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 <sys/ipc.h>
  12 #include <sys/shm.h>
  13 #include <sys/types.h>
  14 #include <sys/stat.h>
  15
  16 #include "fio.h"
  17 #include "parse.h"
  18 #include "smalloc.h"
  19 #include "filehash.h"
  20 #include "verify.h"
  21 #include "profile.h"
  22
  23 #include "lib/getopt.h"
  24
  25 static char fio_version_string[] = "fio 1.58";
  26
  27 #define FIO_RANDSEED            (0xb1899bedUL)
  28
  29 static char **ini_file;
  30 static int max_jobs = FIO_MAX_JOBS;
  31 static int dump_cmdline;
  32
  33 static struct thread_data def_thread;
  34 struct thread_data *threads = NULL;
  35
  36 int exitall_on_terminate = 0;
  37 int terse_output = 0;
  38 int eta_print;
  39 unsigned long long mlock_size = 0;
  40 FILE *f_out = NULL;
  41 FILE *f_err = NULL;
  42 char **job_sections = NULL;
  43 int nr_job_sections = 0;
  44 char *exec_profile = NULL;
  45 int warnings_fatal = 0;
  46 int terse_version = 2;
  47
  48 int write_bw_log = 0;
  49 int read_only = 0;
  50
  51 static int write_lat_log;
  52
  53 static int prev_group_jobs;
  54
  55 unsigned long fio_debug = 0;
  56 unsigned int fio_debug_jobno = -1;
  57 unsigned int *fio_debug_jobp = NULL;
  58
  59 static char cmd_optstr[256];
  60
  61 /*
  62  * Command line options. These will contain the above, plus a few
  63  * extra that only pertain to fio itself and not jobs.
  64  */
  65 static struct option l_opts[FIO_NR_OPTIONS] = {
  66         {
  67                 .name           = (char *) "output",
  68                 .has_arg        = required_argument,
  69                 .val            = 'o',
  70         },
  71         {
  72                 .name           = (char *) "timeout",
  73                 .has_arg        = required_argument,
  74                 .val            = 't',
  75         },
  76         {
  77                 .name           = (char *) "latency-log",
  78                 .has_arg        = required_argument,
  79                 .val            = 'l',
  80         },
  81         {
  82                 .name           = (char *) "bandwidth-log",
  83                 .has_arg        = required_argument,
  84                 .val            = 'b',
  85         },
  86         {
  87                 .name           = (char *) "minimal",
  88                 .has_arg        = optional_argument,
  89                 .val            = 'm',
  90         },
  91         {
  92                 .name           = (char *) "version",
  93                 .has_arg        = no_argument,
  94                 .val            = 'v',
  95         },
  96         {
  97                 .name           = (char *) "help",
  98                 .has_arg        = no_argument,
  99                 .val            = 'h',
 100         },
 101         {
 102                 .name           = (char *) "cmdhelp",
 103                 .has_arg        = optional_argument,
 104                 .val            = 'c',
 105         },
 106         {
 107                 .name           = (char *) "showcmd",
 108                 .has_arg        = no_argument,
 109                 .val            = 's',
 110         },
 111         {
 112                 .name           = (char *) "readonly",
 113                 .has_arg        = no_argument,
 114                 .val            = 'r',
 115         },
 116         {
 117                 .name           = (char *) "eta",
 118                 .has_arg        = required_argument,
 119                 .val            = 'e',
 120         },
 121         {
 122                 .name           = (char *) "debug",
 123                 .has_arg        = required_argument,
 124                 .val            = 'd',
 125         },
 126         {
 127                 .name           = (char *) "section",
 128                 .has_arg        = required_argument,
 129                 .val            = 'x',
 130         },
 131         {
 132                 .name           = (char *) "alloc-size",
 133                 .has_arg        = required_argument,
 134                 .val            = 'a',
 135         },
 136         {
 137                 .name           = (char *) "profile",
 138                 .has_arg        = required_argument,
 139                 .val            = 'p',
 140         },
 141         {
 142                 .name           = (char *) "warnings-fatal",
 143                 .has_arg        = no_argument,
 144                 .val            = 'w',
 145         },
 146         {
 147                 .name           = (char *) "max-jobs",
 148                 .has_arg        = required_argument,
 149                 .val            = 'j',
 150         },
 151         {
 152                 .name           = (char *) "terse-version",
 153                 .has_arg        = required_argument,
 154                 .val            = 'V',
 155         },
 156         {
 157                 .name           = NULL,
 158         },
 159 };
 160
 161 FILE *get_f_out()
 162 {
 163         return f_out;
 164 }
 165
 166 FILE *get_f_err()
 167 {
 168         return f_err;
 169 }
 170
 171 /*
 172  * Return a free job structure.
