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                                                 " %luKiB/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 = td->o.iodepth_batch_complete;
 540                         if (full && !min_events)
 541                                 min_events = 1;
 542
 543                         do {
 544                                 /*
 545                                  * Reap required number of io units, if any,
 546                                  * and do the verification on them through
 547                                  * the callback handler
 548                                  */
 549                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 550                                         ret = -1;
 551                                         break;
 552                                 }
 553                         } while (full && (td->cur_depth > td->o.iodepth_low));
 554                 }
 555                 if (ret < 0)
 556                         break;
 557         }
 558
 559         if (!td->error) {
 560                 min_events = td->cur_depth;
 561
 562                 if (min_events)
 563                         ret = io_u_queued_complete(td, min_events, NULL);
 564         } else
 565                 cleanup_pending_aio(td);
 566
 567         td_set_runstate(td, TD_RUNNING);
 568 }
 569
 570 /*
 571  * Main IO worker function. It retrieves io_u's to process and queues
 572  * and reaps them, checking for rate and errors along the way.
 573  */
 574 static void do_io(struct thread_data *td)
 575 {
 576         unsigned int i;
 577         int ret = 0;
 578
 579         if (in_ramp_time(td))
 580                 td_set_runstate(td, TD_RAMP);
 581         else
 582                 td_set_runstate(td, TD_RUNNING);
 583
 584         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 585                 struct timeval comp_time;
 586                 unsigned long bytes_done[2] = { 0, 0 };
 587                 int min_evts = 0;
 588                 struct io_u *io_u;
 589                 int ret2, full;
 590
 591                 if (td->terminate)
 592                         break;
 593
 594                 update_tv_cache(td);
 595
 596                 if (runtime_exceeded(td, &td->tv_cache)) {
 597                         td->terminate = 1;
 598                         break;
 599                 }
 600
 601                 io_u = get_io_u(td);
 602                 if (!io_u)
 603                         break;
 604
 605                 /*
 606                  * Add verification end_io handler, if asked to verify
 607                  * a previously written file.
 608                  */
 609                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 610                         if (td->o.verify_async)
 611                                 io_u->end_io = verify_io_u_async;
 612                         else
 613                                 io_u->end_io = verify_io_u;
 614                         td_set_runstate(td, TD_VERIFYING);
 615                 } else if (in_ramp_time(td))
 616                         td_set_runstate(td, TD_RAMP);
 617                 else
 618                         td_set_runstate(td, TD_RUNNING);
 619
 620                 ret = td_io_queue(td, io_u);
 621                 switch (ret) {
 622                 case FIO_Q_COMPLETED:
 623                         if (io_u->error) {
 624                                 ret = -io_u->error;
 625                                 clear_io_u(td, io_u);
 626                         } else if (io_u->resid) {
 627                                 int bytes = io_u->xfer_buflen - io_u->resid;
 628                                 struct fio_file *f = io_u->file;
 629
 630                                 /*
 631                                  * zero read, fail
 632                                  */
 633                                 if (!bytes) {
 634                                         td_verror(td, EIO, "full resid");
 635                                         put_io_u(td, io_u);
 636                                         break;
 637                                 }
 638
 639                                 io_u->xfer_buflen = io_u->resid;
 640                                 io_u->xfer_buf += bytes;
 641                                 io_u->offset += bytes;
 642
 643                                 td->ts.short_io_u[io_u->ddir]++;
 644
 645                                 if (io_u->offset == f->real_file_size)
 646                                         goto sync_done;
 647
 648                                 requeue_io_u(td, &io_u);
 649                         } else {
 650 sync_done:
 651                                 if (__should_check_rate(td, 0) ||
 652                                     __should_check_rate(td, 1))
 653                                         fio_gettime(&comp_time, NULL);
 654
 655                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 656                                 if (ret < 0)
 657                                         break;
 658                         }
 659                         break;
 660                 case FIO_Q_QUEUED:
 661                         /*
 662                          * if the engine doesn't have a commit hook,
 663                          * the io_u is really queued. if it does have such
 664                          * a hook, it has to call io_u_queued() itself.
