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                         put_io_u(td, io_u);
 253                         break;
 254                 }
 255
 256                 if (get_next_verify(td, io_u)) {
 257                         put_io_u(td, io_u);
 258                         break;
 259                 }
 260
 261                 if (td_io_prep(td, io_u)) {
 262                         put_io_u(td, io_u);
 263                         break;
 264                 }
 265 requeue:
 266                 ret = td_io_queue(td, io_u);
 267
 268                 switch (ret) {
 269                 case FIO_Q_COMPLETED:
 270                         if (io_u->error)
 271                                 ret = -io_u->error;
 272                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 273                                 int bytes = io_u->xfer_buflen - io_u->resid;
 274
 275                                 io_u->xfer_buflen = io_u->resid;
 276                                 io_u->xfer_buf += bytes;
 277                                 goto requeue;
 278                         }
 279                         ret = io_u_sync_complete(td, io_u, verify_io_u);
 280                         if (ret < 0)
 281                                 break;
 282                         continue;
 283                 case FIO_Q_QUEUED:
 284                         break;
 285                 case FIO_Q_BUSY:
 286                         requeue_io_u(td, &io_u);
 287                         ret = td_io_commit(td);
 288                         break;
 289                 default:
 290                         assert(ret < 0);
 291                         td_verror(td, -ret);
 292                         break;
 293                 }
 294
 295                 if (ret < 0 || td->error)
 296                         break;
 297
 298                 /*
 299                  * if we can queue more, do so. but check if there are
 300                  * completed io_u's first.
 301                  */
 302                 min_events = 0;
 303                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 304                         min_events = 1;
 305
 306                         if (td->cur_depth > td->iodepth_low)
 307                                 min_events = td->cur_depth - td->iodepth_low;
 308                 }
 309
 310                 /*
 311                  * Reap required number of io units, if any, and do the
 312                  * verification on them through the callback handler
 313                  */
 314                 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
 315                         break;
 316         }
 317
 318         if (td->cur_depth)
 319                 cleanup_pending_aio(td);
 320
 321         td_set_runstate(td, TD_RUNNING);
 322 }
 323
 324 /*
 325  * Not really an io thread, all it does is burn CPU cycles in the specified
 326  * manner.
 327  */
 328 static void do_cpuio(struct thread_data *td)
 329 {
 330         struct timeval e;
 331         int split = 100 / td->cpuload;
 332         int i = 0;
 333
 334         while (!td->terminate) {
 335                 fio_gettime(&e, NULL);
 336
 337                 if (runtime_exceeded(td, &e))
 338                         break;
 339
 340                 if (!(i % split))
 341                         __usec_sleep(10000);
 342                 else
 343                         usec_sleep(td, 10000);
 344
 345                 i++;
 346         }
 347 }
 348
 349 /*
 350  * Main IO worker function. It retrieves io_u's to process and queues
 351  * and reaps them, checking for rate and errors along the way.
 352  */
 353 static void do_io(struct thread_data *td)
 354 {
 355         struct timeval s;
 356         unsigned long usec;
 357         int i, ret = 0;
 358
 359         td_set_runstate(td, TD_RUNNING);
 360
 361         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 362                 struct timeval comp_time;
 363                 long bytes_done = 0;
 364                 int min_evts = 0;
 365                 struct io_u *io_u;
 366
 367                 if (td->terminate)
 368                         break;
 369
 370                 io_u = get_io_u(td);
 371                 if (!io_u)
 372                         break;
 373
 374                 memcpy(&s, &io_u->start_time, sizeof(s));
 375
 376                 if (runtime_exceeded(td, &s)) {
 377                         put_io_u(td, io_u);
 378                         break;
 379                 }
 380 requeue:
 381                 ret = td_io_queue(td, io_u);
 382
 383                 switch (ret) {
 384                 case FIO_Q_COMPLETED:
 385                         if (io_u->error) {
 386                                 ret = io_u->error;
 387                                 break;
 388                         }
 389                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 390                                 int bytes = io_u->xfer_buflen - io_u->resid;
 391
 392                                 io_u->xfer_buflen = io_u->resid;
 393                                 io_u->xfer_buf += bytes;
 394                                 goto requeue;
 395                         }
 396                         fio_gettime(&comp_time, NULL);
 397                         bytes_done = io_u_sync_complete(td, io_u, NULL);
 398                         if (bytes_done < 0)
 399                                 ret = bytes_done;
 400                         break;
 401                 case FIO_Q_QUEUED:
 402                         /*
 403                          * if the engine doesn't have a commit hook,
 404                          * the io_u is really queued. if it does have such
 405                          * a hook, it has to call io_u_queued() itself.
