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 nr_process = 0;
  47 int nr_thread = 0;
  48 int shm_id = 0;
  49 int temp_stall_ts;
  50
  51 static struct fio_sem *startup_sem;
  52 static volatile int fio_abort;
  53 static int exit_value;
  54
  55 struct io_log *agg_io_log[2];
  56
  57 #define TERMINATE_ALL           (-1)
  58 #define JOB_START_TIMEOUT       (5 * 1000)
  59
  60 static inline void td_set_runstate(struct thread_data *td, int runstate)
  61 {
  62         td->runstate = runstate;
  63 }
  64
  65 static void terminate_threads(int group_id)
  66 {
  67         struct thread_data *td;
  68         int i;
  69
  70         for_each_td(td, i) {
  71                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  72                         /*
  73                          * if the thread is running, just let it exit
  74                          */
  75                         if (td->runstate < TD_RUNNING)
  76                                 kill(td->pid, SIGQUIT);
  77                         td->terminate = 1;
  78                         td->start_delay = 0;
  79                 }
  80         }
  81 }
  82
  83 static void sig_handler(int sig)
  84 {
  85         switch (sig) {
  86                 case SIGALRM:
  87                         update_io_ticks();
  88                         disk_util_timer_arm();
  89                         print_thread_status();
  90                         break;
  91                 default:
  92                         printf("\nfio: terminating on signal %d\n", sig);
  93                         fflush(stdout);
  94                         terminate_threads(TERMINATE_ALL);
  95                         break;
  96         }
  97 }
  98
  99 /*
 100  * Check if we are above the minimum rate given.
 101  */
 102 static int check_min_rate(struct thread_data *td, struct timeval *now)
 103 {
 104         unsigned long long bytes = 0;
 105         unsigned long spent;
 106         unsigned long rate;
 107
 108         /*
 109          * No minimum rate set, always ok
 110          */
 111         if (!td->ratemin)
 112                 return 0;
 113
 114         /*
 115          * allow a 2 second settle period in the beginning
 116          */
 117         if (mtime_since(&td->start, now) < 2000)
 118                 return 0;
 119
 120         if (td_read(td))
 121                 bytes += td->this_io_bytes[DDIR_READ];
 122         if (td_write(td))
 123                 bytes += td->this_io_bytes[DDIR_WRITE];
 124
 125         /*
 126          * if rate blocks is set, sample is running
 127          */
 128         if (td->rate_bytes) {
 129                 spent = mtime_since(&td->lastrate, now);
 130                 if (spent < td->ratecycle)
 131                         return 0;
 132
 133                 if (bytes < td->rate_bytes) {
 134                         fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
 135                         return 1;
 136                 } else {
 137                         rate = (bytes - td->rate_bytes) / spent;
 138                         if (rate < td->ratemin || bytes < td->rate_bytes) {
 139                                 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
 140                                 return 1;
 141                         }
 142                 }
 143         }
 144
 145         td->rate_bytes = bytes;
 146         memcpy(&td->lastrate, now, sizeof(*now));
 147         return 0;
 148 }
 149
 150 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 151 {
 152         if (!td->timeout)
 153                 return 0;
 154         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 155                 return 1;
 156
 157         return 0;
 158 }
 159
 160 /*
 161  * When job exits, we can cancel the in-flight IO if we are using async
 162  * io. Attempt to do so.
 163  */
 164 static void cleanup_pending_aio(struct thread_data *td)
 165 {
 166         struct list_head *entry, *n;
 167         struct io_u *io_u;
 168         int r;
 169
 170         /*
 171          * get immediately available events, if any
 172          */
 173         r = io_u_queued_complete(td, 0);
 174         if (r < 0)
 175                 return;
 176
 177         /*
 178          * now cancel remaining active events
 179          */
 180         if (td->io_ops->cancel) {
 181                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 182                         io_u = list_entry(entry, struct io_u, list);
 183
 184                         /*
 185                          * if the io_u isn't in flight, then that generally
 186                          * means someone leaked an io_u. complain but fix
 187                          * it up, so we don't stall here.
