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 unsigned int thread_number = 0;
  59 unsigned int nr_process = 0;
  60 unsigned int nr_thread = 0;
  61 int shm_id = 0;
  62 int temp_stall_ts;
  63 unsigned long done_secs = 0;
  64
  65 /*
  66  * Just expose an empty list, if the OS does not support disk util stats
  67  */
  68 #ifndef FIO_HAVE_DISK_UTIL
  69 FLIST_HEAD(disk_list);
  70 #endif
  71
  72 static struct fio_mutex *startup_mutex;
  73 static struct fio_mutex *writeout_mutex;
  74 static volatile int fio_abort;
  75 static int exit_value;
  76 static pthread_t gtod_thread;
  77 static pthread_t disk_util_thread;
  78 static struct flist_head *cgroup_list;
  79 static char *cgroup_mnt;
  80
  81 unsigned long arch_flags = 0;
  82
  83 struct io_log *agg_io_log[2];
  84
  85 #define JOB_START_TIMEOUT       (5 * 1000)
  86
  87 static const char *fio_os_strings[os_nr] = {
  88         "Invalid",
  89         "Linux",
  90         "AIX",
  91         "FreeBSD",
  92         "HP-UX",
  93         "OSX",
  94         "NetBSD",
  95         "Solaris",
  96         "Windows"
  97 };
  98
  99 static const char *fio_arch_strings[arch_nr] = {
 100         "Invalid",
 101         "x86-64",
 102         "x86",
 103         "ppc",
 104         "ia64",
 105         "s390",
 106         "alpha",
 107         "sparc",
 108         "sparc64",
 109         "arm",
 110         "sh",
 111         "hppa",
 112         "generic"
 113 };
 114
 115 const char *fio_get_os_string(int nr)
 116 {
 117         if (nr < os_nr)
 118                 return fio_os_strings[nr];
 119
 120         return NULL;
 121 }
 122
 123 const char *fio_get_arch_string(int nr)
 124 {
 125         if (nr < arch_nr)
 126                 return fio_arch_strings[nr];
 127
 128         return NULL;
 129 }
 130
 131 void td_set_runstate(struct thread_data *td, int runstate)
 132 {
 133         if (td->runstate == runstate)
 134                 return;
 135
 136         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
 137                                                 td->runstate, runstate);
 138         td->runstate = runstate;
 139 }
 140
 141 void fio_terminate_threads(int group_id)
 142 {
 143         struct thread_data *td;
 144         int i;
 145
 146         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
 147
 148         for_each_td(td, i) {
 149                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
 150                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
 151                                                 td->o.name, (int) td->pid);
 152                         td->terminate = 1;
 153                         td->o.start_delay = 0;
 154
 155                         /*
 156                          * if the thread is running, just let it exit
 157                          */
 158                         if (!td->pid)
 159                                 continue;
 160                         else if (td->runstate < TD_RAMP)
 161                                 kill(td->pid, SIGTERM);
 162                         else {
 163                                 struct ioengine_ops *ops = td->io_ops;
 164
 165                                 if (ops && (ops->flags & FIO_SIGTERM))
 166                                         kill(td->pid, SIGTERM);
 167                         }
 168                 }
 169         }
 170 }
 171
 172 static void sig_int(int sig)
 173 {
 174         if (threads) {
 175                 if (is_backend)
 176                         fio_server_got_signal(sig);
 177                 else {
 178                         log_info("\nfio: terminating on signal %d\n", sig);
 179                         fflush(stdout);
 180                         exit_value = 128;
 181                 }
 182
 183                 fio_terminate_threads(TERMINATE_ALL);
 184         }
 185 }
 186
 187 static void *disk_thread_main(void *data)
 188 {
 189         fio_mutex_up(startup_mutex);
 190
 191         while (threads) {
 192                 usleep(DISK_UTIL_MSEC * 1000);
 193                 if (!threads)
 194                         break;
 195                 update_io_ticks();
 196
 197                 if (!is_backend)
 198                         print_thread_status();
 199         }
 200
 201         return NULL;
 202 }
 203
 204 static int create_disk_util_thread(void)
 205 {
 206         int ret;
 207
 208         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 209         if (ret) {
 210                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 211                 return 1;
 212         }
 213
 214         ret = pthread_detach(disk_util_thread);
 215         if (ret) {
 216                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 217                 return 1;
 218         }
 219
 220         dprint(FD_MUTEX, "wait on startup_mutex\n");
 221         fio_mutex_down(startup_mutex);
 222         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 223         return 0;
 224 }
 225
 226 static void set_sig_handlers(void)
 227 {
 228         struct sigaction act;
 229
 230         memset(&act, 0, sizeof(act));
 231         act.sa_handler = sig_int;
 232         act.sa_flags = SA_RESTART;
 233         sigaction(SIGINT, &act, NULL);
 234
 235         memset(&act, 0, sizeof(act));
 236         act.sa_handler = sig_int;
 237         act.sa_flags = SA_RESTART;
 238         sigaction(SIGTERM, &act, NULL);
 239
 240         if (is_backend) {
 241                 memset(&act, 0, sizeof(act));
 242                 act.sa_handler = sig_int;
 243                 act.sa_flags = SA_RESTART;
 244                 sigaction(SIGPIPE, &act, NULL);
 245         }
 246 }
 247
 248 /*
 249  * Check if we are above the minimum rate given.
