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))
 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)
 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))
 310                         break;
 311         }
 312
 313         if (io_u)
 314                 put_io_u(td, io_u);
 315
 316         if (td->cur_depth)
 317                 cleanup_pending_aio(td);
 318
 319         td_set_runstate(td, TD_RUNNING);
 320 }
 321
 322 /*
 323  * Not really an io thread, all it does is burn CPU cycles in the specified
 324  * manner.
 325  */
 326 static void do_cpuio(struct thread_data *td)
 327 {
 328         struct timeval e;
 329         int split = 100 / td->cpuload;
 330         int i = 0;
 331
 332         while (!td->terminate) {
 333                 fio_gettime(&e, NULL);
 334
 335                 if (runtime_exceeded(td, &e))
 336                         break;
 337
 338                 if (!(i % split))
 339                         __usec_sleep(10000);
 340                 else
 341                         usec_sleep(td, 10000);
 342
 343                 i++;
 344         }
 345 }
 346
 347 /*
 348  * Main IO worker function. It retrieves io_u's to process and queues
 349  * and reaps them, checking for rate and errors along the way.
 350  */
 351 static void do_io(struct thread_data *td)
 352 {
 353         struct timeval s;
 354         unsigned long usec;
 355         int i, ret = 0;
 356
 357         td_set_runstate(td, TD_RUNNING);
 358
 359         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 360                 struct timeval comp_time;
 361                 long bytes_done = 0;
 362                 int min_evts = 0;
 363                 struct io_u *io_u;
 364
 365                 if (td->terminate)
 366                         break;
 367
 368                 io_u = get_io_u(td);
 369                 if (!io_u)
 370                         break;
 371
 372                 memcpy(&s, &io_u->start_time, sizeof(s));
 373
 374                 if (runtime_exceeded(td, &s)) {
 375                         put_io_u(td, io_u);
 376                         break;
 377                 }
 378 requeue:
 379                 ret = td_io_queue(td, io_u);
 380
 381                 switch (ret) {
 382                 case FIO_Q_COMPLETED:
 383                         if (io_u->error) {
 384                                 ret = io_u->error;
 385                                 break;
 386                         }
 387                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 388                                 int bytes = io_u->xfer_buflen - io_u->resid;
 389
 390                                 io_u->xfer_buflen = io_u->resid;
 391                                 io_u->xfer_buf += bytes;
 392                                 goto requeue;
 393                         }
 394                         fio_gettime(&comp_time, NULL);
 395                         bytes_done = io_u_sync_complete(td, io_u, NULL);
 396                         if (bytes_done < 0)
 397                                 ret = bytes_done;
 398                         break;
 399                 case FIO_Q_QUEUED:
 400                         /*
 401                          * if the engine doesn't have a commit hook,
 402                          * the io_u is really queued. if it does have such
 403                          * a hook, it has to call io_u_queued() itself.
 404                          */
 405                         if (td->io_ops->commit == NULL)
 406                                 io_u_queued(td, io_u);
 407                         break;
 408                 case FIO_Q_BUSY:
 409                         requeue_io_u(td, &io_u);
 410                         ret = td_io_commit(td);
 411                         break;
 412                 default:
 413                         assert(ret < 0);
 414                         put_io_u(td, io_u);
 415                         break;
 416                 }
 417
 418                 if (ret < 0 || td->error)
 419                         break;
 420
 421                 /*
 422                  * See if we need to complete some commands
 423                  */
 424                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 425                         min_evts = 0;
 426                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 427                                 min_evts = 1;
 428
 429                                 if (td->cur_depth > td->iodepth_low)
 430                                         min_evts = td->cur_depth - td->iodepth_low;
 431                         }
 432
 433                         fio_gettime(&comp_time, NULL);
 434                         bytes_done = io_u_queued_complete(td, min_evts, NULL);
 435                         if (bytes_done < 0)
 436                                 break;
 437                 }
 438
 439                 if (!bytes_done)
 440                         continue;
 441
 442                 /*
 443                  * the rate is batched for now, it should work for batches
 444                  * of completions except the very first one which may look
 445                  * a little bursty
 446                  */
 447                 usec = utime_since(&s, &comp_time);
 448
 449                 rate_throttle(td, usec, bytes_done, td->ddir);
 450
 451                 if (check_min_rate(td, &comp_time)) {
 452                         if (exitall_on_terminate)
 453                                 terminate_threads(td->groupid, 0);
 454                         td_verror(td, ENODATA);
 455                         break;
