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
  49 unsigned long page_mask;
  50 unsigned long page_size;
  51
  52 #define PAGE_ALIGN(buf) \
  53         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  54
  55 int groupid = 0;
  56 int thread_number = 0;
  57 int nr_process = 0;
  58 int nr_thread = 0;
  59 int shm_id = 0;
  60 int temp_stall_ts;
  61 unsigned long done_secs = 0;
  62
  63 static struct fio_mutex *startup_mutex;
  64 static struct fio_mutex *writeout_mutex;
  65 static volatile int fio_abort;
  66 static int exit_value;
  67 static pthread_t gtod_thread;
  68 static pthread_t disk_util_thread;
  69 static struct flist_head *cgroup_list;
  70 static char *cgroup_mnt;
  71
  72 struct io_log *agg_io_log[2];
  73
  74 #define TERMINATE_ALL           (-1)
  75 #define JOB_START_TIMEOUT       (5 * 1000)
  76
  77 void td_set_runstate(struct thread_data *td, int runstate)
  78 {
  79         if (td->runstate == runstate)
  80                 return;
  81
  82         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  83                                                 td->runstate, runstate);
  84         td->runstate = runstate;
  85 }
  86
  87 static void terminate_threads(int group_id)
  88 {
  89         struct thread_data *td;
  90         int i;
  91
  92         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  93
  94         for_each_td(td, i) {
  95                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  96                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  97                                                 td->o.name, (int) td->pid);
  98                         td->terminate = 1;
  99                         td->o.start_delay = 0;
 100
 101                         /*
 102                          * if the thread is running, just let it exit
 103                          */
 104                         if (td->runstate < TD_RUNNING)
 105                                 kill(td->pid, SIGTERM);
 106                         else {
 107                                 struct ioengine_ops *ops = td->io_ops;
 108
 109                                 if (ops && (ops->flags & FIO_SIGTERM))
 110                                         kill(td->pid, SIGTERM);
 111                         }
 112                 }
 113         }
 114 }
 115
 116 /*
 117  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 118  */
 119 static void sig_quit(int sig)
 120 {
 121 }
 122
 123 static void sig_int(int sig)
 124 {
 125         if (threads) {
 126                 log_info("\nfio: terminating on signal %d\n", sig);
 127                 fflush(stdout);
 128                 terminate_threads(TERMINATE_ALL);
 129         }
 130 }
 131
 132 static void *disk_thread_main(void *data)
 133 {
 134         fio_mutex_up(startup_mutex);
 135
 136         while (threads) {
 137                 usleep(DISK_UTIL_MSEC * 1000);
 138                 if (!threads)
 139                         break;
 140                 update_io_ticks();
 141                 print_thread_status();
 142         }
 143
 144         return NULL;
 145 }
 146
 147 static int create_disk_util_thread(void)
 148 {
 149         int ret;
 150
 151         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 152         if (ret) {
 153                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 154                 return 1;
 155         }
 156
 157         ret = pthread_detach(disk_util_thread);
 158         if (ret) {
 159                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 160                 return 1;
 161         }
 162
 163         dprint(FD_MUTEX, "wait on startup_mutex\n");
 164         fio_mutex_down(startup_mutex);
 165         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 166         return 0;
 167 }
 168
 169 static void set_sig_handlers(void)
 170 {
 171         struct sigaction act;
 172
 173         memset(&act, 0, sizeof(act));
 174         act.sa_handler = sig_int;
 175         act.sa_flags = SA_RESTART;
 176         sigaction(SIGINT, &act, NULL);
 177
 178         memset(&act, 0, sizeof(act));
 179         act.sa_handler = sig_quit;
 180         act.sa_flags = SA_RESTART;
 181         sigaction(SIGTERM, &act, NULL);
 182 }
 183
 184 /*
 185  * Check if we are above the minimum rate given.
 186  */
 187 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 188                             enum td_ddir ddir)
 189 {
 190         unsigned long long bytes = 0;
 191         unsigned long iops = 0;
 192         unsigned long spent;
 193         unsigned long rate;
 194         unsigned int ratemin = 0;
 195         unsigned int rate_iops = 0;
 196         unsigned int rate_iops_min = 0;
 197
 198         assert(ddir_rw(ddir));
 199
 200         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 201                 return 0;
 202
 203         /*
 204          * allow a 2 second settle period in the beginning
 205          */
 206         if (mtime_since(&td->start, now) < 2000)
 207                 return 0;
 208
 209         iops += td->io_blocks[ddir];
 210         bytes += td->this_io_bytes[ddir];
 211         ratemin += td->o.ratemin[ddir];
 212         rate_iops += td->o.rate_iops[ddir];
 213         rate_iops_min += td->o.rate_iops_min[ddir];
 214
 215         /*
 216          * if rate blocks is set, sample is running
 217          */
 218         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 219                 spent = mtime_since(&td->lastrate[ddir], now);
 220                 if (spent < td->o.ratecycle)
 221                         return 0;
 222
 223                 if (td->o.rate[ddir]) {
 224                         /*
 225                          * check bandwidth specified rate
 226                          */
 227                         if (bytes < td->rate_bytes[ddir]) {
 228                                 log_err("%s: min rate %u not met\n", td->o.name,
 229                                                                 ratemin);
 230                                 return 1;
 231                         } else {
 232                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 233                                 if (rate < ratemin ||
 234                                     bytes < td->rate_bytes[ddir]) {
 235                                         log_err("%s: min rate %u not met, got"
 236                                                 " %luKB/sec\n", td->o.name,
 237                                                         ratemin, rate);
 238                                         return 1;
 239                                 }
 240                         }
 241                 } else {
 242                         /*
 243                          * checks iops specified rate
 244                          */
 245                         if (iops < rate_iops) {
 246                                 log_err("%s: min iops rate %u not met\n",
 247                                                 td->o.name, rate_iops);
 248                                 return 1;
 249                         } else {
 250                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 251                                 if (rate < rate_iops_min ||
 252                                     iops < td->rate_blocks[ddir]) {
 253                                         log_err("%s: min iops rate %u not met,"
 254                                                 " got %lu\n", td->o.name,
 255                                                         rate_iops_min, rate);
 256                                 }
 257                         }
 258                 }
 259         }
 260
 261         td->rate_bytes[ddir] = bytes;
 262         td->rate_blocks[ddir] = iops;
 263         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 264         return 0;
 265 }
 266
 267 static int check_min_rate(struct thread_data *td, struct timeval *now,
 268                           unsigned long *bytes_done)
 269 {
 270         int ret = 0;
 271
 272         if (bytes_done[0])
 273                 ret |= __check_min_rate(td, now, 0);
 274         if (bytes_done[1])
 275                 ret |= __check_min_rate(td, now, 1);
 276
 277         return ret;
 278 }
 279
 280 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 281 {
 282         if (!td->o.timeout)
 283                 return 0;
 284         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 285                 return 1;
 286
 287         return 0;
 288 }
 289
 290 /*
 291  * When job exits, we can cancel the in-flight IO if we are using async
 292  * io. Attempt to do so.