 250  */
 251 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 252                             enum fio_ddir ddir)
 253 {
 254         unsigned long long bytes = 0;
 255         unsigned long iops = 0;
 256         unsigned long spent;
 257         unsigned long rate;
 258         unsigned int ratemin = 0;
 259         unsigned int rate_iops = 0;
 260         unsigned int rate_iops_min = 0;
 261
 262         assert(ddir_rw(ddir));
 263
 264         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 265                 return 0;
 266
 267         /*
 268          * allow a 2 second settle period in the beginning
 269          */
 270         if (mtime_since(&td->start, now) < 2000)
 271                 return 0;
 272
 273         iops += td->this_io_blocks[ddir];
 274         bytes += td->this_io_bytes[ddir];
 275         ratemin += td->o.ratemin[ddir];
 276         rate_iops += td->o.rate_iops[ddir];
 277         rate_iops_min += td->o.rate_iops_min[ddir];
 278
 279         /*
 280          * if rate blocks is set, sample is running
 281          */
 282         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 283                 spent = mtime_since(&td->lastrate[ddir], now);
 284                 if (spent < td->o.ratecycle)
 285                         return 0;
 286
 287                 if (td->o.rate[ddir]) {
 288                         /*
 289                          * check bandwidth specified rate
 290                          */
 291                         if (bytes < td->rate_bytes[ddir]) {
 292                                 log_err("%s: min rate %u not met\n", td->o.name,
 293                                                                 ratemin);
 294                                 return 1;
 295                         } else {
 296                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 297                                 if (rate < ratemin ||
 298                                     bytes < td->rate_bytes[ddir]) {
 299                                         log_err("%s: min rate %u not met, got"
 300                                                 " %luKB/sec\n", td->o.name,
 301                                                         ratemin, rate);
 302                                         return 1;
 303                                 }
 304                         }
 305                 } else {
 306                         /*
 307                          * checks iops specified rate
 308                          */
 309                         if (iops < rate_iops) {
 310                                 log_err("%s: min iops rate %u not met\n",
 311                                                 td->o.name, rate_iops);
 312                                 return 1;
 313                         } else {
 314                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 315                                 if (rate < rate_iops_min ||
 316                                     iops < td->rate_blocks[ddir]) {
 317                                         log_err("%s: min iops rate %u not met,"
 318                                                 " got %lu\n", td->o.name,
 319                                                         rate_iops_min, rate);
 320                                 }
 321                         }
 322                 }
 323         }
 324
 325         td->rate_bytes[ddir] = bytes;
 326         td->rate_blocks[ddir] = iops;
 327         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 328         return 0;
 329 }
 330
 331 static int check_min_rate(struct thread_data *td, struct timeval *now,
 332                           unsigned long *bytes_done)
 333 {
 334         int ret = 0;
 335
 336         if (bytes_done[0])
 337                 ret |= __check_min_rate(td, now, 0);
 338         if (bytes_done[1])
 339                 ret |= __check_min_rate(td, now, 1);
 340
 341         return ret;
 342 }
 343
 344 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 345 {
 346         if (in_ramp_time(td))
 347                 return 0;
 348         if (!td->o.timeout)
 349                 return 0;
 350         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 351                 return 1;
 352
 353         return 0;
 354 }
 355
 356 /*
 357  * When job exits, we can cancel the in-flight IO if we are using async
 358  * io. Attempt to do so.
 359  */
 360 static void cleanup_pending_aio(struct thread_data *td)
 361 {
 362         struct flist_head *entry, *n;
 363         struct io_u *io_u;
 364         int r;
 365
 366         /*
 367          * get immediately available events, if any
 368          */
 369         r = io_u_queued_complete(td, 0, NULL);
 370         if (r < 0)
 371                 return;
 372
 373         /*
 374          * now cancel remaining active events
 375          */
 376         if (td->io_ops->cancel) {
 377                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 378                         io_u = flist_entry(entry, struct io_u, list);
 379
 380                         /*
 381                          * if the io_u isn't in flight, then that generally
 382                          * means someone leaked an io_u. complain but fix
 383                          * it up, so we don't stall here.
 384                          */
 385                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 386                                 log_err("fio: non-busy IO on busy list\n");
 387                                 put_io_u(td, io_u);
 388                         } else {
 389                                 r = td->io_ops->cancel(td, io_u);
 390                                 if (!r)
 391                                         put_io_u(td, io_u);
 392                         }
 393                 }
 394         }
 395
 396         if (td->cur_depth)
 397                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 398 }
 399
 400 /*
 401  * Helper to handle the final sync of a file. Works just like the normal
 402  * io path, just does everything sync.
 403  */
 404 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 405 {
 406         struct io_u *io_u = __get_io_u(td);
 407         int ret;
 408
 409         if (!io_u)
 410                 return 1;
 411
 412         io_u->ddir = DDIR_SYNC;
 413         io_u->file = f;
 414
 415         if (td_io_prep(td, io_u)) {
 416                 put_io_u(td, io_u);
 417                 return 1;
 418         }
 419
 420 requeue:
 421         ret = td_io_queue(td, io_u);
 422         if (ret < 0) {
 423                 td_verror(td, io_u->error, "td_io_queue");
 424                 put_io_u(td, io_u);
 425                 return 1;
 426         } else if (ret == FIO_Q_QUEUED) {
 427                 if (io_u_queued_complete(td, 1, NULL) < 0)
 428                         return 1;
 429         } else if (ret == FIO_Q_COMPLETED) {
 430                 if (io_u->error) {
 431                         td_verror(td, io_u->error, "td_io_queue");
 432                         return 1;
 433                 }
 434
 435                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 436                         return 1;
 437         } else if (ret == FIO_Q_BUSY) {
 438                 if (td_io_commit(td))
 439                         return 1;
 440                 goto requeue;
 441         }
 442
 443         return 0;
 444 }
 445
 446 static inline void __update_tv_cache(struct thread_data *td)
 447 {
 448         fio_gettime(&td->tv_cache, NULL);
 449 }
 450
 451 static inline void update_tv_cache(struct thread_data *td)
 452 {
 453         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 454                 __update_tv_cache(td);
 455 }
 456
 457 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 458                                int *retptr)
 459 {
 460         int ret = *retptr;
 461
 462         if (ret < 0 || td->error) {
 463                 int err;
 464
 465                 if (ret < 0)
 466                         err = -ret;
 467                 else
 468                         err = td->error;
 469
 470                 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
 471                         return 1;
 472
 473                 if (td_non_fatal_error(err)) {
 474                         /*
 475                          * Continue with the I/Os in case of
 476                          * a non fatal error.
 477                          */
 478                         update_error_count(td, err);
 479                         td_clear_error(td);
 480                         *retptr = 0;
 481                         return 0;
 482                 } else if (td->o.fill_device && err == ENOSPC) {
 483                         /*
 484                          * We expect to hit this error if
 485                          * fill_device option is set.
 486                          */
 487                         td_clear_error(td);
 488                         td->terminate = 1;
 489                         return 1;
 490                 } else {
 491                         /*
 492                          * Stop the I/O in case of a fatal
 493                          * error.
 494                          */
 495                         update_error_count(td, err);
 496                         return 1;
 497                 }
 498         }
 499
 500         return 0;
 501 }
 502
 503 /*
 504  * The main verify engine. Runs over the writes we previously submitted,
 505  * reads the blocks back in, and checks the crc/md5 of the data.
 506  */
 507 static void do_verify(struct thread_data *td)
 508 {
 509         struct fio_file *f;
 510         struct io_u *io_u;
 511         int ret, min_events;
 512         unsigned int i;
 513
 514         dprint(FD_VERIFY, "starting loop\n");
 515
 516         /*
 517          * sync io first and invalidate cache, to make sure we really
 518          * read from disk.