 173  */
 174 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 175 {
 176         struct thread_data *td;
 177
 178         if (global)
 179                 return &def_thread;
 180         if (thread_number >= max_jobs) {
 181                 log_err("error: maximum number of jobs (%d) reached.\n",
 182                                 max_jobs);
 183                 return NULL;
 184         }
 185
 186         td = &threads[thread_number++];
 187         *td = *parent;
 188
 189         td->o.uid = td->o.gid = -1U;
 190
 191         dup_files(td, parent);
 192         options_mem_dupe(td);
 193
 194         profile_add_hooks(td);
 195
 196         td->thread_number = thread_number;
 197         return td;
 198 }
 199
 200 static void put_job(struct thread_data *td)
 201 {
 202         if (td == &def_thread)
 203                 return;
 204
 205         profile_td_exit(td);
 206
 207         if (td->error)
 208                 log_info("fio: %s\n", td->verror);
 209
 210         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 211         thread_number--;
 212 }
 213
 214 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
 215 {
 216         unsigned int bs = td->o.min_bs[ddir];
 217         unsigned long long bytes_per_sec;
 218
 219         assert(ddir_rw(ddir));
 220
 221         if (td->o.rate[ddir])
 222                 bytes_per_sec = td->o.rate[ddir];
 223         else
 224                 bytes_per_sec = td->o.rate_iops[ddir] * bs;
 225
 226         if (!bytes_per_sec) {
 227                 log_err("rate lower than supported\n");
 228                 return -1;
 229         }
 230
 231         td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
 232         td->rate_pending_usleep[ddir] = 0;
 233         return 0;
 234 }
 235
 236 static int setup_rate(struct thread_data *td)
 237 {
 238         int ret = 0;
 239
 240         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
 241                 ret = __setup_rate(td, DDIR_READ);
 242         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
 243                 ret |= __setup_rate(td, DDIR_WRITE);
 244
 245         return ret;
 246 }
 247
 248 static int fixed_block_size(struct thread_options *o)
 249 {
 250         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
 251                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
 252                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
 253 }
 254
 255 /*
 256  * Lazy way of fixing up options that depend on each other. We could also
 257  * define option callback handlers, but this is easier.
 258  */
 259 static int fixup_options(struct thread_data *td)
 260 {
 261         struct thread_options *o = &td->o;
 262         int ret = 0;
 263
 264 #ifndef FIO_HAVE_PSHARED_MUTEX
 265         if (!o->use_thread) {
 266                 log_info("fio: this platform does not support process shared"
 267                          " mutexes, forcing use of threads. Use the 'thread'"
 268                          " option to get rid of this warning.\n");
 269                 o->use_thread = 1;
 270                 ret = warnings_fatal;
 271         }
 272 #endif
 273
 274         if (o->write_iolog_file && o->read_iolog_file) {
 275                 log_err("fio: read iolog overrides write_iolog\n");
 276                 free(o->write_iolog_file);
 277                 o->write_iolog_file = NULL;
 278                 ret = warnings_fatal;
 279         }
 280
 281         /*
 282          * only really works for sequential io for now, and with 1 file
 283          */
 284         if (o->zone_size && td_random(td) && o->open_files == 1)
 285                 o->zone_size = 0;
 286
 287         /*
 288          * Reads can do overwrites, we always need to pre-create the file
 289          */
 290         if (td_read(td) || td_rw(td))
 291                 o->overwrite = 1;
 292
 293         if (!o->min_bs[DDIR_READ])
 294                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
 295         if (!o->max_bs[DDIR_READ])
 296                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
 297         if (!o->min_bs[DDIR_WRITE])
 298                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 299         if (!o->max_bs[DDIR_WRITE])
 300                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 301
 302         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
 303
 304         /*
 305          * For random IO, allow blockalign offset other than min_bs.
 306          */
 307         if (!o->ba[DDIR_READ] || !td_random(td))
 308                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
 309         if (!o->ba[DDIR_WRITE] || !td_random(td))
 310                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
 311
 312         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
 313             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
 314             !o->norandommap) {
 315                 log_err("fio: Any use of blockalign= turns off randommap\n");
 316                 o->norandommap = 1;
 317                 ret = warnings_fatal;
 318         }
 319
 320         if (!o->file_size_high)
 321                 o->file_size_high = o->file_size_low;
 322
 323         if (o->norandommap && o->verify != VERIFY_NONE
 324             && !fixed_block_size(o))  {
 325                 log_err("fio: norandommap given for variable block sizes, "
 326                         "verify disabled\n");
 327                 o->verify = VERIFY_NONE;
 328                 ret = warnings_fatal;
 329         }
 330         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
 331                 log_err("fio: bs_unaligned may not work with raw io\n");
 332
 333         /*
 334          * thinktime_spin must be less than thinktime
 335          */
 336         if (o->thinktime_spin > o->thinktime)
 337                 o->thinktime_spin = o->thinktime;
 338
 339         /*
 340          * The low water mark cannot be bigger than the iodepth
 341          */
 342         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
 343                 /*
 344                  * syslet work around - if the workload is sequential,
 345                  * we want to let the queue drain all the way down to
 346                  * avoid seeking between async threads
 347                  */
 348                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 349                         o->iodepth_low = 1;
 350                 else
