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
 524                 /*
 525                  * when doing I/O (not when verifying),
 526                  * check for any errors that are to be ignored
 527                  */
 528                 if (!from_verify)
 529                         break;
 530
 531                 return 0;
 532         case FIO_Q_QUEUED:
 533                 /*
 534                  * if the engine doesn't have a commit hook,
 535                  * the io_u is really queued. if it does have such
 536                  * a hook, it has to call io_u_queued() itself.
 537                  */
 538                 if (td->io_ops->commit == NULL)
 539                         io_u_queued(td, io_u);
 540                 if (bytes_issued)
 541                         *bytes_issued += io_u->xfer_buflen;
 542                 break;
 543         case FIO_Q_BUSY:
 544                 if (!from_verify)
 545                         unlog_io_piece(td, io_u);
 546                 requeue_io_u(td, &io_u);
 547                 ret2 = td_io_commit(td);
 548                 if (ret2 < 0)
 549                         *ret = ret2;
 550                 break;
 551         default:
 552                 assert(*ret < 0);
 553                 td_verror(td, -(*ret), "td_io_queue");
 554                 break;
 555         }
 556
 557         if (break_on_this_error(td, ddir, ret))
 558                 return 1;
 559
 560         return 0;
 561 }
 562
 563 static inline bool io_in_polling(struct thread_data *td)
 564 {
 565         return !td->o.iodepth_batch_complete_min &&
 566                    !td->o.iodepth_batch_complete_max;
 567 }
 568
 569 /*
 570  * The main verify engine. Runs over the writes we previously submitted,
 571  * reads the blocks back in, and checks the crc/md5 of the data.
 572  */
 573 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 574 {
 575         struct fio_file *f;
 576         struct io_u *io_u;
 577         int ret, min_events;
 578         unsigned int i;
 579
 580         dprint(FD_VERIFY, "starting loop\n");
 581
 582         /*
 583          * sync io first and invalidate cache, to make sure we really
 584          * read from disk.
 585          */
 586         for_each_file(td, f, i) {
 587                 if (!fio_file_open(f))
 588                         continue;
 589                 if (fio_io_sync(td, f))
 590                         break;
 591                 if (file_invalidate_cache(td, f))
 592                         break;
 593         }
 594
 595         check_update_rusage(td);
 596
 597         if (td->error)
 598                 return;
 599
 600         td_set_runstate(td, TD_VERIFYING);
 601
 602         io_u = NULL;
 603         while (!td->terminate) {
 604                 enum fio_ddir ddir;
 605                 int full;
 606
 607                 update_tv_cache(td);
 608                 check_update_rusage(td);
 609
 610                 if (runtime_exceeded(td, &td->tv_cache)) {
 611                         __update_tv_cache(td);
 612                         if (runtime_exceeded(td, &td->tv_cache)) {
 613                                 fio_mark_td_terminate(td);
 614                                 break;
 615                         }
 616                 }
 617
 618                 if (flow_threshold_exceeded(td))
 619                         continue;
 620
 621                 if (!td->o.experimental_verify) {
 622                         io_u = __get_io_u(td);
 623                         if (!io_u)
 624                                 break;
 625
 626                         if (get_next_verify(td, io_u)) {
 627                                 put_io_u(td, io_u);
 628                                 break;
 629                         }
 630
 631                         if (td_io_prep(td, io_u)) {
 632                                 put_io_u(td, io_u);
 633                                 break;
 634                         }
 635                 } else {
 636                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 637                                 break;
 638
 639                         while ((io_u = get_io_u(td)) != NULL) {
 640                                 if (IS_ERR(io_u)) {
 641                                         io_u = NULL;
 642                                         ret = FIO_Q_BUSY;
 643                                         goto reap;
 644                                 }
 645
 646                                 /*
 647                                  * We are only interested in the places where
 648                                  * we wrote or trimmed IOs. Turn those into
 649                                  * reads for verification purposes.
 650                                  */
 651                                 if (io_u->ddir == DDIR_READ) {
 652                                         /*
 653                                          * Pretend we issued it for rwmix
 654                                          * accounting
 655                                          */
 656                                         td->io_issues[DDIR_READ]++;
 657                                         put_io_u(td, io_u);
 658                                         continue;
 659                                 } else if (io_u->ddir == DDIR_TRIM) {
 660                                         io_u->ddir = DDIR_READ;
 661                                         io_u_set(io_u, IO_U_F_TRIMMED);
 662                                         break;
 663                                 } else if (io_u->ddir == DDIR_WRITE) {
 664                                         io_u->ddir = DDIR_READ;
 665                                         break;
 666                                 } else {
 667                                         put_io_u(td, io_u);
 668                                         continue;
 669                                 }
 670                         }
 671
 672                         if (!io_u)
 673                                 break;
 674                 }
 675
 676                 if (verify_state_should_stop(td, io_u)) {
 677                         put_io_u(td, io_u);
 678                         break;
 679                 }
 680
 681                 if (td->o.verify_async)
 682                         io_u->end_io = verify_io_u_async;
 683                 else
 684                         io_u->end_io = verify_io_u;
 685
 686                 ddir = io_u->ddir;
 687                 if (!td->o.disable_slat)
 688                         fio_gettime(&io_u->start_time, NULL);
 689
 690                 ret = td_io_queue(td, io_u);
 691
 692                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 693                         break;
 694
 695                 /*
 696                  * if we can queue more, do so. but check if there are
 697                  * completed io_u's first. Note that we can get BUSY even
 698                  * without IO queued, if the system is resource starved.
