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 #ifdef WIN32
  94 static void sig_break(int sig)
  95 {
  96         struct thread_data *td;
  97         int i;
  98
  99         sig_int(sig);
 100
 101         /**
 102          * Windows terminates all job processes on SIGBREAK after the handler
 103          * returns, so give them time to wrap-up and give stats
 104          */
 105         for_each_td(td, i) {
 106                 while (td->runstate < TD_EXITED)
 107                         sleep(1);
 108         }
 109 }
 110 #endif
 111
 112 void sig_show_status(int sig)
 113 {
 114         show_running_run_stats();
 115 }
 116
 117 static void set_sig_handlers(void)
 118 {
 119         struct sigaction act;
 120
 121         memset(&act, 0, sizeof(act));
 122         act.sa_handler = sig_int;
 123         act.sa_flags = SA_RESTART;
 124         sigaction(SIGINT, &act, NULL);
 125
 126         memset(&act, 0, sizeof(act));
 127         act.sa_handler = sig_int;
 128         act.sa_flags = SA_RESTART;
 129         sigaction(SIGTERM, &act, NULL);
 130
 131 /* Windows uses SIGBREAK as a quit signal from other applications */
 132 #ifdef WIN32
 133         memset(&act, 0, sizeof(act));
 134         act.sa_handler = sig_break;
 135         act.sa_flags = SA_RESTART;
 136         sigaction(SIGBREAK, &act, NULL);
 137 #endif
 138
 139         memset(&act, 0, sizeof(act));
 140         act.sa_handler = sig_show_status;
 141         act.sa_flags = SA_RESTART;
 142         sigaction(SIGUSR1, &act, NULL);
 143
 144         if (is_backend) {
 145                 memset(&act, 0, sizeof(act));
 146                 act.sa_handler = sig_int;
 147                 act.sa_flags = SA_RESTART;
 148                 sigaction(SIGPIPE, &act, NULL);
 149         }
 150 }
 151
 152 /*
 153  * Check if we are above the minimum rate given.
 154  */
 155 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 156                              enum fio_ddir ddir)
 157 {
 158         unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
 159         unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
 160         unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
 161         unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
 162
 163         assert(ddir_rw(ddir));
 164
 165         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 166                 return false;
 167
 168         /*
 169          * allow a 2 second settle period in the beginning
 170          */
 171         if (mtime_since(&td->start, now) < 2000)
 172                 return false;
 173
 174         /*
 175          * if last_rate_check_blocks or last_rate_check_bytes is set,
 176          * we can compute a rate per ratecycle
 177          */
 178         if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
 179                 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
 180                 if (spent < td->o.ratecycle || spent==0)
 181                         return false;
 182
 183                 if (td->o.ratemin[ddir]) {
 184                         /*
 185                          * check bandwidth specified rate
 186                          */
 187                         unsigned long long current_rate_bytes =
 188                                 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
 189                         if (current_rate_bytes < option_rate_bytes_min) {
 190                                 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 191                                         td->o.name, option_rate_bytes_min, current_rate_bytes);
 192                                 return true;
 193                         }
 194                 } else {
 195                         /*
 196                          * checks iops specified rate
 197                          */
 198                         unsigned long long current_rate_iops =
 199                                 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
 200
 201                         if (current_rate_iops < option_rate_iops_min) {
 202                                 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 203                                         td->o.name, option_rate_iops_min, current_rate_iops);
 204                                 return true;
 205                         }
 206                 }
 207         }
 208
 209         td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
 210         td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
 211         memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
 212         return false;
 213 }
 214
 215 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 216 {
 217         bool ret = false;
 218
 219         for_each_rw_ddir(ddir) {
 220                 if (td->bytes_done[ddir])
 221                         ret |= __check_min_rate(td, now, ddir);
 222         }
 223
 224         return ret;
 225 }
 226
 227 /*
 228  * When job exits, we can cancel the in-flight IO if we are using async
 229  * io. Attempt to do so.
 230  */
 231 static void cleanup_pending_aio(struct thread_data *td)
 232 {
 233         int r;
 234
 235         /*
 236          * get immediately available events, if any
 237          */
 238         r = io_u_queued_complete(td, 0);
 239
 240         /*
 241          * now cancel remaining active events
 242          */
 243         if (td->io_ops->cancel) {
 244                 struct io_u *io_u;
 245                 int i;
 246
 247                 io_u_qiter(&td->io_u_all, io_u, i) {
 248                         if (io_u->flags & IO_U_F_FLIGHT) {
 249                                 r = td->io_ops->cancel(td, io_u);
 250                                 if (!r)
 251                                         put_io_u(td, io_u);
 252                         }
 253                 }
 254         }
 255
 256         if (td->cur_depth)
 257                 r = io_u_queued_complete(td, td->cur_depth);
 258 }
 259
 260 /*
 261  * Helper to handle the final sync of a file. Works just like the normal
 262  * io path, just does everything sync.
 263  */
 264 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 265 {
 266         struct io_u *io_u = __get_io_u(td);
 267         enum fio_q_status ret;
 268
 269         if (!io_u)
 270                 return true;
 271
 272         io_u->ddir = DDIR_SYNC;
 273         io_u->file = f;
 274         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 275
 276         if (td_io_prep(td, io_u)) {
 277                 put_io_u(td, io_u);
 278                 return true;
 279         }
 280
 281 requeue:
 282         ret = td_io_queue(td, io_u);
 283         switch (ret) {
 284         case FIO_Q_QUEUED:
 285                 td_io_commit(td);
 286                 if (io_u_queued_complete(td, 1) < 0)
 287                         return true;
 288                 break;
 289         case FIO_Q_COMPLETED:
 290                 if (io_u->error) {
 291                         td_verror(td, io_u->error, "td_io_queue");
 292                         return true;
 293                 }
 294
 295                 if (io_u_sync_complete(td, io_u) < 0)
 296                         return true;
 297                 break;
 298         case FIO_Q_BUSY:
 299                 td_io_commit(td);
 300                 goto requeue;
 301         }
 302
 303         return false;
 304 }
 305
 306 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 307 {
 308         int ret, ret2;
 309
 310         if (fio_file_open(f))
 311                 return fio_io_sync(td, f);
 312
 313         if (td_io_open_file(td, f))
 314                 return 1;
 315
 316         ret = fio_io_sync(td, f);
 317         ret2 = 0;
 318         if (fio_file_open(f))
 319                 ret2 = td_io_close_file(td, f);
 320         return (ret || ret2);
 321 }
 322
 323 static inline void __update_ts_cache(struct thread_data *td)
 324 {
 325         fio_gettime(&td->ts_cache, NULL);
 326 }
 327
 328 static inline void update_ts_cache(struct thread_data *td)
 329 {
 330         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 331                 __update_ts_cache(td);
 332 }
 333
 334 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 335 {
 336         if (in_ramp_time(td))
 337                 return false;
 338         if (!td->o.timeout)
 339                 return false;
 340         if (utime_since(&td->epoch, t) >= td->o.timeout)
 341                 return true;
 342
 343         return false;
 344 }
 345
 346 /*
 347  * We need to update the runtime consistently in ms, but keep a running
 348  * tally of the current elapsed time in microseconds for sub millisecond
 349  * updates.
 350  */
 351 static inline void update_runtime(struct thread_data *td,
 352                                   unsigned long long *elapsed_us,
 353                                   const enum fio_ddir ddir)
 354 {
 355         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 356                 return;
 357
 358         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 359         elapsed_us[ddir] += utime_since_now(&td->start);
 360         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 361 }
 362
 363 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 364                                 int *retptr)
 365 {
 366         int ret = *retptr;
 367
 368         if (ret < 0 || td->error) {
 369                 int err = td->error;
 370                 enum error_type_bit eb;
 371
 372                 if (ret < 0)
 373                         err = -ret;
 374
 375                 eb = td_error_type(ddir, err);
 376                 if (!(td->o.continue_on_error & (1 << eb)))
 377                         return true;
 378
 379                 if (td_non_fatal_error(td, eb, err)) {
 380                         /*
 381                          * Continue with the I/Os in case of
 382                          * a non fatal error.
 383                          */
 384                         update_error_count(td, err);
 385                         td_clear_error(td);
 386                         *retptr = 0;
 387                         return false;
 388                 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
 389                         /*
 390                          * We expect to hit this error if
 391                          * fill_device option is set.
 392                          */
 393                         td_clear_error(td);
 394                         fio_mark_td_terminate(td);
 395                         return true;
 396                 } else {
 397                         /*
 398                          * Stop the I/O in case of a fatal
 399                          * error.
