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 <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <sys/stat.h>
  34 #include <sys/wait.h>
  35 #include <sys/ipc.h>
  36 #include <sys/shm.h>
  37 #include <sys/mman.h>
  38
  39 #include "fio.h"
  40 #include "hash.h"
  41 #include "smalloc.h"
  42 #include "verify.h"
  43 #include "trim.h"
  44 #include "diskutil.h"
  45 #include "cgroup.h"
  46 #include "profile.h"
  47 #include "lib/rand.h"
  48 #include "memalign.h"
  49
  50 unsigned long page_mask;
  51 unsigned long page_size;
  52
  53 #define PAGE_ALIGN(buf) \
  54         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  55
  56 int groupid = 0;
  57 int thread_number = 0;
  58 int nr_process = 0;
  59 int nr_thread = 0;
  60 int shm_id = 0;
  61 int temp_stall_ts;
  62 unsigned long done_secs = 0;
  63
  64 static struct fio_mutex *startup_mutex;
  65 static struct fio_mutex *writeout_mutex;
  66 static volatile int fio_abort;
  67 static int exit_value;
  68 static pthread_t gtod_thread;
  69 static pthread_t disk_util_thread;
  70 static struct flist_head *cgroup_list;
  71 static char *cgroup_mnt;
  72
  73 struct io_log *agg_io_log[2];
  74
  75 #define TERMINATE_ALL           (-1)
  76 #define JOB_START_TIMEOUT       (5 * 1000)
  77
  78 void td_set_runstate(struct thread_data *td, int runstate)
  79 {
  80         if (td->runstate == runstate)
  81                 return;
  82
  83         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  84                                                 td->runstate, runstate);
  85         td->runstate = runstate;
  86 }
  87
  88 static void terminate_threads(int group_id)
  89 {
  90         struct thread_data *td;
  91         int i;
  92
  93         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  94
  95         for_each_td(td, i) {
  96                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  97                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  98                                                 td->o.name, (int) td->pid);
  99                         td->terminate = 1;
 100                         td->o.start_delay = 0;
 101
 102                         /*
 103                          * if the thread is running, just let it exit
 104                          */
 105                         if (td->runstate < TD_RUNNING)
 106                                 kill(td->pid, SIGTERM);
 107                         else {
 108                                 struct ioengine_ops *ops = td->io_ops;
 109
 110                                 if (ops && (ops->flags & FIO_SIGTERM))
 111                                         kill(td->pid, SIGTERM);
 112                         }
 113                 }
 114         }
 115 }
 116
 117 /*
 118  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 119  */
 120 static void sig_quit(int sig)
 121 {
 122 }
 123
 124 static void sig_int(int sig)
 125 {
 126         if (threads) {
 127                 log_info("\nfio: terminating on signal %d\n", sig);
 128                 fflush(stdout);
 129                 terminate_threads(TERMINATE_ALL);
 130         }
 131 }
 132
 133 static void *disk_thread_main(void *data)
 134 {
 135         fio_mutex_up(startup_mutex);
 136
 137         while (threads) {
 138                 usleep(DISK_UTIL_MSEC * 1000);
 139                 if (!threads)
 140                         break;
 141                 update_io_ticks();
 142                 print_thread_status();
 143         }
 144
 145         return NULL;
 146 }
 147
 148 static int create_disk_util_thread(void)
 149 {
 150         int ret;
 151
 152         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 153         if (ret) {
 154                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 155                 return 1;
 156         }
 157
 158         ret = pthread_detach(disk_util_thread);
 159         if (ret) {
 160                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 161                 return 1;
 162         }
 163
 164         dprint(FD_MUTEX, "wait on startup_mutex\n");
 165         fio_mutex_down(startup_mutex);
 166         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 167         return 0;
 168 }
 169
 170 static void set_sig_handlers(void)
 171 {
 172         struct sigaction act;
 173
 174         memset(&act, 0, sizeof(act));
 175         act.sa_handler = sig_int;
 176         act.sa_flags = SA_RESTART;
 177         sigaction(SIGINT, &act, NULL);
 178
 179         memset(&act, 0, sizeof(act));
 180         act.sa_handler = sig_quit;
 181         act.sa_flags = SA_RESTART;
 182         sigaction(SIGTERM, &act, NULL);
 183 }
 184
 185 /*
 186  * Check if we are above the minimum rate given.
 187  */
 188 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 189                             enum td_ddir ddir)
 190 {
 191         unsigned long long bytes = 0;
 192         unsigned long iops = 0;
 193         unsigned long spent;
 194         unsigned long rate;
 195         unsigned int ratemin = 0;
 196         unsigned int rate_iops = 0;
 197         unsigned int rate_iops_min = 0;
 198
 199         assert(ddir_rw(ddir));
 200
 201         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 202                 return 0;
 203
 204         /*
 205          * allow a 2 second settle period in the beginning
 206          */
 207         if (mtime_since(&td->start, now) < 2000)
 208                 return 0;
 209
 210         iops += td->io_blocks[ddir];
 211         bytes += td->this_io_bytes[ddir];
 212         ratemin += td->o.ratemin[ddir];
 213         rate_iops += td->o.rate_iops[ddir];
 214         rate_iops_min += td->o.rate_iops_min[ddir];
 215
 216         /*
 217          * if rate blocks is set, sample is running
 218          */
 219         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 220                 spent = mtime_since(&td->lastrate[ddir], now);
 221                 if (spent < td->o.ratecycle)
 222                         return 0;
 223
 224                 if (td->o.rate[ddir]) {
 225                         /*
 226                          * check bandwidth specified rate
 227                          */
 228                         if (bytes < td->rate_bytes[ddir]) {
 229                                 log_err("%s: min rate %u not met\n", td->o.name,
 230                                                                 ratemin);
 231                                 return 1;
 232                         } else {
 233                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 234                                 if (rate < ratemin ||
 235                                     bytes < td->rate_bytes[ddir]) {
 236                                         log_err("%s: min rate %u not met, got"
 237                                                 " %luKB/sec\n", td->o.name,
 238                                                         ratemin, rate);
 239                                         return 1;
 240                                 }
 241                         }
 242                 } else {
 243                         /*
 244                          * checks iops specified rate
 245                          */
 246                         if (iops < rate_iops) {
 247                                 log_err("%s: min iops rate %u not met\n",
 248                                                 td->o.name, rate_iops);
 249                                 return 1;
 250                         } else {
 251                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 252                                 if (rate < rate_iops_min ||
 253                                     iops < td->rate_blocks[ddir]) {
 254                                         log_err("%s: min iops rate %u not met,"
 255                                                 " got %lu\n", td->o.name,
 256                                                         rate_iops_min, rate);
 257                                 }
 258                         }
 259                 }
 260         }
 261
 262         td->rate_bytes[ddir] = bytes;
 263         td->rate_blocks[ddir] = iops;
 264         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 265         return 0;
 266 }
 267
 268 static int check_min_rate(struct thread_data *td, struct timeval *now,
 269                           unsigned long *bytes_done)
 270 {
 271         int ret = 0;
 272
 273         if (bytes_done[0])
 274                 ret |= __check_min_rate(td, now, 0);
 275         if (bytes_done[1])
 276                 ret |= __check_min_rate(td, now, 1);
 277
 278         return ret;
 279 }
 280
 281 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 282 {
 283         if (!td->o.timeout)
 284                 return 0;
 285         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 286                 return 1;
 287
 288         return 0;
 289 }
 290
 291 /*
 292  * When job exits, we can cancel the in-flight IO if we are using async
 293  * io. Attempt to do so.
