backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38 #include <math.h>
  39
  40 #include "fio.h"
  41 #ifndef FIO_NO_HAVE_SHM_H
  42 #include <sys/shm.h>
  43 #endif
  44 #include "hash.h"
  45 #include "smalloc.h"
  46 #include "verify.h"
  47 #include "trim.h"
  48 #include "diskutil.h"
  49 #include "cgroup.h"
  50 #include "profile.h"
  51 #include "lib/rand.h"
  52 #include "lib/memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59 #include "rate-submit.h"
  60 #include "helper_thread.h"
  61
  62 static struct fio_mutex *startup_mutex;
  63 static struct flist_head *cgroup_list;
  64 static char *cgroup_mnt;
  65 static int exit_value;
  66 static volatile int fio_abort;
  67 static unsigned int nr_process = 0;
  68 static unsigned int nr_thread = 0;
  69
  70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  71
  72 int groupid = 0;
  73 unsigned int thread_number = 0;
  74 unsigned int stat_number = 0;
  75 int shm_id = 0;
  76 int temp_stall_ts;
  77 unsigned long done_secs = 0;
  78
  79 #define JOB_START_TIMEOUT       (5 * 1000)
  80
  81 static void sig_int(int sig)
  82 {
  83         if (threads) {
  84                 if (is_backend)
  85                         fio_server_got_signal(sig);
  86                 else {
  87                         log_info("\nfio: terminating on signal %d\n", sig);
  88                         log_info_flush();
  89                         exit_value = 128;
  90                 }
  91
  92                 fio_terminate_threads(TERMINATE_ALL);
  93         }
  94 }
  95
  96 void sig_show_status(int sig)
  97 {
  98         show_running_run_stats();
  99 }
 100
 101 static void set_sig_handlers(void)
 102 {
 103         struct sigaction act;
 104
 105         memset(&act, 0, sizeof(act));
 106         act.sa_handler = sig_int;
 107         act.sa_flags = SA_RESTART;
 108         sigaction(SIGINT, &act, NULL);
 109
 110         memset(&act, 0, sizeof(act));
 111         act.sa_handler = sig_int;
 112         act.sa_flags = SA_RESTART;
 113         sigaction(SIGTERM, &act, NULL);
 114
 115 /* Windows uses SIGBREAK as a quit signal from other applications */
 116 #ifdef WIN32
 117         memset(&act, 0, sizeof(act));
 118         act.sa_handler = sig_int;
 119         act.sa_flags = SA_RESTART;
 120         sigaction(SIGBREAK, &act, NULL);
 121 #endif
 122
 123         memset(&act, 0, sizeof(act));
 124         act.sa_handler = sig_show_status;
 125         act.sa_flags = SA_RESTART;
 126         sigaction(SIGUSR1, &act, NULL);
 127
 128         if (is_backend) {
 129                 memset(&act, 0, sizeof(act));
 130                 act.sa_handler = sig_int;
 131                 act.sa_flags = SA_RESTART;
 132                 sigaction(SIGPIPE, &act, NULL);
 133         }
 134 }
 135
 136 /*
 137  * Check if we are above the minimum rate given.
 138  */
 139 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 140                              enum fio_ddir ddir)
 141 {
 142         unsigned long long bytes = 0;
 143         unsigned long iops = 0;
 144         unsigned long spent;
 145         unsigned long rate;
 146         unsigned int ratemin = 0;
 147         unsigned int rate_iops = 0;
 148         unsigned int rate_iops_min = 0;
 149
 150         assert(ddir_rw(ddir));
 151
 152         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 153                 return false;
 154
 155         /*
 156          * allow a 2 second settle period in the beginning
 157          */
 158         if (mtime_since(&td->start, now) < 2000)
 159                 return false;
 160
 161         iops += td->this_io_blocks[ddir];
 162         bytes += td->this_io_bytes[ddir];
 163         ratemin += td->o.ratemin[ddir];
 164         rate_iops += td->o.rate_iops[ddir];
 165         rate_iops_min += td->o.rate_iops_min[ddir];
 166
 167         /*
 168          * if rate blocks is set, sample is running
 169          */
 170         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 171                 spent = mtime_since(&td->lastrate[ddir], now);
 172                 if (spent < td->o.ratecycle)
 173                         return false;
 174
 175                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 176                         /*
 177                          * check bandwidth specified rate
 178                          */
 179                         if (bytes < td->rate_bytes[ddir]) {
 180                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 181                                         td->o.name, ratemin, bytes);
 182                                 return true;
 183                         } else {
 184                                 if (spent)
 185                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 186                                 else
 187                                         rate = 0;
 188
 189                                 if (rate < ratemin ||
 190                                     bytes < td->rate_bytes[ddir]) {
 191                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 192                                                 td->o.name, ratemin, rate);
 193                                         return true;
 194                                 }
 195                         }
 196                 } else {
 197                         /*
 198                          * checks iops specified rate
 199                          */
 200                         if (iops < rate_iops) {
 201                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 202                                                 td->o.name, rate_iops, iops);
 203                                 return true;
 204                         } else {
 205                                 if (spent)
 206                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 207                                 else
 208                                         rate = 0;
 209
 210                                 if (rate < rate_iops_min ||
 211                                     iops < td->rate_blocks[ddir]) {
 212                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 213                                                 td->o.name, rate_iops_min, rate);
 214                                         return true;
 215                                 }
 216                         }
 217                 }
 218         }
 219
 220         td->rate_bytes[ddir] = bytes;
 221         td->rate_blocks[ddir] = iops;
 222         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 223         return false;
 224 }
 225
 226 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 227 {
 228         bool ret = false;
 229
 230         if (td->bytes_done[DDIR_READ])
 231                 ret |= __check_min_rate(td, now, DDIR_READ);
 232         if (td->bytes_done[DDIR_WRITE])
 233                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 234         if (td->bytes_done[DDIR_TRIM])
 235                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 236
 237         return ret;
 238 }
 239
 240 /*
 241  * When job exits, we can cancel the in-flight IO if we are using async
 242  * io. Attempt to do so.
 243  */
 244 static void cleanup_pending_aio(struct thread_data *td)
 245 {
 246         int r;
 247
 248         /*
 249          * get immediately available events, if any
 250          */
 251         r = io_u_queued_complete(td, 0);
 252         if (r < 0)
 253                 return;
 254
 255         /*
 256          * now cancel remaining active events
 257          */
 258         if (td->io_ops->cancel) {
 259                 struct io_u *io_u;
 260                 int i;
 261
 262                 io_u_qiter(&td->io_u_all, io_u, i) {
 263                         if (io_u->flags & IO_U_F_FLIGHT) {
 264                                 r = td->io_ops->cancel(td, io_u);
 265                                 if (!r)
 266                                         put_io_u(td, io_u);
 267                         }
 268                 }
 269         }
 270
 271         if (td->cur_depth)
 272                 r = io_u_queued_complete(td, td->cur_depth);
 273 }
 274
 275 /*
 276  * Helper to handle the final sync of a file. Works just like the normal
 277  * io path, just does everything sync.
 278  */
 279 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 280 {
 281         struct io_u *io_u = __get_io_u(td);
 282         int ret;
 283
 284         if (!io_u)
 285                 return true;
 286
 287         io_u->ddir = DDIR_SYNC;
 288         io_u->file = f;
 289
 290         if (td_io_prep(td, io_u)) {
 291                 put_io_u(td, io_u);
 292                 return true;
 293         }
 294
 295 requeue:
 296         ret = td_io_queue(td, io_u);
 297         if (ret < 0) {
 298                 td_verror(td, io_u->error, "td_io_queue");
 299                 put_io_u(td, io_u);
 300                 return true;
 301         } else if (ret == FIO_Q_QUEUED) {
 302                 if (td_io_commit(td))
 303                         return true;
 304                 if (io_u_queued_complete(td, 1) < 0)
 305                         return true;
 306         } else if (ret == FIO_Q_COMPLETED) {
 307                 if (io_u->error) {
 308                         td_verror(td, io_u->error, "td_io_queue");
 309                         return true;
 310                 }
 311
 312                 if (io_u_sync_complete(td, io_u) < 0)
 313                         return true;
 314         } else if (ret == FIO_Q_BUSY) {
 315                 if (td_io_commit(td))
 316                         return true;
 317                 goto requeue;
 318         }
 319
 320         return false;
 321 }
 322
 323 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 324 {
 325         int ret;
 326
 327         if (fio_file_open(f))
 328                 return fio_io_sync(td, f);
 329
 330         if (td_io_open_file(td, f))
 331                 return 1;
 332
 333         ret = fio_io_sync(td, f);
 334         td_io_close_file(td, f);
 335         return ret;
 336 }
 337
 338 static inline void __update_ts_cache(struct thread_data *td)
 339 {
 340         fio_gettime(&td->ts_cache, NULL);
 341 }
 342
 343 static inline void update_ts_cache(struct thread_data *td)
 344 {
 345         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 346                 __update_ts_cache(td);
 347 }
 348
 349 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 350 {
 351         if (in_ramp_time(td))
 352                 return false;
 353         if (!td->o.timeout)
 354                 return false;
 355         if (utime_since(&td->epoch, t) >= td->o.timeout)
 356                 return true;
 357
 358         return false;
 359 }
 360
 361 /*
 362  * We need to update the runtime consistently in ms, but keep a running
 363  * tally of the current elapsed time in microseconds for sub millisecond
 364  * updates.
