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 unsigned long arch_flags = 0;
  74
  75 struct io_log *agg_io_log[2];
  76
  77 #define TERMINATE_ALL           (-1)
  78 #define JOB_START_TIMEOUT       (5 * 1000)
  79
  80 void td_set_runstate(struct thread_data *td, int runstate)
  81 {
  82         if (td->runstate == runstate)
  83                 return;
  84
  85         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  86                                                 td->runstate, runstate);
  87         td->runstate = runstate;
  88 }
  89
  90 static void terminate_threads(int group_id)
  91 {
  92         struct thread_data *td;
  93         int i;
  94
  95         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  96
  97         for_each_td(td, i) {
  98                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  99                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
 100                                                 td->o.name, (int) td->pid);
 101                         td->terminate = 1;
 102                         td->o.start_delay = 0;
 103
 104                         /*
 105                          * if the thread is running, just let it exit
 106                          */
 107                         if (!td->pid)
 108                                 continue;
 109                         else if (td->runstate < TD_RAMP)
 110                                 kill(td->pid, SIGTERM);
 111                         else {
 112                                 struct ioengine_ops *ops = td->io_ops;
 113
 114                                 if (ops && (ops->flags & FIO_SIGTERM))
 115                                         kill(td->pid, SIGTERM);
 116                         }
 117                 }
 118         }
 119 }
 120
 121 static void sig_int(int sig)
 122 {
 123         if (threads) {
 124                 log_info("\nfio: terminating on signal %d\n", sig);
 125                 fflush(stdout);
 126                 exit_value = 128;
 127                 terminate_threads(TERMINATE_ALL);
 128         }
 129 }
 130
 131 static void *disk_thread_main(void *data)
 132 {
 133         fio_mutex_up(startup_mutex);
 134
 135         while (threads) {
 136                 usleep(DISK_UTIL_MSEC * 1000);
 137                 if (!threads)
 138                         break;
 139                 update_io_ticks();
 140                 print_thread_status();
 141         }
 142
 143         return NULL;
 144 }
 145
 146 static int create_disk_util_thread(void)
 147 {
 148         int ret;
 149
 150         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 151         if (ret) {
 152                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 153                 return 1;
 154         }
 155
 156         ret = pthread_detach(disk_util_thread);
 157         if (ret) {
 158                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 159                 return 1;
 160         }
 161
 162         dprint(FD_MUTEX, "wait on startup_mutex\n");
 163         fio_mutex_down(startup_mutex);
 164         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 165         return 0;
 166 }
 167
 168 static void set_sig_handlers(void)
 169 {
 170         struct sigaction act;
 171
 172         memset(&act, 0, sizeof(act));
 173         act.sa_handler = sig_int;
 174         act.sa_flags = SA_RESTART;
 175         sigaction(SIGINT, &act, NULL);
 176
 177         memset(&act, 0, sizeof(act));
 178         act.sa_handler = sig_int;
 179         act.sa_flags = SA_RESTART;
 180         sigaction(SIGTERM, &act, NULL);
 181 }
 182
 183 /*
 184  * Check if we are above the minimum rate given.
 185  */
 186 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 187                             enum fio_ddir ddir)
 188 {
 189         unsigned long long bytes = 0;
 190         unsigned long iops = 0;
 191         unsigned long spent;
 192         unsigned long rate;
 193         unsigned int ratemin = 0;
 194         unsigned int rate_iops = 0;
 195         unsigned int rate_iops_min = 0;
 196
 197         assert(ddir_rw(ddir));
 198
 199         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 200                 return 0;
 201
 202         /*
 203          * allow a 2 second settle period in the beginning
 204          */
 205         if (mtime_since(&td->start, now) < 2000)
 206                 return 0;
 207
 208         iops += td->io_blocks[ddir];
 209         bytes += td->this_io_bytes[ddir];
 210         ratemin += td->o.ratemin[ddir];
 211         rate_iops += td->o.rate_iops[ddir];
 212         rate_iops_min += td->o.rate_iops_min[ddir];
 213
 214         /*
 215          * if rate blocks is set, sample is running
 216          */
 217         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 218                 spent = mtime_since(&td->lastrate[ddir], now);
 219                 if (spent < td->o.ratecycle)
 220                         return 0;
 221
 222                 if (td->o.rate[ddir]) {
 223                         /*
 224                          * check bandwidth specified rate
 225                          */
 226                         if (bytes < td->rate_bytes[ddir]) {
 227                                 log_err("%s: min rate %u not met\n", td->o.name,
 228                                                                 ratemin);
 229                                 return 1;
 230                         } else {
 231                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 232                                 if (rate < ratemin ||
 233                                     bytes < td->rate_bytes[ddir]) {
 234                                         log_err("%s: min rate %u not met, got"
 235                                                 " %luKB/sec\n", td->o.name,
 236                                                         ratemin, rate);
 237                                         return 1;
 238                                 }
 239                         }
 240                 } else {
 241                         /*
 242                          * checks iops specified rate
 243                          */
 244                         if (iops < rate_iops) {
 245                                 log_err("%s: min iops rate %u not met\n",
 246                                                 td->o.name, rate_iops);
 247                                 return 1;
 248                         } else {
 249                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 250                                 if (rate < rate_iops_min ||
 251                                     iops < td->rate_blocks[ddir]) {
 252                                         log_err("%s: min iops rate %u not met,"
 253                                                 " got %lu\n", td->o.name,
 254                                                         rate_iops_min, rate);
 255                                 }
 256                         }
 257                 }
 258         }
 259
 260         td->rate_bytes[ddir] = bytes;
 261         td->rate_blocks[ddir] = iops;
 262         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 263         return 0;
 264 }
 265
 266 static int check_min_rate(struct thread_data *td, struct timeval *now,
 267                           unsigned long *bytes_done)
 268 {
 269         int ret = 0;
 270
 271         if (bytes_done[0])
 272                 ret |= __check_min_rate(td, now, 0);
 273         if (bytes_done[1])
 274                 ret |= __check_min_rate(td, now, 1);
 275
 276         return ret;
 277 }
 278
 279 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 280 {
 281         if (!td->o.timeout)
 282                 return 0;
 283         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 284                 return 1;
 285
 286         return 0;
 287 }
 288
 289 /*
 290  * When job exits, we can cancel the in-flight IO if we are using async
 291  * io. Attempt to do so.
