backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/shm.h>
  38 #include <sys/mman.h>
  39
  40 #include "fio.h"
  41 #include "hash.h"
  42 #include "smalloc.h"
  43 #include "verify.h"
  44 #include "trim.h"
  45 #include "diskutil.h"
  46 #include "cgroup.h"
  47 #include "profile.h"
  48 #include "lib/rand.h"
  49 #include "memalign.h"
  50 #include "server.h"
  51
  52 static pthread_t disk_util_thread;
  53 static struct fio_mutex *startup_mutex;
  54 static struct fio_mutex *writeout_mutex;
  55 static struct flist_head *cgroup_list;
  56 static char *cgroup_mnt;
  57 static int exit_value;
  58 static volatile int fio_abort;
  59
  60 struct io_log *agg_io_log[2];
  61
  62 int groupid = 0;
  63 unsigned int thread_number = 0;
  64 unsigned int nr_process = 0;
  65 unsigned int nr_thread = 0;
  66 int shm_id = 0;
  67 int temp_stall_ts;
  68 unsigned long done_secs = 0;
  69
  70 #define PAGE_ALIGN(buf) \
  71         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  72
  73 #define JOB_START_TIMEOUT       (5 * 1000)
  74
  75 static void sig_int(int sig)
  76 {
  77         if (threads) {
  78                 if (is_backend)
  79                         fio_server_got_signal(sig);
  80                 else {
  81                         log_info("\nfio: terminating on signal %d\n", sig);
  82                         fflush(stdout);
  83                         exit_value = 128;
  84                 }
  85
  86                 fio_terminate_threads(TERMINATE_ALL);
  87         }
  88 }
  89
  90 static void set_sig_handlers(void)
  91 {
  92         struct sigaction act;
  93
  94         memset(&act, 0, sizeof(act));
  95         act.sa_handler = sig_int;
  96         act.sa_flags = SA_RESTART;
  97         sigaction(SIGINT, &act, NULL);
  98
  99         memset(&act, 0, sizeof(act));
 100         act.sa_handler = sig_int;
 101         act.sa_flags = SA_RESTART;
 102         sigaction(SIGTERM, &act, NULL);
 103
 104         if (is_backend) {
 105                 memset(&act, 0, sizeof(act));
 106                 act.sa_handler = sig_int;
 107                 act.sa_flags = SA_RESTART;
 108                 sigaction(SIGPIPE, &act, NULL);
 109         }
 110 }
 111
 112 /*
 113  * Check if we are above the minimum rate given.
 114  */
 115 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 116                             enum fio_ddir ddir)
 117 {
 118         unsigned long long bytes = 0;
 119         unsigned long iops = 0;
 120         unsigned long spent;
 121         unsigned long rate;
 122         unsigned int ratemin = 0;
 123         unsigned int rate_iops = 0;
 124         unsigned int rate_iops_min = 0;
 125
 126         assert(ddir_rw(ddir));
 127
 128         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 129                 return 0;
 130
 131         /*
 132          * allow a 2 second settle period in the beginning
 133          */
 134         if (mtime_since(&td->start, now) < 2000)
 135                 return 0;
 136
 137         iops += td->this_io_blocks[ddir];
 138         bytes += td->this_io_bytes[ddir];
 139         ratemin += td->o.ratemin[ddir];
 140         rate_iops += td->o.rate_iops[ddir];
 141         rate_iops_min += td->o.rate_iops_min[ddir];
 142
 143         /*
 144          * if rate blocks is set, sample is running
 145          */
 146         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 147                 spent = mtime_since(&td->lastrate[ddir], now);
 148                 if (spent < td->o.ratecycle)
 149                         return 0;
 150
 151                 if (td->o.rate[ddir]) {
 152                         /*
 153                          * check bandwidth specified rate
 154                          */
 155                         if (bytes < td->rate_bytes[ddir]) {
 156                                 log_err("%s: min rate %u not met\n", td->o.name,
 157                                                                 ratemin);
 158                                 return 1;
 159                         } else {
 160                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 161                                 if (rate < ratemin ||
 162                                     bytes < td->rate_bytes[ddir]) {
 163                                         log_err("%s: min rate %u not met, got"
 164                                                 " %luKB/sec\n", td->o.name,
 165                                                         ratemin, rate);
 166                                         return 1;
 167                                 }
 168                         }
 169                 } else {
 170                         /*
 171                          * checks iops specified rate
 172                          */
 173                         if (iops < rate_iops) {
 174                                 log_err("%s: min iops rate %u not met\n",
 175                                                 td->o.name, rate_iops);
 176                                 return 1;
 177                         } else {
 178                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 179                                 if (rate < rate_iops_min ||
 180                                     iops < td->rate_blocks[ddir]) {
 181                                         log_err("%s: min iops rate %u not met,"
 182                                                 " got %lu\n", td->o.name,
 183                                                         rate_iops_min, rate);
 184                                 }
 185                         }
 186                 }
 187         }
 188
 189         td->rate_bytes[ddir] = bytes;
 190         td->rate_blocks[ddir] = iops;
 191         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 192         return 0;
 193 }
 194
 195 static int check_min_rate(struct thread_data *td, struct timeval *now,
 196                           unsigned long *bytes_done)
 197 {
 198         int ret = 0;
 199
 200         if (bytes_done[0])
 201                 ret |= __check_min_rate(td, now, 0);
 202         if (bytes_done[1])
 203                 ret |= __check_min_rate(td, now, 1);
 204
 205         return ret;
 206 }
 207
 208 /*
 209  * When job exits, we can cancel the in-flight IO if we are using async
 210  * io. Attempt to do so.
 211  */
 212 static void cleanup_pending_aio(struct thread_data *td)
 213 {
 214         struct flist_head *entry, *n;
 215         struct io_u *io_u;
 216         int r;
 217
 218         /*
 219          * get immediately available events, if any
 220          */
 221         r = io_u_queued_complete(td, 0, NULL);
 222         if (r < 0)
 223                 return;
 224
 225         /*
 226          * now cancel remaining active events
 227          */
 228         if (td->io_ops->cancel) {
 229                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 230                         io_u = flist_entry(entry, struct io_u, list);
 231
 232                         /*
 233                          * if the io_u isn't in flight, then that generally
 234                          * means someone leaked an io_u. complain but fix
 235                          * it up, so we don't stall here.
