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 /*
 530  * Main IO worker function. It retrieves io_u's to process and queues
 531  * and reaps them, checking for rate and errors along the way.
 532  */
 533 static void do_io(struct thread_data *td)
 534 {
 535         unsigned int i;
 536         int ret = 0;
 537
 538         if (in_ramp_time(td))
 539                 td_set_runstate(td, TD_RAMP);
 540         else
 541                 td_set_runstate(td, TD_RUNNING);
 542
 543         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 544                 (!flist_empty(&td->trim_list)) ||
 545                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 546                 struct timeval comp_time;
 547                 unsigned long bytes_done[2] = { 0, 0 };
 548                 int min_evts = 0;
 549                 struct io_u *io_u;
 550                 int ret2, full;
 551                 enum fio_ddir ddir;
 552
 553                 if (td->terminate)
 554                         break;
 555
 556                 update_tv_cache(td);
 557
 558                 if (runtime_exceeded(td, &td->tv_cache)) {
 559                         __update_tv_cache(td);
 560                         if (runtime_exceeded(td, &td->tv_cache)) {
 561                                 td->terminate = 1;
 562                                 break;
 563                         }
 564                 }
 565
 566                 if (flow_threshold_exceeded(td))
 567                         continue;
 568
 569                 io_u = get_io_u(td);
 570                 if (!io_u)
 571                         break;
 572
 573                 ddir = io_u->ddir;
 574
 575                 /*
 576                  * Add verification end_io handler, if asked to verify
 577                  * a previously written file.
 578                  */
 579                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 580                     !td_rw(td)) {
 581                         if (td->o.verify_async)
 582                                 io_u->end_io = verify_io_u_async;
 583                         else
 584                                 io_u->end_io = verify_io_u;
 585                         td_set_runstate(td, TD_VERIFYING);
 586                 } else if (in_ramp_time(td))
 587                         td_set_runstate(td, TD_RAMP);
 588                 else
 589                         td_set_runstate(td, TD_RUNNING);
 590
 591                 ret = td_io_queue(td, io_u);
 592                 switch (ret) {
 593                 case FIO_Q_COMPLETED:
 594                         if (io_u->error) {
 595                                 ret = -io_u->error;
 596                                 clear_io_u(td, io_u);
 597                         } else if (io_u->resid) {
 598                                 int bytes = io_u->xfer_buflen - io_u->resid;
 599                                 struct fio_file *f = io_u->file;
 600
 601                                 /*
 602                                  * zero read, fail
 603                                  */
 604                                 if (!bytes) {
 605                                         td_verror(td, EIO, "full resid");
 606                                         put_io_u(td, io_u);
 607                                         break;
 608                                 }
 609
 610                                 io_u->xfer_buflen = io_u->resid;
 611                                 io_u->xfer_buf += bytes;
 612                                 io_u->offset += bytes;
 613
 614                                 if (ddir_rw(io_u->ddir))
 615                                         td->ts.short_io_u[io_u->ddir]++;
 616
 617                                 if (io_u->offset == f->real_file_size)
 618                                         goto sync_done;
 619
 620                                 requeue_io_u(td, &io_u);
 621                         } else {
 622 sync_done:
 623                                 if (__should_check_rate(td, 0) ||
 624                                     __should_check_rate(td, 1))
 625                                         fio_gettime(&comp_time, NULL);
 626
 627                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 628                                 if (ret < 0)
 629                                         break;
 630                         }
 631                         break;
 632                 case FIO_Q_QUEUED:
 633                         /*
 634                          * if the engine doesn't have a commit hook,
 635                          * the io_u is really queued. if it does have such
 636                          * a hook, it has to call io_u_queued() itself.
 637                          */
 638                         if (td->io_ops->commit == NULL)
 639                                 io_u_queued(td, io_u);
 640                         break;
 641                 case FIO_Q_BUSY:
 642                         requeue_io_u(td, &io_u);
 643                         ret2 = td_io_commit(td);
 644                         if (ret2 < 0)
 645                                 ret = ret2;
 646                         break;
 647                 default:
 648                         assert(ret < 0);
 649                         put_io_u(td, io_u);
 650                         break;
 651                 }
 652
 653                 if (break_on_this_error(td, ddir, &ret))
 654                         break;
 655
 656                 /*
 657                  * See if we need to complete some commands. Note that we
 658                  * can get BUSY even without IO queued, if the system is
 659                  * resource starved.