 292  */
 293 static void cleanup_pending_aio(struct thread_data *td)
 294 {
 295         struct flist_head *entry, *n;
 296         struct io_u *io_u;
 297         int r;
 298
 299         /*
 300          * get immediately available events, if any
 301          */
 302         r = io_u_queued_complete(td, 0, NULL);
 303         if (r < 0)
 304                 return;
 305
 306         /*
 307          * now cancel remaining active events
 308          */
 309         if (td->io_ops->cancel) {
 310                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 311                         io_u = flist_entry(entry, struct io_u, list);
 312
 313                         /*
 314                          * if the io_u isn't in flight, then that generally
 315                          * means someone leaked an io_u. complain but fix
 316                          * it up, so we don't stall here.
 317                          */
 318                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 319                                 log_err("fio: non-busy IO on busy list\n");
 320                                 put_io_u(td, io_u);
 321                         } else {
 322                                 r = td->io_ops->cancel(td, io_u);
 323                                 if (!r)
 324                                         put_io_u(td, io_u);
 325                         }
 326                 }
 327         }
 328
 329         if (td->cur_depth)
 330                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 331 }
 332
 333 /*
 334  * Helper to handle the final sync of a file. Works just like the normal
 335  * io path, just does everything sync.
 336  */
 337 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 338 {
 339         struct io_u *io_u = __get_io_u(td);
 340         int ret;
 341
 342         if (!io_u)
 343                 return 1;
 344
 345         io_u->ddir = DDIR_SYNC;
 346         io_u->file = f;
 347
 348         if (td_io_prep(td, io_u)) {
 349                 put_io_u(td, io_u);
 350                 return 1;
 351         }
 352
 353 requeue:
 354         ret = td_io_queue(td, io_u);
 355         if (ret < 0) {
 356                 td_verror(td, io_u->error, "td_io_queue");
 357                 put_io_u(td, io_u);
 358                 return 1;
 359         } else if (ret == FIO_Q_QUEUED) {
 360                 if (io_u_queued_complete(td, 1, NULL) < 0)
 361                         return 1;
 362         } else if (ret == FIO_Q_COMPLETED) {
 363                 if (io_u->error) {
 364                         td_verror(td, io_u->error, "td_io_queue");
 365                         return 1;
 366                 }
 367
 368                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 369                         return 1;
 370         } else if (ret == FIO_Q_BUSY) {
 371                 if (td_io_commit(td))
 372                         return 1;
 373                 goto requeue;
 374         }
 375
 376         return 0;
 377 }
 378
 379 static inline void __update_tv_cache(struct thread_data *td)
 380 {
 381         fio_gettime(&td->tv_cache, NULL);
 382 }
 383
 384 static inline void update_tv_cache(struct thread_data *td)
 385 {
 386         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 387                 __update_tv_cache(td);
 388 }
 389
 390 static int break_on_this_error(struct thread_data *td, int *retptr)
 391 {
 392         int ret = *retptr;
 393
 394         if (ret < 0 || td->error) {
 395                 int err;
 396
 397                 if (!td->o.continue_on_error)
 398                         return 1;
 399
 400                 if (ret < 0)
 401                         err = -ret;
 402                 else
 403                         err = td->error;
 404
 405                 if (td_non_fatal_error(err)) {
 406                         /*
 407                          * Continue with the I/Os in case of
 408                          * a non fatal error.
 409                          */
 410                         update_error_count(td, err);
 411                         td_clear_error(td);
 412                         *retptr = 0;
 413                         return 0;
 414                 } else if (td->o.fill_device && err == ENOSPC) {
 415                         /*
 416                          * We expect to hit this error if
 417                          * fill_device option is set.
 418                          */
 419                         td_clear_error(td);
 420                         td->terminate = 1;
 421                         return 1;
 422                 } else {
 423                         /*
 424                          * Stop the I/O in case of a fatal
 425                          * error.
 426                          */
 427                         update_error_count(td, err);
 428                         return 1;
 429                 }
 430         }
 431
 432         return 0;
 433 }
 434
 435 /*
 436  * The main verify engine. Runs over the writes we previously submitted,
 437  * reads the blocks back in, and checks the crc/md5 of the data.
 438  */
 439 static void do_verify(struct thread_data *td)
 440 {
 441         struct fio_file *f;
 442         struct io_u *io_u;
 443         int ret, min_events;
 444         unsigned int i;
 445
 446         dprint(FD_VERIFY, "starting loop\n");
 447
 448         /*
 449          * sync io first and invalidate cache, to make sure we really
 450          * read from disk.
 451          */
 452         for_each_file(td, f, i) {
 453                 if (!fio_file_open(f))
 454                         continue;
 455                 if (fio_io_sync(td, f))
 456                         break;
 457                 if (file_invalidate_cache(td, f))
 458                         break;
 459         }
 460
 461         if (td->error)
 462                 return;
 463
 464         td_set_runstate(td, TD_VERIFYING);
 465
 466         io_u = NULL;
 467         while (!td->terminate) {
 468                 int ret2, full;
 469
 470                 update_tv_cache(td);
 471
 472                 if (runtime_exceeded(td, &td->tv_cache)) {
 473                         __update_tv_cache(td);
 474                         if (runtime_exceeded(td, &td->tv_cache)) {
 475                                 td->terminate = 1;
 476                                 break;
 477                         }
 478                 }
 479
 480                 io_u = __get_io_u(td);
 481                 if (!io_u)
 482                         break;
 483
 484                 if (get_next_verify(td, io_u)) {
 485                         put_io_u(td, io_u);
 486                         break;
 487                 }
 488
 489                 if (td_io_prep(td, io_u)) {
 490                         put_io_u(td, io_u);
 491                         break;
 492                 }
 493
 494                 if (td->o.verify_async)
 495                         io_u->end_io = verify_io_u_async;
 496                 else
 497                         io_u->end_io = verify_io_u;
 498
 499                 ret = td_io_queue(td, io_u);
 500                 switch (ret) {
 501                 case FIO_Q_COMPLETED:
 502                         if (io_u->error) {
 503                                 ret = -io_u->error;
 504                                 clear_io_u(td, io_u);
 505                         } else if (io_u->resid) {
 506                                 int bytes = io_u->xfer_buflen - io_u->resid;
 507
 508                                 /*
 509                                  * zero read, fail
 510                                  */
 511                                 if (!bytes) {
 512                                         td_verror(td, EIO, "full resid");
 513                                         put_io_u(td, io_u);
 514                                         break;
 515                                 }
 516
 517                                 io_u->xfer_buflen = io_u->resid;
 518                                 io_u->xfer_buf += bytes;
 519                                 io_u->offset += bytes;
 520
 521                                 if (ddir_rw(io_u->ddir))
 522                                         td->ts.short_io_u[io_u->ddir]++;
 523
 524                                 f = io_u->file;
 525                                 if (io_u->offset == f->real_file_size)
 526                                         goto sync_done;
 527
 528                                 requeue_io_u(td, &io_u);
 529                         } else {
 530 sync_done:
 531                                 ret = io_u_sync_complete(td, io_u, NULL);
 532                                 if (ret < 0)
 533                                         break;
 534                         }
 535                         continue;
 536                 case FIO_Q_QUEUED:
 537                         break;
 538                 case FIO_Q_BUSY:
 539                         requeue_io_u(td, &io_u);
 540                         ret2 = td_io_commit(td);
 541                         if (ret2 < 0)
 542                                 ret = ret2;
 543                         break;
 544                 default:
 545                         assert(ret < 0);
 546                         td_verror(td, -ret, "td_io_queue");
 547                         break;
 548                 }
 549
 550                 if (break_on_this_error(td, &ret))
 551                         break;
 552
 553                 /*
 554                  * if we can queue more, do so. but check if there are
 555                  * completed io_u's first. Note that we can get BUSY even
 556                  * without IO queued, if the system is resource starved.
