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