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 /*
  22  * The default options
  23  */
  24 #define DEF_BS                  (4096)
  25 #define DEF_TIMEOUT             (0)
  26 #define DEF_RATE_CYCLE          (1000)
  27 #define DEF_ODIRECT             (1)
  28 #define DEF_IO_ENGINE           (FIO_SYNCIO)
  29 #define DEF_IO_ENGINE_NAME      "sync"
  30 #define DEF_SEQUENTIAL          (1)
  31 #define DEF_RAND_REPEAT         (1)
  32 #define DEF_OVERWRITE           (0)
  33 #define DEF_INVALIDATE          (1)
  34 #define DEF_SYNCIO              (0)
  35 #define DEF_RANDSEED            (0xb1899bedUL)
  36 #define DEF_BWAVGTIME           (500)
  37 #define DEF_CREATE_SER          (1)
  38 #define DEF_CREATE_FSYNC        (1)
  39 #define DEF_LOOPS               (1)
  40 #define DEF_VERIFY              (0)
  41 #define DEF_STONEWALL           (0)
  42 #define DEF_NUMJOBS             (1)
  43 #define DEF_USE_THREAD          (0)
  44 #define DEF_FILE_SIZE           (1024 * 1024 * 1024UL)
  45 #define DEF_ZONE_SIZE           (0)
  46 #define DEF_ZONE_SKIP           (0)
  47 #define DEF_RWMIX_CYCLE         (500)
  48 #define DEF_RWMIX_READ          (50)
  49 #define DEF_NICE                (0)
  50 #define DEF_NR_FILES            (1)
  51 #define DEF_UNLINK              (1)
  52 #define DEF_WRITE_BW_LOG        (0)
  53 #define DEF_WRITE_LAT_LOG       (0)
  54 #define DEF_NO_RAND_MAP         (0)
  55 #define DEF_HUGEPAGE_SIZE       FIO_HUGE_PAGE
  56 #define DEF_THINKTIME_BLOCKS    (1)
  57
  58 #define td_var_offset(var)      ((size_t) &((struct thread_data *)0)->var)
  59
  60 static int str_rw_cb(void *, const char *);
  61 static int str_ioengine_cb(void *, const char *);
  62 static int str_mem_cb(void *, const char *);
  63 static int str_verify_cb(void *, const char *);
  64 static int str_lockmem_cb(void *, unsigned long *);
  65 #ifdef FIO_HAVE_IOPRIO
  66 static int str_prio_cb(void *, unsigned int *);
  67 static int str_prioclass_cb(void *, unsigned int *);
  68 #endif
  69 static int str_exitall_cb(void);
  70 static int str_cpumask_cb(void *, unsigned int *);
  71
  72 /*
  73  * Map of job/command line options
  74  */
  75 static struct fio_option options[] = {
  76         {
  77                 .name   = "name",
  78                 .type   = FIO_OPT_STR_STORE,
  79                 .off1   = td_var_offset(name),
  80                 .help   = "Name of this job",
  81         },
  82         {
  83                 .name   = "directory",
  84                 .type   = FIO_OPT_STR_STORE,
  85                 .off1   = td_var_offset(directory),
  86                 .help   = "Directory to store files in",
  87         },
  88         {
  89                 .name   = "filename",
  90                 .type   = FIO_OPT_STR_STORE,
  91                 .off1   = td_var_offset(filename),
  92                 .help   = "Force the use of a specific file",
  93         },
  94         {
  95                 .name   = "rw",
  96                 .type   = FIO_OPT_STR,
  97                 .cb     = str_rw_cb,
  98                 .help   = "IO direction (read, write, rw, etc)",
  99         },
 100         {
 101                 .name   = "ioengine",
 102                 .type   = FIO_OPT_STR,
 103                 .cb     = str_ioengine_cb,
 104                 .help   = "IO engine to use (sync, aio, etc)",
 105         },
 106         {
 107                 .name   = "mem",
 108                 .type   = FIO_OPT_STR,
 109                 .cb     = str_mem_cb,
 110                 .help   = "Backing type for IO buffers (malloc, shm, etc)",
 111         },
 112         {
 113                 .name   = "verify",
 114                 .type   = FIO_OPT_STR,
 115                 .cb     = str_verify_cb,
 116                 .help   = "Verify sum function (md5 or crc32)",
 117         },
 118         {
 119                 .name   = "write_iolog",
 120                 .type   = FIO_OPT_STR_STORE,
 121                 .off1   = td_var_offset(write_iolog_file),
 122                 .help   = "Store IO pattern to file",
 123         },
 124         {
 125                 .name   = "read_iolog",
 126                 .type   = FIO_OPT_STR_STORE,
 127                 .off1   = td_var_offset(read_iolog_file),
 128                 .help   = "Playback IO pattern from file",
 129         },
 130         {
 131                 .name   = "exec_prerun",
 132                 .type   = FIO_OPT_STR_STORE,
 133                 .off1   = td_var_offset(exec_prerun),
 134                 .help   = "Execute this file prior to running job",
 135         },
 136         {
 137                 .name   = "exec_postrun",
 138                 .type   = FIO_OPT_STR_STORE,
 139                 .off1   = td_var_offset(exec_postrun),
 140                 .help   = "Execute this file after running job",
 141         },
 142 #ifdef FIO_HAVE_IOSCHED_SWITCH
 143         {
 144                 .name   = "ioscheduler",
 145                 .type   = FIO_OPT_STR_STORE,
 146                 .off1   = td_var_offset(ioscheduler),
 147                 .help   = "Use this IO scheduler on the backing device",
 148         },
 149 #endif
 150         {
 151                 .name   = "size",
 152                 .type   = FIO_OPT_STR_VAL,
 153                 .off1   = td_var_offset(total_file_size),
 154                 .help   = "Size of device or file",
 155         },
 156         {
 157                 .name   = "bs",
 158                 .type   = FIO_OPT_STR_VAL_INT,
 159                 .off1   = td_var_offset(bs[DDIR_READ]),
 160                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 161                 .help   = "Block size unit",
