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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32 #include <pthread.h>
  33
  34 #include "fio.h"
  35 #include "smalloc.h"
  36 #include "verify.h"
  37 #include "diskutil.h"
  38 #include "cgroup.h"
  39 #include "profile.h"
  40 #include "lib/rand.h"
  41 #include "lib/memalign.h"
  42 #include "server.h"
  43 #include "lib/getrusage.h"
  44 #include "idletime.h"
  45 #include "err.h"
  46 #include "workqueue.h"
  47 #include "lib/mountcheck.h"
  48 #include "rate-submit.h"
  49 #include "helper_thread.h"
  50 #include "pshared.h"
  51 #include "zone-dist.h"
  52
  53 static struct fio_sem *startup_sem;
  54 static struct flist_head *cgroup_list;
  55 static struct cgroup_mnt *cgroup_mnt;
  56 static int exit_value;
  57 static volatile bool fio_abort;
  58 static unsigned int nr_process = 0;
  59 static unsigned int nr_thread = 0;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int nr_segments = 0;
  66 unsigned int cur_segment = 0;
  67 unsigned int stat_number = 0;
  68 int temp_stall_ts;
  69 unsigned long done_secs = 0;
  70 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
  71 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
  72 #else
  73 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  74 #endif
  75
  76 #define JOB_START_TIMEOUT       (5 * 1000)
  77
  78 static void sig_int(int sig)
  79 {
  80         if (nr_segments) {
  81                 if (is_backend)
  82                         fio_server_got_signal(sig);
  83                 else {
  84                         log_info("\nfio: terminating on signal %d\n", sig);
  85                         log_info_flush();
  86                         exit_value = 128;
  87                 }
  88
  89                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  90         }
  91 }
  92
  93 void sig_show_status(int sig)
  94 {
  95         show_running_run_stats();
  96 }
  97
  98 static void set_sig_handlers(void)
  99 {
 100         struct sigaction act;
 101
 102         memset(&act, 0, sizeof(act));
 103         act.sa_handler = sig_int;
 104         act.sa_flags = SA_RESTART;
 105         sigaction(SIGINT, &act, NULL);
 106
 107         memset(&act, 0, sizeof(act));
 108         act.sa_handler = sig_int;
 109         act.sa_flags = SA_RESTART;
 110         sigaction(SIGTERM, &act, NULL);
 111
 112 /* Windows uses SIGBREAK as a quit signal from other applications */
 113 #ifdef WIN32
 114         memset(&act, 0, sizeof(act));
 115         act.sa_handler = sig_int;
 116         act.sa_flags = SA_RESTART;
 117         sigaction(SIGBREAK, &act, NULL);
 118 #endif
 119
 120         memset(&act, 0, sizeof(act));
 121         act.sa_handler = sig_show_status;
 122         act.sa_flags = SA_RESTART;
 123         sigaction(SIGUSR1, &act, NULL);
 124
 125         if (is_backend) {
 126                 memset(&act, 0, sizeof(act));
 127                 act.sa_handler = sig_int;
 128                 act.sa_flags = SA_RESTART;
 129                 sigaction(SIGPIPE, &act, NULL);
 130         }
 131 }
 132
 133 /*
 134  * Check if we are above the minimum rate given.
 135  */
 136 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 137                              enum fio_ddir ddir)
 138 {
 139         unsigned long long bytes = 0;
 140         unsigned long iops = 0;
 141         unsigned long spent;
 142         unsigned long long rate;
 143         unsigned long long ratemin = 0;
 144         unsigned int rate_iops = 0;
 145         unsigned int rate_iops_min = 0;
 146
 147         assert(ddir_rw(ddir));
 148
 149         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 150                 return false;
 151
 152         /*
 153          * allow a 2 second settle period in the beginning
 154          */
 155         if (mtime_since(&td->start, now) < 2000)
 156                 return false;
 157
 158         iops += td->this_io_blocks[ddir];
 159         bytes += td->this_io_bytes[ddir];
 160         ratemin += td->o.ratemin[ddir];
 161         rate_iops += td->o.rate_iops[ddir];
 162         rate_iops_min += td->o.rate_iops_min[ddir];
 163
 164         /*
 165          * if rate blocks is set, sample is running
 166          */
 167         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 168                 spent = mtime_since(&td->lastrate[ddir], now);
 169                 if (spent < td->o.ratecycle)
 170                         return false;
 171
 172                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 173                         /*
 174                          * check bandwidth specified rate
 175                          */
 176                         if (bytes < td->rate_bytes[ddir]) {
 177                                 log_err("%s: rate_min=%lluB/s not met, only transferred %lluB\n",
 178                                         td->o.name, ratemin, bytes);
 179                                 return true;
 180                         } else {
 181                                 if (spent)
 182                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 183                                 else
 184                                         rate = 0;
 185
 186                                 if (rate < ratemin ||
 187                                     bytes < td->rate_bytes[ddir]) {
 188                                         log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 189                                                 td->o.name, ratemin, rate);
 190                                         return true;
 191                                 }
 192                         }
 193                 } else {
 194                         /*
 195                          * checks iops specified rate
 196                          */
 197                         if (iops < rate_iops) {
 198                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 199                                                 td->o.name, rate_iops, iops);
 200                                 return true;
 201                         } else {
 202                                 if (spent)
 203                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 204                                 else
 205                                         rate = 0;
 206
 207                                 if (rate < rate_iops_min ||
 208                                     iops < td->rate_blocks[ddir]) {
 209                                         log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 210                                                 td->o.name, rate_iops_min, rate);
 211                                         return true;
 212                                 }
 213                         }
 214                 }
 215         }
 216
 217         td->rate_bytes[ddir] = bytes;
 218         td->rate_blocks[ddir] = iops;
 219         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 220         return false;
 221 }
 222
 223 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 224 {
 225         bool ret = false;
 226
 227         for_each_rw_ddir(ddir) {
 228                 if (td->bytes_done[ddir])
 229                         ret |= __check_min_rate(td, now, ddir);
 230         }
 231
 232         return ret;
 233 }
 234
 235 /*
 236  * When job exits, we can cancel the in-flight IO if we are using async
 237  * io. Attempt to do so.
 238  */
 239 static void cleanup_pending_aio(struct thread_data *td)
 240 {
 241         int r;
 242
 243         /*
 244          * get immediately available events, if any
 245          */
 246         r = io_u_queued_complete(td, 0);
 247
 248         /*
 249          * now cancel remaining active events
 250          */
 251         if (td->io_ops->cancel) {
 252                 struct io_u *io_u;
 253                 int i;
 254
 255                 io_u_qiter(&td->io_u_all, io_u, i) {
 256                         if (io_u->flags & IO_U_F_FLIGHT) {
 257                                 r = td->io_ops->cancel(td, io_u);
 258                                 if (!r)
 259                                         put_io_u(td, io_u);
 260                         }
 261                 }
 262         }
 263
 264         if (td->cur_depth)
 265                 r = io_u_queued_complete(td, td->cur_depth);
 266 }
 267
 268 /*
 269  * Helper to handle the final sync of a file. Works just like the normal
 270  * io path, just does everything sync.
 271  */
 272 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 273 {
 274         struct io_u *io_u = __get_io_u(td);
 275         enum fio_q_status ret;
 276
 277         if (!io_u)
 278                 return true;
 279
 280         io_u->ddir = DDIR_SYNC;
 281         io_u->file = f;
 282         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 283
 284         if (td_io_prep(td, io_u)) {
 285                 put_io_u(td, io_u);
 286                 return true;
 287         }
 288
 289 requeue:
 290         ret = td_io_queue(td, io_u);
 291         switch (ret) {
 292         case FIO_Q_QUEUED:
 293                 td_io_commit(td);
 294                 if (io_u_queued_complete(td, 1) < 0)
 295                         return true;
 296                 break;
 297         case FIO_Q_COMPLETED:
 298                 if (io_u->error) {
 299                         td_verror(td, io_u->error, "td_io_queue");
 300                         return true;
 301                 }
 302
 303                 if (io_u_sync_complete(td, io_u) < 0)
 304                         return true;
 305                 break;
 306         case FIO_Q_BUSY:
 307                 td_io_commit(td);
 308                 goto requeue;
 309         }
 310
 311         return false;
 312 }
 313
 314 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 315 {
 316         int ret, ret2;
 317
 318         if (fio_file_open(f))
 319                 return fio_io_sync(td, f);
 320
 321         if (td_io_open_file(td, f))
 322                 return 1;
 323
 324         ret = fio_io_sync(td, f);
 325         ret2 = 0;
 326         if (fio_file_open(f))
 327                 ret2 = td_io_close_file(td, f);
 328         return (ret || ret2);
 329 }
 330
 331 static inline void __update_ts_cache(struct thread_data *td)
 332 {
 333         fio_gettime(&td->ts_cache, NULL);
 334 }
 335
 336 static inline void update_ts_cache(struct thread_data *td)
 337 {
 338         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 339                 __update_ts_cache(td);
 340 }
 341
 342 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 343 {
 344         if (in_ramp_time(td))
 345                 return false;
 346         if (!td->o.timeout)
 347                 return false;
 348         if (utime_since(&td->epoch, t) >= td->o.timeout)
 349                 return true;
 350
 351         return false;
 352 }
 353
 354 /*
 355  * We need to update the runtime consistently in ms, but keep a running
 356  * tally of the current elapsed time in microseconds for sub millisecond
 357  * updates.