 699                  */
 700 reap:
 701                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 702                 if (full || io_in_polling(td))
 703                         ret = wait_for_completions(td, NULL);
 704
 705                 if (ret < 0)
 706                         break;
 707         }
 708
 709         check_update_rusage(td);
 710
 711         if (!td->error) {
 712                 min_events = td->cur_depth;
 713
 714                 if (min_events)
 715                         ret = io_u_queued_complete(td, min_events);
 716         } else
 717                 cleanup_pending_aio(td);
 718
 719         td_set_runstate(td, TD_RUNNING);
 720
 721         dprint(FD_VERIFY, "exiting loop\n");
 722 }
 723
 724 static bool exceeds_number_ios(struct thread_data *td)
 725 {
 726         unsigned long long number_ios;
 727
 728         if (!td->o.number_ios)
 729                 return false;
 730
 731         number_ios = ddir_rw_sum(td->io_blocks);
 732         number_ios += td->io_u_queued + td->io_u_in_flight;
 733
 734         return number_ios >= (td->o.number_ios * td->loops);
 735 }
 736
 737 static bool io_issue_bytes_exceeded(struct thread_data *td)
 738 {
 739         unsigned long long bytes, limit;
 740
 741         if (td_rw(td))
 742                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 743         else if (td_write(td))
 744                 bytes = td->io_issue_bytes[DDIR_WRITE];
 745         else if (td_read(td))
 746                 bytes = td->io_issue_bytes[DDIR_READ];
 747         else
 748                 bytes = td->io_issue_bytes[DDIR_TRIM];
 749
 750         if (td->o.io_limit)
 751                 limit = td->o.io_limit;
 752         else
 753                 limit = td->o.size;
 754
 755         limit *= td->loops;
 756         return bytes >= limit || exceeds_number_ios(td);
 757 }
 758
 759 static bool io_complete_bytes_exceeded(struct thread_data *td)
 760 {
 761         unsigned long long bytes, limit;
 762
 763         if (td_rw(td))
 764                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 765         else if (td_write(td))
 766                 bytes = td->this_io_bytes[DDIR_WRITE];
 767         else if (td_read(td))
 768                 bytes = td->this_io_bytes[DDIR_READ];
 769         else
 770                 bytes = td->this_io_bytes[DDIR_TRIM];
 771
 772         if (td->o.io_limit)
 773                 limit = td->o.io_limit;
 774         else
 775                 limit = td->o.size;
 776
 777         limit *= td->loops;
 778         return bytes >= limit || exceeds_number_ios(td);
 779 }
 780
 781 /*
 782  * used to calculate the next io time for rate control
 783  *
 784  */
 785 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 786 {
 787         uint64_t secs, remainder, bps, bytes, iops;
 788
 789         assert(!(td->flags & TD_F_CHILD));
 790         bytes = td->rate_io_issue_bytes[ddir];
 791         bps = td->rate_bps[ddir];
 792
 793         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 794                 uint64_t val;
 795                 iops = bps / td->o.bs[ddir];
 796                 val = (int64_t) (1000000 / iops) *
 797                                 -logf(__rand_0_1(&td->poisson_state));
 798                 if (val) {
 799                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 800                                         (unsigned long long) 1000000 / val);
 801                 }
 802                 td->last_usec += val;
 803                 return td->last_usec;
 804         } else if (bps) {
 805                 secs = bytes / bps;
 806                 remainder = bytes % bps;
 807                 return remainder * 1000000 / bps + secs * 1000000;
 808         }
 809
 810         return 0;
 811 }
 812
 813 /*
 814  * Main IO worker function. It retrieves io_u's to process and queues
 815  * and reaps them, checking for rate and errors along the way.
 816  *
 817  * Returns number of bytes written and trimmed.
 818  */
 819 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 820 {
 821         unsigned int i;
 822         int ret = 0;
 823         uint64_t total_bytes, bytes_issued = 0;
 824
 825         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 826                 bytes_done[i] = td->bytes_done[i];
 827
 828         if (in_ramp_time(td))
 829                 td_set_runstate(td, TD_RAMP);
 830         else
 831                 td_set_runstate(td, TD_RUNNING);
 832
 833         lat_target_init(td);
 834
 835         total_bytes = td->o.size;
 836         /*
 837         * Allow random overwrite workloads to write up to io_limit
 838         * before starting verification phase as 'size' doesn't apply.
 839         */
 840         if (td_write(td) && td_random(td) && td->o.norandommap)
 841                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 842         /*
 843          * If verify_backlog is enabled, we'll run the verify in this
 844          * handler as well. For that case, we may need up to twice the
 845          * amount of bytes.
 846          */
 847         if (td->o.verify != VERIFY_NONE &&
 848            (td_write(td) && td->o.verify_backlog))
 849                 total_bytes += td->o.size;
 850
 851         /* In trimwrite mode, each byte is trimmed and then written, so
 852          * allow total_bytes to be twice as big */
 853         if (td_trimwrite(td))
 854                 total_bytes += td->total_io_size;
 855
 856         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 857                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 858                 td->o.time_based) {
 859                 struct timeval comp_time;
 860                 struct io_u *io_u;
 861                 int full;
 862                 enum fio_ddir ddir;
 863
 864                 check_update_rusage(td);
 865
 866                 if (td->terminate || td->done)
 867                         break;
 868
 869                 update_tv_cache(td);
 870
 871                 if (runtime_exceeded(td, &td->tv_cache)) {
 872                         __update_tv_cache(td);
 873                         if (runtime_exceeded(td, &td->tv_cache)) {
 874                                 fio_mark_td_terminate(td);
 875                                 break;
 876                         }
 877                 }
 878
 879                 if (flow_threshold_exceeded(td))
 880                         continue;
 881
 882                 if (!td->o.time_based && bytes_issued >= total_bytes)
 883                         break;
 884
 885                 io_u = get_io_u(td);
 886                 if (IS_ERR_OR_NULL(io_u)) {
 887                         int err = PTR_ERR(io_u);
 888
 889                         io_u = NULL;
 890                         if (err == -EBUSY) {
 891                                 ret = FIO_Q_BUSY;
 892                                 goto reap;
 893                         }
 894                         if (td->o.latency_target)
 895                                 goto reap;
 896                         break;
 897                 }
 898
 899                 ddir = io_u->ddir;
 900
 901                 /*
 902                  * Add verification end_io handler if:
 903                  *      - Asked to verify (!td_rw(td))
 904                  *      - Or the io_u is from our verify list (mixed write/ver)
 905                  */
 906                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 907                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 908
 909                         if (!td->o.verify_pattern_bytes) {
 910                                 io_u->rand_seed = __rand(&td->verify_state);
 911                                 if (sizeof(int) != sizeof(long *))
 912                                         io_u->rand_seed *= __rand(&td->verify_state);
 913                         }
 914
 915                         if (verify_state_should_stop(td, io_u)) {
 916                                 put_io_u(td, io_u);
 917                                 break;
 918                         }
 919
 920                         if (td->o.verify_async)
 921                                 io_u->end_io = verify_io_u_async;
 922                         else
 923                                 io_u->end_io = verify_io_u;
 924                         td_set_runstate(td, TD_VERIFYING);
 925                 } else if (in_ramp_time(td))
 926                         td_set_runstate(td, TD_RAMP);
 927                 else
 928                         td_set_runstate(td, TD_RUNNING);
 929
 930                 /*
 931                  * Always log IO before it's issued, so we know the specific
 932                  * order of it. The logged unit will track when the IO has
 933                  * completed.
 934                  */
 935                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 936                     td->o.do_verify &&
 937                     td->o.verify != VERIFY_NONE &&
 938                     !td->o.experimental_verify)
 939                         log_io_piece(td, io_u);
 940
 941                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 942                         const unsigned long blen = io_u->xfer_buflen;
 943                         const enum fio_ddir ddir = acct_ddir(io_u);
 944
 945                         if (td->error)
 946                                 break;
 947
 948                         workqueue_enqueue(&td->io_wq, &io_u->work);
 949                         ret = FIO_Q_QUEUED;
 950
 951                         if (ddir_rw(ddir)) {
 952                                 td->io_issues[ddir]++;
 953                                 td->io_issue_bytes[ddir] += blen;
 954                                 td->rate_io_issue_bytes[ddir] += blen;
 955                         }
 956
 957                         if (should_check_rate(td))
 958                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 959
 960                 } else {
 961                         ret = td_io_queue(td, io_u);
 962
 963                         if (should_check_rate(td))
 964                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 965
 966                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 967                                 break;
 968
 969                         /*
 970                          * See if we need to complete some commands. Note that
 971                          * we can get BUSY even without IO queued, if the
 972                          * system is resource starved.
