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 #include "server.h"
  50
  51 unsigned long page_mask;
  52 unsigned long page_size;
  53
  54 #define PAGE_ALIGN(buf) \
  55         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  56
  57 int groupid = 0;
  58 int thread_number = 0;
  59 int nr_process = 0;
  60 int nr_thread = 0;
  61 int shm_id = 0;
  62 int temp_stall_ts;
  63 unsigned long done_secs = 0;
  64
  65 static struct fio_mutex *startup_mutex;
  66 static struct fio_mutex *writeout_mutex;
  67 static volatile int fio_abort;
  68 static int exit_value;
  69 static pthread_t gtod_thread;
  70 static pthread_t disk_util_thread;
  71 static struct flist_head *cgroup_list;
  72 static char *cgroup_mnt;
  73
  74 unsigned long arch_flags = 0;
  75
  76 struct io_log *agg_io_log[2];
  77
  78 #define TERMINATE_ALL           (-1)
  79 #define JOB_START_TIMEOUT       (5 * 1000)
  80
  81 void td_set_runstate(struct thread_data *td, int runstate)
  82 {
  83         if (td->runstate == runstate)
  84                 return;
  85
  86         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  87                                                 td->runstate, runstate);
  88         td->runstate = runstate;
  89 }
  90
  91 static void terminate_threads(int group_id)
  92 {
  93         struct thread_data *td;
  94         int i;
  95
  96         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  97
  98         for_each_td(td, i) {
  99                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
 100                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
 101                                                 td->o.name, (int) td->pid);
 102                         td->terminate = 1;
 103                         td->o.start_delay = 0;
 104
 105                         /*
 106                          * if the thread is running, just let it exit
 107                          */
 108                         if (!td->pid)
 109                                 continue;
 110                         else if (td->runstate < TD_RAMP)
 111                                 kill(td->pid, SIGTERM);
 112                         else {
 113                                 struct ioengine_ops *ops = td->io_ops;
 114
 115                                 if (ops && (ops->flags & FIO_SIGTERM))
 116                                         kill(td->pid, SIGTERM);
 117                         }
 118                 }
 119         }
 120 }
 121
 122 static void sig_int(int sig)
 123 {
 124         if (threads) {
 125                 log_info("\nfio: terminating on signal %d\n", sig);
 126                 exit_backend = 1;
 127                 fflush(stdout);
 128                 exit_value = 128;
 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
 143                 if (!is_backend)
 144                         print_thread_status();
 145         }
 146
 147         return NULL;
 148 }
 149
 150 static int create_disk_util_thread(void)
 151 {
 152         int ret;
 153
 154         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 155         if (ret) {
 156                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 157                 return 1;
 158         }
 159
 160         ret = pthread_detach(disk_util_thread);
 161         if (ret) {
 162                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 163                 return 1;
 164         }
 165
 166         dprint(FD_MUTEX, "wait on startup_mutex\n");
 167         fio_mutex_down(startup_mutex);
 168         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 169         return 0;
 170 }
 171
 172 static void set_sig_handlers(void)
 173 {
 174         struct sigaction act;
 175
 176         memset(&act, 0, sizeof(act));
 177         act.sa_handler = sig_int;
 178         act.sa_flags = SA_RESTART;
 179         sigaction(SIGINT, &act, NULL);
 180
 181         memset(&act, 0, sizeof(act));
 182         act.sa_handler = sig_int;
 183         act.sa_flags = SA_RESTART;
 184         sigaction(SIGTERM, &act, NULL);
 185 }
 186
 187 /*
 188  * Check if we are above the minimum rate given.
 189  */
 190 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 191                             enum fio_ddir ddir)
 192 {
 193         unsigned long long bytes = 0;
 194         unsigned long iops = 0;
 195         unsigned long spent;
 196         unsigned long rate;
 197         unsigned int ratemin = 0;
 198         unsigned int rate_iops = 0;
 199         unsigned int rate_iops_min = 0;
 200
 201         assert(ddir_rw(ddir));
 202
 203         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 204                 return 0;
 205
 206         /*
 207          * allow a 2 second settle period in the beginning
 208          */
 209         if (mtime_since(&td->start, now) < 2000)
 210                 return 0;
 211
 212         iops += td->io_blocks[ddir];
 213         bytes += td->this_io_bytes[ddir];
 214         ratemin += td->o.ratemin[ddir];
 215         rate_iops += td->o.rate_iops[ddir];
 216         rate_iops_min += td->o.rate_iops_min[ddir];
 217
 218         /*
 219          * if rate blocks is set, sample is running
 220          */
 221         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 222                 spent = mtime_since(&td->lastrate[ddir], now);
 223                 if (spent < td->o.ratecycle)
 224                         return 0;
 225
 226                 if (td->o.rate[ddir]) {
 227                         /*
 228                          * check bandwidth specified rate
 229                          */
 230                         if (bytes < td->rate_bytes[ddir]) {
 231                                 log_err("%s: min rate %u not met\n", td->o.name,
 232                                                                 ratemin);
 233                                 return 1;
 234                         } else {
 235                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 236                                 if (rate < ratemin ||
 237                                     bytes < td->rate_bytes[ddir]) {
 238                                         log_err("%s: min rate %u not met, got"
 239                                                 " %luKB/sec\n", td->o.name,
 240                                                         ratemin, rate);
 241                                         return 1;
 242                                 }
 243                         }
 244                 } else {
 245                         /*
 246                          * checks iops specified rate
 247                          */
 248                         if (iops < rate_iops) {
 249                                 log_err("%s: min iops rate %u not met\n",
 250                                                 td->o.name, rate_iops);
 251                                 return 1;
 252                         } else {
 253                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 254                                 if (rate < rate_iops_min ||
 255                                     iops < td->rate_blocks[ddir]) {
 256                                         log_err("%s: min iops rate %u not met,"
 257                                                 " got %lu\n", td->o.name,
 258                                                         rate_iops_min, rate);
 259                                 }
 260                         }
 261                 }
 262         }
 263
 264         td->rate_bytes[ddir] = bytes;
 265         td->rate_blocks[ddir] = iops;
 266         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 267         return 0;
 268 }
 269
 270 static int check_min_rate(struct thread_data *td, struct timeval *now,
 271                           unsigned long *bytes_done)
 272 {
 273         int ret = 0;
 274
 275         if (bytes_done[0])
 276                 ret |= __check_min_rate(td, now, 0);
 277         if (bytes_done[1])
 278                 ret |= __check_min_rate(td, now, 1);
 279
 280         return ret;
 281 }
 282
 283 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 284 {
 285         if (!td->o.timeout)
 286                 return 0;
 287         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 288                 return 1;
 289
 290         return 0;
 291 }
 292
 293 /*
 294  * When job exits, we can cancel the in-flight IO if we are using async
 295  * io. Attempt to do so.