 162         },
 163         {
 164                 .name   = "offset",
 165                 .type   = FIO_OPT_STR_VAL,
 166                 .off1   = td_var_offset(start_offset),
 167                 .help   = "Start IO from this offset",
 168         },
 169         {
 170                 .name   = "zonesize",
 171                 .type   = FIO_OPT_STR_VAL,
 172                 .off1   = td_var_offset(zone_size),
 173                 .help   = "Give size of an IO zone",
 174         },
 175         {
 176                 .name   = "zoneskip",
 177                 .type   = FIO_OPT_STR_VAL,
 178                 .off1   = td_var_offset(zone_skip),
 179                 .help   = "Space between IO zones",
 180         },
 181         {
 182                 .name   = "lockmem",
 183                 .type   = FIO_OPT_STR_VAL,
 184                 .cb     = str_lockmem_cb,
 185                 .help   = "Lock down this amount of memory",
 186         },
 187         {
 188                 .name   = "bsrange",
 189                 .type   = FIO_OPT_RANGE,
 190                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 191                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 192                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 193                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 194                 .help   = "Set block size range",
 195         },
 196         {
 197                 .name   = "nrfiles",
 198                 .type   = FIO_OPT_INT,
 199                 .off1   = td_var_offset(nr_files),
 200                 .help   = "Split job workload between this number of files",
 201         },
 202         {
 203                 .name   = "iodepth",
 204                 .type   = FIO_OPT_INT,
 205                 .off1   = td_var_offset(iodepth),
 206                 .help   = "Amount of IO buffers to keep in flight",
 207         },
 208         {
 209                 .name   = "fsync",
 210                 .type   = FIO_OPT_INT,
 211                 .off1   = td_var_offset(fsync_blocks),
 212                 .help   = "Issue fsync for writes every given number of blocks",
 213         },
 214         {
 215                 .name   = "rwmixcycle",
 216                 .type   = FIO_OPT_INT,
 217                 .off1   = td_var_offset(rwmixcycle),
 218                 .help   = "Cycle period for mixed read/write workloads",
 219         },
 220         {
 221                 .name   = "rwmixread",
 222                 .type   = FIO_OPT_INT,
 223                 .off1   = td_var_offset(rwmixread),
 224                 .max_val= 100,
 225                 .help   = "Percentage of mixed workload that is reads",
 226         },
 227         {
 228                 .name   = "rwmixwrite",
 229                 .type   = FIO_OPT_INT,
 230                 .off1   = td_var_offset(rwmixwrite),
 231                 .max_val= 100,
 232                 .help   = "Percentage of mixed workload that is writes",
 233         },
 234         {
 235                 .name   = "nice",
 236                 .type   = FIO_OPT_INT,
 237                 .off1   = td_var_offset(nice),
 238                 .help   = "Set job CPU nice value",
 239         },
 240 #ifdef FIO_HAVE_IOPRIO
 241         {
 242                 .name   = "prio",
 243                 .type   = FIO_OPT_INT,
 244                 .cb     = str_prio_cb,
 245                 .help   = "Set job IO priority value",
 246         },
 247         {
 248                 .name   = "prioclass",
 249                 .type   = FIO_OPT_INT,
 250                 .cb     = str_prioclass_cb,
 251                 .help   = "Set job IO priority class",
 252         },
 253 #endif
 254         {
 255                 .name   = "thinktime",
 256                 .type   = FIO_OPT_INT,
 257                 .off1   = td_var_offset(thinktime),
 258                 .help   = "Idle time between IO buffers",
 259         },
 260         {
 261                 .name   = "thinktime_blocks",
 262                 .type   = FIO_OPT_INT,
 263                 .off1   = td_var_offset(thinktime_blocks),
 264                 .help   = "IO buffer period between 'thinktime'",
 265         },
 266         {
 267                 .name   = "rate",
 268                 .type   = FIO_OPT_INT,
 269                 .off1   = td_var_offset(rate),
 270                 .help   = "Set bandwidth rate",
 271         },
 272         {
 273                 .name   = "ratemin",
 274                 .type   = FIO_OPT_INT,
 275                 .off1   = td_var_offset(ratemin),
 276                 .help   = "The bottom limit accepted",
 277         },
 278         {
 279                 .name   = "ratecycle",
 280                 .type   = FIO_OPT_INT,
 281                 .off1   = td_var_offset(ratecycle),
 282                 .name   = "Window average for rate limits",
 283         },
 284         {
 285                 .name   = "startdelay",
 286                 .type   = FIO_OPT_INT,
 287                 .off1   = td_var_offset(start_delay),
 288                 .help   = "Only start job when this period has passed",
 289         },
 290         {
 291                 .name   = "timeout",
 292                 .type   = FIO_OPT_STR_VAL_TIME,
 293                 .off1   = td_var_offset(timeout),
 294                 .help   = "Stop workload when this amount of time has passed",
 295         },
 296         {
 297                 .name   = "invalidate",
 298                 .type   = FIO_OPT_INT,
 299                 .off1   = td_var_offset(invalidate_cache),
