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 "lib/tp.h"
  58 #include "workqueue.h"
  59 #include "lib/mountcheck.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
 817         if (in_ramp_time(td))
 818                 td_set_runstate(td, TD_RAMP);
 819         else
 820                 td_set_runstate(td, TD_RUNNING);
 821
 822         lat_target_init(td);
 823
 824         total_bytes = td->o.size;
 825         /*
 826         * Allow random overwrite workloads to write up to io_limit
 827         * before starting verification phase as 'size' doesn't apply.
 828         */
 829         if (td_write(td) && td_random(td) && td->o.norandommap)
 830                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 831         /*
 832          * If verify_backlog is enabled, we'll run the verify in this
 833          * handler as well. For that case, we may need up to twice the
 834          * amount of bytes.
 835          */
 836         if (td->o.verify != VERIFY_NONE &&
 837            (td_write(td) && td->o.verify_backlog))
 838                 total_bytes += td->o.size;
 839
 840         /* In trimwrite mode, each byte is trimmed and then written, so
 841          * allow total_bytes to be twice as big */
 842         if (td_trimwrite(td))
 843                 total_bytes += td->total_io_size;
 844
 845         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 846                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 847                 td->o.time_based) {
 848                 struct timeval comp_time;
 849                 struct io_u *io_u;
 850                 int full;
 851                 enum fio_ddir ddir;
 852
 853                 check_update_rusage(td);
 854
 855                 if (td->terminate || td->done)
 856                         break;
 857
 858                 update_tv_cache(td);
 859
 860                 if (runtime_exceeded(td, &td->tv_cache)) {
 861                         __update_tv_cache(td);
 862                         if (runtime_exceeded(td, &td->tv_cache)) {
 863                                 fio_mark_td_terminate(td);
 864                                 break;
 865                         }
 866                 }
 867
 868                 if (flow_threshold_exceeded(td))
 869                         continue;
 870
 871                 if (!td->o.time_based && bytes_issued >= total_bytes)
 872                         break;
 873
 874                 io_u = get_io_u(td);
 875                 if (IS_ERR_OR_NULL(io_u)) {
 876                         int err = PTR_ERR(io_u);
 877
 878                         io_u = NULL;
 879                         if (err == -EBUSY) {
 880                                 ret = FIO_Q_BUSY;
 881                                 goto reap;
 882                         }
 883                         if (td->o.latency_target)
 884                                 goto reap;
 885                         break;
 886                 }
 887
 888                 ddir = io_u->ddir;
 889
 890                 /*
 891                  * Add verification end_io handler if:
 892                  *      - Asked to verify (!td_rw(td))
 893                  *      - Or the io_u is from our verify list (mixed write/ver)
 894                  */
 895                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 896                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 897
 898                         if (!td->o.verify_pattern_bytes) {
 899                                 io_u->rand_seed = __rand(&td->verify_state);
 900                                 if (sizeof(int) != sizeof(long *))
 901                                         io_u->rand_seed *= __rand(&td->verify_state);
 902                         }
 903
 904                         if (verify_state_should_stop(td, io_u)) {
 905                                 put_io_u(td, io_u);
 906                                 break;
 907                         }
 908
 909                         if (td->o.verify_async)
 910                                 io_u->end_io = verify_io_u_async;
 911                         else
 912                                 io_u->end_io = verify_io_u;
 913                         td_set_runstate(td, TD_VERIFYING);
 914                 } else if (in_ramp_time(td))
 915                         td_set_runstate(td, TD_RAMP);
 916                 else
 917                         td_set_runstate(td, TD_RUNNING);
 918
 919                 /*
 920                  * Always log IO before it's issued, so we know the specific
 921                  * order of it. The logged unit will track when the IO has
 922                  * completed.
 923                  */
 924                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 925                     td->o.do_verify &&
 926                     td->o.verify != VERIFY_NONE &&
 927                     !td->o.experimental_verify)
 928                         log_io_piece(td, io_u);
 929
 930                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 931                         const unsigned long blen = io_u->xfer_buflen;
 932                         const enum fio_ddir ddir = acct_ddir(io_u);
 933
 934                         if (td->error)
 935                                 break;
 936
 937                         ret = workqueue_enqueue(&td->io_wq, &io_u->work);
 938                         if (ret)
 939                                 ret = FIO_Q_QUEUED;
 940                         else
 941                                 ret = FIO_Q_BUSY;
 942
 943                         if (ret == FIO_Q_QUEUED && ddir_rw(ddir)) {
 944                                 td->io_issues[ddir]++;
 945                                 td->io_issue_bytes[ddir] += blen;
 946                                 td->rate_io_issue_bytes[ddir] += blen;
 947                         }
 948
 949                         if (should_check_rate(td))
 950                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 951
 952                 } else {
 953                         ret = td_io_queue(td, io_u);
 954
 955                         if (should_check_rate(td))
 956                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 957
 958                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 959                                 break;
 960
 961                         /*
 962                          * See if we need to complete some commands. Note that
 963                          * we can get BUSY even without IO queued, if the
 964                          * system is resource starved.
