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[DDIR_RWDIR_CNT];
  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[DDIR_READ])
 212                 ret |= __check_min_rate(td, now, DDIR_READ);
 213         if (bytes_done[DDIR_WRITE])
 214                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 215         if (bytes_done[DDIR_TRIM])
 216                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 217
 218         return ret;
 219 }
 220
 221 /*
 222  * When job exits, we can cancel the in-flight IO if we are using async
 223  * io. Attempt to do so.
 224  */
 225 static void cleanup_pending_aio(struct thread_data *td)
 226 {
 227         struct flist_head *entry, *n;
 228         struct io_u *io_u;
 229         int r;
 230
 231         /*
 232          * get immediately available events, if any
 233          */
 234         r = io_u_queued_complete(td, 0, NULL);
 235         if (r < 0)
 236                 return;
 237
 238         /*
 239          * now cancel remaining active events
 240          */
 241         if (td->io_ops->cancel) {
 242                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 243                         io_u = flist_entry(entry, struct io_u, list);
 244
 245                         /*
 246                          * if the io_u isn't in flight, then that generally
 247                          * means someone leaked an io_u. complain but fix
 248                          * it up, so we don't stall here.
 249                          */
 250                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 251                                 log_err("fio: non-busy IO on busy list\n");
 252                                 put_io_u(td, io_u);
 253                         } else {
 254                                 r = td->io_ops->cancel(td, io_u);
 255                                 if (!r)
 256                                         put_io_u(td, io_u);
 257                         }
 258                 }
 259         }
 260
 261         if (td->cur_depth)
 262                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 263 }
 264
 265 /*
 266  * Helper to handle the final sync of a file. Works just like the normal
 267  * io path, just does everything sync.
 268  */
 269 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 270 {
 271         struct io_u *io_u = __get_io_u(td);
 272         int ret;
 273
 274         if (!io_u)
 275                 return 1;
 276
 277         io_u->ddir = DDIR_SYNC;
 278         io_u->file = f;
 279
 280         if (td_io_prep(td, io_u)) {
 281                 put_io_u(td, io_u);
 282                 return 1;
 283         }
 284
 285 requeue:
 286         ret = td_io_queue(td, io_u);
 287         if (ret < 0) {
 288                 td_verror(td, io_u->error, "td_io_queue");
 289                 put_io_u(td, io_u);
 290                 return 1;
 291         } else if (ret == FIO_Q_QUEUED) {
 292                 if (io_u_queued_complete(td, 1, NULL) < 0)
 293                         return 1;
 294         } else if (ret == FIO_Q_COMPLETED) {
 295                 if (io_u->error) {
 296                         td_verror(td, io_u->error, "td_io_queue");
 297                         return 1;
 298                 }
 299
 300                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 301                         return 1;
 302         } else if (ret == FIO_Q_BUSY) {
 303                 if (td_io_commit(td))
 304                         return 1;
 305                 goto requeue;
 306         }
 307
 308         return 0;
 309 }
 310
 311 static inline void __update_tv_cache(struct thread_data *td)
 312 {
 313         fio_gettime(&td->tv_cache, NULL);
 314 }
 315
 316 static inline void update_tv_cache(struct thread_data *td)
 317 {
 318         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 319                 __update_tv_cache(td);
 320 }
 321
 322 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 323 {
 324         if (in_ramp_time(td))
 325                 return 0;
 326         if (!td->o.timeout)
 327                 return 0;
 328         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 329                 return 1;
 330
 331         return 0;
 332 }
 333
 334 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 335                                int *retptr)
 336 {
 337         int ret = *retptr;
 338
 339         if (ret < 0 || td->error) {
 340                 int err;
 341
 342                 if (ret < 0)
 343                         err = -ret;
 344                 else
 345                         err = td->error;
 346
 347                 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
 348                         return 1;
 349
 350                 if (td_non_fatal_error(err)) {
 351                         /*
 352                          * Continue with the I/Os in case of
 353                          * a non fatal error.
 354                          */
 355                         update_error_count(td, err);
 356                         td_clear_error(td);
 357                         *retptr = 0;
 358                         return 0;
 359                 } else if (td->o.fill_device && err == ENOSPC) {
 360                         /*
 361                          * We expect to hit this error if
 362                          * fill_device option is set.
 363                          */
 364                         td_clear_error(td);
 365                         td->terminate = 1;
 366                         return 1;
 367                 } else {
 368                         /*
 369                          * Stop the I/O in case of a fatal
 370                          * error.
 371                          */
 372                         update_error_count(td, err);
 373                         return 1;
 374                 }
 375         }
 376
 377         return 0;
 378 }
 379
 380 /*
 381  * The main verify engine. Runs over the writes we previously submitted,
 382  * reads the blocks back in, and checks the crc/md5 of the data.
 383  */
 384 static void do_verify(struct thread_data *td)
 385 {
 386         struct fio_file *f;
 387         struct io_u *io_u;
 388         int ret, min_events;
 389         unsigned int i;
 390
 391         dprint(FD_VERIFY, "starting loop\n");
 392
 393         /*
 394          * sync io first and invalidate cache, to make sure we really
 395          * read from disk.