 519          */
 520         for_each_file(td, f, i) {
 521                 if (!fio_file_open(f))
 522                         continue;
 523                 if (fio_io_sync(td, f))
 524                         break;
 525                 if (file_invalidate_cache(td, f))
 526                         break;
 527         }
 528
 529         if (td->error)
 530                 return;
 531
 532         td_set_runstate(td, TD_VERIFYING);
 533
 534         io_u = NULL;
 535         while (!td->terminate) {
 536                 int ret2, full;
 537
 538                 update_tv_cache(td);
 539
 540                 if (runtime_exceeded(td, &td->tv_cache)) {
 541                         __update_tv_cache(td);
 542                         if (runtime_exceeded(td, &td->tv_cache)) {
 543                                 td->terminate = 1;
 544                                 break;
 545                         }
 546                 }
 547
 548                 io_u = __get_io_u(td);
 549                 if (!io_u)
 550                         break;
 551
 552                 if (get_next_verify(td, io_u)) {
 553                         put_io_u(td, io_u);
 554                         break;
 555                 }
 556
 557                 if (td_io_prep(td, io_u)) {
 558                         put_io_u(td, io_u);
 559                         break;
 560                 }
 561
 562                 if (td->o.verify_async)
 563                         io_u->end_io = verify_io_u_async;
 564                 else
 565                         io_u->end_io = verify_io_u;
 566
 567                 ret = td_io_queue(td, io_u);
 568                 switch (ret) {
 569                 case FIO_Q_COMPLETED:
 570                         if (io_u->error) {
 571                                 ret = -io_u->error;
 572                                 clear_io_u(td, io_u);
 573                         } else if (io_u->resid) {
 574                                 int bytes = io_u->xfer_buflen - io_u->resid;
 575
 576                                 /*
 577                                  * zero read, fail
 578                                  */
 579                                 if (!bytes) {
 580                                         td_verror(td, EIO, "full resid");
 581                                         put_io_u(td, io_u);
 582                                         break;
 583                                 }
 584
 585                                 io_u->xfer_buflen = io_u->resid;
 586                                 io_u->xfer_buf += bytes;
 587                                 io_u->offset += bytes;
 588
 589                                 if (ddir_rw(io_u->ddir))
 590                                         td->ts.short_io_u[io_u->ddir]++;
 591
 592                                 f = io_u->file;
 593                                 if (io_u->offset == f->real_file_size)
 594                                         goto sync_done;
 595
 596                                 requeue_io_u(td, &io_u);
 597                         } else {
 598 sync_done:
 599                                 ret = io_u_sync_complete(td, io_u, NULL);
 600                                 if (ret < 0)
 601                                         break;
 602                         }
 603                         continue;
 604                 case FIO_Q_QUEUED:
 605                         break;
 606                 case FIO_Q_BUSY:
 607                         requeue_io_u(td, &io_u);
 608                         ret2 = td_io_commit(td);
 609                         if (ret2 < 0)
 610                                 ret = ret2;
 611                         break;
 612                 default:
 613                         assert(ret < 0);
 614                         td_verror(td, -ret, "td_io_queue");
 615                         break;
 616                 }
 617
 618                 if (break_on_this_error(td, io_u->ddir, &ret))
 619                         break;
 620
 621                 /*
 622                  * if we can queue more, do so. but check if there are
 623                  * completed io_u's first. Note that we can get BUSY even
 624                  * without IO queued, if the system is resource starved.
 625                  */
 626                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 627                 if (full || !td->o.iodepth_batch_complete) {
 628                         min_events = min(td->o.iodepth_batch_complete,
 629                                          td->cur_depth);
 630                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 631                                 min_events = 1;
 632
 633                         do {
 634                                 /*
 635                                  * Reap required number of io units, if any,
 636                                  * and do the verification on them through
 637                                  * the callback handler
 638                                  */
 639                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 640                                         ret = -1;
 641                                         break;
 642                                 }
 643                         } while (full && (td->cur_depth > td->o.iodepth_low));
 644                 }
 645                 if (ret < 0)
 646                         break;
 647         }
 648
 649         if (!td->error) {
 650                 min_events = td->cur_depth;
 651
 652                 if (min_events)
 653                         ret = io_u_queued_complete(td, min_events, NULL);
 654         } else
 655                 cleanup_pending_aio(td);
 656
 657         td_set_runstate(td, TD_RUNNING);
 658
 659         dprint(FD_VERIFY, "exiting loop\n");
 660 }
 661
 662 /*
 663  * Main IO worker function. It retrieves io_u's to process and queues
 664  * and reaps them, checking for rate and errors along the way.
 665  */
 666 static void do_io(struct thread_data *td)
 667 {
 668         unsigned int i;
 669         int ret = 0;
 670
 671         if (in_ramp_time(td))
 672                 td_set_runstate(td, TD_RAMP);
 673         else
 674                 td_set_runstate(td, TD_RUNNING);
 675
 676         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 677                 (!flist_empty(&td->trim_list)) ||
 678                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 679                 struct timeval comp_time;
 680                 unsigned long bytes_done[2] = { 0, 0 };
 681                 int min_evts = 0;
 682                 struct io_u *io_u;
 683                 int ret2, full;
 684                 enum fio_ddir ddir;
 685
 686                 if (td->terminate)
 687                         break;
 688
 689                 update_tv_cache(td);
 690
 691                 if (runtime_exceeded(td, &td->tv_cache)) {
 692                         __update_tv_cache(td);
 693                         if (runtime_exceeded(td, &td->tv_cache)) {
 694                                 td->terminate = 1;
 695                                 break;
 696                         }
 697                 }
 698
 699                 io_u = get_io_u(td);
 700                 if (!io_u)
 701                         break;
 702
 703                 ddir = io_u->ddir;
 704
 705                 /*
 706                  * Add verification end_io handler, if asked to verify
 707                  * a previously written file.
 708                  */
 709                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 710                     !td_rw(td)) {
 711                         if (td->o.verify_async)
 712                                 io_u->end_io = verify_io_u_async;
 713                         else
 714                                 io_u->end_io = verify_io_u;
 715                         td_set_runstate(td, TD_VERIFYING);
 716                 } else if (in_ramp_time(td))
 717                         td_set_runstate(td, TD_RAMP);
 718                 else
 719                         td_set_runstate(td, TD_RUNNING);
 720
 721                 ret = td_io_queue(td, io_u);
 722                 switch (ret) {
 723                 case FIO_Q_COMPLETED:
 724                         if (io_u->error) {
 725                                 ret = -io_u->error;
 726                                 clear_io_u(td, io_u);
 727                         } else if (io_u->resid) {
 728                                 int bytes = io_u->xfer_buflen - io_u->resid;
 729                                 struct fio_file *f = io_u->file;
 730
 731                                 /*
 732                                  * zero read, fail
 733                                  */
 734                                 if (!bytes) {
 735                                         td_verror(td, EIO, "full resid");
 736                                         put_io_u(td, io_u);
 737                                         break;
 738                                 }
 739
 740                                 io_u->xfer_buflen = io_u->resid;
 741                                 io_u->xfer_buf += bytes;
 742                                 io_u->offset += bytes;
 743
 744                                 if (ddir_rw(io_u->ddir))
 745                                         td->ts.short_io_u[io_u->ddir]++;
 746
 747                                 if (io_u->offset == f->real_file_size)
 748                                         goto sync_done;
 749
 750                                 requeue_io_u(td, &io_u);
 751                         } else {
 752 sync_done:
 753                                 if (__should_check_rate(td, 0) ||
 754                                     __should_check_rate(td, 1))
 755                                         fio_gettime(&comp_time, NULL);
 756
 757                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 758                                 if (ret < 0)
 759                                         break;
 760                         }
 761                         break;
 762                 case FIO_Q_QUEUED:
 763                         /*
 764                          * if the engine doesn't have a commit hook,
 765                          * the io_u is really queued. if it does have such
 766                          * a hook, it has to call io_u_queued() itself.
