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                         fio_server_send_status();
 188                 else
 189                         print_thread_status();
 190         }
 191
 192         return NULL;
 193 }
 194
 195 static int create_disk_util_thread(void)
 196 {
 197         int ret;
 198
 199         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 200         if (ret) {
 201                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 202                 return 1;
 203         }
 204
 205         ret = pthread_detach(disk_util_thread);
 206         if (ret) {
 207                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 208                 return 1;
 209         }
 210
 211         dprint(FD_MUTEX, "wait on startup_mutex\n");
 212         fio_mutex_down(startup_mutex);
 213         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 214         return 0;
 215 }
 216
 217 static void set_sig_handlers(void)
 218 {
 219         struct sigaction act;
 220
 221         memset(&act, 0, sizeof(act));
 222         act.sa_handler = sig_int;
 223         act.sa_flags = SA_RESTART;
 224         sigaction(SIGINT, &act, NULL);
 225
 226         memset(&act, 0, sizeof(act));
 227         act.sa_handler = sig_int;
 228         act.sa_flags = SA_RESTART;
 229         sigaction(SIGTERM, &act, NULL);
 230
 231         if (is_backend) {
 232                 memset(&act, 0, sizeof(act));
 233                 act.sa_handler = sig_int;
 234                 act.sa_flags = SA_RESTART;
 235                 sigaction(SIGPIPE, &act, NULL);
 236         }
 237 }
 238
 239 /*
 240  * Check if we are above the minimum rate given.
 241  */
 242 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 243                             enum fio_ddir ddir)
 244 {
 245         unsigned long long bytes = 0;
 246         unsigned long iops = 0;
 247         unsigned long spent;
 248         unsigned long rate;
 249         unsigned int ratemin = 0;
 250         unsigned int rate_iops = 0;
 251         unsigned int rate_iops_min = 0;
 252
 253         assert(ddir_rw(ddir));
 254
 255         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 256                 return 0;
 257
 258         /*
 259          * allow a 2 second settle period in the beginning
 260          */
 261         if (mtime_since(&td->start, now) < 2000)
 262                 return 0;
 263
 264         iops += td->this_io_blocks[ddir];
 265         bytes += td->this_io_bytes[ddir];
 266         ratemin += td->o.ratemin[ddir];
 267         rate_iops += td->o.rate_iops[ddir];
 268         rate_iops_min += td->o.rate_iops_min[ddir];
 269
 270         /*
 271          * if rate blocks is set, sample is running
 272          */
 273         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 274                 spent = mtime_since(&td->lastrate[ddir], now);
 275                 if (spent < td->o.ratecycle)
 276                         return 0;
 277
 278                 if (td->o.rate[ddir]) {
 279                         /*
 280                          * check bandwidth specified rate
 281                          */
 282                         if (bytes < td->rate_bytes[ddir]) {
 283                                 log_err("%s: min rate %u not met\n", td->o.name,
 284                                                                 ratemin);
 285                                 return 1;
 286                         } else {
 287                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 288                                 if (rate < ratemin ||
 289                                     bytes < td->rate_bytes[ddir]) {
 290                                         log_err("%s: min rate %u not met, got"
 291                                                 " %luKB/sec\n", td->o.name,
 292                                                         ratemin, rate);
 293                                         return 1;
 294                                 }
 295                         }
 296                 } else {
 297                         /*
 298                          * checks iops specified rate
 299                          */
 300                         if (iops < rate_iops) {
 301                                 log_err("%s: min iops rate %u not met\n",
 302                                                 td->o.name, rate_iops);
 303                                 return 1;
 304                         } else {
 305                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 306                                 if (rate < rate_iops_min ||
 307                                     iops < td->rate_blocks[ddir]) {
 308                                         log_err("%s: min iops rate %u not met,"
 309                                                 " got %lu\n", td->o.name,
 310                                                         rate_iops_min, rate);
 311                                 }
 312                         }
 313                 }
 314         }
 315
 316         td->rate_bytes[ddir] = bytes;
 317         td->rate_blocks[ddir] = iops;
 318         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 319         return 0;
 320 }
 321
 322 static int check_min_rate(struct thread_data *td, struct timeval *now,
 323                           unsigned long *bytes_done)
 324 {
 325         int ret = 0;
 326
 327         if (bytes_done[0])
 328                 ret |= __check_min_rate(td, now, 0);
 329         if (bytes_done[1])
 330                 ret |= __check_min_rate(td, now, 1);
 331
 332         return ret;
 333 }
 334
 335 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 336 {
 337         if (!td->o.timeout)
 338                 return 0;
 339         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 340                 return 1;
 341
 342         return 0;
 343 }
 344
 345 /*
 346  * When job exits, we can cancel the in-flight IO if we are using async
 347  * io. Attempt to do so.
 348  */
 349 static void cleanup_pending_aio(struct thread_data *td)
 350 {
 351         struct flist_head *entry, *n;
 352         struct io_u *io_u;
 353         int r;
 354
 355         /*
 356          * get immediately available events, if any
 357          */
 358         r = io_u_queued_complete(td, 0, NULL);
 359         if (r < 0)
 360                 return;
 361
 362         /*
 363          * now cancel remaining active events
 364          */
 365         if (td->io_ops->cancel) {
 366                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 367                         io_u = flist_entry(entry, struct io_u, list);
 368
 369                         /*
 370                          * if the io_u isn't in flight, then that generally
 371                          * means someone leaked an io_u. complain but fix
 372                          * it up, so we don't stall here.
 373                          */
 374                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 375                                 log_err("fio: non-busy IO on busy list\n");
 376                                 put_io_u(td, io_u);
 377                         } else {
 378                                 r = td->io_ops->cancel(td, io_u);
 379                                 if (!r)
 380                                         put_io_u(td, io_u);
 381                         }
 382                 }
 383         }
 384
 385         if (td->cur_depth)
 386                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 387 }
 388
 389 /*
 390  * Helper to handle the final sync of a file. Works just like the normal
 391  * io path, just does everything sync.
