fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 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 <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "hash.h"
  39 #include "smalloc.h"
  40 #include "verify.h"
  41 #include "diskutil.h"
  42 #include "cgroup.h"
  43
  44 unsigned long page_mask;
  45 unsigned long page_size;
  46
  47 #define PAGE_ALIGN(buf) \
  48         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  49
  50 int groupid = 0;
  51 int thread_number = 0;
  52 int nr_process = 0;
  53 int nr_thread = 0;
  54 int shm_id = 0;
  55 int temp_stall_ts;
  56 unsigned long done_secs = 0;
  57
  58 static struct fio_mutex *startup_mutex;
  59 static struct fio_mutex *writeout_mutex;
  60 static volatile int fio_abort;
  61 static int exit_value;
  62 static struct itimerval itimer;
  63 static pthread_t gtod_thread;
  64 static struct flist_head *cgroup_list;
  65 static char *cgroup_mnt;
  66
  67 struct io_log *agg_io_log[2];
  68
  69 #define TERMINATE_ALL           (-1)
  70 #define JOB_START_TIMEOUT       (5 * 1000)
  71
  72 void td_set_runstate(struct thread_data *td, int runstate)
  73 {
  74         if (td->runstate == runstate)
  75                 return;
  76
  77         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  78                                                 td->runstate, runstate);
  79         td->runstate = runstate;
  80 }
  81
  82 static void terminate_threads(int group_id)
  83 {
  84         struct thread_data *td;
  85         int i;
  86
  87         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  88
  89         for_each_td(td, i) {
  90                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  91                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  92                                                 td->o.name, (int) td->pid);
  93                         td->terminate = 1;
  94                         td->o.start_delay = 0;
  95
  96                         /*
  97                          * if the thread is running, just let it exit
  98                          */
  99                         if (td->runstate < TD_RUNNING)
 100                                 kill(td->pid, SIGQUIT);
 101                         else {
 102                                 struct ioengine_ops *ops = td->io_ops;
 103
 104                                 if (ops && (ops->flags & FIO_SIGQUIT))
 105                                         kill(td->pid, SIGQUIT);
 106                         }
 107                 }
 108         }
 109 }
 110
 111 static void status_timer_arm(void)
 112 {
 113         itimer.it_value.tv_sec = 0;
 114         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 115         setitimer(ITIMER_REAL, &itimer, NULL);
 116 }
 117
 118 static void sig_alrm(int fio_unused sig)
 119 {
 120         if (threads) {
 121                 update_io_ticks();
 122                 print_thread_status();
 123                 status_timer_arm();
 124         }
 125 }
 126
 127 /*
 128  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 129  */
 130 static void sig_quit(int sig)
 131 {
 132 }
 133
 134 static void sig_int(int sig)
 135 {
 136         if (threads) {
 137                 printf("\nfio: terminating on signal %d\n", sig);
 138                 fflush(stdout);
 139                 terminate_threads(TERMINATE_ALL);
 140         }
 141 }
 142
 143 static void sig_ill(int fio_unused sig)
 144 {
 145         if (!threads)
 146                 return;
 147
 148         log_err("fio: illegal instruction. your cpu does not support "
 149                 "the sse4.2 instruction for crc32c\n");
 150         terminate_threads(TERMINATE_ALL);
 151         exit(4);
 152 }
 153
 154 static void set_sig_handlers(void)
 155 {
 156         struct sigaction act;
 157
 158         memset(&act, 0, sizeof(act));
 159         act.sa_handler = sig_alrm;
 160         act.sa_flags = SA_RESTART;
 161         sigaction(SIGALRM, &act, NULL);
 162
 163         memset(&act, 0, sizeof(act));
 164         act.sa_handler = sig_int;
 165         act.sa_flags = SA_RESTART;
 166         sigaction(SIGINT, &act, NULL);
 167
 168         memset(&act, 0, sizeof(act));
 169         act.sa_handler = sig_ill;
 170         act.sa_flags = SA_RESTART;
 171         sigaction(SIGILL, &act, NULL);
 172
 173         memset(&act, 0, sizeof(act));
 174         act.sa_handler = sig_quit;
 175         act.sa_flags = SA_RESTART;
 176         sigaction(SIGQUIT, &act, NULL);
 177 }
 178
 179 /*
 180  * Check if we are above the minimum rate given.
 181  */
 182 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 183                             enum td_ddir ddir)
 184 {
 185         unsigned long long bytes = 0;
 186         unsigned long iops = 0;
 187         unsigned long spent;
 188         unsigned long rate;
 189         unsigned int ratemin = 0;
 190         unsigned int rate_iops = 0;
 191         unsigned int rate_iops_min = 0;
 192
 193         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 194                 return 0;
 195
 196         /*
 197          * allow a 2 second settle period in the beginning
 198          */
 199         if (mtime_since(&td->start, now) < 2000)
 200                 return 0;
 201
 202         iops += td->io_blocks[ddir];
 203         bytes += td->this_io_bytes[ddir];
 204         ratemin += td->o.ratemin[ddir];
 205         rate_iops += td->o.rate_iops[ddir];
 206         rate_iops_min += td->o.rate_iops_min[ddir];
 207
 208         /*
 209          * if rate blocks is set, sample is running
 210          */
 211         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 212                 spent = mtime_since(&td->lastrate[ddir], now);
 213                 if (spent < td->o.ratecycle)
 214                         return 0;
 215
 216                 if (td->o.rate[ddir]) {
 217                         /*
 218                          * check bandwidth specified rate
 219                          */
 220                         if (bytes < td->rate_bytes[ddir]) {
 221                                 log_err("%s: min rate %u not met\n", td->o.name,
 222                                                                 ratemin);
 223                                 return 1;
 224                         } else {
 225                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 226                                 if (rate < ratemin ||
 227                                     bytes < td->rate_bytes[ddir]) {
 228                                         log_err("%s: min rate %u not met, got"
 229                                                 " %luKB/sec\n", td->o.name,
 230                                                         ratemin, rate);
 231                                         return 1;
 232                                 }
 233                         }
 234                 } else {
 235                         /*
 236                          * checks iops specified rate
 237                          */
 238                         if (iops < rate_iops) {
 239                                 log_err("%s: min iops rate %u not met\n",
 240                                                 td->o.name, rate_iops);
 241                                 return 1;
 242                         } else {
 243                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 244                                 if (rate < rate_iops_min ||
 245                                     iops < td->rate_blocks[ddir]) {
 246                                         log_err("%s: min iops rate %u not met,"
 247                                                 " got %lu\n", td->o.name,
 248                                                         rate_iops_min, rate);
 249                                 }
 250                         }
 251                 }
 252         }
 253
 254         td->rate_bytes[ddir] = bytes;
 255         td->rate_blocks[ddir] = iops;
 256         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 257         return 0;
 258 }
 259
 260 static int check_min_rate(struct thread_data *td, struct timeval *now,
 261                           unsigned long *bytes_done)
 262 {
 263         int ret = 0;
 264
 265         if (bytes_done[0])
 266                 ret |= __check_min_rate(td, now, 0);
 267         if (bytes_done[1])
 268                 ret |= __check_min_rate(td, now, 1);
 269
 270         return ret;
 271 }
 272
 273 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 274 {
 275         if (!td->o.timeout)
 276                 return 0;
 277         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 278                 return 1;
 279
 280         return 0;
 281 }
 282
 283 /*
 284  * When job exits, we can cancel the in-flight IO if we are using async
 285  * io. Attempt to do so.
