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 *disk_thread_mutex;
  54 static struct fio_mutex *startup_mutex;
  55 static struct fio_mutex *writeout_mutex;
  56 static struct flist_head *cgroup_list;
  57 static char *cgroup_mnt;
  58 static int exit_value;
  59 static volatile int fio_abort;
  60
  61 struct io_log *agg_io_log[2];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int nr_process = 0;
  66 unsigned int nr_thread = 0;
  67 int shm_id = 0;
  68 int temp_stall_ts;
  69 unsigned long done_secs = 0;
  70
  71 #define PAGE_ALIGN(buf) \
  72         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  73
  74 #define JOB_START_TIMEOUT       (5 * 1000)
  75
  76 static void sig_int(int sig)
  77 {
  78         if (threads) {
  79                 if (is_backend)
  80                         fio_server_got_signal(sig);
  81                 else {
  82                         log_info("\nfio: terminating on signal %d\n", sig);
  83                         fflush(stdout);
  84                         exit_value = 128;
  85                 }
  86
  87                 fio_terminate_threads(TERMINATE_ALL);
  88         }
  89 }
  90
  91 static void sig_show_status(int sig)
  92 {
  93         show_running_run_stats();
  94 }
  95
  96 static void set_sig_handlers(void)
  97 {
  98         struct sigaction act;
  99
 100         memset(&act, 0, sizeof(act));
 101         act.sa_handler = sig_int;
 102         act.sa_flags = SA_RESTART;
 103         sigaction(SIGINT, &act, NULL);
 104
 105         memset(&act, 0, sizeof(act));
 106         act.sa_handler = sig_int;
 107         act.sa_flags = SA_RESTART;
 108         sigaction(SIGTERM, &act, NULL);
 109
 110         memset(&act, 0, sizeof(act));
 111         act.sa_handler = sig_show_status;
 112         act.sa_flags = SA_RESTART;
 113         sigaction(SIGUSR1, &act, NULL);
 114
 115         if (is_backend) {
 116                 memset(&act, 0, sizeof(act));
 117                 act.sa_handler = sig_int;
 118                 act.sa_flags = SA_RESTART;
 119                 sigaction(SIGPIPE, &act, NULL);
 120         }
 121 }
 122
 123 /*
 124  * Check if we are above the minimum rate given.
 125  */
 126 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 127                             enum fio_ddir ddir)
 128 {
 129         unsigned long long bytes = 0;
 130         unsigned long iops = 0;
 131         unsigned long spent;
 132         unsigned long rate;
 133         unsigned int ratemin = 0;
 134         unsigned int rate_iops = 0;
 135         unsigned int rate_iops_min = 0;
 136
 137         assert(ddir_rw(ddir));
 138
 139         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 140                 return 0;
 141
 142         /*
 143          * allow a 2 second settle period in the beginning
 144          */
 145         if (mtime_since(&td->start, now) < 2000)
 146                 return 0;
 147
 148         iops += td->this_io_blocks[ddir];
 149         bytes += td->this_io_bytes[ddir];
 150         ratemin += td->o.ratemin[ddir];
 151         rate_iops += td->o.rate_iops[ddir];
 152         rate_iops_min += td->o.rate_iops_min[ddir];
 153
 154         /*
 155          * if rate blocks is set, sample is running
 156          */
 157         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 158                 spent = mtime_since(&td->lastrate[ddir], now);
 159                 if (spent < td->o.ratecycle)
 160                         return 0;
 161
 162                 if (td->o.rate[ddir]) {
 163                         /*
 164                          * check bandwidth specified rate
 165                          */
 166                         if (bytes < td->rate_bytes[ddir]) {
 167                                 log_err("%s: min rate %u not met\n", td->o.name,
 168                                                                 ratemin);
 169                                 return 1;
 170                         } else {
 171                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 172                                 if (rate < ratemin ||
 173                                     bytes < td->rate_bytes[ddir]) {
 174                                         log_err("%s: min rate %u not met, got"
 175                                                 " %luKB/sec\n", td->o.name,
 176                                                         ratemin, rate);
 177                                         return 1;
 178                                 }
 179                         }
 180                 } else {
 181                         /*
 182                          * checks iops specified rate
 183                          */
 184                         if (iops < rate_iops) {
 185                                 log_err("%s: min iops rate %u not met\n",
 186                                                 td->o.name, rate_iops);
 187                                 return 1;
 188                         } else {
 189                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 190                                 if (rate < rate_iops_min ||
 191                                     iops < td->rate_blocks[ddir]) {
 192                                         log_err("%s: min iops rate %u not met,"
 193                                                 " got %lu\n", td->o.name,
 194                                                         rate_iops_min, rate);
 195                                 }
 196                         }
 197                 }
 198         }
 199
 200         td->rate_bytes[ddir] = bytes;
 201         td->rate_blocks[ddir] = iops;
 202         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 203         return 0;
 204 }
 205
 206 static int check_min_rate(struct thread_data *td, struct timeval *now,
 207                           unsigned long *bytes_done)
 208 {
 209         int ret = 0;
 210
 211         if (bytes_done[0])
 212                 ret |= __check_min_rate(td, now, 0);
 213         if (bytes_done[1])
 214                 ret |= __check_min_rate(td, now, 1);
 215
 216         return ret;
 217 }
 218
 219 /*
 220  * When job exits, we can cancel the in-flight IO if we are using async
 221  * io. Attempt to do so.
 222  */
 223 static void cleanup_pending_aio(struct thread_data *td)
 224 {
 225         struct flist_head *entry, *n;
 226         struct io_u *io_u;
 227         int r;
 228
 229         /*
 230          * get immediately available events, if any
 231          */
 232         r = io_u_queued_complete(td, 0, NULL);
 233         if (r < 0)
 234                 return;
 235
 236         /*
 237          * now cancel remaining active events
 238          */
 239         if (td->io_ops->cancel) {
 240                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 241                         io_u = flist_entry(entry, struct io_u, list);
 242
 243                         /*
 244                          * if the io_u isn't in flight, then that generally
 245                          * means someone leaked an io_u. complain but fix
 246                          * it up, so we don't stall here.