 392  */
 393 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 394 {
 395         struct io_u *io_u = __get_io_u(td);
 396         int ret;
 397
 398         if (!io_u)
 399                 return 1;
 400
 401         io_u->ddir = DDIR_SYNC;
 402         io_u->file = f;
 403
 404         if (td_io_prep(td, io_u)) {
 405                 put_io_u(td, io_u);
 406                 return 1;
 407         }
 408
 409 requeue:
 410         ret = td_io_queue(td, io_u);
 411         if (ret < 0) {
 412                 td_verror(td, io_u->error, "td_io_queue");
 413                 put_io_u(td, io_u);
 414                 return 1;
 415         } else if (ret == FIO_Q_QUEUED) {
 416                 if (io_u_queued_complete(td, 1, NULL) < 0)
 417                         return 1;
 418         } else if (ret == FIO_Q_COMPLETED) {
 419                 if (io_u->error) {
 420                         td_verror(td, io_u->error, "td_io_queue");
 421                         return 1;
 422                 }
 423
 424                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 425                         return 1;
 426         } else if (ret == FIO_Q_BUSY) {
 427                 if (td_io_commit(td))
 428                         return 1;
 429                 goto requeue;
 430         }
 431
 432         return 0;
 433 }
 434
 435 static inline void __update_tv_cache(struct thread_data *td)
 436 {
 437         fio_gettime(&td->tv_cache, NULL);
 438 }
 439
 440 static inline void update_tv_cache(struct thread_data *td)
 441 {
 442         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 443                 __update_tv_cache(td);
 444 }
 445
 446 static int break_on_this_error(struct thread_data *td, int *retptr)
 447 {
 448         int ret = *retptr;
 449
 450         if (ret < 0 || td->error) {
 451                 int err;
 452
 453                 if (!td->o.continue_on_error)
 454                         return 1;
 455
 456                 if (ret < 0)
 457                         err = -ret;
 458                 else
 459                         err = td->error;
 460
 461                 if (td_non_fatal_error(err)) {
 462                         /*
 463                          * Continue with the I/Os in case of
 464                          * a non fatal error.
 465                          */
 466                         update_error_count(td, err);
 467                         td_clear_error(td);
 468                         *retptr = 0;
 469                         return 0;
 470                 } else if (td->o.fill_device && err == ENOSPC) {
 471                         /*
 472                          * We expect to hit this error if
 473                          * fill_device option is set.
 474                          */
 475                         td_clear_error(td);
 476                         td->terminate = 1;
 477                         return 1;
 478                 } else {
 479                         /*
 480                          * Stop the I/O in case of a fatal
 481                          * error.
 482                          */
 483                         update_error_count(td, err);
 484                         return 1;
 485                 }
 486         }
 487
 488         return 0;
 489 }
 490
 491 /*
 492  * The main verify engine. Runs over the writes we previously submitted,
 493  * reads the blocks back in, and checks the crc/md5 of the data.
 494  */
 495 static void do_verify(struct thread_data *td)
 496 {
 497         struct fio_file *f;
 498         struct io_u *io_u;
 499         int ret, min_events;
 500         unsigned int i;
 501
 502         dprint(FD_VERIFY, "starting loop\n");
 503
 504         /*
 505          * sync io first and invalidate cache, to make sure we really
 506          * read from disk.
 507          */
 508         for_each_file(td, f, i) {
 509                 if (!fio_file_open(f))
 510                         continue;
 511                 if (fio_io_sync(td, f))
 512                         break;
 513                 if (file_invalidate_cache(td, f))
 514                         break;
 515         }
 516
 517         if (td->error)
 518                 return;
 519
 520         td_set_runstate(td, TD_VERIFYING);
 521
 522         io_u = NULL;
 523         while (!td->terminate) {
 524                 int ret2, full;
 525
 526                 update_tv_cache(td);
 527
 528                 if (runtime_exceeded(td, &td->tv_cache)) {
 529                         __update_tv_cache(td);
 530                         if (runtime_exceeded(td, &td->tv_cache)) {
 531                                 td->terminate = 1;
 532                                 break;
 533                         }
 534                 }
 535
 536                 io_u = __get_io_u(td);
 537                 if (!io_u)
 538                         break;
 539
 540                 if (get_next_verify(td, io_u)) {
 541                         put_io_u(td, io_u);
 542                         break;
 543                 }
 544
 545                 if (td_io_prep(td, io_u)) {
 546                         put_io_u(td, io_u);
 547                         break;
 548                 }
 549
 550                 if (td->o.verify_async)
 551                         io_u->end_io = verify_io_u_async;
 552                 else
 553                         io_u->end_io = verify_io_u;
 554
 555                 ret = td_io_queue(td, io_u);
 556                 switch (ret) {
 557                 case FIO_Q_COMPLETED:
 558                         if (io_u->error) {
 559                                 ret = -io_u->error;
 560                                 clear_io_u(td, io_u);
 561                         } else if (io_u->resid) {
 562                                 int bytes = io_u->xfer_buflen - io_u->resid;
 563
 564                                 /*
 565                                  * zero read, fail
 566                                  */
 567                                 if (!bytes) {
 568                                         td_verror(td, EIO, "full resid");
 569                                         put_io_u(td, io_u);
 570                                         break;
 571                                 }
 572
 573                                 io_u->xfer_buflen = io_u->resid;
 574                                 io_u->xfer_buf += bytes;
 575                                 io_u->offset += bytes;
 576
 577                                 if (ddir_rw(io_u->ddir))
 578                                         td->ts.short_io_u[io_u->ddir]++;
 579
 580                                 f = io_u->file;
 581                                 if (io_u->offset == f->real_file_size)
 582                                         goto sync_done;
 583
 584                                 requeue_io_u(td, &io_u);
 585                         } else {
 586 sync_done:
 587                                 ret = io_u_sync_complete(td, io_u, NULL);
 588                                 if (ret < 0)
 589                                         break;
 590                         }
 591                         continue;
 592                 case FIO_Q_QUEUED:
 593                         break;
 594                 case FIO_Q_BUSY:
 595                         requeue_io_u(td, &io_u);
 596                         ret2 = td_io_commit(td);
 597                         if (ret2 < 0)
 598                                 ret = ret2;
 599                         break;
 600                 default:
 601                         assert(ret < 0);
 602                         td_verror(td, -ret, "td_io_queue");
 603                         break;
 604                 }
 605
 606                 if (break_on_this_error(td, &ret))
 607                         break;
 608
 609                 /*
 610                  * if we can queue more, do so. but check if there are
 611                  * completed io_u's first. Note that we can get BUSY even
 612                  * without IO queued, if the system is resource starved.
