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 #include "profile.h"
  44 #include "lib/rand.h"
  45
  46 unsigned long page_mask;
  47 unsigned long page_size;
  48
  49 #define PAGE_ALIGN(buf) \
  50         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  51
  52 int groupid = 0;
  53 int thread_number = 0;
  54 int nr_process = 0;
  55 int nr_thread = 0;
  56 int shm_id = 0;
  57 int temp_stall_ts;
  58 unsigned long done_secs = 0;
  59
  60 static struct fio_mutex *startup_mutex;
  61 static struct fio_mutex *writeout_mutex;
  62 static volatile int fio_abort;
  63 static int exit_value;
  64 static struct itimerval itimer;
  65 static pthread_t gtod_thread;
  66 static struct flist_head *cgroup_list;
  67 static char *cgroup_mnt;
  68
  69 struct io_log *agg_io_log[2];
  70
  71 #define TERMINATE_ALL           (-1)
  72 #define JOB_START_TIMEOUT       (5 * 1000)
  73
  74 void td_set_runstate(struct thread_data *td, int runstate)
  75 {
  76         if (td->runstate == runstate)
  77                 return;
  78
  79         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  80                                                 td->runstate, runstate);
  81         td->runstate = runstate;
  82 }
  83
  84 static void terminate_threads(int group_id)
  85 {
  86         struct thread_data *td;
  87         int i;
  88
  89         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  90
  91         for_each_td(td, i) {
  92                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  93                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  94                                                 td->o.name, (int) td->pid);
  95                         td->terminate = 1;
  96                         td->o.start_delay = 0;
  97
  98                         /*
  99                          * if the thread is running, just let it exit
 100                          */
 101                         if (td->runstate < TD_RUNNING)
 102                                 kill(td->pid, SIGQUIT);
 103                         else {
 104                                 struct ioengine_ops *ops = td->io_ops;
 105
 106                                 if (ops && (ops->flags & FIO_SIGQUIT))
 107                                         kill(td->pid, SIGQUIT);
 108                         }
 109                 }
 110         }
 111 }
 112
 113 static void status_timer_arm(void)
 114 {
 115         itimer.it_value.tv_sec = 0;
 116         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 117         setitimer(ITIMER_REAL, &itimer, NULL);
 118 }
 119
 120 static void sig_alrm(int fio_unused sig)
 121 {
 122         if (threads) {
 123                 update_io_ticks();
 124                 print_thread_status();
 125                 status_timer_arm();
 126         }
 127 }
 128
 129 /*
 130  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 131  */
 132 static void sig_quit(int sig)
 133 {
 134 }
 135
 136 static void sig_int(int sig)
 137 {
 138         if (threads) {
 139                 log_info("\nfio: terminating on signal %d\n", sig);
 140                 fflush(stdout);
 141                 terminate_threads(TERMINATE_ALL);
 142         }
 143 }
 144
 145 static void sig_ill(int fio_unused sig)
 146 {
 147         if (!threads)
 148                 return;
 149
 150         log_err("fio: illegal instruction. your cpu does not support "
 151                 "the sse4.2 instruction for crc32c\n");
 152         terminate_threads(TERMINATE_ALL);
 153         exit(4);
 154 }
 155
 156 static void set_sig_handlers(void)
 157 {
 158         struct sigaction act;
 159
 160         memset(&act, 0, sizeof(act));
 161         act.sa_handler = sig_alrm;
 162         act.sa_flags = SA_RESTART;
 163         sigaction(SIGALRM, &act, NULL);
 164
 165         memset(&act, 0, sizeof(act));
 166         act.sa_handler = sig_int;
 167         act.sa_flags = SA_RESTART;
 168         sigaction(SIGINT, &act, NULL);
 169
 170         memset(&act, 0, sizeof(act));
 171         act.sa_handler = sig_ill;
 172         act.sa_flags = SA_RESTART;
 173         sigaction(SIGILL, &act, NULL);
 174
 175         memset(&act, 0, sizeof(act));
 176         act.sa_handler = sig_quit;
 177         act.sa_flags = SA_RESTART;
 178         sigaction(SIGQUIT, &act, NULL);
 179 }
 180
 181 /*
 182  * Check if we are above the minimum rate given.
 183  */
 184 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 185                             enum td_ddir ddir)
 186 {
 187         unsigned long long bytes = 0;
 188         unsigned long iops = 0;
 189         unsigned long spent;
 190         unsigned long rate;
 191         unsigned int ratemin = 0;
 192         unsigned int rate_iops = 0;
 193         unsigned int rate_iops_min = 0;
 194
 195         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 196                 return 0;
 197
 198         /*
 199          * allow a 2 second settle period in the beginning
 200          */
 201         if (mtime_since(&td->start, now) < 2000)
 202                 return 0;
 203
 204         iops += td->io_blocks[ddir];
 205         bytes += td->this_io_bytes[ddir];
 206         ratemin += td->o.ratemin[ddir];
 207         rate_iops += td->o.rate_iops[ddir];
 208         rate_iops_min += td->o.rate_iops_min[ddir];
 209
 210         /*
 211          * if rate blocks is set, sample is running
 212          */
 213         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 214                 spent = mtime_since(&td->lastrate[ddir], now);
 215                 if (spent < td->o.ratecycle)
 216                         return 0;
 217
 218                 if (td->o.rate[ddir]) {
 219                         /*
 220                          * check bandwidth specified rate
 221                          */
 222                         if (bytes < td->rate_bytes[ddir]) {
 223                                 log_err("%s: min rate %u not met\n", td->o.name,
 224                                                                 ratemin);
 225                                 return 1;
 226                         } else {
 227                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 228                                 if (rate < ratemin ||
 229                                     bytes < td->rate_bytes[ddir]) {
 230                                         log_err("%s: min rate %u not met, got"
 231                                                 " %luKB/sec\n", td->o.name,
 232                                                         ratemin, rate);
 233                                         return 1;
 234                                 }
 235                         }
 236                 } else {
 237                         /*
 238                          * checks iops specified rate
 239                          */
 240                         if (iops < rate_iops) {
 241                                 log_err("%s: min iops rate %u not met\n",
 242                                                 td->o.name, rate_iops);
 243                                 return 1;
 244                         } else {
 245                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 246                                 if (rate < rate_iops_min ||
 247                                     iops < td->rate_blocks[ddir]) {
 248                                         log_err("%s: min iops rate %u not met,"
 249                                                 " got %lu\n", td->o.name,
 250                                                         rate_iops_min, rate);
 251                                 }
 252                         }
 253                 }
 254         }
 255
 256         td->rate_bytes[ddir] = bytes;
 257         td->rate_blocks[ddir] = iops;
 258         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 259         return 0;
 260 }
 261
 262 static int check_min_rate(struct thread_data *td, struct timeval *now,
 263                           unsigned long *bytes_done)
 264 {
 265         int ret = 0;
 266
 267         if (bytes_done[0])
 268                 ret |= __check_min_rate(td, now, 0);
 269         if (bytes_done[1])
 270                 ret |= __check_min_rate(td, now, 1);
 271
 272         return ret;
 273 }
 274
 275 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 276 {
 277         if (!td->o.timeout)
 278                 return 0;
 279         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 280                 return 1;
 281
 282         return 0;
 283 }
 284
 285 /*
 286  * When job exits, we can cancel the in-flight IO if we are using async
 287  * io. Attempt to do so.