 188                          */
 189                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 190                                 log_err("fio: non-busy IO on busy list\n");
 191                                 put_io_u(td, io_u);
 192                         } else {
 193                                 r = td->io_ops->cancel(td, io_u);
 194                                 if (!r)
 195                                         put_io_u(td, io_u);
 196                         }
 197                 }
 198         }
 199
 200         if (td->cur_depth)
 201                 r = io_u_queued_complete(td, td->cur_depth);
 202 }
 203
 204 /*
 205  * Helper to handle the final sync of a file. Works just like the normal
 206  * io path, just does everything sync.
 207  */
 208 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 209 {
 210         struct io_u *io_u = __get_io_u(td);
 211         int ret;
 212
 213         if (!io_u)
 214                 return 1;
 215
 216         io_u->ddir = DDIR_SYNC;
 217         io_u->file = f;
 218
 219         if (td_io_prep(td, io_u)) {
 220                 put_io_u(td, io_u);
 221                 return 1;
 222         }
 223
 224 requeue:
 225         ret = td_io_queue(td, io_u);
 226         if (ret < 0) {
 227                 td_verror(td, io_u->error, "td_io_queue");
 228                 put_io_u(td, io_u);
 229                 return 1;
 230         } else if (ret == FIO_Q_QUEUED) {
 231                 if (io_u_queued_complete(td, 1) < 0)
 232                         return 1;
 233         } else if (ret == FIO_Q_COMPLETED) {
 234                 if (io_u->error) {
 235                         td_verror(td, io_u->error, "td_io_queue");
 236                         return 1;
 237                 }
 238
 239                 if (io_u_sync_complete(td, io_u) < 0)
 240                         return 1;
 241         } else if (ret == FIO_Q_BUSY) {
 242                 if (td_io_commit(td))
 243                         return 1;
 244                 goto requeue;
 245         }
 246
 247         return 0;
 248 }
 249
 250 /*
 251  * The main verify engine. Runs over the writes we previously submitted,
 252  * reads the blocks back in, and checks the crc/md5 of the data.
 253  */
 254 static void do_verify(struct thread_data *td)
 255 {
 256         struct fio_file *f;
 257         struct io_u *io_u;
 258         int ret, i, min_events;
 259
 260         /*
 261          * sync io first and invalidate cache, to make sure we really
 262          * read from disk.
 263          */
 264         for_each_file(td, f, i) {
 265                 if (fio_io_sync(td, f))
 266                         break;
 267                 if (file_invalidate_cache(td, f))
 268                         break;
 269         }
 270
 271         if (td->error)
 272                 return;
 273
 274         td_set_runstate(td, TD_VERIFYING);
 275
 276         io_u = NULL;
 277         while (!td->terminate) {
 278                 int ret2;
 279
 280                 io_u = __get_io_u(td);
 281                 if (!io_u)
 282                         break;
 283
 284                 if (runtime_exceeded(td, &io_u->start_time)) {
 285                         put_io_u(td, io_u);
 286                         break;
 287                 }
 288
 289                 if (get_next_verify(td, io_u)) {
 290                         put_io_u(td, io_u);
 291                         break;
 292                 }
 293
 294                 if (td_io_prep(td, io_u)) {
 295                         put_io_u(td, io_u);
 296                         break;
 297                 }
 298
 299                 io_u->end_io = verify_io_u;
 300
 301                 ret = td_io_queue(td, io_u);
 302                 switch (ret) {
 303                 case FIO_Q_COMPLETED:
 304                         if (io_u->error)
 305                                 ret = -io_u->error;
 306                         else if (io_u->resid) {
 307                                 int bytes = io_u->xfer_buflen - io_u->resid;
 308
 309                                 /*
 310                                  * zero read, fail
 311                                  */
 312                                 if (!bytes) {
 313                                         td_verror(td, ENODATA, "full resid");
 314                                         put_io_u(td, io_u);
 315                                         break;
 316                                 }
 317                                 io_u->xfer_buflen = io_u->resid;
 318                                 io_u->xfer_buf += bytes;
 319                                 requeue_io_u(td, &io_u);
 320                         } else {
 321                                 ret = io_u_sync_complete(td, io_u);
 322                                 if (ret < 0)
 323                                         break;
 324                         }
 325                         continue;
 326                 case FIO_Q_QUEUED:
 327                         break;
 328                 case FIO_Q_BUSY:
 329                         requeue_io_u(td, &io_u);
 330                         ret2 = td_io_commit(td);
 331                         if (ret2 < 0)
 332                                 ret = ret2;
 333                         break;
 334                 default:
 335                         assert(ret < 0);
 336                         td_verror(td, -ret, "td_io_queue");
 337                         break;
 338                 }
 339
 340                 if (ret < 0 || td->error)
 341                         break;
 342
 343                 /*
 344                  * if we can queue more, do so. but check if there are
 345                  * completed io_u's first.
