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         },
  81         {
  82                 .name   = "directory",
  83                 .type   = FIO_OPT_STR_STORE,
  84                 .off1   = td_var_offset(directory),
  85         },
  86         {
  87                 .name   = "filename",
  88                 .type   = FIO_OPT_STR_STORE,
  89                 .off1   = td_var_offset(filename),
  90         },
  91         {
  92                 .name   = "rw",
  93                 .type   = FIO_OPT_STR,
  94                 .cb     = str_rw_cb,
  95         },
  96         {
  97                 .name   = "ioengine",
  98                 .type   = FIO_OPT_STR,
  99                 .cb     = str_ioengine_cb,
 100         },
 101         {
 102                 .name   = "mem",
 103                 .type   = FIO_OPT_STR,
 104                 .cb     = str_mem_cb,
 105         },
 106         {
 107                 .name   = "verify",
 108                 .type   = FIO_OPT_STR,
 109                 .cb     = str_verify_cb,
 110         },
 111         {
 112                 .name   = "write_iolog",
 113                 .type   = FIO_OPT_STR_STORE,
 114                 .off1   = td_var_offset(write_iolog_file),
 115         },
 116         {
 117                 .name   = "read_iolog",
 118                 .type   = FIO_OPT_STR_STORE,
 119                 .off1   = td_var_offset(read_iolog_file),
 120         },
 121         {
 122                 .name   = "exec_prerun",
 123                 .type   = FIO_OPT_STR_STORE,
 124                 .off1   = td_var_offset(exec_prerun),
 125         },
 126         {
 127                 .name   = "exec_postrun",
 128                 .type   = FIO_OPT_STR_STORE,
 129                 .off1   = td_var_offset(exec_postrun),
 130         },
 131 #ifdef FIO_HAVE_IOSCHED_SWITCH
 132         {
 133                 .name   = "ioscheduler",
 134                 .type   = FIO_OPT_STR_STORE,
 135                 .off1   = td_var_offset(ioscheduler),
 136         },
 137 #endif
 138         {
 139                 .name   = "size",
 140                 .type   = FIO_OPT_STR_VAL,
 141                 .off1   = td_var_offset(total_file_size),
 142         },
 143         {
 144                 .name   = "bs",
 145                 .type   = FIO_OPT_STR_VAL_INT,
 146                 .off1   = td_var_offset(bs[DDIR_READ]),
 147                 .off2   = td_var_offset(bs[DDIR_WRITE]),
 148         },
 149         {
 150                 .name   = "offset",
 151                 .type   = FIO_OPT_STR_VAL,
 152                 .off1   = td_var_offset(start_offset),
 153         },
 154         {
 155                 .name   = "zonesize",
 156                 .type   = FIO_OPT_STR_VAL,
 157                 .off1   = td_var_offset(zone_size),
 158         },
 159         {
 160                 .name   = "zoneskip",
 161                 .type   = FIO_OPT_STR_VAL,
 162                 .off1   = td_var_offset(zone_skip),
 163         },
 164         {
 165                 .name   = "lockmem",
 166                 .type   = FIO_OPT_STR_VAL,
 167                 .cb     = str_lockmem_cb,
 168         },
 169         {
 170                 .name   = "bsrange",
 171                 .type   = FIO_OPT_RANGE,
 172                 .off1   = td_var_offset(min_bs[DDIR_READ]),
 173                 .off2   = td_var_offset(max_bs[DDIR_READ]),
 174                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
 175                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
 176         },
 177         {
 178                 .name   = "nrfiles",
 179                 .type   = FIO_OPT_INT,
 180                 .off1   = td_var_offset(nr_files),
 181         },
 182         {
 183                 .name   = "iodepth",
 184                 .type   = FIO_OPT_INT,
 185                 .off1   = td_var_offset(iodepth),
 186         },
 187         {
 188                 .name   = "fsync",
 189                 .type   = FIO_OPT_INT,
 190                 .off1   = td_var_offset(fsync_blocks),
 191         },
 192         {
 193                 .name   = "rwmixcycle",
 194                 .type   = FIO_OPT_INT,
 195                 .off1   = td_var_offset(rwmixcycle),
 196         },
 197         {
 198                 .name   = "rwmixread",
 199                 .type   = FIO_OPT_INT,
 200                 .off1   = td_var_offset(rwmixread),
 201                 .max_val= 100,
 202         },
 203         {
 204                 .name   = "rwmixwrite",
 205                 .type   = FIO_OPT_INT,
 206                 .off1   = td_var_offset(rwmixwrite),
 207                 .max_val= 100,
 208         },
 209         {
 210                 .name   = "nice",
 211                 .type   = FIO_OPT_INT,
 212                 .off1   = td_var_offset(nice),
 213         },
 214 #ifdef FIO_HAVE_IOPRIO
 215         {
 216                 .name   = "prio",
 217                 .type   = FIO_OPT_INT,
 218                 .cb     = str_prio_cb,
 219         },
 220         {
 221                 .name   = "prioclass",
 222                 .type   = FIO_OPT_INT,
 223                 .cb     = str_prioclass_cb,
 224         },
 225 #endif
 226         {
 227                 .name   = "thinktime",
 228                 .type   = FIO_OPT_INT,
 229                 .off1   = td_var_offset(thinktime),