 288  */
 289 static void cleanup_pending_aio(struct thread_data *td)
 290 {
 291         struct flist_head *entry, *n;
 292         struct io_u *io_u;
 293         int r;
 294
 295         /*
 296          * get immediately available events, if any
 297          */
 298         r = io_u_queued_complete(td, 0, NULL);
 299         if (r < 0)
 300                 return;
 301
 302         /*
 303          * now cancel remaining active events
 304          */
 305         if (td->io_ops->cancel) {
 306                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 307                         io_u = flist_entry(entry, struct io_u, list);
 308
 309                         /*
 310                          * if the io_u isn't in flight, then that generally
 311                          * means someone leaked an io_u. complain but fix
 312                          * it up, so we don't stall here.
 313                          */
 314                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 315                                 log_err("fio: non-busy IO on busy list\n");
 316                                 put_io_u(td, io_u);
 317                         } else {
 318                                 r = td->io_ops->cancel(td, io_u);
 319                                 if (!r)
 320                                         put_io_u(td, io_u);
 321                         }
 322                 }
 323         }
 324
 325         if (td->cur_depth)
 326                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 327 }
 328
 329 /*
 330  * Helper to handle the final sync of a file. Works just like the normal
 331  * io path, just does everything sync.
 332  */
 333 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 334 {
 335         struct io_u *io_u = __get_io_u(td);
 336         int ret;
 337
 338         if (!io_u)
 339                 return 1;
 340
 341         io_u->ddir = DDIR_SYNC;
 342         io_u->file = f;
 343
 344         if (td_io_prep(td, io_u)) {
 345                 put_io_u(td, io_u);
 346                 return 1;
 347         }
 348
 349 requeue:
 350         ret = td_io_queue(td, io_u);
 351         if (ret < 0) {
 352                 td_verror(td, io_u->error, "td_io_queue");
 353                 put_io_u(td, io_u);
 354                 return 1;
 355         } else if (ret == FIO_Q_QUEUED) {
 356                 if (io_u_queued_complete(td, 1, NULL) < 0)
 357                         return 1;
 358         } else if (ret == FIO_Q_COMPLETED) {
 359                 if (io_u->error) {
 360                         td_verror(td, io_u->error, "td_io_queue");
 361                         return 1;
 362                 }
 363
 364                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 365                         return 1;
 366         } else if (ret == FIO_Q_BUSY) {
 367                 if (td_io_commit(td))
 368                         return 1;
 369                 goto requeue;
 370         }
 371
 372         return 0;
 373 }
 374
 375 static inline void update_tv_cache(struct thread_data *td)
 376 {
 377         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 378                 fio_gettime(&td->tv_cache, NULL);
 379 }
 380
 381 static int break_on_this_error(struct thread_data *td, int *retptr)
 382 {
 383         int ret = *retptr;
 384
 385         if (ret < 0 || td->error) {
 386                 int err;
 387
 388                 if (!td->o.continue_on_error)
 389                         return 1;
 390
 391                 if (ret < 0)
 392                         err = -ret;
 393                 else
 394                         err = td->error;
 395
 396                 if (td_non_fatal_error(err)) {
 397                         /*
 398                          * Continue with the I/Os in case of
 399                          * a non fatal error.
 400                          */
 401                         update_error_count(td, err);
 402                         td_clear_error(td);
 403                         *retptr = 0;
 404                         return 0;
 405                 } else if (td->o.fill_device && err == ENOSPC) {
 406                         /*
 407                          * We expect to hit this error if
 408                          * fill_device option is set.
 409                          */
 410                         td_clear_error(td);
 411                         td->terminate = 1;
 412                         return 1;
 413                 } else {
 414                         /*
 415                          * Stop the I/O in case of a fatal
 416                          * error.
 417                          */
 418                         update_error_count(td, err);
 419                         return 1;
 420                 }
 421         }
 422
 423         return 0;
 424 }
 425
 426 /*
 427  * The main verify engine. Runs over the writes we previously submitted,
 428  * reads the blocks back in, and checks the crc/md5 of the data.
 429  */
 430 static void do_verify(struct thread_data *td)
 431 {
 432         struct fio_file *f;
 433         struct io_u *io_u;
 434         int ret, min_events;
 435         unsigned int i;
 436
 437         dprint(FD_VERIFY, "starting loop\n");
 438
 439         /*
 440          * sync io first and invalidate cache, to make sure we really
 441          * read from disk.