 358  */
 359 static inline void update_runtime(struct thread_data *td,
 360                                   unsigned long long *elapsed_us,
 361                                   const enum fio_ddir ddir)
 362 {
 363         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 364                 return;
 365
 366         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 367         elapsed_us[ddir] += utime_since_now(&td->start);
 368         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 369 }
 370
 371 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 372                                 int *retptr)
 373 {
 374         int ret = *retptr;
 375
 376         if (ret < 0 || td->error) {
 377                 int err = td->error;
 378                 enum error_type_bit eb;
 379
 380                 if (ret < 0)
 381                         err = -ret;
 382
 383                 eb = td_error_type(ddir, err);
 384                 if (!(td->o.continue_on_error & (1 << eb)))
 385                         return true;
 386
 387                 if (td_non_fatal_error(td, eb, err)) {
 388                         /*
 389                          * Continue with the I/Os in case of
 390                          * a non fatal error.
 391                          */
 392                         update_error_count(td, err);
 393                         td_clear_error(td);
 394                         *retptr = 0;
 395                         return false;
 396                 } else if (td->o.fill_device && err == ENOSPC) {
 397                         /*
 398                          * We expect to hit this error if
 399                          * fill_device option is set.
 400                          */
 401                         td_clear_error(td);
 402                         fio_mark_td_terminate(td);
 403                         return true;
 404                 } else {
 405                         /*
 406                          * Stop the I/O in case of a fatal
 407                          * error.
 408                          */
 409                         update_error_count(td, err);
 410                         return true;
 411                 }
 412         }
 413
 414         return false;
 415 }
 416
 417 static void check_update_rusage(struct thread_data *td)
 418 {
 419         if (td->update_rusage) {
 420                 td->update_rusage = 0;
 421                 update_rusage_stat(td);
 422                 fio_sem_up(td->rusage_sem);
 423         }
 424 }
 425
 426 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 427 {
 428         const int full = queue_full(td);
 429         int min_evts = 0;
 430         int ret;
 431
 432         if (td->flags & TD_F_REGROW_LOGS)
 433                 return io_u_quiesce(td);
 434
 435         /*
 436          * if the queue is full, we MUST reap at least 1 event
 437          */
 438         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 439         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 440                 min_evts = 1;
 441
 442         if (time && should_check_rate(td))
 443                 fio_gettime(time, NULL);
 444
 445         do {
 446                 ret = io_u_queued_complete(td, min_evts);
 447                 if (ret < 0)
 448                         break;
 449         } while (full && (td->cur_depth > td->o.iodepth_low));
 450
 451         return ret;
 452 }
 453
 454 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 455                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 456                    struct timespec *comp_time)
 457 {
 458         switch (*ret) {
 459         case FIO_Q_COMPLETED:
 460                 if (io_u->error) {
 461                         *ret = -io_u->error;
 462                         clear_io_u(td, io_u);
 463                 } else if (io_u->resid) {
 464                         long long bytes = io_u->xfer_buflen - io_u->resid;
 465                         struct fio_file *f = io_u->file;
 466
 467                         if (bytes_issued)
 468                                 *bytes_issued += bytes;
 469
 470                         if (!from_verify)
 471                                 trim_io_piece(io_u);
 472
 473                         /*
 474                          * zero read, fail
 475                          */
 476                         if (!bytes) {
 477                                 if (!from_verify)
 478                                         unlog_io_piece(td, io_u);
 479                                 td_verror(td, EIO, "full resid");
 480                                 put_io_u(td, io_u);
 481                                 break;
 482                         }
 483
 484                         io_u->xfer_buflen = io_u->resid;
 485                         io_u->xfer_buf += bytes;
 486                         io_u->offset += bytes;
 487
 488                         if (ddir_rw(io_u->ddir))
 489                                 td->ts.short_io_u[io_u->ddir]++;
 490
 491                         if (io_u->offset == f->real_file_size)
 492                                 goto sync_done;
 493
 494                         requeue_io_u(td, &io_u);
 495                 } else {
 496 sync_done:
 497                         if (comp_time && should_check_rate(td))
 498                                 fio_gettime(comp_time, NULL);
 499
 500                         *ret = io_u_sync_complete(td, io_u);
 501                         if (*ret < 0)
 502                                 break;
 503                 }
 504
 505                 if (td->flags & TD_F_REGROW_LOGS)
 506                         regrow_logs(td);
 507
 508                 /*
 509                  * when doing I/O (not when verifying),
 510                  * check for any errors that are to be ignored
 511                  */
 512                 if (!from_verify)
 513                         break;
 514
 515                 return 0;
 516         case FIO_Q_QUEUED:
 517                 /*
 518                  * if the engine doesn't have a commit hook,
 519                  * the io_u is really queued. if it does have such
 520                  * a hook, it has to call io_u_queued() itself.
 521                  */
 522                 if (td->io_ops->commit == NULL)
 523                         io_u_queued(td, io_u);
 524                 if (bytes_issued)
 525                         *bytes_issued += io_u->xfer_buflen;
 526                 break;
 527         case FIO_Q_BUSY:
 528                 if (!from_verify)
 529                         unlog_io_piece(td, io_u);
 530                 requeue_io_u(td, &io_u);
 531                 td_io_commit(td);
 532                 break;
 533         default:
 534                 assert(*ret < 0);
 535                 td_verror(td, -(*ret), "td_io_queue");
 536                 break;
 537         }
 538
 539         if (break_on_this_error(td, ddir, ret))
 540                 return 1;
 541
 542         return 0;
 543 }
 544
 545 static inline bool io_in_polling(struct thread_data *td)
 546 {
 547         return !td->o.iodepth_batch_complete_min &&
 548                    !td->o.iodepth_batch_complete_max;
 549 }
 550 /*
 551  * Unlinks files from thread data fio_file structure
 552  */
 553 static int unlink_all_files(struct thread_data *td)
 554 {
 555         struct fio_file *f;
 556         unsigned int i;
 557         int ret = 0;
 558
 559         for_each_file(td, f, i) {
 560                 if (f->filetype != FIO_TYPE_FILE)
 561                         continue;
 562                 ret = td_io_unlink_file(td, f);
 563                 if (ret)
 564                         break;
 565         }
 566
 567         if (ret)
 568                 td_verror(td, ret, "unlink_all_files");
 569
 570         return ret;
 571 }
 572
 573 /*
 574  * Check if io_u will overlap an in-flight IO in the queue
 575  */
 576 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 577 {
 578         bool overlap;
 579         struct io_u *check_io_u;
 580         unsigned long long x1, x2, y1, y2;
 581         int i;
 582
 583         x1 = io_u->offset;
 584         x2 = io_u->offset + io_u->buflen;
 585         overlap = false;
 586         io_u_qiter(q, check_io_u, i) {
 587                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 588                         y1 = check_io_u->offset;
 589                         y2 = check_io_u->offset + check_io_u->buflen;
 590
 591                         if (x1 < y2 && y1 < x2) {
 592                                 overlap = true;
 593                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 594                                                 x1, io_u->buflen,
 595                                                 y1, check_io_u->buflen);
 596                                 break;
 597                         }
 598                 }
 599         }
 600
 601         return overlap;
 602 }
 603
 604 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 605 {
 606         /*
 607          * Check for overlap if the user asked us to, and we have
 608          * at least one IO in flight besides this one.
 609          */
 610         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 611             in_flight_overlap(&td->io_u_all, io_u))
 612                 return FIO_Q_BUSY;
 613
 614         return td_io_queue(td, io_u);
 615 }
 616
 617 /*
 618  * The main verify engine. Runs over the writes we previously submitted,
 619  * reads the blocks back in, and checks the crc/md5 of the data.
 620  */
 621 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 622 {
 623         struct fio_file *f;
 624         struct io_u *io_u;
 625         int ret, min_events;
 626         unsigned int i;
 627
 628         dprint(FD_VERIFY, "starting loop\n");
 629
 630         /*
 631          * sync io first and invalidate cache, to make sure we really
 632          * read from disk.
 633          */
 634         for_each_file(td, f, i) {
 635                 if (!fio_file_open(f))
 636                         continue;
 637                 if (fio_io_sync(td, f))
 638                         break;
 639                 if (file_invalidate_cache(td, f))
 640                         break;
 641         }
 642
 643         check_update_rusage(td);
 644
 645         if (td->error)
 646                 return;
 647
 648         /*
 649          * verify_state needs to be reset before verification
 650          * proceeds so that expected random seeds match actual
 651          * random seeds in headers. The main loop will reset
 652          * all random number generators if randrepeat is set.