 296  */
 297 static void cleanup_pending_aio(struct thread_data *td)
 298 {
 299         struct flist_head *entry, *n;
 300         struct io_u *io_u;
 301         int r;
 302
 303         /*
 304          * get immediately available events, if any
 305          */
 306         r = io_u_queued_complete(td, 0, NULL);
 307         if (r < 0)
 308                 return;
 309
 310         /*
 311          * now cancel remaining active events
 312          */
 313         if (td->io_ops->cancel) {
 314                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 315                         io_u = flist_entry(entry, struct io_u, list);
 316
 317                         /*
 318                          * if the io_u isn't in flight, then that generally
 319                          * means someone leaked an io_u. complain but fix
 320                          * it up, so we don't stall here.
 321                          */
 322                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 323                                 log_err("fio: non-busy IO on busy list\n");
 324                                 put_io_u(td, io_u);
 325                         } else {
 326                                 r = td->io_ops->cancel(td, io_u);
 327                                 if (!r)
 328                                         put_io_u(td, io_u);
 329                         }
 330                 }
 331         }
 332
 333         if (td->cur_depth)
 334                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 335 }
 336
 337 /*
 338  * Helper to handle the final sync of a file. Works just like the normal
 339  * io path, just does everything sync.
 340  */
 341 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 342 {
 343         struct io_u *io_u = __get_io_u(td);
 344         int ret;
 345
 346         if (!io_u)
 347                 return 1;
 348
 349         io_u->ddir = DDIR_SYNC;
 350         io_u->file = f;
 351
 352         if (td_io_prep(td, io_u)) {
 353                 put_io_u(td, io_u);
 354                 return 1;
 355         }
 356
 357 requeue:
 358         ret = td_io_queue(td, io_u);
 359         if (ret < 0) {
 360                 td_verror(td, io_u->error, "td_io_queue");
 361                 put_io_u(td, io_u);
 362                 return 1;
 363         } else if (ret == FIO_Q_QUEUED) {
 364                 if (io_u_queued_complete(td, 1, NULL) < 0)
 365                         return 1;
 366         } else if (ret == FIO_Q_COMPLETED) {
 367                 if (io_u->error) {
 368                         td_verror(td, io_u->error, "td_io_queue");
 369                         return 1;
 370                 }
 371
 372                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 373                         return 1;
 374         } else if (ret == FIO_Q_BUSY) {
 375                 if (td_io_commit(td))
 376                         return 1;
 377                 goto requeue;
 378         }
 379
 380         return 0;
 381 }
 382
 383 static inline void __update_tv_cache(struct thread_data *td)
 384 {
 385         fio_gettime(&td->tv_cache, NULL);
 386 }
 387
 388 static inline void update_tv_cache(struct thread_data *td)
 389 {
 390         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 391                 __update_tv_cache(td);
 392 }
 393
 394 static int break_on_this_error(struct thread_data *td, int *retptr)
 395 {
 396         int ret = *retptr;
 397
 398         if (ret < 0 || td->error) {
 399                 int err;
 400
 401                 if (!td->o.continue_on_error)
 402                         return 1;
 403
 404                 if (ret < 0)
 405                         err = -ret;
 406                 else
 407                         err = td->error;
 408
 409                 if (td_non_fatal_error(err)) {
 410                         /*
 411                          * Continue with the I/Os in case of
 412                          * a non fatal error.
 413                          */
 414                         update_error_count(td, err);
 415                         td_clear_error(td);
 416                         *retptr = 0;
 417                         return 0;
 418                 } else if (td->o.fill_device && err == ENOSPC) {
 419                         /*
 420                          * We expect to hit this error if
 421                          * fill_device option is set.
 422                          */
 423                         td_clear_error(td);
 424                         td->terminate = 1;
 425                         return 1;
 426                 } else {
 427                         /*
 428                          * Stop the I/O in case of a fatal
 429                          * error.
 430                          */
 431                         update_error_count(td, err);
 432                         return 1;
 433                 }
 434         }
 435
 436         return 0;
 437 }
 438
 439 /*
 440  * The main verify engine. Runs over the writes we previously submitted,
 441  * reads the blocks back in, and checks the crc/md5 of the data.
 442  */
 443 static void do_verify(struct thread_data *td)
 444 {
 445         struct fio_file *f;
 446         struct io_u *io_u;
 447         int ret, min_events;
 448         unsigned int i;
 449
 450         dprint(FD_VERIFY, "starting loop\n");
 451
 452         /*
 453          * sync io first and invalidate cache, to make sure we really
 454          * read from disk.