 973                          */
 974 reap:
 975                         full = queue_full(td) ||
 976                                 (ret == FIO_Q_BUSY && td->cur_depth);
 977                         if (full || io_in_polling(td))
 978                                 ret = wait_for_completions(td, &comp_time);
 979                 }
 980                 if (ret < 0)
 981                         break;
 982                 if (!ddir_rw_sum(td->bytes_done) &&
 983                     !(td->io_ops->flags & FIO_NOIO))
 984                         continue;
 985
 986                 if (!in_ramp_time(td) && should_check_rate(td)) {
 987                         if (check_min_rate(td, &comp_time)) {
 988                                 if (exitall_on_terminate || td->o.exitall_error)
 989                                         fio_terminate_threads(td->groupid);
 990                                 td_verror(td, EIO, "check_min_rate");
 991                                 break;
 992                         }
 993                 }
 994                 if (!in_ramp_time(td) && td->o.latency_target)
 995                         lat_target_check(td);
 996
 997                 if (td->o.thinktime) {
 998                         unsigned long long b;
 999
1000                         b = ddir_rw_sum(td->io_blocks);
1001                         if (!(b % td->o.thinktime_blocks)) {
1002                                 int left;
1003
1004                                 io_u_quiesce(td);
1005
1006                                 if (td->o.thinktime_spin)
1007                                         usec_spin(td->o.thinktime_spin);
1008
1009                                 left = td->o.thinktime - td->o.thinktime_spin;
1010                                 if (left)
1011                                         usec_sleep(td, left);
1012                         }
1013                 }
1014         }
1015
1016         check_update_rusage(td);
1017
1018         if (td->trim_entries)
1019                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1020
1021         if (td->o.fill_device && td->error == ENOSPC) {
1022                 td->error = 0;
1023                 fio_mark_td_terminate(td);
1024         }
1025         if (!td->error) {
1026                 struct fio_file *f;
1027
1028                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1029                         workqueue_flush(&td->io_wq);
1030                         i = 0;
1031                 } else
1032                         i = td->cur_depth;
1033
1034                 if (i) {
1035                         ret = io_u_queued_complete(td, i);
1036                         if (td->o.fill_device && td->error == ENOSPC)
1037                                 td->error = 0;
1038                 }
1039
1040                 if (should_fsync(td) && td->o.end_fsync) {
1041                         td_set_runstate(td, TD_FSYNCING);
1042
1043                         for_each_file(td, f, i) {
1044                                 if (!fio_file_fsync(td, f))
1045                                         continue;
1046
1047                                 log_err("fio: end_fsync failed for file %s\n",
1048                                                                 f->file_name);
1049                         }
1050                 }
1051         } else
1052                 cleanup_pending_aio(td);
1053
1054         /*
1055          * stop job if we failed doing any IO
1056          */
1057         if (!ddir_rw_sum(td->this_io_bytes))
1058                 td->done = 1;
1059
1060         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1061                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1062 }
1063
1064 static void cleanup_io_u(struct thread_data *td)
1065 {
1066         struct io_u *io_u;
1067
1068         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1069
1070                 if (td->io_ops->io_u_free)
1071                         td->io_ops->io_u_free(td, io_u);
1072
1073                 fio_memfree(io_u, sizeof(*io_u));
1074         }
1075
1076         free_io_mem(td);
1077
1078         io_u_rexit(&td->io_u_requeues);
1079         io_u_qexit(&td->io_u_freelist);
1080         io_u_qexit(&td->io_u_all);
1081
1082         if (td->last_write_comp)
1083                 sfree(td->last_write_comp);
1084 }
1085
1086 static int init_io_u(struct thread_data *td)
1087 {
1088         struct io_u *io_u;
1089         unsigned int max_bs, min_write;
1090         int cl_align, i, max_units;
1091         int data_xfer = 1, err;
1092         char *p;
1093
1094         max_units = td->o.iodepth;
1095         max_bs = td_max_bs(td);
1096         min_write = td->o.min_bs[DDIR_WRITE];
1097         td->orig_buffer_size = (unsigned long long) max_bs
1098                                         * (unsigned long long) max_units;
1099
1100         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1101                 data_xfer = 0;
1102
1103         err = 0;
1104         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1105         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1106         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1107
1108         if (err) {
1109                 log_err("fio: failed setting up IO queues\n");
1110                 return 1;
1111         }
1112
1113         /*
1114          * if we may later need to do address alignment, then add any
1115          * possible adjustment here so that we don't cause a buffer
1116          * overflow later. this adjustment may be too much if we get
1117          * lucky and the allocator gives us an aligned address.
1118          */
1119         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1120             (td->io_ops->flags & FIO_RAWIO))
1121                 td->orig_buffer_size += page_mask + td->o.mem_align;
1122
1123         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1124                 unsigned long bs;
1125
1126                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1127                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1128         }
1129
1130         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1131                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1132                 return 1;
1133         }
1134
1135         if (data_xfer && allocate_io_mem(td))
1136                 return 1;
1137
1138         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1139             (td->io_ops->flags & FIO_RAWIO))
1140                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1141         else
1142                 p = td->orig_buffer;
1143
1144         cl_align = os_cache_line_size();
1145
1146         for (i = 0; i < max_units; i++) {
1147                 void *ptr;
1148
1149                 if (td->terminate)
1150                         return 1;
1151
1152                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1153                 if (!ptr) {
1154                         log_err("fio: unable to allocate aligned memory\n");
1155                         break;
1156                 }
1157
1158                 io_u = ptr;
1159                 memset(io_u, 0, sizeof(*io_u));
1160                 INIT_FLIST_HEAD(&io_u->verify_list);
1161                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1162
1163                 if (data_xfer) {
1164                         io_u->buf = p;
1165                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1166
1167                         if (td_write(td))
1168                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1169                         if (td_write(td) && td->o.verify_pattern_bytes) {
1170                                 /*
1171                                  * Fill the buffer with the pattern if we are
1172                                  * going to be doing writes.
1173                                  */
1174                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1175                         }
1176                 }
1177
1178                 io_u->index = i;
1179                 io_u->flags = IO_U_F_FREE;
1180                 io_u_qpush(&td->io_u_freelist, io_u);
1181
1182                 /*
1183                  * io_u never leaves this stack, used for iteration of all
1184                  * io_u buffers.
