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_ioengine_cb(void *, const char *);
  26 static int str_mem_cb(void *, const char *);
  27 static int str_lockmem_cb(void *, unsigned long *);
  28 #ifdef FIO_HAVE_IOPRIO
  29 static int str_prio_cb(void *, unsigned int *);
  30 static int str_prioclass_cb(void *, unsigned int *);
  31 #endif
  32 static int str_exitall_cb(void);
  33 static int str_cpumask_cb(void *, unsigned int *);
  34
  35 #define __stringify_1(x)        #x
  36 #define __stringify(x)          __stringify_1(x)
  37
  38 /*
  39  * Map of job/command line options
  40  */
  41 static struct fio_option options[] = {
  42         {
  43                 .name   = "description",
  44                 .type   = FIO_OPT_STR_STORE,
  45                 .off1   = td_var_offset(description),
  46                 .help   = "Text job description",
  47         },
  48         {
  49                 .name   = "name",
  50                 .type   = FIO_OPT_STR_STORE,
  51                 .off1   = td_var_offset(name),
  52                 .help   = "Name of this job",
  53         },
  54         {
  55                 .name   = "directory",
  56                 .type   = FIO_OPT_STR_STORE,
  57                 .off1   = td_var_offset(directory),
  58                 .help   = "Directory to store files in",
  59         },
  60         {
  61                 .name   = "filename",
  62                 .type   = FIO_OPT_STR_STORE,
  63                 .off1   = td_var_offset(filename),
  64                 .help   = "Force the use of a specific file",
  65         },
  66         {
  67                 .name   = "rw",
  68                 .type   = FIO_OPT_STR,
  69                 .off1   = td_var_offset(td_ddir),
  70                 .help   = "IO direction",
  71                 .def    = "read",
  72                 .posval = {
  73                           { .ival = "read", .oval = TD_DDIR_READ },
  74                           { .ival = "write", .oval = TD_DDIR_WRITE },
  75                           { .ival = "randread", .oval = TD_DDIR_RANDREAD },
  76                           { .ival = "randwrite", .oval = TD_DDIR_RANDWRITE },
  77                           { .ival = "rw", .oval = TD_DDIR_RW },
  78                           { .ival = "randrw", .oval = TD_DDIR_RANDRW },
  79                           },
  80         },
  81         {
  82                 .name   = "ioengine",
  83                 .type   = FIO_OPT_STR,
  84                 .cb     = str_ioengine_cb,
  85                 .help   = "IO engine to use",
  86                 .def    = "sync",
  87                 .posval = {
  88                           { .ival = "sync", },
  89 #ifdef FIO_HAVE_LIBAIO
  90                           { .ival = "libaio", },
  91 #endif
  92 #ifdef FIO_HAVE_POSIXAIO
  93                           { .ival = "posixaio", },
  94 #endif
  95                           { .ival = "mmap", },
  96 #ifdef FIO_HAVE_SPLICE
  97                           { .ival = "splice", },
  98 #endif
  99 #ifdef FIO_HAVE_SGIO
 100                           { .ival = "sg", },
 101 #endif
 102                           { .ival = "null", }, { .ival = "net", },
 103 #ifdef FIO_HAVE_SYSLET
 104                           { .ival = "syslet-rw", },
 105 #endif
 106                           { .ival = "external", },
 107                           },
 108         },
 109         {
 110                 .name   = "iodepth",
 111                 .type   = FIO_OPT_INT,
 112                 .off1   = td_var_offset(iodepth),
 113                 .help   = "Amount of IO buffers to keep in flight",
 114                 .def    = "1",
 115         },
 116         {
 117                 .name   = "iodepth_batch",
 118                 .type   = FIO_OPT_INT,
 119                 .off1   = td_var_offset(iodepth_batch),
 120                 .help   = "Number of IO to submit in one go",
 121         },
 122         {
 123                 .name   = "iodepth_low",
 124                 .type   = FIO_OPT_INT,
 125                 .off1   = td_var_offset(iodepth_low),
 126                 .help   = "Low water mark for queuing depth",
 127         },
 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   = "bsrange",
 144                 .type   = FIO_OPT_RANGE,
 145                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 146                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 147                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 148                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 149                 .help   = "Set block size range (in more detail than bs)",
 150         },
 151         {
 152                 .name   = "bs_unaligned",
 153                 .type   = FIO_OPT_STR_SET,
 154                 .off1   = td_var_offset(bs_unaligned),
 155                 .help   = "Don't sector align IO buffer sizes",
 156         },
 157         {
 158                 .name   = "offset",
 159                 .type   = FIO_OPT_STR_VAL,
 160                 .off1   = td_var_offset(start_offset),
 161                 .help   = "Start IO from this offset",
 162                 .def    = "0",
 163         },
 164         {
 165                 .name   = "randrepeat",
 166                 .type   = FIO_OPT_BOOL,
 167                 .off1   = td_var_offset(rand_repeatable),
 168                 .help   = "Use repeatable random IO pattern",
 169                 .def    = "1",
 170         },
 171         {
 172                 .name   = "norandommap",
 173                 .type   = FIO_OPT_STR_SET,
 174                 .off1   = td_var_offset(norandommap),
 175                 .help   = "Accept potential duplicate random blocks",
 176         },
 177         {
 178                 .name   = "nrfiles",
 179                 .type   = FIO_OPT_INT,
 180                 .off1   = td_var_offset(nr_files),
