backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 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 <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38
  39 #include "fio.h"
  40 #ifndef FIO_NO_HAVE_SHM_H
  41 #include <sys/shm.h>
  42 #endif
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53 #include "lib/getrusage.h"
  54 #include "idletime.h"
  55 #include "err.h"
  56 #include "lib/tp.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59
  60 static pthread_t helper_thread;
  61 static pthread_mutex_t helper_lock;
  62 pthread_cond_t helper_cond;
  63 int helper_do_stat = 0;
  64
  65 static struct fio_mutex *startup_mutex;
  66 static struct flist_head *cgroup_list;
  67 static char *cgroup_mnt;
  68 static int exit_value;
  69 static volatile int fio_abort;
  70 static unsigned int nr_process = 0;
  71 static unsigned int nr_thread = 0;
  72
  73 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  74
  75 int groupid = 0;
  76 unsigned int thread_number = 0;
  77 unsigned int stat_number = 0;
  78 int shm_id = 0;
  79 int temp_stall_ts;
  80 unsigned long done_secs = 0;
  81 volatile int helper_exit = 0;
  82
  83 #define PAGE_ALIGN(buf) \
  84         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  85
  86 #define JOB_START_TIMEOUT       (5 * 1000)
  87
  88 static void sig_int(int sig)
  89 {
  90         if (threads) {
  91                 if (is_backend)
  92                         fio_server_got_signal(sig);
  93                 else {
  94                         log_info("\nfio: terminating on signal %d\n", sig);
  95                         log_info_flush();
  96                         exit_value = 128;
  97                 }
  98
  99                 fio_terminate_threads(TERMINATE_ALL);
 100         }
 101 }
 102
 103 static void sig_show_status(int sig)
 104 {
 105         show_running_run_stats();
 106 }
 107
 108 static void set_sig_handlers(void)
 109 {
 110         struct sigaction act;
 111
 112         memset(&act, 0, sizeof(act));
 113         act.sa_handler = sig_int;
 114         act.sa_flags = SA_RESTART;
 115         sigaction(SIGINT, &act, NULL);
 116
 117         memset(&act, 0, sizeof(act));
 118         act.sa_handler = sig_int;
 119         act.sa_flags = SA_RESTART;
 120         sigaction(SIGTERM, &act, NULL);
 121
 122 /* Windows uses SIGBREAK as a quit signal from other applications */
 123 #ifdef WIN32
 124         memset(&act, 0, sizeof(act));
 125         act.sa_handler = sig_int;
 126         act.sa_flags = SA_RESTART;
 127         sigaction(SIGBREAK, &act, NULL);
 128 #endif
 129
 130         memset(&act, 0, sizeof(act));
 131         act.sa_handler = sig_show_status;
 132         act.sa_flags = SA_RESTART;
 133         sigaction(SIGUSR1, &act, NULL);
 134
 135         if (is_backend) {
 136                 memset(&act, 0, sizeof(act));
 137                 act.sa_handler = sig_int;
 138                 act.sa_flags = SA_RESTART;
 139                 sigaction(SIGPIPE, &act, NULL);
 140         }
 141 }
 142
 143 /*
 144  * Check if we are above the minimum rate given.
 145  */
 146 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 147                             enum fio_ddir ddir)
 148 {
 149         unsigned long long bytes = 0;
 150         unsigned long iops = 0;
 151         unsigned long spent;
 152         unsigned long rate;
 153         unsigned int ratemin = 0;
 154         unsigned int rate_iops = 0;
 155         unsigned int rate_iops_min = 0;
 156
 157         assert(ddir_rw(ddir));
 158
 159         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 160                 return 0;
 161
 162         /*
 163          * allow a 2 second settle period in the beginning
 164          */
 165         if (mtime_since(&td->start, now) < 2000)
 166                 return 0;
 167
 168         iops += td->this_io_blocks[ddir];
 169         bytes += td->this_io_bytes[ddir];
 170         ratemin += td->o.ratemin[ddir];
 171         rate_iops += td->o.rate_iops[ddir];
 172         rate_iops_min += td->o.rate_iops_min[ddir];
 173
 174         /*
 175          * if rate blocks is set, sample is running
 176          */
 177         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 178                 spent = mtime_since(&td->lastrate[ddir], now);
 179                 if (spent < td->o.ratecycle)
 180                         return 0;
 181
 182                 if (td->o.rate[ddir]) {
 183                         /*
 184                          * check bandwidth specified rate
 185                          */
 186                         if (bytes < td->rate_bytes[ddir]) {
 187                                 log_err("%s: min rate %u not met\n", td->o.name,
 188                                                                 ratemin);
 189                                 return 1;
 190                         } else {
 191                                 if (spent)
 192                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 193                                 else
 194                                         rate = 0;
 195
 196                                 if (rate < ratemin ||
 197                                     bytes < td->rate_bytes[ddir]) {
 198                                         log_err("%s: min rate %u not met, got"
 199                                                 " %luKB/sec\n", td->o.name,
 200                                                         ratemin, rate);
 201                                         return 1;
 202                                 }
 203                         }
 204                 } else {
 205                         /*
 206                          * checks iops specified rate
 207                          */
 208                         if (iops < rate_iops) {
 209                                 log_err("%s: min iops rate %u not met\n",
 210                                                 td->o.name, rate_iops);
 211                                 return 1;
 212                         } else {
 213                                 if (spent)
 214                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 215                                 else
 216                                         rate = 0;
 217
 218                                 if (rate < rate_iops_min ||
 219                                     iops < td->rate_blocks[ddir]) {
 220                                         log_err("%s: min iops rate %u not met,"
 221                                                 " got %lu\n", td->o.name,
 222                                                         rate_iops_min, rate);
 223                                 }
 224                         }
 225                 }
 226         }
 227
 228         td->rate_bytes[ddir] = bytes;
 229         td->rate_blocks[ddir] = iops;
 230         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 231         return 0;
 232 }
 233
 234 static int check_min_rate(struct thread_data *td, struct timeval *now)
 235 {
 236         int ret = 0;
 237
 238         if (td->bytes_done[DDIR_READ])
 239                 ret |= __check_min_rate(td, now, DDIR_READ);
 240         if (td->bytes_done[DDIR_WRITE])
 241                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 242         if (td->bytes_done[DDIR_TRIM])
 243                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 244
 245         return ret;
 246 }
 247
 248 /*
 249  * When job exits, we can cancel the in-flight IO if we are using async
 250  * io. Attempt to do so.
 251  */
 252 static void cleanup_pending_aio(struct thread_data *td)
 253 {
 254         int r;
 255
 256         /*
 257          * get immediately available events, if any
 258          */
 259         r = io_u_queued_complete(td, 0);
 260         if (r < 0)
 261                 return;
 262
 263         /*
 264          * now cancel remaining active events
 265          */
 266         if (td->io_ops->cancel) {
 267                 struct io_u *io_u;
 268                 int i;
 269
 270                 io_u_qiter(&td->io_u_all, io_u, i) {
 271                         if (io_u->flags & IO_U_F_FLIGHT) {
 272                                 r = td->io_ops->cancel(td, io_u);
 273                                 if (!r)
 274                                         put_io_u(td, io_u);
 275                         }
 276                 }
 277         }
 278
 279         if (td->cur_depth)
 280                 r = io_u_queued_complete(td, td->cur_depth);
 281 }
 282
 283 /*
 284  * Helper to handle the final sync of a file. Works just like the normal
 285  * io path, just does everything sync.