 247                          */
 248                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 249                                 log_err("fio: non-busy IO on busy list\n");
 250                                 put_io_u(td, io_u);
 251                         } else {
 252                                 r = td->io_ops->cancel(td, io_u);
 253                                 if (!r)
 254                                         put_io_u(td, io_u);
 255                         }
 256                 }
 257         }
 258
 259         if (td->cur_depth)
 260                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 261 }
 262
 263 /*
 264  * Helper to handle the final sync of a file. Works just like the normal
 265  * io path, just does everything sync.
 266  */
 267 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 268 {
 269         struct io_u *io_u = __get_io_u(td);
 270         int ret;
 271
 272         if (!io_u)
 273                 return 1;
 274
 275         io_u->ddir = DDIR_SYNC;
 276         io_u->file = f;
 277
 278         if (td_io_prep(td, io_u)) {
 279                 put_io_u(td, io_u);
 280                 return 1;
 281         }
 282
 283 requeue:
 284         ret = td_io_queue(td, io_u);
 285         if (ret < 0) {
 286                 td_verror(td, io_u->error, "td_io_queue");
 287                 put_io_u(td, io_u);
 288                 return 1;
 289         } else if (ret == FIO_Q_QUEUED) {
 290                 if (io_u_queued_complete(td, 1, NULL) < 0)
 291                         return 1;
 292         } else if (ret == FIO_Q_COMPLETED) {
 293                 if (io_u->error) {
 294                         td_verror(td, io_u->error, "td_io_queue");
 295                         return 1;
 296                 }
 297
 298                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 299                         return 1;
 300         } else if (ret == FIO_Q_BUSY) {
 301                 if (td_io_commit(td))
 302                         return 1;
 303                 goto requeue;
 304         }
 305
 306         return 0;
 307 }
 308
 309 static inline void __update_tv_cache(struct thread_data *td)
 310 {
 311         fio_gettime(&td->tv_cache, NULL);
 312 }
 313
 314 static inline void update_tv_cache(struct thread_data *td)
 315 {
 316         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 317                 __update_tv_cache(td);
 318 }
 319
 320 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 321 {
 322         if (in_ramp_time(td))
 323                 return 0;
 324         if (!td->o.timeout)
 325                 return 0;
 326         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 327                 return 1;
 328
 329         return 0;
 330 }
 331
 332 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 333                                int *retptr)
 334 {
 335         int ret = *retptr;
 336
 337         if (ret < 0 || td->error) {
 338                 int err;
 339
 340                 if (ret < 0)
 341                         err = -ret;
 342                 else
 343                         err = td->error;
 344
 345                 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
 346                         return 1;
 347
 348                 if (td_non_fatal_error(err)) {
 349                         /*
 350                          * Continue with the I/Os in case of
 351                          * a non fatal error.
 352                          */
 353                         update_error_count(td, err);
 354                         td_clear_error(td);
 355                         *retptr = 0;
 356                         return 0;
 357                 } else if (td->o.fill_device && err == ENOSPC) {
 358                         /*
 359                          * We expect to hit this error if
 360                          * fill_device option is set.
 361                          */
 362                         td_clear_error(td);
 363                         td->terminate = 1;
 364                         return 1;
 365                 } else {
 366                         /*
 367                          * Stop the I/O in case of a fatal
 368                          * error.
 369                          */
 370                         update_error_count(td, err);
 371                         return 1;
 372                 }
 373         }
 374
 375         return 0;
 376 }
 377
 378 /*
 379  * The main verify engine. Runs over the writes we previously submitted,
 380  * reads the blocks back in, and checks the crc/md5 of the data.
 381  */
 382 static void do_verify(struct thread_data *td)
 383 {
 384         struct fio_file *f;
 385         struct io_u *io_u;
 386         int ret, min_events;
 387         unsigned int i;
 388
 389         dprint(FD_VERIFY, "starting loop\n");
 390
 391         /*
 392          * sync io first and invalidate cache, to make sure we really
 393          * read from disk.
 394          */
 395         for_each_file(td, f, i) {
 396                 if (!fio_file_open(f))
 397                         continue;
 398                 if (fio_io_sync(td, f))
 399                         break;
 400                 if (file_invalidate_cache(td, f))
 401                         break;
 402         }
 403
 404         if (td->error)
 405                 return;
 406
 407         td_set_runstate(td, TD_VERIFYING);
 408
 409         io_u = NULL;
 410         while (!td->terminate) {
 411                 int ret2, full;
 412
 413                 update_tv_cache(td);
 414
 415                 if (runtime_exceeded(td, &td->tv_cache)) {
 416                         __update_tv_cache(td);
 417                         if (runtime_exceeded(td, &td->tv_cache)) {
 418                                 td->terminate = 1;
 419                                 break;
 420                         }
 421                 }
 422
 423                 if (flow_threshold_exceeded(td))
 424                         continue;
 425
 426                 io_u = __get_io_u(td);
 427                 if (!io_u)
 428                         break;
 429
 430                 if (get_next_verify(td, io_u)) {
 431                         put_io_u(td, io_u);
 432                         break;
 433                 }
 434
 435                 if (td_io_prep(td, io_u)) {
 436                         put_io_u(td, io_u);
 437                         break;
 438                 }
 439
 440                 if (td->o.verify_async)
 441                         io_u->end_io = verify_io_u_async;
 442                 else
 443                         io_u->end_io = verify_io_u;
 444
 445                 ret = td_io_queue(td, io_u);
 446                 switch (ret) {
 447                 case FIO_Q_COMPLETED:
 448                         if (io_u->error) {
 449                                 ret = -io_u->error;
 450                                 clear_io_u(td, io_u);
 451                         } else if (io_u->resid) {
 452                                 int bytes = io_u->xfer_buflen - io_u->resid;
 453
 454                                 /*
 455                                  * zero read, fail
 456                                  */
 457                                 if (!bytes) {
 458                                         td_verror(td, EIO, "full resid");
 459                                         put_io_u(td, io_u);
 460                                         break;
 461                                 }
 462
 463                                 io_u->xfer_buflen = io_u->resid;
 464                                 io_u->xfer_buf += bytes;
 465                                 io_u->offset += bytes;
 466
 467                                 if (ddir_rw(io_u->ddir))
 468                                         td->ts.short_io_u[io_u->ddir]++;
 469
 470                                 f = io_u->file;
 471                                 if (io_u->offset == f->real_file_size)
 472                                         goto sync_done;
 473
 474                                 requeue_io_u(td, &io_u);
 475                         } else {
 476 sync_done:
 477                                 ret = io_u_sync_complete(td, io_u, NULL);
 478                                 if (ret < 0)
 479                                         break;
 480                         }
 481                         continue;
 482                 case FIO_Q_QUEUED:
 483                         break;
 484                 case FIO_Q_BUSY:
 485                         requeue_io_u(td, &io_u);
 486                         ret2 = td_io_commit(td);
 487                         if (ret2 < 0)
 488                                 ret = ret2;
 489                         break;
 490                 default:
 491                         assert(ret < 0);
 492                         td_verror(td, -ret, "td_io_queue");
 493                         break;
 494                 }
 495
 496                 if (break_on_this_error(td, io_u->ddir, &ret))
 497                         break;
 498
 499                 /*
 500                  * if we can queue more, do so. but check if there are
 501                  * completed io_u's first. Note that we can get BUSY even
 502                  * without IO queued, if the system is resource starved.
