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 <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "hash.h"
  39 #include "smalloc.h"
  40 #include "verify.h"
  41 #include "diskutil.h"
  42 #include "cgroup.h"
  43 #include "profile.h"
  44 #include "lib/rand.h"
  45
  46 unsigned long page_mask;
  47 unsigned long page_size;
  48
  49 #define PAGE_ALIGN(buf) \
  50         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  51
  52 int groupid = 0;
  53 int thread_number = 0;
  54 int nr_process = 0;
  55 int nr_thread = 0;
  56 int shm_id = 0;
  57 int temp_stall_ts;
  58 unsigned long done_secs = 0;
  59
  60 static struct fio_mutex *startup_mutex;
  61 static struct fio_mutex *writeout_mutex;
  62 static volatile int fio_abort;
  63 static int exit_value;
  64 static struct itimerval itimer;
  65 static pthread_t gtod_thread;
  66 static struct flist_head *cgroup_list;
  67 static char *cgroup_mnt;
  68
  69 struct io_log *agg_io_log[2];
  70
  71 #define TERMINATE_ALL           (-1)
  72 #define JOB_START_TIMEOUT       (5 * 1000)
  73
  74 void td_set_runstate(struct thread_data *td, int runstate)
  75 {
  76         if (td->runstate == runstate)
  77                 return;
  78
  79         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  80                                                 td->runstate, runstate);
  81         td->runstate = runstate;
  82 }
  83
  84 static void terminate_threads(int group_id)
  85 {
  86         struct thread_data *td;
  87         int i;
  88
  89         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  90
  91         for_each_td(td, i) {
  92                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  93                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  94                                                 td->o.name, (int) td->pid);
  95                         td->terminate = 1;
  96                         td->o.start_delay = 0;
  97
  98                         /*
  99                          * if the thread is running, just let it exit
 100                          */
 101                         if (td->runstate < TD_RUNNING)
 102                                 kill(td->pid, SIGQUIT);
 103                         else {
 104                                 struct ioengine_ops *ops = td->io_ops;
 105
 106                                 if (ops && (ops->flags & FIO_SIGQUIT))
 107                                         kill(td->pid, SIGQUIT);
 108                         }
 109                 }
 110         }
 111 }
 112
 113 static void status_timer_arm(void)
 114 {
 115         itimer.it_value.tv_sec = 0;
 116         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 117         setitimer(ITIMER_REAL, &itimer, NULL);
 118 }
 119
 120 static void sig_alrm(int fio_unused sig)
 121 {
 122         if (threads) {
 123                 update_io_ticks();
 124                 print_thread_status();
 125                 status_timer_arm();
 126         }
 127 }
 128
 129 /*
 130  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 131  */
 132 static void sig_quit(int sig)
 133 {
 134 }
 135
 136 static void sig_int(int sig)
 137 {
 138         if (threads) {
 139                 log_info("\nfio: terminating on signal %d\n", sig);
 140                 fflush(stdout);
 141                 terminate_threads(TERMINATE_ALL);
 142         }
 143 }
 144
 145 static void sig_ill(int fio_unused sig)
 146 {
 147         if (!threads)
 148                 return;
 149
 150         log_err("fio: system does not support the sse4.2 instruction for crc32c-intel.\nUse crc32c instead.");
 151         terminate_threads(TERMINATE_ALL);
 152 }
 153
 154 static void set_sig_handlers(void)
 155 {
 156         struct sigaction act;
 157
 158         memset(&act, 0, sizeof(act));
 159         act.sa_handler = sig_alrm;
 160         act.sa_flags = SA_RESTART;
 161         sigaction(SIGALRM, &act, NULL);
 162
 163         memset(&act, 0, sizeof(act));
 164         act.sa_handler = sig_int;
 165         act.sa_flags = SA_RESTART;
 166         sigaction(SIGINT, &act, NULL);
 167
 168         memset(&act, 0, sizeof(act));
 169         act.sa_handler = sig_ill;
 170         act.sa_flags = SA_RESETHAND;
 171         sigaction(SIGILL, &act, NULL);
 172
 173         memset(&act, 0, sizeof(act));
 174         act.sa_handler = sig_quit;
 175         act.sa_flags = SA_RESTART;
 176         sigaction(SIGQUIT, &act, NULL);
 177 }
 178
 179 /*
 180  * Check if we are above the minimum rate given.
 181  */
 182 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 183                             enum td_ddir ddir)
 184 {
 185         unsigned long long bytes = 0;
 186         unsigned long iops = 0;
 187         unsigned long spent;
 188         unsigned long rate;
 189         unsigned int ratemin = 0;
 190         unsigned int rate_iops = 0;
 191         unsigned int rate_iops_min = 0;
 192
 193         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 194                 return 0;
 195
 196         /*
 197          * allow a 2 second settle period in the beginning
 198          */
 199         if (mtime_since(&td->start, now) < 2000)
 200                 return 0;
 201
 202         iops += td->io_blocks[ddir];
 203         bytes += td->this_io_bytes[ddir];
 204         ratemin += td->o.ratemin[ddir];
 205         rate_iops += td->o.rate_iops[ddir];
 206         rate_iops_min += td->o.rate_iops_min[ddir];
 207
 208         /*
 209          * if rate blocks is set, sample is running
 210          */
 211         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 212                 spent = mtime_since(&td->lastrate[ddir], now);
 213                 if (spent < td->o.ratecycle)
 214                         return 0;
 215
 216                 if (td->o.rate[ddir]) {
 217                         /*
 218                          * check bandwidth specified rate
 219                          */
 220                         if (bytes < td->rate_bytes[ddir]) {
 221                                 log_err("%s: min rate %u not met\n", td->o.name,
 222                                                                 ratemin);
 223                                 return 1;
 224                         } else {
 225                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 226                                 if (rate < ratemin ||
 227                                     bytes < td->rate_bytes[ddir]) {
 228                                         log_err("%s: min rate %u not met, got"
 229                                                 " %luKB/sec\n", td->o.name,
 230                                                         ratemin, rate);
 231                                         return 1;
 232                                 }
 233                         }
 234                 } else {
 235                         /*
 236                          * checks iops specified rate
 237                          */
 238                         if (iops < rate_iops) {
 239                                 log_err("%s: min iops rate %u not met\n",
 240                                                 td->o.name, rate_iops);
 241                                 return 1;
 242                         } else {
 243                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 244                                 if (rate < rate_iops_min ||
 245                                     iops < td->rate_blocks[ddir]) {
 246                                         log_err("%s: min iops rate %u not met,"
 247                                                 " got %lu\n", td->o.name,
 248                                                         rate_iops_min, rate);
 249                                 }
 250                         }
 251                 }
 252         }
 253
 254         td->rate_bytes[ddir] = bytes;
 255         td->rate_blocks[ddir] = iops;
 256         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 257         return 0;
 258 }
 259
 260 static int check_min_rate(struct thread_data *td, struct timeval *now,
 261                           unsigned long *bytes_done)
 262 {
 263         int ret = 0;
 264
 265         if (bytes_done[0])
 266                 ret |= __check_min_rate(td, now, 0);
 267         if (bytes_done[1])
 268                 ret |= __check_min_rate(td, now, 1);
 269
 270         return ret;
 271 }
 272
 273 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 274 {
 275         if (!td->o.timeout)
 276                 return 0;
 277         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 278                 return 1;
 279
 280         return 0;
 281 }
 282
 283 /*
 284  * When job exits, we can cancel the in-flight IO if we are using async
 285  * io. Attempt to do so.