 613                  */
 614                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 615                 if (full || !td->o.iodepth_batch_complete) {
 616                         min_events = min(td->o.iodepth_batch_complete,
 617                                          td->cur_depth);
 618                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 619                                 min_events = 1;
 620
 621                         do {
 622                                 /*
 623                                  * Reap required number of io units, if any,
 624                                  * and do the verification on them through
 625                                  * the callback handler
 626                                  */
 627                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 628                                         ret = -1;
 629                                         break;
 630                                 }
 631                         } while (full && (td->cur_depth > td->o.iodepth_low));
 632                 }
 633                 if (ret < 0)
 634                         break;
 635         }
 636
 637         if (!td->error) {
 638                 min_events = td->cur_depth;
 639
 640                 if (min_events)
 641                         ret = io_u_queued_complete(td, min_events, NULL);
 642         } else
 643                 cleanup_pending_aio(td);
 644
 645         td_set_runstate(td, TD_RUNNING);
 646
 647         dprint(FD_VERIFY, "exiting loop\n");
 648 }
 649
 650 /*
 651  * Main IO worker function. It retrieves io_u's to process and queues
 652  * and reaps them, checking for rate and errors along the way.
 653  */
 654 static void do_io(struct thread_data *td)
 655 {
 656         unsigned int i;
 657         int ret = 0;
 658
 659         if (in_ramp_time(td))
 660                 td_set_runstate(td, TD_RAMP);
 661         else
 662                 td_set_runstate(td, TD_RUNNING);
 663
 664         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 665                 (!flist_empty(&td->trim_list)) ||
 666                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 667                 struct timeval comp_time;
 668                 unsigned long bytes_done[2] = { 0, 0 };
 669                 int min_evts = 0;
 670                 struct io_u *io_u;
 671                 int ret2, full;
 672
 673                 if (td->terminate)
 674                         break;
 675
 676                 update_tv_cache(td);
 677
 678                 if (runtime_exceeded(td, &td->tv_cache)) {
 679                         __update_tv_cache(td);
 680                         if (runtime_exceeded(td, &td->tv_cache)) {
 681                                 td->terminate = 1;
 682                                 break;
 683                         }
 684                 }
 685
 686                 io_u = get_io_u(td);
 687                 if (!io_u)
 688                         break;
 689
 690                 /*
 691                  * Add verification end_io handler, if asked to verify
 692                  * a previously written file.
 693                  */
 694                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 695                     !td_rw(td)) {
 696                         if (td->o.verify_async)
 697                                 io_u->end_io = verify_io_u_async;
 698                         else
 699                                 io_u->end_io = verify_io_u;
 700                         td_set_runstate(td, TD_VERIFYING);
 701                 } else if (in_ramp_time(td))
 702                         td_set_runstate(td, TD_RAMP);
 703                 else
 704                         td_set_runstate(td, TD_RUNNING);
 705
 706                 ret = td_io_queue(td, io_u);
 707                 switch (ret) {
 708                 case FIO_Q_COMPLETED:
 709                         if (io_u->error) {
 710                                 ret = -io_u->error;
 711                                 clear_io_u(td, io_u);
 712                         } else if (io_u->resid) {
 713                                 int bytes = io_u->xfer_buflen - io_u->resid;
 714                                 struct fio_file *f = io_u->file;
 715
 716                                 /*
 717                                  * zero read, fail
 718                                  */
 719                                 if (!bytes) {
 720                                         td_verror(td, EIO, "full resid");
 721                                         put_io_u(td, io_u);
 722                                         break;
 723                                 }
 724
 725                                 io_u->xfer_buflen = io_u->resid;
 726                                 io_u->xfer_buf += bytes;
 727                                 io_u->offset += bytes;
 728
 729                                 if (ddir_rw(io_u->ddir))
 730                                         td->ts.short_io_u[io_u->ddir]++;
 731
 732                                 if (io_u->offset == f->real_file_size)
 733                                         goto sync_done;
 734
 735                                 requeue_io_u(td, &io_u);
 736                         } else {
 737 sync_done:
 738                                 if (__should_check_rate(td, 0) ||
 739                                     __should_check_rate(td, 1))
 740                                         fio_gettime(&comp_time, NULL);
 741
 742                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 743                                 if (ret < 0)
 744                                         break;
 745                         }
 746                         break;
 747                 case FIO_Q_QUEUED:
 748                         /*
 749                          * if the engine doesn't have a commit hook,
 750                          * the io_u is really queued. if it does have such
 751                          * a hook, it has to call io_u_queued() itself.
 752                          */
 753                         if (td->io_ops->commit == NULL)
 754                                 io_u_queued(td, io_u);
 755                         break;
 756                 case FIO_Q_BUSY:
 757                         requeue_io_u(td, &io_u);
 758                         ret2 = td_io_commit(td);
 759                         if (ret2 < 0)
 760                                 ret = ret2;
 761                         break;
 762                 default:
 763                         assert(ret < 0);
 764                         put_io_u(td, io_u);
 765                         break;
 766                 }
 767
 768                 if (break_on_this_error(td, &ret))
 769                         break;
 770
 771                 /*
 772                  * See if we need to complete some commands. Note that we
 773                  * can get BUSY even without IO queued, if the system is
 774                  * resource starved.