 406                          */
 407                         if (td->io_ops->commit == NULL)
 408                                 io_u_queued(td, io_u);
 409                         break;
 410                 case FIO_Q_BUSY:
 411                         requeue_io_u(td, &io_u);
 412                         ret = td_io_commit(td);
 413                         break;
 414                 default:
 415                         assert(ret < 0);
 416                         put_io_u(td, io_u);
 417                         break;
 418                 }
 419
 420                 if (ret < 0 || td->error)
 421                         break;
 422
 423                 /*
 424                  * See if we need to complete some commands
 425                  */
 426                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 427                         min_evts = 0;
 428                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 429                                 min_evts = 1;
 430
 431                                 if (td->cur_depth > td->iodepth_low)
 432                                         min_evts = td->cur_depth - td->iodepth_low;
 433                         }
 434
 435                         fio_gettime(&comp_time, NULL);
 436                         bytes_done = io_u_queued_complete(td, min_evts, NULL);
 437                         if (bytes_done < 0)
 438                                 break;
 439                 }
 440
 441                 if (!bytes_done)
 442                         continue;
 443
 444                 /*
 445                  * the rate is batched for now, it should work for batches
 446                  * of completions except the very first one which may look
 447                  * a little bursty
 448                  */
 449                 usec = utime_since(&s, &comp_time);
 450
 451                 rate_throttle(td, usec, bytes_done, td->ddir);
 452
 453                 if (check_min_rate(td, &comp_time)) {
 454                         if (exitall_on_terminate)
 455                                 terminate_threads(td->groupid, 0);
 456                         td_verror(td, ENODATA);
 457                         break;
 458                 }
 459
 460                 if (td->thinktime) {
 461                         unsigned long long b;
 462
 463                         b = td->io_blocks[0] + td->io_blocks[1];
 464                         if (!(b % td->thinktime_blocks)) {
 465                                 int left;
 466
 467                                 if (td->thinktime_spin)
 468                                         __usec_sleep(td->thinktime_spin);
 469
 470                                 left = td->thinktime - td->thinktime_spin;
 471                                 if (left)
 472                                         usec_sleep(td, left);
 473                         }
 474                 }
 475         }
 476
 477         if (!td->error) {
 478                 struct fio_file *f;
 479
 480                 if (td->cur_depth)
 481                         cleanup_pending_aio(td);
 482
 483                 if (should_fsync(td) && td->end_fsync) {
 484                         td_set_runstate(td, TD_FSYNCING);
 485                         for_each_file(td, f, i)
 486                                 fio_io_sync(td, f);
 487                 }
 488         }
 489 }
 490
 491 static void cleanup_io_u(struct thread_data *td)
 492 {
 493         struct list_head *entry, *n;
 494         struct io_u *io_u;
 495
 496         list_for_each_safe(entry, n, &td->io_u_freelist) {
 497                 io_u = list_entry(entry, struct io_u, list);
 498
 499                 list_del(&io_u->list);
 500                 free(io_u);
 501         }
 502
 503         free_io_mem(td);
 504 }
 505
 506 /*
 507  * "randomly" fill the buffer contents
 508  */
 509 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 510 {
 511         int *ptr = io_u->buf;
 512
 513         while ((void *) ptr - io_u->buf < max_bs) {
 514                 *ptr = rand() * 0x9e370001;
 515                 ptr++;
 516         }
 517 }
 518
 519 static int init_io_u(struct thread_data *td)
 520 {
 521         struct io_u *io_u;
 522         unsigned int max_bs;
 523         int i, max_units;
 524         char *p;
 525
 526         if (td->io_ops->flags & FIO_CPUIO)
 527                 return 0;
 528
 529         if (td->io_ops->flags & FIO_SYNCIO)
 530                 max_units = 1;
 531         else
 532                 max_units = td->iodepth;
 533
 534         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 535         td->orig_buffer_size = max_bs * max_units;
 536
 537         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 538                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 539         else