 351                         o->iodepth_low = o->iodepth;
 352         }
 353
 354         /*
 355          * If batch number isn't set, default to the same as iodepth
 356          */
 357         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
 358                 o->iodepth_batch = o->iodepth;
 359
 360         if (o->nr_files > td->files_index)
 361                 o->nr_files = td->files_index;
 362
 363         if (o->open_files > o->nr_files || !o->open_files)
 364                 o->open_files = o->nr_files;
 365
 366         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
 367             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
 368                 o->rate_iops_min[1]))) {
 369                 log_err("fio: rate and rate_iops are mutually exclusive\n");
 370                 ret = 1;
 371         }
 372         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
 373             (o->rate_iops[0] < o->rate_iops_min[0]) ||
 374             (o->rate_iops[1] < o->rate_iops_min[1])) {
 375                 log_err("fio: minimum rate exceeds rate\n");
 376                 ret = 1;
 377         }
 378
 379         if (!o->timeout && o->time_based) {
 380                 log_err("fio: time_based requires a runtime/timeout setting\n");
 381                 o->time_based = 0;
 382                 ret = warnings_fatal;
 383         }
 384
 385         if (o->fill_device && !o->size)
 386                 o->size = -1ULL;
 387
 388         if (o->verify != VERIFY_NONE) {
 389                 if (td_write(td) && o->do_verify && o->numjobs > 1) {
 390                         log_info("Multiple writers may overwrite blocks that "
 391                                 "belong to other jobs. This can cause "
 392                                 "verification failures.\n");
 393                         ret = warnings_fatal;
 394                 }
 395
 396                 o->refill_buffers = 1;
 397                 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
 398                     !o->verify_interval)
 399                         o->verify_interval = o->min_bs[DDIR_WRITE];
 400         }
 401
 402         if (o->pre_read) {
 403                 o->invalidate_cache = 0;
 404                 if (td->io_ops->flags & FIO_PIPEIO) {
 405                         log_info("fio: cannot pre-read files with an IO engine"
 406                                  " that isn't seekable. Pre-read disabled.\n");
 407                         ret = warnings_fatal;
 408                 }
 409         }
 410
 411 #ifndef FIO_HAVE_FDATASYNC
 412         if (o->fdatasync_blocks) {
 413                 log_info("fio: this platform does not support fdatasync()"
 414                          " falling back to using fsync().  Use the 'fsync'"
 415                          " option instead of 'fdatasync' to get rid of"
 416                          " this warning\n");
 417                 o->fsync_blocks = o->fdatasync_blocks;
 418                 o->fdatasync_blocks = 0;
 419                 ret = warnings_fatal;
 420         }
 421 #endif
 422
 423         return ret;
 424 }
 425
 426 /*
 427  * This function leaks the buffer
 428  */
 429 static char *to_kmg(unsigned int val)
 430 {
 431         char *buf = malloc(32);
 432         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 433         char *p = post;
 434
 435         do {
 436                 if (val & 1023)
 437                         break;
 438
 439                 val >>= 10;
 440                 p++;
 441         } while (*p);
 442
 443         snprintf(buf, 31, "%u%c", val, *p);
 444         return buf;
 445 }
 446
 447 /* External engines are specified by "external:name.o") */
 448 static const char *get_engine_name(const char *str)
 449 {
 450         char *p = strstr(str, ":");
 451
 452         if (!p)
 453                 return str;
 454
 455         p++;
 456         strip_blank_front(&p);
 457         strip_blank_end(p);
 458         return p;
 459 }
 460
 461 static int exists_and_not_file(const char *filename)
 462 {
 463         struct stat sb;
 464
 465         if (lstat(filename, &sb) == -1)
 466                 return 0;
 467
 468         /* \\.\ is the device namespace in Windows, where every file
 469          * is a device node */
 470         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
 471                 return 0;
 472
 473         return 1;
 474 }
 475
 476 static void td_fill_rand_seeds_os(struct thread_data *td)
 477 {
 478         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
 479         os_random_seed(td->rand_seeds[1], &td->verify_state);
 480         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
 481
 482         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
 483                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
 484
 485         os_random_seed(td->rand_seeds[5], &td->file_size_state);
 486         os_random_seed(td->rand_seeds[6], &td->trim_state);
 487
 488         if (!td_random(td))
 489                 return;
 490
 491         if (td->o.rand_repeatable)
 492                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
 493
 494         os_random_seed(td->rand_seeds[4], &td->random_state);
 495 }
 496
 497 static void td_fill_rand_seeds_internal(struct thread_data *td)
 498 {
 499         init_rand_seed(&td->__bsrange_state, td->rand_seeds[0]);
 500         init_rand_seed(&td->__verify_state, td->rand_seeds[1]);
 501         init_rand_seed(&td->__rwmix_state, td->rand_seeds[2]);
 502
 503         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
 504                 init_rand_seed(&td->__next_file_state, td->rand_seeds[3]);
 505
 506         init_rand_seed(&td->__file_size_state, td->rand_seeds[5]);
 507         init_rand_seed(&td->__trim_state, td->rand_seeds[6]);
 508
 509         if (!td_random(td))
 510                 return;
 511
 512         if (td->o.rand_repeatable)
 513                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
 514
 515         init_rand_seed(&td->__random_state, td->rand_seeds[4]);
 516 }
 517
 518 void td_fill_rand_seeds(struct thread_data *td)
 519 {
 520         if (td->o.use_os_rand)
 521                 td_fill_rand_seeds_os(td);
 522         else
 523                 td_fill_rand_seeds_internal(td);
 524
 525         init_rand_seed(&td->buf_state, td->rand_seeds[7]);
 526 }
 527
 528 /*
 529  * Initialize the various random states we need (random io, block size ranges,
 530  * read/write mix, etc).