 965                          */
 966 reap:
 967                         full = queue_full(td) ||
 968                                 (ret == FIO_Q_BUSY && td->cur_depth);
 969                         if (full || io_in_polling(td))
 970                                 ret = wait_for_completions(td, &comp_time);
 971                 }
 972                 if (ret < 0)
 973                         break;
 974                 if (!ddir_rw_sum(td->bytes_done) &&
 975                     !(td->io_ops->flags & FIO_NOIO))
 976                         continue;
 977
 978                 if (!in_ramp_time(td) && should_check_rate(td)) {
 979                         if (check_min_rate(td, &comp_time)) {
 980                                 if (exitall_on_terminate)
 981                                         fio_terminate_threads(td->groupid);
 982                                 td_verror(td, EIO, "check_min_rate");
 983                                 break;
 984                         }
 985                 }
 986                 if (!in_ramp_time(td) && td->o.latency_target)
 987                         lat_target_check(td);
 988
 989                 if (td->o.thinktime) {
 990                         unsigned long long b;
 991
 992                         b = ddir_rw_sum(td->io_blocks);
 993                         if (!(b % td->o.thinktime_blocks)) {
 994                                 int left;
 995
 996                                 io_u_quiesce(td);
 997
 998                                 if (td->o.thinktime_spin)
 999                                         usec_spin(td->o.thinktime_spin);
1000
1001                                 left = td->o.thinktime - td->o.thinktime_spin;
1002                                 if (left)
1003                                         usec_sleep(td, left);
1004                         }
1005                 }
1006         }
1007
1008         check_update_rusage(td);
1009
1010         if (td->trim_entries)
1011                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1012
1013         if (td->o.fill_device && td->error == ENOSPC) {
1014                 td->error = 0;
1015                 fio_mark_td_terminate(td);
1016         }
1017         if (!td->error) {
1018                 struct fio_file *f;
1019
1020                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1021                         workqueue_flush(&td->io_wq);
1022                         i = 0;
1023                 } else
1024                         i = td->cur_depth;
1025
1026                 if (i) {
1027                         ret = io_u_queued_complete(td, i);
1028                         if (td->o.fill_device && td->error == ENOSPC)
1029                                 td->error = 0;
1030                 }
1031
1032                 if (should_fsync(td) && td->o.end_fsync) {
1033                         td_set_runstate(td, TD_FSYNCING);
1034
1035                         for_each_file(td, f, i) {
1036                                 if (!fio_file_fsync(td, f))
1037                                         continue;
1038
1039                                 log_err("fio: end_fsync failed for file %s\n",
1040                                                                 f->file_name);
1041                         }
1042                 }
1043         } else
1044                 cleanup_pending_aio(td);
1045
1046         /*
1047          * stop job if we failed doing any IO
1048          */
1049         if (!ddir_rw_sum(td->this_io_bytes))
1050                 td->done = 1;
1051
1052         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1053 }
1054
1055 static void cleanup_io_u(struct thread_data *td)
1056 {
1057         struct io_u *io_u;
1058
1059         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1060
1061                 if (td->io_ops->io_u_free)
1062                         td->io_ops->io_u_free(td, io_u);
1063
1064                 fio_memfree(io_u, sizeof(*io_u));
1065         }
1066
1067         free_io_mem(td);
1068
1069         io_u_rexit(&td->io_u_requeues);
1070         io_u_qexit(&td->io_u_freelist);
1071         io_u_qexit(&td->io_u_all);
1072
1073         if (td->last_write_comp)
1074                 sfree(td->last_write_comp);
1075 }
1076
1077 static int init_io_u(struct thread_data *td)
1078 {
1079         struct io_u *io_u;
1080         unsigned int max_bs, min_write;
1081         int cl_align, i, max_units;
1082         int data_xfer = 1, err;
1083         char *p;
1084
1085         max_units = td->o.iodepth;
1086         max_bs = td_max_bs(td);
1087         min_write = td->o.min_bs[DDIR_WRITE];
1088         td->orig_buffer_size = (unsigned long long) max_bs
1089                                         * (unsigned long long) max_units;
1090
1091         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1092                 data_xfer = 0;
1093
1094         err = 0;
1095         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1096         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1097         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1098
1099         if (err) {
1100                 log_err("fio: failed setting up IO queues\n");
1101                 return 1;
1102         }
1103
1104         /*
1105          * if we may later need to do address alignment, then add any
1106          * possible adjustment here so that we don't cause a buffer
1107          * overflow later. this adjustment may be too much if we get
1108          * lucky and the allocator gives us an aligned address.
1109          */
1110         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1111             (td->io_ops->flags & FIO_RAWIO))
1112                 td->orig_buffer_size += page_mask + td->o.mem_align;
1113
1114         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1115                 unsigned long bs;
1116
1117                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1118                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1119         }
1120
1121         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1122                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1123                 return 1;
1124         }
1125
1126         if (data_xfer && allocate_io_mem(td))
1127                 return 1;
1128
1129         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1130             (td->io_ops->flags & FIO_RAWIO))
1131                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1132         else
1133                 p = td->orig_buffer;
1134
1135         cl_align = os_cache_line_size();
1136
1137         for (i = 0; i < max_units; i++) {
1138                 void *ptr;
1139
1140                 if (td->terminate)
1141                         return 1;
1142
1143                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1144                 if (!ptr) {
1145                         log_err("fio: unable to allocate aligned memory\n");
1146                         break;
1147                 }
1148
1149                 io_u = ptr;
1150                 memset(io_u, 0, sizeof(*io_u));
1151                 INIT_FLIST_HEAD(&io_u->verify_list);
1152                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1153
1154                 if (data_xfer) {
1155                         io_u->buf = p;
1156                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1157
1158                         if (td_write(td))
1159                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1160                         if (td_write(td) && td->o.verify_pattern_bytes) {
1161                                 /*
1162                                  * Fill the buffer with the pattern if we are
1163                                  * going to be doing writes.
1164                                  */
1165                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1166                         }
1167                 }
1168
1169                 io_u->index = i;
1170                 io_u->flags = IO_U_F_FREE;
1171                 io_u_qpush(&td->io_u_freelist, io_u);
1172
1173                 /*
1174                  * io_u never leaves this stack, used for iteration of all
1175                  * io_u buffers.
1176                  */
1177                 io_u_qpush(&td->io_u_all, io_u);
1178
1179                 if (td->io_ops->io_u_init) {
1180                         int ret = td->io_ops->io_u_init(td, io_u);
1181
1182                         if (ret) {
1183                                 log_err("fio: failed to init engine data: %d\n", ret);
1184                                 return 1;
1185                         }
1186                 }
1187
1188                 p += max_bs;
1189         }
1190
1191         if (td->o.verify != VERIFY_NONE) {
1192                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1193                 if (!td->last_write_comp) {
1194                         log_err("fio: failed to alloc write comp data\n");
1195                         return 1;
1196                 }
1197         }
1198
1199         return 0;
1200 }
1201
1202 static int switch_ioscheduler(struct thread_data *td)
1203 {
1204         char tmp[256], tmp2[128];
1205         FILE *f;
1206         int ret;
1207
1208         if (td->io_ops->flags & FIO_DISKLESSIO)
1209                 return 0;
1210
1211         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1212
1213         f = fopen(tmp, "r+");
1214         if (!f) {
1215                 if (errno == ENOENT) {
1216                         log_err("fio: os or kernel doesn't support IO scheduler"
1217                                 " switching\n");
1218                         return 0;
1219                 }
1220                 td_verror(td, errno, "fopen iosched");
1221                 return 1;
1222         }
1223
1224         /*
1225          * Set io scheduler.