 300                 .help   = "Invalidate buffer/page cache prior to running job",
 301         },
 302         {
 303                 .name   = "sync",
 304                 .type   = FIO_OPT_INT,
 305                 .off1   = td_var_offset(sync_io),
 306                 .help   = "Use O_SYNC for buffered writes",
 307         },
 308         {
 309                 .name   = "bwavgtime",
 310                 .type   = FIO_OPT_INT,
 311                 .off1   = td_var_offset(bw_avg_time),
 312                 .help   = "Time window over which to calculate bandwidth",
 313         },
 314         {
 315                 .name   = "create_serialize",
 316                 .type   = FIO_OPT_INT,
 317                 .off1   = td_var_offset(create_serialize),
 318                 .help   = "Serialize creating of job files",
 319         },
 320         {
 321                 .name   = "create_fsync",
 322                 .type   = FIO_OPT_INT,
 323                 .off1   = td_var_offset(create_fsync),
 324                 .help   = "Fsync file after creation",
 325         },
 326         {
 327                 .name   = "loops",
 328                 .type   = FIO_OPT_INT,
 329                 .off1   = td_var_offset(loops),
 330                 .help   = "Number of times to run the job",
 331         },
 332         {
 333                 .name   = "numjobs",
 334                 .type   = FIO_OPT_INT,
 335                 .off1   = td_var_offset(numjobs),
 336                 .help   = "Duplicate this job this many times",
 337         },
 338         {
 339                 .name   = "cpuload",
 340                 .type   = FIO_OPT_INT,
 341                 .off1   = td_var_offset(cpuload),
 342                 .help   = "Use this percentage of CPU",
 343         },
 344         {
 345                 .name   = "cpuchunks",
 346                 .type   = FIO_OPT_INT,
 347                 .off1   = td_var_offset(cpucycle),
 348                 .help   = "Length of the CPU burn cycles",
 349         },
 350         {
 351                 .name   = "direct",
 352                 .type   = FIO_OPT_INT,
 353                 .off1   = td_var_offset(odirect),
 354                 .help   = "Use O_DIRECT IO",
 355         },
 356         {
 357                 .name   = "overwrite",
 358                 .type   = FIO_OPT_INT,
 359                 .off1   = td_var_offset(overwrite),
 360                 .help   = "When writing, set whether to overwrite current data",
 361         },
 362 #ifdef FIO_HAVE_CPU_AFFINITY
 363         {
 364                 .name   = "cpumask",
 365                 .type   = FIO_OPT_INT,
 366                 .cb     = str_cpumask_cb,
 367                 .help   = "CPU affinity mask",
 368         },
 369 #endif
 370         {
 371                 .name   = "end_fsync",
 372                 .type   = FIO_OPT_INT,
 373                 .off1   = td_var_offset(end_fsync),
 374                 .help   = "Include fsync at the end of job",
 375         },
 376         {
 377                 .name   = "unlink",
 378                 .type   = FIO_OPT_INT,
 379                 .off1   = td_var_offset(unlink),
 380                 .help   = "Unlink files after job has completed",
 381         },
 382         {
 383                 .name   = "exitall",
 384                 .type   = FIO_OPT_STR_SET,
 385                 .cb     = str_exitall_cb,
 386                 .help   = "Terminate all jobs when one exits",
 387         },
 388         {
 389                 .name   = "stonewall",
 390                 .type   = FIO_OPT_STR_SET,
 391                 .off1   = td_var_offset(stonewall),
 392                 .help   = "Insert a hard barrier between this job and previous",
 393         },
 394         {
 395                 .name   = "thread",
 396                 .type   = FIO_OPT_STR_SET,
 397                 .off1   = td_var_offset(thread),
 398                 .help   = "Use threads instead of forks",
 399         },
 400         {
 401                 .name   = "write_bw_log",
 402                 .type   = FIO_OPT_STR_SET,
 403                 .off1   = td_var_offset(write_bw_log),
 404                 .help   = "Write log of bandwidth during run",
 405         },
 406         {
 407                 .name   = "write_lat_log",
 408                 .type   = FIO_OPT_STR_SET,
 409                 .off1   = td_var_offset(write_lat_log),
 410                 .help   = "Write log of latency during run",
 411         },
 412         {
 413                 .name   = "norandommap",
 414                 .type   = FIO_OPT_STR_SET,
 415                 .off1   = td_var_offset(norandommap),
 416                 .help   = "Accept potential duplicate random blocks",
 417         },
 418         {
 419                 .name   = "bs_unaligned",
 420                 .type   = FIO_OPT_STR_SET,
 421                 .off1   = td_var_offset(bs_unaligned),
 422                 .help   = "Don't sector align IO buffer sizes",
 423         },
 424         {
 425                 .name   = "hugepage-size",
 426                 .type   = FIO_OPT_STR_VAL,
 427                 .off1   = td_var_offset(hugepage_size),
 428                 .help   = "When using hugepages, specify size of each page",
 429         },
 430         {
 431                 .name = NULL,
 432         },
 433 };
 434
 435 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 436 #define FIO_CMD_OPTS    (16)
 437 #define FIO_GETOPT_JOB  (0x89988998)
 438
 439 /*
 440  * Command line options. These will contain the above, plus a few
 441  * extra that only pertain to fio itself and not jobs.