 400                          */
 401                         update_error_count(td, err);
 402                         return true;
 403                 }
 404         }
 405
 406         return false;
 407 }
 408
 409 static void check_update_rusage(struct thread_data *td)
 410 {
 411         if (td->update_rusage) {
 412                 td->update_rusage = 0;
 413                 update_rusage_stat(td);
 414                 fio_sem_up(td->rusage_sem);
 415         }
 416 }
 417
 418 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 419 {
 420         const int full = queue_full(td);
 421         int min_evts = 0;
 422         int ret;
 423
 424         if (td->flags & TD_F_REGROW_LOGS)
 425                 return io_u_quiesce(td);
 426
 427         /*
 428          * if the queue is full, we MUST reap at least 1 event
 429          */
 430         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 431         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 432                 min_evts = 1;
 433
 434         if (time && should_check_rate(td))
 435                 fio_gettime(time, NULL);
 436
 437         do {
 438                 ret = io_u_queued_complete(td, min_evts);
 439                 if (ret < 0)
 440                         break;
 441         } while (full && (td->cur_depth > td->o.iodepth_low));
 442
 443         return ret;
 444 }
 445
 446 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 447                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 448                    struct timespec *comp_time)
 449 {
 450         switch (*ret) {
 451         case FIO_Q_COMPLETED:
 452                 if (io_u->error) {
 453                         *ret = -io_u->error;
 454                         clear_io_u(td, io_u);
 455                 } else if (io_u->resid) {
 456                         long long bytes = io_u->xfer_buflen - io_u->resid;
 457                         struct fio_file *f = io_u->file;
 458
 459                         if (bytes_issued)
 460                                 *bytes_issued += bytes;
 461
 462                         if (!from_verify)
 463                                 trim_io_piece(io_u);
 464
 465                         /*
 466                          * zero read, fail
 467                          */
 468                         if (!bytes) {
 469                                 if (!from_verify)
 470                                         unlog_io_piece(td, io_u);
 471                                 td_verror(td, EIO, "full resid");
 472                                 put_io_u(td, io_u);
 473                                 break;
 474                         }
 475
 476                         io_u->xfer_buflen = io_u->resid;
 477                         io_u->xfer_buf += bytes;
 478                         io_u->offset += bytes;
 479
 480                         if (ddir_rw(io_u->ddir))
 481                                 td->ts.short_io_u[io_u->ddir]++;
 482
 483                         if (io_u->offset == f->real_file_size)
 484                                 goto sync_done;
 485
 486                         requeue_io_u(td, &io_u);
 487                 } else {
 488 sync_done:
 489                         if (comp_time && should_check_rate(td))
 490                                 fio_gettime(comp_time, NULL);
 491
 492                         *ret = io_u_sync_complete(td, io_u);
 493                         if (*ret < 0)
 494                                 break;
 495                 }
 496
 497                 if (td->flags & TD_F_REGROW_LOGS)
 498                         regrow_logs(td);
 499
 500                 /*
 501                  * when doing I/O (not when verifying),
 502                  * check for any errors that are to be ignored
 503                  */
 504                 if (!from_verify)
 505                         break;
 506
 507                 return 0;
 508         case FIO_Q_QUEUED:
 509                 /*
 510                  * if the engine doesn't have a commit hook,
 511                  * the io_u is really queued. if it does have such
 512                  * a hook, it has to call io_u_queued() itself.
 513                  */
 514                 if (td->io_ops->commit == NULL)
 515                         io_u_queued(td, io_u);
 516                 if (bytes_issued)
 517                         *bytes_issued += io_u->xfer_buflen;
 518                 break;
 519         case FIO_Q_BUSY:
 520                 if (!from_verify)
 521                         unlog_io_piece(td, io_u);
 522                 requeue_io_u(td, &io_u);
 523                 td_io_commit(td);
 524                 break;
 525         default:
 526                 assert(*ret < 0);
 527                 td_verror(td, -(*ret), "td_io_queue");
 528                 break;
 529         }
 530
 531         if (break_on_this_error(td, ddir, ret))
 532                 return 1;
 533
 534         return 0;
 535 }
 536
 537 static inline bool io_in_polling(struct thread_data *td)
 538 {
 539         return !td->o.iodepth_batch_complete_min &&
 540                    !td->o.iodepth_batch_complete_max;
 541 }
 542 /*
 543  * Unlinks files from thread data fio_file structure
 544  */
 545 static int unlink_all_files(struct thread_data *td)
 546 {
 547         struct fio_file *f;
 548         unsigned int i;
 549         int ret = 0;
 550
 551         for_each_file(td, f, i) {
 552                 if (f->filetype != FIO_TYPE_FILE)
 553                         continue;
 554                 ret = td_io_unlink_file(td, f);
 555                 if (ret)
 556                         break;
 557         }
 558
 559         if (ret)
 560                 td_verror(td, ret, "unlink_all_files");
 561
 562         return ret;
 563 }
 564
 565 /*
 566  * Check if io_u will overlap an in-flight IO in the queue
 567  */
 568 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 569 {
 570         bool overlap;
 571         struct io_u *check_io_u;
 572         unsigned long long x1, x2, y1, y2;
 573         int i;
 574
 575         x1 = io_u->offset;
 576         x2 = io_u->offset + io_u->buflen;
 577         overlap = false;
 578         io_u_qiter(q, check_io_u, i) {
 579                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 580                         y1 = check_io_u->offset;
 581                         y2 = check_io_u->offset + check_io_u->buflen;
 582
 583                         if (x1 < y2 && y1 < x2) {
 584                                 overlap = true;
 585                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 586                                                 x1, io_u->buflen,
 587                                                 y1, check_io_u->buflen);
 588                                 break;
 589                         }
 590                 }
 591         }
 592
 593         return overlap;
 594 }
 595
 596 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 597 {
 598         /*
 599          * Check for overlap if the user asked us to, and we have
 600          * at least one IO in flight besides this one.
 601          */
 602         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 603             in_flight_overlap(&td->io_u_all, io_u))
 604                 return FIO_Q_BUSY;
 605
 606         return td_io_queue(td, io_u);
 607 }
 608
 609 /*
 610  * The main verify engine. Runs over the writes we previously submitted,
 611  * reads the blocks back in, and checks the crc/md5 of the data.
 612  */
 613 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 614 {
 615         struct fio_file *f;
 616         struct io_u *io_u;
 617         int ret, min_events;
 618         unsigned int i;
 619
 620         dprint(FD_VERIFY, "starting loop\n");
 621
 622         /*
 623          * sync io first and invalidate cache, to make sure we really
 624          * read from disk.
 625          */
 626         for_each_file(td, f, i) {
 627                 if (!fio_file_open(f))
 628                         continue;
 629                 if (fio_io_sync(td, f))
 630                         break;
 631                 if (file_invalidate_cache(td, f))
 632                         break;
 633         }
 634
 635         check_update_rusage(td);
 636
 637         if (td->error)
 638                 return;
 639
 640         td_set_runstate(td, TD_VERIFYING);
 641
 642         io_u = NULL;
 643         while (!td->terminate) {
 644                 enum fio_ddir ddir;
 645                 int full;
 646
 647                 update_ts_cache(td);
 648                 check_update_rusage(td);
 649
 650                 if (runtime_exceeded(td, &td->ts_cache)) {
 651                         __update_ts_cache(td);
 652                         if (runtime_exceeded(td, &td->ts_cache)) {
 653                                 fio_mark_td_terminate(td);
 654                                 break;
 655                         }
 656                 }
 657
 658                 if (flow_threshold_exceeded(td))
 659                         continue;
 660
 661                 if (!td->o.experimental_verify) {
 662                         io_u = __get_io_u(td);
 663                         if (!io_u)
 664                                 break;
 665
 666                         if (get_next_verify(td, io_u)) {
 667                                 put_io_u(td, io_u);
 668                                 break;
 669                         }
 670
 671                         if (td_io_prep(td, io_u)) {
 672                                 put_io_u(td, io_u);
 673                                 break;
 674                         }
 675                 } else {
 676                         if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
 677                                 break;
 678
 679                         while ((io_u = get_io_u(td)) != NULL) {
 680                                 if (IS_ERR_OR_NULL(io_u)) {
 681                                         io_u = NULL;
 682                                         ret = FIO_Q_BUSY;
 683                                         goto reap;
 684                                 }
 685
 686                                 /*
 687                                  * We are only interested in the places where
 688                                  * we wrote or trimmed IOs. Turn those into
 689                                  * reads for verification purposes.
