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 1.60.1";
  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 static int def_timeout;
  33
  34 static struct thread_data def_thread;
  35 struct thread_data *threads = NULL;
  36
  37 int exitall_on_terminate = 0;
  38 int terse_output = 0;
  39 int eta_print;
  40 unsigned long long mlock_size = 0;
  41 FILE *f_out = NULL;
  42 FILE *f_err = NULL;
  43 char **job_sections = NULL;
  44 int nr_job_sections = 0;
  45 char *exec_profile = NULL;
  46 int warnings_fatal = 0;
  47 int terse_version = 2;
  48
  49 int write_bw_log = 0;
  50 int read_only = 0;
  51
  52 static int write_lat_log;
  53
  54 static int prev_group_jobs;
  55
  56 unsigned long fio_debug = 0;
  57 unsigned int fio_debug_jobno = -1;
  58 unsigned int *fio_debug_jobp = NULL;
  59
  60 static char cmd_optstr[256];
  61
  62 /*
  63  * Command line options. These will contain the above, plus a few
  64  * extra that only pertain to fio itself and not jobs.
  65  */
  66 static struct option l_opts[FIO_NR_OPTIONS] = {
  67         {
  68                 .name           = (char *) "output",
  69                 .has_arg        = required_argument,
  70                 .val            = 'o',
  71         },
  72         {
  73                 .name           = (char *) "timeout",
  74                 .has_arg        = required_argument,
  75                 .val            = 't',
  76         },
  77         {
  78                 .name           = (char *) "latency-log",
  79                 .has_arg        = required_argument,
  80                 .val            = 'l',
  81         },
  82         {
  83                 .name           = (char *) "bandwidth-log",
  84                 .has_arg        = required_argument,
  85                 .val            = 'b',
  86         },
  87         {
  88                 .name           = (char *) "minimal",
  89                 .has_arg        = optional_argument,
  90                 .val            = 'm',
  91         },
  92         {
  93                 .name           = (char *) "version",
  94                 .has_arg        = no_argument,
  95                 .val            = 'v',
  96         },
  97         {
  98                 .name           = (char *) "help",
  99                 .has_arg        = no_argument,
 100                 .val            = 'h',
 101         },
 102         {
 103                 .name           = (char *) "cmdhelp",
 104                 .has_arg        = optional_argument,
 105                 .val            = 'c',
 106         },
 107         {
 108                 .name           = (char *) "showcmd",
 109                 .has_arg        = no_argument,
 110                 .val            = 's',
 111         },
 112         {
 113                 .name           = (char *) "readonly",
 114                 .has_arg        = no_argument,
 115                 .val            = 'r',
 116         },
 117         {
 118                 .name           = (char *) "eta",
 119                 .has_arg        = required_argument,
 120                 .val            = 'e',
 121         },
 122         {
 123                 .name           = (char *) "debug",
 124                 .has_arg        = required_argument,
 125                 .val            = 'd',
 126         },
 127         {
 128                 .name           = (char *) "section",
 129                 .has_arg        = required_argument,
 130                 .val            = 'x',
 131         },
 132         {
 133                 .name           = (char *) "alloc-size",
 134                 .has_arg        = required_argument,
 135                 .val            = 'a',
 136         },
 137         {
 138                 .name           = (char *) "profile",
 139                 .has_arg        = required_argument,
 140                 .val            = 'p',
 141         },
 142         {
 143                 .name           = (char *) "warnings-fatal",
 144                 .has_arg        = no_argument,
 145                 .val            = 'w',
 146         },
 147         {
 148                 .name           = (char *) "max-jobs",
 149                 .has_arg        = required_argument,
 150                 .val            = 'j',
 151         },
 152         {
 153                 .name           = (char *) "terse-version",
 154                 .has_arg        = required_argument,
 155                 .val            = 'V',
 156         },
 157         {
 158                 .name           = NULL,
 159         },
 160 };
 161
 162 FILE *get_f_out()
 163 {
 164         return f_out;
 165 }
 166
 167 FILE *get_f_err()
 168 {
 169         return f_err;
 170 }
 171
 172 /*
 173  * Return a free job structure.
