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 #include <math.h>
  39
  40 #include "fio.h"
  41 #ifndef FIO_NO_HAVE_SHM_H
  42 #include <sys/shm.h>
  43 #endif
  44 #include "hash.h"
  45 #include "smalloc.h"
  46 #include "verify.h"
  47 #include "trim.h"
  48 #include "diskutil.h"
  49 #include "cgroup.h"
  50 #include "profile.h"
  51 #include "lib/rand.h"
  52 #include "lib/memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59 #include "rate-submit.h"
  60
  61 static pthread_t helper_thread;
  62 static pthread_mutex_t helper_lock;
  63 pthread_cond_t helper_cond;
  64 int helper_do_stat = 0;
  65
  66 static struct fio_mutex *startup_mutex;
  67 static struct flist_head *cgroup_list;
  68 static char *cgroup_mnt;
  69 static int exit_value;
  70 static volatile int fio_abort;
  71 static unsigned int nr_process = 0;
  72 static unsigned int nr_thread = 0;
  73
  74 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  75
  76 int groupid = 0;
  77 unsigned int thread_number = 0;
  78 unsigned int stat_number = 0;
  79 int shm_id = 0;
  80 int temp_stall_ts;
  81 unsigned long done_secs = 0;
  82 volatile int helper_exit = 0;
  83
  84 #define PAGE_ALIGN(buf) \
  85         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  86
  87 #define JOB_START_TIMEOUT       (5 * 1000)
  88
  89 static void sig_int(int sig)
  90 {
  91         if (threads) {
  92                 if (is_backend)
  93                         fio_server_got_signal(sig);
  94                 else {
  95                         log_info("\nfio: terminating on signal %d\n", sig);
  96                         log_info_flush();
  97                         exit_value = 128;
  98                 }
  99
 100                 fio_terminate_threads(TERMINATE_ALL);
 101         }
 102 }
 103
 104 void sig_show_status(int sig)
 105 {
 106         show_running_run_stats();
 107 }
 108
 109 static void set_sig_handlers(void)
 110 {
 111         struct sigaction act;
 112
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGINT, &act, NULL);
 117
 118         memset(&act, 0, sizeof(act));
 119         act.sa_handler = sig_int;
 120         act.sa_flags = SA_RESTART;
 121         sigaction(SIGTERM, &act, NULL);
 122
 123 /* Windows uses SIGBREAK as a quit signal from other applications */
 124 #ifdef WIN32
 125         memset(&act, 0, sizeof(act));
 126         act.sa_handler = sig_int;
 127         act.sa_flags = SA_RESTART;
 128         sigaction(SIGBREAK, &act, NULL);
 129 #endif
 130
 131         memset(&act, 0, sizeof(act));
 132         act.sa_handler = sig_show_status;
 133         act.sa_flags = SA_RESTART;
 134         sigaction(SIGUSR1, &act, NULL);
 135
 136         if (is_backend) {
 137                 memset(&act, 0, sizeof(act));
 138                 act.sa_handler = sig_int;
 139                 act.sa_flags = SA_RESTART;
 140                 sigaction(SIGPIPE, &act, NULL);
 141         }
 142 }
 143
 144 /*
 145  * Check if we are above the minimum rate given.
 146  */
 147 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 148                              enum fio_ddir ddir)
 149 {
 150         unsigned long long bytes = 0;
 151         unsigned long iops = 0;
 152         unsigned long spent;
 153         unsigned long rate;
 154         unsigned int ratemin = 0;
 155         unsigned int rate_iops = 0;
 156         unsigned int rate_iops_min = 0;
 157
 158         assert(ddir_rw(ddir));
 159
 160         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 161                 return false;
 162
 163         /*
 164          * allow a 2 second settle period in the beginning
 165          */
 166         if (mtime_since(&td->start, now) < 2000)
 167                 return false;
 168
 169         iops += td->this_io_blocks[ddir];
 170         bytes += td->this_io_bytes[ddir];
 171         ratemin += td->o.ratemin[ddir];
 172         rate_iops += td->o.rate_iops[ddir];
 173         rate_iops_min += td->o.rate_iops_min[ddir];
 174
 175         /*
 176          * if rate blocks is set, sample is running
 177          */
 178         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 179                 spent = mtime_since(&td->lastrate[ddir], now);
 180                 if (spent < td->o.ratecycle)
 181                         return false;
 182
 183                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 184                         /*
 185                          * check bandwidth specified rate
 186                          */
 187                         if (bytes < td->rate_bytes[ddir]) {
 188                                 log_err("%s: min rate %u not met\n", td->o.name,
 189                                                                 ratemin);
 190                                 return true;
 191                         } else {
 192                                 if (spent)
 193                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 194                                 else
 195                                         rate = 0;
 196
 197                                 if (rate < ratemin ||
 198                                     bytes < td->rate_bytes[ddir]) {
 199                                         log_err("%s: min rate %u not met, got"
 200                                                 " %luKB/sec\n", td->o.name,
 201                                                         ratemin, rate);
 202                                         return true;
 203                                 }
 204                         }
 205                 } else {
 206                         /*
 207                          * checks iops specified rate
 208                          */
 209                         if (iops < rate_iops) {
 210                                 log_err("%s: min iops rate %u not met\n",
 211                                                 td->o.name, rate_iops);
 212                                 return true;
 213                         } else {
 214                                 if (spent)
 215                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 216                                 else
 217                                         rate = 0;
 218
 219                                 if (rate < rate_iops_min ||
 220                                     iops < td->rate_blocks[ddir]) {
 221                                         log_err("%s: min iops rate %u not met,"
 222                                                 " got %lu\n", td->o.name,
 223                                                         rate_iops_min, rate);
 224                                         return true;
 225                                 }
 226                         }
 227                 }
 228         }
 229
 230         td->rate_bytes[ddir] = bytes;
 231         td->rate_blocks[ddir] = iops;
 232         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 233         return false;
 234 }
 235
 236 static bool check_min_rate(struct thread_data *td, struct timeval *now)
 237 {
 238         bool ret = false;
 239
 240         if (td->bytes_done[DDIR_READ])
 241                 ret |= __check_min_rate(td, now, DDIR_READ);
 242         if (td->bytes_done[DDIR_WRITE])
 243                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 244         if (td->bytes_done[DDIR_TRIM])
 245                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 246
 247         return ret;
 248 }
 249
 250 /*
 251  * When job exits, we can cancel the in-flight IO if we are using async
 252  * io. Attempt to do so.
 253  */
 254 static void cleanup_pending_aio(struct thread_data *td)
 255 {
 256         int r;
 257
 258         /*
 259          * get immediately available events, if any
 260          */
 261         r = io_u_queued_complete(td, 0);
 262         if (r < 0)
 263                 return;
 264
 265         /*
 266          * now cancel remaining active events
 267          */
 268         if (td->io_ops->cancel) {
 269                 struct io_u *io_u;
 270                 int i;
 271
 272                 io_u_qiter(&td->io_u_all, io_u, i) {
 273                         if (io_u->flags & IO_U_F_FLIGHT) {
 274                                 r = td->io_ops->cancel(td, io_u);
 275                                 if (!r)
 276                                         put_io_u(td, io_u);
 277                         }
 278                 }
 279         }
 280
 281         if (td->cur_depth)
 282                 r = io_u_queued_complete(td, td->cur_depth);
 283 }
 284
 285 /*
 286  * Helper to handle the final sync of a file. Works just like the normal
 287  * io path, just does everything sync.
 288  */
 289 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 290 {
 291         struct io_u *io_u = __get_io_u(td);
 292         int ret;
 293
 294         if (!io_u)
 295                 return true;
 296
 297         io_u->ddir = DDIR_SYNC;
 298         io_u->file = f;
 299
 300         if (td_io_prep(td, io_u)) {
 301                 put_io_u(td, io_u);
 302                 return true;
 303         }
 304
 305 requeue:
 306         ret = td_io_queue(td, io_u);
 307         if (ret < 0) {
 308                 td_verror(td, io_u->error, "td_io_queue");
 309                 put_io_u(td, io_u);
 310                 return true;
 311         } else if (ret == FIO_Q_QUEUED) {
 312                 if (io_u_queued_complete(td, 1) < 0)
 313                         return true;
 314         } else if (ret == FIO_Q_COMPLETED) {
 315                 if (io_u->error) {
 316                         td_verror(td, io_u->error, "td_io_queue");
 317                         return true;
 318                 }
 319
 320                 if (io_u_sync_complete(td, io_u) < 0)
 321                         return true;
 322         } else if (ret == FIO_Q_BUSY) {
 323                 if (td_io_commit(td))
 324                         return true;
 325                 goto requeue;
 326         }
 327
 328         return false;
 329 }
 330
 331 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 332 {
 333         int ret;
 334
 335         if (fio_file_open(f))
 336                 return fio_io_sync(td, f);
 337
 338         if (td_io_open_file(td, f))
 339                 return 1;
 340
 341         ret = fio_io_sync(td, f);
 342         td_io_close_file(td, f);
 343         return ret;
 344 }
 345
 346 static inline void __update_tv_cache(struct thread_data *td)
 347 {
 348         fio_gettime(&td->tv_cache, NULL);
 349 }
 350
 351 static inline void update_tv_cache(struct thread_data *td)
 352 {
 353         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 354                 __update_tv_cache(td);
 355 }
 356
 357 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 358 {
 359         if (in_ramp_time(td))
 360                 return false;
 361         if (!td->o.timeout)
 362                 return false;
 363         if (utime_since(&td->epoch, t) >= td->o.timeout)
 364                 return true;
 365
 366         return false;
 367 }
 368
 369 /*
 370  * We need to update the runtime consistently in ms, but keep a running
 371  * tally of the current elapsed time in microseconds for sub millisecond
 372  * updates.