 775                  */
 776                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 777                 if (full || !td->o.iodepth_batch_complete) {
 778                         min_evts = min(td->o.iodepth_batch_complete,
 779                                         td->cur_depth);
 780                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 781                                 min_evts = 1;
 782
 783                         if (__should_check_rate(td, 0) ||
 784                             __should_check_rate(td, 1))
 785                                 fio_gettime(&comp_time, NULL);
 786
 787                         do {
 788                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 789                                 if (ret < 0)
 790                                         break;
 791
 792                         } while (full && (td->cur_depth > td->o.iodepth_low));
 793                 }
 794
 795                 if (ret < 0)
 796                         break;
 797                 if (!(bytes_done[0] + bytes_done[1]))
 798                         continue;
 799
 800                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 801                         if (check_min_rate(td, &comp_time, bytes_done)) {
 802                                 if (exitall_on_terminate)
 803                                         fio_terminate_threads(td->groupid);
 804                                 td_verror(td, EIO, "check_min_rate");
 805                                 break;
 806                         }
 807                 }
 808
 809                 if (td->o.thinktime) {
 810                         unsigned long long b;
 811
 812                         b = td->io_blocks[0] + td->io_blocks[1];
 813                         if (!(b % td->o.thinktime_blocks)) {
 814                                 int left;
 815
 816                                 if (td->o.thinktime_spin)
 817                                         usec_spin(td->o.thinktime_spin);
 818
 819                                 left = td->o.thinktime - td->o.thinktime_spin;
 820                                 if (left)
 821                                         usec_sleep(td, left);
 822                         }
 823                 }
 824         }
 825
 826         if (td->trim_entries)
 827                 printf("trim entries %ld\n", td->trim_entries);
 828
 829         if (td->o.fill_device && td->error == ENOSPC) {
 830                 td->error = 0;
 831                 td->terminate = 1;
 832         }
 833         if (!td->error) {
 834                 struct fio_file *f;
 835
 836                 i = td->cur_depth;
 837                 if (i) {
 838                         ret = io_u_queued_complete(td, i, NULL);
 839                         if (td->o.fill_device && td->error == ENOSPC)
 840                                 td->error = 0;
 841                 }
 842
 843                 if (should_fsync(td) && td->o.end_fsync) {
 844                         td_set_runstate(td, TD_FSYNCING);
 845
 846                         for_each_file(td, f, i) {
 847                                 if (!fio_file_open(f))
 848                                         continue;
 849                                 fio_io_sync(td, f);
 850                         }
 851                 }
 852         } else
 853                 cleanup_pending_aio(td);
 854
 855         /*
 856          * stop job if we failed doing any IO
 857          */
 858         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 859                 td->done = 1;
 860 }
 861
 862 static void cleanup_io_u(struct thread_data *td)
 863 {
 864         struct flist_head *entry, *n;
 865         struct io_u *io_u;
 866
 867         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 868                 io_u = flist_entry(entry, struct io_u, list);
 869
 870                 flist_del(&io_u->list);
 871                 fio_memfree(io_u, sizeof(*io_u));
 872         }
 873
 874         free_io_mem(td);
 875 }
 876
 877 static int init_io_u(struct thread_data *td)
 878 {
 879         struct io_u *io_u;
 880         unsigned int max_bs;
 881         int cl_align, i, max_units;
 882         char *p;
 883
 884         max_units = td->o.iodepth;
 885         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 886         td->orig_buffer_size = (unsigned long long) max_bs
 887                                         * (unsigned long long) max_units;
 888
 889         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 890                 unsigned long bs;
 891
 892                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 893                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 894         }
 895
 896         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 897                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 898                 return 1;
 899         }
 900
 901         if (allocate_io_mem(td))
 902                 return 1;
 903
 904         if (td->o.odirect || td->o.mem_align ||
 905             (td->io_ops->flags & FIO_RAWIO))
 906                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 907         else
 908                 p = td->orig_buffer;
 909
 910         cl_align = os_cache_line_size();
 911
 912         for (i = 0; i < max_units; i++) {
 913                 void *ptr;
 914
 915                 if (td->terminate)
 916                         return 1;
 917
 918                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 919                 if (!ptr) {
 920                         log_err("fio: unable to allocate aligned memory\n");
 921                         break;
 922                 }
 923
 924                 io_u = ptr;
 925                 memset(io_u, 0, sizeof(*io_u));
 926                 INIT_FLIST_HEAD(&io_u->list);
 927                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 928
 929                 if (!(td->io_ops->flags & FIO_NOIO)) {
 930                         io_u->buf = p + max_bs * i;
 931                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 932
 933                         if (td_write(td))
 934                                 io_u_fill_buffer(td, io_u, max_bs);
 935                         if (td_write(td) && td->o.verify_pattern_bytes) {
 936                                 /*
 937                                  * Fill the buffer with the pattern if we are
 938                                  * going to be doing writes.
 939                                  */
 940                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 941                         }
 942                 }
 943
 944                 io_u->index = i;
 945                 io_u->flags = IO_U_F_FREE;
 946                 flist_add(&io_u->list, &td->io_u_freelist);
 947         }
 948
 949         return 0;
 950 }
 951
 952 static int switch_ioscheduler(struct thread_data *td)
 953 {
 954         char tmp[256], tmp2[128];
 955         FILE *f;
 956         int ret;
 957
 958         if (td->io_ops->flags & FIO_DISKLESSIO)
 959                 return 0;
 960
 961         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 962
 963         f = fopen(tmp, "r+");
 964         if (!f) {
 965                 if (errno == ENOENT) {
 966                         log_err("fio: os or kernel doesn't support IO scheduler"
 967                                 " switching\n");
 968                         return 0;
 969                 }
 970                 td_verror(td, errno, "fopen iosched");
 971                 return 1;
 972         }
 973
 974         /*
 975          * Set io scheduler.
 976          */
 977         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 978         if (ferror(f) || ret != 1) {
 979                 td_verror(td, errno, "fwrite");
 980                 fclose(f);
 981                 return 1;
 982         }
 983
 984         rewind(f);
 985
 986         /*
 987          * Read back and check that the selected scheduler is now the default.
