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