 455          */
 456         for_each_file(td, f, i) {
 457                 if (!fio_file_open(f))
 458                         continue;
 459                 if (fio_io_sync(td, f))
 460                         break;
 461                 if (file_invalidate_cache(td, f))
 462                         break;
 463         }
 464
 465         if (td->error)
 466                 return;
 467
 468         td_set_runstate(td, TD_VERIFYING);
 469
 470         io_u = NULL;
 471         while (!td->terminate) {
 472                 int ret2, full;
 473
 474                 update_tv_cache(td);
 475
 476                 if (runtime_exceeded(td, &td->tv_cache)) {
 477                         __update_tv_cache(td);
 478                         if (runtime_exceeded(td, &td->tv_cache)) {
 479                                 td->terminate = 1;
 480                                 break;
 481                         }
 482                 }
 483
 484                 io_u = __get_io_u(td);
 485                 if (!io_u)
 486                         break;
 487
 488                 if (get_next_verify(td, io_u)) {
 489                         put_io_u(td, io_u);
 490                         break;
 491                 }
 492
 493                 if (td_io_prep(td, io_u)) {
 494                         put_io_u(td, io_u);
 495                         break;
 496                 }
 497
 498                 if (td->o.verify_async)
 499                         io_u->end_io = verify_io_u_async;
 500                 else
 501                         io_u->end_io = verify_io_u;
 502
 503                 ret = td_io_queue(td, io_u);
 504                 switch (ret) {
 505                 case FIO_Q_COMPLETED:
 506                         if (io_u->error) {
 507                                 ret = -io_u->error;
 508                                 clear_io_u(td, io_u);
 509                         } else if (io_u->resid) {
 510                                 int bytes = io_u->xfer_buflen - io_u->resid;
 511
 512                                 /*
 513                                  * zero read, fail
 514                                  */
 515                                 if (!bytes) {
 516                                         td_verror(td, EIO, "full resid");
 517                                         put_io_u(td, io_u);
 518                                         break;
 519                                 }
 520
 521                                 io_u->xfer_buflen = io_u->resid;
 522                                 io_u->xfer_buf += bytes;
 523                                 io_u->offset += bytes;
 524
 525                                 if (ddir_rw(io_u->ddir))
 526                                         td->ts.short_io_u[io_u->ddir]++;
 527
 528                                 f = io_u->file;
 529                                 if (io_u->offset == f->real_file_size)
 530                                         goto sync_done;
 531
 532                                 requeue_io_u(td, &io_u);
 533                         } else {
 534 sync_done:
 535                                 ret = io_u_sync_complete(td, io_u, NULL);
 536                                 if (ret < 0)
 537                                         break;
 538                         }
 539                         continue;
 540                 case FIO_Q_QUEUED:
 541                         break;
 542                 case FIO_Q_BUSY:
 543                         requeue_io_u(td, &io_u);
 544                         ret2 = td_io_commit(td);
 545                         if (ret2 < 0)
 546                                 ret = ret2;
 547                         break;
 548                 default:
 549                         assert(ret < 0);
 550                         td_verror(td, -ret, "td_io_queue");
 551                         break;
 552                 }
 553
 554                 if (break_on_this_error(td, &ret))
 555                         break;
 556
 557                 /*
 558                  * if we can queue more, do so. but check if there are
 559                  * completed io_u's first. Note that we can get BUSY even
 560                  * without IO queued, if the system is resource starved.
 561                  */
 562                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 563                 if (full || !td->o.iodepth_batch_complete) {
 564                         min_events = min(td->o.iodepth_batch_complete,
 565                                          td->cur_depth);
 566                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 567                                 min_events = 1;
 568
 569                         do {
 570                                 /*
 571                                  * Reap required number of io units, if any,
 572                                  * and do the verification on them through
 573                                  * the callback handler
 574                                  */
 575                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 576                                         ret = -1;
 577                                         break;
 578                                 }
 579                         } while (full && (td->cur_depth > td->o.iodepth_low));
 580                 }
 581                 if (ret < 0)
 582                         break;
 583         }
 584
 585         if (!td->error) {
 586                 min_events = td->cur_depth;
 587
 588                 if (min_events)
 589                         ret = io_u_queued_complete(td, min_events, NULL);
 590         } else
 591                 cleanup_pending_aio(td);
 592
 593         td_set_runstate(td, TD_RUNNING);
 594
 595         dprint(FD_VERIFY, "exiting loop\n");
 596 }
 597
 598 /*
 599  * Main IO worker function. It retrieves io_u's to process and queues
 600  * and reaps them, checking for rate and errors along the way.
 601  */
 602 static void do_io(struct thread_data *td)
 603 {
 604         unsigned int i;
 605         int ret = 0;
 606
 607         if (in_ramp_time(td))
 608                 td_set_runstate(td, TD_RAMP);
 609         else
 610                 td_set_runstate(td, TD_RUNNING);
 611
 612         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 613                 (!flist_empty(&td->trim_list)) ||
 614                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 615                 struct timeval comp_time;
 616                 unsigned long bytes_done[2] = { 0, 0 };
 617                 int min_evts = 0;
 618                 struct io_u *io_u;
 619                 int ret2, full;
 620
 621                 if (td->terminate)
 622                         break;
 623
 624                 update_tv_cache(td);
 625
 626                 if (runtime_exceeded(td, &td->tv_cache)) {
 627                         __update_tv_cache(td);
 628                         if (runtime_exceeded(td, &td->tv_cache)) {
 629                                 td->terminate = 1;
 630                                 break;
 631                         }
 632                 }
 633
 634                 io_u = get_io_u(td);
 635                 if (!io_u)
 636                         break;
 637
 638                 /*
 639                  * Add verification end_io handler, if asked to verify
 640                  * a previously written file.
 641                  */
 642                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 643                     !td_rw(td)) {
 644                         if (td->o.verify_async)
 645                                 io_u->end_io = verify_io_u_async;
 646                         else
 647                                 io_u->end_io = verify_io_u;
 648                         td_set_runstate(td, TD_VERIFYING);
 649                 } else if (in_ramp_time(td))
 650                         td_set_runstate(td, TD_RAMP);
 651                 else
 652                         td_set_runstate(td, TD_RUNNING);
 653
 654                 ret = td_io_queue(td, io_u);
 655                 switch (ret) {
 656                 case FIO_Q_COMPLETED:
 657                         if (io_u->error) {
 658                                 ret = -io_u->error;
 659                                 clear_io_u(td, io_u);
 660                         } else if (io_u->resid) {
 661                                 int bytes = io_u->xfer_buflen - io_u->resid;
 662                                 struct fio_file *f = io_u->file;
 663
 664                                 /*
 665                                  * zero read, fail
 666                                  */
 667                                 if (!bytes) {
 668                                         td_verror(td, EIO, "full resid");
 669                                         put_io_u(td, io_u);
 670                                         break;
 671                                 }
 672
 673                                 io_u->xfer_buflen = io_u->resid;
 674                                 io_u->xfer_buf += bytes;
 675                                 io_u->offset += bytes;
 676
 677                                 if (ddir_rw(io_u->ddir))
 678                                         td->ts.short_io_u[io_u->ddir]++;
 679
 680                                 if (io_u->offset == f->real_file_size)
 681                                         goto sync_done;
 682
 683                                 requeue_io_u(td, &io_u);
 684                         } else {
 685 sync_done:
 686                                 if (__should_check_rate(td, 0) ||
 687                                     __should_check_rate(td, 1))
 688                                         fio_gettime(&comp_time, NULL);
 689
 690                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 691                                 if (ret < 0)
 692                                         break;
 693                         }
 694                         break;
 695                 case FIO_Q_QUEUED:
 696                         /*
 697                          * if the engine doesn't have a commit hook,
 698                          * the io_u is really queued. if it does have such
 699                          * a hook, it has to call io_u_queued() itself.