 181                 .help   = "Split job workload between this number of files",
 182                 .def    = "1",
 183         },
 184         {
 185                 .name   = "file_service_type",
 186                 .type   = FIO_OPT_STR,
 187                 .off1   = td_var_offset(file_service_type),
 188                 .help   = "How to select which file to service next",
 189                 .def    = "roundrobin",
 190                 .posval = {
 191                           { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
 192                           { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
 193                           },
 194         },
 195         {
 196                 .name   = "fsync",
 197                 .type   = FIO_OPT_INT,
 198                 .off1   = td_var_offset(fsync_blocks),
 199                 .help   = "Issue fsync for writes every given number of blocks",
 200                 .def    = "0",
 201         },
 202         {
 203                 .name   = "direct",
 204                 .type   = FIO_OPT_BOOL,
 205                 .off1   = td_var_offset(odirect),
 206                 .help   = "Use O_DIRECT IO (negates buffered)",
 207                 .def    = "0",
 208         },
 209         {
 210                 .name   = "buffered",
 211                 .type   = FIO_OPT_BOOL,
 212                 .off1   = td_var_offset(odirect),
 213                 .neg    = 1,
 214                 .help   = "Use buffered IO (negates direct)",
 215                 .def    = "1",
 216         },
 217         {
 218                 .name   = "overwrite",
 219                 .type   = FIO_OPT_BOOL,
 220                 .off1   = td_var_offset(overwrite),
 221                 .help   = "When writing, set whether to overwrite current data",
 222                 .def    = "0",
 223         },
 224         {
 225                 .name   = "loops",
 226                 .type   = FIO_OPT_INT,
 227                 .off1   = td_var_offset(loops),
 228                 .help   = "Number of times to run the job",
 229                 .def    = "1",
 230         },
 231         {
 232                 .name   = "numjobs",
 233                 .type   = FIO_OPT_INT,
 234                 .off1   = td_var_offset(numjobs),
 235                 .help   = "Duplicate this job this many times",
 236                 .def    = "1",
 237         },
 238         {
 239                 .name   = "startdelay",
 240                 .type   = FIO_OPT_INT,
 241                 .off1   = td_var_offset(start_delay),
 242                 .help   = "Only start job when this period has passed",
 243                 .def    = "0",
 244         },
 245         {
 246                 .name   = "runtime",
 247                 .alias  = "timeout",
 248                 .type   = FIO_OPT_STR_VAL_TIME,
 249                 .off1   = td_var_offset(timeout),
 250                 .help   = "Stop workload when this amount of time has passed",
 251                 .def    = "0",
 252         },
 253         {
 254                 .name   = "mem",
 255                 .type   = FIO_OPT_STR,
 256                 .cb     = str_mem_cb,
 257                 .off1   = td_var_offset(mem_type),
 258                 .help   = "Backing type for IO buffers",
 259                 .def    = "malloc",
 260                 .posval = {
 261                           { .ival = "malloc", .oval = MEM_MALLOC },
 262                           { .ival = "shm", .oval = MEM_SHM },
 263 #ifdef FIO_HAVE_HUGETLB
 264                           { .ival = "shmhuge", .oval = MEM_SHMHUGE },
 265 #endif
 266                           { .ival = "mmap", .oval = MEM_MMAP },
 267 #ifdef FIO_HAVE_HUGETLB
 268                           { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
 269 #endif
 270                           },
 271         },
 272         {
 273                 .name   = "verify",
 274                 .type   = FIO_OPT_STR,
 275                 .off1   = td_var_offset(verify),
 276                 .help   = "Verify sum function",
 277                 .def    = "0",
 278                 .posval = {
 279                           { .ival = "0", .oval = VERIFY_NONE },
 280                           { .ival = "crc32", .oval = VERIFY_CRC32 },
 281                           { .ival = "md5", .oval = VERIFY_MD5 },
 282                           },
 283         },
 284         {
 285                 .name   = "write_iolog",
 286                 .type   = FIO_OPT_STR_STORE,
 287                 .off1   = td_var_offset(write_iolog_file),
 288                 .help   = "Store IO pattern to file",
 289         },
 290         {
 291                 .name   = "read_iolog",
 292                 .type   = FIO_OPT_STR_STORE,
 293                 .off1   = td_var_offset(read_iolog_file),
 294                 .help   = "Playback IO pattern from file",
 295         },
 296         {
 297                 .name   = "exec_prerun",
 298                 .type   = FIO_OPT_STR_STORE,
 299                 .off1   = td_var_offset(exec_prerun),
 300                 .help   = "Execute this file prior to running job",
 301         },
 302         {
 303                 .name   = "exec_postrun",
 304                 .type   = FIO_OPT_STR_STORE,
 305                 .off1   = td_var_offset(exec_postrun),
 306                 .help   = "Execute this file after running job",
 307         },
 308 #ifdef FIO_HAVE_IOSCHED_SWITCH
 309         {
 310                 .name   = "ioscheduler",
 311                 .type   = FIO_OPT_STR_STORE,
 312                 .off1   = td_var_offset(ioscheduler),
 313                 .help   = "Use this IO scheduler on the backing device",
 314         },
 315 #endif
 316         {
 317                 .name   = "zonesize",
 318                 .type   = FIO_OPT_STR_VAL,
 319                 .off1   = td_var_offset(zone_size),
 320                 .help   = "Give size of an IO zone",