 230         },
 231         {
 232                 .name   = "thinktime_blocks",
 233                 .type   = FIO_OPT_INT,
 234                 .off1   = td_var_offset(thinktime_blocks),
 235         },
 236         {
 237                 .name   = "rate",
 238                 .type   = FIO_OPT_INT,
 239                 .off1   = td_var_offset(rate),
 240         },
 241         {
 242                 .name   = "ratemin",
 243                 .type   = FIO_OPT_INT,
 244                 .off1   = td_var_offset(ratemin),
 245         },
 246         {
 247                 .name   = "ratecycle",
 248                 .type   = FIO_OPT_INT,
 249                 .off1   = td_var_offset(ratecycle),
 250         },
 251         {
 252                 .name   = "startdelay",
 253                 .type   = FIO_OPT_INT,
 254                 .off1   = td_var_offset(start_delay),
 255         },
 256         {
 257                 .name   = "timeout",
 258                 .type   = FIO_OPT_STR_VAL_TIME,
 259                 .off1   = td_var_offset(timeout),
 260         },
 261         {
 262                 .name   = "invalidate",
 263                 .type   = FIO_OPT_INT,
 264                 .off1   = td_var_offset(invalidate_cache),
 265         },
 266         {
 267                 .name   = "sync",
 268                 .type   = FIO_OPT_INT,
 269                 .off1   = td_var_offset(sync_io),
 270         },
 271         {
 272                 .name   = "bwavgtime",
 273                 .type   = FIO_OPT_INT,
 274                 .off1   = td_var_offset(bw_avg_time),
 275         },
 276         {
 277                 .name   = "create_serialize",
 278                 .type   = FIO_OPT_INT,
 279                 .off1   = td_var_offset(create_serialize),
 280         },
 281         {
 282                 .name   = "create_fsync",
 283                 .type   = FIO_OPT_INT,
 284                 .off1   = td_var_offset(create_fsync),
 285         },
 286         {
 287                 .name   = "loops",
 288                 .type   = FIO_OPT_INT,
 289                 .off1   = td_var_offset(loops),
 290         },
 291         {
 292                 .name   = "numjobs",
 293                 .type   = FIO_OPT_INT,
 294                 .off1   = td_var_offset(numjobs),
 295         },
 296         {
 297                 .name   = "cpuload",
 298                 .type   = FIO_OPT_INT,
 299                 .off1   = td_var_offset(cpuload),
 300         },
 301         {
 302                 .name   = "cpuchunks",
 303                 .type   = FIO_OPT_INT,
 304                 .off1   = td_var_offset(cpucycle),
 305         },
 306         {
 307                 .name   = "direct",
 308                 .type   = FIO_OPT_INT,
 309                 .off1   = td_var_offset(odirect),
 310         },
 311         {
 312                 .name   = "overwrite",
 313                 .type   = FIO_OPT_INT,
 314                 .off1   = td_var_offset(overwrite),
 315         },
 316 #ifdef FIO_HAVE_CPU_AFFINITY
 317         {
 318                 .name   = "cpumask",
 319                 .type   = FIO_OPT_INT,
 320                 .cb     = str_cpumask_cb,
 321         },
 322 #endif
 323         {
 324                 .name   = "end_fsync",
 325                 .type   = FIO_OPT_INT,
 326                 .off1   = td_var_offset(end_fsync),
 327         },
 328         {
 329                 .name   = "unlink",
 330                 .type   = FIO_OPT_INT,
 331                 .off1   = td_var_offset(unlink),
 332         },
 333         {
 334                 .name   = "exitall",
 335                 .type   = FIO_OPT_STR_SET,
 336                 .cb     = str_exitall_cb,
 337         },
 338         {
 339                 .name   = "stonewall",
 340                 .type   = FIO_OPT_STR_SET,
 341                 .off1   = td_var_offset(stonewall),
 342         },
 343         {
 344                 .name   = "thread",
 345                 .type   = FIO_OPT_STR_SET,
 346                 .off1   = td_var_offset(thread),
 347         },
 348         {
 349                 .name   = "write_bw_log",
 350                 .type   = FIO_OPT_STR_SET,
 351                 .off1   = td_var_offset(write_bw_log),
 352         },
 353         {
 354                 .name   = "write_lat_log",
 355                 .type   = FIO_OPT_STR_SET,
 356                 .off1   = td_var_offset(write_lat_log),
 357         },
 358         {
 359                 .name   = "norandommap",
 360                 .type   = FIO_OPT_STR_SET,
 361                 .off1   = td_var_offset(norandommap),
 362         },
 363         {
 364                 .name   = "bs_unaligned",
 365                 .type   = FIO_OPT_STR_SET,
 366                 .off1   = td_var_offset(bs_unaligned),
 367         },
 368         {
 369                 .name   = "hugepage-size",
 370                 .type   = FIO_OPT_STR_VAL,
 371                 .off1   = td_var_offset(hugepage_size),
 372         },
 373         {
 374                 .name = NULL,
 375         },
 376 };
 377
 378 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
 379 #define FIO_CMD_OPTS    (16)
 380 #define FIO_GETOPT_JOB  (0x89988998)
 381
 382 /*
 383  * Command line options. These will contain the above, plus a few
 384  * extra that only pertain to fio itself and not jobs.