 286  */
 287 static void cleanup_pending_aio(struct thread_data *td)
 288 {
 289         struct flist_head *entry, *n;
 290         struct io_u *io_u;
 291         int r;
 292
 293         /*
 294          * get immediately available events, if any
 295          */
 296         r = io_u_queued_complete(td, 0, NULL);
 297         if (r < 0)
 298                 return;
 299
 300         /*
 301          * now cancel remaining active events
 302          */
 303         if (td->io_ops->cancel) {
 304                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 305                         io_u = flist_entry(entry, struct io_u, list);
 306
 307                         /*
 308                          * if the io_u isn't in flight, then that generally
 309                          * means someone leaked an io_u. complain but fix
 310                          * it up, so we don't stall here.
 311                          */
 312                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 313                                 log_err("fio: non-busy IO on busy list\n");
 314                                 put_io_u(td, io_u);
 315                         } else {
 316                                 r = td->io_ops->cancel(td, io_u);
 317                                 if (!r)
 318                                         put_io_u(td, io_u);
 319                         }
 320                 }
 321         }
 322
 323         if (td->cur_depth)
 324                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 325 }
 326
 327 /*
 328  * Helper to handle the final sync of a file. Works just like the normal
 329  * io path, just does everything sync.
 330  */
 331 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 332 {
 333         struct io_u *io_u = __get_io_u(td);
 334         int ret;
 335
 336         if (!io_u)
 337                 return 1;
 338
 339         io_u->ddir = DDIR_SYNC;
 340         io_u->file = f;
 341
 342         if (td_io_prep(td, io_u)) {
 343                 put_io_u(td, io_u);
 344                 return 1;
 345         }
 346
 347 requeue:
 348         ret = td_io_queue(td, io_u);
 349         if (ret < 0) {
 350                 td_verror(td, io_u->error, "td_io_queue");
 351                 put_io_u(td, io_u);
 352                 return 1;
 353         } else if (ret == FIO_Q_QUEUED) {
 354                 if (io_u_queued_complete(td, 1, NULL) < 0)
 355                         return 1;
 356         } else if (ret == FIO_Q_COMPLETED) {
 357                 if (io_u->error) {
 358                         td_verror(td, io_u->error, "td_io_queue");
 359                         return 1;
 360                 }
 361
 362                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 363                         return 1;
 364         } else if (ret == FIO_Q_BUSY) {
 365                 if (td_io_commit(td))
 366                         return 1;
 367                 goto requeue;
 368         }
 369
 370         return 0;
 371 }
 372
 373 static inline void update_tv_cache(struct thread_data *td)
 374 {
 375         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 376                 fio_gettime(&td->tv_cache, NULL);
 377 }
 378
 379 static int break_on_this_error(struct thread_data *td, int *retptr)
 380 {
 381         int ret = *retptr;
 382
 383         if (ret < 0 || td->error) {
 384                 int err;
 385
 386                 if (!td->o.continue_on_error)
 387                         return 1;
 388
 389                 if (ret < 0)
 390                         err = -ret;
 391                 else
 392                         err = td->error;
 393
 394                 if (td_non_fatal_error(err)) {
 395                         /*
 396                          * Continue with the I/Os in case of
 397                          * a non fatal error.
 398                          */
 399                         update_error_count(td, err);
 400                         td_clear_error(td);
 401                         *retptr = 0;
 402                         return 0;
 403                 } else if (td->o.fill_device && err == ENOSPC) {
 404                         /*
 405                          * We expect to hit this error if
 406                          * fill_device option is set.
 407                          */
 408                         td_clear_error(td);
 409                         td->terminate = 1;
 410                         return 1;
 411                 } else {
 412                         /*
 413                          * Stop the I/O in case of a fatal
 414                          * error.
 415                          */
 416                         update_error_count(td, err);
 417                         return 1;
 418                 }
 419         }
 420
 421         return 0;
 422 }
 423
 424 /*
 425  * The main verify engine. Runs over the writes we previously submitted,
 426  * reads the blocks back in, and checks the crc/md5 of the data.
 427  */
 428 static void do_verify(struct thread_data *td)
 429 {
 430         struct fio_file *f;
 431         struct io_u *io_u;
 432         int ret, min_events;
 433         unsigned int i;
 434
 435         /*
 436          * sync io first and invalidate cache, to make sure we really
 437          * read from disk.
