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, min_events;
 259         unsigned int i;
 260
 261         /*
 262          * sync io first and invalidate cache, to make sure we really
 263          * read from disk.
 264          */
 265         for_each_file(td, f, i) {
 266                 if (fio_io_sync(td, f))
 267                         break;
 268                 if (file_invalidate_cache(td, f))
 269                         break;
 270         }
 271
 272         if (td->error)
 273                 return;
 274
 275         td_set_runstate(td, TD_VERIFYING);
 276
 277         io_u = NULL;
 278         while (!td->terminate) {
 279                 int ret2;
 280
 281                 io_u = __get_io_u(td);
 282                 if (!io_u)
 283                         break;
 284
 285                 if (runtime_exceeded(td, &io_u->start_time)) {
 286                         put_io_u(td, io_u);
 287                         break;
 288                 }
 289
 290                 if (get_next_verify(td, io_u)) {
 291                         put_io_u(td, io_u);
 292                         break;
 293                 }
 294
 295                 if (td_io_prep(td, io_u)) {
 296                         put_io_u(td, io_u);
 297                         break;
 298                 }
 299
 300                 io_u->end_io = verify_io_u;
 301
 302                 ret = td_io_queue(td, io_u);
 303                 switch (ret) {
 304                 case FIO_Q_COMPLETED:
 305                         if (io_u->error)
 306                                 ret = -io_u->error;
 307                         else if (io_u->resid) {
 308                                 int bytes = io_u->xfer_buflen - io_u->resid;
 309
 310                                 /*
 311                                  * zero read, fail
 312                                  */
 313                                 if (!bytes) {
 314                                         td_verror(td, ENODATA, "full resid");
 315                                         put_io_u(td, io_u);
 316                                         break;
 317                                 }
 318                                 io_u->xfer_buflen = io_u->resid;
 319                                 io_u->xfer_buf += bytes;
 320                                 requeue_io_u(td, &io_u);
 321                         } else {
 322                                 ret = io_u_sync_complete(td, io_u);
 323                                 if (ret < 0)
 324                                         break;
 325                         }
 326                         continue;
 327                 case FIO_Q_QUEUED:
 328                         break;
 329                 case FIO_Q_BUSY:
 330                         requeue_io_u(td, &io_u);
 331                         ret2 = td_io_commit(td);
 332                         if (ret2 < 0)
 333                                 ret = ret2;
 334                         break;
 335                 default:
 336                         assert(ret < 0);
 337                         td_verror(td, -ret, "td_io_queue");
 338                         break;
 339                 }
 340
 341                 if (ret < 0 || td->error)
 342                         break;
 343
 344                 /*
 345                  * if we can queue more, do so. but check if there are
 346                  * completed io_u's first.
 347                  */
 348                 min_events = 0;
 349                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 350                         min_events = 1;
 351
 352                         if (td->cur_depth > td->iodepth_low)
 353                                 min_events = td->cur_depth - td->iodepth_low;
 354                 }
 355
 356                 /*
 357                  * Reap required number of io units, if any, and do the
 358                  * verification on them through the callback handler
 359                  */
 360                 if (io_u_queued_complete(td, min_events) < 0)
 361                         break;
 362         }
 363
 364         if (!td->error) {
 365                 min_events = td->cur_depth;
 366
 367                 if (min_events)
 368                         ret = io_u_queued_complete(td, min_events);
 369         } else
 370                 cleanup_pending_aio(td);
 371
 372         td_set_runstate(td, TD_RUNNING);
 373 }
 374
 375 /*
 376  * Main IO worker function. It retrieves io_u's to process and queues
 377  * and reaps them, checking for rate and errors along the way.
