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 JOB_START_TIMEOUT       (5 * 1000)
  79
  80 static const char *fio_os_strings[os_nr] = {
  81         "Invalid",
  82         "Linux",
  83         "AIX",
  84         "FreeBSD",
  85         "HP-UX",
  86         "OSX",
  87         "NetBSD",
  88         "Solaris",
  89         "Windows"
  90 };
  91
  92 static const char *fio_arch_strings[arch_nr] = {
  93         "Invalid",
  94         "x86-64",
  95         "x86",
  96         "ppc",
  97         "ia64",
  98         "s390",
  99         "alpha",
 100         "sparc",
 101         "sparc64",
 102         "arm",
 103         "sh",
 104         "hppa",
 105         "generic"
 106 };
 107
 108 const char *fio_get_os_string(int nr)
 109 {
 110         if (nr < os_nr)
 111                 return fio_os_strings[nr];
 112
 113         return NULL;
 114 }
 115
 116 const char *fio_get_arch_string(int nr)
 117 {
 118         if (nr < arch_nr)
 119                 return fio_arch_strings[nr];
 120
 121         return NULL;
 122 }
 123
 124 void td_set_runstate(struct thread_data *td, int runstate)
 125 {
 126         if (td->runstate == runstate)
 127                 return;
 128
 129         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
 130                                                 td->runstate, runstate);
 131         td->runstate = runstate;
 132 }
 133
 134 void fio_terminate_threads(int group_id)
 135 {
 136         struct thread_data *td;
 137         int i;
 138
 139         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
 140
 141         for_each_td(td, i) {
 142                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
 143                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
 144                                                 td->o.name, (int) td->pid);
 145                         td->terminate = 1;
 146                         td->o.start_delay = 0;
 147
 148                         /*
 149                          * if the thread is running, just let it exit
 150                          */
 151                         if (!td->pid)
 152                                 continue;
 153                         else if (td->runstate < TD_RAMP)
 154                                 kill(td->pid, SIGTERM);
 155                         else {
 156                                 struct ioengine_ops *ops = td->io_ops;
 157
 158                                 if (ops && (ops->flags & FIO_SIGTERM))
 159                                         kill(td->pid, SIGTERM);
 160                         }
 161                 }
 162         }
 163 }
 164
 165 static void sig_int(int sig)
 166 {
 167         if (threads) {
 168                 log_info("\nfio: terminating on signal %d\n", sig);
 169                 exit_backend = 1;
 170                 fflush(stdout);
 171                 exit_value = 128;
 172                 fio_terminate_threads(TERMINATE_ALL);
 173         }
 174 }
 175
 176 static void *disk_thread_main(void *data)
 177 {
 178         fio_mutex_up(startup_mutex);
 179
 180         while (threads) {
 181                 usleep(DISK_UTIL_MSEC * 1000);
 182                 if (!threads)
 183                         break;
 184                 update_io_ticks();
 185
 186                 if (!is_backend)
 187                         print_thread_status();
 188         }
 189
 190         return NULL;
 191 }
 192
 193 static int create_disk_util_thread(void)
 194 {
 195         int ret;
 196
 197         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 198         if (ret) {
 199                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 200                 return 1;
 201         }
 202
 203         ret = pthread_detach(disk_util_thread);
 204         if (ret) {
 205                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 206                 return 1;
 207         }
 208
 209         dprint(FD_MUTEX, "wait on startup_mutex\n");
 210         fio_mutex_down(startup_mutex);
 211         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 212         return 0;
 213 }
 214
 215 static void set_sig_handlers(void)
 216 {
 217         struct sigaction act;
 218
 219         memset(&act, 0, sizeof(act));
 220         act.sa_handler = sig_int;
 221         act.sa_flags = SA_RESTART;
 222         sigaction(SIGINT, &act, NULL);
 223
 224         memset(&act, 0, sizeof(act));
 225         act.sa_handler = sig_int;
 226         act.sa_flags = SA_RESTART;
 227         sigaction(SIGTERM, &act, NULL);
 228
 229         if (is_backend) {
 230                 memset(&act, 0, sizeof(act));
 231                 act.sa_handler = sig_int;
 232                 act.sa_flags = SA_RESTART;
 233                 sigaction(SIGPIPE, &act, NULL);
 234         }
 235 }
 236
 237 /*
 238  * Check if we are above the minimum rate given.
 239  */
 240 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 241                             enum fio_ddir ddir)
 242 {
 243         unsigned long long bytes = 0;
 244         unsigned long iops = 0;
 245         unsigned long spent;
 246         unsigned long rate;
 247         unsigned int ratemin = 0;
 248         unsigned int rate_iops = 0;
 249         unsigned int rate_iops_min = 0;
 250
 251         assert(ddir_rw(ddir));
 252
 253         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 254                 return 0;
 255
 256         /*
 257          * allow a 2 second settle period in the beginning
 258          */
 259         if (mtime_since(&td->start, now) < 2000)
 260                 return 0;
 261
 262         iops += td->this_io_blocks[ddir];
 263         bytes += td->this_io_bytes[ddir];
 264         ratemin += td->o.ratemin[ddir];
 265         rate_iops += td->o.rate_iops[ddir];
 266         rate_iops_min += td->o.rate_iops_min[ddir];
 267
 268         /*
 269          * if rate blocks is set, sample is running
 270          */
 271         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 272                 spent = mtime_since(&td->lastrate[ddir], now);
 273                 if (spent < td->o.ratecycle)
 274                         return 0;
 275
 276                 if (td->o.rate[ddir]) {
 277                         /*
 278                          * check bandwidth specified rate
 279                          */
 280                         if (bytes < td->rate_bytes[ddir]) {
 281                                 log_err("%s: min rate %u not met\n", td->o.name,
 282                                                                 ratemin);
 283                                 return 1;
 284                         } else {
 285                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 286                                 if (rate < ratemin ||
 287                                     bytes < td->rate_bytes[ddir]) {
 288                                         log_err("%s: min rate %u not met, got"
 289                                                 " %luKB/sec\n", td->o.name,
 290                                                         ratemin, rate);
 291                                         return 1;
 292                                 }
 293                         }
 294                 } else {
 295                         /*
 296                          * checks iops specified rate
 297                          */
 298                         if (iops < rate_iops) {
 299                                 log_err("%s: min iops rate %u not met\n",
 300                                                 td->o.name, rate_iops);
 301                                 return 1;
 302                         } else {
 303                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 304                                 if (rate < rate_iops_min ||
 305                                     iops < td->rate_blocks[ddir]) {
 306                                         log_err("%s: min iops rate %u not met,"
 307                                                 " got %lu\n", td->o.name,
 308                                                         rate_iops_min, rate);
 309                                 }
 310                         }
 311                 }
 312         }
 313
 314         td->rate_bytes[ddir] = bytes;
 315         td->rate_blocks[ddir] = iops;
 316         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 317         return 0;
 318 }
 319
 320 static int check_min_rate(struct thread_data *td, struct timeval *now,
 321                           unsigned long *bytes_done)
 322 {
 323         int ret = 0;
 324
 325         if (bytes_done[0])
 326                 ret |= __check_min_rate(td, now, 0);
 327         if (bytes_done[1])
 328                 ret |= __check_min_rate(td, now, 1);
 329
 330         return ret;
 331 }
 332
 333 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 334 {
 335         if (!td->o.timeout)
 336                 return 0;
 337         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 338                 return 1;
 339
 340         return 0;
 341 }
 342
 343 /*
 344  * When job exits, we can cancel the in-flight IO if we are using async
 345  * io. Attempt to do so.
