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