 174  */
 175 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 176 {
 177         struct thread_data *td;
 178
 179         if (global)
 180                 return &def_thread;
 181         if (thread_number >= max_jobs) {
 182                 log_err("error: maximum number of jobs (%d) reached.\n",
 183                                 max_jobs);
 184                 return NULL;
 185         }
 186
 187         td = &threads[thread_number++];
 188         *td = *parent;
 189
 190         td->o.uid = td->o.gid = -1U;
 191
 192         dup_files(td, parent);
 193         options_mem_dupe(td);
 194
 195         profile_add_hooks(td);
 196
 197         td->thread_number = thread_number;
 198         return td;
 199 }
 200
 201 static void put_job(struct thread_data *td)
 202 {
 203         if (td == &def_thread)
 204                 return;
 205
 206         profile_td_exit(td);
 207
 208         if (td->error)
 209                 log_info("fio: %s\n", td->verror);
 210
 211         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 212         thread_number--;
 213 }
 214
 215 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
 216 {
 217         unsigned int bs = td->o.min_bs[ddir];
 218         unsigned long long bytes_per_sec;
 219
 220         assert(ddir_rw(ddir));
 221
 222         if (td->o.rate[ddir])
 223                 bytes_per_sec = td->o.rate[ddir];
 224         else
 225                 bytes_per_sec = td->o.rate_iops[ddir] * bs;
 226
 227         if (!bytes_per_sec) {
 228                 log_err("rate lower than supported\n");
 229                 return -1;
 230         }
 231
 232         td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
 233         td->rate_pending_usleep[ddir] = 0;
 234         return 0;
 235 }
 236
 237 static int setup_rate(struct thread_data *td)
 238 {
 239         int ret = 0;
 240
 241         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
 242                 ret = __setup_rate(td, DDIR_READ);
 243         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
 244                 ret |= __setup_rate(td, DDIR_WRITE);
 245
 246         return ret;
 247 }
 248
 249 static int fixed_block_size(struct thread_options *o)
 250 {
 251         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
 252                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
 253                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
 254 }
 255
 256 /*
 257  * Lazy way of fixing up options that depend on each other. We could also
 258  * define option callback handlers, but this is easier.
 259  */
 260 static int fixup_options(struct thread_data *td)
 261 {
 262         struct thread_options *o = &td->o;
 263         int ret = 0;
 264
 265 #ifndef FIO_HAVE_PSHARED_MUTEX
 266         if (!o->use_thread) {
 267                 log_info("fio: this platform does not support process shared"
 268                          " mutexes, forcing use of threads. Use the 'thread'"
 269                          " option to get rid of this warning.\n");
 270                 o->use_thread = 1;
 271                 ret = warnings_fatal;
 272         }
 273 #endif
 274
 275         if (o->write_iolog_file && o->read_iolog_file) {
 276                 log_err("fio: read iolog overrides write_iolog\n");
 277                 free(o->write_iolog_file);
 278                 o->write_iolog_file = NULL;
 279                 ret = warnings_fatal;
 280         }
 281
 282         /*
 283          * only really works for sequential io for now, and with 1 file
 284          */
 285         if (o->zone_size && td_random(td) && o->open_files == 1)
 286                 o->zone_size = 0;
 287
 288         /*
 289          * Reads can do overwrites, we always need to pre-create the file
 290          */
 291         if (td_read(td) || td_rw(td))
 292                 o->overwrite = 1;
 293
 294         if (!o->min_bs[DDIR_READ])
 295                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
 296         if (!o->max_bs[DDIR_READ])
 297                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
 298         if (!o->min_bs[DDIR_WRITE])
 299                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 300         if (!o->max_bs[DDIR_WRITE])
 301                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 302
 303         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
 304
 305         /*
 306          * For random IO, allow blockalign offset other than min_bs.
 307          */
 308         if (!o->ba[DDIR_READ] || !td_random(td))
 309                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
 310         if (!o->ba[DDIR_WRITE] || !td_random(td))
 311                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
 312
 313         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
 314             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
 315             !o->norandommap) {
 316                 log_err("fio: Any use of blockalign= turns off randommap\n");
 317                 o->norandommap = 1;
 318                 ret = warnings_fatal;
 319         }
 320
 321         if (!o->file_size_high)
 322                 o->file_size_high = o->file_size_low;
 323
 324         if (o->norandommap && o->verify != VERIFY_NONE
 325             && !fixed_block_size(o))  {
 326                 log_err("fio: norandommap given for variable block sizes, "
 327                         "verify disabled\n");
 328                 o->verify = VERIFY_NONE;
 329                 ret = warnings_fatal;
 330         }
 331         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
 332                 log_err("fio: bs_unaligned may not work with raw io\n");
 333
 334         /*
 335          * thinktime_spin must be less than thinktime
 336          */
 337         if (o->thinktime_spin > o->thinktime)
 338                 o->thinktime_spin = o->thinktime;
 339
 340         /*
 341          * The low water mark cannot be bigger than the iodepth
 342          */
 343         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
 344                 /*
 345                  * syslet work around - if the workload is sequential,
 346                  * we want to let the queue drain all the way down to
 347                  * avoid seeking between async threads
 348                  */
 349                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 350                         o->iodepth_low = 1;
 351                 else
