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 #if FIO_PATCH > 0
  27 static char fio_version_string[] =      __fio_stringify(FIO_MAJOR) "."  \
  28                                         __fio_stringify(FIO_MINOR) "."  \
  29                                         __fio_stringify(FIO_PATCH);
  30 #else
  31 static char fio_version_string[] =      __fio_stringify(FIO_MAJOR) "."  \
  32                                         __fio_stringify(FIO_MINOR);
  33 #endif
  34
  35 #define FIO_RANDSEED            (0xb1899bedUL)
  36
  37 static char **ini_file;
  38 static int max_jobs = FIO_MAX_JOBS;
  39 static int dump_cmdline;
  40
  41 static struct thread_data def_thread;
  42 struct thread_data *threads = NULL;
  43
  44 int exitall_on_terminate = 0;
  45 int terse_output = 0;
  46 int eta_print;
  47 unsigned long long mlock_size = 0;
  48 FILE *f_out = NULL;
  49 FILE *f_err = NULL;
  50 char **job_sections = NULL;
  51 int nr_job_sections = 0;
  52 char *exec_profile = NULL;
  53 int warnings_fatal = 0;
  54 int terse_version = 2;
  55 int is_backend = 0;
  56 int nr_clients = 0;
  57 int log_syslog = 0;
  58
  59 int write_bw_log = 0;
  60 int read_only = 0;
  61
  62 static int write_lat_log;
  63
  64 static int prev_group_jobs;
  65
  66 unsigned long fio_debug = 0;
  67 unsigned int fio_debug_jobno = -1;
  68 unsigned int *fio_debug_jobp = NULL;
  69
  70 static char cmd_optstr[256];
  71
  72 /*
  73  * Command line options. These will contain the above, plus a few
  74  * extra that only pertain to fio itself and not jobs.
  75  */
  76 static struct option l_opts[FIO_NR_OPTIONS] = {
  77         {
  78                 .name           = (char *) "output",
  79                 .has_arg        = required_argument,
  80                 .val            = 'o',
  81         },
  82         {
  83                 .name           = (char *) "timeout",
  84                 .has_arg        = required_argument,
  85                 .val            = 't',
  86         },
  87         {
  88                 .name           = (char *) "latency-log",
  89                 .has_arg        = required_argument,
  90                 .val            = 'l',
  91         },
  92         {
  93                 .name           = (char *) "bandwidth-log",
  94                 .has_arg        = required_argument,
  95                 .val            = 'b',
  96         },
  97         {
  98                 .name           = (char *) "minimal",
  99                 .has_arg        = optional_argument,
 100                 .val            = 'm',
 101         },
 102         {
 103                 .name           = (char *) "version",
 104                 .has_arg        = no_argument,
 105                 .val            = 'v',
 106         },
 107         {
 108                 .name           = (char *) "help",
 109                 .has_arg        = no_argument,
 110                 .val            = 'h',
 111         },
 112         {
 113                 .name           = (char *) "cmdhelp",
 114                 .has_arg        = optional_argument,
 115                 .val            = 'c',
 116         },
 117         {
 118                 .name           = (char *) "showcmd",
 119                 .has_arg        = no_argument,
 120                 .val            = 's',
 121         },
 122         {
 123                 .name           = (char *) "readonly",
 124                 .has_arg        = no_argument,
 125                 .val            = 'r',
 126         },
 127         {
 128                 .name           = (char *) "eta",
 129                 .has_arg        = required_argument,
 130                 .val            = 'e',
 131         },
 132         {
 133                 .name           = (char *) "debug",
 134                 .has_arg        = required_argument,
 135                 .val            = 'd',
 136         },
 137         {
 138                 .name           = (char *) "section",
 139                 .has_arg        = required_argument,
 140                 .val            = 'x',
 141         },
 142         {
 143                 .name           = (char *) "alloc-size",
 144                 .has_arg        = required_argument,
 145                 .val            = 'a',
 146         },
 147         {
 148                 .name           = (char *) "profile",
 149                 .has_arg        = required_argument,
 150                 .val            = 'p',
 151         },
 152         {
 153                 .name           = (char *) "warnings-fatal",
 154                 .has_arg        = no_argument,
 155                 .val            = 'w',
 156         },
 157         {
 158                 .name           = (char *) "max-jobs",
 159                 .has_arg        = required_argument,
 160                 .val            = 'j',
 161         },
 162         {
 163                 .name           = (char *) "terse-version",
 164                 .has_arg        = required_argument,
 165                 .val            = 'V',
 166         },
 167         {
 168                 .name           = (char *) "server",
 169                 .has_arg        = no_argument,
 170                 .val            = 'S',
 171         },
 172         {       .name           = (char *) "daemonize",
 173                 .has_arg        = no_argument,
 174                 .val            = 'D',
 175         },
 176         {
 177                 .name           = (char *) "net-port",
 178                 .has_arg        = required_argument,
 179                 .val            = 'P',
 180         },
 181         {
 182                 .name           = (char *) "client",
 183                 .has_arg        = required_argument,
 184                 .val            = 'C',
 185         },
 186         {
 187                 .name           = NULL,
 188         },
 189 };
 190
 191 static void free_shm(void)
 192 {
 193         struct shmid_ds sbuf;
 194
 195         if (threads) {
 196                 void *tp = threads;
 197
 198                 threads = NULL;
 199                 file_hash_exit();
 200                 fio_debug_jobp = NULL;
 201                 shmdt(tp);
 202                 shmctl(shm_id, IPC_RMID, &sbuf);
 203         }
 204
 205         scleanup();
 206 }
 207
 208 /*
 209  * The thread area is shared between the main process and the job
 210  * threads/processes. So setup a shared memory segment that will hold
 211  * all the job info. We use the end of the region for keeping track of
 212  * open files across jobs, for file sharing.
