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 static int str_file_service_cb(void *, const char *);
  37
  38 #define __stringify_1(x)        #x
  39 #define __stringify(x)          __stringify_1(x)
  40
  41 /*
  42  * Map of job/command line options
  43  */
  44 static struct fio_option options[] = {
  45         {
  46                 .name   = "description",
  47                 .type   = FIO_OPT_STR_STORE,
  48                 .off1   = td_var_offset(description),
  49                 .help   = "Text job description",
  50         },
  51         {
  52                 .name   = "name",
  53                 .type   = FIO_OPT_STR_STORE,
  54                 .off1   = td_var_offset(name),
  55                 .help   = "Name of this job",
  56         },
  57         {
  58                 .name   = "directory",
  59                 .type   = FIO_OPT_STR_STORE,
  60                 .off1   = td_var_offset(directory),
  61                 .help   = "Directory to store files in",
  62         },
  63         {
  64                 .name   = "filename",
  65                 .type   = FIO_OPT_STR_STORE,
  66                 .off1   = td_var_offset(filename),
  67                 .help   = "Force the use of a specific file",
  68         },
  69         {
  70                 .name   = "rw",
  71                 .type   = FIO_OPT_STR,
  72                 .cb     = str_rw_cb,
  73                 .help   = "IO direction",
  74                 .def    = "read",
  75                 .posval = { "read", "write", "randwrite", "randread", "rw",
  76                                 "randrw", },
  77         },
  78         {
  79                 .name   = "ioengine",
  80                 .type   = FIO_OPT_STR,
  81                 .cb     = str_ioengine_cb,
  82                 .help   = "IO engine to use",
  83                 .def    = "sync",
  84                 .posval = { "sync", "libaio", "posixaio", "mmap", "splice",
  85                                 "sg", "null", "net", "syslet-rw" },
  86         },
  87         {
  88                 .name   = "iodepth",
  89                 .type   = FIO_OPT_INT,
  90                 .off1   = td_var_offset(iodepth),
  91                 .help   = "Amount of IO buffers to keep in flight",
  92                 .def    = "1",
  93         },
  94         {
  95                 .name   = "iodepth_low",
  96                 .type   = FIO_OPT_INT,
  97                 .off1   = td_var_offset(iodepth_low),
  98                 .help   = "Low water mark for queuing depth",
  99         },
 100         {
 101                 .name   = "size",
 102                 .type   = FIO_OPT_STR_VAL,
 103                 .off1   = td_var_offset(total_file_size),
 104                 .help   = "Size of device or file",
 105         },
 106         {
 107                 .name   = "bs",
 108                 .type   = FIO_OPT_STR_VAL_INT,
 109                 .off1   = td_var_offset(bs[DDIR_READ]),
 110                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 111                 .help   = "Block size unit",
 112                 .def    = "4k",
 113         },
 114         {
 115                 .name   = "bsrange",
 116                 .type   = FIO_OPT_RANGE,
 117                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 118                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 119                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 120                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 121                 .help   = "Set block size range (in more detail than bs)",
 122         },
 123         {
 124                 .name   = "bs_unaligned",
 125                 .type   = FIO_OPT_STR_SET,
 126                 .off1   = td_var_offset(bs_unaligned),
 127                 .help   = "Don't sector align IO buffer sizes",
 128         },
 129         {
 130                 .name   = "offset",
 131                 .type   = FIO_OPT_STR_VAL,
 132                 .off1   = td_var_offset(start_offset),
 133                 .help   = "Start IO from this offset",
 134                 .def    = "0",
 135         },
 136         {
 137                 .name   = "randrepeat",
 138                 .type   = FIO_OPT_BOOL,
 139                 .off1   = td_var_offset(rand_repeatable),
 140                 .help   = "Use repeatable random IO pattern",
 141                 .def    = "1",
 142         },
 143         {
 144                 .name   = "norandommap",
 145                 .type   = FIO_OPT_STR_SET,
 146                 .off1   = td_var_offset(norandommap),
 147                 .help   = "Accept potential duplicate random blocks",
 148         },
 149         {
 150                 .name   = "nrfiles",
 151                 .type   = FIO_OPT_INT,
 152                 .off1   = td_var_offset(nr_files),
 153                 .help   = "Split job workload between this number of files",
 154                 .def    = "1",
 155         },
 156         {
 157                 .name   = "file_service_type",
 158                 .type   = FIO_OPT_STR,
 159                 .cb     = str_file_service_cb,
 160                 .help   = "How to select which file to service next",
 161                 .def    = "roundrobin",
 162                 .posval = { "random", "roundrobin" },
 163         },
 164         {
 165                 .name   = "fsync",
 166                 .type   = FIO_OPT_INT,
 167                 .off1   = td_var_offset(fsync_blocks),
 168                 .help   = "Issue fsync for writes every given number of blocks",
 169                 .def    = "0",
 170         },
 171         {
 172                 .name   = "direct",
 173                 .type   = FIO_OPT_BOOL,
 174                 .off1   = td_var_offset(odirect),
 175                 .help   = "Use O_DIRECT IO (negates buffered)",
 176                 .def    = "0",
 177         },
 178         {
 179                 .name   = "buffered",
 180                 .type   = FIO_OPT_BOOL,
 181                 .off1   = td_var_offset(odirect),
 182                 .neg    = 1,
 183                 .help   = "Use buffered IO (negates direct)",
 184                 .def    = "1",