 767                          */
 768                         if (td->io_ops->commit == NULL)
 769                                 io_u_queued(td, io_u);
 770                         break;
 771                 case FIO_Q_BUSY:
 772                         requeue_io_u(td, &io_u);
 773                         ret2 = td_io_commit(td);
 774                         if (ret2 < 0)
 775                                 ret = ret2;
 776                         break;
 777                 default:
 778                         assert(ret < 0);
 779                         put_io_u(td, io_u);
 780                         break;
 781                 }
 782
 783                 if (break_on_this_error(td, ddir, &ret))
 784                         break;
 785
 786                 /*
 787                  * See if we need to complete some commands. Note that we
 788                  * can get BUSY even without IO queued, if the system is
 789                  * resource starved.
 790                  */
 791                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 792                 if (full || !td->o.iodepth_batch_complete) {
 793                         min_evts = min(td->o.iodepth_batch_complete,
 794                                         td->cur_depth);
 795                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 796                                 min_evts = 1;
 797
 798                         if (__should_check_rate(td, 0) ||
 799                             __should_check_rate(td, 1))
 800                                 fio_gettime(&comp_time, NULL);
 801
 802                         do {
 803                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 804                                 if (ret < 0)
 805                                         break;
 806
 807                         } while (full && (td->cur_depth > td->o.iodepth_low));
 808                 }
 809
 810                 if (ret < 0)
 811                         break;
 812                 if (!(bytes_done[0] + bytes_done[1]))
 813                         continue;
 814
 815                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 816                         if (check_min_rate(td, &comp_time, bytes_done)) {
 817                                 if (exitall_on_terminate)
 818                                         fio_terminate_threads(td->groupid);
 819                                 td_verror(td, EIO, "check_min_rate");
 820                                 break;
 821                         }
 822                 }
 823
 824                 if (td->o.thinktime) {
 825                         unsigned long long b;
 826
 827                         b = td->io_blocks[0] + td->io_blocks[1];
 828                         if (!(b % td->o.thinktime_blocks)) {
 829                                 int left;
 830
 831                                 if (td->o.thinktime_spin)
 832                                         usec_spin(td->o.thinktime_spin);
 833
 834                                 left = td->o.thinktime - td->o.thinktime_spin;
 835                                 if (left)
 836                                         usec_sleep(td, left);
 837                         }
 838                 }
 839         }
 840
 841         if (td->trim_entries)
 842                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 843
 844         if (td->o.fill_device && td->error == ENOSPC) {
 845                 td->error = 0;
 846                 td->terminate = 1;
 847         }
 848         if (!td->error) {
 849                 struct fio_file *f;
 850
 851                 i = td->cur_depth;
 852                 if (i) {
 853                         ret = io_u_queued_complete(td, i, NULL);
 854                         if (td->o.fill_device && td->error == ENOSPC)
 855                                 td->error = 0;
 856                 }
 857
 858                 if (should_fsync(td) && td->o.end_fsync) {
 859                         td_set_runstate(td, TD_FSYNCING);
 860
 861                         for_each_file(td, f, i) {
 862                                 if (!fio_file_open(f))
 863                                         continue;
 864                                 fio_io_sync(td, f);
 865                         }
 866                 }
 867         } else
 868                 cleanup_pending_aio(td);
 869
 870         /*
 871          * stop job if we failed doing any IO
 872          */
 873         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 874                 td->done = 1;
 875 }
 876
 877 static void cleanup_io_u(struct thread_data *td)
 878 {
 879         struct flist_head *entry, *n;
 880         struct io_u *io_u;
 881
 882         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 883                 io_u = flist_entry(entry, struct io_u, list);
 884
 885                 flist_del(&io_u->list);
 886                 fio_memfree(io_u, sizeof(*io_u));
 887         }
 888
 889         free_io_mem(td);
 890 }
 891
 892 static int init_io_u(struct thread_data *td)
 893 {
 894         struct io_u *io_u;
 895         unsigned int max_bs;
 896         int cl_align, i, max_units;
 897         char *p;
 898
 899         max_units = td->o.iodepth;
 900         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 901         td->orig_buffer_size = (unsigned long long) max_bs
 902                                         * (unsigned long long) max_units;
 903
 904         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 905                 unsigned long bs;
 906
 907                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 908                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 909         }
 910
 911         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 912                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 913                 return 1;
 914         }
 915
 916         if (allocate_io_mem(td))
 917                 return 1;
 918
 919         if (td->o.odirect || td->o.mem_align ||
 920             (td->io_ops->flags & FIO_RAWIO))
 921                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 922         else
 923                 p = td->orig_buffer;
 924
 925         cl_align = os_cache_line_size();
 926
 927         for (i = 0; i < max_units; i++) {
 928                 void *ptr;
 929
 930                 if (td->terminate)
 931                         return 1;
 932
 933                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 934                 if (!ptr) {
 935                         log_err("fio: unable to allocate aligned memory\n");
 936                         break;
 937                 }
 938
 939                 io_u = ptr;
 940                 memset(io_u, 0, sizeof(*io_u));
 941                 INIT_FLIST_HEAD(&io_u->list);
 942                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 943
 944                 if (!(td->io_ops->flags & FIO_NOIO)) {
 945                         io_u->buf = p;
 946                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 947
 948                         if (td_write(td))
 949                                 io_u_fill_buffer(td, io_u, max_bs);
 950                         if (td_write(td) && td->o.verify_pattern_bytes) {
 951                                 /*
 952                                  * Fill the buffer with the pattern if we are
 953                                  * going to be doing writes.
 954                                  */
 955                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 956                         }
 957                 }
 958
 959                 io_u->index = i;
 960                 io_u->flags = IO_U_F_FREE;
 961                 flist_add(&io_u->list, &td->io_u_freelist);
 962                 p += max_bs;
 963         }
 964
 965         return 0;
 966 }
 967
 968 static int switch_ioscheduler(struct thread_data *td)
 969 {
 970         char tmp[256], tmp2[128];
 971         FILE *f;
 972         int ret;
 973
 974         if (td->io_ops->flags & FIO_DISKLESSIO)
 975                 return 0;
 976
 977         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 978
 979         f = fopen(tmp, "r+");
 980         if (!f) {
 981                 if (errno == ENOENT) {
 982                         log_err("fio: os or kernel doesn't support IO scheduler"
 983                                 " switching\n");
 984                         return 0;
 985                 }
 986                 td_verror(td, errno, "fopen iosched");
 987                 return 1;
 988         }
 989
 990         /*
 991          * Set io scheduler.