 286  */
 287 static void cleanup_pending_aio(struct thread_data *td)
 288 {
 289         struct flist_head *entry, *n;
 290         struct io_u *io_u;
 291         int r;
 292
 293         /*
 294          * get immediately available events, if any
 295          */
 296         r = io_u_queued_complete(td, 0, NULL);
 297         if (r < 0)
 298                 return;
 299
 300         /*
 301          * now cancel remaining active events
 302          */
 303         if (td->io_ops->cancel) {
 304                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 305                         io_u = flist_entry(entry, struct io_u, list);
 306
 307                         /*
 308                          * if the io_u isn't in flight, then that generally
 309                          * means someone leaked an io_u. complain but fix
 310                          * it up, so we don't stall here.
 311                          */
 312                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 313                                 log_err("fio: non-busy IO on busy list\n");
 314                                 put_io_u(td, io_u);
 315                         } else {
 316                                 r = td->io_ops->cancel(td, io_u);
 317                                 if (!r)
 318                                         put_io_u(td, io_u);
 319                         }
 320                 }
 321         }
 322
 323         if (td->cur_depth)
 324                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 325 }
 326
 327 /*
 328  * Helper to handle the final sync of a file. Works just like the normal
 329  * io path, just does everything sync.
 330  */
 331 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 332 {
 333         struct io_u *io_u = __get_io_u(td);
 334         int ret;
 335
 336         if (!io_u)
 337                 return 1;
 338
 339         io_u->ddir = DDIR_SYNC;
 340         io_u->file = f;
 341
 342         if (td_io_prep(td, io_u)) {
 343                 put_io_u(td, io_u);
 344                 return 1;
 345         }
 346
 347 requeue:
 348         ret = td_io_queue(td, io_u);
 349         if (ret < 0) {
 350                 td_verror(td, io_u->error, "td_io_queue");
 351                 put_io_u(td, io_u);
 352                 return 1;
 353         } else if (ret == FIO_Q_QUEUED) {
 354                 if (io_u_queued_complete(td, 1, NULL) < 0)
 355                         return 1;
 356         } else if (ret == FIO_Q_COMPLETED) {
 357                 if (io_u->error) {
 358                         td_verror(td, io_u->error, "td_io_queue");
 359                         return 1;
 360                 }
 361
 362                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 363                         return 1;
 364         } else if (ret == FIO_Q_BUSY) {
 365                 if (td_io_commit(td))
 366                         return 1;
 367                 goto requeue;
 368         }
 369
 370         return 0;
 371 }
 372
 373 static inline void update_tv_cache(struct thread_data *td)
 374 {
 375         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 376                 fio_gettime(&td->tv_cache, NULL);
 377 }
 378
 379 static int break_on_this_error(struct thread_data *td, int *retptr)
 380 {
 381         int ret = *retptr;
 382
 383         if (ret < 0 || td->error) {
 384                 int err;
 385
 386                 if (!td->o.continue_on_error)
 387                         return 1;
 388
 389                 if (ret < 0)
 390                         err = -ret;
 391                 else
 392                         err = td->error;
 393
 394                 if (td_non_fatal_error(err)) {
 395                         /*
 396                          * Continue with the I/Os in case of
 397                          * a non fatal error.
 398                          */
 399                         update_error_count(td, err);
 400                         td_clear_error(td);
 401                         *retptr = 0;
 402                         return 0;
 403                 } else if (td->o.fill_device && err == ENOSPC) {
 404                         /*
 405                          * We expect to hit this error if
 406                          * fill_device option is set.
 407                          */
 408                         td_clear_error(td);
 409                         td->terminate = 1;
 410                         return 1;
 411                 } else {
 412                         /*
 413                          * Stop the I/O in case of a fatal
 414                          * error.
 415                          */
 416                         update_error_count(td, err);
 417                         return 1;
 418                 }
 419         }
 420
 421         return 0;
 422 }
 423
 424 /*
 425  * The main verify engine. Runs over the writes we previously submitted,
 426  * reads the blocks back in, and checks the crc/md5 of the data.
 427  */
 428 static void do_verify(struct thread_data *td)
 429 {
 430         struct fio_file *f;
 431         struct io_u *io_u;
 432         int ret, min_events;
 433         unsigned int i;
 434
 435         dprint(FD_VERIFY, "starting loop\n");
 436
 437         /*
 438          * sync io first and invalidate cache, to make sure we really
 439          * read from disk.