 185         },
 186         {
 187                 .name   = "overwrite",
 188                 .type   = FIO_OPT_BOOL,
 189                 .off1   = td_var_offset(overwrite),
 190                 .help   = "When writing, set whether to overwrite current data",
 191                 .def    = "0",
 192         },
 193         {
 194                 .name   = "loops",
 195                 .type   = FIO_OPT_INT,
 196                 .off1   = td_var_offset(loops),
 197                 .help   = "Number of times to run the job",
 198                 .def    = "1",
 199         },
 200         {
 201                 .name   = "numjobs",
 202                 .type   = FIO_OPT_INT,
 203                 .off1   = td_var_offset(numjobs),
 204                 .help   = "Duplicate this job this many times",
 205                 .def    = "1",
 206         },
 207         {
 208                 .name   = "startdelay",
 209                 .type   = FIO_OPT_INT,
 210                 .off1   = td_var_offset(start_delay),
 211                 .help   = "Only start job when this period has passed",
 212                 .def    = "0",
 213         },
 214         {
 215                 .name   = "runtime",
 216                 .alias  = "timeout",
 217                 .type   = FIO_OPT_STR_VAL_TIME,
 218                 .off1   = td_var_offset(timeout),
 219                 .help   = "Stop workload when this amount of time has passed",
 220                 .def    = "0",
 221         },
 222         {
 223                 .name   = "mem",
 224                 .type   = FIO_OPT_STR,
 225                 .cb     = str_mem_cb,
 226                 .help   = "Backing type for IO buffers",
 227                 .def    = "malloc",
 228                 .posval =  { "malloc", "shm", "shmhuge", "mmap", "mmaphuge", },
 229         },
 230         {
 231                 .name   = "verify",
 232                 .type   = FIO_OPT_STR,
 233                 .cb     = str_verify_cb,
 234                 .help   = "Verify sum function",
 235                 .def    = "0",
 236                 .posval = { "crc32", "md5", },
 237         },
 238         {
 239                 .name   = "write_iolog",
 240                 .type   = FIO_OPT_STR_STORE,
 241                 .off1   = td_var_offset(write_iolog_file),
 242                 .help   = "Store IO pattern to file",
 243         },
 244         {
 245                 .name   = "read_iolog",
 246                 .type   = FIO_OPT_STR_STORE,
 247                 .off1   = td_var_offset(read_iolog_file),
 248                 .help   = "Playback IO pattern from file",
 249         },
 250         {
 251                 .name   = "exec_prerun",
 252                 .type   = FIO_OPT_STR_STORE,
 253                 .off1   = td_var_offset(exec_prerun),
 254                 .help   = "Execute this file prior to running job",
 255         },
 256         {
 257                 .name   = "exec_postrun",
 258                 .type   = FIO_OPT_STR_STORE,
 259                 .off1   = td_var_offset(exec_postrun),
 260                 .help   = "Execute this file after running job",
 261         },
 262 #ifdef FIO_HAVE_IOSCHED_SWITCH
 263         {
 264                 .name   = "ioscheduler",
 265                 .type   = FIO_OPT_STR_STORE,
 266                 .off1   = td_var_offset(ioscheduler),
 267                 .help   = "Use this IO scheduler on the backing device",
 268         },
 269 #endif
 270         {
 271                 .name   = "zonesize",
 272                 .type   = FIO_OPT_STR_VAL,
 273                 .off1   = td_var_offset(zone_size),
 274                 .help   = "Give size of an IO zone",
 275                 .def    = "0",
 276         },
 277         {
 278                 .name   = "zoneskip",
 279                 .type   = FIO_OPT_STR_VAL,
 280                 .off1   = td_var_offset(zone_skip),
 281                 .help   = "Space between IO zones",
 282                 .def    = "0",
 283         },
 284         {
 285                 .name   = "lockmem",
 286                 .type   = FIO_OPT_STR_VAL,
 287                 .cb     = str_lockmem_cb,
 288                 .help   = "Lock down this amount of memory",
 289                 .def    = "0",
 290         },
 291         {
 292                 .name   = "rwmixcycle",
 293                 .type   = FIO_OPT_INT,
 294                 .off1   = td_var_offset(rwmixcycle),
 295                 .help   = "Cycle period for mixed read/write workloads (msec)",
 296                 .def    = "500",
 297         },
 298         {
 299                 .name   = "rwmixread",
 300                 .type   = FIO_OPT_INT,
 301                 .off1   = td_var_offset(rwmixread),
 302                 .maxval = 100,
 303                 .help   = "Percentage of mixed workload that is reads",
 304                 .def    = "50",
 305         },
 306         {
 307                 .name   = "rwmixwrite",
 308                 .type   = FIO_OPT_INT,
 309                 .off1   = td_var_offset(rwmixwrite),
 310                 .maxval = 100,
 311                 .help   = "Percentage of mixed workload that is writes",
 312                 .def    = "50",
 313         },
 314         {
 315                 .name   = "nice",
 316                 .type   = FIO_OPT_INT,
 317                 .off1   = td_var_offset(nice),
 318                 .help   = "Set job CPU nice value",
 319                 .minval = -19,
 320                 .maxval = 20,
 321                 .def    = "0",
 322         },
 323 #ifdef FIO_HAVE_IOPRIO
 324         {
 325                 .name   = "prio",
 326                 .type   = FIO_OPT_INT,
 327                 .cb     = str_prio_cb,
 328                 .help   = "Set job IO priority value",
 329                 .minval = 0,
 330                 .maxval = 7,
 331         },
 332         {
 333                 .name   = "prioclass",
 334                 .type   = FIO_OPT_INT,
 335                 .cb     = str_prioclass_cb,
 336                 .help   = "Set job IO priority class",
 337                 .minval = 0,
 338                 .maxval = 3,
 339         },
 340 #endif
 341         {
 342                 .name   = "thinktime",