 236                          */
 237                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 238                                 log_err("fio: non-busy IO on busy list\n");
 239                                 put_io_u(td, io_u);
 240                         } else {
 241                                 r = td->io_ops->cancel(td, io_u);
 242                                 if (!r)
 243                                         put_io_u(td, io_u);
 244                         }
 245                 }
 246         }
 247
 248         if (td->cur_depth)
 249                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 250 }
 251
 252 /*
 253  * Helper to handle the final sync of a file. Works just like the normal
 254  * io path, just does everything sync.
 255  */
 256 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 257 {
 258         struct io_u *io_u = __get_io_u(td);
 259         int ret;
 260
 261         if (!io_u)
 262                 return 1;
 263
 264         io_u->ddir = DDIR_SYNC;
 265         io_u->file = f;
 266
 267         if (td_io_prep(td, io_u)) {
 268                 put_io_u(td, io_u);
 269                 return 1;
 270         }
 271
 272 requeue:
 273         ret = td_io_queue(td, io_u);
 274         if (ret < 0) {
 275                 td_verror(td, io_u->error, "td_io_queue");
 276                 put_io_u(td, io_u);
 277                 return 1;
 278         } else if (ret == FIO_Q_QUEUED) {
 279                 if (io_u_queued_complete(td, 1, NULL) < 0)
 280                         return 1;
 281         } else if (ret == FIO_Q_COMPLETED) {
 282                 if (io_u->error) {
 283                         td_verror(td, io_u->error, "td_io_queue");
 284                         return 1;
 285                 }
 286
 287                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 288                         return 1;
 289         } else if (ret == FIO_Q_BUSY) {
 290                 if (td_io_commit(td))
 291                         return 1;
 292                 goto requeue;
 293         }
 294
 295         return 0;
 296 }
 297
 298 static inline void __update_tv_cache(struct thread_data *td)
 299 {
 300         fio_gettime(&td->tv_cache, NULL);
 301 }
 302
 303 static inline void update_tv_cache(struct thread_data *td)
 304 {
 305         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 306                 __update_tv_cache(td);
 307 }
 308
 309 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 310 {
 311         if (in_ramp_time(td))
 312                 return 0;
 313         if (!td->o.timeout)
 314                 return 0;
 315         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 316                 return 1;
 317
 318         return 0;
 319 }
 320
 321 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 322                                int *retptr)
 323 {
 324         int ret = *retptr;
 325
 326         if (ret < 0 || td->error) {
 327                 int err;
 328
 329                 if (ret < 0)
 330                         err = -ret;
 331                 else
 332                         err = td->error;
 333
 334                 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
 335                         return 1;
 336
 337                 if (td_non_fatal_error(err)) {
 338                         /*
 339                          * Continue with the I/Os in case of
 340                          * a non fatal error.
 341                          */
 342                         update_error_count(td, err);
 343                         td_clear_error(td);
 344                         *retptr = 0;
 345                         return 0;
 346                 } else if (td->o.fill_device && err == ENOSPC) {
 347                         /*
 348                          * We expect to hit this error if
 349                          * fill_device option is set.
 350                          */
 351                         td_clear_error(td);
 352                         td->terminate = 1;
 353                         return 1;
 354                 } else {
 355                         /*
 356                          * Stop the I/O in case of a fatal
 357                          * error.
 358                          */
 359                         update_error_count(td, err);
 360                         return 1;
 361                 }
 362         }
 363
 364         return 0;
 365 }
 366
 367 /*
 368  * The main verify engine. Runs over the writes we previously submitted,
 369  * reads the blocks back in, and checks the crc/md5 of the data.
 370  */
 371 static void do_verify(struct thread_data *td)
 372 {
 373         struct fio_file *f;
 374         struct io_u *io_u;
 375         int ret, min_events;
 376         unsigned int i;
 377
 378         dprint(FD_VERIFY, "starting loop\n");
 379
 380         /*
 381          * sync io first and invalidate cache, to make sure we really
 382          * read from disk.
 383          */
 384         for_each_file(td, f, i) {
 385                 if (!fio_file_open(f))
 386                         continue;
 387                 if (fio_io_sync(td, f))
 388                         break;
 389                 if (file_invalidate_cache(td, f))
 390                         break;
 391         }
 392
 393         if (td->error)
 394                 return;
 395
 396         td_set_runstate(td, TD_VERIFYING);
 397
 398         io_u = NULL;
 399         while (!td->terminate) {
 400                 int ret2, full;
 401
 402                 update_tv_cache(td);
 403
 404                 if (runtime_exceeded(td, &td->tv_cache)) {
 405                         __update_tv_cache(td);
 406                         if (runtime_exceeded(td, &td->tv_cache)) {
 407                                 td->terminate = 1;
 408                                 break;
 409                         }
 410                 }
 411
 412                 if (flow_threshold_exceeded(td))
 413                         continue;
 414
 415                 io_u = __get_io_u(td);
 416                 if (!io_u)
 417                         break;
 418
 419                 if (get_next_verify(td, io_u)) {
 420                         put_io_u(td, io_u);
 421                         break;
 422                 }
 423
 424                 if (td_io_prep(td, io_u)) {
 425                         put_io_u(td, io_u);
 426                         break;
 427                 }
 428
 429                 if (td->o.verify_async)
 430                         io_u->end_io = verify_io_u_async;
 431                 else
 432                         io_u->end_io = verify_io_u;
 433
 434                 ret = td_io_queue(td, io_u);
 435                 switch (ret) {
 436                 case FIO_Q_COMPLETED:
 437                         if (io_u->error) {
 438                                 ret = -io_u->error;
 439                                 clear_io_u(td, io_u);
 440                         } else if (io_u->resid) {
 441                                 int bytes = io_u->xfer_buflen - io_u->resid;
 442
 443                                 /*
 444                                  * zero read, fail
 445                                  */
 446                                 if (!bytes) {
 447                                         td_verror(td, EIO, "full resid");
 448                                         put_io_u(td, io_u);
 449                                         break;
 450                                 }
 451
 452                                 io_u->xfer_buflen = io_u->resid;
 453                                 io_u->xfer_buf += bytes;
 454                                 io_u->offset += bytes;
 455
 456                                 if (ddir_rw(io_u->ddir))
 457                                         td->ts.short_io_u[io_u->ddir]++;
 458
 459                                 f = io_u->file;
 460                                 if (io_u->offset == f->real_file_size)
 461                                         goto sync_done;
 462
 463                                 requeue_io_u(td, &io_u);
 464                         } else {
 465 sync_done:
 466                                 ret = io_u_sync_complete(td, io_u, NULL);
 467                                 if (ret < 0)
 468                                         break;
 469                         }
 470                         continue;
 471                 case FIO_Q_QUEUED:
 472                         break;
 473                 case FIO_Q_BUSY:
 474                         requeue_io_u(td, &io_u);
 475                         ret2 = td_io_commit(td);
 476                         if (ret2 < 0)
 477                                 ret = ret2;
 478                         break;
 479                 default:
 480                         assert(ret < 0);
 481                         td_verror(td, -ret, "td_io_queue");
 482                         break;
 483                 }
 484
 485                 if (break_on_this_error(td, io_u->ddir, &ret))
 486                         break;
 487
 488                 /*
 489                  * if we can queue more, do so. but check if there are
 490                  * completed io_u's first. Note that we can get BUSY even
 491                  * without IO queued, if the system is resource starved.