 373  */
 374 static inline void update_runtime(struct thread_data *td,
 375                                   unsigned long long *elapsed_us,
 376                                   const enum fio_ddir ddir)
 377 {
 378         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 379                 return;
 380
 381         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 382         elapsed_us[ddir] += utime_since_now(&td->start);
 383         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 384 }
 385
 386 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 387                                 int *retptr)
 388 {
 389         int ret = *retptr;
 390
 391         if (ret < 0 || td->error) {
 392                 int err = td->error;
 393                 enum error_type_bit eb;
 394
 395                 if (ret < 0)
 396                         err = -ret;
 397
 398                 eb = td_error_type(ddir, err);
 399                 if (!(td->o.continue_on_error & (1 << eb)))
 400                         return true;
 401
 402                 if (td_non_fatal_error(td, eb, err)) {
 403                         /*
 404                          * Continue with the I/Os in case of
 405                          * a non fatal error.
 406                          */
 407                         update_error_count(td, err);
 408                         td_clear_error(td);
 409                         *retptr = 0;
 410                         return false;
 411                 } else if (td->o.fill_device && err == ENOSPC) {
 412                         /*
 413                          * We expect to hit this error if
 414                          * fill_device option is set.
 415                          */
 416                         td_clear_error(td);
 417                         fio_mark_td_terminate(td);
 418                         return true;
 419                 } else {
 420                         /*
 421                          * Stop the I/O in case of a fatal
 422                          * error.
 423                          */
 424                         update_error_count(td, err);
 425                         return true;
 426                 }
 427         }
 428
 429         return false;
 430 }
 431
 432 static void check_update_rusage(struct thread_data *td)
 433 {
 434         if (td->update_rusage) {
 435                 td->update_rusage = 0;
 436                 update_rusage_stat(td);
 437                 fio_mutex_up(td->rusage_sem);
 438         }
 439 }
 440
 441 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 442 {
 443         const int full = queue_full(td);
 444         int min_evts = 0;
 445         int ret;
 446
 447         /*
 448          * if the queue is full, we MUST reap at least 1 event
 449          */
 450         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 451         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 452                 min_evts = 1;
 453
 454         if (time && (__should_check_rate(td, DDIR_READ) ||
 455             __should_check_rate(td, DDIR_WRITE) ||
 456             __should_check_rate(td, DDIR_TRIM)))
 457                 fio_gettime(time, NULL);
 458
 459         do {
 460                 ret = io_u_queued_complete(td, min_evts);
 461                 if (ret < 0)
 462                         break;
 463         } while (full && (td->cur_depth > td->o.iodepth_low));
 464
 465         return ret;
 466 }
 467
 468 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 469                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 470                    struct timeval *comp_time)
 471 {
 472         int ret2;
 473
 474         switch (*ret) {
 475         case FIO_Q_COMPLETED:
 476                 if (io_u->error) {
 477                         *ret = -io_u->error;
 478                         clear_io_u(td, io_u);
 479                 } else if (io_u->resid) {
 480                         int bytes = io_u->xfer_buflen - io_u->resid;
 481                         struct fio_file *f = io_u->file;
 482
 483                         if (bytes_issued)
 484                                 *bytes_issued += bytes;
 485
 486                         if (!from_verify)
 487                                 trim_io_piece(td, io_u);
 488
 489                         /*
 490                          * zero read, fail
 491                          */
 492                         if (!bytes) {
 493                                 if (!from_verify)
 494                                         unlog_io_piece(td, io_u);
 495                                 td_verror(td, EIO, "full resid");
 496                                 put_io_u(td, io_u);
 497                                 break;
 498                         }
 499
 500                         io_u->xfer_buflen = io_u->resid;
 501                         io_u->xfer_buf += bytes;
 502                         io_u->offset += bytes;
 503
 504                         if (ddir_rw(io_u->ddir))
 505                                 td->ts.short_io_u[io_u->ddir]++;
 506
 507                         f = io_u->file;
 508                         if (io_u->offset == f->real_file_size)
 509                                 goto sync_done;
 510
 511                         requeue_io_u(td, &io_u);
 512                 } else {
 513 sync_done:
 514                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 515                             __should_check_rate(td, DDIR_WRITE) ||
 516                             __should_check_rate(td, DDIR_TRIM)))
 517                                 fio_gettime(comp_time, NULL);
 518
 519                         *ret = io_u_sync_complete(td, io_u);
 520                         if (*ret < 0)
 521                                 break;
 522                 }
 523                 return 0;
 524         case FIO_Q_QUEUED:
 525                 /*
 526                  * if the engine doesn't have a commit hook,
 527                  * the io_u is really queued. if it does have such
 528                  * a hook, it has to call io_u_queued() itself.
 529                  */
 530                 if (td->io_ops->commit == NULL)
 531                         io_u_queued(td, io_u);
 532                 if (bytes_issued)
 533                         *bytes_issued += io_u->xfer_buflen;
 534                 break;
 535         case FIO_Q_BUSY:
 536                 if (!from_verify)
 537                         unlog_io_piece(td, io_u);
 538                 requeue_io_u(td, &io_u);
 539                 ret2 = td_io_commit(td);
 540                 if (ret2 < 0)
 541                         *ret = ret2;
 542                 break;
 543         default:
 544                 assert(*ret < 0);
 545                 td_verror(td, -(*ret), "td_io_queue");
 546                 break;
 547         }
 548
 549         if (break_on_this_error(td, ddir, ret))
 550                 return 1;
 551
 552         return 0;
 553 }
 554
 555 static inline bool io_in_polling(struct thread_data *td)
 556 {
 557         return !td->o.iodepth_batch_complete_min &&
 558                    !td->o.iodepth_batch_complete_max;
 559 }
 560
 561 /*
 562  * The main verify engine. Runs over the writes we previously submitted,
 563  * reads the blocks back in, and checks the crc/md5 of the data.
 564  */
 565 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 566 {
 567         struct fio_file *f;
 568         struct io_u *io_u;
 569         int ret, min_events;
 570         unsigned int i;
 571
 572         dprint(FD_VERIFY, "starting loop\n");
 573
 574         /*
 575          * sync io first and invalidate cache, to make sure we really
 576          * read from disk.
 577          */
 578         for_each_file(td, f, i) {
 579                 if (!fio_file_open(f))
 580                         continue;
 581                 if (fio_io_sync(td, f))
 582                         break;
 583                 if (file_invalidate_cache(td, f))
 584                         break;
 585         }
 586
 587         check_update_rusage(td);
 588
 589         if (td->error)
 590                 return;
 591
 592         td_set_runstate(td, TD_VERIFYING);
 593
 594         io_u = NULL;
 595         while (!td->terminate) {
 596                 enum fio_ddir ddir;
 597                 int full;
 598
 599                 update_tv_cache(td);
 600                 check_update_rusage(td);
 601
 602                 if (runtime_exceeded(td, &td->tv_cache)) {
 603                         __update_tv_cache(td);
 604                         if (runtime_exceeded(td, &td->tv_cache)) {
 605                                 fio_mark_td_terminate(td);
 606                                 break;
 607                         }
 608                 }
 609
 610                 if (flow_threshold_exceeded(td))
 611                         continue;
 612
 613                 if (!td->o.experimental_verify) {
 614                         io_u = __get_io_u(td);
 615                         if (!io_u)
 616                                 break;
 617
 618                         if (get_next_verify(td, io_u)) {
 619                                 put_io_u(td, io_u);
 620                                 break;
 621                         }
 622
 623                         if (td_io_prep(td, io_u)) {
 624                                 put_io_u(td, io_u);
 625                                 break;
 626                         }
 627                 } else {
 628                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 629                                 break;
 630
 631                         while ((io_u = get_io_u(td)) != NULL) {
 632                                 if (IS_ERR(io_u)) {
 633                                         io_u = NULL;
 634                                         ret = FIO_Q_BUSY;
 635                                         goto reap;
 636                                 }
 637
 638                                 /*
 639                                  * We are only interested in the places where
 640                                  * we wrote or trimmed IOs. Turn those into
 641                                  * reads for verification purposes.