 660                  */
 661                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 662                 if (full || !td->o.iodepth_batch_complete) {
 663                         min_evts = min(td->o.iodepth_batch_complete,
 664                                         td->cur_depth);
 665                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 666                                 min_evts = 1;
 667
 668                         if (__should_check_rate(td, 0) ||
 669                             __should_check_rate(td, 1))
 670                                 fio_gettime(&comp_time, NULL);
 671
 672                         do {
 673                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 674                                 if (ret < 0)
 675                                         break;
 676
 677                         } while (full && (td->cur_depth > td->o.iodepth_low));
 678                 }
 679
 680                 if (ret < 0)
 681                         break;
 682                 if (!(bytes_done[0] + bytes_done[1]))
 683                         continue;
 684
 685                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 686                         if (check_min_rate(td, &comp_time, bytes_done)) {
 687                                 if (exitall_on_terminate)
 688                                         fio_terminate_threads(td->groupid);
 689                                 td_verror(td, EIO, "check_min_rate");
 690                                 break;
 691                         }
 692                 }
 693
 694                 if (td->o.thinktime) {
 695                         unsigned long long b;
 696
 697                         b = td->io_blocks[0] + td->io_blocks[1];
 698                         if (!(b % td->o.thinktime_blocks)) {
 699                                 int left;
 700
 701                                 if (td->o.thinktime_spin)
 702                                         usec_spin(td->o.thinktime_spin);
 703
 704                                 left = td->o.thinktime - td->o.thinktime_spin;
 705                                 if (left)
 706                                         usec_sleep(td, left);
 707                         }
 708                 }
 709         }
 710
 711         if (td->trim_entries)
 712                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 713
 714         if (td->o.fill_device && td->error == ENOSPC) {
 715                 td->error = 0;
 716                 td->terminate = 1;
 717         }
 718         if (!td->error) {
 719                 struct fio_file *f;
 720
 721                 i = td->cur_depth;
 722                 if (i) {
 723                         ret = io_u_queued_complete(td, i, NULL);
 724                         if (td->o.fill_device && td->error == ENOSPC)
 725                                 td->error = 0;
 726                 }
 727
 728                 if (should_fsync(td) && td->o.end_fsync) {
 729                         td_set_runstate(td, TD_FSYNCING);
 730
 731                         for_each_file(td, f, i) {
 732                                 if (!fio_file_open(f))
 733                                         continue;
 734                                 fio_io_sync(td, f);
 735                         }
 736                 }
 737         } else
 738                 cleanup_pending_aio(td);
 739
 740         /*
 741          * stop job if we failed doing any IO
 742          */
 743         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 744                 td->done = 1;
 745 }
 746
 747 static void cleanup_io_u(struct thread_data *td)
 748 {
 749         struct flist_head *entry, *n;
 750         struct io_u *io_u;
 751
 752         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 753                 io_u = flist_entry(entry, struct io_u, list);
 754
 755                 flist_del(&io_u->list);
 756                 fio_memfree(io_u, sizeof(*io_u));
 757         }
 758
 759         free_io_mem(td);
 760 }
 761
 762 static int init_io_u(struct thread_data *td)
 763 {
 764         struct io_u *io_u;
 765         unsigned int max_bs, min_write;
 766         int cl_align, i, max_units;
 767         char *p;
 768
 769         max_units = td->o.iodepth;
 770         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 771         min_write = td->o.min_bs[DDIR_WRITE];
 772         td->orig_buffer_size = (unsigned long long) max_bs
 773                                         * (unsigned long long) max_units;
 774
 775         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 776                 unsigned long bs;
 777
 778                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 779                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 780         }
 781
 782         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 783                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 784                 return 1;
 785         }
 786
 787         if (allocate_io_mem(td))
 788                 return 1;
 789
 790         if (td->o.odirect || td->o.mem_align ||
 791             (td->io_ops->flags & FIO_RAWIO))
 792                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 793         else
 794                 p = td->orig_buffer;
 795
 796         cl_align = os_cache_line_size();
 797
 798         for (i = 0; i < max_units; i++) {
 799                 void *ptr;
 800
 801                 if (td->terminate)
 802                         return 1;
 803
 804                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 805                 if (!ptr) {
 806                         log_err("fio: unable to allocate aligned memory\n");
 807                         break;
 808                 }
 809
 810                 io_u = ptr;
 811                 memset(io_u, 0, sizeof(*io_u));
 812                 INIT_FLIST_HEAD(&io_u->list);
 813                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 814
 815                 if (!(td->io_ops->flags & FIO_NOIO)) {
 816                         io_u->buf = p;
 817                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 818
 819                         if (td_write(td))
 820                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 821                         if (td_write(td) && td->o.verify_pattern_bytes) {
 822                                 /*
 823                                  * Fill the buffer with the pattern if we are
 824                                  * going to be doing writes.