1185                  */
1186                 io_u_qpush(&td->io_u_all, io_u);
1187
1188                 if (td->io_ops->io_u_init) {
1189                         int ret = td->io_ops->io_u_init(td, io_u);
1190
1191                         if (ret) {
1192                                 log_err("fio: failed to init engine data: %d\n", ret);
1193                                 return 1;
1194                         }
1195                 }
1196
1197                 p += max_bs;
1198         }
1199
1200         if (td->o.verify != VERIFY_NONE) {
1201                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1202                 if (!td->last_write_comp) {
1203                         log_err("fio: failed to alloc write comp data\n");
1204                         return 1;
1205                 }
1206         }
1207
1208         return 0;
1209 }
1210
1211 static int switch_ioscheduler(struct thread_data *td)
1212 {
1213         char tmp[256], tmp2[128];
1214         FILE *f;
1215         int ret;
1216
1217         if (td->io_ops->flags & FIO_DISKLESSIO)
1218                 return 0;
1219
1220         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1221
1222         f = fopen(tmp, "r+");
1223         if (!f) {
1224                 if (errno == ENOENT) {
1225                         log_err("fio: os or kernel doesn't support IO scheduler"
1226                                 " switching\n");
1227                         return 0;
1228                 }
1229                 td_verror(td, errno, "fopen iosched");
1230                 return 1;
1231         }
1232
1233         /*
1234          * Set io scheduler.
1235          */
1236         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1237         if (ferror(f) || ret != 1) {
1238                 td_verror(td, errno, "fwrite");
1239                 fclose(f);
1240                 return 1;
1241         }
1242
1243         rewind(f);
1244
1245         /*
1246          * Read back and check that the selected scheduler is now the default.
1247          */
1248         memset(tmp, 0, sizeof(tmp));
1249         ret = fread(tmp, sizeof(tmp), 1, f);
1250         if (ferror(f) || ret < 0) {
1251                 td_verror(td, errno, "fread");
1252                 fclose(f);
1253                 return 1;
1254         }
1255         /*
1256          * either a list of io schedulers or "none\n" is expected.
1257          */
1258         tmp[strlen(tmp) - 1] = '\0';
1259
1260
1261         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1262         if (!strstr(tmp, tmp2)) {
1263                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1264                 td_verror(td, EINVAL, "iosched_switch");
1265                 fclose(f);
1266                 return 1;
1267         }
1268
1269         fclose(f);
1270         return 0;
1271 }
1272
1273 static bool keep_running(struct thread_data *td)
1274 {
1275         unsigned long long limit;
1276
1277         if (td->done)
1278                 return false;
1279         if (td->o.time_based)
1280                 return true;
1281         if (td->o.loops) {
1282                 td->o.loops--;
1283                 return true;
1284         }
1285         if (exceeds_number_ios(td))
1286                 return false;
1287
1288         if (td->o.io_limit)
1289                 limit = td->o.io_limit;
1290         else
1291                 limit = td->o.size;
1292
1293         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1294                 uint64_t diff;
1295
1296                 /*
1297                  * If the difference is less than the minimum IO size, we
1298                  * are done.
1299                  */
1300                 diff = limit - ddir_rw_sum(td->io_bytes);
1301                 if (diff < td_max_bs(td))
1302                         return false;
1303
1304                 if (fio_files_done(td) && !td->o.io_limit)
1305                         return false;
1306
1307                 return true;
1308         }
1309
1310         return false;
1311 }
1312
1313 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1314 {
1315         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1316         int ret;
1317         char *str;
1318
1319         str = malloc(newlen);
1320         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1321
1322         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1323         ret = system(str);
1324         if (ret == -1)
1325                 log_err("fio: exec of cmd <%s> failed\n", str);
1326
1327         free(str);
1328         return ret;
1329 }
1330
1331 /*
1332  * Dry run to compute correct state of numberio for verification.
1333  */
1334 static uint64_t do_dry_run(struct thread_data *td)
1335 {
1336         td_set_runstate(td, TD_RUNNING);
1337
1338         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1339                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1340                 struct io_u *io_u;
1341                 int ret;
1342
1343                 if (td->terminate || td->done)
1344                         break;
1345
1346                 io_u = get_io_u(td);
1347                 if (!io_u)
1348                         break;
1349
1350                 io_u_set(io_u, IO_U_F_FLIGHT);
1351                 io_u->error = 0;
1352                 io_u->resid = 0;
1353                 if (ddir_rw(acct_ddir(io_u)))
1354                         td->io_issues[acct_ddir(io_u)]++;
1355                 if (ddir_rw(io_u->ddir)) {
1356                         io_u_mark_depth(td, 1);
1357                         td->ts.total_io_u[io_u->ddir]++;
1358                 }
1359
1360                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1361                     td->o.do_verify &&
1362                     td->o.verify != VERIFY_NONE &&
1363                     !td->o.experimental_verify)
1364                         log_io_piece(td, io_u);
1365
1366                 ret = io_u_sync_complete(td, io_u);
1367                 (void) ret;
1368         }
1369
1370         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1371 }
1372
1373 struct fork_data {
1374         struct thread_data *td;
1375         struct sk_out *sk_out;
1376 };
1377
1378 /*
1379  * Entry point for the thread based jobs. The process based jobs end up
1380  * here as well, after a little setup.
1381  */
1382 static void *thread_main(void *data)
1383 {
1384         struct fork_data *fd = data;
1385         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1386         struct thread_data *td = fd->td;
1387         struct thread_options *o = &td->o;
1388         struct sk_out *sk_out = fd->sk_out;
1389         pthread_condattr_t attr;
1390         int clear_state;
1391         int ret;
1392
1393         sk_out_assign(sk_out);
1394         free(fd);
1395
1396         if (!o->use_thread) {
1397                 setsid();
1398                 td->pid = getpid();
1399         } else
1400                 td->pid = gettid();
1401
1402         fio_local_clock_init(o->use_thread);
1403
1404         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1405
1406         if (is_backend)
1407                 fio_server_send_start(td);
1408
1409         INIT_FLIST_HEAD(&td->io_log_list);
1410         INIT_FLIST_HEAD(&td->io_hist_list);
1411         INIT_FLIST_HEAD(&td->verify_list);
1412         INIT_FLIST_HEAD(&td->trim_list);
1413         INIT_FLIST_HEAD(&td->next_rand_list);
1414         pthread_mutex_init(&td->io_u_lock, NULL);
1415         td->io_hist_tree = RB_ROOT;
1416
1417         pthread_condattr_init(&attr);
1418         pthread_cond_init(&td->verify_cond, &attr);
1419         pthread_cond_init(&td->free_cond, &attr);
1420
1421         td_set_runstate(td, TD_INITIALIZED);
1422         dprint(FD_MUTEX, "up startup_mutex\n");
1423         fio_mutex_up(startup_mutex);
1424         dprint(FD_MUTEX, "wait on td->mutex\n");
1425         fio_mutex_down(td->mutex);
1426         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1427
1428         /*
1429          * A new gid requires privilege, so we need to do this before setting
1430          * the uid.
1431          */
1432         if (o->gid != -1U && setgid(o->gid)) {
1433                 td_verror(td, errno, "setgid");
1434                 goto err;
1435         }
1436         if (o->uid != -1U && setuid(o->uid)) {
1437                 td_verror(td, errno, "setuid");
1438                 goto err;
1439         }
1440
1441         /*
1442          * If we have a gettimeofday() thread, make sure we exclude that
1443          * thread from this job
1444          */
1445         if (o->gtod_cpu)
1446                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1447
1448         /*
1449          * Set affinity first, in case it has an impact on the memory
1450          * allocations.