 378  */
 379 static void do_io(struct thread_data *td)
 380 {
 381         struct timeval s;
 382         unsigned long usec;
 383         unsigned int i;
 384         int ret = 0;
 385
 386         td_set_runstate(td, TD_RUNNING);
 387
 388         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 389                 struct timeval comp_time;
 390                 long bytes_done = 0;
 391                 int min_evts = 0;
 392                 struct io_u *io_u;
 393                 int ret2;
 394
 395                 if (td->terminate)
 396                         break;
 397
 398                 io_u = get_io_u(td);
 399                 if (!io_u)
 400                         break;
 401
 402                 memcpy(&s, &io_u->start_time, sizeof(s));
 403
 404                 if (runtime_exceeded(td, &s)) {
 405                         put_io_u(td, io_u);
 406                         break;
 407                 }
 408
 409                 ret = td_io_queue(td, io_u);
 410                 switch (ret) {
 411                 case FIO_Q_COMPLETED:
 412                         if (io_u->error)
 413                                 ret = -io_u->error;
 414                         else if (io_u->resid) {
 415                                 int bytes = io_u->xfer_buflen - io_u->resid;
 416
 417                                 /*
 418                                  * zero read, fail
 419                                  */
 420                                 if (!bytes) {
 421                                         td_verror(td, ENODATA, "full resid");
 422                                         put_io_u(td, io_u);
 423                                         break;
 424                                 }
 425
 426                                 io_u->xfer_buflen = io_u->resid;
 427                                 io_u->xfer_buf += bytes;
 428                                 requeue_io_u(td, &io_u);
 429                         } else {
 430                                 fio_gettime(&comp_time, NULL);
 431                                 bytes_done = io_u_sync_complete(td, io_u);
 432                                 if (bytes_done < 0)
 433                                         ret = bytes_done;
 434                         }
 435                         break;
 436                 case FIO_Q_QUEUED:
 437                         /*
 438                          * if the engine doesn't have a commit hook,
 439                          * the io_u is really queued. if it does have such
 440                          * a hook, it has to call io_u_queued() itself.
 441                          */
 442                         if (td->io_ops->commit == NULL)
 443                                 io_u_queued(td, io_u);
 444                         break;
 445                 case FIO_Q_BUSY:
 446                         requeue_io_u(td, &io_u);
 447                         ret2 = td_io_commit(td);
 448                         if (ret2 < 0)
 449                                 ret = ret2;
 450                         break;
 451                 default:
 452                         assert(ret < 0);
 453                         put_io_u(td, io_u);
 454                         break;
 455                 }
 456
 457                 if (ret < 0 || td->error)
 458                         break;
 459
 460                 /*
 461                  * See if we need to complete some commands
 462                  */
 463                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 464                         min_evts = 0;
 465                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 466                                 min_evts = 1;
 467
 468                                 if (td->cur_depth > td->iodepth_low)
 469                                         min_evts = td->cur_depth - td->iodepth_low;
 470                         }
 471
 472                         fio_gettime(&comp_time, NULL);
 473                         bytes_done = io_u_queued_complete(td, min_evts);
 474                         if (bytes_done < 0)
 475                                 break;
 476                 }
 477
 478                 if (!bytes_done)
 479                         continue;
 480
 481                 /*
 482                  * the rate is batched for now, it should work for batches
 483                  * of completions except the very first one which may look
 484                  * a little bursty
 485                  */
 486                 usec = utime_since(&s, &comp_time);
 487
 488                 rate_throttle(td, usec, bytes_done);
 489