 665                          */
 666                         if (td->io_ops->commit == NULL)
 667                                 io_u_queued(td, io_u);
 668                         break;
 669                 case FIO_Q_BUSY:
 670                         requeue_io_u(td, &io_u);
 671                         ret2 = td_io_commit(td);
 672                         if (ret2 < 0)
 673                                 ret = ret2;
 674                         break;
 675                 default:
 676                         assert(ret < 0);
 677                         put_io_u(td, io_u);
 678                         break;
 679                 }
 680
 681                 if (break_on_this_error(td, &ret))
 682                         break;
 683
 684                 /*
 685                  * See if we need to complete some commands
 686                  */
 687                 full = queue_full(td) || ret == FIO_Q_BUSY;
 688                 if (full || !td->o.iodepth_batch_complete) {
 689                         min_evts = td->o.iodepth_batch_complete;
 690                         if (full && !min_evts)
 691                                 min_evts = 1;
 692
 693                         if (__should_check_rate(td, 0) ||
 694                             __should_check_rate(td, 1))
 695                                 fio_gettime(&comp_time, NULL);
 696
 697                         do {
 698                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 699                                 if (ret < 0)
 700                                         break;
 701
 702                         } while (full && (td->cur_depth > td->o.iodepth_low));
 703                 }
 704
 705                 if (ret < 0)
 706                         break;
 707                 if (!(bytes_done[0] + bytes_done[1]))
 708                         continue;
 709
 710                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 711                         if (check_min_rate(td, &comp_time, bytes_done)) {
 712                                 if (exitall_on_terminate)
 713                                         terminate_threads(td->groupid);
 714                                 td_verror(td, EIO, "check_min_rate");
 715                                 break;
 716                         }
 717                 }
 718
 719                 if (td->o.thinktime) {
 720                         unsigned long long b;
 721
 722                         b = td->io_blocks[0] + td->io_blocks[1];
 723                         if (!(b % td->o.thinktime_blocks)) {
 724                                 int left;
 725
 726                                 if (td->o.thinktime_spin)
 727                                         usec_spin(td->o.thinktime_spin);
 728
 729                                 left = td->o.thinktime - td->o.thinktime_spin;
 730                                 if (left)
 731                                         usec_sleep(td, left);
 732                         }
 733                 }
 734         }
 735
 736         if (td->o.fill_device && td->error == ENOSPC) {
 737                 td->error = 0;
 738                 td->terminate = 1;
 739         }
 740         if (!td->error) {
 741                 struct fio_file *f;
 742
 743                 i = td->cur_depth;
 744                 if (i)
 745                         ret = io_u_queued_complete(td, i, NULL);
 746
 747                 if (should_fsync(td) && td->o.end_fsync) {
 748                         td_set_runstate(td, TD_FSYNCING);
 749
 750                         for_each_file(td, f, i) {
 751                                 if (!fio_file_open(f))
 752                                         continue;
 753                                 fio_io_sync(td, f);
 754                         }
 755                 }
 756         } else
 757                 cleanup_pending_aio(td);
 758
 759         /*
 760          * stop job if we failed doing any IO
 761          */
 762         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 763                 td->done = 1;
 764 }
 765
 766 static void cleanup_io_u(struct thread_data *td)
 767 {
 768         struct flist_head *entry, *n;
 769         struct io_u *io_u;
 770
 771         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 772                 io_u = flist_entry(entry, struct io_u, list);
 773
 774                 flist_del(&io_u->list);
 775                 free(io_u);
 776         }
 777
 778         free_io_mem(td);
 779 }
 780
 781 static int init_io_u(struct thread_data *td)
 782 {
 783         struct io_u *io_u;
 784         unsigned int max_bs;
 785         int cl_align, i, max_units;
 786         char *p;
 787
 788         max_units = td->o.iodepth;
 789         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 790         td->orig_buffer_size = (unsigned long long) max_bs
 791                                         * (unsigned long long) max_units;
 792
 793         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 794                 unsigned long bs;
 795
 796                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 797                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 798         }
 799
 800         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 801                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 802                 return 1;
 803         }
 804
 805         if (allocate_io_mem(td))
 806                 return 1;
 807
 808         if (td->o.odirect || td->o.mem_align)
 809                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 810         else
 811                 p = td->orig_buffer;
 812