 992          */
 993         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 994         if (ferror(f) || ret != 1) {
 995                 td_verror(td, errno, "fwrite");
 996                 fclose(f);
 997                 return 1;
 998         }
 999
1000         rewind(f);
1001
1002         /*
1003          * Read back and check that the selected scheduler is now the default.
1004          */
1005         ret = fread(tmp, 1, sizeof(tmp), f);
1006         if (ferror(f) || ret < 0) {
1007                 td_verror(td, errno, "fread");
1008                 fclose(f);
1009                 return 1;
1010         }
1011
1012         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1013         if (!strstr(tmp, tmp2)) {
1014                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1015                 td_verror(td, EINVAL, "iosched_switch");
1016                 fclose(f);
1017                 return 1;
1018         }
1019
1020         fclose(f);
1021         return 0;
1022 }
1023
1024 static int keep_running(struct thread_data *td)
1025 {
1026         unsigned long long io_done;
1027
1028         if (td->done)
1029                 return 0;
1030         if (td->o.time_based)
1031                 return 1;
1032         if (td->o.loops) {
1033                 td->o.loops--;
1034                 return 1;
1035         }
1036
1037         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1038                         + td->io_skip_bytes;
1039         if (io_done < td->o.size)
1040                 return 1;
1041
1042         return 0;
1043 }
1044
1045 static void reset_io_counters(struct thread_data *td)
1046 {
1047         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1048         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1049         td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1050         td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1051         td->zone_bytes = 0;
1052         td->rate_bytes[0] = td->rate_bytes[1] = 0;
1053         td->rate_blocks[0] = td->rate_blocks[1] = 0;
1054
1055         td->last_was_sync = 0;
1056
1057         /*
1058          * reset file done count if we are to start over
1059          */
1060         if (td->o.time_based || td->o.loops)
1061                 td->nr_done_files = 0;
1062 }
1063
1064 void reset_all_stats(struct thread_data *td)
1065 {
1066         struct timeval tv;
1067         int i;
1068
1069         reset_io_counters(td);
1070
1071         for (i = 0; i < 2; i++) {
1072                 td->io_bytes[i] = 0;
1073                 td->io_blocks[i] = 0;
1074                 td->io_issues[i] = 0;
1075                 td->ts.total_io_u[i] = 0;
1076         }
1077
1078         fio_gettime(&tv, NULL);
1079         td->ts.runtime[0] = 0;
1080         td->ts.runtime[1] = 0;
1081         memcpy(&td->epoch, &tv, sizeof(tv));
1082         memcpy(&td->start, &tv, sizeof(tv));
1083 }
1084
1085 static void clear_io_state(struct thread_data *td)
1086 {
1087         struct fio_file *f;
1088         unsigned int i;
1089
1090         reset_io_counters(td);
1091
1092         close_files(td);
1093         for_each_file(td, f, i)
1094                 fio_file_clear_done(f);
1095
1096         /*
1097          * Set the same seed to get repeatable runs
1098          */
1099         td_fill_rand_seeds(td);
1100 }
1101
1102 static int exec_string(const char *string)
1103 {
1104         int ret, newlen = strlen(string) + 1 + 8;
1105         char *str;
1106
1107         str = malloc(newlen);
1108         sprintf(str, "sh -c %s", string);
1109
1110         ret = system(str);
1111         if (ret == -1)
1112                 log_err("fio: exec of cmd <%s> failed\n", str);
1113
1114         free(str);
1115         return ret;
1116 }
1117
1118 /*
1119  * Entry point for the thread based jobs. The process based jobs end up
1120  * here as well, after a little setup.
1121  */
1122 static void *thread_main(void *data)
1123 {
1124         unsigned long long elapsed;
1125         struct thread_data *td = data;
1126         pthread_condattr_t attr;
1127         int clear_state;
1128
1129         if (!td->o.use_thread) {
1130                 setsid();
1131                 td->pid = getpid();
1132         } else
1133                 td->pid = gettid();
1134
1135         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1136
1137         INIT_FLIST_HEAD(&td->io_u_freelist);
1138         INIT_FLIST_HEAD(&td->io_u_busylist);
1139         INIT_FLIST_HEAD(&td->io_u_requeues);
1140         INIT_FLIST_HEAD(&td->io_log_list);
1141         INIT_FLIST_HEAD(&td->io_hist_list);
1142         INIT_FLIST_HEAD(&td->verify_list);
1143         INIT_FLIST_HEAD(&td->trim_list);
1144         pthread_mutex_init(&td->io_u_lock, NULL);
1145         td->io_hist_tree = RB_ROOT;
1146
1147         pthread_condattr_init(&attr);
1148         pthread_cond_init(&td->verify_cond, &attr);
1149         pthread_cond_init(&td->free_cond, &attr);
1150
1151         td_set_runstate(td, TD_INITIALIZED);
1152         dprint(FD_MUTEX, "up startup_mutex\n");
1153         fio_mutex_up(startup_mutex);
1154         dprint(FD_MUTEX, "wait on td->mutex\n");
1155         fio_mutex_down(td->mutex);
1156         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1157
1158         /*
1159          * the ->mutex mutex is now no longer used, close it to avoid
1160          * eating a file descriptor
1161          */
1162         fio_mutex_remove(td->mutex);
1163
1164         /*
1165          * A new gid requires privilege, so we need to do this before setting
1166          * the uid.
1167          */
1168         if (td->o.gid != -1U && setgid(td->o.gid)) {
1169                 td_verror(td, errno, "setgid");
1170                 goto err;
1171         }
1172         if (td->o.uid != -1U && setuid(td->o.uid)) {
1173                 td_verror(td, errno, "setuid");
1174                 goto err;
1175         }
1176
1177         /*
1178          * If we have a gettimeofday() thread, make sure we exclude that
1179          * thread from this job
1180          */
1181         if (td->o.gtod_cpu)
1182                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1183
1184         /*
1185          * Set affinity first, in case it has an impact on the memory
1186          * allocations.
1187          */
1188         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1189                 td_verror(td, errno, "cpu_set_affinity");
1190                 goto err;
1191         }
1192
1193         /*
1194          * May alter parameters that init_io_u() will use, so we need to
1195          * do this first.