 492                  */
 493                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 494                 if (full || !td->o.iodepth_batch_complete) {
 495                         min_events = min(td->o.iodepth_batch_complete,
 496                                          td->cur_depth);
 497                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 498                                 min_events = 1;
 499
 500                         do {
 501                                 /*
 502                                  * Reap required number of io units, if any,
 503                                  * and do the verification on them through
 504                                  * the callback handler
 505                                  */
 506                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 507                                         ret = -1;
 508                                         break;
 509                                 }
 510                         } while (full && (td->cur_depth > td->o.iodepth_low));
 511                 }
 512                 if (ret < 0)
 513                         break;
 514         }
 515
 516         if (!td->error) {
 517                 min_events = td->cur_depth;
 518
 519                 if (min_events)
 520                         ret = io_u_queued_complete(td, min_events, NULL);
 521         } else
 522                 cleanup_pending_aio(td);
 523
 524         td_set_runstate(td, TD_RUNNING);
 525
 526         dprint(FD_VERIFY, "exiting loop\n");
 527 }
 528
 529 static int io_bytes_exceeded(struct thread_data *td)
 530 {
 531         unsigned long long bytes;
 532
 533         if (td_rw(td))
 534                 bytes = td->this_io_bytes[0] + td->this_io_bytes[1];
 535         else if (td_write(td))
 536                 bytes = td->this_io_bytes[1];
 537         else
 538                 bytes = td->this_io_bytes[0];
 539
 540         return bytes >= td->o.size;
 541 }
 542
 543 /*
 544  * Main IO worker function. It retrieves io_u's to process and queues
 545  * and reaps them, checking for rate and errors along the way.
 546  */
 547 static void do_io(struct thread_data *td)
 548 {
 549         unsigned int i;
 550         int ret = 0;
 551
 552         if (in_ramp_time(td))
 553                 td_set_runstate(td, TD_RAMP);
 554         else
 555                 td_set_runstate(td, TD_RUNNING);
 556
 557         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 558                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
 559                 struct timeval comp_time;
 560                 unsigned long bytes_done[2] = { 0, 0 };
 561                 int min_evts = 0;
 562                 struct io_u *io_u;
 563                 int ret2, full;
 564                 enum fio_ddir ddir;
 565
 566                 if (td->terminate)
 567                         break;
 568
 569                 update_tv_cache(td);
 570
 571                 if (runtime_exceeded(td, &td->tv_cache)) {
 572                         __update_tv_cache(td);
 573                         if (runtime_exceeded(td, &td->tv_cache)) {
 574                                 td->terminate = 1;
 575                                 break;
 576                         }
 577                 }
 578
 579                 if (flow_threshold_exceeded(td))
 580                         continue;
 581
 582                 io_u = get_io_u(td);
 583                 if (!io_u)
 584                         break;
 585
 586                 ddir = io_u->ddir;
 587
 588                 /*
 589                  * Add verification end_io handler, if asked to verify
 590                  * a previously written file.
 591                  */
 592                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 593                     !td_rw(td)) {
 594                         if (td->o.verify_async)
 595                                 io_u->end_io = verify_io_u_async;
 596                         else
 597                                 io_u->end_io = verify_io_u;
 598                         td_set_runstate(td, TD_VERIFYING);
 599                 } else if (in_ramp_time(td))
 600                         td_set_runstate(td, TD_RAMP);
 601                 else
 602                         td_set_runstate(td, TD_RUNNING);
 603
 604                 ret = td_io_queue(td, io_u);
 605                 switch (ret) {
 606                 case FIO_Q_COMPLETED:
 607                         if (io_u->error) {
 608                                 ret = -io_u->error;
 609                                 clear_io_u(td, io_u);
 610                         } else if (io_u->resid) {
 611                                 int bytes = io_u->xfer_buflen - io_u->resid;
 612                                 struct fio_file *f = io_u->file;
 613
 614                                 /*
 615                                  * zero read, fail
 616                                  */
 617                                 if (!bytes) {
 618                                         td_verror(td, EIO, "full resid");
 619                                         put_io_u(td, io_u);
 620                                         break;
 621                                 }
 622
 623                                 io_u->xfer_buflen = io_u->resid;
 624                                 io_u->xfer_buf += bytes;
 625                                 io_u->offset += bytes;
 626
 627                                 if (ddir_rw(io_u->ddir))
 628                                         td->ts.short_io_u[io_u->ddir]++;
 629
 630                                 if (io_u->offset == f->real_file_size)
 631                                         goto sync_done;
 632
 633                                 requeue_io_u(td, &io_u);
 634                         } else {
 635 sync_done:
 636                                 if (__should_check_rate(td, 0) ||
 637                                     __should_check_rate(td, 1))
 638                                         fio_gettime(&comp_time, NULL);
 639
 640                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 641                                 if (ret < 0)
 642                                         break;
 643                         }
 644                         break;
 645                 case FIO_Q_QUEUED:
 646                         /*
 647                          * if the engine doesn't have a commit hook,
 648                          * the io_u is really queued. if it does have such
 649                          * a hook, it has to call io_u_queued() itself.