 346  */
 347 static void cleanup_pending_aio(struct thread_data *td)
 348 {
 349         struct flist_head *entry, *n;
 350         struct io_u *io_u;
 351         int r;
 352
 353         /*
 354          * get immediately available events, if any
 355          */
 356         r = io_u_queued_complete(td, 0, NULL);
 357         if (r < 0)
 358                 return;
 359
 360         /*
 361          * now cancel remaining active events
 362          */
 363         if (td->io_ops->cancel) {
 364                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 365                         io_u = flist_entry(entry, struct io_u, list);
 366
 367                         /*
 368                          * if the io_u isn't in flight, then that generally
 369                          * means someone leaked an io_u. complain but fix
 370                          * it up, so we don't stall here.
 371                          */
 372                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 373                                 log_err("fio: non-busy IO on busy list\n");
 374                                 put_io_u(td, io_u);
 375                         } else {
 376                                 r = td->io_ops->cancel(td, io_u);
 377                                 if (!r)
 378                                         put_io_u(td, io_u);
 379                         }
 380                 }
 381         }
 382
 383         if (td->cur_depth)
 384                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 385 }
 386
 387 /*
 388  * Helper to handle the final sync of a file. Works just like the normal
 389  * io path, just does everything sync.
 390  */
 391 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 392 {
 393         struct io_u *io_u = __get_io_u(td);
 394         int ret;
 395
 396         if (!io_u)
 397                 return 1;
 398
 399         io_u->ddir = DDIR_SYNC;
 400         io_u->file = f;
 401
 402         if (td_io_prep(td, io_u)) {
 403                 put_io_u(td, io_u);
 404                 return 1;
 405         }
 406
 407 requeue:
 408         ret = td_io_queue(td, io_u);
 409         if (ret < 0) {
 410                 td_verror(td, io_u->error, "td_io_queue");
 411                 put_io_u(td, io_u);
 412                 return 1;
 413         } else if (ret == FIO_Q_QUEUED) {
 414                 if (io_u_queued_complete(td, 1, NULL) < 0)
 415                         return 1;
 416         } else if (ret == FIO_Q_COMPLETED) {
 417                 if (io_u->error) {
 418                         td_verror(td, io_u->error, "td_io_queue");
 419                         return 1;
 420                 }
 421
 422                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 423                         return 1;
 424         } else if (ret == FIO_Q_BUSY) {
 425                 if (td_io_commit(td))
 426                         return 1;
 427                 goto requeue;
 428         }
 429
 430         return 0;
 431 }
 432
 433 static inline void __update_tv_cache(struct thread_data *td)
 434 {
 435         fio_gettime(&td->tv_cache, NULL);
 436 }
 437
 438 static inline void update_tv_cache(struct thread_data *td)
 439 {
 440         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 441                 __update_tv_cache(td);
 442 }
 443
 444 static int break_on_this_error(struct thread_data *td, int *retptr)
 445 {
 446         int ret = *retptr;
 447
 448         if (ret < 0 || td->error) {
 449                 int err;
 450
 451                 if (!td->o.continue_on_error)
 452                         return 1;
 453
 454                 if (ret < 0)
 455                         err = -ret;
 456                 else
 457                         err = td->error;
 458
 459                 if (td_non_fatal_error(err)) {
 460                         /*
 461                          * Continue with the I/Os in case of
 462                          * a non fatal error.
 463                          */
 464                         update_error_count(td, err);
 465                         td_clear_error(td);
 466                         *retptr = 0;
 467                         return 0;
 468                 } else if (td->o.fill_device && err == ENOSPC) {
 469                         /*
 470                          * We expect to hit this error if
 471                          * fill_device option is set.
 472                          */
 473                         td_clear_error(td);
 474                         td->terminate = 1;
 475                         return 1;
 476                 } else {
 477                         /*
 478                          * Stop the I/O in case of a fatal
 479                          * error.
 480                          */
 481                         update_error_count(td, err);
 482                         return 1;
 483                 }
 484         }
 485
 486         return 0;
 487 }
 488
 489 /*
 490  * The main verify engine. Runs over the writes we previously submitted,
 491  * reads the blocks back in, and checks the crc/md5 of the data.
 492  */
 493 static void do_verify(struct thread_data *td)
 494 {
 495         struct fio_file *f;
 496         struct io_u *io_u;
 497         int ret, min_events;
 498         unsigned int i;
 499
 500         dprint(FD_VERIFY, "starting loop\n");
 501
 502         /*
 503          * sync io first and invalidate cache, to make sure we really
 504          * read from disk.