 442          */
 443         for_each_file(td, f, i) {
 444                 if (!fio_file_open(f))
 445                         continue;
 446                 if (fio_io_sync(td, f))
 447                         break;
 448                 if (file_invalidate_cache(td, f))
 449                         break;
 450         }
 451
 452         if (td->error)
 453                 return;
 454
 455         td_set_runstate(td, TD_VERIFYING);
 456
 457         io_u = NULL;
 458         while (!td->terminate) {
 459                 int ret2, full;
 460
 461                 update_tv_cache(td);
 462
 463                 if (runtime_exceeded(td, &td->tv_cache)) {
 464                         td->terminate = 1;
 465                         break;
 466                 }
 467
 468                 io_u = __get_io_u(td);
 469                 if (!io_u)
 470                         break;
 471
 472                 if (get_next_verify(td, io_u)) {
 473                         put_io_u(td, io_u);
 474                         break;
 475                 }
 476
 477                 if (td_io_prep(td, io_u)) {
 478                         put_io_u(td, io_u);
 479                         break;
 480                 }
 481
 482                 if (td->o.verify_async)
 483                         io_u->end_io = verify_io_u_async;
 484                 else
 485                         io_u->end_io = verify_io_u;
 486
 487                 ret = td_io_queue(td, io_u);
 488                 switch (ret) {
 489                 case FIO_Q_COMPLETED:
 490                         if (io_u->error) {
 491                                 ret = -io_u->error;
 492                                 clear_io_u(td, io_u);
 493                         } else if (io_u->resid) {
 494                                 int bytes = io_u->xfer_buflen - io_u->resid;
 495                                 struct fio_file *f = io_u->file;
 496
 497                                 /*
 498                                  * zero read, fail
 499                                  */
 500                                 if (!bytes) {
 501                                         td_verror(td, EIO, "full resid");
 502                                         put_io_u(td, io_u);
 503                                         break;
 504                                 }
 505
 506                                 io_u->xfer_buflen = io_u->resid;
 507                                 io_u->xfer_buf += bytes;
 508                                 io_u->offset += bytes;
 509
 510                                 td->ts.short_io_u[io_u->ddir]++;
 511
 512                                 if (io_u->offset == f->real_file_size)
 513                                         goto sync_done;
 514
 515                                 requeue_io_u(td, &io_u);
 516                         } else {
 517 sync_done:
 518                                 ret = io_u_sync_complete(td, io_u, NULL);
 519                                 if (ret < 0)
 520                                         break;
 521                         }
 522                         continue;
 523                 case FIO_Q_QUEUED:
 524                         break;
 525                 case FIO_Q_BUSY:
 526                         requeue_io_u(td, &io_u);
 527                         ret2 = td_io_commit(td);
 528                         if (ret2 < 0)
 529                                 ret = ret2;
 530                         break;
 531                 default:
 532                         assert(ret < 0);
 533                         td_verror(td, -ret, "td_io_queue");
 534                         break;
 535                 }
 536
 537                 if (break_on_this_error(td, &ret))
 538                         break;
 539
 540                 /*
 541                  * if we can queue more, do so. but check if there are
 542                  * completed io_u's first.
 543                  */
 544                 full = queue_full(td) || ret == FIO_Q_BUSY;
 545                 if (full || !td->o.iodepth_batch_complete) {
 546                         min_events = min(td->o.iodepth_batch_complete,
 547                                          td->cur_depth);
 548                         if (full && !min_events)
 549                                 min_events = 1;
 550
 551                         do {
 552                                 /*
 553                                  * Reap required number of io units, if any,
 554                                  * and do the verification on them through
 555                                  * the callback handler
 556                                  */
 557                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 558                                         ret = -1;
 559                                         break;
 560                                 }
 561                         } while (full && (td->cur_depth > td->o.iodepth_low));
 562                 }
 563                 if (ret < 0)
 564                         break;
 565         }
 566
 567         if (!td->error) {
 568                 min_events = td->cur_depth;
 569
 570                 if (min_events)
 571                         ret = io_u_queued_complete(td, min_events, NULL);
 572         } else
 573                 cleanup_pending_aio(td);
 574
 575         td_set_runstate(td, TD_RUNNING);
 576
 577         dprint(FD_VERIFY, "exiting loop\n");
 578 }
 579
 580 /*
 581  * Main IO worker function. It retrieves io_u's to process and queues
 582  * and reaps them, checking for rate and errors along the way.
 583  */
 584 static void do_io(struct thread_data *td)
 585 {
 586         unsigned int i;
 587         int ret = 0;
 588
 589         if (in_ramp_time(td))
 590                 td_set_runstate(td, TD_RAMP);
 591         else
 592                 td_set_runstate(td, TD_RUNNING);
 593
 594         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 595                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 596                 struct timeval comp_time;
 597                 unsigned long bytes_done[2] = { 0, 0 };
 598                 int min_evts = 0;
 599                 struct io_u *io_u;
 600                 int ret2, full;
 601
 602                 if (td->terminate)
 603                         break;
 604
 605                 update_tv_cache(td);
 606
 607                 if (runtime_exceeded(td, &td->tv_cache)) {
 608                         td->terminate = 1;
 609                         break;
 610                 }
 611
 612                 io_u = get_io_u(td);
 613                 if (!io_u)
 614                         break;
 615
 616                 /*
 617                  * Add verification end_io handler, if asked to verify
 618                  * a previously written file.
 619                  */
 620                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 621                     !td_rw(td)) {
 622                         if (td->o.verify_async)
 623                                 io_u->end_io = verify_io_u_async;
 624                         else
 625                                 io_u->end_io = verify_io_u;
 626                         td_set_runstate(td, TD_VERIFYING);
 627                 } else if (in_ramp_time(td))
 628                         td_set_runstate(td, TD_RAMP);
 629                 else
 630                         td_set_runstate(td, TD_RUNNING);
 631
 632                 ret = td_io_queue(td, io_u);
 633                 switch (ret) {
 634                 case FIO_Q_COMPLETED:
 635                         if (io_u->error) {
 636                                 ret = -io_u->error;
 637                                 clear_io_u(td, io_u);
 638                         } else if (io_u->resid) {
 639                                 int bytes = io_u->xfer_buflen - io_u->resid;
 640                                 struct fio_file *f = io_u->file;
 641
 642                                 /*
 643                                  * zero read, fail
 644                                  */
 645                                 if (!bytes) {
 646                                         td_verror(td, EIO, "full resid");
 647                                         put_io_u(td, io_u);
 648                                         break;
 649                                 }
 650
 651                                 io_u->xfer_buflen = io_u->resid;
 652                                 io_u->xfer_buf += bytes;
 653                                 io_u->offset += bytes;
 654
 655                                 td->ts.short_io_u[io_u->ddir]++;
 656
 657                                 if (io_u->offset == f->real_file_size)
 658                                         goto sync_done;
 659
 660                                 requeue_io_u(td, &io_u);
 661                         } else {
 662 sync_done:
 663                                 if (__should_check_rate(td, 0) ||
 664                                     __should_check_rate(td, 1))
 665                                         fio_gettime(&comp_time, NULL);
 666
 667                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 668                                 if (ret < 0)
 669                                         break;
 670                         }
 671                         break;
 672                 case FIO_Q_QUEUED:
 673                         /*
 674                          * if the engine doesn't have a commit hook,
 675                          * the io_u is really queued. if it does have such
 676                          * a hook, it has to call io_u_queued() itself.