 503                  */
 504                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 505                 if (full || !td->o.iodepth_batch_complete) {
 506                         min_events = min(td->o.iodepth_batch_complete,
 507                                          td->cur_depth);
 508                         /*
 509                          * if the queue is full, we MUST reap at least 1 event
 510                          */
 511                         if (full && !min_events)
 512                                 min_events = 1;
 513
 514                         do {
 515                                 /*
 516                                  * Reap required number of io units, if any,
 517                                  * and do the verification on them through
 518                                  * the callback handler
 519                                  */
 520                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 521                                         ret = -1;
 522                                         break;
 523                                 }
 524                         } while (full && (td->cur_depth > td->o.iodepth_low));
 525                 }
 526                 if (ret < 0)
 527                         break;
 528         }
 529
 530         if (!td->error) {
 531                 min_events = td->cur_depth;
 532
 533                 if (min_events)
 534                         ret = io_u_queued_complete(td, min_events, NULL);
 535         } else
 536                 cleanup_pending_aio(td);
 537
 538         td_set_runstate(td, TD_RUNNING);
 539
 540         dprint(FD_VERIFY, "exiting loop\n");
 541 }
 542
 543 static int io_bytes_exceeded(struct thread_data *td)
 544 {
 545         unsigned long long bytes;
 546
 547         if (td_rw(td))
 548                 bytes = td->this_io_bytes[0] + td->this_io_bytes[1];
 549         else if (td_write(td))
 550                 bytes = td->this_io_bytes[1];
 551         else
 552                 bytes = td->this_io_bytes[0];
 553
 554         return bytes >= td->o.size;
 555 }
 556
 557 /*
 558  * Main IO worker function. It retrieves io_u's to process and queues
 559  * and reaps them, checking for rate and errors along the way.
 560  */
 561 static void do_io(struct thread_data *td)
 562 {
 563         unsigned int i;
 564         int ret = 0;
 565
 566         if (in_ramp_time(td))
 567                 td_set_runstate(td, TD_RAMP);
 568         else
 569                 td_set_runstate(td, TD_RUNNING);
 570
 571         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 572                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 573                 td->o.time_based) {
 574                 struct timeval comp_time;
 575                 unsigned long bytes_done[2] = { 0, 0 };
 576                 int min_evts = 0;
 577                 struct io_u *io_u;
 578                 int ret2, full;
 579                 enum fio_ddir ddir;
 580
 581                 if (td->terminate)
 582                         break;
 583
 584                 update_tv_cache(td);
 585
 586                 if (runtime_exceeded(td, &td->tv_cache)) {
 587                         __update_tv_cache(td);
 588                         if (runtime_exceeded(td, &td->tv_cache)) {
 589                                 td->terminate = 1;
 590                                 break;
 591                         }
 592                 }
 593
 594                 if (flow_threshold_exceeded(td))
 595                         continue;
 596
 597                 io_u = get_io_u(td);
 598                 if (!io_u)
 599                         break;
 600
 601                 ddir = io_u->ddir;
 602
 603                 /*
 604                  * Add verification end_io handler if:
 605                  *      - Asked to verify (!td_rw(td))
 606                  *      - Or the io_u is from our verify list (mixed write/ver)
 607                  */
 608                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 609                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 610                         if (td->o.verify_async)
 611                                 io_u->end_io = verify_io_u_async;
 612                         else
 613                                 io_u->end_io = verify_io_u;
 614                         td_set_runstate(td, TD_VERIFYING);
 615                 } else if (in_ramp_time(td))
 616                         td_set_runstate(td, TD_RAMP);
 617                 else
 618                         td_set_runstate(td, TD_RUNNING);
 619
 620                 ret = td_io_queue(td, io_u);
 621                 switch (ret) {
 622                 case FIO_Q_COMPLETED:
 623                         if (io_u->error) {
 624                                 ret = -io_u->error;
 625                                 clear_io_u(td, io_u);
 626                         } else if (io_u->resid) {
 627                                 int bytes = io_u->xfer_buflen - io_u->resid;
 628                                 struct fio_file *f = io_u->file;
 629
 630                                 /*
 631                                  * zero read, fail
 632                                  */
 633                                 if (!bytes) {
 634                                         td_verror(td, EIO, "full resid");
 635                                         put_io_u(td, io_u);
 636                                         break;
 637                                 }
 638
 639                                 io_u->xfer_buflen = io_u->resid;
 640                                 io_u->xfer_buf += bytes;
 641                                 io_u->offset += bytes;
 642
 643                                 if (ddir_rw(io_u->ddir))
 644                                         td->ts.short_io_u[io_u->ddir]++;
 645
 646                                 if (io_u->offset == f->real_file_size)
 647                                         goto sync_done;
 648
 649                                 requeue_io_u(td, &io_u);
 650                         } else {
 651 sync_done:
 652                                 if (__should_check_rate(td, 0) ||
 653                                     __should_check_rate(td, 1))
 654                                         fio_gettime(&comp_time, NULL);
 655
 656                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 657                                 if (ret < 0)
 658                                         break;
 659                         }
 660                         break;
 661                 case FIO_Q_QUEUED:
 662                         /*
 663                          * if the engine doesn't have a commit hook,
 664                          * the io_u is really queued. if it does have such
 665                          * a hook, it has to call io_u_queued() itself.