 505          */
 506         for_each_file(td, f, i) {
 507                 if (!fio_file_open(f))
 508                         continue;
 509                 if (fio_io_sync(td, f))
 510                         break;
 511                 if (file_invalidate_cache(td, f))
 512                         break;
 513         }
 514
 515         if (td->error)
 516                 return;
 517
 518         td_set_runstate(td, TD_VERIFYING);
 519
 520         io_u = NULL;
 521         while (!td->terminate) {
 522                 int ret2, full;
 523
 524                 update_tv_cache(td);
 525
 526                 if (runtime_exceeded(td, &td->tv_cache)) {
 527                         __update_tv_cache(td);
 528                         if (runtime_exceeded(td, &td->tv_cache)) {
 529                                 td->terminate = 1;
 530                                 break;
 531                         }
 532                 }
 533
 534                 io_u = __get_io_u(td);
 535                 if (!io_u)
 536                         break;
 537
 538                 if (get_next_verify(td, io_u)) {
 539                         put_io_u(td, io_u);
 540                         break;
 541                 }
 542
 543                 if (td_io_prep(td, io_u)) {
 544                         put_io_u(td, io_u);
 545                         break;
 546                 }
 547
 548                 if (td->o.verify_async)
 549                         io_u->end_io = verify_io_u_async;
 550                 else
 551                         io_u->end_io = verify_io_u;
 552
 553                 ret = td_io_queue(td, io_u);
 554                 switch (ret) {
 555                 case FIO_Q_COMPLETED:
 556                         if (io_u->error) {
 557                                 ret = -io_u->error;
 558                                 clear_io_u(td, io_u);
 559                         } else if (io_u->resid) {
 560                                 int bytes = io_u->xfer_buflen - io_u->resid;
 561
 562                                 /*
 563                                  * zero read, fail
 564                                  */
 565                                 if (!bytes) {
 566                                         td_verror(td, EIO, "full resid");
 567                                         put_io_u(td, io_u);
 568                                         break;
 569                                 }
 570
 571                                 io_u->xfer_buflen = io_u->resid;
 572                                 io_u->xfer_buf += bytes;
 573                                 io_u->offset += bytes;
 574
 575                                 if (ddir_rw(io_u->ddir))
 576                                         td->ts.short_io_u[io_u->ddir]++;
 577
 578                                 f = io_u->file;
 579                                 if (io_u->offset == f->real_file_size)
 580                                         goto sync_done;
 581
 582                                 requeue_io_u(td, &io_u);
 583                         } else {
 584 sync_done:
 585                                 ret = io_u_sync_complete(td, io_u, NULL);
 586                                 if (ret < 0)
 587                                         break;
 588                         }
 589                         continue;
 590                 case FIO_Q_QUEUED:
 591                         break;
 592                 case FIO_Q_BUSY:
 593                         requeue_io_u(td, &io_u);
 594                         ret2 = td_io_commit(td);
 595                         if (ret2 < 0)
 596                                 ret = ret2;
 597                         break;
 598                 default:
 599                         assert(ret < 0);
 600                         td_verror(td, -ret, "td_io_queue");
 601                         break;
 602                 }
 603
 604                 if (break_on_this_error(td, &ret))
 605                         break;
 606
 607                 /*
 608                  * if we can queue more, do so. but check if there are
 609                  * completed io_u's first. Note that we can get BUSY even
 610                  * without IO queued, if the system is resource starved.
 611                  */
 612                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 613                 if (full || !td->o.iodepth_batch_complete) {
 614                         min_events = min(td->o.iodepth_batch_complete,
 615                                          td->cur_depth);
 616                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 617                                 min_events = 1;
 618
 619                         do {
 620                                 /*
 621                                  * Reap required number of io units, if any,
 622                                  * and do the verification on them through
 623                                  * the callback handler
 624                                  */
 625                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 626                                         ret = -1;
 627                                         break;
 628                                 }
 629                         } while (full && (td->cur_depth > td->o.iodepth_low));
 630                 }
 631                 if (ret < 0)
 632                         break;
 633         }
 634
 635         if (!td->error) {
 636                 min_events = td->cur_depth;
 637
 638                 if (min_events)
 639                         ret = io_u_queued_complete(td, min_events, NULL);
 640         } else
 641                 cleanup_pending_aio(td);
 642
 643         td_set_runstate(td, TD_RUNNING);
 644
 645         dprint(FD_VERIFY, "exiting loop\n");
 646 }
 647
 648 /*
 649  * Main IO worker function. It retrieves io_u's to process and queues
 650  * and reaps them, checking for rate and errors along the way.
 651  */
 652 static void do_io(struct thread_data *td)
 653 {
 654         unsigned int i;
 655         int ret = 0;
 656
 657         if (in_ramp_time(td))
 658                 td_set_runstate(td, TD_RAMP);
 659         else
 660                 td_set_runstate(td, TD_RUNNING);
 661
 662         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 663                 (!flist_empty(&td->trim_list)) ||
 664                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 665                 struct timeval comp_time;
 666                 unsigned long bytes_done[2] = { 0, 0 };
 667                 int min_evts = 0;
 668                 struct io_u *io_u;
 669                 int ret2, full;
 670
 671                 if (td->terminate)
 672                         break;
 673
 674                 update_tv_cache(td);
 675
 676                 if (runtime_exceeded(td, &td->tv_cache)) {
 677                         __update_tv_cache(td);
 678                         if (runtime_exceeded(td, &td->tv_cache)) {
 679                                 td->terminate = 1;
 680                                 break;
 681                         }
 682                 }
 683
 684                 io_u = get_io_u(td);
 685                 if (!io_u)
 686                         break;
 687
 688                 /*
 689                  * Add verification end_io handler, if asked to verify
 690                  * a previously written file.
 691                  */
 692                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 693                     !td_rw(td)) {
 694                         if (td->o.verify_async)
 695                                 io_u->end_io = verify_io_u_async;
 696                         else
 697                                 io_u->end_io = verify_io_u;
 698                         td_set_runstate(td, TD_VERIFYING);
 699                 } else if (in_ramp_time(td))
 700                         td_set_runstate(td, TD_RAMP);
 701                 else
 702                         td_set_runstate(td, TD_RUNNING);
 703
 704                 ret = td_io_queue(td, io_u);
 705                 switch (ret) {
 706                 case FIO_Q_COMPLETED:
 707                         if (io_u->error) {
 708                                 ret = -io_u->error;
 709                                 clear_io_u(td, io_u);
 710                         } else if (io_u->resid) {
 711                                 int bytes = io_u->xfer_buflen - io_u->resid;
 712                                 struct fio_file *f = io_u->file;
 713
 714                                 /*
 715                                  * zero read, fail
 716                                  */
 717                                 if (!bytes) {
 718                                         td_verror(td, EIO, "full resid");
 719                                         put_io_u(td, io_u);
 720                                         break;
 721                                 }
 722
 723                                 io_u->xfer_buflen = io_u->resid;
 724                                 io_u->xfer_buf += bytes;
 725                                 io_u->offset += bytes;
 726
 727                                 if (ddir_rw(io_u->ddir))
 728                                         td->ts.short_io_u[io_u->ddir]++;
 729
 730                                 if (io_u->offset == f->real_file_size)
 731                                         goto sync_done;
 732
 733                                 requeue_io_u(td, &io_u);
 734                         } else {
 735 sync_done:
 736                                 if (__should_check_rate(td, 0) ||
 737                                     __should_check_rate(td, 1))
 738                                         fio_gettime(&comp_time, NULL);
 739
 740                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 741                                 if (ret < 0)
 742                                         break;
 743                         }
 744                         break;
 745                 case FIO_Q_QUEUED:
 746                         /*
 747                          * if the engine doesn't have a commit hook,
 748                          * the io_u is really queued. if it does have such
 749                          * a hook, it has to call io_u_queued() itself.