 650                          */
 651                         if (td->io_ops->commit == NULL)
 652                                 io_u_queued(td, io_u);
 653                         break;
 654                 case FIO_Q_BUSY:
 655                         requeue_io_u(td, &io_u);
 656                         ret2 = td_io_commit(td);
 657                         if (ret2 < 0)
 658                                 ret = ret2;
 659                         break;
 660                 default:
 661                         assert(ret < 0);
 662                         put_io_u(td, io_u);
 663                         break;
 664                 }
 665
 666                 if (break_on_this_error(td, ddir, &ret))
 667                         break;
 668
 669                 /*
 670                  * See if we need to complete some commands. Note that we
 671                  * can get BUSY even without IO queued, if the system is
 672                  * resource starved.
 673                  */
 674                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 675                 if (full || !td->o.iodepth_batch_complete) {
 676                         min_evts = min(td->o.iodepth_batch_complete,
 677                                         td->cur_depth);
 678                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 679                                 min_evts = 1;
 680
 681                         if (__should_check_rate(td, 0) ||
 682                             __should_check_rate(td, 1))
 683                                 fio_gettime(&comp_time, NULL);
 684
 685                         do {
 686                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 687                                 if (ret < 0)
 688                                         break;
 689
 690                         } while (full && (td->cur_depth > td->o.iodepth_low));
 691                 }
 692
 693                 if (ret < 0)
 694                         break;
 695                 if (!(bytes_done[0] + bytes_done[1]))
 696                         continue;
 697
 698                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 699                         if (check_min_rate(td, &comp_time, bytes_done)) {
 700                                 if (exitall_on_terminate)
 701                                         fio_terminate_threads(td->groupid);
 702                                 td_verror(td, EIO, "check_min_rate");
 703                                 break;
 704                         }
 705                 }
 706
 707                 if (td->o.thinktime) {
 708                         unsigned long long b;
 709
 710                         b = td->io_blocks[0] + td->io_blocks[1];
 711                         if (!(b % td->o.thinktime_blocks)) {
 712                                 int left;
 713
 714                                 if (td->o.thinktime_spin)
 715                                         usec_spin(td->o.thinktime_spin);
 716
 717                                 left = td->o.thinktime - td->o.thinktime_spin;
 718                                 if (left)
 719                                         usec_sleep(td, left);
 720                         }
 721                 }
 722         }
 723
 724         if (td->trim_entries)
 725                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 726
 727         if (td->o.fill_device && td->error == ENOSPC) {
 728                 td->error = 0;
 729                 td->terminate = 1;
 730         }
 731         if (!td->error) {
 732                 struct fio_file *f;
 733
 734                 i = td->cur_depth;
 735                 if (i) {
 736                         ret = io_u_queued_complete(td, i, NULL);
 737                         if (td->o.fill_device && td->error == ENOSPC)
 738                                 td->error = 0;
 739                 }
 740
 741                 if (should_fsync(td) && td->o.end_fsync) {
 742                         td_set_runstate(td, TD_FSYNCING);
 743
 744                         for_each_file(td, f, i) {
 745                                 if (!fio_file_open(f))
 746                                         continue;
 747                                 fio_io_sync(td, f);
 748                         }
 749                 }
 750         } else
 751                 cleanup_pending_aio(td);
 752
 753         /*
 754          * stop job if we failed doing any IO
 755          */
 756         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 757                 td->done = 1;
 758 }
 759
 760 static void cleanup_io_u(struct thread_data *td)
 761 {
 762         struct flist_head *entry, *n;
 763         struct io_u *io_u;
 764
 765         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 766                 io_u = flist_entry(entry, struct io_u, list);
 767
 768                 flist_del(&io_u->list);
 769                 fio_memfree(io_u, sizeof(*io_u));
 770         }
 771
 772         free_io_mem(td);
 773 }
 774
 775 static int init_io_u(struct thread_data *td)
 776 {
 777         struct io_u *io_u;
 778         unsigned int max_bs, min_write;
 779         int cl_align, i, max_units;
 780         char *p;
 781
 782         max_units = td->o.iodepth;
 783         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 784         min_write = td->o.min_bs[DDIR_WRITE];
 785         td->orig_buffer_size = (unsigned long long) max_bs
 786                                         * (unsigned long long) max_units;
 787
 788         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 789                 unsigned long bs;
 790
 791                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 792                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 793         }
 794
 795         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 796                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 797                 return 1;
 798         }
 799
 800         if (allocate_io_mem(td))
 801                 return 1;
 802
 803         if (td->o.odirect || td->o.mem_align ||
 804             (td->io_ops->flags & FIO_RAWIO))
 805                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 806         else
 807                 p = td->orig_buffer;
 808
 809         cl_align = os_cache_line_size();
 810
 811         for (i = 0; i < max_units; i++) {
 812                 void *ptr;
 813
 814                 if (td->terminate)
 815                         return 1;
 816
 817                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 818                 if (!ptr) {
 819                         log_err("fio: unable to allocate aligned memory\n");
 820                         break;
 821                 }
 822
 823                 io_u = ptr;
 824                 memset(io_u, 0, sizeof(*io_u));
 825                 INIT_FLIST_HEAD(&io_u->list);
 826                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 827
 828                 if (!(td->io_ops->flags & FIO_NOIO)) {
 829                         io_u->buf = p;
 830                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 831
 832                         if (td_write(td))
 833                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 834                         if (td_write(td) && td->o.verify_pattern_bytes) {
 835                                 /*
 836                                  * Fill the buffer with the pattern if we are
 837                                  * going to be doing writes.