 825                                  */
 826                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 827                         }
 828                 }
 829
 830                 io_u->index = i;
 831                 io_u->flags = IO_U_F_FREE;
 832                 flist_add(&io_u->list, &td->io_u_freelist);
 833                 p += max_bs;
 834         }
 835
 836         return 0;
 837 }
 838
 839 static int switch_ioscheduler(struct thread_data *td)
 840 {
 841         char tmp[256], tmp2[128];
 842         FILE *f;
 843         int ret;
 844
 845         if (td->io_ops->flags & FIO_DISKLESSIO)
 846                 return 0;
 847
 848         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 849
 850         f = fopen(tmp, "r+");
 851         if (!f) {
 852                 if (errno == ENOENT) {
 853                         log_err("fio: os or kernel doesn't support IO scheduler"
 854                                 " switching\n");
 855                         return 0;
 856                 }
 857                 td_verror(td, errno, "fopen iosched");
 858                 return 1;
 859         }
 860
 861         /*
 862          * Set io scheduler.
 863          */
 864         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 865         if (ferror(f) || ret != 1) {
 866                 td_verror(td, errno, "fwrite");
 867                 fclose(f);
 868                 return 1;
 869         }
 870
 871         rewind(f);
 872
 873         /*
 874          * Read back and check that the selected scheduler is now the default.
 875          */
 876         ret = fread(tmp, 1, sizeof(tmp), f);
 877         if (ferror(f) || ret < 0) {
 878                 td_verror(td, errno, "fread");
 879                 fclose(f);
 880                 return 1;
 881         }
 882
 883         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 884         if (!strstr(tmp, tmp2)) {
 885                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 886                 td_verror(td, EINVAL, "iosched_switch");
 887                 fclose(f);
 888                 return 1;
 889         }
 890
 891         fclose(f);
 892         return 0;
 893 }
 894
 895 static int keep_running(struct thread_data *td)
 896 {
 897         unsigned long long io_done;
 898
 899         if (td->done)
 900                 return 0;
 901         if (td->o.time_based)
 902                 return 1;
 903         if (td->o.loops) {
 904                 td->o.loops--;
 905                 return 1;
 906         }
 907
 908         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 909                         + td->io_skip_bytes;
 910         if (io_done < td->o.size)
 911                 return 1;
 912
 913         return 0;
 914 }
 915
 916 static int exec_string(const char *string)
 917 {
 918         int ret, newlen = strlen(string) + 1 + 8;
 919         char *str;
 920
 921         str = malloc(newlen);
 922         sprintf(str, "sh -c %s", string);
 923
 924         ret = system(str);
 925         if (ret == -1)
 926                 log_err("fio: exec of cmd <%s> failed\n", str);
 927
 928         free(str);
 929         return ret;
 930 }
 931
 932 /*
 933  * Entry point for the thread based jobs. The process based jobs end up
 934  * here as well, after a little setup.