 456                 }
 457
 458                 if (td->thinktime) {
 459                         unsigned long long b;
 460
 461                         b = td->io_blocks[0] + td->io_blocks[1];
 462                         if (!(b % td->thinktime_blocks)) {
 463                                 int left;
 464
 465                                 if (td->thinktime_spin)
 466                                         __usec_sleep(td->thinktime_spin);
 467
 468                                 left = td->thinktime - td->thinktime_spin;
 469                                 if (left)
 470                                         usec_sleep(td, left);
 471                         }
 472                 }
 473         }
 474
 475         if (!td->error) {
 476                 struct fio_file *f;
 477
 478                 if (td->cur_depth)
 479                         cleanup_pending_aio(td);
 480
 481                 if (should_fsync(td) && td->end_fsync) {
 482                         td_set_runstate(td, TD_FSYNCING);
 483                         for_each_file(td, f, i)
 484                                 fio_io_sync(td, f);
 485                 }
 486         }
 487 }
 488
 489 static void cleanup_io_u(struct thread_data *td)
 490 {
 491         struct list_head *entry, *n;
 492         struct io_u *io_u;
 493
 494         list_for_each_safe(entry, n, &td->io_u_freelist) {
 495                 io_u = list_entry(entry, struct io_u, list);
 496
 497                 list_del(&io_u->list);
 498                 free(io_u);
 499         }
 500
 501         free_io_mem(td);
 502 }
 503
 504 /*
 505  * "randomly" fill the buffer contents
 506  */
 507 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 508 {
 509         int *ptr = io_u->buf;
 510
 511         while ((void *) ptr - io_u->buf < max_bs) {
 512                 *ptr = rand() * 0x9e370001;
 513                 ptr++;
 514         }
 515 }
 516
 517 static int init_io_u(struct thread_data *td)
 518 {
 519         struct io_u *io_u;
 520         unsigned int max_bs;
 521         int i, max_units;
 522         char *p;
 523
 524         if (td->io_ops->flags & FIO_CPUIO)
 525                 return 0;
 526
 527         if (td->io_ops->flags & FIO_SYNCIO)
 528                 max_units = 1;
 529         else
 530                 max_units = td->iodepth;
 531
 532         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 533         td->orig_buffer_size = max_bs * max_units;
 534
 535         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 536                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 537         else
 538                 td->orig_buffer_size += page_mask;
 539
 540         if (allocate_io_mem(td))
 541                 return 1;
 542
 543         p = ALIGN(td->orig_buffer);
 544         for (i = 0; i < max_units; i++) {
 545                 io_u = malloc(sizeof(*io_u));
 546                 memset(io_u, 0, sizeof(*io_u));
 547                 INIT_LIST_HEAD(&io_u->list);
 548
 549                 io_u->buf = p + max_bs * i;
 550                 if (td_write(td) || td_rw(td))
 551                         fill_rand_buf(io_u, max_bs);
 552
 553                 io_u->index = i;
 554                 list_add(&io_u->list, &td->io_u_freelist);
 555         }
 556
 557         return 0;
 558 }
 559
 560 static int switch_ioscheduler(struct thread_data *td)
 561 {
 562         char tmp[256], tmp2[128];
 563         FILE *f;
 564         int ret;
 565
 566         if (td->io_ops->flags & FIO_CPUIO)
 567                 return 0;
 568
 569         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 570
 571         f = fopen(tmp, "r+");
 572         if (!f) {
 573                 td_verror(td, errno);
 574                 return 1;
 575         }
 576
 577         /*
 578          * Set io scheduler.
 579          */
 580         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 581         if (ferror(f) || ret != 1) {
 582                 td_verror(td, errno);
 583                 fclose(f);
 584                 return 1;
 585         }
 586
 587         rewind(f);
 588
 589         /*
 590          * Read back and check that the selected scheduler is now the default.
 591          */
 592         ret = fread(tmp, 1, sizeof(tmp), f);
 593         if (ferror(f) || ret < 0) {
 594                 td_verror(td, errno);
 595                 fclose(f);
 596                 return 1;
 597         }
 598
 599         sprintf(tmp2, "[%s]", td->ioscheduler);
 600         if (!strstr(tmp, tmp2)) {
 601                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 602                 td_verror(td, EINVAL);
 603                 fclose(f);
 604                 return 1;
 605         }
 606
 607         fclose(f);
 608         return 0;
 609 }
 610
 611 static void clear_io_state(struct thread_data *td)
 612 {
 613         struct fio_file *f;
 614         int i;
 615
 616         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 617         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 618         td->zone_bytes = 0;
 619
 620         td->last_was_sync = 0;
 621
 622         for_each_file(td, f, i) {
 623                 f->last_completed_pos = 0;
 624
 625                 f->last_pos = 0;
 626                 if (td->io_ops->flags & FIO_SYNCIO)
 627                         lseek(f->fd, SEEK_SET, 0);
 628
 629                 if (f->file_map)
 630                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 631         }
 632 }
 633
 634 /*
 635  * Entry point for the thread based jobs. The process based jobs end up
 636  * here as well, after a little setup.