 213  */
 214 static int setup_thread_area(void)
 215 {
 216         void *hash;
 217
 218         if (threads)
 219                 return 0;
 220
 221         /*
 222          * 1024 is too much on some machines, scale max_jobs if
 223          * we get a failure that looks like too large a shm segment
 224          */
 225         do {
 226                 size_t size = max_jobs * sizeof(struct thread_data);
 227
 228                 size += file_hash_size;
 229                 size += sizeof(unsigned int);
 230
 231                 shm_id = shmget(0, size, IPC_CREAT | 0600);
 232                 if (shm_id != -1)
 233                         break;
 234                 if (errno != EINVAL) {
 235                         perror("shmget");
 236                         break;
 237                 }
 238
 239                 max_jobs >>= 1;
 240         } while (max_jobs);
 241
 242         if (shm_id == -1)
 243                 return 1;
 244
 245         threads = shmat(shm_id, NULL, 0);
 246         if (threads == (void *) -1) {
 247                 perror("shmat");
 248                 return 1;
 249         }
 250
 251         memset(threads, 0, max_jobs * sizeof(struct thread_data));
 252         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
 253         fio_debug_jobp = (void *) hash + file_hash_size;
 254         *fio_debug_jobp = -1;
 255         file_hash_init(hash);
 256         return 0;
 257 }
 258
 259 /*
 260  * Return a free job structure.
 261  */
 262 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 263 {
 264         struct thread_data *td;
 265
 266         if (global)
 267                 return &def_thread;
 268         if (setup_thread_area()) {
 269                 log_err("error: failed to setup shm segment\n");
 270                 return NULL;
 271         }
 272         if (thread_number >= max_jobs) {
 273                 log_err("error: maximum number of jobs (%d) reached.\n",
 274                                 max_jobs);
 275                 return NULL;
 276         }
 277
 278         td = &threads[thread_number++];
 279         *td = *parent;
 280
 281         td->o.uid = td->o.gid = -1U;
 282
 283         dup_files(td, parent);
 284         options_mem_dupe(td);
 285
 286         profile_add_hooks(td);
 287
 288         td->thread_number = thread_number;
 289         return td;
 290 }
 291
 292 static void put_job(struct thread_data *td)
 293 {
 294         if (td == &def_thread)
 295                 return;
 296
 297         profile_td_exit(td);
 298
 299         if (td->error)
 300                 log_info("fio: %s\n", td->verror);
 301
 302         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 303         thread_number--;
 304 }
 305
 306 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
 307 {
 308         unsigned int bs = td->o.min_bs[ddir];
 309         unsigned long long bytes_per_sec;
 310
 311         assert(ddir_rw(ddir));
 312
 313         if (td->o.rate[ddir])
 314                 bytes_per_sec = td->o.rate[ddir];
 315         else
 316                 bytes_per_sec = td->o.rate_iops[ddir] * bs;
 317
 318         if (!bytes_per_sec) {
 319                 log_err("rate lower than supported\n");
 320                 return -1;
 321         }
 322
 323         td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
 324         td->rate_pending_usleep[ddir] = 0;
 325         return 0;
 326 }
 327
 328 static int setup_rate(struct thread_data *td)
 329 {
 330         int ret = 0;
 331
 332         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
 333                 ret = __setup_rate(td, DDIR_READ);
 334         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
 335                 ret |= __setup_rate(td, DDIR_WRITE);
 336
 337         return ret;
 338 }
 339
 340 static int fixed_block_size(struct thread_options *o)
 341 {
 342         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
 343                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
 344                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
 345 }
 346
 347 /*
 348  * Lazy way of fixing up options that depend on each other. We could also
 349  * define option callback handlers, but this is easier.
 350  */
 351 static int fixup_options(struct thread_data *td)
 352 {
 353         struct thread_options *o = &td->o;
 354         int ret = 0;
 355
 356 #ifndef FIO_HAVE_PSHARED_MUTEX
 357         if (!o->use_thread) {
 358                 log_info("fio: this platform does not support process shared"
 359                          " mutexes, forcing use of threads. Use the 'thread'"
 360                          " option to get rid of this warning.\n");
 361                 o->use_thread = 1;
 362                 ret = warnings_fatal;
 363         }
 364 #endif
 365
 366         if (o->write_iolog_file && o->read_iolog_file) {
 367                 log_err("fio: read iolog overrides write_iolog\n");
 368                 free(o->write_iolog_file);
 369                 o->write_iolog_file = NULL;
 370                 ret = warnings_fatal;
 371         }
 372
 373         /*
 374          * only really works for sequential io for now, and with 1 file
 375          */
 376         if (o->zone_size && td_random(td) && o->open_files == 1)
 377                 o->zone_size = 0;
 378
 379         /*
 380          * Reads can do overwrites, we always need to pre-create the file
 381          */
 382         if (td_read(td) || td_rw(td))
 383                 o->overwrite = 1;
 384
 385         if (!o->min_bs[DDIR_READ])
 386                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
 387         if (!o->max_bs[DDIR_READ])
 388                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
 389         if (!o->min_bs[DDIR_WRITE])
 390                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 391         if (!o->max_bs[DDIR_WRITE])
 392                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
 393
 394         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
 395
 396         /*
 397          * For random IO, allow blockalign offset other than min_bs.
