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