 531  */
 532 static int init_random_state(struct thread_data *td)
 533 {
 534         int fd;
 535
 536         fd = open("/dev/urandom", O_RDONLY);
 537         if (fd == -1) {
 538                 td_verror(td, errno, "open");
 539                 return 1;
 540         }
 541
 542         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
 543             (int) sizeof(td->rand_seeds)) {
 544                 td_verror(td, EIO, "read");
 545                 close(fd);
 546                 return 1;
 547         }
 548
 549         close(fd);
 550         td_fill_rand_seeds(td);
 551         return 0;
 552 }
 553
 554 /*
 555  * Adds a job to the list of things todo. Sanitizes the various options
 556  * to make sure we don't have conflicts, and initializes various
 557  * members of td.
 558  */
 559 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 560 {
 561         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 562                                    "randread", "randwrite", "randrw" };
 563         unsigned int i;
 564         const char *engine;
 565         char fname[PATH_MAX];
 566         int numjobs, file_alloced;
 567
 568         /*
 569          * the def_thread is just for options, it's not a real job
 570          */
 571         if (td == &def_thread)
 572                 return 0;
 573
 574         /*
 575          * if we are just dumping the output command line, don't add the job
 576          */
 577         if (dump_cmdline) {
 578                 put_job(td);
 579                 return 0;
 580         }
 581
 582         if (profile_td_init(td))
 583                 goto err;
 584
 585         engine = get_engine_name(td->o.ioengine);
 586         td->io_ops = load_ioengine(td, engine);
 587         if (!td->io_ops) {
 588                 log_err("fio: failed to load engine %s\n", engine);
 589                 goto err;
 590         }
 591
 592         if (td->o.use_thread)
 593                 nr_thread++;
 594         else
 595                 nr_process++;
 596
 597         if (td->o.odirect)
 598                 td->io_ops->flags |= FIO_RAWIO;
 599
 600         file_alloced = 0;
 601         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
 602                 file_alloced = 1;
 603
 604                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
 605                         add_file(td, jobname);
 606                 else {
 607                         for (i = 0; i < td->o.nr_files; i++) {
 608                                 sprintf(fname, "%s.%d.%d", jobname,
 609                                                         td->thread_number, i);
 610                                 add_file(td, fname);
 611                         }
 612                 }
 613         }
 614
 615         if (fixup_options(td))
 616                 goto err;
 617
 618         if (td->io_ops->flags & FIO_DISKLESSIO) {
 619                 struct fio_file *f;
 620
 621                 for_each_file(td, f, i)
 622                         f->real_file_size = -1ULL;
 623         }
 624
 625         td->mutex = fio_mutex_init(0);
 626
 627         td->ts.clat_percentiles = td->o.clat_percentiles;
 628         if (td->o.overwrite_plist)
 629                 td->ts.percentile_list = td->o.percentile_list;
 630         else
 631                 td->ts.percentile_list = NULL;
 632
 633         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 634         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 635         td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
 636         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 637         td->ddir_seq_nr = td->o.ddir_seq_nr;
 638
 639         if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
 640                 prev_group_jobs = 0;
 641                 groupid++;
 642         }
 643
 644         td->groupid = groupid;
 645         prev_group_jobs++;
 646
 647         if (init_random_state(td))
 648                 goto err;
 649
 650         if (setup_rate(td))
 651                 goto err;
 652
 653         if (td->o.write_lat_log) {
 654                 setup_log(&td->ts.lat_log);
 655                 setup_log(&td->ts.slat_log);
 656                 setup_log(&td->ts.clat_log);
 657         }
 658         if (td->o.write_bw_log)
 659                 setup_log(&td->ts.bw_log);
 660
 661         if (!td->o.name)
 662                 td->o.name = strdup(jobname);
 663
 664         if (!terse_output) {
 665                 if (!job_add_num) {
 666                         if (!strcmp(td->io_ops->name, "cpuio")) {
 667                                 log_info("%s: ioengine=cpu, cpuload=%u,"
 668                                          " cpucycle=%u\n", td->o.name,
 669                                                         td->o.cpuload,
 670                                                         td->o.cpucycle);
 671                         } else {
 672                                 char *c1, *c2, *c3, *c4;
 673
 674                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
 675                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
 676                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
 677                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
 678
 679                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