 396          */
 397         for_each_file(td, f, i) {
 398                 if (!fio_file_open(f))
 399                         continue;
 400                 if (fio_io_sync(td, f))
 401                         break;
 402                 if (file_invalidate_cache(td, f))
 403                         break;
 404         }
 405
 406         if (td->error)
 407                 return;
 408
 409         td_set_runstate(td, TD_VERIFYING);
 410
 411         io_u = NULL;
 412         while (!td->terminate) {
 413                 int ret2, full;
 414
 415                 update_tv_cache(td);
 416
 417                 if (runtime_exceeded(td, &td->tv_cache)) {
 418                         __update_tv_cache(td);
 419                         if (runtime_exceeded(td, &td->tv_cache)) {
 420                                 td->terminate = 1;
 421                                 break;
 422                         }
 423                 }
 424
 425                 if (flow_threshold_exceeded(td))
 426                         continue;
 427
 428                 io_u = __get_io_u(td);
 429                 if (!io_u)
 430                         break;
 431
 432                 if (get_next_verify(td, io_u)) {
 433                         put_io_u(td, io_u);
 434                         break;
 435                 }
 436
 437                 if (td_io_prep(td, io_u)) {
 438                         put_io_u(td, io_u);
 439                         break;
 440                 }
 441
 442                 if (td->o.verify_async)
 443                         io_u->end_io = verify_io_u_async;
 444                 else
 445                         io_u->end_io = verify_io_u;
 446
 447                 ret = td_io_queue(td, io_u);
 448                 switch (ret) {
 449                 case FIO_Q_COMPLETED:
 450                         if (io_u->error) {
 451                                 ret = -io_u->error;
 452                                 clear_io_u(td, io_u);
 453                         } else if (io_u->resid) {
 454                                 int bytes = io_u->xfer_buflen - io_u->resid;
 455
 456                                 /*
 457                                  * zero read, fail
 458                                  */
 459                                 if (!bytes) {
 460                                         td_verror(td, EIO, "full resid");
 461                                         put_io_u(td, io_u);
 462                                         break;
 463                                 }
 464
 465                                 io_u->xfer_buflen = io_u->resid;
 466                                 io_u->xfer_buf += bytes;
 467                                 io_u->offset += bytes;
 468
 469                                 if (ddir_rw(io_u->ddir))
 470                                         td->ts.short_io_u[io_u->ddir]++;
 471
 472                                 f = io_u->file;
 473                                 if (io_u->offset == f->real_file_size)
 474                                         goto sync_done;
 475
 476                                 requeue_io_u(td, &io_u);
 477                         } else {
 478 sync_done:
 479                                 ret = io_u_sync_complete(td, io_u, NULL);
 480                                 if (ret < 0)
 481                                         break;
 482                         }
 483                         continue;
 484                 case FIO_Q_QUEUED:
 485                         break;
 486                 case FIO_Q_BUSY:
 487                         requeue_io_u(td, &io_u);
 488                         ret2 = td_io_commit(td);
 489                         if (ret2 < 0)
 490                                 ret = ret2;
 491                         break;
 492                 default:
 493                         assert(ret < 0);
 494                         td_verror(td, -ret, "td_io_queue");
 495                         break;
 496                 }
 497
 498                 if (break_on_this_error(td, io_u->ddir, &ret))
 499                         break;
 500
 501                 /*
 502                  * if we can queue more, do so. but check if there are
 503                  * completed io_u's first. Note that we can get BUSY even
 504                  * without IO queued, if the system is resource starved.
 505                  */
 506                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 507                 if (full || !td->o.iodepth_batch_complete) {
 508                         min_events = min(td->o.iodepth_batch_complete,
 509                                          td->cur_depth);
 510                         /*
 511                          * if the queue is full, we MUST reap at least 1 event
 512                          */
 513                         if (full && !min_events)
 514                                 min_events = 1;
 515
 516                         do {
 517                                 /*
 518                                  * Reap required number of io units, if any,
 519                                  * and do the verification on them through
 520                                  * the callback handler
 521                                  */
 522                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 523                                         ret = -1;
 524                                         break;
 525                                 }
 526                         } while (full && (td->cur_depth > td->o.iodepth_low));
 527                 }
 528                 if (ret < 0)
 529                         break;
 530         }
 531
 532         if (!td->error) {
 533                 min_events = td->cur_depth;
 534
 535                 if (min_events)
 536                         ret = io_u_queued_complete(td, min_events, NULL);
 537         } else
 538                 cleanup_pending_aio(td);
 539
 540         td_set_runstate(td, TD_RUNNING);
 541
 542         dprint(FD_VERIFY, "exiting loop\n");
 543 }
 544
 545 static int io_bytes_exceeded(struct thread_data *td)
 546 {
 547         unsigned long long bytes;
 548
 549         if (td_rw(td))
 550                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 551         else if (td_write(td))
 552                 bytes = td->this_io_bytes[DDIR_WRITE];
 553         else if (td_read(td))
 554                 bytes = td->this_io_bytes[DDIR_READ];
 555         else
 556                 bytes = td->this_io_bytes[DDIR_TRIM];
 557
 558         return bytes >= td->o.size;
 559 }
 560
 561 /*
 562  * Main IO worker function. It retrieves io_u's to process and queues
 563  * and reaps them, checking for rate and errors along the way.