 321                 .def    = "0",
 322         },
 323         {
 324                 .name   = "zoneskip",
 325                 .type   = FIO_OPT_STR_VAL,
 326                 .off1   = td_var_offset(zone_skip),
 327                 .help   = "Space between IO zones",
 328                 .def    = "0",
 329         },
 330         {
 331                 .name   = "lockmem",
 332                 .type   = FIO_OPT_STR_VAL,
 333                 .cb     = str_lockmem_cb,
 334                 .help   = "Lock down this amount of memory",
 335                 .def    = "0",
 336         },
 337         {
 338                 .name   = "rwmixcycle",
 339                 .type   = FIO_OPT_INT,
 340                 .off1   = td_var_offset(rwmixcycle),
 341                 .help   = "Cycle period for mixed read/write workloads (msec)",
 342                 .def    = "500",
 343         },
 344         {
 345                 .name   = "rwmixread",
 346                 .type   = FIO_OPT_INT,
 347                 .off1   = td_var_offset(rwmixread),
 348                 .maxval = 100,
 349                 .help   = "Percentage of mixed workload that is reads",
 350                 .def    = "50",
 351         },
 352         {
 353                 .name   = "rwmixwrite",
 354                 .type   = FIO_OPT_INT,
 355                 .off1   = td_var_offset(rwmixwrite),
 356                 .maxval = 100,
 357                 .help   = "Percentage of mixed workload that is writes",
 358                 .def    = "50",
 359         },
 360         {
 361                 .name   = "nice",
 362                 .type   = FIO_OPT_INT,
 363                 .off1   = td_var_offset(nice),
 364                 .help   = "Set job CPU nice value",
 365                 .minval = -19,
 366                 .maxval = 20,
 367                 .def    = "0",
 368         },
 369 #ifdef FIO_HAVE_IOPRIO
 370         {
 371                 .name   = "prio",
 372                 .type   = FIO_OPT_INT,
 373                 .cb     = str_prio_cb,
 374                 .help   = "Set job IO priority value",
 375                 .minval = 0,
 376                 .maxval = 7,
 377         },
 378         {
 379                 .name   = "prioclass",
 380                 .type   = FIO_OPT_INT,
 381                 .cb     = str_prioclass_cb,
 382                 .help   = "Set job IO priority class",
 383                 .minval = 0,
 384                 .maxval = 3,
 385         },
 386 #endif
 387         {
 388                 .name   = "thinktime",
 389                 .type   = FIO_OPT_INT,
 390                 .off1   = td_var_offset(thinktime),
 391                 .help   = "Idle time between IO buffers (usec)",
 392                 .def    = "0",
 393         },
 394         {
 395                 .name   = "thinktime_spin",
 396                 .type   = FIO_OPT_INT,
 397                 .off1   = td_var_offset(thinktime_spin),
 398                 .help   = "Start thinktime by spinning this amount (usec)",
 399                 .def    = "0",
 400         },
 401         {
 402                 .name   = "thinktime_blocks",
 403                 .type   = FIO_OPT_INT,
 404                 .off1   = td_var_offset(thinktime_blocks),
 405                 .help   = "IO buffer period between 'thinktime'",
 406                 .def    = "1",
 407         },
 408         {
 409                 .name   = "rate",
 410                 .type   = FIO_OPT_INT,
 411                 .off1   = td_var_offset(rate),
 412                 .help   = "Set bandwidth rate",
 413         },
 414         {
 415                 .name   = "ratemin",
 416                 .type   = FIO_OPT_INT,
 417                 .off1   = td_var_offset(ratemin),
 418                 .help   = "The bottom limit accepted",
 419         },
 420         {
 421                 .name   = "ratecycle",
 422                 .type   = FIO_OPT_INT,
 423                 .off1   = td_var_offset(ratecycle),
 424                 .help   = "Window average for rate limits (msec)",
 425                 .def    = "1000",
 426         },
 427         {
 428                 .name   = "invalidate",
 429                 .type   = FIO_OPT_BOOL,
 430                 .off1   = td_var_offset(invalidate_cache),
 431                 .help   = "Invalidate buffer/page cache prior to running job",
 432                 .def    = "1",
 433         },
 434         {
 435                 .name   = "sync",
 436                 .type   = FIO_OPT_BOOL,
 437                 .off1   = td_var_offset(sync_io),
 438                 .help   = "Use O_SYNC for buffered writes",
 439                 .def    = "0",
 440         },
 441         {
 442                 .name   = "bwavgtime",
 443                 .type   = FIO_OPT_INT,
 444                 .off1   = td_var_offset(bw_avg_time),
 445                 .help   = "Time window over which to calculate bandwidth (msec)",
 446                 .def    = "500",
 447         },
 448         {
 449                 .name   = "create_serialize",
 450                 .type   = FIO_OPT_BOOL,
 451                 .off1   = td_var_offset(create_serialize),
 452                 .help   = "Serialize creating of job files",
 453                 .def    = "1",
 454         },
 455         {
 456                 .name   = "create_fsync",
 457                 .type   = FIO_OPT_BOOL,
 458                 .off1   = td_var_offset(create_fsync),
 459                 .help   = "Fsync file after creation",
 460                 .def    = "1",
 461         },
 462         {
 463                 .name   = "cpuload",
 464                 .type   = FIO_OPT_INT,
 465                 .off1   = td_var_offset(cpuload),
 466                 .help   = "Use this percentage of CPU",
 467         },
 468         {
 469                 .name   = "cpuchunks",
 470                 .type   = FIO_OPT_INT,
