fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "os.h"
  39
  40 static unsigned long page_mask;
  41 #define ALIGN(buf)      \
  42         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  43
  44 int groupid = 0;
  45 int thread_number = 0;
  46 int shm_id = 0;
  47 int temp_stall_ts;
  48
  49 static volatile int startup_sem;
  50 static volatile int fio_abort;
  51 static int exit_value;
  52
  53 struct io_log *agg_io_log[2];
  54
  55 #define TERMINATE_ALL           (-1)
  56 #define JOB_START_TIMEOUT       (5 * 1000)
  57
  58 static inline void td_set_runstate(struct thread_data *td, int runstate)
  59 {
  60         td->runstate = runstate;
  61 }
  62
  63 static void terminate_threads(int group_id, int forced_kill)
  64 {
  65         struct thread_data *td;
  66         int i;
  67
  68         for_each_td(td, i) {
  69                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  70                         td->terminate = 1;
  71                         td->start_delay = 0;
  72                         if (forced_kill)
  73                                 td_set_runstate(td, TD_EXITED);
  74                 }
  75         }
  76 }
  77
  78 static void sig_handler(int sig)
  79 {
  80         switch (sig) {
  81                 case SIGALRM:
  82                         update_io_ticks();
  83                         disk_util_timer_arm();
  84                         print_thread_status();
  85                         break;
  86                 default:
  87                         printf("\nfio: terminating on signal %d\n", sig);
  88                         fflush(stdout);
  89                         terminate_threads(TERMINATE_ALL, 0);
  90                         break;
  91         }
  92 }
  93
  94 /*
  95  * Check if we are above the minimum rate given.
  96  */
  97 static int check_min_rate(struct thread_data *td, struct timeval *now)
  98 {
  99         unsigned long spent;
 100         unsigned long rate;
 101         int ddir = td->ddir;
 102
 103         /*
 104          * allow a 2 second settle period in the beginning
 105          */
 106         if (mtime_since(&td->start, now) < 2000)
 107                 return 0;
 108
 109         /*
 110          * if rate blocks is set, sample is running
 111          */
 112         if (td->rate_bytes) {
 113                 spent = mtime_since(&td->lastrate, now);
 114                 if (spent < td->ratecycle)
 115                         return 0;
 116
 117                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
 118                 if (rate < td->ratemin) {
 119                         fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
 120                         return 1;
 121                 }
 122         }
 123
 124         td->rate_bytes = td->this_io_bytes[ddir];
 125         memcpy(&td->lastrate, now, sizeof(*now));
 126         return 0;
 127 }
 128
 129 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 130 {
 131         if (!td->timeout)
 132                 return 0;
 133         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 134                 return 1;
 135
 136         return 0;
 137 }
 138
 139 /*
 140  * When job exits, we can cancel the in-flight IO if we are using async
 141  * io. Attempt to do so.
 142  */
 143 static void cleanup_pending_aio(struct thread_data *td)
 144 {
 145         struct list_head *entry, *n;
 146         struct io_u *io_u;
 147         int r;
 148
 149         /*
 150          * get immediately available events, if any
 151          */
 152         r = io_u_queued_complete(td, 0, NULL);
 153         if (r < 0)
 154                 return;
 155
 156         /*
 157          * now cancel remaining active events
 158          */
 159         if (td->io_ops->cancel) {
 160                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 161                         io_u = list_entry(entry, struct io_u, list);
 162
 163                         r = td->io_ops->cancel(td, io_u);
 164                         if (!r)
 165                                 put_io_u(td, io_u);
 166                 }
 167         }
 168
 169         if (td->cur_depth)
 170                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 171 }
 172
 173 /*
 174  * Helper to handle the final sync of a file. Works just like the normal
 175  * io path, just does everything sync.