 385  */
 386 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
 387         {
 388                 .name           = "output",
 389                 .has_arg        = required_argument,
 390                 .val            = 'o',
 391         },
 392         {
 393                 .name           = "timeout",
 394                 .has_arg        = required_argument,
 395                 .val            = 't',
 396         },
 397         {
 398                 .name           = "latency-log",
 399                 .has_arg        = required_argument,
 400                 .val            = 'l',
 401         },
 402         {
 403                 .name           = "bandwidth-log",
 404                 .has_arg        = required_argument,
 405                 .val            = 'b',
 406         },
 407         {
 408                 .name           = "minimal",
 409                 .has_arg        = optional_argument,
 410                 .val            = 'm',
 411         },
 412         {
 413                 .name           = "version",
 414                 .has_arg        = no_argument,
 415                 .val            = 'v',
 416         },
 417         {
 418                 .name           = NULL,
 419         },
 420 };
 421
 422 static int def_timeout = DEF_TIMEOUT;
 423
 424 static char fio_version_string[] = "fio 1.10";
 425
 426 static char **ini_file;
 427 static int max_jobs = MAX_JOBS;
 428
 429 struct thread_data def_thread;
 430 struct thread_data *threads = NULL;
 431
 432 int exitall_on_terminate = 0;
 433 int terse_output = 0;
 434 unsigned long long mlock_size = 0;
 435 FILE *f_out = NULL;
 436 FILE *f_err = NULL;
 437
 438 static int write_lat_log = DEF_WRITE_LAT_LOG;
 439 static int write_bw_log = DEF_WRITE_BW_LOG;
 440
 441 /*
 442  * Return a free job structure.
 443  */
 444 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 445 {
 446         struct thread_data *td;
 447
 448         if (global)
 449                 return &def_thread;
 450         if (thread_number >= max_jobs)
 451                 return NULL;
 452
 453         td = &threads[thread_number++];
 454         *td = *parent;
 455
 456         td->thread_number = thread_number;
 457         return td;
 458 }
 459
 460 static void put_job(struct thread_data *td)
 461 {
 462         if (td == &def_thread)
 463                 return;
 464
 465         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 466         thread_number--;
 467 }
 468
 469 /*
 470  * Lazy way of fixing up options that depend on each other. We could also
 471  * define option callback handlers, but this is easier.
 472  */
 473 static void fixup_options(struct thread_data *td)
 474 {
 475         if (!td->rwmixread && td->rwmixwrite)
 476                 td->rwmixread = 100 - td->rwmixwrite;
 477
 478         if (td->write_iolog_file && td->read_iolog_file) {
 479                 log_err("fio: read iolog overrides write_iolog\n");
 480                 free(td->write_iolog_file);
 481                 td->write_iolog_file = NULL;
 482         }
 483
 484         if (td->io_ops->flags & FIO_SYNCIO)
 485                 td->iodepth = 1;
 486         else {
 487                 if (!td->iodepth)
 488                         td->iodepth = td->nr_files;
 489         }
 490
 491         /*
 492          * only really works for sequential io for now, and with 1 file
 493          */
 494         if (td->zone_size && !td->sequential && td->nr_files == 1)
 495                 td->zone_size = 0;
 496
 497         /*
 498          * Reads can do overwrites, we always need to pre-create the file
 499          */
 500         if (td_read(td) || td_rw(td))
 501                 td->overwrite = 1;
 502
 503         if (!td->min_bs[DDIR_READ])
 504                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 505         if (!td->max_bs[DDIR_READ])
 506                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 507         if (!td->min_bs[DDIR_WRITE])
 508                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 509         if (!td->max_bs[DDIR_WRITE])
 510                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 511
 512         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 513
 514         if (td_read(td) && !td_rw(td))
 515                 td->verify = 0;
 516
 517         if (td->norandommap && td->verify != VERIFY_NONE) {
 518                 log_err("fio: norandommap given, verify disabled\n");
 519                 td->verify = VERIFY_NONE;
 520         }
 521         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 522                 log_err("fio: bs_unaligned may not work with raw io\n");
 523
 524         /*
 525          * O_DIRECT and char doesn't mix, clear that flag if necessary.