 813         cl_align = os_cache_line_size();
 814
 815         for (i = 0; i < max_units; i++) {
 816                 void *ptr;
 817
 818                 if (td->terminate)
 819                         return 1;
 820
 821                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 822                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 823                         break;
 824                 }
 825
 826                 io_u = ptr;
 827                 memset(io_u, 0, sizeof(*io_u));
 828                 INIT_FLIST_HEAD(&io_u->list);
 829                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 830
 831                 if (!(td->io_ops->flags & FIO_NOIO)) {
 832                         io_u->buf = p + max_bs * i;
 833                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 834
 835                         if (td_write(td) && !td->o.refill_buffers)
 836                                 io_u_fill_buffer(td, io_u, max_bs);
 837                 }
 838
 839                 io_u->index = i;
 840                 io_u->flags = IO_U_F_FREE;
 841                 flist_add(&io_u->list, &td->io_u_freelist);
 842         }
 843
 844         return 0;
 845 }
 846
 847 static int switch_ioscheduler(struct thread_data *td)
 848 {
 849         char tmp[256], tmp2[128];
 850         FILE *f;
 851         int ret;
 852
 853         if (td->io_ops->flags & FIO_DISKLESSIO)
 854                 return 0;
 855
 856         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 857
 858         f = fopen(tmp, "r+");
 859         if (!f) {
 860                 if (errno == ENOENT) {
 861                         log_err("fio: os or kernel doesn't support IO scheduler"
 862                                 " switching\n");
 863                         return 0;
 864                 }
 865                 td_verror(td, errno, "fopen iosched");
 866                 return 1;
 867         }
 868
 869         /*
 870          * Set io scheduler.
 871          */
 872         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 873         if (ferror(f) || ret != 1) {
 874                 td_verror(td, errno, "fwrite");
 875                 fclose(f);
 876                 return 1;
 877         }
 878
 879         rewind(f);
 880
 881         /*
 882          * Read back and check that the selected scheduler is now the default.
 883          */
 884         ret = fread(tmp, 1, sizeof(tmp), f);
 885         if (ferror(f) || ret < 0) {
 886                 td_verror(td, errno, "fread");
 887                 fclose(f);
 888                 return 1;
 889         }
 890
 891         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 892         if (!strstr(tmp, tmp2)) {
 893                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 894                 td_verror(td, EINVAL, "iosched_switch");
 895                 fclose(f);
 896                 return 1;
 897         }
 898
 899         fclose(f);
 900         return 0;
 901 }
 902
 903 static int keep_running(struct thread_data *td)
 904 {
 905         unsigned long long io_done;
 906
 907         if (td->done)
 908                 return 0;
 909         if (td->o.time_based)
 910                 return 1;
 911         if (td->o.loops) {
 912                 td->o.loops--;
 913                 return 1;
 914         }
 915
 916         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 917                         + td->io_skip_bytes;
 918         if (io_done < td->o.size)
 919                 return 1;
 920
 921         return 0;
 922 }
 923
 924 static void reset_io_counters(struct thread_data *td)
 925 {
 926         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 927         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 928         td->zone_bytes = 0;
 929         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 930         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 931
 932         td->last_was_sync = 0;
 933
 934         /*
 935          * reset file done count if we are to start over
 936          */
 937         if (td->o.time_based || td->o.loops)
 938                 td->nr_done_files = 0;
 939
 940         /*
 941          * Set the same seed to get repeatable runs
 942          */
 943         td_fill_rand_seeds(td);
 944 }
 945
 946 void reset_all_stats(struct thread_data *td)
 947 {
 948         struct timeval tv;
 949         int i;
 950
 951         reset_io_counters(td);
 952
 953         for (i = 0; i < 2; i++) {
 954                 td->io_bytes[i] = 0;
 955                 td->io_blocks[i] = 0;
 956                 td->io_issues[i] = 0;
 957                 td->ts.total_io_u[i] = 0;
 958         }
 959
 960         fio_gettime(&tv, NULL);
 961         memcpy(&td->epoch, &tv, sizeof(tv));
 962         memcpy(&td->start, &tv, sizeof(tv));
 963 }
 964
 965 static void clear_io_state(struct thread_data *td)
 966 {
 967         struct fio_file *f;
 968         unsigned int i;
 969
 970         reset_io_counters(td);
 971
 972         close_files(td);
 973         for_each_file(td, f, i)
 974                 fio_file_clear_done(f);
 975 }
 976
 977 static int exec_string(const char *string)
 978 {
 979         int ret, newlen = strlen(string) + 1 + 8;
 980         char *str;
 981
 982         str = malloc(newlen);
 983         sprintf(str, "sh -c %s", string);
 984
 985         ret = system(str);
 986         if (ret == -1)
 987                 log_err("fio: exec of cmd <%s> failed\n", str);
 988
 989         free(str);
 990         return ret;
 991 }
 992
 993 /*
 994  * Entry point for the thread based jobs. The process based jobs end up
 995  * here as well, after a little setup.