 176  */
 177 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 178 {
 179         struct io_u *io_u = __get_io_u(td);
 180         int ret;
 181
 182         if (!io_u)
 183                 return 1;
 184
 185         io_u->ddir = DDIR_SYNC;
 186         io_u->file = f;
 187
 188         if (td_io_prep(td, io_u)) {
 189                 put_io_u(td, io_u);
 190                 return 1;
 191         }
 192
 193 requeue:
 194         ret = td_io_queue(td, io_u);
 195         if (ret < 0) {
 196                 td_verror(td, io_u->error);
 197                 put_io_u(td, io_u);
 198                 return 1;
 199         } else if (ret == FIO_Q_QUEUED) {
 200                 if (io_u_queued_complete(td, 1, NULL) < 0)
 201                         return 1;
 202         } else if (ret == FIO_Q_COMPLETED) {
 203                 if (io_u->error) {
 204                         td_verror(td, io_u->error);
 205                         return 1;
 206                 }
 207
 208                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 209                         return 1;
 210         } else if (ret == FIO_Q_BUSY) {
 211                 if (td_io_commit(td))
 212                         return 1;
 213                 goto requeue;
 214         }
 215
 216         return 0;
 217 }
 218
 219 /*
 220  * The main verify engine. Runs over the writes we previusly submitted,
 221  * reads the blocks back in, and checks the crc/md5 of the data.
 222  */
 223 static void do_verify(struct thread_data *td)
 224 {
 225         struct fio_file *f;
 226         struct io_u *io_u;
 227         int ret, i, min_events;
 228
 229         /*
 230          * sync io first and invalidate cache, to make sure we really
 231          * read from disk.
 232          */
 233         for_each_file(td, f, i) {
 234                 if (fio_io_sync(td, f))
 235                         break;
 236                 if (file_invalidate_cache(td, f))
 237                         break;
 238         }
 239
 240         if (td->error)
 241                 return;
 242
 243         td_set_runstate(td, TD_VERIFYING);
 244
 245         io_u = NULL;
 246         while (!td->terminate) {
 247                 io_u = __get_io_u(td);
 248                 if (!io_u)
 249                         break;
 250
 251                 if (runtime_exceeded(td, &io_u->start_time))
 252                         break;
 253
 254                 if (get_next_verify(td, io_u))
 255                         break;
 256
 257                 if (td_io_prep(td, io_u))
 258                         break;
 259
 260 requeue:
 261                 ret = td_io_queue(td, io_u);
 262
 263                 switch (ret) {
 264                 case FIO_Q_COMPLETED:
 265                         if (io_u->error)
 266                                 ret = -io_u->error;
 267                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 268                                 int bytes = io_u->xfer_buflen - io_u->resid;
 269
 270                                 io_u->xfer_buflen = io_u->resid;
 271                                 io_u->xfer_buf += bytes;
 272                                 goto requeue;
 273                         }
 274                         ret = io_u_sync_complete(td, io_u, verify_io_u);
 275                         if (ret < 0)
 276                                 break;
 277                         continue;
 278                 case FIO_Q_QUEUED:
 279                         break;
 280                 case FIO_Q_BUSY:
 281                         requeue_io_u(td, &io_u);
 282                         ret = td_io_commit(td);
 283                         break;
 284                 default:
 285                         assert(ret < 0);
 286                         td_verror(td, -ret);
 287                         break;
 288                 }
 289
 290                 if (ret < 0 || td->error)
 291                         break;
 292
 293                 /*
 294                  * if we can queue more, do so. but check if there are
 295                  * completed io_u's first.
 296                  */
 297                 min_events = 0;
 298                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 299                         min_events = 1;
 300
 301                         if (td->cur_depth > td->iodepth_low)
 302                                 min_events = td->cur_depth - td->iodepth_low;
 303                 }
 304
 305                 /*
 306                  * Reap required number of io units, if any, and do the
 307                  * verification on them through the callback handler
 308                  */
 309                 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
 310                         break;
 311         }
 312
 313         if (td->cur_depth)
 314                 cleanup_pending_aio(td);
 315
 316         td_set_runstate(td, TD_RUNNING);
 317 }
 318
 319 /*
 320  * Not really an io thread, all it does is burn CPU cycles in the specified
 321  * manner.
 322  */
 323 static void do_cpuio(struct thread_data *td)
 324 {
 325         struct timeval e;
 326         int split = 100 / td->cpuload;
 327         int i = 0;
 328
 329         while (!td->terminate) {
 330                 fio_gettime(&e, NULL);
 331
 332                 if (runtime_exceeded(td, &e))
 333                         break;
 334
 335                 if (!(i % split))
 336                         __usec_sleep(10000);
 337                 else
 338                         usec_sleep(td, 10000);
 339
 340                 i++;
 341         }
 342 }
 343
 344 /*
 345  * Main IO worker function. It retrieves io_u's to process and queues
 346  * and reaps them, checking for rate and errors along the way.