 346                  */
 347                 min_events = 0;
 348                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 349                         min_events = 1;
 350
 351                         if (td->cur_depth > td->iodepth_low)
 352                                 min_events = td->cur_depth - td->iodepth_low;
 353                 }
 354
 355                 /*
 356                  * Reap required number of io units, if any, and do the
 357                  * verification on them through the callback handler
 358                  */
 359                 if (io_u_queued_complete(td, min_events) < 0)
 360                         break;
 361         }
 362
 363         if (!td->error) {
 364                 min_events = td->cur_depth;
 365
 366                 if (min_events)
 367                         ret = io_u_queued_complete(td, min_events);
 368         } else
 369                 cleanup_pending_aio(td);
 370
 371         td_set_runstate(td, TD_RUNNING);
 372 }
 373
 374 /*
 375  * Main IO worker function. It retrieves io_u's to process and queues
 376  * and reaps them, checking for rate and errors along the way.
 377  */
 378 static void do_io(struct thread_data *td)
 379 {
 380         struct timeval s;
 381         unsigned long usec;
 382         int i, ret = 0;
 383
 384         td_set_runstate(td, TD_RUNNING);
 385
 386         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 387                 struct timeval comp_time;
 388                 long bytes_done = 0;
 389                 int min_evts = 0;
 390                 struct io_u *io_u;
 391                 int ret2;
 392
 393                 if (td->terminate)
 394                         break;
 395
 396                 io_u = get_io_u(td);
 397                 if (!io_u)
 398                         break;
 399
 400                 memcpy(&s, &io_u->start_time, sizeof(s));
 401
 402                 if (runtime_exceeded(td, &s)) {
 403                         put_io_u(td, io_u);
 404                         break;
 405                 }
 406
 407                 ret = td_io_queue(td, io_u);
 408                 switch (ret) {
 409                 case FIO_Q_COMPLETED:
 410                         if (io_u->error)
 411                                 ret = -io_u->error;
 412                         else if (io_u->resid) {
 413                                 int bytes = io_u->xfer_buflen - io_u->resid;
 414
 415                                 /*
 416                                  * zero read, fail
 417                                  */
 418                                 if (!bytes) {
 419                                         td_verror(td, ENODATA, "full resid");
 420                                         put_io_u(td, io_u);
 421                                         break;
 422                                 }
 423
 424                                 io_u->xfer_buflen = io_u->resid;
 425                                 io_u->xfer_buf += bytes;
 426                                 requeue_io_u(td, &io_u);
 427                         } else {
 428                                 fio_gettime(&comp_time, NULL);
 429                                 bytes_done = io_u_sync_complete(td, io_u);
 430                                 if (bytes_done < 0)
 431                                         ret = bytes_done;
 432                         }
 433                         break;
 434                 case FIO_Q_QUEUED:
 435                         /*
 436                          * if the engine doesn't have a commit hook,
 437                          * the io_u is really queued. if it does have such
 438                          * a hook, it has to call io_u_queued() itself.