 750                          */
 751                         if (td->io_ops->commit == NULL)
 752                                 io_u_queued(td, io_u);
 753                         break;
 754                 case FIO_Q_BUSY:
 755                         requeue_io_u(td, &io_u);
 756                         ret2 = td_io_commit(td);
 757                         if (ret2 < 0)
 758                                 ret = ret2;
 759                         break;
 760                 default:
 761                         assert(ret < 0);
 762                         put_io_u(td, io_u);
 763                         break;
 764                 }
 765
 766                 if (break_on_this_error(td, &ret))
 767                         break;
 768
 769                 /*
 770                  * See if we need to complete some commands. Note that we
 771                  * can get BUSY even without IO queued, if the system is
 772                  * resource starved.
 773                  */
 774                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 775                 if (full || !td->o.iodepth_batch_complete) {
 776                         min_evts = min(td->o.iodepth_batch_complete,
 777                                         td->cur_depth);
 778                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 779                                 min_evts = 1;
 780
 781                         if (__should_check_rate(td, 0) ||
 782                             __should_check_rate(td, 1))
 783                                 fio_gettime(&comp_time, NULL);
 784
 785                         do {
 786                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 787                                 if (ret < 0)
 788                                         break;
 789
 790                         } while (full && (td->cur_depth > td->o.iodepth_low));
 791                 }
 792
 793                 if (ret < 0)
 794                         break;
 795                 if (!(bytes_done[0] + bytes_done[1]))
 796                         continue;
 797
 798                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 799                         if (check_min_rate(td, &comp_time, bytes_done)) {
 800                                 if (exitall_on_terminate)
 801                                         fio_terminate_threads(td->groupid);
 802                                 td_verror(td, EIO, "check_min_rate");
 803                                 break;
 804                         }
 805                 }
 806
 807                 if (td->o.thinktime) {
 808                         unsigned long long b;
 809
 810                         b = td->io_blocks[0] + td->io_blocks[1];
 811                         if (!(b % td->o.thinktime_blocks)) {
 812                                 int left;
 813
 814                                 if (td->o.thinktime_spin)
 815                                         usec_spin(td->o.thinktime_spin);
 816
 817                                 left = td->o.thinktime - td->o.thinktime_spin;
 818                                 if (left)
 819                                         usec_sleep(td, left);
 820                         }
 821                 }
 822         }
 823
 824         if (td->trim_entries)
 825                 printf("trim entries %ld\n", td->trim_entries);
 826
 827         if (td->o.fill_device && td->error == ENOSPC) {
 828                 td->error = 0;
 829                 td->terminate = 1;
 830         }
 831         if (!td->error) {
 832                 struct fio_file *f;
 833
 834                 i = td->cur_depth;
 835                 if (i) {
 836                         ret = io_u_queued_complete(td, i, NULL);
 837                         if (td->o.fill_device && td->error == ENOSPC)
 838                                 td->error = 0;
 839                 }
 840
 841                 if (should_fsync(td) && td->o.end_fsync) {
 842                         td_set_runstate(td, TD_FSYNCING);
 843
 844                         for_each_file(td, f, i) {
 845                                 if (!fio_file_open(f))
 846                                         continue;
 847                                 fio_io_sync(td, f);
 848                         }
 849                 }
 850         } else
 851                 cleanup_pending_aio(td);
 852
 853         /*
 854          * stop job if we failed doing any IO
 855          */
 856         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 857                 td->done = 1;
 858 }
 859
 860 static void cleanup_io_u(struct thread_data *td)
 861 {
 862         struct flist_head *entry, *n;
 863         struct io_u *io_u;
 864
 865         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 866                 io_u = flist_entry(entry, struct io_u, list);
 867
 868                 flist_del(&io_u->list);
 869                 fio_memfree(io_u, sizeof(*io_u));
 870         }
 871
 872         free_io_mem(td);
 873 }
 874
 875 static int init_io_u(struct thread_data *td)
 876 {
 877         struct io_u *io_u;
 878         unsigned int max_bs;
 879         int cl_align, i, max_units;
 880         char *p;
 881
 882         max_units = td->o.iodepth;
 883         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 884         td->orig_buffer_size = (unsigned long long) max_bs
 885                                         * (unsigned long long) max_units;
 886
 887         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 888                 unsigned long bs;
 889
 890                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 891                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 892         }
 893
 894         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 895                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 896                 return 1;
 897         }
 898
 899         if (allocate_io_mem(td))
 900                 return 1;
 901
 902         if (td->o.odirect || td->o.mem_align ||
 903             (td->io_ops->flags & FIO_RAWIO))
 904                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 905         else
 906                 p = td->orig_buffer;
 907
 908         cl_align = os_cache_line_size();
 909
 910         for (i = 0; i < max_units; i++) {
 911                 void *ptr;
 912
 913                 if (td->terminate)
 914                         return 1;
 915
 916                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 917                 if (!ptr) {
 918                         log_err("fio: unable to allocate aligned memory\n");
 919                         break;
 920                 }
 921
 922                 io_u = ptr;
 923                 memset(io_u, 0, sizeof(*io_u));
 924                 INIT_FLIST_HEAD(&io_u->list);
 925                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 926
 927                 if (!(td->io_ops->flags & FIO_NOIO)) {
 928                         io_u->buf = p + max_bs * i;
 929                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 930
 931                         if (td_write(td))
 932                                 io_u_fill_buffer(td, io_u, max_bs);
 933                         if (td_write(td) && td->o.verify_pattern_bytes) {
 934                                 /*
 935                                  * Fill the buffer with the pattern if we are
 936                                  * going to be doing writes.
 937                                  */
 938                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 939                         }
 940                 }
 941
 942                 io_u->index = i;
 943                 io_u->flags = IO_U_F_FREE;
 944                 flist_add(&io_u->list, &td->io_u_freelist);
 945         }
 946
 947         return 0;
 948 }
 949
 950 static int switch_ioscheduler(struct thread_data *td)
 951 {
 952         char tmp[256], tmp2[128];
 953         FILE *f;
 954         int ret;
 955
 956         if (td->io_ops->flags & FIO_DISKLESSIO)
 957                 return 0;
 958
 959         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 960
 961         f = fopen(tmp, "r+");
 962         if (!f) {
 963                 if (errno == ENOENT) {
 964                         log_err("fio: os or kernel doesn't support IO scheduler"
 965                                 " switching\n");
 966                         return 0;
 967                 }
 968                 td_verror(td, errno, "fopen iosched");
 969                 return 1;
 970         }
 971
 972         /*
 973          * Set io scheduler.
 974          */
 975         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 976         if (ferror(f) || ret != 1) {
 977                 td_verror(td, errno, "fwrite");
 978                 fclose(f);
 979                 return 1;
 980         }
 981
 982         rewind(f);
 983
 984         /*
 985          * Read back and check that the selected scheduler is now the default.