 438          */
 439         for_each_file(td, f, i) {
 440                 if (!fio_file_open(f))
 441                         continue;
 442                 if (fio_io_sync(td, f))
 443                         break;
 444                 if (file_invalidate_cache(td, f))
 445                         break;
 446         }
 447
 448         if (td->error)
 449                 return;
 450
 451         td_set_runstate(td, TD_VERIFYING);
 452
 453         io_u = NULL;
 454         while (!td->terminate) {
 455                 int ret2, full;
 456
 457                 update_tv_cache(td);
 458
 459                 if (runtime_exceeded(td, &td->tv_cache)) {
 460                         td->terminate = 1;
 461                         break;
 462                 }
 463
 464                 io_u = __get_io_u(td);
 465                 if (!io_u)
 466                         break;
 467
 468                 if (get_next_verify(td, io_u)) {
 469                         put_io_u(td, io_u);
 470                         break;
 471                 }
 472
 473                 if (td_io_prep(td, io_u)) {
 474                         put_io_u(td, io_u);
 475                         break;
 476                 }
 477
 478                 if (td->o.verify_async)
 479                         io_u->end_io = verify_io_u_async;
 480                 else
 481                         io_u->end_io = verify_io_u;
 482
 483                 ret = td_io_queue(td, io_u);
 484                 switch (ret) {
 485                 case FIO_Q_COMPLETED:
 486                         if (io_u->error) {
 487                                 ret = -io_u->error;
 488                                 clear_io_u(td, io_u);
 489                         } else if (io_u->resid) {
 490                                 int bytes = io_u->xfer_buflen - io_u->resid;
 491                                 struct fio_file *f = io_u->file;
 492
 493                                 /*
 494                                  * zero read, fail
 495                                  */
 496                                 if (!bytes) {
 497                                         td_verror(td, EIO, "full resid");
 498                                         put_io_u(td, io_u);
 499                                         break;
 500                                 }
 501
 502                                 io_u->xfer_buflen = io_u->resid;
 503                                 io_u->xfer_buf += bytes;
 504                                 io_u->offset += bytes;
 505
 506                                 td->ts.short_io_u[io_u->ddir]++;
 507
 508                                 if (io_u->offset == f->real_file_size)
 509                                         goto sync_done;
 510
 511                                 requeue_io_u(td, &io_u);
 512                         } else {
 513 sync_done:
 514                                 ret = io_u_sync_complete(td, io_u, NULL);
 515                                 if (ret < 0)
 516                                         break;
 517                         }
 518                         continue;
 519                 case FIO_Q_QUEUED:
 520                         break;
 521                 case FIO_Q_BUSY:
 522                         requeue_io_u(td, &io_u);
 523                         ret2 = td_io_commit(td);
 524                         if (ret2 < 0)
 525                                 ret = ret2;
 526                         break;
 527                 default:
 528                         assert(ret < 0);
 529                         td_verror(td, -ret, "td_io_queue");
 530                         break;
 531                 }
 532
 533                 if (break_on_this_error(td, &ret))
 534                         break;
 535
 536                 /*
 537                  * if we can queue more, do so. but check if there are
 538                  * completed io_u's first.
 539                  */
 540                 full = queue_full(td) || ret == FIO_Q_BUSY;
 541                 if (full || !td->o.iodepth_batch_complete) {
 542                         min_events = min(td->o.iodepth_batch_complete,
 543                                          td->cur_depth);
 544                         if (full && !min_events)
 545                                 min_events = 1;
 546
 547                         do {
 548                                 /*
 549                                  * Reap required number of io units, if any,
 550                                  * and do the verification on them through
 551                                  * the callback handler
 552                                  */
 553                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 554                                         ret = -1;
 555                                         break;
 556                                 }
 557                         } while (full && (td->cur_depth > td->o.iodepth_low));
 558                 }
 559                 if (ret < 0)
 560                         break;
 561         }
 562
 563         if (!td->error) {
 564                 min_events = td->cur_depth;
 565
 566                 if (min_events)
 567                         ret = io_u_queued_complete(td, min_events, NULL);
 568         } else
 569                 cleanup_pending_aio(td);
 570
 571         td_set_runstate(td, TD_RUNNING);
 572 }
 573
 574 /*
 575  * Main IO worker function. It retrieves io_u's to process and queues
 576  * and reaps them, checking for rate and errors along the way.
 577  */
 578 static void do_io(struct thread_data *td)
 579 {
 580         unsigned int i;
 581         int ret = 0;
 582
 583         if (in_ramp_time(td))
 584                 td_set_runstate(td, TD_RAMP);
 585         else
 586                 td_set_runstate(td, TD_RUNNING);
 587
 588         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 589                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 590                 struct timeval comp_time;
 591                 unsigned long bytes_done[2] = { 0, 0 };
 592                 int min_evts = 0;
 593                 struct io_u *io_u;
 594                 int ret2, full;
 595
 596                 if (td->terminate)
 597                         break;
 598
 599                 update_tv_cache(td);
 600
 601                 if (runtime_exceeded(td, &td->tv_cache)) {
 602                         td->terminate = 1;
 603                         break;
 604                 }
 605
 606                 io_u = get_io_u(td);
 607                 if (!io_u)
 608                         break;
 609
 610                 /*
 611                  * Add verification end_io handler, if asked to verify
 612                  * a previously written file.