1226          */
1227         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1228         if (ferror(f) || ret != 1) {
1229                 td_verror(td, errno, "fwrite");
1230                 fclose(f);
1231                 return 1;
1232         }
1233
1234         rewind(f);
1235
1236         /*
1237          * Read back and check that the selected scheduler is now the default.
1238          */
1239         memset(tmp, 0, sizeof(tmp));
1240         ret = fread(tmp, sizeof(tmp), 1, f);
1241         if (ferror(f) || ret < 0) {
1242                 td_verror(td, errno, "fread");
1243                 fclose(f);
1244                 return 1;
1245         }
1246         /*
1247          * either a list of io schedulers or "none\n" is expected.
1248          */
1249         tmp[strlen(tmp) - 1] = '\0';
1250
1251
1252         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1253         if (!strstr(tmp, tmp2)) {
1254                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1255                 td_verror(td, EINVAL, "iosched_switch");
1256                 fclose(f);
1257                 return 1;
1258         }
1259
1260         fclose(f);
1261         return 0;
1262 }
1263
1264 static bool keep_running(struct thread_data *td)
1265 {
1266         unsigned long long limit;
1267
1268         if (td->done)
1269                 return false;
1270         if (td->o.time_based)
1271                 return true;
1272         if (td->o.loops) {
1273                 td->o.loops--;
1274                 return true;
1275         }
1276         if (exceeds_number_ios(td))
1277                 return false;
1278
1279         if (td->o.io_limit)
1280                 limit = td->o.io_limit;
1281         else
1282                 limit = td->o.size;
1283
1284         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1285                 uint64_t diff;
1286
1287                 /*
1288                  * If the difference is less than the minimum IO size, we
1289                  * are done.
1290                  */
1291                 diff = limit - ddir_rw_sum(td->io_bytes);
1292                 if (diff < td_max_bs(td))
1293                         return false;
1294
1295                 if (fio_files_done(td))
1296                         return false;
1297
1298                 return true;
1299         }
1300
1301         return false;
1302 }
1303
1304 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1305 {
1306         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1307         int ret;
1308         char *str;
1309
1310         str = malloc(newlen);
1311         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1312
1313         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1314         ret = system(str);
1315         if (ret == -1)
1316                 log_err("fio: exec of cmd <%s> failed\n", str);
1317
1318         free(str);
1319         return ret;
1320 }
1321
1322 /*
1323  * Dry run to compute correct state of numberio for verification.
1324  */
1325 static uint64_t do_dry_run(struct thread_data *td)
1326 {
1327         td_set_runstate(td, TD_RUNNING);
1328
1329         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1330                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1331                 struct io_u *io_u;
1332                 int ret;
1333
1334                 if (td->terminate || td->done)
1335                         break;
1336
1337                 io_u = get_io_u(td);
1338                 if (!io_u)
1339                         break;
1340
1341                 io_u_set(io_u, IO_U_F_FLIGHT);
1342                 io_u->error = 0;
1343                 io_u->resid = 0;
1344                 if (ddir_rw(acct_ddir(io_u)))
1345                         td->io_issues[acct_ddir(io_u)]++;
1346                 if (ddir_rw(io_u->ddir)) {
1347                         io_u_mark_depth(td, 1);
1348                         td->ts.total_io_u[io_u->ddir]++;
1349                 }
1350
1351                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1352                     td->o.do_verify &&
1353                     td->o.verify != VERIFY_NONE &&
1354                     !td->o.experimental_verify)
1355                         log_io_piece(td, io_u);
1356
1357                 ret = io_u_sync_complete(td, io_u);
1358                 (void) ret;
1359         }
1360
1361         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1362 }
1363
1364 static void io_workqueue_fn(struct submit_worker *sw,
1365                             struct workqueue_work *work)
1366 {
1367         struct io_u *io_u = container_of(work, struct io_u, work);
1368         const enum fio_ddir ddir = io_u->ddir;
1369         struct thread_data *td = sw->private;
1370         int ret;
1371
1372         dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1373
1374         io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1375
1376         td->cur_depth++;
1377
1378         do {
1379                 ret = td_io_queue(td, io_u);
1380                 if (ret != FIO_Q_BUSY)
1381                         break;
1382                 ret = io_u_queued_complete(td, 1);
1383                 if (ret > 0)
1384                         td->cur_depth -= ret;
1385                 io_u_clear(io_u, IO_U_F_FLIGHT);
1386         } while (1);
1387
1388         dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1389
1390         io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1391
1392         if (ret == FIO_Q_COMPLETED)
1393                 td->cur_depth--;
1394         else if (ret == FIO_Q_QUEUED) {
1395                 unsigned int min_evts;
1396
1397                 if (td->o.iodepth == 1)
1398                         min_evts = 1;
1399                 else
1400                         min_evts = 0;
1401
1402                 ret = io_u_queued_complete(td, min_evts);
1403                 if (ret > 0)
1404                         td->cur_depth -= ret;
1405         } else if (ret == FIO_Q_BUSY) {
1406                 ret = io_u_queued_complete(td, td->cur_depth);
1407                 if (ret > 0)
1408                         td->cur_depth -= ret;
1409         }
1410 }
1411
1412 static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
1413 {
1414         struct thread_data *td = sw->private;
1415
1416         if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
1417                 return true;
1418
1419         return false;
1420 }
1421
1422 static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
1423 {
1424         struct thread_data *td = sw->private;
1425         int ret;
1426
1427         ret = io_u_quiesce(td);
1428         if (ret > 0)
1429                 td->cur_depth -= ret;
1430 }
1431
1432 static int io_workqueue_alloc_fn(struct submit_worker *sw)
1433 {
1434         struct thread_data *td;
1435
1436         td = calloc(1, sizeof(*td));
1437         sw->private = td;
1438         return 0;
1439 }
1440
1441 static void io_workqueue_free_fn(struct submit_worker *sw)
1442 {
1443         free(sw->private);
1444         sw->private = NULL;
1445 }
1446
1447 static int io_workqueue_init_worker_fn(struct submit_worker *sw)
1448 {
1449         struct thread_data *parent = sw->wq->td;
1450         struct thread_data *td = sw->private;
1451         int fio_unused ret;
1452
1453         memcpy(&td->o, &parent->o, sizeof(td->o));
1454         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
1455         td->o.uid = td->o.gid = -1U;
1456         dup_files(td, parent);
1457         td->eo = parent->eo;