 352                         o->iodepth_low = o->iodepth;
 353         }
 354
 355         /*
 356          * If batch number isn't set, default to the same as iodepth
 357          */
 358         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
 359                 o->iodepth_batch = o->iodepth;
 360
 361         if (o->nr_files > td->files_index)
 362                 o->nr_files = td->files_index;
 363
 364         if (o->open_files > o->nr_files || !o->open_files)
 365                 o->open_files = o->nr_files;
 366
 367         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
 368             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
 369                 o->rate_iops_min[1]))) {
 370                 log_err("fio: rate and rate_iops are mutually exclusive\n");
 371                 ret = 1;
 372         }
 373         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
 374             (o->rate_iops[0] < o->rate_iops_min[0]) ||
 375             (o->rate_iops[1] < o->rate_iops_min[1])) {
 376                 log_err("fio: minimum rate exceeds rate\n");
 377                 ret = 1;
 378         }
 379
 380         if (!o->timeout && o->time_based) {
 381                 log_err("fio: time_based requires a runtime/timeout setting\n");
 382                 o->time_based = 0;
 383                 ret = warnings_fatal;
 384         }
 385
 386         if (o->fill_device && !o->size)
 387                 o->size = -1ULL;
 388
 389         if (o->verify != VERIFY_NONE) {
 390                 if (td_write(td) && o->do_verify && o->numjobs > 1) {
 391                         log_info("Multiple writers may overwrite blocks that "
 392                                 "belong to other jobs. This can cause "
 393                                 "verification failures.\n");
 394                         ret = warnings_fatal;
 395                 }
 396
 397                 o->refill_buffers = 1;
 398                 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
 399                     !o->verify_interval)
 400                         o->verify_interval = o->min_bs[DDIR_WRITE];
 401         }
 402
 403         if (o->pre_read) {
 404                 o->invalidate_cache = 0;
 405                 if (td->io_ops->flags & FIO_PIPEIO) {
 406                         log_info("fio: cannot pre-read files with an IO engine"
 407                                  " that isn't seekable. Pre-read disabled.\n");
 408                         ret = warnings_fatal;
 409                 }
 410         }
 411
 412 #ifndef FIO_HAVE_FDATASYNC
 413         if (o->fdatasync_blocks) {
 414                 log_info("fio: this platform does not support fdatasync()"
 415                          " falling back to using fsync().  Use the 'fsync'"
 416                          " option instead of 'fdatasync' to get rid of"
 417                          " this warning\n");
 418                 o->fsync_blocks = o->fdatasync_blocks;
 419                 o->fdatasync_blocks = 0;
 420                 ret = warnings_fatal;
 421         }
 422 #endif
 423
 424         return ret;
 425 }
 426
 427 /*
 428  * This function leaks the buffer
 429  */
 430 static char *to_kmg(unsigned int val)
 431 {
 432         char *buf = malloc(32);
 433         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 434         char *p = post;
 435
 436         do {
 437                 if (val & 1023)
 438                         break;
 439
 440                 val >>= 10;
 441                 p++;
 442         } while (*p);
 443
 444         snprintf(buf, 31, "%u%c", val, *p);
 445         return buf;
 446 }
 447
 448 /* External engines are specified by "external:name.o") */
 449 static const char *get_engine_name(const char *str)
 450 {
 451         char *p = strstr(str, ":");
 452
 453         if (!p)
 454                 return str;
 455
 456         p++;
 457         strip_blank_front(&p);
 458         strip_blank_end(p);
 459         return p;
 460 }
 461
 462 static int exists_and_not_file(const char *filename)
 463 {
 464         struct stat sb;
 465
 466         if (lstat(filename, &sb) == -1)
 467                 return 0;
 468
 469         /* \\.\ is the device namespace in Windows, where every file
 470          * is a device node */
 471         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
 472                 return 0;
 473
 474         return 1;
 475 }
 476
 477 static void td_fill_rand_seeds_os(struct thread_data *td)
 478 {
 479         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
 480         os_random_seed(td->rand_seeds[1], &td->verify_state);
 481         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
 482
 483         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
 484                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
 485
 486         os_random_seed(td->rand_seeds[5], &td->file_size_state);
 487         os_random_seed(td->rand_seeds[6], &td->trim_state);
 488
 489         if (!td_random(td))
 490                 return;
 491
 492         if (td->o.rand_repeatable)
 493                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
 494
 495         os_random_seed(td->rand_seeds[4], &td->random_state);
 496 }
 497
 498 static void td_fill_rand_seeds_internal(struct thread_data *td)
 499 {
 500         init_rand_seed(&td->__bsrange_state, td->rand_seeds[0]);
 501         init_rand_seed(&td->__verify_state, td->rand_seeds[1]);
 502         init_rand_seed(&td->__rwmix_state, td->rand_seeds[2]);
 503
 504         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
 505                 init_rand_seed(&td->__next_file_state, td->rand_seeds[3]);
 506
 507         init_rand_seed(&td->__file_size_state, td->rand_seeds[5]);
 508         init_rand_seed(&td->__trim_state, td->rand_seeds[6]);
 509
 510         if (!td_random(td))
 511                 return;
 512
 513         if (td->o.rand_repeatable)
 514                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
 515
 516         init_rand_seed(&td->__random_state, td->rand_seeds[4]);
 517 }
 518
 519 void td_fill_rand_seeds(struct thread_data *td)
 520 {
 521         if (td->o.use_os_rand)
 522                 td_fill_rand_seeds_os(td);
 523         else
 524                 td_fill_rand_seeds_internal(td);
 525
 526         init_rand_seed(&td->buf_state, td->rand_seeds[7]);
 527 }
 528
 529 /*
 530  * Initialize the various random states we need (random io, block size ranges,
 531  * read/write mix, etc).