 700                          */
 701                         if (td->io_ops->commit == NULL)
 702                                 io_u_queued(td, io_u);
 703                         break;
 704                 case FIO_Q_BUSY:
 705                         requeue_io_u(td, &io_u);
 706                         ret2 = td_io_commit(td);
 707                         if (ret2 < 0)
 708                                 ret = ret2;
 709                         break;
 710                 default:
 711                         assert(ret < 0);
 712                         put_io_u(td, io_u);
 713                         break;
 714                 }
 715
 716                 if (break_on_this_error(td, &ret))
 717                         break;
 718
 719                 /*
 720                  * See if we need to complete some commands. Note that we
 721                  * can get BUSY even without IO queued, if the system is
 722                  * resource starved.
 723                  */
 724                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 725                 if (full || !td->o.iodepth_batch_complete) {
 726                         min_evts = min(td->o.iodepth_batch_complete,
 727                                         td->cur_depth);
 728                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 729                                 min_evts = 1;
 730
 731                         if (__should_check_rate(td, 0) ||
 732                             __should_check_rate(td, 1))
 733                                 fio_gettime(&comp_time, NULL);
 734
 735                         do {
 736                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 737                                 if (ret < 0)
 738                                         break;
 739
 740                         } while (full && (td->cur_depth > td->o.iodepth_low));
 741                 }
 742
 743                 if (ret < 0)
 744                         break;
 745                 if (!(bytes_done[0] + bytes_done[1]))
 746                         continue;
 747
 748                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 749                         if (check_min_rate(td, &comp_time, bytes_done)) {
 750                                 if (exitall_on_terminate)
 751                                         terminate_threads(td->groupid);
 752                                 td_verror(td, EIO, "check_min_rate");
 753                                 break;
 754                         }
 755                 }
 756
 757                 if (td->o.thinktime) {
 758                         unsigned long long b;
 759
 760                         b = td->io_blocks[0] + td->io_blocks[1];
 761                         if (!(b % td->o.thinktime_blocks)) {
 762                                 int left;
 763
 764                                 if (td->o.thinktime_spin)
 765                                         usec_spin(td->o.thinktime_spin);
 766
 767                                 left = td->o.thinktime - td->o.thinktime_spin;
 768                                 if (left)
 769                                         usec_sleep(td, left);
 770                         }
 771                 }
 772         }
 773
 774         if (td->trim_entries)
 775                 printf("trim entries %ld\n", td->trim_entries);
 776
 777         if (td->o.fill_device && td->error == ENOSPC) {
 778                 td->error = 0;
 779                 td->terminate = 1;
 780         }
 781         if (!td->error) {
 782                 struct fio_file *f;
 783
 784                 i = td->cur_depth;
 785                 if (i) {
 786                         ret = io_u_queued_complete(td, i, NULL);
 787                         if (td->o.fill_device && td->error == ENOSPC)
 788                                 td->error = 0;
 789                 }
 790
 791                 if (should_fsync(td) && td->o.end_fsync) {
 792                         td_set_runstate(td, TD_FSYNCING);
 793
 794                         for_each_file(td, f, i) {
 795                                 if (!fio_file_open(f))
 796                                         continue;
 797                                 fio_io_sync(td, f);
 798                         }
 799                 }
 800         } else
 801                 cleanup_pending_aio(td);
 802
 803         /*
 804          * stop job if we failed doing any IO
 805          */
 806         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 807                 td->done = 1;
 808 }
 809
 810 static void cleanup_io_u(struct thread_data *td)
 811 {
 812         struct flist_head *entry, *n;
 813         struct io_u *io_u;
 814
 815         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 816                 io_u = flist_entry(entry, struct io_u, list);
 817
 818                 flist_del(&io_u->list);
 819                 fio_memfree(io_u, sizeof(*io_u));
 820         }
 821
 822         free_io_mem(td);
 823 }
 824
 825 static int init_io_u(struct thread_data *td)
 826 {
 827         struct io_u *io_u;
 828         unsigned int max_bs;
 829         int cl_align, i, max_units;
 830         char *p;
 831
 832         max_units = td->o.iodepth;
 833         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 834         td->orig_buffer_size = (unsigned long long) max_bs
 835                                         * (unsigned long long) max_units;
 836
 837         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 838                 unsigned long bs;
 839
 840                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 841                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 842         }
 843
 844         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 845                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 846                 return 1;
 847         }
 848
 849         if (allocate_io_mem(td))
 850                 return 1;
 851
 852         if (td->o.odirect || td->o.mem_align ||
 853             (td->io_ops->flags & FIO_RAWIO))
 854                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 855         else
 856                 p = td->orig_buffer;
 857
 858         cl_align = os_cache_line_size();
 859
 860         for (i = 0; i < max_units; i++) {
 861                 void *ptr;
 862
 863                 if (td->terminate)
 864                         return 1;
 865
 866                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 867                 if (!ptr) {
 868                         log_err("fio: unable to allocate aligned memory\n");
 869                         break;
 870                 }
 871
 872                 io_u = ptr;
 873                 memset(io_u, 0, sizeof(*io_u));
 874                 INIT_FLIST_HEAD(&io_u->list);
 875                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 876
 877                 if (!(td->io_ops->flags & FIO_NOIO)) {
 878                         io_u->buf = p + max_bs * i;
 879                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 880
 881                         if (td_write(td))
 882                                 io_u_fill_buffer(td, io_u, max_bs);
 883                         if (td_write(td) && td->o.verify_pattern_bytes) {
 884                                 /*
 885                                  * Fill the buffer with the pattern if we are
 886                                  * going to be doing writes.
 887                                  */
 888                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 889                         }
 890                 }
 891
 892                 io_u->index = i;
 893                 io_u->flags = IO_U_F_FREE;
 894                 flist_add(&io_u->list, &td->io_u_freelist);
 895         }
 896
 897         return 0;
 898 }
 899
 900 static int switch_ioscheduler(struct thread_data *td)
 901 {
 902         char tmp[256], tmp2[128];
 903         FILE *f;
 904         int ret;
 905
 906         if (td->io_ops->flags & FIO_DISKLESSIO)
 907                 return 0;
 908
 909         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 910
 911         f = fopen(tmp, "r+");
 912         if (!f) {
 913                 if (errno == ENOENT) {
 914                         log_err("fio: os or kernel doesn't support IO scheduler"
 915                                 " switching\n");
 916                         return 0;
 917                 }
 918                 td_verror(td, errno, "fopen iosched");
 919                 return 1;
 920         }
 921
 922         /*
 923          * Set io scheduler.