 294  */
 295 static void cleanup_pending_aio(struct thread_data *td)
 296 {
 297         struct flist_head *entry, *n;
 298         struct io_u *io_u;
 299         int r;
 300
 301         /*
 302          * get immediately available events, if any
 303          */
 304         r = io_u_queued_complete(td, 0, NULL);
 305         if (r < 0)
 306                 return;
 307
 308         /*
 309          * now cancel remaining active events
 310          */
 311         if (td->io_ops->cancel) {
 312                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 313                         io_u = flist_entry(entry, struct io_u, list);
 314
 315                         /*
 316                          * if the io_u isn't in flight, then that generally
 317                          * means someone leaked an io_u. complain but fix
 318                          * it up, so we don't stall here.
 319                          */
 320                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 321                                 log_err("fio: non-busy IO on busy list\n");
 322                                 put_io_u(td, io_u);
 323                         } else {
 324                                 r = td->io_ops->cancel(td, io_u);
 325                                 if (!r)
 326                                         put_io_u(td, io_u);
 327                         }
 328                 }
 329         }
 330
 331         if (td->cur_depth)
 332                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 333 }
 334
 335 /*
 336  * Helper to handle the final sync of a file. Works just like the normal
 337  * io path, just does everything sync.
 338  */
 339 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 340 {
 341         struct io_u *io_u = __get_io_u(td);
 342         int ret;
 343
 344         if (!io_u)
 345                 return 1;
 346
 347         io_u->ddir = DDIR_SYNC;
 348         io_u->file = f;
 349
 350         if (td_io_prep(td, io_u)) {
 351                 put_io_u(td, io_u);
 352                 return 1;
 353         }
 354
 355 requeue:
 356         ret = td_io_queue(td, io_u);
 357         if (ret < 0) {
 358                 td_verror(td, io_u->error, "td_io_queue");
 359                 put_io_u(td, io_u);
 360                 return 1;
 361         } else if (ret == FIO_Q_QUEUED) {
 362                 if (io_u_queued_complete(td, 1, NULL) < 0)
 363                         return 1;
 364         } else if (ret == FIO_Q_COMPLETED) {
 365                 if (io_u->error) {
 366                         td_verror(td, io_u->error, "td_io_queue");
 367                         return 1;
 368                 }
 369
 370                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 371                         return 1;
 372         } else if (ret == FIO_Q_BUSY) {
 373                 if (td_io_commit(td))
 374                         return 1;
 375                 goto requeue;
 376         }
 377
 378         return 0;
 379 }
 380
 381 static inline void update_tv_cache(struct thread_data *td)
 382 {
 383         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 384                 fio_gettime(&td->tv_cache, NULL);
 385 }
 386
 387 static int break_on_this_error(struct thread_data *td, int *retptr)
 388 {
 389         int ret = *retptr;
 390
 391         if (ret < 0 || td->error) {
 392                 int err;
 393
 394                 if (!td->o.continue_on_error)
 395                         return 1;
 396
 397                 if (ret < 0)
 398                         err = -ret;
 399                 else
 400                         err = td->error;
 401
 402                 if (td_non_fatal_error(err)) {
 403                         /*
 404                          * Continue with the I/Os in case of
 405                          * a non fatal error.
 406                          */
 407                         update_error_count(td, err);
 408                         td_clear_error(td);
 409                         *retptr = 0;
 410                         return 0;
 411                 } else if (td->o.fill_device && err == ENOSPC) {
 412                         /*
 413                          * We expect to hit this error if
 414                          * fill_device option is set.
 415                          */
 416                         td_clear_error(td);
 417                         td->terminate = 1;
 418                         return 1;
 419                 } else {
 420                         /*
 421                          * Stop the I/O in case of a fatal
 422                          * error.
 423                          */
 424                         update_error_count(td, err);
 425                         return 1;
 426                 }
 427         }
 428
 429         return 0;
 430 }
 431
 432 /*
 433  * The main verify engine. Runs over the writes we previously submitted,
 434  * reads the blocks back in, and checks the crc/md5 of the data.
 435  */
 436 static void do_verify(struct thread_data *td)
 437 {
 438         struct fio_file *f;
 439         struct io_u *io_u;
 440         int ret, min_events;
 441         unsigned int i;
 442
 443         dprint(FD_VERIFY, "starting loop\n");
 444
 445         /*
 446          * sync io first and invalidate cache, to make sure we really
 447          * read from disk.
