backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 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 <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/shm.h>
  38 #include <sys/mman.h>
  39
  40 #include "fio.h"
  41 #include "hash.h"
  42 #include "smalloc.h"
  43 #include "verify.h"
  44 #include "trim.h"
  45 #include "diskutil.h"
  46 #include "cgroup.h"
  47 #include "profile.h"
  48 #include "lib/rand.h"
  49 #include "memalign.h"
  50 #include "server.h"
  51
  52 static pthread_t disk_util_thread;
  53 static struct fio_mutex *startup_mutex;
  54 static struct fio_mutex *writeout_mutex;
  55 static struct flist_head *cgroup_list;
  56 static char *cgroup_mnt;
  57 static int exit_value;
  58 static volatile int fio_abort;
  59 static unsigned int nr_process = 0;
  60 static unsigned int nr_thread = 0;
  61
  62 struct io_log *agg_io_log[2];
  63
  64 int groupid = 0;
  65 unsigned int thread_number = 0;
  66 int shm_id = 0;
  67 int temp_stall_ts;
  68 unsigned long done_secs = 0;
  69
  70 #define PAGE_ALIGN(buf) \
  71         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  72
  73 #define JOB_START_TIMEOUT       (5 * 1000)
  74
  75 static void sig_int(int sig)
  76 {
  77         if (threads) {
  78                 if (is_backend)
  79                         fio_server_got_signal(sig);
  80                 else {
  81                         log_info("\nfio: terminating on signal %d\n", sig);
  82                         fflush(stdout);
  83                         exit_value = 128;
  84                 }
  85
  86                 fio_terminate_threads(TERMINATE_ALL);
  87         }
  88 }
  89
  90 static void sig_show_status(int sig)
  91 {
  92         show_running_run_stats();
  93 }
  94
  95 static void set_sig_handlers(void)
  96 {
  97         struct sigaction act;
  98
  99         memset(&act, 0, sizeof(act));
 100         act.sa_handler = sig_int;
 101         act.sa_flags = SA_RESTART;
 102         sigaction(SIGINT, &act, NULL);
 103
 104         memset(&act, 0, sizeof(act));
 105         act.sa_handler = sig_int;
 106         act.sa_flags = SA_RESTART;
 107         sigaction(SIGTERM, &act, NULL);
 108
 109         memset(&act, 0, sizeof(act));
 110         act.sa_handler = sig_show_status;
 111         act.sa_flags = SA_RESTART;
 112         sigaction(SIGUSR1, &act, NULL);
 113
 114         if (is_backend) {
 115                 memset(&act, 0, sizeof(act));
 116                 act.sa_handler = sig_int;
 117                 act.sa_flags = SA_RESTART;
 118                 sigaction(SIGPIPE, &act, NULL);
 119         }
 120 }
 121
 122 /*
 123  * Check if we are above the minimum rate given.
 124  */
 125 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 126                             enum fio_ddir ddir)
 127 {
 128         unsigned long long bytes = 0;
 129         unsigned long iops = 0;
 130         unsigned long spent;
 131         unsigned long rate;
 132         unsigned int ratemin = 0;
 133         unsigned int rate_iops = 0;
 134         unsigned int rate_iops_min = 0;
 135
 136         assert(ddir_rw(ddir));
 137
 138         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 139                 return 0;
 140
 141         /*
 142          * allow a 2 second settle period in the beginning
 143          */
 144         if (mtime_since(&td->start, now) < 2000)
 145                 return 0;
 146
 147         iops += td->this_io_blocks[ddir];
 148         bytes += td->this_io_bytes[ddir];
 149         ratemin += td->o.ratemin[ddir];
 150         rate_iops += td->o.rate_iops[ddir];
 151         rate_iops_min += td->o.rate_iops_min[ddir];
 152
 153         /*
 154          * if rate blocks is set, sample is running
 155          */
 156         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 157                 spent = mtime_since(&td->lastrate[ddir], now);
 158                 if (spent < td->o.ratecycle)
 159                         return 0;
 160
 161                 if (td->o.rate[ddir]) {
 162                         /*
 163                          * check bandwidth specified rate
 164                          */
 165                         if (bytes < td->rate_bytes[ddir]) {
 166                                 log_err("%s: min rate %u not met\n", td->o.name,
 167                                                                 ratemin);
 168                                 return 1;
 169                         } else {
 170                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 171                                 if (rate < ratemin ||
 172                                     bytes < td->rate_bytes[ddir]) {
 173                                         log_err("%s: min rate %u not met, got"
 174                                                 " %luKB/sec\n", td->o.name,
 175                                                         ratemin, rate);
 176                                         return 1;
 177                                 }
 178                         }
 179                 } else {
 180                         /*
 181                          * checks iops specified rate
 182                          */
 183                         if (iops < rate_iops) {
 184                                 log_err("%s: min iops rate %u not met\n",
 185                                                 td->o.name, rate_iops);
 186                                 return 1;
 187                         } else {
 188                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 189                                 if (rate < rate_iops_min ||
 190                                     iops < td->rate_blocks[ddir]) {
 191                                         log_err("%s: min iops rate %u not met,"
 192                                                 " got %lu\n", td->o.name,
 193                                                         rate_iops_min, rate);
 194                                 }
 195                         }
 196                 }
 197         }
 198
 199         td->rate_bytes[ddir] = bytes;
 200         td->rate_blocks[ddir] = iops;
 201         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 202         return 0;
 203 }
 204
 205 static int check_min_rate(struct thread_data *td, struct timeval *now,
 206                           unsigned long *bytes_done)
 207 {
 208         int ret = 0;
 209
 210         if (bytes_done[0])
 211                 ret |= __check_min_rate(td, now, 0);
 212         if (bytes_done[1])
 213                 ret |= __check_min_rate(td, now, 1);
 214
 215         return ret;
 216 }
 217
 218 /*
 219  * When job exits, we can cancel the in-flight IO if we are using async
 220  * io. Attempt to do so.