 439                          */
 440                         if (td->io_ops->commit == NULL)
 441                                 io_u_queued(td, io_u);
 442                         break;
 443                 case FIO_Q_BUSY:
 444                         requeue_io_u(td, &io_u);
 445                         ret2 = td_io_commit(td);
 446                         if (ret2 < 0)
 447                                 ret = ret2;
 448                         break;
 449                 default:
 450                         assert(ret < 0);
 451                         put_io_u(td, io_u);
 452                         break;
 453                 }
 454
 455                 if (ret < 0 || td->error)
 456                         break;
 457
 458                 /*
 459                  * See if we need to complete some commands
 460                  */
 461                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 462                         min_evts = 0;
 463                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 464                                 min_evts = 1;
 465
 466                                 if (td->cur_depth > td->iodepth_low)
 467                                         min_evts = td->cur_depth - td->iodepth_low;
 468                         }
 469
 470                         fio_gettime(&comp_time, NULL);
 471                         bytes_done = io_u_queued_complete(td, min_evts);
 472                         if (bytes_done < 0)
 473                                 break;
 474                 }
 475
 476                 if (!bytes_done)
 477                         continue;
 478
 479                 /*
 480                  * the rate is batched for now, it should work for batches
 481                  * of completions except the very first one which may look
 482                  * a little bursty
 483                  */
 484                 usec = utime_since(&s, &comp_time);
 485
 486                 rate_throttle(td, usec, bytes_done);
 487
 488                 if (check_min_rate(td, &comp_time)) {
 489                         if (exitall_on_terminate)
 490                                 terminate_threads(td->groupid);
 491                         td_verror(td, ENODATA, "check_min_rate");
 492                         break;
 493                 }
 494
 495                 if (td->thinktime) {
 496                         unsigned long long b;
 497
 498                         b = td->io_blocks[0] + td->io_blocks[1];
 499                         if (!(b % td->thinktime_blocks)) {
 500                                 int left;
 501
 502                                 if (td->thinktime_spin)
 503                                         __usec_sleep(td->thinktime_spin);
 504
 505                                 left = td->thinktime - td->thinktime_spin;
 506                                 if (left)
 507                                         usec_sleep(td, left);
 508                         }
 509                 }
 510         }
 511
 512         if (!td->error) {
 513                 struct fio_file *f;
 514
 515                 i = td->cur_depth;
 516                 if (i)
 517                         ret = io_u_queued_complete(td, i);
 518
 519                 if (should_fsync(td) && td->end_fsync) {
 520                         td_set_runstate(td, TD_FSYNCING);
 521                         for_each_file(td, f, i)
 522                                 fio_io_sync(td, f);
 523                 }
 524         } else
 525                 cleanup_pending_aio(td);
 526 }
 527
 528 static void cleanup_io_u(struct thread_data *td)
 529 {
 530         struct list_head *entry, *n;
 531         struct io_u *io_u;
 532
 533         list_for_each_safe(entry, n, &td->io_u_freelist) {
 534                 io_u = list_entry(entry, struct io_u, list);
 535
 536                 list_del(&io_u->list);
 537                 free(io_u);
 538         }
 539
 540         free_io_mem(td);
 541 }
 542
 543 /*
 544  * "randomly" fill the buffer contents
 545  */
 546 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 547 {
 548         int *ptr = io_u->buf;
 549
 550         while ((void *) ptr - io_u->buf < max_bs) {
 551                 *ptr = rand() * 0x9e370001;
 552                 ptr++;
 553         }
 554 }
 555
 556 static int init_io_u(struct thread_data *td)
 557 {
 558         struct io_u *io_u;
 559         unsigned int max_bs;
 560         int i, max_units;
 561         char *p;
 562
 563         if (td->io_ops->flags & FIO_SYNCIO)
 564                 max_units = 1;
 565         else
 566                 max_units = td->iodepth;
 567
 568         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 569         td->orig_buffer_size = max_bs * max_units;
 570
 571         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 572                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 573         else
 574                 td->orig_buffer_size += page_mask;
 575
 576         if (allocate_io_mem(td))
 577                 return 1;
 578
 579         p = ALIGN(td->orig_buffer);
 580         for (i = 0; i < max_units; i++) {
 581                 io_u = malloc(sizeof(*io_u));
 582                 memset(io_u, 0, sizeof(*io_u));
 583                 INIT_LIST_HEAD(&io_u->list);
 584
 585                 io_u->buf = p + max_bs * i;
 586                 if (td_write(td) || td_rw(td))
 587                         fill_rand_buf(io_u, max_bs);
 588
 589                 io_u->index = i;
 590                 io_u->flags = IO_U_F_FREE;
 591                 list_add(&io_u->list, &td->io_u_freelist);
 592         }
 593
 594         io_u_init_timeout();
 595
 596         return 0;
 597 }
 598
 599 static int switch_ioscheduler(struct thread_data *td)
 600 {
 601         char tmp[256], tmp2[128];
 602         FILE *f;
 603         int ret;
 604
 605         if (td->io_ops->flags & FIO_DISKLESSIO)
 606                 return 0;
 607
 608         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 609
 610         f = fopen(tmp, "r+");
 611         if (!f) {
 612                 td_verror(td, errno, "fopen");
 613                 return 1;
 614         }
 615
 616         /*
 617          * Set io scheduler.