 935  */
 936 static void *thread_main(void *data)
 937 {
 938         unsigned long long elapsed;
 939         struct thread_data *td = data;
 940         pthread_condattr_t attr;
 941         int clear_state;
 942
 943         if (!td->o.use_thread) {
 944                 setsid();
 945                 td->pid = getpid();
 946         } else
 947                 td->pid = gettid();
 948
 949         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 950
 951         INIT_FLIST_HEAD(&td->io_u_freelist);
 952         INIT_FLIST_HEAD(&td->io_u_busylist);
 953         INIT_FLIST_HEAD(&td->io_u_requeues);
 954         INIT_FLIST_HEAD(&td->io_log_list);
 955         INIT_FLIST_HEAD(&td->io_hist_list);
 956         INIT_FLIST_HEAD(&td->verify_list);
 957         INIT_FLIST_HEAD(&td->trim_list);
 958         pthread_mutex_init(&td->io_u_lock, NULL);
 959         td->io_hist_tree = RB_ROOT;
 960
 961         pthread_condattr_init(&attr);
 962         pthread_cond_init(&td->verify_cond, &attr);
 963         pthread_cond_init(&td->free_cond, &attr);
 964
 965         td_set_runstate(td, TD_INITIALIZED);
 966         dprint(FD_MUTEX, "up startup_mutex\n");
 967         fio_mutex_up(startup_mutex);
 968         dprint(FD_MUTEX, "wait on td->mutex\n");
 969         fio_mutex_down(td->mutex);
 970         dprint(FD_MUTEX, "done waiting on td->mutex\n");
 971
 972         /*
 973          * the ->mutex mutex is now no longer used, close it to avoid
 974          * eating a file descriptor
 975          */
 976         fio_mutex_remove(td->mutex);
 977
 978         /*
 979          * A new gid requires privilege, so we need to do this before setting
 980          * the uid.
 981          */
 982         if (td->o.gid != -1U && setgid(td->o.gid)) {
 983                 td_verror(td, errno, "setgid");
 984                 goto err;
 985         }
 986         if (td->o.uid != -1U && setuid(td->o.uid)) {
 987                 td_verror(td, errno, "setuid");
 988                 goto err;
 989         }
 990
 991         /*
 992          * If we have a gettimeofday() thread, make sure we exclude that
 993          * thread from this job
 994          */
 995         if (td->o.gtod_cpu)
 996                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
 997
 998         /*
 999          * Set affinity first, in case it has an impact on the memory
1000          * allocations.
1001          */
1002         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1003                 td_verror(td, errno, "cpu_set_affinity");
1004                 goto err;
1005         }
1006
1007         /*
1008          * May alter parameters that init_io_u() will use, so we need to
1009          * do this first.
1010          */
1011         if (init_iolog(td))
1012                 goto err;
1013
1014         if (init_io_u(td))
1015                 goto err;
1016
1017         if (td->o.verify_async && verify_async_init(td))
1018                 goto err;
1019
1020         if (td->ioprio_set) {
1021                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1022                         td_verror(td, errno, "ioprio_set");
1023                         goto err;
1024                 }
1025         }
1026
1027         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1028                 goto err;
1029
1030         errno = 0;
1031         if (nice(td->o.nice) == -1 && errno != 0) {
1032                 td_verror(td, errno, "nice");
1033                 goto err;
1034         }
1035
1036         if (td->o.ioscheduler && switch_ioscheduler(td))
1037                 goto err;
1038
1039         if (!td->o.create_serialize && setup_files(td))
1040                 goto err;
1041
1042         if (td_io_init(td))
1043                 goto err;
1044
1045         if (init_random_map(td))
1046                 goto err;
1047
1048         if (td->o.exec_prerun) {
1049                 if (exec_string(td->o.exec_prerun))
1050                         goto err;
1051         }
1052
1053         if (td->o.pre_read) {
1054                 if (pre_read_files(td) < 0)
1055                         goto err;
1056         }
1057
1058         fio_gettime(&td->epoch, NULL);
1059         getrusage(RUSAGE_SELF, &td->ru_start);
1060
1061         clear_state = 0;
1062         while (keep_running(td)) {
1063                 fio_gettime(&td->start, NULL);
1064                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1065                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1066                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1067
1068                 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1069                         memcpy(&td->lastrate[0], &td->bw_sample_time,
1070                                                 sizeof(td->bw_sample_time));
1071                         memcpy(&td->lastrate[1], &td->bw_sample_time,
1072                                                 sizeof(td->bw_sample_time));
1073                 }
1074
1075                 if (clear_state)