1458         fio_options_mem_dupe(td);
1459
1460         if (ioengine_load(td))
1461                 goto err;
1462
1463         if (td->o.odirect)
1464                 td->io_ops->flags |= FIO_RAWIO;
1465
1466         td->pid = gettid();
1467
1468         INIT_FLIST_HEAD(&td->io_log_list);
1469         INIT_FLIST_HEAD(&td->io_hist_list);
1470         INIT_FLIST_HEAD(&td->verify_list);
1471         INIT_FLIST_HEAD(&td->trim_list);
1472         INIT_FLIST_HEAD(&td->next_rand_list);
1473         td->io_hist_tree = RB_ROOT;
1474
1475         td->o.iodepth = 1;
1476         if (td_io_init(td))
1477                 goto err_io_init;
1478
1479         fio_gettime(&td->epoch, NULL);
1480         fio_getrusage(&td->ru_start);
1481         clear_io_state(td, 1);
1482
1483         td_set_runstate(td, TD_RUNNING);
1484         td->flags |= TD_F_CHILD;
1485         td->parent = parent;
1486         return 0;
1487
1488 err_io_init:
1489         close_ioengine(td);
1490 err:
1491         return 1;
1492
1493 }
1494
1495 static void io_workqueue_exit_worker_fn(struct submit_worker *sw)
1496 {
1497         struct thread_data *td = sw->private;
1498
1499         fio_options_free(td);
1500         close_and_free_files(td);
1501         if (td->io_ops)
1502                 close_ioengine(td);
1503         td_set_runstate(td, TD_EXITED);
1504 }
1505
1506 struct workqueue_ops rated_wq_ops = {
1507         .fn                     = io_workqueue_fn,
1508         .pre_sleep_flush_fn     = io_workqueue_pre_sleep_flush_fn,
1509         .pre_sleep_fn           = io_workqueue_pre_sleep_fn,
1510         .alloc_worker_fn        = io_workqueue_alloc_fn,
1511         .free_worker_fn         = io_workqueue_free_fn,
1512         .init_worker_fn         = io_workqueue_init_worker_fn,
1513         .exit_worker_fn         = io_workqueue_exit_worker_fn,
1514 };
1515
1516 /*
1517  * Entry point for the thread based jobs. The process based jobs end up
1518  * here as well, after a little setup.
1519  */
1520 static void *thread_main(void *data)
1521 {
1522         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1523         struct thread_data *td = data;
1524         struct thread_options *o = &td->o;
1525         pthread_condattr_t attr;
1526         int clear_state;
1527         int ret;
1528
1529         if (!o->use_thread) {
1530                 setsid();
1531                 td->pid = getpid();
1532         } else
1533                 td->pid = gettid();
1534
1535         fio_local_clock_init(o->use_thread);
1536
1537         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1538
1539         if (is_backend)
1540                 fio_server_send_start(td);
1541
1542         INIT_FLIST_HEAD(&td->io_log_list);
1543         INIT_FLIST_HEAD(&td->io_hist_list);
1544         INIT_FLIST_HEAD(&td->verify_list);
1545         INIT_FLIST_HEAD(&td->trim_list);
1546         INIT_FLIST_HEAD(&td->next_rand_list);
1547         pthread_mutex_init(&td->io_u_lock, NULL);
1548         td->io_hist_tree = RB_ROOT;
1549
1550         pthread_condattr_init(&attr);
1551         pthread_cond_init(&td->verify_cond, &attr);
1552         pthread_cond_init(&td->free_cond, &attr);
1553
1554         td_set_runstate(td, TD_INITIALIZED);
1555         dprint(FD_MUTEX, "up startup_mutex\n");
1556         fio_mutex_up(startup_mutex);
1557         dprint(FD_MUTEX, "wait on td->mutex\n");
1558         fio_mutex_down(td->mutex);
1559         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1560
1561         /*
1562          * A new gid requires privilege, so we need to do this before setting
1563          * the uid.
1564          */
1565         if (o->gid != -1U && setgid(o->gid)) {
1566                 td_verror(td, errno, "setgid");
1567                 goto err;
1568         }
1569         if (o->uid != -1U && setuid(o->uid)) {
1570                 td_verror(td, errno, "setuid");
1571                 goto err;
1572         }
1573
1574         /*
1575          * If we have a gettimeofday() thread, make sure we exclude that
1576          * thread from this job
1577          */
1578         if (o->gtod_cpu)
1579                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1580
1581         /*
1582          * Set affinity first, in case it has an impact on the memory
1583          * allocations.
1584          */
1585         if (fio_option_is_set(o, cpumask)) {
1586                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1587                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1588                         if (!ret) {
1589                                 log_err("fio: no CPUs set\n");
1590                                 log_err("fio: Try increasing number of available CPUs\n");
1591                                 td_verror(td, EINVAL, "cpus_split");
1592                                 goto err;
1593                         }
1594                 }
1595                 ret = fio_setaffinity(td->pid, o->cpumask);
1596                 if (ret == -1) {
1597                         td_verror(td, errno, "cpu_set_affinity");
1598                         goto err;
1599                 }
1600         }
1601
1602 #ifdef CONFIG_LIBNUMA
1603         /* numa node setup */
1604         if (fio_option_is_set(o, numa_cpunodes) ||
1605             fio_option_is_set(o, numa_memnodes)) {
1606                 struct bitmask *mask;
1607
1608                 if (numa_available() < 0) {
1609                         td_verror(td, errno, "Does not support NUMA API\n");
1610                         goto err;
1611                 }
1612
1613                 if (fio_option_is_set(o, numa_cpunodes)) {
1614                         mask = numa_parse_nodestring(o->numa_cpunodes);
1615                         ret = numa_run_on_node_mask(mask);
1616                         numa_free_nodemask(mask);
1617                         if (ret == -1) {
1618                                 td_verror(td, errno, \
1619                                         "numa_run_on_node_mask failed\n");
1620                                 goto err;
1621                         }
1622                 }
1623
1624                 if (fio_option_is_set(o, numa_memnodes)) {
1625                         mask = NULL;
1626                         if (o->numa_memnodes)
1627                                 mask = numa_parse_nodestring(o->numa_memnodes);
1628
1629                         switch (o->numa_mem_mode) {
1630                         case MPOL_INTERLEAVE:
1631                                 numa_set_interleave_mask(mask);
1632                                 break;
1633                         case MPOL_BIND:
1634                                 numa_set_membind(mask);
1635                                 break;
1636                         case MPOL_LOCAL:
1637                                 numa_set_localalloc();
1638                                 break;
1639                         case MPOL_PREFERRED:
1640                                 numa_set_preferred(o->numa_mem_prefer_node);
1641                                 break;
1642                         case MPOL_DEFAULT:
1643                         default:
1644                                 break;
1645                         }
1646
1647                         if (mask)
1648                                 numa_free_nodemask(mask);
1649
1650                 }
1651         }
1652 #endif
1653
1654         if (fio_pin_memory(td))
1655                 goto err;
1656
1657         /*
1658          * May alter parameters that init_io_u() will use, so we need to
1659          * do this first.