 286  */
 287 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 288 {
 289         struct io_u *io_u = __get_io_u(td);
 290         int ret;
 291
 292         if (!io_u)
 293                 return 1;
 294
 295         io_u->ddir = DDIR_SYNC;
 296         io_u->file = f;
 297
 298         if (td_io_prep(td, io_u)) {
 299                 put_io_u(td, io_u);
 300                 return 1;
 301         }
 302
 303 requeue:
 304         ret = td_io_queue(td, io_u);
 305         if (ret < 0) {
 306                 td_verror(td, io_u->error, "td_io_queue");
 307                 put_io_u(td, io_u);
 308                 return 1;
 309         } else if (ret == FIO_Q_QUEUED) {
 310                 if (io_u_queued_complete(td, 1) < 0)
 311                         return 1;
 312         } else if (ret == FIO_Q_COMPLETED) {
 313                 if (io_u->error) {
 314                         td_verror(td, io_u->error, "td_io_queue");
 315                         return 1;
 316                 }
 317
 318                 if (io_u_sync_complete(td, io_u) < 0)
 319                         return 1;
 320         } else if (ret == FIO_Q_BUSY) {
 321                 if (td_io_commit(td))
 322                         return 1;
 323                 goto requeue;
 324         }
 325
 326         return 0;
 327 }
 328
 329 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 330 {
 331         int ret;
 332
 333         if (fio_file_open(f))
 334                 return fio_io_sync(td, f);
 335
 336         if (td_io_open_file(td, f))
 337                 return 1;
 338
 339         ret = fio_io_sync(td, f);
 340         td_io_close_file(td, f);
 341         return ret;
 342 }
 343
 344 static inline void __update_tv_cache(struct thread_data *td)
 345 {
 346         fio_gettime(&td->tv_cache, NULL);
 347 }
 348
 349 static inline void update_tv_cache(struct thread_data *td)
 350 {
 351         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 352                 __update_tv_cache(td);
 353 }
 354
 355 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 356 {
 357         if (in_ramp_time(td))
 358                 return 0;
 359         if (!td->o.timeout)
 360                 return 0;
 361         if (utime_since(&td->epoch, t) >= td->o.timeout)
 362                 return 1;
 363
 364         return 0;
 365 }
 366
 367 /*
 368  * We need to update the runtime consistently in ms, but keep a running
 369  * tally of the current elapsed time in microseconds for sub millisecond
 370  * updates.
 371  */
 372 static inline void update_runtime(struct thread_data *td,
 373                                   unsigned long long *elapsed_us,
 374                                   const enum fio_ddir ddir)
 375 {
 376         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 377                 return;
 378
 379         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 380         elapsed_us[ddir] += utime_since_now(&td->start);
 381         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 382 }
 383
 384 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 385                                int *retptr)
 386 {
 387         int ret = *retptr;
 388
 389         if (ret < 0 || td->error) {
 390                 int err = td->error;
 391                 enum error_type_bit eb;
 392
 393                 if (ret < 0)
 394                         err = -ret;
 395
 396                 eb = td_error_type(ddir, err);
 397                 if (!(td->o.continue_on_error & (1 << eb)))
 398                         return 1;
 399
 400                 if (td_non_fatal_error(td, eb, err)) {
 401                         /*
 402                          * Continue with the I/Os in case of
 403                          * a non fatal error.
 404                          */
 405                         update_error_count(td, err);
 406                         td_clear_error(td);
 407                         *retptr = 0;
 408                         return 0;
 409                 } else if (td->o.fill_device && err == ENOSPC) {
 410                         /*
 411                          * We expect to hit this error if
 412                          * fill_device option is set.
 413                          */
 414                         td_clear_error(td);
 415                         fio_mark_td_terminate(td);
 416                         return 1;
 417                 } else {
 418                         /*
 419                          * Stop the I/O in case of a fatal
 420                          * error.
 421                          */
 422                         update_error_count(td, err);
 423                         return 1;
 424                 }
 425         }
 426
 427         return 0;
 428 }
 429
 430 static void check_update_rusage(struct thread_data *td)
 431 {
 432         if (td->update_rusage) {
 433                 td->update_rusage = 0;
 434                 update_rusage_stat(td);
 435                 fio_mutex_up(td->rusage_sem);
 436         }
 437 }
 438
 439 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 440 {
 441         const int full = queue_full(td);
 442         int min_evts = 0;
 443         int ret;
 444
 445         /*
 446          * if the queue is full, we MUST reap at least 1 event
 447          */
 448         min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
 449     if ((full && !min_evts) || !td->o.iodepth_batch_complete)
 450                 min_evts = 1;
 451
 452         if (time && (__should_check_rate(td, DDIR_READ) ||
 453             __should_check_rate(td, DDIR_WRITE) ||
 454             __should_check_rate(td, DDIR_TRIM)))
 455                 fio_gettime(time, NULL);
 456
 457         do {
 458                 ret = io_u_queued_complete(td, min_evts);
 459                 if (ret < 0)
 460                         break;
 461         } while (full && (td->cur_depth > td->o.iodepth_low));
 462
 463         return ret;
 464 }
 465
 466 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 467                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 468                    struct timeval *comp_time)
 469 {
 470         int ret2;
 471
 472         switch (*ret) {
 473         case FIO_Q_COMPLETED:
 474                 if (io_u->error) {
 475                         *ret = -io_u->error;
 476                         clear_io_u(td, io_u);
 477                 } else if (io_u->resid) {
 478                         int bytes = io_u->xfer_buflen - io_u->resid;
 479                         struct fio_file *f = io_u->file;
 480
 481                         if (bytes_issued)
 482                                 *bytes_issued += bytes;
 483
 484                         if (!from_verify)
 485                                 trim_io_piece(td, io_u);
 486
 487                         /*
 488                          * zero read, fail
 489                          */
 490                         if (!bytes) {
 491                                 if (!from_verify)
 492                                         unlog_io_piece(td, io_u);
 493                                 td_verror(td, EIO, "full resid");
 494                                 put_io_u(td, io_u);
 495                                 break;
 496                         }
 497
 498                         io_u->xfer_buflen = io_u->resid;
 499                         io_u->xfer_buf += bytes;
 500                         io_u->offset += bytes;
 501
 502                         if (ddir_rw(io_u->ddir))
 503                                 td->ts.short_io_u[io_u->ddir]++;
 504
 505                         f = io_u->file;
 506                         if (io_u->offset == f->real_file_size)
 507                                 goto sync_done;
 508
 509                         requeue_io_u(td, &io_u);
 510                 } else {
 511 sync_done:
 512                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 513                             __should_check_rate(td, DDIR_WRITE) ||
 514                             __should_check_rate(td, DDIR_TRIM)))
 515                                 fio_gettime(comp_time, NULL);
 516
 517                         *ret = io_u_sync_complete(td, io_u);
 518                         if (*ret < 0)
 519                                 break;
 520                 }
 521                 return 0;
 522         case FIO_Q_QUEUED:
 523                 /*
 524                  * if the engine doesn't have a commit hook,
 525                  * the io_u is really queued. if it does have such
 526                  * a hook, it has to call io_u_queued() itself.
 527                  */
 528                 if (td->io_ops->commit == NULL)
 529                         io_u_queued(td, io_u);
 530                 if (bytes_issued)
 531                         *bytes_issued += io_u->xfer_buflen;
 532                 break;
 533         case FIO_Q_BUSY:
 534                 if (!from_verify)
 535                         unlog_io_piece(td, io_u);
 536                 requeue_io_u(td, &io_u);
 537                 ret2 = td_io_commit(td);
 538                 if (ret2 < 0)
 539                         *ret = ret2;
 540                 break;
 541         default:
 542                 assert(ret < 0);
 543                 td_verror(td, -(*ret), "td_io_queue");
 544                 break;
 545         }
 546
 547         if (break_on_this_error(td, ddir, ret))
 548                 return 1;
 549
 550         return 0;
 551 }
 552
 553 /*
 554  * The main verify engine. Runs over the writes we previously submitted,
 555  * reads the blocks back in, and checks the crc/md5 of the data.
 556  */
 557 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 558 {
 559         struct fio_file *f;
 560         struct io_u *io_u;
 561         int ret, min_events;
 562         unsigned int i;
 563
 564         dprint(FD_VERIFY, "starting loop\n");
 565
 566         /*
 567          * sync io first and invalidate cache, to make sure we really
 568          * read from disk.
 569          */
 570         for_each_file(td, f, i) {
 571                 if (!fio_file_open(f))
 572                         continue;
 573                 if (fio_io_sync(td, f))
 574                         break;
 575                 if (file_invalidate_cache(td, f))
 576                         break;
 577         }
 578
 579         check_update_rusage(td);
 580
 581         if (td->error)
 582                 return;
 583
 584         td_set_runstate(td, TD_VERIFYING);
 585
 586         io_u = NULL;
 587         while (!td->terminate) {
 588                 enum fio_ddir ddir;
 589                 int full;
 590
 591                 update_tv_cache(td);
 592                 check_update_rusage(td);
 593
 594                 if (runtime_exceeded(td, &td->tv_cache)) {
 595                         __update_tv_cache(td);
 596                         if (runtime_exceeded(td, &td->tv_cache)) {
 597                                 fio_mark_td_terminate(td);
 598                                 break;
 599                         }
 600                 }
 601
 602                 if (flow_threshold_exceeded(td))
 603                         continue;
 604
 605                 if (!td->o.experimental_verify) {
 606                         io_u = __get_io_u(td);
 607                         if (!io_u)
 608                                 break;
 609
 610                         if (get_next_verify(td, io_u)) {
 611                                 put_io_u(td, io_u);
 612                                 break;
 613                         }
 614
 615                         if (td_io_prep(td, io_u)) {
 616                                 put_io_u(td, io_u);
 617                                 break;
 618                         }
 619                 } else {
 620                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 621                                 break;
 622
 623                         while ((io_u = get_io_u(td)) != NULL) {
 624                                 if (IS_ERR(io_u)) {
 625                                         io_u = NULL;
 626                                         ret = FIO_Q_BUSY;
 627                                         goto reap;
 628                                 }
 629
 630                                 /*
 631                                  * We are only interested in the places where
 632                                  * we wrote or trimmed IOs. Turn those into
 633                                  * reads for verification purposes.