 490                 if (check_min_rate(td, &comp_time)) {
 491                         if (exitall_on_terminate)
 492                                 terminate_threads(td->groupid);
 493                         td_verror(td, ENODATA, "check_min_rate");
 494                         break;
 495                 }
 496
 497                 if (td->thinktime) {
 498                         unsigned long long b;
 499
 500                         b = td->io_blocks[0] + td->io_blocks[1];
 501                         if (!(b % td->thinktime_blocks)) {
 502                                 int left;
 503
 504                                 if (td->thinktime_spin)
 505                                         __usec_sleep(td->thinktime_spin);
 506
 507                                 left = td->thinktime - td->thinktime_spin;
 508                                 if (left)
 509                                         usec_sleep(td, left);
 510                         }
 511                 }
 512         }
 513
 514         if (!td->error) {
 515                 struct fio_file *f;
 516
 517                 i = td->cur_depth;
 518                 if (i)
 519                         ret = io_u_queued_complete(td, i);
 520
 521                 if (should_fsync(td) && td->end_fsync) {
 522                         td_set_runstate(td, TD_FSYNCING);
 523                         for_each_file(td, f, i)
 524                                 fio_io_sync(td, f);
 525                 }
 526         } else
 527                 cleanup_pending_aio(td);
 528 }
 529
 530 static void cleanup_io_u(struct thread_data *td)
 531 {
 532         struct list_head *entry, *n;
 533         struct io_u *io_u;
 534
 535         list_for_each_safe(entry, n, &td->io_u_freelist) {
 536                 io_u = list_entry(entry, struct io_u, list);
 537
 538                 list_del(&io_u->list);
 539                 free(io_u);
 540         }
 541
 542         free_io_mem(td);
 543 }
 544
 545 /*
 546  * "randomly" fill the buffer contents
 547  */
 548 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 549 {
 550         int *ptr = io_u->buf;
 551
 552         while ((void *) ptr - io_u->buf < max_bs) {
 553                 *ptr = rand() * 0x9e370001;
 554                 ptr++;
 555         }
 556 }
 557
 558 static int init_io_u(struct thread_data *td)
 559 {
 560         struct io_u *io_u;
 561         unsigned int max_bs;
 562         int i, max_units;
 563         char *p;
 564
 565         if (td->io_ops->flags & FIO_SYNCIO)
 566                 max_units = 1;
 567         else
 568                 max_units = td->iodepth;
 569
 570         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 571         td->orig_buffer_size = max_bs * max_units;
 572
 573         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 574                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 575         else
 576                 td->orig_buffer_size += page_mask;
 577
 578         if (allocate_io_mem(td))
 579                 return 1;
 580
 581         p = ALIGN(td->orig_buffer);
 582         for (i = 0; i < max_units; i++) {
 583                 io_u = malloc(sizeof(*io_u));
 584                 memset(io_u, 0, sizeof(*io_u));
 585                 INIT_LIST_HEAD(&io_u->list);
 586
 587                 io_u->buf = p + max_bs * i;
 588                 if (td_write(td) || td_rw(td))
 589                         fill_rand_buf(io_u, max_bs);
 590
 591                 io_u->index = i;
 592                 io_u->flags = IO_U_F_FREE;
 593                 list_add(&io_u->list, &td->io_u_freelist);
 594         }
 595
 596         io_u_init_timeout();
 597
 598         return 0;
 599 }
 600
 601 static int switch_ioscheduler(struct thread_data *td)
 602 {
 603         char tmp[256], tmp2[128];
 604         FILE *f;
 605         int ret;
 606
 607         if (td->io_ops->flags & FIO_DISKLESSIO)
 608                 return 0;
 609
 610         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 611
 612         f = fopen(tmp, "r+");
 613         if (!f) {
 614                 td_verror(td, errno, "fopen");
 615                 return 1;
 616         }
 617
 618         /*
 619          * Set io scheduler.
 620          */
 621         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 622         if (ferror(f) || ret != 1) {
 623                 td_verror(td, errno, "fwrite");
 624                 fclose(f);
 625                 return 1;
 626         }
 627
 628         rewind(f);
 629
 630         /*
 631          * Read back and check that the selected scheduler is now the default.