 690                                  */
 691                                 if (io_u->ddir == DDIR_READ) {
 692                                         /*
 693                                          * Pretend we issued it for rwmix
 694                                          * accounting
 695                                          */
 696                                         td->io_issues[DDIR_READ]++;
 697                                         put_io_u(td, io_u);
 698                                         continue;
 699                                 } else if (io_u->ddir == DDIR_TRIM) {
 700                                         io_u->ddir = DDIR_READ;
 701                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 702                                         break;
 703                                 } else if (io_u->ddir == DDIR_WRITE) {
 704                                         io_u->ddir = DDIR_READ;
 705                                         io_u->numberio = td->verify_read_issues;
 706                                         td->verify_read_issues++;
 707                                         populate_verify_io_u(td, io_u);
 708                                         break;
 709                                 } else {
 710                                         put_io_u(td, io_u);
 711                                         continue;
 712                                 }
 713                         }
 714
 715                         if (!io_u)
 716                                 break;
 717                 }
 718
 719                 if (verify_state_should_stop(td, io_u)) {
 720                         put_io_u(td, io_u);
 721                         break;
 722                 }
 723
 724                 if (td->o.verify_async)
 725                         io_u->end_io = verify_io_u_async;
 726                 else
 727                         io_u->end_io = verify_io_u;
 728
 729                 ddir = io_u->ddir;
 730                 if (!td->o.disable_slat)
 731                         fio_gettime(&io_u->start_time, NULL);
 732
 733                 ret = io_u_submit(td, io_u);
 734
 735                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 736                         break;
 737
 738                 /*
 739                  * if we can queue more, do so. but check if there are
 740                  * completed io_u's first. Note that we can get BUSY even
 741                  * without IO queued, if the system is resource starved.
 742                  */
 743 reap:
 744                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 745                 if (full || io_in_polling(td))
 746                         ret = wait_for_completions(td, NULL);
 747
 748                 if (ret < 0)
 749                         break;
 750         }
 751
 752         check_update_rusage(td);
 753
 754         if (!td->error) {
 755                 min_events = td->cur_depth;
 756
 757                 if (min_events)
 758                         ret = io_u_queued_complete(td, min_events);
 759         } else
 760                 cleanup_pending_aio(td);
 761
 762         td_set_runstate(td, TD_RUNNING);
 763
 764         dprint(FD_VERIFY, "exiting loop\n");
 765 }
 766
 767 static bool exceeds_number_ios(struct thread_data *td)
 768 {
 769         unsigned long long number_ios;
 770
 771         if (!td->o.number_ios)
 772                 return false;
 773
 774         number_ios = ddir_rw_sum(td->io_blocks);
 775         number_ios += td->io_u_queued + td->io_u_in_flight;
 776
 777         return number_ios >= (td->o.number_ios * td->loops);
 778 }
 779
 780 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 781 {
 782         unsigned long long bytes, limit;
 783
 784         if (td_rw(td))
 785                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 786         else if (td_write(td))
 787                 bytes = this_bytes[DDIR_WRITE];
 788         else if (td_read(td))
 789                 bytes = this_bytes[DDIR_READ];
 790         else
 791                 bytes = this_bytes[DDIR_TRIM];
 792
 793         if (td->o.io_size)
 794                 limit = td->o.io_size;
 795         else
 796                 limit = td->o.size;
 797
 798         limit *= td->loops;
 799         return bytes >= limit || exceeds_number_ios(td);
 800 }
 801
 802 static bool io_issue_bytes_exceeded(struct thread_data *td)
 803 {
 804         return io_bytes_exceeded(td, td->io_issue_bytes);
 805 }
 806
 807 static bool io_complete_bytes_exceeded(struct thread_data *td)
 808 {
 809         return io_bytes_exceeded(td, td->this_io_bytes);
 810 }
 811
 812 /*
 813  * used to calculate the next io time for rate control
 814  *
 815  */
 816 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 817 {
 818         uint64_t bps = td->rate_bps[ddir];
 819
 820         assert(!(td->flags & TD_F_CHILD));
 821
 822         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 823                 uint64_t val, iops;
 824
 825                 iops = bps / td->o.min_bs[ddir];
 826                 val = (int64_t) (1000000 / iops) *
 827                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 828                 if (val) {
 829                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 830                                         (unsigned long long) 1000000 / val,
 831                                         ddir);
 832                 }
 833                 td->last_usec[ddir] += val;
 834                 return td->last_usec[ddir];
 835         } else if (bps) {
 836                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 837                 uint64_t secs = bytes / bps;
 838                 uint64_t remainder = bytes % bps;
 839
 840                 return remainder * 1000000 / bps + secs * 1000000;
 841         }
 842
 843         return 0;
 844 }
 845
 846 static void init_thinktime(struct thread_data *td)
 847 {
 848         if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
 849                 td->thinktime_blocks_counter = td->io_blocks;
 850         else
 851                 td->thinktime_blocks_counter = td->io_issues;
 852         td->last_thinktime = td->epoch;
 853         td->last_thinktime_blocks = 0;
 854 }
 855
 856 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
 857                              struct timespec *time)
 858 {
 859         unsigned long long b;
 860         unsigned long long runtime_left;
 861         uint64_t total;
 862         int left;
 863         struct timespec now;
 864         bool stall = false;
 865
 866         if (td->o.thinktime_iotime) {
 867                 fio_gettime(&now, NULL);
 868                 if (utime_since(&td->last_thinktime, &now)
 869                     >= td->o.thinktime_iotime) {
 870                         stall = true;
 871                 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
 872                         /*
 873                          * When thinktime_iotime is set and thinktime_blocks is
 874                          * not set, skip the thinktime_blocks check, since
 875                          * thinktime_blocks default value 1 does not work
 876                          * together with thinktime_iotime.
 877                          */
 878                         return;
 879                 }
 880
 881         }
 882
 883         b = ddir_rw_sum(td->thinktime_blocks_counter);
 884         if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
 885                 stall = true;
 886
 887         if (!stall)
 888                 return;
 889
 890         io_u_quiesce(td);
 891
 892         left = td->o.thinktime_spin;
 893         if (td->o.timeout) {
 894                 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
 895                 if (runtime_left < (unsigned long long)left)
 896                         left = runtime_left;
 897         }
 898
 899         total = 0;
 900         if (left)
 901                 total = usec_spin(left);
 902
 903         left = td->o.thinktime - total;
 904         if (td->o.timeout) {
 905                 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
 906                 if (runtime_left < (unsigned long long)left)
 907                         left = runtime_left;
 908         }
 909
 910         if (left)
 911                 total += usec_sleep(td, left);
 912
 913         /*
 914          * If we're ignoring thinktime for the rate, add the number of bytes
 915          * we would have done while sleeping, minus one block to ensure we
 916          * start issuing immediately after the sleep.
 917          */
 918         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 919                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 920                 uint64_t bs = td->o.min_bs[ddir];
 921                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 922                 uint64_t over;
 923
 924                 if (usperop <= total)
 925                         over = bs;
 926                 else
 927                         over = (usperop - total) / usperop * -bs;
 928
 929                 td->rate_io_issue_bytes[ddir] += (missed - over);
 930                 /* adjust for rate_process=poisson */
 931                 td->last_usec[ddir] += total;
 932         }
 933
 934         if (time && should_check_rate(td))
 935                 fio_gettime(time, NULL);
 936
 937         td->last_thinktime_blocks = b;
 938         if (td->o.thinktime_iotime) {
 939                 fio_gettime(&now, NULL);
 940                 td->last_thinktime = now;
 941         }
 942 }
 943
 944 /*
 945  * Main IO worker function. It retrieves io_u's to process and queues
 946  * and reaps them, checking for rate and errors along the way.
 947  *
 948  * Returns number of bytes written and trimmed.
 949  */
 950 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 951 {
 952         unsigned int i;
 953         int ret = 0;
 954         uint64_t total_bytes, bytes_issued = 0;
 955
 956         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 957                 bytes_done[i] = td->bytes_done[i];
 958
 959         if (in_ramp_time(td))
 960                 td_set_runstate(td, TD_RAMP);
 961         else
 962                 td_set_runstate(td, TD_RUNNING);
 963
 964         lat_target_init(td);
 965
 966         total_bytes = td->o.size;
 967         /*
 968         * Allow random overwrite workloads to write up to io_size
 969         * before starting verification phase as 'size' doesn't apply.
 970         */
 971         if (td_write(td) && td_random(td) && td->o.norandommap)
 972                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 973         /*
 974          * If verify_backlog is enabled, we'll run the verify in this
 975          * handler as well. For that case, we may need up to twice the
 976          * amount of bytes.