 634                                  */
 635                                 if (io_u->ddir == DDIR_READ) {
 636                                         /*
 637                                          * Pretend we issued it for rwmix
 638                                          * accounting
 639                                          */
 640                                         td->io_issues[DDIR_READ]++;
 641                                         put_io_u(td, io_u);
 642                                         continue;
 643                                 } else if (io_u->ddir == DDIR_TRIM) {
 644                                         io_u->ddir = DDIR_READ;
 645                                         io_u_set(io_u, IO_U_F_TRIMMED);
 646                                         break;
 647                                 } else if (io_u->ddir == DDIR_WRITE) {
 648                                         io_u->ddir = DDIR_READ;
 649                                         break;
 650                                 } else {
 651                                         put_io_u(td, io_u);
 652                                         continue;
 653                                 }
 654                         }
 655
 656                         if (!io_u)
 657                                 break;
 658                 }
 659
 660                 if (verify_state_should_stop(td, io_u)) {
 661                         put_io_u(td, io_u);
 662                         break;
 663                 }
 664
 665                 if (td->o.verify_async)
 666                         io_u->end_io = verify_io_u_async;
 667                 else
 668                         io_u->end_io = verify_io_u;
 669
 670                 ddir = io_u->ddir;
 671                 if (!td->o.disable_slat)
 672                         fio_gettime(&io_u->start_time, NULL);
 673
 674                 ret = td_io_queue(td, io_u);
 675
 676                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 677                         break;
 678
 679                 /*
 680                  * if we can queue more, do so. but check if there are
 681                  * completed io_u's first. Note that we can get BUSY even
 682                  * without IO queued, if the system is resource starved.
 683                  */
 684 reap:
 685                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 686                 if (full || !td->o.iodepth_batch_complete)
 687                         ret = wait_for_completions(td, NULL);
 688
 689                 if (ret < 0)
 690                         break;
 691         }
 692
 693         check_update_rusage(td);
 694
 695         if (!td->error) {
 696                 min_events = td->cur_depth;
 697
 698                 if (min_events)
 699                         ret = io_u_queued_complete(td, min_events);
 700         } else
 701                 cleanup_pending_aio(td);
 702
 703         td_set_runstate(td, TD_RUNNING);
 704
 705         dprint(FD_VERIFY, "exiting loop\n");
 706 }
 707
 708 static unsigned int exceeds_number_ios(struct thread_data *td)
 709 {
 710         unsigned long long number_ios;
 711
 712         if (!td->o.number_ios)
 713                 return 0;
 714
 715         number_ios = ddir_rw_sum(td->io_blocks);
 716         number_ios += td->io_u_queued + td->io_u_in_flight;
 717
 718         return number_ios >= (td->o.number_ios * td->loops);
 719 }
 720
 721 static int io_issue_bytes_exceeded(struct thread_data *td)
 722 {
 723         unsigned long long bytes, limit;
 724
 725         if (td_rw(td))
 726                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 727         else if (td_write(td))
 728                 bytes = td->io_issue_bytes[DDIR_WRITE];
 729         else if (td_read(td))
 730                 bytes = td->io_issue_bytes[DDIR_READ];
 731         else
 732                 bytes = td->io_issue_bytes[DDIR_TRIM];
 733
 734         if (td->o.io_limit)
 735                 limit = td->o.io_limit;
 736         else
 737                 limit = td->o.size;
 738
 739         limit *= td->loops;
 740         return bytes >= limit || exceeds_number_ios(td);
 741 }
 742
 743 static int io_complete_bytes_exceeded(struct thread_data *td)
 744 {
 745         unsigned long long bytes, limit;
 746
 747         if (td_rw(td))
 748                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 749         else if (td_write(td))
 750                 bytes = td->this_io_bytes[DDIR_WRITE];
 751         else if (td_read(td))
 752                 bytes = td->this_io_bytes[DDIR_READ];
 753         else
 754                 bytes = td->this_io_bytes[DDIR_TRIM];
 755
 756         if (td->o.io_limit)
 757                 limit = td->o.io_limit;
 758         else
 759                 limit = td->o.size;
 760
 761         limit *= td->loops;
 762         return bytes >= limit || exceeds_number_ios(td);
 763 }
 764
 765 /*
 766  * used to calculate the next io time for rate control
 767  *
 768  */
 769 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 770 {
 771         uint64_t secs, remainder, bps, bytes;
 772
 773         assert(!(td->flags & TD_F_CHILD));
 774         bytes = td->rate_io_issue_bytes[ddir];
 775         bps = td->rate_bps[ddir];
 776         if (bps) {
 777                 secs = bytes / bps;
 778                 remainder = bytes % bps;
 779                 return remainder * 1000000 / bps + secs * 1000000;
 780         } else
 781                 return 0;
 782 }
 783
 784 /*
 785  * Main IO worker function. It retrieves io_u's to process and queues
 786  * and reaps them, checking for rate and errors along the way.
 787  *
 788  * Returns number of bytes written and trimmed.
 789  */
 790 static uint64_t do_io(struct thread_data *td)
 791 {
 792         unsigned int i;
 793         int ret = 0;
 794         uint64_t total_bytes, bytes_issued = 0;
 795
 796         if (in_ramp_time(td))
 797                 td_set_runstate(td, TD_RAMP);
 798         else
 799                 td_set_runstate(td, TD_RUNNING);
 800
 801         lat_target_init(td);
 802
 803         total_bytes = td->o.size;
 804         /*
 805         * Allow random overwrite workloads to write up to io_limit
 806         * before starting verification phase as 'size' doesn't apply.
 807         */
 808         if (td_write(td) && td_random(td) && td->o.norandommap)
 809                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 810         /*
 811          * If verify_backlog is enabled, we'll run the verify in this
 812          * handler as well. For that case, we may need up to twice the
 813          * amount of bytes.
 814          */
 815         if (td->o.verify != VERIFY_NONE &&
 816            (td_write(td) && td->o.verify_backlog))
 817                 total_bytes += td->o.size;
 818
 819         /* In trimwrite mode, each byte is trimmed and then written, so
 820          * allow total_bytes to be twice as big */
 821         if (td_trimwrite(td))
 822                 total_bytes += td->total_io_size;
 823
 824         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 825                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 826                 td->o.time_based) {
 827                 struct timeval comp_time;
 828                 struct io_u *io_u;
 829                 int full;
 830                 enum fio_ddir ddir;
 831
 832                 check_update_rusage(td);
 833
 834                 if (td->terminate || td->done)
 835                         break;
 836
 837                 update_tv_cache(td);
 838
 839                 if (runtime_exceeded(td, &td->tv_cache)) {
 840                         __update_tv_cache(td);
 841                         if (runtime_exceeded(td, &td->tv_cache)) {
 842                                 fio_mark_td_terminate(td);
 843                                 break;
 844                         }
 845                 }
 846
 847                 if (flow_threshold_exceeded(td))
 848                         continue;
 849
 850                 if (bytes_issued >= total_bytes)
 851                         break;
 852
 853                 io_u = get_io_u(td);
 854                 if (IS_ERR_OR_NULL(io_u)) {
 855                         int err = PTR_ERR(io_u);
 856
 857                         io_u = NULL;
 858                         if (err == -EBUSY) {
 859                                 ret = FIO_Q_BUSY;
 860                                 goto reap;
 861                         }
 862                         if (td->o.latency_target)
 863                                 goto reap;
 864                         break;
 865                 }
 866
 867                 ddir = io_u->ddir;
 868
 869                 /*
 870                  * Add verification end_io handler if:
 871                  *      - Asked to verify (!td_rw(td))
 872                  *      - Or the io_u is from our verify list (mixed write/ver)
 873                  */
 874                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 875                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 876
 877                         if (!td->o.verify_pattern_bytes) {
 878                                 io_u->rand_seed = __rand(&td->verify_state);
 879                                 if (sizeof(int) != sizeof(long *))
 880                                         io_u->rand_seed *= __rand(&td->verify_state);
 881                         }
 882
 883                         if (verify_state_should_stop(td, io_u)) {
 884                                 put_io_u(td, io_u);
 885                                 break;
 886                         }
 887
 888                         if (td->o.verify_async)
 889                                 io_u->end_io = verify_io_u_async;
 890                         else
 891                                 io_u->end_io = verify_io_u;
 892                         td_set_runstate(td, TD_VERIFYING);
 893                 } else if (in_ramp_time(td))
 894                         td_set_runstate(td, TD_RAMP);
 895                 else
 896                         td_set_runstate(td, TD_RUNNING);
 897
 898                 /*
 899                  * Always log IO before it's issued, so we know the specific
 900                  * order of it. The logged unit will track when the IO has
 901                  * completed.