 838                                  */
 839                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 840                         }
 841                 }
 842
 843                 io_u->index = i;
 844                 io_u->flags = IO_U_F_FREE;
 845                 flist_add(&io_u->list, &td->io_u_freelist);
 846                 p += max_bs;
 847         }
 848
 849         return 0;
 850 }
 851
 852 static int switch_ioscheduler(struct thread_data *td)
 853 {
 854         char tmp[256], tmp2[128];
 855         FILE *f;
 856         int ret;
 857
 858         if (td->io_ops->flags & FIO_DISKLESSIO)
 859                 return 0;
 860
 861         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 862
 863         f = fopen(tmp, "r+");
 864         if (!f) {
 865                 if (errno == ENOENT) {
 866                         log_err("fio: os or kernel doesn't support IO scheduler"
 867                                 " switching\n");
 868                         return 0;
 869                 }
 870                 td_verror(td, errno, "fopen iosched");
 871                 return 1;
 872         }
 873
 874         /*
 875          * Set io scheduler.
 876          */
 877         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 878         if (ferror(f) || ret != 1) {
 879                 td_verror(td, errno, "fwrite");
 880                 fclose(f);
 881                 return 1;
 882         }
 883
 884         rewind(f);
 885
 886         /*
 887          * Read back and check that the selected scheduler is now the default.
 888          */
 889         ret = fread(tmp, 1, sizeof(tmp), f);
 890         if (ferror(f) || ret < 0) {
 891                 td_verror(td, errno, "fread");
 892                 fclose(f);
 893                 return 1;
 894         }
 895
 896         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 897         if (!strstr(tmp, tmp2)) {
 898                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 899                 td_verror(td, EINVAL, "iosched_switch");
 900                 fclose(f);
 901                 return 1;
 902         }
 903
 904         fclose(f);
 905         return 0;
 906 }
 907
 908 static int keep_running(struct thread_data *td)
 909 {
 910         unsigned long long io_done;
 911
 912         if (td->done)
 913                 return 0;
 914         if (td->o.time_based)
 915                 return 1;
 916         if (td->o.loops) {
 917                 td->o.loops--;
 918                 return 1;
 919         }
 920
 921         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 922                         + td->io_skip_bytes;
 923         if (io_done < td->o.size)
 924                 return 1;
 925
 926         return 0;
 927 }
 928
 929 static int exec_string(const char *string)
 930 {
 931         int ret, newlen = strlen(string) + 1 + 8;
 932         char *str;
 933
 934         str = malloc(newlen);
 935         sprintf(str, "sh -c %s", string);
 936
 937         ret = system(str);
 938         if (ret == -1)
 939                 log_err("fio: exec of cmd <%s> failed\n", str);
 940
 941         free(str);
 942         return ret;
 943 }
 944
 945 /*
 946  * Entry point for the thread based jobs. The process based jobs end up
 947  * here as well, after a little setup.
 948  */
 949 static void *thread_main(void *data)
 950 {
 951         unsigned long long elapsed;
 952         struct thread_data *td = data;
 953         pthread_condattr_t attr;
 954         int clear_state;
 955
 956         if (!td->o.use_thread) {
 957                 setsid();
 958                 td->pid = getpid();
 959         } else
 960                 td->pid = gettid();
 961
 962         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 963
 964         INIT_FLIST_HEAD(&td->io_u_freelist);
 965         INIT_FLIST_HEAD(&td->io_u_busylist);
 966         INIT_FLIST_HEAD(&td->io_u_requeues);
 967         INIT_FLIST_HEAD(&td->io_log_list);
 968         INIT_FLIST_HEAD(&td->io_hist_list);
 969         INIT_FLIST_HEAD(&td->verify_list);
 970         INIT_FLIST_HEAD(&td->trim_list);
 971         pthread_mutex_init(&td->io_u_lock, NULL);
 972         td->io_hist_tree = RB_ROOT;
 973
 974         pthread_condattr_init(&attr);
 975         pthread_cond_init(&td->verify_cond, &attr);
 976         pthread_cond_init(&td->free_cond, &attr);
 977
 978         td_set_runstate(td, TD_INITIALIZED);
 979         dprint(FD_MUTEX, "up startup_mutex\n");
 980         fio_mutex_up(startup_mutex);
 981         dprint(FD_MUTEX, "wait on td->mutex\n");
 982         fio_mutex_down(td->mutex);
 983         dprint(FD_MUTEX, "done waiting on td->mutex\n");
 984
 985         /*
 986          * the ->mutex mutex is now no longer used, close it to avoid
 987          * eating a file descriptor
 988          */
 989         fio_mutex_remove(td->mutex);
 990
 991         /*
 992          * A new gid requires privilege, so we need to do this before setting
 993          * the uid.
 994          */
 995         if (td->o.gid != -1U && setgid(td->o.gid)) {
 996                 td_verror(td, errno, "setgid");
 997                 goto err;
 998         }
 999         if (td->o.uid != -1U && setuid(td->o.uid)) {
1000                 td_verror(td, errno, "setuid");
1001                 goto err;
1002         }
1003
1004         /*
1005          * If we have a gettimeofday() thread, make sure we exclude that
1006          * thread from this job
1007          */
1008         if (td->o.gtod_cpu)
1009                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1010
1011         /*
1012          * Set affinity first, in case it has an impact on the memory
1013          * allocations.
1014          */
1015         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1016                 td_verror(td, errno, "cpu_set_affinity");
1017                 goto err;
1018         }
1019
1020         /*
1021          * May alter parameters that init_io_u() will use, so we need to
1022          * do this first.