 540                 td->orig_buffer_size += page_mask;
 541
 542         if (allocate_io_mem(td))
 543                 return 1;
 544
 545         p = ALIGN(td->orig_buffer);
 546         for (i = 0; i < max_units; i++) {
 547                 io_u = malloc(sizeof(*io_u));
 548                 memset(io_u, 0, sizeof(*io_u));
 549                 INIT_LIST_HEAD(&io_u->list);
 550
 551                 io_u->buf = p + max_bs * i;
 552                 if (td_write(td) || td_rw(td))
 553                         fill_rand_buf(io_u, max_bs);
 554
 555                 io_u->index = i;
 556                 list_add(&io_u->list, &td->io_u_freelist);
 557         }
 558
 559         return 0;
 560 }
 561
 562 static int switch_ioscheduler(struct thread_data *td)
 563 {
 564         char tmp[256], tmp2[128];
 565         FILE *f;
 566         int ret;
 567
 568         if (td->io_ops->flags & FIO_CPUIO)
 569                 return 0;
 570
 571         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 572
 573         f = fopen(tmp, "r+");
 574         if (!f) {
 575                 td_verror(td, errno);
 576                 return 1;
 577         }
 578
 579         /*
 580          * Set io scheduler.
 581          */
 582         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 583         if (ferror(f) || ret != 1) {
 584                 td_verror(td, errno);
 585                 fclose(f);
 586                 return 1;
 587         }
 588
 589         rewind(f);
 590
 591         /*
 592          * Read back and check that the selected scheduler is now the default.
 593          */
 594         ret = fread(tmp, 1, sizeof(tmp), f);
 595         if (ferror(f) || ret < 0) {
 596                 td_verror(td, errno);
 597                 fclose(f);
 598                 return 1;
 599         }
 600
 601         sprintf(tmp2, "[%s]", td->ioscheduler);
 602         if (!strstr(tmp, tmp2)) {
 603                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 604                 td_verror(td, EINVAL);
 605                 fclose(f);
 606                 return 1;
 607         }
 608
 609         fclose(f);
 610         return 0;
 611 }
 612
 613 static void clear_io_state(struct thread_data *td)
 614 {
 615         struct fio_file *f;
 616         int i;
 617
 618         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 619         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 620         td->zone_bytes = 0;
 621
 622         td->last_was_sync = 0;
 623
 624         for_each_file(td, f, i) {
 625                 f->last_completed_pos = 0;
 626
 627                 f->last_pos = 0;
 628                 if (td->io_ops->flags & FIO_SYNCIO)
 629                         lseek(f->fd, SEEK_SET, 0);
 630
 631                 if (f->file_map)
 632                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 633         }
 634 }
 635
 636 /*
 637  * Entry point for the thread based jobs. The process based jobs end up
 638  * here as well, after a little setup.
 639  */
 640 static void *thread_main(void *data)
 641 {
 642         unsigned long long runtime[2];
 643         struct thread_data *td = data;
 644
 645         if (!td->use_thread)
 646                 setsid();
 647
 648         td->pid = getpid();
 649
 650         INIT_LIST_HEAD(&td->io_u_freelist);
 651         INIT_LIST_HEAD(&td->io_u_busylist);
 652         INIT_LIST_HEAD(&td->io_u_requeues);
 653         INIT_LIST_HEAD(&td->io_hist_list);
 654         INIT_LIST_HEAD(&td->io_log_list);
 655
 656         if (init_io_u(td))
 657                 goto err;
 658
 659         if (fio_setaffinity(td) == -1) {
 660                 td_verror(td, errno);
 661                 goto err;
 662         }
 663
 664         if (init_iolog(td))
 665                 goto err;
 666
 667         if (td->ioprio) {
 668                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 669                         td_verror(td, errno);
 670                         goto err;
 671                 }
 672         }
 673
 674         if (nice(td->nice) == -1) {
 675                 td_verror(td, errno);
 676                 goto err;
 677         }
 678
 679         if (init_random_state(td))
 680                 goto err;
 681
 682         if (td->ioscheduler && switch_ioscheduler(td))
 683                 goto err;
 684
 685         td_set_runstate(td, TD_INITIALIZED);
 686         fio_sem_up(&startup_sem);
 687         fio_sem_down(&td->mutex);
 688
 689         if (!td->create_serialize && setup_files(td))
 690                 goto err;
 691         if (open_files(td))
 692                 goto err;
 693
 694         /*
 695          * Do this late, as some IO engines would like to have the
 696          * files setup prior to initializing structures.