1076                         clear_io_state(td);
1077
1078                 prune_io_piece_log(td);
1079
1080                 do_io(td);
1081
1082                 clear_state = 1;
1083
1084                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1085                         elapsed = utime_since_now(&td->start);
1086                         td->ts.runtime[DDIR_READ] += elapsed;
1087                 }
1088                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1089                         elapsed = utime_since_now(&td->start);
1090                         td->ts.runtime[DDIR_WRITE] += elapsed;
1091                 }
1092
1093                 if (td->error || td->terminate)
1094                         break;
1095
1096                 if (!td->o.do_verify ||
1097                     td->o.verify == VERIFY_NONE ||
1098                     (td->io_ops->flags & FIO_UNIDIR))
1099                         continue;
1100
1101                 clear_io_state(td);
1102
1103                 fio_gettime(&td->start, NULL);
1104
1105                 do_verify(td);
1106
1107                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1108
1109                 if (td->error || td->terminate)
1110                         break;
1111         }
1112
1113         update_rusage_stat(td);
1114         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1115         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1116         td->ts.total_run_time = mtime_since_now(&td->epoch);
1117         td->ts.io_bytes[0] = td->io_bytes[0];
1118         td->ts.io_bytes[1] = td->io_bytes[1];
1119
1120         fio_mutex_down(writeout_mutex);
1121         if (td->bw_log) {
1122                 if (td->o.bw_log_file) {
1123                         finish_log_named(td, td->bw_log,
1124                                                 td->o.bw_log_file, "bw");
1125                 } else
1126                         finish_log(td, td->bw_log, "bw");
1127         }
1128         if (td->lat_log) {
1129                 if (td->o.lat_log_file) {
1130                         finish_log_named(td, td->lat_log,
1131                                                 td->o.lat_log_file, "lat");
1132                 } else
1133                         finish_log(td, td->lat_log, "lat");
1134         }
1135         if (td->slat_log) {
1136                 if (td->o.lat_log_file) {
1137                         finish_log_named(td, td->slat_log,
1138                                                 td->o.lat_log_file, "slat");
1139                 } else
1140                         finish_log(td, td->slat_log, "slat");
1141         }
1142         if (td->clat_log) {
1143                 if (td->o.lat_log_file) {
1144                         finish_log_named(td, td->clat_log,
1145                                                 td->o.lat_log_file, "clat");
1146                 } else
1147                         finish_log(td, td->clat_log, "clat");
1148         }
1149         if (td->iops_log) {
1150                 if (td->o.iops_log_file) {
1151                         finish_log_named(td, td->iops_log,
1152                                                 td->o.iops_log_file, "iops");
1153                 } else
1154                         finish_log(td, td->iops_log, "iops");
1155         }
1156
1157         fio_mutex_up(writeout_mutex);
1158         if (td->o.exec_postrun)
1159                 exec_string(td->o.exec_postrun);
1160
1161         if (exitall_on_terminate)
1162                 fio_terminate_threads(td->groupid);
1163
1164 err:
1165         if (td->error)
1166                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1167                                                         td->verror);
1168
1169         if (td->o.verify_async)
1170                 verify_async_exit(td);
1171
1172         close_and_free_files(td);
1173         close_ioengine(td);
1174         cleanup_io_u(td);
1175         cgroup_shutdown(td, &cgroup_mnt);
1176
1177         if (td->o.cpumask_set) {
1178                 int ret = fio_cpuset_exit(&td->o.cpumask);
1179
1180                 td_verror(td, ret, "fio_cpuset_exit");
1181         }
1182
1183         /*
1184          * do this very late, it will log file closing as well
1185          */
1186         if (td->o.write_iolog_file)
1187                 write_iolog_close(td);
1188
1189         td_set_runstate(td, TD_EXITED);
1190         return (void *) (uintptr_t) td->error;
1191 }
1192
1193
1194 /*
1195  * We cannot pass the td data into a forked process, so attach the td and
1196  * pass it to the thread worker.
1197  */
1198 static int fork_main(int shmid, int offset)
1199 {
1200         struct thread_data *td;
1201         void *data, *ret;
1202
1203 #ifndef __hpux
1204         data = shmat(shmid, NULL, 0);
1205         if (data == (void *) -1) {
1206                 int __err = errno;
1207
1208                 perror("shmat");
1209                 return __err;
1210         }
1211 #else
1212         /*
1213          * HP-UX inherits shm mappings?