 996  */
 997 static void *thread_main(void *data)
 998 {
 999         unsigned long long runtime[2], elapsed;
1000         struct thread_data *td = data;
1001         pthread_condattr_t attr;
1002         int clear_state;
1003
1004         if (!td->o.use_thread)
1005                 setsid();
1006
1007         td->pid = getpid();
1008
1009         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1010
1011         INIT_FLIST_HEAD(&td->io_u_freelist);
1012         INIT_FLIST_HEAD(&td->io_u_busylist);
1013         INIT_FLIST_HEAD(&td->io_u_requeues);
1014         INIT_FLIST_HEAD(&td->io_log_list);
1015         INIT_FLIST_HEAD(&td->io_hist_list);
1016         INIT_FLIST_HEAD(&td->verify_list);
1017         pthread_mutex_init(&td->io_u_lock, NULL);
1018         td->io_hist_tree = RB_ROOT;
1019
1020         pthread_condattr_init(&attr);
1021         pthread_cond_init(&td->verify_cond, &attr);
1022         pthread_cond_init(&td->free_cond, &attr);
1023
1024         td_set_runstate(td, TD_INITIALIZED);
1025         dprint(FD_MUTEX, "up startup_mutex\n");
1026         fio_mutex_up(startup_mutex);
1027         dprint(FD_MUTEX, "wait on td->mutex\n");
1028         fio_mutex_down(td->mutex);
1029         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1030
1031         /*
1032          * the ->mutex mutex is now no longer used, close it to avoid
1033          * eating a file descriptor
1034          */
1035         fio_mutex_remove(td->mutex);
1036
1037         /*
1038          * May alter parameters that init_io_u() will use, so we need to
1039          * do this first.
1040          */
1041         if (init_iolog(td))
1042                 goto err;
1043
1044         if (init_io_u(td))
1045                 goto err;
1046
1047         if (td->o.verify_async && verify_async_init(td))
1048                 goto err;
1049
1050         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1051                 td_verror(td, errno, "cpu_set_affinity");
1052                 goto err;
1053         }
1054
1055         /*
1056          * If we have a gettimeofday() thread, make sure we exclude that
1057          * thread from this job
1058          */
1059         if (td->o.gtod_cpu) {
1060                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1061                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1062                         td_verror(td, errno, "cpu_set_affinity");
1063                         goto err;
1064                 }
1065         }
1066
1067         if (td->ioprio_set) {
1068                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1069                         td_verror(td, errno, "ioprio_set");
1070                         goto err;
1071                 }
1072         }
1073
1074         if (nice(td->o.nice) == -1) {
1075                 td_verror(td, errno, "nice");
1076                 goto err;
1077         }
1078
1079         if (td->o.ioscheduler && switch_ioscheduler(td))
1080                 goto err;
1081
1082         if (!td->o.create_serialize && setup_files(td))
1083                 goto err;
1084
1085         if (td_io_init(td))
1086                 goto err;
1087
1088         if (init_random_map(td))
1089                 goto err;
1090
1091         if (td->o.exec_prerun) {
1092                 if (exec_string(td->o.exec_prerun))
1093                         goto err;
1094         }
1095
1096         if (td->o.pre_read) {
1097                 if (pre_read_files(td) < 0)
1098                         goto err;
1099         }
1100
1101         fio_gettime(&td->epoch, NULL);
1102         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1103
1104         runtime[0] = runtime[1] = 0;
1105         clear_state = 0;
1106         while (keep_running(td)) {
1107                 fio_gettime(&td->start, NULL);
1108                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1109                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1110
1111                 if (td->o.ratemin[0] || td->o.ratemin[1])
1112                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1113                                                         sizeof(td->lastrate));
1114
1115                 if (clear_state)
1116                         clear_io_state(td);
1117
1118                 prune_io_piece_log(td);
1119
1120                 do_io(td);
1121
1122                 clear_state = 1;