 680                                          " ioengine=%s, iodepth=%u\n",
 681                                                 td->o.name, td->groupid,
 682                                                 ddir_str[td->o.td_ddir],
 683                                                 c1, c2, c3, c4,
 684                                                 td->io_ops->name,
 685                                                 td->o.iodepth);
 686
 687                                 free(c1);
 688                                 free(c2);
 689                                 free(c3);
 690                                 free(c4);
 691                         }
 692                 } else if (job_add_num == 1)
 693                         log_info("...\n");
 694         }
 695
 696         /*
 697          * recurse add identical jobs, clear numjobs and stonewall options
 698          * as they don't apply to sub-jobs
 699          */
 700         numjobs = td->o.numjobs;
 701         while (--numjobs) {
 702                 struct thread_data *td_new = get_new_job(0, td);
 703
 704                 if (!td_new)
 705                         goto err;
 706
 707                 td_new->o.numjobs = 1;
 708                 td_new->o.stonewall = 0;
 709                 td_new->o.new_group = 0;
 710
 711                 if (file_alloced) {
 712                         td_new->o.filename = NULL;
 713                         td_new->files_index = 0;
 714                         td_new->files_size = 0;
 715                         td_new->files = NULL;
 716                 }
 717
 718                 job_add_num = numjobs - 1;
 719
 720                 if (add_job(td_new, jobname, job_add_num))
 721                         goto err;
 722         }
 723
 724         return 0;
 725 err:
 726         put_job(td);
 727         return -1;
 728 }
 729
 730 /*
 731  * Parse as if 'o' was a command line
 732  */
 733 void add_job_opts(const char **o)
 734 {
 735         struct thread_data *td, *td_parent;
 736         int i, in_global = 1;
 737         char jobname[32];
 738
 739         i = 0;
 740         td_parent = td = NULL;
 741         while (o[i]) {
 742                 if (!strncmp(o[i], "name", 4)) {
 743                         in_global = 0;
 744                         if (td)
 745                                 add_job(td, jobname, 0);
 746                         td = NULL;
 747                         sprintf(jobname, "%s", o[i] + 5);
 748                 }
 749                 if (in_global && !td_parent)
 750                         td_parent = get_new_job(1, &def_thread);
 751                 else if (!in_global && !td) {
 752                         if (!td_parent)
 753                                 td_parent = &def_thread;
 754                         td = get_new_job(0, td_parent);
 755                 }
 756                 if (in_global)
 757                         fio_options_parse(td_parent, (char **) &o[i], 1);
 758                 else
 759                         fio_options_parse(td, (char **) &o[i], 1);
 760                 i++;
 761         }
 762
 763         if (td)
 764                 add_job(td, jobname, 0);
 765 }
 766
 767 static int skip_this_section(const char *name)
 768 {
 769         int i;
 770
 771         if (!nr_job_sections)
 772                 return 0;
 773         if (!strncmp(name, "global", 6))
 774                 return 0;
 775
 776         for (i = 0; i < nr_job_sections; i++)
 777                 if (!strcmp(job_sections[i], name))
 778                         return 0;
 779
 780         return 1;
 781 }
 782
 783 static int is_empty_or_comment(char *line)
 784 {
 785         unsigned int i;
 786
 787         for (i = 0; i < strlen(line); i++) {
 788                 if (line[i] == ';')
 789                         return 1;
 790                 if (line[i] == '#')
 791                         return 1;
 792                 if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
 793                         return 0;
 794         }
 795
 796         return 1;
 797 }
 798
 799 /*
 800  * This is our [ini] type file parser.
 801  */
 802 static int parse_jobs_ini(char *file, int stonewall_flag)
 803 {
 804         unsigned int global;
 805         struct thread_data *td;
 806         char *string, *name;
 807         FILE *f;
 808         char *p;
 809         int ret = 0, stonewall;
 810         int first_sect = 1;
 811         int skip_fgets = 0;
 812         int inside_skip = 0;
 813         char **opts;
 814         int i, alloc_opts, num_opts;
 815
 816         if (!strcmp(file, "-"))
 817                 f = stdin;
 818         else
 819                 f = fopen(file, "r");
 820
 821         if (!f) {
 822                 perror("fopen job file");
 823                 return 1;
 824         }
 825
 826         string = malloc(4096);
 827
 828         /*
 829          * it's really 256 + small bit, 280 should suffice
 830          */
 831         name = malloc(280);
 832         memset(name, 0, 280);
 833
 834         alloc_opts = 8;
 835         opts = malloc(sizeof(char *) * alloc_opts);
 836         num_opts = 0;
 837
 838         stonewall = stonewall_flag;
 839         do {
 840                 /*
 841                  * if skip_fgets is set, we already have loaded a line we
 842                  * haven't handled.