1214          */
1215         data = threads;
1216 #endif
1217
1218         td = data + offset * sizeof(struct thread_data);
1219         ret = thread_main(td);
1220         shmdt(data);
1221         return (int) (uintptr_t) ret;
1222 }
1223
1224 /*
1225  * Run over the job map and reap the threads that have exited, if any.
1226  */
1227 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1228                          unsigned int *m_rate)
1229 {
1230         struct thread_data *td;
1231         unsigned int cputhreads, realthreads, pending;
1232         int i, status, ret;
1233
1234         /*
1235          * reap exited threads (TD_EXITED -> TD_REAPED)
1236          */
1237         realthreads = pending = cputhreads = 0;
1238         for_each_td(td, i) {
1239                 int flags = 0;
1240
1241                 /*
1242                  * ->io_ops is NULL for a thread that has closed its
1243                  * io engine
1244                  */
1245                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1246                         cputhreads++;
1247                 else
1248                         realthreads++;
1249
1250                 if (!td->pid) {
1251                         pending++;
1252                         continue;
1253                 }
1254                 if (td->runstate == TD_REAPED)
1255                         continue;
1256                 if (td->o.use_thread) {
1257                         if (td->runstate == TD_EXITED) {
1258                                 td_set_runstate(td, TD_REAPED);
1259                                 goto reaped;
1260                         }
1261                         continue;
1262                 }
1263
1264                 flags = WNOHANG;
1265                 if (td->runstate == TD_EXITED)
1266                         flags = 0;
1267
1268                 /*
1269                  * check if someone quit or got killed in an unusual way
1270                  */
1271                 ret = waitpid(td->pid, &status, flags);
1272                 if (ret < 0) {
1273                         if (errno == ECHILD) {
1274                                 log_err("fio: pid=%d disappeared %d\n",
1275                                                 (int) td->pid, td->runstate);
1276                                 td_set_runstate(td, TD_REAPED);
1277                                 goto reaped;
1278                         }
1279                         perror("waitpid");
1280                 } else if (ret == td->pid) {
1281                         if (WIFSIGNALED(status)) {
1282                                 int sig = WTERMSIG(status);
1283
1284                                 if (sig != SIGTERM)
1285                                         log_err("fio: pid=%d, got signal=%d\n",
1286                                                         (int) td->pid, sig);
1287                                 td_set_runstate(td, TD_REAPED);
1288                                 goto reaped;
1289                         }
1290                         if (WIFEXITED(status)) {
1291                                 if (WEXITSTATUS(status) && !td->error)
1292                                         td->error = WEXITSTATUS(status);
1293
1294                                 td_set_runstate(td, TD_REAPED);
1295                                 goto reaped;
1296                         }
1297                 }
1298
1299                 /*
1300                  * thread is not dead, continue
1301                  */
1302                 pending++;
1303                 continue;
1304 reaped:
1305                 (*nr_running)--;
1306                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1307                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1308                 if (!td->pid)
1309                         pending--;
1310
1311                 if (td->error)
1312                         exit_value++;
1313
1314                 done_secs += mtime_since_now(&td->epoch) / 1000;
1315         }
1316
1317         if (*nr_running == cputhreads && !pending && realthreads)
1318                 fio_terminate_threads(TERMINATE_ALL);
1319 }
1320
1321 /*
1322  * Main function for kicking off and reaping jobs, as needed.
