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 #include "server.h"
  23
  24 #include "lib/getopt.h"
  25
  26 static char fio_version_string[] = "fio 1.58";
  27
  28 #define FIO_RANDSEED            (0xb1899bedUL)
  29
  30 static char **ini_file;
  31 static int max_jobs = FIO_MAX_JOBS;
  32 static int dump_cmdline;
  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 int is_backend = 0;
  49
  50 int write_bw_log = 0;
  51 int read_only = 0;
  52
  53 static int write_lat_log;
  54
  55 static int prev_group_jobs;
  56
  57 unsigned long fio_debug = 0;
  58 unsigned int fio_debug_jobno = -1;
  59 unsigned int *fio_debug_jobp = NULL;
  60
  61 static char cmd_optstr[256];
  62
  63 /*
  64  * Command line options. These will contain the above, plus a few
  65  * extra that only pertain to fio itself and not jobs.
  66  */
  67 static struct option l_opts[FIO_NR_OPTIONS] = {
  68         {
  69                 .name           = (char *) "output",
  70                 .has_arg        = required_argument,
  71                 .val            = 'o',
  72         },
  73         {
  74                 .name           = (char *) "timeout",
  75                 .has_arg        = required_argument,
  76                 .val            = 't',
  77         },
  78         {
  79                 .name           = (char *) "latency-log",
  80                 .has_arg        = required_argument,
  81                 .val            = 'l',
  82         },
  83         {
  84                 .name           = (char *) "bandwidth-log",
  85                 .has_arg        = required_argument,
  86                 .val            = 'b',
  87         },
  88         {
  89                 .name           = (char *) "minimal",
  90                 .has_arg        = optional_argument,
  91                 .val            = 'm',
  92         },
  93         {
  94                 .name           = (char *) "version",
  95                 .has_arg        = no_argument,
  96                 .val            = 'v',
  97         },
  98         {
  99                 .name           = (char *) "help",
 100                 .has_arg        = no_argument,
 101                 .val            = 'h',
 102         },
 103         {
 104                 .name           = (char *) "cmdhelp",
 105                 .has_arg        = optional_argument,
 106                 .val            = 'c',
 107         },
 108         {
 109                 .name           = (char *) "showcmd",
 110                 .has_arg        = no_argument,
 111                 .val            = 's',
 112         },
 113         {
 114                 .name           = (char *) "readonly",
 115                 .has_arg        = no_argument,
 116                 .val            = 'r',
 117         },
 118         {
 119                 .name           = (char *) "eta",
 120                 .has_arg        = required_argument,
 121                 .val            = 'e',
 122         },
 123         {
 124                 .name           = (char *) "debug",
 125                 .has_arg        = required_argument,
 126                 .val            = 'd',
 127         },
 128         {
 129                 .name           = (char *) "section",
 130                 .has_arg        = required_argument,
 131                 .val            = 'x',
 132         },
 133         {
 134                 .name           = (char *) "alloc-size",
 135                 .has_arg        = required_argument,
 136                 .val            = 'a',
 137         },
 138         {
 139                 .name           = (char *) "profile",
 140                 .has_arg        = required_argument,
 141                 .val            = 'p',
 142         },
 143         {
 144                 .name           = (char *) "warnings-fatal",
 145                 .has_arg        = no_argument,
 146                 .val            = 'w',
 147         },
 148         {
 149                 .name           = (char *) "max-jobs",
 150                 .has_arg        = required_argument,
 151                 .val            = 'j',
 152         },
 153         {
 154                 .name           = (char *) "terse-version",
 155                 .has_arg        = required_argument,
 156                 .val            = 'V',
 157         },
 158         {
 159                 .name           = (char *) "server",
 160                 .has_arg        = no_argument,
 161                 .val            = 'S',
 162         },
 163         {
 164                 .name           = NULL,
 165         },
 166 };
 167
 168 FILE *get_f_out()
 169 {
 170         return f_out;
 171 }
 172
 173 FILE *get_f_err()
 174 {
 175         return f_err;
 176 }
 177
 178 /*
 179  * Return a free job structure.