 902                  */
 903                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 904                     td->o.do_verify &&
 905                     td->o.verify != VERIFY_NONE &&
 906                     !td->o.experimental_verify)
 907                         log_io_piece(td, io_u);
 908
 909                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 910                         if (td->error)
 911                                 break;
 912                         ret = workqueue_enqueue(&td->io_wq, io_u);
 913
 914                         if (should_check_rate(td))
 915                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 916
 917                 } else {
 918                         ret = td_io_queue(td, io_u);
 919
 920                         if (should_check_rate(td))
 921                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 922
 923                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 924                                 break;
 925
 926                         /*
 927                          * See if we need to complete some commands. Note that
 928                          * we can get BUSY even without IO queued, if the
 929                          * system is resource starved.
 930                          */
 931 reap:
 932                         full = queue_full(td) ||
 933                                 (ret == FIO_Q_BUSY && td->cur_depth);
 934                         if (full || !td->o.iodepth_batch_complete)
 935                                 ret = wait_for_completions(td, &comp_time);
 936                 }
 937                 if (ret < 0)
 938                         break;
 939                 if (!ddir_rw_sum(td->bytes_done) &&
 940                     !(td->io_ops->flags & FIO_NOIO))
 941                         continue;
 942
 943                 if (!in_ramp_time(td) && should_check_rate(td)) {
 944                         if (check_min_rate(td, &comp_time)) {
 945                                 if (exitall_on_terminate)
 946                                         fio_terminate_threads(td->groupid);
 947                                 td_verror(td, EIO, "check_min_rate");
 948                                 break;
 949                         }
 950                 }
 951                 if (!in_ramp_time(td) && td->o.latency_target)
 952                         lat_target_check(td);
 953
 954                 if (td->o.thinktime) {
 955                         unsigned long long b;
 956
 957                         b = ddir_rw_sum(td->io_blocks);
 958                         if (!(b % td->o.thinktime_blocks)) {
 959                                 int left;
 960
 961                                 io_u_quiesce(td);
 962
 963                                 if (td->o.thinktime_spin)
 964                                         usec_spin(td->o.thinktime_spin);
 965
 966                                 left = td->o.thinktime - td->o.thinktime_spin;
 967                                 if (left)
 968                                         usec_sleep(td, left);
 969                         }
 970                 }
 971         }
 972
 973         check_update_rusage(td);
 974
 975         if (td->trim_entries)
 976                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 977
 978         if (td->o.fill_device && td->error == ENOSPC) {
 979                 td->error = 0;
 980                 fio_mark_td_terminate(td);
 981         }
 982         if (!td->error) {
 983                 struct fio_file *f;
 984
 985                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 986                         workqueue_flush(&td->io_wq);
 987                         i = 0;
 988                 } else
 989                         i = td->cur_depth;
 990
 991                 if (i) {
 992                         ret = io_u_queued_complete(td, i);
 993                         if (td->o.fill_device && td->error == ENOSPC)
 994                                 td->error = 0;
 995                 }
 996
 997                 if (should_fsync(td) && td->o.end_fsync) {
 998                         td_set_runstate(td, TD_FSYNCING);
 999
1000                         for_each_file(td, f, i) {
1001                                 if (!fio_file_fsync(td, f))
1002                                         continue;
1003
1004                                 log_err("fio: end_fsync failed for file %s\n",
1005                                                                 f->file_name);
1006                         }
1007                 }
1008         } else
1009                 cleanup_pending_aio(td);
1010
1011         /*
1012          * stop job if we failed doing any IO
1013          */
1014         if (!ddir_rw_sum(td->this_io_bytes))
1015                 td->done = 1;
1016
1017         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1018 }
1019
1020 static void cleanup_io_u(struct thread_data *td)
1021 {
1022         struct io_u *io_u;
1023
1024         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1025
1026                 if (td->io_ops->io_u_free)
1027                         td->io_ops->io_u_free(td, io_u);
1028
1029                 fio_memfree(io_u, sizeof(*io_u));
1030         }
1031
1032         free_io_mem(td);
1033
1034         io_u_rexit(&td->io_u_requeues);
1035         io_u_qexit(&td->io_u_freelist);
1036         io_u_qexit(&td->io_u_all);
1037
1038         if (td->last_write_comp)
1039                 sfree(td->last_write_comp);
1040 }
1041
1042 static int init_io_u(struct thread_data *td)
1043 {
1044         struct io_u *io_u;
1045         unsigned int max_bs, min_write;
1046         int cl_align, i, max_units;
1047         int data_xfer = 1, err;
1048         char *p;
1049
1050         max_units = td->o.iodepth;
1051         max_bs = td_max_bs(td);
1052         min_write = td->o.min_bs[DDIR_WRITE];
1053         td->orig_buffer_size = (unsigned long long) max_bs
1054                                         * (unsigned long long) max_units;
1055
1056         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1057                 data_xfer = 0;
1058
1059         err = 0;
1060         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1061         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1062         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1063
1064         if (err) {
1065                 log_err("fio: failed setting up IO queues\n");
1066                 return 1;
1067         }
1068
1069         /*
1070          * if we may later need to do address alignment, then add any
1071          * possible adjustment here so that we don't cause a buffer
1072          * overflow later. this adjustment may be too much if we get
1073          * lucky and the allocator gives us an aligned address.
1074          */
1075         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1076             (td->io_ops->flags & FIO_RAWIO))
1077                 td->orig_buffer_size += page_mask + td->o.mem_align;
1078
1079         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1080                 unsigned long bs;
1081
1082                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1083                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1084         }
1085
1086         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1087                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1088                 return 1;
1089         }
1090
1091         if (data_xfer && allocate_io_mem(td))
1092                 return 1;
1093
1094         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1095             (td->io_ops->flags & FIO_RAWIO))
1096                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1097         else
1098                 p = td->orig_buffer;
1099
1100         cl_align = os_cache_line_size();
1101
1102         for (i = 0; i < max_units; i++) {
1103                 void *ptr;
1104
1105                 if (td->terminate)
1106                         return 1;
1107
1108                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1109                 if (!ptr) {
1110                         log_err("fio: unable to allocate aligned memory\n");
1111                         break;
1112                 }
1113
1114                 io_u = ptr;
1115                 memset(io_u, 0, sizeof(*io_u));
1116                 INIT_FLIST_HEAD(&io_u->verify_list);
1117                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1118
1119                 if (data_xfer) {
1120                         io_u->buf = p;
1121                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1122
1123                         if (td_write(td))
1124                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1125                         if (td_write(td) && td->o.verify_pattern_bytes) {
1126                                 /*
1127                                  * Fill the buffer with the pattern if we are
1128                                  * going to be doing writes.
1129                                  */
1130                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1131                         }
1132                 }
1133
1134                 io_u->index = i;
1135                 io_u->flags = IO_U_F_FREE;
1136                 io_u_qpush(&td->io_u_freelist, io_u);
1137
1138                 /*
1139                  * io_u never leaves this stack, used for iteration of all
1140                  * io_u buffers.