1323  */
1324 static void run_threads(void)
1325 {
1326         struct thread_data *td;
1327         unsigned long spent;
1328         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1329
1330         if (fio_pin_memory())
1331                 return;
1332
1333         if (fio_gtod_offload && fio_start_gtod_thread())
1334                 return;
1335
1336         set_sig_handlers();
1337
1338         if (!terse_output) {
1339                 log_info("Starting ");
1340                 if (nr_thread)
1341                         log_info("%d thread%s", nr_thread,
1342                                                 nr_thread > 1 ? "s" : "");
1343                 if (nr_process) {
1344                         if (nr_thread)
1345                                 log_info(" and ");
1346                         log_info("%d process%s", nr_process,
1347                                                 nr_process > 1 ? "es" : "");
1348                 }
1349                 log_info("\n");
1350                 fflush(stdout);
1351         }
1352
1353         todo = thread_number;
1354         nr_running = 0;
1355         nr_started = 0;
1356         m_rate = t_rate = 0;
1357
1358         for_each_td(td, i) {
1359                 print_status_init(td->thread_number - 1);
1360
1361                 if (!td->o.create_serialize)
1362                         continue;
1363
1364                 /*
1365                  * do file setup here so it happens sequentially,
1366                  * we don't want X number of threads getting their
1367                  * client data interspersed on disk
1368                  */
1369                 if (setup_files(td)) {
1370                         exit_value++;
1371                         if (td->error)
1372                                 log_err("fio: pid=%d, err=%d/%s\n",
1373                                         (int) td->pid, td->error, td->verror);
1374                         td_set_runstate(td, TD_REAPED);
1375                         todo--;
1376                 } else {
1377                         struct fio_file *f;
1378                         unsigned int j;
1379
1380                         /*
1381                          * for sharing to work, each job must always open
1382                          * its own files. so close them, if we opened them
1383                          * for creation
1384                          */
1385                         for_each_file(td, f, j) {
1386                                 if (fio_file_open(f))
1387                                         td_io_close_file(td, f);
1388                         }
1389                 }
1390         }
1391
1392         set_genesis_time();
1393
1394         while (todo) {
1395                 struct thread_data *map[REAL_MAX_JOBS];
1396                 struct timeval this_start;
1397                 int this_jobs = 0, left;
1398
1399                 /*
1400                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1401                  */
1402                 for_each_td(td, i) {
1403                         if (td->runstate != TD_NOT_CREATED)
1404                                 continue;
1405
1406                         /*
1407                          * never got a chance to start, killed by other
1408                          * thread for some reason
1409                          */
1410                         if (td->terminate) {
1411                                 todo--;
1412                                 continue;
1413                         }
1414
1415                         if (td->o.start_delay) {
1416                                 spent = mtime_since_genesis();
1417
1418                                 if (td->o.start_delay * 1000 > spent)
1419                                         continue;
1420                         }
1421
1422                         if (td->o.stonewall && (nr_started || nr_running)) {
1423                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1424                                                         td->o.name);
1425                                 break;
1426                         }
1427
1428                         init_disk_util(td);
1429
1430                         /*
1431                          * Set state to created. Thread will transition
1432                          * to TD_INITIALIZED when it's done setting up.
1433                          */
1434                         td_set_runstate(td, TD_CREATED);
1435                         map[this_jobs++] = td;
1436                         nr_started++;
1437
1438                         if (td->o.use_thread) {
1439                                 int ret;
1440
1441                                 dprint(FD_PROCESS, "will pthread_create\n");
1442                                 ret = pthread_create(&td->thread, NULL,
1443                                                         thread_main, td);
1444                                 if (ret) {
1445                                         log_err("pthread_create: %s\n",
1446                                                         strerror(ret));
1447                                         nr_started--;
1448                                         break;
1449                                 }
1450                                 ret = pthread_detach(td->thread);
1451                                 if (ret)
1452                                         log_err("pthread_detach: %s",
1453                                                         strerror(ret));
1454                         } else {
1455                                 pid_t pid;
1456                                 dprint(FD_PROCESS, "will fork\n");
1457                                 pid = fork();
1458                                 if (!pid) {
1459                                         int ret = fork_main(shm_id, i);
1460
1461                                         _exit(ret);
1462                                 } else if (i == fio_debug_jobno)
1463                                         *fio_debug_jobp = pid;
1464                         }
1465                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1466                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1467                                 log_err("fio: job startup hung? exiting.\n");
1468                                 fio_terminate_threads(TERMINATE_ALL);
1469                                 fio_abort = 1;
1470                                 nr_started--;
1471                                 break;
1472                         }
1473                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1474                 }
1475
1476                 /*
1477                  * Wait for the started threads to transition to
1478                  * TD_INITIALIZED.