 180  */
 181 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 182 {
 183         struct thread_data *td;
 184
 185         if (global)
 186                 return &def_thread;
 187         if (thread_number >= max_jobs) {
 188                 log_err("error: maximum number of jobs (%d) reached.\n",
 189                                 max_jobs);
 190                 return NULL;
 191         }
 192
 193         td = &threads[thread_number++];
 194         *td = *parent;
 195
 196         td->o.uid = td->o.gid = -1U;
 197
 198         dup_files(td, parent);
 199         options_mem_dupe(td);
 200
 201         profile_add_hooks(td);
 202
 203         td->thread_number = thread_number;
 204         return td;
 205 }
 206
 207 static void put_job(struct thread_data *td)
 208 {
 209         if (td == &def_thread)
 210                 return;
 211
 212         profile_td_exit(td);
 213
 214         if (td->error)
 215                 log_info("fio: %s\n", td->verror);
 216
 217         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 218         thread_number--;
 219 }
 220
 221 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
 222 {
 223         unsigned int bs = td->o.min_bs[ddir];
 224         unsigned long long bytes_per_sec;
 225
 226         assert(ddir_rw(ddir));
 227
 228         if (td->o.rate[ddir])
 229                 bytes_per_sec = td->o.rate[ddir];
 230         else
 231                 bytes_per_sec = td->o.rate_iops[ddir] * bs;
 232
 233         if (!bytes_per_sec) {
 234                 log_err("rate lower than supported\n");
 235                 return -1;
 236         }
 237
 238         td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
 239         td->rate_pending_usleep[ddir] = 0;
 240         return 0;
 241 }
 242
 243 static int setup_rate(struct thread_data *td)
 244 {
 245         int ret = 0;
 246
 247         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
 248                 ret = __setup_rate(td, DDIR_READ);
 249         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
 250                 ret |= __setup_rate(td, DDIR_WRITE);
 251
 252         return ret;
 253 }
 254
 255 static int fixed_block_size(struct thread_options *o)
 256 {
 257         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
 258                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
 259                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
 260 }
 261
 262 /*
 263  * Lazy way of fixing up options that depend on each other. We could also
 264  * define option callback handlers, but this is easier.
 265  */
 266 static int fixup_options(struct thread_data *td)
 267 {
 268         struct thread_options *o = &td->o;
 269         int ret = 0;
 270
 271 #ifndef FIO_HAVE_PSHARED_MUTEX
 272         if (!o->use_thread) {
 273                 log_info("fio: this platform does not support process shared"
 274                          " mutexes, forcing use of threads. Use the 'thread'"
 275                          " option to get rid of this warning.\n");
 276                 o->use_thread = 1;
 277                 ret = warnings_fatal;
 278         }
 279 #endif
 280
 281         if (o->write_iolog_file && o->read_iolog_file) {
 282                 log_err("fio: read iolog overrides write_iolog\n");
 283                 free(o->write_iolog_file);
 284                 o->write_iolog_file = NULL;
 285                 ret = warnings_fatal;
 286         }
 287
 288         /*
 289          * only really works for sequential io for now, and with 1 file
 290          */
 291         if (o->zone_size && td_random(td) && o->open_files == 1)
 292                 o->zone_size = 0;
 293
 294         /*
 295          * Reads can do overwrites, we always need to pre-create the file
 296          */
 297         if (td_read(td) || td_rw(td))
 298                 o->overwrite = 1;
 299
 300         if (!o->min_bs[DDIR_READ])
 301                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
 302         if (!o->max_bs[DDIR_READ])
 303                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
 304         if (!o->min_bs[DDIR_WRITE])
 305                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 306         if (!o->max_bs[DDIR_WRITE])
 307                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 308
 309         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
 310
 311         /*
 312          * For random IO, allow blockalign offset other than min_bs.
 313          */
 314         if (!o->ba[DDIR_READ] || !td_random(td))
 315                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
 316         if (!o->ba[DDIR_WRITE] || !td_random(td))
 317                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
 318
 319         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
 320             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
 321             !o->norandommap) {
 322                 log_err("fio: Any use of blockalign= turns off randommap\n");
 323                 o->norandommap = 1;
 324                 ret = warnings_fatal;
 325         }
 326
 327         if (!o->file_size_high)
 328                 o->file_size_high = o->file_size_low;
 329
 330         if (o->norandommap && o->verify != VERIFY_NONE
 331             && !fixed_block_size(o))  {
 332                 log_err("fio: norandommap given for variable block sizes, "
 333                         "verify disabled\n");
 334                 o->verify = VERIFY_NONE;
 335                 ret = warnings_fatal;
 336         }
 337         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
 338                 log_err("fio: bs_unaligned may not work with raw io\n");
 339
 340         /*
 341          * thinktime_spin must be less than thinktime
 342          */
 343         if (o->thinktime_spin > o->thinktime)
 344                 o->thinktime_spin = o->thinktime;
 345
 346         /*
 347          * The low water mark cannot be bigger than the iodepth
 348          */
 349         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
 350                 /*
 351                  * syslet work around - if the workload is sequential,
 352                  * we want to let the queue drain all the way down to
 353                  * avoid seeking between async threads
 354                  */
 355                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 356                         o->iodepth_low = 1;
 357                 else
