init.c

   1 /*
   2  * This file contains job initialization and setup functions.
   3  */
   4 #include <stdio.h>
   5 #include <stdlib.h>
   6 #include <unistd.h>
   7 #include <fcntl.h>
   8 #include <ctype.h>
   9 #include <string.h>
  10 #include <errno.h>
  11 #include <getopt.h>
  12 #include <sys/ipc.h>
  13 #include <sys/shm.h>
  14 #include <sys/types.h>
  15 #include <sys/stat.h>
  16
  17 #include "fio.h"
  18 #include "parse.h"
  19
  20 static char fio_version_string[] = "fio 1.14a";
  21
  22 #define FIO_RANDSEED            (0xb1899bedUL)
  23
  24 static char **ini_file;
  25 static int max_jobs = MAX_JOBS;
  26
  27 struct thread_data def_thread;
  28 struct thread_data *threads = NULL;
  29
  30 int exitall_on_terminate = 0;
  31 int terse_output = 0;
  32 unsigned long long mlock_size = 0;
  33 FILE *f_out = NULL;
  34 FILE *f_err = NULL;
  35
  36 int write_bw_log = 0;
  37
  38 static int def_timeout = 0;
  39 static int write_lat_log = 0;
  40
  41 static int prev_group_jobs;
  42
  43 /*
  44  * Command line options. These will contain the above, plus a few
  45  * extra that only pertain to fio itself and not jobs.
  46  */
  47 static struct option long_options[FIO_NR_OPTIONS] = {
  48         {
  49                 .name           = "output",
  50                 .has_arg        = required_argument,
  51                 .val            = 'o',
  52         },
  53         {
  54                 .name           = "timeout",
  55                 .has_arg        = required_argument,
  56                 .val            = 't',
  57         },
  58         {
  59                 .name           = "latency-log",
  60                 .has_arg        = required_argument,
  61                 .val            = 'l',
  62         },
  63         {
  64                 .name           = "bandwidth-log",
  65                 .has_arg        = required_argument,
  66                 .val            = 'b',
  67         },
  68         {
  69                 .name           = "minimal",
  70                 .has_arg        = optional_argument,
  71                 .val            = 'm',
  72         },
  73         {
  74                 .name           = "version",
  75                 .has_arg        = no_argument,
  76                 .val            = 'v',
  77         },
  78         {
  79                 .name           = "help",
  80                 .has_arg        = no_argument,
  81                 .val            = 'h',
  82         },
  83         {
  84                 .name           = "cmdhelp",
  85                 .has_arg        = optional_argument,
  86                 .val            = 'c',
  87         },
  88         {
  89                 .name           = NULL,
  90         },
  91 };
  92
  93 FILE *get_f_out()
  94 {
  95         return f_out;
  96 }
  97
  98 FILE *get_f_err()
  99 {
 100         return f_err;
 101 }
 102
 103 /*
 104  * Return a free job structure.
 105  */
 106 static struct thread_data *get_new_job(int global, struct thread_data *parent)
 107 {
 108         struct thread_data *td;
 109
 110         if (global)
 111                 return &def_thread;
 112         if (thread_number >= max_jobs)
 113                 return NULL;
 114
 115         td = &threads[thread_number++];
 116         *td = *parent;
 117
 118         td->thread_number = thread_number;
 119         return td;
 120 }
 121
 122 static void put_job(struct thread_data *td)
 123 {
 124         if (td == &def_thread)
 125                 return;
 126
 127         if (td->error)
 128                 log_info("fio: %s\n", td->verror);
 129
 130         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
 131         thread_number--;
 132 }
 133
 134 /*
 135  * Lazy way of fixing up options that depend on each other. We could also
 136  * define option callback handlers, but this is easier.