1479                  */
1480                 fio_gettime(&this_start, NULL);
1481                 left = this_jobs;
1482                 while (left && !fio_abort) {
1483                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1484                                 break;
1485
1486                         usleep(100000);
1487
1488                         for (i = 0; i < this_jobs; i++) {
1489                                 td = map[i];
1490                                 if (!td)
1491                                         continue;
1492                                 if (td->runstate == TD_INITIALIZED) {
1493                                         map[i] = NULL;
1494                                         left--;
1495                                 } else if (td->runstate >= TD_EXITED) {
1496                                         map[i] = NULL;
1497                                         left--;
1498                                         todo--;
1499                                         nr_running++; /* work-around... */
1500                                 }
1501                         }
1502                 }
1503
1504                 if (left) {
1505                         log_err("fio: %d job%s failed to start\n", left,
1506                                         left > 1 ? "s" : "");
1507                         for (i = 0; i < this_jobs; i++) {
1508                                 td = map[i];
1509                                 if (!td)
1510                                         continue;
1511                                 kill(td->pid, SIGTERM);
1512                         }
1513                         break;
1514                 }
1515
1516                 /*
1517                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1518                  */
1519                 for_each_td(td, i) {
1520                         if (td->runstate != TD_INITIALIZED)
1521                                 continue;
1522
1523                         if (in_ramp_time(td))
1524                                 td_set_runstate(td, TD_RAMP);
1525                         else
1526                                 td_set_runstate(td, TD_RUNNING);
1527                         nr_running++;
1528                         nr_started--;
1529                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1530                         t_rate += td->o.rate[0] + td->o.rate[1];
1531                         todo--;
1532                         fio_mutex_up(td->mutex);
1533                 }
1534
1535                 reap_threads(&nr_running, &t_rate, &m_rate);
1536
1537                 if (todo) {
1538                         if (is_backend)
1539                                 fio_server_idle_loop();
1540                         else
1541                                 usleep(100000);
1542                 }
1543         }
1544
1545         while (nr_running) {
1546                 reap_threads(&nr_running, &t_rate, &m_rate);
1547
1548                 if (is_backend)
1549                         fio_server_idle_loop();
1550                 else
1551                         usleep(10000);
1552         }
1553
1554         update_io_ticks();
1555         fio_unpin_memory();
1556 }
1557
1558 static void *disk_thread_main(void *data)
1559 {
1560         fio_mutex_up(startup_mutex);
1561
1562         while (threads) {
1563                 usleep(DISK_UTIL_MSEC * 1000);
1564                 if (!threads)
1565                         break;
1566                 update_io_ticks();
1567
1568                 if (!is_backend)
1569                         print_thread_status();
1570         }
1571
1572         return NULL;
1573 }
1574
1575 static int create_disk_util_thread(void)
1576 {
1577         int ret;
1578
1579         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1580         if (ret) {
1581                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1582                 return 1;
1583         }
1584
1585         ret = pthread_detach(disk_util_thread);
1586         if (ret) {
1587                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1588                 return 1;
1589         }
1590
1591         dprint(FD_MUTEX, "wait on startup_mutex\n");
1592         fio_mutex_down(startup_mutex);
1593         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1594         return 0;
1595 }
1596
1597 int fio_backend(void)
1598 {
1599         struct thread_data *td;
1600         int i;
1601
1602         if (exec_profile) {
1603                 if (load_profile(exec_profile))
1604                         return 1;
1605                 free(exec_profile);
1606                 exec_profile = NULL;
1607         }
1608         if (!thread_number)
1609                 return 0;
1610
1611         if (write_bw_log) {
1612                 setup_log(&agg_io_log[DDIR_READ], 0);
1613                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1614         }
1615
1616         startup_mutex = fio_mutex_init(0);
1617         if (startup_mutex == NULL)
1618                 return 1;
1619         writeout_mutex = fio_mutex_init(1);
1620         if (writeout_mutex == NULL)
1621                 return 1;
1622
1623         set_genesis_time();
1624         create_disk_util_thread();
1625
1626         cgroup_list = smalloc(sizeof(*cgroup_list));
1627         INIT_FLIST_HEAD(cgroup_list);
1628
1629         run_threads();
1630
1631         if (!fio_abort) {
1632                 show_run_stats();
1633                 if (write_bw_log) {
1634                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1635                         __finish_log(agg_io_log[DDIR_WRITE],
1636                                         "agg-write_bw.log");
1637                 }
1638         }
1639
1640         for_each_td(td, i)
1641                 fio_options_free(td);
1642
1643         cgroup_kill(cgroup_list);
1644         sfree(cgroup_list);
1645         sfree(cgroup_mnt);
1646
1647         fio_mutex_remove(startup_mutex);
1648         fio_mutex_remove(writeout_mutex);
1649         return exit_value;
1650 }