 343                 .type   = FIO_OPT_INT,
 344                 .off1   = td_var_offset(thinktime),
 345                 .help   = "Idle time between IO buffers (usec)",
 346                 .def    = "0",
 347         },
 348         {
 349                 .name   = "thinktime_spin",
 350                 .type   = FIO_OPT_INT,
 351                 .off1   = td_var_offset(thinktime_spin),
 352                 .help   = "Start thinktime by spinning this amount (usec)",
 353                 .def    = "0",
 354         },
 355         {
 356                 .name   = "thinktime_blocks",
 357                 .type   = FIO_OPT_INT,
 358                 .off1   = td_var_offset(thinktime_blocks),
 359                 .help   = "IO buffer period between 'thinktime'",
 360                 .def    = "1",
 361         },
 362         {
 363                 .name   = "rate",
 364                 .type   = FIO_OPT_INT,
 365                 .off1   = td_var_offset(rate),
 366                 .help   = "Set bandwidth rate",
 367         },
 368         {
 369                 .name   = "ratemin",
 370                 .type   = FIO_OPT_INT,
 371                 .off1   = td_var_offset(ratemin),
 372                 .help   = "The bottom limit accepted",
 373         },
 374         {
 375                 .name   = "ratecycle",
 376                 .type   = FIO_OPT_INT,
 377                 .off1   = td_var_offset(ratecycle),
 378                 .help   = "Window average for rate limits (msec)",
 379                 .def    = "1000",
 380         },
 381         {
 382                 .name   = "invalidate",
 383                 .type   = FIO_OPT_BOOL,
 384                 .off1   = td_var_offset(invalidate_cache),
 385                 .help   = "Invalidate buffer/page cache prior to running job",
 386                 .def    = "1",
 387         },
 388         {
 389                 .name   = "sync",
 390                 .type   = FIO_OPT_BOOL,
 391                 .off1   = td_var_offset(sync_io),
 392                 .help   = "Use O_SYNC for buffered writes",
 393                 .def    = "0",
 394         },
 395         {
 396                 .name   = "bwavgtime",
 397                 .type   = FIO_OPT_INT,
 398                 .off1   = td_var_offset(bw_avg_time),
 399                 .help   = "Time window over which to calculate bandwidth (msec)",
 400                 .def    = "500",
 401         },
 402         {
 403                 .name   = "create_serialize",
 404                 .type   = FIO_OPT_BOOL,
 405                 .off1   = td_var_offset(create_serialize),
 406                 .help   = "Serialize creating of job files",
 407                 .def    = "1",
 408         },
 409         {
 410                 .name   = "create_fsync",
 411                 .type   = FIO_OPT_BOOL,
 412                 .off1   = td_var_offset(create_fsync),
 413                 .help   = "Fsync file after creation",
 414                 .def    = "1",
 415         },
 416         {
 417                 .name   = "cpuload",
 418                 .type   = FIO_OPT_INT,
 419                 .off1   = td_var_offset(cpuload),
 420                 .help   = "Use this percentage of CPU",
 421         },
 422         {
 423                 .name   = "cpuchunks",
 424                 .type   = FIO_OPT_INT,
 425                 .off1   = td_var_offset(cpucycle),
 426                 .help   = "Length of the CPU burn cycles",
 427         },
 428 #ifdef FIO_HAVE_CPU_AFFINITY
 429         {
 430                 .name   = "cpumask",
 431                 .type   = FIO_OPT_INT,
 432                 .cb     = str_cpumask_cb,
 433                 .help   = "CPU affinity mask",
 434         },
 435 #endif
 436         {
 437                 .name   = "end_fsync",
 438                 .type   = FIO_OPT_BOOL,
 439                 .off1   = td_var_offset(end_fsync),
 440                 .help   = "Include fsync at the end of job",
 441                 .def    = "0",
 442         },
 443         {
 444                 .name   = "unlink",
 445                 .type   = FIO_OPT_BOOL,
 446                 .off1   = td_var_offset(unlink),
 447                 .help   = "Unlink created files after job has completed",
 448                 .def    = "0",
 449         },
 450         {
 451                 .name   = "exitall",
 452                 .type   = FIO_OPT_STR_SET,
 453                 .cb     = str_exitall_cb,
 454                 .help   = "Terminate all jobs when one exits",
 455         },
 456         {
 457                 .name   = "stonewall",
 458                 .type   = FIO_OPT_STR_SET,
 459                 .off1   = td_var_offset(stonewall),
 460                 .help   = "Insert a hard barrier between this job and previous",
 461         },
 462         {
 463                 .name   = "thread",
 464                 .type   = FIO_OPT_STR_SET,
 465                 .off1   = td_var_offset(use_thread),
 466                 .help   = "Use threads instead of forks",
 467         },
 468         {
 469                 .name   = "write_bw_log",
 470                 .type   = FIO_OPT_STR_SET,
 471                 .off1   = td_var_offset(write_bw_log),
 472                 .help   = "Write log of bandwidth during run",
 473         },
 474         {
 475                 .name   = "write_lat_log",
 476                 .type   = FIO_OPT_STR_SET,
 477                 .off1   = td_var_offset(write_lat_log),
 478                 .help   = "Write log of latency during run",
 479         },
 480         {
 481                 .name   = "hugepage-size",
 482                 .type   = FIO_OPT_STR_VAL,
 483                 .off1   = td_var_offset(hugepage_size),
 484                 .help   = "When using hugepages, specify size of each page",
 485                 .def    = __stringify(FIO_HUGE_PAGE),
 486         },
 487         {
 488                 .name = NULL,
 489         },
 490 };
 491
 492 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 493 #define FIO_CMD_OPTS    (16)
 494 #define FIO_GETOPT_JOB  (0x89988998)
 495
 496 /*
 497  * Command line options. These will contain the above, plus a few
 498  * extra that only pertain to fio itself and not jobs.