 137  */
 138 static void fixup_options(struct thread_data *td)
 139 {
 140         if (!td->rwmixread && td->rwmixwrite)
 141                 td->rwmixread = 100 - td->rwmixwrite;
 142
 143         if (td->write_iolog_file && td->read_iolog_file) {
 144                 log_err("fio: read iolog overrides write_iolog\n");
 145                 free(td->write_iolog_file);
 146                 td->write_iolog_file = NULL;
 147         }
 148
 149         if (td->io_ops->flags & FIO_SYNCIO)
 150                 td->iodepth = 1;
 151         else {
 152                 if (!td->iodepth)
 153                         td->iodepth = td->open_files;
 154         }
 155
 156         /*
 157          * only really works for sequential io for now, and with 1 file
 158          */
 159         if (td->zone_size && td_random(td) && td->open_files == 1)
 160                 td->zone_size = 0;
 161
 162         /*
 163          * Reads can do overwrites, we always need to pre-create the file
 164          */
 165         if (td_read(td) || td_rw(td))
 166                 td->overwrite = 1;
 167
 168         if (!td->min_bs[DDIR_READ])
 169                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
 170         if (!td->max_bs[DDIR_READ])
 171                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
 172         if (!td->min_bs[DDIR_WRITE])
 173                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
 174         if (!td->max_bs[DDIR_WRITE])
 175                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
 176
 177         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
 178
 179         if (!td->file_size_high)
 180                 td->file_size_high = td->file_size_low;
 181
 182         if (td_read(td) && !td_rw(td))
 183                 td->verify = 0;
 184
 185         if (td->norandommap && td->verify != VERIFY_NONE) {
 186                 log_err("fio: norandommap given, verify disabled\n");
 187                 td->verify = VERIFY_NONE;
 188         }
 189         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
 190                 log_err("fio: bs_unaligned may not work with raw io\n");
 191
 192         /*
 193          * thinktime_spin must be less than thinktime
 194          */
 195         if (td->thinktime_spin > td->thinktime)
 196                 td->thinktime_spin = td->thinktime;
 197
 198         /*
 199          * The low water mark cannot be bigger than the iodepth
 200          */
 201         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
 202                 /*
 203                  * syslet work around - if the workload is sequential,
 204                  * we want to let the queue drain all the way down to
 205                  * avoid seeking between async threads
 206                  */
 207                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
 208                         td->iodepth_low = 1;
 209                 else
 210                         td->iodepth_low = td->iodepth;
 211         }
 212
 213         /*
 214          * If batch number isn't set, default to the same as iodepth
 215          */
 216         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
 217                 td->iodepth_batch = td->iodepth;
 218
 219         if (td->nr_files > td->files_index)
 220                 td->nr_files = td->files_index;
 221
 222         if (td->open_files > td->nr_files || !td->open_files)
 223                 td->open_files = td->nr_files;
 224 }
 225
 226 /*
 227  * This function leaks the buffer
 228  */
 229 static char *to_kmg(unsigned int val)
 230 {
 231         char *buf = malloc(32);
 232         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
 233         char *p = post;
 234
 235         do {
 236                 if (val & 1023)
 237                         break;
 238
 239                 val >>= 10;
 240                 p++;
 241         } while (*p);
 242
 243         snprintf(buf, 31, "%u%c", val, *p);
 244         return buf;
 245 }
 246
 247 /* External engines are specified by "external:name.o") */
 248 static const char *get_engine_name(const char *str)
 249 {
 250         char *p = strstr(str, ":");
 251
 252         if (!p)
 253                 return str;
 254
 255         p++;
 256         strip_blank_front(&p);
 257         strip_blank_end(p);
 258         return p;
 259 }
 260
 261 static int exists_and_not_file(const char *filename)
 262 {
 263         struct stat sb;
 264
 265         if (lstat(filename, &sb) == -1)
 266                 return 0;
 267
 268         if (S_ISREG(sb.st_mode))
 269                 return 0;
 270
 271         return 1;
 272 }
 273
 274 /*
 275  * Initialize the various random states we need (random io, block size ranges,
 276  * read/write mix, etc).
 277  */
 278 static int init_random_state(struct thread_data *td)
 279 {
 280         unsigned long seeds[6];
 281         int fd, num_maps, blocks;
 282         struct fio_file *f;
 283         unsigned int i;
 284
 285         fd = open("/dev/urandom", O_RDONLY);
 286         if (fd == -1) {
 287                 td_verror(td, errno, "open");
 288                 return 1;
 289         }
 290
 291         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
 292                 td_verror(td, EIO, "read");
 293                 close(fd);
 294                 return 1;
 295         }
 296
 297         close(fd);
 298
 299         os_random_seed(seeds[0], &td->bsrange_state);
 300         os_random_seed(seeds[1], &td->verify_state);
 301         os_random_seed(seeds[2], &td->rwmix_state);
 302
 303         if (td->file_service_type == FIO_FSERVICE_RANDOM)
 304                 os_random_seed(seeds[3], &td->next_file_state);
 305
 306         os_random_seed(seeds[5], &td->file_size_state);
 307
 308         if (!td_random(td))
 309                 return 0;
 310
 311         if (td->rand_repeatable)
 312                 seeds[4] = FIO_RANDSEED * td->thread_number;
 313
 314         if (!td->norandommap) {
 315                 for_each_file(td, f, i) {
 316                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
 317                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
 318                         f->file_map = malloc(num_maps * sizeof(long));
 319                         if (!f->file_map) {
 320                                 log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
 321                                 return 1;
 322                         }
 323                         f->num_maps = num_maps;
 324                         memset(f->file_map, 0, num_maps * sizeof(long));
 325                 }
 326         }
 327
 328         os_random_seed(seeds[4], &td->random_state);
 329         return 0;
 330 }
 331
 332
 333 /*
 334  * Adds a job to the list of things todo. Sanitizes the various options
 335  * to make sure we don't have conflicts, and initializes various
 336  * members of td.