 471                 .off1   = td_var_offset(cpucycle),
 472                 .help   = "Length of the CPU burn cycles",
 473         },
 474 #ifdef FIO_HAVE_CPU_AFFINITY
 475         {
 476                 .name   = "cpumask",
 477                 .type   = FIO_OPT_INT,
 478                 .cb     = str_cpumask_cb,
 479                 .help   = "CPU affinity mask",
 480         },
 481 #endif
 482         {
 483                 .name   = "end_fsync",
 484                 .type   = FIO_OPT_BOOL,
 485                 .off1   = td_var_offset(end_fsync),
 486                 .help   = "Include fsync at the end of job",
 487                 .def    = "0",
 488         },
 489         {
 490                 .name   = "unlink",
 491                 .type   = FIO_OPT_BOOL,
 492                 .off1   = td_var_offset(unlink),
 493                 .help   = "Unlink created files after job has completed",
 494                 .def    = "0",
 495         },
 496         {
 497                 .name   = "exitall",
 498                 .type   = FIO_OPT_STR_SET,
 499                 .cb     = str_exitall_cb,
 500                 .help   = "Terminate all jobs when one exits",
 501         },
 502         {
 503                 .name   = "stonewall",
 504                 .type   = FIO_OPT_STR_SET,
 505                 .off1   = td_var_offset(stonewall),
 506                 .help   = "Insert a hard barrier between this job and previous",
 507         },
 508         {
 509                 .name   = "thread",
 510                 .type   = FIO_OPT_STR_SET,
 511                 .off1   = td_var_offset(use_thread),
 512                 .help   = "Use threads instead of forks",
 513         },
 514         {
 515                 .name   = "write_bw_log",
 516                 .type   = FIO_OPT_STR_SET,
 517                 .off1   = td_var_offset(write_bw_log),
 518                 .help   = "Write log of bandwidth during run",
 519         },
 520         {
 521                 .name   = "write_lat_log",
 522                 .type   = FIO_OPT_STR_SET,
 523                 .off1   = td_var_offset(write_lat_log),
 524                 .help   = "Write log of latency during run",
 525         },
 526         {
 527                 .name   = "hugepage-size",
 528                 .type   = FIO_OPT_STR_VAL,
 529                 .off1   = td_var_offset(hugepage_size),
 530                 .help   = "When using hugepages, specify size of each page",
 531                 .def    = __stringify(FIO_HUGE_PAGE),
 532         },
 533         {
 534                 .name = NULL,
 535         },
 536 };
 537
 538 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 539 #define FIO_CMD_OPTS    (16)
 540 #define FIO_GETOPT_JOB  (0x89988998)
 541
 542 /*
 543  * Command line options. These will contain the above, plus a few
 544  * extra that only pertain to fio itself and not jobs.
 545  */
 546 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 547         {
 548                 .name           = "output",
 549                 .has_arg        = required_argument,
 550                 .val            = 'o',
 551         },
 552         {
 553                 .name           = "timeout",
 554                 .has_arg        = required_argument,
 555                 .val            = 't',
 556         },
 557         {
 558                 .name           = "latency-log",
 559                 .has_arg        = required_argument,
 560                 .val            = 'l',
 561         },
 562         {
 563                 .name           = "bandwidth-log",
 564                 .has_arg        = required_argument,
 565                 .val            = 'b',
 566         },
 567         {
 568                 .name           = "minimal",
 569                 .has_arg        = optional_argument,
 570                 .val            = 'm',
 571         },
 572         {
 573                 .name           = "version",
 574                 .has_arg        = no_argument,
 575                 .val            = 'v',
 576         },
 577         {
 578                 .name           = "help",
 579                 .has_arg        = no_argument,
 580                 .val            = 'h',
 581         },
 582         {
 583                 .name           = "cmdhelp",
 584                 .has_arg        = required_argument,
 585                 .val            = 'c',
 586         },
 587         {
 588                 .name           = NULL,
 589         },
 590 };
 591
 592 static int def_timeout = 0;
 593
 594 static char fio_version_string[] = "fio 1.12";
 595
 596 static char **ini_file;
 597 static int max_jobs = MAX_JOBS;
 598
 599 struct thread_data def_thread;
 600 struct thread_data *threads = NULL;
 601
 602 int exitall_on_terminate = 0;
 603 int terse_output = 0;
 604 unsigned long long mlock_size = 0;
 605 FILE *f_out = NULL;
 606 FILE *f_err = NULL;
 607
 608 static int write_lat_log = 0;
 609 int write_bw_log = 0;
 610
 611 /*
 612  * Return a free job structure.
 613  */
 614 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 615 {
 616         struct thread_data *td;
 617
 618         if (global)
 619                 return &def_thread;
 620         if (thread_number >= max_jobs)
 621                 return NULL;
 622
 623         td = &threads[thread_number++];
 624         *td = *parent;
 625
 626         td->thread_number = thread_number;
 627         return td;
 628 }
 629
 630 static void put_job(struct thread_data *td)
 631 {
 632         if (td == &def_thread)
 633                 return;
 634
 635         if (td->error)
 636                 fprintf(f_out, "fio: %s\n", td->verror);
 637
 638         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 639         thread_number--;
 640 }
 641
 642 /*
 643  * Lazy way of fixing up options that depend on each other. We could also
 644  * define option callback handlers, but this is easier.