 448          */
 449         for_each_file(td, f, i) {
 450                 if (!fio_file_open(f))
 451                         continue;
 452                 if (fio_io_sync(td, f))
 453                         break;
 454                 if (file_invalidate_cache(td, f))
 455                         break;
 456         }
 457
 458         if (td->error)
 459                 return;
 460
 461         td_set_runstate(td, TD_VERIFYING);
 462
 463         io_u = NULL;
 464         while (!td->terminate) {
 465                 int ret2, full;
 466
 467                 update_tv_cache(td);
 468
 469                 if (runtime_exceeded(td, &td->tv_cache)) {
 470                         td->terminate = 1;
 471                         break;
 472                 }
 473
 474                 io_u = __get_io_u(td);
 475                 if (!io_u)
 476                         break;
 477
 478                 if (get_next_verify(td, io_u)) {
 479                         put_io_u(td, io_u);
 480                         break;
 481                 }
 482
 483                 if (td_io_prep(td, io_u)) {
 484                         put_io_u(td, io_u);
 485                         break;
 486                 }
 487
 488                 if (td->o.verify_async)
 489                         io_u->end_io = verify_io_u_async;
 490                 else
 491                         io_u->end_io = verify_io_u;
 492
 493                 ret = td_io_queue(td, io_u);
 494                 switch (ret) {
 495                 case FIO_Q_COMPLETED:
 496                         if (io_u->error) {
 497                                 ret = -io_u->error;
 498                                 clear_io_u(td, io_u);
 499                         } else if (io_u->resid) {
 500                                 int bytes = io_u->xfer_buflen - io_u->resid;
 501
 502                                 /*
 503                                  * zero read, fail
 504                                  */
 505                                 if (!bytes) {
 506                                         td_verror(td, EIO, "full resid");
 507                                         put_io_u(td, io_u);
 508                                         break;
 509                                 }
 510
 511                                 io_u->xfer_buflen = io_u->resid;
 512                                 io_u->xfer_buf += bytes;
 513                                 io_u->offset += bytes;
 514
 515                                 if (ddir_rw(io_u->ddir))
 516                                         td->ts.short_io_u[io_u->ddir]++;
 517
 518                                 f = io_u->file;
 519                                 if (io_u->offset == f->real_file_size)
 520                                         goto sync_done;
 521
 522                                 requeue_io_u(td, &io_u);
 523                         } else {
 524 sync_done:
 525                                 ret = io_u_sync_complete(td, io_u, NULL);
 526                                 if (ret < 0)
 527                                         break;
 528                         }
 529                         continue;
 530                 case FIO_Q_QUEUED:
 531                         break;
 532                 case FIO_Q_BUSY:
 533                         requeue_io_u(td, &io_u);
 534                         ret2 = td_io_commit(td);
 535                         if (ret2 < 0)
 536                                 ret = ret2;
 537                         break;
 538                 default:
 539                         assert(ret < 0);
 540                         td_verror(td, -ret, "td_io_queue");
 541                         break;
 542                 }
 543
 544                 if (break_on_this_error(td, &ret))
 545                         break;
 546
 547                 /*
 548                  * if we can queue more, do so. but check if there are
 549                  * completed io_u's first.
 550                  */
 551                 full = queue_full(td) || ret == FIO_Q_BUSY;
 552                 if (full || !td->o.iodepth_batch_complete) {
 553                         min_events = min(td->o.iodepth_batch_complete,
 554                                          td->cur_depth);
 555                         if (full && !min_events)
 556                                 min_events = 1;
 557
 558                         do {
 559                                 /*
 560                                  * Reap required number of io units, if any,
 561                                  * and do the verification on them through
 562                                  * the callback handler
 563                                  */
 564                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 565                                         ret = -1;
 566                                         break;
 567                                 }
 568                         } while (full && (td->cur_depth > td->o.iodepth_low));
 569                 }
 570                 if (ret < 0)
 571                         break;
 572         }
 573
 574         if (!td->error) {
 575                 min_events = td->cur_depth;
 576
 577                 if (min_events)
 578                         ret = io_u_queued_complete(td, min_events, NULL);
 579         } else
 580                 cleanup_pending_aio(td);
 581
 582         td_set_runstate(td, TD_RUNNING);
 583
 584         dprint(FD_VERIFY, "exiting loop\n");
 585 }
 586
 587 /*
 588  * Main IO worker function. It retrieves io_u's to process and queues
 589  * and reaps them, checking for rate and errors along the way.
 590  */
 591 static void do_io(struct thread_data *td)
 592 {
 593         unsigned int i;
 594         int ret = 0;
 595
 596         if (in_ramp_time(td))
 597                 td_set_runstate(td, TD_RAMP);
 598         else
 599                 td_set_runstate(td, TD_RUNNING);
 600
 601         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 602                 (!flist_empty(&td->trim_list)) ||
 603                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 604                 struct timeval comp_time;
 605                 unsigned long bytes_done[2] = { 0, 0 };
 606                 int min_evts = 0;
 607                 struct io_u *io_u;
 608                 int ret2, full;
 609
 610                 if (td->terminate)
 611                         break;
 612
 613                 update_tv_cache(td);
 614
 615                 if (runtime_exceeded(td, &td->tv_cache)) {
 616                         td->terminate = 1;
 617                         break;
 618                 }
 619
 620                 io_u = get_io_u(td);
 621                 if (!io_u)
 622                         break;
 623
 624                 /*
 625                  * Add verification end_io handler, if asked to verify
 626                  * a previously written file.
 627                  */
 628                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 629                     !td_rw(td)) {
 630                         if (td->o.verify_async)
 631                                 io_u->end_io = verify_io_u_async;
 632                         else
 633                                 io_u->end_io = verify_io_u;
 634                         td_set_runstate(td, TD_VERIFYING);
 635                 } else if (in_ramp_time(td))
 636                         td_set_runstate(td, TD_RAMP);
 637                 else
 638                         td_set_runstate(td, TD_RUNNING);
 639
 640                 ret = td_io_queue(td, io_u);
 641                 switch (ret) {
 642                 case FIO_Q_COMPLETED:
 643                         if (io_u->error) {
 644                                 ret = -io_u->error;
 645                                 clear_io_u(td, io_u);
 646                         } else if (io_u->resid) {
 647                                 int bytes = io_u->xfer_buflen - io_u->resid;
 648                                 struct fio_file *f = io_u->file;
 649
 650                                 /*
 651                                  * zero read, fail
 652                                  */
 653                                 if (!bytes) {
 654                                         td_verror(td, EIO, "full resid");
 655                                         put_io_u(td, io_u);
 656                                         break;
 657                                 }
 658
 659                                 io_u->xfer_buflen = io_u->resid;
 660                                 io_u->xfer_buf += bytes;
 661                                 io_u->offset += bytes;
 662
 663                                 if (ddir_rw(io_u->ddir))
 664                                         td->ts.short_io_u[io_u->ddir]++;
 665
 666                                 if (io_u->offset == f->real_file_size)
 667                                         goto sync_done;
 668
 669                                 requeue_io_u(td, &io_u);
 670                         } else {
 671 sync_done:
 672                                 if (__should_check_rate(td, 0) ||
 673                                     __should_check_rate(td, 1))
 674                                         fio_gettime(&comp_time, NULL);
 675
 676                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 677                                 if (ret < 0)
 678                                         break;
 679                         }
 680                         break;
 681                 case FIO_Q_QUEUED:
 682                         /*
 683                          * if the engine doesn't have a commit hook,
 684                          * the io_u is really queued. if it does have such
 685                          * a hook, it has to call io_u_queued() itself.
