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