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