 358                         o->iodepth_low = o->iodepth;
 359         }
 360
 361         /*
 362          * If batch number isn't set, default to the same as iodepth
 363          */
 364         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
 365                 o->iodepth_batch = o->iodepth;
 366
 367         if (o->nr_files > td->files_index)
 368                 o->nr_files = td->files_index;
 369
 370         if (o->open_files > o->nr_files || !o->open_files)
 371                 o->open_files = o->nr_files;
 372
 373         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
 374             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
 375                 o->rate_iops_min[1]))) {
 376                 log_err("fio: rate and rate_iops are mutually exclusive\n");
 377                 ret = 1;
 378         }
 379         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
 380             (o->rate_iops[0] < o->rate_iops_min[0]) ||
 381             (o->rate_iops[1] < o->rate_iops_min[1])) {
 382                 log_err("fio: minimum rate exceeds rate\n");
 383                 ret = 1;
 384         }
 385
 386         if (!o->timeout && o->time_based) {
 387                 log_err("fio: time_based requires a runtime/timeout setting\n");
 388                 o->time_based = 0;
 389                 ret = warnings_fatal;
 390         }
 391
 392         if (o->fill_device && !o->size)
 393                 o->size = -1ULL;
 394
 395         if (o->verify != VERIFY_NONE) {
 396                 if (td_write(td) && o->do_verify && o->numjobs > 1) {
 397                         log_info("Multiple writers may overwrite blocks that "
 398                                 "belong to other jobs. This can cause "
 399                                 "verification failures.\n");
 400                         ret = warnings_fatal;
 401                 }
 402
 403                 o->refill_buffers = 1;
 404                 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
 405                     !o->verify_interval)
 406                         o->verify_interval = o->min_bs[DDIR_WRITE];
 407         }
 408
 409         if (o->pre_read) {
 410                 o->invalidate_cache = 0;
 411                 if (td->io_ops->flags & FIO_PIPEIO) {
 412                         log_info("fio: cannot pre-read files with an IO engine"
 413                                  " that isn't seekable. Pre-read disabled.\n");
 414                         ret = warnings_fatal;
 415                 }
 416         }
 417
 418 #ifndef FIO_HAVE_FDATASYNC
 419         if (o->fdatasync_blocks) {
 420                 log_info("fio: this platform does not support fdatasync()"
 421                          " falling back to using fsync().  Use the 'fsync'"
 422                          " option instead of 'fdatasync' to get rid of"
 423                          " this warning\n");
 424                 o->fsync_blocks = o->fdatasync_blocks;
 425                 o->fdatasync_blocks = 0;
 426                 ret = warnings_fatal;
 427         }
 428 #endif
 429
 430         return ret;
 431 }
 432
 433 /*
 434  * This function leaks the buffer
 435  */
 436 static char *to_kmg(unsigned int val)
 437 {
 438         char *buf = malloc(32);
 439         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 440         char *p = post;
 441
 442         do {
 443                 if (val & 1023)
 444                         break;
 445
 446                 val >>= 10;
 447                 p++;
 448         } while (*p);
 449
 450         snprintf(buf, 31, "%u%c", val, *p);
 451         return buf;
 452 }
 453
 454 /* External engines are specified by "external:name.o") */
 455 static const char *get_engine_name(const char *str)
 456 {
 457         char *p = strstr(str, ":");
 458
 459         if (!p)
 460                 return str;
 461
 462         p++;
 463         strip_blank_front(&p);
 464         strip_blank_end(p);
 465         return p;
 466 }
 467
 468 static int exists_and_not_file(const char *filename)
 469 {
 470         struct stat sb;
 471
 472         if (lstat(filename, &sb) == -1)
 473                 return 0;
 474
 475         /* \\.\ is the device namespace in Windows, where every file
 476          * is a device node */
 477         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
 478                 return 0;
 479
 480         return 1;
 481 }
 482
 483 static void td_fill_rand_seeds_os(struct thread_data *td)
 484 {
 485         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
 486         os_random_seed(td->rand_seeds[1], &td->verify_state);
 487         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
 488
 489         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
 490                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
 491
 492         os_random_seed(td->rand_seeds[5], &td->file_size_state);
 493         os_random_seed(td->rand_seeds[6], &td->trim_state);
 494
 495         if (!td_random(td))
 496                 return;
 497
 498         if (td->o.rand_repeatable)
 499                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
 500
 501         os_random_seed(td->rand_seeds[4], &td->random_state);
 502 }
 503
 504 static void td_fill_rand_seeds_internal(struct thread_data *td)
 505 {
 506         init_rand_seed(&td->__bsrange_state, td->rand_seeds[0]);
 507         init_rand_seed(&td->__verify_state, td->rand_seeds[1]);
 508         init_rand_seed(&td->__rwmix_state, td->rand_seeds[2]);
 509
 510         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
 511                 init_rand_seed(&td->__next_file_state, td->rand_seeds[3]);
 512
 513         init_rand_seed(&td->__file_size_state, td->rand_seeds[5]);
 514         init_rand_seed(&td->__trim_state, td->rand_seeds[6]);
 515
 516         if (!td_random(td))
 517                 return;
 518
 519         if (td->o.rand_repeatable)
 520                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
 521
 522         init_rand_seed(&td->__random_state, td->rand_seeds[4]);
 523 }
 524
 525 void td_fill_rand_seeds(struct thread_data *td)
 526 {
 527         if (td->o.use_os_rand)
 528                 td_fill_rand_seeds_os(td);
 529         else
 530                 td_fill_rand_seeds_internal(td);
 531
 532         init_rand_seed(&td->buf_state, td->rand_seeds[7]);
 533 }
 534
 535 /*
 536  * Initialize the various random states we need (random io, block size ranges,
 537  * read/write mix, etc).