 642                                  */
 643                                 if (io_u->ddir == DDIR_READ) {
 644                                         /*
 645                                          * Pretend we issued it for rwmix
 646                                          * accounting
 647                                          */
 648                                         td->io_issues[DDIR_READ]++;
 649                                         put_io_u(td, io_u);
 650                                         continue;
 651                                 } else if (io_u->ddir == DDIR_TRIM) {
 652                                         io_u->ddir = DDIR_READ;
 653                                         io_u_set(io_u, IO_U_F_TRIMMED);
 654                                         break;
 655                                 } else if (io_u->ddir == DDIR_WRITE) {
 656                                         io_u->ddir = DDIR_READ;
 657                                         break;
 658                                 } else {
 659                                         put_io_u(td, io_u);
 660                                         continue;
 661                                 }
 662                         }
 663
 664                         if (!io_u)
 665                                 break;
 666                 }
 667
 668                 if (verify_state_should_stop(td, io_u)) {
 669                         put_io_u(td, io_u);
 670                         break;
 671                 }
 672
 673                 if (td->o.verify_async)
 674                         io_u->end_io = verify_io_u_async;
 675                 else
 676                         io_u->end_io = verify_io_u;
 677
 678                 ddir = io_u->ddir;
 679                 if (!td->o.disable_slat)
 680                         fio_gettime(&io_u->start_time, NULL);
 681
 682                 ret = td_io_queue(td, io_u);
 683
 684                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 685                         break;
 686
 687                 /*
 688                  * if we can queue more, do so. but check if there are
 689                  * completed io_u's first. Note that we can get BUSY even
 690                  * without IO queued, if the system is resource starved.
 691                  */
 692 reap:
 693                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 694                 if (full || io_in_polling(td))
 695                         ret = wait_for_completions(td, NULL);
 696
 697                 if (ret < 0)
 698                         break;
 699         }
 700
 701         check_update_rusage(td);
 702
 703         if (!td->error) {
 704                 min_events = td->cur_depth;
 705
 706                 if (min_events)
 707                         ret = io_u_queued_complete(td, min_events);
 708         } else
 709                 cleanup_pending_aio(td);
 710
 711         td_set_runstate(td, TD_RUNNING);
 712
 713         dprint(FD_VERIFY, "exiting loop\n");
 714 }
 715
 716 static bool exceeds_number_ios(struct thread_data *td)
 717 {
 718         unsigned long long number_ios;
 719
 720         if (!td->o.number_ios)
 721                 return false;
 722
 723         number_ios = ddir_rw_sum(td->io_blocks);
 724         number_ios += td->io_u_queued + td->io_u_in_flight;
 725
 726         return number_ios >= (td->o.number_ios * td->loops);
 727 }
 728
 729 static bool io_issue_bytes_exceeded(struct thread_data *td)
 730 {
 731         unsigned long long bytes, limit;
 732
 733         if (td_rw(td))
 734                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 735         else if (td_write(td))
 736                 bytes = td->io_issue_bytes[DDIR_WRITE];
 737         else if (td_read(td))
 738                 bytes = td->io_issue_bytes[DDIR_READ];
 739         else
 740                 bytes = td->io_issue_bytes[DDIR_TRIM];
 741
 742         if (td->o.io_limit)
 743                 limit = td->o.io_limit;
 744         else
 745                 limit = td->o.size;
 746
 747         limit *= td->loops;
 748         return bytes >= limit || exceeds_number_ios(td);
 749 }
 750
 751 static bool io_complete_bytes_exceeded(struct thread_data *td)
 752 {
 753         unsigned long long bytes, limit;
 754
 755         if (td_rw(td))
 756                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 757         else if (td_write(td))
 758                 bytes = td->this_io_bytes[DDIR_WRITE];
 759         else if (td_read(td))
 760                 bytes = td->this_io_bytes[DDIR_READ];
 761         else
 762                 bytes = td->this_io_bytes[DDIR_TRIM];
 763
 764         if (td->o.io_limit)
 765                 limit = td->o.io_limit;
 766         else
 767                 limit = td->o.size;
 768
 769         limit *= td->loops;
 770         return bytes >= limit || exceeds_number_ios(td);
 771 }
 772
 773 /*
 774  * used to calculate the next io time for rate control
 775  *
 776  */
 777 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 778 {
 779         uint64_t secs, remainder, bps, bytes, iops;
 780
 781         assert(!(td->flags & TD_F_CHILD));
 782         bytes = td->rate_io_issue_bytes[ddir];
 783         bps = td->rate_bps[ddir];
 784
 785         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 786                 uint64_t val;
 787                 iops = bps / td->o.bs[ddir];
 788                 val = (int64_t) (1000000 / iops) *
 789                                 -logf(__rand_0_1(&td->poisson_state));
 790                 if (val) {
 791                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 792                                         (unsigned long long) 1000000 / val);
 793                 }
 794                 td->last_usec += val;
 795                 return td->last_usec;
 796         } else if (bps) {
 797                 secs = bytes / bps;
 798                 remainder = bytes % bps;
 799                 return remainder * 1000000 / bps + secs * 1000000;
 800         }
 801
 802         return 0;
 803 }
 804
 805 /*
 806  * Main IO worker function. It retrieves io_u's to process and queues
 807  * and reaps them, checking for rate and errors along the way.
 808  *
 809  * Returns number of bytes written and trimmed.
 810  */
 811 static uint64_t do_io(struct thread_data *td)
 812 {
 813         unsigned int i;
 814         int ret = 0;
 815         uint64_t total_bytes, bytes_issued = 0;
 816         uint64_t this_bytes[2];
 817
 818         this_bytes[0] = td->bytes_done[DDIR_WRITE];
 819         this_bytes[1] = td->bytes_done[DDIR_TRIM];
 820
 821         if (in_ramp_time(td))
 822                 td_set_runstate(td, TD_RAMP);
 823         else
 824                 td_set_runstate(td, TD_RUNNING);
 825
 826         lat_target_init(td);
 827
 828         total_bytes = td->o.size;
 829         /*
 830         * Allow random overwrite workloads to write up to io_limit
 831         * before starting verification phase as 'size' doesn't apply.
 832         */
 833         if (td_write(td) && td_random(td) && td->o.norandommap)
 834                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 835         /*
 836          * If verify_backlog is enabled, we'll run the verify in this
 837          * handler as well. For that case, we may need up to twice the
 838          * amount of bytes.
 839          */
 840         if (td->o.verify != VERIFY_NONE &&
 841            (td_write(td) && td->o.verify_backlog))
 842                 total_bytes += td->o.size;
 843
 844         /* In trimwrite mode, each byte is trimmed and then written, so
 845          * allow total_bytes to be twice as big */
 846         if (td_trimwrite(td))
 847                 total_bytes += td->total_io_size;
 848
 849         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 850                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 851                 td->o.time_based) {
 852                 struct timeval comp_time;
 853                 struct io_u *io_u;
 854                 int full;
 855                 enum fio_ddir ddir;
 856
 857                 check_update_rusage(td);
 858
 859                 if (td->terminate || td->done)
 860                         break;
 861
 862                 update_tv_cache(td);
 863
 864                 if (runtime_exceeded(td, &td->tv_cache)) {
 865                         __update_tv_cache(td);
 866                         if (runtime_exceeded(td, &td->tv_cache)) {
 867                                 fio_mark_td_terminate(td);
 868                                 break;
 869                         }
 870                 }
 871
 872                 if (flow_threshold_exceeded(td))
 873                         continue;
 874
 875                 if (!td->o.time_based && bytes_issued >= total_bytes)
 876                         break;
 877
 878                 io_u = get_io_u(td);
 879                 if (IS_ERR_OR_NULL(io_u)) {
 880                         int err = PTR_ERR(io_u);
 881
 882                         io_u = NULL;
 883                         if (err == -EBUSY) {
 884                                 ret = FIO_Q_BUSY;
 885                                 goto reap;
 886                         }
 887                         if (td->o.latency_target)
 888                                 goto reap;
 889                         break;
 890                 }
 891
 892                 ddir = io_u->ddir;
 893
 894                 /*
 895                  * Add verification end_io handler if:
 896                  *      - Asked to verify (!td_rw(td))
 897                  *      - Or the io_u is from our verify list (mixed write/ver)
 898                  */
 899                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 900                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 901
 902                         if (!td->o.verify_pattern_bytes) {
 903                                 io_u->rand_seed = __rand(&td->verify_state);
 904                                 if (sizeof(int) != sizeof(long *))
 905                                         io_u->rand_seed *= __rand(&td->verify_state);
 906                         }
 907
 908                         if (verify_state_should_stop(td, io_u)) {
 909                                 put_io_u(td, io_u);
 910                                 break;
 911                         }
 912
 913                         if (td->o.verify_async)
 914                                 io_u->end_io = verify_io_u_async;
 915                         else
 916                                 io_u->end_io = verify_io_u;
 917                         td_set_runstate(td, TD_VERIFYING);
 918                 } else if (in_ramp_time(td))
 919                         td_set_runstate(td, TD_RAMP);
 920                 else
 921                         td_set_runstate(td, TD_RUNNING);
 922
 923                 /*
 924                  * Always log IO before it's issued, so we know the specific
 925                  * order of it. The logged unit will track when the IO has
 926                  * completed.