1451          */
1452         if (fio_option_is_set(o, cpumask)) {
1453                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1454                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1455                         if (!ret) {
1456                                 log_err("fio: no CPUs set\n");
1457                                 log_err("fio: Try increasing number of available CPUs\n");
1458                                 td_verror(td, EINVAL, "cpus_split");
1459                                 goto err;
1460                         }
1461                 }
1462                 ret = fio_setaffinity(td->pid, o->cpumask);
1463                 if (ret == -1) {
1464                         td_verror(td, errno, "cpu_set_affinity");
1465                         goto err;
1466                 }
1467         }
1468
1469 #ifdef CONFIG_LIBNUMA
1470         /* numa node setup */
1471         if (fio_option_is_set(o, numa_cpunodes) ||
1472             fio_option_is_set(o, numa_memnodes)) {
1473                 struct bitmask *mask;
1474
1475                 if (numa_available() < 0) {
1476                         td_verror(td, errno, "Does not support NUMA API\n");
1477                         goto err;
1478                 }
1479
1480                 if (fio_option_is_set(o, numa_cpunodes)) {
1481                         mask = numa_parse_nodestring(o->numa_cpunodes);
1482                         ret = numa_run_on_node_mask(mask);
1483                         numa_free_nodemask(mask);
1484                         if (ret == -1) {
1485                                 td_verror(td, errno, \
1486                                         "numa_run_on_node_mask failed\n");
1487                                 goto err;
1488                         }
1489                 }
1490
1491                 if (fio_option_is_set(o, numa_memnodes)) {
1492                         mask = NULL;
1493                         if (o->numa_memnodes)
1494                                 mask = numa_parse_nodestring(o->numa_memnodes);
1495
1496                         switch (o->numa_mem_mode) {
1497                         case MPOL_INTERLEAVE:
1498                                 numa_set_interleave_mask(mask);
1499                                 break;
1500                         case MPOL_BIND:
1501                                 numa_set_membind(mask);
1502                                 break;
1503                         case MPOL_LOCAL:
1504                                 numa_set_localalloc();
1505                                 break;
1506                         case MPOL_PREFERRED:
1507                                 numa_set_preferred(o->numa_mem_prefer_node);
1508                                 break;
1509                         case MPOL_DEFAULT:
1510                         default:
1511                                 break;
1512                         }
1513
1514                         if (mask)
1515                                 numa_free_nodemask(mask);
1516
1517                 }
1518         }
1519 #endif
1520
1521         if (fio_pin_memory(td))
1522                 goto err;
1523
1524         /*
1525          * May alter parameters that init_io_u() will use, so we need to
1526          * do this first.
1527          */
1528         if (init_iolog(td))
1529                 goto err;
1530
1531         if (init_io_u(td))
1532                 goto err;
1533
1534         if (o->verify_async && verify_async_init(td))
1535                 goto err;
1536
1537         if (fio_option_is_set(o, ioprio) ||
1538             fio_option_is_set(o, ioprio_class)) {
1539                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1540                 if (ret == -1) {
1541                         td_verror(td, errno, "ioprio_set");
1542                         goto err;
1543                 }
1544         }
1545
1546         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1547                 goto err;
1548
1549         errno = 0;
1550         if (nice(o->nice) == -1 && errno != 0) {
1551                 td_verror(td, errno, "nice");
1552                 goto err;
1553         }
1554
1555         if (o->ioscheduler && switch_ioscheduler(td))
1556                 goto err;
1557
1558         if (!o->create_serialize && setup_files(td))
1559                 goto err;
1560
1561         if (td_io_init(td))
1562                 goto err;
1563
1564         if (init_random_map(td))
1565                 goto err;
1566
1567         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1568                 goto err;
1569
1570         if (o->pre_read) {
1571                 if (pre_read_files(td) < 0)
1572                         goto err;
1573         }
1574
1575         if (iolog_compress_init(td, sk_out))
1576                 goto err;
1577
1578         fio_verify_init(td);
1579
1580         if (rate_submit_init(td, sk_out))
1581                 goto err;
1582
1583         fio_gettime(&td->epoch, NULL);
1584         fio_getrusage(&td->ru_start);
1585         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1586         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1587
1588         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1589                         o->ratemin[DDIR_TRIM]) {
1590                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1591                                         sizeof(td->bw_sample_time));
1592                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1593                                         sizeof(td->bw_sample_time));
1594                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1595                                         sizeof(td->bw_sample_time));
1596         }
1597
1598         clear_state = 0;
1599         while (keep_running(td)) {
1600                 uint64_t verify_bytes;
1601
1602                 fio_gettime(&td->start, NULL);
1603                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1604
1605                 if (clear_state)
1606                         clear_io_state(td, 0);
1607
1608                 prune_io_piece_log(td);
1609
1610                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1611                         verify_bytes = do_dry_run(td);
1612                 else {
1613                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1614
1615                         do_io(td, bytes_done);
1616
1617                         if (!ddir_rw_sum(bytes_done)) {
1618                                 fio_mark_td_terminate(td);
1619                                 verify_bytes = 0;
1620                         } else {
1621                                 verify_bytes = bytes_done[DDIR_WRITE] +
1622                                                 bytes_done[DDIR_TRIM];
1623                         }
1624                 }
1625
1626                 clear_state = 1;
1627
1628                 /*
1629                  * Make sure we've successfully updated the rusage stats
1630                  * before waiting on the stat mutex. Otherwise we could have
1631                  * the stat thread holding stat mutex and waiting for
1632                  * the rusage_sem, which would never get upped because
1633                  * this thread is waiting for the stat mutex.
1634                  */
1635                 check_update_rusage(td);
1636
1637                 fio_mutex_down(stat_mutex);
1638                 if (td_read(td) && td->io_bytes[DDIR_READ])
1639                         update_runtime(td, elapsed_us, DDIR_READ);
1640                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1641                         update_runtime(td, elapsed_us, DDIR_WRITE);
1642                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1643                         update_runtime(td, elapsed_us, DDIR_TRIM);
1644                 fio_gettime(&td->start, NULL);
1645                 fio_mutex_up(stat_mutex);
1646
1647                 if (td->error || td->terminate)
1648                         break;
1649
1650                 if (!o->do_verify ||
1651                     o->verify == VERIFY_NONE ||
1652                     (td->io_ops->flags & FIO_UNIDIR))
1653                         continue;
1654
1655                 clear_io_state(td, 0);
1656
1657                 fio_gettime(&td->start, NULL);
1658
1659                 do_verify(td, verify_bytes);
1660
1661                 /*
1662                  * See comment further up for why this is done here.