1123
1124                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1125                         elapsed = utime_since_now(&td->start);
1126                         runtime[DDIR_READ] += elapsed;
1127                 }
1128                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1129                         elapsed = utime_since_now(&td->start);
1130                         runtime[DDIR_WRITE] += elapsed;
1131                 }
1132
1133                 if (td->error || td->terminate)
1134                         break;
1135
1136                 if (!td->o.do_verify ||
1137                     td->o.verify == VERIFY_NONE ||
1138                     (td->io_ops->flags & FIO_UNIDIR))
1139                         continue;
1140
1141                 clear_io_state(td);
1142
1143                 fio_gettime(&td->start, NULL);
1144
1145                 do_verify(td);
1146
1147                 runtime[DDIR_READ] += utime_since_now(&td->start);
1148
1149                 if (td->error || td->terminate)
1150                         break;
1151         }
1152
1153         update_rusage_stat(td);
1154         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1155         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1156         td->ts.total_run_time = mtime_since_now(&td->epoch);
1157         td->ts.io_bytes[0] = td->io_bytes[0];
1158         td->ts.io_bytes[1] = td->io_bytes[1];
1159
1160         fio_mutex_down(writeout_mutex);
1161         if (td->ts.bw_log) {
1162                 if (td->o.bw_log_file) {
1163                         finish_log_named(td, td->ts.bw_log,
1164                                                 td->o.bw_log_file, "bw");
1165                 } else
1166                         finish_log(td, td->ts.bw_log, "bw");
1167         }
1168         if (td->ts.slat_log) {
1169                 if (td->o.lat_log_file) {
1170                         finish_log_named(td, td->ts.slat_log,
1171                                                 td->o.lat_log_file, "slat");
1172                 } else
1173                         finish_log(td, td->ts.slat_log, "slat");
1174         }
1175         if (td->ts.clat_log) {
1176                 if (td->o.lat_log_file) {
1177                         finish_log_named(td, td->ts.clat_log,
1178                                                 td->o.lat_log_file, "clat");
1179                 } else
1180                         finish_log(td, td->ts.clat_log, "clat");
1181         }
1182         fio_mutex_up(writeout_mutex);
1183         if (td->o.exec_postrun)
1184                 exec_string(td->o.exec_postrun);
1185
1186         if (exitall_on_terminate)
1187                 terminate_threads(td->groupid);
1188
1189 err:
1190         if (td->error)
1191                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1192                                                         td->verror);
1193         close_and_free_files(td);
1194         close_ioengine(td);
1195         cleanup_io_u(td);
1196
1197         if (td->o.cpumask_set) {
1198                 int ret = fio_cpuset_exit(&td->o.cpumask);
1199
1200                 td_verror(td, ret, "fio_cpuset_exit");
1201         }
1202
1203         if (td->o.verify_async)
1204                 verify_async_exit(td);
1205
1206         /*
1207          * do this very late, it will log file closing as well
1208          */
1209         if (td->o.write_iolog_file)
1210                 write_iolog_close(td);
1211
1212         options_mem_free(td);
1213         td_set_runstate(td, TD_EXITED);
1214         return (void *) (unsigned long) td->error;
1215 }
1216
1217 /*
1218  * We cannot pass the td data into a forked process, so attach the td and
1219  * pass it to the thread worker.
1220  */
1221 static int fork_main(int shmid, int offset)
1222 {
1223         struct thread_data *td;
1224         void *data, *ret;
1225
1226         data = shmat(shmid, NULL, 0);
1227         if (data == (void *) -1) {
1228                 int __err = errno;
1229
1230                 perror("shmat");
1231                 return __err;
1232         }
1233
1234         td = data + offset * sizeof(struct thread_data);
1235         ret = thread_main(td);
1236         shmdt(data);
1237         return (int) (unsigned long) ret;
1238 }
1239
1240 /*
1241  * Run over the job map and reap the threads that have exited, if any.