 666                          */
 667                         if (td->io_ops->commit == NULL)
 668                                 io_u_queued(td, io_u);
 669                         break;
 670                 case FIO_Q_BUSY:
 671                         requeue_io_u(td, &io_u);
 672                         ret2 = td_io_commit(td);
 673                         if (ret2 < 0)
 674                                 ret = ret2;
 675                         break;
 676                 default:
 677                         assert(ret < 0);
 678                         put_io_u(td, io_u);
 679                         break;
 680                 }
 681
 682                 if (break_on_this_error(td, ddir, &ret))
 683                         break;
 684
 685                 /*
 686                  * See if we need to complete some commands. Note that we
 687                  * can get BUSY even without IO queued, if the system is
 688                  * resource starved.
 689                  */
 690                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 691                 if (full || !td->o.iodepth_batch_complete) {
 692                         min_evts = min(td->o.iodepth_batch_complete,
 693                                         td->cur_depth);
 694                         /*
 695                          * if the queue is full, we MUST reap at least 1 event
 696                          */
 697                         if (full && !min_evts)
 698                                 min_evts = 1;
 699
 700                         if (__should_check_rate(td, 0) ||
 701                             __should_check_rate(td, 1))
 702                                 fio_gettime(&comp_time, NULL);
 703
 704                         do {
 705                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 706                                 if (ret < 0)
 707                                         break;
 708
 709                         } while (full && (td->cur_depth > td->o.iodepth_low));
 710                 }
 711
 712                 if (ret < 0)
 713                         break;
 714                 if (!(bytes_done[0] + bytes_done[1]))
 715                         continue;
 716
 717                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 718                         if (check_min_rate(td, &comp_time, bytes_done)) {
 719                                 if (exitall_on_terminate)
 720                                         fio_terminate_threads(td->groupid);
 721                                 td_verror(td, EIO, "check_min_rate");
 722                                 break;
 723                         }
 724                 }
 725
 726                 if (td->o.thinktime) {
 727                         unsigned long long b;
 728
 729                         b = td->io_blocks[0] + td->io_blocks[1];
 730                         if (!(b % td->o.thinktime_blocks)) {
 731                                 int left;
 732
 733                                 if (td->o.thinktime_spin)
 734                                         usec_spin(td->o.thinktime_spin);
 735
 736                                 left = td->o.thinktime - td->o.thinktime_spin;
 737                                 if (left)
 738                                         usec_sleep(td, left);
 739                         }
 740                 }
 741         }
 742
 743         if (td->trim_entries)
 744                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 745
 746         if (td->o.fill_device && td->error == ENOSPC) {
 747                 td->error = 0;
 748                 td->terminate = 1;
 749         }
 750         if (!td->error) {
 751                 struct fio_file *f;
 752
 753                 i = td->cur_depth;
 754                 if (i) {
 755                         ret = io_u_queued_complete(td, i, NULL);
 756                         if (td->o.fill_device && td->error == ENOSPC)
 757                                 td->error = 0;
 758                 }
 759
 760                 if (should_fsync(td) && td->o.end_fsync) {
 761                         td_set_runstate(td, TD_FSYNCING);
 762
 763                         for_each_file(td, f, i) {
 764                                 if (!fio_file_open(f))
 765                                         continue;
 766                                 fio_io_sync(td, f);
 767                         }
 768                 }
 769         } else
 770                 cleanup_pending_aio(td);
 771
 772         /*
 773          * stop job if we failed doing any IO
 774          */
 775         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 776                 td->done = 1;
 777 }
 778
 779 static void cleanup_io_u(struct thread_data *td)
 780 {
 781         struct flist_head *entry, *n;
 782         struct io_u *io_u;
 783
 784         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 785                 io_u = flist_entry(entry, struct io_u, list);
 786
 787                 flist_del(&io_u->list);
 788                 fio_memfree(io_u, sizeof(*io_u));
 789         }
 790
 791         free_io_mem(td);
 792 }
 793
 794 static int init_io_u(struct thread_data *td)
 795 {
 796         struct io_u *io_u;
 797         unsigned int max_bs, min_write;
 798         int cl_align, i, max_units;
 799         char *p;
 800
 801         max_units = td->o.iodepth;
 802         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 803         min_write = td->o.min_bs[DDIR_WRITE];
 804         td->orig_buffer_size = (unsigned long long) max_bs
 805                                         * (unsigned long long) max_units;
 806
 807         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 808                 unsigned long bs;
 809
 810                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 811                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 812         }
 813
 814         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 815                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 816                 return 1;
 817         }
 818
 819         if (allocate_io_mem(td))
 820                 return 1;
 821
 822         if (td->o.odirect || td->o.mem_align ||
 823             (td->io_ops->flags & FIO_RAWIO))
 824                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 825         else
 826                 p = td->orig_buffer;
 827
 828         cl_align = os_cache_line_size();
 829
 830         for (i = 0; i < max_units; i++) {
 831                 void *ptr;
 832
 833                 if (td->terminate)
 834                         return 1;
 835
 836                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 837                 if (!ptr) {
 838                         log_err("fio: unable to allocate aligned memory\n");
 839                         break;
 840                 }
 841
 842                 io_u = ptr;
 843                 memset(io_u, 0, sizeof(*io_u));
 844                 INIT_FLIST_HEAD(&io_u->list);
 845                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 846
 847                 if (!(td->io_ops->flags & FIO_NOIO)) {
 848                         io_u->buf = p;
 849                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 850
 851                         if (td_write(td))
 852                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 853                         if (td_write(td) && td->o.verify_pattern_bytes) {
 854                                 /*
 855                                  * Fill the buffer with the pattern if we are
 856                                  * going to be doing writes.