 337  */
 338 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 339 {
 340         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
 341                                    "randread", "randwrite", "randrw" };
 342         unsigned int i;
 343         struct fio_file *f;
 344         const char *engine;
 345         char fname[PATH_MAX];
 346         int numjobs, file_alloced;
 347
 348         /*
 349          * the def_thread is just for options, it's not a real job
 350          */
 351         if (td == &def_thread)
 352                 return 0;
 353
 354         engine = get_engine_name(td->ioengine);
 355         td->io_ops = load_ioengine(td, engine);
 356         if (!td->io_ops) {
 357                 log_err("fio: failed to load engine %s\n", engine);
 358                 goto err;
 359         }
 360
 361         if (td->use_thread)
 362                 nr_thread++;
 363         else
 364                 nr_process++;
 365
 366         if (td->odirect)
 367                 td->io_ops->flags |= FIO_RAWIO;
 368
 369         file_alloced = 0;
 370         if (!td->filename && !td->files_index) {
 371                 file_alloced = 1;
 372
 373                 if (td->nr_files == 1 && exists_and_not_file(jobname))
 374                         add_file(td, jobname);
 375                 else {
 376                         for (i = 0; i < td->nr_files; i++) {
 377                                 sprintf(fname, "%s.%d.%d", jobname, td->thread_number, i);
 378                                 add_file(td, fname);
 379                         }
 380                 }
 381         }
 382
 383         fixup_options(td);
 384
 385         for_each_file(td, f, i) {
 386                 if (td->directory && f->filetype == FIO_TYPE_FILE) {
 387                         sprintf(fname, "%s/%s", td->directory, f->file_name);
 388                         f->file_name = strdup(fname);
 389                 }
 390         }
 391
 392         td->mutex = fio_sem_init(0);
 393
 394         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
 395         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
 396         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
 397
 398         if ((td->stonewall || td->numjobs > 1) && prev_group_jobs) {
 399                 prev_group_jobs = 0;
 400                 groupid++;
 401         }
 402
 403         td->groupid = groupid;
 404         prev_group_jobs++;
 405
 406         if (init_random_state(td))
 407                 goto err;
 408
 409         if (setup_rate(td))
 410                 goto err;
 411
 412         if (td->write_lat_log) {
 413                 setup_log(&td->ts.slat_log);
 414                 setup_log(&td->ts.clat_log);
 415         }
 416         if (td->write_bw_log)
 417                 setup_log(&td->ts.bw_log);
 418
 419         if (!td->name)
 420                 td->name = strdup(jobname);
 421
 422         if (!terse_output) {
 423                 if (!job_add_num) {
 424                         if (!strcmp(td->io_ops->name, "cpuio"))
 425                                 log_info("%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
 426                         else {
 427                                 char *c1, *c2, *c3, *c4;
 428
 429                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
 430                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
 431                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
 432                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
 433
 434                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[td->td_ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
 435
 436                                 free(c1);
 437                                 free(c2);
 438                                 free(c3);
 439                                 free(c4);
 440                         }
 441                 } else if (job_add_num == 1)
 442                         log_info("...\n");
 443         }
 444
 445         /*
 446          * recurse add identical jobs, clear numjobs and stonewall options
 447          * as they don't apply to sub-jobs
 448          */
 449         numjobs = td->numjobs;
 450         while (--numjobs) {
 451                 struct thread_data *td_new = get_new_job(0, td);
 452
 453                 if (!td_new)
 454                         goto err;
 455
 456                 td_new->numjobs = 1;
 457                 td_new->stonewall = 0;
 458
 459                 if (file_alloced) {
 460                         td_new->filename = NULL;
 461                         td_new->files_index = 0;
 462                         td_new->files = NULL;
 463                 }
 464
 465                 job_add_num = numjobs - 1;
 466
 467                 if (add_job(td_new, jobname, job_add_num))
 468                         goto err;
 469         }
 470
 471         if (td->numjobs > 1) {
 472                 groupid++;
 473                 prev_group_jobs = 0;
 474         }
 475
 476         return 0;
 477 err:
 478         put_job(td);
 479         return -1;
 480 }
 481
 482 static int is_empty_or_comment(char *line)
 483 {
 484         unsigned int i;
 485
 486         for (i = 0; i < strlen(line); i++) {
 487                 if (line[i] == ';')
 488                         return 1;
 489                 if (line[i] == '#')
 490                         return 1;
 491                 if (!isspace(line[i]) && !iscntrl(line[i]))
 492                         return 0;
 493         }
 494
 495         return 1;
 496 }
 497
 498 /*
 499  * This is our [ini] type file parser.
