 910                 td->sequential = 0;
 911                 return 0;
 912         } else if (!strncmp(mem, "rw", 2)) {
 913                 td->ddir = DDIR_READ;
 914                 td->iomix = 1;
 915                 td->sequential = 1;
 916                 return 0;
 917         } else if (!strncmp(mem, "randrw", 6)) {
 918                 td->ddir = DDIR_READ;
 919                 td->iomix = 1;
 920                 td->sequential = 0;
 921                 return 0;
 922         }
 923
 924         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 925         return 1;
 926 }
 927
 928 static int str_verify_cb(void *data, const char *mem)
 929 {
 930         struct thread_data *td = data;
 931
 932         if (!strncmp(mem, "0", 1)) {
 933                 td->verify = VERIFY_NONE;
 934                 return 0;
 935         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 936                 td->verify = VERIFY_MD5;
 937                 return 0;
 938         } else if (!strncmp(mem, "crc32", 5)) {
 939                 td->verify = VERIFY_CRC32;
 940                 return 0;
 941         }
 942
 943         log_err("fio: verify types: md5, crc32\n");
 944         return 1;
 945 }
 946
 947 /*
 948  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
 949  */
 950 static char *get_mmap_file(const char *str)
 951 {
 952         char *p = strstr(str, ":");
 953
 954         if (!p)
 955                 return NULL;
 956
 957         p++;
 958         strip_blank_front(&p);
 959         strip_blank_end(p);
 960         return strdup(p);
 961 }
 962
 963 static int str_mem_cb(void *data, const char *mem)
 964 {
 965         struct thread_data *td = data;
 966
 967         if (!strncmp(mem, "malloc", 6)) {
 968                 td->mem_type = MEM_MALLOC;
 969                 return 0;
 970         } else if (!strncmp(mem, "mmaphuge", 8)) {
 971 #ifdef FIO_HAVE_HUGETLB
 972                 /*
 973                  * mmaphuge must be appended with the actual file
 974                  */
 975                 td->mmapfile = get_mmap_file(mem);
 976                 if (!td->mmapfile) {
 977                         log_err("fio: mmaphuge:/path/to/file\n");
 978                         return 1;
 979                 }
 980
 981                 td->mem_type = MEM_MMAPHUGE;
 982                 return 0;
 983 #else
 984                 log_err("fio: mmaphuge not available\n");
 985                 return 1;
 986 #endif
 987         } else if (!strncmp(mem, "mmap", 4)) {
 988                 /*
 989                  * Check if the user wants file backed memory. It's ok
 990                  * if there's no file given, we'll just use anon mamp then.
 991                  */
 992                 td->mmapfile = get_mmap_file(mem);
 993                 td->mem_type = MEM_MMAP;
 994                 return 0;
 995         } else if (!strncmp(mem, "shmhuge", 7)) {
 996 #ifdef FIO_HAVE_HUGETLB
 997                 td->mem_type = MEM_SHMHUGE;
 998                 return 0;
 999 #else
1000                 log_err("fio: shmhuge not available\n");
1001                 return 1;
1002 #endif
1003         } else if (!strncmp(mem, "shm", 3)) {
1004                 td->mem_type = MEM_SHM;
1005                 return 0;
1006         }
1007
1008         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
1009         return 1;
1010 }
1011
1012 static int str_ioengine_cb(void *data, const char *str)
1013 {
1014         struct thread_data *td = data;
1015
1016         td->io_ops = load_ioengine(td, str);
1017         if (td->io_ops)
1018                 return 0;
1019
1020         log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
1021         log_err("fio: or specify path to dynamic ioengine module\n");
1022         return 1;
1023 }
1024
1025 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1026 {
1027         mlock_size = *val;
1028         return 0;
1029 }
1030
1031 #ifdef FIO_HAVE_IOPRIO
1032 static int str_prioclass_cb(void *data, unsigned int *val)
1033 {
1034         struct thread_data *td = data;
1035
1036         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1037         return 0;
1038 }
1039
1040 static int str_prio_cb(void *data, unsigned int *val)
1041 {
1042         struct thread_data *td = data;
1043
1044         td->ioprio |= *val;
1045         return 0;
1046 }
1047 #endif
1048
1049 static int str_exitall_cb(void)
1050 {
1051         exitall_on_terminate = 1;
1052         return 0;
1053 }
1054
1055 static int str_cpumask_cb(void *data, unsigned int *val)
1056 {
1057         struct thread_data *td = data;
1058
1059         fill_cpu_mask(td->cpumask, *val);
1060         return 0;
1061 }
1062
1063 /*
1064  * This is our [ini] type file parser.