 988          */
 989         ret = fread(tmp, 1, sizeof(tmp), f);
 990         if (ferror(f) || ret < 0) {
 991                 td_verror(td, errno, "fread");
 992                 fclose(f);
 993                 return 1;
 994         }
 995
 996         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 997         if (!strstr(tmp, tmp2)) {
 998                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 999                 td_verror(td, EINVAL, "iosched_switch");
1000                 fclose(f);
1001                 return 1;
1002         }
1003
1004         fclose(f);
1005         return 0;
1006 }
1007
1008 static int keep_running(struct thread_data *td)
1009 {
1010         unsigned long long io_done;
1011
1012         if (td->done)
1013                 return 0;
1014         if (td->o.time_based)
1015                 return 1;
1016         if (td->o.loops) {
1017                 td->o.loops--;
1018                 return 1;
1019         }
1020
1021         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1022                         + td->io_skip_bytes;
1023         if (io_done < td->o.size)
1024                 return 1;
1025
1026         return 0;
1027 }
1028
1029 static void reset_io_counters(struct thread_data *td)
1030 {
1031         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1032         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1033         td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1034         td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1035         td->zone_bytes = 0;
1036         td->rate_bytes[0] = td->rate_bytes[1] = 0;
1037         td->rate_blocks[0] = td->rate_blocks[1] = 0;
1038
1039         td->last_was_sync = 0;
1040
1041         /*
1042          * reset file done count if we are to start over
1043          */
1044         if (td->o.time_based || td->o.loops)
1045                 td->nr_done_files = 0;
1046 }
1047
1048 void reset_all_stats(struct thread_data *td)
1049 {
1050         struct timeval tv;
1051         int i;
1052
1053         reset_io_counters(td);
1054
1055         for (i = 0; i < 2; i++) {
1056                 td->io_bytes[i] = 0;
1057                 td->io_blocks[i] = 0;
1058                 td->io_issues[i] = 0;
1059                 td->ts.total_io_u[i] = 0;
1060         }
1061
1062         fio_gettime(&tv, NULL);
1063         td->ts.runtime[0] = 0;
1064         td->ts.runtime[1] = 0;
1065         memcpy(&td->epoch, &tv, sizeof(tv));
1066         memcpy(&td->start, &tv, sizeof(tv));
1067 }
1068
1069 static void clear_io_state(struct thread_data *td)
1070 {
1071         struct fio_file *f;
1072         unsigned int i;
1073
1074         reset_io_counters(td);
1075
1076         close_files(td);
1077         for_each_file(td, f, i)
1078                 fio_file_clear_done(f);
1079
1080         /*
1081          * Set the same seed to get repeatable runs
1082          */
1083         td_fill_rand_seeds(td);
1084 }
1085
1086 static int exec_string(const char *string)
1087 {
1088         int ret, newlen = strlen(string) + 1 + 8;
1089         char *str;
1090
1091         str = malloc(newlen);
1092         sprintf(str, "sh -c %s", string);
1093
1094         ret = system(str);
1095         if (ret == -1)
1096                 log_err("fio: exec of cmd <%s> failed\n", str);
1097
1098         free(str);
1099         return ret;
1100 }
1101
1102 /*
1103  * Entry point for the thread based jobs. The process based jobs end up
1104  * here as well, after a little setup.
1105  */
1106 static void *thread_main(void *data)
1107 {
1108         unsigned long long elapsed;
1109         struct thread_data *td = data;
1110         pthread_condattr_t attr;
1111         int clear_state;
1112
1113         if (!td->o.use_thread) {
1114                 setsid();
1115                 td->pid = getpid();
1116         } else
1117                 td->pid = gettid();
1118
1119         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1120
1121         INIT_FLIST_HEAD(&td->io_u_freelist);
1122         INIT_FLIST_HEAD(&td->io_u_busylist);
1123         INIT_FLIST_HEAD(&td->io_u_requeues);
1124         INIT_FLIST_HEAD(&td->io_log_list);
1125         INIT_FLIST_HEAD(&td->io_hist_list);
1126         INIT_FLIST_HEAD(&td->verify_list);
1127         INIT_FLIST_HEAD(&td->trim_list);
1128         pthread_mutex_init(&td->io_u_lock, NULL);
1129         td->io_hist_tree = RB_ROOT;
1130
1131         pthread_condattr_init(&attr);
1132         pthread_cond_init(&td->verify_cond, &attr);
1133         pthread_cond_init(&td->free_cond, &attr);
1134
1135         td_set_runstate(td, TD_INITIALIZED);
1136         dprint(FD_MUTEX, "up startup_mutex\n");
1137         fio_mutex_up(startup_mutex);
1138         dprint(FD_MUTEX, "wait on td->mutex\n");
1139         fio_mutex_down(td->mutex);
1140         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1141
1142         /*
1143          * the ->mutex mutex is now no longer used, close it to avoid
1144          * eating a file descriptor
1145          */
1146         fio_mutex_remove(td->mutex);
1147
1148         /*
1149          * A new gid requires privilege, so we need to do this before setting
1150          * the uid.
1151          */
1152         if (td->o.gid != -1U && setgid(td->o.gid)) {
1153                 td_verror(td, errno, "setgid");
1154                 goto err;
1155         }
1156         if (td->o.uid != -1U && setuid(td->o.uid)) {
1157                 td_verror(td, errno, "setuid");
1158                 goto err;
1159         }
1160
1161         /*
1162          * If we have a gettimeofday() thread, make sure we exclude that
1163          * thread from this job
1164          */
1165         if (td->o.gtod_cpu)
1166                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1167
1168         /*
1169          * Set affinity first, in case it has an impact on the memory
1170          * allocations.
1171          */
1172         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1173                 td_verror(td, errno, "cpu_set_affinity");
1174                 goto err;
1175         }
1176
1177         /*
1178          * May alter parameters that init_io_u() will use, so we need to
1179          * do this first.
1180          */
1181         if (init_iolog(td))
1182                 goto err;
1183
1184         if (init_io_u(td))
1185                 goto err;
1186
1187         if (td->o.verify_async && verify_async_init(td))
1188                 goto err;
1189
1190         if (td->ioprio_set) {
1191                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1192                         td_verror(td, errno, "ioprio_set");
1193                         goto err;
1194                 }
1195         }
1196
1197         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1198                 goto err;
1199
1200         if (nice(td->o.nice) == -1) {
1201                 td_verror(td, errno, "nice");
1202                 goto err;
1203         }
1204
1205         if (td->o.ioscheduler && switch_ioscheduler(td))
1206                 goto err;
1207
1208         if (!td->o.create_serialize && setup_files(td))
1209                 goto err;
1210
1211         if (td_io_init(td))
1212                 goto err;
1213
1214         if (init_random_map(td))
1215                 goto err;
1216
1217         if (td->o.exec_prerun) {
1218                 if (exec_string(td->o.exec_prerun))
1219                         goto err;
1220         }
1221
1222         if (td->o.pre_read) {
1223                 if (pre_read_files(td) < 0)
1224                         goto err;
1225         }
1226
1227         fio_gettime(&td->epoch, NULL);
1228         getrusage(RUSAGE_SELF, &td->ru_start);
1229
1230         clear_state = 0;
1231         while (keep_running(td)) {
1232                 fio_gettime(&td->start, NULL);
1233                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1234                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1235                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1236
1237                 if (td->o.ratemin[0] || td->o.ratemin[1])
1238                         memcpy(&td->lastrate, &td->bw_sample_time,
1239                                                         sizeof(td->lastrate));
1240
1241                 if (clear_state)
1242                         clear_io_state(td);
1243
1244                 prune_io_piece_log(td);
1245
1246                 do_io(td);
1247
1248                 clear_state = 1;
1249
1250                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1251                         elapsed = utime_since_now(&td->start);
1252                         td->ts.runtime[DDIR_READ] += elapsed;
1253                 }
1254                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1255                         elapsed = utime_since_now(&td->start);
1256                         td->ts.runtime[DDIR_WRITE] += elapsed;
1257                 }