 618          */
 619         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 620         if (ferror(f) || ret != 1) {
 621                 td_verror(td, errno, "fwrite");
 622                 fclose(f);
 623                 return 1;
 624         }
 625
 626         rewind(f);
 627
 628         /*
 629          * Read back and check that the selected scheduler is now the default.
 630          */
 631         ret = fread(tmp, 1, sizeof(tmp), f);
 632         if (ferror(f) || ret < 0) {
 633                 td_verror(td, errno, "fread");
 634                 fclose(f);
 635                 return 1;
 636         }
 637
 638         sprintf(tmp2, "[%s]", td->ioscheduler);
 639         if (!strstr(tmp, tmp2)) {
 640                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 641                 td_verror(td, EINVAL, "iosched_switch");
 642                 fclose(f);
 643                 return 1;
 644         }
 645
 646         fclose(f);
 647         return 0;
 648 }
 649
 650 static int clear_io_state(struct thread_data *td)
 651 {
 652         struct fio_file *f;
 653         int i, ret;
 654
 655         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 656         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 657         td->zone_bytes = 0;
 658         td->rate_bytes = 0;
 659
 660         td->last_was_sync = 0;
 661
 662         for_each_file(td, f, i)
 663                 td_io_close_file(td, f);
 664
 665         ret = 0;
 666         for_each_file(td, f, i) {
 667                 ret = td_io_open_file(td, f);
 668                 if (ret)
 669                         break;
 670         }
 671
 672         return ret;
 673 }
 674
 675 /*
 676  * Entry point for the thread based jobs. The process based jobs end up
 677  * here as well, after a little setup.
 678  */
 679 static void *thread_main(void *data)
 680 {
 681         unsigned long long runtime[2];
 682         struct thread_data *td = data;
 683         int clear_state;
 684
 685         if (!td->use_thread)
 686                 setsid();
 687
 688         td->pid = getpid();
 689
 690         INIT_LIST_HEAD(&td->io_u_freelist);
 691         INIT_LIST_HEAD(&td->io_u_busylist);
 692         INIT_LIST_HEAD(&td->io_u_requeues);
 693         INIT_LIST_HEAD(&td->io_hist_list);
 694         INIT_LIST_HEAD(&td->io_log_list);
 695
 696         if (init_io_u(td))
 697                 goto err_sem;
 698
 699         if (fio_setaffinity(td) == -1) {
 700                 td_verror(td, errno, "cpu_set_affinity");
 701                 goto err_sem;
 702         }
 703
 704         if (init_iolog(td))
 705                 goto err_sem;
 706
 707         if (td->ioprio) {
 708                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 709                         td_verror(td, errno, "ioprio_set");
 710                         goto err_sem;
 711                 }
 712         }
 713
 714         if (nice(td->nice) == -1) {
 715                 td_verror(td, errno, "nice");
 716                 goto err_sem;
 717         }
 718
 719         if (init_random_state(td))
 720                 goto err_sem;
 721
 722         if (td->ioscheduler && switch_ioscheduler(td))
 723                 goto err_sem;
 724
 725         td_set_runstate(td, TD_INITIALIZED);
 726         fio_sem_up(startup_sem);
 727         fio_sem_down(td->mutex);
 728
 729         /*
 730          * the ->mutex semaphore is now no longer used, close it to avoid
 731          * eating a file descriptor
 732          */
 733         fio_sem_remove(td->mutex);
 734
 735         if (!td->create_serialize && setup_files(td))
 736                 goto err;
 737
 738         if (td_io_init(td))
 739                 goto err;
 740
 741         if (open_files(td))
 742                 goto err;
 743
 744         if (td->exec_prerun) {