 927                  */
 928                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 929                     td->o.do_verify &&
 930                     td->o.verify != VERIFY_NONE &&
 931                     !td->o.experimental_verify)
 932                         log_io_piece(td, io_u);
 933
 934                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 935                         const unsigned long blen = io_u->xfer_buflen;
 936                         const enum fio_ddir ddir = acct_ddir(io_u);
 937
 938                         if (td->error)
 939                                 break;
 940
 941                         workqueue_enqueue(&td->io_wq, &io_u->work);
 942                         ret = FIO_Q_QUEUED;
 943
 944                         if (ddir_rw(ddir)) {
 945                                 td->io_issues[ddir]++;
 946                                 td->io_issue_bytes[ddir] += blen;
 947                                 td->rate_io_issue_bytes[ddir] += blen;
 948                         }
 949
 950                         if (should_check_rate(td))
 951                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 952
 953                 } else {
 954                         ret = td_io_queue(td, io_u);
 955
 956                         if (should_check_rate(td))
 957                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 958
 959                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 960                                 break;
 961
 962                         /*
 963                          * See if we need to complete some commands. Note that
 964                          * we can get BUSY even without IO queued, if the
 965                          * system is resource starved.
 966                          */
 967 reap:
 968                         full = queue_full(td) ||
 969                                 (ret == FIO_Q_BUSY && td->cur_depth);
 970                         if (full || io_in_polling(td))
 971                                 ret = wait_for_completions(td, &comp_time);
 972                 }
 973                 if (ret < 0)
 974                         break;
 975                 if (!ddir_rw_sum(td->bytes_done) &&
 976                     !(td->io_ops->flags & FIO_NOIO))
 977                         continue;
 978
 979                 if (!in_ramp_time(td) && should_check_rate(td)) {
 980                         if (check_min_rate(td, &comp_time)) {
 981                                 if (exitall_on_terminate || td->o.exitall_error)
 982                                         fio_terminate_threads(td->groupid);
 983                                 td_verror(td, EIO, "check_min_rate");
 984                                 break;
 985                         }
 986                 }
 987                 if (!in_ramp_time(td) && td->o.latency_target)
 988                         lat_target_check(td);
 989
 990                 if (td->o.thinktime) {
 991                         unsigned long long b;
 992
 993                         b = ddir_rw_sum(td->io_blocks);
 994                         if (!(b % td->o.thinktime_blocks)) {
 995                                 int left;
 996
 997                                 io_u_quiesce(td);
 998
 999                                 if (td->o.thinktime_spin)
1000                                         usec_spin(td->o.thinktime_spin);
1001
1002                                 left = td->o.thinktime - td->o.thinktime_spin;
1003                                 if (left)
1004                                         usec_sleep(td, left);
1005                         }
1006                 }
1007         }
1008
1009         check_update_rusage(td);
1010
1011         if (td->trim_entries)
1012                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1013
1014         if (td->o.fill_device && td->error == ENOSPC) {
1015                 td->error = 0;
1016                 fio_mark_td_terminate(td);
1017         }
1018         if (!td->error) {
1019                 struct fio_file *f;
1020
1021                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1022                         workqueue_flush(&td->io_wq);
1023                         i = 0;
1024                 } else
1025                         i = td->cur_depth;
1026
1027                 if (i) {
1028                         ret = io_u_queued_complete(td, i);
1029                         if (td->o.fill_device && td->error == ENOSPC)
1030                                 td->error = 0;
1031                 }
1032
1033                 if (should_fsync(td) && td->o.end_fsync) {
1034                         td_set_runstate(td, TD_FSYNCING);
1035
1036                         for_each_file(td, f, i) {
1037                                 if (!fio_file_fsync(td, f))
1038                                         continue;
1039
1040                                 log_err("fio: end_fsync failed for file %s\n",
1041                                                                 f->file_name);
1042                         }
1043                 }
1044         } else
1045                 cleanup_pending_aio(td);
1046
1047         /*
1048          * stop job if we failed doing any IO
1049          */
1050         if (!ddir_rw_sum(td->this_io_bytes))
1051                 td->done = 1;
1052
1053         return (td->bytes_done[DDIR_WRITE] - this_bytes[0]) +
1054                 (td->bytes_done[DDIR_TRIM] - this_bytes[1]);
1055 }
1056
1057 static void cleanup_io_u(struct thread_data *td)
1058 {
1059         struct io_u *io_u;
1060
1061         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1062
1063                 if (td->io_ops->io_u_free)
1064                         td->io_ops->io_u_free(td, io_u);
1065
1066                 fio_memfree(io_u, sizeof(*io_u));
1067         }
1068
1069         free_io_mem(td);
1070
1071         io_u_rexit(&td->io_u_requeues);
1072         io_u_qexit(&td->io_u_freelist);
1073         io_u_qexit(&td->io_u_all);
1074
1075         if (td->last_write_comp)
1076                 sfree(td->last_write_comp);
1077 }
1078
1079 static int init_io_u(struct thread_data *td)
1080 {
1081         struct io_u *io_u;
1082         unsigned int max_bs, min_write;
1083         int cl_align, i, max_units;
1084         int data_xfer = 1, err;
1085         char *p;
1086
1087         max_units = td->o.iodepth;
1088         max_bs = td_max_bs(td);
1089         min_write = td->o.min_bs[DDIR_WRITE];
1090         td->orig_buffer_size = (unsigned long long) max_bs
1091                                         * (unsigned long long) max_units;
1092
1093         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1094                 data_xfer = 0;
1095
1096         err = 0;
1097         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1098         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1099         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1100
1101         if (err) {
1102                 log_err("fio: failed setting up IO queues\n");
1103                 return 1;
1104         }
1105
1106         /*
1107          * if we may later need to do address alignment, then add any
1108          * possible adjustment here so that we don't cause a buffer
1109          * overflow later. this adjustment may be too much if we get
1110          * lucky and the allocator gives us an aligned address.
1111          */
1112         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1113             (td->io_ops->flags & FIO_RAWIO))
1114                 td->orig_buffer_size += page_mask + td->o.mem_align;
1115
1116         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1117                 unsigned long bs;
1118
1119                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1120                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1121         }
1122
1123         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1124                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1125                 return 1;
1126         }
1127
1128         if (data_xfer && allocate_io_mem(td))
1129                 return 1;
1130
1131         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1132             (td->io_ops->flags & FIO_RAWIO))
1133                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1134         else
1135                 p = td->orig_buffer;
1136
1137         cl_align = os_cache_line_size();
1138
1139         for (i = 0; i < max_units; i++) {
1140                 void *ptr;
1141
1142                 if (td->terminate)
1143                         return 1;
1144
1145                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1146                 if (!ptr) {
1147                         log_err("fio: unable to allocate aligned memory\n");
1148                         break;
1149                 }
1150
1151                 io_u = ptr;
1152                 memset(io_u, 0, sizeof(*io_u));
1153                 INIT_FLIST_HEAD(&io_u->verify_list);
1154                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1155
1156                 if (data_xfer) {
1157                         io_u->buf = p;
1158                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1159
1160                         if (td_write(td))
1161                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1162                         if (td_write(td) && td->o.verify_pattern_bytes) {
1163                                 /*
1164                                  * Fill the buffer with the pattern if we are
1165                                  * going to be doing writes.
1166                                  */
1167                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1168                         }
1169                 }
1170
1171                 io_u->index = i;
1172                 io_u->flags = IO_U_F_FREE;
1173                 io_u_qpush(&td->io_u_freelist, io_u);
1174
1175                 /*
1176                  * io_u never leaves this stack, used for iteration of all
1177                  * io_u buffers.