 677                          */
 678                         if (td->io_ops->commit == NULL)
 679                                 io_u_queued(td, io_u);
 680                         break;
 681                 case FIO_Q_BUSY:
 682                         requeue_io_u(td, &io_u);
 683                         ret2 = td_io_commit(td);
 684                         if (ret2 < 0)
 685                                 ret = ret2;
 686                         break;
 687                 default:
 688                         assert(ret < 0);
 689                         put_io_u(td, io_u);
 690                         break;
 691                 }
 692
 693                 if (break_on_this_error(td, &ret))
 694                         break;
 695
 696                 /*
 697                  * See if we need to complete some commands
 698                  */
 699                 full = queue_full(td) || ret == FIO_Q_BUSY;
 700                 if (full || !td->o.iodepth_batch_complete) {
 701                         min_evts = min(td->o.iodepth_batch_complete,
 702                                         td->cur_depth);
 703                         if (full && !min_evts)
 704                                 min_evts = 1;
 705
 706                         if (__should_check_rate(td, 0) ||
 707                             __should_check_rate(td, 1))
 708                                 fio_gettime(&comp_time, NULL);
 709
 710                         do {
 711                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 712                                 if (ret < 0)
 713                                         break;
 714
 715                         } while (full && (td->cur_depth > td->o.iodepth_low));
 716                 }
 717
 718                 if (ret < 0)
 719                         break;
 720                 if (!(bytes_done[0] + bytes_done[1]))
 721                         continue;
 722
 723                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 724                         if (check_min_rate(td, &comp_time, bytes_done)) {
 725                                 if (exitall_on_terminate)
 726                                         terminate_threads(td->groupid);
 727                                 td_verror(td, EIO, "check_min_rate");
 728                                 break;
 729                         }
 730                 }
 731
 732                 if (td->o.thinktime) {
 733                         unsigned long long b;
 734
 735                         b = td->io_blocks[0] + td->io_blocks[1];
 736                         if (!(b % td->o.thinktime_blocks)) {
 737                                 int left;
 738
 739                                 if (td->o.thinktime_spin)
 740                                         usec_spin(td->o.thinktime_spin);
 741
 742                                 left = td->o.thinktime - td->o.thinktime_spin;
 743                                 if (left)
 744                                         usec_sleep(td, left);
 745                         }
 746                 }
 747         }
 748
 749         if (td->o.fill_device && td->error == ENOSPC) {
 750                 td->error = 0;
 751                 td->terminate = 1;
 752         }
 753         if (!td->error) {
 754                 struct fio_file *f;
 755
 756                 i = td->cur_depth;
 757                 if (i)
 758                         ret = io_u_queued_complete(td, i, NULL);
 759
 760                 if (should_fsync(td) && td->o.end_fsync) {
 761                         td_set_runstate(td, TD_FSYNCING);
 762
 763                         for_each_file(td, f, i) {
 764                                 if (!fio_file_open(f))
 765                                         continue;
 766                                 fio_io_sync(td, f);
 767                         }
 768                 }
 769         } else
 770                 cleanup_pending_aio(td);
 771
 772         /*
 773          * stop job if we failed doing any IO
 774          */
 775         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 776                 td->done = 1;
 777 }
 778
 779 static void cleanup_io_u(struct thread_data *td)
 780 {
 781         struct flist_head *entry, *n;
 782         struct io_u *io_u;
 783
 784         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 785                 io_u = flist_entry(entry, struct io_u, list);
 786
 787                 flist_del(&io_u->list);
 788                 free(io_u);
 789         }
 790
 791         free_io_mem(td);
 792 }
 793
 794 static int init_io_u(struct thread_data *td)
 795 {
 796         struct io_u *io_u;
 797         unsigned int max_bs;
 798         int cl_align, i, max_units;
 799         char *p;
 800
 801         max_units = td->o.iodepth;
 802         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 803         td->orig_buffer_size = (unsigned long long) max_bs
 804                                         * (unsigned long long) max_units;
 805
 806         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 807                 unsigned long bs;
 808
 809                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 810                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 811         }
 812
 813         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 814                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 815                 return 1;
 816         }
 817
 818         if (allocate_io_mem(td))
 819                 return 1;
 820
 821         if (td->o.odirect || td->o.mem_align)
 822                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 823         else
 824                 p = td->orig_buffer;
 825
 826         cl_align = os_cache_line_size();
 827
 828         for (i = 0; i < max_units; i++) {
 829                 void *ptr;
 830
 831                 if (td->terminate)
 832                         return 1;
 833
 834                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 835                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 836                         break;
 837                 }
 838
 839                 io_u = ptr;
 840                 memset(io_u, 0, sizeof(*io_u));
 841                 INIT_FLIST_HEAD(&io_u->list);
 842                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 843
 844                 if (!(td->io_ops->flags & FIO_NOIO)) {
 845                         io_u->buf = p + max_bs * i;
 846                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 847
 848                         if (td_write(td) && !td->o.refill_buffers)
 849                                 io_u_fill_buffer(td, io_u, max_bs);
 850                 }
 851
 852                 io_u->index = i;
 853                 io_u->flags = IO_U_F_FREE;
 854                 flist_add(&io_u->list, &td->io_u_freelist);
 855         }
 856
 857         return 0;
 858 }
 859
 860 static int switch_ioscheduler(struct thread_data *td)
 861 {
 862         char tmp[256], tmp2[128];
 863         FILE *f;
 864         int ret;
 865
 866         if (td->io_ops->flags & FIO_DISKLESSIO)
 867                 return 0;
 868
 869         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 870
 871         f = fopen(tmp, "r+");
 872         if (!f) {
 873                 if (errno == ENOENT) {
 874                         log_err("fio: os or kernel doesn't support IO scheduler"
 875                                 " switching\n");
 876                         return 0;
 877                 }
 878                 td_verror(td, errno, "fopen iosched");
 879                 return 1;
 880         }
 881
 882         /*
 883          * Set io scheduler.