 557                  */
 558                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 559                 if (full || !td->o.iodepth_batch_complete) {
 560                         min_events = min(td->o.iodepth_batch_complete,
 561                                          td->cur_depth);
 562                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 563                                 min_events = 1;
 564
 565                         do {
 566                                 /*
 567                                  * Reap required number of io units, if any,
 568                                  * and do the verification on them through
 569                                  * the callback handler
 570                                  */
 571                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 572                                         ret = -1;
 573                                         break;
 574                                 }
 575                         } while (full && (td->cur_depth > td->o.iodepth_low));
 576                 }
 577                 if (ret < 0)
 578                         break;
 579         }
 580
 581         if (!td->error) {
 582                 min_events = td->cur_depth;
 583
 584                 if (min_events)
 585                         ret = io_u_queued_complete(td, min_events, NULL);
 586         } else
 587                 cleanup_pending_aio(td);
 588
 589         td_set_runstate(td, TD_RUNNING);
 590
 591         dprint(FD_VERIFY, "exiting loop\n");
 592 }
 593
 594 /*
 595  * Main IO worker function. It retrieves io_u's to process and queues
 596  * and reaps them, checking for rate and errors along the way.
 597  */
 598 static void do_io(struct thread_data *td)
 599 {
 600         unsigned int i;
 601         int ret = 0;
 602
 603         if (in_ramp_time(td))
 604                 td_set_runstate(td, TD_RAMP);
 605         else
 606                 td_set_runstate(td, TD_RUNNING);
 607
 608         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 609                 (!flist_empty(&td->trim_list)) ||
 610                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 611                 struct timeval comp_time;
 612                 unsigned long bytes_done[2] = { 0, 0 };
 613                 int min_evts = 0;
 614                 struct io_u *io_u;
 615                 int ret2, full;
 616
 617                 if (td->terminate)
 618                         break;
 619
 620                 update_tv_cache(td);
 621
 622                 if (runtime_exceeded(td, &td->tv_cache)) {
 623                         __update_tv_cache(td);
 624                         if (runtime_exceeded(td, &td->tv_cache)) {
 625                                 td->terminate = 1;
 626                                 break;
 627                         }
 628                 }
 629
 630                 io_u = get_io_u(td);
 631                 if (!io_u)
 632                         break;
 633
 634                 /*
 635                  * Add verification end_io handler, if asked to verify
 636                  * a previously written file.
 637                  */
 638                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 639                     !td_rw(td)) {
 640                         if (td->o.verify_async)
 641                                 io_u->end_io = verify_io_u_async;
 642                         else
 643                                 io_u->end_io = verify_io_u;
 644                         td_set_runstate(td, TD_VERIFYING);
 645                 } else if (in_ramp_time(td))
 646                         td_set_runstate(td, TD_RAMP);
 647                 else
 648                         td_set_runstate(td, TD_RUNNING);
 649
 650                 ret = td_io_queue(td, io_u);
 651                 switch (ret) {
 652                 case FIO_Q_COMPLETED:
 653                         if (io_u->error) {
 654                                 ret = -io_u->error;
 655                                 clear_io_u(td, io_u);
 656                         } else if (io_u->resid) {
 657                                 int bytes = io_u->xfer_buflen - io_u->resid;
 658                                 struct fio_file *f = io_u->file;
 659
 660                                 /*
 661                                  * zero read, fail
 662                                  */
 663                                 if (!bytes) {
 664                                         td_verror(td, EIO, "full resid");
 665                                         put_io_u(td, io_u);
 666                                         break;
 667                                 }
 668
 669                                 io_u->xfer_buflen = io_u->resid;
 670                                 io_u->xfer_buf += bytes;
 671                                 io_u->offset += bytes;
 672
 673                                 if (ddir_rw(io_u->ddir))
 674                                         td->ts.short_io_u[io_u->ddir]++;
 675
 676                                 if (io_u->offset == f->real_file_size)
 677                                         goto sync_done;
 678
 679                                 requeue_io_u(td, &io_u);
 680                         } else {
 681 sync_done:
 682                                 if (__should_check_rate(td, 0) ||
 683                                     __should_check_rate(td, 1))
 684                                         fio_gettime(&comp_time, NULL);
 685
 686                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 687                                 if (ret < 0)
 688                                         break;
 689                         }
 690                         break;
 691                 case FIO_Q_QUEUED:
 692                         /*
 693                          * if the engine doesn't have a commit hook,
 694                          * the io_u is really queued. if it does have such
 695                          * a hook, it has to call io_u_queued() itself.