 221  */
 222 static void cleanup_pending_aio(struct thread_data *td)
 223 {
 224         struct flist_head *entry, *n;
 225         struct io_u *io_u;
 226         int r;
 227
 228         /*
 229          * get immediately available events, if any
 230          */
 231         r = io_u_queued_complete(td, 0, NULL);
 232         if (r < 0)
 233                 return;
 234
 235         /*
 236          * now cancel remaining active events
 237          */
 238         if (td->io_ops->cancel) {
 239                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 240                         io_u = flist_entry(entry, struct io_u, list);
 241
 242                         /*
 243                          * if the io_u isn't in flight, then that generally
 244                          * means someone leaked an io_u. complain but fix
 245                          * it up, so we don't stall here.
 246                          */
 247                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 248                                 log_err("fio: non-busy IO on busy list\n");
 249                                 put_io_u(td, io_u);
 250                         } else {
 251                                 r = td->io_ops->cancel(td, io_u);
 252                                 if (!r)
 253                                         put_io_u(td, io_u);
 254                         }
 255                 }
 256         }
 257
 258         if (td->cur_depth)
 259                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 260 }
 261
 262 /*
 263  * Helper to handle the final sync of a file. Works just like the normal
 264  * io path, just does everything sync.
 265  */
 266 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 267 {
 268         struct io_u *io_u = __get_io_u(td);
 269         int ret;
 270
 271         if (!io_u)
 272                 return 1;
 273
 274         io_u->ddir = DDIR_SYNC;
 275         io_u->file = f;
 276
 277         if (td_io_prep(td, io_u)) {
 278                 put_io_u(td, io_u);
 279                 return 1;
 280         }
 281
 282 requeue:
 283         ret = td_io_queue(td, io_u);
 284         if (ret < 0) {
 285                 td_verror(td, io_u->error, "td_io_queue");
 286                 put_io_u(td, io_u);
 287                 return 1;
 288         } else if (ret == FIO_Q_QUEUED) {
 289                 if (io_u_queued_complete(td, 1, NULL) < 0)
 290                         return 1;
 291         } else if (ret == FIO_Q_COMPLETED) {
 292                 if (io_u->error) {
 293                         td_verror(td, io_u->error, "td_io_queue");
 294                         return 1;
 295                 }
 296
 297                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 298                         return 1;
 299         } else if (ret == FIO_Q_BUSY) {
 300                 if (td_io_commit(td))
 301                         return 1;
 302                 goto requeue;
 303         }
 304
 305         return 0;
 306 }
 307
 308 static inline void __update_tv_cache(struct thread_data *td)
 309 {
 310         fio_gettime(&td->tv_cache, NULL);
 311 }
 312
 313 static inline void update_tv_cache(struct thread_data *td)
 314 {
 315         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 316                 __update_tv_cache(td);
 317 }
 318
 319 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 320 {
 321         if (in_ramp_time(td))
 322                 return 0;
 323         if (!td->o.timeout)
 324                 return 0;
 325         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 326                 return 1;
 327
 328         return 0;
 329 }
 330
 331 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 332                                int *retptr)
 333 {
 334         int ret = *retptr;
 335
 336         if (ret < 0 || td->error) {
 337                 int err;
 338
 339                 if (ret < 0)
 340                         err = -ret;
 341                 else
 342                         err = td->error;
 343
 344                 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
 345                         return 1;
 346
 347                 if (td_non_fatal_error(err)) {
 348                         /*
 349                          * Continue with the I/Os in case of
 350                          * a non fatal error.
 351                          */
 352                         update_error_count(td, err);
 353                         td_clear_error(td);
 354                         *retptr = 0;
 355                         return 0;
 356                 } else if (td->o.fill_device && err == ENOSPC) {
 357                         /*
 358                          * We expect to hit this error if
 359                          * fill_device option is set.
 360                          */
 361                         td_clear_error(td);
 362                         td->terminate = 1;
 363                         return 1;
 364                 } else {
 365                         /*
 366                          * Stop the I/O in case of a fatal
 367                          * error.
 368                          */
 369                         update_error_count(td, err);
 370                         return 1;
 371                 }
 372         }
 373
 374         return 0;
 375 }
 376
 377 /*
 378  * The main verify engine. Runs over the writes we previously submitted,
 379  * reads the blocks back in, and checks the crc/md5 of the data.
 380  */
 381 static void do_verify(struct thread_data *td)
 382 {
 383         struct fio_file *f;
 384         struct io_u *io_u;
 385         int ret, min_events;
 386         unsigned int i;
 387
 388         dprint(FD_VERIFY, "starting loop\n");
 389
 390         /*
 391          * sync io first and invalidate cache, to make sure we really
 392          * read from disk.
 393          */
 394         for_each_file(td, f, i) {
 395                 if (!fio_file_open(f))
 396                         continue;
 397                 if (fio_io_sync(td, f))
 398                         break;
 399                 if (file_invalidate_cache(td, f))
 400                         break;
 401         }
 402
 403         if (td->error)
 404                 return;
 405
 406         td_set_runstate(td, TD_VERIFYING);
 407
 408         io_u = NULL;
 409         while (!td->terminate) {
 410                 int ret2, full;
 411
 412                 update_tv_cache(td);
 413
 414                 if (runtime_exceeded(td, &td->tv_cache)) {
 415                         __update_tv_cache(td);
 416                         if (runtime_exceeded(td, &td->tv_cache)) {
 417                                 td->terminate = 1;
 418                                 break;
 419                         }
 420                 }
 421
 422                 if (flow_threshold_exceeded(td))
 423                         continue;
 424
 425                 io_u = __get_io_u(td);
 426                 if (!io_u)
 427                         break;
 428
 429                 if (get_next_verify(td, io_u)) {
 430                         put_io_u(td, io_u);
 431                         break;
 432                 }
 433
 434                 if (td_io_prep(td, io_u)) {
 435                         put_io_u(td, io_u);
 436                         break;
 437                 }
 438
 439                 if (td->o.verify_async)
 440                         io_u->end_io = verify_io_u_async;
 441                 else
 442                         io_u->end_io = verify_io_u;
 443
 444                 ret = td_io_queue(td, io_u);
 445                 switch (ret) {
 446                 case FIO_Q_COMPLETED:
 447                         if (io_u->error) {
 448                                 ret = -io_u->error;
 449                                 clear_io_u(td, io_u);
 450                         } else if (io_u->resid) {
 451                                 int bytes = io_u->xfer_buflen - io_u->resid;
 452
 453                                 /*
 454                                  * zero read, fail
 455                                  */
 456                                 if (!bytes) {
 457                                         td_verror(td, EIO, "full resid");
 458                                         put_io_u(td, io_u);
 459                                         break;
 460                                 }
 461
 462                                 io_u->xfer_buflen = io_u->resid;
 463                                 io_u->xfer_buf += bytes;
 464                                 io_u->offset += bytes;
 465
 466                                 if (ddir_rw(io_u->ddir))
 467                                         td->ts.short_io_u[io_u->ddir]++;
 468
 469                                 f = io_u->file;
 470                                 if (io_u->offset == f->real_file_size)
 471                                         goto sync_done;
 472
 473                                 requeue_io_u(td, &io_u);
 474                         } else {
 475 sync_done:
 476                                 ret = io_u_sync_complete(td, io_u, NULL);
 477                                 if (ret < 0)
 478                                         break;
 479                         }
 480                         continue;
 481                 case FIO_Q_QUEUED:
 482                         break;
 483                 case FIO_Q_BUSY:
 484                         requeue_io_u(td, &io_u);
 485                         ret2 = td_io_commit(td);
 486                         if (ret2 < 0)
 487                                 ret = ret2;
 488                         break;
 489                 default:
 490                         assert(ret < 0);
 491                         td_verror(td, -ret, "td_io_queue");
 492                         break;
 493                 }
 494
 495                 if (break_on_this_error(td, io_u->ddir, &ret))
 496                         break;
 497
 498                 /*
 499                  * if we can queue more, do so. but check if there are
 500                  * completed io_u's first. Note that we can get BUSY even
 501                  * without IO queued, if the system is resource starved.