 564  */
 565 static void do_io(struct thread_data *td)
 566 {
 567         unsigned int i;
 568         int ret = 0;
 569
 570         if (in_ramp_time(td))
 571                 td_set_runstate(td, TD_RAMP);
 572         else
 573                 td_set_runstate(td, TD_RUNNING);
 574
 575         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 576                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 577                 td->o.time_based) {
 578                 struct timeval comp_time;
 579                 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 580                 int min_evts = 0;
 581                 struct io_u *io_u;
 582                 int ret2, full;
 583                 enum fio_ddir ddir;
 584
 585                 if (td->terminate)
 586                         break;
 587
 588                 update_tv_cache(td);
 589
 590                 if (runtime_exceeded(td, &td->tv_cache)) {
 591                         __update_tv_cache(td);
 592                         if (runtime_exceeded(td, &td->tv_cache)) {
 593                                 td->terminate = 1;
 594                                 break;
 595                         }
 596                 }
 597
 598                 if (flow_threshold_exceeded(td))
 599                         continue;
 600
 601                 io_u = get_io_u(td);
 602                 if (!io_u)
 603                         break;
 604
 605                 ddir = io_u->ddir;
 606
 607                 /*
 608                  * Add verification end_io handler if:
 609                  *      - Asked to verify (!td_rw(td))
 610                  *      - Or the io_u is from our verify list (mixed write/ver)
 611                  */
 612                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 613                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 614                         if (td->o.verify_async)
 615                                 io_u->end_io = verify_io_u_async;
 616                         else
 617                                 io_u->end_io = verify_io_u;
 618                         td_set_runstate(td, TD_VERIFYING);
 619                 } else if (in_ramp_time(td))
 620                         td_set_runstate(td, TD_RAMP);
 621                 else
 622                         td_set_runstate(td, TD_RUNNING);
 623
 624                 ret = td_io_queue(td, io_u);
 625                 switch (ret) {
 626                 case FIO_Q_COMPLETED:
 627                         if (io_u->error) {
 628                                 ret = -io_u->error;
 629                                 clear_io_u(td, io_u);
 630                         } else if (io_u->resid) {
 631                                 int bytes = io_u->xfer_buflen - io_u->resid;
 632                                 struct fio_file *f = io_u->file;
 633
 634                                 /*
 635                                  * zero read, fail
 636                                  */
 637                                 if (!bytes) {
 638                                         td_verror(td, EIO, "full resid");
 639                                         put_io_u(td, io_u);
 640                                         break;
 641                                 }
 642
 643                                 io_u->xfer_buflen = io_u->resid;
 644                                 io_u->xfer_buf += bytes;
 645                                 io_u->offset += bytes;
 646
 647                                 if (ddir_rw(io_u->ddir))
 648                                         td->ts.short_io_u[io_u->ddir]++;
 649
 650                                 if (io_u->offset == f->real_file_size)
 651                                         goto sync_done;
 652
 653                                 requeue_io_u(td, &io_u);
 654                         } else {
 655 sync_done:
 656                                 if (__should_check_rate(td, DDIR_READ) ||
 657                                     __should_check_rate(td, DDIR_WRITE) ||
 658                                     __should_check_rate(td, DDIR_TRIM))
 659                                         fio_gettime(&comp_time, NULL);
 660
 661                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 662                                 if (ret < 0)
 663                                         break;
 664                         }
 665                         break;
 666                 case FIO_Q_QUEUED:
 667                         /*
 668                          * if the engine doesn't have a commit hook,
 669                          * the io_u is really queued. if it does have such
 670                          * a hook, it has to call io_u_queued() itself.
 671                          */
 672                         if (td->io_ops->commit == NULL)
 673                                 io_u_queued(td, io_u);
 674                         break;
 675                 case FIO_Q_BUSY:
 676                         requeue_io_u(td, &io_u);
 677                         ret2 = td_io_commit(td);
 678                         if (ret2 < 0)
 679                                 ret = ret2;
 680                         break;
 681                 default:
 682                         assert(ret < 0);
 683                         put_io_u(td, io_u);
 684                         break;
 685                 }
 686
 687                 if (break_on_this_error(td, ddir, &ret))
 688                         break;
 689
 690                 /*
 691                  * See if we need to complete some commands. Note that we
 692                  * can get BUSY even without IO queued, if the system is
 693                  * resource starved.
 694                  */
 695                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 696                 if (full || !td->o.iodepth_batch_complete) {
 697                         min_evts = min(td->o.iodepth_batch_complete,
 698                                         td->cur_depth);
 699                         /*
 700                          * if the queue is full, we MUST reap at least 1 event
 701                          */
 702                         if (full && !min_evts)
 703                                 min_evts = 1;
 704
 705                         if (__should_check_rate(td, DDIR_READ) ||
 706                             __should_check_rate(td, DDIR_WRITE) ||
 707                             __should_check_rate(td, DDIR_TRIM))
 708                                 fio_gettime(&comp_time, NULL);
 709
 710                         do {
 711                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 712                                 if (ret < 0)
 713                                         break;
 714
 715                         } while (full && (td->cur_depth > td->o.iodepth_low));
 716                 }
 717
 718                 if (ret < 0)
 719                         break;
 720                 if (!ddir_rw_sum(bytes_done))
 721                         continue;
 722
 723                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 724                         if (check_min_rate(td, &comp_time, bytes_done)) {
 725                                 if (exitall_on_terminate)
 726                                         fio_terminate_threads(td->groupid);
 727                                 td_verror(td, EIO, "check_min_rate");
 728                                 break;
 729                         }
 730                 }
 731
 732                 if (td->o.thinktime) {
 733                         unsigned long long b;
 734
 735                         b = ddir_rw_sum(td->io_blocks);
 736                         if (!(b % td->o.thinktime_blocks)) {
 737                                 int left;
 738
 739                                 if (td->o.thinktime_spin)