 645  */
 646 static void fixup_options(struct thread_data *td)
 647 {
 648         if (!td->rwmixread && td->rwmixwrite)
 649                 td->rwmixread = 100 - td->rwmixwrite;
 650
 651         if (td->write_iolog_file && td->read_iolog_file) {
 652                 log_err("fio: read iolog overrides write_iolog\n");
 653                 free(td->write_iolog_file);
 654                 td->write_iolog_file = NULL;
 655         }
 656
 657         if (td->io_ops->flags & FIO_SYNCIO)
 658                 td->iodepth = 1;
 659         else {
 660                 if (!td->iodepth)
 661                         td->iodepth = td->nr_files;
 662         }
 663
 664         /*
 665          * only really works for sequential io for now, and with 1 file
 666          */
 667         if (td->zone_size && td_random(td) && td->nr_files == 1)
 668                 td->zone_size = 0;
 669
 670         /*
 671          * Reads can do overwrites, we always need to pre-create the file
 672          */
 673         if (td_read(td) || td_rw(td))
 674                 td->overwrite = 1;
 675
 676         if (!td->min_bs[DDIR_READ])
 677                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 678         if (!td->max_bs[DDIR_READ])
 679                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 680         if (!td->min_bs[DDIR_WRITE])
 681                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 682         if (!td->max_bs[DDIR_WRITE])
 683                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 684
 685         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 686
 687         if (td_read(td) && !td_rw(td))
 688                 td->verify = 0;
 689
 690         if (td->norandommap && td->verify != VERIFY_NONE) {
 691                 log_err("fio: norandommap given, verify disabled\n");
 692                 td->verify = VERIFY_NONE;
 693         }
 694         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 695                 log_err("fio: bs_unaligned may not work with raw io\n");
 696
 697         /*
 698          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 699          */
 700         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 701                 td->odirect = 0;
 702
 703         /*
 704          * thinktime_spin must be less than thinktime
 705          */
 706         if (td->thinktime_spin > td->thinktime)
 707                 td->thinktime_spin = td->thinktime;
 708
 709         /*
 710          * The low water mark cannot be bigger than the iodepth
 711          */
 712         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
 713                 /*
 714                  * syslet work around - if the workload is sequential,
 715                  * we want to let the queue drain all the way down to
 716                  * avoid seeking between async threads
 717                  */
 718                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 719                         td->iodepth_low = 1;
 720                 else
 721                         td->iodepth_low = td->iodepth;
 722         }
 723
 724         /*
 725          * If batch number isn't set, default to the same as iodepth
 726          */
 727         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
 728                 td->iodepth_batch = td->iodepth;
 729 }
 730
 731 /*
 732  * This function leaks the buffer
 733  */
 734 static char *to_kmg(unsigned int val)
 735 {
 736         char *buf = malloc(32);
 737         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 738         char *p = post;
 739
 740         do {
 741                 if (val & 1023)
 742                         break;
 743
 744                 val >>= 10;
 745                 p++;
 746         } while (*p);
 747
 748         snprintf(buf, 31, "%u%c", val, *p);
 749         return buf;
 750 }
 751
 752 /*
 753  * Adds a job to the list of things todo. Sanitizes the various options
 754  * to make sure we don't have conflicts, and initializes various
 755  * members of td.
 756  */
 757 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 758 {
 759         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 760                                    "randread", "randwrite", "randrw" };
 761         struct stat sb;
 762         int numjobs, i;
 763         struct fio_file *f;
 764
 765         /*
 766          * the def_thread is just for options, it's not a real job
 767          */
 768         if (td == &def_thread)
 769                 return 0;
 770
 771         assert(td->io_ops);
 772
 773         if (td->odirect)
 774                 td->io_ops->flags |= FIO_RAWIO;
 775
 776         td->filetype = FIO_TYPE_FILE;
 777         if (td->filename && !lstat(td->filename, &sb)) {
 778                 if (S_ISBLK(sb.st_mode))
 779                         td->filetype = FIO_TYPE_BD;
 780                 else if (S_ISCHR(sb.st_mode))
 781                         td->filetype = FIO_TYPE_CHAR;
 782         }
 783
 784         fixup_options(td);
 785
 786         if (td->filename)
 787                 td->nr_uniq_files = 1;
 788         else
 789                 td->nr_uniq_files = td->nr_files;
 790
 791         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 792                 char tmp[PATH_MAX];
 793                 int len = 0;
 794
 795                 if (td->directory && td->directory[0] != '\0') {
 796                         if (lstat(td->directory, &sb) < 0) {
 797                                 log_err("fio: %s is not a directory\n", td->directory);
 798                                 td_verror(td, errno, "lstat");
 799                                 return 1;