 745                 if (system(td->exec_prerun) < 0)
 746                         goto err;
 747         }
 748
 749         fio_gettime(&td->epoch, NULL);
 750         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 751         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 752
 753         runtime[0] = runtime[1] = 0;
 754         clear_state = 0;
 755         while (td->loops--) {
 756                 fio_gettime(&td->start, NULL);
 757                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 758
 759                 if (td->ratemin)
 760                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 761
 762                 if (clear_state && clear_io_state(td))
 763                         break;
 764
 765                 prune_io_piece_log(td);
 766
 767                 do_io(td);
 768
 769                 clear_state = 1;
 770
 771                 if (td_read(td) && td->io_bytes[DDIR_READ])
 772                         runtime[DDIR_READ] += utime_since_now(&td->start);
 773                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
 774                         runtime[DDIR_WRITE] += utime_since_now(&td->start);
 775
 776                 if (td->error || td->terminate)
 777                         break;
 778
 779                 if (td->verify == VERIFY_NONE)
 780                         continue;
 781
 782                 if (clear_io_state(td))
 783                         break;
 784
 785                 fio_gettime(&td->start, NULL);
 786
 787                 do_verify(td);
 788
 789                 runtime[DDIR_READ] += utime_since_now(&td->start);
 790
 791                 if (td->error || td->terminate)
 792                         break;
 793         }
 794
 795         update_rusage_stat(td);
 796         td->ts.runtime[0] = runtime[0] / 1000;
 797         td->ts.runtime[1] = runtime[1] / 1000;
 798         td->ts.total_run_time = mtime_since_now(&td->epoch);
 799         td->ts.io_bytes[0] = td->io_bytes[0];
 800         td->ts.io_bytes[1] = td->io_bytes[1];
 801
 802         if (td->ts.bw_log)
 803                 finish_log(td, td->ts.bw_log, "bw");
 804         if (td->ts.slat_log)
 805                 finish_log(td, td->ts.slat_log, "slat");
 806         if (td->ts.clat_log)
 807                 finish_log(td, td->ts.clat_log, "clat");
 808         if (td->write_iolog_file)
 809                 write_iolog_close(td);
 810         if (td->exec_postrun) {
 811                 if (system(td->exec_postrun) < 0)
 812                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 813         }
 814
 815         if (exitall_on_terminate)
 816                 terminate_threads(td->groupid);
 817
 818 err:
 819         if (td->error)
 820                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 821         close_files(td);
 822         close_ioengine(td);
 823         cleanup_io_u(td);
 824         td_set_runstate(td, TD_EXITED);
 825         return (void *) (unsigned long) td->error;
 826 err_sem:
 827         fio_sem_up(startup_sem);
 828         goto err;
 829 }
 830
 831 /*
 832  * We cannot pass the td data into a forked process, so attach the td and
 833  * pass it to the thread worker.
 834  */
 835 static int fork_main(int shmid, int offset)
 836 {
 837         struct thread_data *td;
 838         void *data, *ret;
 839
 840         data = shmat(shmid, NULL, 0);
 841         if (data == (void *) -1) {
 842                 int __err = errno;
 843
 844                 perror("shmat");
 845                 return __err;
 846         }
 847
 848         td = data + offset * sizeof(struct thread_data);
 849         ret = thread_main(td);
 850         shmdt(data);
 851         return (int) (unsigned long) ret;
 852 }
 853
 854 /*
 855  * Run over the job map and reap the threads that have exited, if any.