 696                          */
 697                         if (td->io_ops->commit == NULL)
 698                                 io_u_queued(td, io_u);
 699                         break;
 700                 case FIO_Q_BUSY:
 701                         requeue_io_u(td, &io_u);
 702                         ret2 = td_io_commit(td);
 703                         if (ret2 < 0)
 704                                 ret = ret2;
 705                         break;
 706                 default:
 707                         assert(ret < 0);
 708                         put_io_u(td, io_u);
 709                         break;
 710                 }
 711
 712                 if (break_on_this_error(td, &ret))
 713                         break;
 714
 715                 /*
 716                  * See if we need to complete some commands. Note that we
 717                  * can get BUSY even without IO queued, if the system is
 718                  * resource starved.
 719                  */
 720                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 721                 if (full || !td->o.iodepth_batch_complete) {
 722                         min_evts = min(td->o.iodepth_batch_complete,
 723                                         td->cur_depth);
 724                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 725                                 min_evts = 1;
 726
 727                         if (__should_check_rate(td, 0) ||
 728                             __should_check_rate(td, 1))
 729                                 fio_gettime(&comp_time, NULL);
 730
 731                         do {
 732                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 733                                 if (ret < 0)
 734                                         break;
 735
 736                         } while (full && (td->cur_depth > td->o.iodepth_low));
 737                 }
 738
 739                 if (ret < 0)
 740                         break;
 741                 if (!(bytes_done[0] + bytes_done[1]))
 742                         continue;
 743
 744                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 745                         if (check_min_rate(td, &comp_time, bytes_done)) {
 746                                 if (exitall_on_terminate)
 747                                         terminate_threads(td->groupid);
 748                                 td_verror(td, EIO, "check_min_rate");
 749                                 break;
 750                         }
 751                 }
 752
 753                 if (td->o.thinktime) {
 754                         unsigned long long b;
 755
 756                         b = td->io_blocks[0] + td->io_blocks[1];
 757                         if (!(b % td->o.thinktime_blocks)) {
 758                                 int left;
 759
 760                                 if (td->o.thinktime_spin)
 761                                         usec_spin(td->o.thinktime_spin);
 762
 763                                 left = td->o.thinktime - td->o.thinktime_spin;
 764                                 if (left)
 765                                         usec_sleep(td, left);
 766                         }
 767                 }
 768         }
 769
 770         if (td->trim_entries)
 771                 printf("trim entries %ld\n", td->trim_entries);
 772
 773         if (td->o.fill_device && td->error == ENOSPC) {
 774                 td->error = 0;
 775                 td->terminate = 1;
 776         }
 777         if (!td->error) {
 778                 struct fio_file *f;
 779
 780                 i = td->cur_depth;
 781                 if (i) {
 782                         ret = io_u_queued_complete(td, i, NULL);
 783                         if (td->o.fill_device && td->error == ENOSPC)
 784                                 td->error = 0;
 785                 }
 786
 787                 if (should_fsync(td) && td->o.end_fsync) {
 788                         td_set_runstate(td, TD_FSYNCING);
 789
 790                         for_each_file(td, f, i) {
 791                                 if (!fio_file_open(f))
 792                                         continue;
 793                                 fio_io_sync(td, f);
 794                         }
 795                 }
 796         } else
 797                 cleanup_pending_aio(td);
 798
 799         /*
 800          * stop job if we failed doing any IO
 801          */
 802         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 803                 td->done = 1;
 804 }
 805
 806 static void cleanup_io_u(struct thread_data *td)
 807 {
 808         struct flist_head *entry, *n;
 809         struct io_u *io_u;
 810
 811         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 812                 io_u = flist_entry(entry, struct io_u, list);
 813
 814                 flist_del(&io_u->list);
 815                 fio_memfree(io_u, sizeof(*io_u));
 816         }
 817
 818         free_io_mem(td);
 819 }
 820
 821 static int init_io_u(struct thread_data *td)
 822 {
 823         struct io_u *io_u;
 824         unsigned int max_bs;
 825         int cl_align, i, max_units;
 826         char *p;
 827
 828         max_units = td->o.iodepth;
 829         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 830         td->orig_buffer_size = (unsigned long long) max_bs
 831                                         * (unsigned long long) max_units;
 832
 833         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 834                 unsigned long bs;
 835
 836                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 837                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 838         }
 839
 840         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 841                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 842                 return 1;
 843         }
 844
 845         if (allocate_io_mem(td))
 846                 return 1;
 847
 848         if (td->o.odirect || td->o.mem_align ||
 849             (td->io_ops->flags & FIO_RAWIO))
 850                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 851         else
 852                 p = td->orig_buffer;
 853
 854         cl_align = os_cache_line_size();
 855
 856         for (i = 0; i < max_units; i++) {
 857                 void *ptr;
 858
 859                 if (td->terminate)
 860                         return 1;
 861
 862                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 863                 if (!ptr) {
 864                         log_err("fio: unable to allocate aligned memory\n");
 865                         break;
 866                 }
 867
 868                 io_u = ptr;
 869                 memset(io_u, 0, sizeof(*io_u));
 870                 INIT_FLIST_HEAD(&io_u->list);
 871                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 872
 873                 if (!(td->io_ops->flags & FIO_NOIO)) {
 874                         io_u->buf = p;
 875                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 876
 877                         if (td_write(td))
 878                                 io_u_fill_buffer(td, io_u, max_bs);
 879                         if (td_write(td) && td->o.verify_pattern_bytes) {
 880                                 /*
 881                                  * Fill the buffer with the pattern if we are
 882                                  * going to be doing writes.
 883                                  */
 884                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 885                         }
 886                 }
 887
 888                 io_u->index = i;
 889                 io_u->flags = IO_U_F_FREE;
 890                 flist_add(&io_u->list, &td->io_u_freelist);
 891                 p += max_bs;
 892         }
 893
 894         return 0;
 895 }
 896
 897 static int switch_ioscheduler(struct thread_data *td)
 898 {
 899         char tmp[256], tmp2[128];
 900         FILE *f;
 901         int ret;
 902
 903         if (td->io_ops->flags & FIO_DISKLESSIO)
 904                 return 0;
 905
 906         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 907
 908         f = fopen(tmp, "r+");
 909         if (!f) {
 910                 if (errno == ENOENT) {
 911                         log_err("fio: os or kernel doesn't support IO scheduler"
 912                                 " switching\n");
 913                         return 0;
 914                 }
 915                 td_verror(td, errno, "fopen iosched");
 916                 return 1;
 917         }
 918
 919         /*
 920          * Set io scheduler.