1242  */
1243 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1244 {
1245         struct thread_data *td;
1246         int i, cputhreads, realthreads, pending, status, ret;
1247
1248         /*
1249          * reap exited threads (TD_EXITED -> TD_REAPED)
1250          */
1251         realthreads = pending = cputhreads = 0;
1252         for_each_td(td, i) {
1253                 int flags = 0;
1254
1255                 /*
1256                  * ->io_ops is NULL for a thread that has closed its
1257                  * io engine
1258                  */
1259                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1260                         cputhreads++;
1261                 else
1262                         realthreads++;
1263
1264                 if (!td->pid) {
1265                         pending++;
1266                         continue;
1267                 }
1268                 if (td->runstate == TD_REAPED)
1269                         continue;
1270                 if (td->o.use_thread) {
1271                         if (td->runstate == TD_EXITED) {
1272                                 td_set_runstate(td, TD_REAPED);
1273                                 goto reaped;
1274                         }
1275                         continue;
1276                 }
1277
1278                 flags = WNOHANG;
1279                 if (td->runstate == TD_EXITED)
1280                         flags = 0;
1281
1282                 /*
1283                  * check if someone quit or got killed in an unusual way
1284                  */
1285                 ret = waitpid(td->pid, &status, flags);
1286                 if (ret < 0) {
1287                         if (errno == ECHILD) {
1288                                 log_err("fio: pid=%d disappeared %d\n",
1289                                                 (int) td->pid, td->runstate);
1290                                 td_set_runstate(td, TD_REAPED);
1291                                 goto reaped;
1292                         }
1293                         perror("waitpid");
1294                 } else if (ret == td->pid) {
1295                         if (WIFSIGNALED(status)) {
1296                                 int sig = WTERMSIG(status);
1297
1298                                 if (sig != SIGQUIT)
1299                                         log_err("fio: pid=%d, got signal=%d\n",
1300                                                         (int) td->pid, sig);
1301                                 td_set_runstate(td, TD_REAPED);
1302                                 goto reaped;
1303                         }
1304                         if (WIFEXITED(status)) {
1305                                 if (WEXITSTATUS(status) && !td->error)
1306                                         td->error = WEXITSTATUS(status);
1307
1308                                 td_set_runstate(td, TD_REAPED);
1309                                 goto reaped;
1310                         }
1311                 }
1312
1313                 /*
1314                  * thread is not dead, continue
1315                  */
1316                 pending++;
1317                 continue;
1318 reaped:
1319                 (*nr_running)--;
1320                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1321                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1322                 if (!td->pid)
1323                         pending--;
1324
1325                 if (td->error)
1326                         exit_value++;
1327
1328                 done_secs += mtime_since_now(&td->epoch) / 1000;
1329         }
1330
1331         if (*nr_running == cputhreads && !pending && realthreads)
1332                 terminate_threads(TERMINATE_ALL);
1333 }
1334
1335 static void *gtod_thread_main(void *data)
1336 {
1337         fio_mutex_up(startup_mutex);
1338
1339         /*
1340          * As long as we have jobs around, update the clock. It would be nice
1341          * to have some way of NOT hammering that CPU with gettimeofday(),
1342          * but I'm not sure what to use outside of a simple CPU nop to relax
1343          * it - we don't want to lose precision.
1344          */
1345         while (threads) {
1346                 fio_gtod_update();
1347                 nop;
1348         }
1349
1350         return NULL;
1351 }
1352
1353 static int fio_start_gtod_thread(void)
1354 {
1355         int ret;
1356
1357         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1358         if (ret) {
1359                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1360                 return 1;
1361         }
1362
1363         ret = pthread_detach(gtod_thread);
1364         if (ret) {
1365                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1366                 return 1;
1367         }
1368
1369         dprint(FD_MUTEX, "wait on startup_mutex\n");
1370         fio_mutex_down(startup_mutex);
1371         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1372         return 0;
1373 }
1374
1375 /*
1376  * Main function for kicking off and reaping jobs, as needed.