 440          */
 441         for_each_file(td, f, i) {
 442                 if (!fio_file_open(f))
 443                         continue;
 444                 if (fio_io_sync(td, f))
 445                         break;
 446                 if (file_invalidate_cache(td, f))
 447                         break;
 448         }
 449
 450         if (td->error)
 451                 return;
 452
 453         td_set_runstate(td, TD_VERIFYING);
 454
 455         io_u = NULL;
 456         while (!td->terminate) {
 457                 int ret2, full;
 458
 459                 update_tv_cache(td);
 460
 461                 if (runtime_exceeded(td, &td->tv_cache)) {
 462                         td->terminate = 1;
 463                         break;
 464                 }
 465
 466                 io_u = __get_io_u(td);
 467                 if (!io_u)
 468                         break;
 469
 470                 if (get_next_verify(td, io_u)) {
 471                         put_io_u(td, io_u);
 472                         break;
 473                 }
 474
 475                 if (td_io_prep(td, io_u)) {
 476                         put_io_u(td, io_u);
 477                         break;
 478                 }
 479
 480                 if (td->o.verify_async)
 481                         io_u->end_io = verify_io_u_async;
 482                 else
 483                         io_u->end_io = verify_io_u;
 484
 485                 ret = td_io_queue(td, io_u);
 486                 switch (ret) {
 487                 case FIO_Q_COMPLETED:
 488                         if (io_u->error) {
 489                                 ret = -io_u->error;
 490                                 clear_io_u(td, io_u);
 491                         } else if (io_u->resid) {
 492                                 int bytes = io_u->xfer_buflen - io_u->resid;
 493                                 struct fio_file *f = io_u->file;
 494
 495                                 /*
 496                                  * zero read, fail
 497                                  */
 498                                 if (!bytes) {
 499                                         td_verror(td, EIO, "full resid");
 500                                         put_io_u(td, io_u);
 501                                         break;
 502                                 }
 503
 504                                 io_u->xfer_buflen = io_u->resid;
 505                                 io_u->xfer_buf += bytes;
 506                                 io_u->offset += bytes;
 507
 508                                 td->ts.short_io_u[io_u->ddir]++;
 509
 510                                 if (io_u->offset == f->real_file_size)
 511                                         goto sync_done;
 512
 513                                 requeue_io_u(td, &io_u);
 514                         } else {
 515 sync_done:
 516                                 ret = io_u_sync_complete(td, io_u, NULL);
 517                                 if (ret < 0)
 518                                         break;
 519                         }
 520                         continue;
 521                 case FIO_Q_QUEUED:
 522                         break;
 523                 case FIO_Q_BUSY:
 524                         requeue_io_u(td, &io_u);
 525                         ret2 = td_io_commit(td);
 526                         if (ret2 < 0)
 527                                 ret = ret2;
 528                         break;
 529                 default:
 530                         assert(ret < 0);
 531                         td_verror(td, -ret, "td_io_queue");
 532                         break;
 533                 }
 534
 535                 if (break_on_this_error(td, &ret))
 536                         break;
 537
 538                 /*
 539                  * if we can queue more, do so. but check if there are
 540                  * completed io_u's first.
 541                  */
 542                 full = queue_full(td) || ret == FIO_Q_BUSY;
 543                 if (full || !td->o.iodepth_batch_complete) {
 544                         min_events = min(td->o.iodepth_batch_complete,
 545                                          td->cur_depth);
 546                         if (full && !min_events)
 547                                 min_events = 1;
 548
 549                         do {
 550                                 /*
 551                                  * Reap required number of io units, if any,
 552                                  * and do the verification on them through
 553                                  * the callback handler
 554                                  */
 555                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 556                                         ret = -1;
 557                                         break;
 558                                 }
 559                         } while (full && (td->cur_depth > td->o.iodepth_low));
 560                 }
 561                 if (ret < 0)
 562                         break;
 563         }
 564
 565         if (!td->error) {
 566                 min_events = td->cur_depth;
 567
 568                 if (min_events)
 569                         ret = io_u_queued_complete(td, min_events, NULL);
 570         } else
 571                 cleanup_pending_aio(td);
 572
 573         td_set_runstate(td, TD_RUNNING);
 574
 575         dprint(FD_VERIFY, "exiting loop\n");
 576 }
 577
 578 /*
 579  * Main IO worker function. It retrieves io_u's to process and queues
 580  * and reaps them, checking for rate and errors along the way.
 581  */
 582 static void do_io(struct thread_data *td)
 583 {
 584         unsigned int i;
 585         int ret = 0;
 586
 587         if (in_ramp_time(td))
 588                 td_set_runstate(td, TD_RAMP);
 589         else
 590                 td_set_runstate(td, TD_RUNNING);
 591
 592         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 593                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 594                 struct timeval comp_time;
 595                 unsigned long bytes_done[2] = { 0, 0 };
 596                 int min_evts = 0;
 597                 struct io_u *io_u;
 598                 int ret2, full;
 599
 600                 if (td->terminate)
 601                         break;
 602
 603                 update_tv_cache(td);
 604
 605                 if (runtime_exceeded(td, &td->tv_cache)) {
 606                         td->terminate = 1;
 607                         break;
 608                 }
 609
 610                 io_u = get_io_u(td);
 611                 if (!io_u)
 612                         break;
 613
 614                 /*
 615                  * Add verification end_io handler, if asked to verify
 616                  * a previously written file.
 617                  */
 618                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 619                     !td_rw(td)) {
 620                         if (td->o.verify_async)
 621                                 io_u->end_io = verify_io_u_async;
 622                         else
 623                                 io_u->end_io = verify_io_u;
 624                         td_set_runstate(td, TD_VERIFYING);
 625                 } else if (in_ramp_time(td))
 626                         td_set_runstate(td, TD_RAMP);
 627                 else
 628                         td_set_runstate(td, TD_RUNNING);
 629
 630                 ret = td_io_queue(td, io_u);
 631                 switch (ret) {
 632                 case FIO_Q_COMPLETED:
 633                         if (io_u->error) {
 634                                 ret = -io_u->error;
 635                                 clear_io_u(td, io_u);
 636                         } else if (io_u->resid) {
 637                                 int bytes = io_u->xfer_buflen - io_u->resid;
 638                                 struct fio_file *f = io_u->file;
 639
 640                                 /*
 641                                  * zero read, fail
 642                                  */
 643                                 if (!bytes) {
 644                                         td_verror(td, EIO, "full resid");
 645                                         put_io_u(td, io_u);
 646                                         break;
 647                                 }
 648
 649                                 io_u->xfer_buflen = io_u->resid;
 650                                 io_u->xfer_buf += bytes;
 651                                 io_u->offset += bytes;
 652
 653                                 td->ts.short_io_u[io_u->ddir]++;
 654
 655                                 if (io_u->offset == f->real_file_size)
 656                                         goto sync_done;
 657
 658                                 requeue_io_u(td, &io_u);
 659                         } else {
 660 sync_done:
 661                                 if (__should_check_rate(td, 0) ||
 662                                     __should_check_rate(td, 1))
 663                                         fio_gettime(&comp_time, NULL);
 664
 665                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 666                                 if (ret < 0)
 667                                         break;
 668                         }
 669                         break;
 670                 case FIO_Q_QUEUED:
 671                         /*
 672                          * if the engine doesn't have a commit hook,
 673                          * the io_u is really queued. if it does have such
 674                          * a hook, it has to call io_u_queued() itself.