 653          */
 654         if (!td->o.rand_repeatable)
 655                 td_fill_verify_state_seed(td);
 656
 657         td_set_runstate(td, TD_VERIFYING);
 658
 659         io_u = NULL;
 660         while (!td->terminate) {
 661                 enum fio_ddir ddir;
 662                 int full;
 663
 664                 update_ts_cache(td);
 665                 check_update_rusage(td);
 666
 667                 if (runtime_exceeded(td, &td->ts_cache)) {
 668                         __update_ts_cache(td);
 669                         if (runtime_exceeded(td, &td->ts_cache)) {
 670                                 fio_mark_td_terminate(td);
 671                                 break;
 672                         }
 673                 }
 674
 675                 if (flow_threshold_exceeded(td))
 676                         continue;
 677
 678                 if (!td->o.experimental_verify) {
 679                         io_u = __get_io_u(td);
 680                         if (!io_u)
 681                                 break;
 682
 683                         if (get_next_verify(td, io_u)) {
 684                                 put_io_u(td, io_u);
 685                                 break;
 686                         }
 687
 688                         if (td_io_prep(td, io_u)) {
 689                                 put_io_u(td, io_u);
 690                                 break;
 691                         }
 692                 } else {
 693                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 694                                 break;
 695
 696                         while ((io_u = get_io_u(td)) != NULL) {
 697                                 if (IS_ERR_OR_NULL(io_u)) {
 698                                         io_u = NULL;
 699                                         ret = FIO_Q_BUSY;
 700                                         goto reap;
 701                                 }
 702
 703                                 /*
 704                                  * We are only interested in the places where
 705                                  * we wrote or trimmed IOs. Turn those into
 706                                  * reads for verification purposes.
 707                                  */
 708                                 if (io_u->ddir == DDIR_READ) {
 709                                         /*
 710                                          * Pretend we issued it for rwmix
 711                                          * accounting
 712                                          */
 713                                         td->io_issues[DDIR_READ]++;
 714                                         put_io_u(td, io_u);
 715                                         continue;
 716                                 } else if (io_u->ddir == DDIR_TRIM) {
 717                                         io_u->ddir = DDIR_READ;
 718                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 719                                         break;
 720                                 } else if (io_u->ddir == DDIR_WRITE) {
 721                                         io_u->ddir = DDIR_READ;
 722                                         populate_verify_io_u(td, io_u);
 723                                         break;
 724                                 } else {
 725                                         put_io_u(td, io_u);
 726                                         continue;
 727                                 }
 728                         }
 729
 730                         if (!io_u)
 731                                 break;
 732                 }
 733
 734                 if (verify_state_should_stop(td, io_u)) {
 735                         put_io_u(td, io_u);
 736                         break;
 737                 }
 738
 739                 if (td->o.verify_async)
 740                         io_u->end_io = verify_io_u_async;
 741                 else
 742                         io_u->end_io = verify_io_u;
 743
 744                 ddir = io_u->ddir;
 745                 if (!td->o.disable_slat)
 746                         fio_gettime(&io_u->start_time, NULL);
 747
 748                 ret = io_u_submit(td, io_u);
 749
 750                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 751                         break;
 752
 753                 /*
 754                  * if we can queue more, do so. but check if there are
 755                  * completed io_u's first. Note that we can get BUSY even
 756                  * without IO queued, if the system is resource starved.
 757                  */
 758 reap:
 759                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 760                 if (full || io_in_polling(td))
 761                         ret = wait_for_completions(td, NULL);
 762
 763                 if (ret < 0)
 764                         break;
 765         }
 766
 767         check_update_rusage(td);
 768
 769         if (!td->error) {
 770                 min_events = td->cur_depth;
 771
 772                 if (min_events)
 773                         ret = io_u_queued_complete(td, min_events);
 774         } else
 775                 cleanup_pending_aio(td);
 776
 777         td_set_runstate(td, TD_RUNNING);
 778
 779         dprint(FD_VERIFY, "exiting loop\n");
 780 }
 781
 782 static bool exceeds_number_ios(struct thread_data *td)
 783 {
 784         unsigned long long number_ios;
 785
 786         if (!td->o.number_ios)
 787                 return false;
 788
 789         number_ios = ddir_rw_sum(td->io_blocks);
 790         number_ios += td->io_u_queued + td->io_u_in_flight;
 791
 792         return number_ios >= (td->o.number_ios * td->loops);
 793 }
 794
 795 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 796 {
 797         unsigned long long bytes, limit;
 798
 799         if (td_rw(td))
 800                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 801         else if (td_write(td))
 802                 bytes = this_bytes[DDIR_WRITE];
 803         else if (td_read(td))
 804                 bytes = this_bytes[DDIR_READ];
 805         else
 806                 bytes = this_bytes[DDIR_TRIM];
 807
 808         if (td->o.io_size)
 809                 limit = td->o.io_size;
 810         else
 811                 limit = td->o.size;
 812
 813         limit *= td->loops;
 814         return bytes >= limit || exceeds_number_ios(td);
 815 }
 816
 817 static bool io_issue_bytes_exceeded(struct thread_data *td)
 818 {
 819         return io_bytes_exceeded(td, td->io_issue_bytes);
 820 }
 821
 822 static bool io_complete_bytes_exceeded(struct thread_data *td)
 823 {
 824         return io_bytes_exceeded(td, td->this_io_bytes);
 825 }
 826
 827 /*
 828  * used to calculate the next io time for rate control
 829  *
 830  */
 831 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 832 {
 833         uint64_t bps = td->rate_bps[ddir];
 834
 835         assert(!(td->flags & TD_F_CHILD));
 836
 837         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 838                 uint64_t val, iops;
 839
 840                 iops = bps / td->o.bs[ddir];
 841                 val = (int64_t) (1000000 / iops) *
 842                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 843                 if (val) {
 844                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 845                                         (unsigned long long) 1000000 / val,
 846                                         ddir);
 847                 }
 848                 td->last_usec[ddir] += val;
 849                 return td->last_usec[ddir];
 850         } else if (bps) {
 851                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 852                 uint64_t secs = bytes / bps;
 853                 uint64_t remainder = bytes % bps;
 854
 855                 return remainder * 1000000 / bps + secs * 1000000;
 856         }
 857
 858         return 0;
 859 }
 860
 861 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
 862                              struct timespec *time)
 863 {
 864         unsigned long long b;
 865         uint64_t total;
 866         int left;
 867
 868         b = ddir_rw_sum(td->thinktime_blocks_counter);
 869         if (b % td->o.thinktime_blocks || !b)
 870                 return;
 871
 872         io_u_quiesce(td);
 873
 874         total = 0;
 875         if (td->o.thinktime_spin)
 876                 total = usec_spin(td->o.thinktime_spin);
 877
 878         left = td->o.thinktime - total;
 879         if (left)
 880                 total += usec_sleep(td, left);
 881
 882         /*
 883          * If we're ignoring thinktime for the rate, add the number of bytes
 884          * we would have done while sleeping, minus one block to ensure we
 885          * start issuing immediately after the sleep.
 886          */
 887         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 888                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 889                 uint64_t bs = td->o.min_bs[ddir];
 890                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 891                 uint64_t over;
 892
 893                 if (usperop <= total)
 894                         over = bs;
 895                 else
 896                         over = (usperop - total) / usperop * -bs;
 897
 898                 td->rate_io_issue_bytes[ddir] += (missed - over);
 899                 /* adjust for rate_process=poisson */
 900                 td->last_usec[ddir] += total;
 901         }
 902
 903         if (time && should_check_rate(td))
 904                 fio_gettime(time, NULL);
 905 }
 906
 907 /*
 908  * Main IO worker function. It retrieves io_u's to process and queues
 909  * and reaps them, checking for rate and errors along the way.
 910  *
 911  * Returns number of bytes written and trimmed.
 912  */
 913 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 914 {
 915         unsigned int i;
 916         int ret = 0;
 917         uint64_t total_bytes, bytes_issued = 0;
 918
 919         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 920                 bytes_done[i] = td->bytes_done[i];
 921
 922         if (in_ramp_time(td))
 923                 td_set_runstate(td, TD_RAMP);
 924         else
 925                 td_set_runstate(td, TD_RUNNING);
 926
 927         lat_target_init(td);
 928
 929         total_bytes = td->o.size;
 930         /*
 931         * Allow random overwrite workloads to write up to io_size
 932         * before starting verification phase as 'size' doesn't apply.
 933         */
 934         if (td_write(td) && td_random(td) && td->o.norandommap)
 935                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 936         /*
 937          * If verify_backlog is enabled, we'll run the verify in this
 938          * handler as well. For that case, we may need up to twice the
 939          * amount of bytes.
 940          */
 941         if (td->o.verify != VERIFY_NONE &&
 942            (td_write(td) && td->o.verify_backlog))
 943                 total_bytes += td->o.size;
 944
 945         /* In trimwrite mode, each byte is trimmed and then written, so
 946          * allow total_bytes to be twice as big */
 947         if (td_trimwrite(td))
 948                 total_bytes += td->total_io_size;
 949
 950         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 951                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 952                 td->o.time_based) {
 953                 struct timespec comp_time;
 954                 struct io_u *io_u;
 955                 int full;
 956                 enum fio_ddir ddir;
 957
 958                 check_update_rusage(td);
 959
 960                 if (td->terminate || td->done)
 961                         break;
 962
 963                 update_ts_cache(td);
 964
 965                 if (runtime_exceeded(td, &td->ts_cache)) {
 966                         __update_ts_cache(td);
 967                         if (runtime_exceeded(td, &td->ts_cache)) {
 968                                 fio_mark_td_terminate(td);
 969                                 break;
 970                         }
 971                 }
 972
 973                 if (flow_threshold_exceeded(td))
 974                         continue;
 975
 976                 /*
 977                  * Break if we exceeded the bytes. The exception is time
 978                  * based runs, but we still need to break out of the loop
 979                  * for those to run verification, if enabled.