 857                                  */
 858                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 859                         }
 860                 }
 861
 862                 io_u->index = i;
 863                 io_u->flags = IO_U_F_FREE;
 864                 flist_add(&io_u->list, &td->io_u_freelist);
 865                 p += max_bs;
 866         }
 867
 868         return 0;
 869 }
 870
 871 static int switch_ioscheduler(struct thread_data *td)
 872 {
 873         char tmp[256], tmp2[128];
 874         FILE *f;
 875         int ret;
 876
 877         if (td->io_ops->flags & FIO_DISKLESSIO)
 878                 return 0;
 879
 880         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 881
 882         f = fopen(tmp, "r+");
 883         if (!f) {
 884                 if (errno == ENOENT) {
 885                         log_err("fio: os or kernel doesn't support IO scheduler"
 886                                 " switching\n");
 887                         return 0;
 888                 }
 889                 td_verror(td, errno, "fopen iosched");
 890                 return 1;
 891         }
 892
 893         /*
 894          * Set io scheduler.
 895          */
 896         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 897         if (ferror(f) || ret != 1) {
 898                 td_verror(td, errno, "fwrite");
 899                 fclose(f);
 900                 return 1;
 901         }
 902
 903         rewind(f);
 904
 905         /*
 906          * Read back and check that the selected scheduler is now the default.
 907          */
 908         ret = fread(tmp, 1, sizeof(tmp), f);
 909         if (ferror(f) || ret < 0) {
 910                 td_verror(td, errno, "fread");
 911                 fclose(f);
 912                 return 1;
 913         }
 914
 915         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 916         if (!strstr(tmp, tmp2)) {
 917                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 918                 td_verror(td, EINVAL, "iosched_switch");
 919                 fclose(f);
 920                 return 1;
 921         }
 922
 923         fclose(f);
 924         return 0;
 925 }
 926
 927 static int keep_running(struct thread_data *td)
 928 {
 929         unsigned long long io_done;
 930
 931         if (td->done)
 932                 return 0;
 933         if (td->o.time_based)
 934                 return 1;
 935         if (td->o.loops) {
 936                 td->o.loops--;
 937                 return 1;
 938         }
 939
 940         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 941                         + td->io_skip_bytes;
 942         if (io_done < td->o.size)
 943                 return 1;
 944
 945         return 0;
 946 }
 947
 948 static int exec_string(const char *string)
 949 {
 950         int ret, newlen = strlen(string) + 1 + 8;
 951         char *str;
 952
 953         str = malloc(newlen);
 954         sprintf(str, "sh -c %s", string);
 955
 956         ret = system(str);
 957         if (ret == -1)
 958                 log_err("fio: exec of cmd <%s> failed\n", str);
 959
 960         free(str);
 961         return ret;
 962 }
 963
 964 /*
 965  * Entry point for the thread based jobs. The process based jobs end up
 966  * here as well, after a little setup.
 967  */
 968 static void *thread_main(void *data)
 969 {
 970         unsigned long long elapsed;
 971         struct thread_data *td = data;
 972         pthread_condattr_t attr;
 973         int clear_state;
 974
 975         if (!td->o.use_thread) {
 976                 setsid();
 977                 td->pid = getpid();
 978         } else
 979                 td->pid = gettid();
 980
 981         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 982
 983         INIT_FLIST_HEAD(&td->io_u_freelist);
 984         INIT_FLIST_HEAD(&td->io_u_busylist);
 985         INIT_FLIST_HEAD(&td->io_u_requeues);
 986         INIT_FLIST_HEAD(&td->io_log_list);
 987         INIT_FLIST_HEAD(&td->io_hist_list);
 988         INIT_FLIST_HEAD(&td->verify_list);
 989         INIT_FLIST_HEAD(&td->trim_list);
 990         pthread_mutex_init(&td->io_u_lock, NULL);
 991         td->io_hist_tree = RB_ROOT;
 992
 993         pthread_condattr_init(&attr);
 994         pthread_cond_init(&td->verify_cond, &attr);
 995         pthread_cond_init(&td->free_cond, &attr);
 996
 997         td_set_runstate(td, TD_INITIALIZED);
 998         dprint(FD_MUTEX, "up startup_mutex\n");
 999         fio_mutex_up(startup_mutex);
1000         dprint(FD_MUTEX, "wait on td->mutex\n");
1001         fio_mutex_down(td->mutex);
1002         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1003
1004         /*
1005          * the ->mutex mutex is now no longer used, close it to avoid
1006          * eating a file descriptor
1007          */
1008         fio_mutex_remove(td->mutex);
1009
1010         /*
1011          * A new gid requires privilege, so we need to do this before setting
1012          * the uid.
1013          */
1014         if (td->o.gid != -1U && setgid(td->o.gid)) {
1015                 td_verror(td, errno, "setgid");
1016                 goto err;
1017         }
1018         if (td->o.uid != -1U && setuid(td->o.uid)) {
1019                 td_verror(td, errno, "setuid");
1020                 goto err;
1021         }
1022
1023         /*
1024          * If we have a gettimeofday() thread, make sure we exclude that
1025          * thread from this job
1026          */
1027         if (td->o.gtod_cpu)
1028                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1029
1030         /*
1031          * Set affinity first, in case it has an impact on the memory
1032          * allocations.
1033          */
1034         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1035                 td_verror(td, errno, "cpu_set_affinity");
1036                 goto err;
1037         }
1038
1039         /*
1040          * May alter parameters that init_io_u() will use, so we need to
1041          * do this first.