 502                  */
 503                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 504                 if (full || !td->o.iodepth_batch_complete) {
 505                         min_events = min(td->o.iodepth_batch_complete,
 506                                          td->cur_depth);
 507                         /*
 508                          * if the queue is full, we MUST reap at least 1 event
 509                          */
 510                         if (full && !min_events)
 511                                 min_events = 1;
 512
 513                         do {
 514                                 /*
 515                                  * Reap required number of io units, if any,
 516                                  * and do the verification on them through
 517                                  * the callback handler
 518                                  */
 519                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 520                                         ret = -1;
 521                                         break;
 522                                 }
 523                         } while (full && (td->cur_depth > td->o.iodepth_low));
 524                 }
 525                 if (ret < 0)
 526                         break;
 527         }
 528
 529         if (!td->error) {
 530                 min_events = td->cur_depth;
 531
 532                 if (min_events)
 533                         ret = io_u_queued_complete(td, min_events, NULL);
 534         } else
 535                 cleanup_pending_aio(td);
 536
 537         td_set_runstate(td, TD_RUNNING);
 538
 539         dprint(FD_VERIFY, "exiting loop\n");
 540 }
 541
 542 static int io_bytes_exceeded(struct thread_data *td)
 543 {
 544         unsigned long long bytes;
 545
 546         if (td_rw(td))
 547                 bytes = td->this_io_bytes[0] + td->this_io_bytes[1];
 548         else if (td_write(td))
 549                 bytes = td->this_io_bytes[1];
 550         else
 551                 bytes = td->this_io_bytes[0];
 552
 553         return bytes >= td->o.size;
 554 }
 555
 556 /*
 557  * Main IO worker function. It retrieves io_u's to process and queues
 558  * and reaps them, checking for rate and errors along the way.
 559  */
 560 static void do_io(struct thread_data *td)
 561 {
 562         unsigned int i;
 563         int ret = 0;
 564
 565         if (in_ramp_time(td))
 566                 td_set_runstate(td, TD_RAMP);
 567         else
 568                 td_set_runstate(td, TD_RUNNING);
 569
 570         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 571                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 572                 td->o.time_based) {
 573                 struct timeval comp_time;
 574                 unsigned long bytes_done[2] = { 0, 0 };
 575                 int min_evts = 0;
 576                 struct io_u *io_u;
 577                 int ret2, full;
 578                 enum fio_ddir ddir;
 579
 580                 if (td->terminate)
 581                         break;
 582
 583                 update_tv_cache(td);
 584
 585                 if (runtime_exceeded(td, &td->tv_cache)) {
 586                         __update_tv_cache(td);
 587                         if (runtime_exceeded(td, &td->tv_cache)) {
 588                                 td->terminate = 1;
 589                                 break;
 590                         }
 591                 }
 592
 593                 if (flow_threshold_exceeded(td))
 594                         continue;
 595
 596                 io_u = get_io_u(td);
 597                 if (!io_u)
 598                         break;
 599
 600                 ddir = io_u->ddir;
 601
 602                 /*
 603                  * Add verification end_io handler if:
 604                  *      - Asked to verify (!td_rw(td))
 605                  *      - Or the io_u is from our verify list (mixed write/ver)
 606                  */
 607                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 608                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 609                         if (td->o.verify_async)
 610                                 io_u->end_io = verify_io_u_async;
 611                         else
 612                                 io_u->end_io = verify_io_u;
 613                         td_set_runstate(td, TD_VERIFYING);
 614                 } else if (in_ramp_time(td))
 615                         td_set_runstate(td, TD_RAMP);
 616                 else
 617                         td_set_runstate(td, TD_RUNNING);
 618
 619                 ret = td_io_queue(td, io_u);
 620                 switch (ret) {
 621                 case FIO_Q_COMPLETED:
 622                         if (io_u->error) {
 623                                 ret = -io_u->error;
 624                                 clear_io_u(td, io_u);
 625                         } else if (io_u->resid) {
 626                                 int bytes = io_u->xfer_buflen - io_u->resid;
 627                                 struct fio_file *f = io_u->file;
 628
 629                                 /*
 630                                  * zero read, fail
 631                                  */
 632                                 if (!bytes) {
 633                                         td_verror(td, EIO, "full resid");
 634                                         put_io_u(td, io_u);
 635                                         break;
 636                                 }
 637
 638                                 io_u->xfer_buflen = io_u->resid;
 639                                 io_u->xfer_buf += bytes;
 640                                 io_u->offset += bytes;
 641
 642                                 if (ddir_rw(io_u->ddir))
 643                                         td->ts.short_io_u[io_u->ddir]++;
 644
 645                                 if (io_u->offset == f->real_file_size)
 646                                         goto sync_done;
 647
 648                                 requeue_io_u(td, &io_u);
 649                         } else {
 650 sync_done:
 651                                 if (__should_check_rate(td, 0) ||
 652                                     __should_check_rate(td, 1))
 653                                         fio_gettime(&comp_time, NULL);
 654
 655                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 656                                 if (ret < 0)
 657                                         break;
 658                         }
 659                         break;
 660                 case FIO_Q_QUEUED:
 661                         /*
 662                          * if the engine doesn't have a commit hook,
 663                          * the io_u is really queued. if it does have such
 664                          * a hook, it has to call io_u_queued() itself.