 632          */
 633         ret = fread(tmp, 1, sizeof(tmp), f);
 634         if (ferror(f) || ret < 0) {
 635                 td_verror(td, errno, "fread");
 636                 fclose(f);
 637                 return 1;
 638         }
 639
 640         sprintf(tmp2, "[%s]", td->ioscheduler);
 641         if (!strstr(tmp, tmp2)) {
 642                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 643                 td_verror(td, EINVAL, "iosched_switch");
 644                 fclose(f);
 645                 return 1;
 646         }
 647
 648         fclose(f);
 649         return 0;
 650 }
 651
 652 static int clear_io_state(struct thread_data *td)
 653 {
 654         struct fio_file *f;
 655         unsigned int i;
 656         int ret;
 657
 658         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 659         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 660         td->zone_bytes = 0;
 661         td->rate_bytes = 0;
 662
 663         td->last_was_sync = 0;
 664
 665         for_each_file(td, f, i)
 666                 td_io_close_file(td, f);
 667
 668         ret = 0;
 669         for_each_file(td, f, i) {
 670                 ret = td_io_open_file(td, f);
 671                 if (ret)
 672                         break;
 673         }
 674
 675         return ret;
 676 }
 677
 678 /*
 679  * Entry point for the thread based jobs. The process based jobs end up
 680  * here as well, after a little setup.
 681  */
 682 static void *thread_main(void *data)
 683 {
 684         unsigned long long runtime[2];
 685         struct thread_data *td = data;
 686         int clear_state;
 687
 688         if (!td->use_thread)
 689                 setsid();
 690
 691         td->pid = getpid();
 692
 693         INIT_LIST_HEAD(&td->io_u_freelist);
 694         INIT_LIST_HEAD(&td->io_u_busylist);
 695         INIT_LIST_HEAD(&td->io_u_requeues);
 696         INIT_LIST_HEAD(&td->io_hist_list);
 697         INIT_LIST_HEAD(&td->io_log_list);
 698
 699         if (init_io_u(td))
 700                 goto err_sem;
 701
 702         if (fio_setaffinity(td) == -1) {
 703                 td_verror(td, errno, "cpu_set_affinity");
 704                 goto err_sem;
 705         }
 706
 707         if (init_iolog(td))
 708                 goto err_sem;
 709
 710         if (td->ioprio) {
 711                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 712                         td_verror(td, errno, "ioprio_set");
 713                         goto err_sem;
 714                 }
 715         }
 716
 717         if (nice(td->nice) == -1) {
 718                 td_verror(td, errno, "nice");
 719                 goto err_sem;
 720         }
 721
 722         if (init_random_state(td))
 723                 goto err_sem;
 724
 725         if (td->ioscheduler && switch_ioscheduler(td))
 726                 goto err_sem;
 727
 728         td_set_runstate(td, TD_INITIALIZED);
 729         fio_sem_up(startup_sem);
 730         fio_sem_down(td->mutex);
 731
 732         /*
 733          * the ->mutex semaphore is now no longer used, close it to avoid
 734          * eating a file descriptor
 735          */
 736         fio_sem_remove(td->mutex);
 737
 738         if (!td->create_serialize && setup_files(td))
 739                 goto err;
 740
 741         if (td_io_init(td))
 742                 goto err;
 743
 744         if (open_files(td))
 745                 goto err;
 746
 747         if (td->exec_prerun) {
 748                 if (system(td->exec_prerun) < 0)
 749                         goto err;
 750         }
 751
 752         fio_gettime(&td->epoch, NULL);
 753         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 754         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 755
 756         runtime[0] = runtime[1] = 0;
 757         clear_state = 0;
 758         while (td->loops--) {
 759                 fio_gettime(&td->start, NULL);
 760                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 761
 762                 if (td->ratemin)
 763                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 764
 765                 if (clear_state && clear_io_state(td))
 766                         break;
 767
 768                 prune_io_piece_log(td);
 769
 770                 do_io(td);
 771
 772                 clear_state = 1;
 773
 774                 if (td_read(td) && td->io_bytes[DDIR_READ])
 775                         runtime[DDIR_READ] += utime_since_now(&td->start);
 776                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
 777                         runtime[DDIR_WRITE] += utime_since_now(&td->start);
 778
 779                 if (td->error || td->terminate)