 856  */
 857 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 858 {
 859         struct thread_data *td;
 860         int i, cputhreads, pending, status, ret;
 861
 862         /*
 863          * reap exited threads (TD_EXITED -> TD_REAPED)
 864          */
 865         pending = cputhreads = 0;
 866         for_each_td(td, i) {
 867                 int flags = 0;
 868
 869                 /*
 870                  * ->io_ops is NULL for a thread that has closed its
 871                  * io engine
 872                  */
 873                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 874                         cputhreads++;
 875
 876                 if (!td->pid || td->runstate == TD_REAPED)
 877                         continue;
 878                 if (td->use_thread) {
 879                         if (td->runstate == TD_EXITED) {
 880                                 td_set_runstate(td, TD_REAPED);
 881                                 goto reaped;
 882                         }
 883                         continue;
 884                 }
 885
 886                 flags = WNOHANG;
 887                 if (td->runstate == TD_EXITED)
 888                         flags = 0;
 889
 890                 /*
 891                  * check if someone quit or got killed in an unusual way
 892                  */
 893                 ret = waitpid(td->pid, &status, flags);
 894                 if (ret < 0) {
 895                         if (errno == ECHILD) {
 896                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 897                                 td_set_runstate(td, TD_REAPED);
 898                                 goto reaped;
 899                         }
 900                         perror("waitpid");
 901                 } else if (ret == td->pid) {
 902                         if (WIFSIGNALED(status)) {
 903                                 int sig = WTERMSIG(status);
 904
 905                                 if (sig != SIGQUIT)
 906                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 907                                 td_set_runstate(td, TD_REAPED);
 908                                 goto reaped;
 909                         }
 910                         if (WIFEXITED(status)) {
 911                                 if (WEXITSTATUS(status) && !td->error)
 912                                         td->error = WEXITSTATUS(status);
 913
 914                                 td_set_runstate(td, TD_REAPED);
 915                                 goto reaped;
 916                         }
 917                 }
 918
 919                 /*
 920                  * thread is not dead, continue
 921                  */
 922                 continue;
 923 reaped:
 924                 if (td->use_thread) {
 925                         long ret;
 926
 927                         if (pthread_join(td->thread, (void *) &ret))
 928                                 perror("pthread_join");
 929                 }
 930
 931                 (*nr_running)--;
 932                 (*m_rate) -= td->ratemin;
 933                 (*t_rate) -= td->rate;
 934
 935                 if (td->error)
 936                         exit_value++;
 937         }
 938
 939         if (*nr_running == cputhreads && !pending)
 940                 terminate_threads(TERMINATE_ALL);
 941 }
 942
 943 /*
 944  * Main function for kicking off and reaping jobs, as needed.
 945  */
 946 static void run_threads(void)
 947 {
 948         struct thread_data *td;
 949         unsigned long spent;
 950         int i, todo, nr_running, m_rate, t_rate, nr_started;
 951
 952         if (fio_pin_memory())
 953                 return;
 954
 955         if (!terse_output) {
 956                 printf("Starting ");
 957                 if (nr_thread)
 958                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
 959                 if (nr_process) {
 960                         if (nr_thread)
 961                                 printf(" and ");
 962                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
 963                 }
 964                 printf("\n");
 965                 fflush(stdout);
 966         }
 967
 968         signal(SIGINT, sig_handler);
 969         signal(SIGALRM, sig_handler);
 970
 971         todo = thread_number;
 972         nr_running = 0;
 973         nr_started = 0;
 974         m_rate = t_rate = 0;
 975
 976         for_each_td(td, i) {
 977                 print_status_init(td->thread_number - 1);
 978
 979                 if (!td->create_serialize) {
 980                         init_disk_util(td);
 981                         continue;
 982                 }
 983
 984                 /*
 985                  * do file setup here so it happens sequentially,
 986                  * we don't want X number of threads getting their
 987                  * client data interspersed on disk
 988                  */
 989                 if (setup_files(td)) {
 990                         exit_value++;
 991                         if (td->error)
 992                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 993                         td_set_runstate(td, TD_REAPED);
 994                         todo--;
 995                 }
 996
 997                 init_disk_util(td);
 998         }
 999
1000         set_genesis_time();
1001
1002         while (todo) {
1003                 struct thread_data *map[MAX_JOBS];
1004                 struct timeval this_start;
1005                 int this_jobs = 0, left;
1006
1007                 /*
1008                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1009                  */
1010                 for_each_td(td, i) {
1011                         if (td->runstate != TD_NOT_CREATED)
1012                                 continue;
1013
1014                         /*
1015                          * never got a chance to start, killed by other
1016                          * thread for some reason
1017                          */
1018                         if (td->terminate) {
1019                                 todo--;
1020                                 continue;
1021                         }
1022
1023                         if (td->start_delay) {
1024                                 spent = mtime_since_genesis();
1025
1026                                 if (td->start_delay * 1000 > spent)
1027                                         continue;
1028                         }
1029
1030                         if (td->stonewall && (nr_started || nr_running))
1031                                 break;
1032
1033                         /*
1034                          * Set state to created. Thread will transition
1035                          * to TD_INITIALIZED when it's done setting up.