1023          */
1024         if (init_iolog(td))
1025                 goto err;
1026
1027         if (init_io_u(td))
1028                 goto err;
1029
1030         if (td->o.verify_async && verify_async_init(td))
1031                 goto err;
1032
1033         if (td->ioprio_set) {
1034                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1035                         td_verror(td, errno, "ioprio_set");
1036                         goto err;
1037                 }
1038         }
1039
1040         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1041                 goto err;
1042
1043         errno = 0;
1044         if (nice(td->o.nice) == -1 && errno != 0) {
1045                 td_verror(td, errno, "nice");
1046                 goto err;
1047         }
1048
1049         if (td->o.ioscheduler && switch_ioscheduler(td))
1050                 goto err;
1051
1052         if (!td->o.create_serialize && setup_files(td))
1053                 goto err;
1054
1055         if (td_io_init(td))
1056                 goto err;
1057
1058         if (init_random_map(td))
1059                 goto err;
1060
1061         if (td->o.exec_prerun) {
1062                 if (exec_string(td->o.exec_prerun))
1063                         goto err;
1064         }
1065
1066         if (td->o.pre_read) {
1067                 if (pre_read_files(td) < 0)
1068                         goto err;
1069         }
1070
1071         fio_gettime(&td->epoch, NULL);
1072         getrusage(RUSAGE_SELF, &td->ru_start);
1073
1074         clear_state = 0;
1075         while (keep_running(td)) {
1076                 fio_gettime(&td->start, NULL);
1077                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1078                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1079                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1080
1081                 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1082                         memcpy(&td->lastrate[0], &td->bw_sample_time,
1083                                                 sizeof(td->bw_sample_time));
1084                         memcpy(&td->lastrate[1], &td->bw_sample_time,
1085                                                 sizeof(td->bw_sample_time));
1086                 }
1087
1088                 if (clear_state)
1089                         clear_io_state(td);
1090
1091                 prune_io_piece_log(td);
1092
1093                 do_io(td);
1094
1095                 clear_state = 1;
1096
1097                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1098                         elapsed = utime_since_now(&td->start);
1099                         td->ts.runtime[DDIR_READ] += elapsed;
1100                 }
1101                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1102                         elapsed = utime_since_now(&td->start);
1103                         td->ts.runtime[DDIR_WRITE] += elapsed;
1104                 }
1105
1106                 if (td->error || td->terminate)
1107                         break;
1108
1109                 if (!td->o.do_verify ||
1110                     td->o.verify == VERIFY_NONE ||
1111                     (td->io_ops->flags & FIO_UNIDIR))
1112                         continue;
1113
1114                 clear_io_state(td);
1115
1116                 fio_gettime(&td->start, NULL);
1117
1118                 do_verify(td);
1119
1120                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1121
1122                 if (td->error || td->terminate)
1123                         break;
1124         }
1125
1126         update_rusage_stat(td);
1127         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1128         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1129         td->ts.total_run_time = mtime_since_now(&td->epoch);
1130         td->ts.io_bytes[0] = td->io_bytes[0];
1131         td->ts.io_bytes[1] = td->io_bytes[1];
1132
1133         fio_mutex_down(writeout_mutex);
1134         if (td->bw_log) {
1135                 if (td->o.bw_log_file) {
1136                         finish_log_named(td, td->bw_log,
1137                                                 td->o.bw_log_file, "bw");
1138                 } else
1139                         finish_log(td, td->bw_log, "bw");
1140         }
1141         if (td->lat_log) {
1142                 if (td->o.lat_log_file) {
1143                         finish_log_named(td, td->lat_log,
1144                                                 td->o.lat_log_file, "lat");
1145                 } else
1146                         finish_log(td, td->lat_log, "lat");
1147         }
1148         if (td->slat_log) {
1149                 if (td->o.lat_log_file) {
1150                         finish_log_named(td, td->slat_log,
1151                                                 td->o.lat_log_file, "slat");
1152                 } else
1153                         finish_log(td, td->slat_log, "slat");
1154         }
1155         if (td->clat_log) {
1156                 if (td->o.lat_log_file) {
1157                         finish_log_named(td, td->clat_log,
1158                                                 td->o.lat_log_file, "clat");
1159                 } else
1160                         finish_log(td, td->clat_log, "clat");
1161         }
1162         if (td->iops_log) {
1163                 if (td->o.iops_log_file) {
1164                         finish_log_named(td, td->iops_log,
1165                                                 td->o.iops_log_file, "iops");
1166                 } else
1167                         finish_log(td, td->iops_log, "iops");
1168         }
1169
1170         fio_mutex_up(writeout_mutex);
1171         if (td->o.exec_postrun)
1172                 exec_string(td->o.exec_postrun);
1173
1174         if (exitall_on_terminate)
1175                 fio_terminate_threads(td->groupid);
1176
1177 err:
1178         if (td->error)
1179                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1180                                                         td->verror);
1181
1182         if (td->o.verify_async)
1183                 verify_async_exit(td);
1184
1185         close_and_free_files(td);
1186         close_ioengine(td);
1187         cleanup_io_u(td);
1188         cgroup_shutdown(td, &cgroup_mnt);
1189
1190         if (td->o.cpumask_set) {
1191                 int ret = fio_cpuset_exit(&td->o.cpumask);
1192
1193                 td_verror(td, ret, "fio_cpuset_exit");
1194         }
1195
1196         /*
1197          * do this very late, it will log file closing as well
1198          */
1199         if (td->o.write_iolog_file)
1200                 write_iolog_close(td);
1201
1202         td_set_runstate(td, TD_EXITED);
1203         return (void *) (uintptr_t) td->error;
1204 }
1205
1206
1207 /*
1208  * We cannot pass the td data into a forked process, so attach the td and
1209  * pass it to the thread worker.
1210  */
1211 static int fork_main(int shmid, int offset)
1212 {
1213         struct thread_data *td;
1214         void *data, *ret;
1215
1216 #ifndef __hpux
1217         data = shmat(shmid, NULL, 0);
1218         if (data == (void *) -1) {
1219                 int __err = errno;
1220
1221                 perror("shmat");
1222                 return __err;
1223         }
1224 #else
1225         /*
1226          * HP-UX inherits shm mappings?
1227          */
1228         data = threads;
1229 #endif
1230
1231         td = data + offset * sizeof(struct thread_data);
1232         ret = thread_main(td);
1233         shmdt(data);
1234         return (int) (uintptr_t) ret;
1235 }
1236
1237 /*
1238  * Run over the job map and reap the threads that have exited, if any.