 800                         }
 801                         if (!S_ISDIR(sb.st_mode)) {
 802                                 log_err("fio: %s is not a directory\n", td->directory);
 803                                 return 1;
 804                         }
 805                         len = sprintf(tmp, "%s/", td->directory);
 806                 }
 807
 808                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 809
 810                 for_each_file(td, f, i) {
 811                         memset(f, 0, sizeof(*f));
 812                         f->fd = -1;
 813
 814                         if (td->filename)
 815                                 sprintf(tmp + len, "%s", td->filename);
 816                         else
 817                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 818                         f->file_name = strdup(tmp);
 819                 }
 820         } else {
 821                 td->nr_files = 1;
 822                 td->files = malloc(sizeof(struct fio_file));
 823                 f = &td->files[0];
 824
 825                 memset(f, 0, sizeof(*f));
 826                 f->fd = -1;
 827                 f->file_name = strdup(jobname);
 828         }
 829
 830         for_each_file(td, f, i) {
 831                 f->file_size = td->total_file_size / td->nr_files;
 832                 f->file_offset = td->start_offset;
 833         }
 834
 835         fio_sem_init(&td->mutex, 0);
 836
 837         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 838         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 839         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 840
 841         if (td->stonewall && td->thread_number > 1)
 842                 groupid++;
 843
 844         td->groupid = groupid;
 845
 846         if (setup_rate(td))
 847                 goto err;
 848
 849         if (td->write_lat_log) {
 850                 setup_log(&td->ts.slat_log);
 851                 setup_log(&td->ts.clat_log);
 852         }
 853         if (td->write_bw_log)
 854                 setup_log(&td->ts.bw_log);
 855
 856         if (!td->name)
 857                 td->name = strdup(jobname);
 858
 859         if (!terse_output) {
 860                 if (!job_add_num) {
 861                         if (td->io_ops->flags & FIO_CPUIO)
 862                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 863                         else {
 864                                 char *c1, *c2, *c3, *c4;
 865
 866                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 867                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 868                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 869                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 870
 871                                 fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[td->td_ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
 872
 873                                 free(c1);
 874                                 free(c2);
 875                                 free(c3);
 876                                 free(c4);
 877                         }
 878                 } else if (job_add_num == 1)
 879                         fprintf(f_out, "...\n");
 880         }
 881
 882         /*
 883          * recurse add identical jobs, clear numjobs and stonewall options
 884          * as they don't apply to sub-jobs
 885          */
 886         numjobs = td->numjobs;
 887         while (--numjobs) {
 888                 struct thread_data *td_new = get_new_job(0, td);
 889
 890                 if (!td_new)
 891                         goto err;
 892
 893                 td_new->numjobs = 1;
 894                 td_new->stonewall = 0;
 895                 job_add_num = numjobs - 1;
 896
 897                 if (add_job(td_new, jobname, job_add_num))
 898                         goto err;
 899         }
 900         return 0;
 901 err:
 902         put_job(td);
 903         return -1;
 904 }
 905
 906 /*
 907  * Initialize the various random states we need (random io, block size ranges,
 908  * read/write mix, etc).
 909  */
 910 int init_random_state(struct thread_data *td)
 911 {
 912         unsigned long seeds[5];
 913         int fd, num_maps, blocks, i;
 914         struct fio_file *f;
 915
 916         if (td->io_ops->flags & FIO_CPUIO)
 917                 return 0;
 918
 919         fd = open("/dev/urandom", O_RDONLY);
 920         if (fd == -1) {
 921                 td_verror(td, errno, "open");
 922                 return 1;
 923         }
 924
 925         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 926                 td_verror(td, EIO, "read");
 927                 close(fd);
 928                 return 1;
 929         }
 930
 931         close(fd);
 932
 933         os_random_seed(seeds[0], &td->bsrange_state);
 934         os_random_seed(seeds[1], &td->verify_state);
 935         os_random_seed(seeds[2], &td->rwmix_state);
 936
 937         if (td->file_service_type == FIO_FSERVICE_RANDOM)
 938                 os_random_seed(seeds[3], &td->next_file_state);
 939
 940         if (!td_random(td))
 941                 return 0;
 942
 943         if (td->rand_repeatable)
 944                 seeds[4] = FIO_RANDSEED * td->thread_number;
 945
 946         if (!td->norandommap) {
 947                 for_each_file(td, f, i) {
 948                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 949                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 950                         f->file_map = malloc(num_maps * sizeof(long));
 951                         f->num_maps = num_maps;
 952                         memset(f->file_map, 0, num_maps * sizeof(long));