 293  */
 294 static void cleanup_pending_aio(struct thread_data *td)
 295 {
 296         struct flist_head *entry, *n;
 297         struct io_u *io_u;
 298         int r;
 299
 300         /*
 301          * get immediately available events, if any
 302          */
 303         r = io_u_queued_complete(td, 0, NULL);
 304         if (r < 0)
 305                 return;
 306
 307         /*
 308          * now cancel remaining active events
 309          */
 310         if (td->io_ops->cancel) {
 311                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 312                         io_u = flist_entry(entry, struct io_u, list);
 313
 314                         /*
 315                          * if the io_u isn't in flight, then that generally
 316                          * means someone leaked an io_u. complain but fix
 317                          * it up, so we don't stall here.
 318                          */
 319                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 320                                 log_err("fio: non-busy IO on busy list\n");
 321                                 put_io_u(td, io_u);
 322                         } else {
 323                                 r = td->io_ops->cancel(td, io_u);
 324                                 if (!r)
 325                                         put_io_u(td, io_u);
 326                         }
 327                 }
 328         }
 329
 330         if (td->cur_depth)
 331                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 332 }
 333
 334 /*
 335  * Helper to handle the final sync of a file. Works just like the normal
 336  * io path, just does everything sync.
 337  */
 338 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 339 {
 340         struct io_u *io_u = __get_io_u(td);
 341         int ret;
 342
 343         if (!io_u)
 344                 return 1;
 345
 346         io_u->ddir = DDIR_SYNC;
 347         io_u->file = f;
 348
 349         if (td_io_prep(td, io_u)) {
 350                 put_io_u(td, io_u);
 351                 return 1;
 352         }
 353
 354 requeue:
 355         ret = td_io_queue(td, io_u);
 356         if (ret < 0) {
 357                 td_verror(td, io_u->error, "td_io_queue");
 358                 put_io_u(td, io_u);
 359                 return 1;
 360         } else if (ret == FIO_Q_QUEUED) {
 361                 if (io_u_queued_complete(td, 1, NULL) < 0)
 362                         return 1;
 363         } else if (ret == FIO_Q_COMPLETED) {
 364                 if (io_u->error) {
 365                         td_verror(td, io_u->error, "td_io_queue");
 366                         return 1;
 367                 }
 368
 369                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 370                         return 1;
 371         } else if (ret == FIO_Q_BUSY) {
 372                 if (td_io_commit(td))
 373                         return 1;
 374                 goto requeue;
 375         }
 376
 377         return 0;
 378 }
 379
 380 static inline void update_tv_cache(struct thread_data *td)
 381 {
 382         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 383                 fio_gettime(&td->tv_cache, NULL);
 384 }
 385
 386 static int break_on_this_error(struct thread_data *td, int *retptr)
 387 {
 388         int ret = *retptr;
 389
 390         if (ret < 0 || td->error) {
 391                 int err;
 392
 393                 if (!td->o.continue_on_error)
 394                         return 1;
 395
 396                 if (ret < 0)
 397                         err = -ret;
 398                 else
 399                         err = td->error;
 400
 401                 if (td_non_fatal_error(err)) {
 402                         /*
 403                          * Continue with the I/Os in case of
 404                          * a non fatal error.
 405                          */
 406                         update_error_count(td, err);
 407                         td_clear_error(td);
 408                         *retptr = 0;
 409                         return 0;
 410                 } else if (td->o.fill_device && err == ENOSPC) {
 411                         /*
 412                          * We expect to hit this error if
 413                          * fill_device option is set.
 414                          */
 415                         td_clear_error(td);
 416                         td->terminate = 1;
 417                         return 1;
 418                 } else {
 419                         /*
 420                          * Stop the I/O in case of a fatal
 421                          * error.
 422                          */
 423                         update_error_count(td, err);
 424                         return 1;
 425                 }
 426         }
 427
 428         return 0;
 429 }
 430
 431 /*
 432  * The main verify engine. Runs over the writes we previously submitted,
 433  * reads the blocks back in, and checks the crc/md5 of the data.
 434  */
 435 static void do_verify(struct thread_data *td)
 436 {
 437         struct fio_file *f;
 438         struct io_u *io_u;
 439         int ret, min_events;
 440         unsigned int i;
 441
 442         dprint(FD_VERIFY, "starting loop\n");
 443
 444         /*
 445          * sync io first and invalidate cache, to make sure we really
 446          * read from disk.