 665                          */
 666                         if (td->io_ops->commit == NULL)
 667                                 io_u_queued(td, io_u);
 668                         break;
 669                 case FIO_Q_BUSY:
 670                         requeue_io_u(td, &io_u);
 671                         ret2 = td_io_commit(td);
 672                         if (ret2 < 0)
 673                                 ret = ret2;
 674                         break;
 675                 default:
 676                         assert(ret < 0);
 677                         put_io_u(td, io_u);
 678                         break;
 679                 }
 680
 681                 if (break_on_this_error(td, ddir, &ret))
 682                         break;
 683
 684                 /*
 685                  * See if we need to complete some commands. Note that we
 686                  * can get BUSY even without IO queued, if the system is
 687                  * resource starved.
 688                  */
 689                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 690                 if (full || !td->o.iodepth_batch_complete) {
 691                         min_evts = min(td->o.iodepth_batch_complete,
 692                                         td->cur_depth);
 693                         /*
 694                          * if the queue is full, we MUST reap at least 1 event
 695                          */
 696                         if (full && !min_evts)
 697                                 min_evts = 1;
 698
 699                         if (__should_check_rate(td, 0) ||
 700                             __should_check_rate(td, 1))
 701                                 fio_gettime(&comp_time, NULL);
 702
 703                         do {
 704                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 705                                 if (ret < 0)
 706                                         break;
 707
 708                         } while (full && (td->cur_depth > td->o.iodepth_low));
 709                 }
 710
 711                 if (ret < 0)
 712                         break;
 713                 if (!(bytes_done[0] + bytes_done[1]))
 714                         continue;
 715
 716                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 717                         if (check_min_rate(td, &comp_time, bytes_done)) {
 718                                 if (exitall_on_terminate)
 719                                         fio_terminate_threads(td->groupid);
 720                                 td_verror(td, EIO, "check_min_rate");
 721                                 break;
 722                         }
 723                 }
 724
 725                 if (td->o.thinktime) {
 726                         unsigned long long b;
 727
 728                         b = td->io_blocks[0] + td->io_blocks[1];
 729                         if (!(b % td->o.thinktime_blocks)) {
 730                                 int left;
 731
 732                                 if (td->o.thinktime_spin)
 733                                         usec_spin(td->o.thinktime_spin);
 734
 735                                 left = td->o.thinktime - td->o.thinktime_spin;
 736                                 if (left)
 737                                         usec_sleep(td, left);
 738                         }
 739                 }
 740         }
 741
 742         if (td->trim_entries)
 743                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 744
 745         if (td->o.fill_device && td->error == ENOSPC) {
 746                 td->error = 0;
 747                 td->terminate = 1;
 748         }
 749         if (!td->error) {
 750                 struct fio_file *f;
 751
 752                 i = td->cur_depth;
 753                 if (i) {
 754                         ret = io_u_queued_complete(td, i, NULL);
 755                         if (td->o.fill_device && td->error == ENOSPC)
 756                                 td->error = 0;
 757                 }
 758
 759                 if (should_fsync(td) && td->o.end_fsync) {
 760                         td_set_runstate(td, TD_FSYNCING);
 761
 762                         for_each_file(td, f, i) {
 763                                 if (!fio_file_open(f))
 764                                         continue;
 765                                 fio_io_sync(td, f);
 766                         }
 767                 }
 768         } else
 769                 cleanup_pending_aio(td);
 770
 771         /*
 772          * stop job if we failed doing any IO
 773          */
 774         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 775                 td->done = 1;
 776 }
 777
 778 static void cleanup_io_u(struct thread_data *td)
 779 {
 780         struct flist_head *entry, *n;
 781         struct io_u *io_u;
 782
 783         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 784                 io_u = flist_entry(entry, struct io_u, list);
 785
 786                 flist_del(&io_u->list);
 787                 fio_memfree(io_u, sizeof(*io_u));
 788         }
 789
 790         free_io_mem(td);
 791 }
 792
 793 static int init_io_u(struct thread_data *td)
 794 {
 795         struct io_u *io_u;
 796         unsigned int max_bs, min_write;
 797         int cl_align, i, max_units;
 798         char *p;
 799
 800         max_units = td->o.iodepth;
 801         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 802         min_write = td->o.min_bs[DDIR_WRITE];
 803         td->orig_buffer_size = (unsigned long long) max_bs
 804                                         * (unsigned long long) max_units;
 805
 806         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 807                 unsigned long bs;
 808
 809                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 810                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 811         }
 812
 813         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 814                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 815                 return 1;
 816         }
 817
 818         if (allocate_io_mem(td))
 819                 return 1;
 820
 821         if (td->o.odirect || td->o.mem_align ||
 822             (td->io_ops->flags & FIO_RAWIO))
 823                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 824         else
 825                 p = td->orig_buffer;
 826
 827         cl_align = os_cache_line_size();
 828
 829         for (i = 0; i < max_units; i++) {
 830                 void *ptr;
 831
 832                 if (td->terminate)
 833                         return 1;
 834
 835                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 836                 if (!ptr) {
 837                         log_err("fio: unable to allocate aligned memory\n");
 838                         break;
 839                 }
 840
 841                 io_u = ptr;
 842                 memset(io_u, 0, sizeof(*io_u));
 843                 INIT_FLIST_HEAD(&io_u->list);
 844                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 845
 846                 if (!(td->io_ops->flags & FIO_NOIO)) {
 847                         io_u->buf = p;
 848                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 849
 850                         if (td_write(td))
 851                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 852                         if (td_write(td) && td->o.verify_pattern_bytes) {
 853                                 /*
 854                                  * Fill the buffer with the pattern if we are
 855                                  * going to be doing writes.