 980                  * Jobs read from iolog do not use this stop condition.
 981                  */
 982                 if (bytes_issued >= total_bytes &&
 983                     !td->o.read_iolog_file &&
 984                     (!td->o.time_based ||
 985                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 986                         break;
 987
 988                 io_u = get_io_u(td);
 989                 if (IS_ERR_OR_NULL(io_u)) {
 990                         int err = PTR_ERR(io_u);
 991
 992                         io_u = NULL;
 993                         ddir = DDIR_INVAL;
 994                         if (err == -EBUSY) {
 995                                 ret = FIO_Q_BUSY;
 996                                 goto reap;
 997                         }
 998                         if (td->o.latency_target)
 999                                 goto reap;
1000                         break;
1001                 }
1002
1003                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1004                         populate_verify_io_u(td, io_u);
1005
1006                 ddir = io_u->ddir;
1007
1008                 /*
1009                  * Add verification end_io handler if:
1010                  *      - Asked to verify (!td_rw(td))
1011                  *      - Or the io_u is from our verify list (mixed write/ver)
1012                  */
1013                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1014                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1015
1016                         if (verify_state_should_stop(td, io_u)) {
1017                                 put_io_u(td, io_u);
1018                                 break;
1019                         }
1020
1021                         if (td->o.verify_async)
1022                                 io_u->end_io = verify_io_u_async;
1023                         else
1024                                 io_u->end_io = verify_io_u;
1025                         td_set_runstate(td, TD_VERIFYING);
1026                 } else if (in_ramp_time(td))
1027                         td_set_runstate(td, TD_RAMP);
1028                 else
1029                         td_set_runstate(td, TD_RUNNING);
1030
1031                 /*
1032                  * Always log IO before it's issued, so we know the specific
1033                  * order of it. The logged unit will track when the IO has
1034                  * completed.
1035                  */
1036                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1037                     td->o.do_verify &&
1038                     td->o.verify != VERIFY_NONE &&
1039                     !td->o.experimental_verify)
1040                         log_io_piece(td, io_u);
1041
1042                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1043                         const unsigned long long blen = io_u->xfer_buflen;
1044                         const enum fio_ddir __ddir = acct_ddir(io_u);
1045
1046                         if (td->error)
1047                                 break;
1048
1049                         workqueue_enqueue(&td->io_wq, &io_u->work);
1050                         ret = FIO_Q_QUEUED;
1051
1052                         if (ddir_rw(__ddir)) {
1053                                 td->io_issues[__ddir]++;
1054                                 td->io_issue_bytes[__ddir] += blen;
1055                                 td->rate_io_issue_bytes[__ddir] += blen;
1056                         }
1057
1058                         if (should_check_rate(td))
1059                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1060
1061                 } else {
1062                         ret = io_u_submit(td, io_u);
1063
1064                         if (should_check_rate(td))
1065                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1066
1067                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1068                                 break;
1069
1070                         /*
1071                          * See if we need to complete some commands. Note that
1072                          * we can get BUSY even without IO queued, if the
1073                          * system is resource starved.
1074                          */
1075 reap:
1076                         full = queue_full(td) ||
1077                                 (ret == FIO_Q_BUSY && td->cur_depth);
1078                         if (full || io_in_polling(td))
1079                                 ret = wait_for_completions(td, &comp_time);
1080                 }
1081                 if (ret < 0)
1082                         break;
1083
1084                 if (ddir_rw(ddir) && td->o.thinktime)
1085                         handle_thinktime(td, ddir, &comp_time);
1086
1087                 if (!ddir_rw_sum(td->bytes_done) &&
1088                     !td_ioengine_flagged(td, FIO_NOIO))
1089                         continue;
1090
1091                 if (!in_ramp_time(td) && should_check_rate(td)) {
1092                         if (check_min_rate(td, &comp_time)) {
1093                                 if (exitall_on_terminate || td->o.exitall_error)
1094                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1095                                 td_verror(td, EIO, "check_min_rate");
1096                                 break;
1097                         }
1098                 }
1099                 if (!in_ramp_time(td) && td->o.latency_target)
1100                         lat_target_check(td);
1101         }
1102
1103         check_update_rusage(td);
1104
1105         if (td->trim_entries)
1106                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1107
1108         if (td->o.fill_device && td->error == ENOSPC) {
1109                 td->error = 0;
1110                 fio_mark_td_terminate(td);
1111         }
1112         if (!td->error) {
1113                 struct fio_file *f;
1114
1115                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1116                         workqueue_flush(&td->io_wq);
1117                         i = 0;
1118                 } else
1119                         i = td->cur_depth;
1120
1121                 if (i) {
1122                         ret = io_u_queued_complete(td, i);
1123                         if (td->o.fill_device && td->error == ENOSPC)
1124                                 td->error = 0;
1125                 }
1126
1127                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1128                         td_set_runstate(td, TD_FSYNCING);
1129
1130                         for_each_file(td, f, i) {
1131                                 if (!fio_file_fsync(td, f))
1132                                         continue;
1133
1134                                 log_err("fio: end_fsync failed for file %s\n",
1135                                                                 f->file_name);
1136                         }
1137                 }
1138         } else
1139                 cleanup_pending_aio(td);
1140
1141         /*
1142          * stop job if we failed doing any IO
1143          */
1144         if (!ddir_rw_sum(td->this_io_bytes))
1145                 td->done = 1;
1146
1147         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1148                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1149 }
1150
1151 static void free_file_completion_logging(struct thread_data *td)
1152 {
1153         struct fio_file *f;
1154         unsigned int i;
1155
1156         for_each_file(td, f, i) {
1157                 if (!f->last_write_comp)
1158                         break;
1159                 sfree(f->last_write_comp);
1160         }
1161 }
1162
1163 static int init_file_completion_logging(struct thread_data *td,
1164                                         unsigned int depth)
1165 {
1166         struct fio_file *f;
1167         unsigned int i;
1168
1169         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1170                 return 0;
1171
1172         for_each_file(td, f, i) {
1173                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1174                 if (!f->last_write_comp)
1175                         goto cleanup;
1176         }
1177
1178         return 0;
1179
1180 cleanup:
1181         free_file_completion_logging(td);
1182         log_err("fio: failed to alloc write comp data\n");
1183         return 1;
1184 }
1185
1186 static void cleanup_io_u(struct thread_data *td)
1187 {
1188         struct io_u *io_u;
1189
1190         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1191
1192                 if (td->io_ops->io_u_free)
1193                         td->io_ops->io_u_free(td, io_u);
1194
1195                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1196         }
1197
1198         free_io_mem(td);
1199
1200         io_u_rexit(&td->io_u_requeues);
1201         io_u_qexit(&td->io_u_freelist, false);
1202         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1203
1204         free_file_completion_logging(td);
1205 }
1206
1207 static int init_io_u(struct thread_data *td)
1208 {
1209         struct io_u *io_u;
1210         int cl_align, i, max_units;
1211         int err;
1212
1213         max_units = td->o.iodepth;
1214
1215         err = 0;
1216         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1217         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1218         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1219
1220         if (err) {
1221                 log_err("fio: failed setting up IO queues\n");
1222                 return 1;
1223         }
1224
1225         cl_align = os_cache_line_size();
1226
1227         for (i = 0; i < max_units; i++) {
1228                 void *ptr;
1229
1230                 if (td->terminate)
1231                         return 1;
1232
1233                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1234                 if (!ptr) {
1235                         log_err("fio: unable to allocate aligned memory\n");
1236                         return 1;
1237                 }
1238
1239                 io_u = ptr;
1240                 memset(io_u, 0, sizeof(*io_u));
1241                 INIT_FLIST_HEAD(&io_u->verify_list);
1242                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1243
1244                 io_u->index = i;
1245                 io_u->flags = IO_U_F_FREE;
1246                 io_u_qpush(&td->io_u_freelist, io_u);
1247
1248                 /*
1249                  * io_u never leaves this stack, used for iteration of all
1250                  * io_u buffers.
1251                  */
1252                 io_u_qpush(&td->io_u_all, io_u);
1253
1254                 if (td->io_ops->io_u_init) {
1255                         int ret = td->io_ops->io_u_init(td, io_u);
1256
1257                         if (ret) {
1258                                 log_err("fio: failed to init engine data: %d\n", ret);
1259                                 return 1;
1260                         }
1261                 }
1262         }
1263
1264         init_io_u_buffers(td);
1265
1266         if (init_file_completion_logging(td, max_units))
1267                 return 1;
1268
1269         return 0;
1270 }
1271
1272 int init_io_u_buffers(struct thread_data *td)
1273 {
1274         struct io_u *io_u;
1275         unsigned long long max_bs, min_write;
1276         int i, max_units;
1277         int data_xfer = 1;
1278         char *p;
1279
1280         max_units = td->o.iodepth;
1281         max_bs = td_max_bs(td);
1282         min_write = td->o.min_bs[DDIR_WRITE];
1283         td->orig_buffer_size = (unsigned long long) max_bs
1284                                         * (unsigned long long) max_units;
1285
1286         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1287                 data_xfer = 0;
1288
1289         /*
1290          * if we may later need to do address alignment, then add any
1291          * possible adjustment here so that we don't cause a buffer
1292          * overflow later. this adjustment may be too much if we get
1293          * lucky and the allocator gives us an aligned address.