1036                          */
1037                         td_set_runstate(td, TD_CREATED);
1038                         map[this_jobs++] = td;
1039                         nr_started++;
1040
1041                         if (td->use_thread) {
1042                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1043                                         perror("thread_create");
1044                                         nr_started--;
1045                                         break;
1046                                 }
1047                         } else {
1048                                 if (!fork()) {
1049                                         int ret = fork_main(shm_id, i);
1050
1051                                         exit(ret);
1052                                 }
1053                         }
1054                         fio_sem_down(startup_sem);
1055                 }
1056
1057                 /*
1058                  * Wait for the started threads to transition to
1059                  * TD_INITIALIZED.
1060                  */
1061                 fio_gettime(&this_start, NULL);
1062                 left = this_jobs;
1063                 while (left && !fio_abort) {
1064                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1065                                 break;
1066
1067                         usleep(100000);
1068
1069                         for (i = 0; i < this_jobs; i++) {
1070                                 td = map[i];
1071                                 if (!td)
1072                                         continue;
1073                                 if (td->runstate == TD_INITIALIZED) {
1074                                         map[i] = NULL;
1075                                         left--;
1076                                 } else if (td->runstate >= TD_EXITED) {
1077                                         map[i] = NULL;
1078                                         left--;
1079                                         todo--;
1080                                         nr_running++; /* work-around... */
1081                                 }
1082                         }
1083                 }
1084
1085                 if (left) {
1086                         log_err("fio: %d jobs failed to start\n", left);
1087                         for (i = 0; i < this_jobs; i++) {
1088                                 td = map[i];
1089                                 if (!td)
1090                                         continue;
1091                                 kill(td->pid, SIGTERM);
1092                         }
1093                         break;
1094                 }
1095
1096                 /*
1097                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1098                  */
1099                 for_each_td(td, i) {
1100                         if (td->runstate != TD_INITIALIZED)
1101                                 continue;
1102
1103                         td_set_runstate(td, TD_RUNNING);
1104                         nr_running++;
1105                         nr_started--;
1106                         m_rate += td->ratemin;
1107                         t_rate += td->rate;
1108                         todo--;
1109                         fio_sem_up(td->mutex);
1110                 }
1111
1112                 reap_threads(&nr_running, &t_rate, &m_rate);
1113
1114                 if (todo)
1115                         usleep(100000);
1116         }
1117
1118         while (nr_running) {
1119                 reap_threads(&nr_running, &t_rate, &m_rate);
1120                 usleep(10000);
1121         }
1122
1123         update_io_ticks();
1124         fio_unpin_memory();
1125 }
1126
1127 int main(int argc, char *argv[])
1128 {
1129         long ps;
1130
1131         /*
1132          * We need locale for number printing, if it isn't set then just
1133          * go with the US format.
1134          */
1135         if (!getenv("LC_NUMERIC"))
1136                 setlocale(LC_NUMERIC, "en_US");
1137
1138         if (parse_options(argc, argv))
1139                 return 1;
1140
1141         if (!thread_number) {
1142                 log_err("Nothing to do\n");
1143                 return 1;
1144         }
1145
1146         ps = sysconf(_SC_PAGESIZE);
1147         if (ps < 0) {
1148                 log_err("Failed to get page size\n");
1149                 return 1;
1150         }
1151
1152         page_mask = ps - 1;
1153
1154         if (write_bw_log) {
1155                 setup_log(&agg_io_log[DDIR_READ]);
1156                 setup_log(&agg_io_log[DDIR_WRITE]);
1157         }
1158
1159         startup_sem = fio_sem_init(0);
1160
1161         set_genesis_time();
1162
1163         disk_util_timer_arm();
1164
1165         run_threads();
1166
1167         if (!fio_abort) {
1168                 show_run_stats();
1169                 if (write_bw_log) {
1170                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1171                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1172                 }
1173         }
1174
1175         fio_sem_remove(startup_sem);
1176         return exit_value;
1177 }