 843                  */
 844                 if (!skip_fgets) {
 845                         p = fgets(string, 4095, f);
 846                         if (!p)
 847                                 break;
 848                 }
 849
 850                 skip_fgets = 0;
 851                 strip_blank_front(&p);
 852                 strip_blank_end(p);
 853
 854                 if (is_empty_or_comment(p))
 855                         continue;
 856                 if (sscanf(p, "[%255[^\n]]", name) != 1) {
 857                         if (inside_skip)
 858                                 continue;
 859                         log_err("fio: option <%s> outside of [] job section\n",
 860                                                                         p);
 861                         break;
 862                 }
 863
 864                 name[strlen(name) - 1] = '\0';
 865
 866                 if (skip_this_section(name)) {
 867                         inside_skip = 1;
 868                         continue;
 869                 } else
 870                         inside_skip = 0;
 871
 872                 global = !strncmp(name, "global", 6);
 873
 874                 if (dump_cmdline) {
 875                         if (first_sect)
 876                                 log_info("fio ");
 877                         if (!global)
 878                                 log_info("--name=%s ", name);
 879                         first_sect = 0;
 880                 }
 881
 882                 td = get_new_job(global, &def_thread);
 883                 if (!td) {
 884                         ret = 1;
 885                         break;
 886                 }
 887
 888                 /*
 889                  * Seperate multiple job files by a stonewall
 890                  */
 891                 if (!global && stonewall) {
 892                         td->o.stonewall = stonewall;
 893                         stonewall = 0;
 894                 }
 895
 896                 num_opts = 0;
 897                 memset(opts, 0, alloc_opts * sizeof(char *));
 898
 899                 while ((p = fgets(string, 4096, f)) != NULL) {
 900                         if (is_empty_or_comment(p))
 901                                 continue;
 902
 903                         strip_blank_front(&p);
 904
 905                         /*
 906                          * new section, break out and make sure we don't
 907                          * fgets() a new line at the top.
 908                          */
 909                         if (p[0] == '[') {
 910                                 skip_fgets = 1;
 911                                 break;
 912                         }
 913
 914                         strip_blank_end(p);
 915
 916                         if (num_opts == alloc_opts) {
 917                                 alloc_opts <<= 1;
 918                                 opts = realloc(opts,
 919                                                 alloc_opts * sizeof(char *));
 920                         }
 921
 922                         opts[num_opts] = strdup(p);
 923                         num_opts++;
 924                 }
 925
 926                 ret = fio_options_parse(td, opts, num_opts);
 927                 if (!ret) {
 928                         if (dump_cmdline)
 929                                 for (i = 0; i < num_opts; i++)
 930                                         log_info("--%s ", opts[i]);
 931
 932                         ret = add_job(td, name, 0);
 933                 } else {
 934                         log_err("fio: job %s dropped\n", name);
 935                         put_job(td);
 936                 }
 937
 938                 for (i = 0; i < num_opts; i++)
 939                         free(opts[i]);
 940                 num_opts = 0;
 941         } while (!ret);
 942
 943         if (dump_cmdline)
 944                 log_info("\n");
 945
 946         for (i = 0; i < num_opts; i++)
 947                 free(opts[i]);
 948
 949         free(string);
 950         free(name);
 951         free(opts);
 952         if (f != stdin)
 953                 fclose(f);
 954         return ret;
 955 }
 956
 957 static int fill_def_thread(void)
 958 {
 959         memset(&def_thread, 0, sizeof(def_thread));
 960
 961         fio_getaffinity(getpid(), &def_thread.o.cpumask);
 962
 963         /*
 964          * fill default options
 965          */
 966         fio_fill_default_options(&def_thread);
 967         return 0;
 968 }
 969
 970 static void free_shm(void)
 971 {
 972         struct shmid_ds sbuf;
 973
 974         if (threads) {
 975                 void *tp = threads;
 976
 977                 threads = NULL;
 978                 file_hash_exit();
 979                 fio_debug_jobp = NULL;
 980                 shmdt(tp);
 981                 shmctl(shm_id, IPC_RMID, &sbuf);
 982         }
 983
 984         scleanup();
 985 }
 986
 987 /*
 988  * The thread area is shared between the main process and the job
 989  * threads/processes. So setup a shared memory segment that will hold
 990  * all the job info. We use the end of the region for keeping track of
 991  * open files across jobs, for file sharing.
 992  */
 993 static int setup_thread_area(void)
 994 {
 995         void *hash;
 996
 997         /*
 998          * 1024 is too much on some machines, scale max_jobs if
 999          * we get a failure that looks like too large a shm segment
1000          */
1001         do {
1002                 size_t size = max_jobs * sizeof(struct thread_data);
1003
1004                 size += file_hash_size;
1005                 size += sizeof(unsigned int);
1006
1007                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1008                 if (shm_id != -1)
1009                         break;
1010                 if (errno != EINVAL) {
1011                         perror("shmget");
1012                         break;
1013                 }
1014
1015                 max_jobs >>= 1;
1016         } while (max_jobs);
1017
1018         if (shm_id == -1)
1019                 return 1;
1020
1021         threads = shmat(shm_id, NULL, 0);
1022         if (threads == (void *) -1) {
1023                 perror("shmat");
1024                 return 1;
1025         }
1026
1027         memset(threads, 0, max_jobs * sizeof(struct thread_data));
1028         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
1029         fio_debug_jobp = (void *) hash + file_hash_size;
1030         *fio_debug_jobp = -1;
1031         file_hash_init(hash);
1032         atexit(free_shm);
1033         return 0;