1196          */
1197         if (init_iolog(td))
1198                 goto err;
1199
1200         if (init_io_u(td))
1201                 goto err;
1202
1203         if (td->o.verify_async && verify_async_init(td))
1204                 goto err;
1205
1206         if (td->ioprio_set) {
1207                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1208                         td_verror(td, errno, "ioprio_set");
1209                         goto err;
1210                 }
1211         }
1212
1213         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1214                 goto err;
1215
1216         if (nice(td->o.nice) == -1) {
1217                 td_verror(td, errno, "nice");
1218                 goto err;
1219         }
1220
1221         if (td->o.ioscheduler && switch_ioscheduler(td))
1222                 goto err;
1223
1224         if (!td->o.create_serialize && setup_files(td))
1225                 goto err;
1226
1227         if (td_io_init(td))
1228                 goto err;
1229
1230         if (init_random_map(td))
1231                 goto err;
1232
1233         if (td->o.exec_prerun) {
1234                 if (exec_string(td->o.exec_prerun))
1235                         goto err;
1236         }
1237
1238         if (td->o.pre_read) {
1239                 if (pre_read_files(td) < 0)
1240                         goto err;
1241         }
1242
1243         fio_gettime(&td->epoch, NULL);
1244         getrusage(RUSAGE_SELF, &td->ru_start);
1245
1246         clear_state = 0;
1247         while (keep_running(td)) {
1248                 fio_gettime(&td->start, NULL);
1249                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1250                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1251                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1252
1253                 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1254                         memcpy(&td->lastrate[0], &td->bw_sample_time,
1255                                                 sizeof(td->bw_sample_time));
1256                         memcpy(&td->lastrate[1], &td->bw_sample_time,
1257                                                 sizeof(td->bw_sample_time));
1258                 }
1259
1260                 if (clear_state)
1261                         clear_io_state(td);
1262
1263                 prune_io_piece_log(td);
1264
1265                 do_io(td);
1266
1267                 clear_state = 1;
1268
1269                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1270                         elapsed = utime_since_now(&td->start);
1271                         td->ts.runtime[DDIR_READ] += elapsed;
1272                 }
1273                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1274                         elapsed = utime_since_now(&td->start);
1275                         td->ts.runtime[DDIR_WRITE] += elapsed;
1276                 }
1277
1278                 if (td->error || td->terminate)
1279                         break;
1280
1281                 if (!td->o.do_verify ||
1282                     td->o.verify == VERIFY_NONE ||
1283                     (td->io_ops->flags & FIO_UNIDIR))
1284                         continue;
1285
1286                 clear_io_state(td);
1287
1288                 fio_gettime(&td->start, NULL);
1289
1290                 do_verify(td);
1291
1292                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1293
1294                 if (td->error || td->terminate)
1295                         break;
1296         }
1297
1298         update_rusage_stat(td);
1299         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1300         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1301         td->ts.total_run_time = mtime_since_now(&td->epoch);
1302         td->ts.io_bytes[0] = td->io_bytes[0];
1303         td->ts.io_bytes[1] = td->io_bytes[1];
1304
1305         fio_mutex_down(writeout_mutex);
1306         if (td->bw_log) {
1307                 if (td->o.bw_log_file) {
1308                         finish_log_named(td, td->bw_log,
1309                                                 td->o.bw_log_file, "bw");
1310                 } else
1311                         finish_log(td, td->bw_log, "bw");
1312         }
1313         if (td->lat_log) {
1314                 if (td->o.lat_log_file) {
1315                         finish_log_named(td, td->lat_log,
1316                                                 td->o.lat_log_file, "lat");
1317                 } else
1318                         finish_log(td, td->lat_log, "lat");
1319         }
1320         if (td->slat_log) {
1321                 if (td->o.lat_log_file) {
1322                         finish_log_named(td, td->slat_log,
1323                                                 td->o.lat_log_file, "slat");
1324                 } else
1325                         finish_log(td, td->slat_log, "slat");
1326         }
1327         if (td->clat_log) {
1328                 if (td->o.lat_log_file) {
1329                         finish_log_named(td, td->clat_log,
1330                                                 td->o.lat_log_file, "clat");
1331                 } else
1332                         finish_log(td, td->clat_log, "clat");
1333         }
1334         if (td->iops_log) {
1335                 if (td->o.iops_log_file) {
1336                         finish_log_named(td, td->iops_log,
1337                                                 td->o.iops_log_file, "iops");
1338                 } else
1339                         finish_log(td, td->iops_log, "iops");
1340         }
1341
1342         fio_mutex_up(writeout_mutex);
1343         if (td->o.exec_postrun)
1344                 exec_string(td->o.exec_postrun);
1345
1346         if (exitall_on_terminate)
1347                 fio_terminate_threads(td->groupid);
1348
1349 err:
1350         if (td->error)
1351                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1352                                                         td->verror);
1353
1354         if (td->o.verify_async)
1355                 verify_async_exit(td);
1356
1357         close_and_free_files(td);
1358         close_ioengine(td);
1359         cleanup_io_u(td);
1360         cgroup_shutdown(td, &cgroup_mnt);
1361
1362         if (td->o.cpumask_set) {
1363                 int ret = fio_cpuset_exit(&td->o.cpumask);
1364
1365                 td_verror(td, ret, "fio_cpuset_exit");
1366         }
1367
1368         /*
1369          * do this very late, it will log file closing as well
1370          */
1371         if (td->o.write_iolog_file)
1372                 write_iolog_close(td);
1373
1374         td_set_runstate(td, TD_EXITED);
1375         return (void *) (unsigned long) td->error;
1376 }
1377
1378 /*
1379  * We cannot pass the td data into a forked process, so attach the td and
1380  * pass it to the thread worker.
1381  */
1382 static int fork_main(int shmid, int offset)
1383 {
1384         struct thread_data *td;
1385         void *data, *ret;
1386
1387 #ifndef __hpux
1388         data = shmat(shmid, NULL, 0);
1389         if (data == (void *) -1) {
1390                 int __err = errno;
1391
1392                 perror("shmat");
1393                 return __err;
1394         }
1395 #else
1396         /*
1397          * HP-UX inherits shm mappings?
1398          */
1399         data = threads;
1400 #endif
1401
1402         td = data + offset * sizeof(struct thread_data);
1403         ret = thread_main(td);
1404         shmdt(data);
1405         return (int) (unsigned long) ret;
1406 }
1407
1408 /*
1409  * Run over the job map and reap the threads that have exited, if any.