 986          */
 987         ret = fread(tmp, 1, sizeof(tmp), f);
 988         if (ferror(f) || ret < 0) {
 989                 td_verror(td, errno, "fread");
 990                 fclose(f);
 991                 return 1;
 992         }
 993
 994         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 995         if (!strstr(tmp, tmp2)) {
 996                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 997                 td_verror(td, EINVAL, "iosched_switch");
 998                 fclose(f);
 999                 return 1;
1000         }
1001
1002         fclose(f);
1003         return 0;
1004 }
1005
1006 static int keep_running(struct thread_data *td)
1007 {
1008         unsigned long long io_done;
1009
1010         if (td->done)
1011                 return 0;
1012         if (td->o.time_based)
1013                 return 1;
1014         if (td->o.loops) {
1015                 td->o.loops--;
1016                 return 1;
1017         }
1018
1019         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1020                         + td->io_skip_bytes;
1021         if (io_done < td->o.size)
1022                 return 1;
1023
1024         return 0;
1025 }
1026
1027 static void reset_io_counters(struct thread_data *td)
1028 {
1029         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1030         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1031         td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1032         td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1033         td->zone_bytes = 0;
1034         td->rate_bytes[0] = td->rate_bytes[1] = 0;
1035         td->rate_blocks[0] = td->rate_blocks[1] = 0;
1036
1037         td->last_was_sync = 0;
1038
1039         /*
1040          * reset file done count if we are to start over
1041          */
1042         if (td->o.time_based || td->o.loops)
1043                 td->nr_done_files = 0;
1044 }
1045
1046 void reset_all_stats(struct thread_data *td)
1047 {
1048         struct timeval tv;
1049         int i;
1050
1051         reset_io_counters(td);
1052
1053         for (i = 0; i < 2; i++) {
1054                 td->io_bytes[i] = 0;
1055                 td->io_blocks[i] = 0;
1056                 td->io_issues[i] = 0;
1057                 td->ts.total_io_u[i] = 0;
1058         }
1059
1060         fio_gettime(&tv, NULL);
1061         td->ts.runtime[0] = 0;
1062         td->ts.runtime[1] = 0;
1063         memcpy(&td->epoch, &tv, sizeof(tv));
1064         memcpy(&td->start, &tv, sizeof(tv));
1065 }
1066
1067 static void clear_io_state(struct thread_data *td)
1068 {
1069         struct fio_file *f;
1070         unsigned int i;
1071
1072         reset_io_counters(td);
1073
1074         close_files(td);
1075         for_each_file(td, f, i)
1076                 fio_file_clear_done(f);
1077
1078         /*
1079          * Set the same seed to get repeatable runs
1080          */
1081         td_fill_rand_seeds(td);
1082 }
1083
1084 static int exec_string(const char *string)
1085 {
1086         int ret, newlen = strlen(string) + 1 + 8;
1087         char *str;
1088
1089         str = malloc(newlen);
1090         sprintf(str, "sh -c %s", string);
1091
1092         ret = system(str);
1093         if (ret == -1)
1094                 log_err("fio: exec of cmd <%s> failed\n", str);
1095
1096         free(str);
1097         return ret;
1098 }
1099
1100 /*
1101  * Entry point for the thread based jobs. The process based jobs end up
1102  * here as well, after a little setup.
1103  */
1104 static void *thread_main(void *data)
1105 {
1106         unsigned long long elapsed;
1107         struct thread_data *td = data;
1108         pthread_condattr_t attr;
1109         int clear_state;
1110
1111         if (!td->o.use_thread) {
1112                 setsid();
1113                 td->pid = getpid();
1114         } else
1115                 td->pid = gettid();
1116
1117         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1118
1119         INIT_FLIST_HEAD(&td->io_u_freelist);
1120         INIT_FLIST_HEAD(&td->io_u_busylist);
1121         INIT_FLIST_HEAD(&td->io_u_requeues);
1122         INIT_FLIST_HEAD(&td->io_log_list);
1123         INIT_FLIST_HEAD(&td->io_hist_list);
1124         INIT_FLIST_HEAD(&td->verify_list);
1125         INIT_FLIST_HEAD(&td->trim_list);
1126         pthread_mutex_init(&td->io_u_lock, NULL);
1127         td->io_hist_tree = RB_ROOT;
1128
1129         pthread_condattr_init(&attr);
1130         pthread_cond_init(&td->verify_cond, &attr);
1131         pthread_cond_init(&td->free_cond, &attr);
1132
1133         td_set_runstate(td, TD_INITIALIZED);
1134         dprint(FD_MUTEX, "up startup_mutex\n");
1135         fio_mutex_up(startup_mutex);
1136         dprint(FD_MUTEX, "wait on td->mutex\n");
1137         fio_mutex_down(td->mutex);
1138         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1139
1140         /*
1141          * the ->mutex mutex is now no longer used, close it to avoid
1142          * eating a file descriptor
1143          */
1144         fio_mutex_remove(td->mutex);
1145
1146         /*
1147          * A new gid requires privilege, so we need to do this before setting
1148          * the uid.
1149          */
1150         if (td->o.gid != -1U && setgid(td->o.gid)) {
1151                 td_verror(td, errno, "setgid");
1152                 goto err;
1153         }
1154         if (td->o.uid != -1U && setuid(td->o.uid)) {
1155                 td_verror(td, errno, "setuid");
1156                 goto err;
1157         }
1158
1159         /*
1160          * If we have a gettimeofday() thread, make sure we exclude that
1161          * thread from this job
1162          */
1163         if (td->o.gtod_cpu)
1164                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1165
1166         /*
1167          * Set affinity first, in case it has an impact on the memory
1168          * allocations.
1169          */
1170         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1171                 td_verror(td, errno, "cpu_set_affinity");
1172                 goto err;
1173         }
1174
1175         /*
1176          * May alter parameters that init_io_u() will use, so we need to
1177          * do this first.