 856                                  */
 857                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 858                         }
 859                 }
 860
 861                 io_u->index = i;
 862                 io_u->flags = IO_U_F_FREE;
 863                 flist_add(&io_u->list, &td->io_u_freelist);
 864                 p += max_bs;
 865         }
 866
 867         return 0;
 868 }
 869
 870 static int switch_ioscheduler(struct thread_data *td)
 871 {
 872         char tmp[256], tmp2[128];
 873         FILE *f;
 874         int ret;
 875
 876         if (td->io_ops->flags & FIO_DISKLESSIO)
 877                 return 0;
 878
 879         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 880
 881         f = fopen(tmp, "r+");
 882         if (!f) {
 883                 if (errno == ENOENT) {
 884                         log_err("fio: os or kernel doesn't support IO scheduler"
 885                                 " switching\n");
 886                         return 0;
 887                 }
 888                 td_verror(td, errno, "fopen iosched");
 889                 return 1;
 890         }
 891
 892         /*
 893          * Set io scheduler.
 894          */
 895         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 896         if (ferror(f) || ret != 1) {
 897                 td_verror(td, errno, "fwrite");
 898                 fclose(f);
 899                 return 1;
 900         }
 901
 902         rewind(f);
 903
 904         /*
 905          * Read back and check that the selected scheduler is now the default.
 906          */
 907         ret = fread(tmp, 1, sizeof(tmp), f);
 908         if (ferror(f) || ret < 0) {
 909                 td_verror(td, errno, "fread");
 910                 fclose(f);
 911                 return 1;
 912         }
 913
 914         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 915         if (!strstr(tmp, tmp2)) {
 916                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 917                 td_verror(td, EINVAL, "iosched_switch");
 918                 fclose(f);
 919                 return 1;
 920         }
 921
 922         fclose(f);
 923         return 0;
 924 }
 925
 926 static int keep_running(struct thread_data *td)
 927 {
 928         unsigned long long io_done;
 929
 930         if (td->done)
 931                 return 0;
 932         if (td->o.time_based)
 933                 return 1;
 934         if (td->o.loops) {
 935                 td->o.loops--;
 936                 return 1;
 937         }
 938
 939         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 940                         + td->io_skip_bytes;
 941         if (io_done < td->o.size)
 942                 return 1;
 943
 944         return 0;
 945 }
 946
 947 static int exec_string(const char *string)
 948 {
 949         int ret, newlen = strlen(string) + 1 + 8;
 950         char *str;
 951
 952         str = malloc(newlen);
 953         sprintf(str, "sh -c %s", string);
 954
 955         ret = system(str);
 956         if (ret == -1)
 957                 log_err("fio: exec of cmd <%s> failed\n", str);
 958
 959         free(str);
 960         return ret;
 961 }
 962
 963 /*
 964  * Entry point for the thread based jobs. The process based jobs end up
 965  * here as well, after a little setup.
 966  */
 967 static void *thread_main(void *data)
 968 {
 969         unsigned long long elapsed;
 970         struct thread_data *td = data;
 971         struct thread_options *o = &td->o;
 972         pthread_condattr_t attr;
 973         int clear_state;
 974         int ret;
 975
 976         if (!o->use_thread) {
 977                 setsid();
 978                 td->pid = getpid();
 979         } else
 980                 td->pid = gettid();
 981
 982         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 983
 984         if (is_backend)
 985                 fio_server_send_start(td);
 986
 987         INIT_FLIST_HEAD(&td->io_u_freelist);
 988         INIT_FLIST_HEAD(&td->io_u_busylist);
 989         INIT_FLIST_HEAD(&td->io_u_requeues);
 990         INIT_FLIST_HEAD(&td->io_log_list);
 991         INIT_FLIST_HEAD(&td->io_hist_list);
 992         INIT_FLIST_HEAD(&td->verify_list);
 993         INIT_FLIST_HEAD(&td->trim_list);
 994         pthread_mutex_init(&td->io_u_lock, NULL);
 995         td->io_hist_tree = RB_ROOT;
 996
 997         pthread_condattr_init(&attr);
 998         pthread_cond_init(&td->verify_cond, &attr);
 999         pthread_cond_init(&td->free_cond, &attr);
1000
1001         td_set_runstate(td, TD_INITIALIZED);
1002         dprint(FD_MUTEX, "up startup_mutex\n");
1003         fio_mutex_up(startup_mutex);
1004         dprint(FD_MUTEX, "wait on td->mutex\n");
1005         fio_mutex_down(td->mutex);
1006         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1007
1008         /*
1009          * the ->mutex mutex is now no longer used, close it to avoid
1010          * eating a file descriptor
1011          */
1012         fio_mutex_remove(td->mutex);
1013         td->mutex = NULL;
1014
1015         /*
1016          * A new gid requires privilege, so we need to do this before setting
1017          * the uid.
1018          */
1019         if (o->gid != -1U && setgid(o->gid)) {
1020                 td_verror(td, errno, "setgid");
1021                 goto err;
1022         }
1023         if (o->uid != -1U && setuid(o->uid)) {
1024                 td_verror(td, errno, "setuid");
1025                 goto err;
1026         }
1027
1028         /*
1029          * If we have a gettimeofday() thread, make sure we exclude that
1030          * thread from this job
1031          */
1032         if (o->gtod_cpu)
1033                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1034
1035         /*
1036          * Set affinity first, in case it has an impact on the memory
1037          * allocations.
1038          */
1039         if (o->cpumask_set) {
1040                 ret = fio_setaffinity(td->pid, o->cpumask);
1041                 if (ret == -1) {
1042                         td_verror(td, errno, "cpu_set_affinity");
1043                         goto err;
1044                 }
1045         }
1046
1047         if (fio_pin_memory(td))
1048                 goto err;
1049
1050         /*
1051          * May alter parameters that init_io_u() will use, so we need to
1052          * do this first.