 532  */
 533 static int init_random_state(struct thread_data *td)
 534 {
 535         int fd;
 536
 537         fd = open("/dev/urandom", O_RDONLY);
 538         if (fd == -1) {
 539                 td_verror(td, errno, "open");
 540                 return 1;
 541         }
 542
 543         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
 544             (int) sizeof(td->rand_seeds)) {
 545                 td_verror(td, EIO, "read");
 546                 close(fd);
 547                 return 1;
 548         }
 549
 550         close(fd);
 551         td_fill_rand_seeds(td);
 552         return 0;
 553 }
 554
 555 /*
 556  * Adds a job to the list of things todo. Sanitizes the various options
 557  * to make sure we don't have conflicts, and initializes various
 558  * members of td.
 559  */
 560 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 561 {
 562         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 563                                    "randread", "randwrite", "randrw" };
 564         unsigned int i;
 565         const char *engine;
 566         char fname[PATH_MAX];
 567         int numjobs, file_alloced;
 568
 569         /*
 570          * the def_thread is just for options, it's not a real job
 571          */
 572         if (td == &def_thread)
 573                 return 0;
 574
 575         /*
 576          * if we are just dumping the output command line, don't add the job
 577          */
 578         if (dump_cmdline) {
 579                 put_job(td);
 580                 return 0;
 581         }
 582
 583         if (profile_td_init(td))
 584                 goto err;
 585
 586         engine = get_engine_name(td->o.ioengine);
 587         td->io_ops = load_ioengine(td, engine);
 588         if (!td->io_ops) {
 589                 log_err("fio: failed to load engine %s\n", engine);
 590                 goto err;
 591         }
 592
 593         if (td->o.use_thread)
 594                 nr_thread++;
 595         else
 596                 nr_process++;
 597
 598         if (td->o.odirect)
 599                 td->io_ops->flags |= FIO_RAWIO;
 600
 601         file_alloced = 0;
 602         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
 603                 file_alloced = 1;
 604
 605                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
 606                         add_file(td, jobname);
 607                 else {
 608                         for (i = 0; i < td->o.nr_files; i++) {
 609                                 sprintf(fname, "%s.%d.%d", jobname,
 610                                                         td->thread_number, i);
 611                                 add_file(td, fname);
 612                         }
 613                 }
 614         }
 615
 616         if (fixup_options(td))
 617                 goto err;
 618
 619         if (td->io_ops->flags & FIO_DISKLESSIO) {
 620                 struct fio_file *f;
 621
 622                 for_each_file(td, f, i)
 623                         f->real_file_size = -1ULL;
 624         }
 625
 626         td->mutex = fio_mutex_init(0);
 627
 628         td->ts.clat_percentiles = td->o.clat_percentiles;
 629         if (td->o.overwrite_plist)
 630                 td->ts.percentile_list = td->o.percentile_list;
 631         else
 632                 td->ts.percentile_list = NULL;
 633
 634         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 635         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 636         td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
 637         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 638         td->ddir_seq_nr = td->o.ddir_seq_nr;
 639
 640         if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
 641                 prev_group_jobs = 0;
 642                 groupid++;
 643         }
 644
 645         td->groupid = groupid;
 646         prev_group_jobs++;
 647
 648         if (init_random_state(td))
 649                 goto err;
 650
 651         if (setup_rate(td))
 652                 goto err;
 653
 654         if (td->o.write_lat_log) {
 655                 setup_log(&td->ts.lat_log);
 656                 setup_log(&td->ts.slat_log);
 657                 setup_log(&td->ts.clat_log);
 658         }
 659         if (td->o.write_bw_log)
 660                 setup_log(&td->ts.bw_log);
 661
 662         if (!td->o.name)
 663                 td->o.name = strdup(jobname);
 664
 665         if (!terse_output) {
 666                 if (!job_add_num) {
 667                         if (!strcmp(td->io_ops->name, "cpuio")) {
 668                                 log_info("%s: ioengine=cpu, cpuload=%u,"
 669                                          " cpucycle=%u\n", td->o.name,
 670                                                         td->o.cpuload,
 671                                                         td->o.cpucycle);
 672                         } else {
 673                                 char *c1, *c2, *c3, *c4;
 674
 675                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
 676                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
 677                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
 678                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
 679
 680                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
 681                                          " ioengine=%s, iodepth=%u\n",
 682                                                 td->o.name, td->groupid,
 683                                                 ddir_str[td->o.td_ddir],
 684                                                 c1, c2, c3, c4,
 685                                                 td->io_ops->name,
 686                                                 td->o.iodepth);
 687
 688                                 free(c1);
 689                                 free(c2);
 690                                 free(c3);
 691                                 free(c4);
 692                         }
 693                 } else if (job_add_num == 1)
 694                         log_info("...\n");
 695         }
 696
 697         /*
 698          * recurse add identical jobs, clear numjobs and stonewall options
 699          * as they don't apply to sub-jobs
 700          */
 701         numjobs = td->o.numjobs;
 702         while (--numjobs) {
 703                 struct thread_data *td_new = get_new_job(0, td);
 704
 705                 if (!td_new)
 706                         goto err;
 707
 708                 td_new->o.numjobs = 1;
 709                 td_new->o.stonewall = 0;