 442  */
 443 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 444         {
 445                 .name           = "output",
 446                 .has_arg        = required_argument,
 447                 .val            = 'o',
 448         },
 449         {
 450                 .name           = "timeout",
 451                 .has_arg        = required_argument,
 452                 .val            = 't',
 453         },
 454         {
 455                 .name           = "latency-log",
 456                 .has_arg        = required_argument,
 457                 .val            = 'l',
 458         },
 459         {
 460                 .name           = "bandwidth-log",
 461                 .has_arg        = required_argument,
 462                 .val            = 'b',
 463         },
 464         {
 465                 .name           = "minimal",
 466                 .has_arg        = optional_argument,
 467                 .val            = 'm',
 468         },
 469         {
 470                 .name           = "version",
 471                 .has_arg        = no_argument,
 472                 .val            = 'v',
 473         },
 474         {
 475                 .name           = "help",
 476                 .has_arg        = no_argument,
 477                 .val            = 'h',
 478         },
 479         {
 480                 .name           = "cmdhelp",
 481                 .has_arg        = required_argument,
 482                 .val            = 'c',
 483         },
 484         {
 485                 .name           = NULL,
 486         },
 487 };
 488
 489 static int def_timeout = DEF_TIMEOUT;
 490
 491 static char fio_version_string[] = "fio 1.11";
 492
 493 static char **ini_file;
 494 static int max_jobs = MAX_JOBS;
 495
 496 struct thread_data def_thread;
 497 struct thread_data *threads = NULL;
 498
 499 int exitall_on_terminate = 0;
 500 int terse_output = 0;
 501 unsigned long long mlock_size = 0;
 502 FILE *f_out = NULL;
 503 FILE *f_err = NULL;
 504
 505 static int write_lat_log = DEF_WRITE_LAT_LOG;
 506 static int write_bw_log = DEF_WRITE_BW_LOG;
 507
 508 /*
 509  * Return a free job structure.
 510  */
 511 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 512 {
 513         struct thread_data *td;
 514
 515         if (global)
 516                 return &def_thread;
 517         if (thread_number >= max_jobs)
 518                 return NULL;
 519
 520         td = &threads[thread_number++];
 521         *td = *parent;
 522
 523         td->thread_number = thread_number;
 524         return td;
 525 }
 526
 527 static void put_job(struct thread_data *td)
 528 {
 529         if (td == &def_thread)
 530                 return;
 531
 532         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 533         thread_number--;
 534 }
 535
 536 /*
 537  * Lazy way of fixing up options that depend on each other. We could also
 538  * define option callback handlers, but this is easier.
 539  */
 540 static void fixup_options(struct thread_data *td)
 541 {
 542         if (!td->rwmixread && td->rwmixwrite)
 543                 td->rwmixread = 100 - td->rwmixwrite;
 544
 545         if (td->write_iolog_file && td->read_iolog_file) {
 546                 log_err("fio: read iolog overrides write_iolog\n");
 547                 free(td->write_iolog_file);
 548                 td->write_iolog_file = NULL;
 549         }
 550
 551         if (td->io_ops->flags & FIO_SYNCIO)
 552                 td->iodepth = 1;
 553         else {
 554                 if (!td->iodepth)
 555                         td->iodepth = td->nr_files;
 556         }
 557
 558         /*
 559          * only really works for sequential io for now, and with 1 file
 560          */
 561         if (td->zone_size && !td->sequential && td->nr_files == 1)
 562                 td->zone_size = 0;
 563
 564         /*
 565          * Reads can do overwrites, we always need to pre-create the file
 566          */
 567         if (td_read(td) || td_rw(td))
 568                 td->overwrite = 1;
 569
 570         if (!td->min_bs[DDIR_READ])
 571                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 572         if (!td->max_bs[DDIR_READ])
 573                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 574         if (!td->min_bs[DDIR_WRITE])
 575                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 576         if (!td->max_bs[DDIR_WRITE])
 577                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 578
 579         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 580
 581         if (td_read(td) && !td_rw(td))
 582                 td->verify = 0;
 583
 584         if (td->norandommap && td->verify != VERIFY_NONE) {
 585                 log_err("fio: norandommap given, verify disabled\n");
 586                 td->verify = VERIFY_NONE;
 587         }
 588         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 589                 log_err("fio: bs_unaligned may not work with raw io\n");
 590
 591         /*
 592          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 593          */
 594         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 595                 td->odirect = 0;
 596 }
 597
 598 /*
 599  * This function leaks the buffer
 600  */
 601 static char *to_kmg(unsigned int val)
 602 {
 603         char *buf = malloc(32);
 604         char post[] = { 0, 'K', 'M', 'G', 'P', 0 };
 605         char *p = post;
 606
 607         do {
 608                 if (val & 1023)
 609                         break;
 610
 611                 val >>= 10;
 612                 p++;
 613         } while (*p);
 614
 615         snprintf(buf, 31, "%u%c", val, *p);
 616         return buf;
 617 }
 618
 619 /*
 620  * Adds a job to the list of things todo. Sanitizes the various options
 621  * to make sure we don't have conflicts, and initializes various
 622  * members of td.