1034 }
1035
1036 static void usage(const char *name)
1037 {
1038         printf("%s\n", fio_version_string);
1039         printf("%s [options] [job options] <job file(s)>\n", name);
1040         printf("\t--debug=options\tEnable debug logging\n");
1041         printf("\t--output\tWrite output to file\n");
1042         printf("\t--timeout\tRuntime in seconds\n");
1043         printf("\t--latency-log\tGenerate per-job latency logs\n");
1044         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1045         printf("\t--minimal\tMinimal (terse) output\n");
1046         printf("\t--version\tPrint version info and exit\n");
1047         printf("\t--terse-version=x Terse version output format\n");
1048         printf("\t--help\t\tPrint this page\n");
1049         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
1050                 " them\n");
1051         printf("\t--showcmd\tTurn a job file into command line options\n");
1052         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
1053         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
1054         printf("\t--readonly\tTurn on safety read-only checks, preventing"
1055                 " writes\n");
1056         printf("\t--section=name\tOnly run specified section in job file\n");
1057         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
1058                 " (def 1024)\n");
1059         printf("\t--warnings-fatal Fio parser warnings are fatal\n");
1060         printf("\t--max-jobs\tMaximum number of threads/processes to support\n");
1061         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1062         printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
1063 }
1064
1065 #ifdef FIO_INC_DEBUG
1066 struct debug_level debug_levels[] = {
1067         { .name = "process",    .shift = FD_PROCESS, },
1068         { .name = "file",       .shift = FD_FILE, },
1069         { .name = "io",         .shift = FD_IO, },
1070         { .name = "mem",        .shift = FD_MEM, },
1071         { .name = "blktrace",   .shift = FD_BLKTRACE },
1072         { .name = "verify",     .shift = FD_VERIFY },
1073         { .name = "random",     .shift = FD_RANDOM },
1074         { .name = "parse",      .shift = FD_PARSE },
1075         { .name = "diskutil",   .shift = FD_DISKUTIL },
1076         { .name = "job",        .shift = FD_JOB },
1077         { .name = "mutex",      .shift = FD_MUTEX },
1078         { .name = "profile",    .shift = FD_PROFILE },
1079         { .name = "time",       .shift = FD_TIME },
1080         { .name = NULL, },
1081 };
1082
1083 static int set_debug(const char *string)
1084 {
1085         struct debug_level *dl;
1086         char *p = (char *) string;
1087         char *opt;
1088         int i;
1089
1090         if (!strcmp(string, "?") || !strcmp(string, "help")) {
1091                 log_info("fio: dumping debug options:");
1092                 for (i = 0; debug_levels[i].name; i++) {
1093                         dl = &debug_levels[i];
1094                         log_info("%s,", dl->name);
1095                 }
1096                 log_info("all\n");
1097                 return 1;
1098         }
1099
1100         while ((opt = strsep(&p, ",")) != NULL) {
1101                 int found = 0;
1102
1103                 if (!strncmp(opt, "all", 3)) {
1104                         log_info("fio: set all debug options\n");
1105                         fio_debug = ~0UL;
1106                         continue;
1107                 }
1108
1109                 for (i = 0; debug_levels[i].name; i++) {
1110                         dl = &debug_levels[i];
1111                         found = !strncmp(opt, dl->name, strlen(dl->name));
1112                         if (!found)
1113                                 continue;
1114
1115                         if (dl->shift == FD_JOB) {
1116                                 opt = strchr(opt, ':');
1117                                 if (!opt) {
1118                                         log_err("fio: missing job number\n");
1119                                         break;
1120                                 }
1121                                 opt++;
1122                                 fio_debug_jobno = atoi(opt);
1123                                 log_info("fio: set debug jobno %d\n",
1124                                                         fio_debug_jobno);
1125                         } else {
1126                                 log_info("fio: set debug option %s\n", opt);
1127                                 fio_debug |= (1UL << dl->shift);
1128                         }
1129                         break;
1130                 }
1131
1132                 if (!found)
1133                         log_err("fio: debug mask %s not found\n", opt);
1134         }
1135         return 0;
1136 }
1137 #else
1138 static int set_debug(const char *string)
1139 {
1140         log_err("fio: debug tracing not included in build\n");
1141         return 1;
1142 }
1143 #endif
1144
1145 static void fio_options_fill_optstring(void)
1146 {
1147         char *ostr = cmd_optstr;
1148         int i, c;
1149
1150         c = i = 0;
1151         while (l_opts[i].name) {
1152                 ostr[c++] = l_opts[i].val;
1153                 if (l_opts[i].has_arg == required_argument)
1154                         ostr[c++] = ':';
1155                 else if (l_opts[i].has_arg == optional_argument) {
1156                         ostr[c++] = ':';
1157                         ostr[c++] = ':';
1158                 }
1159                 i++;
1160         }
1161         ostr[c] = '\0';
1162 }
1163
1164 static int parse_cmd_line(int argc, char *argv[])
1165 {
1166         struct thread_data *td = NULL;
1167         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1168         char *ostr = cmd_optstr;
1169
1170         while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1171                 switch (c) {
1172                 case 'a':
1173                         smalloc_pool_size = atoi(optarg);
1174                         break;
1175                 case 't':
1176                         def_thread.o.timeout = atoi(optarg);
1177                         break;
1178                 case 'l':
1179                         write_lat_log = 1;
1180                         break;
1181                 case 'b':
1182                         write_bw_log = 1;
1183                         break;
1184                 case 'o':
1185                         f_out = fopen(optarg, "w+");
1186                         if (!f_out) {
1187                                 perror("fopen output");