 499  */
 500 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 501         {
 502                 .name           = "output",
 503                 .has_arg        = required_argument,
 504                 .val            = 'o',
 505         },
 506         {
 507                 .name           = "timeout",
 508                 .has_arg        = required_argument,
 509                 .val            = 't',
 510         },
 511         {
 512                 .name           = "latency-log",
 513                 .has_arg        = required_argument,
 514                 .val            = 'l',
 515         },
 516         {
 517                 .name           = "bandwidth-log",
 518                 .has_arg        = required_argument,
 519                 .val            = 'b',
 520         },
 521         {
 522                 .name           = "minimal",
 523                 .has_arg        = optional_argument,
 524                 .val            = 'm',
 525         },
 526         {
 527                 .name           = "version",
 528                 .has_arg        = no_argument,
 529                 .val            = 'v',
 530         },
 531         {
 532                 .name           = "help",
 533                 .has_arg        = no_argument,
 534                 .val            = 'h',
 535         },
 536         {
 537                 .name           = "cmdhelp",
 538                 .has_arg        = required_argument,
 539                 .val            = 'c',
 540         },
 541         {
 542                 .name           = NULL,
 543         },
 544 };
 545
 546 static int def_timeout = 0;
 547
 548 static char fio_version_string[] = "fio 1.11";
 549
 550 static char **ini_file;
 551 static int max_jobs = MAX_JOBS;
 552
 553 struct thread_data def_thread;
 554 struct thread_data *threads = NULL;
 555
 556 int exitall_on_terminate = 0;
 557 int terse_output = 0;
 558 unsigned long long mlock_size = 0;
 559 FILE *f_out = NULL;
 560 FILE *f_err = NULL;
 561
 562 static int write_lat_log = 0;
 563 int write_bw_log = 0;
 564
 565 /*
 566  * Return a free job structure.
 567  */
 568 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 569 {
 570         struct thread_data *td;
 571
 572         if (global)
 573                 return &def_thread;
 574         if (thread_number >= max_jobs)
 575                 return NULL;
 576
 577         td = &threads[thread_number++];
 578         *td = *parent;
 579
 580         td->thread_number = thread_number;
 581         return td;
 582 }
 583
 584 static void put_job(struct thread_data *td)
 585 {
 586         if (td == &def_thread)
 587                 return;
 588
 589         if (td->error)
 590                 fprintf(f_out, "fio: %s\n", td->verror);
 591
 592         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 593         thread_number--;
 594 }
 595
 596 /*
 597  * Lazy way of fixing up options that depend on each other. We could also
 598  * define option callback handlers, but this is easier.
 599  */
 600 static void fixup_options(struct thread_data *td)
 601 {
 602         if (!td->rwmixread && td->rwmixwrite)
 603                 td->rwmixread = 100 - td->rwmixwrite;
 604
 605         if (td->write_iolog_file && td->read_iolog_file) {
 606                 log_err("fio: read iolog overrides write_iolog\n");
 607                 free(td->write_iolog_file);
 608                 td->write_iolog_file = NULL;
 609         }
 610
 611         if (td->io_ops->flags & FIO_SYNCIO)
 612                 td->iodepth = 1;
 613         else {
 614                 if (!td->iodepth)
 615                         td->iodepth = td->nr_files;
 616         }
 617
 618         /*
 619          * only really works for sequential io for now, and with 1 file
 620          */
 621         if (td->zone_size && !td->sequential && td->nr_files == 1)
 622                 td->zone_size = 0;
 623
 624         /*
 625          * Reads can do overwrites, we always need to pre-create the file
 626          */
 627         if (td_read(td) || td_rw(td))
 628                 td->overwrite = 1;
 629
 630         if (!td->min_bs[DDIR_READ])
 631                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 632         if (!td->max_bs[DDIR_READ])
 633                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 634         if (!td->min_bs[DDIR_WRITE])
 635                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 636         if (!td->max_bs[DDIR_WRITE])
 637                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 638
 639         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 640
 641         if (td_read(td) && !td_rw(td))
 642                 td->verify = 0;
 643
 644         if (td->norandommap && td->verify != VERIFY_NONE) {
 645                 log_err("fio: norandommap given, verify disabled\n");
 646                 td->verify = VERIFY_NONE;
 647         }
 648         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 649                 log_err("fio: bs_unaligned may not work with raw io\n");
 650
 651         /*
 652          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 653          */
 654         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 655                 td->odirect = 0;
 656
 657         /*
 658          * thinktime_spin must be less than thinktime
 659          */
 660         if (td->thinktime_spin > td->thinktime)
 661                 td->thinktime_spin = td->thinktime;
 662
 663         /*
 664          * The low water mark cannot be bigger than the iodepth
 665          */
 666         if (td->iodepth_low > td->iodepth || !td->iodepth_low)
 667                 td->iodepth_low = td->iodepth;
 668 }
 669
 670 /*
 671  * This function leaks the buffer
 672  */
 673 static char *to_kmg(unsigned int val)
 674 {
 675         char *buf = malloc(32);
 676         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 677         char *p = post;
 678
 679         do {
 680                 if (val & 1023)
 681                         break;
 682
 683                 val >>= 10;
 684                 p++;
 685         } while (*p);
 686
 687         snprintf(buf, 31, "%u%c", val, *p);
 688         return buf;
 689 }
 690
 691 /*
 692  * Adds a job to the list of things todo. Sanitizes the various options
 693  * to make sure we don't have conflicts, and initializes various
 694  * members of td.