 613                  */
 614                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 615                     !td_rw(td)) {
 616                         if (td->o.verify_async)
 617                                 io_u->end_io = verify_io_u_async;
 618                         else
 619                                 io_u->end_io = verify_io_u;
 620                         td_set_runstate(td, TD_VERIFYING);
 621                 } else if (in_ramp_time(td))
 622                         td_set_runstate(td, TD_RAMP);
 623                 else
 624                         td_set_runstate(td, TD_RUNNING);
 625
 626                 ret = td_io_queue(td, io_u);
 627                 switch (ret) {
 628                 case FIO_Q_COMPLETED:
 629                         if (io_u->error) {
 630                                 ret = -io_u->error;
 631                                 clear_io_u(td, io_u);
 632                         } else if (io_u->resid) {
 633                                 int bytes = io_u->xfer_buflen - io_u->resid;
 634                                 struct fio_file *f = io_u->file;
 635
 636                                 /*
 637                                  * zero read, fail
 638                                  */
 639                                 if (!bytes) {
 640                                         td_verror(td, EIO, "full resid");
 641                                         put_io_u(td, io_u);
 642                                         break;
 643                                 }
 644
 645                                 io_u->xfer_buflen = io_u->resid;
 646                                 io_u->xfer_buf += bytes;
 647                                 io_u->offset += bytes;
 648
 649                                 td->ts.short_io_u[io_u->ddir]++;
 650
 651                                 if (io_u->offset == f->real_file_size)
 652                                         goto sync_done;
 653
 654                                 requeue_io_u(td, &io_u);
 655                         } else {
 656 sync_done:
 657                                 if (__should_check_rate(td, 0) ||
 658                                     __should_check_rate(td, 1))
 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, &ret))
 688                         break;
 689
 690                 /*
 691                  * See if we need to complete some commands
 692                  */
 693                 full = queue_full(td) || ret == FIO_Q_BUSY;
 694                 if (full || !td->o.iodepth_batch_complete) {
 695                         min_evts = min(td->o.iodepth_batch_complete,
 696                                         td->cur_depth);
 697                         if (full && !min_evts)
 698                                 min_evts = 1;
 699
 700                         if (__should_check_rate(td, 0) ||
 701                             __should_check_rate(td, 1))
 702                                 fio_gettime(&comp_time, NULL);
 703
 704                         do {
 705                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 706                                 if (ret < 0)
 707                                         break;
 708
 709                         } while (full && (td->cur_depth > td->o.iodepth_low));
 710                 }
 711
 712                 if (ret < 0)
 713                         break;
 714                 if (!(bytes_done[0] + bytes_done[1]))
 715                         continue;
 716
 717                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 718                         if (check_min_rate(td, &comp_time, bytes_done)) {
 719                                 if (exitall_on_terminate)
 720                                         terminate_threads(td->groupid);
 721                                 td_verror(td, EIO, "check_min_rate");
 722                                 break;
 723                         }
 724                 }
 725
 726                 if (td->o.thinktime) {
 727                         unsigned long long b;
 728
 729                         b = td->io_blocks[0] + td->io_blocks[1];
 730                         if (!(b % td->o.thinktime_blocks)) {
 731                                 int left;
 732
 733                                 if (td->o.thinktime_spin)
 734                                         usec_spin(td->o.thinktime_spin);
 735
 736                                 left = td->o.thinktime - td->o.thinktime_spin;
 737                                 if (left)
 738                                         usec_sleep(td, left);
 739                         }
 740                 }
 741         }
 742
 743         if (td->o.fill_device && td->error == ENOSPC) {
 744                 td->error = 0;
 745                 td->terminate = 1;
 746         }
 747         if (!td->error) {
 748                 struct fio_file *f;
 749
 750                 i = td->cur_depth;
 751                 if (i)
 752                         ret = io_u_queued_complete(td, i, NULL);
 753
 754                 if (should_fsync(td) && td->o.end_fsync) {
 755                         td_set_runstate(td, TD_FSYNCING);
 756
 757                         for_each_file(td, f, i) {
 758                                 if (!fio_file_open(f))
 759                                         continue;
 760                                 fio_io_sync(td, f);
 761                         }
 762                 }
 763         } else
 764                 cleanup_pending_aio(td);
 765
 766         /*
 767          * stop job if we failed doing any IO
 768          */
 769         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 770                 td->done = 1;
 771 }
 772
 773 static void cleanup_io_u(struct thread_data *td)
 774 {
 775         struct flist_head *entry, *n;
 776         struct io_u *io_u;
 777
 778         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 779                 io_u = flist_entry(entry, struct io_u, list);
 780
 781                 flist_del(&io_u->list);
 782                 free(io_u);
 783         }
 784
 785         free_io_mem(td);
 786 }
 787
 788 static int init_io_u(struct thread_data *td)
 789 {
 790         struct io_u *io_u;
 791         unsigned int max_bs;
 792         int cl_align, i, max_units;
 793         char *p;
 794
 795         max_units = td->o.iodepth;
 796         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 797         td->orig_buffer_size = (unsigned long long) max_bs
 798                                         * (unsigned long long) max_units;
 799
 800         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 801                 unsigned long bs;
 802
 803                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 804                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 805         }
 806
 807         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 808                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 809                 return 1;
 810         }
 811
 812         if (allocate_io_mem(td))
 813                 return 1;
 814
 815         if (td->o.odirect || td->o.mem_align)
 816                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 817         else
 818                 p = td->orig_buffer;
 819
 820         cl_align = os_cache_line_size();
 821
 822         for (i = 0; i < max_units; i++) {
 823                 void *ptr;
 824
 825                 if (td->terminate)
 826                         return 1;
 827
 828                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 829                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 830                         break;
 831                 }
 832
 833                 io_u = ptr;
 834                 memset(io_u, 0, sizeof(*io_u));
 835                 INIT_FLIST_HEAD(&io_u->list);
 836                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 837
 838                 if (!(td->io_ops->flags & FIO_NOIO)) {
 839                         io_u->buf = p + max_bs * i;
 840                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 841
 842                         if (td_write(td) && !td->o.refill_buffers)
 843                                 io_u_fill_buffer(td, io_u, max_bs);
 844                 }
 845
 846                 io_u->index = i;
 847                 io_u->flags = IO_U_F_FREE;
 848                 flist_add(&io_u->list, &td->io_u_freelist);
 849         }
 850
 851         return 0;
 852 }
 853
 854 static int switch_ioscheduler(struct thread_data *td)
 855 {
 856         char tmp[256], tmp2[128];
 857         FILE *f;
 858         int ret;
 859
 860         if (td->io_ops->flags & FIO_DISKLESSIO)
 861                 return 0;
 862
 863         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 864
 865         f = fopen(tmp, "r+");
 866         if (!f) {
 867                 if (errno == ENOENT) {
 868                         log_err("fio: os or kernel doesn't support IO scheduler"
 869                                 " switching\n");
 870                         return 0;
 871                 }
 872                 td_verror(td, errno, "fopen iosched");
 873                 return 1;
 874         }
 875
 876         /*
 877          * Set io scheduler.