1178          */
1179         if (init_iolog(td))
1180                 goto err;
1181
1182         if (init_io_u(td))
1183                 goto err;
1184
1185         if (td->o.verify_async && verify_async_init(td))
1186                 goto err;
1187
1188         if (td->ioprio_set) {
1189                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1190                         td_verror(td, errno, "ioprio_set");
1191                         goto err;
1192                 }
1193         }
1194
1195         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1196                 goto err;
1197
1198         if (nice(td->o.nice) == -1) {
1199                 td_verror(td, errno, "nice");
1200                 goto err;
1201         }
1202
1203         if (td->o.ioscheduler && switch_ioscheduler(td))
1204                 goto err;
1205
1206         if (!td->o.create_serialize && setup_files(td))
1207                 goto err;
1208
1209         if (td_io_init(td))
1210                 goto err;
1211
1212         if (init_random_map(td))
1213                 goto err;
1214
1215         if (td->o.exec_prerun) {
1216                 if (exec_string(td->o.exec_prerun))
1217                         goto err;
1218         }
1219
1220         if (td->o.pre_read) {
1221                 if (pre_read_files(td) < 0)
1222                         goto err;
1223         }
1224
1225         fio_gettime(&td->epoch, NULL);
1226         getrusage(RUSAGE_SELF, &td->ru_start);
1227
1228         clear_state = 0;
1229         while (keep_running(td)) {
1230                 fio_gettime(&td->start, NULL);
1231                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1232                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1233                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1234
1235                 if (td->o.ratemin[0] || td->o.ratemin[1])
1236                         memcpy(&td->lastrate, &td->bw_sample_time,
1237                                                         sizeof(td->lastrate));
1238
1239                 if (clear_state)
1240                         clear_io_state(td);
1241
1242                 prune_io_piece_log(td);
1243
1244                 do_io(td);
1245
1246                 clear_state = 1;
1247
1248                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1249                         elapsed = utime_since_now(&td->start);
1250                         td->ts.runtime[DDIR_READ] += elapsed;
1251                 }
1252                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1253                         elapsed = utime_since_now(&td->start);
1254                         td->ts.runtime[DDIR_WRITE] += elapsed;
1255                 }
1256
1257                 if (td->error || td->terminate)
1258                         break;
1259
1260                 if (!td->o.do_verify ||
1261                     td->o.verify == VERIFY_NONE ||
1262                     (td->io_ops->flags & FIO_UNIDIR))
1263                         continue;
1264
1265                 clear_io_state(td);
1266
1267                 fio_gettime(&td->start, NULL);
1268
1269                 do_verify(td);
1270
1271                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1272
1273                 if (td->error || td->terminate)
1274                         break;
1275         }
1276
1277         update_rusage_stat(td);
1278         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1279         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1280         td->ts.total_run_time = mtime_since_now(&td->epoch);
1281         td->ts.io_bytes[0] = td->io_bytes[0];
1282         td->ts.io_bytes[1] = td->io_bytes[1];
1283
1284         fio_mutex_down(writeout_mutex);
1285         if (td->bw_log) {
1286                 if (td->o.bw_log_file) {
1287                         finish_log_named(td, td->bw_log,
1288                                                 td->o.bw_log_file, "bw");
1289                 } else
1290                         finish_log(td, td->bw_log, "bw");
1291         }
1292         if (td->lat_log) {
1293                 if (td->o.lat_log_file) {
1294                         finish_log_named(td, td->lat_log,
1295                                                 td->o.lat_log_file, "lat");
1296                 } else
1297                         finish_log(td, td->lat_log, "lat");
1298         }
1299         if (td->slat_log) {
1300                 if (td->o.lat_log_file) {
1301                         finish_log_named(td, td->slat_log,
1302                                                 td->o.lat_log_file, "slat");
1303                 } else
1304                         finish_log(td, td->slat_log, "slat");
1305         }
1306         if (td->clat_log) {
1307                 if (td->o.lat_log_file) {
1308                         finish_log_named(td, td->clat_log,
1309                                                 td->o.lat_log_file, "clat");
1310                 } else
1311                         finish_log(td, td->clat_log, "clat");
1312         }
1313         if (td->iops_log) {
1314                 if (td->o.iops_log_file) {
1315                         finish_log_named(td, td->iops_log,
1316                                                 td->o.iops_log_file, "iops");
1317                 } else
1318                         finish_log(td, td->iops_log, "iops");
1319         }
1320
1321         fio_mutex_up(writeout_mutex);
1322         if (td->o.exec_postrun)
1323                 exec_string(td->o.exec_postrun);
1324
1325         if (exitall_on_terminate)
1326                 fio_terminate_threads(td->groupid);
1327
1328 err:
1329         if (td->error)
1330                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1331                                                         td->verror);
1332
1333         if (td->o.verify_async)
1334                 verify_async_exit(td);
1335
1336         close_and_free_files(td);
1337         close_ioengine(td);
1338         cleanup_io_u(td);
1339         cgroup_shutdown(td, &cgroup_mnt);
1340
1341         if (td->o.cpumask_set) {
1342                 int ret = fio_cpuset_exit(&td->o.cpumask);
1343
1344                 td_verror(td, ret, "fio_cpuset_exit");
1345         }
1346
1347         /*
1348          * do this very late, it will log file closing as well
1349          */
1350         if (td->o.write_iolog_file)
1351                 write_iolog_close(td);
1352
1353         options_mem_free(td);
1354         td_set_runstate(td, TD_EXITED);
1355         return (void *) (unsigned long) td->error;
1356 }
1357
1358 /*
1359  * We cannot pass the td data into a forked process, so attach the td and
1360  * pass it to the thread worker.
1361  */
1362 static int fork_main(int shmid, int offset)
1363 {
1364         struct thread_data *td;
1365         void *data, *ret;
1366
1367 #ifndef __hpux
1368         data = shmat(shmid, NULL, 0);
1369         if (data == (void *) -1) {
1370                 int __err = errno;
1371
1372                 perror("shmat");
1373                 return __err;
1374         }
1375 #else
1376         /*
1377          * HP-UX inherits shm mappings?
1378          */
1379         data = threads;
1380 #endif
1381
1382         td = data + offset * sizeof(struct thread_data);
1383         ret = thread_main(td);
1384         shmdt(data);
1385         return (int) (unsigned long) ret;
1386 }
1387
1388 /*
1389  * Run over the job map and reap the threads that have exited, if any.