 365  */
 366 static inline void update_runtime(struct thread_data *td,
 367                                   unsigned long long *elapsed_us,
 368                                   const enum fio_ddir ddir)
 369 {
 370         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 371                 return;
 372
 373         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 374         elapsed_us[ddir] += utime_since_now(&td->start);
 375         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 376 }
 377
 378 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 379                                 int *retptr)
 380 {
 381         int ret = *retptr;
 382
 383         if (ret < 0 || td->error) {
 384                 int err = td->error;
 385                 enum error_type_bit eb;
 386
 387                 if (ret < 0)
 388                         err = -ret;
 389
 390                 eb = td_error_type(ddir, err);
 391                 if (!(td->o.continue_on_error & (1 << eb)))
 392                         return true;
 393
 394                 if (td_non_fatal_error(td, eb, err)) {
 395                         /*
 396                          * Continue with the I/Os in case of
 397                          * a non fatal error.
 398                          */
 399                         update_error_count(td, err);
 400                         td_clear_error(td);
 401                         *retptr = 0;
 402                         return false;
 403                 } else if (td->o.fill_device && err == ENOSPC) {
 404                         /*
 405                          * We expect to hit this error if
 406                          * fill_device option is set.
 407                          */
 408                         td_clear_error(td);
 409                         fio_mark_td_terminate(td);
 410                         return true;
 411                 } else {
 412                         /*
 413                          * Stop the I/O in case of a fatal
 414                          * error.
 415                          */
 416                         update_error_count(td, err);
 417                         return true;
 418                 }
 419         }
 420
 421         return false;
 422 }
 423
 424 static void check_update_rusage(struct thread_data *td)
 425 {
 426         if (td->update_rusage) {
 427                 td->update_rusage = 0;
 428                 update_rusage_stat(td);
 429                 fio_mutex_up(td->rusage_sem);
 430         }
 431 }
 432
 433 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 434 {
 435         const int full = queue_full(td);
 436         int min_evts = 0;
 437         int ret;
 438
 439         if (td->flags & TD_F_REGROW_LOGS)
 440                 return io_u_quiesce(td);
 441
 442         /*
 443          * if the queue is full, we MUST reap at least 1 event
 444          */
 445         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 446         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 447                 min_evts = 1;
 448
 449         if (time && (__should_check_rate(td, DDIR_READ) ||
 450             __should_check_rate(td, DDIR_WRITE) ||
 451             __should_check_rate(td, DDIR_TRIM)))
 452                 fio_gettime(time, NULL);
 453
 454         do {
 455                 ret = io_u_queued_complete(td, min_evts);
 456                 if (ret < 0)
 457                         break;
 458         } while (full && (td->cur_depth > td->o.iodepth_low));
 459
 460         return ret;
 461 }
 462
 463 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 464                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 465                    struct timespec *comp_time)
 466 {
 467         int ret2;
 468
 469         switch (*ret) {
 470         case FIO_Q_COMPLETED:
 471                 if (io_u->error) {
 472                         *ret = -io_u->error;
 473                         clear_io_u(td, io_u);
 474                 } else if (io_u->resid) {
 475                         int bytes = io_u->xfer_buflen - io_u->resid;
 476                         struct fio_file *f = io_u->file;
 477
 478                         if (bytes_issued)
 479                                 *bytes_issued += bytes;
 480
 481                         if (!from_verify)
 482                                 trim_io_piece(td, io_u);
 483
 484                         /*
 485                          * zero read, fail
 486                          */
 487                         if (!bytes) {
 488                                 if (!from_verify)
 489                                         unlog_io_piece(td, io_u);
 490                                 td_verror(td, EIO, "full resid");
 491                                 put_io_u(td, io_u);
 492                                 break;
 493                         }
 494
 495                         io_u->xfer_buflen = io_u->resid;
 496                         io_u->xfer_buf += bytes;
 497                         io_u->offset += bytes;
 498
 499                         if (ddir_rw(io_u->ddir))
 500                                 td->ts.short_io_u[io_u->ddir]++;
 501
 502                         if (io_u->offset == f->real_file_size)
 503                                 goto sync_done;
 504
 505                         requeue_io_u(td, &io_u);
 506                 } else {
 507 sync_done:
 508                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 509                             __should_check_rate(td, DDIR_WRITE) ||
 510                             __should_check_rate(td, DDIR_TRIM)))
 511                                 fio_gettime(comp_time, NULL);
 512
 513                         *ret = io_u_sync_complete(td, io_u);
 514                         if (*ret < 0)
 515                                 break;
 516                 }
 517
 518                 if (td->flags & TD_F_REGROW_LOGS)
 519                         regrow_logs(td);
 520
 521                 /*
 522                  * when doing I/O (not when verifying),
 523                  * check for any errors that are to be ignored
 524                  */
 525                 if (!from_verify)
 526                         break;
 527
 528                 return 0;
 529         case FIO_Q_QUEUED:
 530                 /*
 531                  * if the engine doesn't have a commit hook,
 532                  * the io_u is really queued. if it does have such
 533                  * a hook, it has to call io_u_queued() itself.
 534                  */
 535                 if (td->io_ops->commit == NULL)
 536                         io_u_queued(td, io_u);
 537                 if (bytes_issued)
 538                         *bytes_issued += io_u->xfer_buflen;
 539                 break;
 540         case FIO_Q_BUSY:
 541                 if (!from_verify)
 542                         unlog_io_piece(td, io_u);
 543                 requeue_io_u(td, &io_u);
 544                 ret2 = td_io_commit(td);
 545                 if (ret2 < 0)
 546                         *ret = ret2;
 547                 break;
 548         default:
 549                 assert(*ret < 0);
 550                 td_verror(td, -(*ret), "td_io_queue");
 551                 break;
 552         }
 553
 554         if (break_on_this_error(td, ddir, ret))
 555                 return 1;
 556
 557         return 0;
 558 }
 559
 560 static inline bool io_in_polling(struct thread_data *td)
 561 {
 562         return !td->o.iodepth_batch_complete_min &&
 563                    !td->o.iodepth_batch_complete_max;
 564 }
 565 /*
 566  * Unlinks files from thread data fio_file structure
 567  */
 568 static int unlink_all_files(struct thread_data *td)
 569 {
 570         struct fio_file *f;
 571         unsigned int i;
 572         int ret = 0;
 573
 574         for_each_file(td, f, i) {
 575                 if (f->filetype != FIO_TYPE_FILE)
 576                         continue;
 577                 ret = td_io_unlink_file(td, f);
 578                 if (ret)
 579                         break;
 580         }
 581
 582         if (ret)
 583                 td_verror(td, ret, "unlink_all_files");
 584
 585         return ret;
 586 }
 587
 588 /*
 589  * Check if io_u will overlap an in-flight IO in the queue
 590  */
 591 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 592 {
 593         bool overlap;
 594         struct io_u *check_io_u;
 595         unsigned long long x1, x2, y1, y2;
 596         int i;
 597
 598         x1 = io_u->offset;
 599         x2 = io_u->offset + io_u->buflen;
 600         overlap = false;
 601         io_u_qiter(q, check_io_u, i) {
 602                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 603                         y1 = check_io_u->offset;
 604                         y2 = check_io_u->offset + check_io_u->buflen;
 605
 606                         if (x1 < y2 && y1 < x2) {
 607                                 overlap = true;
 608                                 dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
 609                                                 x1, io_u->buflen,
 610                                                 y1, check_io_u->buflen);
 611                                 break;
 612                         }
 613                 }
 614         }
 615
 616         return overlap;
 617 }
 618
 619 static int io_u_submit(struct thread_data *td, struct io_u *io_u)
 620 {
 621         /*
 622          * Check for overlap if the user asked us to, and we have
 623          * at least one IO in flight besides this one.
 624          */
 625         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 626             in_flight_overlap(&td->io_u_all, io_u))
 627                 return FIO_Q_BUSY;
 628
 629         return td_io_queue(td, io_u);
 630 }
 631
 632 /*
 633  * The main verify engine. Runs over the writes we previously submitted,
 634  * reads the blocks back in, and checks the crc/md5 of the data.
 635  */
 636 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 637 {
 638         struct fio_file *f;
 639         struct io_u *io_u;
 640         int ret, min_events;
 641         unsigned int i;
 642
 643         dprint(FD_VERIFY, "starting loop\n");
 644
 645         /*
 646          * sync io first and invalidate cache, to make sure we really
 647          * read from disk.
 648          */
 649         for_each_file(td, f, i) {
 650                 if (!fio_file_open(f))
 651                         continue;
 652                 if (fio_io_sync(td, f))
 653                         break;
 654                 if (file_invalidate_cache(td, f))
 655                         break;
 656         }
 657
 658         check_update_rusage(td);
 659
 660         if (td->error)
 661                 return;
 662
 663         /*
 664          * verify_state needs to be reset before verification
 665          * proceeds so that expected random seeds match actual
 666          * random seeds in headers. The main loop will reset
 667          * all random number generators if randrepeat is set.