1042          */
1043         if (init_iolog(td))
1044                 goto err;
1045
1046         if (init_io_u(td))
1047                 goto err;
1048
1049         if (td->o.verify_async && verify_async_init(td))
1050                 goto err;
1051
1052         if (td->ioprio_set) {
1053                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1054                         td_verror(td, errno, "ioprio_set");
1055                         goto err;
1056                 }
1057         }
1058
1059         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1060                 goto err;
1061
1062         errno = 0;
1063         if (nice(td->o.nice) == -1 && errno != 0) {
1064                 td_verror(td, errno, "nice");
1065                 goto err;
1066         }
1067
1068         if (td->o.ioscheduler && switch_ioscheduler(td))
1069                 goto err;
1070
1071         if (!td->o.create_serialize && setup_files(td))
1072                 goto err;
1073
1074         if (td_io_init(td))
1075                 goto err;
1076
1077         if (init_random_map(td))
1078                 goto err;
1079
1080         if (td->o.exec_prerun) {
1081                 if (exec_string(td->o.exec_prerun))
1082                         goto err;
1083         }
1084
1085         if (td->o.pre_read) {
1086                 if (pre_read_files(td) < 0)
1087                         goto err;
1088         }
1089
1090         fio_gettime(&td->epoch, NULL);
1091         getrusage(RUSAGE_SELF, &td->ru_start);
1092
1093         clear_state = 0;
1094         while (keep_running(td)) {
1095                 fio_gettime(&td->start, NULL);
1096                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1097                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1098                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1099
1100                 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1101                         memcpy(&td->lastrate[0], &td->bw_sample_time,
1102                                                 sizeof(td->bw_sample_time));
1103                         memcpy(&td->lastrate[1], &td->bw_sample_time,
1104                                                 sizeof(td->bw_sample_time));
1105                 }
1106
1107                 if (clear_state)
1108                         clear_io_state(td);
1109
1110                 prune_io_piece_log(td);
1111
1112                 do_io(td);
1113
1114                 clear_state = 1;
1115
1116                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1117                         elapsed = utime_since_now(&td->start);
1118                         td->ts.runtime[DDIR_READ] += elapsed;
1119                 }
1120                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1121                         elapsed = utime_since_now(&td->start);
1122                         td->ts.runtime[DDIR_WRITE] += elapsed;
1123                 }
1124
1125                 if (td->error || td->terminate)
1126                         break;
1127
1128                 if (!td->o.do_verify ||
1129                     td->o.verify == VERIFY_NONE ||
1130                     (td->io_ops->flags & FIO_UNIDIR))
1131                         continue;
1132
1133                 clear_io_state(td);
1134
1135                 fio_gettime(&td->start, NULL);
1136
1137                 do_verify(td);
1138
1139                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1140
1141                 if (td->error || td->terminate)
1142                         break;
1143         }
1144
1145         update_rusage_stat(td);
1146         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1147         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1148         td->ts.total_run_time = mtime_since_now(&td->epoch);
1149         td->ts.io_bytes[0] = td->io_bytes[0];
1150         td->ts.io_bytes[1] = td->io_bytes[1];
1151
1152         fio_mutex_down(writeout_mutex);
1153         if (td->bw_log) {
1154                 if (td->o.bw_log_file) {
1155                         finish_log_named(td, td->bw_log,
1156                                                 td->o.bw_log_file, "bw");
1157                 } else
1158                         finish_log(td, td->bw_log, "bw");
1159         }
1160         if (td->lat_log) {
1161                 if (td->o.lat_log_file) {
1162                         finish_log_named(td, td->lat_log,
1163                                                 td->o.lat_log_file, "lat");
1164                 } else
1165                         finish_log(td, td->lat_log, "lat");
1166         }
1167         if (td->slat_log) {
1168                 if (td->o.lat_log_file) {
1169                         finish_log_named(td, td->slat_log,
1170                                                 td->o.lat_log_file, "slat");
1171                 } else
1172                         finish_log(td, td->slat_log, "slat");
1173         }
1174         if (td->clat_log) {
1175                 if (td->o.lat_log_file) {
1176                         finish_log_named(td, td->clat_log,
1177                                                 td->o.lat_log_file, "clat");
1178                 } else
1179                         finish_log(td, td->clat_log, "clat");
1180         }
1181         if (td->iops_log) {
1182                 if (td->o.iops_log_file) {
1183                         finish_log_named(td, td->iops_log,
1184                                                 td->o.iops_log_file, "iops");
1185                 } else
1186                         finish_log(td, td->iops_log, "iops");
1187         }
1188
1189         fio_mutex_up(writeout_mutex);
1190         if (td->o.exec_postrun)
1191                 exec_string(td->o.exec_postrun);
1192
1193         if (exitall_on_terminate)
1194                 fio_terminate_threads(td->groupid);
1195
1196 err:
1197         if (td->error)
1198                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1199                                                         td->verror);
1200
1201         if (td->o.verify_async)
1202                 verify_async_exit(td);
1203
1204         close_and_free_files(td);
1205         close_ioengine(td);
1206         cleanup_io_u(td);
1207         cgroup_shutdown(td, &cgroup_mnt);
1208
1209         if (td->o.cpumask_set) {
1210                 int ret = fio_cpuset_exit(&td->o.cpumask);
1211
1212                 td_verror(td, ret, "fio_cpuset_exit");
1213         }
1214
1215         /*
1216          * do this very late, it will log file closing as well
1217          */
1218         if (td->o.write_iolog_file)
1219                 write_iolog_close(td);
1220
1221         td_set_runstate(td, TD_EXITED);
1222         return (void *) (uintptr_t) td->error;
1223 }
1224
1225
1226 /*
1227  * We cannot pass the td data into a forked process, so attach the td and
1228  * pass it to the thread worker.
1229  */
1230 static int fork_main(int shmid, int offset)
1231 {
1232         struct thread_data *td;
1233         void *data, *ret;
1234
1235 #ifndef __hpux
1236         data = shmat(shmid, NULL, 0);
1237         if (data == (void *) -1) {
1238                 int __err = errno;
1239
1240                 perror("shmat");
1241                 return __err;
1242         }
1243 #else
1244         /*
1245          * HP-UX inherits shm mappings?
1246          */
1247         data = threads;
1248 #endif
1249
1250         td = data + offset * sizeof(struct thread_data);
1251         ret = thread_main(td);
1252         shmdt(data);
1253         return (int) (uintptr_t) ret;
1254 }
1255
1256 /*
1257  * Run over the job map and reap the threads that have exited, if any.