 447          */
 448         for_each_file(td, f, i) {
 449                 if (!fio_file_open(f))
 450                         continue;
 451                 if (fio_io_sync(td, f))
 452                         break;
 453                 if (file_invalidate_cache(td, f))
 454                         break;
 455         }
 456
 457         if (td->error)
 458                 return;
 459
 460         td_set_runstate(td, TD_VERIFYING);
 461
 462         io_u = NULL;
 463         while (!td->terminate) {
 464                 int ret2, full;
 465
 466                 update_tv_cache(td);
 467
 468                 if (runtime_exceeded(td, &td->tv_cache)) {
 469                         td->terminate = 1;
 470                         break;
 471                 }
 472
 473                 io_u = __get_io_u(td);
 474                 if (!io_u)
 475                         break;
 476
 477                 if (get_next_verify(td, io_u)) {
 478                         put_io_u(td, io_u);
 479                         break;
 480                 }
 481
 482                 if (td_io_prep(td, io_u)) {
 483                         put_io_u(td, io_u);
 484                         break;
 485                 }
 486
 487                 if (td->o.verify_async)
 488                         io_u->end_io = verify_io_u_async;
 489                 else
 490                         io_u->end_io = verify_io_u;
 491
 492                 ret = td_io_queue(td, io_u);
 493                 switch (ret) {
 494                 case FIO_Q_COMPLETED:
 495                         if (io_u->error) {
 496                                 ret = -io_u->error;
 497                                 clear_io_u(td, io_u);
 498                         } else if (io_u->resid) {
 499                                 int bytes = io_u->xfer_buflen - io_u->resid;
 500
 501                                 /*
 502                                  * zero read, fail
 503                                  */
 504                                 if (!bytes) {
 505                                         td_verror(td, EIO, "full resid");
 506                                         put_io_u(td, io_u);
 507                                         break;
 508                                 }
 509
 510                                 io_u->xfer_buflen = io_u->resid;
 511                                 io_u->xfer_buf += bytes;
 512                                 io_u->offset += bytes;
 513
 514                                 if (ddir_rw(io_u->ddir))
 515                                         td->ts.short_io_u[io_u->ddir]++;
 516
 517                                 f = io_u->file;
 518                                 if (io_u->offset == f->real_file_size)
 519                                         goto sync_done;
 520
 521                                 requeue_io_u(td, &io_u);
 522                         } else {
 523 sync_done:
 524                                 ret = io_u_sync_complete(td, io_u, NULL);
 525                                 if (ret < 0)
 526                                         break;
 527                         }
 528                         continue;
 529                 case FIO_Q_QUEUED:
 530                         break;
 531                 case FIO_Q_BUSY:
 532                         requeue_io_u(td, &io_u);
 533                         ret2 = td_io_commit(td);
 534                         if (ret2 < 0)
 535                                 ret = ret2;
 536                         break;
 537                 default:
 538                         assert(ret < 0);
 539                         td_verror(td, -ret, "td_io_queue");
 540                         break;
 541                 }
 542
 543                 if (break_on_this_error(td, &ret))
 544                         break;
 545
 546                 /*
 547                  * if we can queue more, do so. but check if there are
 548                  * completed io_u's first.
 549                  */
 550                 full = queue_full(td) || ret == FIO_Q_BUSY;
 551                 if (full || !td->o.iodepth_batch_complete) {
 552                         min_events = min(td->o.iodepth_batch_complete,
 553                                          td->cur_depth);
 554                         if (full && !min_events)
 555                                 min_events = 1;
 556
 557                         do {
 558                                 /*
 559                                  * Reap required number of io units, if any,
 560                                  * and do the verification on them through
 561                                  * the callback handler
 562                                  */
 563                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 564                                         ret = -1;
 565                                         break;
 566                                 }
 567                         } while (full && (td->cur_depth > td->o.iodepth_low));
 568                 }
 569                 if (ret < 0)
 570                         break;
 571         }
 572
 573         if (!td->error) {
 574                 min_events = td->cur_depth;
 575
 576                 if (min_events)
 577                         ret = io_u_queued_complete(td, min_events, NULL);
 578         } else
 579                 cleanup_pending_aio(td);
 580
 581         td_set_runstate(td, TD_RUNNING);
 582
 583         dprint(FD_VERIFY, "exiting loop\n");
 584 }
 585
 586 /*
 587  * Main IO worker function. It retrieves io_u's to process and queues
 588  * and reaps them, checking for rate and errors along the way.
 589  */
 590 static void do_io(struct thread_data *td)
 591 {
 592         unsigned int i;
 593         int ret = 0;
 594
 595         if (in_ramp_time(td))
 596                 td_set_runstate(td, TD_RAMP);
 597         else
 598                 td_set_runstate(td, TD_RUNNING);
 599
 600         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 601                 (!flist_empty(&td->trim_list)) ||
 602                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 603                 struct timeval comp_time;
 604                 unsigned long bytes_done[2] = { 0, 0 };
 605                 int min_evts = 0;
 606                 struct io_u *io_u;
 607                 int ret2, full;
 608
 609                 if (td->terminate)
 610                         break;
 611
 612                 update_tv_cache(td);
 613
 614                 if (runtime_exceeded(td, &td->tv_cache)) {
 615                         td->terminate = 1;
 616                         break;
 617                 }
 618
 619                 io_u = get_io_u(td);
 620                 if (!io_u)
 621                         break;
 622
 623                 /*
 624                  * Add verification end_io handler, if asked to verify
 625                  * a previously written file.
 626                  */
 627                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 628                     !td_rw(td)) {
 629                         if (td->o.verify_async)
 630                                 io_u->end_io = verify_io_u_async;
 631                         else
 632                                 io_u->end_io = verify_io_u;
 633                         td_set_runstate(td, TD_VERIFYING);
 634                 } else if (in_ramp_time(td))
 635                         td_set_runstate(td, TD_RAMP);
 636                 else
 637                         td_set_runstate(td, TD_RUNNING);
 638
 639                 ret = td_io_queue(td, io_u);
 640                 switch (ret) {
 641                 case FIO_Q_COMPLETED:
 642                         if (io_u->error) {
 643                                 ret = -io_u->error;
 644                                 clear_io_u(td, io_u);
 645                         } else if (io_u->resid) {
 646                                 int bytes = io_u->xfer_buflen - io_u->resid;
 647                                 struct fio_file *f = io_u->file;
 648
 649                                 /*
 650                                  * zero read, fail
 651                                  */
 652                                 if (!bytes) {
 653                                         td_verror(td, EIO, "full resid");
 654                                         put_io_u(td, io_u);
 655                                         break;
 656                                 }
 657
 658                                 io_u->xfer_buflen = io_u->resid;
 659                                 io_u->xfer_buf += bytes;
 660                                 io_u->offset += bytes;
 661
 662                                 if (ddir_rw(io_u->ddir))
 663                                         td->ts.short_io_u[io_u->ddir]++;
 664
 665                                 if (io_u->offset == f->real_file_size)
 666                                         goto sync_done;
 667
 668                                 requeue_io_u(td, &io_u);
 669                         } else {
 670 sync_done:
 671                                 if (__should_check_rate(td, 0) ||
 672                                     __should_check_rate(td, 1))
 673                                         fio_gettime(&comp_time, NULL);
 674
 675                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 676                                 if (ret < 0)
 677                                         break;
 678                         }
 679                         break;
 680                 case FIO_Q_QUEUED:
 681                         /*
 682                          * if the engine doesn't have a commit hook,
 683                          * the io_u is really queued. if it does have such
 684                          * a hook, it has to call io_u_queued() itself.