 884          */
 885         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 886         if (ferror(f) || ret != 1) {
 887                 td_verror(td, errno, "fwrite");
 888                 fclose(f);
 889                 return 1;
 890         }
 891
 892         rewind(f);
 893
 894         /*
 895          * Read back and check that the selected scheduler is now the default.
 896          */
 897         ret = fread(tmp, 1, sizeof(tmp), f);
 898         if (ferror(f) || ret < 0) {
 899                 td_verror(td, errno, "fread");
 900                 fclose(f);
 901                 return 1;
 902         }
 903
 904         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 905         if (!strstr(tmp, tmp2)) {
 906                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 907                 td_verror(td, EINVAL, "iosched_switch");
 908                 fclose(f);
 909                 return 1;
 910         }
 911
 912         fclose(f);
 913         return 0;
 914 }
 915
 916 static int keep_running(struct thread_data *td)
 917 {
 918         unsigned long long io_done;
 919
 920         if (td->done)
 921                 return 0;
 922         if (td->o.time_based)
 923                 return 1;
 924         if (td->o.loops) {
 925                 td->o.loops--;
 926                 return 1;
 927         }
 928
 929         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 930                         + td->io_skip_bytes;
 931         if (io_done < td->o.size)
 932                 return 1;
 933
 934         return 0;
 935 }
 936
 937 static void reset_io_counters(struct thread_data *td)
 938 {
 939         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 940         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 941         td->zone_bytes = 0;
 942         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 943         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 944
 945         td->last_was_sync = 0;
 946
 947         /*
 948          * reset file done count if we are to start over
 949          */
 950         if (td->o.time_based || td->o.loops)
 951                 td->nr_done_files = 0;
 952
 953         /*
 954          * Set the same seed to get repeatable runs
 955          */
 956         td_fill_rand_seeds(td);
 957 }
 958
 959 void reset_all_stats(struct thread_data *td)
 960 {
 961         struct timeval tv;
 962         int i;
 963
 964         reset_io_counters(td);
 965
 966         for (i = 0; i < 2; i++) {
 967                 td->io_bytes[i] = 0;
 968                 td->io_blocks[i] = 0;
 969                 td->io_issues[i] = 0;
 970                 td->ts.total_io_u[i] = 0;
 971         }
 972
 973         fio_gettime(&tv, NULL);
 974         memcpy(&td->epoch, &tv, sizeof(tv));
 975         memcpy(&td->start, &tv, sizeof(tv));
 976 }
 977
 978 static void clear_io_state(struct thread_data *td)
 979 {
 980         struct fio_file *f;
 981         unsigned int i;
 982
 983         reset_io_counters(td);
 984
 985         close_files(td);
 986         for_each_file(td, f, i)
 987                 fio_file_clear_done(f);
 988 }
 989
 990 static int exec_string(const char *string)
 991 {
 992         int ret, newlen = strlen(string) + 1 + 8;
 993         char *str;
 994
 995         str = malloc(newlen);
 996         sprintf(str, "sh -c %s", string);
 997
 998         ret = system(str);
 999         if (ret == -1)
1000                 log_err("fio: exec of cmd <%s> failed\n", str);
1001
1002         free(str);
1003         return ret;
1004 }
1005
1006 /*
1007  * Entry point for the thread based jobs. The process based jobs end up
1008  * here as well, after a little setup.
1009  */
1010 static void *thread_main(void *data)
1011 {
1012         unsigned long long runtime[2], elapsed;
1013         struct thread_data *td = data;
1014         pthread_condattr_t attr;
1015         int clear_state;
1016
1017         if (!td->o.use_thread)
1018                 setsid();
1019
1020         td->pid = getpid();
1021
1022         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1023
1024         INIT_FLIST_HEAD(&td->io_u_freelist);
1025         INIT_FLIST_HEAD(&td->io_u_busylist);
1026         INIT_FLIST_HEAD(&td->io_u_requeues);
1027         INIT_FLIST_HEAD(&td->io_log_list);
1028         INIT_FLIST_HEAD(&td->io_hist_list);
1029         INIT_FLIST_HEAD(&td->verify_list);
1030         pthread_mutex_init(&td->io_u_lock, NULL);
1031         td->io_hist_tree = RB_ROOT;
1032
1033         pthread_condattr_init(&attr);
1034         pthread_cond_init(&td->verify_cond, &attr);
1035         pthread_cond_init(&td->free_cond, &attr);
1036
1037         td_set_runstate(td, TD_INITIALIZED);
1038         dprint(FD_MUTEX, "up startup_mutex\n");
1039         fio_mutex_up(startup_mutex);
1040         dprint(FD_MUTEX, "wait on td->mutex\n");
1041         fio_mutex_down(td->mutex);
1042         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1043
1044         /*
1045          * the ->mutex mutex is now no longer used, close it to avoid
1046          * eating a file descriptor
1047          */
1048         fio_mutex_remove(td->mutex);
1049
1050         if (td->o.uid != -1U && setuid(td->o.uid)) {
1051                 td_verror(td, errno, "setuid");
1052                 goto err;
1053         }
1054         if (td->o.gid != -1U && setgid(td->o.gid)) {
1055                 td_verror(td, errno, "setgid");
1056                 goto err;
1057         }
1058
1059         /*
1060          * May alter parameters that init_io_u() will use, so we need to
1061          * do this first.