 921          */
 922         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 923         if (ferror(f) || ret != 1) {
 924                 td_verror(td, errno, "fwrite");
 925                 fclose(f);
 926                 return 1;
 927         }
 928
 929         rewind(f);
 930
 931         /*
 932          * Read back and check that the selected scheduler is now the default.
 933          */
 934         ret = fread(tmp, 1, sizeof(tmp), f);
 935         if (ferror(f) || ret < 0) {
 936                 td_verror(td, errno, "fread");
 937                 fclose(f);
 938                 return 1;
 939         }
 940
 941         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 942         if (!strstr(tmp, tmp2)) {
 943                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 944                 td_verror(td, EINVAL, "iosched_switch");
 945                 fclose(f);
 946                 return 1;
 947         }
 948
 949         fclose(f);
 950         return 0;
 951 }
 952
 953 static int keep_running(struct thread_data *td)
 954 {
 955         unsigned long long io_done;
 956
 957         if (td->done)
 958                 return 0;
 959         if (td->o.time_based)
 960                 return 1;
 961         if (td->o.loops) {
 962                 td->o.loops--;
 963                 return 1;
 964         }
 965
 966         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 967                         + td->io_skip_bytes;
 968         if (io_done < td->o.size)
 969                 return 1;
 970
 971         return 0;
 972 }
 973
 974 static void reset_io_counters(struct thread_data *td)
 975 {
 976         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 977         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 978         td->zone_bytes = 0;
 979         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 980         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 981
 982         td->last_was_sync = 0;
 983
 984         /*
 985          * reset file done count if we are to start over
 986          */
 987         if (td->o.time_based || td->o.loops)
 988                 td->nr_done_files = 0;
 989 }
 990
 991 void reset_all_stats(struct thread_data *td)
 992 {
 993         struct timeval tv;
 994         int i;
 995
 996         reset_io_counters(td);
 997
 998         for (i = 0; i < 2; i++) {
 999                 td->io_bytes[i] = 0;
1000                 td->io_blocks[i] = 0;
1001                 td->io_issues[i] = 0;
1002                 td->ts.total_io_u[i] = 0;
1003         }
1004
1005         fio_gettime(&tv, NULL);
1006         td->ts.runtime[0] = 0;
1007         td->ts.runtime[1] = 0;
1008         memcpy(&td->epoch, &tv, sizeof(tv));
1009         memcpy(&td->start, &tv, sizeof(tv));
1010 }
1011
1012 static void clear_io_state(struct thread_data *td)
1013 {
1014         struct fio_file *f;
1015         unsigned int i;
1016
1017         reset_io_counters(td);
1018
1019         close_files(td);
1020         for_each_file(td, f, i)
1021                 fio_file_clear_done(f);
1022
1023         /*
1024          * Set the same seed to get repeatable runs
1025          */
1026         td_fill_rand_seeds(td);
1027 }
1028
1029 static int exec_string(const char *string)
1030 {
1031         int ret, newlen = strlen(string) + 1 + 8;
1032         char *str;
1033
1034         str = malloc(newlen);
1035         sprintf(str, "sh -c %s", string);
1036
1037         ret = system(str);
1038         if (ret == -1)
1039                 log_err("fio: exec of cmd <%s> failed\n", str);
1040
1041         free(str);
1042         return ret;
1043 }
1044
1045 /*
1046  * Entry point for the thread based jobs. The process based jobs end up
1047  * here as well, after a little setup.
1048  */
1049 static void *thread_main(void *data)
1050 {
1051         unsigned long long elapsed;
1052         struct thread_data *td = data;
1053         pthread_condattr_t attr;
1054         int clear_state;
1055
1056         if (!td->o.use_thread) {
1057                 setsid();
1058                 td->pid = getpid();
1059         } else
1060                 td->pid = gettid();
1061
1062         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1063
1064         INIT_FLIST_HEAD(&td->io_u_freelist);
1065         INIT_FLIST_HEAD(&td->io_u_busylist);
1066         INIT_FLIST_HEAD(&td->io_u_requeues);
1067         INIT_FLIST_HEAD(&td->io_log_list);
1068         INIT_FLIST_HEAD(&td->io_hist_list);
1069         INIT_FLIST_HEAD(&td->verify_list);
1070         INIT_FLIST_HEAD(&td->trim_list);
1071         pthread_mutex_init(&td->io_u_lock, NULL);
1072         td->io_hist_tree = RB_ROOT;
1073
1074         pthread_condattr_init(&attr);
1075         pthread_cond_init(&td->verify_cond, &attr);
1076         pthread_cond_init(&td->free_cond, &attr);
1077
1078         td_set_runstate(td, TD_INITIALIZED);
1079         dprint(FD_MUTEX, "up startup_mutex\n");
1080         fio_mutex_up(startup_mutex);
1081         dprint(FD_MUTEX, "wait on td->mutex\n");
1082         fio_mutex_down(td->mutex);
1083         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1084
1085         /*
1086          * the ->mutex mutex is now no longer used, close it to avoid
1087          * eating a file descriptor
1088          */
1089         fio_mutex_remove(td->mutex);
1090
1091         /*
1092          * A new gid requires privilege, so we need to do this before setting
1093          * the uid.
1094          */
1095         if (td->o.gid != -1U && setgid(td->o.gid)) {
1096                 td_verror(td, errno, "setgid");
1097                 goto err;
1098         }
1099         if (td->o.uid != -1U && setuid(td->o.uid)) {
1100                 td_verror(td, errno, "setuid");
1101                 goto err;
1102         }
1103
1104         /*
1105          * If we have a gettimeofday() thread, make sure we exclude that
1106          * thread from this job
1107          */
1108         if (td->o.gtod_cpu)
1109                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1110
1111         /*
1112          * Set affinity first, in case it has an impact on the memory
1113          * allocations.
1114          */
1115         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1116                 td_verror(td, errno, "cpu_set_affinity");
1117                 goto err;
1118         }
1119
1120         /*
1121          * May alter parameters that init_io_u() will use, so we need to
1122          * do this first.