 526          */
 527         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
 528                 td->odirect = 0;
 529 }
 530
 531 /*
 532  * This function leaks the buffer
 533  */
 534 static char *to_kmg(unsigned int val)
 535 {
 536         char *buf = malloc(32);
 537         char post[] = { 0, 'K', 'M', 'G', 'P', 0 };
 538         char *p = post;
 539
 540         do {
 541                 if (val & 1023)
 542                         break;
 543
 544                 val >>= 10;
 545                 p++;
 546         } while (*p);
 547
 548         snprintf(buf, 31, "%u%c", val, *p);
 549         return buf;
 550 }
 551
 552 /*
 553  * Adds a job to the list of things todo. Sanitizes the various options
 554  * to make sure we don't have conflicts, and initializes various
 555  * members of td.
 556  */
 557 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 558 {
 559         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
 560                                    "rw", NULL, "randrw" };
 561         struct stat sb;
 562         int numjobs, ddir, i;
 563         struct fio_file *f;
 564
 565         /*
 566          * the def_thread is just for options, it's not a real job
 567          */
 568         if (td == &def_thread)
 569                 return 0;
 570
 571         /*
 572          * Set default io engine, if none set
 573          */
 574         if (!td->io_ops) {
 575                 td->io_ops = load_ioengine(td, DEF_IO_ENGINE_NAME);
 576                 if (!td->io_ops) {
 577                         log_err("default engine %s not there?\n", DEF_IO_ENGINE_NAME);
 578                         return 1;
 579                 }
 580         }
 581
 582         if (td->odirect)
 583                 td->io_ops->flags |= FIO_RAWIO;
 584
 585         td->filetype = FIO_TYPE_FILE;
 586         if (!stat(jobname, &sb)) {
 587                 if (S_ISBLK(sb.st_mode))
 588                         td->filetype = FIO_TYPE_BD;
 589                 else if (S_ISCHR(sb.st_mode))
 590                         td->filetype = FIO_TYPE_CHAR;
 591         }
 592
 593         fixup_options(td);
 594
 595         if (td->filename)
 596                 td->nr_uniq_files = 1;
 597         else
 598                 td->nr_uniq_files = td->nr_files;
 599
 600         if (td->filetype == FIO_TYPE_FILE || td->filename) {
 601                 char tmp[PATH_MAX];
 602                 int len = 0;
 603
 604                 if (td->directory && td->directory[0] != '\0')
 605                         len = sprintf(tmp, "%s/", td->directory);
 606
 607                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
 608
 609                 for_each_file(td, f, i) {
 610                         memset(f, 0, sizeof(*f));
 611                         f->fd = -1;
 612
 613                         if (td->filename)
 614                                 sprintf(tmp + len, "%s", td->filename);
 615                         else
 616                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
 617                         f->file_name = strdup(tmp);
 618                         fprintf(stderr, "%s\n", f->file_name);
 619                 }
 620         } else {
 621                 td->nr_files = 1;
 622                 td->files = malloc(sizeof(struct fio_file));
 623                 f = &td->files[0];
 624
 625                 memset(f, 0, sizeof(*f));
 626                 f->fd = -1;
 627                 f->file_name = strdup(jobname);
 628         }
 629
 630         for_each_file(td, f, i) {
 631                 f->file_size = td->total_file_size / td->nr_files;
 632                 f->file_offset = td->start_offset;
 633         }
 634
 635         fio_sem_init(&td->mutex, 0);
 636
 637         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
 638         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
 639         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
 640
 641         if (td->stonewall && td->thread_number > 1)
 642                 groupid++;
 643
 644         td->groupid = groupid;
 645
 646         if (setup_rate(td))
 647                 goto err;
 648
 649         if (td->write_lat_log) {
 650                 setup_log(&td->slat_log);
 651                 setup_log(&td->clat_log);
 652         }
 653         if (td->write_bw_log)
 654                 setup_log(&td->bw_log);
 655
 656         if (!td->name)
 657                 td->name = strdup(jobname);
 658
 659         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
 660
 661         if (!terse_output) {
 662                 if (!job_add_num) {
 663                         if (td->io_ops->flags & FIO_CPUIO)
 664                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 665                         else {
 666                                 char *c1, *c2, *c3, *c4;
 667
 668                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 669                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 670                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 671                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 672
 673                                 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);
 674
 675                                 free(c1);
 676                                 free(c2);
 677                                 free(c3);
 678                                 free(c4);
 679                         }
 680                 } else if (job_add_num == 1)
 681                         fprintf(f_out, "...\n");
 682         }
 683
 684         /*
 685          * recurse add identical jobs, clear numjobs and stonewall options
 686          * as they don't apply to sub-jobs
 687          */
 688         numjobs = td->numjobs;
 689         while (--numjobs) {
 690                 struct thread_data *td_new = get_new_job(0, td);
 691
 692                 if (!td_new)
 693                         goto err;
 694
 695                 td_new->numjobs = 1;
 696                 td_new->stonewall = 0;
 697                 job_add_num = numjobs - 1;
 698
 699                 if (add_job(td_new, jobname, job_add_num))
 700                         goto err;
 701         }
 702         return 0;
 703 err:
 704         put_job(td);
 705         return -1;
 706 }
 707
 708 /*
 709  * Initialize the various random states we need (random io, block size ranges,
 710  * read/write mix, etc).