 740                                         usec_spin(td->o.thinktime_spin);
 741
 742                                 left = td->o.thinktime - td->o.thinktime_spin;
 743                                 if (left)
 744                                         usec_sleep(td, left);
 745                         }
 746                 }
 747         }
 748
 749         if (td->trim_entries)
 750                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 751
 752         if (td->o.fill_device && td->error == ENOSPC) {
 753                 td->error = 0;
 754                 td->terminate = 1;
 755         }
 756         if (!td->error) {
 757                 struct fio_file *f;
 758
 759                 i = td->cur_depth;
 760                 if (i) {
 761                         ret = io_u_queued_complete(td, i, NULL);
 762                         if (td->o.fill_device && td->error == ENOSPC)
 763                                 td->error = 0;
 764                 }
 765
 766                 if (should_fsync(td) && td->o.end_fsync) {
 767                         td_set_runstate(td, TD_FSYNCING);
 768
 769                         for_each_file(td, f, i) {
 770                                 if (!fio_file_open(f))
 771                                         continue;
 772                                 fio_io_sync(td, f);
 773                         }
 774                 }
 775         } else
 776                 cleanup_pending_aio(td);
 777
 778         /*
 779          * stop job if we failed doing any IO
 780          */
 781         if (!ddir_rw_sum(td->this_io_bytes))
 782                 td->done = 1;
 783 }
 784
 785 static void cleanup_io_u(struct thread_data *td)
 786 {
 787         struct flist_head *entry, *n;
 788         struct io_u *io_u;
 789
 790         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 791                 io_u = flist_entry(entry, struct io_u, list);
 792
 793                 flist_del(&io_u->list);
 794                 fio_memfree(io_u, sizeof(*io_u));
 795         }
 796
 797         free_io_mem(td);
 798 }
 799
 800 static int init_io_u(struct thread_data *td)
 801 {
 802         struct io_u *io_u;
 803         unsigned int max_bs, min_write;
 804         int cl_align, i, max_units;
 805         char *p;
 806
 807         max_units = td->o.iodepth;
 808         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 809         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
 810         min_write = td->o.min_bs[DDIR_WRITE];
 811         td->orig_buffer_size = (unsigned long long) max_bs
 812                                         * (unsigned long long) max_units;
 813
 814         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 815                 unsigned long bs;
 816
 817                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 818                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 819         }
 820
 821         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 822                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 823                 return 1;
 824         }
 825
 826         if (allocate_io_mem(td))
 827                 return 1;
 828
 829         if (td->o.odirect || td->o.mem_align ||
 830             (td->io_ops->flags & FIO_RAWIO))
 831                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 832         else
 833                 p = td->orig_buffer;
 834
 835         cl_align = os_cache_line_size();
 836
 837         for (i = 0; i < max_units; i++) {
 838                 void *ptr;
 839
 840                 if (td->terminate)
 841                         return 1;
 842
 843                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 844                 if (!ptr) {
 845                         log_err("fio: unable to allocate aligned memory\n");
 846                         break;
 847                 }
 848
 849                 io_u = ptr;
 850                 memset(io_u, 0, sizeof(*io_u));
 851                 INIT_FLIST_HEAD(&io_u->list);
 852                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 853
 854                 if (!(td->io_ops->flags & FIO_NOIO)) {
 855                         io_u->buf = p;
 856                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 857
 858                         if (td_write(td))
 859                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 860                         if (td_write(td) && td->o.verify_pattern_bytes) {
 861                                 /*
 862                                  * Fill the buffer with the pattern if we are
 863                                  * going to be doing writes.
 864                                  */
 865                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 866                         }
 867                 }
 868
 869                 io_u->index = i;
 870                 io_u->flags = IO_U_F_FREE;
 871                 flist_add(&io_u->list, &td->io_u_freelist);
 872                 p += max_bs;
 873         }
 874
 875         return 0;
 876 }
 877
 878 static int switch_ioscheduler(struct thread_data *td)
 879 {
 880         char tmp[256], tmp2[128];
 881         FILE *f;
 882         int ret;
 883
 884         if (td->io_ops->flags & FIO_DISKLESSIO)
 885                 return 0;
 886
 887         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 888
 889         f = fopen(tmp, "r+");
 890         if (!f) {
 891                 if (errno == ENOENT) {
 892                         log_err("fio: os or kernel doesn't support IO scheduler"
 893                                 " switching\n");
 894                         return 0;
 895                 }
 896                 td_verror(td, errno, "fopen iosched");
 897                 return 1;
 898         }
 899
 900         /*
 901          * Set io scheduler.
 902          */
 903         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 904         if (ferror(f) || ret != 1) {
 905                 td_verror(td, errno, "fwrite");
 906                 fclose(f);
 907                 return 1;
 908         }
 909
 910         rewind(f);
 911
 912         /*
 913          * Read back and check that the selected scheduler is now the default.
 914          */
 915         ret = fread(tmp, 1, sizeof(tmp), f);
 916         if (ferror(f) || ret < 0) {
 917                 td_verror(td, errno, "fread");
 918                 fclose(f);
 919                 return 1;
 920         }
 921
 922         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 923         if (!strstr(tmp, tmp2)) {
 924                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 925                 td_verror(td, EINVAL, "iosched_switch");
 926                 fclose(f);
 927                 return 1;
 928         }
 929
 930         fclose(f);
 931         return 0;
 932 }
 933
 934 static int keep_running(struct thread_data *td)
 935 {
 936         if (td->done)
 937                 return 0;
 938         if (td->o.time_based)
 939                 return 1;
 940         if (td->o.loops) {
 941                 td->o.loops--;
 942                 return 1;
 943         }
 944
 945         if (ddir_rw_sum(td->io_bytes) < td->o.size)
 946                 return 1;
 947
 948         return 0;
 949 }
 950
 951 static int exec_string(const char *string)
 952 {
 953         int ret, newlen = strlen(string) + 1 + 8;
 954         char *str;
 955
 956         str = malloc(newlen);
 957         sprintf(str, "sh -c %s", string);
 958
 959         ret = system(str);
 960         if (ret == -1)
 961                 log_err("fio: exec of cmd <%s> failed\n", str);
 962
 963         free(str);
 964         return ret;
 965 }
 966
 967 /*
 968  * Entry point for the thread based jobs. The process based jobs end up
 969  * here as well, after a little setup.