 695  */
 696 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 697 {
 698         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
 699                                    "rw", NULL, "randrw" };
 700         struct stat sb;
 701         int numjobs, ddir, i;
 702         struct fio_file *f;
 703
 704         /*
 705          * the def_thread is just for options, it's not a real job
 706          */
 707         if (td == &def_thread)
 708                 return 0;
 709
 710         assert(td->io_ops);
 711
 712         if (td->odirect)
 713                 td->io_ops->flags |= FIO_RAWIO;
 714
 715         td->filetype = FIO_TYPE_FILE;
 716         if (td->filename && !lstat(td->filename, &sb)) {
 717                 if (S_ISBLK(sb.st_mode))
 718                         td->filetype = FIO_TYPE_BD;
 719                 else if (S_ISCHR(sb.st_mode))
 720                         td->filetype = FIO_TYPE_CHAR;
 721         }
 722
 723         fixup_options(td);
 724
 725         if (td->filename)
 726                 td->nr_uniq_files = 1;
 727         else
 728                 td->nr_uniq_files = td->nr_files;
 729
 730         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 731                 char tmp[PATH_MAX];
 732                 int len = 0;
 733
 734                 if (td->directory && td->directory[0] != '\0') {
 735                         if (lstat(td->directory, &sb) < 0) {
 736                                 log_err("fio: %s is not a directory\n", td->directory);
 737                                 td_verror(td, errno, "lstat");
 738                                 return 1;
 739                         }
 740                         if (!S_ISDIR(sb.st_mode)) {
 741                                 log_err("fio: %s is not a directory\n", td->directory);
 742                                 return 1;
 743                         }
 744                         len = sprintf(tmp, "%s/", td->directory);
 745                 }
 746
 747                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 748
 749                 for_each_file(td, f, i) {
 750                         memset(f, 0, sizeof(*f));
 751                         f->fd = -1;
 752
 753                         if (td->filename)
 754                                 sprintf(tmp + len, "%s", td->filename);
 755                         else
 756                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 757                         f->file_name = strdup(tmp);
 758                 }
 759         } else {
 760                 td->nr_files = 1;
 761                 td->files = malloc(sizeof(struct fio_file));
 762                 f = &td->files[0];
 763
 764                 memset(f, 0, sizeof(*f));
 765                 f->fd = -1;
 766                 f->file_name = strdup(jobname);
 767         }
 768
 769         for_each_file(td, f, i) {
 770                 f->file_size = td->total_file_size / td->nr_files;
 771                 f->file_offset = td->start_offset;
 772         }
 773
 774         fio_sem_init(&td->mutex, 0);
 775
 776         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 777         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 778         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 779
 780         if (td->stonewall && td->thread_number > 1)
 781                 groupid++;
 782
 783         td->groupid = groupid;
 784
 785         if (setup_rate(td))
 786                 goto err;
 787
 788         if (td->write_lat_log) {
 789                 setup_log(&td->ts.slat_log);
 790                 setup_log(&td->ts.clat_log);
 791         }
 792         if (td->write_bw_log)
 793                 setup_log(&td->ts.bw_log);
 794
 795         if (!td->name)
 796                 td->name = strdup(jobname);
 797
 798         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 799
 800         if (!terse_output) {
 801                 if (!job_add_num) {
 802                         if (td->io_ops->flags & FIO_CPUIO)
 803                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 804                         else {
 805                                 char *c1, *c2, *c3, *c4;
 806
 807                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 808                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 809                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 810                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 811
 812                                 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);
 813
 814                                 free(c1);
 815                                 free(c2);
 816                                 free(c3);
 817                                 free(c4);
 818                         }
 819                 } else if (job_add_num == 1)
 820                         fprintf(f_out, "...\n");
 821         }
 822
 823         /*
 824          * recurse add identical jobs, clear numjobs and stonewall options
 825          * as they don't apply to sub-jobs
 826          */
 827         numjobs = td->numjobs;
 828         while (--numjobs) {
 829                 struct thread_data *td_new = get_new_job(0, td);
 830
 831                 if (!td_new)
 832                         goto err;
 833
 834                 td_new->numjobs = 1;
 835                 td_new->stonewall = 0;
 836                 job_add_num = numjobs - 1;
 837
 838                 if (add_job(td_new, jobname, job_add_num))
 839                         goto err;
 840         }
 841         return 0;
 842 err:
 843         put_job(td);
 844         return -1;
 845 }
 846
 847 /*
 848  * Initialize the various random states we need (random io, block size ranges,
 849  * read/write mix, etc).