 398          */
 399         if (!o->ba[DDIR_READ] || !td_random(td))
 400                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
 401         if (!o->ba[DDIR_WRITE] || !td_random(td))
 402                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
 403
 404         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
 405             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
 406             !o->norandommap) {
 407                 log_err("fio: Any use of blockalign= turns off randommap\n");
 408                 o->norandommap = 1;
 409                 ret = warnings_fatal;
 410         }
 411
 412         if (!o->file_size_high)
 413                 o->file_size_high = o->file_size_low;
 414
 415         if (o->norandommap && o->verify != VERIFY_NONE
 416             && !fixed_block_size(o))  {
 417                 log_err("fio: norandommap given for variable block sizes, "
 418                         "verify disabled\n");
 419                 o->verify = VERIFY_NONE;
 420                 ret = warnings_fatal;
 421         }
 422         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
 423                 log_err("fio: bs_unaligned may not work with raw io\n");
 424
 425         /*
 426          * thinktime_spin must be less than thinktime
 427          */
 428         if (o->thinktime_spin > o->thinktime)
 429                 o->thinktime_spin = o->thinktime;
 430
 431         /*
 432          * The low water mark cannot be bigger than the iodepth
 433          */
 434         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
 435                 /*
 436                  * syslet work around - if the workload is sequential,
 437                  * we want to let the queue drain all the way down to
 438                  * avoid seeking between async threads
 439                  */
 440                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 441                         o->iodepth_low = 1;
 442                 else
 443                         o->iodepth_low = o->iodepth;
 444         }
 445
 446         /*
 447          * If batch number isn't set, default to the same as iodepth
 448          */
 449         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
 450                 o->iodepth_batch = o->iodepth;
 451
 452         if (o->nr_files > td->files_index)
 453                 o->nr_files = td->files_index;
 454
 455         if (o->open_files > o->nr_files || !o->open_files)
 456                 o->open_files = o->nr_files;
 457
 458         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
 459             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
 460                 o->rate_iops_min[1]))) {
 461                 log_err("fio: rate and rate_iops are mutually exclusive\n");
 462                 ret = 1;
 463         }
 464         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
 465             (o->rate_iops[0] < o->rate_iops_min[0]) ||
 466             (o->rate_iops[1] < o->rate_iops_min[1])) {
 467                 log_err("fio: minimum rate exceeds rate\n");
 468                 ret = 1;
 469         }
 470
 471         if (!o->timeout && o->time_based) {
 472                 log_err("fio: time_based requires a runtime/timeout setting\n");
 473                 o->time_based = 0;
 474                 ret = warnings_fatal;
 475         }
 476
 477         if (o->fill_device && !o->size)
 478                 o->size = -1ULL;
 479
 480         if (o->verify != VERIFY_NONE) {
 481                 if (td_write(td) && o->do_verify && o->numjobs > 1) {
 482                         log_info("Multiple writers may overwrite blocks that "
 483                                 "belong to other jobs. This can cause "
 484                                 "verification failures.\n");
 485                         ret = warnings_fatal;
 486                 }
 487
 488                 o->refill_buffers = 1;
 489                 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
 490                     !o->verify_interval)
 491                         o->verify_interval = o->min_bs[DDIR_WRITE];
 492         }
 493
 494         if (o->pre_read) {
 495                 o->invalidate_cache = 0;
 496                 if (td->io_ops->flags & FIO_PIPEIO) {
 497                         log_info("fio: cannot pre-read files with an IO engine"
 498                                  " that isn't seekable. Pre-read disabled.\n");
 499                         ret = warnings_fatal;
 500                 }
 501         }
 502
 503 #ifndef FIO_HAVE_FDATASYNC
 504         if (o->fdatasync_blocks) {
 505                 log_info("fio: this platform does not support fdatasync()"
 506                          " falling back to using fsync().  Use the 'fsync'"
 507                          " option instead of 'fdatasync' to get rid of"
 508                          " this warning\n");
 509                 o->fsync_blocks = o->fdatasync_blocks;
 510                 o->fdatasync_blocks = 0;
 511                 ret = warnings_fatal;
 512         }
 513 #endif
 514
 515         return ret;
 516 }
 517
 518 /*
 519  * This function leaks the buffer
 520  */
 521 static char *to_kmg(unsigned int val)
 522 {
 523         char *buf = malloc(32);
 524         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 525         char *p = post;
 526
 527         do {
 528                 if (val & 1023)
 529                         break;
 530
 531                 val >>= 10;
 532                 p++;
 533         } while (*p);
 534
 535         snprintf(buf, 31, "%u%c", val, *p);
 536         return buf;
 537 }
 538
 539 /* External engines are specified by "external:name.o") */
 540 static const char *get_engine_name(const char *str)
 541 {
 542         char *p = strstr(str, ":");
 543
 544         if (!p)
 545                 return str;
 546
 547         p++;
 548         strip_blank_front(&p);
 549         strip_blank_end(p);
 550         return p;
 551 }
 552
 553 static int exists_and_not_file(const char *filename)
 554 {
 555         struct stat sb;
 556
 557         if (lstat(filename, &sb) == -1)
 558                 return 0;
 559
 560         /* \\.\ is the device namespace in Windows, where every file
 561          * is a device node */
 562         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
 563                 return 0;
 564
 565         return 1;
 566 }
 567
 568 static void td_fill_rand_seeds_os(struct thread_data *td)
 569 {
 570         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
 571         os_random_seed(td->rand_seeds[1], &td->verify_state);
 572         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
 573
 574         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
 575                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
 576
 577         os_random_seed(td->rand_seeds[5], &td->file_size_state);
 578         os_random_seed(td->rand_seeds[6], &td->trim_state);
 579
 580         if (!td_random(td))
 581                 return;
 582
 583         if (td->o.rand_repeatable)
 584                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
 585
 586         os_random_seed(td->rand_seeds[4], &td->random_state);
 587 }
 588
 589 static void td_fill_rand_seeds_internal(struct thread_data *td)
 590 {
 591         init_rand_seed(&td->__bsrange_state, td->rand_seeds[0]);
 592         init_rand_seed(&td->__verify_state, td->rand_seeds[1]);
 593         init_rand_seed(&td->__rwmix_state, td->rand_seeds[2]);
 594
 595         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
 596                 init_rand_seed(&td->__next_file_state, td->rand_seeds[3]);
 597
 598         init_rand_seed(&td->__file_size_state, td->rand_seeds[5]);
 599         init_rand_seed(&td->__trim_state, td->rand_seeds[6]);
 600
 601         if (!td_random(td))
 602                 return;
 603
 604         if (td->o.rand_repeatable)
 605                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
 606
 607         init_rand_seed(&td->__random_state, td->rand_seeds[4]);
 608 }
 609
 610 void td_fill_rand_seeds(struct thread_data *td)
 611 {
 612         if (td->o.use_os_rand)
 613                 td_fill_rand_seeds_os(td);
 614         else
 615                 td_fill_rand_seeds_internal(td);
 616
 617         init_rand_seed(&td->buf_state, td->rand_seeds[7]);
 618 }
 619
 620 /*
 621  * Initialize the various random states we need (random io, block size ranges,
 622  * read/write mix, etc).