 500  */
 501 static int parse_jobs_ini(char *file, int stonewall_flag)
 502 {
 503         unsigned int global;
 504         struct thread_data *td;
 505         char *string, *name;
 506         fpos_t off;
 507         FILE *f;
 508         char *p;
 509         int ret = 0, stonewall;
 510
 511         f = fopen(file, "r");
 512         if (!f) {
 513                 perror("fopen job file");
 514                 return 1;
 515         }
 516
 517         string = malloc(4096);
 518         name = malloc(256);
 519         memset(name, 0, 256);
 520
 521         stonewall = stonewall_flag;
 522         do {
 523                 p = fgets(string, 4095, f);
 524                 if (!p)
 525                         break;
 526                 if (is_empty_or_comment(p))
 527                         continue;
 528                 if (sscanf(p, "[%255s]", name) != 1)
 529                         continue;
 530
 531                 global = !strncmp(name, "global", 6);
 532
 533                 name[strlen(name) - 1] = '\0';
 534
 535                 td = get_new_job(global, &def_thread);
 536                 if (!td) {
 537                         ret = 1;
 538                         break;
 539                 }
 540
 541                 /*
 542                  * Seperate multiple job files by a stonewall
 543                  */
 544                 if (!global && stonewall) {
 545                         td->stonewall = stonewall;
 546                         stonewall = 0;
 547                 }
 548
 549                 fgetpos(f, &off);
 550                 while ((p = fgets(string, 4096, f)) != NULL) {
 551                         if (is_empty_or_comment(p))
 552                                 continue;
 553
 554                         strip_blank_front(&p);
 555
 556                         if (p[0] == '[')
 557                                 break;
 558
 559                         strip_blank_end(p);
 560
 561                         fgetpos(f, &off);
 562
 563                         /*
 564                          * Don't break here, continue parsing options so we
 565                          * dump all the bad ones. Makes trial/error fixups
 566                          * easier on the user.
 567                          */
 568                         ret |= fio_option_parse(td, p);
 569                 }
 570
 571                 if (!ret) {
 572                         fsetpos(f, &off);
 573                         ret = add_job(td, name, 0);
 574                 } else {
 575                         log_err("fio: job %s dropped\n", name);
 576                         put_job(td);
 577                 }
 578         } while (!ret);
 579
 580         free(string);
 581         free(name);
 582         fclose(f);
 583         return ret;
 584 }
 585
 586 static int fill_def_thread(void)
 587 {
 588         memset(&def_thread, 0, sizeof(def_thread));
 589
 590         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
 591                 perror("sched_getaffinity");
 592                 return 1;
 593         }
 594
 595         /*
 596          * fill default options
 597          */
 598         fio_fill_default_options(&def_thread);
 599
 600         def_thread.timeout = def_timeout;
 601         def_thread.write_bw_log = write_bw_log;
 602         def_thread.write_lat_log = write_lat_log;
 603
 604 #ifdef FIO_HAVE_DISK_UTIL
 605         def_thread.do_disk_util = 1;
 606 #endif
 607
 608         return 0;
 609 }
 610
 611 static void free_shm(void)
 612 {
 613         struct shmid_ds sbuf;
 614
 615         if (threads) {
 616                 shmdt((void *) threads);
 617                 threads = NULL;
 618                 shmctl(shm_id, IPC_RMID, &sbuf);
 619         }
 620 }
 621
 622 /*
 623  * The thread area is shared between the main process and the job
 624  * threads/processes. So setup a shared memory segment that will hold
 625  * all the job info.