 685                          */
 686                         if (td->io_ops->commit == NULL)
 687                                 io_u_queued(td, io_u);
 688                         break;
 689                 case FIO_Q_BUSY:
 690                         requeue_io_u(td, &io_u);
 691                         ret2 = td_io_commit(td);
 692                         if (ret2 < 0)
 693                                 ret = ret2;
 694                         break;
 695                 default:
 696                         assert(ret < 0);
 697                         put_io_u(td, io_u);
 698                         break;
 699                 }
 700
 701                 if (break_on_this_error(td, &ret))
 702                         break;
 703
 704                 /*
 705                  * See if we need to complete some commands
 706                  */
 707                 full = queue_full(td) || ret == FIO_Q_BUSY;
 708                 if (full || !td->o.iodepth_batch_complete) {
 709                         min_evts = min(td->o.iodepth_batch_complete,
 710                                         td->cur_depth);
 711                         if (full && !min_evts)
 712                                 min_evts = 1;
 713
 714                         if (__should_check_rate(td, 0) ||
 715                             __should_check_rate(td, 1))
 716                                 fio_gettime(&comp_time, NULL);
 717
 718                         do {
 719                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 720                                 if (ret < 0)
 721                                         break;
 722
 723                         } while (full && (td->cur_depth > td->o.iodepth_low));
 724                 }
 725
 726                 if (ret < 0)
 727                         break;
 728                 if (!(bytes_done[0] + bytes_done[1]))
 729                         continue;
 730
 731                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 732                         if (check_min_rate(td, &comp_time, bytes_done)) {
 733                                 if (exitall_on_terminate)
 734                                         terminate_threads(td->groupid);
 735                                 td_verror(td, EIO, "check_min_rate");
 736                                 break;
 737                         }
 738                 }
 739
 740                 if (td->o.thinktime) {
 741                         unsigned long long b;
 742
 743                         b = td->io_blocks[0] + td->io_blocks[1];
 744                         if (!(b % td->o.thinktime_blocks)) {
 745                                 int left;
 746
 747                                 if (td->o.thinktime_spin)
 748                                         usec_spin(td->o.thinktime_spin);
 749
 750                                 left = td->o.thinktime - td->o.thinktime_spin;
 751                                 if (left)
 752                                         usec_sleep(td, left);
 753                         }
 754                 }
 755         }
 756
 757         if (td->trim_entries)
 758                 printf("trim entries %ld\n", td->trim_entries);
 759
 760         if (td->o.fill_device && td->error == ENOSPC) {
 761                 td->error = 0;
 762                 td->terminate = 1;
 763         }
 764         if (!td->error) {
 765                 struct fio_file *f;
 766
 767                 i = td->cur_depth;
 768                 if (i)
 769                         ret = io_u_queued_complete(td, i, NULL);
 770
 771                 if (should_fsync(td) && td->o.end_fsync) {
 772                         td_set_runstate(td, TD_FSYNCING);
 773
 774                         for_each_file(td, f, i) {
 775                                 if (!fio_file_open(f))
 776                                         continue;
 777                                 fio_io_sync(td, f);
 778                         }
 779                 }
 780         } else
 781                 cleanup_pending_aio(td);
 782
 783         /*
 784          * stop job if we failed doing any IO
 785          */
 786         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 787                 td->done = 1;
 788 }
 789
 790 static void cleanup_io_u(struct thread_data *td)
 791 {
 792         struct flist_head *entry, *n;
 793         struct io_u *io_u;
 794
 795         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 796                 io_u = flist_entry(entry, struct io_u, list);
 797
 798                 flist_del(&io_u->list);
 799                 free(io_u);
 800         }
 801
 802         free_io_mem(td);
 803 }
 804
 805 static int init_io_u(struct thread_data *td)
 806 {
 807         struct io_u *io_u;
 808         unsigned int max_bs;
 809         int cl_align, i, max_units;
 810         char *p;
 811
 812         max_units = td->o.iodepth;
 813         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 814         td->orig_buffer_size = (unsigned long long) max_bs
 815                                         * (unsigned long long) max_units;
 816
 817         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 818                 unsigned long bs;
 819
 820                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 821                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 822         }
 823
 824         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 825                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 826                 return 1;
 827         }
 828
 829         if (allocate_io_mem(td))
 830                 return 1;
 831
 832         if (td->o.odirect || td->o.mem_align ||
 833             (td->io_ops->flags & FIO_RAWIO))
 834                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 835         else
 836                 p = td->orig_buffer;
 837
 838         cl_align = os_cache_line_size();
 839
 840         for (i = 0; i < max_units; i++) {
 841                 void *ptr;
 842
 843                 if (td->terminate)
 844                         return 1;
 845
 846                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
 847                         log_err("fio: posix_memalign=%s\n", strerror(errno));
 848                         break;
 849                 }
 850
 851                 io_u = ptr;
 852                 memset(io_u, 0, sizeof(*io_u));
 853                 INIT_FLIST_HEAD(&io_u->list);
 854                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 855
 856                 if (!(td->io_ops->flags & FIO_NOIO)) {
 857                         io_u->buf = p + max_bs * i;
 858                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 859
 860                         if (td_write(td) && !td->o.refill_buffers)
 861                                 io_u_fill_buffer(td, io_u, max_bs);
 862                         else if (td_write(td) && td->o.verify_pattern_bytes) {
 863                                 /*
 864                                  * Fill the buffer with the pattern if we are
 865                                  * going to be doing writes.
 866                                  */
 867                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 868                         }
 869                 }
 870
 871                 io_u->index = i;
 872                 io_u->flags = IO_U_F_FREE;
 873                 flist_add(&io_u->list, &td->io_u_freelist);
 874         }
 875
 876         return 0;
 877 }
 878
 879 static int switch_ioscheduler(struct thread_data *td)
 880 {
 881         char tmp[256], tmp2[128];
 882         FILE *f;
 883         int ret;
 884
 885         if (td->io_ops->flags & FIO_DISKLESSIO)
 886                 return 0;
 887
 888         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 889
 890         f = fopen(tmp, "r+");
 891         if (!f) {
 892                 if (errno == ENOENT) {
 893                         log_err("fio: os or kernel doesn't support IO scheduler"
 894                                 " switching\n");
 895                         return 0;
 896                 }
 897                 td_verror(td, errno, "fopen iosched");
 898                 return 1;
 899         }
 900
 901         /*
 902          * Set io scheduler.