 697          */
 698         if (td_io_init(td))
 699                 goto err;
 700
 701         if (td->exec_prerun) {
 702                 if (system(td->exec_prerun) < 0)
 703                         goto err;
 704         }
 705
 706         fio_gettime(&td->epoch, NULL);
 707         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 708
 709         runtime[0] = runtime[1] = 0;
 710         while (td->loops--) {
 711                 fio_gettime(&td->start, NULL);
 712                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 713
 714                 if (td->ratemin)
 715                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 716
 717                 clear_io_state(td);
 718                 prune_io_piece_log(td);
 719
 720                 if (td->io_ops->flags & FIO_CPUIO)
 721                         do_cpuio(td);
 722                 else
 723                         do_io(td);
 724
 725                 runtime[td->ddir] += utime_since_now(&td->start);
 726                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 727                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 728
 729                 if (td->error || td->terminate)
 730                         break;
 731
 732                 if (td->verify == VERIFY_NONE)
 733                         continue;
 734
 735                 clear_io_state(td);
 736                 fio_gettime(&td->start, NULL);
 737
 738                 do_verify(td);
 739
 740                 runtime[DDIR_READ] += utime_since_now(&td->start);
 741
 742                 if (td->error || td->terminate)
 743                         break;
 744         }
 745
 746         update_rusage_stat(td);
 747         fio_gettime(&td->end_time, NULL);
 748         td->runtime[0] = runtime[0] / 1000;
 749         td->runtime[1] = runtime[1] / 1000;
 750
 751         if (td->ts.bw_log)
 752                 finish_log(td, td->ts.bw_log, "bw");
 753         if (td->ts.slat_log)
 754                 finish_log(td, td->ts.slat_log, "slat");
 755         if (td->ts.clat_log)
 756                 finish_log(td, td->ts.clat_log, "clat");
 757         if (td->write_iolog_file)
 758                 write_iolog_close(td);
 759         if (td->exec_postrun) {
 760                 if (system(td->exec_postrun) < 0)
 761                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 762         }
 763
 764         if (exitall_on_terminate)
 765                 terminate_threads(td->groupid, 0);
 766
 767 err:
 768         if (td->error)
 769                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 770         close_files(td);
 771         close_ioengine(td);
 772         cleanup_io_u(td);
 773         td_set_runstate(td, TD_EXITED);
 774         return (void *) (unsigned long) td->error;
 775 }
 776
 777 /*
 778  * We cannot pass the td data into a forked process, so attach the td and
 779  * pass it to the thread worker.
 780  */
 781 static int fork_main(int shmid, int offset)
 782 {
 783         struct thread_data *td;
 784         void *data, *ret;
 785
 786         data = shmat(shmid, NULL, 0);
 787         if (data == (void *) -1) {
 788                 int __err = errno;
 789
 790                 perror("shmat");
 791                 return __err;
 792         }
 793
 794         td = data + offset * sizeof(struct thread_data);
 795         ret = thread_main(td);
 796         shmdt(data);
 797         return (int) (unsigned long) ret;
 798 }
 799
 800 /*
 801  * Run over the job map and reap the threads that have exited, if any.