1377  */
1378 static void run_threads(void)
1379 {
1380         struct thread_data *td;
1381         unsigned long spent;
1382         int i, todo, nr_running, m_rate, t_rate, nr_started;
1383
1384         if (fio_pin_memory())
1385                 return;
1386
1387         if (fio_gtod_offload && fio_start_gtod_thread())
1388                 return;
1389
1390         if (!terse_output) {
1391                 printf("Starting ");
1392                 if (nr_thread)
1393                         printf("%d thread%s", nr_thread,
1394                                                 nr_thread > 1 ? "s" : "");
1395                 if (nr_process) {
1396                         if (nr_thread)
1397                                 printf(" and ");
1398                         printf("%d process%s", nr_process,
1399                                                 nr_process > 1 ? "es" : "");
1400                 }
1401                 printf("\n");
1402                 fflush(stdout);
1403         }
1404
1405         set_sig_handlers();
1406
1407         todo = thread_number;
1408         nr_running = 0;
1409         nr_started = 0;
1410         m_rate = t_rate = 0;
1411
1412         for_each_td(td, i) {
1413                 print_status_init(td->thread_number - 1);
1414
1415                 if (!td->o.create_serialize) {
1416                         init_disk_util(td);
1417                         continue;
1418                 }
1419
1420                 /*
1421                  * do file setup here so it happens sequentially,
1422                  * we don't want X number of threads getting their
1423                  * client data interspersed on disk
1424                  */
1425                 if (setup_files(td)) {
1426                         exit_value++;
1427                         if (td->error)
1428                                 log_err("fio: pid=%d, err=%d/%s\n",
1429                                         (int) td->pid, td->error, td->verror);
1430                         td_set_runstate(td, TD_REAPED);
1431                         todo--;
1432                 } else {
1433                         struct fio_file *f;
1434                         unsigned int i;
1435
1436                         /*
1437                          * for sharing to work, each job must always open
1438                          * its own files. so close them, if we opened them
1439                          * for creation
1440                          */
1441                         for_each_file(td, f, i) {
1442                                 if (fio_file_open(f))
1443                                         td_io_close_file(td, f);
1444                         }
1445                 }
1446
1447                 init_disk_util(td);
1448         }
1449
1450         set_genesis_time();
1451
1452         while (todo) {
1453                 struct thread_data *map[MAX_JOBS];
1454                 struct timeval this_start;
1455                 int this_jobs = 0, left;
1456
1457                 /*
1458                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1459                  */
1460                 for_each_td(td, i) {
1461                         if (td->runstate != TD_NOT_CREATED)
1462                                 continue;
1463
1464                         /*
1465                          * never got a chance to start, killed by other
1466                          * thread for some reason
1467                          */
1468                         if (td->terminate) {
1469                                 todo--;
1470                                 continue;
1471                         }
1472
1473                         if (td->o.start_delay) {
1474                                 spent = mtime_since_genesis();
1475
1476                                 if (td->o.start_delay * 1000 > spent)
1477                                         continue;
1478                         }
1479
1480                         if (td->o.stonewall && (nr_started || nr_running)) {
1481                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1482                                                         td->o.name);
1483                                 break;
1484                         }
1485
1486                         /*
1487                          * Set state to created. Thread will transition
1488                          * to TD_INITIALIZED when it's done setting up.
1489                          */
1490                         td_set_runstate(td, TD_CREATED);
1491                         map[this_jobs++] = td;
1492                         nr_started++;
1493
1494                         if (td->o.use_thread) {
1495                                 int ret;
1496
1497                                 dprint(FD_PROCESS, "will pthread_create\n");
1498                                 ret = pthread_create(&td->thread, NULL,
1499                                                         thread_main, td);
1500                                 if (ret) {
1501                                         log_err("pthread_create: %s\n",
1502                                                         strerror(ret));
1503                                         nr_started--;
1504                                         break;
1505                                 }
1506                                 ret = pthread_detach(td->thread);
1507                                 if (ret)
1508                                         log_err("pthread_detach: %s",
1509                                                         strerror(ret));
1510                         } else {
1511                                 pid_t pid;
1512                                 dprint(FD_PROCESS, "will fork\n");
1513                                 pid = fork();
1514                                 if (!pid) {
1515                                         int ret = fork_main(shm_id, i);
1516
1517                                         _exit(ret);
1518                                 } else if (i == fio_debug_jobno)
1519                                         *fio_debug_jobp = pid;
1520                         }
1521                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1522                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1523                                 log_err("fio: job startup hung? exiting.\n");
1524                                 terminate_threads(TERMINATE_ALL);
1525                                 fio_abort = 1;
1526                                 nr_started--;
1527                                 break;
1528                         }
1529                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1530                 }
1531
1532                 /*
1533                  * Wait for the started threads to transition to
1534                  * TD_INITIALIZED.