 347  */
 348 static void do_io(struct thread_data *td)
 349 {
 350         struct timeval s;
 351         unsigned long usec;
 352         int i, ret = 0;
 353
 354         td_set_runstate(td, TD_RUNNING);
 355
 356         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 357                 struct timeval comp_time;
 358                 long bytes_done = 0;
 359                 int min_evts = 0;
 360                 struct io_u *io_u;
 361
 362                 if (td->terminate)
 363                         break;
 364
 365                 io_u = get_io_u(td);
 366                 if (!io_u)
 367                         break;
 368
 369                 memcpy(&s, &io_u->start_time, sizeof(s));
 370
 371                 if (runtime_exceeded(td, &s)) {
 372                         put_io_u(td, io_u);
 373                         break;
 374                 }
 375 requeue:
 376                 ret = td_io_queue(td, io_u);
 377
 378                 switch (ret) {
 379                 case FIO_Q_COMPLETED:
 380                         if (io_u->error) {
 381                                 ret = io_u->error;
 382                                 break;
 383                         }
 384                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 385                                 int bytes = io_u->xfer_buflen - io_u->resid;
 386
 387                                 io_u->xfer_buflen = io_u->resid;
 388                                 io_u->xfer_buf += bytes;
 389                                 goto requeue;
 390                         }
 391                         fio_gettime(&comp_time, NULL);
 392                         bytes_done = io_u_sync_complete(td, io_u, NULL);
 393                         if (bytes_done < 0)
 394                                 ret = bytes_done;
 395                         break;
 396                 case FIO_Q_QUEUED:
 397                         /*
 398                          * if the engine doesn't have a commit hook,
 399                          * the io_u is really queued. if it does have such
 400                          * a hook, it has to call io_u_queued() itself.
 401                          */
 402                         if (td->io_ops->commit == NULL)
 403                                 io_u_queued(td, io_u);
 404                         break;
 405                 case FIO_Q_BUSY:
 406                         requeue_io_u(td, &io_u);
 407                         ret = td_io_commit(td);
 408                         break;
 409                 default:
 410                         assert(ret < 0);
 411                         put_io_u(td, io_u);
 412                         break;
 413                 }
 414
 415                 if (ret < 0 || td->error)
 416                         break;
 417
 418                 /*
 419                  * See if we need to complete some commands
 420                  */
 421                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 422                         min_evts = 0;
 423                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 424                                 min_evts = 1;
 425
 426                                 if (td->cur_depth > td->iodepth_low)
 427                                         min_evts = td->cur_depth - td->iodepth_low;
 428                         }
 429
 430                         fio_gettime(&comp_time, NULL);
 431                         bytes_done = io_u_queued_complete(td, min_evts, NULL);
 432                         if (bytes_done < 0)
 433                                 break;
 434                 }
 435
 436                 if (!bytes_done)
 437                         continue;
 438
 439                 /*
 440                  * the rate is batched for now, it should work for batches
 441                  * of completions except the very first one which may look
 442                  * a little bursty
 443                  */
 444                 usec = utime_since(&s, &comp_time);
 445
 446                 rate_throttle(td, usec, bytes_done, td->ddir);
 447
 448                 if (check_min_rate(td, &comp_time)) {
 449                         if (exitall_on_terminate)
 450                                 terminate_threads(td->groupid, 0);
 451                         td_verror(td, ENODATA);
 452                         break;
 453                 }
 454
 455                 if (td->thinktime) {
 456                         unsigned long long b;
 457
 458                         b = td->io_blocks[0] + td->io_blocks[1];
 459                         if (!(b % td->thinktime_blocks)) {
 460                                 int left;
 461
 462                                 if (td->thinktime_spin)
 463                                         __usec_sleep(td->thinktime_spin);
 464
 465                                 left = td->thinktime - td->thinktime_spin;
 466                                 if (left)
 467                                         usec_sleep(td, left);