 686                          */
 687                         if (td->io_ops->commit == NULL)
 688                                 io_u_queued(td, io_u);
 689                         break;
 690                 case FIO_Q_BUSY:
 691                         requeue_io_u(td, &io_u);
 692                         ret2 = td_io_commit(td);
 693                         if (ret2 < 0)
 694                                 ret = ret2;
 695                         break;
 696                 default:
 697                         assert(ret < 0);
 698                         put_io_u(td, io_u);
 699                         break;
 700                 }
 701
 702                 if (break_on_this_error(td, &ret))
 703                         break;
 704
 705                 /*
 706                  * See if we need to complete some commands
 707                  */
 708                 full = queue_full(td) || ret == FIO_Q_BUSY;
 709                 if (full || !td->o.iodepth_batch_complete) {
 710                         min_evts = min(td->o.iodepth_batch_complete,
 711                                         td->cur_depth);
 712                         if (full && !min_evts)
 713                                 min_evts = 1;
 714
 715                         if (__should_check_rate(td, 0) ||
 716                             __should_check_rate(td, 1))
 717                                 fio_gettime(&comp_time, NULL);
 718
 719                         do {
 720                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 721                                 if (ret < 0)
 722                                         break;
 723
 724                         } while (full && (td->cur_depth > td->o.iodepth_low));
 725                 }
 726
 727                 if (ret < 0)
 728                         break;
 729                 if (!(bytes_done[0] + bytes_done[1]))
 730                         continue;
 731
 732                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 733                         if (check_min_rate(td, &comp_time, bytes_done)) {
 734                                 if (exitall_on_terminate)
 735                                         terminate_threads(td->groupid);
 736                                 td_verror(td, EIO, "check_min_rate");
 737                                 break;
 738                         }
 739                 }
 740
 741                 if (td->o.thinktime) {
 742                         unsigned long long b;
 743
 744                         b = td->io_blocks[0] + td->io_blocks[1];
 745                         if (!(b % td->o.thinktime_blocks)) {
 746                                 int left;
 747
 748                                 if (td->o.thinktime_spin)
 749                                         usec_spin(td->o.thinktime_spin);
 750
 751                                 left = td->o.thinktime - td->o.thinktime_spin;
 752                                 if (left)
 753                                         usec_sleep(td, left);
 754                         }
 755                 }
 756         }
 757
 758         if (td->trim_entries)
 759                 printf("trim entries %ld\n", td->trim_entries);
 760
 761         if (td->o.fill_device && td->error == ENOSPC) {
 762                 td->error = 0;
 763                 td->terminate = 1;
 764         }
 765         if (!td->error) {
 766                 struct fio_file *f;
 767
 768                 i = td->cur_depth;
 769                 if (i)
 770                         ret = io_u_queued_complete(td, i, NULL);
 771
 772                 if (should_fsync(td) && td->o.end_fsync) {
 773                         td_set_runstate(td, TD_FSYNCING);
 774
 775                         for_each_file(td, f, i) {
 776                                 if (!fio_file_open(f))
 777                                         continue;
 778                                 fio_io_sync(td, f);
 779                         }
 780                 }
 781         } else
 782                 cleanup_pending_aio(td);
 783
 784         /*
 785          * stop job if we failed doing any IO
 786          */
 787         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 788                 td->done = 1;
 789 }
 790
 791 static void cleanup_io_u(struct thread_data *td)
 792 {
 793         struct flist_head *entry, *n;
 794         struct io_u *io_u;
 795
 796         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 797                 io_u = flist_entry(entry, struct io_u, list);
 798
 799                 flist_del(&io_u->list);
 800                 fio_memfree(io_u, sizeof(*io_u));
 801         }
 802
 803         free_io_mem(td);
 804 }
 805
 806 static int init_io_u(struct thread_data *td)
 807 {
 808         struct io_u *io_u;
 809         unsigned int max_bs;
 810         int cl_align, i, max_units;
 811         char *p;
 812
 813         max_units = td->o.iodepth;
 814         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 815         td->orig_buffer_size = (unsigned long long) max_bs
 816                                         * (unsigned long long) max_units;
 817
 818         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 819                 unsigned long bs;
 820
 821                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 822                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 823         }
 824
 825         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 826                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 827                 return 1;
 828         }
 829
 830         if (allocate_io_mem(td))
 831                 return 1;
 832
 833         if (td->o.odirect || td->o.mem_align ||
 834             (td->io_ops->flags & FIO_RAWIO))
 835                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 836         else
 837                 p = td->orig_buffer;
 838
 839         cl_align = os_cache_line_size();
 840
 841         for (i = 0; i < max_units; i++) {
 842                 void *ptr;
 843
 844                 if (td->terminate)
 845                         return 1;
 846
 847                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 848                 if (!ptr) {
 849                         log_err("fio: unable to allocate aligned memory\n");
 850                         break;
 851                 }
 852
 853                 io_u = ptr;
 854                 memset(io_u, 0, sizeof(*io_u));
 855                 INIT_FLIST_HEAD(&io_u->list);
 856                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 857
 858                 if (!(td->io_ops->flags & FIO_NOIO)) {
 859                         io_u->buf = p + max_bs * i;
 860                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 861
 862                         if (td_write(td) && !td->o.refill_buffers)
 863                                 io_u_fill_buffer(td, io_u, max_bs);
 864                         else if (td_write(td) && td->o.verify_pattern_bytes) {
 865                                 /*
 866                                  * Fill the buffer with the pattern if we are
 867                                  * going to be doing writes.