 924          */
 925         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 926         if (ferror(f) || ret != 1) {
 927                 td_verror(td, errno, "fwrite");
 928                 fclose(f);
 929                 return 1;
 930         }
 931
 932         rewind(f);
 933
 934         /*
 935          * Read back and check that the selected scheduler is now the default.
 936          */
 937         ret = fread(tmp, 1, sizeof(tmp), f);
 938         if (ferror(f) || ret < 0) {
 939                 td_verror(td, errno, "fread");
 940                 fclose(f);
 941                 return 1;
 942         }
 943
 944         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 945         if (!strstr(tmp, tmp2)) {
 946                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 947                 td_verror(td, EINVAL, "iosched_switch");
 948                 fclose(f);
 949                 return 1;
 950         }
 951
 952         fclose(f);
 953         return 0;
 954 }
 955
 956 static int keep_running(struct thread_data *td)
 957 {
 958         unsigned long long io_done;
 959
 960         if (td->done)
 961                 return 0;
 962         if (td->o.time_based)
 963                 return 1;
 964         if (td->o.loops) {
 965                 td->o.loops--;
 966                 return 1;
 967         }
 968
 969         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 970                         + td->io_skip_bytes;
 971         if (io_done < td->o.size)
 972                 return 1;
 973
 974         return 0;
 975 }
 976
 977 static void reset_io_counters(struct thread_data *td)
 978 {
 979         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 980         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 981         td->zone_bytes = 0;
 982         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 983         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 984
 985         td->last_was_sync = 0;
 986
 987         /*
 988          * reset file done count if we are to start over
 989          */
 990         if (td->o.time_based || td->o.loops)
 991                 td->nr_done_files = 0;
 992 }
 993
 994 void reset_all_stats(struct thread_data *td)
 995 {
 996         struct timeval tv;
 997         int i;
 998
 999         reset_io_counters(td);
1000
1001         for (i = 0; i < 2; i++) {
1002                 td->io_bytes[i] = 0;
1003                 td->io_blocks[i] = 0;
1004                 td->io_issues[i] = 0;
1005                 td->ts.total_io_u[i] = 0;
1006         }
1007
1008         fio_gettime(&tv, NULL);
1009         td->ts.runtime[0] = 0;
1010         td->ts.runtime[1] = 0;
1011         memcpy(&td->epoch, &tv, sizeof(tv));
1012         memcpy(&td->start, &tv, sizeof(tv));
1013 }
1014
1015 static void clear_io_state(struct thread_data *td)
1016 {
1017         struct fio_file *f;
1018         unsigned int i;
1019
1020         reset_io_counters(td);
1021
1022         close_files(td);
1023         for_each_file(td, f, i)
1024                 fio_file_clear_done(f);
1025
1026         /*
1027          * Set the same seed to get repeatable runs
1028          */
1029         td_fill_rand_seeds(td);
1030 }
1031
1032 static int exec_string(const char *string)
1033 {
1034         int ret, newlen = strlen(string) + 1 + 8;
1035         char *str;
1036
1037         str = malloc(newlen);
1038         sprintf(str, "sh -c %s", string);
1039
1040         ret = system(str);
1041         if (ret == -1)
1042                 log_err("fio: exec of cmd <%s> failed\n", str);
1043
1044         free(str);
1045         return ret;
1046 }
1047
1048 /*
1049  * Entry point for the thread based jobs. The process based jobs end up
1050  * here as well, after a little setup.
1051  */
1052 static void *thread_main(void *data)
1053 {
1054         unsigned long long elapsed;
1055         struct thread_data *td = data;
1056         pthread_condattr_t attr;
1057         int clear_state;
1058
1059         if (!td->o.use_thread) {
1060                 setsid();
1061                 td->pid = getpid();
1062         } else
1063                 td->pid = gettid();
1064
1065         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1066
1067         INIT_FLIST_HEAD(&td->io_u_freelist);
1068         INIT_FLIST_HEAD(&td->io_u_busylist);
1069         INIT_FLIST_HEAD(&td->io_u_requeues);
1070         INIT_FLIST_HEAD(&td->io_log_list);
1071         INIT_FLIST_HEAD(&td->io_hist_list);
1072         INIT_FLIST_HEAD(&td->verify_list);
1073         INIT_FLIST_HEAD(&td->trim_list);
1074         pthread_mutex_init(&td->io_u_lock, NULL);
1075         td->io_hist_tree = RB_ROOT;
1076
1077         pthread_condattr_init(&attr);
1078         pthread_cond_init(&td->verify_cond, &attr);
1079         pthread_cond_init(&td->free_cond, &attr);
1080
1081         td_set_runstate(td, TD_INITIALIZED);
1082         dprint(FD_MUTEX, "up startup_mutex\n");
1083         fio_mutex_up(startup_mutex);
1084         dprint(FD_MUTEX, "wait on td->mutex\n");
1085         fio_mutex_down(td->mutex);
1086         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1087
1088         /*
1089          * the ->mutex mutex is now no longer used, close it to avoid
1090          * eating a file descriptor
1091          */
1092         fio_mutex_remove(td->mutex);
1093
1094         /*
1095          * A new gid requires privilege, so we need to do this before setting
1096          * the uid.
1097          */
1098         if (td->o.gid != -1U && setgid(td->o.gid)) {
1099                 td_verror(td, errno, "setgid");
1100                 goto err;
1101         }
1102         if (td->o.uid != -1U && setuid(td->o.uid)) {
1103                 td_verror(td, errno, "setuid");
1104                 goto err;
1105         }
1106
1107         /*
1108          * If we have a gettimeofday() thread, make sure we exclude that
1109          * thread from this job
1110          */
1111         if (td->o.gtod_cpu)
1112                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1113
1114         /*
1115          * Set affinity first, in case it has an impact on the memory
1116          * allocations.
1117          */
1118         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1119                 td_verror(td, errno, "cpu_set_affinity");
1120                 goto err;
1121         }
1122
1123         /*
1124          * May alter parameters that init_io_u() will use, so we need to
1125          * do this first.