 970  */
 971 static void *thread_main(void *data)
 972 {
 973         unsigned long long elapsed;
 974         struct thread_data *td = data;
 975         pthread_condattr_t attr;
 976         int clear_state;
 977
 978         if (!td->o.use_thread) {
 979                 setsid();
 980                 td->pid = getpid();
 981         } else
 982                 td->pid = gettid();
 983
 984         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 985
 986         INIT_FLIST_HEAD(&td->io_u_freelist);
 987         INIT_FLIST_HEAD(&td->io_u_busylist);
 988         INIT_FLIST_HEAD(&td->io_u_requeues);
 989         INIT_FLIST_HEAD(&td->io_log_list);
 990         INIT_FLIST_HEAD(&td->io_hist_list);
 991         INIT_FLIST_HEAD(&td->verify_list);
 992         INIT_FLIST_HEAD(&td->trim_list);
 993         pthread_mutex_init(&td->io_u_lock, NULL);
 994         td->io_hist_tree = RB_ROOT;
 995
 996         pthread_condattr_init(&attr);
 997         pthread_cond_init(&td->verify_cond, &attr);
 998         pthread_cond_init(&td->free_cond, &attr);
 999
1000         td_set_runstate(td, TD_INITIALIZED);
1001         dprint(FD_MUTEX, "up startup_mutex\n");
1002         fio_mutex_up(startup_mutex);
1003         dprint(FD_MUTEX, "wait on td->mutex\n");
1004         fio_mutex_down(td->mutex);
1005         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1006
1007         /*
1008          * the ->mutex mutex is now no longer used, close it to avoid
1009          * eating a file descriptor
1010          */
1011         fio_mutex_remove(td->mutex);
1012
1013         /*
1014          * A new gid requires privilege, so we need to do this before setting
1015          * the uid.
1016          */
1017         if (td->o.gid != -1U && setgid(td->o.gid)) {
1018                 td_verror(td, errno, "setgid");
1019                 goto err;
1020         }
1021         if (td->o.uid != -1U && setuid(td->o.uid)) {
1022                 td_verror(td, errno, "setuid");
1023                 goto err;
1024         }
1025
1026         /*
1027          * If we have a gettimeofday() thread, make sure we exclude that
1028          * thread from this job
1029          */
1030         if (td->o.gtod_cpu)
1031                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1032
1033         /*
1034          * Set affinity first, in case it has an impact on the memory
1035          * allocations.
1036          */
1037         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1038                 td_verror(td, errno, "cpu_set_affinity");
1039                 goto err;
1040         }
1041
1042         /*
1043          * May alter parameters that init_io_u() will use, so we need to
1044          * do this first.
1045          */
1046         if (init_iolog(td))
1047                 goto err;
1048
1049         if (init_io_u(td))
1050                 goto err;
1051
1052         if (td->o.verify_async && verify_async_init(td))
1053                 goto err;
1054
1055         if (td->ioprio_set) {
1056                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1057                         td_verror(td, errno, "ioprio_set");
1058                         goto err;
1059                 }
1060         }
1061
1062         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1063                 goto err;
1064
1065         errno = 0;
1066         if (nice(td->o.nice) == -1 && errno != 0) {
1067                 td_verror(td, errno, "nice");
1068                 goto err;
1069         }
1070
1071         if (td->o.ioscheduler && switch_ioscheduler(td))
1072                 goto err;
1073
1074         if (!td->o.create_serialize && setup_files(td))
1075                 goto err;
1076
1077         if (td_io_init(td))
1078                 goto err;
1079
1080         if (init_random_map(td))
1081                 goto err;
1082
1083         if (td->o.exec_prerun) {
1084                 if (exec_string(td->o.exec_prerun))
1085                         goto err;
1086         }
1087
1088         if (td->o.pre_read) {
1089                 if (pre_read_files(td) < 0)
1090                         goto err;
1091         }
1092
1093         fio_gettime(&td->epoch, NULL);
1094         getrusage(RUSAGE_SELF, &td->ru_start);
1095
1096         clear_state = 0;
1097         while (keep_running(td)) {
1098                 fio_gettime(&td->start, NULL);
1099                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1100                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1101                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1102
1103                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1104                                 td->o.ratemin[DDIR_TRIM]) {
1105                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1106                                                 sizeof(td->bw_sample_time));
1107                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1108                                                 sizeof(td->bw_sample_time));
1109                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1110                                                 sizeof(td->bw_sample_time));
1111                 }
1112
1113                 if (clear_state)
1114                         clear_io_state(td);
1115
1116                 prune_io_piece_log(td);
1117
1118                 do_io(td);
1119
1120                 clear_state = 1;
1121
1122                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1123                         elapsed = utime_since_now(&td->start);
1124                         td->ts.runtime[DDIR_READ] += elapsed;
1125                 }
1126                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1127                         elapsed = utime_since_now(&td->start);
1128                         td->ts.runtime[DDIR_WRITE] += elapsed;
1129                 }
1130                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1131                         elapsed = utime_since_now(&td->start);
1132                         td->ts.runtime[DDIR_TRIM] += elapsed;
1133                 }
1134
1135                 if (td->error || td->terminate)
1136                         break;
1137
1138                 if (!td->o.do_verify ||
1139                     td->o.verify == VERIFY_NONE ||
1140                     (td->io_ops->flags & FIO_UNIDIR))
1141                         continue;
1142
1143                 clear_io_state(td);
1144
1145                 fio_gettime(&td->start, NULL);
1146
1147                 do_verify(td);
1148
1149                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1150
1151                 if (td->error || td->terminate)
1152                         break;
1153         }
1154
1155         update_rusage_stat(td);
1156         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1157         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1158         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1159         td->ts.total_run_time = mtime_since_now(&td->epoch);
1160         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1161         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1162         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1163
1164         fio_mutex_down(writeout_mutex);
1165         if (td->bw_log) {
1166                 if (td->o.bw_log_file) {
1167                         finish_log_named(td, td->bw_log,
1168                                                 td->o.bw_log_file, "bw");