1660          */
1661         if (init_iolog(td))
1662                 goto err;
1663
1664         if (init_io_u(td))
1665                 goto err;
1666
1667         if (o->verify_async && verify_async_init(td))
1668                 goto err;
1669
1670         if (fio_option_is_set(o, ioprio) ||
1671             fio_option_is_set(o, ioprio_class)) {
1672                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1673                 if (ret == -1) {
1674                         td_verror(td, errno, "ioprio_set");
1675                         goto err;
1676                 }
1677         }
1678
1679         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1680                 goto err;
1681
1682         errno = 0;
1683         if (nice(o->nice) == -1 && errno != 0) {
1684                 td_verror(td, errno, "nice");
1685                 goto err;
1686         }
1687
1688         if (o->ioscheduler && switch_ioscheduler(td))
1689                 goto err;
1690
1691         if (!o->create_serialize && setup_files(td))
1692                 goto err;
1693
1694         if (td_io_init(td))
1695                 goto err;
1696
1697         if (init_random_map(td))
1698                 goto err;
1699
1700         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1701                 goto err;
1702
1703         if (o->pre_read) {
1704                 if (pre_read_files(td) < 0)
1705                         goto err;
1706         }
1707
1708         if (td->flags & TD_F_COMPRESS_LOG)
1709                 tp_init(&td->tp_data);
1710
1711         fio_verify_init(td);
1712
1713         if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1714             workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth))
1715                 goto err;
1716
1717         fio_gettime(&td->epoch, NULL);
1718         fio_getrusage(&td->ru_start);
1719         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1720         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1721
1722         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1723                         o->ratemin[DDIR_TRIM]) {
1724                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1725                                         sizeof(td->bw_sample_time));
1726                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1727                                         sizeof(td->bw_sample_time));
1728                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1729                                         sizeof(td->bw_sample_time));
1730         }
1731
1732         clear_state = 0;
1733         while (keep_running(td)) {
1734                 uint64_t verify_bytes;
1735
1736                 fio_gettime(&td->start, NULL);
1737                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1738
1739                 if (clear_state)
1740                         clear_io_state(td, 0);
1741
1742                 prune_io_piece_log(td);
1743
1744                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1745                         verify_bytes = do_dry_run(td);
1746                 else
1747                         verify_bytes = do_io(td);
1748
1749                 clear_state = 1;
1750
1751                 /*
1752                  * Make sure we've successfully updated the rusage stats
1753                  * before waiting on the stat mutex. Otherwise we could have
1754                  * the stat thread holding stat mutex and waiting for
1755                  * the rusage_sem, which would never get upped because
1756                  * this thread is waiting for the stat mutex.
1757                  */
1758                 check_update_rusage(td);
1759
1760                 fio_mutex_down(stat_mutex);
1761                 if (td_read(td) && td->io_bytes[DDIR_READ])
1762                         update_runtime(td, elapsed_us, DDIR_READ);
1763                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1764                         update_runtime(td, elapsed_us, DDIR_WRITE);
1765                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1766                         update_runtime(td, elapsed_us, DDIR_TRIM);
1767                 fio_gettime(&td->start, NULL);
1768                 fio_mutex_up(stat_mutex);
1769
1770                 if (td->error || td->terminate)
1771                         break;
1772
1773                 if (!o->do_verify ||
1774                     o->verify == VERIFY_NONE ||
1775                     (td->io_ops->flags & FIO_UNIDIR))
1776                         continue;
1777
1778                 clear_io_state(td, 0);
1779
1780                 fio_gettime(&td->start, NULL);
1781
1782                 do_verify(td, verify_bytes);
1783
1784                 /*
1785                  * See comment further up for why this is done here.
1786                  */
1787                 check_update_rusage(td);
1788
1789                 fio_mutex_down(stat_mutex);
1790                 update_runtime(td, elapsed_us, DDIR_READ);
1791                 fio_gettime(&td->start, NULL);
1792                 fio_mutex_up(stat_mutex);
1793
1794                 if (td->error || td->terminate)
1795                         break;
1796         }
1797
1798         update_rusage_stat(td);
1799         td->ts.total_run_time = mtime_since_now(&td->epoch);
1800         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1801         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1802         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1803
1804         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1805             (td->o.verify != VERIFY_NONE && td_write(td)))
1806                 verify_save_state(td->thread_number);
1807
1808         fio_unpin_memory(td);
1809
1810         fio_writeout_logs(td);
1811
1812         if (o->io_submit_mode == IO_MODE_OFFLOAD)
1813                 workqueue_exit(&td->io_wq);
1814
1815         if (td->flags & TD_F_COMPRESS_LOG)
1816                 tp_exit(&td->tp_data);
1817
1818         if (o->exec_postrun)
1819                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1820
1821         if (exitall_on_terminate)
1822                 fio_terminate_threads(td->groupid);
1823
1824 err:
1825         if (td->error)
1826                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1827                                                         td->verror);
1828
1829         if (o->verify_async)
1830                 verify_async_exit(td);
1831
1832         close_and_free_files(td);
1833         cleanup_io_u(td);
1834         close_ioengine(td);
1835         cgroup_shutdown(td, &cgroup_mnt);
1836         verify_free_state(td);
1837
1838         if (fio_option_is_set(o, cpumask)) {
1839                 ret = fio_cpuset_exit(&o->cpumask);
1840                 if (ret)
1841                         td_verror(td, ret, "fio_cpuset_exit");
1842         }
1843
1844         /*
1845          * do this very late, it will log file closing as well
1846          */
1847         if (o->write_iolog_file)
1848                 write_iolog_close(td);
1849
1850         fio_mutex_remove(td->mutex);
1851         td->mutex = NULL;
1852
1853         td_set_runstate(td, TD_EXITED);
1854
1855         /*
1856          * Do this last after setting our runstate to exited, so we
1857          * know that the stat thread is signaled.