 780                         break;
 781
 782                 if (td->verify == VERIFY_NONE)
 783                         continue;
 784
 785                 if (clear_io_state(td))
 786                         break;
 787
 788                 fio_gettime(&td->start, NULL);
 789
 790                 do_verify(td);
 791
 792                 runtime[DDIR_READ] += utime_since_now(&td->start);
 793
 794                 if (td->error || td->terminate)
 795                         break;
 796         }
 797
 798         update_rusage_stat(td);
 799         td->ts.runtime[0] = runtime[0] / 1000;
 800         td->ts.runtime[1] = runtime[1] / 1000;
 801         td->ts.total_run_time = mtime_since_now(&td->epoch);
 802         td->ts.io_bytes[0] = td->io_bytes[0];
 803         td->ts.io_bytes[1] = td->io_bytes[1];
 804
 805         if (td->ts.bw_log)
 806                 finish_log(td, td->ts.bw_log, "bw");
 807         if (td->ts.slat_log)
 808                 finish_log(td, td->ts.slat_log, "slat");
 809         if (td->ts.clat_log)
 810                 finish_log(td, td->ts.clat_log, "clat");
 811         if (td->write_iolog_file)
 812                 write_iolog_close(td);
 813         if (td->exec_postrun) {
 814                 if (system(td->exec_postrun) < 0)
 815                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 816         }
 817
 818         if (exitall_on_terminate)
 819                 terminate_threads(td->groupid);
 820
 821 err:
 822         if (td->error)
 823                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 824         close_files(td);
 825         close_ioengine(td);
 826         cleanup_io_u(td);
 827         td_set_runstate(td, TD_EXITED);
 828         return (void *) (unsigned long) td->error;
 829 err_sem:
 830         fio_sem_up(startup_sem);
 831         goto err;
 832 }
 833
 834 /*
 835  * We cannot pass the td data into a forked process, so attach the td and
 836  * pass it to the thread worker.
 837  */
 838 static int fork_main(int shmid, int offset)
 839 {
 840         struct thread_data *td;
 841         void *data, *ret;
 842
 843         data = shmat(shmid, NULL, 0);
 844         if (data == (void *) -1) {
 845                 int __err = errno;
 846
 847                 perror("shmat");
 848                 return __err;
 849         }
 850
 851         td = data + offset * sizeof(struct thread_data);
 852         ret = thread_main(td);
 853         shmdt(data);
 854         return (int) (unsigned long) ret;
 855 }
 856
 857 /*
 858  * Run over the job map and reap the threads that have exited, if any.
 859  */
 860 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 861 {
 862         struct thread_data *td;
 863         int i, cputhreads, pending, status, ret;
 864
 865         /*
 866          * reap exited threads (TD_EXITED -> TD_REAPED)
 867          */
 868         pending = cputhreads = 0;
 869         for_each_td(td, i) {
 870                 int flags = 0;
 871
 872                 /*
 873                  * ->io_ops is NULL for a thread that has closed its
 874                  * io engine
 875                  */
 876                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 877                         cputhreads++;
 878
 879                 if (!td->pid || td->runstate == TD_REAPED)
 880                         continue;
 881                 if (td->use_thread) {
 882                         if (td->runstate == TD_EXITED) {
 883                                 td_set_runstate(td, TD_REAPED);
 884                                 goto reaped;
 885                         }
 886                         continue;
 887                 }
 888
 889                 flags = WNOHANG;
 890                 if (td->runstate == TD_EXITED)
 891                         flags = 0;
 892
 893                 /*
 894                  * check if someone quit or got killed in an unusual way
 895                  */
 896                 ret = waitpid(td->pid, &status, flags);
 897                 if (ret < 0) {
 898                         if (errno == ECHILD) {
 899                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 900                                 td_set_runstate(td, TD_REAPED);
 901                                 goto reaped;
 902                         }
 903                         perror("waitpid");
 904                 } else if (ret == td->pid) {
 905                         if (WIFSIGNALED(status)) {
 906                                 int sig = WTERMSIG(status);
 907
 908                                 if (sig != SIGQUIT)
 909                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 910                                 td_set_runstate(td, TD_REAPED);