 468                         }
 469                 }
 470         }
 471
 472         if (!td->error) {
 473                 struct fio_file *f;
 474
 475                 if (td->cur_depth)
 476                         cleanup_pending_aio(td);
 477
 478                 if (should_fsync(td) && td->end_fsync) {
 479                         td_set_runstate(td, TD_FSYNCING);
 480                         for_each_file(td, f, i)
 481                                 fio_io_sync(td, f);
 482                 }
 483         }
 484 }
 485
 486 static void cleanup_io_u(struct thread_data *td)
 487 {
 488         struct list_head *entry, *n;
 489         struct io_u *io_u;
 490
 491         list_for_each_safe(entry, n, &td->io_u_freelist) {
 492                 io_u = list_entry(entry, struct io_u, list);
 493
 494                 list_del(&io_u->list);
 495                 free(io_u);
 496         }
 497
 498         free_io_mem(td);
 499 }
 500
 501 /*
 502  * "randomly" fill the buffer contents
 503  */
 504 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 505 {
 506         int *ptr = io_u->buf;
 507
 508         while ((void *) ptr - io_u->buf < max_bs) {
 509                 *ptr = rand() * 0x9e370001;
 510                 ptr++;
 511         }
 512 }
 513
 514 static int init_io_u(struct thread_data *td)
 515 {
 516         struct io_u *io_u;
 517         unsigned int max_bs;
 518         int i, max_units;
 519         char *p;
 520
 521         if (td->io_ops->flags & FIO_CPUIO)
 522                 return 0;
 523
 524         if (td->io_ops->flags & FIO_SYNCIO)
 525                 max_units = 1;
 526         else
 527                 max_units = td->iodepth;
 528
 529         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 530         td->orig_buffer_size = max_bs * max_units;
 531
 532         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 533                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 534         else
 535                 td->orig_buffer_size += page_mask;
 536
 537         if (allocate_io_mem(td))
 538                 return 1;
 539
 540         p = ALIGN(td->orig_buffer);
 541         for (i = 0; i < max_units; i++) {
 542                 io_u = malloc(sizeof(*io_u));
 543                 memset(io_u, 0, sizeof(*io_u));
 544                 INIT_LIST_HEAD(&io_u->list);
 545
 546                 io_u->buf = p + max_bs * i;
 547                 if (td_write(td) || td_rw(td))
 548                         fill_rand_buf(io_u, max_bs);
 549
 550                 io_u->index = i;
 551                 list_add(&io_u->list, &td->io_u_freelist);
 552         }
 553
 554         return 0;
 555 }
 556
 557 static int switch_ioscheduler(struct thread_data *td)
 558 {
 559         char tmp[256], tmp2[128];
 560         FILE *f;
 561         int ret;
 562
 563         if (td->io_ops->flags & FIO_CPUIO)
 564                 return 0;
 565
 566         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 567
 568         f = fopen(tmp, "r+");
 569         if (!f) {
 570                 td_verror(td, errno);
 571                 return 1;
 572         }
 573
 574         /*
 575          * Set io scheduler.
 576          */
 577         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 578         if (ferror(f) || ret != 1) {
 579                 td_verror(td, errno);
 580                 fclose(f);
 581                 return 1;
 582         }
 583
 584         rewind(f);
 585
 586         /*
 587          * Read back and check that the selected scheduler is now the default.
 588          */
 589         ret = fread(tmp, 1, sizeof(tmp), f);
 590         if (ferror(f) || ret < 0) {
 591                 td_verror(td, errno);
 592                 fclose(f);
 593                 return 1;
 594         }
 595
 596         sprintf(tmp2, "[%s]", td->ioscheduler);
 597         if (!strstr(tmp, tmp2)) {
 598                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 599                 td_verror(td, EINVAL);
 600                 fclose(f);
 601                 return 1;
 602         }
 603
 604         fclose(f);
 605         return 0;
 606 }
 607
 608 static void clear_io_state(struct thread_data *td)
 609 {
 610         struct fio_file *f;
 611         int i;
 612
 613         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 614         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 615         td->zone_bytes = 0;
 616
 617         td->last_was_sync = 0;
 618
 619         for_each_file(td, f, i) {
 620                 f->last_completed_pos = 0;
 621
 622                 f->last_pos = 0;
 623                 if (td->io_ops->flags & FIO_SYNCIO)
 624                         lseek(f->fd, SEEK_SET, 0);
 625
 626                 if (f->file_map)
 627                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 628         }
 629 }
 630
 631 /*
 632  * Entry point for the thread based jobs. The process based jobs end up
 633  * here as well, after a little setup.