1258
1259                 if (td->error || td->terminate)
1260                         break;
1261
1262                 if (!td->o.do_verify ||
1263                     td->o.verify == VERIFY_NONE ||
1264                     (td->io_ops->flags & FIO_UNIDIR))
1265                         continue;
1266
1267                 clear_io_state(td);
1268
1269                 fio_gettime(&td->start, NULL);
1270
1271                 do_verify(td);
1272
1273                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1274
1275                 if (td->error || td->terminate)
1276                         break;
1277         }
1278
1279         update_rusage_stat(td);
1280         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1281         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1282         td->ts.total_run_time = mtime_since_now(&td->epoch);
1283         td->ts.io_bytes[0] = td->io_bytes[0];
1284         td->ts.io_bytes[1] = td->io_bytes[1];
1285
1286         fio_mutex_down(writeout_mutex);
1287         if (td->bw_log) {
1288                 if (td->o.bw_log_file) {
1289                         finish_log_named(td, td->bw_log,
1290                                                 td->o.bw_log_file, "bw");
1291                 } else
1292                         finish_log(td, td->bw_log, "bw");
1293         }
1294         if (td->lat_log) {
1295                 if (td->o.lat_log_file) {
1296                         finish_log_named(td, td->lat_log,
1297                                                 td->o.lat_log_file, "lat");
1298                 } else
1299                         finish_log(td, td->lat_log, "lat");
1300         }
1301         if (td->slat_log) {
1302                 if (td->o.lat_log_file) {
1303                         finish_log_named(td, td->slat_log,
1304                                                 td->o.lat_log_file, "slat");
1305                 } else
1306                         finish_log(td, td->slat_log, "slat");
1307         }
1308         if (td->clat_log) {
1309                 if (td->o.lat_log_file) {
1310                         finish_log_named(td, td->clat_log,
1311                                                 td->o.lat_log_file, "clat");
1312                 } else
1313                         finish_log(td, td->clat_log, "clat");
1314         }
1315         if (td->iops_log) {
1316                 if (td->o.iops_log_file) {
1317                         finish_log_named(td, td->iops_log,
1318                                                 td->o.iops_log_file, "iops");
1319                 } else
1320                         finish_log(td, td->iops_log, "iops");
1321         }
1322
1323         fio_mutex_up(writeout_mutex);
1324         if (td->o.exec_postrun)
1325                 exec_string(td->o.exec_postrun);
1326
1327         if (exitall_on_terminate)
1328                 fio_terminate_threads(td->groupid);
1329
1330 err:
1331         if (td->error)
1332                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1333                                                         td->verror);
1334
1335         if (td->o.verify_async)
1336                 verify_async_exit(td);
1337
1338         close_and_free_files(td);
1339         close_ioengine(td);
1340         cleanup_io_u(td);
1341         cgroup_shutdown(td, &cgroup_mnt);
1342
1343         if (td->o.cpumask_set) {
1344                 int ret = fio_cpuset_exit(&td->o.cpumask);
1345
1346                 td_verror(td, ret, "fio_cpuset_exit");
1347         }
1348
1349         /*
1350          * do this very late, it will log file closing as well
1351          */
1352         if (td->o.write_iolog_file)
1353                 write_iolog_close(td);
1354
1355         options_mem_free(td);
1356         td_set_runstate(td, TD_EXITED);
1357         return (void *) (unsigned long) td->error;
1358 }
1359
1360 /*
1361  * We cannot pass the td data into a forked process, so attach the td and
1362  * pass it to the thread worker.
1363  */
1364 static int fork_main(int shmid, int offset)
1365 {
1366         struct thread_data *td;
1367         void *data, *ret;
1368
1369 #ifndef __hpux
1370         data = shmat(shmid, NULL, 0);
1371         if (data == (void *) -1) {
1372                 int __err = errno;
1373
1374                 perror("shmat");
1375                 return __err;
1376         }
1377 #else
1378         /*
1379          * HP-UX inherits shm mappings?
1380          */
1381         data = threads;
1382 #endif
1383
1384         td = data + offset * sizeof(struct thread_data);
1385         ret = thread_main(td);
1386         shmdt(data);
1387         return (int) (unsigned long) ret;
1388 }
1389
1390 /*
1391  * Run over the job map and reap the threads that have exited, if any.
1392  */
1393 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1394 {
1395         struct thread_data *td;
1396         int i, cputhreads, realthreads, pending, status, ret;
1397
1398         /*
1399          * reap exited threads (TD_EXITED -> TD_REAPED)
1400          */
1401         realthreads = pending = cputhreads = 0;
1402         for_each_td(td, i) {
1403                 int flags = 0;
1404
1405                 /*
1406                  * ->io_ops is NULL for a thread that has closed its
1407                  * io engine
1408                  */
1409                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1410                         cputhreads++;
1411                 else
1412                         realthreads++;
1413
1414                 if (!td->pid) {
1415                         pending++;
1416                         continue;
1417                 }
1418                 if (td->runstate == TD_REAPED)
1419                         continue;
1420                 if (td->o.use_thread) {
1421                         if (td->runstate == TD_EXITED) {
1422                                 td_set_runstate(td, TD_REAPED);
1423                                 goto reaped;
1424                         }
1425                         continue;
1426                 }
1427
1428                 flags = WNOHANG;
1429                 if (td->runstate == TD_EXITED)
1430                         flags = 0;
1431
1432                 /*
1433                  * check if someone quit or got killed in an unusual way
1434                  */
1435                 ret = waitpid(td->pid, &status, flags);
1436                 if (ret < 0) {
1437                         if (errno == ECHILD) {
1438                                 log_err("fio: pid=%d disappeared %d\n",
1439                                                 (int) td->pid, td->runstate);
1440                                 td_set_runstate(td, TD_REAPED);
1441                                 goto reaped;
1442                         }
1443                         perror("waitpid");
1444                 } else if (ret == td->pid) {
1445                         if (WIFSIGNALED(status)) {
1446                                 int sig = WTERMSIG(status);
1447
1448                                 if (sig != SIGTERM)
1449                                         log_err("fio: pid=%d, got signal=%d\n",
1450                                                         (int) td->pid, sig);
1451                                 td_set_runstate(td, TD_REAPED);
1452                                 goto reaped;
1453                         }
1454                         if (WIFEXITED(status)) {
1455                                 if (WEXITSTATUS(status) && !td->error)
1456                                         td->error = WEXITSTATUS(status);
1457
1458                                 td_set_runstate(td, TD_REAPED);
1459                                 goto reaped;
1460                         }
1461                 }
1462
1463                 /*
1464                  * thread is not dead, continue
1465                  */
1466                 pending++;
1467                 continue;
1468 reaped:
1469                 (*nr_running)--;
1470                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1471                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1472                 if (!td->pid)
1473                         pending--;
1474
1475                 if (td->error)
1476                         exit_value++;
1477
1478                 done_secs += mtime_since_now(&td->epoch) / 1000;
1479         }
1480
1481         if (*nr_running == cputhreads && !pending && realthreads)
1482                 fio_terminate_threads(TERMINATE_ALL);
1483 }
1484
1485 static void *gtod_thread_main(void *data)
1486 {
1487         fio_mutex_up(startup_mutex);
1488
1489         /*
1490          * As long as we have jobs around, update the clock. It would be nice
1491          * to have some way of NOT hammering that CPU with gettimeofday(),
1492          * but I'm not sure what to use outside of a simple CPU nop to relax
1493          * it - we don't want to lose precision.