 911                                 goto reaped;
 912                         }
 913                         if (WIFEXITED(status)) {
 914                                 if (WEXITSTATUS(status) && !td->error)
 915                                         td->error = WEXITSTATUS(status);
 916
 917                                 td_set_runstate(td, TD_REAPED);
 918                                 goto reaped;
 919                         }
 920                 }
 921
 922                 /*
 923                  * thread is not dead, continue
 924                  */
 925                 continue;
 926 reaped:
 927                 if (td->use_thread) {
 928                         long ret;
 929
 930                         if (pthread_join(td->thread, (void *) &ret))
 931                                 perror("pthread_join");
 932                 }
 933
 934                 (*nr_running)--;
 935                 (*m_rate) -= td->ratemin;
 936                 (*t_rate) -= td->rate;
 937
 938                 if (td->error)
 939                         exit_value++;
 940         }
 941
 942         if (*nr_running == cputhreads && !pending)
 943                 terminate_threads(TERMINATE_ALL);
 944 }
 945
 946 /*
 947  * Main function for kicking off and reaping jobs, as needed.
 948  */
 949 static void run_threads(void)
 950 {
 951         struct thread_data *td;
 952         unsigned long spent;
 953         int i, todo, nr_running, m_rate, t_rate, nr_started;
 954
 955         if (fio_pin_memory())
 956                 return;
 957
 958         if (!terse_output) {
 959                 printf("Starting ");
 960                 if (nr_thread)
 961                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
 962                 if (nr_process) {
 963                         if (nr_thread)
 964                                 printf(" and ");
 965                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
 966                 }
 967                 printf("\n");
 968                 fflush(stdout);
 969         }
 970
 971         signal(SIGINT, sig_handler);
 972         signal(SIGALRM, sig_handler);
 973
 974         todo = thread_number;
 975         nr_running = 0;
 976         nr_started = 0;
 977         m_rate = t_rate = 0;
 978
 979         for_each_td(td, i) {
 980                 print_status_init(td->thread_number - 1);
 981
 982                 if (!td->create_serialize) {
 983                         init_disk_util(td);
 984                         continue;
 985                 }
 986
 987                 /*
 988                  * do file setup here so it happens sequentially,
 989                  * we don't want X number of threads getting their
 990                  * client data interspersed on disk
 991                  */
 992                 if (setup_files(td)) {
 993                         exit_value++;
 994                         if (td->error)
 995                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 996                         td_set_runstate(td, TD_REAPED);
 997                         todo--;
 998                 }
 999
1000                 init_disk_util(td);
1001         }
1002
1003         set_genesis_time();
1004
1005         while (todo) {
1006                 struct thread_data *map[MAX_JOBS];
1007                 struct timeval this_start;
1008                 int this_jobs = 0, left;
1009
1010                 /*
1011                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1012                  */
1013                 for_each_td(td, i) {
1014                         if (td->runstate != TD_NOT_CREATED)
1015                                 continue;
1016
1017                         /*
1018                          * never got a chance to start, killed by other
1019                          * thread for some reason
1020                          */
1021                         if (td->terminate) {
1022                                 todo--;
1023                                 continue;
1024                         }
1025
1026                         if (td->start_delay) {
1027                                 spent = mtime_since_genesis();
1028
1029                                 if (td->start_delay * 1000 > spent)
1030                                         continue;
1031                         }
1032
1033                         if (td->stonewall && (nr_started || nr_running))
1034                                 break;
1035
1036                         /*
1037                          * Set state to created. Thread will transition
1038                          * to TD_INITIALIZED when it's done setting up.