 634  */
 635 static void *thread_main(void *data)
 636 {
 637         unsigned long long runtime[2];
 638         struct thread_data *td = data;
 639
 640         if (!td->use_thread)
 641                 setsid();
 642
 643         td->pid = getpid();
 644
 645         INIT_LIST_HEAD(&td->io_u_freelist);
 646         INIT_LIST_HEAD(&td->io_u_busylist);
 647         INIT_LIST_HEAD(&td->io_u_requeues);
 648         INIT_LIST_HEAD(&td->io_hist_list);
 649         INIT_LIST_HEAD(&td->io_log_list);
 650
 651         if (init_io_u(td))
 652                 goto err;
 653
 654         if (fio_setaffinity(td) == -1) {
 655                 td_verror(td, errno);
 656                 goto err;
 657         }
 658
 659         if (init_iolog(td))
 660                 goto err;
 661
 662         if (td->ioprio) {
 663                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 664                         td_verror(td, errno);
 665                         goto err;
 666                 }
 667         }
 668
 669         if (nice(td->nice) == -1) {
 670                 td_verror(td, errno);
 671                 goto err;
 672         }
 673
 674         if (init_random_state(td))
 675                 goto err;
 676
 677         if (td->ioscheduler && switch_ioscheduler(td))
 678                 goto err;
 679
 680         td_set_runstate(td, TD_INITIALIZED);
 681         fio_sem_up(&startup_sem);
 682         fio_sem_down(&td->mutex);
 683
 684         if (!td->create_serialize && setup_files(td))
 685                 goto err;
 686         if (open_files(td))
 687                 goto err;
 688
 689         /*
 690          * Do this late, as some IO engines would like to have the
 691          * files setup prior to initializing structures.
 692          */
 693         if (td_io_init(td))
 694                 goto err;
 695
 696         if (td->exec_prerun) {
 697                 if (system(td->exec_prerun) < 0)
 698                         goto err;
 699         }
 700
 701         fio_gettime(&td->epoch, NULL);
 702         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 703
 704         runtime[0] = runtime[1] = 0;
 705         while (td->loops--) {
 706                 fio_gettime(&td->start, NULL);
 707                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 708
 709                 if (td->ratemin)
 710                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 711
 712                 clear_io_state(td);
 713                 prune_io_piece_log(td);
 714
 715                 if (td->io_ops->flags & FIO_CPUIO)
 716                         do_cpuio(td);
 717                 else
 718                         do_io(td);
 719
 720                 runtime[td->ddir] += utime_since_now(&td->start);
 721                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 722                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 723
 724                 if (td->error || td->terminate)
 725                         break;
 726
 727                 if (td->verify == VERIFY_NONE)
 728                         continue;
 729
 730                 clear_io_state(td);
 731                 fio_gettime(&td->start, NULL);
 732
 733                 do_verify(td);
 734
 735                 runtime[DDIR_READ] += utime_since_now(&td->start);
 736
 737                 if (td->error || td->terminate)
 738                         break;
 739         }
 740
 741         update_rusage_stat(td);
 742         fio_gettime(&td->end_time, NULL);
 743         td->runtime[0] = runtime[0] / 1000;
 744         td->runtime[1] = runtime[1] / 1000;
 745
 746         if (td->ts.bw_log)
 747                 finish_log(td, td->ts.bw_log, "bw");
 748         if (td->ts.slat_log)
 749                 finish_log(td, td->ts.slat_log, "slat");
 750         if (td->ts.clat_log)
 751                 finish_log(td, td->ts.clat_log, "clat");
 752         if (td->write_iolog_file)
 753                 write_iolog_close(td);
 754         if (td->exec_postrun) {
 755                 if (system(td->exec_postrun) < 0)
 756                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 757         }
 758
 759         if (exitall_on_terminate)
 760                 terminate_threads(td->groupid, 0);
 761
 762 err:
 763         if (td->error)
 764                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 765         close_files(td);
 766         close_ioengine(td);
 767         cleanup_io_u(td);
 768         td_set_runstate(td, TD_EXITED);
 769         return (void *) (unsigned long) td->error;
 770 }
 771
 772 /*
 773  * We cannot pass the td data into a forked process, so attach the td and
 774  * pass it to the thread worker.