 675                          */
 676                         if (td->io_ops->commit == NULL)
 677                                 io_u_queued(td, io_u);
 678                         break;
 679                 case FIO_Q_BUSY:
 680                         requeue_io_u(td, &io_u);
 681                         ret2 = td_io_commit(td);
 682                         if (ret2 < 0)
 683                                 ret = ret2;
 684                         break;
 685                 default:
 686                         assert(ret < 0);
 687                         put_io_u(td, io_u);
 688                         break;
 689                 }
 690
 691                 if (break_on_this_error(td, &ret))
 692                         break;
 693
 694                 /*
 695                  * See if we need to complete some commands
 696                  */
 697                 full = queue_full(td) || ret == FIO_Q_BUSY;
 698                 if (full || !td->o.iodepth_batch_complete) {
 699                         min_evts = min(td->o.iodepth_batch_complete,
 700                                         td->cur_depth);
 701                         if (full && !min_evts)
 702                                 min_evts = 1;
 703
 704                         if (__should_check_rate(td, 0) ||
 705                             __should_check_rate(td, 1))
 706                                 fio_gettime(&comp_time, NULL);
 707
 708                         do {
 709                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 710                                 if (ret < 0)
 711                                         break;
 712
 713                         } while (full && (td->cur_depth > td->o.iodepth_low));
 714                 }
 715
 716                 if (ret < 0)
 717                         break;
 718                 if (!(bytes_done[0] + bytes_done[1]))
 719                         continue;
 720
 721                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 722                         if (check_min_rate(td, &comp_time, bytes_done)) {
 723                                 if (exitall_on_terminate)
 724                                         terminate_threads(td->groupid);
 725                                 td_verror(td, EIO, "check_min_rate");
 726                                 break;
 727                         }
 728                 }
 729
 730                 if (td->o.thinktime) {
 731                         unsigned long long b;
 732
 733                         b = td->io_blocks[0] + td->io_blocks[1];
 734                         if (!(b % td->o.thinktime_blocks)) {
 735                                 int left;
 736
 737                                 if (td->o.thinktime_spin)
 738                                         usec_spin(td->o.thinktime_spin);
 739
 740                                 left = td->o.thinktime - td->o.thinktime_spin;
 741                                 if (left)
 742                                         usec_sleep(td, left);
 743                         }
 744                 }
 745         }
 746
 747         if (td->o.fill_device && td->error == ENOSPC) {
 748                 td->error = 0;
 749                 td->terminate = 1;
 750         }
 751         if (!td->error) {
 752                 struct fio_file *f;
 753
 754                 i = td->cur_depth;
 755                 if (i)
 756                         ret = io_u_queued_complete(td, i, NULL);
 757
 758                 if (should_fsync(td) && td->o.end_fsync) {
 759                         td_set_runstate(td, TD_FSYNCING);
 760
 761                         for_each_file(td, f, i) {
 762                                 if (!fio_file_open(f))
 763                                         continue;
 764                                 fio_io_sync(td, f);
 765                         }
 766                 }
 767         } else
 768                 cleanup_pending_aio(td);
 769
 770         /*
 771          * stop job if we failed doing any IO
 772          */
 773         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 774                 td->done = 1;
 775 }
 776
 777 static void cleanup_io_u(struct thread_data *td)
 778 {
 779         struct flist_head *entry, *n;
 780         struct io_u *io_u;
 781
 782         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 783                 io_u = flist_entry(entry, struct io_u, list);
 784
 785                 flist_del(&io_u->list);
 786                 free(io_u);
 787         }
 788
 789         free_io_mem(td);
 790 }
 791
 792 static int init_io_u(struct thread_data *td)
 793 {
 794         struct io_u *io_u;
 795         unsigned int max_bs;
 796         int cl_align, i, max_units;
 797         char *p;
 798
 799         max_units = td->o.iodepth;
 800         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 801         td->orig_buffer_size = (unsigned long long) max_bs
 802                                         * (unsigned long long) max_units;
 803
 804         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 805                 unsigned long bs;
 806
 807                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 808                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 809         }
 810
 811         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 812                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 813                 return 1;
 814         }
 815
 816         if (allocate_io_mem(td))
 817                 return 1;
 818
 819         if (td->o.odirect || td->o.mem_align)
 820                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 821         else
 822                 p = td->orig_buffer;
 823
 824         cl_align = os_cache_line_size();
 825
 826         for (i = 0; i < max_units; i++) {
 827                 void *ptr;
 828
 829                 if (td->terminate)
 830                         return 1;
 831
 832                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 833                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 834                         break;
 835                 }
 836
 837                 io_u = ptr;
 838                 memset(io_u, 0, sizeof(*io_u));
 839                 INIT_FLIST_HEAD(&io_u->list);
 840                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 841
 842                 if (!(td->io_ops->flags & FIO_NOIO)) {
 843                         io_u->buf = p + max_bs * i;
 844                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 845
 846                         if (td_write(td) && !td->o.refill_buffers)
 847                                 io_u_fill_buffer(td, io_u, max_bs);
 848                 }
 849
 850                 io_u->index = i;
 851                 io_u->flags = IO_U_F_FREE;
 852                 flist_add(&io_u->list, &td->io_u_freelist);
 853         }
 854
 855         return 0;
 856 }
 857
 858 static int switch_ioscheduler(struct thread_data *td)
 859 {
 860         char tmp[256], tmp2[128];
 861         FILE *f;
 862         int ret;
 863
 864         if (td->io_ops->flags & FIO_DISKLESSIO)
 865                 return 0;
 866
 867         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 868
 869         f = fopen(tmp, "r+");
 870         if (!f) {
 871                 if (errno == ENOENT) {
 872                         log_err("fio: os or kernel doesn't support IO scheduler"
 873                                 " switching\n");
 874                         return 0;
 875                 }
 876                 td_verror(td, errno, "fopen iosched");
 877                 return 1;
 878         }
 879
 880         /*
 881          * Set io scheduler.