1141                  */
1142                 io_u_qpush(&td->io_u_all, io_u);
1143
1144                 if (td->io_ops->io_u_init) {
1145                         int ret = td->io_ops->io_u_init(td, io_u);
1146
1147                         if (ret) {
1148                                 log_err("fio: failed to init engine data: %d\n", ret);
1149                                 return 1;
1150                         }
1151                 }
1152
1153                 p += max_bs;
1154         }
1155
1156         if (td->o.verify != VERIFY_NONE) {
1157                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1158                 if (!td->last_write_comp) {
1159                         log_err("fio: failed to alloc write comp data\n");
1160                         return 1;
1161                 }
1162         }
1163
1164         return 0;
1165 }
1166
1167 static int switch_ioscheduler(struct thread_data *td)
1168 {
1169         char tmp[256], tmp2[128];
1170         FILE *f;
1171         int ret;
1172
1173         if (td->io_ops->flags & FIO_DISKLESSIO)
1174                 return 0;
1175
1176         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1177
1178         f = fopen(tmp, "r+");
1179         if (!f) {
1180                 if (errno == ENOENT) {
1181                         log_err("fio: os or kernel doesn't support IO scheduler"
1182                                 " switching\n");
1183                         return 0;
1184                 }
1185                 td_verror(td, errno, "fopen iosched");
1186                 return 1;
1187         }
1188
1189         /*
1190          * Set io scheduler.
1191          */
1192         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1193         if (ferror(f) || ret != 1) {
1194                 td_verror(td, errno, "fwrite");
1195                 fclose(f);
1196                 return 1;
1197         }
1198
1199         rewind(f);
1200
1201         /*
1202          * Read back and check that the selected scheduler is now the default.
1203          */
1204         memset(tmp, 0, sizeof(tmp));
1205         ret = fread(tmp, sizeof(tmp), 1, f);
1206         if (ferror(f) || ret < 0) {
1207                 td_verror(td, errno, "fread");
1208                 fclose(f);
1209                 return 1;
1210         }
1211         /*
1212          * either a list of io schedulers or "none\n" is expected.
1213          */
1214         tmp[strlen(tmp) - 1] = '\0';
1215
1216
1217         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1218         if (!strstr(tmp, tmp2)) {
1219                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1220                 td_verror(td, EINVAL, "iosched_switch");
1221                 fclose(f);
1222                 return 1;
1223         }
1224
1225         fclose(f);
1226         return 0;
1227 }
1228
1229 static int keep_running(struct thread_data *td)
1230 {
1231         unsigned long long limit;
1232
1233         if (td->done)
1234                 return 0;
1235         if (td->o.time_based)
1236                 return 1;
1237         if (td->o.loops) {
1238                 td->o.loops--;
1239                 return 1;
1240         }
1241         if (exceeds_number_ios(td))
1242                 return 0;
1243
1244         if (td->o.io_limit)
1245                 limit = td->o.io_limit;
1246         else
1247                 limit = td->o.size;
1248
1249         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1250                 uint64_t diff;
1251
1252                 /*
1253                  * If the difference is less than the minimum IO size, we
1254                  * are done.
1255                  */
1256                 diff = limit - ddir_rw_sum(td->io_bytes);
1257                 if (diff < td_max_bs(td))
1258                         return 0;
1259
1260                 if (fio_files_done(td))
1261                         return 0;
1262
1263                 return 1;
1264         }
1265
1266         return 0;
1267 }
1268
1269 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1270 {
1271         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1272         int ret;
1273         char *str;
1274
1275         str = malloc(newlen);
1276         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1277
1278         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1279         ret = system(str);
1280         if (ret == -1)
1281                 log_err("fio: exec of cmd <%s> failed\n", str);
1282
1283         free(str);
1284         return ret;
1285 }
1286
1287 /*
1288  * Dry run to compute correct state of numberio for verification.
1289  */
1290 static uint64_t do_dry_run(struct thread_data *td)
1291 {
1292         td_set_runstate(td, TD_RUNNING);
1293
1294         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1295                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1296                 struct io_u *io_u;
1297                 int ret;
1298
1299                 if (td->terminate || td->done)
1300                         break;
1301
1302                 io_u = get_io_u(td);
1303                 if (!io_u)
1304                         break;
1305
1306                 io_u_set(io_u, IO_U_F_FLIGHT);
1307                 io_u->error = 0;
1308                 io_u->resid = 0;
1309                 if (ddir_rw(acct_ddir(io_u)))
1310                         td->io_issues[acct_ddir(io_u)]++;
1311                 if (ddir_rw(io_u->ddir)) {
1312                         io_u_mark_depth(td, 1);
1313                         td->ts.total_io_u[io_u->ddir]++;
1314                 }
1315
1316                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1317                     td->o.do_verify &&
1318                     td->o.verify != VERIFY_NONE &&
1319                     !td->o.experimental_verify)
1320                         log_io_piece(td, io_u);
1321
1322                 ret = io_u_sync_complete(td, io_u);
1323                 (void) ret;
1324         }
1325
1326         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1327 }
1328
1329 static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1330 {
1331         const enum fio_ddir ddir = io_u->ddir;
1332         int ret;
1333
1334         dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1335
1336         io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1337
1338         td->cur_depth++;
1339
1340         ret = td_io_queue(td, io_u);
1341
1342         dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1343
1344         io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1345
1346         if (ret == FIO_Q_QUEUED)
1347                 ret = io_u_queued_complete(td, 1);
1348
1349         td->cur_depth--;
1350 }
1351
1352 /*
1353  * Entry point for the thread based jobs. The process based jobs end up
1354  * here as well, after a little setup.
1355  */
1356 static void *thread_main(void *data)
1357 {
1358         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1359         struct thread_data *td = data;
1360         struct thread_options *o = &td->o;
1361         pthread_condattr_t attr;
1362         int clear_state;
1363         int ret;
1364
1365         if (!o->use_thread) {
1366                 setsid();
1367                 td->pid = getpid();
1368         } else
1369                 td->pid = gettid();
1370
1371         fio_local_clock_init(o->use_thread);
1372
1373         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1374
1375         if (is_backend)
1376                 fio_server_send_start(td);
1377
1378         INIT_FLIST_HEAD(&td->io_log_list);
1379         INIT_FLIST_HEAD(&td->io_hist_list);
1380         INIT_FLIST_HEAD(&td->verify_list);
1381         INIT_FLIST_HEAD(&td->trim_list);
1382         INIT_FLIST_HEAD(&td->next_rand_list);
1383         pthread_mutex_init(&td->io_u_lock, NULL);
1384         td->io_hist_tree = RB_ROOT;
1385
1386         pthread_condattr_init(&attr);
1387         pthread_cond_init(&td->verify_cond, &attr);
1388         pthread_cond_init(&td->free_cond, &attr);
1389
1390         td_set_runstate(td, TD_INITIALIZED);
1391         dprint(FD_MUTEX, "up startup_mutex\n");
1392         fio_mutex_up(startup_mutex);
1393         dprint(FD_MUTEX, "wait on td->mutex\n");
1394         fio_mutex_down(td->mutex);
1395         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1396
1397         /*
1398          * A new gid requires privilege, so we need to do this before setting
1399          * the uid.
1400          */
1401         if (o->gid != -1U && setgid(o->gid)) {
1402                 td_verror(td, errno, "setgid");
1403                 goto err;
1404         }
1405         if (o->uid != -1U && setuid(o->uid)) {
1406                 td_verror(td, errno, "setuid");
1407                 goto err;
1408         }
1409
1410         /*
1411          * If we have a gettimeofday() thread, make sure we exclude that
1412          * thread from this job
1413          */
1414         if (o->gtod_cpu)
1415                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1416
1417         /*
1418          * Set affinity first, in case it has an impact on the memory
1419          * allocations.