 623  */
 624 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 625 {
 626         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
 627                                    "rw", NULL, "randrw" };
 628         struct stat sb;
 629         int numjobs, ddir, i;
 630         struct fio_file *f;
 631
 632         /*
 633          * the def_thread is just for options, it's not a real job
 634          */
 635         if (td == &def_thread)
 636                 return 0;
 637
 638         /*
 639          * Set default io engine, if none set
 640          */
 641         if (!td->io_ops) {
 642                 td->io_ops = load_ioengine(td, DEF_IO_ENGINE_NAME);
 643                 if (!td->io_ops) {
 644                         log_err("default engine %s not there?\n", DEF_IO_ENGINE_NAME);
 645                         return 1;
 646                 }
 647         }
 648
 649         if (td->odirect)
 650                 td->io_ops->flags |= FIO_RAWIO;
 651
 652         td->filetype = FIO_TYPE_FILE;
 653         if (!stat(jobname, &sb)) {
 654                 if (S_ISBLK(sb.st_mode))
 655                         td->filetype = FIO_TYPE_BD;
 656                 else if (S_ISCHR(sb.st_mode))
 657                         td->filetype = FIO_TYPE_CHAR;
 658         }
 659
 660         fixup_options(td);
 661
 662         if (td->filename)
 663                 td->nr_uniq_files = 1;
 664         else
 665                 td->nr_uniq_files = td->nr_files;
 666
 667         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 668                 char tmp[PATH_MAX];
 669                 int len = 0;
 670
 671                 if (td->directory && td->directory[0] != '\0')
 672                         len = sprintf(tmp, "%s/", td->directory);
 673
 674                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 675
 676                 for_each_file(td, f, i) {
 677                         memset(f, 0, sizeof(*f));
 678                         f->fd = -1;
 679
 680                         if (td->filename)
 681                                 sprintf(tmp + len, "%s", td->filename);
 682                         else
 683                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 684                         f->file_name = strdup(tmp);
 685                 }
 686         } else {
 687                 td->nr_files = 1;
 688                 td->files = malloc(sizeof(struct fio_file));
 689                 f = &td->files[0];
 690
 691                 memset(f, 0, sizeof(*f));
 692                 f->fd = -1;
 693                 f->file_name = strdup(jobname);
 694         }
 695
 696         for_each_file(td, f, i) {
 697                 f->file_size = td->total_file_size / td->nr_files;
 698                 f->file_offset = td->start_offset;
 699         }
 700
 701         fio_sem_init(&td->mutex, 0);
 702
 703         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
 704         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
 705         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
 706
 707         if (td->stonewall && td->thread_number > 1)
 708                 groupid++;
 709
 710         td->groupid = groupid;
 711
 712         if (setup_rate(td))
 713                 goto err;
 714
 715         if (td->write_lat_log) {
 716                 setup_log(&td->slat_log);
 717                 setup_log(&td->clat_log);
 718         }
 719         if (td->write_bw_log)
 720                 setup_log(&td->bw_log);
 721
 722         if (!td->name)
 723                 td->name = strdup(jobname);
 724
 725         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 726
 727         if (!terse_output) {
 728                 if (!job_add_num) {
 729                         if (td->io_ops->flags & FIO_CPUIO)
 730                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 731                         else {
 732                                 char *c1, *c2, *c3, *c4;
 733
 734                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 735                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 736                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 737                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 738
 739                                 fprintf(f_out, "%s: (g=%d): rw=%s, odir=%u, bs=%s-%s/%s-%s, rate=%u, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[ddir], td->odirect, c1, c2, c3, c4, td->rate, td->io_ops->name, td->iodepth);
 740
 741                                 free(c1);
 742                                 free(c2);
 743                                 free(c3);
 744                                 free(c4);
 745                         }
 746                 } else if (job_add_num == 1)
 747                         fprintf(f_out, "...\n");
 748         }
 749
 750         /*
 751          * recurse add identical jobs, clear numjobs and stonewall options
 752          * as they don't apply to sub-jobs
 753          */
 754         numjobs = td->numjobs;
 755         while (--numjobs) {
 756                 struct thread_data *td_new = get_new_job(0, td);
 757
 758                 if (!td_new)
 759                         goto err;
 760
 761                 td_new->numjobs = 1;
 762                 td_new->stonewall = 0;
 763                 job_add_num = numjobs - 1;
 764
 765                 if (add_job(td_new, jobname, job_add_num))
 766                         goto err;
 767         }
 768         return 0;
 769 err:
 770         put_job(td);
 771         return -1;
 772 }
 773
 774 /*
 775  * Initialize the various random states we need (random io, block size ranges,
 776  * read/write mix, etc).