 868                                  */
 869                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 870                         }
 871                 }
 872
 873                 io_u->index = i;
 874                 io_u->flags = IO_U_F_FREE;
 875                 flist_add(&io_u->list, &td->io_u_freelist);
 876         }
 877
 878         return 0;
 879 }
 880
 881 static int switch_ioscheduler(struct thread_data *td)
 882 {
 883         char tmp[256], tmp2[128];
 884         FILE *f;
 885         int ret;
 886
 887         if (td->io_ops->flags & FIO_DISKLESSIO)
 888                 return 0;
 889
 890         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 891
 892         f = fopen(tmp, "r+");
 893         if (!f) {
 894                 if (errno == ENOENT) {
 895                         log_err("fio: os or kernel doesn't support IO scheduler"
 896                                 " switching\n");
 897                         return 0;
 898                 }
 899                 td_verror(td, errno, "fopen iosched");
 900                 return 1;
 901         }
 902
 903         /*
 904          * Set io scheduler.
 905          */
 906         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 907         if (ferror(f) || ret != 1) {
 908                 td_verror(td, errno, "fwrite");
 909                 fclose(f);
 910                 return 1;
 911         }
 912
 913         rewind(f);
 914
 915         /*
 916          * Read back and check that the selected scheduler is now the default.
 917          */
 918         ret = fread(tmp, 1, sizeof(tmp), f);
 919         if (ferror(f) || ret < 0) {
 920                 td_verror(td, errno, "fread");
 921                 fclose(f);
 922                 return 1;
 923         }
 924
 925         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 926         if (!strstr(tmp, tmp2)) {
 927                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 928                 td_verror(td, EINVAL, "iosched_switch");
 929                 fclose(f);
 930                 return 1;
 931         }
 932
 933         fclose(f);
 934         return 0;
 935 }
 936
 937 static int keep_running(struct thread_data *td)
 938 {
 939         unsigned long long io_done;
 940
 941         if (td->done)
 942                 return 0;
 943         if (td->o.time_based)
 944                 return 1;
 945         if (td->o.loops) {
 946                 td->o.loops--;
 947                 return 1;
 948         }
 949
 950         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 951                         + td->io_skip_bytes;
 952         if (io_done < td->o.size)
 953                 return 1;
 954
 955         return 0;
 956 }
 957
 958 static void reset_io_counters(struct thread_data *td)
 959 {
 960         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 961         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 962         td->zone_bytes = 0;
 963         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 964         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 965
 966         td->last_was_sync = 0;
 967
 968         /*
 969          * reset file done count if we are to start over
 970          */
 971         if (td->o.time_based || td->o.loops)
 972                 td->nr_done_files = 0;
 973
 974         /*
 975          * Set the same seed to get repeatable runs
 976          */
 977         td_fill_rand_seeds(td);
 978 }
 979
 980 void reset_all_stats(struct thread_data *td)
 981 {
 982         struct timeval tv;
 983         int i;
 984
 985         reset_io_counters(td);
 986
 987         for (i = 0; i < 2; i++) {
 988                 td->io_bytes[i] = 0;
 989                 td->io_blocks[i] = 0;
 990                 td->io_issues[i] = 0;
 991                 td->ts.total_io_u[i] = 0;
 992         }
 993
 994         fio_gettime(&tv, NULL);
 995         td->ts.runtime[0] = 0;
 996         td->ts.runtime[1] = 0;
 997         memcpy(&td->epoch, &tv, sizeof(tv));
 998         memcpy(&td->start, &tv, sizeof(tv));
 999 }
1000
1001 static void clear_io_state(struct thread_data *td)
1002 {
1003         struct fio_file *f;
1004         unsigned int i;
1005
1006         reset_io_counters(td);
1007
1008         close_files(td);
1009         for_each_file(td, f, i)
1010                 fio_file_clear_done(f);
1011 }
1012
1013 static int exec_string(const char *string)
1014 {
1015         int ret, newlen = strlen(string) + 1 + 8;
1016         char *str;
1017
1018         str = malloc(newlen);
1019         sprintf(str, "sh -c %s", string);
1020
1021         ret = system(str);
1022         if (ret == -1)
1023                 log_err("fio: exec of cmd <%s> failed\n", str);
1024
1025         free(str);
1026         return ret;
1027 }
1028
1029 /*
1030  * Entry point for the thread based jobs. The process based jobs end up
1031  * here as well, after a little setup.
1032  */
1033 static void *thread_main(void *data)
1034 {
1035         unsigned long long elapsed;
1036         struct thread_data *td = data;
1037         pthread_condattr_t attr;
1038         int clear_state;
1039
1040         if (!td->o.use_thread)
1041                 setsid();
1042
1043         td->pid = getpid();
1044
1045         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1046
1047         INIT_FLIST_HEAD(&td->io_u_freelist);
1048         INIT_FLIST_HEAD(&td->io_u_busylist);
1049         INIT_FLIST_HEAD(&td->io_u_requeues);
1050         INIT_FLIST_HEAD(&td->io_log_list);
1051         INIT_FLIST_HEAD(&td->io_hist_list);
1052         INIT_FLIST_HEAD(&td->verify_list);
1053         INIT_FLIST_HEAD(&td->trim_list);
1054         pthread_mutex_init(&td->io_u_lock, NULL);
1055         td->io_hist_tree = RB_ROOT;
1056
1057         pthread_condattr_init(&attr);
1058         pthread_cond_init(&td->verify_cond, &attr);
1059         pthread_cond_init(&td->free_cond, &attr);
1060
1061         td_set_runstate(td, TD_INITIALIZED);
1062         dprint(FD_MUTEX, "up startup_mutex\n");
1063         fio_mutex_up(startup_mutex);
1064         dprint(FD_MUTEX, "wait on td->mutex\n");
1065         fio_mutex_down(td->mutex);
1066         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1067
1068         /*
1069          * the ->mutex mutex is now no longer used, close it to avoid
1070          * eating a file descriptor
1071          */
1072         fio_mutex_remove(td->mutex);
1073
1074         /*
1075          * A new gid requires privilege, so we need to do this before setting
1076          * the uid.
1077          */
1078         if (td->o.gid != -1U && setgid(td->o.gid)) {
1079                 td_verror(td, errno, "setgid");
1080                 goto err;
1081         }
1082         if (td->o.uid != -1U && setuid(td->o.uid)) {
1083                 td_verror(td, errno, "setuid");
1084                 goto err;
1085         }
1086
1087         /*
1088          * May alter parameters that init_io_u() will use, so we need to
1089          * do this first.