 623  */
 624 static int init_random_state(struct thread_data *td)
 625 {
 626         int fd;
 627
 628         fd = open("/dev/urandom", O_RDONLY);
 629         if (fd == -1) {
 630                 td_verror(td, errno, "open");
 631                 return 1;
 632         }
 633
 634         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
 635             (int) sizeof(td->rand_seeds)) {
 636                 td_verror(td, EIO, "read");
 637                 close(fd);
 638                 return 1;
 639         }
 640
 641         close(fd);
 642         td_fill_rand_seeds(td);
 643         return 0;
 644 }
 645
 646 /*
 647  * Adds a job to the list of things todo. Sanitizes the various options
 648  * to make sure we don't have conflicts, and initializes various
 649  * members of td.
 650  */
 651 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 652 {
 653         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 654                                    "randread", "randwrite", "randrw" };
 655         unsigned int i;
 656         const char *engine;
 657         char fname[PATH_MAX];
 658         int numjobs, file_alloced;
 659
 660         /*
 661          * the def_thread is just for options, it's not a real job
 662          */
 663         if (td == &def_thread)
 664                 return 0;
 665
 666         /*
 667          * if we are just dumping the output command line, don't add the job
 668          */
 669         if (dump_cmdline) {
 670                 put_job(td);
 671                 return 0;
 672         }
 673
 674         if (profile_td_init(td))
 675                 goto err;
 676
 677         engine = get_engine_name(td->o.ioengine);
 678         td->io_ops = load_ioengine(td, engine);
 679         if (!td->io_ops) {
 680                 log_err("fio: failed to load engine %s\n", engine);
 681                 goto err;
 682         }
 683
 684         if (td->o.use_thread)
 685                 nr_thread++;
 686         else
 687                 nr_process++;
 688
 689         if (td->o.odirect)
 690                 td->io_ops->flags |= FIO_RAWIO;
 691
 692         file_alloced = 0;
 693         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
 694                 file_alloced = 1;
 695
 696                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
 697                         add_file(td, jobname);
 698                 else {
 699                         for (i = 0; i < td->o.nr_files; i++) {
 700                                 sprintf(fname, "%s.%d.%d", jobname,
 701                                                         td->thread_number, i);
 702                                 add_file(td, fname);
 703                         }
 704                 }
 705         }
 706
 707         if (fixup_options(td))
 708                 goto err;
 709
 710         if (td->io_ops->flags & FIO_DISKLESSIO) {
 711                 struct fio_file *f;
 712
 713                 for_each_file(td, f, i)
 714                         f->real_file_size = -1ULL;
 715         }
 716
 717         td->mutex = fio_mutex_init(0);
 718
 719         td->ts.clat_percentiles = td->o.clat_percentiles;
 720         if (td->o.overwrite_plist)
 721                 td->ts.percentile_list = td->o.percentile_list;
 722         else
 723                 td->ts.percentile_list = NULL;
 724
 725         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 726         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 727         td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
 728         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 729         td->ddir_seq_nr = td->o.ddir_seq_nr;
 730
 731         if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
 732                 prev_group_jobs = 0;
 733                 groupid++;
 734         }
 735
 736         td->groupid = groupid;
 737         prev_group_jobs++;
 738
 739         if (init_random_state(td))
 740                 goto err;
 741
 742         if (setup_rate(td))
 743                 goto err;
 744
 745         if (td->o.write_lat_log) {
 746                 setup_log(&td->lat_log);
 747                 setup_log(&td->slat_log);
 748                 setup_log(&td->clat_log);
 749         }
 750         if (td->o.write_bw_log)
 751                 setup_log(&td->bw_log);
 752
 753         if (!td->o.name)
 754                 td->o.name = strdup(jobname);
 755
 756         if (!terse_output) {
 757                 if (!job_add_num) {
 758                         if (!strcmp(td->io_ops->name, "cpuio")) {
 759                                 log_info("%s: ioengine=cpu, cpuload=%u,"
 760                                          " cpucycle=%u\n", td->o.name,
 761                                                         td->o.cpuload,
 762                                                         td->o.cpucycle);
 763                         } else {
 764                                 char *c1, *c2, *c3, *c4;
 765
 766                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
 767                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
 768                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
 769                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
 770
 771                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
 772                                          " ioengine=%s, iodepth=%u\n",