1126          */
1127         if (init_iolog(td))
1128                 goto err;
1129
1130         if (init_io_u(td))
1131                 goto err;
1132
1133         if (td->o.verify_async && verify_async_init(td))
1134                 goto err;
1135
1136         if (td->ioprio_set) {
1137                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1138                         td_verror(td, errno, "ioprio_set");
1139                         goto err;
1140                 }
1141         }
1142
1143         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1144                 goto err;
1145
1146         if (nice(td->o.nice) == -1) {
1147                 td_verror(td, errno, "nice");
1148                 goto err;
1149         }
1150
1151         if (td->o.ioscheduler && switch_ioscheduler(td))
1152                 goto err;
1153
1154         if (!td->o.create_serialize && setup_files(td))
1155                 goto err;
1156
1157         if (td_io_init(td))
1158                 goto err;
1159
1160         if (init_random_map(td))
1161                 goto err;
1162
1163         if (td->o.exec_prerun) {
1164                 if (exec_string(td->o.exec_prerun))
1165                         goto err;
1166         }
1167
1168         if (td->o.pre_read) {
1169                 if (pre_read_files(td) < 0)
1170                         goto err;
1171         }
1172
1173         fio_gettime(&td->epoch, NULL);
1174         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1175
1176         clear_state = 0;
1177         while (keep_running(td)) {
1178                 fio_gettime(&td->start, NULL);
1179                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1180                                 sizeof(td->start));
1181                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1182                                 sizeof(td->start));
1183                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1184
1185                 if (td->o.ratemin[0] || td->o.ratemin[1])
1186                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1187                                                         sizeof(td->lastrate));
1188
1189                 if (clear_state)
1190                         clear_io_state(td);
1191
1192                 prune_io_piece_log(td);
1193
1194                 do_io(td);
1195
1196                 clear_state = 1;
1197
1198                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1199                         elapsed = utime_since_now(&td->start);
1200                         td->ts.runtime[DDIR_READ] += elapsed;
1201                 }
1202                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1203                         elapsed = utime_since_now(&td->start);
1204                         td->ts.runtime[DDIR_WRITE] += elapsed;
1205                 }
1206
1207                 if (td->error || td->terminate)
1208                         break;
1209
1210                 if (!td->o.do_verify ||
1211                     td->o.verify == VERIFY_NONE ||
1212                     (td->io_ops->flags & FIO_UNIDIR))
1213                         continue;
1214
1215                 clear_io_state(td);
1216
1217                 fio_gettime(&td->start, NULL);
1218
1219                 do_verify(td);
1220
1221                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1222
1223                 if (td->error || td->terminate)
1224                         break;
1225         }
1226
1227         update_rusage_stat(td);
1228         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1229         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1230         td->ts.total_run_time = mtime_since_now(&td->epoch);
1231         td->ts.io_bytes[0] = td->io_bytes[0];
1232         td->ts.io_bytes[1] = td->io_bytes[1];
1233
1234         fio_mutex_down(writeout_mutex);
1235         if (td->ts.bw_log) {
1236                 if (td->o.bw_log_file) {
1237                         finish_log_named(td, td->ts.bw_log,
1238                                                 td->o.bw_log_file, "bw");
1239                 } else
1240                         finish_log(td, td->ts.bw_log, "bw");
1241         }
1242         if (td->ts.lat_log) {
1243                 if (td->o.lat_log_file) {
1244                         finish_log_named(td, td->ts.lat_log,
1245                                                 td->o.lat_log_file, "lat");
1246                 } else
1247                         finish_log(td, td->ts.lat_log, "lat");
1248         }
1249         if (td->ts.slat_log) {
1250                 if (td->o.lat_log_file) {
1251                         finish_log_named(td, td->ts.slat_log,
1252                                                 td->o.lat_log_file, "slat");
1253                 } else
1254                         finish_log(td, td->ts.slat_log, "slat");
1255         }
1256         if (td->ts.clat_log) {
1257                 if (td->o.lat_log_file) {
1258                         finish_log_named(td, td->ts.clat_log,
1259                                                 td->o.lat_log_file, "clat");
1260                 } else
1261                         finish_log(td, td->ts.clat_log, "clat");
1262         }
1263         fio_mutex_up(writeout_mutex);
1264         if (td->o.exec_postrun)
1265                 exec_string(td->o.exec_postrun);
1266
1267         if (exitall_on_terminate)
1268                 terminate_threads(td->groupid);
1269
1270 err:
1271         if (td->error)
1272                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1273                                                         td->verror);
1274
1275         if (td->o.verify_async)
1276                 verify_async_exit(td);
1277
1278         close_and_free_files(td);
1279         close_ioengine(td);
1280         cleanup_io_u(td);
1281         cgroup_shutdown(td, &cgroup_mnt);
1282
1283         if (td->o.cpumask_set) {
1284                 int ret = fio_cpuset_exit(&td->o.cpumask);
1285
1286                 td_verror(td, ret, "fio_cpuset_exit");
1287         }
1288
1289         /*
1290          * do this very late, it will log file closing as well
1291          */
1292         if (td->o.write_iolog_file)
1293                 write_iolog_close(td);
1294
1295         options_mem_free(td);
1296         td_set_runstate(td, TD_EXITED);
1297         return (void *) (unsigned long) td->error;
1298 }
1299
1300 /*
1301  * We cannot pass the td data into a forked process, so attach the td and
1302  * pass it to the thread worker.
1303  */
1304 static int fork_main(int shmid, int offset)
1305 {
1306         struct thread_data *td;
1307         void *data, *ret;
1308
1309 #ifndef __hpux
1310         data = shmat(shmid, NULL, 0);
1311         if (data == (void *) -1) {
1312                 int __err = errno;
1313
1314                 perror("shmat");
1315                 return __err;
1316         }
1317 #else
1318         /*
1319          * HP-UX inherits shm mappings?
1320          */
1321         data = threads;
1322 #endif
1323
1324         td = data + offset * sizeof(struct thread_data);
1325         ret = thread_main(td);
1326         shmdt(data);
1327         return (int) (unsigned long) ret;
1328 }
1329
1330 /*
1331  * Run over the job map and reap the threads that have exited, if any.