1090          */
1091         if (init_iolog(td))
1092                 goto err;
1093
1094         if (init_io_u(td))
1095                 goto err;
1096
1097         if (td->o.verify_async && verify_async_init(td))
1098                 goto err;
1099
1100         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1101                 td_verror(td, errno, "cpu_set_affinity");
1102                 goto err;
1103         }
1104
1105         /*
1106          * If we have a gettimeofday() thread, make sure we exclude that
1107          * thread from this job
1108          */
1109         if (td->o.gtod_cpu) {
1110                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1111                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1112                         td_verror(td, errno, "cpu_set_affinity");
1113                         goto err;
1114                 }
1115         }
1116
1117         if (td->ioprio_set) {
1118                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1119                         td_verror(td, errno, "ioprio_set");
1120                         goto err;
1121                 }
1122         }
1123
1124         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1125                 goto err;
1126
1127         if (nice(td->o.nice) == -1) {
1128                 td_verror(td, errno, "nice");
1129                 goto err;
1130         }
1131
1132         if (td->o.ioscheduler && switch_ioscheduler(td))
1133                 goto err;
1134
1135         if (!td->o.create_serialize && setup_files(td))
1136                 goto err;
1137
1138         if (td_io_init(td))
1139                 goto err;
1140
1141         if (init_random_map(td))
1142                 goto err;
1143
1144         if (td->o.exec_prerun) {
1145                 if (exec_string(td->o.exec_prerun))
1146                         goto err;
1147         }
1148
1149         if (td->o.pre_read) {
1150                 if (pre_read_files(td) < 0)
1151                         goto err;
1152         }
1153
1154         fio_gettime(&td->epoch, NULL);
1155         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1156
1157         clear_state = 0;
1158         while (keep_running(td)) {
1159                 fio_gettime(&td->start, NULL);
1160                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1161                                 sizeof(td->start));
1162                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1163                                 sizeof(td->start));
1164                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1165
1166                 if (td->o.ratemin[0] || td->o.ratemin[1])
1167                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1168                                                         sizeof(td->lastrate));
1169
1170                 if (clear_state)
1171                         clear_io_state(td);
1172
1173                 prune_io_piece_log(td);
1174
1175                 do_io(td);
1176
1177                 clear_state = 1;
1178
1179                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1180                         elapsed = utime_since_now(&td->start);
1181                         td->ts.runtime[DDIR_READ] += elapsed;
1182                 }
1183                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1184                         elapsed = utime_since_now(&td->start);
1185                         td->ts.runtime[DDIR_WRITE] += elapsed;
1186                 }
1187
1188                 if (td->error || td->terminate)
1189                         break;
1190
1191                 if (!td->o.do_verify ||
1192                     td->o.verify == VERIFY_NONE ||
1193                     (td->io_ops->flags & FIO_UNIDIR))
1194                         continue;
1195
1196                 clear_io_state(td);
1197
1198                 fio_gettime(&td->start, NULL);
1199
1200                 do_verify(td);
1201
1202                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1203
1204                 if (td->error || td->terminate)
1205                         break;
1206         }
1207
1208         update_rusage_stat(td);
1209         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1210         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1211         td->ts.total_run_time = mtime_since_now(&td->epoch);
1212         td->ts.io_bytes[0] = td->io_bytes[0];
1213         td->ts.io_bytes[1] = td->io_bytes[1];
1214
1215         fio_mutex_down(writeout_mutex);
1216         if (td->ts.bw_log) {
1217                 if (td->o.bw_log_file) {
1218                         finish_log_named(td, td->ts.bw_log,
1219                                                 td->o.bw_log_file, "bw");
1220                 } else
1221                         finish_log(td, td->ts.bw_log, "bw");
1222         }
1223         if (td->ts.lat_log) {
1224                 if (td->o.lat_log_file) {
1225                         finish_log_named(td, td->ts.lat_log,
1226                                                 td->o.lat_log_file, "lat");
1227                 } else
1228                         finish_log(td, td->ts.lat_log, "lat");
1229         }
1230         if (td->ts.slat_log) {
1231                 if (td->o.lat_log_file) {
1232                         finish_log_named(td, td->ts.slat_log,
1233                                                 td->o.lat_log_file, "slat");
1234                 } else
1235                         finish_log(td, td->ts.slat_log, "slat");
1236         }
1237         if (td->ts.clat_log) {
1238                 if (td->o.lat_log_file) {
1239                         finish_log_named(td, td->ts.clat_log,
1240                                                 td->o.lat_log_file, "clat");
1241                 } else
1242                         finish_log(td, td->ts.clat_log, "clat");
1243         }
1244         fio_mutex_up(writeout_mutex);
1245         if (td->o.exec_postrun)
1246                 exec_string(td->o.exec_postrun);
1247
1248         if (exitall_on_terminate)
1249                 terminate_threads(td->groupid);
1250
1251 err:
1252         if (td->error)
1253                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1254                                                         td->verror);
1255
1256         if (td->o.verify_async)
1257                 verify_async_exit(td);
1258
1259         close_and_free_files(td);
1260         close_ioengine(td);
1261         cleanup_io_u(td);
1262         cgroup_shutdown(td, &cgroup_mnt);
1263
1264         if (td->o.cpumask_set) {
1265                 int ret = fio_cpuset_exit(&td->o.cpumask);
1266
1267                 td_verror(td, ret, "fio_cpuset_exit");
1268         }
1269
1270         /*
1271          * do this very late, it will log file closing as well
1272          */
1273         if (td->o.write_iolog_file)
1274                 write_iolog_close(td);
1275
1276         options_mem_free(td);
1277         td_set_runstate(td, TD_EXITED);
1278         return (void *) (unsigned long) td->error;
1279 }
1280
1281 /*
1282  * We cannot pass the td data into a forked process, so attach the td and
1283  * pass it to the thread worker.
1284  */
1285 static int fork_main(int shmid, int offset)
1286 {
1287         struct thread_data *td;
1288         void *data, *ret;
1289
1290         data = shmat(shmid, NULL, 0);
1291         if (data == (void *) -1) {
1292                 int __err = errno;
1293
1294                 perror("shmat");
1295                 return __err;
1296         }
1297
1298         td = data + offset * sizeof(struct thread_data);
1299         ret = thread_main(td);
1300         shmdt(data);
1301         return (int) (unsigned long) ret;
1302 }
1303
1304 /*
1305  * Run over the job map and reap the threads that have exited, if any.