 850  */
 851 int init_random_state(struct thread_data *td)
 852 {
 853         unsigned long seeds[5];
 854         int fd, num_maps, blocks, i;
 855         struct fio_file *f;
 856
 857         if (td->io_ops->flags & FIO_CPUIO)
 858                 return 0;
 859
 860         fd = open("/dev/urandom", O_RDONLY);
 861         if (fd == -1) {
 862                 td_verror(td, errno, "open");
 863                 return 1;
 864         }
 865
 866         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 867                 td_verror(td, EIO, "read");
 868                 close(fd);
 869                 return 1;
 870         }
 871
 872         close(fd);
 873
 874         os_random_seed(seeds[0], &td->bsrange_state);
 875         os_random_seed(seeds[1], &td->verify_state);
 876         os_random_seed(seeds[2], &td->rwmix_state);
 877         os_random_seed(seeds[3], &td->next_file_state);
 878
 879         if (td->sequential)
 880                 return 0;
 881
 882         if (td->rand_repeatable)
 883                 seeds[4] = FIO_RANDSEED * td->thread_number;
 884
 885         if (!td->norandommap) {
 886                 for_each_file(td, f, i) {
 887                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 888                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 889                         f->file_map = malloc(num_maps * sizeof(long));
 890                         f->num_maps = num_maps;
 891                         memset(f->file_map, 0, num_maps * sizeof(long));
 892                 }
 893         }
 894
 895         os_random_seed(seeds[4], &td->random_state);
 896         return 0;
 897 }
 898
 899 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 900 {
 901 #ifdef FIO_HAVE_CPU_AFFINITY
 902         unsigned int i;
 903
 904         CPU_ZERO(&cpumask);
 905
 906         for (i = 0; i < sizeof(int) * 8; i++) {
 907                 if ((1 << i) & cpu)
 908                         CPU_SET(i, &cpumask);
 909         }
 910 #endif
 911 }
 912
 913 static int is_empty_or_comment(char *line)
 914 {
 915         unsigned int i;
 916
 917         for (i = 0; i < strlen(line); i++) {
 918                 if (line[i] == ';')
 919                         return 1;
 920                 if (line[i] == '#')
 921                         return 1;
 922                 if (!isspace(line[i]) && !iscntrl(line[i]))
 923                         return 0;
 924         }
 925
 926         return 1;
 927 }
 928
 929 static int str_rw_cb(void *data, const char *mem)
 930 {
 931         struct thread_data *td = data;
 932
 933         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 934                 td->ddir = DDIR_READ;
 935                 td->sequential = 1;
 936                 return 0;
 937         } else if (!strncmp(mem, "randread", 8)) {
 938                 td->ddir = DDIR_READ;
 939                 td->sequential = 0;
 940                 return 0;
 941         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 942                 td->ddir = DDIR_WRITE;
 943                 td->sequential = 1;
 944                 return 0;
 945         } else if (!strncmp(mem, "randwrite", 9)) {
 946                 td->ddir = DDIR_WRITE;
 947                 td->sequential = 0;
 948                 return 0;
 949         } else if (!strncmp(mem, "rw", 2)) {
 950                 td->ddir = DDIR_READ;
 951                 td->iomix = 1;
 952                 td->sequential = 1;
 953                 return 0;
 954         } else if (!strncmp(mem, "randrw", 6)) {
 955                 td->ddir = DDIR_READ;
 956                 td->iomix = 1;
 957                 td->sequential = 0;
 958                 return 0;
 959         }
 960
 961         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 962         return 1;
 963 }
 964
 965 static int str_verify_cb(void *data, const char *mem)
 966 {
 967         struct thread_data *td = data;
 968
 969         if (!strncmp(mem, "0", 1)) {
 970                 td->verify = VERIFY_NONE;
 971                 return 0;
 972         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 973                 td->verify = VERIFY_MD5;
 974                 return 0;
 975         } else if (!strncmp(mem, "crc32", 5)) {
 976                 td->verify = VERIFY_CRC32;
 977                 return 0;
 978         }
 979
 980         log_err("fio: verify types: md5, crc32\n");
 981         return 1;
 982 }
 983
 984 /*
 985  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
 986  */
 987 static char *get_mmap_file(const char *str)
 988 {
 989         char *p = strstr(str, ":");
 990
 991         if (!p)
 992                 return NULL;
 993
 994         p++;
 995         strip_blank_front(&p);
 996         strip_blank_end(p);
 997         return strdup(p);
 998 }
 999
1000 static int str_mem_cb(void *data, const char *mem)
1001 {
1002         struct thread_data *td = data;
1003
1004         if (!strncmp(mem, "malloc", 6)) {
1005                 td->mem_type = MEM_MALLOC;
1006                 return 0;
1007         } else if (!strncmp(mem, "mmaphuge", 8)) {
1008 #ifdef FIO_HAVE_HUGETLB
1009                 /*
1010                  * mmaphuge must be appended with the actual file
1011                  */
1012                 td->mmapfile = get_mmap_file(mem);
1013                 if (!td->mmapfile) {
1014                         log_err("fio: mmaphuge:/path/to/file\n");
1015                         return 1;
1016                 }
1017
1018                 td->mem_type = MEM_MMAPHUGE;
1019                 return 0;
1020 #else
1021                 log_err("fio: mmaphuge not available\n");
1022                 return 1;
1023 #endif
1024         } else if (!strncmp(mem, "mmap", 4)) {
1025                 /*
1026                  * Check if the user wants file backed memory. It's ok
1027                  * if there's no file given, we'll just use anon mamp then.