 711  */
 712 int init_random_state(struct thread_data *td)
 713 {
 714         unsigned long seeds[4];
 715         int fd, num_maps, blocks, i;
 716         struct fio_file *f;
 717
 718         if (td->io_ops->flags & FIO_CPUIO)
 719                 return 0;
 720
 721         fd = open("/dev/urandom", O_RDONLY);
 722         if (fd == -1) {
 723                 td_verror(td, errno);
 724                 return 1;
 725         }
 726
 727         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 728                 td_verror(td, EIO);
 729                 close(fd);
 730                 return 1;
 731         }
 732
 733         close(fd);
 734
 735         os_random_seed(seeds[0], &td->bsrange_state);
 736         os_random_seed(seeds[1], &td->verify_state);
 737         os_random_seed(seeds[2], &td->rwmix_state);
 738
 739         if (td->sequential)
 740                 return 0;
 741
 742         if (td->rand_repeatable)
 743                 seeds[3] = DEF_RANDSEED;
 744
 745         if (!td->norandommap) {
 746                 for_each_file(td, f, i) {
 747                         blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 748                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 749                         f->file_map = malloc(num_maps * sizeof(long));
 750                         f->num_maps = num_maps;
 751                         memset(f->file_map, 0, num_maps * sizeof(long));
 752                 }
 753         }
 754
 755         os_random_seed(seeds[3], &td->random_state);
 756         return 0;
 757 }
 758
 759 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
 760 {
 761 #ifdef FIO_HAVE_CPU_AFFINITY
 762         unsigned int i;
 763
 764         CPU_ZERO(&cpumask);
 765
 766         for (i = 0; i < sizeof(int) * 8; i++) {
 767                 if ((1 << i) & cpu)
 768                         CPU_SET(i, &cpumask);
 769         }
 770 #endif
 771 }
 772
 773 static int is_empty_or_comment(char *line)
 774 {
 775         unsigned int i;
 776
 777         for (i = 0; i < strlen(line); i++) {
 778                 if (line[i] == ';')
 779                         return 1;
 780                 if (!isspace(line[i]) && !iscntrl(line[i]))
 781                         return 0;
 782         }
 783
 784         return 1;
 785 }
 786
 787 static int str_rw_cb(void *data, const char *mem)
 788 {
 789         struct thread_data *td = data;
 790
 791         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
 792                 td->ddir = DDIR_READ;
 793                 td->sequential = 1;
 794                 return 0;
 795         } else if (!strncmp(mem, "randread", 8)) {
 796                 td->ddir = DDIR_READ;
 797                 td->sequential = 0;
 798                 return 0;
 799         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
 800                 td->ddir = DDIR_WRITE;
 801                 td->sequential = 1;
 802                 return 0;
 803         } else if (!strncmp(mem, "randwrite", 9)) {
 804                 td->ddir = DDIR_WRITE;
 805                 td->sequential = 0;
 806                 return 0;
 807         } else if (!strncmp(mem, "rw", 2)) {
 808                 td->ddir = DDIR_READ;
 809                 td->iomix = 1;
 810                 td->sequential = 1;
 811                 return 0;
 812         } else if (!strncmp(mem, "randrw", 6)) {
 813                 td->ddir = DDIR_READ;
 814                 td->iomix = 1;
 815                 td->sequential = 0;
 816                 return 0;
 817         }
 818
 819         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
 820         return 1;
 821 }
 822
 823 static int str_verify_cb(void *data, const char *mem)
 824 {
 825         struct thread_data *td = data;
 826
 827         if (!strncmp(mem, "0", 1)) {
 828                 td->verify = VERIFY_NONE;
 829                 return 0;
 830         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
 831                 td->verify = VERIFY_MD5;
 832                 return 0;
 833         } else if (!strncmp(mem, "crc32", 5)) {
 834                 td->verify = VERIFY_CRC32;
 835                 return 0;
 836         }
 837
 838         log_err("fio: verify types: md5, crc32\n");
 839         return 1;
 840 }
 841
 842 /*
 843  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
 844  */
 845 static char *get_mmap_file(const char *str)
 846 {
 847         char *p = strstr(str, ":");
 848
 849         if (!p)
 850                 return NULL;
 851
 852         p++;
 853         strip_blank_front(&p);
 854         strip_blank_end(p);
 855         return strdup(p);
 856 }
 857
 858 static int str_mem_cb(void *data, const char *mem)
 859 {
 860         struct thread_data *td = data;
 861
 862         if (!strncmp(mem, "malloc", 6)) {
 863                 td->mem_type = MEM_MALLOC;
 864                 return 0;
 865         } else if (!strncmp(mem, "mmaphuge", 8)) {
 866 #ifdef FIO_HAVE_HUGETLB
 867                 /*