 878          */
 879         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 880         if (ferror(f) || ret != 1) {
 881                 td_verror(td, errno, "fwrite");
 882                 fclose(f);
 883                 return 1;
 884         }
 885
 886         rewind(f);
 887
 888         /*
 889          * Read back and check that the selected scheduler is now the default.
 890          */
 891         ret = fread(tmp, 1, sizeof(tmp), f);
 892         if (ferror(f) || ret < 0) {
 893                 td_verror(td, errno, "fread");
 894                 fclose(f);
 895                 return 1;
 896         }
 897
 898         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 899         if (!strstr(tmp, tmp2)) {
 900                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 901                 td_verror(td, EINVAL, "iosched_switch");
 902                 fclose(f);
 903                 return 1;
 904         }
 905
 906         fclose(f);
 907         return 0;
 908 }
 909
 910 static int keep_running(struct thread_data *td)
 911 {
 912         unsigned long long io_done;
 913
 914         if (td->done)
 915                 return 0;
 916         if (td->o.time_based)
 917                 return 1;
 918         if (td->o.loops) {
 919                 td->o.loops--;
 920                 return 1;
 921         }
 922
 923         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 924                         + td->io_skip_bytes;
 925         if (io_done < td->o.size)
 926                 return 1;
 927
 928         return 0;
 929 }
 930
 931 static void reset_io_counters(struct thread_data *td)
 932 {
 933         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 934         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 935         td->zone_bytes = 0;
 936         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 937         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 938
 939         td->last_was_sync = 0;
 940
 941         /*
 942          * reset file done count if we are to start over
 943          */
 944         if (td->o.time_based || td->o.loops)
 945                 td->nr_done_files = 0;
 946
 947         /*
 948          * Set the same seed to get repeatable runs
 949          */
 950         td_fill_rand_seeds(td);
 951 }
 952
 953 void reset_all_stats(struct thread_data *td)
 954 {
 955         struct timeval tv;
 956         int i;
 957
 958         reset_io_counters(td);
 959
 960         for (i = 0; i < 2; i++) {
 961                 td->io_bytes[i] = 0;
 962                 td->io_blocks[i] = 0;
 963                 td->io_issues[i] = 0;
 964                 td->ts.total_io_u[i] = 0;
 965         }
 966
 967         fio_gettime(&tv, NULL);
 968         memcpy(&td->epoch, &tv, sizeof(tv));
 969         memcpy(&td->start, &tv, sizeof(tv));
 970 }
 971
 972 static void clear_io_state(struct thread_data *td)
 973 {
 974         struct fio_file *f;
 975         unsigned int i;
 976
 977         reset_io_counters(td);
 978
 979         close_files(td);
 980         for_each_file(td, f, i)
 981                 fio_file_clear_done(f);
 982 }
 983
 984 static int exec_string(const char *string)
 985 {
 986         int ret, newlen = strlen(string) + 1 + 8;
 987         char *str;
 988
 989         str = malloc(newlen);
 990         sprintf(str, "sh -c %s", string);
 991
 992         ret = system(str);
 993         if (ret == -1)
 994                 log_err("fio: exec of cmd <%s> failed\n", str);
 995
 996         free(str);
 997         return ret;
 998 }
 999
1000 /*
1001  * Entry point for the thread based jobs. The process based jobs end up
1002  * here as well, after a little setup.
1003  */
1004 static void *thread_main(void *data)
1005 {
1006         unsigned long long runtime[2], elapsed;
1007         struct thread_data *td = data;
1008         pthread_condattr_t attr;
1009         int clear_state;
1010
1011         if (!td->o.use_thread)
1012                 setsid();
1013
1014         td->pid = getpid();
1015
1016         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1017
1018         INIT_FLIST_HEAD(&td->io_u_freelist);
1019         INIT_FLIST_HEAD(&td->io_u_busylist);
1020         INIT_FLIST_HEAD(&td->io_u_requeues);
1021         INIT_FLIST_HEAD(&td->io_log_list);
1022         INIT_FLIST_HEAD(&td->io_hist_list);
1023         INIT_FLIST_HEAD(&td->verify_list);
1024         pthread_mutex_init(&td->io_u_lock, NULL);
1025         td->io_hist_tree = RB_ROOT;
1026
1027         pthread_condattr_init(&attr);
1028         pthread_cond_init(&td->verify_cond, &attr);
1029         pthread_cond_init(&td->free_cond, &attr);
1030
1031         td_set_runstate(td, TD_INITIALIZED);
1032         dprint(FD_MUTEX, "up startup_mutex\n");
1033         fio_mutex_up(startup_mutex);
1034         dprint(FD_MUTEX, "wait on td->mutex\n");
1035         fio_mutex_down(td->mutex);
1036         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1037
1038         /*
1039          * the ->mutex mutex is now no longer used, close it to avoid
1040          * eating a file descriptor
1041          */
1042         fio_mutex_remove(td->mutex);
1043
1044         if (td->o.uid != -1U && setuid(td->o.uid)) {
1045                 td_verror(td, errno, "setuid");
1046                 goto err;
1047         }
1048         if (td->o.gid != -1U && setgid(td->o.gid)) {
1049                 td_verror(td, errno, "setgid");
1050                 goto err;
1051         }
1052
1053         /*
1054          * May alter parameters that init_io_u() will use, so we need to
1055          * do this first.