1663                  */
1664                 check_update_rusage(td);
1665
1666                 fio_mutex_down(stat_mutex);
1667                 update_runtime(td, elapsed_us, DDIR_READ);
1668                 fio_gettime(&td->start, NULL);
1669                 fio_mutex_up(stat_mutex);
1670
1671                 if (td->error || td->terminate)
1672                         break;
1673         }
1674
1675         update_rusage_stat(td);
1676         td->ts.total_run_time = mtime_since_now(&td->epoch);
1677         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1678         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1679         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1680
1681         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1682             (td->o.verify != VERIFY_NONE && td_write(td)))
1683                 verify_save_state(td->thread_number);
1684
1685         fio_unpin_memory(td);
1686
1687         fio_writeout_logs(td);
1688
1689         iolog_compress_exit(td);
1690         rate_submit_exit(td);
1691
1692         if (o->exec_postrun)
1693                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1694
1695         if (exitall_on_terminate || (o->exitall_error && td->error))
1696                 fio_terminate_threads(td->groupid);
1697
1698 err:
1699         if (td->error)
1700                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1701                                                         td->verror);
1702
1703         if (o->verify_async)
1704                 verify_async_exit(td);
1705
1706         close_and_free_files(td);
1707         cleanup_io_u(td);
1708         close_ioengine(td);
1709         cgroup_shutdown(td, &cgroup_mnt);
1710         verify_free_state(td);
1711
1712         if (fio_option_is_set(o, cpumask)) {
1713                 ret = fio_cpuset_exit(&o->cpumask);
1714                 if (ret)
1715                         td_verror(td, ret, "fio_cpuset_exit");
1716         }
1717
1718         /*
1719          * do this very late, it will log file closing as well
1720          */
1721         if (o->write_iolog_file)
1722                 write_iolog_close(td);
1723
1724         fio_mutex_remove(td->mutex);
1725         td->mutex = NULL;
1726
1727         td_set_runstate(td, TD_EXITED);
1728
1729         /*
1730          * Do this last after setting our runstate to exited, so we
1731          * know that the stat thread is signaled.
1732          */
1733         check_update_rusage(td);
1734
1735         sk_out_drop();
1736         return (void *) (uintptr_t) td->error;
1737 }
1738
1739
1740 /*
1741  * We cannot pass the td data into a forked process, so attach the td and
1742  * pass it to the thread worker.
1743  */
1744 static int fork_main(struct sk_out *sk_out, int shmid, int offset)
1745 {
1746         struct fork_data *fd;
1747         void *data, *ret;
1748
1749 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1750         data = shmat(shmid, NULL, 0);
1751         if (data == (void *) -1) {
1752                 int __err = errno;
1753
1754                 perror("shmat");
1755                 return __err;
1756         }
1757 #else
1758         /*
1759          * HP-UX inherits shm mappings?
1760          */
1761         data = threads;
1762 #endif
1763
1764         fd = calloc(1, sizeof(*fd));
1765         fd->td = data + offset * sizeof(struct thread_data);
1766         fd->sk_out = sk_out;
1767         ret = thread_main(fd);
1768         shmdt(data);
1769         return (int) (uintptr_t) ret;
1770 }
1771
1772 static void dump_td_info(struct thread_data *td)
1773 {
1774         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1775                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1776                         (unsigned long) time_since_now(&td->terminate_time));
1777 }
1778
1779 /*
1780  * Run over the job map and reap the threads that have exited, if any.
1781  */
1782 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1783                          unsigned int *m_rate)
1784 {
1785         struct thread_data *td;
1786         unsigned int cputhreads, realthreads, pending;
1787         int i, status, ret;
1788
1789         /*
1790          * reap exited threads (TD_EXITED -> TD_REAPED)
1791          */
1792         realthreads = pending = cputhreads = 0;
1793         for_each_td(td, i) {
1794                 int flags = 0;
1795
1796                 /*
1797                  * ->io_ops is NULL for a thread that has closed its
1798                  * io engine
1799                  */
1800                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1801                         cputhreads++;
1802                 else
1803                         realthreads++;
1804
1805                 if (!td->pid) {
1806                         pending++;
1807                         continue;
1808                 }
1809                 if (td->runstate == TD_REAPED)
1810                         continue;
1811                 if (td->o.use_thread) {
1812                         if (td->runstate == TD_EXITED) {
1813                                 td_set_runstate(td, TD_REAPED);
1814                                 goto reaped;
1815                         }
1816                         continue;
1817                 }
1818
1819                 flags = WNOHANG;
1820                 if (td->runstate == TD_EXITED)
1821                         flags = 0;
1822
1823                 /*
1824                  * check if someone quit or got killed in an unusual way
1825                  */
1826                 ret = waitpid(td->pid, &status, flags);
1827                 if (ret < 0) {
1828                         if (errno == ECHILD) {
1829                                 log_err("fio: pid=%d disappeared %d\n",
1830                                                 (int) td->pid, td->runstate);
1831                                 td->sig = ECHILD;
1832                                 td_set_runstate(td, TD_REAPED);
1833                                 goto reaped;
1834                         }
1835                         perror("waitpid");
1836                 } else if (ret == td->pid) {
1837                         if (WIFSIGNALED(status)) {
1838                                 int sig = WTERMSIG(status);
1839
1840                                 if (sig != SIGTERM && sig != SIGUSR2)
1841                                         log_err("fio: pid=%d, got signal=%d\n",
1842                                                         (int) td->pid, sig);
1843                                 td->sig = sig;
1844                                 td_set_runstate(td, TD_REAPED);
1845                                 goto reaped;
1846                         }
1847                         if (WIFEXITED(status)) {
1848                                 if (WEXITSTATUS(status) && !td->error)
1849                                         td->error = WEXITSTATUS(status);
1850
1851                                 td_set_runstate(td, TD_REAPED);
1852                                 goto reaped;
1853                         }
1854                 }
1855
1856                 /*
1857                  * If the job is stuck, do a forceful timeout of it and
1858                  * move on.