 637  */
 638 static void *thread_main(void *data)
 639 {
 640         unsigned long long runtime[2];
 641         struct thread_data *td = data;
 642
 643         if (!td->use_thread)
 644                 setsid();
 645
 646         td->pid = getpid();
 647
 648         INIT_LIST_HEAD(&td->io_u_freelist);
 649         INIT_LIST_HEAD(&td->io_u_busylist);
 650         INIT_LIST_HEAD(&td->io_u_requeues);
 651         INIT_LIST_HEAD(&td->io_hist_list);
 652         INIT_LIST_HEAD(&td->io_log_list);
 653
 654         if (init_io_u(td))
 655                 goto err;
 656
 657         if (fio_setaffinity(td) == -1) {
 658                 td_verror(td, errno);
 659                 goto err;
 660         }
 661
 662         if (init_iolog(td))
 663                 goto err;
 664
 665         if (td->ioprio) {
 666                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 667                         td_verror(td, errno);
 668                         goto err;
 669                 }
 670         }
 671
 672         if (nice(td->nice) == -1) {
 673                 td_verror(td, errno);
 674                 goto err;
 675         }
 676
 677         if (init_random_state(td))
 678                 goto err;
 679
 680         if (td->ioscheduler && switch_ioscheduler(td))
 681                 goto err;
 682
 683         td_set_runstate(td, TD_INITIALIZED);
 684         fio_sem_up(&startup_sem);
 685         fio_sem_down(&td->mutex);
 686
 687         if (!td->create_serialize && setup_files(td))
 688                 goto err;
 689         if (open_files(td))
 690                 goto err;
 691
 692         /*
 693          * Do this late, as some IO engines would like to have the
 694          * files setup prior to initializing structures.
 695          */
 696         if (td_io_init(td))
 697                 goto err;
 698
 699         if (td->exec_prerun) {
 700                 if (system(td->exec_prerun) < 0)
 701                         goto err;
 702         }
 703
 704         fio_gettime(&td->epoch, NULL);
 705         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 706
 707         runtime[0] = runtime[1] = 0;
 708         while (td->loops--) {
 709                 fio_gettime(&td->start, NULL);
 710                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 711
 712                 if (td->ratemin)
 713                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 714
 715                 clear_io_state(td);
 716                 prune_io_piece_log(td);
 717
 718                 if (td->io_ops->flags & FIO_CPUIO)
 719                         do_cpuio(td);
 720                 else
 721                         do_io(td);
 722
 723                 runtime[td->ddir] += utime_since_now(&td->start);
 724                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 725                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 726
 727                 if (td->error || td->terminate)
 728                         break;
 729
 730                 if (td->verify == VERIFY_NONE)
 731                         continue;
 732
 733                 clear_io_state(td);
 734                 fio_gettime(&td->start, NULL);
 735
 736                 do_verify(td);
 737
 738                 runtime[DDIR_READ] += utime_since_now(&td->start);
 739
 740                 if (td->error || td->terminate)
 741                         break;
 742         }
 743
 744         update_rusage_stat(td);
 745         fio_gettime(&td->end_time, NULL);
 746         td->runtime[0] = runtime[0] / 1000;
 747         td->runtime[1] = runtime[1] / 1000;
 748
 749         if (td->ts.bw_log)
 750                 finish_log(td, td->ts.bw_log, "bw");
 751         if (td->ts.slat_log)
 752                 finish_log(td, td->ts.slat_log, "slat");
 753         if (td->ts.clat_log)
 754                 finish_log(td, td->ts.clat_log, "clat");
 755         if (td->write_iolog_file)
 756                 write_iolog_close(td);
 757         if (td->exec_postrun) {
 758                 if (system(td->exec_postrun) < 0)
 759                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 760         }
 761
 762         if (exitall_on_terminate)
 763                 terminate_threads(td->groupid, 0);
 764
 765 err:
 766         if (td->error)
 767                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 768         close_files(td);
 769         close_ioengine(td);
 770         cleanup_io_u(td);
 771         td_set_runstate(td, TD_EXITED);
 772         return (void *) td->error;
 773 }
 774
 775 /*
 776  * We cannot pass the td data into a forked process, so attach the td and
 777  * pass it to the thread worker.