 668          */
 669         if (!td->o.rand_repeatable)
 670                 td_fill_verify_state_seed(td);
 671
 672         td_set_runstate(td, TD_VERIFYING);
 673
 674         io_u = NULL;
 675         while (!td->terminate) {
 676                 enum fio_ddir ddir;
 677                 int full;
 678
 679                 update_ts_cache(td);
 680                 check_update_rusage(td);
 681
 682                 if (runtime_exceeded(td, &td->ts_cache)) {
 683                         __update_ts_cache(td);
 684                         if (runtime_exceeded(td, &td->ts_cache)) {
 685                                 fio_mark_td_terminate(td);
 686                                 break;
 687                         }
 688                 }
 689
 690                 if (flow_threshold_exceeded(td))
 691                         continue;
 692
 693                 if (!td->o.experimental_verify) {
 694                         io_u = __get_io_u(td);
 695                         if (!io_u)
 696                                 break;
 697
 698                         if (get_next_verify(td, io_u)) {
 699                                 put_io_u(td, io_u);
 700                                 break;
 701                         }
 702
 703                         if (td_io_prep(td, io_u)) {
 704                                 put_io_u(td, io_u);
 705                                 break;
 706                         }
 707                 } else {
 708                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 709                                 break;
 710
 711                         while ((io_u = get_io_u(td)) != NULL) {
 712                                 if (IS_ERR_OR_NULL(io_u)) {
 713                                         io_u = NULL;
 714                                         ret = FIO_Q_BUSY;
 715                                         goto reap;
 716                                 }
 717
 718                                 /*
 719                                  * We are only interested in the places where
 720                                  * we wrote or trimmed IOs. Turn those into
 721                                  * reads for verification purposes.
 722                                  */
 723                                 if (io_u->ddir == DDIR_READ) {
 724                                         /*
 725                                          * Pretend we issued it for rwmix
 726                                          * accounting
 727                                          */
 728                                         td->io_issues[DDIR_READ]++;
 729                                         put_io_u(td, io_u);
 730                                         continue;
 731                                 } else if (io_u->ddir == DDIR_TRIM) {
 732                                         io_u->ddir = DDIR_READ;
 733                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 734                                         break;
 735                                 } else if (io_u->ddir == DDIR_WRITE) {
 736                                         io_u->ddir = DDIR_READ;
 737                                         break;
 738                                 } else {
 739                                         put_io_u(td, io_u);
 740                                         continue;
 741                                 }
 742                         }
 743
 744                         if (!io_u)
 745                                 break;
 746                 }
 747
 748                 if (verify_state_should_stop(td, io_u)) {
 749                         put_io_u(td, io_u);
 750                         break;
 751                 }
 752
 753                 if (td->o.verify_async)
 754                         io_u->end_io = verify_io_u_async;
 755                 else
 756                         io_u->end_io = verify_io_u;
 757
 758                 ddir = io_u->ddir;
 759                 if (!td->o.disable_slat)
 760                         fio_gettime(&io_u->start_time, NULL);
 761
 762                 ret = io_u_submit(td, io_u);
 763
 764                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 765                         break;
 766
 767                 /*
 768                  * if we can queue more, do so. but check if there are
 769                  * completed io_u's first. Note that we can get BUSY even
 770                  * without IO queued, if the system is resource starved.
 771                  */
 772 reap:
 773                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 774                 if (full || io_in_polling(td))
 775                         ret = wait_for_completions(td, NULL);
 776
 777                 if (ret < 0)
 778                         break;
 779         }
 780
 781         check_update_rusage(td);
 782
 783         if (!td->error) {
 784                 min_events = td->cur_depth;
 785
 786                 if (min_events)
 787                         ret = io_u_queued_complete(td, min_events);
 788         } else
 789                 cleanup_pending_aio(td);
 790
 791         td_set_runstate(td, TD_RUNNING);
 792
 793         dprint(FD_VERIFY, "exiting loop\n");
 794 }
 795
 796 static bool exceeds_number_ios(struct thread_data *td)
 797 {
 798         unsigned long long number_ios;
 799
 800         if (!td->o.number_ios)
 801                 return false;
 802
 803         number_ios = ddir_rw_sum(td->io_blocks);
 804         number_ios += td->io_u_queued + td->io_u_in_flight;
 805
 806         return number_ios >= (td->o.number_ios * td->loops);
 807 }
 808
 809 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 810 {
 811         unsigned long long bytes, limit;
 812
 813         if (td_rw(td))
 814                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 815         else if (td_write(td))
 816                 bytes = this_bytes[DDIR_WRITE];
 817         else if (td_read(td))
 818                 bytes = this_bytes[DDIR_READ];
 819         else
 820                 bytes = this_bytes[DDIR_TRIM];
 821
 822         if (td->o.io_size)
 823                 limit = td->o.io_size;
 824         else
 825                 limit = td->o.size;
 826
 827         limit *= td->loops;
 828         return bytes >= limit || exceeds_number_ios(td);
 829 }
 830
 831 static bool io_issue_bytes_exceeded(struct thread_data *td)
 832 {
 833         return io_bytes_exceeded(td, td->io_issue_bytes);
 834 }
 835
 836 static bool io_complete_bytes_exceeded(struct thread_data *td)
 837 {
 838         return io_bytes_exceeded(td, td->this_io_bytes);
 839 }
 840
 841 /*
 842  * used to calculate the next io time for rate control
 843  *
 844  */
 845 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 846 {
 847         uint64_t secs, remainder, bps, bytes, iops;
 848
 849         assert(!(td->flags & TD_F_CHILD));
 850         bytes = td->rate_io_issue_bytes[ddir];
 851         bps = td->rate_bps[ddir];
 852
 853         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 854                 uint64_t val;
 855                 iops = bps / td->o.bs[ddir];
 856                 val = (int64_t) (1000000 / iops) *
 857                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 858                 if (val) {
 859                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 860                                         (unsigned long long) 1000000 / val,
 861                                         ddir);
 862                 }
 863                 td->last_usec[ddir] += val;
 864                 return td->last_usec[ddir];
 865         } else if (bps) {
 866                 secs = bytes / bps;
 867                 remainder = bytes % bps;
 868                 return remainder * 1000000 / bps + secs * 1000000;
 869         }
 870
 871         return 0;
 872 }
 873
 874 /*
 875  * Main IO worker function. It retrieves io_u's to process and queues
 876  * and reaps them, checking for rate and errors along the way.
 877  *
 878  * Returns number of bytes written and trimmed.
 879  */
 880 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 881 {
 882         unsigned int i;
 883         int ret = 0;
 884         uint64_t total_bytes, bytes_issued = 0;
 885
 886         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 887                 bytes_done[i] = td->bytes_done[i];
 888
 889         if (in_ramp_time(td))
 890                 td_set_runstate(td, TD_RAMP);
 891         else
 892                 td_set_runstate(td, TD_RUNNING);
 893
 894         lat_target_init(td);
 895
 896         total_bytes = td->o.size;
 897         /*
 898         * Allow random overwrite workloads to write up to io_size
 899         * before starting verification phase as 'size' doesn't apply.
 900         */
 901         if (td_write(td) && td_random(td) && td->o.norandommap)
 902                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 903         /*
 904          * If verify_backlog is enabled, we'll run the verify in this
 905          * handler as well. For that case, we may need up to twice the
 906          * amount of bytes.
 907          */
 908         if (td->o.verify != VERIFY_NONE &&
 909            (td_write(td) && td->o.verify_backlog))
 910                 total_bytes += td->o.size;
 911
 912         /* In trimwrite mode, each byte is trimmed and then written, so
 913          * allow total_bytes to be twice as big */
 914         if (td_trimwrite(td))
 915                 total_bytes += td->total_io_size;
 916
 917         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 918                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 919                 td->o.time_based) {
 920                 struct timespec comp_time;
 921                 struct io_u *io_u;
 922                 int full;
 923                 enum fio_ddir ddir;
 924
 925                 check_update_rusage(td);
 926
 927                 if (td->terminate || td->done)
 928                         break;
 929
 930                 update_ts_cache(td);
 931
 932                 if (runtime_exceeded(td, &td->ts_cache)) {
 933                         __update_ts_cache(td);
 934                         if (runtime_exceeded(td, &td->ts_cache)) {
 935                                 fio_mark_td_terminate(td);
 936                                 break;
 937                         }
 938                 }
 939
 940                 if (flow_threshold_exceeded(td))
 941                         continue;
 942
 943                 /*
 944                  * Break if we exceeded the bytes. The exception is time
 945                  * based runs, but we still need to break out of the loop
 946                  * for those to run verification, if enabled.