1056          */
1057         if (init_iolog(td))
1058                 goto err;
1059
1060         if (init_io_u(td))
1061                 goto err;
1062
1063         if (td->o.verify_async && verify_async_init(td))
1064                 goto err;
1065
1066         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1067                 td_verror(td, errno, "cpu_set_affinity");
1068                 goto err;
1069         }
1070
1071         /*
1072          * If we have a gettimeofday() thread, make sure we exclude that
1073          * thread from this job
1074          */
1075         if (td->o.gtod_cpu) {
1076                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1077                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1078                         td_verror(td, errno, "cpu_set_affinity");
1079                         goto err;
1080                 }
1081         }
1082
1083         if (td->ioprio_set) {
1084                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1085                         td_verror(td, errno, "ioprio_set");
1086                         goto err;
1087                 }
1088         }
1089
1090         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1091                 goto err;
1092
1093         if (nice(td->o.nice) == -1) {
1094                 td_verror(td, errno, "nice");
1095                 goto err;
1096         }
1097
1098         if (td->o.ioscheduler && switch_ioscheduler(td))
1099                 goto err;
1100
1101         if (!td->o.create_serialize && setup_files(td))
1102                 goto err;
1103
1104         if (td_io_init(td))
1105                 goto err;
1106
1107         if (init_random_map(td))
1108                 goto err;
1109
1110         if (td->o.exec_prerun) {
1111                 if (exec_string(td->o.exec_prerun))
1112                         goto err;
1113         }
1114
1115         if (td->o.pre_read) {
1116                 if (pre_read_files(td) < 0)
1117                         goto err;
1118         }
1119
1120         fio_gettime(&td->epoch, NULL);
1121         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1122
1123         runtime[0] = runtime[1] = 0;
1124         clear_state = 0;
1125         while (keep_running(td)) {
1126                 fio_gettime(&td->start, NULL);
1127                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1128                                 sizeof(td->start));
1129                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1130                                 sizeof(td->start));
1131                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1132
1133                 if (td->o.ratemin[0] || td->o.ratemin[1])
1134                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1135                                                         sizeof(td->lastrate));
1136
1137                 if (clear_state)
1138                         clear_io_state(td);
1139
1140                 prune_io_piece_log(td);
1141
1142                 do_io(td);
1143
1144                 clear_state = 1;
1145
1146                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1147                         elapsed = utime_since_now(&td->start);
1148                         runtime[DDIR_READ] += elapsed;
1149                 }
1150                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1151                         elapsed = utime_since_now(&td->start);
1152                         runtime[DDIR_WRITE] += elapsed;
1153                 }
1154
1155                 if (td->error || td->terminate)
1156                         break;
1157
1158                 if (!td->o.do_verify ||
1159                     td->o.verify == VERIFY_NONE ||
1160                     (td->io_ops->flags & FIO_UNIDIR))
1161                         continue;
1162
1163                 clear_io_state(td);
1164
1165                 fio_gettime(&td->start, NULL);
1166
1167                 do_verify(td);
1168
1169                 runtime[DDIR_READ] += utime_since_now(&td->start);
1170
1171                 if (td->error || td->terminate)
1172                         break;
1173         }
1174
1175         update_rusage_stat(td);
1176         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1177         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1178         td->ts.total_run_time = mtime_since_now(&td->epoch);
1179         td->ts.io_bytes[0] = td->io_bytes[0];
1180         td->ts.io_bytes[1] = td->io_bytes[1];
1181
1182         fio_mutex_down(writeout_mutex);
1183         if (td->ts.bw_log) {
1184                 if (td->o.bw_log_file) {
1185                         finish_log_named(td, td->ts.bw_log,
1186                                                 td->o.bw_log_file, "bw");
1187                 } else
1188                         finish_log(td, td->ts.bw_log, "bw");
1189         }
1190         if (td->ts.slat_log) {
1191                 if (td->o.lat_log_file) {
1192                         finish_log_named(td, td->ts.slat_log,
1193                                                 td->o.lat_log_file, "slat");
1194                 } else
1195                         finish_log(td, td->ts.slat_log, "slat");
1196         }
1197         if (td->ts.clat_log) {
1198                 if (td->o.lat_log_file) {
1199                         finish_log_named(td, td->ts.clat_log,
1200                                                 td->o.lat_log_file, "clat");
1201                 } else
1202                         finish_log(td, td->ts.clat_log, "clat");
1203         }
1204         fio_mutex_up(writeout_mutex);
1205         if (td->o.exec_postrun)
1206                 exec_string(td->o.exec_postrun);
1207
1208         if (exitall_on_terminate)
1209                 terminate_threads(td->groupid);
1210
1211 err:
1212         if (td->error)
1213                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1214                                                         td->verror);
1215
1216         if (td->o.verify_async)
1217                 verify_async_exit(td);
1218
1219         close_and_free_files(td);
1220         close_ioengine(td);
1221         cleanup_io_u(td);
1222         cgroup_shutdown(td, &cgroup_mnt);
1223
1224         if (td->o.cpumask_set) {
1225                 int ret = fio_cpuset_exit(&td->o.cpumask);
1226
1227                 td_verror(td, ret, "fio_cpuset_exit");
1228         }
1229
1230         /*
1231          * do this very late, it will log file closing as well
1232          */
1233         if (td->o.write_iolog_file)
1234                 write_iolog_close(td);
1235
1236         options_mem_free(td);
1237         td_set_runstate(td, TD_EXITED);
1238         return (void *) (unsigned long) td->error;
1239 }
1240
1241 /*
1242  * We cannot pass the td data into a forked process, so attach the td and
1243  * pass it to the thread worker.
1244  */
1245 static int fork_main(int shmid, int offset)
1246 {
1247         struct thread_data *td;
1248         void *data, *ret;
1249
1250         data = shmat(shmid, NULL, 0);
1251         if (data == (void *) -1) {
1252                 int __err = errno;
1253
1254                 perror("shmat");
1255                 return __err;
1256         }
1257
1258         td = data + offset * sizeof(struct thread_data);
1259         ret = thread_main(td);
1260         shmdt(data);
1261         return (int) (unsigned long) ret;
1262 }
1263
1264 /*
1265  * Run over the job map and reap the threads that have exited, if any.