 903          */
 904         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 905         if (ferror(f) || ret != 1) {
 906                 td_verror(td, errno, "fwrite");
 907                 fclose(f);
 908                 return 1;
 909         }
 910
 911         rewind(f);
 912
 913         /*
 914          * Read back and check that the selected scheduler is now the default.
 915          */
 916         ret = fread(tmp, 1, sizeof(tmp), f);
 917         if (ferror(f) || ret < 0) {
 918                 td_verror(td, errno, "fread");
 919                 fclose(f);
 920                 return 1;
 921         }
 922
 923         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 924         if (!strstr(tmp, tmp2)) {
 925                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 926                 td_verror(td, EINVAL, "iosched_switch");
 927                 fclose(f);
 928                 return 1;
 929         }
 930
 931         fclose(f);
 932         return 0;
 933 }
 934
 935 static int keep_running(struct thread_data *td)
 936 {
 937         unsigned long long io_done;
 938
 939         if (td->done)
 940                 return 0;
 941         if (td->o.time_based)
 942                 return 1;
 943         if (td->o.loops) {
 944                 td->o.loops--;
 945                 return 1;
 946         }
 947
 948         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 949                         + td->io_skip_bytes;
 950         if (io_done < td->o.size)
 951                 return 1;
 952
 953         return 0;
 954 }
 955
 956 static void reset_io_counters(struct thread_data *td)
 957 {
 958         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 959         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 960         td->zone_bytes = 0;
 961         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 962         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 963
 964         td->last_was_sync = 0;
 965
 966         /*
 967          * reset file done count if we are to start over
 968          */
 969         if (td->o.time_based || td->o.loops)
 970                 td->nr_done_files = 0;
 971
 972         /*
 973          * Set the same seed to get repeatable runs
 974          */
 975         td_fill_rand_seeds(td);
 976 }
 977
 978 void reset_all_stats(struct thread_data *td)
 979 {
 980         struct timeval tv;
 981         int i;
 982
 983         reset_io_counters(td);
 984
 985         for (i = 0; i < 2; i++) {
 986                 td->io_bytes[i] = 0;
 987                 td->io_blocks[i] = 0;
 988                 td->io_issues[i] = 0;
 989                 td->ts.total_io_u[i] = 0;
 990         }
 991
 992         fio_gettime(&tv, NULL);
 993         td->ts.runtime[0] = 0;
 994         td->ts.runtime[1] = 0;
 995         memcpy(&td->epoch, &tv, sizeof(tv));
 996         memcpy(&td->start, &tv, sizeof(tv));
 997 }
 998
 999 static void clear_io_state(struct thread_data *td)
1000 {
1001         struct fio_file *f;
1002         unsigned int i;
1003
1004         reset_io_counters(td);
1005
1006         close_files(td);
1007         for_each_file(td, f, i)
1008                 fio_file_clear_done(f);
1009 }
1010
1011 static int exec_string(const char *string)
1012 {
1013         int ret, newlen = strlen(string) + 1 + 8;
1014         char *str;
1015
1016         str = malloc(newlen);
1017         sprintf(str, "sh -c %s", string);
1018
1019         ret = system(str);
1020         if (ret == -1)
1021                 log_err("fio: exec of cmd <%s> failed\n", str);
1022
1023         free(str);
1024         return ret;
1025 }
1026
1027 /*
1028  * Entry point for the thread based jobs. The process based jobs end up
1029  * here as well, after a little setup.
1030  */
1031 static void *thread_main(void *data)
1032 {
1033         unsigned long long elapsed;
1034         struct thread_data *td = data;
1035         pthread_condattr_t attr;
1036         int clear_state;
1037
1038         if (!td->o.use_thread)
1039                 setsid();
1040
1041         td->pid = getpid();
1042
1043         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1044
1045         INIT_FLIST_HEAD(&td->io_u_freelist);
1046         INIT_FLIST_HEAD(&td->io_u_busylist);
1047         INIT_FLIST_HEAD(&td->io_u_requeues);
1048         INIT_FLIST_HEAD(&td->io_log_list);
1049         INIT_FLIST_HEAD(&td->io_hist_list);
1050         INIT_FLIST_HEAD(&td->verify_list);
1051         INIT_FLIST_HEAD(&td->trim_list);
1052         pthread_mutex_init(&td->io_u_lock, NULL);
1053         td->io_hist_tree = RB_ROOT;
1054
1055         pthread_condattr_init(&attr);
1056         pthread_cond_init(&td->verify_cond, &attr);
1057         pthread_cond_init(&td->free_cond, &attr);
1058
1059         td_set_runstate(td, TD_INITIALIZED);
1060         dprint(FD_MUTEX, "up startup_mutex\n");
1061         fio_mutex_up(startup_mutex);
1062         dprint(FD_MUTEX, "wait on td->mutex\n");
1063         fio_mutex_down(td->mutex);
1064         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1065
1066         /*
1067          * the ->mutex mutex is now no longer used, close it to avoid
1068          * eating a file descriptor
1069          */
1070         fio_mutex_remove(td->mutex);
1071
1072         /*
1073          * A new gid requires privilege, so we need to do this before setting
1074          * the uid.
1075          */
1076         if (td->o.gid != -1U && setgid(td->o.gid)) {
1077                 td_verror(td, errno, "setgid");
1078                 goto err;
1079         }
1080         if (td->o.uid != -1U && setuid(td->o.uid)) {
1081                 td_verror(td, errno, "setuid");
1082                 goto err;
1083         }
1084
1085         /*
1086          * May alter parameters that init_io_u() will use, so we need to
1087          * do this first.