1410  */
1411 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1412                          unsigned int *m_rate)
1413 {
1414         struct thread_data *td;
1415         unsigned int cputhreads, realthreads, pending;
1416         int i, status, ret;
1417
1418         /*
1419          * reap exited threads (TD_EXITED -> TD_REAPED)
1420          */
1421         realthreads = pending = cputhreads = 0;
1422         for_each_td(td, i) {
1423                 int flags = 0;
1424
1425                 /*
1426                  * ->io_ops is NULL for a thread that has closed its
1427                  * io engine
1428                  */
1429                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1430                         cputhreads++;
1431                 else
1432                         realthreads++;
1433
1434                 if (!td->pid) {
1435                         pending++;
1436                         continue;
1437                 }
1438                 if (td->runstate == TD_REAPED)
1439                         continue;
1440                 if (td->o.use_thread) {
1441                         if (td->runstate == TD_EXITED) {
1442                                 td_set_runstate(td, TD_REAPED);
1443                                 goto reaped;
1444                         }
1445                         continue;
1446                 }
1447
1448                 flags = WNOHANG;
1449                 if (td->runstate == TD_EXITED)
1450                         flags = 0;
1451
1452                 /*
1453                  * check if someone quit or got killed in an unusual way
1454                  */
1455                 ret = waitpid(td->pid, &status, flags);
1456                 if (ret < 0) {
1457                         if (errno == ECHILD) {
1458                                 log_err("fio: pid=%d disappeared %d\n",
1459                                                 (int) td->pid, td->runstate);
1460                                 td_set_runstate(td, TD_REAPED);
1461                                 goto reaped;
1462                         }
1463                         perror("waitpid");
1464                 } else if (ret == td->pid) {
1465                         if (WIFSIGNALED(status)) {
1466                                 int sig = WTERMSIG(status);
1467
1468                                 if (sig != SIGTERM)
1469                                         log_err("fio: pid=%d, got signal=%d\n",
1470                                                         (int) td->pid, sig);
1471                                 td_set_runstate(td, TD_REAPED);
1472                                 goto reaped;
1473                         }
1474                         if (WIFEXITED(status)) {
1475                                 if (WEXITSTATUS(status) && !td->error)
1476                                         td->error = WEXITSTATUS(status);
1477
1478                                 td_set_runstate(td, TD_REAPED);
1479                                 goto reaped;
1480                         }
1481                 }
1482
1483                 /*
1484                  * thread is not dead, continue
1485                  */
1486                 pending++;
1487                 continue;
1488 reaped:
1489                 (*nr_running)--;
1490                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1491                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1492                 if (!td->pid)
1493                         pending--;
1494
1495                 if (td->error)
1496                         exit_value++;
1497
1498                 done_secs += mtime_since_now(&td->epoch) / 1000;
1499         }
1500
1501         if (*nr_running == cputhreads && !pending && realthreads)
1502                 fio_terminate_threads(TERMINATE_ALL);
1503 }
1504
1505 static void *gtod_thread_main(void *data)
1506 {
1507         fio_mutex_up(startup_mutex);
1508
1509         /*
1510          * As long as we have jobs around, update the clock. It would be nice
1511          * to have some way of NOT hammering that CPU with gettimeofday(),
1512          * but I'm not sure what to use outside of a simple CPU nop to relax
1513          * it - we don't want to lose precision.
1514          */
1515         while (threads) {
1516                 fio_gtod_update();
1517                 nop;
1518         }
1519
1520         return NULL;
1521 }
1522
1523 static int fio_start_gtod_thread(void)
1524 {
1525         pthread_attr_t attr;
1526         int ret;
1527
1528         pthread_attr_init(&attr);
1529         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1530         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1531         pthread_attr_destroy(&attr);
1532         if (ret) {
1533                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1534                 return 1;
1535         }
1536
1537         ret = pthread_detach(gtod_thread);
1538         if (ret) {
1539                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1540                 return 1;
1541         }
1542
1543         dprint(FD_MUTEX, "wait on startup_mutex\n");
1544         fio_mutex_down(startup_mutex);
1545         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1546         return 0;
1547 }
1548
1549 /*
1550  * Main function for kicking off and reaping jobs, as needed.
1551  */
1552 static void run_threads(void)
1553 {
1554         struct thread_data *td;
1555         unsigned long spent;
1556         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1557
1558         if (fio_pin_memory())
1559                 return;
1560
1561         if (fio_gtod_offload && fio_start_gtod_thread())
1562                 return;
1563
1564         set_sig_handlers();
1565
1566         if (!terse_output) {
1567                 log_info("Starting ");
1568                 if (nr_thread)
1569                         log_info("%d thread%s", nr_thread,
1570                                                 nr_thread > 1 ? "s" : "");
1571                 if (nr_process) {
1572                         if (nr_thread)
1573                                 log_info(" and ");
1574                         log_info("%d process%s", nr_process,
1575                                                 nr_process > 1 ? "es" : "");
1576                 }
1577                 log_info("\n");
1578                 fflush(stdout);
1579         }
1580
1581         todo = thread_number;
1582         nr_running = 0;
1583         nr_started = 0;
1584         m_rate = t_rate = 0;
1585
1586         for_each_td(td, i) {
1587                 print_status_init(td->thread_number - 1);
1588
1589                 if (!td->o.create_serialize)
1590                         continue;
1591
1592                 /*
1593                  * do file setup here so it happens sequentially,
1594                  * we don't want X number of threads getting their
1595                  * client data interspersed on disk
1596                  */
1597                 if (setup_files(td)) {
1598                         exit_value++;
1599                         if (td->error)
1600                                 log_err("fio: pid=%d, err=%d/%s\n",
1601                                         (int) td->pid, td->error, td->verror);
1602                         td_set_runstate(td, TD_REAPED);
1603                         todo--;
1604                 } else {
1605                         struct fio_file *f;
1606                         unsigned int j;
1607
1608                         /*
1609                          * for sharing to work, each job must always open
1610                          * its own files. so close them, if we opened them
1611                          * for creation
1612                          */
1613                         for_each_file(td, f, j) {
1614                                 if (fio_file_open(f))
1615                                         td_io_close_file(td, f);
1616                         }
1617                 }
1618         }
1619
1620         set_genesis_time();
1621
1622         while (todo) {
1623                 struct thread_data *map[REAL_MAX_JOBS];
1624                 struct timeval this_start;
1625                 int this_jobs = 0, left;
1626
1627                 /*
1628                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1629                  */
1630                 for_each_td(td, i) {
1631                         if (td->runstate != TD_NOT_CREATED)
1632                                 continue;
1633
1634                         /*
1635                          * never got a chance to start, killed by other
1636                          * thread for some reason
1637                          */
1638                         if (td->terminate) {
1639                                 todo--;
1640                                 continue;
1641                         }
1642
1643                         if (td->o.start_delay) {
1644                                 spent = mtime_since_genesis();
1645
1646                                 if (td->o.start_delay * 1000 > spent)
1647                                         continue;
1648                         }
1649
1650                         if (td->o.stonewall && (nr_started || nr_running)) {
1651                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1652                                                         td->o.name);
1653                                 break;
1654                         }
1655
1656                         init_disk_util(td);
1657
1658                         /*
1659                          * Set state to created. Thread will transition
1660                          * to TD_INITIALIZED when it's done setting up.