1266  */
1267 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1268 {
1269         struct thread_data *td;
1270         int i, cputhreads, realthreads, pending, status, ret;
1271
1272         /*
1273          * reap exited threads (TD_EXITED -> TD_REAPED)
1274          */
1275         realthreads = pending = cputhreads = 0;
1276         for_each_td(td, i) {
1277                 int flags = 0;
1278
1279                 /*
1280                  * ->io_ops is NULL for a thread that has closed its
1281                  * io engine
1282                  */
1283                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1284                         cputhreads++;
1285                 else
1286                         realthreads++;
1287
1288                 if (!td->pid) {
1289                         pending++;
1290                         continue;
1291                 }
1292                 if (td->runstate == TD_REAPED)
1293                         continue;
1294                 if (td->o.use_thread) {
1295                         if (td->runstate == TD_EXITED) {
1296                                 td_set_runstate(td, TD_REAPED);
1297                                 goto reaped;
1298                         }
1299                         continue;
1300                 }
1301
1302                 flags = WNOHANG;
1303                 if (td->runstate == TD_EXITED)
1304                         flags = 0;
1305
1306                 /*
1307                  * check if someone quit or got killed in an unusual way
1308                  */
1309                 ret = waitpid(td->pid, &status, flags);
1310                 if (ret < 0) {
1311                         if (errno == ECHILD) {
1312                                 log_err("fio: pid=%d disappeared %d\n",
1313                                                 (int) td->pid, td->runstate);
1314                                 td_set_runstate(td, TD_REAPED);
1315                                 goto reaped;
1316                         }
1317                         perror("waitpid");
1318                 } else if (ret == td->pid) {
1319                         if (WIFSIGNALED(status)) {
1320                                 int sig = WTERMSIG(status);
1321
1322                                 if (sig != SIGQUIT)
1323                                         log_err("fio: pid=%d, got signal=%d\n",
1324                                                         (int) td->pid, sig);
1325                                 td_set_runstate(td, TD_REAPED);
1326                                 goto reaped;
1327                         }
1328                         if (WIFEXITED(status)) {
1329                                 if (WEXITSTATUS(status) && !td->error)
1330                                         td->error = WEXITSTATUS(status);
1331
1332                                 td_set_runstate(td, TD_REAPED);
1333                                 goto reaped;
1334                         }
1335                 }
1336
1337                 /*
1338                  * thread is not dead, continue
1339                  */
1340                 pending++;
1341                 continue;
1342 reaped:
1343                 (*nr_running)--;
1344                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1345                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1346                 if (!td->pid)
1347                         pending--;
1348
1349                 if (td->error)
1350                         exit_value++;
1351
1352                 done_secs += mtime_since_now(&td->epoch) / 1000;
1353         }
1354
1355         if (*nr_running == cputhreads && !pending && realthreads)
1356                 terminate_threads(TERMINATE_ALL);
1357 }
1358
1359 static void *gtod_thread_main(void *data)
1360 {
1361         fio_mutex_up(startup_mutex);
1362
1363         /*
1364          * As long as we have jobs around, update the clock. It would be nice
1365          * to have some way of NOT hammering that CPU with gettimeofday(),
1366          * but I'm not sure what to use outside of a simple CPU nop to relax
1367          * it - we don't want to lose precision.
1368          */
1369         while (threads) {
1370                 fio_gtod_update();
1371                 nop;
1372         }
1373
1374         return NULL;
1375 }
1376
1377 static int fio_start_gtod_thread(void)
1378 {
1379         int ret;
1380
1381         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1382         if (ret) {
1383                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1384                 return 1;
1385         }
1386
1387         ret = pthread_detach(gtod_thread);
1388         if (ret) {
1389                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1390                 return 1;
1391         }
1392
1393         dprint(FD_MUTEX, "wait on startup_mutex\n");
1394         fio_mutex_down(startup_mutex);
1395         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1396         return 0;
1397 }
1398
1399 /*
1400  * Main function for kicking off and reaping jobs, as needed.
1401  */
1402 static void run_threads(void)
1403 {
1404         struct thread_data *td;
1405         unsigned long spent;
1406         int i, todo, nr_running, m_rate, t_rate, nr_started;
1407
1408         if (fio_pin_memory())
1409                 return;
1410
1411         if (fio_gtod_offload && fio_start_gtod_thread())
1412                 return;
1413
1414         if (!terse_output) {
1415                 printf("Starting ");
1416                 if (nr_thread)
1417                         printf("%d thread%s", nr_thread,
1418                                                 nr_thread > 1 ? "s" : "");
1419                 if (nr_process) {
1420                         if (nr_thread)
1421                                 printf(" and ");
1422                         printf("%d process%s", nr_process,
1423                                                 nr_process > 1 ? "es" : "");
1424                 }
1425                 printf("\n");
1426                 fflush(stdout);
1427         }
1428
1429         set_sig_handlers();
1430
1431         todo = thread_number;
1432         nr_running = 0;
1433         nr_started = 0;
1434         m_rate = t_rate = 0;
1435
1436         for_each_td(td, i) {
1437                 print_status_init(td->thread_number - 1);
1438
1439                 if (!td->o.create_serialize) {
1440                         init_disk_util(td);
1441                         continue;
1442                 }
1443
1444                 /*
1445                  * do file setup here so it happens sequentially,
1446                  * we don't want X number of threads getting their
1447                  * client data interspersed on disk
1448                  */
1449                 if (setup_files(td)) {
1450                         exit_value++;
1451                         if (td->error)
1452                                 log_err("fio: pid=%d, err=%d/%s\n",
1453                                         (int) td->pid, td->error, td->verror);
1454                         td_set_runstate(td, TD_REAPED);
1455                         todo--;
1456                 } else {
1457                         struct fio_file *f;
1458                         unsigned int i;
1459
1460                         /*
1461                          * for sharing to work, each job must always open
1462                          * its own files. so close them, if we opened them
1463                          * for creation
1464                          */
1465                         for_each_file(td, f, i) {
1466                                 if (fio_file_open(f))
1467                                         td_io_close_file(td, f);
1468                         }
1469                 }
1470
1471                 init_disk_util(td);
1472         }
1473
1474         set_genesis_time();
1475
1476         while (todo) {
1477                 struct thread_data *map[MAX_JOBS];
1478                 struct timeval this_start;
1479                 int this_jobs = 0, left;
1480
1481                 /*
1482                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1483                  */
1484                 for_each_td(td, i) {
1485                         if (td->runstate != TD_NOT_CREATED)
1486                                 continue;
1487
1488                         /*
1489                          * never got a chance to start, killed by other
1490                          * thread for some reason
1491                          */
1492                         if (td->terminate) {
1493                                 todo--;
1494                                 continue;
1495                         }
1496
1497                         if (td->o.start_delay) {
1498                                 spent = mtime_since_genesis();
1499
1500                                 if (td->o.start_delay * 1000 > spent)
1501                                         continue;
1502                         }
1503
1504                         if (td->o.stonewall && (nr_started || nr_running)) {
1505                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1506                                                         td->o.name);
1507                                 break;
1508                         }
1509
1510                         /*
1511                          * Set state to created. Thread will transition
1512                          * to TD_INITIALIZED when it's done setting up.