1123          */
1124         if (init_iolog(td))
1125                 goto err;
1126
1127         if (init_io_u(td))
1128                 goto err;
1129
1130         if (td->o.verify_async && verify_async_init(td))
1131                 goto err;
1132
1133         if (td->ioprio_set) {
1134                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1135                         td_verror(td, errno, "ioprio_set");
1136                         goto err;
1137                 }
1138         }
1139
1140         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1141                 goto err;
1142
1143         if (nice(td->o.nice) == -1) {
1144                 td_verror(td, errno, "nice");
1145                 goto err;
1146         }
1147
1148         if (td->o.ioscheduler && switch_ioscheduler(td))
1149                 goto err;
1150
1151         if (!td->o.create_serialize && setup_files(td))
1152                 goto err;
1153
1154         if (td_io_init(td))
1155                 goto err;
1156
1157         if (init_random_map(td))
1158                 goto err;
1159
1160         if (td->o.exec_prerun) {
1161                 if (exec_string(td->o.exec_prerun))
1162                         goto err;
1163         }
1164
1165         if (td->o.pre_read) {
1166                 if (pre_read_files(td) < 0)
1167                         goto err;
1168         }
1169
1170         fio_gettime(&td->epoch, NULL);
1171         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1172
1173         clear_state = 0;
1174         while (keep_running(td)) {
1175                 fio_gettime(&td->start, NULL);
1176                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1177                                 sizeof(td->start));
1178                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1179                                 sizeof(td->start));
1180                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1181
1182                 if (td->o.ratemin[0] || td->o.ratemin[1])
1183                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1184                                                         sizeof(td->lastrate));
1185
1186                 if (clear_state)
1187                         clear_io_state(td);
1188
1189                 prune_io_piece_log(td);
1190
1191                 do_io(td);
1192
1193                 clear_state = 1;
1194
1195                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1196                         elapsed = utime_since_now(&td->start);
1197                         td->ts.runtime[DDIR_READ] += elapsed;
1198                 }
1199                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1200                         elapsed = utime_since_now(&td->start);
1201                         td->ts.runtime[DDIR_WRITE] += elapsed;
1202                 }
1203
1204                 if (td->error || td->terminate)
1205                         break;
1206
1207                 if (!td->o.do_verify ||
1208                     td->o.verify == VERIFY_NONE ||
1209                     (td->io_ops->flags & FIO_UNIDIR))
1210                         continue;
1211
1212                 clear_io_state(td);
1213
1214                 fio_gettime(&td->start, NULL);
1215
1216                 do_verify(td);
1217
1218                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1219
1220                 if (td->error || td->terminate)
1221                         break;
1222         }
1223
1224         update_rusage_stat(td);
1225         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1226         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1227         td->ts.total_run_time = mtime_since_now(&td->epoch);
1228         td->ts.io_bytes[0] = td->io_bytes[0];
1229         td->ts.io_bytes[1] = td->io_bytes[1];
1230
1231         fio_mutex_down(writeout_mutex);
1232         if (td->ts.bw_log) {
1233                 if (td->o.bw_log_file) {
1234                         finish_log_named(td, td->ts.bw_log,
1235                                                 td->o.bw_log_file, "bw");
1236                 } else
1237                         finish_log(td, td->ts.bw_log, "bw");
1238         }
1239         if (td->ts.lat_log) {
1240                 if (td->o.lat_log_file) {
1241                         finish_log_named(td, td->ts.lat_log,
1242                                                 td->o.lat_log_file, "lat");
1243                 } else
1244                         finish_log(td, td->ts.lat_log, "lat");
1245         }
1246         if (td->ts.slat_log) {
1247                 if (td->o.lat_log_file) {
1248                         finish_log_named(td, td->ts.slat_log,
1249                                                 td->o.lat_log_file, "slat");
1250                 } else
1251                         finish_log(td, td->ts.slat_log, "slat");
1252         }
1253         if (td->ts.clat_log) {
1254                 if (td->o.lat_log_file) {
1255                         finish_log_named(td, td->ts.clat_log,
1256                                                 td->o.lat_log_file, "clat");
1257                 } else
1258                         finish_log(td, td->ts.clat_log, "clat");
1259         }
1260         fio_mutex_up(writeout_mutex);
1261         if (td->o.exec_postrun)
1262                 exec_string(td->o.exec_postrun);
1263
1264         if (exitall_on_terminate)
1265                 terminate_threads(td->groupid);
1266
1267 err:
1268         if (td->error)
1269                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1270                                                         td->verror);
1271
1272         if (td->o.verify_async)
1273                 verify_async_exit(td);
1274
1275         close_and_free_files(td);
1276         close_ioengine(td);
1277         cleanup_io_u(td);
1278         cgroup_shutdown(td, &cgroup_mnt);
1279
1280         if (td->o.cpumask_set) {
1281                 int ret = fio_cpuset_exit(&td->o.cpumask);
1282
1283                 td_verror(td, ret, "fio_cpuset_exit");
1284         }
1285
1286         /*
1287          * do this very late, it will log file closing as well
1288          */
1289         if (td->o.write_iolog_file)
1290                 write_iolog_close(td);
1291
1292         options_mem_free(td);
1293         td_set_runstate(td, TD_EXITED);
1294         return (void *) (unsigned long) td->error;
1295 }
1296
1297 /*
1298  * We cannot pass the td data into a forked process, so attach the td and
1299  * pass it to the thread worker.
1300  */
1301 static int fork_main(int shmid, int offset)
1302 {
1303         struct thread_data *td;
1304         void *data, *ret;
1305
1306 #ifndef __hpux
1307         data = shmat(shmid, NULL, 0);
1308         if (data == (void *) -1) {
1309                 int __err = errno;
1310
1311                 perror("shmat");
1312                 return __err;
1313         }
1314 #else
1315         /*
1316          * HP-UX inherits shm mappings?
1317          */
1318         data = threads;
1319 #endif
1320
1321         td = data + offset * sizeof(struct thread_data);
1322         ret = thread_main(td);
1323         shmdt(data);
1324         return (int) (unsigned long) ret;
1325 }
1326
1327 /*
1328  * Run over the job map and reap the threads that have exited, if any.