1258  */
1259 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1260                          unsigned int *m_rate)
1261 {
1262         struct thread_data *td;
1263         unsigned int cputhreads, realthreads, pending;
1264         int i, status, ret;
1265
1266         /*
1267          * reap exited threads (TD_EXITED -> TD_REAPED)
1268          */
1269         realthreads = pending = cputhreads = 0;
1270         for_each_td(td, i) {
1271                 int flags = 0;
1272
1273                 /*
1274                  * ->io_ops is NULL for a thread that has closed its
1275                  * io engine
1276                  */
1277                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1278                         cputhreads++;
1279                 else
1280                         realthreads++;
1281
1282                 if (!td->pid) {
1283                         pending++;
1284                         continue;
1285                 }
1286                 if (td->runstate == TD_REAPED)
1287                         continue;
1288                 if (td->o.use_thread) {
1289                         if (td->runstate == TD_EXITED) {
1290                                 td_set_runstate(td, TD_REAPED);
1291                                 goto reaped;
1292                         }
1293                         continue;
1294                 }
1295
1296                 flags = WNOHANG;
1297                 if (td->runstate == TD_EXITED)
1298                         flags = 0;
1299
1300                 /*
1301                  * check if someone quit or got killed in an unusual way
1302                  */
1303                 ret = waitpid(td->pid, &status, flags);
1304                 if (ret < 0) {
1305                         if (errno == ECHILD) {
1306                                 log_err("fio: pid=%d disappeared %d\n",
1307                                                 (int) td->pid, td->runstate);
1308                                 td->sig = ECHILD;
1309                                 td_set_runstate(td, TD_REAPED);
1310                                 goto reaped;
1311                         }
1312                         perror("waitpid");
1313                 } else if (ret == td->pid) {
1314                         if (WIFSIGNALED(status)) {
1315                                 int sig = WTERMSIG(status);
1316
1317                                 if (sig != SIGTERM)
1318                                         log_err("fio: pid=%d, got signal=%d\n",
1319                                                         (int) td->pid, sig);
1320                                 td->sig = sig;
1321                                 td_set_runstate(td, TD_REAPED);
1322                                 goto reaped;
1323                         }
1324                         if (WIFEXITED(status)) {
1325                                 if (WEXITSTATUS(status) && !td->error)
1326                                         td->error = WEXITSTATUS(status);
1327
1328                                 td_set_runstate(td, TD_REAPED);
1329                                 goto reaped;
1330                         }
1331                 }
1332
1333                 /*
1334                  * thread is not dead, continue
1335                  */
1336                 pending++;
1337                 continue;
1338 reaped:
1339                 (*nr_running)--;
1340                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1341                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1342                 if (!td->pid)
1343                         pending--;
1344
1345                 if (td->error)
1346                         exit_value++;
1347
1348                 done_secs += mtime_since_now(&td->epoch) / 1000;
1349         }
1350
1351         if (*nr_running == cputhreads && !pending && realthreads)
1352                 fio_terminate_threads(TERMINATE_ALL);
1353 }
1354
1355 /*
1356  * Main function for kicking off and reaping jobs, as needed.
1357  */
1358 static void run_threads(void)
1359 {
1360         struct thread_data *td;
1361         unsigned long spent;
1362         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1363
1364         if (fio_pin_memory())
1365                 return;
1366
1367         if (fio_gtod_offload && fio_start_gtod_thread())
1368                 return;
1369
1370         set_sig_handlers();
1371
1372         if (!terse_output) {
1373                 log_info("Starting ");
1374                 if (nr_thread)
1375                         log_info("%d thread%s", nr_thread,
1376                                                 nr_thread > 1 ? "s" : "");
1377                 if (nr_process) {
1378                         if (nr_thread)
1379                                 log_info(" and ");
1380                         log_info("%d process%s", nr_process,
1381                                                 nr_process > 1 ? "es" : "");
1382                 }
1383                 log_info("\n");
1384                 fflush(stdout);
1385         }
1386
1387         todo = thread_number;
1388         nr_running = 0;
1389         nr_started = 0;
1390         m_rate = t_rate = 0;
1391
1392         for_each_td(td, i) {
1393                 print_status_init(td->thread_number - 1);
1394
1395                 if (!td->o.create_serialize)
1396                         continue;
1397
1398                 /*
1399                  * do file setup here so it happens sequentially,
1400                  * we don't want X number of threads getting their
1401                  * client data interspersed on disk
1402                  */
1403                 if (setup_files(td)) {
1404                         exit_value++;
1405                         if (td->error)
1406                                 log_err("fio: pid=%d, err=%d/%s\n",
1407                                         (int) td->pid, td->error, td->verror);
1408                         td_set_runstate(td, TD_REAPED);
1409                         todo--;
1410                 } else {
1411                         struct fio_file *f;
1412                         unsigned int j;
1413
1414                         /*
1415                          * for sharing to work, each job must always open
1416                          * its own files. so close them, if we opened them
1417                          * for creation
1418                          */
1419                         for_each_file(td, f, j) {
1420                                 if (fio_file_open(f))
1421                                         td_io_close_file(td, f);
1422                         }
1423                 }
1424         }
1425
1426         set_genesis_time();
1427
1428         while (todo) {
1429                 struct thread_data *map[REAL_MAX_JOBS];
1430                 struct timeval this_start;
1431                 int this_jobs = 0, left;
1432
1433                 /*
1434                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1435                  */
1436                 for_each_td(td, i) {
1437                         if (td->runstate != TD_NOT_CREATED)
1438                                 continue;
1439
1440                         /*
1441                          * never got a chance to start, killed by other
1442                          * thread for some reason
1443                          */
1444                         if (td->terminate) {
1445                                 todo--;
1446                                 continue;
1447                         }
1448
1449                         if (td->o.start_delay) {
1450                                 spent = mtime_since_genesis();
1451
1452                                 if (td->o.start_delay * 1000 > spent)
1453                                         continue;
1454                         }
1455
1456                         if (td->o.stonewall && (nr_started || nr_running)) {
1457                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1458                                                         td->o.name);
1459                                 break;
1460                         }
1461
1462                         init_disk_util(td);
1463
1464                         /*
1465                          * Set state to created. Thread will transition
1466                          * to TD_INITIALIZED when it's done setting up.