1858          */
1859         check_update_rusage(td);
1860
1861         return (void *) (uintptr_t) td->error;
1862 }
1863
1864
1865 /*
1866  * We cannot pass the td data into a forked process, so attach the td and
1867  * pass it to the thread worker.
1868  */
1869 static int fork_main(int shmid, int offset)
1870 {
1871         struct thread_data *td;
1872         void *data, *ret;
1873
1874 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1875         data = shmat(shmid, NULL, 0);
1876         if (data == (void *) -1) {
1877                 int __err = errno;
1878
1879                 perror("shmat");
1880                 return __err;
1881         }
1882 #else
1883         /*
1884          * HP-UX inherits shm mappings?
1885          */
1886         data = threads;
1887 #endif
1888
1889         td = data + offset * sizeof(struct thread_data);
1890         ret = thread_main(td);
1891         shmdt(data);
1892         return (int) (uintptr_t) ret;
1893 }
1894
1895 static void dump_td_info(struct thread_data *td)
1896 {
1897         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1898                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1899                         (unsigned long) time_since_now(&td->terminate_time));
1900 }
1901
1902 /*
1903  * Run over the job map and reap the threads that have exited, if any.
1904  */
1905 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1906                          unsigned int *m_rate)
1907 {
1908         struct thread_data *td;
1909         unsigned int cputhreads, realthreads, pending;
1910         int i, status, ret;
1911
1912         /*
1913          * reap exited threads (TD_EXITED -> TD_REAPED)
1914          */
1915         realthreads = pending = cputhreads = 0;
1916         for_each_td(td, i) {
1917                 int flags = 0;
1918
1919                 /*
1920                  * ->io_ops is NULL for a thread that has closed its
1921                  * io engine
1922                  */
1923                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1924                         cputhreads++;
1925                 else
1926                         realthreads++;
1927
1928                 if (!td->pid) {
1929                         pending++;
1930                         continue;
1931                 }
1932                 if (td->runstate == TD_REAPED)
1933                         continue;
1934                 if (td->o.use_thread) {
1935                         if (td->runstate == TD_EXITED) {
1936                                 td_set_runstate(td, TD_REAPED);
1937                                 goto reaped;
1938                         }
1939                         continue;
1940                 }
1941
1942                 flags = WNOHANG;
1943                 if (td->runstate == TD_EXITED)
1944                         flags = 0;
1945
1946                 /*
1947                  * check if someone quit or got killed in an unusual way
1948                  */
1949                 ret = waitpid(td->pid, &status, flags);
1950                 if (ret < 0) {
1951                         if (errno == ECHILD) {
1952                                 log_err("fio: pid=%d disappeared %d\n",
1953                                                 (int) td->pid, td->runstate);
1954                                 td->sig = ECHILD;
1955                                 td_set_runstate(td, TD_REAPED);
1956                                 goto reaped;
1957                         }
1958                         perror("waitpid");
1959                 } else if (ret == td->pid) {
1960                         if (WIFSIGNALED(status)) {
1961                                 int sig = WTERMSIG(status);
1962
1963                                 if (sig != SIGTERM && sig != SIGUSR2)
1964                                         log_err("fio: pid=%d, got signal=%d\n",
1965                                                         (int) td->pid, sig);
1966                                 td->sig = sig;
1967                                 td_set_runstate(td, TD_REAPED);
1968                                 goto reaped;
1969                         }
1970                         if (WIFEXITED(status)) {
1971                                 if (WEXITSTATUS(status) && !td->error)
1972                                         td->error = WEXITSTATUS(status);
1973
1974                                 td_set_runstate(td, TD_REAPED);
1975                                 goto reaped;
1976                         }
1977                 }
1978
1979                 /*
1980                  * If the job is stuck, do a forceful timeout of it and
1981                  * move on.
1982                  */
1983                 if (td->terminate &&
1984                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1985                         dump_td_info(td);
1986                         td_set_runstate(td, TD_REAPED);
1987                         goto reaped;
1988                 }
1989
1990                 /*
1991                  * thread is not dead, continue
1992                  */
1993                 pending++;
1994                 continue;
1995 reaped:
1996                 (*nr_running)--;
1997                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1998                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1999                 if (!td->pid)
2000                         pending--;
2001
2002                 if (td->error)
2003                         exit_value++;
2004
2005                 done_secs += mtime_since_now(&td->epoch) / 1000;
2006                 profile_td_exit(td);
2007         }
2008
2009         if (*nr_running == cputhreads && !pending && realthreads)
2010                 fio_terminate_threads(TERMINATE_ALL);
2011 }
2012
2013 static bool __check_trigger_file(void)
2014 {
2015         struct stat sb;
2016
2017         if (!trigger_file)
2018                 return false;
2019
2020         if (stat(trigger_file, &sb))
2021                 return false;
2022
2023         if (unlink(trigger_file) < 0)
2024                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2025                                                         strerror(errno));
2026
2027         return true;
2028 }
2029
2030 static bool trigger_timedout(void)
2031 {
2032         if (trigger_timeout)
2033                 return time_since_genesis() >= trigger_timeout;
2034
2035         return false;
2036 }
2037
2038 void exec_trigger(const char *cmd)
2039 {
2040         int ret;
2041
2042         if (!cmd)
2043                 return;
2044
2045         ret = system(cmd);
2046         if (ret == -1)
2047                 log_err("fio: failed executing %s trigger\n", cmd);
2048 }
2049
2050 void check_trigger_file(void)
2051 {
2052         if (__check_trigger_file() || trigger_timedout()) {
2053                 if (nr_clients)
2054                         fio_clients_send_trigger(trigger_remote_cmd);
2055                 else {
2056                         verify_save_state(IO_LIST_ALL);
2057                         fio_terminate_threads(TERMINATE_ALL);
2058                         exec_trigger(trigger_cmd);
2059                 }
2060         }
2061 }
2062
2063 static int fio_verify_load_state(struct thread_data *td)
2064 {
2065         int ret;
2066
2067         if (!td->o.verify_state)
2068                 return 0;
2069
2070         if (is_backend) {
2071                 void *data;
2072                 int ver;
2073
2074                 ret = fio_server_get_verify_state(td->o.name,
2075                                         td->thread_number - 1, &data, &ver);
2076                 if (!ret)
2077                         verify_convert_assign_state(td, data, ver);
2078         } else
2079                 ret = verify_load_state(td, "local");
2080
2081         return ret;
2082 }
2083
2084 static void do_usleep(unsigned int usecs)
2085 {
2086         check_for_running_stats();
2087         check_trigger_file();
2088         usleep(usecs);
2089 }
2090
2091 static bool check_mount_writes(struct thread_data *td)
2092 {
2093         struct fio_file *f;
2094         unsigned int i;
2095
2096         if (!td_write(td) || td->o.allow_mounted_write)
2097                 return false;
2098
2099         for_each_file(td, f, i) {
2100                 if (f->filetype != FIO_TYPE_BD)
2101                         continue;
2102                 if (device_is_mounted(f->file_name))
2103                         goto mounted;
2104         }
2105
2106         return false;
2107 mounted:
2108         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2109         return true;
2110 }
2111
2112 /*
2113  * Main function for kicking off and reaping jobs, as needed.