1062          */
1063         if (init_iolog(td))
1064                 goto err;
1065
1066         if (init_io_u(td))
1067                 goto err;
1068
1069         if (td->o.verify_async && verify_async_init(td))
1070                 goto err;
1071
1072         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1073                 td_verror(td, errno, "cpu_set_affinity");
1074                 goto err;
1075         }
1076
1077         /*
1078          * If we have a gettimeofday() thread, make sure we exclude that
1079          * thread from this job
1080          */
1081         if (td->o.gtod_cpu) {
1082                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1083                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1084                         td_verror(td, errno, "cpu_set_affinity");
1085                         goto err;
1086                 }
1087         }
1088
1089         if (td->ioprio_set) {
1090                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1091                         td_verror(td, errno, "ioprio_set");
1092                         goto err;
1093                 }
1094         }
1095
1096         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1097                 goto err;
1098
1099         if (nice(td->o.nice) == -1) {
1100                 td_verror(td, errno, "nice");
1101                 goto err;
1102         }
1103
1104         if (td->o.ioscheduler && switch_ioscheduler(td))
1105                 goto err;
1106
1107         if (!td->o.create_serialize && setup_files(td))
1108                 goto err;
1109
1110         if (td_io_init(td))
1111                 goto err;
1112
1113         if (init_random_map(td))
1114                 goto err;
1115
1116         if (td->o.exec_prerun) {
1117                 if (exec_string(td->o.exec_prerun))
1118                         goto err;
1119         }
1120
1121         if (td->o.pre_read) {
1122                 if (pre_read_files(td) < 0)
1123                         goto err;
1124         }
1125
1126         fio_gettime(&td->epoch, NULL);
1127         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1128
1129         runtime[0] = runtime[1] = 0;
1130         clear_state = 0;
1131         while (keep_running(td)) {
1132                 fio_gettime(&td->start, NULL);
1133                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1134                                 sizeof(td->start));
1135                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1136                                 sizeof(td->start));
1137                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1138
1139                 if (td->o.ratemin[0] || td->o.ratemin[1])
1140                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1141                                                         sizeof(td->lastrate));
1142
1143                 if (clear_state)
1144                         clear_io_state(td);
1145
1146                 prune_io_piece_log(td);
1147
1148                 do_io(td);
1149
1150                 clear_state = 1;
1151
1152                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1153                         elapsed = utime_since_now(&td->start);
1154                         runtime[DDIR_READ] += elapsed;
1155                 }
1156                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1157                         elapsed = utime_since_now(&td->start);
1158                         runtime[DDIR_WRITE] += elapsed;
1159                 }
1160
1161                 if (td->error || td->terminate)
1162                         break;
1163
1164                 if (!td->o.do_verify ||
1165                     td->o.verify == VERIFY_NONE ||
1166                     (td->io_ops->flags & FIO_UNIDIR))
1167                         continue;
1168
1169                 clear_io_state(td);
1170
1171                 fio_gettime(&td->start, NULL);
1172
1173                 do_verify(td);
1174
1175                 runtime[DDIR_READ] += utime_since_now(&td->start);
1176
1177                 if (td->error || td->terminate)
1178                         break;
1179         }
1180
1181         update_rusage_stat(td);
1182         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1183         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1184         td->ts.total_run_time = mtime_since_now(&td->epoch);
1185         td->ts.io_bytes[0] = td->io_bytes[0];
1186         td->ts.io_bytes[1] = td->io_bytes[1];
1187
1188         fio_mutex_down(writeout_mutex);
1189         if (td->ts.bw_log) {
1190                 if (td->o.bw_log_file) {
1191                         finish_log_named(td, td->ts.bw_log,
1192                                                 td->o.bw_log_file, "bw");
1193                 } else
1194                         finish_log(td, td->ts.bw_log, "bw");
1195         }
1196         if (td->ts.slat_log) {
1197                 if (td->o.lat_log_file) {
1198                         finish_log_named(td, td->ts.slat_log,
1199                                                 td->o.lat_log_file, "slat");
1200                 } else
1201                         finish_log(td, td->ts.slat_log, "slat");
1202         }
1203         if (td->ts.clat_log) {
1204                 if (td->o.lat_log_file) {
1205                         finish_log_named(td, td->ts.clat_log,
1206                                                 td->o.lat_log_file, "clat");
1207                 } else
1208                         finish_log(td, td->ts.clat_log, "clat");
1209         }
1210         fio_mutex_up(writeout_mutex);
1211         if (td->o.exec_postrun)
1212                 exec_string(td->o.exec_postrun);
1213
1214         if (exitall_on_terminate)
1215                 terminate_threads(td->groupid);
1216
1217 err:
1218         if (td->error)
1219                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1220                                                         td->verror);
1221
1222         if (td->o.verify_async)
1223                 verify_async_exit(td);
1224
1225         close_and_free_files(td);
1226         close_ioengine(td);
1227         cleanup_io_u(td);
1228         cgroup_shutdown(td, &cgroup_mnt);
1229
1230         if (td->o.cpumask_set) {
1231                 int ret = fio_cpuset_exit(&td->o.cpumask);
1232
1233                 td_verror(td, ret, "fio_cpuset_exit");
1234         }
1235
1236         /*
1237          * do this very late, it will log file closing as well
1238          */
1239         if (td->o.write_iolog_file)
1240                 write_iolog_close(td);
1241
1242         options_mem_free(td);
1243         td_set_runstate(td, TD_EXITED);
1244         return (void *) (unsigned long) td->error;
1245 }
1246
1247 /*
1248  * We cannot pass the td data into a forked process, so attach the td and
1249  * pass it to the thread worker.
1250  */
1251 static int fork_main(int shmid, int offset)
1252 {
1253         struct thread_data *td;
1254         void *data, *ret;
1255
1256         data = shmat(shmid, NULL, 0);
1257         if (data == (void *) -1) {
1258                 int __err = errno;
1259
1260                 perror("shmat");
1261                 return __err;
1262         }
1263
1264         td = data + offset * sizeof(struct thread_data);
1265         ret = thread_main(td);
1266         shmdt(data);
1267         return (int) (unsigned long) ret;
1268 }
1269
1270 /*
1271  * Run over the job map and reap the threads that have exited, if any.