1169                 } else
1170                         finish_log(td, td->bw_log, "bw");
1171         }
1172         if (td->lat_log) {
1173                 if (td->o.lat_log_file) {
1174                         finish_log_named(td, td->lat_log,
1175                                                 td->o.lat_log_file, "lat");
1176                 } else
1177                         finish_log(td, td->lat_log, "lat");
1178         }
1179         if (td->slat_log) {
1180                 if (td->o.lat_log_file) {
1181                         finish_log_named(td, td->slat_log,
1182                                                 td->o.lat_log_file, "slat");
1183                 } else
1184                         finish_log(td, td->slat_log, "slat");
1185         }
1186         if (td->clat_log) {
1187                 if (td->o.lat_log_file) {
1188                         finish_log_named(td, td->clat_log,
1189                                                 td->o.lat_log_file, "clat");
1190                 } else
1191                         finish_log(td, td->clat_log, "clat");
1192         }
1193         if (td->iops_log) {
1194                 if (td->o.iops_log_file) {
1195                         finish_log_named(td, td->iops_log,
1196                                                 td->o.iops_log_file, "iops");
1197                 } else
1198                         finish_log(td, td->iops_log, "iops");
1199         }
1200
1201         fio_mutex_up(writeout_mutex);
1202         if (td->o.exec_postrun)
1203                 exec_string(td->o.exec_postrun);
1204
1205         if (exitall_on_terminate)
1206                 fio_terminate_threads(td->groupid);
1207
1208 err:
1209         if (td->error)
1210                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1211                                                         td->verror);
1212
1213         if (td->o.verify_async)
1214                 verify_async_exit(td);
1215
1216         close_and_free_files(td);
1217         close_ioengine(td);
1218         cleanup_io_u(td);
1219         cgroup_shutdown(td, &cgroup_mnt);
1220
1221         if (td->o.cpumask_set) {
1222                 int ret = fio_cpuset_exit(&td->o.cpumask);
1223
1224                 td_verror(td, ret, "fio_cpuset_exit");
1225         }
1226
1227         /*
1228          * do this very late, it will log file closing as well
1229          */
1230         if (td->o.write_iolog_file)
1231                 write_iolog_close(td);
1232
1233         td_set_runstate(td, TD_EXITED);
1234         return (void *) (uintptr_t) td->error;
1235 }
1236
1237
1238 /*
1239  * We cannot pass the td data into a forked process, so attach the td and
1240  * pass it to the thread worker.
1241  */
1242 static int fork_main(int shmid, int offset)
1243 {
1244         struct thread_data *td;
1245         void *data, *ret;
1246
1247 #ifndef __hpux
1248         data = shmat(shmid, NULL, 0);
1249         if (data == (void *) -1) {
1250                 int __err = errno;
1251
1252                 perror("shmat");
1253                 return __err;
1254         }
1255 #else
1256         /*
1257          * HP-UX inherits shm mappings?
1258          */
1259         data = threads;
1260 #endif
1261
1262         td = data + offset * sizeof(struct thread_data);
1263         ret = thread_main(td);
1264         shmdt(data);
1265         return (int) (uintptr_t) ret;
1266 }
1267
1268 /*
1269  * Run over the job map and reap the threads that have exited, if any.
1270  */
1271 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1272                          unsigned int *m_rate)
1273 {
1274         struct thread_data *td;
1275         unsigned int cputhreads, realthreads, pending;
1276         int i, status, ret;
1277
1278         /*
1279          * reap exited threads (TD_EXITED -> TD_REAPED)
1280          */
1281         realthreads = pending = cputhreads = 0;
1282         for_each_td(td, i) {
1283                 int flags = 0;
1284
1285                 /*
1286                  * ->io_ops is NULL for a thread that has closed its
1287                  * io engine
1288                  */
1289                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1290                         cputhreads++;
1291                 else
1292                         realthreads++;
1293
1294                 if (!td->pid) {
1295                         pending++;
1296                         continue;
1297                 }
1298                 if (td->runstate == TD_REAPED)
1299                         continue;
1300                 if (td->o.use_thread) {
1301                         if (td->runstate == TD_EXITED) {
1302                                 td_set_runstate(td, TD_REAPED);
1303                                 goto reaped;
1304                         }
1305                         continue;
1306                 }
1307
1308                 flags = WNOHANG;
1309                 if (td->runstate == TD_EXITED)
1310                         flags = 0;
1311
1312                 /*
1313                  * check if someone quit or got killed in an unusual way
1314                  */
1315                 ret = waitpid(td->pid, &status, flags);
1316                 if (ret < 0) {
1317                         if (errno == ECHILD) {
1318                                 log_err("fio: pid=%d disappeared %d\n",
1319                                                 (int) td->pid, td->runstate);
1320                                 td->sig = ECHILD;
1321                                 td_set_runstate(td, TD_REAPED);
1322                                 goto reaped;
1323                         }
1324                         perror("waitpid");
1325                 } else if (ret == td->pid) {
1326                         if (WIFSIGNALED(status)) {
1327                                 int sig = WTERMSIG(status);
1328
1329                                 if (sig != SIGTERM)
1330                                         log_err("fio: pid=%d, got signal=%d\n",
1331                                                         (int) td->pid, sig);
1332                                 td->sig = sig;
1333                                 td_set_runstate(td, TD_REAPED);
1334                                 goto reaped;
1335                         }
1336                         if (WIFEXITED(status)) {
1337                                 if (WEXITSTATUS(status) && !td->error)
1338                                         td->error = WEXITSTATUS(status);
1339
1340                                 td_set_runstate(td, TD_REAPED);
1341                                 goto reaped;
1342                         }
1343                 }
1344
1345                 /*
1346                  * thread is not dead, continue
1347                  */
1348                 pending++;
1349                 continue;
1350 reaped:
1351                 (*nr_running)--;
1352                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1353                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1354                 if (!td->pid)
1355                         pending--;
1356
1357                 if (td->error)
1358                         exit_value++;
1359
1360                 done_secs += mtime_since_now(&td->epoch) / 1000;
1361         }
1362
1363         if (*nr_running == cputhreads && !pending && realthreads)
1364                 fio_terminate_threads(TERMINATE_ALL);
1365 }
1366
1367 /*
1368  * Main function for kicking off and reaping jobs, as needed.