 775  */
 776 static int fork_main(int shmid, int offset)
 777 {
 778         struct thread_data *td;
 779         void *data, *ret;
 780
 781         data = shmat(shmid, NULL, 0);
 782         if (data == (void *) -1) {
 783                 int __err = errno;
 784
 785                 perror("shmat");
 786                 return __err;
 787         }
 788
 789         td = data + offset * sizeof(struct thread_data);
 790         ret = thread_main(td);
 791         shmdt(data);
 792         return (int) (unsigned long) ret;
 793 }
 794
 795 /*
 796  * Run over the job map and reap the threads that have exited, if any.
 797  */
 798 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 799 {
 800         struct thread_data *td;
 801         int i, cputhreads, pending, status, ret;
 802
 803         /*
 804          * reap exited threads (TD_EXITED -> TD_REAPED)
 805          */
 806         pending = cputhreads = 0;
 807         for_each_td(td, i) {
 808                 /*
 809                  * ->io_ops is NULL for a thread that has closed its
 810                  * io engine
 811                  */
 812                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 813                         cputhreads++;
 814
 815                 if (td->runstate < TD_EXITED) {
 816                         /*
 817                          * check if someone quit or got killed in an unusual way
 818                          */
 819                         ret = waitpid(td->pid, &status, WNOHANG);
 820                         if (ret < 0)
 821                                 perror("waitpid");
 822                         else if ((ret == td->pid) && WIFSIGNALED(status)) {
 823                                 int sig = WTERMSIG(status);
 824
 825                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 826                                 td_set_runstate(td, TD_REAPED);
 827                                 goto reaped;
 828                         }
 829                 }
 830
 831                 if (td->runstate != TD_EXITED) {
 832                         if (td->runstate < TD_RUNNING)
 833                                 pending++;
 834
 835                         continue;
 836                 }
 837
 838                 if (td->error)
 839                         exit_value++;
 840
 841                 td_set_runstate(td, TD_REAPED);
 842
 843                 if (td->use_thread) {
 844                         long ret;
 845
 846                         if (pthread_join(td->thread, (void *) &ret))
 847                                 perror("thread_join");
 848                 } else {
 849                         int status;
 850
 851                         ret = waitpid(td->pid, &status, 0);
 852                         if (ret < 0)
 853                                 perror("waitpid");
 854                         else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 855                                 if (!exit_value)
 856                                         exit_value++;
 857                         }
 858                 }
 859
 860 reaped:
 861                 (*nr_running)--;
 862                 (*m_rate) -= td->ratemin;
 863                 (*t_rate) -= td->rate;
 864         }
 865
 866         if (*nr_running == cputhreads && !pending)
 867                 terminate_threads(TERMINATE_ALL, 0);
 868 }
 869
 870 /*
 871  * Main function for kicking off and reaping jobs, as needed.
 872  */
 873 static void run_threads(void)
 874 {
 875         struct thread_data *td;
 876         unsigned long spent;
 877         int i, todo, nr_running, m_rate, t_rate, nr_started;
 878
 879         if (fio_pin_memory())
 880                 return;
 881
 882         if (!terse_output) {
 883                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 884                 fflush(stdout);
 885         }
 886
 887         signal(SIGINT, sig_handler);
 888         signal(SIGALRM, sig_handler);
 889
 890         todo = thread_number;
 891         nr_running = 0;
 892         nr_started = 0;
 893         m_rate = t_rate = 0;
 894
 895         for_each_td(td, i) {
 896                 print_status_init(td->thread_number - 1);
 897
 898                 if (!td->create_serialize) {
 899                         init_disk_util(td);
 900                         continue;
 901                 }
 902
 903                 /*
 904                  * do file setup here so it happens sequentially,
 905                  * we don't want X number of threads getting their
 906                  * client data interspersed on disk
 907                  */
 908                 if (setup_files(td)) {
 909                         exit_value++;
 910                         if (td->error)
 911                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 912                         td_set_runstate(td, TD_REAPED);
 913                         todo--;
 914                 }
 915
 916                 init_disk_util(td);
 917         }
 918
 919         while (todo) {
 920                 struct thread_data *map[MAX_JOBS];
 921                 struct timeval this_start;
 922                 int this_jobs = 0, left;
 923
 924                 /*
 925                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 926                  */
 927                 for_each_td(td, i) {
 928                         if (td->runstate != TD_NOT_CREATED)
 929                                 continue;
 930
 931                         /*
 932                          * never got a chance to start, killed by other
 933                          * thread for some reason
 934                          */
 935                         if (td->terminate) {
 936                                 todo--;
 937                                 continue;
 938                         }
 939
 940                         if (td->start_delay) {
 941                                 spent = mtime_since_genesis();
 942
 943                                 if (td->start_delay * 1000 > spent)
 944                                         continue;
 945                         }
 946
 947                         if (td->stonewall && (nr_started || nr_running))
 948                                 break;
 949
 950                         /*
 951                          * Set state to created. Thread will transition
 952                          * to TD_INITIALIZED when it's done setting up.