 977          */
 978         if (td->o.verify != VERIFY_NONE &&
 979            (td_write(td) && td->o.verify_backlog))
 980                 total_bytes += td->o.size;
 981
 982         /* In trimwrite mode, each byte is trimmed and then written, so
 983          * allow total_bytes or number of ios to be twice as big */
 984         if (td_trimwrite(td)) {
 985                 total_bytes += td->total_io_size;
 986                 td->o.number_ios *= 2;
 987         }
 988
 989         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 990                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 991                 td->o.time_based) {
 992                 struct timespec comp_time;
 993                 struct io_u *io_u;
 994                 int full;
 995                 enum fio_ddir ddir;
 996
 997                 check_update_rusage(td);
 998
 999                 if (td->terminate || td->done)
1000                         break;
1001
1002                 update_ts_cache(td);
1003
1004                 if (runtime_exceeded(td, &td->ts_cache)) {
1005                         __update_ts_cache(td);
1006                         if (runtime_exceeded(td, &td->ts_cache)) {
1007                                 fio_mark_td_terminate(td);
1008                                 break;
1009                         }
1010                 }
1011
1012                 if (flow_threshold_exceeded(td))
1013                         continue;
1014
1015                 /*
1016                  * Break if we exceeded the bytes. The exception is time
1017                  * based runs, but we still need to break out of the loop
1018                  * for those to run verification, if enabled.
1019                  * Jobs read from iolog do not use this stop condition.
1020                  */
1021                 if (bytes_issued >= total_bytes &&
1022                     !td->o.read_iolog_file &&
1023                     (!td->o.time_based ||
1024                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
1025                         break;
1026
1027                 io_u = get_io_u(td);
1028                 if (IS_ERR_OR_NULL(io_u)) {
1029                         int err = PTR_ERR(io_u);
1030
1031                         io_u = NULL;
1032                         ddir = DDIR_INVAL;
1033                         if (err == -EBUSY) {
1034                                 ret = FIO_Q_BUSY;
1035                                 goto reap;
1036                         }
1037                         if (td->o.latency_target)
1038                                 goto reap;
1039                         break;
1040                 }
1041
1042                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
1043                         if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
1044                                 io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
1045                                 io_u->numberio = td->io_issues[io_u->ddir];
1046                                 populate_verify_io_u(td, io_u);
1047                         }
1048                 }
1049
1050                 ddir = io_u->ddir;
1051
1052                 /*
1053                  * Add verification end_io handler if:
1054                  *      - Asked to verify (!td_rw(td))
1055                  *      - Or the io_u is from our verify list (mixed write/ver)
1056                  */
1057                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1058                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1059
1060                         if (verify_state_should_stop(td, io_u)) {
1061                                 put_io_u(td, io_u);
1062                                 break;
1063                         }
1064
1065                         if (td->o.verify_async)
1066                                 io_u->end_io = verify_io_u_async;
1067                         else
1068                                 io_u->end_io = verify_io_u;
1069                         td_set_runstate(td, TD_VERIFYING);
1070                 } else if (in_ramp_time(td))
1071                         td_set_runstate(td, TD_RAMP);
1072                 else
1073                         td_set_runstate(td, TD_RUNNING);
1074
1075                 /*
1076                  * Always log IO before it's issued, so we know the specific
1077                  * order of it. The logged unit will track when the IO has
1078                  * completed.
1079                  */
1080                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1081                     td->o.do_verify &&
1082                     td->o.verify != VERIFY_NONE &&
1083                     !td->o.experimental_verify)
1084                         log_io_piece(td, io_u);
1085
1086                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1087                         const unsigned long long blen = io_u->xfer_buflen;
1088                         const enum fio_ddir __ddir = acct_ddir(io_u);
1089
1090                         if (td->error)
1091                                 break;
1092
1093                         workqueue_enqueue(&td->io_wq, &io_u->work);
1094                         ret = FIO_Q_QUEUED;
1095
1096                         if (ddir_rw(__ddir)) {
1097                                 td->io_issues[__ddir]++;
1098                                 td->io_issue_bytes[__ddir] += blen;
1099                                 td->rate_io_issue_bytes[__ddir] += blen;
1100                         }
1101
1102                         if (should_check_rate(td)) {
1103                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1104                                 fio_gettime(&comp_time, NULL);
1105                         }
1106
1107                 } else {
1108                         ret = io_u_submit(td, io_u);
1109
1110                         if (should_check_rate(td))
1111                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1112
1113                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1114                                 break;
1115
1116                         /*
1117                          * See if we need to complete some commands. Note that
1118                          * we can get BUSY even without IO queued, if the
1119                          * system is resource starved.
1120                          */
1121 reap:
1122                         full = queue_full(td) ||
1123                                 (ret == FIO_Q_BUSY && td->cur_depth);
1124                         if (full || io_in_polling(td))
1125                                 ret = wait_for_completions(td, &comp_time);
1126                 }
1127                 if (ret < 0)
1128                         break;
1129
1130                 if (ddir_rw(ddir) && td->o.thinktime)
1131                         handle_thinktime(td, ddir, &comp_time);
1132
1133                 if (!ddir_rw_sum(td->bytes_done) &&
1134                     !td_ioengine_flagged(td, FIO_NOIO))
1135                         continue;
1136
1137                 if (!in_ramp_time(td) && should_check_rate(td)) {
1138                         if (check_min_rate(td, &comp_time)) {
1139                                 if (exitall_on_terminate || td->o.exitall_error)
1140                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1141                                 td_verror(td, EIO, "check_min_rate");
1142                                 break;
1143                         }
1144                 }
1145                 if (!in_ramp_time(td) && td->o.latency_target)
1146                         lat_target_check(td);
1147         }
1148
1149         check_update_rusage(td);
1150
1151         if (td->trim_entries)
1152                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1153
1154         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1155                 td->error = 0;
1156                 fio_mark_td_terminate(td);
1157         }
1158         if (!td->error) {
1159                 struct fio_file *f;
1160
1161                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1162                         workqueue_flush(&td->io_wq);
1163                         i = 0;
1164                 } else
1165                         i = td->cur_depth;
1166
1167                 if (i) {
1168                         ret = io_u_queued_complete(td, i);
1169                         if (td->o.fill_device &&
1170                             (td->error == ENOSPC || td->error == EDQUOT))
1171                                 td->error = 0;
1172                 }
1173
1174                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1175                         td_set_runstate(td, TD_FSYNCING);
1176
1177                         for_each_file(td, f, i) {
1178                                 if (!fio_file_fsync(td, f))
1179                                         continue;
1180
1181                                 log_err("fio: end_fsync failed for file %s\n",
1182                                                                 f->file_name);
1183                         }
1184                 }
1185         } else {
1186                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1187                         workqueue_flush(&td->io_wq);
1188                 cleanup_pending_aio(td);
1189         }
1190
1191         /*
1192          * stop job if we failed doing any IO
1193          */
1194         if (!ddir_rw_sum(td->this_io_bytes))
1195                 td->done = 1;
1196
1197         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1198                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1199 }
1200
1201 static void free_file_completion_logging(struct thread_data *td)
1202 {
1203         struct fio_file *f;
1204         unsigned int i;
1205
1206         for_each_file(td, f, i) {
1207                 if (!f->last_write_comp)
1208                         break;
1209                 sfree(f->last_write_comp);
1210         }
1211 }
1212
1213 static int init_file_completion_logging(struct thread_data *td,
1214                                         unsigned int depth)
1215 {
1216         struct fio_file *f;
1217         unsigned int i;
1218
1219         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1220                 return 0;
1221
1222         for_each_file(td, f, i) {
1223                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1224                 if (!f->last_write_comp)
1225                         goto cleanup;
1226         }
1227
1228         return 0;
1229
1230 cleanup:
1231         free_file_completion_logging(td);
1232         log_err("fio: failed to alloc write comp data\n");
1233         return 1;
1234 }
1235
1236 static void cleanup_io_u(struct thread_data *td)
1237 {
1238         struct io_u *io_u;
1239
1240         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1241
1242                 if (td->io_ops->io_u_free)
1243                         td->io_ops->io_u_free(td, io_u);
1244
1245                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1246         }
1247
1248         free_io_mem(td);
1249
1250         io_u_rexit(&td->io_u_requeues);
1251         io_u_qexit(&td->io_u_freelist, false);
1252         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1253
1254         free_file_completion_logging(td);
1255 }
1256
1257 static int init_io_u(struct thread_data *td)
1258 {
1259         struct io_u *io_u;
1260         int cl_align, i, max_units;
1261         int err;
1262
1263         max_units = td->o.iodepth;
1264
1265         err = 0;
1266         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1267         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1268         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1269
1270         if (err) {
1271                 log_err("fio: failed setting up IO queues\n");
1272                 return 1;
1273         }
1274
1275         cl_align = os_cache_line_size();
1276
1277         for (i = 0; i < max_units; i++) {
1278                 void *ptr;
1279
1280                 if (td->terminate)
1281                         return 1;
1282
1283                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1284                 if (!ptr) {
1285                         log_err("fio: unable to allocate aligned memory\n");
1286                         return 1;
1287                 }
1288
1289                 io_u = ptr;
1290                 memset(io_u, 0, sizeof(*io_u));
1291                 INIT_FLIST_HEAD(&io_u->verify_list);
1292                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1293
1294                 io_u->index = i;
1295                 io_u->flags = IO_U_F_FREE;
1296                 io_u_qpush(&td->io_u_freelist, io_u);
1297
1298                 /*
1299                  * io_u never leaves this stack, used for iteration of all
1300                  * io_u buffers.