1188                                 exit(1);
1189                         }
1190                         f_err = f_out;
1191                         break;
1192                 case 'm':
1193                         terse_output = 1;
1194                         break;
1195                 case 'h':
1196                         usage(argv[0]);
1197                         exit(0);
1198                 case 'c':
1199                         exit(fio_show_option_help(optarg));
1200                 case 's':
1201                         dump_cmdline = 1;
1202                         break;
1203                 case 'r':
1204                         read_only = 1;
1205                         break;
1206                 case 'v':
1207                         log_info("%s\n", fio_version_string);
1208                         exit(0);
1209                 case 'V':
1210                         terse_version = atoi(optarg);
1211                         if (terse_version != 2) {
1212                                 log_err("fio: bad terse version format\n");
1213                                 exit_val = 1;
1214                                 do_exit++;
1215                         }
1216                         break;
1217                 case 'e':
1218                         if (!strcmp("always", optarg))
1219                                 eta_print = FIO_ETA_ALWAYS;
1220                         else if (!strcmp("never", optarg))
1221                                 eta_print = FIO_ETA_NEVER;
1222                         break;
1223                 case 'd':
1224                         if (set_debug(optarg))
1225                                 do_exit++;
1226                         break;
1227                 case 'x': {
1228                         size_t new_size;
1229
1230                         if (!strcmp(optarg, "global")) {
1231                                 log_err("fio: can't use global as only "
1232                                         "section\n");
1233                                 do_exit++;
1234                                 exit_val = 1;
1235                                 break;
1236                         }
1237                         new_size = (nr_job_sections + 1) * sizeof(char *);
1238                         job_sections = realloc(job_sections, new_size);
1239                         job_sections[nr_job_sections] = strdup(optarg);
1240                         nr_job_sections++;
1241                         break;
1242                         }
1243                 case 'p':
1244                         exec_profile = strdup(optarg);
1245                         break;
1246                 case FIO_GETOPT_JOB: {
1247                         const char *opt = l_opts[lidx].name;
1248                         char *val = optarg;
1249
1250                         if (!strncmp(opt, "name", 4) && td) {
1251                                 ret = add_job(td, td->o.name ?: "fio", 0);
1252                                 if (ret)
1253                                         return 0;
1254                                 td = NULL;
1255                         }
1256                         if (!td) {
1257                                 int is_section = !strncmp(opt, "name", 4);
1258                                 int global = 0;
1259
1260                                 if (!is_section || !strncmp(val, "global", 6))
1261                                         global = 1;
1262
1263                                 if (is_section && skip_this_section(val))
1264                                         continue;
1265
1266                                 td = get_new_job(global, &def_thread);
1267                                 if (!td)
1268                                         return 0;
1269                         }
1270
1271                         ret = fio_cmd_option_parse(td, opt, val);
1272                         break;
1273                 }
1274                 case 'w':
1275                         warnings_fatal = 1;
1276                         break;
1277                 case 'j':
1278                         max_jobs = atoi(optarg);
1279                         if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1280                                 log_err("fio: invalid max jobs: %d\n", max_jobs);
1281                                 do_exit++;
1282                                 exit_val = 1;
1283                         }
1284                         break;
1285                 default:
1286                         do_exit++;
1287                         exit_val = 1;
1288                         break;
1289                 }
1290         }
1291
1292         if (do_exit)
1293                 exit(exit_val);
1294
1295         if (td) {
1296                 if (!ret)
1297                         ret = add_job(td, td->o.name ?: "fio", 0);
1298         }
1299
1300         while (optind < argc) {
1301                 ini_idx++;
1302                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1303                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1304                 optind++;
1305         }
1306
1307         return ini_idx;
1308 }
1309
1310 int parse_options(int argc, char *argv[])
1311 {
1312         int job_files, i;
1313
1314         f_out = stdout;
1315         f_err = stderr;
1316
1317         fio_options_fill_optstring();
1318         fio_options_dup_and_init(l_opts);
1319
1320         if (setup_thread_area())
1321                 return 1;
1322         if (fill_def_thread())
1323                 return 1;
1324
1325         job_files = parse_cmd_line(argc, argv);
1326
1327         for (i = 0; i < job_files; i++) {
1328                 if (fill_def_thread())
1329                         return 1;
1330                 if (parse_jobs_ini(ini_file[i], i))
1331                         return 1;
1332                 free(ini_file[i]);
1333         }
1334
1335         free(ini_file);
1336         options_mem_free(&def_thread);
1337
1338         if (!thread_number) {
1339                 if (dump_cmdline)
1340                         return 0;
1341                 if (exec_profile)
1342                         return 0;
1343
1344                 log_err("No jobs(s) defined\n\n");
1345                 usage(argv[0]);
1346                 return 1;
1347         }
1348
1349         if (def_thread.o.gtod_offload) {
1350                 fio_gtod_init();
1351                 fio_gtod_offload = 1;
1352                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1353         }
1354
1355         log_info("%s\n", fio_version_string);
1356         return 0;
1357 }