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