1369  */
1370 static void run_threads(void)
1371 {
1372         struct thread_data *td;
1373         unsigned long spent;
1374         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1375
1376         if (fio_pin_memory())
1377                 return;
1378
1379         if (fio_gtod_offload && fio_start_gtod_thread())
1380                 return;
1381
1382         set_sig_handlers();
1383
1384         if (output_format == FIO_OUTPUT_NORMAL) {
1385                 log_info("Starting ");
1386                 if (nr_thread)
1387                         log_info("%d thread%s", nr_thread,
1388                                                 nr_thread > 1 ? "s" : "");
1389                 if (nr_process) {
1390                         if (nr_thread)
1391                                 log_info(" and ");
1392                         log_info("%d process%s", nr_process,
1393                                                 nr_process > 1 ? "es" : "");
1394                 }
1395                 log_info("\n");
1396                 fflush(stdout);
1397         }
1398
1399         todo = thread_number;
1400         nr_running = 0;
1401         nr_started = 0;
1402         m_rate = t_rate = 0;
1403
1404         for_each_td(td, i) {
1405                 print_status_init(td->thread_number - 1);
1406
1407                 if (!td->o.create_serialize)
1408                         continue;
1409
1410                 /*
1411                  * do file setup here so it happens sequentially,
1412                  * we don't want X number of threads getting their
1413                  * client data interspersed on disk
1414                  */
1415                 if (setup_files(td)) {
1416                         exit_value++;
1417                         if (td->error)
1418                                 log_err("fio: pid=%d, err=%d/%s\n",
1419                                         (int) td->pid, td->error, td->verror);
1420                         td_set_runstate(td, TD_REAPED);
1421                         todo--;
1422                 } else {
1423                         struct fio_file *f;
1424                         unsigned int j;
1425
1426                         /*
1427                          * for sharing to work, each job must always open
1428                          * its own files. so close them, if we opened them
1429                          * for creation
1430                          */
1431                         for_each_file(td, f, j) {
1432                                 if (fio_file_open(f))
1433                                         td_io_close_file(td, f);
1434                         }
1435                 }
1436         }
1437
1438         set_genesis_time();
1439
1440         while (todo) {
1441                 struct thread_data *map[REAL_MAX_JOBS];
1442                 struct timeval this_start;
1443                 int this_jobs = 0, left;
1444
1445                 /*
1446                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1447                  */
1448                 for_each_td(td, i) {
1449                         if (td->runstate != TD_NOT_CREATED)
1450                                 continue;
1451
1452                         /*
1453                          * never got a chance to start, killed by other
1454                          * thread for some reason
1455                          */
1456                         if (td->terminate) {
1457                                 todo--;
1458                                 continue;
1459                         }
1460
1461                         if (td->o.start_delay) {
1462                                 spent = mtime_since_genesis();
1463
1464                                 if (td->o.start_delay * 1000 > spent)
1465                                         continue;
1466                         }
1467
1468                         if (td->o.stonewall && (nr_started || nr_running)) {
1469                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1470                                                         td->o.name);
1471                                 break;
1472                         }
1473
1474                         init_disk_util(td);
1475
1476                         /*
1477                          * Set state to created. Thread will transition
1478                          * to TD_INITIALIZED when it's done setting up.
1479                          */
1480                         td_set_runstate(td, TD_CREATED);
1481                         map[this_jobs++] = td;
1482                         nr_started++;
1483
1484                         if (td->o.use_thread) {
1485                                 int ret;
1486
1487                                 dprint(FD_PROCESS, "will pthread_create\n");
1488                                 ret = pthread_create(&td->thread, NULL,
1489                                                         thread_main, td);
1490                                 if (ret) {
1491                                         log_err("pthread_create: %s\n",
1492                                                         strerror(ret));
1493                                         nr_started--;
1494                                         break;
1495                                 }
1496                                 ret = pthread_detach(td->thread);
1497                                 if (ret)
1498                                         log_err("pthread_detach: %s",
1499                                                         strerror(ret));
1500                         } else {
1501                                 pid_t pid;
1502                                 dprint(FD_PROCESS, "will fork\n");
1503                                 pid = fork();
1504                                 if (!pid) {
1505                                         int ret = fork_main(shm_id, i);
1506
1507                                         _exit(ret);
1508                                 } else if (i == fio_debug_jobno)
1509                                         *fio_debug_jobp = pid;
1510                         }
1511                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1512                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1513                                 log_err("fio: job startup hung? exiting.\n");
1514                                 fio_terminate_threads(TERMINATE_ALL);
1515                                 fio_abort = 1;
1516                                 nr_started--;
1517                                 break;
1518                         }
1519                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1520                 }
1521
1522                 /*
1523                  * Wait for the started threads to transition to
1524                  * TD_INITIALIZED.