1420          */
1421         if (fio_option_is_set(o, cpumask)) {
1422                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1423                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1424                         if (!ret) {
1425                                 log_err("fio: no CPUs set\n");
1426                                 log_err("fio: Try increasing number of available CPUs\n");
1427                                 td_verror(td, EINVAL, "cpus_split");
1428                                 goto err;
1429                         }
1430                 }
1431                 ret = fio_setaffinity(td->pid, o->cpumask);
1432                 if (ret == -1) {
1433                         td_verror(td, errno, "cpu_set_affinity");
1434                         goto err;
1435                 }
1436         }
1437
1438 #ifdef CONFIG_LIBNUMA
1439         /* numa node setup */
1440         if (fio_option_is_set(o, numa_cpunodes) ||
1441             fio_option_is_set(o, numa_memnodes)) {
1442                 struct bitmask *mask;
1443
1444                 if (numa_available() < 0) {
1445                         td_verror(td, errno, "Does not support NUMA API\n");
1446                         goto err;
1447                 }
1448
1449                 if (fio_option_is_set(o, numa_cpunodes)) {
1450                         mask = numa_parse_nodestring(o->numa_cpunodes);
1451                         ret = numa_run_on_node_mask(mask);
1452                         numa_free_nodemask(mask);
1453                         if (ret == -1) {
1454                                 td_verror(td, errno, \
1455                                         "numa_run_on_node_mask failed\n");
1456                                 goto err;
1457                         }
1458                 }
1459
1460                 if (fio_option_is_set(o, numa_memnodes)) {
1461                         mask = NULL;
1462                         if (o->numa_memnodes)
1463                                 mask = numa_parse_nodestring(o->numa_memnodes);
1464
1465                         switch (o->numa_mem_mode) {
1466                         case MPOL_INTERLEAVE:
1467                                 numa_set_interleave_mask(mask);
1468                                 break;
1469                         case MPOL_BIND:
1470                                 numa_set_membind(mask);
1471                                 break;
1472                         case MPOL_LOCAL:
1473                                 numa_set_localalloc();
1474                                 break;
1475                         case MPOL_PREFERRED:
1476                                 numa_set_preferred(o->numa_mem_prefer_node);
1477                                 break;
1478                         case MPOL_DEFAULT:
1479                         default:
1480                                 break;
1481                         }
1482
1483                         if (mask)
1484                                 numa_free_nodemask(mask);
1485
1486                 }
1487         }
1488 #endif
1489
1490         if (fio_pin_memory(td))
1491                 goto err;
1492
1493         /*
1494          * May alter parameters that init_io_u() will use, so we need to
1495          * do this first.
1496          */
1497         if (init_iolog(td))
1498                 goto err;
1499
1500         if (init_io_u(td))
1501                 goto err;
1502
1503         if (o->verify_async && verify_async_init(td))
1504                 goto err;
1505
1506         if (fio_option_is_set(o, ioprio) ||
1507             fio_option_is_set(o, ioprio_class)) {
1508                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1509                 if (ret == -1) {
1510                         td_verror(td, errno, "ioprio_set");
1511                         goto err;
1512                 }
1513         }
1514
1515         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1516                 goto err;
1517
1518         errno = 0;
1519         if (nice(o->nice) == -1 && errno != 0) {
1520                 td_verror(td, errno, "nice");
1521                 goto err;
1522         }
1523
1524         if (o->ioscheduler && switch_ioscheduler(td))
1525                 goto err;
1526
1527         if (!o->create_serialize && setup_files(td))
1528                 goto err;
1529
1530         if (td_io_init(td))
1531                 goto err;
1532
1533         if (init_random_map(td))
1534                 goto err;
1535
1536         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1537                 goto err;
1538
1539         if (o->pre_read) {
1540                 if (pre_read_files(td) < 0)
1541                         goto err;
1542         }
1543
1544         if (td->flags & TD_F_COMPRESS_LOG)
1545                 tp_init(&td->tp_data);
1546
1547         fio_verify_init(td);
1548
1549         if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1550             workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1551                 goto err;
1552
1553         fio_gettime(&td->epoch, NULL);
1554         fio_getrusage(&td->ru_start);
1555         clear_state = 0;
1556         while (keep_running(td)) {
1557                 uint64_t verify_bytes;
1558
1559                 fio_gettime(&td->start, NULL);
1560                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1561                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1562                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1563
1564                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1565                                 o->ratemin[DDIR_TRIM]) {
1566                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1567                                                 sizeof(td->bw_sample_time));
1568                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1569                                                 sizeof(td->bw_sample_time));
1570                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1571                                                 sizeof(td->bw_sample_time));
1572                 }
1573
1574                 if (clear_state)
1575                         clear_io_state(td);
1576
1577                 prune_io_piece_log(td);
1578
1579                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1580                         verify_bytes = do_dry_run(td);
1581                 else
1582                         verify_bytes = do_io(td);
1583
1584                 clear_state = 1;
1585
1586                 /*
1587                  * Make sure we've successfully updated the rusage stats
1588                  * before waiting on the stat mutex. Otherwise we could have
1589                  * the stat thread holding stat mutex and waiting for
1590                  * the rusage_sem, which would never get upped because
1591                  * this thread is waiting for the stat mutex.
1592                  */
1593                 check_update_rusage(td);
1594
1595                 fio_mutex_down(stat_mutex);
1596                 if (td_read(td) && td->io_bytes[DDIR_READ])
1597                         update_runtime(td, elapsed_us, DDIR_READ);
1598                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1599                         update_runtime(td, elapsed_us, DDIR_WRITE);
1600                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1601                         update_runtime(td, elapsed_us, DDIR_TRIM);
1602                 fio_gettime(&td->start, NULL);
1603                 fio_mutex_up(stat_mutex);
1604
1605                 if (td->error || td->terminate)
1606                         break;
1607
1608                 if (!o->do_verify ||
1609                     o->verify == VERIFY_NONE ||
1610                     (td->io_ops->flags & FIO_UNIDIR))
1611                         continue;
1612
1613                 clear_io_state(td);
1614
1615                 fio_gettime(&td->start, NULL);
1616
1617                 do_verify(td, verify_bytes);
1618
1619                 /*
1620                  * See comment further up for why this is done here.
1621                  */
1622                 check_update_rusage(td);
1623
1624                 fio_mutex_down(stat_mutex);
1625                 update_runtime(td, elapsed_us, DDIR_READ);
1626                 fio_gettime(&td->start, NULL);
1627                 fio_mutex_up(stat_mutex);
1628
1629                 if (td->error || td->terminate)
1630                         break;
1631         }
1632
1633         update_rusage_stat(td);
1634         td->ts.total_run_time = mtime_since_now(&td->epoch);
1635         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1636         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1637         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1638
1639         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1640             (td->o.verify != VERIFY_NONE && td_write(td))) {
1641                 struct all_io_list *state;
1642                 size_t sz;
1643
1644                 state = get_all_io_list(td->thread_number, &sz);
1645                 if (state) {
1646                         __verify_save_state(state, "local");
1647                         free(state);
1648                 }
1649         }
1650
1651         fio_unpin_memory(td);
1652
1653         fio_writeout_logs(td);
1654
1655         if (o->io_submit_mode == IO_MODE_OFFLOAD)
1656                 workqueue_exit(&td->io_wq);
1657
1658         if (td->flags & TD_F_COMPRESS_LOG)
1659                 tp_exit(&td->tp_data);
1660
1661         if (o->exec_postrun)
1662                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1663
1664         if (exitall_on_terminate)
1665                 fio_terminate_threads(td->groupid);
1666
1667 err:
1668         if (td->error)
1669                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1670                                                         td->verror);
1671
1672         if (o->verify_async)
1673                 verify_async_exit(td);
1674
1675         close_and_free_files(td);
1676         cleanup_io_u(td);
1677         close_ioengine(td);
1678         cgroup_shutdown(td, &cgroup_mnt);
1679         verify_free_state(td);
1680
1681         if (fio_option_is_set(o, cpumask)) {
1682                 ret = fio_cpuset_exit(&o->cpumask);
1683                 if (ret)
1684                         td_verror(td, ret, "fio_cpuset_exit");
1685         }
1686
1687         /*
1688          * do this very late, it will log file closing as well
1689          */
1690         if (o->write_iolog_file)
1691                 write_iolog_close(td);
1692
1693         fio_mutex_remove(td->mutex);
1694         td->mutex = NULL;
1695
1696         td_set_runstate(td, TD_EXITED);
1697
1698         /*
1699          * Do this last after setting our runstate to exited, so we
1700          * know that the stat thread is signaled.
1701          */
1702         check_update_rusage(td);
1703
1704         return (void *) (uintptr_t) td->error;
1705 }
1706
1707
1708 /*
1709  * We cannot pass the td data into a forked process, so attach the td and
1710  * pass it to the thread worker.
1711  */
1712 static int fork_main(int shmid, int offset)
1713 {
1714         struct thread_data *td;
1715         void *data, *ret;
1716
1717 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1718         data = shmat(shmid, NULL, 0);
1719         if (data == (void *) -1) {
1720                 int __err = errno;
1721
1722                 perror("shmat");
1723                 return __err;
1724         }
1725 #else
1726         /*
1727          * HP-UX inherits shm mappings?
1728          */
1729         data = threads;
1730 #endif
1731
1732         td = data + offset * sizeof(struct thread_data);
1733         ret = thread_main(td);
1734         shmdt(data);
1735         return (int) (uintptr_t) ret;
1736 }
1737
1738 static void dump_td_info(struct thread_data *td)
1739 {
1740         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1741                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1742                         (unsigned long) time_since_now(&td->terminate_time));
1743 }
1744
1745 /*
1746  * Run over the job map and reap the threads that have exited, if any.