 538  */
 539 static int init_random_state(struct thread_data *td)
 540 {
 541         int fd;
 542
 543         fd = open("/dev/urandom", O_RDONLY);
 544         if (fd == -1) {
 545                 td_verror(td, errno, "open");
 546                 return 1;
 547         }
 548
 549         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
 550             (int) sizeof(td->rand_seeds)) {
 551                 td_verror(td, EIO, "read");
 552                 close(fd);
 553                 return 1;
 554         }
 555
 556         close(fd);
 557         td_fill_rand_seeds(td);
 558         return 0;
 559 }
 560
 561 /*
 562  * Adds a job to the list of things todo. Sanitizes the various options
 563  * to make sure we don't have conflicts, and initializes various
 564  * members of td.
 565  */
 566 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 567 {
 568         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 569                                    "randread", "randwrite", "randrw" };
 570         unsigned int i;
 571         const char *engine;
 572         char fname[PATH_MAX];
 573         int numjobs, file_alloced;
 574
 575         /*
 576          * the def_thread is just for options, it's not a real job
 577          */
 578         if (td == &def_thread)
 579                 return 0;
 580
 581         /*
 582          * if we are just dumping the output command line, don't add the job
 583          */
 584         if (dump_cmdline) {
 585                 put_job(td);
 586                 return 0;
 587         }
 588
 589         if (profile_td_init(td))
 590                 goto err;
 591
 592         engine = get_engine_name(td->o.ioengine);
 593         td->io_ops = load_ioengine(td, engine);
 594         if (!td->io_ops) {
 595                 log_err("fio: failed to load engine %s\n", engine);
 596                 goto err;
 597         }
 598
 599         if (td->o.use_thread)
 600                 nr_thread++;
 601         else
 602                 nr_process++;
 603
 604         if (td->o.odirect)
 605                 td->io_ops->flags |= FIO_RAWIO;
 606
 607         file_alloced = 0;
 608         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
 609                 file_alloced = 1;
 610
 611                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
 612                         add_file(td, jobname);
 613                 else {
 614                         for (i = 0; i < td->o.nr_files; i++) {
 615                                 sprintf(fname, "%s.%d.%d", jobname,
 616                                                         td->thread_number, i);
 617                                 add_file(td, fname);
 618                         }
 619                 }
 620         }
 621
 622         if (fixup_options(td))
 623                 goto err;
 624
 625         if (td->io_ops->flags & FIO_DISKLESSIO) {
 626                 struct fio_file *f;
 627
 628                 for_each_file(td, f, i)
 629                         f->real_file_size = -1ULL;
 630         }
 631
 632         td->mutex = fio_mutex_init(0);
 633
 634         td->ts.clat_percentiles = td->o.clat_percentiles;
 635         if (td->o.overwrite_plist)
 636                 td->ts.percentile_list = td->o.percentile_list;
 637         else
 638                 td->ts.percentile_list = NULL;
 639
 640         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 641         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 642         td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
 643         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 644         td->ddir_seq_nr = td->o.ddir_seq_nr;
 645
 646         if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
 647                 prev_group_jobs = 0;
 648                 groupid++;
 649         }
 650
 651         td->groupid = groupid;
 652         prev_group_jobs++;
 653
 654         if (init_random_state(td))
 655                 goto err;
 656
 657         if (setup_rate(td))
 658                 goto err;
 659
 660         if (td->o.write_lat_log) {
 661                 setup_log(&td->ts.lat_log);
 662                 setup_log(&td->ts.slat_log);
 663                 setup_log(&td->ts.clat_log);
 664         }
 665         if (td->o.write_bw_log)
 666                 setup_log(&td->ts.bw_log);
 667
 668         if (!td->o.name)
 669                 td->o.name = strdup(jobname);
 670
 671         if (!terse_output) {
 672                 if (!job_add_num) {
 673                         if (!strcmp(td->io_ops->name, "cpuio")) {
 674                                 log_info("%s: ioengine=cpu, cpuload=%u,"
 675                                          " cpucycle=%u\n", td->o.name,
 676                                                         td->o.cpuload,
 677                                                         td->o.cpucycle);
 678                         } else {
 679                                 char *c1, *c2, *c3, *c4;
 680
 681                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
 682                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
 683                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
 684                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
 685
 686                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
 687                                          " ioengine=%s, iodepth=%u\n",
 688                                                 td->o.name, td->groupid,
 689                                                 ddir_str[td->o.td_ddir],
 690                                                 c1, c2, c3, c4,
 691                                                 td->io_ops->name,
 692                                                 td->o.iodepth);
 693
 694                                 free(c1);
 695                                 free(c2);
 696                                 free(c3);
 697                                 free(c4);
 698                         }
 699                 } else if (job_add_num == 1)
 700                         log_info("...\n");
 701         }
 702
 703         /*
 704          * recurse add identical jobs, clear numjobs and stonewall options
 705          * as they don't apply to sub-jobs
 706          */
 707         numjobs = td->o.numjobs;
 708         while (--numjobs) {
 709                 struct thread_data *td_new = get_new_job(0, td);
 710
 711                 if (!td_new)
 712                         goto err;
 713
 714                 td_new->o.numjobs = 1;
 715                 td_new->o.stonewall = 0;
 716                 td_new->o.new_group = 0;
 717
 718                 if (file_alloced) {
 719                         td_new->o.filename = NULL;
 720                         td_new->files_index = 0;
 721                         td_new->files_size = 0;