 882          */
 883         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 884         if (ferror(f) || ret != 1) {
 885                 td_verror(td, errno, "fwrite");
 886                 fclose(f);
 887                 return 1;
 888         }
 889
 890         rewind(f);
 891
 892         /*
 893          * Read back and check that the selected scheduler is now the default.
 894          */
 895         ret = fread(tmp, 1, sizeof(tmp), f);
 896         if (ferror(f) || ret < 0) {
 897                 td_verror(td, errno, "fread");
 898                 fclose(f);
 899                 return 1;
 900         }
 901
 902         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 903         if (!strstr(tmp, tmp2)) {
 904                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 905                 td_verror(td, EINVAL, "iosched_switch");
 906                 fclose(f);
 907                 return 1;
 908         }
 909
 910         fclose(f);
 911         return 0;
 912 }
 913
 914 static int keep_running(struct thread_data *td)
 915 {
 916         unsigned long long io_done;
 917
 918         if (td->done)
 919                 return 0;
 920         if (td->o.time_based)
 921                 return 1;
 922         if (td->o.loops) {
 923                 td->o.loops--;
 924                 return 1;
 925         }
 926
 927         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 928                         + td->io_skip_bytes;
 929         if (io_done < td->o.size)
 930                 return 1;
 931
 932         return 0;
 933 }
 934
 935 static void reset_io_counters(struct thread_data *td)
 936 {
 937         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 938         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 939         td->zone_bytes = 0;
 940         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 941         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 942
 943         td->last_was_sync = 0;
 944
 945         /*
 946          * reset file done count if we are to start over
 947          */
 948         if (td->o.time_based || td->o.loops)
 949                 td->nr_done_files = 0;
 950
 951         /*
 952          * Set the same seed to get repeatable runs
 953          */
 954         td_fill_rand_seeds(td);
 955 }
 956
 957 void reset_all_stats(struct thread_data *td)
 958 {
 959         struct timeval tv;
 960         int i;
 961
 962         reset_io_counters(td);
 963
 964         for (i = 0; i < 2; i++) {
 965                 td->io_bytes[i] = 0;
 966                 td->io_blocks[i] = 0;
 967                 td->io_issues[i] = 0;
 968                 td->ts.total_io_u[i] = 0;
 969         }
 970
 971         fio_gettime(&tv, NULL);
 972         memcpy(&td->epoch, &tv, sizeof(tv));
 973         memcpy(&td->start, &tv, sizeof(tv));
 974 }
 975
 976 static void clear_io_state(struct thread_data *td)
 977 {
 978         struct fio_file *f;
 979         unsigned int i;
 980
 981         reset_io_counters(td);
 982
 983         close_files(td);
 984         for_each_file(td, f, i)
 985                 fio_file_clear_done(f);
 986 }
 987
 988 static int exec_string(const char *string)
 989 {
 990         int ret, newlen = strlen(string) + 1 + 8;
 991         char *str;
 992
 993         str = malloc(newlen);
 994         sprintf(str, "sh -c %s", string);
 995
 996         ret = system(str);
 997         if (ret == -1)
 998                 log_err("fio: exec of cmd <%s> failed\n", str);
 999
1000         free(str);
1001         return ret;
1002 }
1003
1004 /*
1005  * Entry point for the thread based jobs. The process based jobs end up
1006  * here as well, after a little setup.
1007  */
1008 static void *thread_main(void *data)
1009 {
1010         unsigned long long runtime[2], elapsed;
1011         struct thread_data *td = data;
1012         pthread_condattr_t attr;
1013         int clear_state;
1014
1015         if (!td->o.use_thread)
1016                 setsid();
1017
1018         td->pid = getpid();
1019
1020         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1021
1022         INIT_FLIST_HEAD(&td->io_u_freelist);
1023         INIT_FLIST_HEAD(&td->io_u_busylist);
1024         INIT_FLIST_HEAD(&td->io_u_requeues);
1025         INIT_FLIST_HEAD(&td->io_log_list);
1026         INIT_FLIST_HEAD(&td->io_hist_list);
1027         INIT_FLIST_HEAD(&td->verify_list);
1028         pthread_mutex_init(&td->io_u_lock, NULL);
1029         td->io_hist_tree = RB_ROOT;
1030
1031         pthread_condattr_init(&attr);
1032         pthread_cond_init(&td->verify_cond, &attr);
1033         pthread_cond_init(&td->free_cond, &attr);
1034
1035         td_set_runstate(td, TD_INITIALIZED);
1036         dprint(FD_MUTEX, "up startup_mutex\n");
1037         fio_mutex_up(startup_mutex);
1038         dprint(FD_MUTEX, "wait on td->mutex\n");
1039         fio_mutex_down(td->mutex);
1040         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1041
1042         /*
1043          * the ->mutex mutex is now no longer used, close it to avoid
1044          * eating a file descriptor
1045          */
1046         fio_mutex_remove(td->mutex);
1047
1048         if (td->o.uid != -1U && setuid(td->o.uid)) {
1049                 td_verror(td, errno, "setuid");
1050                 goto err;
1051         }
1052         if (td->o.gid != -1U && setgid(td->o.gid)) {
1053                 td_verror(td, errno, "setgid");
1054                 goto err;
1055         }
1056
1057         /*
1058          * May alter parameters that init_io_u() will use, so we need to
1059          * do this first.