 953                 }
 954         }
 955
 956         os_random_seed(seeds[4], &td->random_state);
 957         return 0;
 958 }
 959
 960 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 961 {
 962 #ifdef FIO_HAVE_CPU_AFFINITY
 963         unsigned int i;
 964
 965         CPU_ZERO(&cpumask);
 966
 967         for (i = 0; i < sizeof(int) * 8; i++) {
 968                 if ((1 << i) & cpu)
 969                         CPU_SET(i, &cpumask);
 970         }
 971 #endif
 972 }
 973
 974 static int is_empty_or_comment(char *line)
 975 {
 976         unsigned int i;
 977
 978         for (i = 0; i < strlen(line); i++) {
 979                 if (line[i] == ';')
 980                         return 1;
 981                 if (line[i] == '#')
 982                         return 1;
 983                 if (!isspace(line[i]) && !iscntrl(line[i]))
 984                         return 0;
 985         }
 986
 987         return 1;
 988 }
 989
 990 /*
 991  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
 992  */
 993 static char *get_mmap_file(const char *str)
 994 {
 995         char *p = strstr(str, ":");
 996
 997         if (!p)
 998                 return NULL;
 999
1000         p++;
1001         strip_blank_front(&p);
1002         strip_blank_end(p);
1003         return strdup(p);
1004 }
1005
1006 static int str_mem_cb(void *data, const char *mem)
1007 {
1008         struct thread_data *td = data;
1009
1010         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1011                 td->mmapfile = get_mmap_file(mem);
1012                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1013                         log_err("fio: mmaphuge:/path/to/file\n");
1014                         return 1;
1015                 }
1016         }
1017
1018         return 0;
1019 }
1020
1021 /* External engines are specified by "external:name.o") */
1022 static const char *get_engine_name(const char *str)
1023 {
1024         char *p = strstr(str, ":");
1025
1026         if (!p)
1027                 return str;
1028
1029         p++;
1030         strip_blank_front(&p);
1031         strip_blank_end(p);
1032         return p;
1033 }
1034
1035 static int str_ioengine_cb(void *data, const char *str)
1036 {
1037         struct thread_data *td = data;
1038         const char *name = get_engine_name(str);
1039
1040         td->io_ops = load_ioengine(td, name);
1041         if (td->io_ops)
1042                 return 0;
1043
1044         return 1;
1045 }
1046
1047 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1048 {
1049         mlock_size = *val;
1050         return 0;
1051 }
1052
1053 #ifdef FIO_HAVE_IOPRIO
1054 static int str_prioclass_cb(void *data, unsigned int *val)
1055 {
1056         struct thread_data *td = data;
1057
1058         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1059         return 0;
1060 }
1061
1062 static int str_prio_cb(void *data, unsigned int *val)
1063 {
1064         struct thread_data *td = data;
1065
1066         td->ioprio |= *val;
1067         return 0;
1068 }
1069 #endif
1070
1071 static int str_exitall_cb(void)
1072 {
1073         exitall_on_terminate = 1;
1074         return 0;
1075 }
1076
1077 static int str_cpumask_cb(void *data, unsigned int *val)
1078 {
1079         struct thread_data *td = data;
1080
1081         fill_cpu_mask(td->cpumask, *val);
1082         return 0;
1083 }
1084
1085 /*
1086  * This is our [ini] type file parser.
1087  */
1088 static int parse_jobs_ini(char *file, int stonewall_flag)
1089 {
1090         unsigned int global;
1091         struct thread_data *td;
1092         char *string, *name;
1093         fpos_t off;
1094         FILE *f;
1095         char *p;
1096         int ret = 0, stonewall;
1097
1098         f = fopen(file, "r");
1099         if (!f) {
1100                 perror("fopen job file");
1101                 return 1;
1102         }
1103
1104         string = malloc(4096);
1105         name = malloc(256);
1106         memset(name, 0, 256);
1107
1108         stonewall = stonewall_flag;
1109         do {
1110                 p = fgets(string, 4095, f);
1111                 if (!p)
1112                         break;
1113                 if (is_empty_or_comment(p))
1114                         continue;
1115                 if (sscanf(p, "[%255s]", name) != 1)
1116                         continue;
1117
1118                 global = !strncmp(name, "global", 6);
1119
1120                 name[strlen(name) - 1] = '\0';
1121
1122                 td = get_new_job(global, &def_thread);
1123                 if (!td) {
1124                         ret = 1;
1125                         break;
1126                 }
1127
1128                 /*
1129                  * Seperate multiple job files by a stonewall
1130                  */
1131                 if (!global && stonewall) {
1132                         td->stonewall = stonewall;
1133                         stonewall = 0;
1134                 }
1135
1136                 fgetpos(f, &off);
1137                 while ((p = fgets(string, 4096, f)) != NULL) {
1138                         if (is_empty_or_comment(p))
1139                                 continue;
1140
1141                         strip_blank_front(&p);
1142
1143                         if (p[0] == '[')
1144                                 break;
1145
1146                         strip_blank_end(p);
1147
1148                         fgetpos(f, &off);
1149
1150                         /*
1151                          * Don't break here, continue parsing options so we
1152                          * dump all the bad ones. Makes trial/error fixups
1153                          * easier on the user.