 947                  */
 948                 if (bytes_issued >= total_bytes &&
 949                     (!td->o.time_based ||
 950                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 951                         break;
 952
 953                 io_u = get_io_u(td);
 954                 if (IS_ERR_OR_NULL(io_u)) {
 955                         int err = PTR_ERR(io_u);
 956
 957                         io_u = NULL;
 958                         if (err == -EBUSY) {
 959                                 ret = FIO_Q_BUSY;
 960                                 goto reap;
 961                         }
 962                         if (td->o.latency_target)
 963                                 goto reap;
 964                         break;
 965                 }
 966
 967                 ddir = io_u->ddir;
 968
 969                 /*
 970                  * Add verification end_io handler if:
 971                  *      - Asked to verify (!td_rw(td))
 972                  *      - Or the io_u is from our verify list (mixed write/ver)
 973                  */
 974                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 975                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 976
 977                         if (!td->o.verify_pattern_bytes) {
 978                                 io_u->rand_seed = __rand(&td->verify_state);
 979                                 if (sizeof(int) != sizeof(long *))
 980                                         io_u->rand_seed *= __rand(&td->verify_state);
 981                         }
 982
 983                         if (verify_state_should_stop(td, io_u)) {
 984                                 put_io_u(td, io_u);
 985                                 break;
 986                         }
 987
 988                         if (td->o.verify_async)
 989                                 io_u->end_io = verify_io_u_async;
 990                         else
 991                                 io_u->end_io = verify_io_u;
 992                         td_set_runstate(td, TD_VERIFYING);
 993                 } else if (in_ramp_time(td))
 994                         td_set_runstate(td, TD_RAMP);
 995                 else
 996                         td_set_runstate(td, TD_RUNNING);
 997
 998                 /*
 999                  * Always log IO before it's issued, so we know the specific
1000                  * order of it. The logged unit will track when the IO has
1001                  * completed.
1002                  */
1003                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1004                     td->o.do_verify &&
1005                     td->o.verify != VERIFY_NONE &&
1006                     !td->o.experimental_verify)
1007                         log_io_piece(td, io_u);
1008
1009                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1010                         const unsigned long blen = io_u->xfer_buflen;
1011                         const enum fio_ddir ddir = acct_ddir(io_u);
1012
1013                         if (td->error)
1014                                 break;
1015
1016                         workqueue_enqueue(&td->io_wq, &io_u->work);
1017                         ret = FIO_Q_QUEUED;
1018
1019                         if (ddir_rw(ddir)) {
1020                                 td->io_issues[ddir]++;
1021                                 td->io_issue_bytes[ddir] += blen;
1022                                 td->rate_io_issue_bytes[ddir] += blen;
1023                         }
1024
1025                         if (should_check_rate(td))
1026                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1027
1028                 } else {
1029                         ret = io_u_submit(td, io_u);
1030
1031                         if (should_check_rate(td))
1032                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1033
1034                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1035                                 break;
1036
1037                         /*
1038                          * See if we need to complete some commands. Note that
1039                          * we can get BUSY even without IO queued, if the
1040                          * system is resource starved.
1041                          */
1042 reap:
1043                         full = queue_full(td) ||
1044                                 (ret == FIO_Q_BUSY && td->cur_depth);
1045                         if (full || io_in_polling(td))
1046                                 ret = wait_for_completions(td, &comp_time);
1047                 }
1048                 if (ret < 0)
1049                         break;
1050                 if (!ddir_rw_sum(td->bytes_done) &&
1051                     !td_ioengine_flagged(td, FIO_NOIO))
1052                         continue;
1053
1054                 if (!in_ramp_time(td) && should_check_rate(td)) {
1055                         if (check_min_rate(td, &comp_time)) {
1056                                 if (exitall_on_terminate || td->o.exitall_error)
1057                                         fio_terminate_threads(td->groupid);
1058                                 td_verror(td, EIO, "check_min_rate");
1059                                 break;
1060                         }
1061                 }
1062                 if (!in_ramp_time(td) && td->o.latency_target)
1063                         lat_target_check(td);
1064
1065                 if (td->o.thinktime) {
1066                         unsigned long long b;
1067
1068                         b = ddir_rw_sum(td->io_blocks);
1069                         if (!(b % td->o.thinktime_blocks)) {
1070                                 int left;
1071
1072                                 io_u_quiesce(td);
1073
1074                                 if (td->o.thinktime_spin)
1075                                         usec_spin(td->o.thinktime_spin);
1076
1077                                 left = td->o.thinktime - td->o.thinktime_spin;
1078                                 if (left)
1079                                         usec_sleep(td, left);
1080                         }
1081                 }
1082         }
1083
1084         check_update_rusage(td);
1085
1086         if (td->trim_entries)
1087                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1088
1089         if (td->o.fill_device && td->error == ENOSPC) {
1090                 td->error = 0;
1091                 fio_mark_td_terminate(td);
1092         }
1093         if (!td->error) {
1094                 struct fio_file *f;
1095
1096                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1097                         workqueue_flush(&td->io_wq);
1098                         i = 0;
1099                 } else
1100                         i = td->cur_depth;
1101
1102                 if (i) {
1103                         ret = io_u_queued_complete(td, i);
1104                         if (td->o.fill_device && td->error == ENOSPC)
1105                                 td->error = 0;
1106                 }
1107
1108                 if (should_fsync(td) && td->o.end_fsync) {
1109                         td_set_runstate(td, TD_FSYNCING);
1110
1111                         for_each_file(td, f, i) {
1112                                 if (!fio_file_fsync(td, f))
1113                                         continue;
1114
1115                                 log_err("fio: end_fsync failed for file %s\n",
1116                                                                 f->file_name);
1117                         }
1118                 }
1119         } else
1120                 cleanup_pending_aio(td);
1121
1122         /*
1123          * stop job if we failed doing any IO
1124          */
1125         if (!ddir_rw_sum(td->this_io_bytes))
1126                 td->done = 1;
1127
1128         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1129                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1130 }
1131
1132 static void free_file_completion_logging(struct thread_data *td)
1133 {
1134         struct fio_file *f;
1135         unsigned int i;
1136
1137         for_each_file(td, f, i) {
1138                 if (!f->last_write_comp)
1139                         break;
1140                 sfree(f->last_write_comp);
1141         }
1142 }
1143
1144 static int init_file_completion_logging(struct thread_data *td,
1145                                         unsigned int depth)
1146 {
1147         struct fio_file *f;
1148         unsigned int i;
1149
1150         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1151                 return 0;
1152
1153         for_each_file(td, f, i) {
1154                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1155                 if (!f->last_write_comp)
1156                         goto cleanup;
1157         }
1158
1159         return 0;
1160
1161 cleanup:
1162         free_file_completion_logging(td);
1163         log_err("fio: failed to alloc write comp data\n");
1164         return 1;
1165 }
1166
1167 static void cleanup_io_u(struct thread_data *td)
1168 {
1169         struct io_u *io_u;
1170
1171         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1172
1173                 if (td->io_ops->io_u_free)
1174                         td->io_ops->io_u_free(td, io_u);
1175
1176                 fio_memfree(io_u, sizeof(*io_u));
1177         }
1178
1179         free_io_mem(td);
1180
1181         io_u_rexit(&td->io_u_requeues);
1182         io_u_qexit(&td->io_u_freelist);
1183         io_u_qexit(&td->io_u_all);
1184
1185         free_file_completion_logging(td);
1186 }
1187
1188 static int init_io_u(struct thread_data *td)
1189 {
1190         struct io_u *io_u;
1191         unsigned int max_bs, min_write;
1192         int cl_align, i, max_units;
1193         int data_xfer = 1, err;
1194         char *p;
1195
1196         max_units = td->o.iodepth;
1197         max_bs = td_max_bs(td);
1198         min_write = td->o.min_bs[DDIR_WRITE];
1199         td->orig_buffer_size = (unsigned long long) max_bs
1200                                         * (unsigned long long) max_units;
1201
1202         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1203                 data_xfer = 0;
1204
1205         err = 0;
1206         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1207         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1208         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1209
1210         if (err) {
1211                 log_err("fio: failed setting up IO queues\n");
1212                 return 1;
1213         }
1214
1215         /*
1216          * if we may later need to do address alignment, then add any
1217          * possible adjustment here so that we don't cause a buffer
1218          * overflow later. this adjustment may be too much if we get
1219          * lucky and the allocator gives us an aligned address.
1220          */
1221         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1222             td_ioengine_flagged(td, FIO_RAWIO))
1223                 td->orig_buffer_size += page_mask + td->o.mem_align;
1224
1225         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1226                 unsigned long bs;
1227
1228                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1229                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1230         }
1231
1232         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1233                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1234                 return 1;
1235         }
1236
1237         if (data_xfer && allocate_io_mem(td))
1238                 return 1;
1239
1240         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1241             td_ioengine_flagged(td, FIO_RAWIO))
1242                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1243         else
1244                 p = td->orig_buffer;
1245
1246         cl_align = os_cache_line_size();
1247
1248         for (i = 0; i < max_units; i++) {
1249                 void *ptr;
1250
1251                 if (td->terminate)
1252                         return 1;
1253
1254                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1255                 if (!ptr) {
1256                         log_err("fio: unable to allocate aligned memory\n");
1257                         break;
1258                 }
1259
1260                 io_u = ptr;
1261                 memset(io_u, 0, sizeof(*io_u));
1262                 INIT_FLIST_HEAD(&io_u->verify_list);
1263                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1264
1265                 if (data_xfer) {
1266                         io_u->buf = p;
1267                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1268
1269                         if (td_write(td))
1270                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1271                         if (td_write(td) && td->o.verify_pattern_bytes) {
1272                                 /*
1273                                  * Fill the buffer with the pattern if we are
1274                                  * going to be doing writes.