1747  */
1748 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1749                          unsigned int *m_rate)
1750 {
1751         struct thread_data *td;
1752         unsigned int cputhreads, realthreads, pending;
1753         int i, status, ret;
1754
1755         /*
1756          * reap exited threads (TD_EXITED -> TD_REAPED)
1757          */
1758         realthreads = pending = cputhreads = 0;
1759         for_each_td(td, i) {
1760                 int flags = 0;
1761
1762                 /*
1763                  * ->io_ops is NULL for a thread that has closed its
1764                  * io engine
1765                  */
1766                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1767                         cputhreads++;
1768                 else
1769                         realthreads++;
1770
1771                 if (!td->pid) {
1772                         pending++;
1773                         continue;
1774                 }
1775                 if (td->runstate == TD_REAPED)
1776                         continue;
1777                 if (td->o.use_thread) {
1778                         if (td->runstate == TD_EXITED) {
1779                                 td_set_runstate(td, TD_REAPED);
1780                                 goto reaped;
1781                         }
1782                         continue;
1783                 }
1784
1785                 flags = WNOHANG;
1786                 if (td->runstate == TD_EXITED)
1787                         flags = 0;
1788
1789                 /*
1790                  * check if someone quit or got killed in an unusual way
1791                  */
1792                 ret = waitpid(td->pid, &status, flags);
1793                 if (ret < 0) {
1794                         if (errno == ECHILD) {
1795                                 log_err("fio: pid=%d disappeared %d\n",
1796                                                 (int) td->pid, td->runstate);
1797                                 td->sig = ECHILD;
1798                                 td_set_runstate(td, TD_REAPED);
1799                                 goto reaped;
1800                         }
1801                         perror("waitpid");
1802                 } else if (ret == td->pid) {
1803                         if (WIFSIGNALED(status)) {
1804                                 int sig = WTERMSIG(status);
1805
1806                                 if (sig != SIGTERM && sig != SIGUSR2)
1807                                         log_err("fio: pid=%d, got signal=%d\n",
1808                                                         (int) td->pid, sig);
1809                                 td->sig = sig;
1810                                 td_set_runstate(td, TD_REAPED);
1811                                 goto reaped;
1812                         }
1813                         if (WIFEXITED(status)) {
1814                                 if (WEXITSTATUS(status) && !td->error)
1815                                         td->error = WEXITSTATUS(status);
1816
1817                                 td_set_runstate(td, TD_REAPED);
1818                                 goto reaped;
1819                         }
1820                 }
1821
1822                 /*
1823                  * If the job is stuck, do a forceful timeout of it and
1824                  * move on.
1825                  */
1826                 if (td->terminate &&
1827                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1828                         dump_td_info(td);
1829                         td_set_runstate(td, TD_REAPED);
1830                         goto reaped;
1831                 }
1832
1833                 /*
1834                  * thread is not dead, continue
1835                  */
1836                 pending++;
1837                 continue;
1838 reaped:
1839                 (*nr_running)--;
1840                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1841                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1842                 if (!td->pid)
1843                         pending--;
1844
1845                 if (td->error)
1846                         exit_value++;
1847
1848                 done_secs += mtime_since_now(&td->epoch) / 1000;
1849                 profile_td_exit(td);
1850         }
1851
1852         if (*nr_running == cputhreads && !pending && realthreads)
1853                 fio_terminate_threads(TERMINATE_ALL);
1854 }
1855
1856 static int __check_trigger_file(void)
1857 {
1858         struct stat sb;
1859
1860         if (!trigger_file)
1861                 return 0;
1862
1863         if (stat(trigger_file, &sb))
1864                 return 0;
1865
1866         if (unlink(trigger_file) < 0)
1867                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1868                                                         strerror(errno));
1869
1870         return 1;
1871 }
1872
1873 static int trigger_timedout(void)
1874 {
1875         if (trigger_timeout)
1876                 return time_since_genesis() >= trigger_timeout;
1877
1878         return 0;
1879 }
1880
1881 void exec_trigger(const char *cmd)
1882 {
1883         int ret;
1884
1885         if (!cmd)
1886                 return;
1887
1888         ret = system(cmd);
1889         if (ret == -1)
1890                 log_err("fio: failed executing %s trigger\n", cmd);
1891 }
1892
1893 void check_trigger_file(void)
1894 {
1895         if (__check_trigger_file() || trigger_timedout()) {
1896                 if (nr_clients)
1897                         fio_clients_send_trigger(trigger_remote_cmd);
1898                 else {
1899                         verify_save_state();
1900                         fio_terminate_threads(TERMINATE_ALL);
1901                         exec_trigger(trigger_cmd);
1902                 }
1903         }
1904 }
1905
1906 static int fio_verify_load_state(struct thread_data *td)
1907 {
1908         int ret;
1909
1910         if (!td->o.verify_state)
1911                 return 0;
1912
1913         if (is_backend) {
1914                 void *data;
1915                 int ver;
1916
1917                 ret = fio_server_get_verify_state(td->o.name,
1918                                         td->thread_number - 1, &data, &ver);
1919                 if (!ret)
1920                         verify_convert_assign_state(td, data, ver);
1921         } else
1922                 ret = verify_load_state(td, "local");
1923
1924         return ret;
1925 }
1926
1927 static void do_usleep(unsigned int usecs)
1928 {
1929         check_for_running_stats();
1930         check_trigger_file();
1931         usleep(usecs);
1932 }
1933
1934 static int check_mount_writes(struct thread_data *td)
1935 {
1936         struct fio_file *f;
1937         unsigned int i;
1938
1939         if (!td_write(td) || td->o.allow_mounted_write)
1940                 return 0;
1941
1942         for_each_file(td, f, i) {
1943                 if (f->filetype != FIO_TYPE_BD)
1944                         continue;
1945                 if (device_is_mounted(f->file_name))
1946                         goto mounted;
1947         }
1948
1949         return 0;
1950 mounted:
1951         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1952         return 1;
1953 }
1954
1955 /*
1956  * Main function for kicking off and reaping jobs, as needed.
1957  */
1958 static void run_threads(void)
1959 {
1960         struct thread_data *td;
1961         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1962         uint64_t spent;
1963
1964         if (fio_gtod_offload && fio_start_gtod_thread())
1965                 return;
1966
1967         fio_idle_prof_init();
1968
1969         set_sig_handlers();
1970
1971         nr_thread = nr_process = 0;
1972         for_each_td(td, i) {
1973                 if (check_mount_writes(td))
1974                         return;
1975                 if (td->o.use_thread)
1976                         nr_thread++;
1977                 else
1978                         nr_process++;
1979         }
1980
1981         if (output_format == FIO_OUTPUT_NORMAL) {
1982                 log_info("Starting ");
1983                 if (nr_thread)
1984                         log_info("%d thread%s", nr_thread,
1985                                                 nr_thread > 1 ? "s" : "");
1986                 if (nr_process) {
1987                         if (nr_thread)
1988                                 log_info(" and ");
1989                         log_info("%d process%s", nr_process,
1990                                                 nr_process > 1 ? "es" : "");
1991                 }
1992                 log_info("\n");
1993                 log_info_flush();
1994         }
1995
1996         todo = thread_number;
1997         nr_running = 0;
1998         nr_started = 0;
1999         m_rate = t_rate = 0;
2000
2001         for_each_td(td, i) {
2002                 print_status_init(td->thread_number - 1);
2003
2004                 if (!td->o.create_serialize)
2005                         continue;
2006
2007                 if (fio_verify_load_state(td))
2008                         goto reap;
2009
2010                 /*
2011                  * do file setup here so it happens sequentially,
2012                  * we don't want X number of threads getting their
2013                  * client data interspersed on disk
2014                  */
2015                 if (setup_files(td)) {
2016 reap:
2017                         exit_value++;
2018                         if (td->error)
2019                                 log_err("fio: pid=%d, err=%d/%s\n",
2020                                         (int) td->pid, td->error, td->verror);
2021                         td_set_runstate(td, TD_REAPED);
2022                         todo--;
2023                 } else {
2024                         struct fio_file *f;
2025                         unsigned int j;
2026
2027                         /*
2028                          * for sharing to work, each job must always open
2029                          * its own files. so close them, if we opened them
2030                          * for creation
2031                          */
2032                         for_each_file(td, f, j) {
2033                                 if (fio_file_open(f))
2034                                         td_io_close_file(td, f);
2035                         }
2036                 }
2037         }
2038
2039         /* start idle threads before io threads start to run */
2040         fio_idle_prof_start();
2041
2042         set_genesis_time();
2043
2044         while (todo) {
2045                 struct thread_data *map[REAL_MAX_JOBS];
2046                 struct timeval this_start;
2047                 int this_jobs = 0, left;
2048
2049                 /*
2050                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2051                  */
2052                 for_each_td(td, i) {
2053                         if (td->runstate != TD_NOT_CREATED)
2054                                 continue;
2055
2056                         /*
2057                          * never got a chance to start, killed by other
2058                          * thread for some reason
2059                          */
2060                         if (td->terminate) {
2061                                 todo--;
2062                                 continue;
2063                         }
2064
2065                         if (td->o.start_delay) {
2066                                 spent = utime_since_genesis();
2067
2068                                 if (td->o.start_delay > spent)
2069                                         continue;
2070                         }
2071
2072                         if (td->o.stonewall && (nr_started || nr_running)) {
2073                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2074                                                         td->o.name);
2075                                 break;
2076                         }
2077
2078                         init_disk_util(td);
2079
2080                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2081                         td->update_rusage = 0;
2082
2083                         /*
2084                          * Set state to created. Thread will transition
2085                          * to TD_INITIALIZED when it's done setting up.