1513                          */
1514                         td_set_runstate(td, TD_CREATED);
1515                         map[this_jobs++] = td;
1516                         nr_started++;
1517
1518                         if (td->o.use_thread) {
1519                                 int ret;
1520
1521                                 dprint(FD_PROCESS, "will pthread_create\n");
1522                                 ret = pthread_create(&td->thread, NULL,
1523                                                         thread_main, td);
1524                                 if (ret) {
1525                                         log_err("pthread_create: %s\n",
1526                                                         strerror(ret));
1527                                         nr_started--;
1528                                         break;
1529                                 }
1530                                 ret = pthread_detach(td->thread);
1531                                 if (ret)
1532                                         log_err("pthread_detach: %s",
1533                                                         strerror(ret));
1534                         } else {
1535                                 pid_t pid;
1536                                 dprint(FD_PROCESS, "will fork\n");
1537                                 pid = fork();
1538                                 if (!pid) {
1539                                         int ret = fork_main(shm_id, i);
1540
1541                                         _exit(ret);
1542                                 } else if (i == fio_debug_jobno)
1543                                         *fio_debug_jobp = pid;
1544                         }
1545                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1546                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1547                                 log_err("fio: job startup hung? exiting.\n");
1548                                 terminate_threads(TERMINATE_ALL);
1549                                 fio_abort = 1;
1550                                 nr_started--;
1551                                 break;
1552                         }
1553                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1554                 }
1555
1556                 /*
1557                  * Wait for the started threads to transition to
1558                  * TD_INITIALIZED.
1559                  */
1560                 fio_gettime(&this_start, NULL);
1561                 left = this_jobs;
1562                 while (left && !fio_abort) {
1563                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1564                                 break;
1565
1566                         usleep(100000);
1567
1568                         for (i = 0; i < this_jobs; i++) {
1569                                 td = map[i];
1570                                 if (!td)
1571                                         continue;
1572                                 if (td->runstate == TD_INITIALIZED) {
1573                                         map[i] = NULL;
1574                                         left--;
1575                                 } else if (td->runstate >= TD_EXITED) {
1576                                         map[i] = NULL;
1577                                         left--;
1578                                         todo--;
1579                                         nr_running++; /* work-around... */
1580                                 }
1581                         }
1582                 }
1583
1584                 if (left) {
1585                         log_err("fio: %d jobs failed to start\n", left);
1586                         for (i = 0; i < this_jobs; i++) {
1587                                 td = map[i];
1588                                 if (!td)
1589                                         continue;
1590                                 kill(td->pid, SIGTERM);
1591                         }
1592                         break;
1593                 }
1594
1595                 /*
1596                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1597                  */
1598                 for_each_td(td, i) {
1599                         if (td->runstate != TD_INITIALIZED)
1600                                 continue;
1601
1602                         if (in_ramp_time(td))
1603                                 td_set_runstate(td, TD_RAMP);
1604                         else
1605                                 td_set_runstate(td, TD_RUNNING);
1606                         nr_running++;
1607                         nr_started--;
1608                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1609                         t_rate += td->o.rate[0] + td->o.rate[1];
1610                         todo--;
1611                         fio_mutex_up(td->mutex);
1612                 }
1613
1614                 reap_threads(&nr_running, &t_rate, &m_rate);
1615
1616                 if (todo)
1617                         usleep(100000);
1618         }
1619
1620         while (nr_running) {
1621                 reap_threads(&nr_running, &t_rate, &m_rate);
1622                 usleep(10000);
1623         }
1624
1625         update_io_ticks();
1626         fio_unpin_memory();
1627 }
1628
1629 int main(int argc, char *argv[])
1630 {
1631         long ps;
1632
1633         sinit();
1634
1635         /*
1636          * We need locale for number printing, if it isn't set then just
1637          * go with the US format.
1638          */
1639         if (!getenv("LC_NUMERIC"))
1640                 setlocale(LC_NUMERIC, "en_US");
1641
1642         ps = sysconf(_SC_PAGESIZE);
1643         if (ps < 0) {
1644                 log_err("Failed to get page size\n");
1645                 return 1;
1646         }
1647
1648         page_size = ps;
1649         page_mask = ps - 1;
1650
1651         fio_keywords_init();
1652
1653         if (parse_options(argc, argv))
1654                 return 1;
1655
1656         if (!thread_number)
1657                 return 0;
1658
1659         if (write_bw_log) {
1660                 setup_log(&agg_io_log[DDIR_READ]);
1661                 setup_log(&agg_io_log[DDIR_WRITE]);
1662         }
1663
1664         startup_mutex = fio_mutex_init(0);
1665         writeout_mutex = fio_mutex_init(1);
1666
1667         set_genesis_time();
1668
1669         status_timer_arm();
1670
1671         cgroup_list = smalloc(sizeof(*cgroup_list));
1672         INIT_FLIST_HEAD(cgroup_list);
1673
1674         run_threads();
1675
1676         if (!fio_abort) {
1677                 show_run_stats();
1678                 if (write_bw_log) {
1679                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1680                         __finish_log(agg_io_log[DDIR_WRITE],
1681                                         "agg-write_bw.log");
1682                 }
1683         }
1684
1685         cgroup_kill(cgroup_list);
1686         sfree(cgroup_list);
1687         if (cgroup_mnt)
1688                 sfree(cgroup_mnt);
1689
1690         fio_mutex_remove(startup_mutex);
1691         fio_mutex_remove(writeout_mutex);
1692         return exit_value;
1693 }