1494          */
1495         while (threads) {
1496                 fio_gtod_update();
1497                 nop;
1498         }
1499
1500         return NULL;
1501 }
1502
1503 static int fio_start_gtod_thread(void)
1504 {
1505         pthread_attr_t attr;
1506         int ret;
1507
1508         pthread_attr_init(&attr);
1509         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1510         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1511         pthread_attr_destroy(&attr);
1512         if (ret) {
1513                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1514                 return 1;
1515         }
1516
1517         ret = pthread_detach(gtod_thread);
1518         if (ret) {
1519                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1520                 return 1;
1521         }
1522
1523         dprint(FD_MUTEX, "wait on startup_mutex\n");
1524         fio_mutex_down(startup_mutex);
1525         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1526         return 0;
1527 }
1528
1529 /*
1530  * Main function for kicking off and reaping jobs, as needed.
1531  */
1532 static void run_threads(void)
1533 {
1534         struct thread_data *td;
1535         unsigned long spent;
1536         int i, todo, nr_running, m_rate, t_rate, nr_started;
1537
1538         if (fio_pin_memory())
1539                 return;
1540
1541         if (fio_gtod_offload && fio_start_gtod_thread())
1542                 return;
1543
1544         set_sig_handlers();
1545
1546         if (!terse_output) {
1547                 log_info("Starting ");
1548                 if (nr_thread)
1549                         log_info("%d thread%s", nr_thread,
1550                                                 nr_thread > 1 ? "s" : "");
1551                 if (nr_process) {
1552                         if (nr_thread)
1553                                 printf(" and ");
1554                         log_info("%d process%s", nr_process,
1555                                                 nr_process > 1 ? "es" : "");
1556                 }
1557                 log_info("\n");
1558                 fflush(stdout);
1559         }
1560
1561         todo = thread_number;
1562         nr_running = 0;
1563         nr_started = 0;
1564         m_rate = t_rate = 0;
1565
1566         for_each_td(td, i) {
1567                 print_status_init(td->thread_number - 1);
1568
1569                 if (!td->o.create_serialize)
1570                         continue;
1571
1572                 /*
1573                  * do file setup here so it happens sequentially,
1574                  * we don't want X number of threads getting their
1575                  * client data interspersed on disk
1576                  */
1577                 if (setup_files(td)) {
1578                         exit_value++;
1579                         if (td->error)
1580                                 log_err("fio: pid=%d, err=%d/%s\n",
1581                                         (int) td->pid, td->error, td->verror);
1582                         td_set_runstate(td, TD_REAPED);
1583                         todo--;
1584                 } else {
1585                         struct fio_file *f;
1586                         unsigned int j;
1587
1588                         /*
1589                          * for sharing to work, each job must always open
1590                          * its own files. so close them, if we opened them
1591                          * for creation
1592                          */
1593                         for_each_file(td, f, j) {
1594                                 if (fio_file_open(f))
1595                                         td_io_close_file(td, f);
1596                         }
1597                 }
1598         }
1599
1600         set_genesis_time();
1601
1602         while (todo) {
1603                 struct thread_data *map[REAL_MAX_JOBS];
1604                 struct timeval this_start;
1605                 int this_jobs = 0, left;
1606
1607                 /*
1608                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1609                  */
1610                 for_each_td(td, i) {
1611                         if (td->runstate != TD_NOT_CREATED)
1612                                 continue;
1613
1614                         /*
1615                          * never got a chance to start, killed by other
1616                          * thread for some reason
1617                          */
1618                         if (td->terminate) {
1619                                 todo--;
1620                                 continue;
1621                         }
1622
1623                         if (td->o.start_delay) {
1624                                 spent = mtime_since_genesis();
1625
1626                                 if (td->o.start_delay * 1000 > spent)
1627                                         continue;
1628                         }
1629
1630                         if (td->o.stonewall && (nr_started || nr_running)) {
1631                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1632                                                         td->o.name);
1633                                 break;
1634                         }
1635
1636                         init_disk_util(td);
1637
1638                         /*
1639                          * Set state to created. Thread will transition
1640                          * to TD_INITIALIZED when it's done setting up.
1641                          */
1642                         td_set_runstate(td, TD_CREATED);
1643                         map[this_jobs++] = td;
1644                         nr_started++;
1645
1646                         if (td->o.use_thread) {
1647                                 int ret;
1648
1649                                 dprint(FD_PROCESS, "will pthread_create\n");
1650                                 ret = pthread_create(&td->thread, NULL,
1651                                                         thread_main, td);
1652                                 if (ret) {
1653                                         log_err("pthread_create: %s\n",
1654                                                         strerror(ret));
1655                                         nr_started--;
1656                                         break;
1657                                 }
1658                                 ret = pthread_detach(td->thread);
1659                                 if (ret)
1660                                         log_err("pthread_detach: %s",
1661                                                         strerror(ret));
1662                         } else {
1663                                 pid_t pid;
1664                                 dprint(FD_PROCESS, "will fork\n");
1665                                 pid = fork();
1666                                 if (!pid) {
1667                                         int ret = fork_main(shm_id, i);
1668
1669                                         _exit(ret);
1670                                 } else if (i == fio_debug_jobno)
1671                                         *fio_debug_jobp = pid;
1672                         }
1673                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1674                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1675                                 log_err("fio: job startup hung? exiting.\n");
1676                                 fio_terminate_threads(TERMINATE_ALL);
1677                                 fio_abort = 1;
1678                                 nr_started--;
1679                                 break;
1680                         }
1681                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1682                 }
1683
1684                 /*
1685                  * Wait for the started threads to transition to
1686                  * TD_INITIALIZED.