1088          */
1089         if (init_iolog(td))
1090                 goto err;
1091
1092         if (init_io_u(td))
1093                 goto err;
1094
1095         if (td->o.verify_async && verify_async_init(td))
1096                 goto err;
1097
1098         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1099                 td_verror(td, errno, "cpu_set_affinity");
1100                 goto err;
1101         }
1102
1103         /*
1104          * If we have a gettimeofday() thread, make sure we exclude that
1105          * thread from this job
1106          */
1107         if (td->o.gtod_cpu) {
1108                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1109                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1110                         td_verror(td, errno, "cpu_set_affinity");
1111                         goto err;
1112                 }
1113         }
1114
1115         if (td->ioprio_set) {
1116                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1117                         td_verror(td, errno, "ioprio_set");
1118                         goto err;
1119                 }
1120         }
1121
1122         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1123                 goto err;
1124
1125         if (nice(td->o.nice) == -1) {
1126                 td_verror(td, errno, "nice");
1127                 goto err;
1128         }
1129
1130         if (td->o.ioscheduler && switch_ioscheduler(td))
1131                 goto err;
1132
1133         if (!td->o.create_serialize && setup_files(td))
1134                 goto err;
1135
1136         if (td_io_init(td))
1137                 goto err;
1138
1139         if (init_random_map(td))
1140                 goto err;
1141
1142         if (td->o.exec_prerun) {
1143                 if (exec_string(td->o.exec_prerun))
1144                         goto err;
1145         }
1146
1147         if (td->o.pre_read) {
1148                 if (pre_read_files(td) < 0)
1149                         goto err;
1150         }
1151
1152         fio_gettime(&td->epoch, NULL);
1153         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1154
1155         clear_state = 0;
1156         while (keep_running(td)) {
1157                 fio_gettime(&td->start, NULL);
1158                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1159                                 sizeof(td->start));
1160                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1161                                 sizeof(td->start));
1162                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1163
1164                 if (td->o.ratemin[0] || td->o.ratemin[1])
1165                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1166                                                         sizeof(td->lastrate));
1167
1168                 if (clear_state)
1169                         clear_io_state(td);
1170
1171                 prune_io_piece_log(td);
1172
1173                 do_io(td);
1174
1175                 clear_state = 1;
1176
1177                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1178                         elapsed = utime_since_now(&td->start);
1179                         td->ts.runtime[DDIR_READ] += elapsed;
1180                 }
1181                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1182                         elapsed = utime_since_now(&td->start);
1183                         td->ts.runtime[DDIR_WRITE] += elapsed;
1184                 }
1185
1186                 if (td->error || td->terminate)
1187                         break;
1188
1189                 if (!td->o.do_verify ||
1190                     td->o.verify == VERIFY_NONE ||
1191                     (td->io_ops->flags & FIO_UNIDIR))
1192                         continue;
1193
1194                 clear_io_state(td);
1195
1196                 fio_gettime(&td->start, NULL);
1197
1198                 do_verify(td);
1199
1200                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1201
1202                 if (td->error || td->terminate)
1203                         break;
1204         }
1205
1206         update_rusage_stat(td);
1207         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1208         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1209         td->ts.total_run_time = mtime_since_now(&td->epoch);
1210         td->ts.io_bytes[0] = td->io_bytes[0];
1211         td->ts.io_bytes[1] = td->io_bytes[1];
1212
1213         fio_mutex_down(writeout_mutex);
1214         if (td->ts.bw_log) {
1215                 if (td->o.bw_log_file) {
1216                         finish_log_named(td, td->ts.bw_log,
1217                                                 td->o.bw_log_file, "bw");
1218                 } else
1219                         finish_log(td, td->ts.bw_log, "bw");
1220         }
1221         if (td->ts.lat_log) {
1222                 if (td->o.lat_log_file) {
1223                         finish_log_named(td, td->ts.lat_log,
1224                                                 td->o.lat_log_file, "lat");
1225                 } else
1226                         finish_log(td, td->ts.lat_log, "lat");
1227         }
1228         if (td->ts.slat_log) {
1229                 if (td->o.lat_log_file) {
1230                         finish_log_named(td, td->ts.slat_log,
1231                                                 td->o.lat_log_file, "slat");
1232                 } else
1233                         finish_log(td, td->ts.slat_log, "slat");
1234         }
1235         if (td->ts.clat_log) {
1236                 if (td->o.lat_log_file) {
1237                         finish_log_named(td, td->ts.clat_log,
1238                                                 td->o.lat_log_file, "clat");
1239                 } else
1240                         finish_log(td, td->ts.clat_log, "clat");
1241         }
1242         fio_mutex_up(writeout_mutex);
1243         if (td->o.exec_postrun)
1244                 exec_string(td->o.exec_postrun);
1245
1246         if (exitall_on_terminate)
1247                 terminate_threads(td->groupid);
1248
1249 err:
1250         if (td->error)
1251                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1252                                                         td->verror);
1253
1254         if (td->o.verify_async)
1255                 verify_async_exit(td);
1256
1257         close_and_free_files(td);
1258         close_ioengine(td);
1259         cleanup_io_u(td);
1260         cgroup_shutdown(td, &cgroup_mnt);
1261
1262         if (td->o.cpumask_set) {
1263                 int ret = fio_cpuset_exit(&td->o.cpumask);
1264
1265                 td_verror(td, ret, "fio_cpuset_exit");
1266         }
1267
1268         /*
1269          * do this very late, it will log file closing as well
1270          */
1271         if (td->o.write_iolog_file)
1272                 write_iolog_close(td);
1273
1274         options_mem_free(td);
1275         td_set_runstate(td, TD_EXITED);
1276         return (void *) (unsigned long) td->error;
1277 }
1278
1279 /*
1280  * We cannot pass the td data into a forked process, so attach the td and
1281  * pass it to the thread worker.
1282  */
1283 static int fork_main(int shmid, int offset)
1284 {
1285         struct thread_data *td;
1286         void *data, *ret;
1287
1288         data = shmat(shmid, NULL, 0);
1289         if (data == (void *) -1) {
1290                 int __err = errno;
1291
1292                 perror("shmat");
1293                 return __err;
1294         }
1295
1296         td = data + offset * sizeof(struct thread_data);
1297         ret = thread_main(td);
1298         shmdt(data);
1299         return (int) (unsigned long) ret;
1300 }
1301
1302 /*
1303  * Run over the job map and reap the threads that have exited, if any.