1301                  */
1302                 io_u_qpush(&td->io_u_all, io_u);
1303
1304                 if (td->io_ops->io_u_init) {
1305                         int ret = td->io_ops->io_u_init(td, io_u);
1306
1307                         if (ret) {
1308                                 log_err("fio: failed to init engine data: %d\n", ret);
1309                                 return 1;
1310                         }
1311                 }
1312         }
1313
1314         if (init_io_u_buffers(td))
1315                 return 1;
1316
1317         if (init_file_completion_logging(td, max_units))
1318                 return 1;
1319
1320         return 0;
1321 }
1322
1323 int init_io_u_buffers(struct thread_data *td)
1324 {
1325         struct io_u *io_u;
1326         unsigned long long max_bs, min_write;
1327         int i, max_units;
1328         int data_xfer = 1;
1329         char *p;
1330
1331         max_units = td->o.iodepth;
1332         max_bs = td_max_bs(td);
1333         min_write = td->o.min_bs[DDIR_WRITE];
1334         td->orig_buffer_size = (unsigned long long) max_bs
1335                                         * (unsigned long long) max_units;
1336
1337         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1338                 data_xfer = 0;
1339
1340         /*
1341          * if we may later need to do address alignment, then add any
1342          * possible adjustment here so that we don't cause a buffer
1343          * overflow later. this adjustment may be too much if we get
1344          * lucky and the allocator gives us an aligned address.
1345          */
1346         if (td->o.odirect || td->o.mem_align ||
1347             td_ioengine_flagged(td, FIO_RAWIO))
1348                 td->orig_buffer_size += page_mask + td->o.mem_align;
1349
1350         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1351                 unsigned long long bs;
1352
1353                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1354                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1355         }
1356
1357         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1358                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1359                 return 1;
1360         }
1361
1362         if (data_xfer && allocate_io_mem(td))
1363                 return 1;
1364
1365         if (td->o.odirect || td->o.mem_align ||
1366             td_ioengine_flagged(td, FIO_RAWIO))
1367                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1368         else
1369                 p = td->orig_buffer;
1370
1371         for (i = 0; i < max_units; i++) {
1372                 io_u = td->io_u_all.io_us[i];
1373                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1374
1375                 if (data_xfer) {
1376                         io_u->buf = p;
1377                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1378
1379                         if (td_write(td))
1380                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1381                         if (td_write(td) && td->o.verify_pattern_bytes) {
1382                                 /*
1383                                  * Fill the buffer with the pattern if we are
1384                                  * going to be doing writes.
1385                                  */
1386                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1387                         }
1388                 }
1389                 p += max_bs;
1390         }
1391
1392         return 0;
1393 }
1394
1395 #ifdef FIO_HAVE_IOSCHED_SWITCH
1396 /*
1397  * These functions are Linux specific.
1398  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1399  */
1400 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1401 {
1402         char tmp[256], tmp2[128], *p;
1403         FILE *f;
1404         int ret;
1405
1406         assert(file->du && file->du->sysfs_root);
1407         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1408
1409         f = fopen(tmp, "r+");
1410         if (!f) {
1411                 if (errno == ENOENT) {
1412                         log_err("fio: os or kernel doesn't support IO scheduler"
1413                                 " switching\n");
1414                         return 0;
1415                 }
1416                 td_verror(td, errno, "fopen iosched");
1417                 return 1;
1418         }
1419
1420         /*
1421          * Set io scheduler.
1422          */
1423         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1424         if (ferror(f) || ret != 1) {
1425                 td_verror(td, errno, "fwrite");
1426                 fclose(f);
1427                 return 1;
1428         }
1429
1430         rewind(f);
1431
1432         /*
1433          * Read back and check that the selected scheduler is now the default.
1434          */
1435         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1436         if (ferror(f) || ret < 0) {
1437                 td_verror(td, errno, "fread");
1438                 fclose(f);
1439                 return 1;
1440         }
1441         tmp[ret] = '\0';
1442         /*
1443          * either a list of io schedulers or "none\n" is expected. Strip the
1444          * trailing newline.
1445          */
1446         p = tmp;
1447         strsep(&p, "\n");
1448
1449         /*
1450          * Write to "none" entry doesn't fail, so check the result here.
1451          */
1452         if (!strcmp(tmp, "none")) {
1453                 log_err("fio: io scheduler is not tunable\n");
1454                 fclose(f);
1455                 return 0;
1456         }
1457
1458         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1459         if (!strstr(tmp, tmp2)) {
1460                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1461                 td_verror(td, EINVAL, "iosched_switch");
1462                 fclose(f);
1463                 return 1;
1464         }
1465
1466         fclose(f);
1467         return 0;
1468 }
1469
1470 static int switch_ioscheduler(struct thread_data *td)
1471 {
1472         struct fio_file *f;
1473         unsigned int i;
1474         int ret = 0;
1475
1476         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1477                 return 0;
1478
1479         assert(td->files && td->files[0]);
1480
1481         for_each_file(td, f, i) {
1482
1483                 /* Only consider regular files and block device files */
1484                 switch (f->filetype) {
1485                 case FIO_TYPE_FILE:
1486                 case FIO_TYPE_BLOCK:
1487                         /*
1488                          * Make sure that the device hosting the file could
1489                          * be determined.
1490                          */
1491                         if (!f->du)
1492                                 continue;
1493                         break;
1494                 case FIO_TYPE_CHAR:
1495                 case FIO_TYPE_PIPE:
1496                 default:
1497                         continue;
1498                 }
1499
1500                 ret = set_ioscheduler(td, f);
1501                 if (ret)
1502                         return ret;
1503         }
1504
1505         return 0;
1506 }
1507
1508 #else
1509
1510 static int switch_ioscheduler(struct thread_data *td)
1511 {
1512         return 0;
1513 }
1514
1515 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1516
1517 static bool keep_running(struct thread_data *td)
1518 {
1519         unsigned long long limit;
1520
1521         if (td->done)
1522                 return false;
1523         if (td->terminate)
1524                 return false;
1525         if (td->o.time_based)
1526                 return true;
1527         if (td->o.loops) {
1528                 td->o.loops--;
1529                 return true;
1530         }
1531         if (exceeds_number_ios(td))
1532                 return false;
1533
1534         if (td->o.io_size)
1535                 limit = td->o.io_size;
1536         else
1537                 limit = td->o.size;
1538
1539         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1540                 uint64_t diff;
1541
1542                 /*
1543                  * If the difference is less than the maximum IO size, we
1544                  * are done.
1545                  */
1546                 diff = limit - ddir_rw_sum(td->io_bytes);
1547                 if (diff < td_max_bs(td))
1548                         return false;
1549
1550                 if (fio_files_done(td) && !td->o.io_size)
1551                         return false;
1552
1553                 return true;
1554         }
1555
1556         return false;
1557 }
1558
1559 static int exec_string(struct thread_options *o, const char *string,
1560                        const char *mode)
1561 {
1562         int ret;
1563         char *str;
1564
1565         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1566                 return -1;
1567
1568         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1569                  o->name, mode);
1570         ret = system(str);
1571         if (ret == -1)
1572                 log_err("fio: exec of cmd <%s> failed\n", str);
1573
1574         free(str);
1575         return ret;
1576 }
1577
1578 /*
1579  * Dry run to compute correct state of numberio for verification.