2086                          */
2087                         td_set_runstate(td, TD_CREATED);
2088                         map[this_jobs++] = td;
2089                         nr_started++;
2090
2091                         if (td->o.use_thread) {
2092                                 int ret;
2093
2094                                 dprint(FD_PROCESS, "will pthread_create\n");
2095                                 ret = pthread_create(&td->thread, NULL,
2096                                                         thread_main, td);
2097                                 if (ret) {
2098                                         log_err("pthread_create: %s\n",
2099                                                         strerror(ret));
2100                                         nr_started--;
2101                                         break;
2102                                 }
2103                                 ret = pthread_detach(td->thread);
2104                                 if (ret)
2105                                         log_err("pthread_detach: %s",
2106                                                         strerror(ret));
2107                         } else {
2108                                 pid_t pid;
2109                                 dprint(FD_PROCESS, "will fork\n");
2110                                 pid = fork();
2111                                 if (!pid) {
2112                                         int ret = fork_main(shm_id, i);
2113
2114                                         _exit(ret);
2115                                 } else if (i == fio_debug_jobno)
2116                                         *fio_debug_jobp = pid;
2117                         }
2118                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2119                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2120                                 log_err("fio: job startup hung? exiting.\n");
2121                                 fio_terminate_threads(TERMINATE_ALL);
2122                                 fio_abort = 1;
2123                                 nr_started--;
2124                                 break;
2125                         }
2126                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2127                 }
2128
2129                 /*
2130                  * Wait for the started threads to transition to
2131                  * TD_INITIALIZED.
2132                  */
2133                 fio_gettime(&this_start, NULL);
2134                 left = this_jobs;
2135                 while (left && !fio_abort) {
2136                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2137                                 break;
2138
2139                         do_usleep(100000);
2140
2141                         for (i = 0; i < this_jobs; i++) {
2142                                 td = map[i];
2143                                 if (!td)
2144                                         continue;
2145                                 if (td->runstate == TD_INITIALIZED) {
2146                                         map[i] = NULL;
2147                                         left--;
2148                                 } else if (td->runstate >= TD_EXITED) {
2149                                         map[i] = NULL;
2150                                         left--;
2151                                         todo--;
2152                                         nr_running++; /* work-around... */
2153                                 }
2154                         }
2155                 }
2156
2157                 if (left) {
2158                         log_err("fio: %d job%s failed to start\n", left,
2159                                         left > 1 ? "s" : "");
2160                         for (i = 0; i < this_jobs; i++) {
2161                                 td = map[i];
2162                                 if (!td)
2163                                         continue;
2164                                 kill(td->pid, SIGTERM);
2165                         }
2166                         break;
2167                 }
2168
2169                 /*
2170                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2171                  */
2172                 for_each_td(td, i) {
2173                         if (td->runstate != TD_INITIALIZED)
2174                                 continue;
2175
2176                         if (in_ramp_time(td))
2177                                 td_set_runstate(td, TD_RAMP);
2178                         else
2179                                 td_set_runstate(td, TD_RUNNING);
2180                         nr_running++;
2181                         nr_started--;
2182                         m_rate += ddir_rw_sum(td->o.ratemin);
2183                         t_rate += ddir_rw_sum(td->o.rate);
2184                         todo--;
2185                         fio_mutex_up(td->mutex);
2186                 }
2187
2188                 reap_threads(&nr_running, &t_rate, &m_rate);
2189
2190                 if (todo)
2191                         do_usleep(100000);
2192         }
2193
2194         while (nr_running) {
2195                 reap_threads(&nr_running, &t_rate, &m_rate);
2196                 do_usleep(10000);
2197         }
2198
2199         fio_idle_prof_stop();
2200
2201         update_io_ticks();
2202 }
2203
2204 static void wait_for_helper_thread_exit(void)
2205 {
2206         void *ret;
2207
2208         helper_exit = 1;
2209         pthread_cond_signal(&helper_cond);
2210         pthread_join(helper_thread, &ret);
2211 }
2212
2213 static void free_disk_util(void)
2214 {
2215         disk_util_prune_entries();
2216
2217         pthread_cond_destroy(&helper_cond);
2218 }
2219
2220 static void *helper_thread_main(void *data)
2221 {
2222         int ret = 0;
2223
2224         fio_mutex_up(startup_mutex);
2225
2226         while (!ret) {
2227                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2228                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2229                 struct timespec ts;
2230                 struct timeval tv;
2231
2232                 gettimeofday(&tv, NULL);
2233                 ts.tv_sec = tv.tv_sec + sec;
2234                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2235
2236                 if (ts.tv_nsec >= 1000000000ULL) {
2237                         ts.tv_nsec -= 1000000000ULL;
2238                         ts.tv_sec++;
2239                 }
2240
2241                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2242
2243                 ret = update_io_ticks();
2244
2245                 if (helper_do_stat) {
2246                         helper_do_stat = 0;
2247                         __show_running_run_stats();
2248                 }
2249
2250                 if (!is_backend)
2251                         print_thread_status();
2252         }
2253
2254         return NULL;
2255 }
2256
2257 static int create_helper_thread(void)
2258 {
2259         int ret;
2260
2261         setup_disk_util();
2262
2263         pthread_cond_init(&helper_cond, NULL);
2264         pthread_mutex_init(&helper_lock, NULL);
2265
2266         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2267         if (ret) {
2268                 log_err("Can't create helper thread: %s\n", strerror(ret));
2269                 return 1;
2270         }
2271
2272         dprint(FD_MUTEX, "wait on startup_mutex\n");
2273         fio_mutex_down(startup_mutex);
2274         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2275         return 0;
2276 }
2277
2278 int fio_backend(void)
2279 {
2280         struct thread_data *td;
2281         int i;
2282
2283         if (exec_profile) {
2284                 if (load_profile(exec_profile))
2285                         return 1;
2286                 free(exec_profile);
2287                 exec_profile = NULL;
2288         }
2289         if (!thread_number)
2290                 return 0;
2291
2292         if (write_bw_log) {
2293                 struct log_params p = {
2294                         .log_type = IO_LOG_TYPE_BW,
2295                 };
2296
2297                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2298                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2299                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2300         }
2301
2302         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2303         if (startup_mutex == NULL)
2304                 return 1;
2305
2306         set_genesis_time();
2307         stat_init();
2308         create_helper_thread();
2309
2310         cgroup_list = smalloc(sizeof(*cgroup_list));
2311         INIT_FLIST_HEAD(cgroup_list);
2312
2313         run_threads();
2314
2315         wait_for_helper_thread_exit();
2316
2317         if (!fio_abort) {
2318                 __show_run_stats();
2319                 if (write_bw_log) {
2320                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2321                                 struct io_log *log = agg_io_log[i];
2322
2323                                 flush_log(log, 0);
2324                                 free_log(log);
2325                         }
2326                 }
2327         }
2328
2329         for_each_td(td, i) {
2330                 fio_options_free(td);
2331                 if (td->rusage_sem) {
2332                         fio_mutex_remove(td->rusage_sem);
2333                         td->rusage_sem = NULL;
2334                 }
2335         }
2336
2337         free_disk_util();
2338         cgroup_kill(cgroup_list);
2339         sfree(cgroup_list);
2340         sfree(cgroup_mnt);
2341
2342         fio_mutex_remove(startup_mutex);
2343         stat_exit();
2344         return exit_value;
2345 }