 773                                                 td->o.name, td->groupid,
 774                                                 ddir_str[td->o.td_ddir],
 775                                                 c1, c2, c3, c4,
 776                                                 td->io_ops->name,
 777                                                 td->o.iodepth);
 778
 779                                 free(c1);
 780                                 free(c2);
 781                                 free(c3);
 782                                 free(c4);
 783                         }
 784                 } else if (job_add_num == 1)
 785                         log_info("...\n");
 786         }
 787
 788         /*
 789          * recurse add identical jobs, clear numjobs and stonewall options
 790          * as they don't apply to sub-jobs
 791          */
 792         numjobs = td->o.numjobs;
 793         while (--numjobs) {
 794                 struct thread_data *td_new = get_new_job(0, td);
 795
 796                 if (!td_new)
 797                         goto err;
 798
 799                 td_new->o.numjobs = 1;
 800                 td_new->o.stonewall = 0;
 801                 td_new->o.new_group = 0;
 802
 803                 if (file_alloced) {
 804                         td_new->o.filename = NULL;
 805                         td_new->files_index = 0;
 806                         td_new->files_size = 0;
 807                         td_new->files = NULL;
 808                 }
 809
 810                 job_add_num = numjobs - 1;
 811
 812                 if (add_job(td_new, jobname, job_add_num))
 813                         goto err;
 814         }
 815
 816         return 0;
 817 err:
 818         put_job(td);
 819         return -1;
 820 }
 821
 822 /*
 823  * Parse as if 'o' was a command line
 824  */
 825 void add_job_opts(const char **o)
 826 {
 827         struct thread_data *td, *td_parent;
 828         int i, in_global = 1;
 829         char jobname[32];
 830
 831         i = 0;
 832         td_parent = td = NULL;
 833         while (o[i]) {
 834                 if (!strncmp(o[i], "name", 4)) {
 835                         in_global = 0;
 836                         if (td)
 837                                 add_job(td, jobname, 0);
 838                         td = NULL;
 839                         sprintf(jobname, "%s", o[i] + 5);
 840                 }
 841                 if (in_global && !td_parent)
 842                         td_parent = get_new_job(1, &def_thread);
 843                 else if (!in_global && !td) {
 844                         if (!td_parent)
 845                                 td_parent = &def_thread;
 846                         td = get_new_job(0, td_parent);
 847                 }
 848                 if (in_global)
 849                         fio_options_parse(td_parent, (char **) &o[i], 1);
 850                 else
 851                         fio_options_parse(td, (char **) &o[i], 1);
 852                 i++;
 853         }
 854
 855         if (td)
 856                 add_job(td, jobname, 0);
 857 }
 858
 859 static int skip_this_section(const char *name)
 860 {
 861         int i;
 862
 863         if (!nr_job_sections)
 864                 return 0;
 865         if (!strncmp(name, "global", 6))
 866                 return 0;
 867
 868         for (i = 0; i < nr_job_sections; i++)
 869                 if (!strcmp(job_sections[i], name))
 870                         return 0;
 871
 872         return 1;
 873 }
 874
 875 static int is_empty_or_comment(char *line)
 876 {
 877         unsigned int i;
 878
 879         for (i = 0; i < strlen(line); i++) {
 880                 if (line[i] == ';')
 881                         return 1;
 882                 if (line[i] == '#')
 883                         return 1;
 884                 if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
 885                         return 0;
 886         }
 887
 888         return 1;
 889 }
 890
 891 /*
 892  * This is our [ini] type file parser.
 893  */
 894 int parse_jobs_ini(char *file, int is_buf, int stonewall_flag)
 895 {
 896         unsigned int global;
 897         struct thread_data *td;
 898         char *string, *name;
 899         FILE *f;
 900         char *p;
 901         int ret = 0, stonewall;
 902         int first_sect = 1;
 903         int skip_fgets = 0;
 904         int inside_skip = 0;
 905         char **opts;
 906         int i, alloc_opts, num_opts;
 907
 908         if (is_buf)
 909                 f = NULL;
 910         else {
 911                 if (!strcmp(file, "-"))
 912                         f = stdin;
 913                 else
 914                         f = fopen(file, "r");
 915
 916                 if (!f) {
 917                         perror("fopen job file");
 918                         return 1;
 919                 }
 920         }
 921
 922         string = malloc(4096);
 923
 924         /*
 925          * it's really 256 + small bit, 280 should suffice
 926          */
 927         name = malloc(280);
 928         memset(name, 0, 280);
 929
 930         alloc_opts = 8;
 931         opts = malloc(sizeof(char *) * alloc_opts);
 932         num_opts = 0;
 933
 934         stonewall = stonewall_flag;
 935         do {
 936                 /*
 937                  * if skip_fgets is set, we already have loaded a line we
 938                  * haven't handled.