1065  */
1066 static int parse_jobs_ini(char *file, int stonewall_flag)
1067 {
1068         unsigned int global;
1069         struct thread_data *td;
1070         char *string, *name;
1071         fpos_t off;
1072         FILE *f;
1073         char *p;
1074         int ret = 0, stonewall;
1075
1076         f = fopen(file, "r");
1077         if (!f) {
1078                 perror("fopen job file");
1079                 return 1;
1080         }
1081
1082         string = malloc(4096);
1083         name = malloc(256);
1084         memset(name, 0, 256);
1085
1086         stonewall = stonewall_flag;
1087         do {
1088                 p = fgets(string, 4095, f);
1089                 if (!p)
1090                         break;
1091                 if (is_empty_or_comment(p))
1092                         continue;
1093                 if (sscanf(p, "[%255s]", name) != 1)
1094                         continue;
1095
1096                 global = !strncmp(name, "global", 6);
1097
1098                 name[strlen(name) - 1] = '\0';
1099
1100                 td = get_new_job(global, &def_thread);
1101                 if (!td) {
1102                         ret = 1;
1103                         break;
1104                 }
1105
1106                 /*
1107                  * Seperate multiple job files by a stonewall
1108                  */
1109                 if (!global && stonewall) {
1110                         td->stonewall = stonewall;
1111                         stonewall = 0;
1112                 }
1113
1114                 fgetpos(f, &off);
1115                 while ((p = fgets(string, 4096, f)) != NULL) {
1116                         if (is_empty_or_comment(p))
1117                                 continue;
1118
1119                         strip_blank_front(&p);
1120
1121                         if (p[0] == '[')
1122                                 break;
1123
1124                         strip_blank_end(p);
1125
1126                         fgetpos(f, &off);
1127
1128                         /*
1129                          * Don't break here, continue parsing options so we
1130                          * dump all the bad ones. Makes trial/error fixups
1131                          * easier on the user.
1132                          */
1133                         ret |= parse_option(p, options, td);
1134                 }
1135
1136                 if (!ret) {
1137                         fsetpos(f, &off);
1138                         ret = add_job(td, name, 0);
1139                 } else {
1140                         log_err("fio: job %s dropped\n", name);
1141                         put_job(td);
1142                 }
1143         } while (!ret);
1144
1145         free(string);
1146         free(name);
1147         fclose(f);
1148         return ret;
1149 }
1150
1151 static int fill_def_thread(void)
1152 {
1153         memset(&def_thread, 0, sizeof(def_thread));
1154
1155         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1156                 perror("sched_getaffinity");
1157                 return 1;
1158         }
1159
1160         /*
1161          * fill default options
1162          */
1163         fill_default_options(&def_thread, options);
1164
1165         def_thread.timeout = def_timeout;
1166         def_thread.write_bw_log = write_bw_log;
1167         def_thread.write_lat_log = write_lat_log;
1168
1169 #ifdef FIO_HAVE_DISK_UTIL
1170         def_thread.do_disk_util = 1;
1171 #endif
1172
1173         return 0;
1174 }
1175
1176 static void usage(void)
1177 {
1178         printf("%s\n", fio_version_string);
1179         printf("\t--output\tWrite output to file\n");
1180         printf("\t--timeout\tRuntime in seconds\n");
1181         printf("\t--latency-log\tGenerate per-job latency logs\n");
1182         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1183         printf("\t--minimal\tMinimal (terse) output\n");
1184         printf("\t--version\tPrint version info and exit\n");
1185         printf("\t--help\t\tPrint this page\n");
1186         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1187 }
1188
1189 static int parse_cmd_line(int argc, char *argv[])
1190 {
1191         struct thread_data *td = NULL;
1192         int c, ini_idx = 0, lidx, ret;
1193
1194         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1195                 switch (c) {
1196                 case 't':
1197                         def_timeout = atoi(optarg);
1198                         break;
1199                 case 'l':
1200                         write_lat_log = 1;
1201                         break;
1202                 case 'w':
1203                         write_bw_log = 1;
1204                         break;
1205                 case 'o':
1206                         f_out = fopen(optarg, "w+");
1207                         if (!f_out) {
1208                                 perror("fopen output");
1209                                 exit(1);
1210                         }
1211                         f_err = f_out;
1212                         break;
1213                 case 'm':
1214                         terse_output = 1;
1215                         break;
1216                 case 'h':
1217                         usage();
1218                         exit(0);
1219                 case 'c':
1220                         ret = show_cmd_help(options, optarg);