 868                  * mmaphuge must be appended with the actual file
 869                  */
 870                 td->mmapfile = get_mmap_file(mem);
 871                 if (!td->mmapfile) {
 872                         log_err("fio: mmaphuge:/path/to/file\n");
 873                         return 1;
 874                 }
 875
 876                 td->mem_type = MEM_MMAPHUGE;
 877                 return 0;
 878 #else
 879                 log_err("fio: mmaphuge not available\n");
 880                 return 1;
 881 #endif
 882         } else if (!strncmp(mem, "mmap", 4)) {
 883                 /*
 884                  * Check if the user wants file backed memory. It's ok
 885                  * if there's no file given, we'll just use anon mamp then.
 886                  */
 887                 td->mmapfile = get_mmap_file(mem);
 888                 td->mem_type = MEM_MMAP;
 889                 return 0;
 890         } else if (!strncmp(mem, "shmhuge", 7)) {
 891 #ifdef FIO_HAVE_HUGETLB
 892                 td->mem_type = MEM_SHMHUGE;
 893                 return 0;
 894 #else
 895                 log_err("fio: shmhuge not available\n");
 896                 return 1;
 897 #endif
 898         } else if (!strncmp(mem, "shm", 3)) {
 899                 td->mem_type = MEM_SHM;
 900                 return 0;
 901         }
 902
 903         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
 904         return 1;
 905 }
 906
 907 static int str_ioengine_cb(void *data, const char *str)
 908 {
 909         struct thread_data *td = data;
 910
 911         td->io_ops = load_ioengine(td, str);
 912         if (td->io_ops)
 913                 return 0;
 914
 915         log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
 916         log_err("fio: or specify path to dynamic ioengine module\n");
 917         return 1;
 918 }
 919
 920 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
 921 {
 922         mlock_size = *val;
 923         return 0;
 924 }
 925
 926 #ifdef FIO_HAVE_IOPRIO
 927 static int str_prioclass_cb(void *data, unsigned int *val)
 928 {
 929         struct thread_data *td = data;
 930
 931         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
 932         return 0;
 933 }
 934
 935 static int str_prio_cb(void *data, unsigned int *val)
 936 {
 937         struct thread_data *td = data;
 938
 939         td->ioprio |= *val;
 940         return 0;
 941 }
 942 #endif
 943
 944 static int str_exitall_cb(void)
 945 {
 946         exitall_on_terminate = 1;
 947         return 0;
 948 }
 949
 950 static int str_cpumask_cb(void *data, unsigned int *val)
 951 {
 952         struct thread_data *td = data;
 953
 954         fill_cpu_mask(td->cpumask, *val);
 955         return 0;
 956 }
 957
 958 /*
 959  * This is our [ini] type file parser.
 960  */
 961 static int parse_jobs_ini(char *file, int stonewall_flag)
 962 {
 963         unsigned int global;
 964         struct thread_data *td;
 965         char *string, *name;
 966         fpos_t off;
 967         FILE *f;
 968         char *p;
 969         int ret = 0, stonewall;
 970
 971         f = fopen(file, "r");
 972         if (!f) {
 973                 perror("fopen job file");
 974                 return 1;
 975         }
 976
 977         string = malloc(4096);
 978         name = malloc(256);
 979         memset(name, 0, 256);
 980
 981         stonewall = stonewall_flag;
 982         do {
 983                 p = fgets(string, 4095, f);
 984                 if (!p)
 985                         break;
 986                 if (is_empty_or_comment(p))
 987                         continue;
 988                 if (sscanf(p, "[%255s]", name) != 1)
 989                         continue;
 990
 991                 global = !strncmp(name, "global", 6);
 992
 993                 name[strlen(name) - 1] = '\0';
 994
 995                 td = get_new_job(global, &def_thread);
 996                 if (!td) {
 997                         ret = 1;
 998                         break;
 999                 }
1000
1001                 /*
1002                  * Seperate multiple job files by a stonewall
1003                  */
1004                 if (!global && stonewall) {
1005                         td->stonewall = stonewall;
1006                         stonewall = 0;
1007                 }
1008
1009                 fgetpos(f, &off);
1010                 while ((p = fgets(string, 4096, f)) != NULL) {
1011                         if (is_empty_or_comment(p))
1012                                 continue;
1013
1014                         strip_blank_front(&p);
1015
1016                         if (p[0] == '[')
1017                                 break;
1018
1019                         strip_blank_end(p);
1020
1021                         fgetpos(f, &off);
1022
1023                         /*
1024                          * Don't break here, continue parsing options so we
1025                          * dump all the bad ones. Makes trial/error fixups
1026                          * easier on the user.