1275                                  */
1276                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1277                         }
1278                 }
1279
1280                 io_u->index = i;
1281                 io_u->flags = IO_U_F_FREE;
1282                 io_u_qpush(&td->io_u_freelist, io_u);
1283
1284                 /*
1285                  * io_u never leaves this stack, used for iteration of all
1286                  * io_u buffers.
1287                  */
1288                 io_u_qpush(&td->io_u_all, io_u);
1289
1290                 if (td->io_ops->io_u_init) {
1291                         int ret = td->io_ops->io_u_init(td, io_u);
1292
1293                         if (ret) {
1294                                 log_err("fio: failed to init engine data: %d\n", ret);
1295                                 return 1;
1296                         }
1297                 }
1298
1299                 p += max_bs;
1300         }
1301
1302         if (init_file_completion_logging(td, max_units))
1303                 return 1;
1304
1305         return 0;
1306 }
1307
1308 /*
1309  * This function is Linux specific.
1310  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1311  */
1312 static int switch_ioscheduler(struct thread_data *td)
1313 {
1314 #ifdef FIO_HAVE_IOSCHED_SWITCH
1315         char tmp[256], tmp2[128];
1316         FILE *f;
1317         int ret;
1318
1319         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1320                 return 0;
1321
1322         assert(td->files && td->files[0]);
1323         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1324
1325         f = fopen(tmp, "r+");
1326         if (!f) {
1327                 if (errno == ENOENT) {
1328                         log_err("fio: os or kernel doesn't support IO scheduler"
1329                                 " switching\n");
1330                         return 0;
1331                 }
1332                 td_verror(td, errno, "fopen iosched");
1333                 return 1;
1334         }
1335
1336         /*
1337          * Set io scheduler.
1338          */
1339         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1340         if (ferror(f) || ret != 1) {
1341                 td_verror(td, errno, "fwrite");
1342                 fclose(f);
1343                 return 1;
1344         }
1345
1346         rewind(f);
1347
1348         /*
1349          * Read back and check that the selected scheduler is now the default.
1350          */
1351         memset(tmp, 0, sizeof(tmp));
1352         ret = fread(tmp, sizeof(tmp), 1, f);
1353         if (ferror(f) || ret < 0) {
1354                 td_verror(td, errno, "fread");
1355                 fclose(f);
1356                 return 1;
1357         }
1358         /*
1359          * either a list of io schedulers or "none\n" is expected.
1360          */
1361         tmp[strlen(tmp) - 1] = '\0';
1362
1363         /*
1364          * Write to "none" entry doesn't fail, so check the result here.
1365          */
1366         if (!strcmp(tmp, "none")) {
1367                 log_err("fio: io scheduler is not tunable\n");
1368                 fclose(f);
1369                 return 0;
1370         }
1371
1372         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1373         if (!strstr(tmp, tmp2)) {
1374                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1375                 td_verror(td, EINVAL, "iosched_switch");
1376                 fclose(f);
1377                 return 1;
1378         }
1379
1380         fclose(f);
1381         return 0;
1382 #else
1383         return 0;
1384 #endif
1385 }
1386
1387 static bool keep_running(struct thread_data *td)
1388 {
1389         unsigned long long limit;
1390
1391         if (td->done)
1392                 return false;
1393         if (td->terminate)
1394                 return false;
1395         if (td->o.time_based)
1396                 return true;
1397         if (td->o.loops) {
1398                 td->o.loops--;
1399                 return true;
1400         }
1401         if (exceeds_number_ios(td))
1402                 return false;
1403
1404         if (td->o.io_size)
1405                 limit = td->o.io_size;
1406         else
1407                 limit = td->o.size;
1408
1409         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1410                 uint64_t diff;
1411
1412                 /*
1413                  * If the difference is less than the maximum IO size, we
1414                  * are done.
1415                  */
1416                 diff = limit - ddir_rw_sum(td->io_bytes);
1417                 if (diff < td_max_bs(td))
1418                         return false;
1419
1420                 if (fio_files_done(td) && !td->o.io_size)
1421                         return false;
1422
1423                 return true;
1424         }
1425
1426         return false;
1427 }
1428
1429 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1430 {
1431         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1432         int ret;
1433         char *str;
1434
1435         str = malloc(newlen);
1436         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1437
1438         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1439         ret = system(str);
1440         if (ret == -1)
1441                 log_err("fio: exec of cmd <%s> failed\n", str);
1442
1443         free(str);
1444         return ret;
1445 }
1446
1447 /*
1448  * Dry run to compute correct state of numberio for verification.
1449  */
1450 static uint64_t do_dry_run(struct thread_data *td)
1451 {
1452         td_set_runstate(td, TD_RUNNING);
1453
1454         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1455                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1456                 struct io_u *io_u;
1457                 int ret;
1458
1459                 if (td->terminate || td->done)
1460                         break;
1461
1462                 io_u = get_io_u(td);
1463                 if (IS_ERR_OR_NULL(io_u))
1464                         break;
1465
1466                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1467                 io_u->error = 0;
1468                 io_u->resid = 0;
1469                 if (ddir_rw(acct_ddir(io_u)))
1470                         td->io_issues[acct_ddir(io_u)]++;
1471                 if (ddir_rw(io_u->ddir)) {
1472                         io_u_mark_depth(td, 1);
1473                         td->ts.total_io_u[io_u->ddir]++;
1474                 }
1475
1476                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1477                     td->o.do_verify &&
1478                     td->o.verify != VERIFY_NONE &&
1479                     !td->o.experimental_verify)
1480                         log_io_piece(td, io_u);
1481
1482                 ret = io_u_sync_complete(td, io_u);
1483                 (void) ret;
1484         }
1485
1486         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1487 }
1488
1489 struct fork_data {
1490         struct thread_data *td;
1491         struct sk_out *sk_out;
1492 };
1493
1494 /*
1495  * Entry point for the thread based jobs. The process based jobs end up
1496  * here as well, after a little setup.
1497  */
1498 static void *thread_main(void *data)
1499 {
1500         struct fork_data *fd = data;
1501         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1502         struct thread_data *td = fd->td;
1503         struct thread_options *o = &td->o;
1504         struct sk_out *sk_out = fd->sk_out;
1505         uint64_t bytes_done[DDIR_RWDIR_CNT];
1506         int deadlock_loop_cnt;
1507         bool clear_state, did_some_io;
1508         int ret;
1509
1510         sk_out_assign(sk_out);
1511         free(fd);
1512
1513         if (!o->use_thread) {
1514                 setsid();
1515                 td->pid = getpid();
1516         } else
1517                 td->pid = gettid();
1518
1519         fio_local_clock_init(o->use_thread);
1520
1521         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1522
1523         if (is_backend)
1524                 fio_server_send_start(td);
1525
1526         INIT_FLIST_HEAD(&td->io_log_list);
1527         INIT_FLIST_HEAD(&td->io_hist_list);
1528         INIT_FLIST_HEAD(&td->verify_list);
1529         INIT_FLIST_HEAD(&td->trim_list);
1530         INIT_FLIST_HEAD(&td->next_rand_list);
1531         td->io_hist_tree = RB_ROOT;
1532
1533         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1534         if (ret) {
1535                 td_verror(td, ret, "mutex_cond_init_pshared");
1536                 goto err;
1537         }
1538         ret = cond_init_pshared(&td->verify_cond);
1539         if (ret) {
1540                 td_verror(td, ret, "mutex_cond_pshared");
1541                 goto err;
1542         }
1543
1544         td_set_runstate(td, TD_INITIALIZED);
1545         dprint(FD_MUTEX, "up startup_mutex\n");
1546         fio_mutex_up(startup_mutex);
1547         dprint(FD_MUTEX, "wait on td->mutex\n");
1548         fio_mutex_down(td->mutex);
1549         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1550
1551         /*
1552          * A new gid requires privilege, so we need to do this before setting
1553          * the uid.
1554          */
1555         if (o->gid != -1U && setgid(o->gid)) {
1556                 td_verror(td, errno, "setgid");
1557                 goto err;
1558         }
1559         if (o->uid != -1U && setuid(o->uid)) {
1560                 td_verror(td, errno, "setuid");
1561                 goto err;
1562         }
1563
1564         /*
1565          * Do this early, we don't want the compress threads to be limited
1566          * to the same CPUs as the IO workers. So do this before we set
1567          * any potential CPU affinity
1568          */
1569         if (iolog_compress_init(td, sk_out))
1570                 goto err;
1571
1572         /*
1573          * If we have a gettimeofday() thread, make sure we exclude that
1574          * thread from this job
1575          */
1576         if (o->gtod_cpu)
1577                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1578
1579         /*
1580          * Set affinity first, in case it has an impact on the memory
1581          * allocations.