2114  */
2115 static void run_threads(void)
2116 {
2117         struct thread_data *td;
2118         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2119         uint64_t spent;
2120
2121         if (fio_gtod_offload && fio_start_gtod_thread())
2122                 return;
2123
2124         fio_idle_prof_init();
2125
2126         set_sig_handlers();
2127
2128         nr_thread = nr_process = 0;
2129         for_each_td(td, i) {
2130                 if (check_mount_writes(td))
2131                         return;
2132                 if (td->o.use_thread)
2133                         nr_thread++;
2134                 else
2135                         nr_process++;
2136         }
2137
2138         if (output_format & FIO_OUTPUT_NORMAL) {
2139                 log_info("Starting ");
2140                 if (nr_thread)
2141                         log_info("%d thread%s", nr_thread,
2142                                                 nr_thread > 1 ? "s" : "");
2143                 if (nr_process) {
2144                         if (nr_thread)
2145                                 log_info(" and ");
2146                         log_info("%d process%s", nr_process,
2147                                                 nr_process > 1 ? "es" : "");
2148                 }
2149                 log_info("\n");
2150                 log_info_flush();
2151         }
2152
2153         todo = thread_number;
2154         nr_running = 0;
2155         nr_started = 0;
2156         m_rate = t_rate = 0;
2157
2158         for_each_td(td, i) {
2159                 print_status_init(td->thread_number - 1);
2160
2161                 if (!td->o.create_serialize)
2162                         continue;
2163
2164                 if (fio_verify_load_state(td))
2165                         goto reap;
2166
2167                 /*
2168                  * do file setup here so it happens sequentially,
2169                  * we don't want X number of threads getting their
2170                  * client data interspersed on disk
2171                  */
2172                 if (setup_files(td)) {
2173 reap:
2174                         exit_value++;
2175                         if (td->error)
2176                                 log_err("fio: pid=%d, err=%d/%s\n",
2177                                         (int) td->pid, td->error, td->verror);
2178                         td_set_runstate(td, TD_REAPED);
2179                         todo--;
2180                 } else {
2181                         struct fio_file *f;
2182                         unsigned int j;
2183
2184                         /*
2185                          * for sharing to work, each job must always open
2186                          * its own files. so close them, if we opened them
2187                          * for creation
2188                          */
2189                         for_each_file(td, f, j) {
2190                                 if (fio_file_open(f))
2191                                         td_io_close_file(td, f);
2192                         }
2193                 }
2194         }
2195
2196         /* start idle threads before io threads start to run */
2197         fio_idle_prof_start();
2198
2199         set_genesis_time();
2200
2201         while (todo) {
2202                 struct thread_data *map[REAL_MAX_JOBS];
2203                 struct timeval this_start;
2204                 int this_jobs = 0, left;
2205
2206                 /*
2207                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2208                  */
2209                 for_each_td(td, i) {
2210                         if (td->runstate != TD_NOT_CREATED)
2211                                 continue;
2212
2213                         /*
2214                          * never got a chance to start, killed by other
2215                          * thread for some reason
2216                          */
2217                         if (td->terminate) {
2218                                 todo--;
2219                                 continue;
2220                         }
2221
2222                         if (td->o.start_delay) {
2223                                 spent = utime_since_genesis();
2224
2225                                 if (td->o.start_delay > spent)
2226                                         continue;
2227                         }
2228
2229                         if (td->o.stonewall && (nr_started || nr_running)) {
2230                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2231                                                         td->o.name);
2232                                 break;
2233                         }
2234
2235                         init_disk_util(td);
2236
2237                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2238                         td->update_rusage = 0;
2239
2240                         /*
2241                          * Set state to created. Thread will transition
2242                          * to TD_INITIALIZED when it's done setting up.
2243                          */
2244                         td_set_runstate(td, TD_CREATED);
2245                         map[this_jobs++] = td;
2246                         nr_started++;
2247
2248                         if (td->o.use_thread) {
2249                                 int ret;
2250
2251                                 dprint(FD_PROCESS, "will pthread_create\n");
2252                                 ret = pthread_create(&td->thread, NULL,
2253                                                         thread_main, td);
2254                                 if (ret) {
2255                                         log_err("pthread_create: %s\n",
2256                                                         strerror(ret));
2257                                         nr_started--;
2258                                         break;
2259                                 }
2260                                 ret = pthread_detach(td->thread);
2261                                 if (ret)
2262                                         log_err("pthread_detach: %s",
2263                                                         strerror(ret));
2264                         } else {
2265                                 pid_t pid;
2266                                 dprint(FD_PROCESS, "will fork\n");
2267                                 pid = fork();
2268                                 if (!pid) {
2269                                         int ret = fork_main(shm_id, i);
2270
2271                                         _exit(ret);
2272                                 } else if (i == fio_debug_jobno)
2273                                         *fio_debug_jobp = pid;
2274                         }
2275                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2276                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2277                                 log_err("fio: job startup hung? exiting.\n");
2278                                 fio_terminate_threads(TERMINATE_ALL);
2279                                 fio_abort = 1;
2280                                 nr_started--;
2281                                 break;
2282                         }
2283                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2284                 }
2285
2286                 /*
2287                  * Wait for the started threads to transition to
2288                  * TD_INITIALIZED.