 953                          */
 954                         td_set_runstate(td, TD_CREATED);
 955                         map[this_jobs++] = td;
 956                         fio_sem_init(&startup_sem, 1);
 957                         nr_started++;
 958
 959                         if (td->use_thread) {
 960                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 961                                         perror("thread_create");
 962                                         nr_started--;
 963                                 }
 964                         } else {
 965                                 if (fork())
 966                                         fio_sem_down(&startup_sem);
 967                                 else {
 968                                         int ret = fork_main(shm_id, i);
 969
 970                                         exit(ret);
 971                                 }
 972                         }
 973                 }
 974
 975                 /*
 976                  * Wait for the started threads to transition to
 977                  * TD_INITIALIZED.
 978                  */
 979                 fio_gettime(&this_start, NULL);
 980                 left = this_jobs;
 981                 while (left && !fio_abort) {
 982                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 983                                 break;
 984
 985                         usleep(100000);
 986
 987                         for (i = 0; i < this_jobs; i++) {
 988                                 td = map[i];
 989                                 if (!td)
 990                                         continue;
 991                                 if (td->runstate == TD_INITIALIZED) {
 992                                         map[i] = NULL;
 993                                         left--;
 994                                 } else if (td->runstate >= TD_EXITED) {
 995                                         map[i] = NULL;
 996                                         left--;
 997                                         todo--;
 998                                         nr_running++; /* work-around... */
 999                                 }
1000                         }
1001                 }
1002
1003                 if (left) {
1004                         log_err("fio: %d jobs failed to start\n", left);
1005                         for (i = 0; i < this_jobs; i++) {
1006                                 td = map[i];
1007                                 if (!td)
1008                                         continue;
1009                                 kill(td->pid, SIGTERM);
1010                         }
1011                         break;
1012                 }
1013
1014                 /*
1015                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1016                  */
1017                 for_each_td(td, i) {
1018                         if (td->runstate != TD_INITIALIZED)
1019                                 continue;
1020
1021                         td_set_runstate(td, TD_RUNNING);
1022                         nr_running++;
1023                         nr_started--;
1024                         m_rate += td->ratemin;
1025                         t_rate += td->rate;
1026                         todo--;
1027                         fio_sem_up(&td->mutex);
1028                 }
1029
1030                 reap_threads(&nr_running, &t_rate, &m_rate);
1031
1032                 if (todo)
1033                         usleep(100000);
1034         }
1035
1036         while (nr_running) {
1037                 reap_threads(&nr_running, &t_rate, &m_rate);
1038                 usleep(10000);
1039         }
1040
1041         update_io_ticks();
1042         fio_unpin_memory();
1043 }
1044
1045 int main(int argc, char *argv[])
1046 {
1047         long ps;
1048
1049         /*
1050          * We need locale for number printing, if it isn't set then just
1051          * go with the US format.
1052          */
1053         if (!getenv("LC_NUMERIC"))
1054                 setlocale(LC_NUMERIC, "en_US");
1055
1056         if (parse_options(argc, argv))
1057                 return 1;
1058
1059         if (!thread_number) {
1060                 log_err("Nothing to do\n");
1061                 return 1;
1062         }
1063
1064         ps = sysconf(_SC_PAGESIZE);
1065         if (ps < 0) {
1066                 log_err("Failed to get page size\n");
1067                 return 1;
1068         }
1069
1070         page_mask = ps - 1;
1071
1072         if (write_bw_log) {
1073                 setup_log(&agg_io_log[DDIR_READ]);
1074                 setup_log(&agg_io_log[DDIR_WRITE]);
1075         }
1076
1077         disk_util_timer_arm();
1078
1079         run_threads();
1080
1081         if (!fio_abort) {
1082                 show_run_stats();
1083                 if (write_bw_log) {
1084                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1085                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1086                 }
1087         }
1088
1089         return exit_value;
1090 }