1329  */
1330 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1331 {
1332         struct thread_data *td;
1333         int i, cputhreads, realthreads, pending, status, ret;
1334
1335         /*
1336          * reap exited threads (TD_EXITED -> TD_REAPED)
1337          */
1338         realthreads = pending = cputhreads = 0;
1339         for_each_td(td, i) {
1340                 int flags = 0;
1341
1342                 /*
1343                  * ->io_ops is NULL for a thread that has closed its
1344                  * io engine
1345                  */
1346                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1347                         cputhreads++;
1348                 else
1349                         realthreads++;
1350
1351                 if (!td->pid) {
1352                         pending++;
1353                         continue;
1354                 }
1355                 if (td->runstate == TD_REAPED)
1356                         continue;
1357                 if (td->o.use_thread) {
1358                         if (td->runstate == TD_EXITED) {
1359                                 td_set_runstate(td, TD_REAPED);
1360                                 goto reaped;
1361                         }
1362                         continue;
1363                 }
1364
1365                 flags = WNOHANG;
1366                 if (td->runstate == TD_EXITED)
1367                         flags = 0;
1368
1369                 /*
1370                  * check if someone quit or got killed in an unusual way
1371                  */
1372                 ret = waitpid(td->pid, &status, flags);
1373                 if (ret < 0) {
1374                         if (errno == ECHILD) {
1375                                 log_err("fio: pid=%d disappeared %d\n",
1376                                                 (int) td->pid, td->runstate);
1377                                 td_set_runstate(td, TD_REAPED);
1378                                 goto reaped;
1379                         }
1380                         perror("waitpid");
1381                 } else if (ret == td->pid) {
1382                         if (WIFSIGNALED(status)) {
1383                                 int sig = WTERMSIG(status);
1384
1385                                 if (sig != SIGTERM)
1386                                         log_err("fio: pid=%d, got signal=%d\n",
1387                                                         (int) td->pid, sig);
1388                                 td_set_runstate(td, TD_REAPED);
1389                                 goto reaped;
1390                         }
1391                         if (WIFEXITED(status)) {
1392                                 if (WEXITSTATUS(status) && !td->error)
1393                                         td->error = WEXITSTATUS(status);
1394
1395                                 td_set_runstate(td, TD_REAPED);
1396                                 goto reaped;
1397                         }
1398                 }
1399
1400                 /*
1401                  * thread is not dead, continue
1402                  */
1403                 pending++;
1404                 continue;
1405 reaped:
1406                 (*nr_running)--;
1407                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1408                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1409                 if (!td->pid)
1410                         pending--;
1411
1412                 if (td->error)
1413                         exit_value++;
1414
1415                 done_secs += mtime_since_now(&td->epoch) / 1000;
1416         }
1417
1418         if (*nr_running == cputhreads && !pending && realthreads)
1419                 terminate_threads(TERMINATE_ALL);
1420 }
1421
1422 static void *gtod_thread_main(void *data)
1423 {
1424         fio_mutex_up(startup_mutex);
1425
1426         /*
1427          * As long as we have jobs around, update the clock. It would be nice
1428          * to have some way of NOT hammering that CPU with gettimeofday(),
1429          * but I'm not sure what to use outside of a simple CPU nop to relax
1430          * it - we don't want to lose precision.
1431          */
1432         while (threads) {
1433                 fio_gtod_update();
1434                 nop;
1435         }
1436
1437         return NULL;
1438 }
1439
1440 static int fio_start_gtod_thread(void)
1441 {
1442         pthread_attr_t attr;
1443         int ret;
1444
1445         pthread_attr_init(&attr);
1446         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1447         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1448         pthread_attr_destroy(&attr);
1449         if (ret) {
1450                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1451                 return 1;
1452         }
1453
1454         ret = pthread_detach(gtod_thread);
1455         if (ret) {
1456                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1457                 return 1;
1458         }
1459
1460         dprint(FD_MUTEX, "wait on startup_mutex\n");
1461         fio_mutex_down(startup_mutex);
1462         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1463         return 0;
1464 }
1465
1466 /*
1467  * Main function for kicking off and reaping jobs, as needed.
1468  */
1469 static void run_threads(void)
1470 {
1471         struct thread_data *td;
1472         unsigned long spent;
1473         int i, todo, nr_running, m_rate, t_rate, nr_started;
1474
1475         if (fio_pin_memory())
1476                 return;
1477
1478         if (fio_gtod_offload && fio_start_gtod_thread())
1479                 return;
1480
1481         if (!terse_output) {
1482                 log_info("Starting ");
1483                 if (nr_thread)
1484                         log_info("%d thread%s", nr_thread,
1485                                                 nr_thread > 1 ? "s" : "");
1486                 if (nr_process) {
1487                         if (nr_thread)
1488                                 printf(" and ");
1489                         log_info("%d process%s", nr_process,
1490                                                 nr_process > 1 ? "es" : "");
1491                 }
1492                 log_info("\n");
1493                 fflush(stdout);
1494         }
1495
1496         set_sig_handlers();
1497
1498         todo = thread_number;
1499         nr_running = 0;
1500         nr_started = 0;
1501         m_rate = t_rate = 0;
1502
1503         for_each_td(td, i) {
1504                 print_status_init(td->thread_number - 1);
1505
1506                 if (!td->o.create_serialize)
1507                         continue;
1508
1509                 /*
1510                  * do file setup here so it happens sequentially,
1511                  * we don't want X number of threads getting their
1512                  * client data interspersed on disk
1513                  */
1514                 if (setup_files(td)) {
1515                         exit_value++;
1516                         if (td->error)
1517                                 log_err("fio: pid=%d, err=%d/%s\n",
1518                                         (int) td->pid, td->error, td->verror);
1519                         td_set_runstate(td, TD_REAPED);
1520                         todo--;
1521                 } else {
1522                         struct fio_file *f;
1523                         unsigned int j;
1524
1525                         /*
1526                          * for sharing to work, each job must always open
1527                          * its own files. so close them, if we opened them
1528                          * for creation
1529                          */
1530                         for_each_file(td, f, j) {
1531                                 if (fio_file_open(f))
1532                                         td_io_close_file(td, f);
1533                         }
1534                 }
1535         }
1536
1537         set_genesis_time();
1538
1539         while (todo) {
1540                 struct thread_data *map[REAL_MAX_JOBS];
1541                 struct timeval this_start;
1542                 int this_jobs = 0, left;
1543
1544                 /*
1545                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1546                  */
1547                 for_each_td(td, i) {
1548                         if (td->runstate != TD_NOT_CREATED)
1549                                 continue;
1550
1551                         /*
1552                          * never got a chance to start, killed by other
1553                          * thread for some reason
1554                          */
1555                         if (td->terminate) {
1556                                 todo--;
1557                                 continue;
1558                         }
1559
1560                         if (td->o.start_delay) {
1561                                 spent = mtime_since_genesis();
1562
1563                                 if (td->o.start_delay * 1000 > spent)
1564                                         continue;
1565                         }
1566
1567                         if (td->o.stonewall && (nr_started || nr_running)) {
1568                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1569                                                         td->o.name);
1570                                 break;
1571                         }
1572
1573                         init_disk_util(td);
1574
1575                         /*
1576                          * Set state to created. Thread will transition
1577                          * to TD_INITIALIZED when it's done setting up.