1294          */
1295         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1296             td_ioengine_flagged(td, FIO_RAWIO))
1297                 td->orig_buffer_size += page_mask + td->o.mem_align;
1298
1299         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1300                 unsigned long long bs;
1301
1302                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1303                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1304         }
1305
1306         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1307                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1308                 return 1;
1309         }
1310
1311         if (data_xfer && allocate_io_mem(td))
1312                 return 1;
1313
1314         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1315             td_ioengine_flagged(td, FIO_RAWIO))
1316                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1317         else
1318                 p = td->orig_buffer;
1319
1320         for (i = 0; i < max_units; i++) {
1321                 io_u = td->io_u_all.io_us[i];
1322                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1323
1324                 if (data_xfer) {
1325                         io_u->buf = p;
1326                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1327
1328                         if (td_write(td))
1329                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1330                         if (td_write(td) && td->o.verify_pattern_bytes) {
1331                                 /*
1332                                  * Fill the buffer with the pattern if we are
1333                                  * going to be doing writes.
1334                                  */
1335                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1336                         }
1337                 }
1338                 p += max_bs;
1339         }
1340
1341         return 0;
1342 }
1343
1344 /*
1345  * This function is Linux specific.
1346  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1347  */
1348 static int switch_ioscheduler(struct thread_data *td)
1349 {
1350 #ifdef FIO_HAVE_IOSCHED_SWITCH
1351         char tmp[256], tmp2[128], *p;
1352         FILE *f;
1353         int ret;
1354
1355         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1356                 return 0;
1357
1358         assert(td->files && td->files[0]);
1359         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1360
1361         f = fopen(tmp, "r+");
1362         if (!f) {
1363                 if (errno == ENOENT) {
1364                         log_err("fio: os or kernel doesn't support IO scheduler"
1365                                 " switching\n");
1366                         return 0;
1367                 }
1368                 td_verror(td, errno, "fopen iosched");
1369                 return 1;
1370         }
1371
1372         /*
1373          * Set io scheduler.
1374          */
1375         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1376         if (ferror(f) || ret != 1) {
1377                 td_verror(td, errno, "fwrite");
1378                 fclose(f);
1379                 return 1;
1380         }
1381
1382         rewind(f);
1383
1384         /*
1385          * Read back and check that the selected scheduler is now the default.
1386          */
1387         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1388         if (ferror(f) || ret < 0) {
1389                 td_verror(td, errno, "fread");
1390                 fclose(f);
1391                 return 1;
1392         }
1393         tmp[ret] = '\0';
1394         /*
1395          * either a list of io schedulers or "none\n" is expected. Strip the
1396          * trailing newline.
1397          */
1398         p = tmp;
1399         strsep(&p, "\n");
1400
1401         /*
1402          * Write to "none" entry doesn't fail, so check the result here.
1403          */
1404         if (!strcmp(tmp, "none")) {
1405                 log_err("fio: io scheduler is not tunable\n");
1406                 fclose(f);
1407                 return 0;
1408         }
1409
1410         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1411         if (!strstr(tmp, tmp2)) {
1412                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1413                 td_verror(td, EINVAL, "iosched_switch");
1414                 fclose(f);
1415                 return 1;
1416         }
1417
1418         fclose(f);
1419         return 0;
1420 #else
1421         return 0;
1422 #endif
1423 }
1424
1425 static bool keep_running(struct thread_data *td)
1426 {
1427         unsigned long long limit;
1428
1429         if (td->done)
1430                 return false;
1431         if (td->terminate)
1432                 return false;
1433         if (td->o.time_based)
1434                 return true;
1435         if (td->o.loops) {
1436                 td->o.loops--;
1437                 return true;
1438         }
1439         if (exceeds_number_ios(td))
1440                 return false;
1441
1442         if (td->o.io_size)
1443                 limit = td->o.io_size;
1444         else
1445                 limit = td->o.size;
1446
1447         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1448                 uint64_t diff;
1449
1450                 /*
1451                  * If the difference is less than the maximum IO size, we
1452                  * are done.
1453                  */
1454                 diff = limit - ddir_rw_sum(td->io_bytes);
1455                 if (diff < td_max_bs(td))
1456                         return false;
1457
1458                 if (fio_files_done(td) && !td->o.io_size)
1459                         return false;
1460
1461                 return true;
1462         }
1463
1464         return false;
1465 }
1466
1467 static int exec_string(struct thread_options *o, const char *string,
1468                        const char *mode)
1469 {
1470         int ret;
1471         char *str;
1472
1473         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1474                 return -1;
1475
1476         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1477                  o->name, mode);
1478         ret = system(str);
1479         if (ret == -1)
1480                 log_err("fio: exec of cmd <%s> failed\n", str);
1481
1482         free(str);
1483         return ret;
1484 }
1485
1486 /*
1487  * Dry run to compute correct state of numberio for verification.
1488  */
1489 static uint64_t do_dry_run(struct thread_data *td)
1490 {
1491         td_set_runstate(td, TD_RUNNING);
1492
1493         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1494                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1495                 struct io_u *io_u;
1496                 int ret;
1497
1498                 if (td->terminate || td->done)
1499                         break;
1500
1501                 io_u = get_io_u(td);
1502                 if (IS_ERR_OR_NULL(io_u))
1503                         break;
1504
1505                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1506                 io_u->error = 0;
1507                 io_u->resid = 0;
1508                 if (ddir_rw(acct_ddir(io_u)))
1509                         td->io_issues[acct_ddir(io_u)]++;
1510                 if (ddir_rw(io_u->ddir)) {
1511                         io_u_mark_depth(td, 1);
1512                         td->ts.total_io_u[io_u->ddir]++;
1513                 }
1514
1515                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1516                     td->o.do_verify &&
1517                     td->o.verify != VERIFY_NONE &&
1518                     !td->o.experimental_verify)
1519                         log_io_piece(td, io_u);
1520
1521                 ret = io_u_sync_complete(td, io_u);
1522                 (void) ret;
1523         }
1524
1525         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1526 }
1527
1528 struct fork_data {
1529         struct thread_data *td;
1530         struct sk_out *sk_out;
1531 };
1532
1533 /*
1534  * Entry point for the thread based jobs. The process based jobs end up
1535  * here as well, after a little setup.
1536  */
1537 static void *thread_main(void *data)
1538 {
1539         struct fork_data *fd = data;
1540         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1541         struct thread_data *td = fd->td;
1542         struct thread_options *o = &td->o;
1543         struct sk_out *sk_out = fd->sk_out;
1544         uint64_t bytes_done[DDIR_RWDIR_CNT];
1545         int deadlock_loop_cnt;
1546         bool clear_state;
1547         int res, ret;
1548
1549         sk_out_assign(sk_out);
1550         free(fd);
1551
1552         if (!o->use_thread) {
1553                 setsid();
1554                 td->pid = getpid();
1555         } else
1556                 td->pid = gettid();
1557
1558         fio_local_clock_init();
1559
1560         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1561
1562         if (is_backend)
1563                 fio_server_send_start(td);
1564
1565         INIT_FLIST_HEAD(&td->io_log_list);
1566         INIT_FLIST_HEAD(&td->io_hist_list);
1567         INIT_FLIST_HEAD(&td->verify_list);
1568         INIT_FLIST_HEAD(&td->trim_list);
1569         td->io_hist_tree = RB_ROOT;
1570
1571         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1572         if (ret) {
1573                 td_verror(td, ret, "mutex_cond_init_pshared");
1574                 goto err;
1575         }
1576         ret = cond_init_pshared(&td->verify_cond);
1577         if (ret) {
1578                 td_verror(td, ret, "mutex_cond_pshared");
1579                 goto err;
1580         }
1581
1582         td_set_runstate(td, TD_INITIALIZED);
1583         dprint(FD_MUTEX, "up startup_sem\n");
1584         fio_sem_up(startup_sem);
1585         dprint(FD_MUTEX, "wait on td->sem\n");
1586         fio_sem_down(td->sem);
1587         dprint(FD_MUTEX, "done waiting on td->sem\n");
1588
1589         /*
1590          * A new gid requires privilege, so we need to do this before setting
1591          * the uid.
1592          */
1593         if (o->gid != -1U && setgid(o->gid)) {
1594                 td_verror(td, errno, "setgid");
1595                 goto err;
1596         }
1597         if (o->uid != -1U && setuid(o->uid)) {
1598                 td_verror(td, errno, "setuid");
1599                 goto err;
1600         }
1601
1602         td_zone_gen_index(td);
1603
1604         /*
1605          * Do this early, we don't want the compress threads to be limited
1606          * to the same CPUs as the IO workers. So do this before we set
1607          * any potential CPU affinity
1608          */
1609         if (iolog_compress_init(td, sk_out))
1610                 goto err;
1611
1612         /*
1613          * If we have a gettimeofday() thread, make sure we exclude that
1614          * thread from this job
1615          */
1616         if (o->gtod_cpu)
1617                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1618
1619         /*
1620          * Set affinity first, in case it has an impact on the memory
1621          * allocations.