1060          */
1061         if (init_iolog(td))
1062                 goto err;
1063
1064         if (init_io_u(td))
1065                 goto err;
1066
1067         if (td->o.verify_async && verify_async_init(td))
1068                 goto err;
1069
1070         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1071                 td_verror(td, errno, "cpu_set_affinity");
1072                 goto err;
1073         }
1074
1075         /*
1076          * If we have a gettimeofday() thread, make sure we exclude that
1077          * thread from this job
1078          */
1079         if (td->o.gtod_cpu) {
1080                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1081                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1082                         td_verror(td, errno, "cpu_set_affinity");
1083                         goto err;
1084                 }
1085         }
1086
1087         if (td->ioprio_set) {
1088                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1089                         td_verror(td, errno, "ioprio_set");
1090                         goto err;
1091                 }
1092         }
1093
1094         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1095                 goto err;
1096
1097         if (nice(td->o.nice) == -1) {
1098                 td_verror(td, errno, "nice");
1099                 goto err;
1100         }
1101
1102         if (td->o.ioscheduler && switch_ioscheduler(td))
1103                 goto err;
1104
1105         if (!td->o.create_serialize && setup_files(td))
1106                 goto err;
1107
1108         if (td_io_init(td))
1109                 goto err;
1110
1111         if (init_random_map(td))
1112                 goto err;
1113
1114         if (td->o.exec_prerun) {
1115                 if (exec_string(td->o.exec_prerun))
1116                         goto err;
1117         }
1118
1119         if (td->o.pre_read) {
1120                 if (pre_read_files(td) < 0)
1121                         goto err;
1122         }
1123
1124         fio_gettime(&td->epoch, NULL);
1125         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1126
1127         runtime[0] = runtime[1] = 0;
1128         clear_state = 0;
1129         while (keep_running(td)) {
1130                 fio_gettime(&td->start, NULL);
1131                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1132                                 sizeof(td->start));
1133                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1134                                 sizeof(td->start));
1135                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1136
1137                 if (td->o.ratemin[0] || td->o.ratemin[1])
1138                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1139                                                         sizeof(td->lastrate));
1140
1141                 if (clear_state)
1142                         clear_io_state(td);
1143
1144                 prune_io_piece_log(td);
1145
1146                 do_io(td);
1147
1148                 clear_state = 1;
1149
1150                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1151                         elapsed = utime_since_now(&td->start);
1152                         runtime[DDIR_READ] += elapsed;
1153                 }
1154                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1155                         elapsed = utime_since_now(&td->start);
1156                         runtime[DDIR_WRITE] += elapsed;
1157                 }
1158
1159                 if (td->error || td->terminate)
1160                         break;
1161
1162                 if (!td->o.do_verify ||
1163                     td->o.verify == VERIFY_NONE ||
1164                     (td->io_ops->flags & FIO_UNIDIR))
1165                         continue;
1166
1167                 clear_io_state(td);
1168
1169                 fio_gettime(&td->start, NULL);
1170
1171                 do_verify(td);
1172
1173                 runtime[DDIR_READ] += utime_since_now(&td->start);
1174
1175                 if (td->error || td->terminate)
1176                         break;
1177         }
1178
1179         update_rusage_stat(td);
1180         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1181         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1182         td->ts.total_run_time = mtime_since_now(&td->epoch);
1183         td->ts.io_bytes[0] = td->io_bytes[0];
1184         td->ts.io_bytes[1] = td->io_bytes[1];
1185
1186         fio_mutex_down(writeout_mutex);
1187         if (td->ts.bw_log) {
1188                 if (td->o.bw_log_file) {
1189                         finish_log_named(td, td->ts.bw_log,
1190                                                 td->o.bw_log_file, "bw");
1191                 } else
1192                         finish_log(td, td->ts.bw_log, "bw");
1193         }
1194         if (td->ts.slat_log) {
1195                 if (td->o.lat_log_file) {
1196                         finish_log_named(td, td->ts.slat_log,
1197                                                 td->o.lat_log_file, "slat");
1198                 } else
1199                         finish_log(td, td->ts.slat_log, "slat");
1200         }
1201         if (td->ts.clat_log) {
1202                 if (td->o.lat_log_file) {
1203                         finish_log_named(td, td->ts.clat_log,
1204                                                 td->o.lat_log_file, "clat");
1205                 } else
1206                         finish_log(td, td->ts.clat_log, "clat");
1207         }
1208         fio_mutex_up(writeout_mutex);
1209         if (td->o.exec_postrun)
1210                 exec_string(td->o.exec_postrun);
1211
1212         if (exitall_on_terminate)
1213                 terminate_threads(td->groupid);
1214
1215 err:
1216         if (td->error)
1217                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1218                                                         td->verror);
1219
1220         if (td->o.verify_async)
1221                 verify_async_exit(td);
1222
1223         close_and_free_files(td);
1224         close_ioengine(td);
1225         cleanup_io_u(td);
1226         cgroup_shutdown(td, &cgroup_mnt);
1227
1228         if (td->o.cpumask_set) {
1229                 int ret = fio_cpuset_exit(&td->o.cpumask);
1230
1231                 td_verror(td, ret, "fio_cpuset_exit");
1232         }
1233
1234         /*
1235          * do this very late, it will log file closing as well
1236          */
1237         if (td->o.write_iolog_file)
1238                 write_iolog_close(td);
1239
1240         options_mem_free(td);
1241         td_set_runstate(td, TD_EXITED);
1242         return (void *) (unsigned long) td->error;
1243 }
1244
1245 /*
1246  * We cannot pass the td data into a forked process, so attach the td and
1247  * pass it to the thread worker.
1248  */
1249 static int fork_main(int shmid, int offset)
1250 {
1251         struct thread_data *td;
1252         void *data, *ret;
1253
1254         data = shmat(shmid, NULL, 0);
1255         if (data == (void *) -1) {
1256                 int __err = errno;
1257
1258                 perror("shmat");
1259                 return __err;
1260         }
1261
1262         td = data + offset * sizeof(struct thread_data);
1263         ret = thread_main(td);
1264         shmdt(data);
1265         return (int) (unsigned long) ret;
1266 }
1267
1268 /*
1269  * Run over the job map and reap the threads that have exited, if any.