1053          */
1054         if (init_iolog(td))
1055                 goto err;
1056
1057         if (init_io_u(td))
1058                 goto err;
1059
1060         if (o->verify_async && verify_async_init(td))
1061                 goto err;
1062
1063         if (o->ioprio) {
1064                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1065                 if (ret == -1) {
1066                         td_verror(td, errno, "ioprio_set");
1067                         goto err;
1068                 }
1069         }
1070
1071         if (o->cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1072                 goto err;
1073
1074         errno = 0;
1075         if (nice(o->nice) == -1 && errno != 0) {
1076                 td_verror(td, errno, "nice");
1077                 goto err;
1078         }
1079
1080         if (o->ioscheduler && switch_ioscheduler(td))
1081                 goto err;
1082
1083         if (!o->create_serialize && setup_files(td))
1084                 goto err;
1085
1086         if (td_io_init(td))
1087                 goto err;
1088
1089         if (init_random_map(td))
1090                 goto err;
1091
1092         if (o->exec_prerun && exec_string(o->exec_prerun))
1093                 goto err;
1094
1095         if (o->pre_read) {
1096                 if (pre_read_files(td) < 0)
1097                         goto err;
1098         }
1099
1100         fio_verify_init(td);
1101
1102         fio_gettime(&td->epoch, NULL);
1103         getrusage(RUSAGE_SELF, &td->ru_start);
1104
1105         clear_state = 0;
1106         while (keep_running(td)) {
1107                 fio_gettime(&td->start, NULL);
1108                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1109                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1110                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1111
1112                 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1113                         memcpy(&td->lastrate[0], &td->bw_sample_time,
1114                                                 sizeof(td->bw_sample_time));
1115                         memcpy(&td->lastrate[1], &td->bw_sample_time,
1116                                                 sizeof(td->bw_sample_time));
1117                 }
1118
1119                 if (clear_state)
1120                         clear_io_state(td);
1121
1122                 prune_io_piece_log(td);
1123
1124                 do_io(td);
1125
1126                 clear_state = 1;
1127
1128                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1129                         elapsed = utime_since_now(&td->start);
1130                         td->ts.runtime[DDIR_READ] += elapsed;
1131                 }
1132                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1133                         elapsed = utime_since_now(&td->start);
1134                         td->ts.runtime[DDIR_WRITE] += elapsed;
1135                 }
1136
1137                 if (td->error || td->terminate)
1138                         break;
1139
1140                 if (!td->o.do_verify ||
1141                     td->o.verify == VERIFY_NONE ||
1142                     (td->io_ops->flags & FIO_UNIDIR))
1143                         continue;
1144
1145                 clear_io_state(td);
1146
1147                 fio_gettime(&td->start, NULL);
1148
1149                 do_verify(td);
1150
1151                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1152
1153                 if (td->error || td->terminate)
1154                         break;
1155         }
1156
1157         update_rusage_stat(td);
1158         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1159         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1160         td->ts.total_run_time = mtime_since_now(&td->epoch);
1161         td->ts.io_bytes[0] = td->io_bytes[0];
1162         td->ts.io_bytes[1] = td->io_bytes[1];
1163
1164         fio_unpin_memory(td);
1165
1166         fio_mutex_down(writeout_mutex);
1167         if (td->bw_log) {
1168                 if (td->o.bw_log_file) {
1169                         finish_log_named(td, td->bw_log,
1170                                                 td->o.bw_log_file, "bw");
1171                 } else
1172                         finish_log(td, td->bw_log, "bw");
1173         }
1174         if (td->lat_log) {
1175                 if (td->o.lat_log_file) {
1176                         finish_log_named(td, td->lat_log,
1177                                                 td->o.lat_log_file, "lat");
1178                 } else
1179                         finish_log(td, td->lat_log, "lat");
1180         }
1181         if (td->slat_log) {
1182                 if (td->o.lat_log_file) {
1183                         finish_log_named(td, td->slat_log,
1184                                                 td->o.lat_log_file, "slat");
1185                 } else
1186                         finish_log(td, td->slat_log, "slat");
1187         }
1188         if (td->clat_log) {
1189                 if (td->o.lat_log_file) {
1190                         finish_log_named(td, td->clat_log,
1191                                                 td->o.lat_log_file, "clat");
1192                 } else
1193                         finish_log(td, td->clat_log, "clat");
1194         }
1195         if (td->iops_log) {
1196                 if (td->o.iops_log_file) {
1197                         finish_log_named(td, td->iops_log,
1198                                                 td->o.iops_log_file, "iops");
1199                 } else
1200                         finish_log(td, td->iops_log, "iops");
1201         }
1202
1203         fio_mutex_up(writeout_mutex);
1204         if (td->o.exec_postrun)
1205                 exec_string(td->o.exec_postrun);
1206
1207         if (exitall_on_terminate)
1208                 fio_terminate_threads(td->groupid);
1209
1210 err:
1211         if (td->error)
1212                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1213                                                         td->verror);
1214
1215         if (td->o.verify_async)
1216                 verify_async_exit(td);
1217
1218         close_and_free_files(td);
1219         close_ioengine(td);
1220         cleanup_io_u(td);
1221         cgroup_shutdown(td, &cgroup_mnt);
1222
1223         if (o->cpumask_set) {
1224                 int ret = fio_cpuset_exit(&o->cpumask);
1225
1226                 td_verror(td, ret, "fio_cpuset_exit");
1227         }
1228
1229         /*
1230          * do this very late, it will log file closing as well
1231          */
1232         if (td->o.write_iolog_file)
1233                 write_iolog_close(td);
1234
1235         td_set_runstate(td, TD_EXITED);
1236         return (void *) (uintptr_t) td->error;
1237 }
1238
1239
1240 /*
1241  * We cannot pass the td data into a forked process, so attach the td and
1242  * pass it to the thread worker.
1243  */
1244 static int fork_main(int shmid, int offset)
1245 {
1246         struct thread_data *td;
1247         void *data, *ret;
1248
1249 #ifndef __hpux
1250         data = shmat(shmid, NULL, 0);
1251         if (data == (void *) -1) {
1252                 int __err = errno;
1253
1254                 perror("shmat");
1255                 return __err;
1256         }
1257 #else
1258         /*
1259          * HP-UX inherits shm mappings?
1260          */
1261         data = threads;
1262 #endif
1263
1264         td = data + offset * sizeof(struct thread_data);
1265         ret = thread_main(td);
1266         shmdt(data);
1267         return (int) (uintptr_t) ret;
1268 }
1269
1270 /*
1271  * Run over the job map and reap the threads that have exited, if any.