1622          */
1623         if (fio_option_is_set(o, cpumask)) {
1624                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1625                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1626                         if (!ret) {
1627                                 log_err("fio: no CPUs set\n");
1628                                 log_err("fio: Try increasing number of available CPUs\n");
1629                                 td_verror(td, EINVAL, "cpus_split");
1630                                 goto err;
1631                         }
1632                 }
1633                 ret = fio_setaffinity(td->pid, o->cpumask);
1634                 if (ret == -1) {
1635                         td_verror(td, errno, "cpu_set_affinity");
1636                         goto err;
1637                 }
1638         }
1639
1640 #ifdef CONFIG_LIBNUMA
1641         /* numa node setup */
1642         if (fio_option_is_set(o, numa_cpunodes) ||
1643             fio_option_is_set(o, numa_memnodes)) {
1644                 struct bitmask *mask;
1645
1646                 if (numa_available() < 0) {
1647                         td_verror(td, errno, "Does not support NUMA API\n");
1648                         goto err;
1649                 }
1650
1651                 if (fio_option_is_set(o, numa_cpunodes)) {
1652                         mask = numa_parse_nodestring(o->numa_cpunodes);
1653                         ret = numa_run_on_node_mask(mask);
1654                         numa_free_nodemask(mask);
1655                         if (ret == -1) {
1656                                 td_verror(td, errno, \
1657                                         "numa_run_on_node_mask failed\n");
1658                                 goto err;
1659                         }
1660                 }
1661
1662                 if (fio_option_is_set(o, numa_memnodes)) {
1663                         mask = NULL;
1664                         if (o->numa_memnodes)
1665                                 mask = numa_parse_nodestring(o->numa_memnodes);
1666
1667                         switch (o->numa_mem_mode) {
1668                         case MPOL_INTERLEAVE:
1669                                 numa_set_interleave_mask(mask);
1670                                 break;
1671                         case MPOL_BIND:
1672                                 numa_set_membind(mask);
1673                                 break;
1674                         case MPOL_LOCAL:
1675                                 numa_set_localalloc();
1676                                 break;
1677                         case MPOL_PREFERRED:
1678                                 numa_set_preferred(o->numa_mem_prefer_node);
1679                                 break;
1680                         case MPOL_DEFAULT:
1681                         default:
1682                                 break;
1683                         }
1684
1685                         if (mask)
1686                                 numa_free_nodemask(mask);
1687
1688                 }
1689         }
1690 #endif
1691
1692         if (fio_pin_memory(td))
1693                 goto err;
1694
1695         /*
1696          * May alter parameters that init_io_u() will use, so we need to
1697          * do this first.
1698          */
1699         if (!init_iolog(td))
1700                 goto err;
1701
1702         if (td_io_init(td))
1703                 goto err;
1704
1705         if (init_io_u(td))
1706                 goto err;
1707
1708         if (td->io_ops->post_init && td->io_ops->post_init(td))
1709                 goto err;
1710
1711         if (o->verify_async && verify_async_init(td))
1712                 goto err;
1713
1714         if (fio_option_is_set(o, ioprio) ||
1715             fio_option_is_set(o, ioprio_class)) {
1716                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1717                 if (ret == -1) {
1718                         td_verror(td, errno, "ioprio_set");
1719                         goto err;
1720                 }
1721         }
1722
1723         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1724                 goto err;
1725
1726         errno = 0;
1727         if (nice(o->nice) == -1 && errno != 0) {
1728                 td_verror(td, errno, "nice");
1729                 goto err;
1730         }
1731
1732         if (o->ioscheduler && switch_ioscheduler(td))
1733                 goto err;
1734
1735         if (!o->create_serialize && setup_files(td))
1736                 goto err;
1737
1738         if (!init_random_map(td))
1739                 goto err;
1740
1741         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1742                 goto err;
1743
1744         if (o->pre_read && !pre_read_files(td))
1745                 goto err;
1746
1747         fio_verify_init(td);
1748
1749         if (rate_submit_init(td, sk_out))
1750                 goto err;
1751
1752         if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
1753                 td->thinktime_blocks_counter = td->io_blocks;
1754         else
1755                 td->thinktime_blocks_counter = td->io_issues;
1756
1757         set_epoch_time(td, o->log_unix_epoch);
1758         fio_getrusage(&td->ru_start);
1759         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1760         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1761         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1762
1763         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1764                         o->ratemin[DDIR_TRIM]) {
1765                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1766                                         sizeof(td->bw_sample_time));
1767                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1768                                         sizeof(td->bw_sample_time));
1769                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1770                                         sizeof(td->bw_sample_time));
1771         }
1772
1773         memset(bytes_done, 0, sizeof(bytes_done));
1774         clear_state = false;
1775
1776         while (keep_running(td)) {
1777                 uint64_t verify_bytes;
1778
1779                 fio_gettime(&td->start, NULL);
1780                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1781
1782                 if (clear_state) {
1783                         clear_io_state(td, 0);
1784
1785                         if (o->unlink_each_loop && unlink_all_files(td))
1786                                 break;
1787                 }
1788
1789                 prune_io_piece_log(td);
1790
1791                 if (td->o.verify_only && td_write(td))
1792                         verify_bytes = do_dry_run(td);
1793                 else {
1794                         do_io(td, bytes_done);
1795
1796                         if (!ddir_rw_sum(bytes_done)) {
1797                                 fio_mark_td_terminate(td);
1798                                 verify_bytes = 0;
1799                         } else {
1800                                 verify_bytes = bytes_done[DDIR_WRITE] +
1801                                                 bytes_done[DDIR_TRIM];
1802                         }
1803                 }
1804
1805                 /*
1806                  * If we took too long to shut down, the main thread could
1807                  * already consider us reaped/exited. If that happens, break
1808                  * out and clean up.
1809                  */
1810                 if (td->runstate >= TD_EXITED)
1811                         break;
1812
1813                 clear_state = true;
1814
1815                 /*
1816                  * Make sure we've successfully updated the rusage stats
1817                  * before waiting on the stat mutex. Otherwise we could have
1818                  * the stat thread holding stat mutex and waiting for
1819                  * the rusage_sem, which would never get upped because
1820                  * this thread is waiting for the stat mutex.
1821                  */
1822                 deadlock_loop_cnt = 0;
1823                 do {
1824                         check_update_rusage(td);
1825                         if (!fio_sem_down_trylock(stat_sem))
1826                                 break;
1827                         usleep(1000);
1828                         if (deadlock_loop_cnt++ > 5000) {
1829                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1830                                 td->error = EDEADLK;
1831                                 goto err;
1832                         }
1833                 } while (1);
1834
1835                 if (td_read(td) && td->io_bytes[DDIR_READ])
1836                         update_runtime(td, elapsed_us, DDIR_READ);
1837                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1838                         update_runtime(td, elapsed_us, DDIR_WRITE);
1839                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1840                         update_runtime(td, elapsed_us, DDIR_TRIM);
1841                 fio_gettime(&td->start, NULL);
1842                 fio_sem_up(stat_sem);
1843
1844                 if (td->error || td->terminate)
1845                         break;
1846
1847                 if (!o->do_verify ||
1848                     o->verify == VERIFY_NONE ||
1849                     td_ioengine_flagged(td, FIO_UNIDIR))
1850                         continue;
1851
1852                 clear_io_state(td, 0);
1853
1854                 fio_gettime(&td->start, NULL);
1855
1856                 do_verify(td, verify_bytes);
1857
1858                 /*
1859                  * See comment further up for why this is done here.
1860                  */
1861                 check_update_rusage(td);
1862
1863                 fio_sem_down(stat_sem);
1864                 update_runtime(td, elapsed_us, DDIR_READ);
1865                 fio_gettime(&td->start, NULL);
1866                 fio_sem_up(stat_sem);
1867
1868                 if (td->error || td->terminate)
1869                         break;
1870         }
1871
1872         /*
1873          * Acquire this lock if we were doing overlap checking in
1874          * offload mode so that we don't clean up this job while
1875          * another thread is checking its io_u's for overlap
1876          */
1877         if (td_offload_overlap(td)) {
1878                 int res = pthread_mutex_lock(&overlap_check);
1879                 assert(res == 0);
1880         }
1881         td_set_runstate(td, TD_FINISHING);
1882         if (td_offload_overlap(td)) {
1883                 res = pthread_mutex_unlock(&overlap_check);
1884                 assert(res == 0);
1885         }
1886
1887         update_rusage_stat(td);
1888         td->ts.total_run_time = mtime_since_now(&td->epoch);
1889         for_each_rw_ddir(ddir) {
1890                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1891         }
1892
1893         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1894             (td->o.verify != VERIFY_NONE && td_write(td)))
1895                 verify_save_state(td->thread_number);
1896
1897         fio_unpin_memory(td);
1898
1899         td_writeout_logs(td, true);
1900
1901         iolog_compress_exit(td);
1902         rate_submit_exit(td);
1903
1904         if (o->exec_postrun)
1905                 exec_string(o, o->exec_postrun, "postrun");
1906
1907         if (exitall_on_terminate || (o->exitall_error && td->error))
1908                 fio_terminate_threads(td->groupid, td->o.exit_what);
1909
1910 err:
1911         if (td->error)
1912                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1913                                                         td->verror);
1914
1915         if (o->verify_async)
1916                 verify_async_exit(td);
1917
1918         close_and_free_files(td);
1919         cleanup_io_u(td);
1920         close_ioengine(td);
1921         cgroup_shutdown(td, cgroup_mnt);
1922         verify_free_state(td);
1923         td_zone_free_index(td);
1924
1925         if (fio_option_is_set(o, cpumask)) {
1926                 ret = fio_cpuset_exit(&o->cpumask);
1927                 if (ret)
1928                         td_verror(td, ret, "fio_cpuset_exit");
1929         }
1930
1931         /*
1932          * do this very late, it will log file closing as well
1933          */
1934         if (o->write_iolog_file)
1935                 write_iolog_close(td);
1936         if (td->io_log_rfile)
1937                 fclose(td->io_log_rfile);
1938
1939         td_set_runstate(td, TD_EXITED);
1940
1941         /*
1942          * Do this last after setting our runstate to exited, so we
1943          * know that the stat thread is signaled.