 777  */
 778 int init_random_state(struct thread_data *td)
 779 {
 780         unsigned long seeds[4];
 781         int fd, num_maps, blocks, i;
 782         struct fio_file *f;
 783
 784         if (td->io_ops->flags & FIO_CPUIO)
 785                 return 0;
 786
 787         fd = open("/dev/urandom", O_RDONLY);
 788         if (fd == -1) {
 789                 td_verror(td, errno);
 790                 return 1;
 791         }
 792
 793         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 794                 td_verror(td, EIO);
 795                 close(fd);
 796                 return 1;
 797         }
 798
 799         close(fd);
 800
 801         os_random_seed(seeds[0], &td->bsrange_state);
 802         os_random_seed(seeds[1], &td->verify_state);
 803         os_random_seed(seeds[2], &td->rwmix_state);
 804
 805         if (td->sequential)
 806                 return 0;
 807
 808         if (td->rand_repeatable)
 809                 seeds[3] = DEF_RANDSEED;
 810
 811         if (!td->norandommap) {
 812                 for_each_file(td, f, i) {
 813                         blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 814                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 815                         f->file_map = malloc(num_maps * sizeof(long));
 816                         f->num_maps = num_maps;
 817                         memset(f->file_map, 0, num_maps * sizeof(long));
 818                 }
 819         }
 820
 821         os_random_seed(seeds[3], &td->random_state);
 822         return 0;
 823 }
 824
 825 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 826 {
 827 #ifdef FIO_HAVE_CPU_AFFINITY
 828         unsigned int i;
 829
 830         CPU_ZERO(&cpumask);
 831
 832         for (i = 0; i < sizeof(int) * 8; i++) {
 833                 if ((1 << i) & cpu)
 834                         CPU_SET(i, &cpumask);
 835         }
 836 #endif
 837 }
 838
 839 static int is_empty_or_comment(char *line)
 840 {
 841         unsigned int i;
 842
 843         for (i = 0; i < strlen(line); i++) {
 844                 if (line[i] == ';')
 845                         return 1;
 846                 if (!isspace(line[i]) && !iscntrl(line[i]))
 847                         return 0;
 848         }
 849
 850         return 1;
 851 }
 852
 853 static int str_rw_cb(void *data, const char *mem)
 854 {
 855         struct thread_data *td = data;
 856
 857         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 858                 td->ddir = DDIR_READ;
 859                 td->sequential = 1;
 860                 return 0;
 861         } else if (!strncmp(mem, "randread", 8)) {
 862                 td->ddir = DDIR_READ;
 863                 td->sequential = 0;
 864                 return 0;
 865         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 866                 td->ddir = DDIR_WRITE;
 867                 td->sequential = 1;
 868                 return 0;
 869         } else if (!strncmp(mem, "randwrite", 9)) {
 870                 td->ddir = DDIR_WRITE;
 871                 td->sequential = 0;
 872                 return 0;
 873         } else if (!strncmp(mem, "rw", 2)) {
 874                 td->ddir = DDIR_READ;
 875                 td->iomix = 1;
 876                 td->sequential = 1;
 877                 return 0;
 878         } else if (!strncmp(mem, "randrw", 6)) {
 879                 td->ddir = DDIR_READ;
 880                 td->iomix = 1;
 881                 td->sequential = 0;
 882                 return 0;
 883         }
 884
 885         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 886         return 1;
 887 }
 888
 889 static int str_verify_cb(void *data, const char *mem)
 890 {
 891         struct thread_data *td = data;
 892
 893         if (!strncmp(mem, "0", 1)) {
 894                 td->verify = VERIFY_NONE;
 895                 return 0;
 896         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 897                 td->verify = VERIFY_MD5;
 898                 return 0;
 899         } else if (!strncmp(mem, "crc32", 5)) {
 900                 td->verify = VERIFY_CRC32;
 901                 return 0;
 902         }
 903
 904         log_err("fio: verify types: md5, crc32\n");
 905         return 1;
 906 }
 907
 908 /*
 909  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
 910  */
 911 static char *get_mmap_file(const char *str)
 912 {
 913         char *p = strstr(str, ":");
 914
 915         if (!p)
 916                 return NULL;
 917
 918         p++;
 919         strip_blank_front(&p);
 920         strip_blank_end(p);
 921         return strdup(p);
 922 }
 923
 924 static int str_mem_cb(void *data, const char *mem)
 925 {
 926         struct thread_data *td = data;
 927
 928         if (!strncmp(mem, "malloc", 6)) {
 929                 td->mem_type = MEM_MALLOC;
 930                 return 0;
 931         } else if (!strncmp(mem, "mmaphuge", 8)) {
 932 #ifdef FIO_HAVE_HUGETLB
 933                 /*
 934                  * mmaphuge must be appended with the actual file
 935                  */
 936                 td->mmapfile = get_mmap_file(mem);
 937                 if (!td->mmapfile) {
 938                         log_err("fio: mmaphuge:/path/to/file\n");
 939                         return 1;
 940                 }
 941
 942                 td->mem_type = MEM_MMAPHUGE;
 943                 return 0;
 944 #else
 945                 log_err("fio: mmaphuge not available\n");
 946                 return 1;
 947 #endif
 948         } else if (!strncmp(mem, "mmap", 4)) {
 949                 /*
 950                  * Check if the user wants file backed memory. It's ok
 951                  * if there's no file given, we'll just use anon mamp then.
 952                  */
 953                 td->mmapfile = get_mmap_file(mem);
 954                 td->mem_type = MEM_MMAP;
 955                 return 0;
 956         } else if (!strncmp(mem, "shmhuge", 7)) {
 957 #ifdef FIO_HAVE_HUGETLB
 958                 td->mem_type = MEM_SHMHUGE;
 959                 return 0;
 960 #else
 961                 log_err("fio: shmhuge not available\n");
 962                 return 1;
 963 #endif
 964         } else if (!strncmp(mem, "shm", 3)) {
 965                 td->mem_type = MEM_SHM;
 966                 return 0;
 967         }
 968
 969         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
 970         return 1;
 971 }
 972
 973 static int str_ioengine_cb(void *data, const char *str)
 974 {
 975         struct thread_data *td = data;
 976
 977         td->io_ops = load_ioengine(td, str);
 978         if (td->io_ops)
 979                 return 0;
 980
 981         log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
 982         log_err("fio: or specify path to dynamic ioengine module\n");
 983         return 1;
 984 }
 985
 986 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
 987 {
 988         mlock_size = *val;
 989         return 0;
 990 }
 991
 992 #ifdef FIO_HAVE_IOPRIO
 993 static int str_prioclass_cb(void *data, unsigned int *val)
 994 {
 995         struct thread_data *td = data;
 996
 997         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
 998         return 0;
 999 }
1000
1001 static int str_prio_cb(void *data, unsigned int *val)
1002 {
1003         struct thread_data *td = data;
1004
1005         td->ioprio |= *val;
1006         return 0;
1007 }
1008 #endif
1009
1010 static int str_exitall_cb(void)
1011 {
1012         exitall_on_terminate = 1;
1013         return 0;
1014 }
1015
1016 static int str_cpumask_cb(void *data, unsigned int *val)
1017 {
1018         struct thread_data *td = data;
1019
1020         fill_cpu_mask(td->cpumask, *val);
1021         return 0;
1022 }
1023
1024 /*
1025  * This is our [ini] type file parser.
