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