1859                  */
1860                 if (td->terminate &&
1861                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1862                         dump_td_info(td);
1863                         td_set_runstate(td, TD_REAPED);
1864                         goto reaped;
1865                 }
1866
1867                 /*
1868                  * thread is not dead, continue
1869                  */
1870                 pending++;
1871                 continue;
1872 reaped:
1873                 (*nr_running)--;
1874                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1875                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1876                 if (!td->pid)
1877                         pending--;
1878
1879                 if (td->error)
1880                         exit_value++;
1881
1882                 done_secs += mtime_since_now(&td->epoch) / 1000;
1883                 profile_td_exit(td);
1884         }
1885
1886         if (*nr_running == cputhreads && !pending && realthreads)
1887                 fio_terminate_threads(TERMINATE_ALL);
1888 }
1889
1890 static bool __check_trigger_file(void)
1891 {
1892         struct stat sb;
1893
1894         if (!trigger_file)
1895                 return false;
1896
1897         if (stat(trigger_file, &sb))
1898                 return false;
1899
1900         if (unlink(trigger_file) < 0)
1901                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1902                                                         strerror(errno));
1903
1904         return true;
1905 }
1906
1907 static bool trigger_timedout(void)
1908 {
1909         if (trigger_timeout)
1910                 return time_since_genesis() >= trigger_timeout;
1911
1912         return false;
1913 }
1914
1915 void exec_trigger(const char *cmd)
1916 {
1917         int ret;
1918
1919         if (!cmd)
1920                 return;
1921
1922         ret = system(cmd);
1923         if (ret == -1)
1924                 log_err("fio: failed executing %s trigger\n", cmd);
1925 }
1926
1927 void check_trigger_file(void)
1928 {
1929         if (__check_trigger_file() || trigger_timedout()) {
1930                 if (nr_clients)
1931                         fio_clients_send_trigger(trigger_remote_cmd);
1932                 else {
1933                         verify_save_state(IO_LIST_ALL);
1934                         fio_terminate_threads(TERMINATE_ALL);
1935                         exec_trigger(trigger_cmd);
1936                 }
1937         }
1938 }
1939
1940 static int fio_verify_load_state(struct thread_data *td)
1941 {
1942         int ret;
1943
1944         if (!td->o.verify_state)
1945                 return 0;
1946
1947         if (is_backend) {
1948                 void *data;
1949                 int ver;
1950
1951                 ret = fio_server_get_verify_state(td->o.name,
1952                                         td->thread_number - 1, &data, &ver);
1953                 if (!ret)
1954                         verify_convert_assign_state(td, data, ver);
1955         } else
1956                 ret = verify_load_state(td, "local");
1957
1958         return ret;
1959 }
1960
1961 static void do_usleep(unsigned int usecs)
1962 {
1963         check_for_running_stats();
1964         check_trigger_file();
1965         usleep(usecs);
1966 }
1967
1968 static bool check_mount_writes(struct thread_data *td)
1969 {
1970         struct fio_file *f;
1971         unsigned int i;
1972
1973         if (!td_write(td) || td->o.allow_mounted_write)
1974                 return false;
1975
1976         for_each_file(td, f, i) {
1977                 if (f->filetype != FIO_TYPE_BD)
1978                         continue;
1979                 if (device_is_mounted(f->file_name))
1980                         goto mounted;
1981         }
1982
1983         return false;
1984 mounted:
1985         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1986         return true;
1987 }
1988
1989 static bool waitee_running(struct thread_data *me)
1990 {
1991         const char *waitee = me->o.wait_for;
1992         const char *self = me->o.name;
1993         struct thread_data *td;
1994         int i;
1995
1996         if (!waitee)
1997                 return false;
1998
1999         for_each_td(td, i) {
2000                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2001                         continue;
2002
2003                 if (td->runstate < TD_EXITED) {
2004                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2005                                         self, td->o.name,
2006                                         runstate_to_name(td->runstate));
2007                         return true;
2008                 }
2009         }
2010
2011         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2012         return false;
2013 }
2014
2015 /*
2016  * Main function for kicking off and reaping jobs, as needed.
2017  */
2018 static void run_threads(struct sk_out *sk_out)
2019 {
2020         struct thread_data *td;
2021         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2022         uint64_t spent;
2023
2024         if (fio_gtod_offload && fio_start_gtod_thread())
2025                 return;
2026
2027         fio_idle_prof_init();
2028
2029         set_sig_handlers();
2030
2031         nr_thread = nr_process = 0;
2032         for_each_td(td, i) {
2033                 if (check_mount_writes(td))
2034                         return;
2035                 if (td->o.use_thread)
2036                         nr_thread++;
2037                 else
2038                         nr_process++;
2039         }
2040
2041         if (output_format & FIO_OUTPUT_NORMAL) {
2042                 log_info("Starting ");
2043                 if (nr_thread)
2044                         log_info("%d thread%s", nr_thread,
2045                                                 nr_thread > 1 ? "s" : "");
2046                 if (nr_process) {
2047                         if (nr_thread)
2048                                 log_info(" and ");
2049                         log_info("%d process%s", nr_process,
2050                                                 nr_process > 1 ? "es" : "");
2051                 }
2052                 log_info("\n");
2053                 log_info_flush();
2054         }
2055
2056         todo = thread_number;
2057         nr_running = 0;
2058         nr_started = 0;
2059         m_rate = t_rate = 0;
2060
2061         for_each_td(td, i) {
2062                 print_status_init(td->thread_number - 1);
2063
2064                 if (!td->o.create_serialize)
2065                         continue;
2066
2067                 if (fio_verify_load_state(td))
2068                         goto reap;
2069
2070                 /*
2071                  * do file setup here so it happens sequentially,
2072                  * we don't want X number of threads getting their
2073                  * client data interspersed on disk
2074                  */
2075                 if (setup_files(td)) {
2076 reap:
2077                         exit_value++;
2078                         if (td->error)
2079                                 log_err("fio: pid=%d, err=%d/%s\n",
2080                                         (int) td->pid, td->error, td->verror);
2081                         td_set_runstate(td, TD_REAPED);
2082                         todo--;
2083                 } else {
2084                         struct fio_file *f;
2085                         unsigned int j;
2086
2087                         /*
2088                          * for sharing to work, each job must always open
2089                          * its own files. so close them, if we opened them
2090                          * for creation
2091                          */
2092                         for_each_file(td, f, j) {
2093                                 if (fio_file_open(f))
2094                                         td_io_close_file(td, f);
2095                         }
2096                 }
2097         }
2098
2099         /* start idle threads before io threads start to run */
2100         fio_idle_prof_start();
2101
2102         set_genesis_time();
2103
2104         while (todo) {
2105                 struct thread_data *map[REAL_MAX_JOBS];
2106                 struct timeval this_start;
2107                 int this_jobs = 0, left;
2108
2109                 /*
2110                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2111                  */
2112                 for_each_td(td, i) {
2113                         if (td->runstate != TD_NOT_CREATED)
2114                                 continue;
2115
2116                         /*
2117                          * never got a chance to start, killed by other
2118                          * thread for some reason
2119                          */
2120                         if (td->terminate) {
2121                                 todo--;
2122                                 continue;
2123                         }
2124
2125                         if (td->o.start_delay) {
2126                                 spent = utime_since_genesis();
2127
2128                                 if (td->o.start_delay > spent)
2129                                         continue;
2130                         }
2131
2132                         if (td->o.stonewall && (nr_started || nr_running)) {
2133                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2134                                                         td->o.name);
2135                                 break;
2136                         }
2137
2138                         if (waitee_running(td)) {
2139                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2140                                                 td->o.name, td->o.wait_for);
2141                                 continue;
2142                         }
2143
2144                         init_disk_util(td);
2145
2146                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2147                         td->update_rusage = 0;
2148
2149                         /*
2150                          * Set state to created. Thread will transition
2151                          * to TD_INITIALIZED when it's done setting up.