1028                  */
1029                 td->mmapfile = get_mmap_file(mem);
1030                 td->mem_type = MEM_MMAP;
1031                 return 0;
1032         } else if (!strncmp(mem, "shmhuge", 7)) {
1033 #ifdef FIO_HAVE_HUGETLB
1034                 td->mem_type = MEM_SHMHUGE;
1035                 return 0;
1036 #else
1037                 log_err("fio: shmhuge not available\n");
1038                 return 1;
1039 #endif
1040         } else if (!strncmp(mem, "shm", 3)) {
1041                 td->mem_type = MEM_SHM;
1042                 return 0;
1043         }
1044
1045         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
1046         return 1;
1047 }
1048
1049 static int str_ioengine_cb(void *data, const char *str)
1050 {
1051         struct thread_data *td = data;
1052
1053         td->io_ops = load_ioengine(td, str);
1054         if (td->io_ops)
1055                 return 0;
1056
1057         log_err("fio: ioengine= libaio, posixaio, sync, syslet-rw, mmap, sgio, splice, cpu, null\n");
1058         log_err("fio: or specify path to dynamic ioengine module\n");
1059         return 1;
1060 }
1061
1062 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1063 {
1064         mlock_size = *val;
1065         return 0;
1066 }
1067
1068 #ifdef FIO_HAVE_IOPRIO
1069 static int str_prioclass_cb(void *data, unsigned int *val)
1070 {
1071         struct thread_data *td = data;
1072
1073         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1074         return 0;
1075 }
1076
1077 static int str_prio_cb(void *data, unsigned int *val)
1078 {
1079         struct thread_data *td = data;
1080
1081         td->ioprio |= *val;
1082         return 0;
1083 }
1084 #endif
1085
1086 static int str_exitall_cb(void)
1087 {
1088         exitall_on_terminate = 1;
1089         return 0;
1090 }
1091
1092 static int str_cpumask_cb(void *data, unsigned int *val)
1093 {
1094         struct thread_data *td = data;
1095
1096         fill_cpu_mask(td->cpumask, *val);
1097         return 0;
1098 }
1099
1100 static int str_file_service_cb(void *data, const char *str)
1101 {
1102         struct thread_data *td = data;
1103
1104         if (!strncmp(str, "random", 6)) {
1105                 td->file_service_type = FIO_FSERVICE_RANDOM;
1106                 return 0;
1107         } else if (!strncmp(str, "roundrobin", 10)) {
1108                 td->file_service_type = FIO_FSERVICE_RR;
1109                 return 0;
1110         }
1111
1112         log_err("fio: file_service= random, roundrobin\n");
1113         return 1;
1114 }
1115
1116 /*
1117  * This is our [ini] type file parser.
1118  */
1119 static int parse_jobs_ini(char *file, int stonewall_flag)
1120 {
1121         unsigned int global;
1122         struct thread_data *td;
1123         char *string, *name;
1124         fpos_t off;
1125         FILE *f;
1126         char *p;
1127         int ret = 0, stonewall;
1128
1129         f = fopen(file, "r");
1130         if (!f) {
1131                 perror("fopen job file");
1132                 return 1;
1133         }
1134
1135         string = malloc(4096);
1136         name = malloc(256);
1137         memset(name, 0, 256);
1138
1139         stonewall = stonewall_flag;
1140         do {
1141                 p = fgets(string, 4095, f);
1142                 if (!p)
1143                         break;
1144                 if (is_empty_or_comment(p))
1145                         continue;
1146                 if (sscanf(p, "[%255s]", name) != 1)
1147                         continue;
1148
1149                 global = !strncmp(name, "global", 6);
1150
1151                 name[strlen(name) - 1] = '\0';
1152
1153                 td = get_new_job(global, &def_thread);
1154                 if (!td) {
1155                         ret = 1;
1156                         break;
1157                 }
1158
1159                 /*
1160                  * Seperate multiple job files by a stonewall
1161                  */
1162                 if (!global && stonewall) {
1163                         td->stonewall = stonewall;
1164                         stonewall = 0;
1165                 }
1166
1167                 fgetpos(f, &off);
1168                 while ((p = fgets(string, 4096, f)) != NULL) {
1169                         if (is_empty_or_comment(p))
1170                                 continue;
1171
1172                         strip_blank_front(&p);
1173
1174                         if (p[0] == '[')
1175                                 break;
1176
1177                         strip_blank_end(p);
1178
1179                         fgetpos(f, &off);
1180
1181                         /*
1182                          * Don't break here, continue parsing options so we
1183                          * dump all the bad ones. Makes trial/error fixups
1184                          * easier on the user.