 939                  */
 940                 if (!skip_fgets) {
 941                         if (is_buf)
 942                                 p = strsep(&file, "\n");
 943                         else
 944                                 p = fgets(string, 4095, f);
 945                         if (!p)
 946                                 break;
 947                 }
 948
 949                 skip_fgets = 0;
 950                 strip_blank_front(&p);
 951                 strip_blank_end(p);
 952
 953                 if (is_empty_or_comment(p))
 954                         continue;
 955                 if (sscanf(p, "[%255[^\n]]", name) != 1) {
 956                         if (inside_skip)
 957                                 continue;
 958                         log_err("fio: option <%s> outside of [] job section\n",
 959                                                                         p);
 960                         break;
 961                 }
 962
 963                 name[strlen(name) - 1] = '\0';
 964
 965                 if (skip_this_section(name)) {
 966                         inside_skip = 1;
 967                         continue;
 968                 } else
 969                         inside_skip = 0;
 970
 971                 global = !strncmp(name, "global", 6);
 972
 973                 if (dump_cmdline) {
 974                         if (first_sect)
 975                                 log_info("fio ");
 976                         if (!global)
 977                                 log_info("--name=%s ", name);
 978                         first_sect = 0;
 979                 }
 980
 981                 td = get_new_job(global, &def_thread);
 982                 if (!td) {
 983                         ret = 1;
 984                         break;
 985                 }
 986
 987                 /*
 988                  * Seperate multiple job files by a stonewall
 989                  */
 990                 if (!global && stonewall) {
 991                         td->o.stonewall = stonewall;
 992                         stonewall = 0;
 993                 }
 994
 995                 num_opts = 0;
 996                 memset(opts, 0, alloc_opts * sizeof(char *));
 997
 998                 while (1) {
 999                         if (is_buf)
1000                                 p = strsep(&file, "\n");
1001                         else
1002                                 p = fgets(string, 4096, f);
1003                         if (!p)
1004                                 break;
1005
1006                         if (is_empty_or_comment(p))
1007                                 continue;
1008
1009                         strip_blank_front(&p);
1010
1011                         /*
1012                          * new section, break out and make sure we don't
1013                          * fgets() a new line at the top.
1014                          */
1015                         if (p[0] == '[') {
1016                                 skip_fgets = 1;
1017                                 break;
1018                         }
1019
1020                         strip_blank_end(p);
1021
1022                         if (num_opts == alloc_opts) {
1023                                 alloc_opts <<= 1;
1024                                 opts = realloc(opts,
1025                                                 alloc_opts * sizeof(char *));
1026                         }
1027
1028                         opts[num_opts] = strdup(p);
1029                         num_opts++;
1030                 }
1031
1032                 ret = fio_options_parse(td, opts, num_opts);
1033                 if (!ret) {
1034                         if (dump_cmdline)
1035                                 for (i = 0; i < num_opts; i++)
1036                                         log_info("--%s ", opts[i]);
1037
1038                         ret = add_job(td, name, 0);
1039                 } else {
1040                         log_err("fio: job %s dropped\n", name);
1041                         put_job(td);
1042                 }
1043
1044                 for (i = 0; i < num_opts; i++)
1045                         free(opts[i]);
1046                 num_opts = 0;
1047         } while (!ret);
1048
1049         if (dump_cmdline)
1050                 log_info("\n");
1051
1052         for (i = 0; i < num_opts; i++)
1053                 free(opts[i]);
1054
1055         free(string);
1056         free(name);
1057         free(opts);
1058         if (!is_buf && f != stdin)
1059                 fclose(f);
1060         return ret;
1061 }
1062
1063 static int fill_def_thread(void)
1064 {
1065         memset(&def_thread, 0, sizeof(def_thread));
1066
1067         fio_getaffinity(getpid(), &def_thread.o.cpumask);
1068
1069         /*
1070          * fill default options
1071          */
1072         fio_fill_default_options(&def_thread);
1073         return 0;
1074 }
1075
1076 static void usage(const char *name)
1077 {
1078         printf("%s\n", fio_version_string);
1079         printf("%s [options] [job options] <job file(s)>\n", name);
1080         printf("\t--debug=options\tEnable debug logging\n");
1081         printf("\t--output\tWrite output to file\n");
1082         printf("\t--timeout\tRuntime in seconds\n");
1083         printf("\t--latency-log\tGenerate per-job latency logs\n");
1084         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1085         printf("\t--minimal\tMinimal (terse) output\n");
1086         printf("\t--version\tPrint version info and exit\n");
1087         printf("\t--terse-version=x Terse version output format\n");
1088         printf("\t--help\t\tPrint this page\n");
1089         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
1090                 " them\n");
1091         printf("\t--showcmd\tTurn a job file into command line options\n");
1092         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
1093         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
1094         printf("\t--readonly\tTurn on safety read-only checks, preventing"
1095                 " writes\n");
1096         printf("\t--section=name\tOnly run specified section in job file\n");
1097         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
1098                 " (def 1024)\n");
1099         printf("\t--warnings-fatal Fio parser warnings are fatal\n");
1100         printf("\t--max-jobs\tMaximum number of threads/processes to support\n");
1101         printf("\t--server\tStart a backend fio server\n");
1102         printf("\t--client=hostname Talk to remove backend fio server at hostname\n");
1103         printf("\t--net-port=port\tUse specified port for client/server connection\n");
1104         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1105         printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
1106 }
1107
1108 #ifdef FIO_INC_DEBUG
1109 struct debug_level debug_levels[] = {
1110         { .name = "process",    .shift = FD_PROCESS, },
1111         { .name = "file",       .shift = FD_FILE, },
1112         { .name = "io",         .shift = FD_IO, },
1113         { .name = "mem",        .shift = FD_MEM, },
1114         { .name = "blktrace",   .shift = FD_BLKTRACE },
1115         { .name = "verify",     .shift = FD_VERIFY },
1116         { .name = "random",     .shift = FD_RANDOM },