2152                          */
2153                         td_set_runstate(td, TD_CREATED);
2154                         map[this_jobs++] = td;
2155                         nr_started++;
2156
2157                         if (td->o.use_thread) {
2158                                 struct fork_data *fd;
2159                                 int ret;
2160
2161                                 fd = calloc(1, sizeof(*fd));
2162                                 fd->td = td;
2163                                 fd->sk_out = sk_out;
2164
2165                                 dprint(FD_PROCESS, "will pthread_create\n");
2166                                 ret = pthread_create(&td->thread, NULL,
2167                                                         thread_main, fd);
2168                                 if (ret) {
2169                                         log_err("pthread_create: %s\n",
2170                                                         strerror(ret));
2171                                         free(fd);
2172                                         nr_started--;
2173                                         break;
2174                                 }
2175                                 ret = pthread_detach(td->thread);
2176                                 if (ret)
2177                                         log_err("pthread_detach: %s",
2178                                                         strerror(ret));
2179                         } else {
2180                                 pid_t pid;
2181                                 dprint(FD_PROCESS, "will fork\n");
2182                                 pid = fork();
2183                                 if (!pid) {
2184                                         int ret = fork_main(sk_out, shm_id, i);
2185
2186                                         _exit(ret);
2187                                 } else if (i == fio_debug_jobno)
2188                                         *fio_debug_jobp = pid;
2189                         }
2190                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2191                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2192                                 log_err("fio: job startup hung? exiting.\n");
2193                                 fio_terminate_threads(TERMINATE_ALL);
2194                                 fio_abort = 1;
2195                                 nr_started--;
2196                                 break;
2197                         }
2198                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2199                 }
2200
2201                 /*
2202                  * Wait for the started threads to transition to
2203                  * TD_INITIALIZED.
2204                  */
2205                 fio_gettime(&this_start, NULL);
2206                 left = this_jobs;
2207                 while (left && !fio_abort) {
2208                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2209                                 break;
2210
2211                         do_usleep(100000);
2212
2213                         for (i = 0; i < this_jobs; i++) {
2214                                 td = map[i];
2215                                 if (!td)
2216                                         continue;
2217                                 if (td->runstate == TD_INITIALIZED) {
2218                                         map[i] = NULL;
2219                                         left--;
2220                                 } else if (td->runstate >= TD_EXITED) {
2221                                         map[i] = NULL;
2222                                         left--;
2223                                         todo--;
2224                                         nr_running++; /* work-around... */
2225                                 }
2226                         }
2227                 }
2228
2229                 if (left) {
2230                         log_err("fio: %d job%s failed to start\n", left,
2231                                         left > 1 ? "s" : "");
2232                         for (i = 0; i < this_jobs; i++) {
2233                                 td = map[i];
2234                                 if (!td)
2235                                         continue;
2236                                 kill(td->pid, SIGTERM);
2237                         }
2238                         break;
2239                 }
2240
2241                 /*
2242                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2243                  */
2244                 for_each_td(td, i) {
2245                         if (td->runstate != TD_INITIALIZED)
2246                                 continue;
2247
2248                         if (in_ramp_time(td))
2249                                 td_set_runstate(td, TD_RAMP);
2250                         else
2251                                 td_set_runstate(td, TD_RUNNING);
2252                         nr_running++;
2253                         nr_started--;
2254                         m_rate += ddir_rw_sum(td->o.ratemin);
2255                         t_rate += ddir_rw_sum(td->o.rate);
2256                         todo--;
2257                         fio_mutex_up(td->mutex);
2258                 }
2259
2260                 reap_threads(&nr_running, &t_rate, &m_rate);
2261
2262                 if (todo)
2263                         do_usleep(100000);
2264         }
2265
2266         while (nr_running) {
2267                 reap_threads(&nr_running, &t_rate, &m_rate);
2268                 do_usleep(10000);
2269         }
2270
2271         fio_idle_prof_stop();
2272
2273         update_io_ticks();
2274 }
2275
2276 static void wait_for_helper_thread_exit(void)
2277 {
2278         void *ret;
2279
2280         helper_exit = 1;
2281         pthread_cond_signal(&helper_cond);
2282         pthread_join(helper_thread, &ret);
2283 }
2284
2285 static void free_disk_util(void)
2286 {
2287         disk_util_prune_entries();
2288
2289         pthread_cond_destroy(&helper_cond);
2290 }
2291
2292 static void *helper_thread_main(void *data)
2293 {
2294         struct sk_out *sk_out = data;
2295         int ret = 0;
2296
2297         sk_out_assign(sk_out);
2298
2299         fio_mutex_up(startup_mutex);
2300
2301         while (!ret) {
2302                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2303                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2304                 struct timespec ts;
2305                 struct timeval tv;
2306
2307                 gettimeofday(&tv, NULL);
2308                 ts.tv_sec = tv.tv_sec + sec;
2309                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2310
2311                 if (ts.tv_nsec >= 1000000000ULL) {
2312                         ts.tv_nsec -= 1000000000ULL;
2313                         ts.tv_sec++;
2314                 }
2315
2316                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2317
2318                 ret = update_io_ticks();
2319
2320                 if (helper_do_stat) {
2321                         helper_do_stat = 0;
2322                         __show_running_run_stats();
2323                 }
2324
2325                 if (!is_backend)
2326                         print_thread_status();
2327         }
2328
2329         sk_out_drop();
2330         return NULL;
2331 }
2332
2333 static int create_helper_thread(struct sk_out *sk_out)
2334 {
2335         int ret;
2336
2337         setup_disk_util();
2338
2339         pthread_cond_init(&helper_cond, NULL);
2340         pthread_mutex_init(&helper_lock, NULL);
2341
2342         ret = pthread_create(&helper_thread, NULL, helper_thread_main, sk_out);
2343         if (ret) {
2344                 log_err("Can't create helper thread: %s\n", strerror(ret));
2345                 return 1;
2346         }
2347
2348         dprint(FD_MUTEX, "wait on startup_mutex\n");
2349         fio_mutex_down(startup_mutex);
2350         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2351         return 0;
2352 }
2353
2354 int fio_backend(struct sk_out *sk_out)
2355 {
2356         struct thread_data *td;
2357         int i;
2358
2359         if (exec_profile) {
2360                 if (load_profile(exec_profile))
2361                         return 1;
2362                 free(exec_profile);
2363                 exec_profile = NULL;
2364         }
2365         if (!thread_number)
2366                 return 0;
2367
2368         if (write_bw_log) {
2369                 struct log_params p = {
2370                         .log_type = IO_LOG_TYPE_BW,
2371                 };
2372
2373                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2374                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2375                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2376         }
2377
2378         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2379         if (startup_mutex == NULL)
2380                 return 1;
2381
2382         set_genesis_time();
2383         stat_init();
2384         create_helper_thread(sk_out);
2385
2386         cgroup_list = smalloc(sizeof(*cgroup_list));
2387         INIT_FLIST_HEAD(cgroup_list);
2388
2389         run_threads(sk_out);
2390
2391         wait_for_helper_thread_exit();
2392
2393         if (!fio_abort) {
2394                 __show_run_stats();
2395                 if (write_bw_log) {
2396                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2397                                 struct io_log *log = agg_io_log[i];
2398
2399                                 flush_log(log, 0);
2400                                 free_log(log);
2401                         }
2402                 }
2403         }
2404
2405         for_each_td(td, i) {
2406                 fio_options_free(td);
2407                 if (td->rusage_sem) {
2408                         fio_mutex_remove(td->rusage_sem);
2409                         td->rusage_sem = NULL;
2410                 }
2411         }
2412
2413         free_disk_util();
2414         cgroup_kill(cgroup_list);
2415         sfree(cgroup_list);
2416         sfree(cgroup_mnt);
2417
2418         fio_mutex_remove(startup_mutex);
2419         stat_exit();
2420         return exit_value;
2421 }