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