1117         { .name = "parse",      .shift = FD_PARSE },
1118         { .name = "diskutil",   .shift = FD_DISKUTIL },
1119         { .name = "job",        .shift = FD_JOB },
1120         { .name = "mutex",      .shift = FD_MUTEX },
1121         { .name = "profile",    .shift = FD_PROFILE },
1122         { .name = "time",       .shift = FD_TIME },
1123         { .name = "net",        .shift = FD_NET },
1124         { .name = NULL, },
1125 };
1126
1127 static int set_debug(const char *string)
1128 {
1129         struct debug_level *dl;
1130         char *p = (char *) string;
1131         char *opt;
1132         int i;
1133
1134         if (!strcmp(string, "?") || !strcmp(string, "help")) {
1135                 log_info("fio: dumping debug options:");
1136                 for (i = 0; debug_levels[i].name; i++) {
1137                         dl = &debug_levels[i];
1138                         log_info("%s,", dl->name);
1139                 }
1140                 log_info("all\n");
1141                 return 1;
1142         }
1143
1144         while ((opt = strsep(&p, ",")) != NULL) {
1145                 int found = 0;
1146
1147                 if (!strncmp(opt, "all", 3)) {
1148                         log_info("fio: set all debug options\n");
1149                         fio_debug = ~0UL;
1150                         continue;
1151                 }
1152
1153                 for (i = 0; debug_levels[i].name; i++) {
1154                         dl = &debug_levels[i];
1155                         found = !strncmp(opt, dl->name, strlen(dl->name));
1156                         if (!found)
1157                                 continue;
1158
1159                         if (dl->shift == FD_JOB) {
1160                                 opt = strchr(opt, ':');
1161                                 if (!opt) {
1162                                         log_err("fio: missing job number\n");
1163                                         break;
1164                                 }
1165                                 opt++;
1166                                 fio_debug_jobno = atoi(opt);
1167                                 log_info("fio: set debug jobno %d\n",
1168                                                         fio_debug_jobno);
1169                         } else {
1170                                 log_info("fio: set debug option %s\n", opt);
1171                                 fio_debug |= (1UL << dl->shift);
1172                         }
1173                         break;
1174                 }
1175
1176                 if (!found)
1177                         log_err("fio: debug mask %s not found\n", opt);
1178         }
1179         return 0;
1180 }
1181 #else
1182 static int set_debug(const char *string)
1183 {
1184         log_err("fio: debug tracing not included in build\n");
1185         return 1;
1186 }
1187 #endif
1188
1189 static void fio_options_fill_optstring(void)
1190 {
1191         char *ostr = cmd_optstr;
1192         int i, c;
1193
1194         c = i = 0;
1195         while (l_opts[i].name) {
1196                 ostr[c++] = l_opts[i].val;
1197                 if (l_opts[i].has_arg == required_argument)
1198                         ostr[c++] = ':';
1199                 else if (l_opts[i].has_arg == optional_argument) {
1200                         ostr[c++] = ':';
1201                         ostr[c++] = ':';
1202                 }
1203                 i++;
1204         }
1205         ostr[c] = '\0';
1206 }
1207
1208 int parse_cmd_line(int argc, char *argv[])
1209 {
1210         struct thread_data *td = NULL;
1211         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1212         char *ostr = cmd_optstr;
1213         int daemonize_server = 0;
1214         char *cur_client = NULL;
1215         int backend = 0;
1216
1217         while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1218                 switch (c) {
1219                 case 'a':
1220                         fio_client_add_cmd_option(cur_client, argv[optind]);
1221                         smalloc_pool_size = atoi(optarg);
1222                         break;
1223                 case 't':
1224                         fio_client_add_cmd_option(cur_client, argv[optind]);
1225                         def_thread.o.timeout = atoi(optarg);
1226                         break;
1227                 case 'l':
1228                         fio_client_add_cmd_option(cur_client, argv[optind]);
1229                         write_lat_log = 1;
1230                         break;
1231                 case 'b':
1232                         fio_client_add_cmd_option(cur_client, argv[optind]);
1233                         write_bw_log = 1;
1234                         break;
1235                 case 'o':
1236                         fio_client_add_cmd_option(cur_client, argv[optind]);
1237                         f_out = fopen(optarg, "w+");
1238                         if (!f_out) {
1239                                 perror("fopen output");
1240                                 exit(1);
1241                         }
1242                         f_err = f_out;
1243                         break;
1244                 case 'm':
1245                         fio_client_add_cmd_option(cur_client, argv[optind]);
1246                         terse_output = 1;
1247                         break;
1248                 case 'h':
1249                         usage(argv[0]);
1250                         exit(0);
1251                 case 'c':
1252                         exit(fio_show_option_help(optarg));
1253                 case 's':
1254                         fio_client_add_cmd_option(cur_client, argv[optind]);
1255                         dump_cmdline = 1;
1256                         break;
1257                 case 'r':
1258                         fio_client_add_cmd_option(cur_client, argv[optind]);
1259                         read_only = 1;
1260                         break;
1261                 case 'v':
1262                         log_info("%s\n", fio_version_string);
1263                         exit(0);
1264                 case 'V':
1265                         fio_client_add_cmd_option(cur_client, argv[optind]);
1266                         terse_version = atoi(optarg);
1267                         if (terse_version != 2) {
1268                                 log_err("fio: bad terse version format\n");
1269                                 exit_val = 1;
1270                                 do_exit++;
1271                         }
1272                         break;
1273                 case 'e':
1274                         fio_client_add_cmd_option(cur_client, argv[optind]);
1275                         if (!strcmp("always", optarg))
1276                                 eta_print = FIO_ETA_ALWAYS;
1277                         else if (!strcmp("never", optarg))
1278                                 eta_print = FIO_ETA_NEVER;
1279                         break;
1280                 case 'd':
1281                         fio_client_add_cmd_option(cur_client, argv[optind]);
1282                         if (set_debug(optarg))
1283                                 do_exit++;
1284                         break;
1285                 case 'x': {
1286                         size_t new_size;
1287
1288                         fio_client_add_cmd_option(cur_client, argv[optind]);