1178                  */
1179                 io_u_qpush(&td->io_u_all, io_u);
1180
1181                 if (td->io_ops->io_u_init) {
1182                         int ret = td->io_ops->io_u_init(td, io_u);
1183
1184                         if (ret) {
1185                                 log_err("fio: failed to init engine data: %d\n", ret);
1186                                 return 1;
1187                         }
1188                 }
1189
1190                 p += max_bs;
1191         }
1192
1193         if (td->o.verify != VERIFY_NONE) {
1194                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1195                 if (!td->last_write_comp) {
1196                         log_err("fio: failed to alloc write comp data\n");
1197                         return 1;
1198                 }
1199         }
1200
1201         return 0;
1202 }
1203
1204 static int switch_ioscheduler(struct thread_data *td)
1205 {
1206         char tmp[256], tmp2[128];
1207         FILE *f;
1208         int ret;
1209
1210         if (td->io_ops->flags & FIO_DISKLESSIO)
1211                 return 0;
1212
1213         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1214
1215         f = fopen(tmp, "r+");
1216         if (!f) {
1217                 if (errno == ENOENT) {
1218                         log_err("fio: os or kernel doesn't support IO scheduler"
1219                                 " switching\n");
1220                         return 0;
1221                 }
1222                 td_verror(td, errno, "fopen iosched");
1223                 return 1;
1224         }
1225
1226         /*
1227          * Set io scheduler.
1228          */
1229         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1230         if (ferror(f) || ret != 1) {
1231                 td_verror(td, errno, "fwrite");
1232                 fclose(f);
1233                 return 1;
1234         }
1235
1236         rewind(f);
1237
1238         /*
1239          * Read back and check that the selected scheduler is now the default.
1240          */
1241         memset(tmp, 0, sizeof(tmp));
1242         ret = fread(tmp, sizeof(tmp), 1, f);
1243         if (ferror(f) || ret < 0) {
1244                 td_verror(td, errno, "fread");
1245                 fclose(f);
1246                 return 1;
1247         }
1248         /*
1249          * either a list of io schedulers or "none\n" is expected.
1250          */
1251         tmp[strlen(tmp) - 1] = '\0';
1252
1253
1254         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1255         if (!strstr(tmp, tmp2)) {
1256                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1257                 td_verror(td, EINVAL, "iosched_switch");
1258                 fclose(f);
1259                 return 1;
1260         }
1261
1262         fclose(f);
1263         return 0;
1264 }
1265
1266 static bool keep_running(struct thread_data *td)
1267 {
1268         unsigned long long limit;
1269
1270         if (td->done)
1271                 return false;
1272         if (td->o.time_based)
1273                 return true;
1274         if (td->o.loops) {
1275                 td->o.loops--;
1276                 return true;
1277         }
1278         if (exceeds_number_ios(td))
1279                 return false;
1280
1281         if (td->o.io_limit)
1282                 limit = td->o.io_limit;
1283         else
1284                 limit = td->o.size;
1285
1286         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1287                 uint64_t diff;
1288
1289                 /*
1290                  * If the difference is less than the minimum IO size, we
1291                  * are done.
1292                  */
1293                 diff = limit - ddir_rw_sum(td->io_bytes);
1294                 if (diff < td_max_bs(td))
1295                         return false;
1296
1297                 if (fio_files_done(td))
1298                         return false;
1299
1300                 return true;
1301         }
1302
1303         return false;
1304 }
1305
1306 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1307 {
1308         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1309         int ret;
1310         char *str;
1311
1312         str = malloc(newlen);
1313         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1314
1315         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1316         ret = system(str);
1317         if (ret == -1)
1318                 log_err("fio: exec of cmd <%s> failed\n", str);
1319
1320         free(str);
1321         return ret;
1322 }
1323
1324 /*
1325  * Dry run to compute correct state of numberio for verification.
1326  */
1327 static uint64_t do_dry_run(struct thread_data *td)
1328 {
1329         td_set_runstate(td, TD_RUNNING);
1330
1331         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1332                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1333                 struct io_u *io_u;
1334                 int ret;
1335
1336                 if (td->terminate || td->done)
1337                         break;
1338
1339                 io_u = get_io_u(td);
1340                 if (!io_u)
1341                         break;
1342
1343                 io_u_set(io_u, IO_U_F_FLIGHT);
1344                 io_u->error = 0;
1345                 io_u->resid = 0;
1346                 if (ddir_rw(acct_ddir(io_u)))
1347                         td->io_issues[acct_ddir(io_u)]++;
1348                 if (ddir_rw(io_u->ddir)) {
1349                         io_u_mark_depth(td, 1);
1350                         td->ts.total_io_u[io_u->ddir]++;
1351                 }
1352
1353                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1354                     td->o.do_verify &&
1355                     td->o.verify != VERIFY_NONE &&
1356                     !td->o.experimental_verify)
1357                         log_io_piece(td, io_u);
1358
1359                 ret = io_u_sync_complete(td, io_u);
1360                 (void) ret;
1361         }
1362
1363         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1364 }
1365
1366 struct fork_data {
1367         struct thread_data *td;
1368         struct sk_out *sk_out;
1369 };
1370
1371 /*
1372  * Entry point for the thread based jobs. The process based jobs end up
1373  * here as well, after a little setup.
1374  */
1375 static void *thread_main(void *data)
1376 {
1377         struct fork_data *fd = data;
1378         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1379         struct thread_data *td = fd->td;
1380         struct thread_options *o = &td->o;
1381         struct sk_out *sk_out = fd->sk_out;
1382         pthread_condattr_t attr;
1383         int clear_state;
1384         int ret;
1385
1386         sk_out_assign(sk_out);
1387         free(fd);
1388
1389         if (!o->use_thread) {
1390                 setsid();
1391                 td->pid = getpid();
1392         } else
1393                 td->pid = gettid();
1394
1395         fio_local_clock_init(o->use_thread);
1396
1397         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1398
1399         if (is_backend)
1400                 fio_server_send_start(td);
1401
1402         INIT_FLIST_HEAD(&td->io_log_list);
1403         INIT_FLIST_HEAD(&td->io_hist_list);
1404         INIT_FLIST_HEAD(&td->verify_list);
1405         INIT_FLIST_HEAD(&td->trim_list);
1406         INIT_FLIST_HEAD(&td->next_rand_list);
1407         pthread_mutex_init(&td->io_u_lock, NULL);
1408         td->io_hist_tree = RB_ROOT;
1409
1410         pthread_condattr_init(&attr);
1411         pthread_cond_init(&td->verify_cond, &attr);
1412         pthread_cond_init(&td->free_cond, &attr);
1413
1414         td_set_runstate(td, TD_INITIALIZED);
1415         dprint(FD_MUTEX, "up startup_mutex\n");
1416         fio_mutex_up(startup_mutex);
1417         dprint(FD_MUTEX, "wait on td->mutex\n");
1418         fio_mutex_down(td->mutex);
1419         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1420
1421         /*
1422          * A new gid requires privilege, so we need to do this before setting
1423          * the uid.
1424          */
1425         if (o->gid != -1U && setgid(o->gid)) {
1426                 td_verror(td, errno, "setgid");
1427                 goto err;
1428         }
1429         if (o->uid != -1U && setuid(o->uid)) {
1430                 td_verror(td, errno, "setuid");
1431                 goto err;
1432         }
1433
1434         /*
1435          * If we have a gettimeofday() thread, make sure we exclude that
1436          * thread from this job
1437          */
1438         if (o->gtod_cpu)
1439                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1440
1441         /*
1442          * Set affinity first, in case it has an impact on the memory
1443          * allocations.