 722                         td_new->files = NULL;
 723                 }
 724
 725                 job_add_num = numjobs - 1;
 726
 727                 if (add_job(td_new, jobname, job_add_num))
 728                         goto err;
 729         }
 730
 731         return 0;
 732 err:
 733         put_job(td);
 734         return -1;
 735 }
 736
 737 /*
 738  * Parse as if 'o' was a command line
 739  */
 740 void add_job_opts(const char **o)
 741 {
 742         struct thread_data *td, *td_parent;
 743         int i, in_global = 1;
 744         char jobname[32];
 745
 746         i = 0;
 747         td_parent = td = NULL;
 748         while (o[i]) {
 749                 if (!strncmp(o[i], "name", 4)) {
 750                         in_global = 0;
 751                         if (td)
 752                                 add_job(td, jobname, 0);
 753                         td = NULL;
 754                         sprintf(jobname, "%s", o[i] + 5);
 755                 }
 756                 if (in_global && !td_parent)
 757                         td_parent = get_new_job(1, &def_thread);
 758                 else if (!in_global && !td) {
 759                         if (!td_parent)
 760                                 td_parent = &def_thread;
 761                         td = get_new_job(0, td_parent);
 762                 }
 763                 if (in_global)
 764                         fio_options_parse(td_parent, (char **) &o[i], 1);
 765                 else
 766                         fio_options_parse(td, (char **) &o[i], 1);
 767                 i++;
 768         }
 769
 770         if (td)
 771                 add_job(td, jobname, 0);
 772 }
 773
 774 static int skip_this_section(const char *name)
 775 {
 776         int i;
 777
 778         if (!nr_job_sections)
 779                 return 0;
 780         if (!strncmp(name, "global", 6))
 781                 return 0;
 782
 783         for (i = 0; i < nr_job_sections; i++)
 784                 if (!strcmp(job_sections[i], name))
 785                         return 0;
 786
 787         return 1;
 788 }
 789
 790 static int is_empty_or_comment(char *line)
 791 {
 792         unsigned int i;
 793
 794         for (i = 0; i < strlen(line); i++) {
 795                 if (line[i] == ';')
 796                         return 1;
 797                 if (line[i] == '#')
 798                         return 1;
 799                 if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
 800                         return 0;
 801         }
 802
 803         return 1;
 804 }
 805
 806 /*
 807  * This is our [ini] type file parser.
 808  */
 809 int parse_jobs_ini(char *file, int is_buf, int stonewall_flag)
 810 {
 811         unsigned int global;
 812         struct thread_data *td;
 813         char *string, *name;
 814         FILE *f;
 815         char *p;
 816         int ret = 0, stonewall;
 817         int first_sect = 1;
 818         int skip_fgets = 0;
 819         int inside_skip = 0;
 820         char **opts;
 821         int i, alloc_opts, num_opts;
 822
 823         if (is_buf)
 824                 f = NULL;
 825         else {
 826                 if (!strcmp(file, "-"))
 827                         f = stdin;
 828                 else
 829                         f = fopen(file, "r");
 830
 831                 if (!f) {
 832                         perror("fopen job file");
 833                         return 1;
 834                 }
 835         }
 836
 837         string = malloc(4096);
 838
 839         /*
 840          * it's really 256 + small bit, 280 should suffice
 841          */
 842         name = malloc(280);
 843         memset(name, 0, 280);
 844
 845         alloc_opts = 8;
 846         opts = malloc(sizeof(char *) * alloc_opts);
 847         num_opts = 0;
 848
 849         stonewall = stonewall_flag;
 850         do {
 851                 /*
 852                  * if skip_fgets is set, we already have loaded a line we
 853                  * haven't handled.
 854                  */
 855                 if (!skip_fgets) {
 856                         if (is_buf)
 857                                 p = strsep(&file, "\n");
 858                         else
 859                                 p = fgets(string, 4095, f);
 860                         if (!p)
 861                                 break;
 862                 }
 863
 864                 skip_fgets = 0;
 865                 strip_blank_front(&p);
 866                 strip_blank_end(p);
 867
 868                 if (is_empty_or_comment(p))
 869                         continue;
 870                 if (sscanf(p, "[%255[^\n]]", name) != 1) {
 871                         if (inside_skip)
 872                                 continue;
 873                         log_err("fio: option <%s> outside of [] job section\n",
 874                                                                         p);
 875                         break;
 876                 }
 877
 878                 name[strlen(name) - 1] = '\0';
 879
 880                 if (skip_this_section(name)) {
 881                         inside_skip = 1;
 882                         continue;
 883                 } else
 884                         inside_skip = 0;
 885
 886                 global = !strncmp(name, "global", 6);
 887
 888                 if (dump_cmdline) {
 889                         if (first_sect)
 890                                 log_info("fio ");
 891                         if (!global)
 892                                 log_info("--name=%s ", name);
 893                         first_sect = 0;
 894                 }
 895
 896                 td = get_new_job(global, &def_thread);
 897                 if (!td) {
 898                         ret = 1;
 899                         break;
 900                 }
 901
 902                 /*
 903                  * Seperate multiple job files by a stonewall
 904                  */
 905                 if (!global && stonewall) {
 906                         td->o.stonewall = stonewall;
 907                         stonewall = 0;
 908                 }
 909
 910                 num_opts = 0;
 911                 memset(opts, 0, alloc_opts * sizeof(char *));
 912
 913                 while (1) {
 914                         if (is_buf)
 915                                 p = strsep(&file, "\n");
 916                         else
 917                                 p = fgets(string, 4096, f);
 918                         if (!p)
 919                                 break;
 920
 921                         if (is_empty_or_comment(p))
 922                                 continue;
 923
 924                         strip_blank_front(&p);
 925
 926                         /*
 927                          * new section, break out and make sure we don't
 928                          * fgets() a new line at the top.