1944          */
1945         check_update_rusage(td);
1946
1947         sk_out_drop();
1948         return (void *) (uintptr_t) td->error;
1949 }
1950
1951 /*
1952  * Run over the job map and reap the threads that have exited, if any.
1953  */
1954 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1955                          uint64_t *m_rate)
1956 {
1957         struct thread_data *td;
1958         unsigned int cputhreads, realthreads, pending;
1959         int i, status, ret;
1960
1961         /*
1962          * reap exited threads (TD_EXITED -> TD_REAPED)
1963          */
1964         realthreads = pending = cputhreads = 0;
1965         for_each_td(td, i) {
1966                 int flags = 0;
1967
1968                  if (!strcmp(td->o.ioengine, "cpuio"))
1969                         cputhreads++;
1970                 else
1971                         realthreads++;
1972
1973                 if (!td->pid) {
1974                         pending++;
1975                         continue;
1976                 }
1977                 if (td->runstate == TD_REAPED)
1978                         continue;
1979                 if (td->o.use_thread) {
1980                         if (td->runstate == TD_EXITED) {
1981                                 td_set_runstate(td, TD_REAPED);
1982                                 goto reaped;
1983                         }
1984                         continue;
1985                 }
1986
1987                 flags = WNOHANG;
1988                 if (td->runstate == TD_EXITED)
1989                         flags = 0;
1990
1991                 /*
1992                  * check if someone quit or got killed in an unusual way
1993                  */
1994                 ret = waitpid(td->pid, &status, flags);
1995                 if (ret < 0) {
1996                         if (errno == ECHILD) {
1997                                 log_err("fio: pid=%d disappeared %d\n",
1998                                                 (int) td->pid, td->runstate);
1999                                 td->sig = ECHILD;
2000                                 td_set_runstate(td, TD_REAPED);
2001                                 goto reaped;
2002                         }
2003                         perror("waitpid");
2004                 } else if (ret == td->pid) {
2005                         if (WIFSIGNALED(status)) {
2006                                 int sig = WTERMSIG(status);
2007
2008                                 if (sig != SIGTERM && sig != SIGUSR2)
2009                                         log_err("fio: pid=%d, got signal=%d\n",
2010                                                         (int) td->pid, sig);
2011                                 td->sig = sig;
2012                                 td_set_runstate(td, TD_REAPED);
2013                                 goto reaped;
2014                         }
2015                         if (WIFEXITED(status)) {
2016                                 if (WEXITSTATUS(status) && !td->error)
2017                                         td->error = WEXITSTATUS(status);
2018
2019                                 td_set_runstate(td, TD_REAPED);
2020                                 goto reaped;
2021                         }
2022                 }
2023
2024                 /*
2025                  * If the job is stuck, do a forceful timeout of it and
2026                  * move on.
2027                  */
2028                 if (td->terminate &&
2029                     td->runstate < TD_FSYNCING &&
2030                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2031                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2032                                 "%lu seconds, it appears to be stuck. Doing "
2033                                 "forceful exit of this job.\n",
2034                                 td->o.name, td->runstate,
2035                                 (unsigned long) time_since_now(&td->terminate_time));
2036                         td_set_runstate(td, TD_REAPED);
2037                         goto reaped;
2038                 }
2039
2040                 /*
2041                  * thread is not dead, continue
2042                  */
2043                 pending++;
2044                 continue;
2045 reaped:
2046                 (*nr_running)--;
2047                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2048                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2049                 if (!td->pid)
2050                         pending--;
2051
2052                 if (td->error)
2053                         exit_value++;
2054
2055                 done_secs += mtime_since_now(&td->epoch) / 1000;
2056                 profile_td_exit(td);
2057                 flow_exit_job(td);
2058         }
2059
2060         if (*nr_running == cputhreads && !pending && realthreads)
2061                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2062 }
2063
2064 static bool __check_trigger_file(void)
2065 {
2066         struct stat sb;
2067
2068         if (!trigger_file)
2069                 return false;
2070
2071         if (stat(trigger_file, &sb))
2072                 return false;
2073
2074         if (unlink(trigger_file) < 0)
2075                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2076                                                         strerror(errno));
2077
2078         return true;
2079 }
2080
2081 static bool trigger_timedout(void)
2082 {
2083         if (trigger_timeout)
2084                 if (time_since_genesis() >= trigger_timeout) {
2085                         trigger_timeout = 0;
2086                         return true;
2087                 }
2088
2089         return false;
2090 }
2091
2092 void exec_trigger(const char *cmd)
2093 {
2094         int ret;
2095
2096         if (!cmd || cmd[0] == '\0')
2097                 return;
2098
2099         ret = system(cmd);
2100         if (ret == -1)
2101                 log_err("fio: failed executing %s trigger\n", cmd);
2102 }
2103
2104 void check_trigger_file(void)
2105 {
2106         if (__check_trigger_file() || trigger_timedout()) {
2107                 if (nr_clients)
2108                         fio_clients_send_trigger(trigger_remote_cmd);
2109                 else {
2110                         verify_save_state(IO_LIST_ALL);
2111                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2112                         exec_trigger(trigger_cmd);
2113                 }
2114         }
2115 }
2116
2117 static int fio_verify_load_state(struct thread_data *td)
2118 {
2119         int ret;
2120
2121         if (!td->o.verify_state)
2122                 return 0;
2123
2124         if (is_backend) {
2125                 void *data;
2126
2127                 ret = fio_server_get_verify_state(td->o.name,
2128                                         td->thread_number - 1, &data);
2129                 if (!ret)
2130                         verify_assign_state(td, data);
2131         } else {
2132                 char prefix[PATH_MAX];
2133
2134                 if (aux_path)
2135                         sprintf(prefix, "%s%clocal", aux_path,
2136                                         FIO_OS_PATH_SEPARATOR);
2137                 else
2138                         strcpy(prefix, "local");
2139                 ret = verify_load_state(td, prefix);
2140         }
2141
2142         return ret;
2143 }
2144
2145 static void do_usleep(unsigned int usecs)
2146 {
2147         check_for_running_stats();
2148         check_trigger_file();
2149         usleep(usecs);
2150 }
2151
2152 static bool check_mount_writes(struct thread_data *td)
2153 {
2154         struct fio_file *f;
2155         unsigned int i;
2156
2157         if (!td_write(td) || td->o.allow_mounted_write)
2158                 return false;
2159
2160         /*
2161          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2162          * are mkfs'd and mounted.
2163          */
2164         for_each_file(td, f, i) {
2165 #ifdef FIO_HAVE_CHARDEV_SIZE
2166                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2167 #else
2168                 if (f->filetype != FIO_TYPE_BLOCK)
2169 #endif
2170                         continue;
2171                 if (device_is_mounted(f->file_name))
2172                         goto mounted;
2173         }
2174
2175         return false;
2176 mounted:
2177         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2178         return true;
2179 }
2180
2181 static bool waitee_running(struct thread_data *me)
2182 {
2183         const char *waitee = me->o.wait_for;
2184         const char *self = me->o.name;
2185         struct thread_data *td;
2186         int i;
2187
2188         if (!waitee)
2189                 return false;
2190
2191         for_each_td(td, i) {
2192                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2193                         continue;
2194
2195                 if (td->runstate < TD_EXITED) {
2196                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2197                                         self, td->o.name,
2198                                         runstate_to_name(td->runstate));
2199                         return true;
2200                 }
2201         }
2202
2203         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2204         return false;
2205 }
2206
2207 /*
2208  * Main function for kicking off and reaping jobs, as needed.