1444          */
1445         if (fio_option_is_set(o, cpumask)) {
1446                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1447                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1448                         if (!ret) {
1449                                 log_err("fio: no CPUs set\n");
1450                                 log_err("fio: Try increasing number of available CPUs\n");
1451                                 td_verror(td, EINVAL, "cpus_split");
1452                                 goto err;
1453                         }
1454                 }
1455                 ret = fio_setaffinity(td->pid, o->cpumask);
1456                 if (ret == -1) {
1457                         td_verror(td, errno, "cpu_set_affinity");
1458                         goto err;
1459                 }
1460         }
1461
1462 #ifdef CONFIG_LIBNUMA
1463         /* numa node setup */
1464         if (fio_option_is_set(o, numa_cpunodes) ||
1465             fio_option_is_set(o, numa_memnodes)) {
1466                 struct bitmask *mask;
1467
1468                 if (numa_available() < 0) {
1469                         td_verror(td, errno, "Does not support NUMA API\n");
1470                         goto err;
1471                 }
1472
1473                 if (fio_option_is_set(o, numa_cpunodes)) {
1474                         mask = numa_parse_nodestring(o->numa_cpunodes);
1475                         ret = numa_run_on_node_mask(mask);
1476                         numa_free_nodemask(mask);
1477                         if (ret == -1) {
1478                                 td_verror(td, errno, \
1479                                         "numa_run_on_node_mask failed\n");
1480                                 goto err;
1481                         }
1482                 }
1483
1484                 if (fio_option_is_set(o, numa_memnodes)) {
1485                         mask = NULL;
1486                         if (o->numa_memnodes)
1487                                 mask = numa_parse_nodestring(o->numa_memnodes);
1488
1489                         switch (o->numa_mem_mode) {
1490                         case MPOL_INTERLEAVE:
1491                                 numa_set_interleave_mask(mask);
1492                                 break;
1493                         case MPOL_BIND:
1494                                 numa_set_membind(mask);
1495                                 break;
1496                         case MPOL_LOCAL:
1497                                 numa_set_localalloc();
1498                                 break;
1499                         case MPOL_PREFERRED:
1500                                 numa_set_preferred(o->numa_mem_prefer_node);
1501                                 break;
1502                         case MPOL_DEFAULT:
1503                         default:
1504                                 break;
1505                         }
1506
1507                         if (mask)
1508                                 numa_free_nodemask(mask);
1509
1510                 }
1511         }
1512 #endif
1513
1514         if (fio_pin_memory(td))
1515                 goto err;
1516
1517         /*
1518          * May alter parameters that init_io_u() will use, so we need to
1519          * do this first.
1520          */
1521         if (init_iolog(td))
1522                 goto err;
1523
1524         if (init_io_u(td))
1525                 goto err;
1526
1527         if (o->verify_async && verify_async_init(td))
1528                 goto err;
1529
1530         if (fio_option_is_set(o, ioprio) ||
1531             fio_option_is_set(o, ioprio_class)) {
1532                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1533                 if (ret == -1) {
1534                         td_verror(td, errno, "ioprio_set");
1535                         goto err;
1536                 }
1537         }
1538
1539         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1540                 goto err;
1541
1542         errno = 0;
1543         if (nice(o->nice) == -1 && errno != 0) {
1544                 td_verror(td, errno, "nice");
1545                 goto err;
1546         }
1547
1548         if (o->ioscheduler && switch_ioscheduler(td))
1549                 goto err;
1550
1551         if (!o->create_serialize && setup_files(td))
1552                 goto err;
1553
1554         if (td_io_init(td))
1555                 goto err;
1556
1557         if (init_random_map(td))
1558                 goto err;
1559
1560         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1561                 goto err;
1562
1563         if (o->pre_read) {
1564                 if (pre_read_files(td) < 0)
1565                         goto err;
1566         }
1567
1568         if (iolog_compress_init(td, sk_out))
1569                 goto err;
1570
1571         fio_verify_init(td);
1572
1573         if (rate_submit_init(td, sk_out))
1574                 goto err;
1575
1576         fio_gettime(&td->epoch, NULL);
1577         fio_getrusage(&td->ru_start);
1578         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1579         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1580
1581         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1582                         o->ratemin[DDIR_TRIM]) {
1583                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1584                                         sizeof(td->bw_sample_time));
1585                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1586                                         sizeof(td->bw_sample_time));
1587                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1588                                         sizeof(td->bw_sample_time));
1589         }
1590
1591         clear_state = 0;
1592         while (keep_running(td)) {
1593                 uint64_t verify_bytes;
1594
1595                 fio_gettime(&td->start, NULL);
1596                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1597
1598                 if (clear_state)
1599                         clear_io_state(td, 0);
1600
1601                 prune_io_piece_log(td);
1602
1603                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1604                         verify_bytes = do_dry_run(td);
1605                 else {
1606                         verify_bytes = do_io(td);
1607                         if (!verify_bytes)
1608                                 fio_mark_td_terminate(td);
1609                 }
1610
1611                 clear_state = 1;
1612
1613                 /*
1614                  * Make sure we've successfully updated the rusage stats
1615                  * before waiting on the stat mutex. Otherwise we could have
1616                  * the stat thread holding stat mutex and waiting for
1617                  * the rusage_sem, which would never get upped because
1618                  * this thread is waiting for the stat mutex.
1619                  */
1620                 check_update_rusage(td);
1621
1622                 fio_mutex_down(stat_mutex);
1623                 if (td_read(td) && td->io_bytes[DDIR_READ])
1624                         update_runtime(td, elapsed_us, DDIR_READ);
1625                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1626                         update_runtime(td, elapsed_us, DDIR_WRITE);
1627                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1628                         update_runtime(td, elapsed_us, DDIR_TRIM);
1629                 fio_gettime(&td->start, NULL);
1630                 fio_mutex_up(stat_mutex);
1631
1632                 if (td->error || td->terminate)
1633                         break;
1634
1635                 if (!o->do_verify ||
1636                     o->verify == VERIFY_NONE ||
1637                     (td->io_ops->flags & FIO_UNIDIR))
1638                         continue;
1639
1640                 clear_io_state(td, 0);
1641
1642                 fio_gettime(&td->start, NULL);
1643
1644                 do_verify(td, verify_bytes);
1645
1646                 /*
1647                  * See comment further up for why this is done here.
1648                  */
1649                 check_update_rusage(td);
1650
1651                 fio_mutex_down(stat_mutex);
1652                 update_runtime(td, elapsed_us, DDIR_READ);
1653                 fio_gettime(&td->start, NULL);
1654                 fio_mutex_up(stat_mutex);
1655
1656                 if (td->error || td->terminate)
1657                         break;
1658         }
1659
1660         update_rusage_stat(td);
1661         td->ts.total_run_time = mtime_since_now(&td->epoch);
1662         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1663         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1664         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1665
1666         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1667             (td->o.verify != VERIFY_NONE && td_write(td)))
1668                 verify_save_state(td->thread_number);
1669
1670         fio_unpin_memory(td);
1671
1672         fio_writeout_logs(td);
1673
1674         iolog_compress_exit(td);
1675         rate_submit_exit(td);
1676
1677         if (o->exec_postrun)
1678                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1679
1680         if (exitall_on_terminate || (o->exitall_error && td->error))
1681                 fio_terminate_threads(td->groupid);
1682
1683 err:
1684         if (td->error)
1685                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1686                                                         td->verror);
1687
1688         if (o->verify_async)
1689                 verify_async_exit(td);
1690
1691         close_and_free_files(td);
1692         cleanup_io_u(td);
1693         close_ioengine(td);
1694         cgroup_shutdown(td, &cgroup_mnt);
1695         verify_free_state(td);
1696
1697         if (fio_option_is_set(o, cpumask)) {
1698                 ret = fio_cpuset_exit(&o->cpumask);
1699                 if (ret)
1700                         td_verror(td, ret, "fio_cpuset_exit");
1701         }
1702
1703         /*
1704          * do this very late, it will log file closing as well
1705          */
1706         if (o->write_iolog_file)
1707                 write_iolog_close(td);
1708
1709         fio_mutex_remove(td->mutex);
1710         td->mutex = NULL;
1711
1712         td_set_runstate(td, TD_EXITED);
1713
1714         /*
1715          * Do this last after setting our runstate to exited, so we
1716          * know that the stat thread is signaled.
1717          */
1718         check_update_rusage(td);
1719
1720         sk_out_drop();
1721         return (void *) (uintptr_t) td->error;
1722 }
1723
1724
1725 /*
1726  * We cannot pass the td data into a forked process, so attach the td and
1727  * pass it to the thread worker.
1728  */
1729 static int fork_main(struct sk_out *sk_out, int shmid, int offset)
1730 {
1731         struct fork_data *fd;
1732         void *data, *ret;
1733
1734 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1735         data = shmat(shmid, NULL, 0);
1736         if (data == (void *) -1) {
1737                 int __err = errno;
1738
1739                 perror("shmat");
1740                 return __err;
1741         }
1742 #else
1743         /*
1744          * HP-UX inherits shm mappings?
1745          */
1746         data = threads;
1747 #endif
1748
1749         fd = calloc(1, sizeof(*fd));
1750         fd->td = data + offset * sizeof(struct thread_data);
1751         fd->sk_out = sk_out;
1752         ret = thread_main(fd);
1753         shmdt(data);
1754         return (int) (uintptr_t) ret;
1755 }
1756
1757 static void dump_td_info(struct thread_data *td)
1758 {
1759         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1760                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1761                         (unsigned long) time_since_now(&td->terminate_time));
1762 }
1763
1764 /*
1765  * Run over the job map and reap the threads that have exited, if any.