1272  */
1273 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1274 {
1275         struct thread_data *td;
1276         int i, cputhreads, realthreads, pending, status, ret;
1277
1278         /*
1279          * reap exited threads (TD_EXITED -> TD_REAPED)
1280          */
1281         realthreads = pending = cputhreads = 0;
1282         for_each_td(td, i) {
1283                 int flags = 0;
1284
1285                 /*
1286                  * ->io_ops is NULL for a thread that has closed its
1287                  * io engine
1288                  */
1289                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1290                         cputhreads++;
1291                 else
1292                         realthreads++;
1293
1294                 if (!td->pid) {
1295                         pending++;
1296                         continue;
1297                 }
1298                 if (td->runstate == TD_REAPED)
1299                         continue;
1300                 if (td->o.use_thread) {
1301                         if (td->runstate == TD_EXITED) {
1302                                 td_set_runstate(td, TD_REAPED);
1303                                 goto reaped;
1304                         }
1305                         continue;
1306                 }
1307
1308                 flags = WNOHANG;
1309                 if (td->runstate == TD_EXITED)
1310                         flags = 0;
1311
1312                 /*
1313                  * check if someone quit or got killed in an unusual way
1314                  */
1315                 ret = waitpid(td->pid, &status, flags);
1316                 if (ret < 0) {
1317                         if (errno == ECHILD) {
1318                                 log_err("fio: pid=%d disappeared %d\n",
1319                                                 (int) td->pid, td->runstate);
1320                                 td_set_runstate(td, TD_REAPED);
1321                                 goto reaped;
1322                         }
1323                         perror("waitpid");
1324                 } else if (ret == td->pid) {
1325                         if (WIFSIGNALED(status)) {
1326                                 int sig = WTERMSIG(status);
1327
1328                                 if (sig != SIGQUIT)
1329                                         log_err("fio: pid=%d, got signal=%d\n",
1330                                                         (int) td->pid, sig);
1331                                 td_set_runstate(td, TD_REAPED);
1332                                 goto reaped;
1333                         }
1334                         if (WIFEXITED(status)) {
1335                                 if (WEXITSTATUS(status) && !td->error)
1336                                         td->error = WEXITSTATUS(status);
1337
1338                                 td_set_runstate(td, TD_REAPED);
1339                                 goto reaped;
1340                         }
1341                 }
1342
1343                 /*
1344                  * thread is not dead, continue
1345                  */
1346                 pending++;
1347                 continue;
1348 reaped:
1349                 (*nr_running)--;
1350                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1351                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1352                 if (!td->pid)
1353                         pending--;
1354
1355                 if (td->error)
1356                         exit_value++;
1357
1358                 done_secs += mtime_since_now(&td->epoch) / 1000;
1359         }
1360
1361         if (*nr_running == cputhreads && !pending && realthreads)
1362                 terminate_threads(TERMINATE_ALL);
1363 }
1364
1365 static void *gtod_thread_main(void *data)
1366 {
1367         fio_mutex_up(startup_mutex);
1368
1369         /*
1370          * As long as we have jobs around, update the clock. It would be nice
1371          * to have some way of NOT hammering that CPU with gettimeofday(),
1372          * but I'm not sure what to use outside of a simple CPU nop to relax
1373          * it - we don't want to lose precision.
1374          */
1375         while (threads) {
1376                 fio_gtod_update();
1377                 nop;
1378         }
1379
1380         return NULL;
1381 }
1382
1383 static int fio_start_gtod_thread(void)
1384 {
1385         int ret;
1386
1387         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1388         if (ret) {
1389                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1390                 return 1;
1391         }
1392
1393         ret = pthread_detach(gtod_thread);
1394         if (ret) {
1395                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1396                 return 1;
1397         }
1398
1399         dprint(FD_MUTEX, "wait on startup_mutex\n");
1400         fio_mutex_down(startup_mutex);
1401         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1402         return 0;
1403 }
1404
1405 /*
1406  * Main function for kicking off and reaping jobs, as needed.
1407  */
1408 static void run_threads(void)
1409 {
1410         struct thread_data *td;
1411         unsigned long spent;
1412         int i, todo, nr_running, m_rate, t_rate, nr_started;
1413
1414         if (fio_pin_memory())
1415                 return;
1416
1417         if (fio_gtod_offload && fio_start_gtod_thread())
1418                 return;
1419
1420         if (!terse_output) {
1421                 log_info("Starting ");
1422                 if (nr_thread)
1423                         log_info("%d thread%s", nr_thread,
1424                                                 nr_thread > 1 ? "s" : "");
1425                 if (nr_process) {
1426                         if (nr_thread)
1427                                 printf(" and ");
1428                         log_info("%d process%s", nr_process,
1429                                                 nr_process > 1 ? "es" : "");
1430                 }
1431                 log_info("\n");
1432                 fflush(stdout);
1433         }
1434
1435         set_sig_handlers();
1436
1437         todo = thread_number;
1438         nr_running = 0;
1439         nr_started = 0;
1440         m_rate = t_rate = 0;
1441
1442         for_each_td(td, i) {
1443                 print_status_init(td->thread_number - 1);
1444
1445                 if (!td->o.create_serialize) {
1446                         init_disk_util(td);
1447                         continue;
1448                 }
1449
1450                 /*
1451                  * do file setup here so it happens sequentially,
1452                  * we don't want X number of threads getting their
1453                  * client data interspersed on disk
1454                  */
1455                 if (setup_files(td)) {
1456                         exit_value++;
1457                         if (td->error)
1458                                 log_err("fio: pid=%d, err=%d/%s\n",
1459                                         (int) td->pid, td->error, td->verror);
1460                         td_set_runstate(td, TD_REAPED);
1461                         todo--;
1462                 } else {
1463                         struct fio_file *f;
1464                         unsigned int i;
1465
1466                         /*
1467                          * for sharing to work, each job must always open
1468                          * its own files. so close them, if we opened them
1469                          * for creation
1470                          */
1471                         for_each_file(td, f, i) {
1472                                 if (fio_file_open(f))
1473                                         td_io_close_file(td, f);
1474                         }
1475                 }
1476
1477                 init_disk_util(td);
1478         }
1479
1480         set_genesis_time();
1481
1482         while (todo) {
1483                 struct thread_data *map[MAX_JOBS];
1484                 struct timeval this_start;
1485                 int this_jobs = 0, left;
1486
1487                 /*
1488                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1489                  */
1490                 for_each_td(td, i) {
1491                         if (td->runstate != TD_NOT_CREATED)
1492                                 continue;
1493
1494                         /*
1495                          * never got a chance to start, killed by other
1496                          * thread for some reason
1497                          */
1498                         if (td->terminate) {
1499                                 todo--;
1500                                 continue;
1501                         }
1502
1503                         if (td->o.start_delay) {
1504                                 spent = mtime_since_genesis();
1505
1506                                 if (td->o.start_delay * 1000 > spent)
1507                                         continue;
1508                         }
1509
1510                         if (td->o.stonewall && (nr_started || nr_running)) {
1511                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1512                                                         td->o.name);
1513                                 break;
1514                         }
1515
1516                         /*
1517                          * Set state to created. Thread will transition
1518                          * to TD_INITIALIZED when it's done setting up.