1582          */
1583         if (fio_option_is_set(o, cpumask)) {
1584                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1585                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1586                         if (!ret) {
1587                                 log_err("fio: no CPUs set\n");
1588                                 log_err("fio: Try increasing number of available CPUs\n");
1589                                 td_verror(td, EINVAL, "cpus_split");
1590                                 goto err;
1591                         }
1592                 }
1593                 ret = fio_setaffinity(td->pid, o->cpumask);
1594                 if (ret == -1) {
1595                         td_verror(td, errno, "cpu_set_affinity");
1596                         goto err;
1597                 }
1598         }
1599
1600 #ifdef CONFIG_LIBNUMA
1601         /* numa node setup */
1602         if (fio_option_is_set(o, numa_cpunodes) ||
1603             fio_option_is_set(o, numa_memnodes)) {
1604                 struct bitmask *mask;
1605
1606                 if (numa_available() < 0) {
1607                         td_verror(td, errno, "Does not support NUMA API\n");
1608                         goto err;
1609                 }
1610
1611                 if (fio_option_is_set(o, numa_cpunodes)) {
1612                         mask = numa_parse_nodestring(o->numa_cpunodes);
1613                         ret = numa_run_on_node_mask(mask);
1614                         numa_free_nodemask(mask);
1615                         if (ret == -1) {
1616                                 td_verror(td, errno, \
1617                                         "numa_run_on_node_mask failed\n");
1618                                 goto err;
1619                         }
1620                 }
1621
1622                 if (fio_option_is_set(o, numa_memnodes)) {
1623                         mask = NULL;
1624                         if (o->numa_memnodes)
1625                                 mask = numa_parse_nodestring(o->numa_memnodes);
1626
1627                         switch (o->numa_mem_mode) {
1628                         case MPOL_INTERLEAVE:
1629                                 numa_set_interleave_mask(mask);
1630                                 break;
1631                         case MPOL_BIND:
1632                                 numa_set_membind(mask);
1633                                 break;
1634                         case MPOL_LOCAL:
1635                                 numa_set_localalloc();
1636                                 break;
1637                         case MPOL_PREFERRED:
1638                                 numa_set_preferred(o->numa_mem_prefer_node);
1639                                 break;
1640                         case MPOL_DEFAULT:
1641                         default:
1642                                 break;
1643                         }
1644
1645                         if (mask)
1646                                 numa_free_nodemask(mask);
1647
1648                 }
1649         }
1650 #endif
1651
1652         if (fio_pin_memory(td))
1653                 goto err;
1654
1655         /*
1656          * May alter parameters that init_io_u() will use, so we need to
1657          * do this first.
1658          */
1659         if (init_iolog(td))
1660                 goto err;
1661
1662         if (init_io_u(td))
1663                 goto err;
1664
1665         if (o->verify_async && verify_async_init(td))
1666                 goto err;
1667
1668         if (fio_option_is_set(o, ioprio) ||
1669             fio_option_is_set(o, ioprio_class)) {
1670                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1671                 if (ret == -1) {
1672                         td_verror(td, errno, "ioprio_set");
1673                         goto err;
1674                 }
1675         }
1676
1677         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1678                 goto err;
1679
1680         errno = 0;
1681         if (nice(o->nice) == -1 && errno != 0) {
1682                 td_verror(td, errno, "nice");
1683                 goto err;
1684         }
1685
1686         if (o->ioscheduler && switch_ioscheduler(td))
1687                 goto err;
1688
1689         if (!o->create_serialize && setup_files(td))
1690                 goto err;
1691
1692         if (td_io_init(td))
1693                 goto err;
1694
1695         if (init_random_map(td))
1696                 goto err;
1697
1698         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1699                 goto err;
1700
1701         if (o->pre_read) {
1702                 if (pre_read_files(td) < 0)
1703                         goto err;
1704         }
1705
1706         fio_verify_init(td);
1707
1708         if (rate_submit_init(td, sk_out))
1709                 goto err;
1710
1711         set_epoch_time(td, o->log_unix_epoch);
1712         fio_getrusage(&td->ru_start);
1713         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1714         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1715         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1716
1717         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1718                         o->ratemin[DDIR_TRIM]) {
1719                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1720                                         sizeof(td->bw_sample_time));
1721                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1722                                         sizeof(td->bw_sample_time));
1723                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1724                                         sizeof(td->bw_sample_time));
1725         }
1726
1727         memset(bytes_done, 0, sizeof(bytes_done));
1728         clear_state = false;
1729         did_some_io = false;
1730
1731         while (keep_running(td)) {
1732                 uint64_t verify_bytes;
1733
1734                 fio_gettime(&td->start, NULL);
1735                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1736
1737                 if (clear_state) {
1738                         clear_io_state(td, 0);
1739
1740                         if (o->unlink_each_loop && unlink_all_files(td))
1741                                 break;
1742                 }
1743
1744                 prune_io_piece_log(td);
1745
1746                 if (td->o.verify_only && td_write(td))
1747                         verify_bytes = do_dry_run(td);
1748                 else {
1749                         do_io(td, bytes_done);
1750
1751                         if (!ddir_rw_sum(bytes_done)) {
1752                                 fio_mark_td_terminate(td);
1753                                 verify_bytes = 0;
1754                         } else {
1755                                 verify_bytes = bytes_done[DDIR_WRITE] +
1756                                                 bytes_done[DDIR_TRIM];
1757                         }
1758                 }
1759
1760                 /*
1761                  * If we took too long to shut down, the main thread could
1762                  * already consider us reaped/exited. If that happens, break
1763                  * out and clean up.
1764                  */
1765                 if (td->runstate >= TD_EXITED)
1766                         break;
1767
1768                 clear_state = true;
1769
1770                 /*
1771                  * Make sure we've successfully updated the rusage stats
1772                  * before waiting on the stat mutex. Otherwise we could have
1773                  * the stat thread holding stat mutex and waiting for
1774                  * the rusage_sem, which would never get upped because
1775                  * this thread is waiting for the stat mutex.
1776                  */
1777                 deadlock_loop_cnt = 0;
1778                 do {
1779                         check_update_rusage(td);
1780                         if (!fio_mutex_down_trylock(stat_mutex))
1781                                 break;
1782                         usleep(1000);
1783                         if (deadlock_loop_cnt++ > 5000) {
1784                                 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1785                                 td->error = EDEADLK;
1786                                 goto err;
1787                         }
1788                 } while (1);
1789
1790                 if (td_read(td) && td->io_bytes[DDIR_READ])
1791                         update_runtime(td, elapsed_us, DDIR_READ);
1792                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1793                         update_runtime(td, elapsed_us, DDIR_WRITE);
1794                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1795                         update_runtime(td, elapsed_us, DDIR_TRIM);
1796                 fio_gettime(&td->start, NULL);
1797                 fio_mutex_up(stat_mutex);
1798
1799                 if (td->error || td->terminate)
1800                         break;
1801
1802                 if (!o->do_verify ||
1803                     o->verify == VERIFY_NONE ||
1804                     td_ioengine_flagged(td, FIO_UNIDIR))
1805                         continue;
1806
1807                 if (ddir_rw_sum(bytes_done))
1808                         did_some_io = true;
1809
1810                 clear_io_state(td, 0);
1811
1812                 fio_gettime(&td->start, NULL);
1813
1814                 do_verify(td, verify_bytes);
1815
1816                 /*
1817                  * See comment further up for why this is done here.
1818                  */
1819                 check_update_rusage(td);
1820
1821                 fio_mutex_down(stat_mutex);
1822                 update_runtime(td, elapsed_us, DDIR_READ);
1823                 fio_gettime(&td->start, NULL);
1824                 fio_mutex_up(stat_mutex);
1825
1826                 if (td->error || td->terminate)
1827                         break;
1828         }
1829
1830         /*
1831          * If td ended up with no I/O when it should have had,
1832          * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1833          * (Are we not missing other flags that can be ignored ?)
1834          */
1835         if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1836             !did_some_io && !td->o.create_only &&
1837             !(td_ioengine_flagged(td, FIO_NOIO) ||
1838               td_ioengine_flagged(td, FIO_DISKLESSIO)))
1839                 log_err("%s: No I/O performed by %s, "
1840                          "perhaps try --debug=io option for details?\n",
1841                          td->o.name, td->io_ops->name);
1842
1843         td_set_runstate(td, TD_FINISHING);
1844
1845         update_rusage_stat(td);
1846         td->ts.total_run_time = mtime_since_now(&td->epoch);
1847         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1848         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1849         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1850
1851         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1852             (td->o.verify != VERIFY_NONE && td_write(td)))
1853                 verify_save_state(td->thread_number);
1854
1855         fio_unpin_memory(td);
1856
1857         td_writeout_logs(td, true);
1858
1859         iolog_compress_exit(td);
1860         rate_submit_exit(td);
1861
1862         if (o->exec_postrun)
1863                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1864
1865         if (exitall_on_terminate || (o->exitall_error && td->error))
1866                 fio_terminate_threads(td->groupid);
1867
1868 err:
1869         if (td->error)
1870                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1871                                                         td->verror);
1872
1873         if (o->verify_async)
1874                 verify_async_exit(td);
1875
1876         close_and_free_files(td);
1877         cleanup_io_u(td);
1878         close_ioengine(td);
1879         cgroup_shutdown(td, &cgroup_mnt);
1880         verify_free_state(td);
1881
1882         if (td->zone_state_index) {
1883                 int i;
1884
1885                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1886                         free(td->zone_state_index[i]);
1887                 free(td->zone_state_index);
1888                 td->zone_state_index = NULL;
1889         }
1890
1891         if (fio_option_is_set(o, cpumask)) {
1892                 ret = fio_cpuset_exit(&o->cpumask);
1893                 if (ret)
1894                         td_verror(td, ret, "fio_cpuset_exit");
1895         }
1896
1897         /*
1898          * do this very late, it will log file closing as well
1899          */
1900         if (o->write_iolog_file)
1901                 write_iolog_close(td);
1902
1903         td_set_runstate(td, TD_EXITED);
1904
1905         /*
1906          * Do this last after setting our runstate to exited, so we
1907          * know that the stat thread is signaled.