 710                 td_new->o.new_group = 0;
 711
 712                 if (file_alloced) {
 713                         td_new->o.filename = NULL;
 714                         td_new->files_index = 0;
 715                         td_new->files_size = 0;
 716                         td_new->files = NULL;
 717                 }
 718
 719                 job_add_num = numjobs - 1;
 720
 721                 if (add_job(td_new, jobname, job_add_num))
 722                         goto err;
 723         }
 724
 725         return 0;
 726 err:
 727         put_job(td);
 728         return -1;
 729 }
 730
 731 /*
 732  * Parse as if 'o' was a command line
 733  */
 734 void add_job_opts(const char **o)
 735 {
 736         struct thread_data *td, *td_parent;
 737         int i, in_global = 1;
 738         char jobname[32];
 739
 740         i = 0;
 741         td_parent = td = NULL;
 742         while (o[i]) {
 743                 if (!strncmp(o[i], "name", 4)) {
 744                         in_global = 0;
 745                         if (td)
 746                                 add_job(td, jobname, 0);
 747                         td = NULL;
 748                         sprintf(jobname, "%s", o[i] + 5);
 749                 }
 750                 if (in_global && !td_parent)
 751                         td_parent = get_new_job(1, &def_thread);
 752                 else if (!in_global && !td) {
 753                         if (!td_parent)
 754                                 td_parent = &def_thread;
 755                         td = get_new_job(0, td_parent);
 756                 }
 757                 if (in_global)
 758                         fio_options_parse(td_parent, (char **) &o[i], 1);
 759                 else
 760                         fio_options_parse(td, (char **) &o[i], 1);
 761                 i++;
 762         }
 763
 764         if (td)
 765                 add_job(td, jobname, 0);
 766 }
 767
 768 static int skip_this_section(const char *name)
 769 {
 770         int i;
 771
 772         if (!nr_job_sections)
 773                 return 0;
 774         if (!strncmp(name, "global", 6))
 775                 return 0;
 776
 777         for (i = 0; i < nr_job_sections; i++)
 778                 if (!strcmp(job_sections[i], name))
 779                         return 0;
 780
 781         return 1;
 782 }
 783
 784 static int is_empty_or_comment(char *line)
 785 {
 786         unsigned int i;
 787
 788         for (i = 0; i < strlen(line); i++) {
 789                 if (line[i] == ';')
 790                         return 1;
 791                 if (line[i] == '#')
 792                         return 1;
 793                 if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
 794                         return 0;
 795         }
 796
 797         return 1;
 798 }
 799
 800 /*
 801  * This is our [ini] type file parser.
 802  */
 803 static int parse_jobs_ini(char *file, int stonewall_flag)
 804 {
 805         unsigned int global;
 806         struct thread_data *td;
 807         char *string, *name;
 808         FILE *f;
 809         char *p;
 810         int ret = 0, stonewall;
 811         int first_sect = 1;
 812         int skip_fgets = 0;
 813         int inside_skip = 0;
 814         char **opts;
 815         int i, alloc_opts, num_opts;
 816
 817         if (!strcmp(file, "-"))
 818                 f = stdin;
 819         else
 820                 f = fopen(file, "r");
 821
 822         if (!f) {
 823                 perror("fopen job file");
 824                 return 1;
 825         }
 826
 827         string = malloc(4096);
 828
 829         /*
 830          * it's really 256 + small bit, 280 should suffice
 831          */
 832         name = malloc(280);
 833         memset(name, 0, 280);
 834
 835         alloc_opts = 8;
 836         opts = malloc(sizeof(char *) * alloc_opts);
 837         num_opts = 0;
 838
 839         stonewall = stonewall_flag;
 840         do {
 841                 /*
 842                  * if skip_fgets is set, we already have loaded a line we
 843                  * haven't handled.
 844                  */
 845                 if (!skip_fgets) {
 846                         p = fgets(string, 4095, f);
 847                         if (!p)
 848                                 break;
 849                 }
 850
 851                 skip_fgets = 0;
 852                 strip_blank_front(&p);
 853                 strip_blank_end(p);
 854
 855                 if (is_empty_or_comment(p))
 856                         continue;
 857                 if (sscanf(p, "[%255[^\n]]", name) != 1) {
 858                         if (inside_skip)
 859                                 continue;
 860                         log_err("fio: option <%s> outside of [] job section\n",
 861                                                                         p);
 862                         break;
 863                 }
 864
 865                 name[strlen(name) - 1] = '\0';
 866
 867                 if (skip_this_section(name)) {
 868                         inside_skip = 1;
 869                         continue;
 870                 } else
 871                         inside_skip = 0;
 872
 873                 global = !strncmp(name, "global", 6);
 874
 875                 if (dump_cmdline) {
 876                         if (first_sect)
 877                                 log_info("fio ");
 878                         if (!global)
 879                                 log_info("--name=%s ", name);
 880                         first_sect = 0;
 881                 }
 882
 883                 td = get_new_job(global, &def_thread);
 884                 if (!td) {
 885                         ret = 1;
 886                         break;
 887                 }
 888
 889                 /*
 890                  * Seperate multiple job files by a stonewall
 891                  */
 892                 if (!global && stonewall) {
 893                         td->o.stonewall = stonewall;
 894                         stonewall = 0;
 895                 }
 896
 897                 num_opts = 0;
 898                 memset(opts, 0, alloc_opts * sizeof(char *));
 899
 900                 while ((p = fgets(string, 4096, f)) != NULL) {
 901                         if (is_empty_or_comment(p))
 902                                 continue;
 903
 904                         strip_blank_front(&p);
 905
 906                         /*
 907                          * new section, break out and make sure we don't
 908                          * fgets() a new line at the top.