1306  */
1307 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1308 {
1309         struct thread_data *td;
1310         int i, cputhreads, realthreads, pending, status, ret;
1311
1312         /*
1313          * reap exited threads (TD_EXITED -> TD_REAPED)
1314          */
1315         realthreads = pending = cputhreads = 0;
1316         for_each_td(td, i) {
1317                 int flags = 0;
1318
1319                 /*
1320                  * ->io_ops is NULL for a thread that has closed its
1321                  * io engine
1322                  */
1323                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1324                         cputhreads++;
1325                 else
1326                         realthreads++;
1327
1328                 if (!td->pid) {
1329                         pending++;
1330                         continue;
1331                 }
1332                 if (td->runstate == TD_REAPED)
1333                         continue;
1334                 if (td->o.use_thread) {
1335                         if (td->runstate == TD_EXITED) {
1336                                 td_set_runstate(td, TD_REAPED);
1337                                 goto reaped;
1338                         }
1339                         continue;
1340                 }
1341
1342                 flags = WNOHANG;
1343                 if (td->runstate == TD_EXITED)
1344                         flags = 0;
1345
1346                 /*
1347                  * check if someone quit or got killed in an unusual way
1348                  */
1349                 ret = waitpid(td->pid, &status, flags);
1350                 if (ret < 0) {
1351                         if (errno == ECHILD) {
1352                                 log_err("fio: pid=%d disappeared %d\n",
1353                                                 (int) td->pid, td->runstate);
1354                                 td_set_runstate(td, TD_REAPED);
1355                                 goto reaped;
1356                         }
1357                         perror("waitpid");
1358                 } else if (ret == td->pid) {
1359                         if (WIFSIGNALED(status)) {
1360                                 int sig = WTERMSIG(status);
1361
1362                                 if (sig != SIGTERM)
1363                                         log_err("fio: pid=%d, got signal=%d\n",
1364                                                         (int) td->pid, sig);
1365                                 td_set_runstate(td, TD_REAPED);
1366                                 goto reaped;
1367                         }
1368                         if (WIFEXITED(status)) {
1369                                 if (WEXITSTATUS(status) && !td->error)
1370                                         td->error = WEXITSTATUS(status);
1371
1372                                 td_set_runstate(td, TD_REAPED);
1373                                 goto reaped;
1374                         }
1375                 }
1376
1377                 /*
1378                  * thread is not dead, continue
1379                  */
1380                 pending++;
1381                 continue;
1382 reaped:
1383                 (*nr_running)--;
1384                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1385                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1386                 if (!td->pid)
1387                         pending--;
1388
1389                 if (td->error)
1390                         exit_value++;
1391
1392                 done_secs += mtime_since_now(&td->epoch) / 1000;
1393         }
1394
1395         if (*nr_running == cputhreads && !pending && realthreads)
1396                 terminate_threads(TERMINATE_ALL);
1397 }
1398
1399 static void *gtod_thread_main(void *data)
1400 {
1401         fio_mutex_up(startup_mutex);
1402
1403         /*
1404          * As long as we have jobs around, update the clock. It would be nice
1405          * to have some way of NOT hammering that CPU with gettimeofday(),
1406          * but I'm not sure what to use outside of a simple CPU nop to relax
1407          * it - we don't want to lose precision.
1408          */
1409         while (threads) {
1410                 fio_gtod_update();
1411                 nop;
1412         }
1413
1414         return NULL;
1415 }
1416
1417 static int fio_start_gtod_thread(void)
1418 {
1419         pthread_attr_t attr;
1420         int ret;
1421
1422         pthread_attr_init(&attr);
1423         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1424         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1425         pthread_attr_destroy(&attr);
1426         if (ret) {
1427                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1428                 return 1;
1429         }
1430
1431         ret = pthread_detach(gtod_thread);
1432         if (ret) {
1433                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1434                 return 1;
1435         }
1436
1437         dprint(FD_MUTEX, "wait on startup_mutex\n");
1438         fio_mutex_down(startup_mutex);
1439         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1440         return 0;
1441 }
1442
1443 /*
1444  * Main function for kicking off and reaping jobs, as needed.
1445  */
1446 static void run_threads(void)
1447 {
1448         struct thread_data *td;
1449         unsigned long spent;
1450         int i, todo, nr_running, m_rate, t_rate, nr_started;
1451
1452         if (fio_pin_memory())
1453                 return;
1454
1455         if (fio_gtod_offload && fio_start_gtod_thread())
1456                 return;
1457
1458         if (!terse_output) {
1459                 log_info("Starting ");
1460                 if (nr_thread)
1461                         log_info("%d thread%s", nr_thread,
1462                                                 nr_thread > 1 ? "s" : "");
1463                 if (nr_process) {
1464                         if (nr_thread)
1465                                 printf(" and ");
1466                         log_info("%d process%s", nr_process,
1467                                                 nr_process > 1 ? "es" : "");
1468                 }
1469                 log_info("\n");
1470                 fflush(stdout);
1471         }
1472
1473         set_sig_handlers();
1474
1475         todo = thread_number;
1476         nr_running = 0;
1477         nr_started = 0;
1478         m_rate = t_rate = 0;
1479
1480         for_each_td(td, i) {
1481                 print_status_init(td->thread_number - 1);
1482
1483                 if (!td->o.create_serialize) {
1484                         init_disk_util(td);
1485                         continue;
1486                 }
1487
1488                 /*
1489                  * do file setup here so it happens sequentially,
1490                  * we don't want X number of threads getting their
1491                  * client data interspersed on disk
1492                  */
1493                 if (setup_files(td)) {
1494                         exit_value++;
1495                         if (td->error)
1496                                 log_err("fio: pid=%d, err=%d/%s\n",
1497                                         (int) td->pid, td->error, td->verror);
1498                         td_set_runstate(td, TD_REAPED);
1499                         todo--;
1500                 } else {
1501                         struct fio_file *f;
1502                         unsigned int j;
1503
1504                         /*
1505                          * for sharing to work, each job must always open
1506                          * its own files. so close them, if we opened them
1507                          * for creation
1508                          */
1509                         for_each_file(td, f, j) {
1510                                 if (fio_file_open(f))
1511                                         td_io_close_file(td, f);
1512                         }
1513                 }
1514
1515                 init_disk_util(td);
1516         }
1517
1518         set_genesis_time();
1519
1520         while (todo) {
1521                 struct thread_data *map[MAX_JOBS];
1522                 struct timeval this_start;
1523                 int this_jobs = 0, left;
1524
1525                 /*
1526                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1527                  */
1528                 for_each_td(td, i) {
1529                         if (td->runstate != TD_NOT_CREATED)
1530                                 continue;
1531
1532                         /*
1533                          * never got a chance to start, killed by other
1534                          * thread for some reason
1535                          */
1536                         if (td->terminate) {
1537                                 todo--;
1538                                 continue;
1539                         }
1540
1541                         if (td->o.start_delay) {
1542                                 spent = mtime_since_genesis();
1543
1544                                 if (td->o.start_delay * 1000 > spent)
1545                                         continue;
1546                         }
1547
1548                         if (td->o.stonewall && (nr_started || nr_running)) {
1549                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1550                                                         td->o.name);
1551                                 break;
1552                         }
1553
1554                         /*
1555                          * Set state to created. Thread will transition
1556                          * to TD_INITIALIZED when it's done setting up.