1525                  */
1526                 fio_gettime(&this_start, NULL);
1527                 left = this_jobs;
1528                 while (left && !fio_abort) {
1529                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1530                                 break;
1531
1532                         usleep(100000);
1533
1534                         for (i = 0; i < this_jobs; i++) {
1535                                 td = map[i];
1536                                 if (!td)
1537                                         continue;
1538                                 if (td->runstate == TD_INITIALIZED) {
1539                                         map[i] = NULL;
1540                                         left--;
1541                                 } else if (td->runstate >= TD_EXITED) {
1542                                         map[i] = NULL;
1543                                         left--;
1544                                         todo--;
1545                                         nr_running++; /* work-around... */
1546                                 }
1547                         }
1548                 }
1549
1550                 if (left) {
1551                         log_err("fio: %d job%s failed to start\n", left,
1552                                         left > 1 ? "s" : "");
1553                         for (i = 0; i < this_jobs; i++) {
1554                                 td = map[i];
1555                                 if (!td)
1556                                         continue;
1557                                 kill(td->pid, SIGTERM);
1558                         }
1559                         break;
1560                 }
1561
1562                 /*
1563                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1564                  */
1565                 for_each_td(td, i) {
1566                         if (td->runstate != TD_INITIALIZED)
1567                                 continue;
1568
1569                         if (in_ramp_time(td))
1570                                 td_set_runstate(td, TD_RAMP);
1571                         else
1572                                 td_set_runstate(td, TD_RUNNING);
1573                         nr_running++;
1574                         nr_started--;
1575                         m_rate += ddir_rw_sum(td->o.ratemin);
1576                         t_rate += ddir_rw_sum(td->o.rate);
1577                         todo--;
1578                         fio_mutex_up(td->mutex);
1579                 }
1580
1581                 reap_threads(&nr_running, &t_rate, &m_rate);
1582
1583                 if (todo) {
1584                         if (is_backend)
1585                                 fio_server_idle_loop();
1586                         else
1587                                 usleep(100000);
1588                 }
1589         }
1590
1591         while (nr_running) {
1592                 reap_threads(&nr_running, &t_rate, &m_rate);
1593
1594                 if (is_backend)
1595                         fio_server_idle_loop();
1596                 else
1597                         usleep(10000);
1598         }
1599
1600         update_io_ticks();
1601         fio_unpin_memory();
1602 }
1603
1604 void wait_for_disk_thread_exit(void)
1605 {
1606         fio_mutex_down(disk_thread_mutex);
1607 }
1608
1609 static void *disk_thread_main(void *data)
1610 {
1611         int ret = 0;
1612
1613         fio_mutex_up(startup_mutex);
1614
1615         while (threads && !ret) {
1616                 usleep(DISK_UTIL_MSEC * 1000);
1617                 if (!threads)
1618                         break;
1619                 ret = update_io_ticks();
1620
1621                 if (!is_backend)
1622                         print_thread_status();
1623         }
1624
1625         fio_mutex_up(disk_thread_mutex);
1626         return NULL;
1627 }
1628
1629 static int create_disk_util_thread(void)
1630 {
1631         int ret;
1632
1633         setup_disk_util();
1634
1635         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1636
1637         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1638         if (ret) {
1639                 fio_mutex_remove(disk_thread_mutex);
1640                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1641                 return 1;
1642         }
1643
1644         ret = pthread_detach(disk_util_thread);
1645         if (ret) {
1646                 fio_mutex_remove(disk_thread_mutex);
1647                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1648                 return 1;
1649         }
1650
1651         dprint(FD_MUTEX, "wait on startup_mutex\n");
1652         fio_mutex_down(startup_mutex);
1653         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1654         return 0;
1655 }
1656
1657 int fio_backend(void)
1658 {
1659         struct thread_data *td;
1660         int i;
1661
1662         if (exec_profile) {
1663                 if (load_profile(exec_profile))
1664                         return 1;
1665                 free(exec_profile);
1666                 exec_profile = NULL;
1667         }
1668         if (!thread_number)
1669                 return 0;
1670
1671         if (write_bw_log) {
1672                 setup_log(&agg_io_log[DDIR_READ], 0);
1673                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1674                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1675         }
1676
1677         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1678         if (startup_mutex == NULL)
1679                 return 1;
1680         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1681         if (writeout_mutex == NULL)
1682                 return 1;
1683
1684         set_genesis_time();
1685         create_disk_util_thread();
1686
1687         cgroup_list = smalloc(sizeof(*cgroup_list));
1688         INIT_FLIST_HEAD(cgroup_list);
1689
1690         run_threads();
1691
1692         if (!fio_abort) {
1693                 show_run_stats();
1694                 if (write_bw_log) {
1695                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1696                         __finish_log(agg_io_log[DDIR_WRITE],
1697                                         "agg-write_bw.log");
1698                         __finish_log(agg_io_log[DDIR_TRIM],
1699                                         "agg-write_bw.log");
1700                 }
1701         }
1702
1703         for_each_td(td, i)
1704                 fio_options_free(td);
1705
1706         free_disk_util();
1707         cgroup_kill(cgroup_list);
1708         sfree(cgroup_list);
1709         sfree(cgroup_mnt);
1710
1711         fio_mutex_remove(startup_mutex);
1712         fio_mutex_remove(writeout_mutex);
1713         fio_mutex_remove(disk_thread_mutex);
1714         return exit_value;
1715 }