 909                          */
 910                         if (p[0] == '[') {
 911                                 skip_fgets = 1;
 912                                 break;
 913                         }
 914
 915                         strip_blank_end(p);
 916
 917                         if (num_opts == alloc_opts) {
 918                                 alloc_opts <<= 1;
 919                                 opts = realloc(opts,
 920                                                 alloc_opts * sizeof(char *));
 921                         }
 922
 923                         opts[num_opts] = strdup(p);
 924                         num_opts++;
 925                 }
 926
 927                 ret = fio_options_parse(td, opts, num_opts);
 928                 if (!ret) {
 929                         if (dump_cmdline)
 930                                 for (i = 0; i < num_opts; i++)
 931                                         log_info("--%s ", opts[i]);
 932
 933                         ret = add_job(td, name, 0);
 934                 } else {
 935                         log_err("fio: job %s dropped\n", name);
 936                         put_job(td);
 937                 }
 938
 939                 for (i = 0; i < num_opts; i++)
 940                         free(opts[i]);
 941                 num_opts = 0;
 942         } while (!ret);
 943
 944         if (dump_cmdline)
 945                 log_info("\n");
 946
 947         for (i = 0; i < num_opts; i++)
 948                 free(opts[i]);
 949
 950         free(string);
 951         free(name);
 952         free(opts);
 953         if (f != stdin)
 954                 fclose(f);
 955         return ret;
 956 }
 957
 958 static int fill_def_thread(void)
 959 {
 960         memset(&def_thread, 0, sizeof(def_thread));
 961
 962         fio_getaffinity(getpid(), &def_thread.o.cpumask);
 963         def_thread.o.timeout = def_timeout;
 964
 965         /*
 966          * fill default options
 967          */
 968         fio_fill_default_options(&def_thread);
 969         return 0;
 970 }
 971
 972 static void free_shm(void)
 973 {
 974         struct shmid_ds sbuf;
 975
 976         if (threads) {
 977                 void *tp = threads;
 978
 979                 threads = NULL;
 980                 file_hash_exit();
 981                 fio_debug_jobp = NULL;
 982                 shmdt(tp);
 983                 shmctl(shm_id, IPC_RMID, &sbuf);
 984         }
 985
 986         scleanup();
 987 }
 988
 989 /*
 990  * The thread area is shared between the main process and the job
 991  * threads/processes. So setup a shared memory segment that will hold
 992  * all the job info. We use the end of the region for keeping track of
 993  * open files across jobs, for file sharing.
 994  */
 995 static int setup_thread_area(void)
 996 {
 997         void *hash;
 998
 999         /*
1000          * 1024 is too much on some machines, scale max_jobs if
1001          * we get a failure that looks like too large a shm segment
1002          */
1003         do {
1004                 size_t size = max_jobs * sizeof(struct thread_data);
1005
1006                 size += file_hash_size;
1007                 size += sizeof(unsigned int);
1008
1009                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1010                 if (shm_id != -1)
1011                         break;
1012                 if (errno != EINVAL) {
1013                         perror("shmget");
1014                         break;
1015                 }
1016
1017                 max_jobs >>= 1;
1018         } while (max_jobs);
1019
1020         if (shm_id == -1)
1021                 return 1;
1022
1023         threads = shmat(shm_id, NULL, 0);
1024         if (threads == (void *) -1) {
1025                 perror("shmat");
1026                 return 1;
1027         }
1028
1029         memset(threads, 0, max_jobs * sizeof(struct thread_data));
1030         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
1031         fio_debug_jobp = (void *) hash + file_hash_size;
1032         *fio_debug_jobp = -1;
1033         file_hash_init(hash);
1034         atexit(free_shm);
1035         return 0;
1036 }
1037
1038 static void usage(const char *name)
1039 {
1040         printf("%s\n", fio_version_string);
1041         printf("%s [options] [job options] <job file(s)>\n", name);
1042         printf("\t--debug=options\tEnable debug logging\n");
1043         printf("\t--output\tWrite output to file\n");
1044         printf("\t--timeout\tRuntime in seconds\n");
1045         printf("\t--latency-log\tGenerate per-job latency logs\n");
1046         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1047         printf("\t--minimal\tMinimal (terse) output\n");
1048         printf("\t--version\tPrint version info and exit\n");
1049         printf("\t--terse-version=x Terse version output format\n");
1050         printf("\t--help\t\tPrint this page\n");
1051         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
1052                 " them\n");
1053         printf("\t--showcmd\tTurn a job file into command line options\n");
1054         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
1055         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
1056         printf("\t--readonly\tTurn on safety read-only checks, preventing"
1057                 " writes\n");
1058         printf("\t--section=name\tOnly run specified section in job file\n");
1059         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
1060                 " (def 1024)\n");
1061         printf("\t--warnings-fatal Fio parser warnings are fatal\n");
1062         printf("\t--max-jobs\tMaximum number of threads/processes to support\n");
1063         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1064         printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