2289                  */
2290                 fio_gettime(&this_start, NULL);
2291                 left = this_jobs;
2292                 while (left && !fio_abort) {
2293                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2294                                 break;
2295
2296                         do_usleep(100000);
2297
2298                         for (i = 0; i < this_jobs; i++) {
2299                                 td = map[i];
2300                                 if (!td)
2301                                         continue;
2302                                 if (td->runstate == TD_INITIALIZED) {
2303                                         map[i] = NULL;
2304                                         left--;
2305                                 } else if (td->runstate >= TD_EXITED) {
2306                                         map[i] = NULL;
2307                                         left--;
2308                                         todo--;
2309                                         nr_running++; /* work-around... */
2310                                 }
2311                         }
2312                 }
2313
2314                 if (left) {
2315                         log_err("fio: %d job%s failed to start\n", left,
2316                                         left > 1 ? "s" : "");
2317                         for (i = 0; i < this_jobs; i++) {
2318                                 td = map[i];
2319                                 if (!td)
2320                                         continue;
2321                                 kill(td->pid, SIGTERM);
2322                         }
2323                         break;
2324                 }
2325
2326                 /*
2327                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2328                  */
2329                 for_each_td(td, i) {
2330                         if (td->runstate != TD_INITIALIZED)
2331                                 continue;
2332
2333                         if (in_ramp_time(td))
2334                                 td_set_runstate(td, TD_RAMP);
2335                         else
2336                                 td_set_runstate(td, TD_RUNNING);
2337                         nr_running++;
2338                         nr_started--;
2339                         m_rate += ddir_rw_sum(td->o.ratemin);
2340                         t_rate += ddir_rw_sum(td->o.rate);
2341                         todo--;
2342                         fio_mutex_up(td->mutex);
2343                 }
2344
2345                 reap_threads(&nr_running, &t_rate, &m_rate);
2346
2347                 if (todo)
2348                         do_usleep(100000);
2349         }
2350
2351         while (nr_running) {
2352                 reap_threads(&nr_running, &t_rate, &m_rate);
2353                 do_usleep(10000);
2354         }
2355
2356         fio_idle_prof_stop();
2357
2358         update_io_ticks();
2359 }
2360
2361 static void wait_for_helper_thread_exit(void)
2362 {
2363         void *ret;
2364
2365         helper_exit = 1;
2366         pthread_cond_signal(&helper_cond);
2367         pthread_join(helper_thread, &ret);
2368 }
2369
2370 static void free_disk_util(void)
2371 {
2372         disk_util_prune_entries();
2373
2374         pthread_cond_destroy(&helper_cond);
2375 }
2376
2377 static void *helper_thread_main(void *data)
2378 {
2379         int ret = 0;
2380
2381         fio_mutex_up(startup_mutex);
2382
2383         while (!ret) {
2384                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2385                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2386                 struct timespec ts;
2387                 struct timeval tv;
2388
2389                 gettimeofday(&tv, NULL);
2390                 ts.tv_sec = tv.tv_sec + sec;
2391                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2392
2393                 if (ts.tv_nsec >= 1000000000ULL) {
2394                         ts.tv_nsec -= 1000000000ULL;
2395                         ts.tv_sec++;
2396                 }
2397
2398                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2399
2400                 ret = update_io_ticks();
2401
2402                 if (helper_do_stat) {
2403                         helper_do_stat = 0;
2404                         __show_running_run_stats();
2405                 }
2406
2407                 if (!is_backend)
2408                         print_thread_status();
2409         }
2410
2411         return NULL;
2412 }
2413
2414 static int create_helper_thread(void)
2415 {
2416         int ret;
2417
2418         setup_disk_util();
2419
2420         pthread_cond_init(&helper_cond, NULL);
2421         pthread_mutex_init(&helper_lock, NULL);
2422
2423         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2424         if (ret) {
2425                 log_err("Can't create helper thread: %s\n", strerror(ret));
2426                 return 1;
2427         }
2428
2429         dprint(FD_MUTEX, "wait on startup_mutex\n");
2430         fio_mutex_down(startup_mutex);
2431         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2432         return 0;
2433 }
2434
2435 int fio_backend(void)
2436 {
2437         struct thread_data *td;
2438         int i;
2439
2440         if (exec_profile) {
2441                 if (load_profile(exec_profile))
2442                         return 1;
2443                 free(exec_profile);
2444                 exec_profile = NULL;
2445         }
2446         if (!thread_number)
2447                 return 0;
2448
2449         if (write_bw_log) {
2450                 struct log_params p = {
2451                         .log_type = IO_LOG_TYPE_BW,
2452                 };
2453
2454                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2455                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2456                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2457         }
2458
2459         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2460         if (startup_mutex == NULL)
2461                 return 1;
2462
2463         set_genesis_time();
2464         stat_init();
2465         create_helper_thread();
2466
2467         cgroup_list = smalloc(sizeof(*cgroup_list));
2468         INIT_FLIST_HEAD(cgroup_list);
2469
2470         run_threads();
2471
2472         wait_for_helper_thread_exit();
2473
2474         if (!fio_abort) {
2475                 __show_run_stats();
2476                 if (write_bw_log) {
2477                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2478                                 struct io_log *log = agg_io_log[i];
2479
2480                                 flush_log(log, 0);
2481                                 free_log(log);
2482                         }
2483                 }
2484         }
2485
2486         for_each_td(td, i) {
2487                 fio_options_free(td);
2488                 if (td->rusage_sem) {
2489                         fio_mutex_remove(td->rusage_sem);
2490                         td->rusage_sem = NULL;
2491                 }
2492         }
2493
2494         free_disk_util();
2495         cgroup_kill(cgroup_list);
2496         sfree(cgroup_list);
2497         sfree(cgroup_mnt);
2498
2499         fio_mutex_remove(startup_mutex);
2500         stat_exit();
2501         return exit_value;
2502 }