1272  */
1273 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1274                          unsigned int *m_rate)
1275 {
1276         struct thread_data *td;
1277         unsigned int cputhreads, realthreads, pending;
1278         int i, status, ret;
1279
1280         /*
1281          * reap exited threads (TD_EXITED -> TD_REAPED)
1282          */
1283         realthreads = pending = cputhreads = 0;
1284         for_each_td(td, i) {
1285                 int flags = 0;
1286
1287                 /*
1288                  * ->io_ops is NULL for a thread that has closed its
1289                  * io engine
1290                  */
1291                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1292                         cputhreads++;
1293                 else
1294                         realthreads++;
1295
1296                 if (!td->pid) {
1297                         pending++;
1298                         continue;
1299                 }
1300                 if (td->runstate == TD_REAPED)
1301                         continue;
1302                 if (td->o.use_thread) {
1303                         if (td->runstate == TD_EXITED) {
1304                                 td_set_runstate(td, TD_REAPED);
1305                                 goto reaped;
1306                         }
1307                         continue;
1308                 }
1309
1310                 flags = WNOHANG;
1311                 if (td->runstate == TD_EXITED)
1312                         flags = 0;
1313
1314                 /*
1315                  * check if someone quit or got killed in an unusual way
1316                  */
1317                 ret = waitpid(td->pid, &status, flags);
1318                 if (ret < 0) {
1319                         if (errno == ECHILD) {
1320                                 log_err("fio: pid=%d disappeared %d\n",
1321                                                 (int) td->pid, td->runstate);
1322                                 td->sig = ECHILD;
1323                                 td_set_runstate(td, TD_REAPED);
1324                                 goto reaped;
1325                         }
1326                         perror("waitpid");
1327                 } else if (ret == td->pid) {
1328                         if (WIFSIGNALED(status)) {
1329                                 int sig = WTERMSIG(status);
1330
1331                                 if (sig != SIGTERM)
1332                                         log_err("fio: pid=%d, got signal=%d\n",
1333                                                         (int) td->pid, sig);
1334                                 td->sig = sig;
1335                                 td_set_runstate(td, TD_REAPED);
1336                                 goto reaped;
1337                         }
1338                         if (WIFEXITED(status)) {
1339                                 if (WEXITSTATUS(status) && !td->error)
1340                                         td->error = WEXITSTATUS(status);
1341
1342                                 td_set_runstate(td, TD_REAPED);
1343                                 goto reaped;
1344                         }
1345                 }
1346
1347                 /*
1348                  * thread is not dead, continue
1349                  */
1350                 pending++;
1351                 continue;
1352 reaped:
1353                 (*nr_running)--;
1354                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1355                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1356                 if (!td->pid)
1357                         pending--;
1358
1359                 if (td->error)
1360                         exit_value++;
1361
1362                 done_secs += mtime_since_now(&td->epoch) / 1000;
1363         }
1364
1365         if (*nr_running == cputhreads && !pending && realthreads)
1366                 fio_terminate_threads(TERMINATE_ALL);
1367 }
1368
1369 /*
1370  * Main function for kicking off and reaping jobs, as needed.
1371  */
1372 static void run_threads(void)
1373 {
1374         struct thread_data *td;
1375         unsigned long spent;
1376         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1377
1378         if (fio_gtod_offload && fio_start_gtod_thread())
1379                 return;
1380
1381         set_sig_handlers();
1382
1383         nr_thread = nr_process = 0;
1384         for_each_td(td, i) {
1385                 if (td->o.use_thread)
1386                         nr_thread++;
1387                 else
1388                         nr_process++;
1389         }
1390
1391         if (!terse_output) {
1392                 log_info("Starting ");
1393                 if (nr_thread)
1394                         log_info("%d thread%s", nr_thread,
1395                                                 nr_thread > 1 ? "s" : "");
1396                 if (nr_process) {
1397                         if (nr_thread)
1398                                 log_info(" and ");
1399                         log_info("%d process%s", nr_process,
1400                                                 nr_process > 1 ? "es" : "");
1401                 }
1402                 log_info("\n");
1403                 fflush(stdout);
1404         }
1405
1406         todo = thread_number;
1407         nr_running = 0;
1408         nr_started = 0;
1409         m_rate = t_rate = 0;
1410
1411         for_each_td(td, i) {
1412                 print_status_init(td->thread_number - 1);
1413
1414                 if (!td->o.create_serialize)
1415                         continue;
1416
1417                 /*
1418                  * do file setup here so it happens sequentially,
1419                  * we don't want X number of threads getting their
1420                  * client data interspersed on disk
1421                  */
1422                 if (setup_files(td)) {
1423                         exit_value++;
1424                         if (td->error)
1425                                 log_err("fio: pid=%d, err=%d/%s\n",
1426                                         (int) td->pid, td->error, td->verror);
1427                         td_set_runstate(td, TD_REAPED);
1428                         todo--;
1429                 } else {
1430                         struct fio_file *f;
1431                         unsigned int j;
1432
1433                         /*
1434                          * for sharing to work, each job must always open
1435                          * its own files. so close them, if we opened them
1436                          * for creation
1437                          */
1438                         for_each_file(td, f, j) {
1439                                 if (fio_file_open(f))
1440                                         td_io_close_file(td, f);
1441                         }
1442                 }
1443         }
1444
1445         set_genesis_time();
1446
1447         while (todo) {
1448                 struct thread_data *map[REAL_MAX_JOBS];
1449                 struct timeval this_start;
1450                 int this_jobs = 0, left;
1451
1452                 /*
1453                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1454                  */
1455                 for_each_td(td, i) {
1456                         if (td->runstate != TD_NOT_CREATED)
1457                                 continue;
1458
1459                         /*
1460                          * never got a chance to start, killed by other
1461                          * thread for some reason
1462                          */
1463                         if (td->terminate) {
1464                                 todo--;
1465                                 continue;
1466                         }
1467
1468                         if (td->o.start_delay) {
1469                                 spent = mtime_since_genesis();
1470
1471                                 if (td->o.start_delay * 1000 > spent)
1472                                         continue;
1473                         }
1474
1475                         if (td->o.stonewall && (nr_started || nr_running)) {
1476                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1477                                                         td->o.name);
1478                                 break;
1479                         }
1480
1481                         init_disk_util(td);
1482
1483                         /*
1484                          * Set state to created. Thread will transition
1485                          * to TD_INITIALIZED when it's done setting up.