1065 }
1066
1067 #ifdef FIO_INC_DEBUG
1068 struct debug_level debug_levels[] = {
1069         { .name = "process",    .shift = FD_PROCESS, },
1070         { .name = "file",       .shift = FD_FILE, },
1071         { .name = "io",         .shift = FD_IO, },
1072         { .name = "mem",        .shift = FD_MEM, },
1073         { .name = "blktrace",   .shift = FD_BLKTRACE },
1074         { .name = "verify",     .shift = FD_VERIFY },
1075         { .name = "random",     .shift = FD_RANDOM },
1076         { .name = "parse",      .shift = FD_PARSE },
1077         { .name = "diskutil",   .shift = FD_DISKUTIL },
1078         { .name = "job",        .shift = FD_JOB },
1079         { .name = "mutex",      .shift = FD_MUTEX },
1080         { .name = "profile",    .shift = FD_PROFILE },
1081         { .name = "time",       .shift = FD_TIME },
1082         { .name = NULL, },
1083 };
1084
1085 static int set_debug(const char *string)
1086 {
1087         struct debug_level *dl;
1088         char *p = (char *) string;
1089         char *opt;
1090         int i;
1091
1092         if (!strcmp(string, "?") || !strcmp(string, "help")) {
1093                 log_info("fio: dumping debug options:");
1094                 for (i = 0; debug_levels[i].name; i++) {
1095                         dl = &debug_levels[i];
1096                         log_info("%s,", dl->name);
1097                 }
1098                 log_info("all\n");
1099                 return 1;
1100         }
1101
1102         while ((opt = strsep(&p, ",")) != NULL) {
1103                 int found = 0;
1104
1105                 if (!strncmp(opt, "all", 3)) {
1106                         log_info("fio: set all debug options\n");
1107                         fio_debug = ~0UL;
1108                         continue;
1109                 }
1110
1111                 for (i = 0; debug_levels[i].name; i++) {
1112                         dl = &debug_levels[i];
1113                         found = !strncmp(opt, dl->name, strlen(dl->name));
1114                         if (!found)
1115                                 continue;
1116
1117                         if (dl->shift == FD_JOB) {
1118                                 opt = strchr(opt, ':');
1119                                 if (!opt) {
1120                                         log_err("fio: missing job number\n");
1121                                         break;
1122                                 }
1123                                 opt++;
1124                                 fio_debug_jobno = atoi(opt);
1125                                 log_info("fio: set debug jobno %d\n",
1126                                                         fio_debug_jobno);
1127                         } else {
1128                                 log_info("fio: set debug option %s\n", opt);
1129                                 fio_debug |= (1UL << dl->shift);
1130                         }
1131                         break;
1132                 }
1133
1134                 if (!found)
1135                         log_err("fio: debug mask %s not found\n", opt);
1136         }
1137         return 0;
1138 }
1139 #else
1140 static int set_debug(const char *string)
1141 {
1142         log_err("fio: debug tracing not included in build\n");
1143         return 1;
1144 }
1145 #endif
1146
1147 static void fio_options_fill_optstring(void)
1148 {
1149         char *ostr = cmd_optstr;
1150         int i, c;
1151
1152         c = i = 0;
1153         while (l_opts[i].name) {
1154                 ostr[c++] = l_opts[i].val;
1155                 if (l_opts[i].has_arg == required_argument)
1156                         ostr[c++] = ':';
1157                 else if (l_opts[i].has_arg == optional_argument) {
1158                         ostr[c++] = ':';
1159                         ostr[c++] = ':';
1160                 }
1161                 i++;
1162         }
1163         ostr[c] = '\0';
1164 }
1165
1166 static int parse_cmd_line(int argc, char *argv[])
1167 {
1168         struct thread_data *td = NULL;
1169         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1170         char *ostr = cmd_optstr;
1171
1172         while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1173                 switch (c) {
1174                 case 'a':
1175                         smalloc_pool_size = atoi(optarg);
1176                         break;
1177                 case 't':
1178                         def_timeout = atoi(optarg);
1179                         break;
1180                 case 'l':
1181                         write_lat_log = 1;
1182                         break;
1183                 case 'b':
1184                         write_bw_log = 1;
1185                         break;
1186                 case 'o':
1187                         f_out = fopen(optarg, "w+");
1188                         if (!f_out) {
1189                                 perror("fopen output");
1190                                 exit(1);
1191                         }
1192                         f_err = f_out;
1193                         break;
1194                 case 'm':
1195                         terse_output = 1;
1196                         break;
1197                 case 'h':
1198                         usage(argv[0]);
1199                         exit(0);
1200                 case 'c':
1201                         exit(fio_show_option_help(optarg));
1202                 case 's':
1203                         dump_cmdline = 1;
1204                         break;
1205                 case 'r':
1206                         read_only = 1;
1207                         break;
1208                 case 'v':
1209                         log_info("%s\n", fio_version_string);
1210                         exit(0);