1390  */
1391 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1392 {
1393         struct thread_data *td;
1394         int i, cputhreads, realthreads, pending, status, ret;
1395
1396         /*
1397          * reap exited threads (TD_EXITED -> TD_REAPED)
1398          */
1399         realthreads = pending = cputhreads = 0;
1400         for_each_td(td, i) {
1401                 int flags = 0;
1402
1403                 /*
1404                  * ->io_ops is NULL for a thread that has closed its
1405                  * io engine
1406                  */
1407                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1408                         cputhreads++;
1409                 else
1410                         realthreads++;
1411
1412                 if (!td->pid) {
1413                         pending++;
1414                         continue;
1415                 }
1416                 if (td->runstate == TD_REAPED)
1417                         continue;
1418                 if (td->o.use_thread) {
1419                         if (td->runstate == TD_EXITED) {
1420                                 td_set_runstate(td, TD_REAPED);
1421                                 goto reaped;
1422                         }
1423                         continue;
1424                 }
1425
1426                 flags = WNOHANG;
1427                 if (td->runstate == TD_EXITED)
1428                         flags = 0;
1429
1430                 /*
1431                  * check if someone quit or got killed in an unusual way
1432                  */
1433                 ret = waitpid(td->pid, &status, flags);
1434                 if (ret < 0) {
1435                         if (errno == ECHILD) {
1436                                 log_err("fio: pid=%d disappeared %d\n",
1437                                                 (int) td->pid, td->runstate);
1438                                 td_set_runstate(td, TD_REAPED);
1439                                 goto reaped;
1440                         }
1441                         perror("waitpid");
1442                 } else if (ret == td->pid) {
1443                         if (WIFSIGNALED(status)) {
1444                                 int sig = WTERMSIG(status);
1445
1446                                 if (sig != SIGTERM)
1447                                         log_err("fio: pid=%d, got signal=%d\n",
1448                                                         (int) td->pid, sig);
1449                                 td_set_runstate(td, TD_REAPED);
1450                                 goto reaped;
1451                         }
1452                         if (WIFEXITED(status)) {
1453                                 if (WEXITSTATUS(status) && !td->error)
1454                                         td->error = WEXITSTATUS(status);
1455
1456                                 td_set_runstate(td, TD_REAPED);
1457                                 goto reaped;
1458                         }
1459                 }
1460
1461                 /*
1462                  * thread is not dead, continue
1463                  */
1464                 pending++;
1465                 continue;
1466 reaped:
1467                 (*nr_running)--;
1468                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1469                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1470                 if (!td->pid)
1471                         pending--;
1472
1473                 if (td->error)
1474                         exit_value++;
1475
1476                 done_secs += mtime_since_now(&td->epoch) / 1000;
1477         }
1478
1479         if (*nr_running == cputhreads && !pending && realthreads)
1480                 fio_terminate_threads(TERMINATE_ALL);
1481 }
1482
1483 static void *gtod_thread_main(void *data)
1484 {
1485         fio_mutex_up(startup_mutex);
1486
1487         /*
1488          * As long as we have jobs around, update the clock. It would be nice
1489          * to have some way of NOT hammering that CPU with gettimeofday(),
1490          * but I'm not sure what to use outside of a simple CPU nop to relax
1491          * it - we don't want to lose precision.
1492          */
1493         while (threads) {
1494                 fio_gtod_update();
1495                 nop;
1496         }
1497
1498         return NULL;
1499 }
1500
1501 static int fio_start_gtod_thread(void)
1502 {
1503         pthread_attr_t attr;
1504         int ret;
1505
1506         pthread_attr_init(&attr);
1507         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1508         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1509         pthread_attr_destroy(&attr);
1510         if (ret) {
1511                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1512                 return 1;
1513         }
1514
1515         ret = pthread_detach(gtod_thread);
1516         if (ret) {
1517                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1518                 return 1;
1519         }
1520
1521         dprint(FD_MUTEX, "wait on startup_mutex\n");
1522         fio_mutex_down(startup_mutex);
1523         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1524         return 0;
1525 }
1526
1527 /*
1528  * Main function for kicking off and reaping jobs, as needed.
1529  */
1530 static void run_threads(void)
1531 {
1532         struct thread_data *td;
1533         unsigned long spent;
1534         int i, todo, nr_running, m_rate, t_rate, nr_started;
1535
1536         if (fio_pin_memory())
1537                 return;
1538
1539         if (fio_gtod_offload && fio_start_gtod_thread())
1540                 return;
1541
1542         set_sig_handlers();
1543
1544         if (!terse_output) {
1545                 log_info("Starting ");
1546                 if (nr_thread)
1547                         log_info("%d thread%s", nr_thread,
1548                                                 nr_thread > 1 ? "s" : "");
1549                 if (nr_process) {
1550                         if (nr_thread)
1551                                 printf(" and ");
1552                         log_info("%d process%s", nr_process,
1553                                                 nr_process > 1 ? "es" : "");
1554                 }
1555                 log_info("\n");
1556                 fflush(stdout);
1557         }
1558
1559         todo = thread_number;
1560         nr_running = 0;
1561         nr_started = 0;
1562         m_rate = t_rate = 0;
1563
1564         for_each_td(td, i) {
1565                 print_status_init(td->thread_number - 1);
1566
1567                 if (!td->o.create_serialize)
1568                         continue;
1569
1570                 /*
1571                  * do file setup here so it happens sequentially,
1572                  * we don't want X number of threads getting their
1573                  * client data interspersed on disk
1574                  */
1575                 if (setup_files(td)) {
1576                         exit_value++;
1577                         if (td->error)
1578                                 log_err("fio: pid=%d, err=%d/%s\n",
1579                                         (int) td->pid, td->error, td->verror);
1580                         td_set_runstate(td, TD_REAPED);
1581                         todo--;
1582                 } else {
1583                         struct fio_file *f;
1584                         unsigned int j;
1585
1586                         /*
1587                          * for sharing to work, each job must always open
1588                          * its own files. so close them, if we opened them
1589                          * for creation
1590                          */
1591                         for_each_file(td, f, j) {
1592                                 if (fio_file_open(f))
1593                                         td_io_close_file(td, f);
1594                         }
1595                 }
1596         }
1597
1598         set_genesis_time();
1599
1600         while (todo) {
1601                 struct thread_data *map[REAL_MAX_JOBS];
1602                 struct timeval this_start;
1603                 int this_jobs = 0, left;
1604
1605                 /*
1606                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1607                  */
1608                 for_each_td(td, i) {
1609                         if (td->runstate != TD_NOT_CREATED)
1610                                 continue;
1611
1612                         /*
1613                          * never got a chance to start, killed by other
1614                          * thread for some reason
1615                          */
1616                         if (td->terminate) {
1617                                 todo--;
1618                                 continue;
1619                         }
1620
1621                         if (td->o.start_delay) {
1622                                 spent = mtime_since_genesis();
1623
1624                                 if (td->o.start_delay * 1000 > spent)
1625                                         continue;
1626                         }
1627
1628                         if (td->o.stonewall && (nr_started || nr_running)) {
1629                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1630                                                         td->o.name);
1631                                 break;
1632                         }
1633
1634                         init_disk_util(td);
1635
1636                         /*
1637                          * Set state to created. Thread will transition
1638                          * to TD_INITIALIZED when it's done setting up.