1908          */
1909         check_update_rusage(td);
1910
1911         sk_out_drop();
1912         return (void *) (uintptr_t) td->error;
1913 }
1914
1915 /*
1916  * Run over the job map and reap the threads that have exited, if any.
1917  */
1918 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1919                          uint64_t *m_rate)
1920 {
1921         struct thread_data *td;
1922         unsigned int cputhreads, realthreads, pending;
1923         int i, status, ret;
1924
1925         /*
1926          * reap exited threads (TD_EXITED -> TD_REAPED)
1927          */
1928         realthreads = pending = cputhreads = 0;
1929         for_each_td(td, i) {
1930                 int flags = 0;
1931
1932                  if (!strcmp(td->o.ioengine, "cpuio"))
1933                         cputhreads++;
1934                 else
1935                         realthreads++;
1936
1937                 if (!td->pid) {
1938                         pending++;
1939                         continue;
1940                 }
1941                 if (td->runstate == TD_REAPED)
1942                         continue;
1943                 if (td->o.use_thread) {
1944                         if (td->runstate == TD_EXITED) {
1945                                 td_set_runstate(td, TD_REAPED);
1946                                 goto reaped;
1947                         }
1948                         continue;
1949                 }
1950
1951                 flags = WNOHANG;
1952                 if (td->runstate == TD_EXITED)
1953                         flags = 0;
1954
1955                 /*
1956                  * check if someone quit or got killed in an unusual way
1957                  */
1958                 ret = waitpid(td->pid, &status, flags);
1959                 if (ret < 0) {
1960                         if (errno == ECHILD) {
1961                                 log_err("fio: pid=%d disappeared %d\n",
1962                                                 (int) td->pid, td->runstate);
1963                                 td->sig = ECHILD;
1964                                 td_set_runstate(td, TD_REAPED);
1965                                 goto reaped;
1966                         }
1967                         perror("waitpid");
1968                 } else if (ret == td->pid) {
1969                         if (WIFSIGNALED(status)) {
1970                                 int sig = WTERMSIG(status);
1971
1972                                 if (sig != SIGTERM && sig != SIGUSR2)
1973                                         log_err("fio: pid=%d, got signal=%d\n",
1974                                                         (int) td->pid, sig);
1975                                 td->sig = sig;
1976                                 td_set_runstate(td, TD_REAPED);
1977                                 goto reaped;
1978                         }
1979                         if (WIFEXITED(status)) {
1980                                 if (WEXITSTATUS(status) && !td->error)
1981                                         td->error = WEXITSTATUS(status);
1982
1983                                 td_set_runstate(td, TD_REAPED);
1984                                 goto reaped;
1985                         }
1986                 }
1987
1988                 /*
1989                  * If the job is stuck, do a forceful timeout of it and
1990                  * move on.
1991                  */
1992                 if (td->terminate &&
1993                     td->runstate < TD_FSYNCING &&
1994                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1995                         log_err("fio: job '%s' (state=%d) hasn't exited in "
1996                                 "%lu seconds, it appears to be stuck. Doing "
1997                                 "forceful exit of this job.\n",
1998                                 td->o.name, td->runstate,
1999                                 (unsigned long) time_since_now(&td->terminate_time));
2000                         td_set_runstate(td, TD_REAPED);
2001                         goto reaped;
2002                 }
2003
2004                 /*
2005                  * thread is not dead, continue
2006                  */
2007                 pending++;
2008                 continue;
2009 reaped:
2010                 (*nr_running)--;
2011                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2012                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2013                 if (!td->pid)
2014                         pending--;
2015
2016                 if (td->error)
2017                         exit_value++;
2018
2019                 done_secs += mtime_since_now(&td->epoch) / 1000;
2020                 profile_td_exit(td);
2021         }
2022
2023         if (*nr_running == cputhreads && !pending && realthreads)
2024                 fio_terminate_threads(TERMINATE_ALL);
2025 }
2026
2027 static bool __check_trigger_file(void)
2028 {
2029         struct stat sb;
2030
2031         if (!trigger_file)
2032                 return false;
2033
2034         if (stat(trigger_file, &sb))
2035                 return false;
2036
2037         if (unlink(trigger_file) < 0)
2038                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2039                                                         strerror(errno));
2040
2041         return true;
2042 }
2043
2044 static bool trigger_timedout(void)
2045 {
2046         if (trigger_timeout)
2047                 if (time_since_genesis() >= trigger_timeout) {
2048                         trigger_timeout = 0;
2049                         return true;
2050                 }
2051
2052         return false;
2053 }
2054
2055 void exec_trigger(const char *cmd)
2056 {
2057         int ret;
2058
2059         if (!cmd || cmd[0] == '\0')
2060                 return;
2061
2062         ret = system(cmd);
2063         if (ret == -1)
2064                 log_err("fio: failed executing %s trigger\n", cmd);
2065 }
2066
2067 void check_trigger_file(void)
2068 {
2069         if (__check_trigger_file() || trigger_timedout()) {
2070                 if (nr_clients)
2071                         fio_clients_send_trigger(trigger_remote_cmd);
2072                 else {
2073                         verify_save_state(IO_LIST_ALL);
2074                         fio_terminate_threads(TERMINATE_ALL);
2075                         exec_trigger(trigger_cmd);
2076                 }
2077         }
2078 }
2079
2080 static int fio_verify_load_state(struct thread_data *td)
2081 {
2082         int ret;
2083
2084         if (!td->o.verify_state)
2085                 return 0;
2086
2087         if (is_backend) {
2088                 void *data;
2089
2090                 ret = fio_server_get_verify_state(td->o.name,
2091                                         td->thread_number - 1, &data);
2092                 if (!ret)
2093                         verify_assign_state(td, data);
2094         } else
2095                 ret = verify_load_state(td, "local");
2096
2097         return ret;
2098 }
2099
2100 static void do_usleep(unsigned int usecs)
2101 {
2102         check_for_running_stats();
2103         check_trigger_file();
2104         usleep(usecs);
2105 }
2106
2107 static bool check_mount_writes(struct thread_data *td)
2108 {
2109         struct fio_file *f;
2110         unsigned int i;
2111
2112         if (!td_write(td) || td->o.allow_mounted_write)
2113                 return false;
2114
2115         /*
2116          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2117          * are mkfs'd and mounted.
2118          */
2119         for_each_file(td, f, i) {
2120 #ifdef FIO_HAVE_CHARDEV_SIZE
2121                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2122 #else
2123                 if (f->filetype != FIO_TYPE_BLOCK)
2124 #endif
2125                         continue;
2126                 if (device_is_mounted(f->file_name))
2127                         goto mounted;
2128         }
2129
2130         return false;
2131 mounted:
2132         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2133         return true;
2134 }
2135
2136 static bool waitee_running(struct thread_data *me)
2137 {
2138         const char *waitee = me->o.wait_for;
2139         const char *self = me->o.name;
2140         struct thread_data *td;
2141         int i;
2142
2143         if (!waitee)
2144                 return false;
2145
2146         for_each_td(td, i) {
2147                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2148                         continue;
2149
2150                 if (td->runstate < TD_EXITED) {
2151                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2152                                         self, td->o.name,
2153                                         runstate_to_name(td->runstate));
2154                         return true;
2155                 }
2156         }
2157
2158         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2159         return false;
2160 }
2161
2162 /*
2163  * Main function for kicking off and reaping jobs, as needed.
