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