1766  */
1767 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1768                          unsigned int *m_rate)
1769 {
1770         struct thread_data *td;
1771         unsigned int cputhreads, realthreads, pending;
1772         int i, status, ret;
1773
1774         /*
1775          * reap exited threads (TD_EXITED -> TD_REAPED)
1776          */
1777         realthreads = pending = cputhreads = 0;
1778         for_each_td(td, i) {
1779                 int flags = 0;
1780
1781                 /*
1782                  * ->io_ops is NULL for a thread that has closed its
1783                  * io engine
1784                  */
1785                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1786                         cputhreads++;
1787                 else
1788                         realthreads++;
1789
1790                 if (!td->pid) {
1791                         pending++;
1792                         continue;
1793                 }
1794                 if (td->runstate == TD_REAPED)
1795                         continue;
1796                 if (td->o.use_thread) {
1797                         if (td->runstate == TD_EXITED) {
1798                                 td_set_runstate(td, TD_REAPED);
1799                                 goto reaped;
1800                         }
1801                         continue;
1802                 }
1803
1804                 flags = WNOHANG;
1805                 if (td->runstate == TD_EXITED)
1806                         flags = 0;
1807
1808                 /*
1809                  * check if someone quit or got killed in an unusual way
1810                  */
1811                 ret = waitpid(td->pid, &status, flags);
1812                 if (ret < 0) {
1813                         if (errno == ECHILD) {
1814                                 log_err("fio: pid=%d disappeared %d\n",
1815                                                 (int) td->pid, td->runstate);
1816                                 td->sig = ECHILD;
1817                                 td_set_runstate(td, TD_REAPED);
1818                                 goto reaped;
1819                         }
1820                         perror("waitpid");
1821                 } else if (ret == td->pid) {
1822                         if (WIFSIGNALED(status)) {
1823                                 int sig = WTERMSIG(status);
1824
1825                                 if (sig != SIGTERM && sig != SIGUSR2)
1826                                         log_err("fio: pid=%d, got signal=%d\n",
1827                                                         (int) td->pid, sig);
1828                                 td->sig = sig;
1829                                 td_set_runstate(td, TD_REAPED);
1830                                 goto reaped;
1831                         }
1832                         if (WIFEXITED(status)) {
1833                                 if (WEXITSTATUS(status) && !td->error)
1834                                         td->error = WEXITSTATUS(status);
1835
1836                                 td_set_runstate(td, TD_REAPED);
1837                                 goto reaped;
1838                         }
1839                 }
1840
1841                 /*
1842                  * If the job is stuck, do a forceful timeout of it and
1843                  * move on.
1844                  */
1845                 if (td->terminate &&
1846                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1847                         dump_td_info(td);
1848                         td_set_runstate(td, TD_REAPED);
1849                         goto reaped;
1850                 }
1851
1852                 /*
1853                  * thread is not dead, continue
1854                  */
1855                 pending++;
1856                 continue;
1857 reaped:
1858                 (*nr_running)--;
1859                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1860                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1861                 if (!td->pid)
1862                         pending--;
1863
1864                 if (td->error)
1865                         exit_value++;
1866
1867                 done_secs += mtime_since_now(&td->epoch) / 1000;
1868                 profile_td_exit(td);
1869         }
1870
1871         if (*nr_running == cputhreads && !pending && realthreads)
1872                 fio_terminate_threads(TERMINATE_ALL);
1873 }
1874
1875 static bool __check_trigger_file(void)
1876 {
1877         struct stat sb;
1878
1879         if (!trigger_file)
1880                 return false;
1881
1882         if (stat(trigger_file, &sb))
1883                 return false;
1884
1885         if (unlink(trigger_file) < 0)
1886                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1887                                                         strerror(errno));
1888
1889         return true;
1890 }
1891
1892 static bool trigger_timedout(void)
1893 {
1894         if (trigger_timeout)
1895                 return time_since_genesis() >= trigger_timeout;
1896
1897         return false;
1898 }
1899
1900 void exec_trigger(const char *cmd)
1901 {
1902         int ret;
1903
1904         if (!cmd)
1905                 return;
1906
1907         ret = system(cmd);
1908         if (ret == -1)
1909                 log_err("fio: failed executing %s trigger\n", cmd);
1910 }
1911
1912 void check_trigger_file(void)
1913 {
1914         if (__check_trigger_file() || trigger_timedout()) {
1915                 if (nr_clients)
1916                         fio_clients_send_trigger(trigger_remote_cmd);
1917                 else {
1918                         verify_save_state(IO_LIST_ALL);
1919                         fio_terminate_threads(TERMINATE_ALL);
1920                         exec_trigger(trigger_cmd);
1921                 }
1922         }
1923 }
1924
1925 static int fio_verify_load_state(struct thread_data *td)
1926 {
1927         int ret;
1928
1929         if (!td->o.verify_state)
1930                 return 0;
1931
1932         if (is_backend) {
1933                 void *data;
1934                 int ver;
1935
1936                 ret = fio_server_get_verify_state(td->o.name,
1937                                         td->thread_number - 1, &data, &ver);
1938                 if (!ret)
1939                         verify_convert_assign_state(td, data, ver);
1940         } else
1941                 ret = verify_load_state(td, "local");
1942
1943         return ret;
1944 }
1945
1946 static void do_usleep(unsigned int usecs)
1947 {
1948         check_for_running_stats();
1949         check_trigger_file();
1950         usleep(usecs);
1951 }
1952
1953 static bool check_mount_writes(struct thread_data *td)
1954 {
1955         struct fio_file *f;
1956         unsigned int i;
1957
1958         if (!td_write(td) || td->o.allow_mounted_write)
1959                 return false;
1960
1961         for_each_file(td, f, i) {
1962                 if (f->filetype != FIO_TYPE_BD)
1963                         continue;
1964                 if (device_is_mounted(f->file_name))
1965                         goto mounted;
1966         }
1967
1968         return false;
1969 mounted:
1970         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1971         return true;
1972 }
1973
1974 /*
1975  * Main function for kicking off and reaping jobs, as needed.
1976  */
1977 static void run_threads(struct sk_out *sk_out)
1978 {
1979         struct thread_data *td;
1980         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1981         uint64_t spent;
1982
1983         if (fio_gtod_offload && fio_start_gtod_thread())
1984                 return;
1985
1986         fio_idle_prof_init();
1987
1988         set_sig_handlers();
1989
1990         nr_thread = nr_process = 0;
1991         for_each_td(td, i) {
1992                 if (check_mount_writes(td))
1993                         return;
1994                 if (td->o.use_thread)
1995                         nr_thread++;
1996                 else
1997                         nr_process++;
1998         }
1999
2000         if (output_format & FIO_OUTPUT_NORMAL) {
2001                 log_info("Starting ");
2002                 if (nr_thread)
2003                         log_info("%d thread%s", nr_thread,
2004                                                 nr_thread > 1 ? "s" : "");
2005                 if (nr_process) {
2006                         if (nr_thread)
2007                                 log_info(" and ");
2008                         log_info("%d process%s", nr_process,
2009                                                 nr_process > 1 ? "es" : "");
2010                 }
2011                 log_info("\n");
2012                 log_info_flush();
2013         }
2014
2015         todo = thread_number;
2016         nr_running = 0;
2017         nr_started = 0;
2018         m_rate = t_rate = 0;
2019
2020         for_each_td(td, i) {
2021                 print_status_init(td->thread_number - 1);
2022
2023                 if (!td->o.create_serialize)
2024                         continue;
2025
2026                 if (fio_verify_load_state(td))
2027                         goto reap;
2028
2029                 /*
2030                  * do file setup here so it happens sequentially,
2031                  * we don't want X number of threads getting their
2032                  * client data interspersed on disk
2033                  */
2034                 if (setup_files(td)) {
2035 reap:
2036                         exit_value++;
2037                         if (td->error)
2038                                 log_err("fio: pid=%d, err=%d/%s\n",
2039                                         (int) td->pid, td->error, td->verror);
2040                         td_set_runstate(td, TD_REAPED);
2041                         todo--;
2042                 } else {
2043                         struct fio_file *f;
2044                         unsigned int j;
2045
2046                         /*
2047                          * for sharing to work, each job must always open
2048                          * its own files. so close them, if we opened them
2049                          * for creation
2050                          */
2051                         for_each_file(td, f, j) {
2052                                 if (fio_file_open(f))
2053                                         td_io_close_file(td, f);
2054                         }
2055                 }
2056         }
2057
2058         /* start idle threads before io threads start to run */
2059         fio_idle_prof_start();
2060
2061         set_genesis_time();
2062
2063         while (todo) {
2064                 struct thread_data *map[REAL_MAX_JOBS];
2065                 struct timeval this_start;
2066                 int this_jobs = 0, left;
2067
2068                 /*
2069                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2070                  */
2071                 for_each_td(td, i) {
2072                         if (td->runstate != TD_NOT_CREATED)
2073                                 continue;
2074
2075                         /*
2076                          * never got a chance to start, killed by other
2077                          * thread for some reason
2078                          */
2079                         if (td->terminate) {
2080                                 todo--;
2081                                 continue;
2082                         }
2083
2084                         if (td->o.start_delay) {
2085                                 spent = utime_since_genesis();
2086
2087                                 if (td->o.start_delay > spent)
2088                                         continue;
2089                         }
2090
2091                         if (td->o.stonewall && (nr_started || nr_running)) {
2092                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2093                                                         td->o.name);
2094                                 break;
2095                         }
2096
2097                         init_disk_util(td);
2098
2099                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2100                         td->update_rusage = 0;
2101
2102                         /*
2103                          * Set state to created. Thread will transition
2104                          * to TD_INITIALIZED when it's done setting up.