1289
1290                         if (!strcmp(optarg, "global")) {
1291                                 log_err("fio: can't use global as only "
1292                                         "section\n");
1293                                 do_exit++;
1294                                 exit_val = 1;
1295                                 break;
1296                         }
1297                         new_size = (nr_job_sections + 1) * sizeof(char *);
1298                         job_sections = realloc(job_sections, new_size);
1299                         job_sections[nr_job_sections] = strdup(optarg);
1300                         nr_job_sections++;
1301                         break;
1302                         }
1303                 case 'p':
1304                         fio_client_add_cmd_option(cur_client, argv[optind]);
1305                         exec_profile = strdup(optarg);
1306                         break;
1307                 case FIO_GETOPT_JOB: {
1308                         const char *opt = l_opts[lidx].name;
1309                         char *val = optarg;
1310
1311                         fio_client_add_cmd_option(cur_client, argv[optind]);
1312
1313                         if (!strncmp(opt, "name", 4) && td) {
1314                                 ret = add_job(td, td->o.name ?: "fio", 0);
1315                                 if (ret)
1316                                         return 0;
1317                                 td = NULL;
1318                         }
1319                         if (!td) {
1320                                 int is_section = !strncmp(opt, "name", 4);
1321                                 int global = 0;
1322
1323                                 if (!is_section || !strncmp(val, "global", 6))
1324                                         global = 1;
1325
1326                                 if (is_section && skip_this_section(val))
1327                                         continue;
1328
1329                                 td = get_new_job(global, &def_thread);
1330                                 if (!td)
1331                                         return 0;
1332                         }
1333
1334                         ret = fio_cmd_option_parse(td, opt, val);
1335                         break;
1336                 }
1337                 case 'w':
1338                         fio_client_add_cmd_option(cur_client, argv[optind]);
1339                         warnings_fatal = 1;
1340                         break;
1341                 case 'j':
1342                         fio_client_add_cmd_option(cur_client, argv[optind]);
1343                         max_jobs = atoi(optarg);
1344                         if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1345                                 log_err("fio: invalid max jobs: %d\n", max_jobs);
1346                                 do_exit++;
1347                                 exit_val = 1;
1348                         }
1349                         break;
1350                 case 'S':
1351                         if (nr_clients) {
1352                                 log_err("fio: can't be both client and server\n");
1353                                 do_exit++;
1354                                 exit_val = 1;
1355                                 break;
1356                         }
1357                         is_backend = 1;
1358                         backend = 1;
1359                         break;
1360                 case 'D':
1361                         daemonize_server = 1;
1362                         break;
1363                 case 'P':
1364                         fio_net_port = atoi(optarg);
1365                         break;
1366                 case 'C':
1367                         if (is_backend) {
1368                                 log_err("fio: can't be both client and server\n");
1369                                 do_exit++;
1370                                 exit_val = 1;
1371                                 break;
1372                         }
1373                         fio_client_add(optarg);
1374                         cur_client = optarg;
1375                         break;
1376                 default:
1377                         do_exit++;
1378                         exit_val = 1;
1379                         break;
1380                 }
1381                 if (do_exit)
1382                         break;
1383         }
1384
1385         if (do_exit)
1386                 exit(exit_val);
1387
1388         if (nr_clients && fio_clients_connect()) {
1389                 do_exit++;
1390                 exit_val = 1;
1391                 return -1;
1392         }
1393
1394         if (is_backend && backend)
1395                 return fio_start_server(daemonize_server);
1396
1397         if (td) {
1398                 if (!ret)
1399                         ret = add_job(td, td->o.name ?: "fio", 0);
1400         }
1401
1402         while (optind < argc) {
1403                 ini_idx++;
1404                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1405                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1406                 optind++;
1407         }
1408
1409         return ini_idx;
1410 }
1411
1412 int parse_options(int argc, char *argv[])
1413 {
1414         int job_files, i;
1415
1416         f_out = stdout;
1417         f_err = stderr;
1418
1419         fio_options_fill_optstring();
1420         fio_options_dup_and_init(l_opts);
1421
1422         atexit(free_shm);
1423
1424         if (fill_def_thread())
1425                 return 1;
1426
1427         job_files = parse_cmd_line(argc, argv);
1428
1429         if (job_files > 0) {
1430                 for (i = 0; i < job_files; i++) {
1431                         if (fill_def_thread())
1432                                 return 1;
1433                         if (nr_clients) {
1434                                 if (fio_clients_send_ini(ini_file[i]))
1435                                         return 1;
1436                                 free(ini_file[i]);
1437                         } else if (!is_backend) {
1438                                 if (parse_jobs_ini(ini_file[i], 0, i))
1439                                         return 1;
1440                                 free(ini_file[i]);
1441                         }
1442                 }
1443         }
1444
1445         free(ini_file);
1446         options_mem_free(&def_thread);
1447
1448         if (!thread_number) {
1449                 if (dump_cmdline)
1450                         return 0;
1451                 if (exec_profile)
1452                         return 0;
1453                 if (is_backend || nr_clients)
1454                         return 0;
1455
1456                 if (job_files > 0) {
1457                         log_err("No jobs(s) defined\n\n");
1458                         usage(argv[0]);
1459                 }
1460                 return 1;
1461         }
1462
1463         if (def_thread.o.gtod_offload) {
1464                 fio_gtod_init();
1465                 fio_gtod_offload = 1;
1466                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1467         }
1468
1469         log_info("%s\n", fio_version_string);
1470         return 0;
1471 }