1639                          */
1640                         td_set_runstate(td, TD_CREATED);
1641                         map[this_jobs++] = td;
1642                         nr_started++;
1643
1644                         if (td->o.use_thread) {
1645                                 int ret;
1646
1647                                 dprint(FD_PROCESS, "will pthread_create\n");
1648                                 ret = pthread_create(&td->thread, NULL,
1649                                                         thread_main, td);
1650                                 if (ret) {
1651                                         log_err("pthread_create: %s\n",
1652                                                         strerror(ret));
1653                                         nr_started--;
1654                                         break;
1655                                 }
1656                                 ret = pthread_detach(td->thread);
1657                                 if (ret)
1658                                         log_err("pthread_detach: %s",
1659                                                         strerror(ret));
1660                         } else {
1661                                 pid_t pid;
1662                                 dprint(FD_PROCESS, "will fork\n");
1663                                 pid = fork();
1664                                 if (!pid) {
1665                                         int ret = fork_main(shm_id, i);
1666
1667                                         _exit(ret);
1668                                 } else if (i == fio_debug_jobno)
1669                                         *fio_debug_jobp = pid;
1670                         }
1671                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1672                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1673                                 log_err("fio: job startup hung? exiting.\n");
1674                                 fio_terminate_threads(TERMINATE_ALL);
1675                                 fio_abort = 1;
1676                                 nr_started--;
1677                                 break;
1678                         }
1679                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1680                 }
1681
1682                 /*
1683                  * Wait for the started threads to transition to
1684                  * TD_INITIALIZED.
1685                  */
1686                 fio_gettime(&this_start, NULL);
1687                 left = this_jobs;
1688                 while (left && !fio_abort) {
1689                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1690                                 break;
1691
1692                         usleep(100000);
1693
1694                         for (i = 0; i < this_jobs; i++) {
1695                                 td = map[i];
1696                                 if (!td)
1697                                         continue;
1698                                 if (td->runstate == TD_INITIALIZED) {
1699                                         map[i] = NULL;
1700                                         left--;
1701                                 } else if (td->runstate >= TD_EXITED) {
1702                                         map[i] = NULL;
1703                                         left--;
1704                                         todo--;
1705                                         nr_running++; /* work-around... */
1706                                 }
1707                         }
1708                 }
1709
1710                 if (left) {
1711                         log_err("fio: %d jobs failed to start\n", left);
1712                         for (i = 0; i < this_jobs; i++) {
1713                                 td = map[i];
1714                                 if (!td)
1715                                         continue;
1716                                 kill(td->pid, SIGTERM);
1717                         }
1718                         break;
1719                 }
1720
1721                 /*
1722                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1723                  */
1724                 for_each_td(td, i) {
1725                         if (td->runstate != TD_INITIALIZED)
1726                                 continue;
1727
1728                         if (in_ramp_time(td))
1729                                 td_set_runstate(td, TD_RAMP);
1730                         else
1731                                 td_set_runstate(td, TD_RUNNING);
1732                         nr_running++;
1733                         nr_started--;
1734                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1735                         t_rate += td->o.rate[0] + td->o.rate[1];
1736                         todo--;
1737                         fio_mutex_up(td->mutex);
1738                 }
1739
1740                 reap_threads(&nr_running, &t_rate, &m_rate);
1741
1742                 if (todo) {
1743                         if (is_backend)
1744                                 fio_server_idle_loop();
1745                         else
1746                                 usleep(100000);
1747                 }
1748         }
1749
1750         while (nr_running) {
1751                 reap_threads(&nr_running, &t_rate, &m_rate);
1752
1753                 if (is_backend)
1754                         fio_server_idle_loop();
1755                 else
1756                         usleep(10000);
1757         }
1758
1759         update_io_ticks();
1760         fio_unpin_memory();
1761 }
1762
1763 int exec_run(void)
1764 {
1765         if (nr_clients)
1766                 return fio_handle_clients();
1767         if (exec_profile && load_profile(exec_profile))
1768                 return 1;
1769
1770         if (!thread_number)
1771                 return 0;
1772
1773         if (write_bw_log) {
1774                 setup_log(&agg_io_log[DDIR_READ]);
1775                 setup_log(&agg_io_log[DDIR_WRITE]);
1776         }
1777
1778         startup_mutex = fio_mutex_init(0);
1779         if (startup_mutex == NULL)
1780                 return 1;
1781         writeout_mutex = fio_mutex_init(1);
1782         if (writeout_mutex == NULL)
1783                 return 1;
1784
1785         set_genesis_time();
1786         create_disk_util_thread();
1787
1788         cgroup_list = smalloc(sizeof(*cgroup_list));
1789         INIT_FLIST_HEAD(cgroup_list);
1790
1791         run_threads();
1792
1793         if (!fio_abort) {
1794                 show_run_stats();
1795                 if (write_bw_log) {
1796                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1797                         __finish_log(agg_io_log[DDIR_WRITE],
1798                                         "agg-write_bw.log");
1799                 }
1800         }
1801
1802         cgroup_kill(cgroup_list);
1803         sfree(cgroup_list);
1804         sfree(cgroup_mnt);
1805
1806         fio_mutex_remove(startup_mutex);
1807         fio_mutex_remove(writeout_mutex);
1808         return exit_value;
1809 }
1810
1811 void reset_fio_state(void)
1812 {
1813         groupid = 0;
1814         thread_number = 0;
1815         nr_process = 0;
1816         nr_thread = 0;
1817         done_secs = 0;
1818 }
1819
1820 static int endian_check(void)
1821 {
1822         union {
1823                 uint8_t c[8];
1824                 uint64_t v;
1825         } u;
1826         int le = 0, be = 0;
1827
1828         u.v = 0x12;
1829         if (u.c[7] == 0x12)
1830                 be = 1;
1831         else if (u.c[0] == 0x12)
1832                 le = 1;
1833
1834 #if defined(FIO_LITTLE_ENDIAN)
1835         if (be)
1836                 return 1;
1837 #elif defined(FIO_BIG_ENDIAN)
1838         if (le)
1839                 return 1;
1840 #else
1841         return 1;
1842 #endif
1843
1844         if (!le && !be)
1845                 return 1;
1846
1847         return 0;
1848 }
1849
1850 int main(int argc, char *argv[], char *envp[])
1851 {
1852         long ps;
1853
1854         if (endian_check()) {
1855                 log_err("fio: endianness settings appear wrong.\n");
1856                 log_err("fio: please report this to fio@vger.kernel.org\n");
1857                 return 1;
1858         }
1859
1860         arch_init(envp);
1861
1862         sinit();
1863
1864         /*
1865          * We need locale for number printing, if it isn't set then just
1866          * go with the US format.
1867          */
1868         if (!getenv("LC_NUMERIC"))
1869                 setlocale(LC_NUMERIC, "en_US");
1870
1871         ps = sysconf(_SC_PAGESIZE);
1872         if (ps < 0) {
1873                 log_err("Failed to get page size\n");
1874                 return 1;
1875         }
1876
1877         page_size = ps;
1878         page_mask = ps - 1;
1879
1880         fio_keywords_init();
1881
1882         if (parse_options(argc, argv))
1883                 return 1;
1884
1885         return exec_run();
1886 }