1557                          */
1558                         td_set_runstate(td, TD_CREATED);
1559                         map[this_jobs++] = td;
1560                         nr_started++;
1561
1562                         if (td->o.use_thread) {
1563                                 int ret;
1564
1565                                 dprint(FD_PROCESS, "will pthread_create\n");
1566                                 ret = pthread_create(&td->thread, NULL,
1567                                                         thread_main, td);
1568                                 if (ret) {
1569                                         log_err("pthread_create: %s\n",
1570                                                         strerror(ret));
1571                                         nr_started--;
1572                                         break;
1573                                 }
1574                                 ret = pthread_detach(td->thread);
1575                                 if (ret)
1576                                         log_err("pthread_detach: %s",
1577                                                         strerror(ret));
1578                         } else {
1579                                 pid_t pid;
1580                                 dprint(FD_PROCESS, "will fork\n");
1581                                 pid = fork();
1582                                 if (!pid) {
1583                                         int ret = fork_main(shm_id, i);
1584
1585                                         _exit(ret);
1586                                 } else if (i == fio_debug_jobno)
1587                                         *fio_debug_jobp = pid;
1588                         }
1589                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1590                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1591                                 log_err("fio: job startup hung? exiting.\n");
1592                                 terminate_threads(TERMINATE_ALL);
1593                                 fio_abort = 1;
1594                                 nr_started--;
1595                                 break;
1596                         }
1597                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1598                 }
1599
1600                 /*
1601                  * Wait for the started threads to transition to
1602                  * TD_INITIALIZED.
1603                  */
1604                 fio_gettime(&this_start, NULL);
1605                 left = this_jobs;
1606                 while (left && !fio_abort) {
1607                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1608                                 break;
1609
1610                         usleep(100000);
1611
1612                         for (i = 0; i < this_jobs; i++) {
1613                                 td = map[i];
1614                                 if (!td)
1615                                         continue;
1616                                 if (td->runstate == TD_INITIALIZED) {
1617                                         map[i] = NULL;
1618                                         left--;
1619                                 } else if (td->runstate >= TD_EXITED) {
1620                                         map[i] = NULL;
1621                                         left--;
1622                                         todo--;
1623                                         nr_running++; /* work-around... */
1624                                 }
1625                         }
1626                 }
1627
1628                 if (left) {
1629                         log_err("fio: %d jobs failed to start\n", left);
1630                         for (i = 0; i < this_jobs; i++) {
1631                                 td = map[i];
1632                                 if (!td)
1633                                         continue;
1634                                 kill(td->pid, SIGTERM);
1635                         }
1636                         break;
1637                 }
1638
1639                 /*
1640                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1641                  */
1642                 for_each_td(td, i) {
1643                         if (td->runstate != TD_INITIALIZED)
1644                                 continue;
1645
1646                         if (in_ramp_time(td))
1647                                 td_set_runstate(td, TD_RAMP);
1648                         else
1649                                 td_set_runstate(td, TD_RUNNING);
1650                         nr_running++;
1651                         nr_started--;
1652                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1653                         t_rate += td->o.rate[0] + td->o.rate[1];
1654                         todo--;
1655                         fio_mutex_up(td->mutex);
1656                 }
1657
1658                 reap_threads(&nr_running, &t_rate, &m_rate);
1659
1660                 if (todo)
1661                         usleep(100000);
1662         }
1663
1664         while (nr_running) {
1665                 reap_threads(&nr_running, &t_rate, &m_rate);
1666                 usleep(10000);
1667         }
1668
1669         update_io_ticks();
1670         fio_unpin_memory();
1671 }
1672
1673 int main(int argc, char *argv[])
1674 {
1675         long ps;
1676
1677         sinit();
1678         init_rand(&__fio_rand_state);
1679
1680         /*
1681          * We need locale for number printing, if it isn't set then just
1682          * go with the US format.
1683          */
1684         if (!getenv("LC_NUMERIC"))
1685                 setlocale(LC_NUMERIC, "en_US");
1686
1687         ps = sysconf(_SC_PAGESIZE);
1688         if (ps < 0) {
1689                 log_err("Failed to get page size\n");
1690                 return 1;
1691         }
1692
1693         page_size = ps;
1694         page_mask = ps - 1;
1695
1696         fio_keywords_init();
1697
1698         if (parse_options(argc, argv))
1699                 return 1;
1700
1701         if (exec_profile && load_profile(exec_profile))
1702                 return 1;
1703
1704         if (!thread_number)
1705                 return 0;
1706
1707         if (write_bw_log) {
1708                 setup_log(&agg_io_log[DDIR_READ]);
1709                 setup_log(&agg_io_log[DDIR_WRITE]);
1710         }
1711
1712         startup_mutex = fio_mutex_init(0);
1713         if (startup_mutex == NULL)
1714                 return 1;
1715         writeout_mutex = fio_mutex_init(1);
1716         if (writeout_mutex == NULL)
1717                 return 1;
1718
1719         set_genesis_time();
1720         create_disk_util_thread();
1721
1722         cgroup_list = smalloc(sizeof(*cgroup_list));
1723         INIT_FLIST_HEAD(cgroup_list);
1724
1725         run_threads();
1726
1727         if (!fio_abort) {
1728                 show_run_stats();
1729                 if (write_bw_log) {
1730                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1731                         __finish_log(agg_io_log[DDIR_WRITE],
1732                                         "agg-write_bw.log");
1733                 }
1734         }
1735
1736         cgroup_kill(cgroup_list);
1737         sfree(cgroup_list);
1738         sfree(cgroup_mnt);
1739
1740         fio_mutex_remove(startup_mutex);
1741         fio_mutex_remove(writeout_mutex);
1742         return exit_value;
1743 }