 802  */
 803 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 804 {
 805         struct thread_data *td;
 806         int i, cputhreads, pending, status, ret;
 807
 808         /*
 809          * reap exited threads (TD_EXITED -> TD_REAPED)
 810          */
 811         pending = cputhreads = 0;
 812         for_each_td(td, i) {
 813                 /*
 814                  * ->io_ops is NULL for a thread that has closed its
 815                  * io engine
 816                  */
 817                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 818                         cputhreads++;
 819
 820                 if (td->runstate < TD_EXITED) {
 821                         /*
 822                          * check if someone quit or got killed in an unusual way
 823                          */
 824                         ret = waitpid(td->pid, &status, WNOHANG);
 825                         if (ret < 0)
 826                                 perror("waitpid");
 827                         else if ((ret == td->pid) && WIFSIGNALED(status)) {
 828                                 int sig = WTERMSIG(status);
 829
 830                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 831                                 td_set_runstate(td, TD_REAPED);
 832                                 goto reaped;
 833                         }
 834                 }
 835
 836                 if (td->runstate != TD_EXITED) {
 837                         if (td->runstate < TD_RUNNING)
 838                                 pending++;
 839
 840                         continue;
 841                 }
 842
 843                 if (td->error)
 844                         exit_value++;
 845
 846                 td_set_runstate(td, TD_REAPED);
 847
 848                 if (td->use_thread) {
 849                         long ret;
 850
 851                         if (pthread_join(td->thread, (void *) &ret))
 852                                 perror("thread_join");
 853                 } else {
 854                         int status;
 855
 856                         ret = waitpid(td->pid, &status, 0);
 857                         if (ret < 0)
 858                                 perror("waitpid");
 859                         else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 860                                 if (!exit_value)
 861                                         exit_value++;
 862                         }
 863                 }
 864
 865 reaped:
 866                 (*nr_running)--;
 867                 (*m_rate) -= td->ratemin;
 868                 (*t_rate) -= td->rate;
 869         }
 870
 871         if (*nr_running == cputhreads && !pending)
 872                 terminate_threads(TERMINATE_ALL, 0);
 873 }
 874
 875 /*
 876  * Main function for kicking off and reaping jobs, as needed.
 877  */
 878 static void run_threads(void)
 879 {
 880         struct thread_data *td;
 881         unsigned long spent;
 882         int i, todo, nr_running, m_rate, t_rate, nr_started;
 883
 884         if (fio_pin_memory())
 885                 return;
 886
 887         if (!terse_output) {
 888                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 889                 fflush(stdout);
 890         }
 891
 892         signal(SIGINT, sig_handler);
 893         signal(SIGALRM, sig_handler);
 894
 895         todo = thread_number;
 896         nr_running = 0;
 897         nr_started = 0;
 898         m_rate = t_rate = 0;
 899
 900         for_each_td(td, i) {
 901                 print_status_init(td->thread_number - 1);
 902
 903                 if (!td->create_serialize) {
 904                         init_disk_util(td);
 905                         continue;
 906                 }
 907
 908                 /*
 909                  * do file setup here so it happens sequentially,
 910                  * we don't want X number of threads getting their
 911                  * client data interspersed on disk
 912                  */
 913                 if (setup_files(td)) {
 914                         exit_value++;
 915                         if (td->error)
 916                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 917                         td_set_runstate(td, TD_REAPED);
 918                         todo--;
 919                 }
 920
 921                 init_disk_util(td);
 922         }
 923
 924         while (todo) {
 925                 struct thread_data *map[MAX_JOBS];
 926                 struct timeval this_start;
 927                 int this_jobs = 0, left;
 928
 929                 /*
 930                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 931                  */
 932                 for_each_td(td, i) {
 933                         if (td->runstate != TD_NOT_CREATED)
 934                                 continue;
 935
 936                         /*
 937                          * never got a chance to start, killed by other
 938                          * thread for some reason
 939                          */
 940                         if (td->terminate) {
 941                                 todo--;
 942                                 continue;
 943                         }
 944
 945                         if (td->start_delay) {
 946                                 spent = mtime_since_genesis();
 947
 948                                 if (td->start_delay * 1000 > spent)
 949                                         continue;
 950                         }
 951
 952                         if (td->stonewall && (nr_started || nr_running))
 953                                 break;
 954
 955                         /*
 956                          * Set state to created. Thread will transition
 957                          * to TD_INITIALIZED when it's done setting up.