1270  */
1271 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1272 {
1273         struct thread_data *td;
1274         int i, cputhreads, realthreads, pending, status, ret;
1275
1276         /*
1277          * reap exited threads (TD_EXITED -> TD_REAPED)
1278          */
1279         realthreads = pending = cputhreads = 0;
1280         for_each_td(td, i) {
1281                 int flags = 0;
1282
1283                 /*
1284                  * ->io_ops is NULL for a thread that has closed its
1285                  * io engine
1286                  */
1287                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1288                         cputhreads++;
1289                 else
1290                         realthreads++;
1291
1292                 if (!td->pid) {
1293                         pending++;
1294                         continue;
1295                 }
1296                 if (td->runstate == TD_REAPED)
1297                         continue;
1298                 if (td->o.use_thread) {
1299                         if (td->runstate == TD_EXITED) {
1300                                 td_set_runstate(td, TD_REAPED);
1301                                 goto reaped;
1302                         }
1303                         continue;
1304                 }
1305
1306                 flags = WNOHANG;
1307                 if (td->runstate == TD_EXITED)
1308                         flags = 0;
1309
1310                 /*
1311                  * check if someone quit or got killed in an unusual way
1312                  */
1313                 ret = waitpid(td->pid, &status, flags);
1314                 if (ret < 0) {
1315                         if (errno == ECHILD) {
1316                                 log_err("fio: pid=%d disappeared %d\n",
1317                                                 (int) td->pid, td->runstate);
1318                                 td_set_runstate(td, TD_REAPED);
1319                                 goto reaped;
1320                         }
1321                         perror("waitpid");
1322                 } else if (ret == td->pid) {
1323                         if (WIFSIGNALED(status)) {
1324                                 int sig = WTERMSIG(status);
1325
1326                                 if (sig != SIGQUIT)
1327                                         log_err("fio: pid=%d, got signal=%d\n",
1328                                                         (int) td->pid, sig);
1329                                 td_set_runstate(td, TD_REAPED);
1330                                 goto reaped;
1331                         }
1332                         if (WIFEXITED(status)) {
1333                                 if (WEXITSTATUS(status) && !td->error)
1334                                         td->error = WEXITSTATUS(status);
1335
1336                                 td_set_runstate(td, TD_REAPED);
1337                                 goto reaped;
1338                         }
1339                 }
1340
1341                 /*
1342                  * thread is not dead, continue
1343                  */
1344                 pending++;
1345                 continue;
1346 reaped:
1347                 (*nr_running)--;
1348                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1349                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1350                 if (!td->pid)
1351                         pending--;
1352
1353                 if (td->error)
1354                         exit_value++;
1355
1356                 done_secs += mtime_since_now(&td->epoch) / 1000;
1357         }
1358
1359         if (*nr_running == cputhreads && !pending && realthreads)
1360                 terminate_threads(TERMINATE_ALL);
1361 }
1362
1363 static void *gtod_thread_main(void *data)
1364 {
1365         fio_mutex_up(startup_mutex);
1366
1367         /*
1368          * As long as we have jobs around, update the clock. It would be nice
1369          * to have some way of NOT hammering that CPU with gettimeofday(),
1370          * but I'm not sure what to use outside of a simple CPU nop to relax
1371          * it - we don't want to lose precision.
1372          */
1373         while (threads) {
1374                 fio_gtod_update();
1375                 nop;
1376         }
1377
1378         return NULL;
1379 }
1380
1381 static int fio_start_gtod_thread(void)
1382 {
1383         int ret;
1384
1385         ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
1386         if (ret) {
1387                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1388                 return 1;
1389         }
1390
1391         ret = pthread_detach(gtod_thread);
1392         if (ret) {
1393                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1394                 return 1;
1395         }
1396
1397         dprint(FD_MUTEX, "wait on startup_mutex\n");
1398         fio_mutex_down(startup_mutex);
1399         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1400         return 0;
1401 }
1402
1403 /*
1404  * Main function for kicking off and reaping jobs, as needed.
1405  */
1406 static void run_threads(void)
1407 {
1408         struct thread_data *td;
1409         unsigned long spent;
1410         int i, todo, nr_running, m_rate, t_rate, nr_started;
1411
1412         if (fio_pin_memory())
1413                 return;
1414
1415         if (fio_gtod_offload && fio_start_gtod_thread())
1416                 return;
1417
1418         if (!terse_output) {
1419                 log_info("Starting ");
1420                 if (nr_thread)
1421                         log_info("%d thread%s", nr_thread,
1422                                                 nr_thread > 1 ? "s" : "");
1423                 if (nr_process) {
1424                         if (nr_thread)
1425                                 printf(" and ");
1426                         log_info("%d process%s", nr_process,
1427                                                 nr_process > 1 ? "es" : "");
1428                 }
1429                 log_info("\n");
1430                 fflush(stdout);
1431         }
1432
1433         set_sig_handlers();
1434
1435         todo = thread_number;
1436         nr_running = 0;
1437         nr_started = 0;
1438         m_rate = t_rate = 0;
1439
1440         for_each_td(td, i) {
1441                 print_status_init(td->thread_number - 1);
1442
1443                 if (!td->o.create_serialize) {
1444                         init_disk_util(td);
1445                         continue;
1446                 }
1447
1448                 /*
1449                  * do file setup here so it happens sequentially,
1450                  * we don't want X number of threads getting their
1451                  * client data interspersed on disk
1452                  */
1453                 if (setup_files(td)) {
1454                         exit_value++;
1455                         if (td->error)
1456                                 log_err("fio: pid=%d, err=%d/%s\n",
1457                                         (int) td->pid, td->error, td->verror);
1458                         td_set_runstate(td, TD_REAPED);
1459                         todo--;
1460                 } else {
1461                         struct fio_file *f;
1462                         unsigned int i;
1463
1464                         /*
1465                          * for sharing to work, each job must always open
1466                          * its own files. so close them, if we opened them
1467                          * for creation
1468                          */
1469                         for_each_file(td, f, i) {
1470                                 if (fio_file_open(f))
1471                                         td_io_close_file(td, f);
1472                         }
1473                 }
1474
1475                 init_disk_util(td);
1476         }
1477
1478         set_genesis_time();
1479
1480         while (todo) {
1481                 struct thread_data *map[MAX_JOBS];
1482                 struct timeval this_start;
1483                 int this_jobs = 0, left;
1484
1485                 /*
1486                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1487                  */
1488                 for_each_td(td, i) {
1489                         if (td->runstate != TD_NOT_CREATED)
1490                                 continue;
1491
1492                         /*
1493                          * never got a chance to start, killed by other
1494                          * thread for some reason
1495                          */
1496                         if (td->terminate) {
1497                                 todo--;
1498                                 continue;
1499                         }
1500
1501                         if (td->o.start_delay) {
1502                                 spent = mtime_since_genesis();
1503
1504                                 if (td->o.start_delay * 1000 > spent)
1505                                         continue;
1506                         }
1507
1508                         if (td->o.stonewall && (nr_started || nr_running)) {
1509                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1510                                                         td->o.name);
1511                                 break;
1512                         }
1513
1514                         /*
1515                          * Set state to created. Thread will transition
1516                          * to TD_INITIALIZED when it's done setting up.