1239  */
1240 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1241                          unsigned int *m_rate)
1242 {
1243         struct thread_data *td;
1244         unsigned int cputhreads, realthreads, pending;
1245         int i, status, ret;
1246
1247         /*
1248          * reap exited threads (TD_EXITED -> TD_REAPED)
1249          */
1250         realthreads = pending = cputhreads = 0;
1251         for_each_td(td, i) {
1252                 int flags = 0;
1253
1254                 /*
1255                  * ->io_ops is NULL for a thread that has closed its
1256                  * io engine
1257                  */
1258                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1259                         cputhreads++;
1260                 else
1261                         realthreads++;
1262
1263                 if (!td->pid) {
1264                         pending++;
1265                         continue;
1266                 }
1267                 if (td->runstate == TD_REAPED)
1268                         continue;
1269                 if (td->o.use_thread) {
1270                         if (td->runstate == TD_EXITED) {
1271                                 td_set_runstate(td, TD_REAPED);
1272                                 goto reaped;
1273                         }
1274                         continue;
1275                 }
1276
1277                 flags = WNOHANG;
1278                 if (td->runstate == TD_EXITED)
1279                         flags = 0;
1280
1281                 /*
1282                  * check if someone quit or got killed in an unusual way
1283                  */
1284                 ret = waitpid(td->pid, &status, flags);
1285                 if (ret < 0) {
1286                         if (errno == ECHILD) {
1287                                 log_err("fio: pid=%d disappeared %d\n",
1288                                                 (int) td->pid, td->runstate);
1289                                 td_set_runstate(td, TD_REAPED);
1290                                 goto reaped;
1291                         }
1292                         perror("waitpid");
1293                 } else if (ret == td->pid) {
1294                         if (WIFSIGNALED(status)) {
1295                                 int sig = WTERMSIG(status);
1296
1297                                 if (sig != SIGTERM)
1298                                         log_err("fio: pid=%d, got signal=%d\n",
1299                                                         (int) td->pid, sig);
1300                                 td_set_runstate(td, TD_REAPED);
1301                                 goto reaped;
1302                         }
1303                         if (WIFEXITED(status)) {
1304                                 if (WEXITSTATUS(status) && !td->error)
1305                                         td->error = WEXITSTATUS(status);
1306
1307                                 td_set_runstate(td, TD_REAPED);
1308                                 goto reaped;
1309                         }
1310                 }
1311
1312                 /*
1313                  * thread is not dead, continue
1314                  */
1315                 pending++;
1316                 continue;
1317 reaped:
1318                 (*nr_running)--;
1319                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1320                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1321                 if (!td->pid)
1322                         pending--;
1323
1324                 if (td->error)
1325                         exit_value++;
1326
1327                 done_secs += mtime_since_now(&td->epoch) / 1000;
1328         }
1329
1330         if (*nr_running == cputhreads && !pending && realthreads)
1331                 fio_terminate_threads(TERMINATE_ALL);
1332 }
1333
1334 /*
1335  * Main function for kicking off and reaping jobs, as needed.
1336  */
1337 static void run_threads(void)
1338 {
1339         struct thread_data *td;
1340         unsigned long spent;
1341         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1342
1343         if (fio_pin_memory())
1344                 return;
1345
1346         if (fio_gtod_offload && fio_start_gtod_thread())
1347                 return;
1348
1349         set_sig_handlers();
1350
1351         if (!terse_output) {
1352                 log_info("Starting ");
1353                 if (nr_thread)
1354                         log_info("%d thread%s", nr_thread,
1355                                                 nr_thread > 1 ? "s" : "");
1356                 if (nr_process) {
1357                         if (nr_thread)
1358                                 log_info(" and ");
1359                         log_info("%d process%s", nr_process,
1360                                                 nr_process > 1 ? "es" : "");
1361                 }
1362                 log_info("\n");
1363                 fflush(stdout);
1364         }
1365
1366         todo = thread_number;
1367         nr_running = 0;
1368         nr_started = 0;
1369         m_rate = t_rate = 0;
1370
1371         for_each_td(td, i) {
1372                 print_status_init(td->thread_number - 1);
1373
1374                 if (!td->o.create_serialize)
1375                         continue;
1376
1377                 /*
1378                  * do file setup here so it happens sequentially,
1379                  * we don't want X number of threads getting their
1380                  * client data interspersed on disk
1381                  */
1382                 if (setup_files(td)) {
1383                         exit_value++;
1384                         if (td->error)
1385                                 log_err("fio: pid=%d, err=%d/%s\n",
1386                                         (int) td->pid, td->error, td->verror);
1387                         td_set_runstate(td, TD_REAPED);
1388                         todo--;
1389                 } else {
1390                         struct fio_file *f;
1391                         unsigned int j;
1392
1393                         /*
1394                          * for sharing to work, each job must always open
1395                          * its own files. so close them, if we opened them
1396                          * for creation
1397                          */
1398                         for_each_file(td, f, j) {
1399                                 if (fio_file_open(f))
1400                                         td_io_close_file(td, f);
1401                         }
1402                 }
1403         }
1404
1405         set_genesis_time();
1406
1407         while (todo) {
1408                 struct thread_data *map[REAL_MAX_JOBS];
1409                 struct timeval this_start;
1410                 int this_jobs = 0, left;
1411
1412                 /*
1413                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1414                  */
1415                 for_each_td(td, i) {
1416                         if (td->runstate != TD_NOT_CREATED)
1417                                 continue;
1418
1419                         /*
1420                          * never got a chance to start, killed by other
1421                          * thread for some reason
1422                          */
1423                         if (td->terminate) {
1424                                 todo--;
1425                                 continue;
1426                         }
1427
1428                         if (td->o.start_delay) {
1429                                 spent = mtime_since_genesis();
1430
1431                                 if (td->o.start_delay * 1000 > spent)
1432                                         continue;
1433                         }
1434
1435                         if (td->o.stonewall && (nr_started || nr_running)) {
1436                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1437                                                         td->o.name);
1438                                 break;
1439                         }
1440
1441                         init_disk_util(td);
1442
1443                         /*
1444                          * Set state to created. Thread will transition
1445                          * to TD_INITIALIZED when it's done setting up.
