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