1154                          */
1155                         ret |= parse_option(p, options, td);
1156                 }
1157
1158                 if (!ret) {
1159                         fsetpos(f, &off);
1160                         ret = add_job(td, name, 0);
1161                 } else {
1162                         log_err("fio: job %s dropped\n", name);
1163                         put_job(td);
1164                 }
1165         } while (!ret);
1166
1167         free(string);
1168         free(name);
1169         fclose(f);
1170         return ret;
1171 }
1172
1173 static int fill_def_thread(void)
1174 {
1175         memset(&def_thread, 0, sizeof(def_thread));
1176
1177         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1178                 perror("sched_getaffinity");
1179                 return 1;
1180         }
1181
1182         /*
1183          * fill default options
1184          */
1185         fill_default_options(&def_thread, options);
1186
1187         def_thread.timeout = def_timeout;
1188         def_thread.write_bw_log = write_bw_log;
1189         def_thread.write_lat_log = write_lat_log;
1190
1191 #ifdef FIO_HAVE_DISK_UTIL
1192         def_thread.do_disk_util = 1;
1193 #endif
1194
1195         return 0;
1196 }
1197
1198 static void usage(void)
1199 {
1200         printf("%s\n", fio_version_string);
1201         printf("\t--output\tWrite output to file\n");
1202         printf("\t--timeout\tRuntime in seconds\n");
1203         printf("\t--latency-log\tGenerate per-job latency logs\n");
1204         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1205         printf("\t--minimal\tMinimal (terse) output\n");
1206         printf("\t--version\tPrint version info and exit\n");
1207         printf("\t--help\t\tPrint this page\n");
1208         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1209 }
1210
1211 static int parse_cmd_line(int argc, char *argv[])
1212 {
1213         struct thread_data *td = NULL;
1214         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1215
1216         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1217                 switch (c) {
1218                 case 't':
1219                         def_timeout = atoi(optarg);
1220                         break;
1221                 case 'l':
1222                         write_lat_log = 1;
1223                         break;
1224                 case 'w':
1225                         write_bw_log = 1;
1226                         break;
1227                 case 'o':
1228                         f_out = fopen(optarg, "w+");
1229                         if (!f_out) {
1230                                 perror("fopen output");
1231                                 exit(1);
1232                         }
1233                         f_err = f_out;
1234                         break;
1235                 case 'm':
1236                         terse_output = 1;
1237                         break;
1238                 case 'h':
1239                         usage();
1240                         exit(0);
1241                 case 'c':
1242                         ret = show_cmd_help(options, optarg);
1243                         exit(ret);
1244                 case 'v':
1245                         printf("%s\n", fio_version_string);
1246                         exit(0);
1247                 case FIO_GETOPT_JOB: {
1248                         const char *opt = long_options[lidx].name;
1249                         char *val = optarg;
1250
1251                         if (!strncmp(opt, "name", 4) && td) {
1252                                 ret = add_job(td, td->name ?: "fio", 0);
1253                                 if (ret) {
1254                                         put_job(td);
1255                                         return 0;
1256                                 }
1257                                 td = NULL;
1258                         }
1259                         if (!td) {
1260                                 int global = !strncmp(val, "global", 6);
1261
1262                                 td = get_new_job(global, &def_thread);
1263                                 if (!td)
1264                                         return 0;
1265                         }
1266
1267                         ret = parse_cmd_option(opt, val, options, td);
1268                         if (ret)
1269                                 dont_add_job = 1;
1270                         break;
1271                 }
1272                 default:
1273                         break;
1274                 }
1275         }
1276
1277         if (td) {
1278                 if (dont_add_job)
1279                         put_job(td);
1280                 else {
1281                         ret = add_job(td, td->name ?: "fio", 0);
1282                         if (ret)
1283                                 put_job(td);
1284                 }
1285         }
1286
1287         while (optind < argc) {
1288                 ini_idx++;
1289                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1290                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1291                 optind++;
1292         }
1293
1294         return ini_idx;
1295 }
1296
1297 static void free_shm(void)
1298 {
1299         struct shmid_ds sbuf;
1300
1301         if (threads) {
1302                 shmdt((void *) threads);
1303                 threads = NULL;
1304                 shmctl(shm_id, IPC_RMID, &sbuf);
1305         }
1306 }
1307
1308 /*
1309  * The thread area is shared between the main process and the job
1310  * threads/processes. So setup a shared memory segment that will hold
1311  * all the job info.
1312  */
1313 static int setup_thread_area(void)
1314 {
1315         /*
1316          * 1024 is too much on some machines, scale max_jobs if
1317          * we get a failure that looks like too large a shm segment
1318          */
1319         do {
1320                 size_t size = max_jobs * sizeof(struct thread_data);
1321
1322                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1323                 if (shm_id != -1)
1324                         break;
1325                 if (errno != EINVAL) {
1326                         perror("shmget");
1327                         break;
1328                 }
1329
1330                 max_jobs >>= 1;
1331         } while (max_jobs);
1332
1333         if (shm_id == -1)
1334                 return 1;
1335
1336         threads = shmat(shm_id, NULL, 0);
1337         if (threads == (void *) -1) {
1338                 perror("shmat");
1339                 return 1;
1340         }
1341
1342         atexit(free_shm);
1343         return 0;
1344 }
1345
1346 /*
1347  * Copy the fio options into the long options map, so we mirror
1348  * job and cmd line options.
1349  */
1350 static void dupe_job_options(void)
1351 {
1352         struct fio_option *o;
1353         unsigned int i;
1354
1355         i = 0;
1356         while (long_options[i].name)
1357                 i++;
1358
1359         o = &options[0];
1360         while (o->name) {
1361                 long_options[i].name = o->name;
1362                 long_options[i].val = FIO_GETOPT_JOB;
1363                 if (o->type == FIO_OPT_STR_SET)
1364                         long_options[i].has_arg = no_argument;
1365                 else
1366                         long_options[i].has_arg = required_argument;
1367
1368                 i++;
1369                 o++;
1370                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1371         }
1372 }
1373
1374 int parse_options(int argc, char *argv[])
1375 {
1376         int job_files, i;
1377
1378         f_out = stdout;
1379         f_err = stderr;
1380
1381         options_init(options);
1382
1383         dupe_job_options();
1384
1385         if (setup_thread_area())
1386                 return 1;
1387         if (fill_def_thread())
1388                 return 1;
1389
1390         job_files = parse_cmd_line(argc, argv);
1391
1392         for (i = 0; i < job_files; i++) {
1393                 if (fill_def_thread())
1394                         return 1;
1395                 if (parse_jobs_ini(ini_file[i], i))
1396                         return 1;
1397                 free(ini_file[i]);
1398         }
1399
1400         free(ini_file);
1401
1402         if (!thread_number) {
1403                 log_err("No jobs defined(s)\n");
1404                 return 1;
1405         }
1406
1407         return 0;
1408 }