2209  */
2210 static void run_threads(struct sk_out *sk_out)
2211 {
2212         struct thread_data *td;
2213         unsigned int i, todo, nr_running, nr_started;
2214         uint64_t m_rate, t_rate;
2215         uint64_t spent;
2216
2217         if (fio_gtod_offload && fio_start_gtod_thread())
2218                 return;
2219
2220         fio_idle_prof_init();
2221
2222         set_sig_handlers();
2223
2224         nr_thread = nr_process = 0;
2225         for_each_td(td, i) {
2226                 if (check_mount_writes(td))
2227                         return;
2228                 if (td->o.use_thread)
2229                         nr_thread++;
2230                 else
2231                         nr_process++;
2232         }
2233
2234         if (output_format & FIO_OUTPUT_NORMAL) {
2235                 struct buf_output out;
2236
2237                 buf_output_init(&out);
2238                 __log_buf(&out, "Starting ");
2239                 if (nr_thread)
2240                         __log_buf(&out, "%d thread%s", nr_thread,
2241                                                 nr_thread > 1 ? "s" : "");
2242                 if (nr_process) {
2243                         if (nr_thread)
2244                                 __log_buf(&out, " and ");
2245                         __log_buf(&out, "%d process%s", nr_process,
2246                                                 nr_process > 1 ? "es" : "");
2247                 }
2248                 __log_buf(&out, "\n");
2249                 log_info_buf(out.buf, out.buflen);
2250                 buf_output_free(&out);
2251         }
2252
2253         todo = thread_number;
2254         nr_running = 0;
2255         nr_started = 0;
2256         m_rate = t_rate = 0;
2257
2258         for_each_td(td, i) {
2259                 print_status_init(td->thread_number - 1);
2260
2261                 if (!td->o.create_serialize)
2262                         continue;
2263
2264                 if (fio_verify_load_state(td))
2265                         goto reap;
2266
2267                 /*
2268                  * do file setup here so it happens sequentially,
2269                  * we don't want X number of threads getting their
2270                  * client data interspersed on disk
2271                  */
2272                 if (setup_files(td)) {
2273 reap:
2274                         exit_value++;
2275                         if (td->error)
2276                                 log_err("fio: pid=%d, err=%d/%s\n",
2277                                         (int) td->pid, td->error, td->verror);
2278                         td_set_runstate(td, TD_REAPED);
2279                         todo--;
2280                 } else {
2281                         struct fio_file *f;
2282                         unsigned int j;
2283
2284                         /*
2285                          * for sharing to work, each job must always open
2286                          * its own files. so close them, if we opened them
2287                          * for creation
2288                          */
2289                         for_each_file(td, f, j) {
2290                                 if (fio_file_open(f))
2291                                         td_io_close_file(td, f);
2292                         }
2293                 }
2294         }
2295
2296         /* start idle threads before io threads start to run */
2297         fio_idle_prof_start();
2298
2299         set_genesis_time();
2300
2301         while (todo) {
2302                 struct thread_data *map[REAL_MAX_JOBS];
2303                 struct timespec this_start;
2304                 int this_jobs = 0, left;
2305                 struct fork_data *fd;
2306
2307                 /*
2308                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2309                  */
2310                 for_each_td(td, i) {
2311                         if (td->runstate != TD_NOT_CREATED)
2312                                 continue;
2313
2314                         /*
2315                          * never got a chance to start, killed by other
2316                          * thread for some reason
2317                          */
2318                         if (td->terminate) {
2319                                 todo--;
2320                                 continue;
2321                         }
2322
2323                         if (td->o.start_delay) {
2324                                 spent = utime_since_genesis();
2325
2326                                 if (td->o.start_delay > spent)
2327                                         continue;
2328                         }
2329
2330                         if (td->o.stonewall && (nr_started || nr_running)) {
2331                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2332                                                         td->o.name);
2333                                 break;
2334                         }
2335
2336                         if (waitee_running(td)) {
2337                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2338                                                 td->o.name, td->o.wait_for);
2339                                 continue;
2340                         }
2341
2342                         init_disk_util(td);
2343
2344                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2345                         td->update_rusage = 0;
2346
2347                         /*
2348                          * Set state to created. Thread will transition
2349                          * to TD_INITIALIZED when it's done setting up.
2350                          */
2351                         td_set_runstate(td, TD_CREATED);
2352                         map[this_jobs++] = td;
2353                         nr_started++;
2354
2355                         fd = calloc(1, sizeof(*fd));
2356                         fd->td = td;
2357                         fd->sk_out = sk_out;
2358
2359                         if (td->o.use_thread) {
2360                                 int ret;
2361
2362                                 dprint(FD_PROCESS, "will pthread_create\n");
2363                                 ret = pthread_create(&td->thread, NULL,
2364                                                         thread_main, fd);
2365                                 if (ret) {
2366                                         log_err("pthread_create: %s\n",
2367                                                         strerror(ret));
2368                                         free(fd);
2369                                         nr_started--;
2370                                         break;
2371                                 }
2372                                 fd = NULL;
2373                                 ret = pthread_detach(td->thread);
2374                                 if (ret)
2375                                         log_err("pthread_detach: %s",
2376                                                         strerror(ret));
2377                         } else {
2378                                 pid_t pid;
2379                                 dprint(FD_PROCESS, "will fork\n");
2380                                 pid = fork();
2381                                 if (!pid) {
2382                                         int ret;
2383
2384                                         ret = (int)(uintptr_t)thread_main(fd);
2385                                         _exit(ret);
2386                                 } else if (i == fio_debug_jobno)
2387                                         *fio_debug_jobp = pid;
2388                         }
2389                         dprint(FD_MUTEX, "wait on startup_sem\n");
2390                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2391                                 log_err("fio: job startup hung? exiting.\n");
2392                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2393                                 fio_abort = true;
2394                                 nr_started--;
2395                                 free(fd);
2396                                 break;
2397                         }
2398                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2399                 }
2400
2401                 /*
2402                  * Wait for the started threads to transition to
2403                  * TD_INITIALIZED.
2404                  */
2405                 fio_gettime(&this_start, NULL);
2406                 left = this_jobs;
2407                 while (left && !fio_abort) {
2408                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2409                                 break;
2410
2411                         do_usleep(100000);
2412
2413                         for (i = 0; i < this_jobs; i++) {
2414                                 td = map[i];
2415                                 if (!td)
2416                                         continue;
2417                                 if (td->runstate == TD_INITIALIZED) {
2418                                         map[i] = NULL;
2419                                         left--;
2420                                 } else if (td->runstate >= TD_EXITED) {
2421                                         map[i] = NULL;
2422                                         left--;
2423                                         todo--;
2424                                         nr_running++; /* work-around... */
2425                                 }
2426                         }
2427                 }
2428
2429                 if (left) {
2430                         log_err("fio: %d job%s failed to start\n", left,
2431                                         left > 1 ? "s" : "");
2432                         for (i = 0; i < this_jobs; i++) {
2433                                 td = map[i];
2434                                 if (!td)
2435                                         continue;
2436                                 kill(td->pid, SIGTERM);
2437                         }
2438                         break;
2439                 }
2440
2441                 /*
2442                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2443                  */
2444                 for_each_td(td, i) {
2445                         if (td->runstate != TD_INITIALIZED)
2446                                 continue;
2447
2448                         if (in_ramp_time(td))
2449                                 td_set_runstate(td, TD_RAMP);
2450                         else
2451                                 td_set_runstate(td, TD_RUNNING);
2452                         nr_running++;
2453                         nr_started--;
2454                         m_rate += ddir_rw_sum(td->o.ratemin);
2455                         t_rate += ddir_rw_sum(td->o.rate);
2456                         todo--;
2457                         fio_sem_up(td->sem);
2458                 }
2459
2460                 reap_threads(&nr_running, &t_rate, &m_rate);
2461
2462                 if (todo)
2463                         do_usleep(100000);
2464         }
2465
2466         while (nr_running) {
2467                 reap_threads(&nr_running, &t_rate, &m_rate);
2468                 do_usleep(10000);
2469         }
2470
2471         fio_idle_prof_stop();
2472
2473         update_io_ticks();
2474 }
2475
2476 static void free_disk_util(void)
2477 {
2478         disk_util_prune_entries();
2479         helper_thread_destroy();
2480 }
2481
2482 int fio_backend(struct sk_out *sk_out)
2483 {
2484         struct thread_data *td;
2485         int i;
2486
2487         if (exec_profile) {
2488                 if (load_profile(exec_profile))
2489                         return 1;
2490                 free(exec_profile);
2491                 exec_profile = NULL;
2492         }
2493         if (!thread_number)
2494                 return 0;
2495
2496         if (write_bw_log) {
2497                 struct log_params p = {
2498                         .log_type = IO_LOG_TYPE_BW,
2499                 };
2500
2501                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2502                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2503                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2504         }
2505
2506         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2507         if (!sk_out)
2508                 is_local_backend = true;
2509         if (startup_sem == NULL)
2510                 return 1;
2511
2512         set_genesis_time();
2513         stat_init();
2514         if (helper_thread_create(startup_sem, sk_out))
2515                 log_err("fio: failed to create helper thread\n");
2516
2517         cgroup_list = smalloc(sizeof(*cgroup_list));
2518         if (cgroup_list)
2519                 INIT_FLIST_HEAD(cgroup_list);
2520
2521         run_threads(sk_out);
2522
2523         helper_thread_exit();
2524
2525         if (!fio_abort) {
2526                 __show_run_stats();
2527                 if (write_bw_log) {
2528                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2529                                 struct io_log *log = agg_io_log[i];
2530
2531                                 flush_log(log, false);
2532                                 free_log(log);
2533                         }
2534                 }
2535         }
2536
2537         for_each_td(td, i) {
2538                 steadystate_free(td);
2539                 fio_options_free(td);
2540                 fio_dump_options_free(td);
2541                 if (td->rusage_sem) {
2542                         fio_sem_remove(td->rusage_sem);
2543                         td->rusage_sem = NULL;
2544                 }
2545                 fio_sem_remove(td->sem);
2546                 td->sem = NULL;
2547         }
2548
2549         free_disk_util();
2550         if (cgroup_list) {
2551                 cgroup_kill(cgroup_list);
2552                 sfree(cgroup_list);
2553         }
2554
2555         fio_sem_remove(startup_sem);
2556         stat_exit();
2557         return exit_value;
2558 }