1446                          */
1447                         td_set_runstate(td, TD_CREATED);
1448                         map[this_jobs++] = td;
1449                         nr_started++;
1450
1451                         if (td->o.use_thread) {
1452                                 int ret;
1453
1454                                 dprint(FD_PROCESS, "will pthread_create\n");
1455                                 ret = pthread_create(&td->thread, NULL,
1456                                                         thread_main, td);
1457                                 if (ret) {
1458                                         log_err("pthread_create: %s\n",
1459                                                         strerror(ret));
1460                                         nr_started--;
1461                                         break;
1462                                 }
1463                                 ret = pthread_detach(td->thread);
1464                                 if (ret)
1465                                         log_err("pthread_detach: %s",
1466                                                         strerror(ret));
1467                         } else {
1468                                 pid_t pid;
1469                                 dprint(FD_PROCESS, "will fork\n");
1470                                 pid = fork();
1471                                 if (!pid) {
1472                                         int ret = fork_main(shm_id, i);
1473
1474                                         _exit(ret);
1475                                 } else if (i == fio_debug_jobno)
1476                                         *fio_debug_jobp = pid;
1477                         }
1478                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1479                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1480                                 log_err("fio: job startup hung? exiting.\n");
1481                                 fio_terminate_threads(TERMINATE_ALL);
1482                                 fio_abort = 1;
1483                                 nr_started--;
1484                                 break;
1485                         }
1486                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1487                 }
1488
1489                 /*
1490                  * Wait for the started threads to transition to
1491                  * TD_INITIALIZED.
1492                  */
1493                 fio_gettime(&this_start, NULL);
1494                 left = this_jobs;
1495                 while (left && !fio_abort) {
1496                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1497                                 break;
1498
1499                         usleep(100000);
1500
1501                         for (i = 0; i < this_jobs; i++) {
1502                                 td = map[i];
1503                                 if (!td)
1504                                         continue;
1505                                 if (td->runstate == TD_INITIALIZED) {
1506                                         map[i] = NULL;
1507                                         left--;
1508                                 } else if (td->runstate >= TD_EXITED) {
1509                                         map[i] = NULL;
1510                                         left--;
1511                                         todo--;
1512                                         nr_running++; /* work-around... */
1513                                 }
1514                         }
1515                 }
1516
1517                 if (left) {
1518                         log_err("fio: %d job%s failed to start\n", left,
1519                                         left > 1 ? "s" : "");
1520                         for (i = 0; i < this_jobs; i++) {
1521                                 td = map[i];
1522                                 if (!td)
1523                                         continue;
1524                                 kill(td->pid, SIGTERM);
1525                         }
1526                         break;
1527                 }
1528
1529                 /*
1530                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1531                  */
1532                 for_each_td(td, i) {
1533                         if (td->runstate != TD_INITIALIZED)
1534                                 continue;
1535
1536                         if (in_ramp_time(td))
1537                                 td_set_runstate(td, TD_RAMP);
1538                         else
1539                                 td_set_runstate(td, TD_RUNNING);
1540                         nr_running++;
1541                         nr_started--;
1542                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1543                         t_rate += td->o.rate[0] + td->o.rate[1];
1544                         todo--;
1545                         fio_mutex_up(td->mutex);
1546                 }
1547
1548                 reap_threads(&nr_running, &t_rate, &m_rate);
1549
1550                 if (todo) {
1551                         if (is_backend)
1552                                 fio_server_idle_loop();
1553                         else
1554                                 usleep(100000);
1555                 }
1556         }
1557
1558         while (nr_running) {
1559                 reap_threads(&nr_running, &t_rate, &m_rate);
1560
1561                 if (is_backend)
1562                         fio_server_idle_loop();
1563                 else
1564                         usleep(10000);
1565         }
1566
1567         update_io_ticks();
1568         fio_unpin_memory();
1569 }
1570
1571 static void *disk_thread_main(void *data)
1572 {
1573         fio_mutex_up(startup_mutex);
1574
1575         while (threads) {
1576                 usleep(DISK_UTIL_MSEC * 1000);
1577                 if (!threads)
1578                         break;
1579                 update_io_ticks();
1580
1581                 if (!is_backend)
1582                         print_thread_status();
1583         }
1584
1585         return NULL;
1586 }
1587
1588 static int create_disk_util_thread(void)
1589 {
1590         int ret;
1591
1592         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1593         if (ret) {
1594                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1595                 return 1;
1596         }
1597
1598         ret = pthread_detach(disk_util_thread);
1599         if (ret) {
1600                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1601                 return 1;
1602         }
1603
1604         dprint(FD_MUTEX, "wait on startup_mutex\n");
1605         fio_mutex_down(startup_mutex);
1606         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1607         return 0;
1608 }
1609
1610 int fio_backend(void)
1611 {
1612         struct thread_data *td;
1613         int i;
1614
1615         if (exec_profile) {
1616                 if (load_profile(exec_profile))
1617                         return 1;
1618                 free(exec_profile);
1619                 exec_profile = NULL;
1620         }
1621         if (!thread_number)
1622                 return 0;
1623
1624         if (write_bw_log) {
1625                 setup_log(&agg_io_log[DDIR_READ], 0);
1626                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1627         }
1628
1629         startup_mutex = fio_mutex_init(0);
1630         if (startup_mutex == NULL)
1631                 return 1;
1632         writeout_mutex = fio_mutex_init(1);
1633         if (writeout_mutex == NULL)
1634                 return 1;
1635
1636         set_genesis_time();
1637         create_disk_util_thread();
1638
1639         cgroup_list = smalloc(sizeof(*cgroup_list));
1640         INIT_FLIST_HEAD(cgroup_list);
1641
1642         run_threads();
1643
1644         if (!fio_abort) {
1645                 show_run_stats();
1646                 if (write_bw_log) {
1647                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1648                         __finish_log(agg_io_log[DDIR_WRITE],
1649                                         "agg-write_bw.log");
1650                 }
1651         }
1652
1653         for_each_td(td, i)
1654                 fio_options_free(td);
1655
1656         cgroup_kill(cgroup_list);
1657         sfree(cgroup_list);
1658         sfree(cgroup_mnt);
1659
1660         fio_mutex_remove(startup_mutex);
1661         fio_mutex_remove(writeout_mutex);
1662         return exit_value;
1663 }