1027                          */
1028                         ret |= parse_option(p, options, td);
1029                 }
1030
1031                 if (!ret) {
1032                         fsetpos(f, &off);
1033                         ret = add_job(td, name, 0);
1034                 } else {
1035                         log_err("fio: job %s dropped\n", name);
1036                         put_job(td);
1037                 }
1038         } while (!ret);
1039
1040         free(string);
1041         free(name);
1042         fclose(f);
1043         return ret;
1044 }
1045
1046 static int fill_def_thread(void)
1047 {
1048         memset(&def_thread, 0, sizeof(def_thread));
1049
1050         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1051                 perror("sched_getaffinity");
1052                 return 1;
1053         }
1054
1055         /*
1056          * fill globals
1057          */
1058         def_thread.ddir = DDIR_READ;
1059         def_thread.iomix = 0;
1060         def_thread.bs[DDIR_READ] = DEF_BS;
1061         def_thread.bs[DDIR_WRITE] = DEF_BS;
1062         def_thread.min_bs[DDIR_READ] = def_thread.min_bs[DDIR_WRITE] = 0;
1063         def_thread.max_bs[DDIR_READ] = def_thread.max_bs[DDIR_WRITE] = 0;
1064         def_thread.odirect = DEF_ODIRECT;
1065         def_thread.ratecycle = DEF_RATE_CYCLE;
1066         def_thread.sequential = DEF_SEQUENTIAL;
1067         def_thread.timeout = def_timeout;
1068         def_thread.overwrite = DEF_OVERWRITE;
1069         def_thread.invalidate_cache = DEF_INVALIDATE;
1070         def_thread.sync_io = DEF_SYNCIO;
1071         def_thread.mem_type = MEM_MALLOC;
1072         def_thread.bw_avg_time = DEF_BWAVGTIME;
1073         def_thread.create_serialize = DEF_CREATE_SER;
1074         def_thread.create_fsync = DEF_CREATE_FSYNC;
1075         def_thread.loops = DEF_LOOPS;
1076         def_thread.verify = DEF_VERIFY;
1077         def_thread.stonewall = DEF_STONEWALL;
1078         def_thread.numjobs = DEF_NUMJOBS;
1079         def_thread.use_thread = DEF_USE_THREAD;
1080         def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
1081         def_thread.rwmixread = DEF_RWMIX_READ;
1082         def_thread.nice = DEF_NICE;
1083         def_thread.rand_repeatable = DEF_RAND_REPEAT;
1084         def_thread.nr_files = DEF_NR_FILES;
1085         def_thread.unlink = DEF_UNLINK;
1086         def_thread.write_bw_log = write_bw_log;
1087         def_thread.write_lat_log = write_lat_log;
1088         def_thread.norandommap = DEF_NO_RAND_MAP;
1089         def_thread.hugepage_size = DEF_HUGEPAGE_SIZE;
1090         def_thread.thinktime_blocks = DEF_THINKTIME_BLOCKS;
1091 #ifdef FIO_HAVE_DISK_UTIL
1092         def_thread.do_disk_util = 1;
1093 #endif
1094
1095         return 0;
1096 }
1097
1098 static void usage(void)
1099 {
1100         printf("%s\n", fio_version_string);
1101         printf("\t--output\tWrite output to file\n");
1102         printf("\t--timeout\tRuntime in seconds\n");
1103         printf("\t--latency-log\tGenerate per-job latency logs\n");
1104         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1105         printf("\t--minimal\tMinimal (terse) output\n");
1106         printf("\t--version\tPrint version info and exit\n");
1107 }
1108
1109 static int parse_cmd_line(int argc, char *argv[])
1110 {
1111         struct thread_data *td = NULL;
1112         int c, ini_idx = 0, lidx, ret;
1113
1114         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1115                 switch (c) {
1116                 case 't':
1117                         def_timeout = atoi(optarg);
1118                         break;
1119                 case 'l':
1120                         write_lat_log = 1;
1121                         break;
1122                 case 'w':
1123                         write_bw_log = 1;
1124                         break;
1125                 case 'o':
1126                         f_out = fopen(optarg, "w+");
1127                         if (!f_out) {
1128                                 perror("fopen output");
1129                                 exit(1);
1130                         }
1131                         f_err = f_out;
1132                         break;
1133                 case 'm':
1134                         terse_output = 1;
1135                         break;
1136                 case 'h':
1137                         usage();
1138                         exit(0);
1139                 case 'v':
1140                         printf("%s\n", fio_version_string);
1141                         exit(0);
1142                 case FIO_GETOPT_JOB: {
1143                         const char *opt = long_options[lidx].name;
1144                         char *val = optarg;
1145
1146                         if (!strncmp(opt, "name", 4) && td) {