2105                          */
2106                         td_set_runstate(td, TD_CREATED);
2107                         map[this_jobs++] = td;
2108                         nr_started++;
2109
2110                         if (td->o.use_thread) {
2111                                 struct fork_data *fd;
2112                                 int ret;
2113
2114                                 fd = calloc(1, sizeof(*fd));
2115                                 fd->td = td;
2116                                 fd->sk_out = sk_out;
2117
2118                                 dprint(FD_PROCESS, "will pthread_create\n");
2119                                 ret = pthread_create(&td->thread, NULL,
2120                                                         thread_main, fd);
2121                                 if (ret) {
2122                                         log_err("pthread_create: %s\n",
2123                                                         strerror(ret));
2124                                         free(fd);
2125                                         nr_started--;
2126                                         break;
2127                                 }
2128                                 ret = pthread_detach(td->thread);
2129                                 if (ret)
2130                                         log_err("pthread_detach: %s",
2131                                                         strerror(ret));
2132                         } else {
2133                                 pid_t pid;
2134                                 dprint(FD_PROCESS, "will fork\n");
2135                                 pid = fork();
2136                                 if (!pid) {
2137                                         int ret = fork_main(sk_out, shm_id, i);
2138
2139                                         _exit(ret);
2140                                 } else if (i == fio_debug_jobno)
2141                                         *fio_debug_jobp = pid;
2142                         }
2143                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2144                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2145                                 log_err("fio: job startup hung? exiting.\n");
2146                                 fio_terminate_threads(TERMINATE_ALL);
2147                                 fio_abort = 1;
2148                                 nr_started--;
2149                                 break;
2150                         }
2151                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2152                 }
2153
2154                 /*
2155                  * Wait for the started threads to transition to
2156                  * TD_INITIALIZED.
2157                  */
2158                 fio_gettime(&this_start, NULL);
2159                 left = this_jobs;
2160                 while (left && !fio_abort) {
2161                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2162                                 break;
2163
2164                         do_usleep(100000);
2165
2166                         for (i = 0; i < this_jobs; i++) {
2167                                 td = map[i];
2168                                 if (!td)
2169                                         continue;
2170                                 if (td->runstate == TD_INITIALIZED) {
2171                                         map[i] = NULL;
2172                                         left--;
2173                                 } else if (td->runstate >= TD_EXITED) {
2174                                         map[i] = NULL;
2175                                         left--;
2176                                         todo--;
2177                                         nr_running++; /* work-around... */
2178                                 }
2179                         }
2180                 }
2181
2182                 if (left) {
2183                         log_err("fio: %d job%s failed to start\n", left,
2184                                         left > 1 ? "s" : "");
2185                         for (i = 0; i < this_jobs; i++) {
2186                                 td = map[i];
2187                                 if (!td)
2188                                         continue;
2189                                 kill(td->pid, SIGTERM);
2190                         }
2191                         break;
2192                 }
2193
2194                 /*
2195                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2196                  */
2197                 for_each_td(td, i) {
2198                         if (td->runstate != TD_INITIALIZED)
2199                                 continue;
2200
2201                         if (in_ramp_time(td))
2202                                 td_set_runstate(td, TD_RAMP);
2203                         else
2204                                 td_set_runstate(td, TD_RUNNING);
2205                         nr_running++;
2206                         nr_started--;
2207                         m_rate += ddir_rw_sum(td->o.ratemin);
2208                         t_rate += ddir_rw_sum(td->o.rate);
2209                         todo--;
2210                         fio_mutex_up(td->mutex);
2211                 }
2212
2213                 reap_threads(&nr_running, &t_rate, &m_rate);
2214
2215                 if (todo)
2216                         do_usleep(100000);
2217         }
2218
2219         while (nr_running) {
2220                 reap_threads(&nr_running, &t_rate, &m_rate);
2221                 do_usleep(10000);
2222         }
2223
2224         fio_idle_prof_stop();
2225
2226         update_io_ticks();
2227 }
2228
2229 static void wait_for_helper_thread_exit(void)
2230 {
2231         void *ret;
2232
2233         helper_exit = 1;
2234         pthread_cond_signal(&helper_cond);
2235         pthread_join(helper_thread, &ret);
2236 }
2237
2238 static void free_disk_util(void)
2239 {
2240         disk_util_prune_entries();
2241
2242         pthread_cond_destroy(&helper_cond);
2243 }
2244
2245 static void *helper_thread_main(void *data)
2246 {
2247         struct sk_out *sk_out = data;
2248         int ret = 0;
2249
2250         sk_out_assign(sk_out);
2251
2252         fio_mutex_up(startup_mutex);
2253
2254         while (!ret) {
2255                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2256                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2257                 struct timespec ts;
2258                 struct timeval tv;
2259
2260                 gettimeofday(&tv, NULL);
2261                 ts.tv_sec = tv.tv_sec + sec;
2262                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2263
2264                 if (ts.tv_nsec >= 1000000000ULL) {
2265                         ts.tv_nsec -= 1000000000ULL;
2266                         ts.tv_sec++;
2267                 }
2268
2269                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2270
2271                 ret = update_io_ticks();
2272
2273                 if (helper_do_stat) {
2274                         helper_do_stat = 0;
2275                         __show_running_run_stats();
2276                 }
2277
2278                 if (!is_backend)
2279                         print_thread_status();
2280         }
2281
2282         sk_out_drop();
2283         return NULL;
2284 }
2285
2286 static int create_helper_thread(struct sk_out *sk_out)
2287 {
2288         int ret;
2289
2290         setup_disk_util();
2291
2292         pthread_cond_init(&helper_cond, NULL);
2293         pthread_mutex_init(&helper_lock, NULL);
2294
2295         ret = pthread_create(&helper_thread, NULL, helper_thread_main, sk_out);
2296         if (ret) {
2297                 log_err("Can't create helper thread: %s\n", strerror(ret));
2298                 return 1;
2299         }
2300
2301         dprint(FD_MUTEX, "wait on startup_mutex\n");
2302         fio_mutex_down(startup_mutex);
2303         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2304         return 0;
2305 }
2306
2307 int fio_backend(struct sk_out *sk_out)
2308 {
2309         struct thread_data *td;
2310         int i;
2311
2312         if (exec_profile) {
2313                 if (load_profile(exec_profile))
2314                         return 1;
2315                 free(exec_profile);
2316                 exec_profile = NULL;
2317         }
2318         if (!thread_number)
2319                 return 0;
2320
2321         if (write_bw_log) {
2322                 struct log_params p = {
2323                         .log_type = IO_LOG_TYPE_BW,
2324                 };
2325
2326                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2327                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2328                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2329         }
2330
2331         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2332         if (startup_mutex == NULL)
2333                 return 1;
2334
2335         set_genesis_time();
2336         stat_init();
2337         create_helper_thread(sk_out);
2338
2339         cgroup_list = smalloc(sizeof(*cgroup_list));
2340         INIT_FLIST_HEAD(cgroup_list);
2341
2342         run_threads(sk_out);
2343
2344         wait_for_helper_thread_exit();
2345
2346         if (!fio_abort) {
2347                 __show_run_stats();
2348                 if (write_bw_log) {
2349                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2350                                 struct io_log *log = agg_io_log[i];
2351
2352                                 flush_log(log, 0);
2353                                 free_log(log);
2354                         }
2355                 }
2356         }
2357
2358         for_each_td(td, i) {
2359                 fio_options_free(td);
2360                 if (td->rusage_sem) {
2361                         fio_mutex_remove(td->rusage_sem);
2362                         td->rusage_sem = NULL;
2363                 }
2364         }
2365
2366         free_disk_util();
2367         cgroup_kill(cgroup_list);
2368         sfree(cgroup_list);
2369         sfree(cgroup_mnt);
2370
2371         fio_mutex_remove(startup_mutex);
2372         stat_exit();
2373         return exit_value;
2374 }