 958                          */
 959                         td_set_runstate(td, TD_CREATED);
 960                         map[this_jobs++] = td;
 961                         fio_sem_init(&startup_sem, 1);
 962                         nr_started++;
 963
 964                         if (td->use_thread) {
 965                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 966                                         perror("thread_create");
 967                                         nr_started--;
 968                                 }
 969                         } else {
 970                                 if (fork())
 971                                         fio_sem_down(&startup_sem);
 972                                 else {
 973                                         int ret = fork_main(shm_id, i);
 974
 975                                         exit(ret);
 976                                 }
 977                         }
 978                 }
 979
 980                 /*
 981                  * Wait for the started threads to transition to
 982                  * TD_INITIALIZED.
 983                  */
 984                 fio_gettime(&this_start, NULL);
 985                 left = this_jobs;
 986                 while (left && !fio_abort) {
 987                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 988                                 break;
 989
 990                         usleep(100000);
 991
 992                         for (i = 0; i < this_jobs; i++) {
 993                                 td = map[i];
 994                                 if (!td)
 995                                         continue;
 996                                 if (td->runstate == TD_INITIALIZED) {
 997                                         map[i] = NULL;
 998                                         left--;
 999                                 } else if (td->runstate >= TD_EXITED) {
1000                                         map[i] = NULL;
1001                                         left--;
1002                                         todo--;
1003                                         nr_running++; /* work-around... */
1004                                 }
1005                         }
1006                 }
1007
1008                 if (left) {
1009                         log_err("fio: %d jobs failed to start\n", left);
1010                         for (i = 0; i < this_jobs; i++) {
1011                                 td = map[i];
1012                                 if (!td)
1013                                         continue;
1014                                 kill(td->pid, SIGTERM);
1015                         }
1016                         break;
1017                 }
1018
1019                 /*
1020                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1021                  */
1022                 for_each_td(td, i) {
1023                         if (td->runstate != TD_INITIALIZED)
1024                                 continue;
1025
1026                         td_set_runstate(td, TD_RUNNING);
1027                         nr_running++;
1028                         nr_started--;
1029                         m_rate += td->ratemin;
1030                         t_rate += td->rate;
1031                         todo--;
1032                         fio_sem_up(&td->mutex);
1033                 }
1034
1035                 reap_threads(&nr_running, &t_rate, &m_rate);
1036
1037                 if (todo)
1038                         usleep(100000);
1039         }
1040
1041         while (nr_running) {
1042                 reap_threads(&nr_running, &t_rate, &m_rate);
1043                 usleep(10000);
1044         }
1045
1046         update_io_ticks();
1047         fio_unpin_memory();
1048 }
1049
1050 int main(int argc, char *argv[])
1051 {
1052         long ps;
1053
1054         /*
1055          * We need locale for number printing, if it isn't set then just
1056          * go with the US format.
1057          */
1058         if (!getenv("LC_NUMERIC"))
1059                 setlocale(LC_NUMERIC, "en_US");
1060
1061         if (parse_options(argc, argv))
1062                 return 1;
1063
1064         if (!thread_number) {
1065                 log_err("Nothing to do\n");
1066                 return 1;
1067         }
1068
1069         ps = sysconf(_SC_PAGESIZE);
1070         if (ps < 0) {
1071                 log_err("Failed to get page size\n");
1072                 return 1;
1073         }
1074
1075         page_mask = ps - 1;
1076
1077         if (write_bw_log) {
1078                 setup_log(&agg_io_log[DDIR_READ]);
1079                 setup_log(&agg_io_log[DDIR_WRITE]);
1080         }
1081
1082         disk_util_timer_arm();
1083
1084         run_threads();
1085
1086         if (!fio_abort) {
1087                 show_run_stats();
1088                 if (write_bw_log) {
1089                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1090                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1091                 }
1092         }
1093
1094         return exit_value;
1095 }