1517                          */
1518                         td_set_runstate(td, TD_CREATED);
1519                         map[this_jobs++] = td;
1520                         nr_started++;
1521
1522                         if (td->o.use_thread) {
1523                                 int ret;
1524
1525                                 dprint(FD_PROCESS, "will pthread_create\n");
1526                                 ret = pthread_create(&td->thread, NULL,
1527                                                         thread_main, td);
1528                                 if (ret) {
1529                                         log_err("pthread_create: %s\n",
1530                                                         strerror(ret));
1531                                         nr_started--;
1532                                         break;
1533                                 }
1534                                 ret = pthread_detach(td->thread);
1535                                 if (ret)
1536                                         log_err("pthread_detach: %s",
1537                                                         strerror(ret));
1538                         } else {
1539                                 pid_t pid;
1540                                 dprint(FD_PROCESS, "will fork\n");
1541                                 pid = fork();
1542                                 if (!pid) {
1543                                         int ret = fork_main(shm_id, i);
1544
1545                                         _exit(ret);
1546                                 } else if (i == fio_debug_jobno)
1547                                         *fio_debug_jobp = pid;
1548                         }
1549                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1550                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1551                                 log_err("fio: job startup hung? exiting.\n");
1552                                 terminate_threads(TERMINATE_ALL);
1553                                 fio_abort = 1;
1554                                 nr_started--;
1555                                 break;
1556                         }
1557                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1558                 }
1559
1560                 /*
1561                  * Wait for the started threads to transition to
1562                  * TD_INITIALIZED.
1563                  */
1564                 fio_gettime(&this_start, NULL);
1565                 left = this_jobs;
1566                 while (left && !fio_abort) {
1567                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1568                                 break;
1569
1570                         usleep(100000);
1571
1572                         for (i = 0; i < this_jobs; i++) {
1573                                 td = map[i];
1574                                 if (!td)
1575                                         continue;
1576                                 if (td->runstate == TD_INITIALIZED) {
1577                                         map[i] = NULL;
1578                                         left--;
1579                                 } else if (td->runstate >= TD_EXITED) {
1580                                         map[i] = NULL;
1581                                         left--;
1582                                         todo--;
1583                                         nr_running++; /* work-around... */
1584                                 }
1585                         }
1586                 }
1587
1588                 if (left) {
1589                         log_err("fio: %d jobs failed to start\n", left);
1590                         for (i = 0; i < this_jobs; i++) {
1591                                 td = map[i];
1592                                 if (!td)
1593                                         continue;
1594                                 kill(td->pid, SIGTERM);
1595                         }
1596                         break;
1597                 }
1598
1599                 /*
1600                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1601                  */
1602                 for_each_td(td, i) {
1603                         if (td->runstate != TD_INITIALIZED)
1604                                 continue;
1605
1606                         if (in_ramp_time(td))
1607                                 td_set_runstate(td, TD_RAMP);
1608                         else
1609                                 td_set_runstate(td, TD_RUNNING);
1610                         nr_running++;
1611                         nr_started--;
1612                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1613                         t_rate += td->o.rate[0] + td->o.rate[1];
1614                         todo--;
1615                         fio_mutex_up(td->mutex);
1616                 }
1617
1618                 reap_threads(&nr_running, &t_rate, &m_rate);
1619
1620                 if (todo)
1621                         usleep(100000);
1622         }
1623
1624         while (nr_running) {
1625                 reap_threads(&nr_running, &t_rate, &m_rate);
1626                 usleep(10000);
1627         }
1628
1629         update_io_ticks();
1630         fio_unpin_memory();
1631 }
1632
1633 int main(int argc, char *argv[])
1634 {
1635         long ps;
1636
1637         sinit();
1638         init_rand(&__fio_rand_state);
1639
1640         /*
1641          * We need locale for number printing, if it isn't set then just
1642          * go with the US format.
1643          */
1644         if (!getenv("LC_NUMERIC"))
1645                 setlocale(LC_NUMERIC, "en_US");
1646
1647         ps = sysconf(_SC_PAGESIZE);
1648         if (ps < 0) {
1649                 log_err("Failed to get page size\n");
1650                 return 1;
1651         }
1652
1653         page_size = ps;
1654         page_mask = ps - 1;
1655
1656         fio_keywords_init();
1657
1658         if (parse_options(argc, argv))
1659                 return 1;
1660
1661         if (exec_profile && load_profile(exec_profile))
1662                 return 1;
1663
1664         if (!thread_number)
1665                 return 0;
1666
1667         if (write_bw_log) {
1668                 setup_log(&agg_io_log[DDIR_READ]);
1669                 setup_log(&agg_io_log[DDIR_WRITE]);
1670         }
1671
1672         startup_mutex = fio_mutex_init(0);
1673         writeout_mutex = fio_mutex_init(1);
1674
1675         set_genesis_time();
1676
1677         status_timer_arm();
1678
1679         cgroup_list = smalloc(sizeof(*cgroup_list));
1680         INIT_FLIST_HEAD(cgroup_list);
1681
1682         run_threads();
1683
1684         if (!fio_abort) {
1685                 show_run_stats();
1686                 if (write_bw_log) {
1687                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1688                         __finish_log(agg_io_log[DDIR_WRITE],
1689                                         "agg-write_bw.log");
1690                 }
1691         }
1692
1693         cgroup_kill(cgroup_list);
1694         sfree(cgroup_list);
1695         sfree(cgroup_mnt);
1696
1697         fio_mutex_remove(startup_mutex);
1698         fio_mutex_remove(writeout_mutex);
1699         return exit_value;
1700 }