1535                  */
1536                 fio_gettime(&this_start, NULL);
1537                 left = this_jobs;
1538                 while (left && !fio_abort) {
1539                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1540                                 break;
1541
1542                         usleep(100000);
1543
1544                         for (i = 0; i < this_jobs; i++) {
1545                                 td = map[i];
1546                                 if (!td)
1547                                         continue;
1548                                 if (td->runstate == TD_INITIALIZED) {
1549                                         map[i] = NULL;
1550                                         left--;
1551                                 } else if (td->runstate >= TD_EXITED) {
1552                                         map[i] = NULL;
1553                                         left--;
1554                                         todo--;
1555                                         nr_running++; /* work-around... */
1556                                 }
1557                         }
1558                 }
1559
1560                 if (left) {
1561                         log_err("fio: %d jobs failed to start\n", left);
1562                         for (i = 0; i < this_jobs; i++) {
1563                                 td = map[i];
1564                                 if (!td)
1565                                         continue;
1566                                 kill(td->pid, SIGTERM);
1567                         }
1568                         break;
1569                 }
1570
1571                 /*
1572                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1573                  */
1574                 for_each_td(td, i) {
1575                         if (td->runstate != TD_INITIALIZED)
1576                                 continue;
1577
1578                         if (in_ramp_time(td))
1579                                 td_set_runstate(td, TD_RAMP);
1580                         else
1581                                 td_set_runstate(td, TD_RUNNING);
1582                         nr_running++;
1583                         nr_started--;
1584                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1585                         t_rate += td->o.rate[0] + td->o.rate[1];
1586                         todo--;
1587                         fio_mutex_up(td->mutex);
1588                 }
1589
1590                 reap_threads(&nr_running, &t_rate, &m_rate);
1591
1592                 if (todo)
1593                         usleep(100000);
1594         }
1595
1596         while (nr_running) {
1597                 reap_threads(&nr_running, &t_rate, &m_rate);
1598                 usleep(10000);
1599         }
1600
1601         update_io_ticks();
1602         fio_unpin_memory();
1603 }
1604
1605 int main(int argc, char *argv[])
1606 {
1607         long ps;
1608
1609         sinit();
1610
1611         /*
1612          * We need locale for number printing, if it isn't set then just
1613          * go with the US format.
1614          */
1615         if (!getenv("LC_NUMERIC"))
1616                 setlocale(LC_NUMERIC, "en_US");
1617
1618         if (parse_options(argc, argv))
1619                 return 1;
1620
1621         if (!thread_number)
1622                 return 0;
1623
1624         ps = sysconf(_SC_PAGESIZE);
1625         if (ps < 0) {
1626                 log_err("Failed to get page size\n");
1627                 return 1;
1628         }
1629
1630         page_size = ps;
1631         page_mask = ps - 1;
1632
1633         if (write_bw_log) {
1634                 setup_log(&agg_io_log[DDIR_READ]);
1635                 setup_log(&agg_io_log[DDIR_WRITE]);
1636         }
1637
1638         startup_mutex = fio_mutex_init(0);
1639         writeout_mutex = fio_mutex_init(1);
1640
1641         set_genesis_time();
1642
1643         status_timer_arm();
1644
1645         run_threads();
1646
1647         if (!fio_abort) {
1648                 show_run_stats();
1649                 if (write_bw_log) {
1650                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1651                         __finish_log(agg_io_log[DDIR_WRITE],
1652                                         "agg-write_bw.log");
1653                 }
1654         }
1655
1656         fio_mutex_remove(startup_mutex);
1657         fio_mutex_remove(writeout_mutex);
1658         return exit_value;
1659 }