1221                         exit(ret);
1222                 case 'v':
1223                         printf("%s\n", fio_version_string);
1224                         exit(0);
1225                 case FIO_GETOPT_JOB: {
1226                         const char *opt = long_options[lidx].name;
1227                         char *val = optarg;
1228
1229                         if (!strncmp(opt, "name", 4) && td) {
1230                                 ret = add_job(td, td->name ?: "fio", 0);
1231                                 if (ret) {
1232                                         put_job(td);
1233                                         return 0;
1234                                 }
1235                                 td = NULL;
1236                         }
1237                         if (!td) {
1238                                 int global = !strncmp(val, "global", 6);
1239
1240                                 td = get_new_job(global, &def_thread);
1241                                 if (!td)
1242                                         return 0;
1243                         }
1244
1245                         ret = parse_cmd_option(opt, val, options, td);
1246                         if (ret) {
1247                                 log_err("fio: job dropped\n");
1248                                 put_job(td);
1249                                 td = NULL;
1250                         }
1251                         break;
1252                 }
1253                 default:
1254                         break;
1255                 }
1256         }
1257
1258         if (td) {
1259                 ret = add_job(td, td->name ?: "fio", 0);
1260                 if (ret)
1261                         put_job(td);
1262         }
1263
1264         while (optind < argc) {
1265                 ini_idx++;
1266                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1267                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1268                 optind++;
1269         }
1270
1271         return ini_idx;
1272 }
1273
1274 static void free_shm(void)
1275 {
1276         struct shmid_ds sbuf;
1277
1278         if (threads) {
1279                 shmdt((void *) threads);
1280                 threads = NULL;
1281                 shmctl(shm_id, IPC_RMID, &sbuf);
1282         }
1283 }
1284
1285 /*
1286  * The thread area is shared between the main process and the job
1287  * threads/processes. So setup a shared memory segment that will hold
1288  * all the job info.
1289  */
1290 static int setup_thread_area(void)
1291 {
1292         /*
1293          * 1024 is too much on some machines, scale max_jobs if
1294          * we get a failure that looks like too large a shm segment
1295          */
1296         do {
1297                 size_t size = max_jobs * sizeof(struct thread_data);
1298
1299                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1300                 if (shm_id != -1)
1301                         break;
1302                 if (errno != EINVAL) {
1303                         perror("shmget");
1304                         break;
1305                 }
1306
1307                 max_jobs >>= 1;
1308         } while (max_jobs);
1309
1310         if (shm_id == -1)
1311                 return 1;
1312
1313         threads = shmat(shm_id, NULL, 0);
1314         if (threads == (void *) -1) {
1315                 perror("shmat");
1316                 return 1;
1317         }
1318
1319         atexit(free_shm);
1320         return 0;
1321 }
1322
1323 /*
1324  * Copy the fio options into the long options map, so we mirror
1325  * job and cmd line options.
1326  */
1327 static void dupe_job_options(void)
1328 {
1329         struct fio_option *o;
1330         unsigned int i;
1331
1332         i = 0;
1333         while (long_options[i].name)
1334                 i++;
1335
1336         o = &options[0];
1337         while (o->name) {
1338                 long_options[i].name = o->name;
1339                 long_options[i].val = FIO_GETOPT_JOB;
1340                 if (o->type == FIO_OPT_STR_SET)
1341                         long_options[i].has_arg = no_argument;
1342                 else
1343                         long_options[i].has_arg = required_argument;
1344
1345                 i++;
1346                 o++;
1347                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1348         }
1349 }
1350
1351 int parse_options(int argc, char *argv[])
1352 {
1353         int job_files, i;
1354
1355         f_out = stdout;
1356         f_err = stderr;
1357
1358         options_init(options);
1359
1360         dupe_job_options();
1361
1362         if (setup_thread_area())
1363                 return 1;
1364         if (fill_def_thread())
1365                 return 1;
1366
1367         job_files = parse_cmd_line(argc, argv);
1368
1369         for (i = 0; i < job_files; i++) {
1370                 if (fill_def_thread())
1371                         return 1;
1372                 if (parse_jobs_ini(ini_file[i], i))
1373                         return 1;
1374                 free(ini_file[i]);
1375         }
1376
1377         free(ini_file);
1378
1379         if (!thread_number) {
1380                 log_err("No jobs defined(s)\n");
1381                 return 1;
1382         }
1383
1384         return 0;
1385 }