 778  */
 779 static int fork_main(int shmid, int offset)
 780 {
 781         struct thread_data *td;
 782         void *data, *ret;
 783
 784         data = shmat(shmid, NULL, 0);
 785         if (data == (void *) -1) {
 786                 int __err = errno;
 787
 788                 perror("shmat");
 789                 return __err;
 790         }
 791
 792         td = data + offset * sizeof(struct thread_data);
 793         ret = thread_main(td);
 794         shmdt(data);
 795         return (int) ret;
 796 }
 797
 798 /*
 799  * Run over the job map and reap the threads that have exited, if any.
 800  */
 801 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 802 {
 803         struct thread_data *td;
 804         int i, cputhreads, pending, status, ret;
 805
 806         /*
 807          * reap exited threads (TD_EXITED -> TD_REAPED)
 808          */
 809         pending = cputhreads = 0;
 810         for_each_td(td, i) {
 811                 /*
 812                  * ->io_ops is NULL for a thread that has closed its
 813                  * io engine
 814                  */
 815                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 816                         cputhreads++;
 817
 818                 if (td->runstate < TD_EXITED) {
 819                         /*
 820                          * check if someone quit or got killed in an unusual way
 821                          */
 822                         ret = waitpid(td->pid, &status, WNOHANG);
 823                         if (ret < 0)
 824                                 perror("waitpid");
 825                         else if ((ret == td->pid) && WIFSIGNALED(status)) {
 826                                 int sig = WTERMSIG(status);
 827
 828                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 829                                 td_set_runstate(td, TD_REAPED);
 830                                 goto reaped;
 831                         }
 832                 }
 833
 834                 if (td->runstate != TD_EXITED) {
 835                         if (td->runstate < TD_RUNNING)
 836                                 pending++;
 837
 838                         continue;
 839                 }
 840
 841                 if (td->error)
 842                         exit_value++;
 843
 844                 td_set_runstate(td, TD_REAPED);
 845
 846                 if (td->use_thread) {
 847                         long ret;
 848
 849                         if (pthread_join(td->thread, (void *) &ret))
 850                                 perror("thread_join");
 851                 } else {
 852                         int status;
 853
 854                         ret = waitpid(td->pid, &status, 0);
 855                         if (ret < 0)
 856                                 perror("waitpid");
 857                         else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 858                                 if (!exit_value)
 859                                         exit_value++;
 860                         }
 861                 }
 862
 863 reaped:
 864                 (*nr_running)--;
 865                 (*m_rate) -= td->ratemin;
 866                 (*t_rate) -= td->rate;
 867         }
 868
 869         if (*nr_running == cputhreads && !pending)
 870                 terminate_threads(TERMINATE_ALL, 0);
 871 }
 872
 873 /*
 874  * Main function for kicking off and reaping jobs, as needed.
 875  */
 876 static void run_threads(void)
 877 {
 878         struct thread_data *td;
 879         unsigned long spent;
 880         int i, todo, nr_running, m_rate, t_rate, nr_started;
 881
 882         if (fio_pin_memory())
 883                 return;
 884
 885         if (!terse_output) {
 886                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 887                 fflush(stdout);
 888         }
 889
 890         signal(SIGINT, sig_handler);
 891         signal(SIGALRM, sig_handler);
 892
 893         todo = thread_number;
 894         nr_running = 0;
 895         nr_started = 0;
 896         m_rate = t_rate = 0;
 897
 898         for_each_td(td, i) {
 899                 print_status_init(td->thread_number - 1);
 900
 901                 if (!td->create_serialize) {
 902                         init_disk_util(td);
 903                         continue;
 904                 }
 905
 906                 /*
 907                  * do file setup here so it happens sequentially,
 908                  * we don't want X number of threads getting their
 909                  * client data interspersed on disk
 910                  */
 911                 if (setup_files(td)) {
 912                         exit_value++;
 913                         if (td->error)
 914                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 915                         td_set_runstate(td, TD_REAPED);
 916                         todo--;
 917                 }
 918
 919                 init_disk_util(td);
 920         }
 921
 922         while (todo) {
 923                 struct thread_data *map[MAX_JOBS];
 924                 struct timeval this_start;
 925                 int this_jobs = 0, left;
 926
 927                 /*
 928                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 929                  */
 930                 for_each_td(td, i) {
 931                         if (td->runstate != TD_NOT_CREATED)
 932                                 continue;
 933
 934                         /*
 935                          * never got a chance to start, killed by other
 936                          * thread for some reason
 937                          */
 938                         if (td->terminate) {
 939                                 todo--;
 940                                 continue;
 941                         }
 942
 943                         if (td->start_delay) {
 944                                 spent = mtime_since_genesis();
 945
 946                                 if (td->start_delay * 1000 > spent)
 947                                         continue;
 948                         }
 949
 950                         if (td->stonewall && (nr_started || nr_running))
 951                                 break;
 952
 953                         /*
 954                          * Set state to created. Thread will transition
 955                          * to TD_INITIALIZED when it's done setting up.