 929                          */
 930                         if (p[0] == '[') {
 931                                 skip_fgets = 1;
 932                                 break;
 933                         }
 934
 935                         strip_blank_end(p);
 936
 937                         if (num_opts == alloc_opts) {
 938                                 alloc_opts <<= 1;
 939                                 opts = realloc(opts,
 940                                                 alloc_opts * sizeof(char *));
 941                         }
 942
 943                         opts[num_opts] = strdup(p);
 944                         num_opts++;
 945                 }
 946
 947                 ret = fio_options_parse(td, opts, num_opts);
 948                 if (!ret) {
 949                         if (dump_cmdline)
 950                                 for (i = 0; i < num_opts; i++)
 951                                         log_info("--%s ", opts[i]);
 952
 953                         ret = add_job(td, name, 0);
 954                 } else {
 955                         log_err("fio: job %s dropped\n", name);
 956                         put_job(td);
 957                 }
 958
 959                 for (i = 0; i < num_opts; i++)
 960                         free(opts[i]);
 961                 num_opts = 0;
 962         } while (!ret);
 963
 964         if (dump_cmdline)
 965                 log_info("\n");
 966
 967         for (i = 0; i < num_opts; i++)
 968                 free(opts[i]);
 969
 970         free(string);
 971         free(name);
 972         free(opts);
 973         if (!is_buf && f != stdin)
 974                 fclose(f);
 975         return ret;
 976 }
 977
 978 static int fill_def_thread(void)
 979 {
 980         memset(&def_thread, 0, sizeof(def_thread));
 981
 982         fio_getaffinity(getpid(), &def_thread.o.cpumask);
 983
 984         /*
 985          * fill default options
 986          */
 987         fio_fill_default_options(&def_thread);
 988         return 0;
 989 }
 990
 991 static void free_shm(void)
 992 {
 993         struct shmid_ds sbuf;
 994
 995         if (threads) {
 996                 void *tp = threads;
 997
 998                 threads = NULL;
 999                 file_hash_exit();
1000                 fio_debug_jobp = NULL;
1001                 shmdt(tp);
1002                 shmctl(shm_id, IPC_RMID, &sbuf);
1003         }
1004
1005         scleanup();
1006 }
1007
1008 /*
1009  * The thread area is shared between the main process and the job
1010  * threads/processes. So setup a shared memory segment that will hold
1011  * all the job info. We use the end of the region for keeping track of
1012  * open files across jobs, for file sharing.
1013  */
1014 static int setup_thread_area(void)
1015 {
1016         void *hash;
1017
1018         /*
1019          * 1024 is too much on some machines, scale max_jobs if
1020          * we get a failure that looks like too large a shm segment
1021          */
1022         do {
1023                 size_t size = max_jobs * sizeof(struct thread_data);
1024
1025                 size += file_hash_size;
1026                 size += sizeof(unsigned int);
1027
1028                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1029                 if (shm_id != -1)
1030                         break;
1031                 if (errno != EINVAL) {
1032                         perror("shmget");
1033                         break;
1034                 }
1035
1036                 max_jobs >>= 1;
1037         } while (max_jobs);
1038
1039         if (shm_id == -1)
1040                 return 1;
1041
1042         threads = shmat(shm_id, NULL, 0);
1043         if (threads == (void *) -1) {
1044                 perror("shmat");
1045                 return 1;
1046         }
1047
1048         memset(threads, 0, max_jobs * sizeof(struct thread_data));
1049         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
1050         fio_debug_jobp = (void *) hash + file_hash_size;
1051         *fio_debug_jobp = -1;
1052         file_hash_init(hash);
1053         atexit(free_shm);
1054         return 0;
1055 }
1056
1057 static void usage(const char *name)
1058 {
1059         printf("%s\n", fio_version_string);
1060         printf("%s [options] [job options] <job file(s)>\n", name);
1061         printf("\t--debug=options\tEnable debug logging\n");
1062         printf("\t--output\tWrite output to file\n");
1063         printf("\t--timeout\tRuntime in seconds\n");
1064         printf("\t--latency-log\tGenerate per-job latency logs\n");
1065         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1066         printf("\t--minimal\tMinimal (terse) output\n");
1067         printf("\t--version\tPrint version info and exit\n");
1068         printf("\t--terse-version=x Terse version output format\n");
1069         printf("\t--help\t\tPrint this page\n");
1070         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
1071                 " them\n");
1072         printf("\t--showcmd\tTurn a job file into command line options\n");
1073         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
1074         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
1075         printf("\t--readonly\tTurn on safety read-only checks, preventing"
1076                 " writes\n");
1077         printf("\t--section=name\tOnly run specified section in job file\n");
1078         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
1079                 " (def 1024)\n");
1080         printf("\t--warnings-fatal Fio parser warnings are fatal\n");
1081         printf("\t--max-jobs\tMaximum number of threads/processes to support\n");
1082         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1083         printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
1084 }
1085
1086 #ifdef FIO_INC_DEBUG
1087 struct debug_level debug_levels[] = {