1661                          */
1662                         td_set_runstate(td, TD_CREATED);
1663                         map[this_jobs++] = td;
1664                         nr_started++;
1665
1666                         if (td->o.use_thread) {
1667                                 int ret;
1668
1669                                 dprint(FD_PROCESS, "will pthread_create\n");
1670                                 ret = pthread_create(&td->thread, NULL,
1671                                                         thread_main, td);
1672                                 if (ret) {
1673                                         log_err("pthread_create: %s\n",
1674                                                         strerror(ret));
1675                                         nr_started--;
1676                                         break;
1677                                 }
1678                                 ret = pthread_detach(td->thread);
1679                                 if (ret)
1680                                         log_err("pthread_detach: %s",
1681                                                         strerror(ret));
1682                         } else {
1683                                 pid_t pid;
1684                                 dprint(FD_PROCESS, "will fork\n");
1685                                 pid = fork();
1686                                 if (!pid) {
1687                                         int ret = fork_main(shm_id, i);
1688
1689                                         _exit(ret);
1690                                 } else if (i == fio_debug_jobno)
1691                                         *fio_debug_jobp = pid;
1692                         }
1693                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1694                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1695                                 log_err("fio: job startup hung? exiting.\n");
1696                                 fio_terminate_threads(TERMINATE_ALL);
1697                                 fio_abort = 1;
1698                                 nr_started--;
1699                                 break;
1700                         }
1701                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1702                 }
1703
1704                 /*
1705                  * Wait for the started threads to transition to
1706                  * TD_INITIALIZED.
1707                  */
1708                 fio_gettime(&this_start, NULL);
1709                 left = this_jobs;
1710                 while (left && !fio_abort) {
1711                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1712                                 break;
1713
1714                         usleep(100000);
1715
1716                         for (i = 0; i < this_jobs; i++) {
1717                                 td = map[i];
1718                                 if (!td)
1719                                         continue;
1720                                 if (td->runstate == TD_INITIALIZED) {
1721                                         map[i] = NULL;
1722                                         left--;
1723                                 } else if (td->runstate >= TD_EXITED) {
1724                                         map[i] = NULL;
1725                                         left--;
1726                                         todo--;
1727                                         nr_running++; /* work-around... */
1728                                 }
1729                         }
1730                 }
1731
1732                 if (left) {
1733                         log_err("fio: %d jobs failed to start\n", left);
1734                         for (i = 0; i < this_jobs; i++) {
1735                                 td = map[i];
1736                                 if (!td)
1737                                         continue;
1738                                 kill(td->pid, SIGTERM);
1739                         }
1740                         break;
1741                 }
1742
1743                 /*
1744                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1745                  */
1746                 for_each_td(td, i) {
1747                         if (td->runstate != TD_INITIALIZED)
1748                                 continue;
1749
1750                         if (in_ramp_time(td))
1751                                 td_set_runstate(td, TD_RAMP);
1752                         else
1753                                 td_set_runstate(td, TD_RUNNING);
1754                         nr_running++;
1755                         nr_started--;
1756                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1757                         t_rate += td->o.rate[0] + td->o.rate[1];
1758                         todo--;
1759                         fio_mutex_up(td->mutex);
1760                 }
1761
1762                 reap_threads(&nr_running, &t_rate, &m_rate);
1763
1764                 if (todo) {
1765                         if (is_backend)
1766                                 fio_server_idle_loop();
1767                         else
1768                                 usleep(100000);
1769                 }
1770         }
1771
1772         while (nr_running) {
1773                 reap_threads(&nr_running, &t_rate, &m_rate);
1774
1775                 if (is_backend)
1776                         fio_server_idle_loop();
1777                 else
1778                         usleep(10000);
1779         }
1780
1781         update_io_ticks();
1782         fio_unpin_memory();
1783 }
1784
1785 int exec_run(void)
1786 {
1787         struct thread_data *td;
1788         int i;
1789
1790         if (nr_clients)
1791                 return fio_handle_clients();
1792         if (exec_profile) {
1793                 if (load_profile(exec_profile))
1794                         return 1;
1795                 free(exec_profile);
1796                 exec_profile = NULL;
1797         }
1798         if (!thread_number)
1799                 return 0;
1800
1801         if (write_bw_log) {
1802                 setup_log(&agg_io_log[DDIR_READ], 0);
1803                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1804         }
1805
1806         startup_mutex = fio_mutex_init(0);
1807         if (startup_mutex == NULL)
1808                 return 1;
1809         writeout_mutex = fio_mutex_init(1);
1810         if (writeout_mutex == NULL)
1811                 return 1;
1812
1813         set_genesis_time();
1814         create_disk_util_thread();
1815
1816         cgroup_list = smalloc(sizeof(*cgroup_list));
1817         INIT_FLIST_HEAD(cgroup_list);
1818
1819         run_threads();
1820
1821         if (!fio_abort) {
1822                 show_run_stats();
1823                 if (write_bw_log) {
1824                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1825                         __finish_log(agg_io_log[DDIR_WRITE],
1826                                         "agg-write_bw.log");
1827                 }
1828         }
1829
1830         for_each_td(td, i)
1831                 fio_options_free(td);
1832
1833         cgroup_kill(cgroup_list);
1834         sfree(cgroup_list);
1835         sfree(cgroup_mnt);
1836
1837         fio_mutex_remove(startup_mutex);
1838         fio_mutex_remove(writeout_mutex);
1839         return exit_value;
1840 }
1841
1842 void reset_fio_state(void)
1843 {
1844         groupid = 0;
1845         thread_number = 0;
1846         nr_process = 0;
1847         nr_thread = 0;
1848         done_secs = 0;
1849 }
1850
1851 static int endian_check(void)
1852 {
1853         union {
1854                 uint8_t c[8];
1855                 uint64_t v;
1856         } u;
1857         int le = 0, be = 0;
1858
1859         u.v = 0x12;
1860         if (u.c[7] == 0x12)
1861                 be = 1;
1862         else if (u.c[0] == 0x12)
1863                 le = 1;
1864
1865 #if defined(FIO_LITTLE_ENDIAN)
1866         if (be)
1867                 return 1;
1868 #elif defined(FIO_BIG_ENDIAN)
1869         if (le)
1870                 return 1;
1871 #else
1872         return 1;
1873 #endif
1874
1875         if (!le && !be)
1876                 return 1;
1877
1878         return 0;
1879 }
1880
1881 int main(int argc, char *argv[], char *envp[])
1882 {
1883         long ps;
1884
1885         if (endian_check()) {
1886                 log_err("fio: endianness settings appear wrong.\n");
1887                 log_err("fio: please report this to fio@vger.kernel.org\n");
1888                 return 1;
1889         }
1890
1891         arch_init(envp);
1892
1893         sinit();
1894
1895         /*
1896          * We need locale for number printing, if it isn't set then just
1897          * go with the US format.
1898          */
1899         if (!getenv("LC_NUMERIC"))
1900                 setlocale(LC_NUMERIC, "en_US");
1901
1902         ps = sysconf(_SC_PAGESIZE);
1903         if (ps < 0) {
1904                 log_err("Failed to get page size\n");
1905                 return 1;
1906         }
1907
1908         page_size = ps;
1909         page_mask = ps - 1;
1910
1911         fio_keywords_init();
1912
1913         if (parse_options(argc, argv))
1914                 return 1;
1915
1916         return exec_run();
1917 }