1185                          */
1186                         ret |= parse_option(p, options, td);
1187                 }
1188
1189                 if (!ret) {
1190                         fsetpos(f, &off);
1191                         ret = add_job(td, name, 0);
1192                 } else {
1193                         log_err("fio: job %s dropped\n", name);
1194                         put_job(td);
1195                 }
1196         } while (!ret);
1197
1198         free(string);
1199         free(name);
1200         fclose(f);
1201         return ret;
1202 }
1203
1204 static int fill_def_thread(void)
1205 {
1206         memset(&def_thread, 0, sizeof(def_thread));
1207
1208         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1209                 perror("sched_getaffinity");
1210                 return 1;
1211         }
1212
1213         /*
1214          * fill default options
1215          */
1216         fill_default_options(&def_thread, options);
1217
1218         def_thread.timeout = def_timeout;
1219         def_thread.write_bw_log = write_bw_log;
1220         def_thread.write_lat_log = write_lat_log;
1221
1222 #ifdef FIO_HAVE_DISK_UTIL
1223         def_thread.do_disk_util = 1;
1224 #endif
1225
1226         return 0;
1227 }
1228
1229 static void usage(void)
1230 {
1231         printf("%s\n", fio_version_string);
1232         printf("\t--output\tWrite output to file\n");
1233         printf("\t--timeout\tRuntime in seconds\n");
1234         printf("\t--latency-log\tGenerate per-job latency logs\n");
1235         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1236         printf("\t--minimal\tMinimal (terse) output\n");
1237         printf("\t--version\tPrint version info and exit\n");
1238         printf("\t--help\t\tPrint this page\n");
1239         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1240 }
1241
1242 static int parse_cmd_line(int argc, char *argv[])
1243 {
1244         struct thread_data *td = NULL;
1245         int c, ini_idx = 0, lidx, ret;
1246
1247         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1248                 switch (c) {
1249                 case 't':
1250                         def_timeout = atoi(optarg);
1251                         break;
1252                 case 'l':
1253                         write_lat_log = 1;
1254                         break;
1255                 case 'w':
1256                         write_bw_log = 1;
1257                         break;
1258                 case 'o':
1259                         f_out = fopen(optarg, "w+");
1260                         if (!f_out) {
1261                                 perror("fopen output");
1262                                 exit(1);
1263                         }
1264                         f_err = f_out;
1265                         break;
1266                 case 'm':
1267                         terse_output = 1;
1268                         break;
1269                 case 'h':
1270                         usage();
1271                         exit(0);
1272                 case 'c':
1273                         ret = show_cmd_help(options, optarg);
1274                         exit(ret);
1275                 case 'v':
1276                         printf("%s\n", fio_version_string);
1277                         exit(0);
1278                 case FIO_GETOPT_JOB: {
1279                         const char *opt = long_options[lidx].name;
1280                         char *val = optarg;
1281
1282                         if (!strncmp(opt, "name", 4) && td) {
1283                                 ret = add_job(td, td->name ?: "fio", 0);
1284                                 if (ret) {
1285                                         put_job(td);
1286                                         return 0;
1287                                 }
1288                                 td = NULL;
1289                         }
1290                         if (!td) {
1291                                 int global = !strncmp(val, "global", 6);
1292
1293                                 td = get_new_job(global, &def_thread);
1294                                 if (!td)
1295                                         return 0;
1296                         }
1297
1298                         ret = parse_cmd_option(opt, val, options, td);
1299                         if (ret) {
1300                                 log_err("fio: job dropped\n");
1301                                 put_job(td);
1302                                 td = NULL;
1303                         }
1304                         break;
1305                 }
1306                 default:
1307                         break;
1308                 }
1309         }
1310
1311         if (td) {
1312                 ret = add_job(td, td->name ?: "fio", 0);
1313                 if (ret)
1314                         put_job(td);
1315         }
1316
1317         while (optind < argc) {
1318                 ini_idx++;
1319                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1320                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1321                 optind++;
1322         }
1323
1324         return ini_idx;
1325 }
1326
1327 static void free_shm(void)
1328 {
1329         struct shmid_ds sbuf;
1330
1331         if (threads) {
1332                 shmdt((void *) threads);
1333                 threads = NULL;
1334                 shmctl(shm_id, IPC_RMID, &sbuf);
1335         }
1336 }
1337
1338 /*
1339  * The thread area is shared between the main process and the job
1340  * threads/processes. So setup a shared memory segment that will hold
1341  * all the job info.
1342  */
1343 static int setup_thread_area(void)
1344 {
1345         /*
1346          * 1024 is too much on some machines, scale max_jobs if
1347          * we get a failure that looks like too large a shm segment
1348          */
1349         do {
1350                 size_t size = max_jobs * sizeof(struct thread_data);
1351
1352                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1353                 if (shm_id != -1)
1354                         break;
1355                 if (errno != EINVAL) {
1356                         perror("shmget");
1357                         break;
1358                 }
1359
1360                 max_jobs >>= 1;
1361         } while (max_jobs);
1362
1363         if (shm_id == -1)
1364                 return 1;
1365
1366         threads = shmat(shm_id, NULL, 0);
1367         if (threads == (void *) -1) {
1368                 perror("shmat");
1369                 return 1;
1370         }
1371
1372         atexit(free_shm);
1373         return 0;
1374 }
1375
1376 /*
1377  * Copy the fio options into the long options map, so we mirror
1378  * job and cmd line options.
1379  */
1380 static void dupe_job_options(void)
1381 {
1382         struct fio_option *o;
1383         unsigned int i;
1384
1385         i = 0;
1386         while (long_options[i].name)
1387                 i++;
1388
1389         o = &options[0];
1390         while (o->name) {
1391                 long_options[i].name = o->name;
1392                 long_options[i].val = FIO_GETOPT_JOB;
1393                 if (o->type == FIO_OPT_STR_SET)
1394                         long_options[i].has_arg = no_argument;
1395                 else
1396                         long_options[i].has_arg = required_argument;
1397
1398                 i++;
1399                 o++;
1400                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1401         }
1402 }
1403
1404 int parse_options(int argc, char *argv[])
1405 {
1406         int job_files, i;
1407
1408         f_out = stdout;
1409         f_err = stderr;
1410
1411         options_init(options);
1412
1413         dupe_job_options();
1414
1415         if (setup_thread_area())
1416                 return 1;
1417         if (fill_def_thread())
1418                 return 1;
1419
1420         job_files = parse_cmd_line(argc, argv);
1421
1422         for (i = 0; i < job_files; i++) {
1423                 if (fill_def_thread())
1424                         return 1;
1425                 if (parse_jobs_ini(ini_file[i], i))
1426                         return 1;
1427                 free(ini_file[i]);
1428         }
1429
1430         free(ini_file);
1431
1432         if (!thread_number) {
1433                 log_err("No jobs defined(s)\n");
1434                 return 1;
1435         }
1436
1437         return 0;
1438 }