1088         { .name = "process",    .shift = FD_PROCESS, },
1089         { .name = "file",       .shift = FD_FILE, },
1090         { .name = "io",         .shift = FD_IO, },
1091         { .name = "mem",        .shift = FD_MEM, },
1092         { .name = "blktrace",   .shift = FD_BLKTRACE },
1093         { .name = "verify",     .shift = FD_VERIFY },
1094         { .name = "random",     .shift = FD_RANDOM },
1095         { .name = "parse",      .shift = FD_PARSE },
1096         { .name = "diskutil",   .shift = FD_DISKUTIL },
1097         { .name = "job",        .shift = FD_JOB },
1098         { .name = "mutex",      .shift = FD_MUTEX },
1099         { .name = "profile",    .shift = FD_PROFILE },
1100         { .name = "time",       .shift = FD_TIME },
1101         { .name = NULL, },
1102 };
1103
1104 static int set_debug(const char *string)
1105 {
1106         struct debug_level *dl;
1107         char *p = (char *) string;
1108         char *opt;
1109         int i;
1110
1111         if (!strcmp(string, "?") || !strcmp(string, "help")) {
1112                 log_info("fio: dumping debug options:");
1113                 for (i = 0; debug_levels[i].name; i++) {
1114                         dl = &debug_levels[i];
1115                         log_info("%s,", dl->name);
1116                 }
1117                 log_info("all\n");
1118                 return 1;
1119         }
1120
1121         while ((opt = strsep(&p, ",")) != NULL) {
1122                 int found = 0;
1123
1124                 if (!strncmp(opt, "all", 3)) {
1125                         log_info("fio: set all debug options\n");
1126                         fio_debug = ~0UL;
1127                         continue;
1128                 }
1129
1130                 for (i = 0; debug_levels[i].name; i++) {
1131                         dl = &debug_levels[i];
1132                         found = !strncmp(opt, dl->name, strlen(dl->name));
1133                         if (!found)
1134                                 continue;
1135
1136                         if (dl->shift == FD_JOB) {
1137                                 opt = strchr(opt, ':');
1138                                 if (!opt) {
1139                                         log_err("fio: missing job number\n");
1140                                         break;
1141                                 }
1142                                 opt++;
1143                                 fio_debug_jobno = atoi(opt);
1144                                 log_info("fio: set debug jobno %d\n",
1145                                                         fio_debug_jobno);
1146                         } else {
1147                                 log_info("fio: set debug option %s\n", opt);
1148                                 fio_debug |= (1UL << dl->shift);
1149                         }
1150                         break;
1151                 }
1152
1153                 if (!found)
1154                         log_err("fio: debug mask %s not found\n", opt);
1155         }
1156         return 0;
1157 }
1158 #else
1159 static int set_debug(const char *string)
1160 {
1161         log_err("fio: debug tracing not included in build\n");
1162         return 1;
1163 }
1164 #endif
1165
1166 static void fio_options_fill_optstring(void)
1167 {
1168         char *ostr = cmd_optstr;
1169         int i, c;
1170
1171         c = i = 0;
1172         while (l_opts[i].name) {
1173                 ostr[c++] = l_opts[i].val;
1174                 if (l_opts[i].has_arg == required_argument)
1175                         ostr[c++] = ':';
1176                 else if (l_opts[i].has_arg == optional_argument) {
1177                         ostr[c++] = ':';
1178                         ostr[c++] = ':';
1179                 }
1180                 i++;
1181         }
1182         ostr[c] = '\0';
1183 }
1184
1185 static int parse_cmd_line(int argc, char *argv[])
1186 {
1187         struct thread_data *td = NULL;
1188         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1189         char *ostr = cmd_optstr;
1190
1191         while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1192                 switch (c) {
1193                 case 'a':
1194                         smalloc_pool_size = atoi(optarg);
1195                         break;
1196                 case 't':
1197                         def_thread.o.timeout = atoi(optarg);
1198                         break;
1199                 case 'l':
1200                         write_lat_log = 1;
1201                         break;
1202                 case 'b':
1203                         write_bw_log = 1;
1204                         break;
1205                 case 'o':
1206                         f_out = fopen(optarg, "w+");
1207                         if (!f_out) {
1208                                 perror("fopen output");
1209                                 exit(1);
1210                         }
1211                         f_err = f_out;
1212                         break;
1213                 case 'm':
1214                         terse_output = 1;
1215                         break;
1216                 case 'h':
1217                         usage(argv[0]);
1218                         exit(0);
1219                 case 'c':
1220                         exit(fio_show_option_help(optarg));
1221                 case 's':
1222                         dump_cmdline = 1;
1223                         break;
1224                 case 'r':
1225                         read_only = 1;
1226                         break;
1227                 case 'v':
1228                         log_info("%s\n", fio_version_string);
1229                         exit(0);
1230                 case 'V':
1231                         terse_version = atoi(optarg);
1232                         if (terse_version != 2) {