 956                          */
 957                         td_set_runstate(td, TD_CREATED);
 958                         map[this_jobs++] = td;
 959                         fio_sem_init(&startup_sem, 1);
 960                         nr_started++;
 961
 962                         if (td->use_thread) {
 963                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 964                                         perror("thread_create");
 965                                         nr_started--;
 966                                 }
 967                         } else {
 968                                 if (fork())
 969                                         fio_sem_down(&startup_sem);
 970                                 else {
 971                                         int ret = fork_main(shm_id, i);
 972
 973                                         exit(ret);
 974                                 }
 975                         }
 976                 }
 977
 978                 /*
 979                  * Wait for the started threads to transition to
 980                  * TD_INITIALIZED.
 981                  */
 982                 fio_gettime(&this_start, NULL);
 983                 left = this_jobs;
 984                 while (left && !fio_abort) {
 985                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 986                                 break;
 987
 988                         usleep(100000);
 989
 990                         for (i = 0; i < this_jobs; i++) {
 991                                 td = map[i];
 992                                 if (!td)
 993                                         continue;
 994                                 if (td->runstate == TD_INITIALIZED) {
 995                                         map[i] = NULL;
 996                                         left--;
 997                                 } else if (td->runstate >= TD_EXITED) {
 998                                         map[i] = NULL;
 999                                         left--;
1000                                         todo--;
1001                                         nr_running++; /* work-around... */
1002                                 }
1003                         }
1004                 }
1005
1006                 if (left) {
1007                         log_err("fio: %d jobs failed to start\n", left);
1008                         for (i = 0; i < this_jobs; i++) {
1009                                 td = map[i];
1010                                 if (!td)
1011                                         continue;
1012                                 kill(td->pid, SIGTERM);
1013                         }
1014                         break;
1015                 }
1016
1017                 /*
1018                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1019                  */
1020                 for_each_td(td, i) {
1021                         if (td->runstate != TD_INITIALIZED)
1022                                 continue;
1023
1024                         td_set_runstate(td, TD_RUNNING);
1025                         nr_running++;
1026                         nr_started--;
1027                         m_rate += td->ratemin;
1028                         t_rate += td->rate;
1029                         todo--;
1030                         fio_sem_up(&td->mutex);
1031                 }
1032
1033                 reap_threads(&nr_running, &t_rate, &m_rate);
1034
1035                 if (todo)
1036                         usleep(100000);
1037         }
1038
1039         while (nr_running) {
1040                 reap_threads(&nr_running, &t_rate, &m_rate);
1041                 usleep(10000);
1042         }
1043
1044         update_io_ticks();
1045         fio_unpin_memory();
1046 }
1047
1048 int main(int argc, char *argv[])
1049 {
1050         long ps;
1051
1052         /*
1053          * We need locale for number printing, if it isn't set then just
1054          * go with the US format.
1055          */
1056         if (!getenv("LC_NUMERIC"))
1057                 setlocale(LC_NUMERIC, "en_US");
1058
1059         if (parse_options(argc, argv))
1060                 return 1;
1061
1062         if (!thread_number) {
1063                 log_err("Nothing to do\n");
1064                 return 1;
1065         }
1066
1067         ps = sysconf(_SC_PAGESIZE);
1068         if (ps < 0) {
1069                 log_err("Failed to get page size\n");
1070                 return 1;
1071         }
1072
1073         page_mask = ps - 1;
1074
1075         if (write_bw_log) {
1076                 setup_log(&agg_io_log[DDIR_READ]);
1077                 setup_log(&agg_io_log[DDIR_WRITE]);
1078         }
1079
1080         disk_util_timer_arm();
1081
1082         run_threads();
1083
1084         if (!fio_abort) {
1085                 show_run_stats();
1086                 if (write_bw_log) {
1087                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1088                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1089                 }
1090         }
1091
1092         return exit_value;
1093 }