1026  */
1027 static int parse_jobs_ini(char *file, int stonewall_flag)
1028 {
1029         unsigned int global;
1030         struct thread_data *td;
1031         char *string, *name;
1032         fpos_t off;
1033         FILE *f;
1034         char *p;
1035         int ret = 0, stonewall;
1036
1037         f = fopen(file, "r");
1038         if (!f) {
1039                 perror("fopen job file");
1040                 return 1;
1041         }
1042
1043         string = malloc(4096);
1044         name = malloc(256);
1045         memset(name, 0, 256);
1046
1047         stonewall = stonewall_flag;
1048         do {
1049                 p = fgets(string, 4095, f);
1050                 if (!p)
1051                         break;
1052                 if (is_empty_or_comment(p))
1053                         continue;
1054                 if (sscanf(p, "[%255s]", name) != 1)
1055                         continue;
1056
1057                 global = !strncmp(name, "global", 6);
1058
1059                 name[strlen(name) - 1] = '\0';
1060
1061                 td = get_new_job(global, &def_thread);
1062                 if (!td) {
1063                         ret = 1;
1064                         break;
1065                 }
1066
1067                 /*
1068                  * Seperate multiple job files by a stonewall
1069                  */
1070                 if (!global && stonewall) {
1071                         td->stonewall = stonewall;
1072                         stonewall = 0;
1073                 }
1074
1075                 fgetpos(f, &off);
1076                 while ((p = fgets(string, 4096, f)) != NULL) {
1077                         if (is_empty_or_comment(p))
1078                                 continue;
1079
1080                         strip_blank_front(&p);
1081
1082                         if (p[0] == '[')
1083                                 break;
1084
1085                         strip_blank_end(p);
1086
1087                         fgetpos(f, &off);
1088
1089                         /*
1090                          * Don't break here, continue parsing options so we
1091                          * dump all the bad ones. Makes trial/error fixups
1092                          * easier on the user.
1093                          */
1094                         ret |= parse_option(p, options, td);
1095                 }
1096
1097                 if (!ret) {
1098                         fsetpos(f, &off);
1099                         ret = add_job(td, name, 0);
1100                 } else {
1101                         log_err("fio: job %s dropped\n", name);
1102                         put_job(td);
1103                 }
1104         } while (!ret);
1105
1106         free(string);
1107         free(name);
1108         fclose(f);
1109         return ret;
1110 }
1111
1112 static int fill_def_thread(void)
1113 {
1114         memset(&def_thread, 0, sizeof(def_thread));
1115
1116         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1117                 perror("sched_getaffinity");
1118                 return 1;
1119         }
1120
1121         /*
1122          * fill globals
1123          */
1124         def_thread.ddir = DDIR_READ;
1125         def_thread.iomix = 0;
1126         def_thread.bs[DDIR_READ] = DEF_BS;
1127         def_thread.bs[DDIR_WRITE] = DEF_BS;
1128         def_thread.min_bs[DDIR_READ] = def_thread.min_bs[DDIR_WRITE] = 0;
1129         def_thread.max_bs[DDIR_READ] = def_thread.max_bs[DDIR_WRITE] = 0;
1130         def_thread.odirect = DEF_ODIRECT;
1131         def_thread.ratecycle = DEF_RATE_CYCLE;
1132         def_thread.sequential = DEF_SEQUENTIAL;
1133         def_thread.timeout = def_timeout;
1134         def_thread.overwrite = DEF_OVERWRITE;
1135         def_thread.invalidate_cache = DEF_INVALIDATE;
1136         def_thread.sync_io = DEF_SYNCIO;
1137         def_thread.mem_type = MEM_MALLOC;
1138         def_thread.bw_avg_time = DEF_BWAVGTIME;
1139         def_thread.create_serialize = DEF_CREATE_SER;
1140         def_thread.create_fsync = DEF_CREATE_FSYNC;
1141         def_thread.loops = DEF_LOOPS;
1142         def_thread.verify = DEF_VERIFY;
1143         def_thread.stonewall = DEF_STONEWALL;
1144         def_thread.numjobs = DEF_NUMJOBS;
1145         def_thread.use_thread = DEF_USE_THREAD;
1146         def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
1147         def_thread.rwmixread = DEF_RWMIX_READ;
1148         def_thread.nice = DEF_NICE;
1149         def_thread.rand_repeatable = DEF_RAND_REPEAT;
1150         def_thread.nr_files = DEF_NR_FILES;
1151         def_thread.unlink = DEF_UNLINK;
1152         def_thread.write_bw_log = write_bw_log;
1153         def_thread.write_lat_log = write_lat_log;
1154         def_thread.norandommap = DEF_NO_RAND_MAP;
1155         def_thread.hugepage_size = DEF_HUGEPAGE_SIZE;
1156         def_thread.thinktime_blocks = DEF_THINKTIME_BLOCKS;
1157 #ifdef FIO_HAVE_DISK_UTIL
1158         def_thread.do_disk_util = 1;
1159 #endif
1160
1161         return 0;
1162 }
1163
1164 static void usage(void)
1165 {
1166         printf("%s\n", fio_version_string);
1167         printf("\t--output\tWrite output to file\n");
1168         printf("\t--timeout\tRuntime in seconds\n");
1169         printf("\t--latency-log\tGenerate per-job latency logs\n");
1170         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1171         printf("\t--minimal\tMinimal (terse) output\n");
1172         printf("\t--version\tPrint version info and exit\n");
1173         printf("\t--help\t\tPrint this page\n");
1174         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1175 }
1176
1177 static int parse_cmd_line(int argc, char *argv[])
1178 {
1179         struct thread_data *td = NULL;
1180         int c, ini_idx = 0, lidx, ret;
1181
1182         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1183                 switch (c) {
1184                 case 't':
1185                         def_timeout = atoi(optarg);
1186                         break;
1187                 case 'l':
1188                         write_lat_log = 1;