1233                                 log_err("fio: bad terse version format\n");
1234                                 exit_val = 1;
1235                                 do_exit++;
1236                         }
1237                         break;
1238                 case 'e':
1239                         if (!strcmp("always", optarg))
1240                                 eta_print = FIO_ETA_ALWAYS;
1241                         else if (!strcmp("never", optarg))
1242                                 eta_print = FIO_ETA_NEVER;
1243                         break;
1244                 case 'd':
1245                         if (set_debug(optarg))
1246                                 do_exit++;
1247                         break;
1248                 case 'x': {
1249                         size_t new_size;
1250
1251                         if (!strcmp(optarg, "global")) {
1252                                 log_err("fio: can't use global as only "
1253                                         "section\n");
1254                                 do_exit++;
1255                                 exit_val = 1;
1256                                 break;
1257                         }
1258                         new_size = (nr_job_sections + 1) * sizeof(char *);
1259                         job_sections = realloc(job_sections, new_size);
1260                         job_sections[nr_job_sections] = strdup(optarg);
1261                         nr_job_sections++;
1262                         break;
1263                         }
1264                 case 'p':
1265                         exec_profile = strdup(optarg);
1266                         break;
1267                 case FIO_GETOPT_JOB: {
1268                         const char *opt = l_opts[lidx].name;
1269                         char *val = optarg;
1270
1271                         if (!strncmp(opt, "name", 4) && td) {
1272                                 ret = add_job(td, td->o.name ?: "fio", 0);
1273                                 if (ret)
1274                                         return 0;
1275                                 td = NULL;
1276                         }
1277                         if (!td) {
1278                                 int is_section = !strncmp(opt, "name", 4);
1279                                 int global = 0;
1280
1281                                 if (!is_section || !strncmp(val, "global", 6))
1282                                         global = 1;
1283
1284                                 if (is_section && skip_this_section(val))
1285                                         continue;
1286
1287                                 td = get_new_job(global, &def_thread);
1288                                 if (!td)
1289                                         return 0;
1290                         }
1291
1292                         ret = fio_cmd_option_parse(td, opt, val);
1293                         break;
1294                 }
1295                 case 'w':
1296                         warnings_fatal = 1;
1297                         break;
1298                 case 'j':
1299                         max_jobs = atoi(optarg);
1300                         if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1301                                 log_err("fio: invalid max jobs: %d\n", max_jobs);
1302                                 do_exit++;
1303                                 exit_val = 1;
1304                         }
1305                         break;
1306                 case 'S':
1307                         is_backend = 1;
1308                         break;
1309                 default:
1310                         do_exit++;
1311                         exit_val = 1;
1312                         break;
1313                 }
1314         }
1315
1316         if (do_exit)
1317                 exit(exit_val);
1318
1319         if (is_backend)
1320                 return fio_server();
1321
1322         if (td) {
1323                 if (!ret)
1324                         ret = add_job(td, td->o.name ?: "fio", 0);
1325         }
1326
1327         while (optind < argc) {
1328                 ini_idx++;
1329                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1330                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1331                 optind++;
1332         }
1333
1334         return ini_idx;
1335 }
1336
1337 int parse_options(int argc, char *argv[])
1338 {
1339         int job_files, i;
1340
1341         f_out = stdout;
1342         f_err = stderr;
1343
1344         fio_options_fill_optstring();
1345         fio_options_dup_and_init(l_opts);
1346
1347         if (setup_thread_area())
1348                 return 1;
1349         if (fill_def_thread())
1350                 return 1;
1351
1352         job_files = parse_cmd_line(argc, argv);
1353
1354         for (i = 0; i < job_files; i++) {
1355                 if (fill_def_thread())
1356                         return 1;
1357                 if (parse_jobs_ini(ini_file[i], 0, i))
1358                         return 1;
1359                 free(ini_file[i]);
1360         }
1361
1362         free(ini_file);
1363         options_mem_free(&def_thread);
1364
1365         if (!thread_number) {
1366                 if (dump_cmdline)
1367                         return 0;
1368                 if (exec_profile)
1369                         return 0;
1370                 if (is_backend)
1371                         return 0;
1372
1373                 log_err("No jobs(s) defined\n\n");
1374                 usage(argv[0]);
1375                 return 1;
1376         }
1377
1378         if (def_thread.o.gtod_offload) {
1379                 fio_gtod_init();
1380                 fio_gtod_offload = 1;
1381                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1382         }
1383
1384         log_info("%s\n", fio_version_string);
1385         return 0;
1386 }