1467                          */
1468                         td_set_runstate(td, TD_CREATED);
1469                         map[this_jobs++] = td;
1470                         nr_started++;
1471
1472                         if (td->o.use_thread) {
1473                                 int ret;
1474
1475                                 dprint(FD_PROCESS, "will pthread_create\n");
1476                                 ret = pthread_create(&td->thread, NULL,
1477                                                         thread_main, td);
1478                                 if (ret) {
1479                                         log_err("pthread_create: %s\n",
1480                                                         strerror(ret));
1481                                         nr_started--;
1482                                         break;
1483                                 }
1484                                 ret = pthread_detach(td->thread);
1485                                 if (ret)
1486                                         log_err("pthread_detach: %s",
1487                                                         strerror(ret));
1488                         } else {
1489                                 pid_t pid;
1490                                 dprint(FD_PROCESS, "will fork\n");
1491                                 pid = fork();
1492                                 if (!pid) {
1493                                         int ret = fork_main(shm_id, i);
1494
1495                                         _exit(ret);
1496                                 } else if (i == fio_debug_jobno)
1497                                         *fio_debug_jobp = pid;
1498                         }
1499                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1500                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1501                                 log_err("fio: job startup hung? exiting.\n");
1502                                 fio_terminate_threads(TERMINATE_ALL);
1503                                 fio_abort = 1;
1504                                 nr_started--;
1505                                 break;
1506                         }
1507                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1508                 }
1509
1510                 /*
1511                  * Wait for the started threads to transition to
1512                  * TD_INITIALIZED.
1513                  */
1514                 fio_gettime(&this_start, NULL);
1515                 left = this_jobs;
1516                 while (left && !fio_abort) {
1517                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1518                                 break;
1519
1520                         usleep(100000);
1521
1522                         for (i = 0; i < this_jobs; i++) {
1523                                 td = map[i];
1524                                 if (!td)
1525                                         continue;
1526                                 if (td->runstate == TD_INITIALIZED) {
1527                                         map[i] = NULL;
1528                                         left--;
1529                                 } else if (td->runstate >= TD_EXITED) {
1530                                         map[i] = NULL;
1531                                         left--;
1532                                         todo--;
1533                                         nr_running++; /* work-around... */
1534                                 }
1535                         }
1536                 }
1537
1538                 if (left) {
1539                         log_err("fio: %d job%s failed to start\n", left,
1540                                         left > 1 ? "s" : "");
1541                         for (i = 0; i < this_jobs; i++) {
1542                                 td = map[i];
1543                                 if (!td)
1544                                         continue;
1545                                 kill(td->pid, SIGTERM);
1546                         }
1547                         break;
1548                 }
1549
1550                 /*
1551                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1552                  */
1553                 for_each_td(td, i) {
1554                         if (td->runstate != TD_INITIALIZED)
1555                                 continue;
1556
1557                         if (in_ramp_time(td))
1558                                 td_set_runstate(td, TD_RAMP);
1559                         else
1560                                 td_set_runstate(td, TD_RUNNING);
1561                         nr_running++;
1562                         nr_started--;
1563                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1564                         t_rate += td->o.rate[0] + td->o.rate[1];
1565                         todo--;
1566                         fio_mutex_up(td->mutex);
1567                 }
1568
1569                 reap_threads(&nr_running, &t_rate, &m_rate);
1570
1571                 if (todo) {
1572                         if (is_backend)
1573                                 fio_server_idle_loop();
1574                         else
1575                                 usleep(100000);
1576                 }
1577         }
1578
1579         while (nr_running) {
1580                 reap_threads(&nr_running, &t_rate, &m_rate);
1581
1582                 if (is_backend)
1583                         fio_server_idle_loop();
1584                 else
1585                         usleep(10000);
1586         }
1587
1588         update_io_ticks();
1589         fio_unpin_memory();
1590 }
1591
1592 void wait_for_disk_thread_exit(void)
1593 {
1594         fio_mutex_down(disk_thread_mutex);
1595 }
1596
1597 static void *disk_thread_main(void *data)
1598 {
1599         int ret = 0;
1600
1601         fio_mutex_up(startup_mutex);
1602
1603         while (threads && !ret) {
1604                 usleep(DISK_UTIL_MSEC * 1000);
1605                 if (!threads)
1606                         break;
1607                 ret = update_io_ticks();
1608
1609                 if (!is_backend)
1610                         print_thread_status();
1611         }
1612
1613         fio_mutex_up(disk_thread_mutex);
1614         return NULL;
1615 }
1616
1617 static int create_disk_util_thread(void)
1618 {
1619         int ret;
1620
1621         setup_disk_util();
1622
1623         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1624
1625         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1626         if (ret) {
1627                 fio_mutex_remove(disk_thread_mutex);
1628                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1629                 return 1;
1630         }
1631
1632         ret = pthread_detach(disk_util_thread);
1633         if (ret) {
1634                 fio_mutex_remove(disk_thread_mutex);
1635                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1636                 return 1;
1637         }
1638
1639         dprint(FD_MUTEX, "wait on startup_mutex\n");
1640         fio_mutex_down(startup_mutex);
1641         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1642         return 0;
1643 }
1644
1645 int fio_backend(void)
1646 {
1647         struct thread_data *td;
1648         int i;
1649
1650         if (exec_profile) {
1651                 if (load_profile(exec_profile))
1652                         return 1;
1653                 free(exec_profile);
1654                 exec_profile = NULL;
1655         }
1656         if (!thread_number)
1657                 return 0;
1658
1659         if (write_bw_log) {
1660                 setup_log(&agg_io_log[DDIR_READ], 0);
1661                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1662         }
1663
1664         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1665         if (startup_mutex == NULL)
1666                 return 1;
1667         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1668         if (writeout_mutex == NULL)
1669                 return 1;
1670
1671         set_genesis_time();
1672         create_disk_util_thread();
1673
1674         cgroup_list = smalloc(sizeof(*cgroup_list));
1675         INIT_FLIST_HEAD(cgroup_list);
1676
1677         run_threads();
1678
1679         if (!fio_abort) {
1680                 show_run_stats();
1681                 if (write_bw_log) {
1682                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1683                         __finish_log(agg_io_log[DDIR_WRITE],
1684                                         "agg-write_bw.log");
1685                 }
1686         }
1687
1688         for_each_td(td, i)
1689                 fio_options_free(td);
1690
1691         free_disk_util();
1692         cgroup_kill(cgroup_list);
1693         sfree(cgroup_list);
1694         sfree(cgroup_mnt);
1695
1696         fio_mutex_remove(startup_mutex);
1697         fio_mutex_remove(writeout_mutex);
1698         fio_mutex_remove(disk_thread_mutex);
1699         return exit_value;
1700 }