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