1486                          */
1487                         td_set_runstate(td, TD_CREATED);
1488                         map[this_jobs++] = td;
1489                         nr_started++;
1490
1491                         if (td->o.use_thread) {
1492                                 int ret;
1493
1494                                 dprint(FD_PROCESS, "will pthread_create\n");
1495                                 ret = pthread_create(&td->thread, NULL,
1496                                                         thread_main, td);
1497                                 if (ret) {
1498                                         log_err("pthread_create: %s\n",
1499                                                         strerror(ret));
1500                                         nr_started--;
1501                                         break;
1502                                 }
1503                                 ret = pthread_detach(td->thread);
1504                                 if (ret)
1505                                         log_err("pthread_detach: %s",
1506                                                         strerror(ret));
1507                         } else {
1508                                 pid_t pid;
1509                                 dprint(FD_PROCESS, "will fork\n");
1510                                 pid = fork();
1511                                 if (!pid) {
1512                                         int ret = fork_main(shm_id, i);
1513
1514                                         _exit(ret);
1515                                 } else if (i == fio_debug_jobno)
1516                                         *fio_debug_jobp = pid;
1517                         }
1518                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1519                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1520                                 log_err("fio: job startup hung? exiting.\n");
1521                                 fio_terminate_threads(TERMINATE_ALL);
1522                                 fio_abort = 1;
1523                                 nr_started--;
1524                                 break;
1525                         }
1526                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1527                 }
1528
1529                 /*
1530                  * Wait for the started threads to transition to
1531                  * TD_INITIALIZED.
1532                  */
1533                 fio_gettime(&this_start, NULL);
1534                 left = this_jobs;
1535                 while (left && !fio_abort) {
1536                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1537                                 break;
1538
1539                         usleep(100000);
1540
1541                         for (i = 0; i < this_jobs; i++) {
1542                                 td = map[i];
1543                                 if (!td)
1544                                         continue;
1545                                 if (td->runstate == TD_INITIALIZED) {
1546                                         map[i] = NULL;
1547                                         left--;
1548                                 } else if (td->runstate >= TD_EXITED) {
1549                                         map[i] = NULL;
1550                                         left--;
1551                                         todo--;
1552                                         nr_running++; /* work-around... */
1553                                 }
1554                         }
1555                 }
1556
1557                 if (left) {
1558                         log_err("fio: %d job%s failed to start\n", left,
1559                                         left > 1 ? "s" : "");
1560                         for (i = 0; i < this_jobs; i++) {
1561                                 td = map[i];
1562                                 if (!td)
1563                                         continue;
1564                                 kill(td->pid, SIGTERM);
1565                         }
1566                         break;
1567                 }
1568
1569                 /*
1570                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1571                  */
1572                 for_each_td(td, i) {
1573                         if (td->runstate != TD_INITIALIZED)
1574                                 continue;
1575
1576                         if (in_ramp_time(td))
1577                                 td_set_runstate(td, TD_RAMP);
1578                         else
1579                                 td_set_runstate(td, TD_RUNNING);
1580                         nr_running++;
1581                         nr_started--;
1582                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1583                         t_rate += td->o.rate[0] + td->o.rate[1];
1584                         todo--;
1585                         fio_mutex_up(td->mutex);
1586                 }
1587
1588                 reap_threads(&nr_running, &t_rate, &m_rate);
1589
1590                 if (todo)
1591                         usleep(100000);
1592         }
1593
1594         while (nr_running) {
1595                 reap_threads(&nr_running, &t_rate, &m_rate);
1596                 usleep(10000);
1597         }
1598
1599         update_io_ticks();
1600 }
1601
1602 static void *disk_thread_main(void *data)
1603 {
1604         fio_mutex_up(startup_mutex);
1605
1606         while (threads) {
1607                 usleep(DISK_UTIL_MSEC * 1000);
1608                 if (!threads)
1609                         break;
1610                 update_io_ticks();
1611
1612                 if (!is_backend)
1613                         print_thread_status();
1614         }
1615
1616         return NULL;
1617 }
1618
1619 static int create_disk_util_thread(void)
1620 {
1621         int ret;
1622
1623         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1624         if (ret) {
1625                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1626                 return 1;
1627         }
1628
1629         ret = pthread_detach(disk_util_thread);
1630         if (ret) {
1631                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1632                 return 1;
1633         }
1634
1635         dprint(FD_MUTEX, "wait on startup_mutex\n");
1636         fio_mutex_down(startup_mutex);
1637         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1638         return 0;
1639 }
1640
1641 int fio_backend(void)
1642 {
1643         struct thread_data *td;
1644         int i;
1645
1646         if (exec_profile) {
1647                 if (load_profile(exec_profile))
1648                         return 1;
1649                 free(exec_profile);
1650                 exec_profile = NULL;
1651         }
1652         if (!thread_number)
1653                 return 0;
1654
1655         if (write_bw_log) {
1656                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1657                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1658         }
1659
1660         startup_mutex = fio_mutex_init(0);
1661         if (startup_mutex == NULL)
1662                 return 1;
1663         writeout_mutex = fio_mutex_init(1);
1664         if (writeout_mutex == NULL)
1665                 return 1;
1666
1667         set_genesis_time();
1668         create_disk_util_thread();
1669
1670         cgroup_list = smalloc(sizeof(*cgroup_list));
1671         INIT_FLIST_HEAD(cgroup_list);
1672
1673         run_threads();
1674
1675         if (!fio_abort) {
1676                 show_run_stats();
1677                 if (write_bw_log) {
1678                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1679                         __finish_log(agg_io_log[DDIR_WRITE],
1680                                         "agg-write_bw.log");
1681                 }
1682         }
1683
1684         for_each_td(td, i)
1685                 fio_options_free(td);
1686
1687         free_disk_util();
1688         cgroup_kill(cgroup_list);
1689         sfree(cgroup_list);
1690         sfree(cgroup_mnt);
1691
1692         fio_mutex_remove(startup_mutex);
1693         fio_mutex_remove(writeout_mutex);
1694         return exit_value;
1695 }