1580  */
1581 static uint64_t do_dry_run(struct thread_data *td)
1582 {
1583         td_set_runstate(td, TD_RUNNING);
1584
1585         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1586                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1587                 struct io_u *io_u;
1588                 int ret;
1589
1590                 if (td->terminate || td->done)
1591                         break;
1592
1593                 io_u = get_io_u(td);
1594                 if (IS_ERR_OR_NULL(io_u))
1595                         break;
1596
1597                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1598                 io_u->error = 0;
1599                 io_u->resid = 0;
1600                 if (ddir_rw(acct_ddir(io_u)))
1601                         td->io_issues[acct_ddir(io_u)]++;
1602                 if (ddir_rw(io_u->ddir)) {
1603                         io_u_mark_depth(td, 1);
1604                         td->ts.total_io_u[io_u->ddir]++;
1605                 }
1606
1607                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1608                     td->o.do_verify &&
1609                     td->o.verify != VERIFY_NONE &&
1610                     !td->o.experimental_verify)
1611                         log_io_piece(td, io_u);
1612
1613                 ret = io_u_sync_complete(td, io_u);
1614                 (void) ret;
1615         }
1616
1617         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1618 }
1619
1620 struct fork_data {
1621         struct thread_data *td;
1622         struct sk_out *sk_out;
1623 };
1624
1625 /*
1626  * Entry point for the thread based jobs. The process based jobs end up
1627  * here as well, after a little setup.
1628  */
1629 static void *thread_main(void *data)
1630 {
1631         struct fork_data *fd = data;
1632         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1633         struct thread_data *td = fd->td;
1634         struct thread_options *o = &td->o;
1635         struct sk_out *sk_out = fd->sk_out;
1636         uint64_t bytes_done[DDIR_RWDIR_CNT];
1637         int deadlock_loop_cnt;
1638         bool clear_state;
1639         int res, ret;
1640
1641         sk_out_assign(sk_out);
1642         free(fd);
1643
1644         if (!o->use_thread) {
1645                 setsid();
1646                 td->pid = getpid();
1647         } else
1648                 td->pid = gettid();
1649
1650         fio_local_clock_init();
1651
1652         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1653
1654         if (is_backend)
1655                 fio_server_send_start(td);
1656
1657         INIT_FLIST_HEAD(&td->io_log_list);
1658         INIT_FLIST_HEAD(&td->io_hist_list);
1659         INIT_FLIST_HEAD(&td->verify_list);
1660         INIT_FLIST_HEAD(&td->trim_list);
1661         td->io_hist_tree = RB_ROOT;
1662
1663         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1664         if (ret) {
1665                 td_verror(td, ret, "mutex_cond_init_pshared");
1666                 goto err;
1667         }
1668         ret = cond_init_pshared(&td->verify_cond);
1669         if (ret) {
1670                 td_verror(td, ret, "mutex_cond_pshared");
1671                 goto err;
1672         }
1673
1674         td_set_runstate(td, TD_INITIALIZED);
1675         dprint(FD_MUTEX, "up startup_sem\n");
1676         fio_sem_up(startup_sem);
1677         dprint(FD_MUTEX, "wait on td->sem\n");
1678         fio_sem_down(td->sem);
1679         dprint(FD_MUTEX, "done waiting on td->sem\n");
1680
1681         /*
1682          * A new gid requires privilege, so we need to do this before setting
1683          * the uid.
1684          */
1685         if (o->gid != -1U && setgid(o->gid)) {
1686                 td_verror(td, errno, "setgid");
1687                 goto err;
1688         }
1689         if (o->uid != -1U && setuid(o->uid)) {
1690                 td_verror(td, errno, "setuid");
1691                 goto err;
1692         }
1693
1694         td_zone_gen_index(td);
1695
1696         /*
1697          * Do this early, we don't want the compress threads to be limited
1698          * to the same CPUs as the IO workers. So do this before we set
1699          * any potential CPU affinity
1700          */
1701         if (iolog_compress_init(td, sk_out))
1702                 goto err;
1703
1704         /*
1705          * If we have a gettimeofday() thread, make sure we exclude that
1706          * thread from this job
1707          */
1708         if (o->gtod_cpu)
1709                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1710
1711         /*
1712          * Set affinity first, in case it has an impact on the memory
1713          * allocations.
1714          */
1715         if (fio_option_is_set(o, cpumask)) {
1716                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1717                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1718                         if (!ret) {
1719                                 log_err("fio: no CPUs set\n");
1720                                 log_err("fio: Try increasing number of available CPUs\n");
1721                                 td_verror(td, EINVAL, "cpus_split");
1722                                 goto err;
1723                         }
1724                 }
1725                 ret = fio_setaffinity(td->pid, o->cpumask);
1726                 if (ret == -1) {
1727                         td_verror(td, errno, "cpu_set_affinity");
1728                         goto err;
1729                 }
1730         }
1731
1732 #ifdef CONFIG_LIBNUMA
1733         /* numa node setup */
1734         if (fio_option_is_set(o, numa_cpunodes) ||
1735             fio_option_is_set(o, numa_memnodes)) {
1736                 struct bitmask *mask;
1737
1738                 if (numa_available() < 0) {
1739                         td_verror(td, errno, "Does not support NUMA API\n");
1740                         goto err;
1741                 }
1742
1743                 if (fio_option_is_set(o, numa_cpunodes)) {
1744                         mask = numa_parse_nodestring(o->numa_cpunodes);
1745                         ret = numa_run_on_node_mask(mask);
1746                         numa_free_nodemask(mask);
1747                         if (ret == -1) {
1748                                 td_verror(td, errno, \
1749                                         "numa_run_on_node_mask failed\n");
1750                                 goto err;
1751                         }
1752                 }
1753
1754                 if (fio_option_is_set(o, numa_memnodes)) {
1755                         mask = NULL;
1756                         if (o->numa_memnodes)
1757                                 mask = numa_parse_nodestring(o->numa_memnodes);
1758
1759                         switch (o->numa_mem_mode) {
1760                         case MPOL_INTERLEAVE:
1761                                 numa_set_interleave_mask(mask);
1762                                 break;
1763                         case MPOL_BIND:
1764                                 numa_set_membind(mask);
1765                                 break;
1766                         case MPOL_LOCAL:
1767                                 numa_set_localalloc();
1768                                 break;
1769                         case MPOL_PREFERRED:
1770                                 numa_set_preferred(o->numa_mem_prefer_node);
1771                                 break;
1772                         case MPOL_DEFAULT:
1773                         default:
1774                                 break;
1775                         }
1776
1777                         if (mask)
1778                                 numa_free_nodemask(mask);
1779
1780                 }
1781         }
1782 #endif
1783
1784         if (fio_pin_memory(td))
1785                 goto err;
1786
1787         /*
1788          * May alter parameters that init_io_u() will use, so we need to
1789          * do this first.
1790          */
1791         if (!init_iolog(td))
1792                 goto err;
1793
1794         /* ioprio_set() has to be done before td_io_init() */
1795         if (fio_option_is_set(o, ioprio) ||
1796             fio_option_is_set(o, ioprio_class)) {
1797                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1798                 if (ret == -1) {
1799                         td_verror(td, errno, "ioprio_set");
1800                         goto err;
1801                 }
1802                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1803                 td->ts.ioprio = td->ioprio;
1804         }
1805
1806         if (td_io_init(td))
1807                 goto err;
1808
1809         if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
1810                 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1811                          "are selected, queue depth will be capped at 1\n");
1812         }
1813
1814         if (init_io_u(td))
1815                 goto err;
1816
1817         if (td->io_ops->post_init && td->io_ops->post_init(td))
1818                 goto err;
1819
1820         if (o->verify_async && verify_async_init(td))
1821                 goto err;
1822
1823         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1824                 goto err;
1825
1826         errno = 0;
1827         if (nice(o->nice) == -1 && errno != 0) {
1828                 td_verror(td, errno, "nice");
1829                 goto err;
1830         }
1831
1832         if (o->ioscheduler && switch_ioscheduler(td))
1833                 goto err;
1834
1835         if (!o->create_serialize && setup_files(td))
1836                 goto err;
1837
1838         if (!init_random_map(td))
1839                 goto err;
1840
1841         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1842                 goto err;
1843
1844         if (o->pre_read && !pre_read_files(td))
1845                 goto err;
1846
1847         fio_verify_init(td);
1848
1849         if (rate_submit_init(td, sk_out))
1850                 goto err;
1851
1852         set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1853         fio_getrusage(&td->ru_start);
1854         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1855         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1856         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1857
1858         init_thinktime(td);
1859
1860         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1861                         o->ratemin[DDIR_TRIM]) {
1862                 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1863                                         sizeof(td->bw_sample_time));
1864                 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1865                                         sizeof(td->bw_sample_time));
1866                 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1867                                         sizeof(td->bw_sample_time));
1868         }
1869
1870         memset(bytes_done, 0, sizeof(bytes_done));
1871         clear_state = false;
1872
1873         while (keep_running(td)) {
1874                 uint64_t verify_bytes;
1875
1876                 fio_gettime(&td->start, NULL);
1877                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1878
1879                 if (clear_state) {
1880                         clear_io_state(td, 0);
1881
1882                         if (o->unlink_each_loop && unlink_all_files(td))
1883                                 break;
1884                 }
1885
1886                 prune_io_piece_log(td);
1887
1888                 if (td->o.verify_only && td_write(td))
1889                         verify_bytes = do_dry_run(td);
1890                 else {
1891                         if (!td->o.rand_repeatable)
1892                                 /* save verify rand state to replay hdr seeds later at verify */
1893                                 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
1894                         do_io(td, bytes_done);
1895                         if (!td->o.rand_repeatable)
1896                                 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
1897                         if (!ddir_rw_sum(bytes_done)) {
1898                                 fio_mark_td_terminate(td);
1899                                 verify_bytes = 0;
1900                         } else {
1901                                 verify_bytes = bytes_done[DDIR_WRITE] +
1902                                                 bytes_done[DDIR_TRIM];
1903                         }
1904                 }
1905
1906                 /*
1907                  * If we took too long to shut down, the main thread could
1908                  * already consider us reaped/exited. If that happens, break
1909                  * out and clean up.