1578                          */
1579                         td_set_runstate(td, TD_CREATED);
1580                         map[this_jobs++] = td;
1581                         nr_started++;
1582
1583                         if (td->o.use_thread) {
1584                                 int ret;
1585
1586                                 dprint(FD_PROCESS, "will pthread_create\n");
1587                                 ret = pthread_create(&td->thread, NULL,
1588                                                         thread_main, td);
1589                                 if (ret) {
1590                                         log_err("pthread_create: %s\n",
1591                                                         strerror(ret));
1592                                         nr_started--;
1593                                         break;
1594                                 }
1595                                 ret = pthread_detach(td->thread);
1596                                 if (ret)
1597                                         log_err("pthread_detach: %s",
1598                                                         strerror(ret));
1599                         } else {
1600                                 pid_t pid;
1601                                 dprint(FD_PROCESS, "will fork\n");
1602                                 pid = fork();
1603                                 if (!pid) {
1604                                         int ret = fork_main(shm_id, i);
1605
1606                                         _exit(ret);
1607                                 } else if (i == fio_debug_jobno)
1608                                         *fio_debug_jobp = pid;
1609                         }
1610                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1611                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1612                                 log_err("fio: job startup hung? exiting.\n");
1613                                 terminate_threads(TERMINATE_ALL);
1614                                 fio_abort = 1;
1615                                 nr_started--;
1616                                 break;
1617                         }
1618                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1619                 }
1620
1621                 /*
1622                  * Wait for the started threads to transition to
1623                  * TD_INITIALIZED.
1624                  */
1625                 fio_gettime(&this_start, NULL);
1626                 left = this_jobs;
1627                 while (left && !fio_abort) {
1628                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1629                                 break;
1630
1631                         usleep(100000);
1632
1633                         for (i = 0; i < this_jobs; i++) {
1634                                 td = map[i];
1635                                 if (!td)
1636                                         continue;
1637                                 if (td->runstate == TD_INITIALIZED) {
1638                                         map[i] = NULL;
1639                                         left--;
1640                                 } else if (td->runstate >= TD_EXITED) {
1641                                         map[i] = NULL;
1642                                         left--;
1643                                         todo--;
1644                                         nr_running++; /* work-around... */
1645                                 }
1646                         }
1647                 }
1648
1649                 if (left) {
1650                         log_err("fio: %d jobs failed to start\n", left);
1651                         for (i = 0; i < this_jobs; i++) {
1652                                 td = map[i];
1653                                 if (!td)
1654                                         continue;
1655                                 kill(td->pid, SIGTERM);
1656                         }
1657                         break;
1658                 }
1659
1660                 /*
1661                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1662                  */
1663                 for_each_td(td, i) {
1664                         if (td->runstate != TD_INITIALIZED)
1665                                 continue;
1666
1667                         if (in_ramp_time(td))
1668                                 td_set_runstate(td, TD_RAMP);
1669                         else
1670                                 td_set_runstate(td, TD_RUNNING);
1671                         nr_running++;
1672                         nr_started--;
1673                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1674                         t_rate += td->o.rate[0] + td->o.rate[1];
1675                         todo--;
1676                         fio_mutex_up(td->mutex);
1677                 }
1678
1679                 reap_threads(&nr_running, &t_rate, &m_rate);
1680
1681                 if (todo)
1682                         usleep(100000);
1683         }
1684
1685         while (nr_running) {
1686                 reap_threads(&nr_running, &t_rate, &m_rate);
1687                 usleep(10000);
1688         }
1689
1690         update_io_ticks();
1691         fio_unpin_memory();
1692 }
1693
1694 int main(int argc, char *argv[], char *envp[])
1695 {
1696         long ps;
1697
1698         arch_init(envp);
1699
1700         sinit();
1701
1702         /*
1703          * We need locale for number printing, if it isn't set then just
1704          * go with the US format.
1705          */
1706         if (!getenv("LC_NUMERIC"))
1707                 setlocale(LC_NUMERIC, "en_US");
1708
1709         ps = sysconf(_SC_PAGESIZE);
1710         if (ps < 0) {
1711                 log_err("Failed to get page size\n");
1712                 return 1;
1713         }
1714
1715         page_size = ps;
1716         page_mask = ps - 1;
1717
1718         fio_keywords_init();
1719
1720         if (parse_options(argc, argv))
1721                 return 1;
1722
1723         if (exec_profile && load_profile(exec_profile))
1724                 return 1;
1725
1726         if (!thread_number)
1727                 return 0;
1728
1729         if (write_bw_log) {
1730                 setup_log(&agg_io_log[DDIR_READ]);
1731                 setup_log(&agg_io_log[DDIR_WRITE]);
1732         }
1733
1734         startup_mutex = fio_mutex_init(0);
1735         if (startup_mutex == NULL)
1736                 return 1;
1737         writeout_mutex = fio_mutex_init(1);
1738         if (writeout_mutex == NULL)
1739                 return 1;
1740
1741         set_genesis_time();
1742         create_disk_util_thread();
1743
1744         cgroup_list = smalloc(sizeof(*cgroup_list));
1745         INIT_FLIST_HEAD(cgroup_list);
1746
1747         run_threads();
1748
1749         if (!fio_abort) {
1750                 show_run_stats();
1751                 if (write_bw_log) {
1752                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1753                         __finish_log(agg_io_log[DDIR_WRITE],
1754                                         "agg-write_bw.log");
1755                 }
1756         }
1757
1758         cgroup_kill(cgroup_list);
1759         sfree(cgroup_list);
1760         sfree(cgroup_mnt);
1761
1762         fio_mutex_remove(startup_mutex);
1763         fio_mutex_remove(writeout_mutex);
1764         return exit_value;
1765 }