1147                                 ret = add_job(td, td->name ?: "fio", 0);
1148                                 if (ret) {
1149                                         put_job(td);
1150                                         return 0;
1151                                 }
1152                                 td = NULL;
1153                         }
1154                         if (!td) {
1155                                 int global = !strncmp(val, "global", 6);
1156
1157                                 td = get_new_job(global, &def_thread);
1158                                 if (!td)
1159                                         return 0;
1160                         }
1161
1162                         ret = parse_cmd_option(opt, val, options, td);
1163                         if (ret) {
1164                                 log_err("fio: job dropped\n");
1165                                 put_job(td);
1166                                 td = NULL;
1167                         }
1168                         break;
1169                 }
1170                 default:
1171                         break;
1172                 }
1173         }
1174
1175         if (td) {
1176                 ret = add_job(td, td->name ?: "fio", 0);
1177                 if (ret)
1178                         put_job(td);
1179         }
1180
1181         while (optind < argc) {
1182                 ini_idx++;
1183                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1184                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1185                 optind++;
1186         }
1187
1188         return ini_idx;
1189 }
1190
1191 static void free_shm(void)
1192 {
1193         struct shmid_ds sbuf;
1194
1195         if (threads) {
1196                 shmdt((void *) threads);
1197                 threads = NULL;
1198                 shmctl(shm_id, IPC_RMID, &sbuf);
1199         }
1200 }
1201
1202 /*
1203  * The thread area is shared between the main process and the job
1204  * threads/processes. So setup a shared memory segment that will hold
1205  * all the job info.
1206  */
1207 static int setup_thread_area(void)
1208 {
1209         /*
1210          * 1024 is too much on some machines, scale max_jobs if
1211          * we get a failure that looks like too large a shm segment
1212          */
1213         do {
1214                 size_t size = max_jobs * sizeof(struct thread_data);
1215
1216                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1217                 if (shm_id != -1)
1218                         break;
1219                 if (errno != EINVAL) {
1220                         perror("shmget");
1221                         break;
1222                 }
1223
1224                 max_jobs >>= 1;
1225         } while (max_jobs);
1226
1227         if (shm_id == -1)
1228                 return 1;
1229
1230         threads = shmat(shm_id, NULL, 0);
1231         if (threads == (void *) -1) {
1232                 perror("shmat");
1233                 return 1;
1234         }
1235
1236         atexit(free_shm);
1237         return 0;
1238 }
1239
1240 /*
1241  * Copy the fio options into the long options map, so we mirror
1242  * job and cmd line options.
1243  */
1244 static void dupe_job_options(void)
1245 {
1246         struct fio_option *o;
1247         unsigned int i;
1248
1249         i = 0;
1250         while (long_options[i].name)
1251                 i++;
1252
1253         o = &options[0];
1254         while (o->name) {
1255                 long_options[i].name = o->name;
1256                 long_options[i].val = FIO_GETOPT_JOB;
1257                 if (o->type == FIO_OPT_STR_SET)
1258                         long_options[i].has_arg = no_argument;
1259                 else
1260                         long_options[i].has_arg = required_argument;
1261
1262                 i++;
1263                 o++;
1264                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1265         }
1266 }
1267
1268 int parse_options(int argc, char *argv[])
1269 {
1270         int job_files, i;
1271
1272         f_out = stdout;
1273         f_err = stderr;
1274
1275         dupe_job_options();
1276
1277         if (setup_thread_area())
1278                 return 1;
1279         if (fill_def_thread())
1280                 return 1;
1281
1282         job_files = parse_cmd_line(argc, argv);
1283
1284         for (i = 0; i < job_files; i++) {
1285                 if (fill_def_thread())
1286                         return 1;
1287                 if (parse_jobs_ini(ini_file[i], i))
1288                         return 1;
1289                 free(ini_file[i]);
1290         }
1291
1292         free(ini_file);
1293
1294         if (!thread_number) {
1295                 log_err("No jobs defined(s)\n");
1296                 return 1;
1297         }
1298
1299         return 0;
1300 }