1211                 case 'V':
1212                         terse_version = atoi(optarg);
1213                         if (terse_version != 2) {
1214                                 log_err("fio: bad terse version format\n");
1215                                 exit_val = 1;
1216                                 do_exit++;
1217                         }
1218                         break;
1219                 case 'e':
1220                         if (!strcmp("always", optarg))
1221                                 eta_print = FIO_ETA_ALWAYS;
1222                         else if (!strcmp("never", optarg))
1223                                 eta_print = FIO_ETA_NEVER;
1224                         break;
1225                 case 'd':
1226                         if (set_debug(optarg))
1227                                 do_exit++;
1228                         break;
1229                 case 'x': {
1230                         size_t new_size;
1231
1232                         if (!strcmp(optarg, "global")) {
1233                                 log_err("fio: can't use global as only "
1234                                         "section\n");
1235                                 do_exit++;
1236                                 exit_val = 1;
1237                                 break;
1238                         }
1239                         new_size = (nr_job_sections + 1) * sizeof(char *);
1240                         job_sections = realloc(job_sections, new_size);
1241                         job_sections[nr_job_sections] = strdup(optarg);
1242                         nr_job_sections++;
1243                         break;
1244                         }
1245                 case 'p':
1246                         exec_profile = strdup(optarg);
1247                         break;
1248                 case FIO_GETOPT_JOB: {
1249                         const char *opt = l_opts[lidx].name;
1250                         char *val = optarg;
1251
1252                         if (!strncmp(opt, "name", 4) && td) {
1253                                 ret = add_job(td, td->o.name ?: "fio", 0);
1254                                 if (ret)
1255                                         return 0;
1256                                 td = NULL;
1257                         }
1258                         if (!td) {
1259                                 int is_section = !strncmp(opt, "name", 4);
1260                                 int global = 0;
1261
1262                                 if (!is_section || !strncmp(val, "global", 6))
1263                                         global = 1;
1264
1265                                 if (is_section && skip_this_section(val))
1266                                         continue;
1267
1268                                 td = get_new_job(global, &def_thread);
1269                                 if (!td)
1270                                         return 0;
1271                         }
1272
1273                         ret = fio_cmd_option_parse(td, opt, val);
1274                         break;
1275                 }
1276                 case 'w':
1277                         warnings_fatal = 1;
1278                         break;
1279                 case 'j':
1280                         max_jobs = atoi(optarg);
1281                         if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1282                                 log_err("fio: invalid max jobs: %d\n", max_jobs);
1283                                 do_exit++;
1284                                 exit_val = 1;
1285                         }
1286                         break;
1287                 default:
1288                         do_exit++;
1289                         exit_val = 1;
1290                         break;
1291                 }
1292         }
1293
1294         if (do_exit)
1295                 exit(exit_val);
1296
1297         if (td) {
1298                 if (!ret)
1299                         ret = add_job(td, td->o.name ?: "fio", 0);
1300         }
1301
1302         while (optind < argc) {
1303                 ini_idx++;
1304                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1305                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1306                 optind++;
1307         }
1308
1309         return ini_idx;
1310 }
1311
1312 int parse_options(int argc, char *argv[])
1313 {
1314         int job_files, i;
1315
1316         f_out = stdout;
1317         f_err = stderr;
1318
1319         fio_options_fill_optstring();
1320         fio_options_dup_and_init(l_opts);
1321
1322         if (setup_thread_area())
1323                 return 1;
1324         if (fill_def_thread())
1325                 return 1;
1326
1327         job_files = parse_cmd_line(argc, argv);
1328
1329         for (i = 0; i < job_files; i++) {
1330                 if (fill_def_thread())
1331                         return 1;
1332                 if (parse_jobs_ini(ini_file[i], i))
1333                         return 1;
1334                 free(ini_file[i]);
1335         }
1336
1337         free(ini_file);
1338         options_mem_free(&def_thread);
1339
1340         if (!thread_number) {
1341                 if (dump_cmdline)
1342                         return 0;
1343                 if (exec_profile)
1344                         return 0;
1345
1346                 log_err("No jobs(s) defined\n\n");
1347                 usage(argv[0]);
1348                 return 1;
1349         }
1350
1351         if (def_thread.o.gtod_offload) {
1352                 fio_gtod_init();
1353                 fio_gtod_offload = 1;
1354                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1355         }
1356
1357         log_info("%s\n", fio_version_string);
1358         return 0;
1359 }