 626  */
 627 static int setup_thread_area(void)
 628 {
 629         /*
 630          * 1024 is too much on some machines, scale max_jobs if
 631          * we get a failure that looks like too large a shm segment
 632          */
 633         do {
 634                 size_t size = max_jobs * sizeof(struct thread_data);
 635
 636                 shm_id = shmget(0, size, IPC_CREAT | 0600);
 637                 if (shm_id != -1)
 638                         break;
 639                 if (errno != EINVAL) {
 640                         perror("shmget");
 641                         break;
 642                 }
 643
 644                 max_jobs >>= 1;
 645         } while (max_jobs);
 646
 647         if (shm_id == -1)
 648                 return 1;
 649
 650         threads = shmat(shm_id, NULL, 0);
 651         if (threads == (void *) -1) {
 652                 perror("shmat");
 653                 return 1;
 654         }
 655
 656         atexit(free_shm);
 657         return 0;
 658 }
 659
 660 static void usage(void)
 661 {
 662         printf("%s\n", fio_version_string);
 663         printf("\t--output\tWrite output to file\n");
 664         printf("\t--timeout\tRuntime in seconds\n");
 665         printf("\t--latency-log\tGenerate per-job latency logs\n");
 666         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
 667         printf("\t--minimal\tMinimal (terse) output\n");
 668         printf("\t--version\tPrint version info and exit\n");
 669         printf("\t--help\t\tPrint this page\n");
 670         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
 671 }
 672
 673 static int parse_cmd_line(int argc, char *argv[])
 674 {
 675         struct thread_data *td = NULL;
 676         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
 677
 678         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
 679                 switch (c) {
 680                 case 't':
 681                         def_timeout = atoi(optarg);
 682                         break;
 683                 case 'l':
 684                         write_lat_log = 1;
 685                         break;
 686                 case 'w':
 687                         write_bw_log = 1;
 688                         break;
 689                 case 'o':
 690                         f_out = fopen(optarg, "w+");
 691                         if (!f_out) {
 692                                 perror("fopen output");
 693                                 exit(1);
 694                         }
 695                         f_err = f_out;
 696                         break;
 697                 case 'm':
 698                         terse_output = 1;
 699                         break;
 700                 case 'h':
 701                         usage();
 702                         exit(0);
 703                 case 'c':
 704                         exit(fio_show_option_help(optarg));
 705                 case 'v':
 706                         printf("%s\n", fio_version_string);
 707                         exit(0);
 708                 case FIO_GETOPT_JOB: {
 709                         const char *opt = long_options[lidx].name;
 710                         char *val = optarg;
 711
 712                         if (!strncmp(opt, "name", 4) && td) {
 713                                 ret = add_job(td, td->name ?: "fio", 0);
 714                                 if (ret) {
 715                                         put_job(td);
 716                                         return 0;
 717                                 }
 718                                 td = NULL;
 719                         }
 720                         if (!td) {
 721                                 int global = !strncmp(val, "global", 6);
 722
 723                                 td = get_new_job(global, &def_thread);
 724                                 if (!td)
 725                                         return 0;
 726                         }
 727
 728                         ret = fio_cmd_option_parse(td, opt, val);
 729                         if (ret)
 730                                 dont_add_job = 1;
 731                         break;
 732                 }
 733                 default:
 734                         break;
 735                 }
 736         }
 737
 738         if (td) {
 739                 if (dont_add_job)
 740                         put_job(td);
 741                 else {
 742                         ret = add_job(td, td->name ?: "fio", 0);
 743                         if (ret)
 744                                 put_job(td);
 745                 }
 746         }
 747
 748         while (optind < argc) {
 749                 ini_idx++;
 750                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
 751                 ini_file[ini_idx - 1] = strdup(argv[optind]);
 752                 optind++;
 753         }
 754
 755         return ini_idx;
 756 }
 757
 758
 759 int parse_options(int argc, char *argv[])
 760 {
 761         int job_files, i;
 762
 763         f_out = stdout;
 764         f_err = stderr;
 765
 766         fio_options_dup_and_init(long_options);
 767
 768         if (setup_thread_area())
 769                 return 1;
 770         if (fill_def_thread())
 771                 return 1;
 772
 773         job_files = parse_cmd_line(argc, argv);
 774
 775         for (i = 0; i < job_files; i++) {
 776                 if (fill_def_thread())
 777                         return 1;
 778                 if (parse_jobs_ini(ini_file[i], i))
 779                         return 1;
 780                 free(ini_file[i]);
 781         }
 782
 783         free(ini_file);
 784
 785         if (!thread_number) {
 786                 log_err("No jobs defined(s)\n");
 787                 return 1;
 788         }
 789
 790         return 0;
 791 }