1304  */
1305 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1306 {
1307         struct thread_data *td;
1308         int i, cputhreads, realthreads, pending, status, ret;
1309
1310         /*
1311          * reap exited threads (TD_EXITED -> TD_REAPED)
1312          */
1313         realthreads = pending = cputhreads = 0;
1314         for_each_td(td, i) {
1315                 int flags = 0;
1316
1317                 /*
1318                  * ->io_ops is NULL for a thread that has closed its
1319                  * io engine
1320                  */
1321                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1322                         cputhreads++;
1323                 else
1324                         realthreads++;
1325
1326                 if (!td->pid) {
1327                         pending++;
1328                         continue;
1329                 }
1330                 if (td->runstate == TD_REAPED)
1331                         continue;
1332                 if (td->o.use_thread) {
1333                         if (td->runstate == TD_EXITED) {
1334                                 td_set_runstate(td, TD_REAPED);
1335                                 goto reaped;
1336                         }
1337                         continue;
1338                 }
1339
1340                 flags = WNOHANG;
1341                 if (td->runstate == TD_EXITED)
1342                         flags = 0;
1343
1344                 /*
1345                  * check if someone quit or got killed in an unusual way
1346                  */
1347                 ret = waitpid(td->pid, &status, flags);
1348                 if (ret < 0) {
1349                         if (errno == ECHILD) {
1350                                 log_err("fio: pid=%d disappeared %d\n",
1351                                                 (int) td->pid, td->runstate);
1352                                 td_set_runstate(td, TD_REAPED);
1353                                 goto reaped;
1354                         }
1355                         perror("waitpid");
1356                 } else if (ret == td->pid) {
1357                         if (WIFSIGNALED(status)) {
1358                                 int sig = WTERMSIG(status);
1359
1360                                 if (sig != SIGTERM)
1361                                         log_err("fio: pid=%d, got signal=%d\n",
1362                                                         (int) td->pid, sig);
1363                                 td_set_runstate(td, TD_REAPED);
1364                                 goto reaped;
1365                         }
1366                         if (WIFEXITED(status)) {
1367                                 if (WEXITSTATUS(status) && !td->error)
1368                                         td->error = WEXITSTATUS(status);
1369
1370                                 td_set_runstate(td, TD_REAPED);
1371                                 goto reaped;
1372                         }
1373                 }
1374
1375                 /*
1376                  * thread is not dead, continue
1377                  */
1378                 pending++;
1379                 continue;
1380 reaped:
1381                 (*nr_running)--;
1382                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1383                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1384                 if (!td->pid)
1385                         pending--;
1386
1387                 if (td->error)
1388                         exit_value++;
1389
1390                 done_secs += mtime_since_now(&td->epoch) / 1000;
1391         }
1392
1393         if (*nr_running == cputhreads && !pending && realthreads)
1394                 terminate_threads(TERMINATE_ALL);
1395 }
1396
1397 static void *gtod_thread_main(void *data)
1398 {
1399         fio_mutex_up(startup_mutex);
1400
1401         /*
1402          * As long as we have jobs around, update the clock. It would be nice
1403          * to have some way of NOT hammering that CPU with gettimeofday(),
1404          * but I'm not sure what to use outside of a simple CPU nop to relax
1405          * it - we don't want to lose precision.
1406          */
1407         while (threads) {
1408                 fio_gtod_update();
1409                 nop;
1410         }
1411
1412         return NULL;
1413 }
1414
1415 static int fio_start_gtod_thread(void)
1416 {
1417         pthread_attr_t attr;
1418         int ret;
1419
1420         pthread_attr_init(&attr);
1421         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1422         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1423         pthread_attr_destroy(&attr);
1424         if (ret) {
1425                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1426                 return 1;
1427         }
1428
1429         ret = pthread_detach(gtod_thread);
1430         if (ret) {
1431                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1432                 return 1;
1433         }
1434
1435         dprint(FD_MUTEX, "wait on startup_mutex\n");
1436         fio_mutex_down(startup_mutex);
1437         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1438         return 0;
1439 }
1440
1441 /*
1442  * Main function for kicking off and reaping jobs, as needed.
1443  */
1444 static void run_threads(void)
1445 {
1446         struct thread_data *td;
1447         unsigned long spent;
1448         int i, todo, nr_running, m_rate, t_rate, nr_started;
1449
1450         if (fio_pin_memory())
1451                 return;
1452
1453         if (fio_gtod_offload && fio_start_gtod_thread())
1454                 return;
1455
1456         if (!terse_output) {
1457                 log_info("Starting ");
1458                 if (nr_thread)
1459                         log_info("%d thread%s", nr_thread,
1460                                                 nr_thread > 1 ? "s" : "");
1461                 if (nr_process) {
1462                         if (nr_thread)
1463                                 printf(" and ");
1464                         log_info("%d process%s", nr_process,
1465                                                 nr_process > 1 ? "es" : "");
1466                 }
1467                 log_info("\n");
1468                 fflush(stdout);
1469         }
1470
1471         set_sig_handlers();
1472
1473         todo = thread_number;
1474         nr_running = 0;
1475         nr_started = 0;
1476         m_rate = t_rate = 0;
1477
1478         for_each_td(td, i) {
1479                 print_status_init(td->thread_number - 1);
1480
1481                 if (!td->o.create_serialize) {
1482                         init_disk_util(td);
1483                         continue;
1484                 }
1485
1486                 /*
1487                  * do file setup here so it happens sequentially,
1488                  * we don't want X number of threads getting their
1489                  * client data interspersed on disk
1490                  */
1491                 if (setup_files(td)) {
1492                         exit_value++;
1493                         if (td->error)
1494                                 log_err("fio: pid=%d, err=%d/%s\n",
1495                                         (int) td->pid, td->error, td->verror);
1496                         td_set_runstate(td, TD_REAPED);
1497                         todo--;
1498                 } else {
1499                         struct fio_file *f;
1500                         unsigned int j;
1501
1502                         /*
1503                          * for sharing to work, each job must always open
1504                          * its own files. so close them, if we opened them
1505                          * for creation
1506                          */
1507                         for_each_file(td, f, j) {
1508                                 if (fio_file_open(f))
1509                                         td_io_close_file(td, f);
1510                         }
1511                 }
1512
1513                 init_disk_util(td);
1514         }
1515
1516         set_genesis_time();
1517
1518         while (todo) {
1519                 struct thread_data *map[MAX_JOBS];
1520                 struct timeval this_start;
1521                 int this_jobs = 0, left;
1522
1523                 /*
1524                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1525                  */
1526                 for_each_td(td, i) {
1527                         if (td->runstate != TD_NOT_CREATED)
1528                                 continue;
1529
1530                         /*
1531                          * never got a chance to start, killed by other
1532                          * thread for some reason
1533                          */
1534                         if (td->terminate) {
1535                                 todo--;
1536                                 continue;
1537                         }
1538
1539                         if (td->o.start_delay) {
1540                                 spent = mtime_since_genesis();
1541
1542                                 if (td->o.start_delay * 1000 > spent)
1543                                         continue;
1544                         }
1545
1546                         if (td->o.stonewall && (nr_started || nr_running)) {
1547                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1548                                                         td->o.name);
1549                                 break;
1550                         }
1551
1552                         /*
1553                          * Set state to created. Thread will transition
1554                          * to TD_INITIALIZED when it's done setting up.