1332  */
1333 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1334 {
1335         struct thread_data *td;
1336         int i, cputhreads, realthreads, pending, status, ret;
1337
1338         /*
1339          * reap exited threads (TD_EXITED -> TD_REAPED)
1340          */
1341         realthreads = pending = cputhreads = 0;
1342         for_each_td(td, i) {
1343                 int flags = 0;
1344
1345                 /*
1346                  * ->io_ops is NULL for a thread that has closed its
1347                  * io engine
1348                  */
1349                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1350                         cputhreads++;
1351                 else
1352                         realthreads++;
1353
1354                 if (!td->pid) {
1355                         pending++;
1356                         continue;
1357                 }
1358                 if (td->runstate == TD_REAPED)
1359                         continue;
1360                 if (td->o.use_thread) {
1361                         if (td->runstate == TD_EXITED) {
1362                                 td_set_runstate(td, TD_REAPED);
1363                                 goto reaped;
1364                         }
1365                         continue;
1366                 }
1367
1368                 flags = WNOHANG;
1369                 if (td->runstate == TD_EXITED)
1370                         flags = 0;
1371
1372                 /*
1373                  * check if someone quit or got killed in an unusual way
1374                  */
1375                 ret = waitpid(td->pid, &status, flags);
1376                 if (ret < 0) {
1377                         if (errno == ECHILD) {
1378                                 log_err("fio: pid=%d disappeared %d\n",
1379                                                 (int) td->pid, td->runstate);
1380                                 td_set_runstate(td, TD_REAPED);
1381                                 goto reaped;
1382                         }
1383                         perror("waitpid");
1384                 } else if (ret == td->pid) {
1385                         if (WIFSIGNALED(status)) {
1386                                 int sig = WTERMSIG(status);
1387
1388                                 if (sig != SIGTERM)
1389                                         log_err("fio: pid=%d, got signal=%d\n",
1390                                                         (int) td->pid, sig);
1391                                 td_set_runstate(td, TD_REAPED);
1392                                 goto reaped;
1393                         }
1394                         if (WIFEXITED(status)) {
1395                                 if (WEXITSTATUS(status) && !td->error)
1396                                         td->error = WEXITSTATUS(status);
1397
1398                                 td_set_runstate(td, TD_REAPED);
1399                                 goto reaped;
1400                         }
1401                 }
1402
1403                 /*
1404                  * thread is not dead, continue
1405                  */
1406                 pending++;
1407                 continue;
1408 reaped:
1409                 (*nr_running)--;
1410                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1411                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1412                 if (!td->pid)
1413                         pending--;
1414
1415                 if (td->error)
1416                         exit_value++;
1417
1418                 done_secs += mtime_since_now(&td->epoch) / 1000;
1419         }
1420
1421         if (*nr_running == cputhreads && !pending && realthreads)
1422                 terminate_threads(TERMINATE_ALL);
1423 }
1424
1425 static void *gtod_thread_main(void *data)
1426 {
1427         fio_mutex_up(startup_mutex);
1428
1429         /*
1430          * As long as we have jobs around, update the clock. It would be nice
1431          * to have some way of NOT hammering that CPU with gettimeofday(),
1432          * but I'm not sure what to use outside of a simple CPU nop to relax
1433          * it - we don't want to lose precision.
1434          */
1435         while (threads) {
1436                 fio_gtod_update();
1437                 nop;
1438         }
1439
1440         return NULL;
1441 }
1442
1443 static int fio_start_gtod_thread(void)
1444 {
1445         pthread_attr_t attr;
1446         int ret;
1447
1448         pthread_attr_init(&attr);
1449         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1450         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1451         pthread_attr_destroy(&attr);
1452         if (ret) {
1453                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1454                 return 1;
1455         }
1456
1457         ret = pthread_detach(gtod_thread);
1458         if (ret) {
1459                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1460                 return 1;
1461         }
1462
1463         dprint(FD_MUTEX, "wait on startup_mutex\n");
1464         fio_mutex_down(startup_mutex);
1465         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1466         return 0;
1467 }
1468
1469 /*
1470  * Main function for kicking off and reaping jobs, as needed.
1471  */
1472 static void run_threads(void)
1473 {
1474         struct thread_data *td;
1475         unsigned long spent;
1476         int i, todo, nr_running, m_rate, t_rate, nr_started;
1477
1478         if (fio_pin_memory())
1479                 return;
1480
1481         if (fio_gtod_offload && fio_start_gtod_thread())
1482                 return;
1483
1484         if (!terse_output) {
1485                 log_info("Starting ");
1486                 if (nr_thread)
1487                         log_info("%d thread%s", nr_thread,
1488                                                 nr_thread > 1 ? "s" : "");
1489                 if (nr_process) {
1490                         if (nr_thread)
1491                                 printf(" and ");
1492                         log_info("%d process%s", nr_process,
1493                                                 nr_process > 1 ? "es" : "");
1494                 }
1495                 log_info("\n");
1496                 fflush(stdout);
1497         }
1498
1499         set_sig_handlers();
1500
1501         todo = thread_number;
1502         nr_running = 0;
1503         nr_started = 0;
1504         m_rate = t_rate = 0;
1505
1506         for_each_td(td, i) {
1507                 print_status_init(td->thread_number - 1);
1508
1509                 if (!td->o.create_serialize)
1510                         continue;
1511
1512                 /*
1513                  * do file setup here so it happens sequentially,
1514                  * we don't want X number of threads getting their
1515                  * client data interspersed on disk
1516                  */
1517                 if (setup_files(td)) {
1518                         exit_value++;
1519                         if (td->error)
1520                                 log_err("fio: pid=%d, err=%d/%s\n",
1521                                         (int) td->pid, td->error, td->verror);
1522                         td_set_runstate(td, TD_REAPED);
1523                         todo--;
1524                 } else {
1525                         struct fio_file *f;
1526                         unsigned int j;
1527
1528                         /*
1529                          * for sharing to work, each job must always open
1530                          * its own files. so close them, if we opened them
1531                          * for creation
1532                          */
1533                         for_each_file(td, f, j) {
1534                                 if (fio_file_open(f))
1535                                         td_io_close_file(td, f);
1536                         }
1537                 }
1538         }
1539
1540         set_genesis_time();
1541
1542         while (todo) {
1543                 struct thread_data *map[REAL_MAX_JOBS];
1544                 struct timeval this_start;
1545                 int this_jobs = 0, left;
1546
1547                 /*
1548                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1549                  */
1550                 for_each_td(td, i) {
1551                         if (td->runstate != TD_NOT_CREATED)
1552                                 continue;
1553
1554                         /*
1555                          * never got a chance to start, killed by other
1556                          * thread for some reason
1557                          */
1558                         if (td->terminate) {
1559                                 todo--;
1560                                 continue;
1561                         }
1562
1563                         if (td->o.start_delay) {
1564                                 spent = mtime_since_genesis();
1565
1566                                 if (td->o.start_delay * 1000 > spent)
1567                                         continue;
1568                         }
1569
1570                         if (td->o.stonewall && (nr_started || nr_running)) {
1571                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1572                                                         td->o.name);
1573                                 break;
1574                         }
1575
1576                         init_disk_util(td);
1577
1578                         /*
1579                          * Set state to created. Thread will transition
1580                          * to TD_INITIALIZED when it's done setting up.