1189                         break;
1190                 case 'w':
1191                         write_bw_log = 1;
1192                         break;
1193                 case 'o':
1194                         f_out = fopen(optarg, "w+");
1195                         if (!f_out) {
1196                                 perror("fopen output");
1197                                 exit(1);
1198                         }
1199                         f_err = f_out;
1200                         break;
1201                 case 'm':
1202                         terse_output = 1;
1203                         break;
1204                 case 'h':
1205                         usage();
1206                         exit(0);
1207                 case 'c':
1208                         ret = show_cmd_help(options, optarg);
1209                         exit(ret);
1210                 case 'v':
1211                         printf("%s\n", fio_version_string);
1212                         exit(0);
1213                 case FIO_GETOPT_JOB: {
1214                         const char *opt = long_options[lidx].name;
1215                         char *val = optarg;
1216
1217                         if (!strncmp(opt, "name", 4) && td) {
1218                                 ret = add_job(td, td->name ?: "fio", 0);
1219                                 if (ret) {
1220                                         put_job(td);
1221                                         return 0;
1222                                 }
1223                                 td = NULL;
1224                         }
1225                         if (!td) {
1226                                 int global = !strncmp(val, "global", 6);
1227
1228                                 td = get_new_job(global, &def_thread);
1229                                 if (!td)
1230                                         return 0;
1231                         }
1232
1233                         ret = parse_cmd_option(opt, val, options, td);
1234                         if (ret) {
1235                                 log_err("fio: job dropped\n");
1236                                 put_job(td);
1237                                 td = NULL;
1238                         }
1239                         break;
1240                 }
1241                 default:
1242                         break;
1243                 }
1244         }
1245
1246         if (td) {
1247                 ret = add_job(td, td->name ?: "fio", 0);
1248                 if (ret)
1249                         put_job(td);
1250         }
1251
1252         while (optind < argc) {
1253                 ini_idx++;
1254                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1255                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1256                 optind++;
1257         }
1258
1259         return ini_idx;
1260 }
1261
1262 static void free_shm(void)
1263 {
1264         struct shmid_ds sbuf;
1265
1266         if (threads) {
1267                 shmdt((void *) threads);
1268                 threads = NULL;
1269                 shmctl(shm_id, IPC_RMID, &sbuf);
1270         }
1271 }
1272
1273 /*
1274  * The thread area is shared between the main process and the job
1275  * threads/processes. So setup a shared memory segment that will hold
1276  * all the job info.
1277  */
1278 static int setup_thread_area(void)
1279 {
1280         /*
1281          * 1024 is too much on some machines, scale max_jobs if
1282          * we get a failure that looks like too large a shm segment
1283          */
1284         do {
1285                 size_t size = max_jobs * sizeof(struct thread_data);
1286
1287                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1288                 if (shm_id != -1)
1289                         break;
1290                 if (errno != EINVAL) {
1291                         perror("shmget");
1292                         break;
1293                 }
1294
1295                 max_jobs >>= 1;
1296         } while (max_jobs);
1297
1298         if (shm_id == -1)
1299                 return 1;
1300
1301         threads = shmat(shm_id, NULL, 0);
1302         if (threads == (void *) -1) {
1303                 perror("shmat");
1304                 return 1;
1305         }
1306
1307         atexit(free_shm);
1308         return 0;
1309 }
1310
1311 /*
1312  * Copy the fio options into the long options map, so we mirror
1313  * job and cmd line options.
1314  */
1315 static void dupe_job_options(void)
1316 {
1317         struct fio_option *o;
1318         unsigned int i;
1319
1320         i = 0;
1321         while (long_options[i].name)
1322                 i++;
1323
1324         o = &options[0];
1325         while (o->name) {
1326                 long_options[i].name = o->name;
1327                 long_options[i].val = FIO_GETOPT_JOB;
1328                 if (o->type == FIO_OPT_STR_SET)
1329                         long_options[i].has_arg = no_argument;
1330                 else
1331                         long_options[i].has_arg = required_argument;
1332
1333                 i++;
1334                 o++;
1335                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1336         }
1337 }
1338
1339 int parse_options(int argc, char *argv[])
1340 {
1341         int job_files, i;
1342
1343         f_out = stdout;
1344         f_err = stderr;
1345
1346         dupe_job_options();
1347
1348         if (setup_thread_area())
1349                 return 1;
1350         if (fill_def_thread())
1351                 return 1;
1352
1353         job_files = parse_cmd_line(argc, argv);
1354
1355         for (i = 0; i < job_files; i++) {
1356                 if (fill_def_thread())
1357                         return 1;
1358                 if (parse_jobs_ini(ini_file[i], i))
1359                         return 1;
1360                 free(ini_file[i]);
1361         }
1362
1363         free(ini_file);
1364
1365         if (!thread_number) {
1366                 log_err("No jobs defined(s)\n");
1367                 return 1;
1368         }
1369
1370         return 0;
1371 }