1910                  */
1911                 if (td->runstate >= TD_EXITED)
1912                         break;
1913
1914                 clear_state = true;
1915
1916                 /*
1917                  * Make sure we've successfully updated the rusage stats
1918                  * before waiting on the stat mutex. Otherwise we could have
1919                  * the stat thread holding stat mutex and waiting for
1920                  * the rusage_sem, which would never get upped because
1921                  * this thread is waiting for the stat mutex.
1922                  */
1923                 deadlock_loop_cnt = 0;
1924                 do {
1925                         check_update_rusage(td);
1926                         if (!fio_sem_down_trylock(stat_sem))
1927                                 break;
1928                         usleep(1000);
1929                         if (deadlock_loop_cnt++ > 5000) {
1930                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1931                                 td->error = EDEADLK;
1932                                 goto err;
1933                         }
1934                 } while (1);
1935
1936                 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1937                         ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
1938                         update_runtime(td, elapsed_us, DDIR_READ);
1939                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1940                         update_runtime(td, elapsed_us, DDIR_WRITE);
1941                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1942                         update_runtime(td, elapsed_us, DDIR_TRIM);
1943                 fio_gettime(&td->start, NULL);
1944                 fio_sem_up(stat_sem);
1945
1946                 if (td->error || td->terminate)
1947                         break;
1948
1949                 if (!o->do_verify ||
1950                     o->verify == VERIFY_NONE ||
1951                     td_ioengine_flagged(td, FIO_UNIDIR))
1952                         continue;
1953
1954                 clear_io_state(td, 0);
1955
1956                 fio_gettime(&td->start, NULL);
1957
1958                 do_verify(td, verify_bytes);
1959
1960                 /*
1961                  * See comment further up for why this is done here.
1962                  */
1963                 check_update_rusage(td);
1964
1965                 fio_sem_down(stat_sem);
1966                 update_runtime(td, elapsed_us, DDIR_READ);
1967                 fio_gettime(&td->start, NULL);
1968                 fio_sem_up(stat_sem);
1969
1970                 if (td->error || td->terminate)
1971                         break;
1972         }
1973
1974         /*
1975          * Acquire this lock if we were doing overlap checking in
1976          * offload mode so that we don't clean up this job while
1977          * another thread is checking its io_u's for overlap
1978          */
1979         if (td_offload_overlap(td)) {
1980                 int res = pthread_mutex_lock(&overlap_check);
1981                 assert(res == 0);
1982         }
1983         td_set_runstate(td, TD_FINISHING);
1984         if (td_offload_overlap(td)) {
1985                 res = pthread_mutex_unlock(&overlap_check);
1986                 assert(res == 0);
1987         }
1988
1989         update_rusage_stat(td);
1990         td->ts.total_run_time = mtime_since_now(&td->epoch);
1991         for_each_rw_ddir(ddir) {
1992                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1993         }
1994
1995         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1996             (td->o.verify != VERIFY_NONE && td_write(td)))
1997                 verify_save_state(td->thread_number);
1998
1999         fio_unpin_memory(td);
2000
2001         td_writeout_logs(td, true);
2002
2003         iolog_compress_exit(td);
2004         rate_submit_exit(td);
2005
2006         if (o->exec_postrun)
2007                 exec_string(o, o->exec_postrun, "postrun");
2008
2009         if (exitall_on_terminate || (o->exitall_error && td->error))
2010                 fio_terminate_threads(td->groupid, td->o.exit_what);
2011
2012 err:
2013         if (td->error)
2014                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
2015                                                         td->verror);
2016
2017         if (o->verify_async)
2018                 verify_async_exit(td);
2019
2020         close_and_free_files(td);
2021         cleanup_io_u(td);
2022         close_ioengine(td);
2023         cgroup_shutdown(td, cgroup_mnt);
2024         verify_free_state(td);
2025         td_zone_free_index(td);
2026
2027         if (fio_option_is_set(o, cpumask)) {
2028                 ret = fio_cpuset_exit(&o->cpumask);
2029                 if (ret)
2030                         td_verror(td, ret, "fio_cpuset_exit");
2031         }
2032
2033         /*
2034          * do this very late, it will log file closing as well
2035          */
2036         if (o->write_iolog_file)
2037                 write_iolog_close(td);
2038         if (td->io_log_rfile)
2039                 fclose(td->io_log_rfile);
2040
2041         td_set_runstate(td, TD_EXITED);
2042
2043         /*
2044          * Do this last after setting our runstate to exited, so we
2045          * know that the stat thread is signaled.
2046          */
2047         check_update_rusage(td);
2048
2049         sk_out_drop();
2050         return (void *) (uintptr_t) td->error;
2051 }
2052
2053 /*
2054  * Run over the job map and reap the threads that have exited, if any.
2055  */
2056 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2057                          uint64_t *m_rate)
2058 {
2059         struct thread_data *td;
2060         unsigned int cputhreads, realthreads, pending;
2061         int i, status, ret;
2062
2063         /*
2064          * reap exited threads (TD_EXITED -> TD_REAPED)
2065          */
2066         realthreads = pending = cputhreads = 0;
2067         for_each_td(td, i) {
2068                 int flags = 0;
2069
2070                 if (!strcmp(td->o.ioengine, "cpuio"))
2071                         cputhreads++;
2072                 else
2073                         realthreads++;
2074
2075                 if (!td->pid) {
2076                         pending++;
2077                         continue;
2078                 }
2079                 if (td->runstate == TD_REAPED)
2080                         continue;
2081                 if (td->o.use_thread) {
2082                         if (td->runstate == TD_EXITED) {
2083                                 td_set_runstate(td, TD_REAPED);
2084                                 goto reaped;
2085                         }
2086                         continue;
2087                 }
2088
2089                 flags = WNOHANG;
2090                 if (td->runstate == TD_EXITED)
2091                         flags = 0;
2092
2093                 /*
2094                  * check if someone quit or got killed in an unusual way
2095                  */
2096                 ret = waitpid(td->pid, &status, flags);
2097                 if (ret < 0) {
2098                         if (errno == ECHILD) {
2099                                 log_err("fio: pid=%d disappeared %d\n",
2100                                                 (int) td->pid, td->runstate);
2101                                 td->sig = ECHILD;
2102                                 td_set_runstate(td, TD_REAPED);
2103                                 goto reaped;
2104                         }
2105                         perror("waitpid");
2106                 } else if (ret == td->pid) {
2107                         if (WIFSIGNALED(status)) {
2108                                 int sig = WTERMSIG(status);
2109
2110                                 if (sig != SIGTERM && sig != SIGUSR2)
2111                                         log_err("fio: pid=%d, got signal=%d\n",
2112                                                         (int) td->pid, sig);
2113                                 td->sig = sig;
2114                                 td_set_runstate(td, TD_REAPED);
2115                                 goto reaped;
2116                         }
2117                         if (WIFEXITED(status)) {
2118                                 if (WEXITSTATUS(status) && !td->error)
2119                                         td->error = WEXITSTATUS(status);
2120
2121                                 td_set_runstate(td, TD_REAPED);
2122                                 goto reaped;
2123                         }
2124                 }
2125
2126                 /*
2127                  * If the job is stuck, do a forceful timeout of it and
2128                  * move on.