1581                          */
1582                         td_set_runstate(td, TD_CREATED);
1583                         map[this_jobs++] = td;
1584                         nr_started++;
1585
1586                         if (td->o.use_thread) {
1587                                 int ret;
1588
1589                                 dprint(FD_PROCESS, "will pthread_create\n");
1590                                 ret = pthread_create(&td->thread, NULL,
1591                                                         thread_main, td);
1592                                 if (ret) {
1593                                         log_err("pthread_create: %s\n",
1594                                                         strerror(ret));
1595                                         nr_started--;
1596                                         break;
1597                                 }
1598                                 ret = pthread_detach(td->thread);
1599                                 if (ret)
1600                                         log_err("pthread_detach: %s",
1601                                                         strerror(ret));
1602                         } else {
1603                                 pid_t pid;
1604                                 dprint(FD_PROCESS, "will fork\n");
1605                                 pid = fork();
1606                                 if (!pid) {
1607                                         int ret = fork_main(shm_id, i);
1608
1609                                         _exit(ret);
1610                                 } else if (i == fio_debug_jobno)
1611                                         *fio_debug_jobp = pid;
1612                         }
1613                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1614                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1615                                 log_err("fio: job startup hung? exiting.\n");
1616                                 terminate_threads(TERMINATE_ALL);
1617                                 fio_abort = 1;
1618                                 nr_started--;
1619                                 break;
1620                         }
1621                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1622                 }
1623
1624                 /*
1625                  * Wait for the started threads to transition to
1626                  * TD_INITIALIZED.
1627                  */
1628                 fio_gettime(&this_start, NULL);
1629                 left = this_jobs;
1630                 while (left && !fio_abort) {
1631                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1632                                 break;
1633
1634                         usleep(100000);
1635
1636                         for (i = 0; i < this_jobs; i++) {
1637                                 td = map[i];
1638                                 if (!td)
1639                                         continue;
1640                                 if (td->runstate == TD_INITIALIZED) {
1641                                         map[i] = NULL;
1642                                         left--;
1643                                 } else if (td->runstate >= TD_EXITED) {
1644                                         map[i] = NULL;
1645                                         left--;
1646                                         todo--;
1647                                         nr_running++; /* work-around... */
1648                                 }
1649                         }
1650                 }
1651
1652                 if (left) {
1653                         log_err("fio: %d jobs failed to start\n", left);
1654                         for (i = 0; i < this_jobs; i++) {
1655                                 td = map[i];
1656                                 if (!td)
1657                                         continue;
1658                                 kill(td->pid, SIGTERM);
1659                         }
1660                         break;
1661                 }
1662
1663                 /*
1664                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1665                  */
1666                 for_each_td(td, i) {
1667                         if (td->runstate != TD_INITIALIZED)
1668                                 continue;
1669
1670                         if (in_ramp_time(td))
1671                                 td_set_runstate(td, TD_RAMP);
1672                         else
1673                                 td_set_runstate(td, TD_RUNNING);
1674                         nr_running++;
1675                         nr_started--;
1676                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1677                         t_rate += td->o.rate[0] + td->o.rate[1];
1678                         todo--;
1679                         fio_mutex_up(td->mutex);
1680                 }
1681
1682                 reap_threads(&nr_running, &t_rate, &m_rate);
1683
1684                 if (todo)
1685                         usleep(100000);
1686         }
1687
1688         while (nr_running) {
1689                 reap_threads(&nr_running, &t_rate, &m_rate);
1690                 usleep(10000);
1691         }
1692
1693         update_io_ticks();
1694         fio_unpin_memory();
1695 }
1696
1697 int exec_run(void)
1698 {
1699         if (exec_profile && load_profile(exec_profile))
1700                 return 1;
1701
1702         if (!thread_number)
1703                 return 0;
1704
1705         if (write_bw_log) {
1706                 setup_log(&agg_io_log[DDIR_READ]);
1707                 setup_log(&agg_io_log[DDIR_WRITE]);
1708         }
1709
1710         startup_mutex = fio_mutex_init(0);
1711         if (startup_mutex == NULL)
1712                 return 1;
1713         writeout_mutex = fio_mutex_init(1);
1714         if (writeout_mutex == NULL)
1715                 return 1;
1716
1717         set_genesis_time();
1718         create_disk_util_thread();
1719
1720         cgroup_list = smalloc(sizeof(*cgroup_list));
1721         INIT_FLIST_HEAD(cgroup_list);
1722
1723         run_threads();
1724
1725         if (!fio_abort) {
1726                 show_run_stats();
1727                 if (write_bw_log) {
1728                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1729                         __finish_log(agg_io_log[DDIR_WRITE],
1730                                         "agg-write_bw.log");
1731                 }
1732         }
1733
1734         cgroup_kill(cgroup_list);
1735         sfree(cgroup_list);
1736         sfree(cgroup_mnt);
1737
1738         fio_mutex_remove(startup_mutex);
1739         fio_mutex_remove(writeout_mutex);
1740         return exit_value;
1741 }
1742
1743 void reset_fio_state(void)
1744 {
1745         groupid = 0;
1746         thread_number = 0;
1747         nr_process = 0;
1748         nr_thread = 0;
1749         done_secs = 0;
1750 }
1751
1752 int main(int argc, char *argv[], char *envp[])
1753 {
1754         long ps;
1755
1756         arch_init(envp);
1757
1758         sinit();
1759
1760         /*
1761          * We need locale for number printing, if it isn't set then just
1762          * go with the US format.
1763          */
1764         if (!getenv("LC_NUMERIC"))
1765                 setlocale(LC_NUMERIC, "en_US");
1766
1767         ps = sysconf(_SC_PAGESIZE);
1768         if (ps < 0) {
1769                 log_err("Failed to get page size\n");
1770                 return 1;
1771         }
1772
1773         page_size = ps;
1774         page_mask = ps - 1;
1775
1776         fio_keywords_init();
1777
1778         if (parse_options(argc, argv))
1779                 return 1;
1780
1781         return exec_run();
1782 }