2164  */
2165 static void run_threads(struct sk_out *sk_out)
2166 {
2167         struct thread_data *td;
2168         unsigned int i, todo, nr_running, nr_started;
2169         uint64_t m_rate, t_rate;
2170         uint64_t spent;
2171
2172         if (fio_gtod_offload && fio_start_gtod_thread())
2173                 return;
2174
2175         fio_idle_prof_init();
2176
2177         set_sig_handlers();
2178
2179         nr_thread = nr_process = 0;
2180         for_each_td(td, i) {
2181                 if (check_mount_writes(td))
2182                         return;
2183                 if (td->o.use_thread)
2184                         nr_thread++;
2185                 else
2186                         nr_process++;
2187         }
2188
2189         if (output_format & FIO_OUTPUT_NORMAL) {
2190                 log_info("Starting ");
2191                 if (nr_thread)
2192                         log_info("%d thread%s", nr_thread,
2193                                                 nr_thread > 1 ? "s" : "");
2194                 if (nr_process) {
2195                         if (nr_thread)
2196                                 log_info(" and ");
2197                         log_info("%d process%s", nr_process,
2198                                                 nr_process > 1 ? "es" : "");
2199                 }
2200                 log_info("\n");
2201                 log_info_flush();
2202         }
2203
2204         todo = thread_number;
2205         nr_running = 0;
2206         nr_started = 0;
2207         m_rate = t_rate = 0;
2208
2209         for_each_td(td, i) {
2210                 print_status_init(td->thread_number - 1);
2211
2212                 if (!td->o.create_serialize)
2213                         continue;
2214
2215                 if (fio_verify_load_state(td))
2216                         goto reap;
2217
2218                 /*
2219                  * do file setup here so it happens sequentially,
2220                  * we don't want X number of threads getting their
2221                  * client data interspersed on disk
2222                  */
2223                 if (setup_files(td)) {
2224 reap:
2225                         exit_value++;
2226                         if (td->error)
2227                                 log_err("fio: pid=%d, err=%d/%s\n",
2228                                         (int) td->pid, td->error, td->verror);
2229                         td_set_runstate(td, TD_REAPED);
2230                         todo--;
2231                 } else {
2232                         struct fio_file *f;
2233                         unsigned int j;
2234
2235                         /*
2236                          * for sharing to work, each job must always open
2237                          * its own files. so close them, if we opened them
2238                          * for creation
2239                          */
2240                         for_each_file(td, f, j) {
2241                                 if (fio_file_open(f))
2242                                         td_io_close_file(td, f);
2243                         }
2244                 }
2245         }
2246
2247         /* start idle threads before io threads start to run */
2248         fio_idle_prof_start();
2249
2250         set_genesis_time();
2251
2252         while (todo) {
2253                 struct thread_data *map[REAL_MAX_JOBS];
2254                 struct timespec this_start;
2255                 int this_jobs = 0, left;
2256                 struct fork_data *fd;
2257
2258                 /*
2259                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2260                  */
2261                 for_each_td(td, i) {
2262                         if (td->runstate != TD_NOT_CREATED)
2263                                 continue;
2264
2265                         /*
2266                          * never got a chance to start, killed by other
2267                          * thread for some reason
2268                          */
2269                         if (td->terminate) {
2270                                 todo--;
2271                                 continue;
2272                         }
2273
2274                         if (td->o.start_delay) {
2275                                 spent = utime_since_genesis();
2276
2277                                 if (td->o.start_delay > spent)
2278                                         continue;
2279                         }
2280
2281                         if (td->o.stonewall && (nr_started || nr_running)) {
2282                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2283                                                         td->o.name);
2284                                 break;
2285                         }
2286
2287                         if (waitee_running(td)) {
2288                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2289                                                 td->o.name, td->o.wait_for);
2290                                 continue;
2291                         }
2292
2293                         init_disk_util(td);
2294
2295                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2296                         td->update_rusage = 0;
2297
2298                         /*
2299                          * Set state to created. Thread will transition
2300                          * to TD_INITIALIZED when it's done setting up.
2301                          */
2302                         td_set_runstate(td, TD_CREATED);
2303                         map[this_jobs++] = td;
2304                         nr_started++;
2305
2306                         fd = calloc(1, sizeof(*fd));
2307                         fd->td = td;
2308                         fd->sk_out = sk_out;
2309
2310                         if (td->o.use_thread) {
2311                                 int ret;
2312
2313                                 dprint(FD_PROCESS, "will pthread_create\n");
2314                                 ret = pthread_create(&td->thread, NULL,
2315                                                         thread_main, fd);
2316                                 if (ret) {
2317                                         log_err("pthread_create: %s\n",
2318                                                         strerror(ret));
2319                                         free(fd);
2320                                         nr_started--;
2321                                         break;
2322                                 }
2323                                 ret = pthread_detach(td->thread);
2324                                 if (ret)
2325                                         log_err("pthread_detach: %s",
2326                                                         strerror(ret));
2327                         } else {
2328                                 pid_t pid;
2329                                 dprint(FD_PROCESS, "will fork\n");
2330                                 pid = fork();
2331                                 if (!pid) {
2332                                         int ret;
2333
2334                                         ret = (int)(uintptr_t)thread_main(fd);
2335                                         _exit(ret);
2336                                 } else if (i == fio_debug_jobno)
2337                                         *fio_debug_jobp = pid;
2338                         }
2339                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2340                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2341                                 log_err("fio: job startup hung? exiting.\n");
2342                                 fio_terminate_threads(TERMINATE_ALL);
2343                                 fio_abort = 1;
2344                                 nr_started--;
2345                                 free(fd);
2346                                 break;
2347                         }
2348                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2349                 }
2350
2351                 /*
2352                  * Wait for the started threads to transition to
2353                  * TD_INITIALIZED.
2354                  */
2355                 fio_gettime(&this_start, NULL);
2356                 left = this_jobs;
2357                 while (left && !fio_abort) {
2358                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2359                                 break;
2360
2361                         do_usleep(100000);
2362
2363                         for (i = 0; i < this_jobs; i++) {
2364                                 td = map[i];
2365                                 if (!td)
2366                                         continue;
2367                                 if (td->runstate == TD_INITIALIZED) {
2368                                         map[i] = NULL;
2369                                         left--;
2370                                 } else if (td->runstate >= TD_EXITED) {
2371                                         map[i] = NULL;
2372                                         left--;
2373                                         todo--;
2374                                         nr_running++; /* work-around... */
2375                                 }
2376                         }
2377                 }
2378
2379                 if (left) {
2380                         log_err("fio: %d job%s failed to start\n", left,
2381                                         left > 1 ? "s" : "");
2382                         for (i = 0; i < this_jobs; i++) {
2383                                 td = map[i];
2384                                 if (!td)
2385                                         continue;
2386                                 kill(td->pid, SIGTERM);
2387                         }
2388                         break;
2389                 }
2390
2391                 /*
2392                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2393                  */
2394                 for_each_td(td, i) {
2395                         if (td->runstate != TD_INITIALIZED)
2396                                 continue;
2397
2398                         if (in_ramp_time(td))
2399                                 td_set_runstate(td, TD_RAMP);
2400                         else
2401                                 td_set_runstate(td, TD_RUNNING);
2402                         nr_running++;
2403                         nr_started--;
2404                         m_rate += ddir_rw_sum(td->o.ratemin);
2405                         t_rate += ddir_rw_sum(td->o.rate);
2406                         todo--;
2407                         fio_mutex_up(td->mutex);
2408                 }
2409
2410                 reap_threads(&nr_running, &t_rate, &m_rate);
2411
2412                 if (todo)
2413                         do_usleep(100000);
2414         }
2415
2416         while (nr_running) {
2417                 reap_threads(&nr_running, &t_rate, &m_rate);
2418                 do_usleep(10000);
2419         }
2420
2421         fio_idle_prof_stop();
2422
2423         update_io_ticks();
2424 }
2425
2426 static void free_disk_util(void)
2427 {
2428         disk_util_prune_entries();
2429         helper_thread_destroy();
2430 }
2431
2432 int fio_backend(struct sk_out *sk_out)
2433 {
2434         struct thread_data *td;
2435         int i;
2436
2437         if (exec_profile) {
2438                 if (load_profile(exec_profile))
2439                         return 1;
2440                 free(exec_profile);
2441                 exec_profile = NULL;
2442         }
2443         if (!thread_number)
2444                 return 0;
2445
2446         if (write_bw_log) {
2447                 struct log_params p = {
2448                         .log_type = IO_LOG_TYPE_BW,
2449                 };
2450
2451                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2452                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2453                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2454         }
2455
2456         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2457         if (startup_mutex == NULL)
2458                 return 1;
2459
2460         set_genesis_time();
2461         stat_init();
2462         helper_thread_create(startup_mutex, sk_out);
2463
2464         cgroup_list = smalloc(sizeof(*cgroup_list));
2465         INIT_FLIST_HEAD(cgroup_list);
2466
2467         run_threads(sk_out);
2468
2469         helper_thread_exit();
2470
2471         if (!fio_abort) {
2472                 __show_run_stats();
2473                 if (write_bw_log) {
2474                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2475                                 struct io_log *log = agg_io_log[i];
2476
2477                                 flush_log(log, false);
2478                                 free_log(log);
2479                         }
2480                 }
2481         }
2482
2483         for_each_td(td, i) {
2484                 if (td->ss.dur) {
2485                         if (td->ss.iops_data != NULL) {
2486                                 free(td->ss.iops_data);
2487                                 free(td->ss.bw_data);
2488                         }
2489                 }
2490                 fio_options_free(td);
2491                 if (td->rusage_sem) {
2492                         fio_mutex_remove(td->rusage_sem);
2493                         td->rusage_sem = NULL;
2494                 }
2495                 fio_mutex_remove(td->mutex);
2496                 td->mutex = NULL;
2497         }
2498
2499         free_disk_util();
2500         cgroup_kill(cgroup_list);
2501         sfree(cgroup_list);
2502         sfree(cgroup_mnt);
2503
2504         fio_mutex_remove(startup_mutex);
2505         stat_exit();
2506         return exit_value;
2507 }