1039                          */
1040                         td_set_runstate(td, TD_CREATED);
1041                         map[this_jobs++] = td;
1042                         nr_started++;
1043
1044                         if (td->use_thread) {
1045                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1046                                         perror("thread_create");
1047                                         nr_started--;
1048                                         break;
1049                                 }
1050                         } else {
1051                                 if (!fork()) {
1052                                         int ret = fork_main(shm_id, i);
1053
1054                                         exit(ret);
1055                                 }
1056                         }
1057                         fio_sem_down(startup_sem);
1058                 }
1059
1060                 /*
1061                  * Wait for the started threads to transition to
1062                  * TD_INITIALIZED.
1063                  */
1064                 fio_gettime(&this_start, NULL);
1065                 left = this_jobs;
1066                 while (left && !fio_abort) {
1067                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1068                                 break;
1069
1070                         usleep(100000);
1071
1072                         for (i = 0; i < this_jobs; i++) {
1073                                 td = map[i];
1074                                 if (!td)
1075                                         continue;
1076                                 if (td->runstate == TD_INITIALIZED) {
1077                                         map[i] = NULL;
1078                                         left--;
1079                                 } else if (td->runstate >= TD_EXITED) {
1080                                         map[i] = NULL;
1081                                         left--;
1082                                         todo--;
1083                                         nr_running++; /* work-around... */
1084                                 }
1085                         }
1086                 }
1087
1088                 if (left) {
1089                         log_err("fio: %d jobs failed to start\n", left);
1090                         for (i = 0; i < this_jobs; i++) {
1091                                 td = map[i];
1092                                 if (!td)
1093                                         continue;
1094                                 kill(td->pid, SIGTERM);
1095                         }
1096                         break;
1097                 }
1098
1099                 /*
1100                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1101                  */
1102                 for_each_td(td, i) {
1103                         if (td->runstate != TD_INITIALIZED)
1104                                 continue;
1105
1106                         td_set_runstate(td, TD_RUNNING);
1107                         nr_running++;
1108                         nr_started--;
1109                         m_rate += td->ratemin;
1110                         t_rate += td->rate;
1111                         todo--;
1112                         fio_sem_up(td->mutex);
1113                 }
1114
1115                 reap_threads(&nr_running, &t_rate, &m_rate);
1116
1117                 if (todo)
1118                         usleep(100000);
1119         }
1120
1121         while (nr_running) {
1122                 reap_threads(&nr_running, &t_rate, &m_rate);
1123                 usleep(10000);
1124         }
1125
1126         update_io_ticks();
1127         fio_unpin_memory();
1128 }
1129
1130 int main(int argc, char *argv[])
1131 {
1132         long ps;
1133
1134         /*
1135          * We need locale for number printing, if it isn't set then just
1136          * go with the US format.
1137          */
1138         if (!getenv("LC_NUMERIC"))
1139                 setlocale(LC_NUMERIC, "en_US");
1140
1141         if (parse_options(argc, argv))
1142                 return 1;
1143
1144         if (!thread_number) {
1145                 log_err("Nothing to do\n");
1146                 return 1;
1147         }
1148
1149         ps = sysconf(_SC_PAGESIZE);
1150         if (ps < 0) {
1151                 log_err("Failed to get page size\n");
1152                 return 1;
1153         }
1154
1155         page_mask = ps - 1;
1156
1157         if (write_bw_log) {
1158                 setup_log(&agg_io_log[DDIR_READ]);
1159                 setup_log(&agg_io_log[DDIR_WRITE]);
1160         }
1161
1162         startup_sem = fio_sem_init(0);
1163
1164         set_genesis_time();
1165
1166         disk_util_timer_arm();
1167
1168         run_threads();
1169
1170         if (!fio_abort) {
1171                 show_run_stats();
1172                 if (write_bw_log) {
1173                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1174                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1175                 }
1176         }
1177
1178         fio_sem_remove(startup_sem);
1179         return exit_value;
1180 }