1687                  */
1688                 fio_gettime(&this_start, NULL);
1689                 left = this_jobs;
1690                 while (left && !fio_abort) {
1691                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1692                                 break;
1693
1694                         usleep(100000);
1695
1696                         for (i = 0; i < this_jobs; i++) {
1697                                 td = map[i];
1698                                 if (!td)
1699                                         continue;
1700                                 if (td->runstate == TD_INITIALIZED) {
1701                                         map[i] = NULL;
1702                                         left--;
1703                                 } else if (td->runstate >= TD_EXITED) {
1704                                         map[i] = NULL;
1705                                         left--;
1706                                         todo--;
1707                                         nr_running++; /* work-around... */
1708                                 }
1709                         }
1710                 }
1711
1712                 if (left) {
1713                         log_err("fio: %d jobs failed to start\n", left);
1714                         for (i = 0; i < this_jobs; i++) {
1715                                 td = map[i];
1716                                 if (!td)
1717                                         continue;
1718                                 kill(td->pid, SIGTERM);
1719                         }
1720                         break;
1721                 }
1722
1723                 /*
1724                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1725                  */
1726                 for_each_td(td, i) {
1727                         if (td->runstate != TD_INITIALIZED)
1728                                 continue;
1729
1730                         if (in_ramp_time(td))
1731                                 td_set_runstate(td, TD_RAMP);
1732                         else
1733                                 td_set_runstate(td, TD_RUNNING);
1734                         nr_running++;
1735                         nr_started--;
1736                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1737                         t_rate += td->o.rate[0] + td->o.rate[1];
1738                         todo--;
1739                         fio_mutex_up(td->mutex);
1740                 }
1741
1742                 reap_threads(&nr_running, &t_rate, &m_rate);
1743
1744                 if (todo) {
1745                         if (is_backend)
1746                                 fio_server_idle_loop();
1747                         else
1748                                 usleep(100000);
1749                 }
1750         }
1751
1752         while (nr_running) {
1753                 reap_threads(&nr_running, &t_rate, &m_rate);
1754
1755                 if (is_backend)
1756                         fio_server_idle_loop();
1757                 else
1758                         usleep(10000);
1759         }
1760
1761         update_io_ticks();
1762         fio_unpin_memory();
1763 }
1764
1765 int exec_run(void)
1766 {
1767         if (nr_clients)
1768                 return fio_handle_clients();
1769         if (exec_profile && load_profile(exec_profile))
1770                 return 1;
1771
1772         if (!thread_number)
1773                 return 0;
1774
1775         if (write_bw_log) {
1776                 setup_log(&agg_io_log[DDIR_READ]);
1777                 setup_log(&agg_io_log[DDIR_WRITE]);
1778         }
1779
1780         startup_mutex = fio_mutex_init(0);
1781         if (startup_mutex == NULL)
1782                 return 1;
1783         writeout_mutex = fio_mutex_init(1);
1784         if (writeout_mutex == NULL)
1785                 return 1;
1786
1787         set_genesis_time();
1788         create_disk_util_thread();
1789
1790         cgroup_list = smalloc(sizeof(*cgroup_list));
1791         INIT_FLIST_HEAD(cgroup_list);
1792
1793         run_threads();
1794
1795         if (!fio_abort) {
1796                 show_run_stats();
1797                 if (write_bw_log) {
1798                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1799                         __finish_log(agg_io_log[DDIR_WRITE],
1800                                         "agg-write_bw.log");
1801                 }
1802         }
1803
1804         cgroup_kill(cgroup_list);
1805         sfree(cgroup_list);
1806         sfree(cgroup_mnt);
1807
1808         fio_mutex_remove(startup_mutex);
1809         fio_mutex_remove(writeout_mutex);
1810         return exit_value;
1811 }
1812
1813 void reset_fio_state(void)
1814 {
1815         groupid = 0;
1816         thread_number = 0;
1817         nr_process = 0;
1818         nr_thread = 0;
1819         done_secs = 0;
1820 }
1821
1822 static int endian_check(void)
1823 {
1824         union {
1825                 uint8_t c[8];
1826                 uint64_t v;
1827         } u;
1828         int le = 0, be = 0;
1829
1830         u.v = 0x12;
1831         if (u.c[7] == 0x12)
1832                 be = 1;
1833         else if (u.c[0] == 0x12)
1834                 le = 1;
1835
1836 #if defined(FIO_LITTLE_ENDIAN)
1837         if (be)
1838                 return 1;
1839 #elif defined(FIO_BIG_ENDIAN)
1840         if (le)
1841                 return 1;
1842 #else
1843         return 1;
1844 #endif
1845
1846         if (!le && !be)
1847                 return 1;
1848
1849         return 0;
1850 }
1851
1852 int main(int argc, char *argv[], char *envp[])
1853 {
1854         long ps;
1855
1856         if (endian_check()) {
1857                 log_err("fio: endianness settings appear wrong.\n");
1858                 log_err("fio: please report this to fio@vger.kernel.org\n");
1859                 return 1;
1860         }
1861
1862         arch_init(envp);
1863
1864         sinit();
1865
1866         /*
1867          * We need locale for number printing, if it isn't set then just
1868          * go with the US format.
1869          */
1870         if (!getenv("LC_NUMERIC"))
1871                 setlocale(LC_NUMERIC, "en_US");
1872
1873         ps = sysconf(_SC_PAGESIZE);
1874         if (ps < 0) {
1875                 log_err("Failed to get page size\n");
1876                 return 1;
1877         }
1878
1879         page_size = ps;
1880         page_mask = ps - 1;
1881
1882         fio_keywords_init();
1883
1884         if (parse_options(argc, argv))
1885                 return 1;
1886
1887         return exec_run();
1888 }