1519                          */
1520                         td_set_runstate(td, TD_CREATED);
1521                         map[this_jobs++] = td;
1522                         nr_started++;
1523
1524                         if (td->o.use_thread) {
1525                                 int ret;
1526
1527                                 dprint(FD_PROCESS, "will pthread_create\n");
1528                                 ret = pthread_create(&td->thread, NULL,
1529                                                         thread_main, td);
1530                                 if (ret) {
1531                                         log_err("pthread_create: %s\n",
1532                                                         strerror(ret));
1533                                         nr_started--;
1534                                         break;
1535                                 }
1536                                 ret = pthread_detach(td->thread);
1537                                 if (ret)
1538                                         log_err("pthread_detach: %s",
1539                                                         strerror(ret));
1540                         } else {
1541                                 pid_t pid;
1542                                 dprint(FD_PROCESS, "will fork\n");
1543                                 pid = fork();
1544                                 if (!pid) {
1545                                         int ret = fork_main(shm_id, i);
1546
1547                                         _exit(ret);
1548                                 } else if (i == fio_debug_jobno)
1549                                         *fio_debug_jobp = pid;
1550                         }
1551                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1552                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1553                                 log_err("fio: job startup hung? exiting.\n");
1554                                 terminate_threads(TERMINATE_ALL);
1555                                 fio_abort = 1;
1556                                 nr_started--;
1557                                 break;
1558                         }
1559                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1560                 }
1561
1562                 /*
1563                  * Wait for the started threads to transition to
1564                  * TD_INITIALIZED.
1565                  */
1566                 fio_gettime(&this_start, NULL);
1567                 left = this_jobs;
1568                 while (left && !fio_abort) {
1569                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1570                                 break;
1571
1572                         usleep(100000);
1573
1574                         for (i = 0; i < this_jobs; i++) {
1575                                 td = map[i];
1576                                 if (!td)
1577                                         continue;
1578                                 if (td->runstate == TD_INITIALIZED) {
1579                                         map[i] = NULL;
1580                                         left--;
1581                                 } else if (td->runstate >= TD_EXITED) {
1582                                         map[i] = NULL;
1583                                         left--;
1584                                         todo--;
1585                                         nr_running++; /* work-around... */
1586                                 }
1587                         }
1588                 }
1589
1590                 if (left) {
1591                         log_err("fio: %d jobs failed to start\n", left);
1592                         for (i = 0; i < this_jobs; i++) {
1593                                 td = map[i];
1594                                 if (!td)
1595                                         continue;
1596                                 kill(td->pid, SIGTERM);
1597                         }
1598                         break;
1599                 }
1600
1601                 /*
1602                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1603                  */
1604                 for_each_td(td, i) {
1605                         if (td->runstate != TD_INITIALIZED)
1606                                 continue;
1607
1608                         if (in_ramp_time(td))
1609                                 td_set_runstate(td, TD_RAMP);
1610                         else
1611                                 td_set_runstate(td, TD_RUNNING);
1612                         nr_running++;
1613                         nr_started--;
1614                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1615                         t_rate += td->o.rate[0] + td->o.rate[1];
1616                         todo--;
1617                         fio_mutex_up(td->mutex);
1618                 }
1619
1620                 reap_threads(&nr_running, &t_rate, &m_rate);
1621
1622                 if (todo)
1623                         usleep(100000);
1624         }
1625
1626         while (nr_running) {
1627                 reap_threads(&nr_running, &t_rate, &m_rate);
1628                 usleep(10000);
1629         }
1630
1631         update_io_ticks();
1632         fio_unpin_memory();
1633 }
1634
1635 int main(int argc, char *argv[])
1636 {
1637         long ps;
1638
1639         sinit();
1640         init_rand(&__fio_rand_state);
1641
1642         /*
1643          * We need locale for number printing, if it isn't set then just
1644          * go with the US format.
1645          */
1646         if (!getenv("LC_NUMERIC"))
1647                 setlocale(LC_NUMERIC, "en_US");
1648
1649         ps = sysconf(_SC_PAGESIZE);
1650         if (ps < 0) {
1651                 log_err("Failed to get page size\n");
1652                 return 1;
1653         }
1654
1655         page_size = ps;
1656         page_mask = ps - 1;
1657
1658         fio_keywords_init();
1659
1660         if (parse_options(argc, argv))
1661                 return 1;
1662
1663         if (exec_profile && load_profile(exec_profile))
1664                 return 1;
1665
1666         if (!thread_number)
1667                 return 0;
1668
1669         if (write_bw_log) {
1670                 setup_log(&agg_io_log[DDIR_READ]);
1671                 setup_log(&agg_io_log[DDIR_WRITE]);
1672         }
1673
1674         startup_mutex = fio_mutex_init(0);
1675         writeout_mutex = fio_mutex_init(1);
1676
1677         set_genesis_time();
1678
1679         status_timer_arm();
1680
1681         cgroup_list = smalloc(sizeof(*cgroup_list));
1682         INIT_FLIST_HEAD(cgroup_list);
1683
1684         run_threads();
1685
1686         if (!fio_abort) {
1687                 show_run_stats();
1688                 if (write_bw_log) {
1689                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1690                         __finish_log(agg_io_log[DDIR_WRITE],
1691                                         "agg-write_bw.log");
1692                 }
1693         }
1694
1695         cgroup_kill(cgroup_list);
1696         sfree(cgroup_list);
1697         sfree(cgroup_mnt);
1698
1699         fio_mutex_remove(startup_mutex);
1700         fio_mutex_remove(writeout_mutex);
1701         return exit_value;
1702 }