1555                          */
1556                         td_set_runstate(td, TD_CREATED);
1557                         map[this_jobs++] = td;
1558                         nr_started++;
1559
1560                         if (td->o.use_thread) {
1561                                 int ret;
1562
1563                                 dprint(FD_PROCESS, "will pthread_create\n");
1564                                 ret = pthread_create(&td->thread, NULL,
1565                                                         thread_main, td);
1566                                 if (ret) {
1567                                         log_err("pthread_create: %s\n",
1568                                                         strerror(ret));
1569                                         nr_started--;
1570                                         break;
1571                                 }
1572                                 ret = pthread_detach(td->thread);
1573                                 if (ret)
1574                                         log_err("pthread_detach: %s",
1575                                                         strerror(ret));
1576                         } else {
1577                                 pid_t pid;
1578                                 dprint(FD_PROCESS, "will fork\n");
1579                                 pid = fork();
1580                                 if (!pid) {
1581                                         int ret = fork_main(shm_id, i);
1582
1583                                         _exit(ret);
1584                                 } else if (i == fio_debug_jobno)
1585                                         *fio_debug_jobp = pid;
1586                         }
1587                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1588                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1589                                 log_err("fio: job startup hung? exiting.\n");
1590                                 terminate_threads(TERMINATE_ALL);
1591                                 fio_abort = 1;
1592                                 nr_started--;
1593                                 break;
1594                         }
1595                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1596                 }
1597
1598                 /*
1599                  * Wait for the started threads to transition to
1600                  * TD_INITIALIZED.
1601                  */
1602                 fio_gettime(&this_start, NULL);
1603                 left = this_jobs;
1604                 while (left && !fio_abort) {
1605                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1606                                 break;
1607
1608                         usleep(100000);
1609
1610                         for (i = 0; i < this_jobs; i++) {
1611                                 td = map[i];
1612                                 if (!td)
1613                                         continue;
1614                                 if (td->runstate == TD_INITIALIZED) {
1615                                         map[i] = NULL;
1616                                         left--;
1617                                 } else if (td->runstate >= TD_EXITED) {
1618                                         map[i] = NULL;
1619                                         left--;
1620                                         todo--;
1621                                         nr_running++; /* work-around... */
1622                                 }
1623                         }
1624                 }
1625
1626                 if (left) {
1627                         log_err("fio: %d jobs failed to start\n", left);
1628                         for (i = 0; i < this_jobs; i++) {
1629                                 td = map[i];
1630                                 if (!td)
1631                                         continue;
1632                                 kill(td->pid, SIGTERM);
1633                         }
1634                         break;
1635                 }
1636
1637                 /*
1638                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1639                  */
1640                 for_each_td(td, i) {
1641                         if (td->runstate != TD_INITIALIZED)
1642                                 continue;
1643
1644                         if (in_ramp_time(td))
1645                                 td_set_runstate(td, TD_RAMP);
1646                         else
1647                                 td_set_runstate(td, TD_RUNNING);
1648                         nr_running++;
1649                         nr_started--;
1650                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1651                         t_rate += td->o.rate[0] + td->o.rate[1];
1652                         todo--;
1653                         fio_mutex_up(td->mutex);
1654                 }
1655
1656                 reap_threads(&nr_running, &t_rate, &m_rate);
1657
1658                 if (todo)
1659                         usleep(100000);
1660         }
1661
1662         while (nr_running) {
1663                 reap_threads(&nr_running, &t_rate, &m_rate);
1664                 usleep(10000);
1665         }
1666
1667         update_io_ticks();
1668         fio_unpin_memory();
1669 }
1670
1671 int main(int argc, char *argv[])
1672 {
1673         long ps;
1674
1675         sinit();
1676         init_rand(&__fio_rand_state);
1677
1678         /*
1679          * We need locale for number printing, if it isn't set then just
1680          * go with the US format.
1681          */
1682         if (!getenv("LC_NUMERIC"))
1683                 setlocale(LC_NUMERIC, "en_US");
1684
1685         ps = sysconf(_SC_PAGESIZE);
1686         if (ps < 0) {
1687                 log_err("Failed to get page size\n");
1688                 return 1;
1689         }
1690
1691         page_size = ps;
1692         page_mask = ps - 1;
1693
1694         fio_keywords_init();
1695
1696         if (parse_options(argc, argv))
1697                 return 1;
1698
1699         if (exec_profile && load_profile(exec_profile))
1700                 return 1;
1701
1702         if (!thread_number)
1703                 return 0;
1704
1705         if (write_bw_log) {
1706                 setup_log(&agg_io_log[DDIR_READ]);
1707                 setup_log(&agg_io_log[DDIR_WRITE]);
1708         }
1709
1710         startup_mutex = fio_mutex_init(0);
1711         if (startup_mutex == NULL)
1712                 return 1;
1713         writeout_mutex = fio_mutex_init(1);
1714         if (writeout_mutex == NULL)
1715                 return 1;
1716
1717         set_genesis_time();
1718         create_disk_util_thread();
1719
1720         cgroup_list = smalloc(sizeof(*cgroup_list));
1721         INIT_FLIST_HEAD(cgroup_list);
1722
1723         run_threads();
1724
1725         if (!fio_abort) {
1726                 show_run_stats();
1727                 if (write_bw_log) {
1728                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1729                         __finish_log(agg_io_log[DDIR_WRITE],
1730                                         "agg-write_bw.log");
1731                 }
1732         }
1733
1734         cgroup_kill(cgroup_list);
1735         sfree(cgroup_list);
1736         sfree(cgroup_mnt);
1737
1738         fio_mutex_remove(startup_mutex);
1739         fio_mutex_remove(writeout_mutex);
1740         return exit_value;
1741 }