2129                  */
2130                 if (td->terminate &&
2131                     td->runstate < TD_FSYNCING &&
2132                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2133                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2134                                 "%lu seconds, it appears to be stuck. Doing "
2135                                 "forceful exit of this job.\n",
2136                                 td->o.name, td->runstate,
2137                                 (unsigned long) time_since_now(&td->terminate_time));
2138                         td_set_runstate(td, TD_REAPED);
2139                         goto reaped;
2140                 }
2141
2142                 /*
2143                  * thread is not dead, continue
2144                  */
2145                 pending++;
2146                 continue;
2147 reaped:
2148                 (*nr_running)--;
2149                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2150                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2151                 if (!td->pid)
2152                         pending--;
2153
2154                 if (td->error)
2155                         exit_value++;
2156
2157                 done_secs += mtime_since_now(&td->epoch) / 1000;
2158                 profile_td_exit(td);
2159                 flow_exit_job(td);
2160         }
2161
2162         if (*nr_running == cputhreads && !pending && realthreads)
2163                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2164 }
2165
2166 static bool __check_trigger_file(void)
2167 {
2168         struct stat sb;
2169
2170         if (!trigger_file)
2171                 return false;
2172
2173         if (stat(trigger_file, &sb))
2174                 return false;
2175
2176         if (unlink(trigger_file) < 0)
2177                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2178                                                         strerror(errno));
2179
2180         return true;
2181 }
2182
2183 static bool trigger_timedout(void)
2184 {
2185         if (trigger_timeout)
2186                 if (time_since_genesis() >= trigger_timeout) {
2187                         trigger_timeout = 0;
2188                         return true;
2189                 }
2190
2191         return false;
2192 }
2193
2194 void exec_trigger(const char *cmd)
2195 {
2196         int ret;
2197
2198         if (!cmd || cmd[0] == '\0')
2199                 return;
2200
2201         ret = system(cmd);
2202         if (ret == -1)
2203                 log_err("fio: failed executing %s trigger\n", cmd);
2204 }
2205
2206 void check_trigger_file(void)
2207 {
2208         if (__check_trigger_file() || trigger_timedout()) {
2209                 if (nr_clients)
2210                         fio_clients_send_trigger(trigger_remote_cmd);
2211                 else {
2212                         verify_save_state(IO_LIST_ALL);
2213                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2214                         exec_trigger(trigger_cmd);
2215                 }
2216         }
2217 }
2218
2219 static int fio_verify_load_state(struct thread_data *td)
2220 {
2221         int ret;
2222
2223         if (!td->o.verify_state)
2224                 return 0;
2225
2226         if (is_backend) {
2227                 void *data;
2228
2229                 ret = fio_server_get_verify_state(td->o.name,
2230                                         td->thread_number - 1, &data);
2231                 if (!ret)
2232                         verify_assign_state(td, data);
2233         } else {
2234                 char prefix[PATH_MAX];
2235
2236                 if (aux_path)
2237                         sprintf(prefix, "%s%clocal", aux_path,
2238                                         FIO_OS_PATH_SEPARATOR);
2239                 else
2240                         strcpy(prefix, "local");
2241                 ret = verify_load_state(td, prefix);
2242         }
2243
2244         return ret;
2245 }
2246
2247 static void do_usleep(unsigned int usecs)
2248 {
2249         check_for_running_stats();
2250         check_trigger_file();
2251         usleep(usecs);
2252 }
2253
2254 static bool check_mount_writes(struct thread_data *td)
2255 {
2256         struct fio_file *f;
2257         unsigned int i;
2258
2259         if (!td_write(td) || td->o.allow_mounted_write)
2260                 return false;
2261
2262         /*
2263          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2264          * are mkfs'd and mounted.
2265          */
2266         for_each_file(td, f, i) {
2267 #ifdef FIO_HAVE_CHARDEV_SIZE
2268                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2269 #else
2270                 if (f->filetype != FIO_TYPE_BLOCK)
2271 #endif
2272                         continue;
2273                 if (device_is_mounted(f->file_name))
2274                         goto mounted;
2275         }
2276
2277         return false;
2278 mounted:
2279         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2280         return true;
2281 }
2282
2283 static bool waitee_running(struct thread_data *me)
2284 {
2285         const char *waitee = me->o.wait_for;
2286         const char *self = me->o.name;
2287         struct thread_data *td;
2288         int i;
2289
2290         if (!waitee)
2291                 return false;
2292
2293         for_each_td(td, i) {
2294                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2295                         continue;
2296
2297                 if (td->runstate < TD_EXITED) {
2298                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2299                                         self, td->o.name,
2300                                         runstate_to_name(td->runstate));
2301                         return true;
2302                 }
2303         }
2304
2305         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2306         return false;
2307 }
2308
2309 /*
2310  * Main function for kicking off and reaping jobs, as needed.
2311  */
2312 static void run_threads(struct sk_out *sk_out)
2313 {
2314         struct thread_data *td;
2315         unsigned int i, todo, nr_running, nr_started;
2316         uint64_t m_rate, t_rate;
2317         uint64_t spent;
2318
2319         if (fio_gtod_offload && fio_start_gtod_thread())
2320                 return;
2321
2322         fio_idle_prof_init();
2323
2324         set_sig_handlers();
2325
2326         nr_thread = nr_process = 0;
2327         for_each_td(td, i) {
2328                 if (check_mount_writes(td))
2329                         return;
2330                 if (td->o.use_thread)
2331                         nr_thread++;
2332                 else
2333                         nr_process++;
2334         }
2335
2336         if (output_format & FIO_OUTPUT_NORMAL) {
2337                 struct buf_output out;
2338
2339                 buf_output_init(&out);
2340                 __log_buf(&out, "Starting ");
2341                 if (nr_thread)
2342                         __log_buf(&out, "%d thread%s", nr_thread,
2343                                                 nr_thread > 1 ? "s" : "");
2344                 if (nr_process) {
2345                         if (nr_thread)
2346                                 __log_buf(&out, " and ");
2347                         __log_buf(&out, "%d process%s", nr_process,
2348                                                 nr_process > 1 ? "es" : "");
2349                 }
2350                 __log_buf(&out, "\n");
2351                 log_info_buf(out.buf, out.buflen);
2352                 buf_output_free(&out);
2353         }
2354
2355         todo = thread_number;
2356         nr_running = 0;
2357         nr_started = 0;
2358         m_rate = t_rate = 0;
2359
2360         for_each_td(td, i) {
2361                 print_status_init(td->thread_number - 1);
2362
2363                 if (!td->o.create_serialize)
2364                         continue;
2365
2366                 if (fio_verify_load_state(td))
2367                         goto reap;
2368
2369                 /*
2370                  * do file setup here so it happens sequentially,
2371                  * we don't want X number of threads getting their
2372                  * client data interspersed on disk
2373                  */
2374                 if (setup_files(td)) {
2375 reap:
2376                         exit_value++;
2377                         if (td->error)
2378                                 log_err("fio: pid=%d, err=%d/%s\n",
2379                                         (int) td->pid, td->error, td->verror);
2380                         td_set_runstate(td, TD_REAPED);
2381                         todo--;
2382                 } else {
2383                         struct fio_file *f;
2384                         unsigned int j;
2385
2386                         /*
2387                          * for sharing to work, each job must always open
2388                          * its own files. so close them, if we opened them
2389                          * for creation
2390                          */
2391                         for_each_file(td, f, j) {
2392                                 if (fio_file_open(f))
2393                                         td_io_close_file(td, f);
2394                         }
2395                 }
2396         }
2397
2398         /* start idle threads before io threads start to run */
2399         fio_idle_prof_start();
2400
2401         set_genesis_time();
2402
2403         while (todo) {
2404                 struct thread_data *map[REAL_MAX_JOBS];
2405                 struct timespec this_start;
2406                 int this_jobs = 0, left;
2407                 struct fork_data *fd;
2408
2409                 /*
2410                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2411                  */
2412                 for_each_td(td, i) {
2413                         if (td->runstate != TD_NOT_CREATED)
2414                                 continue;
2415
2416                         /*
2417                          * never got a chance to start, killed by other
2418                          * thread for some reason
2419                          */
2420                         if (td->terminate) {
2421                                 todo--;
2422                                 continue;
2423                         }
2424
2425                         if (td->o.start_delay) {
2426                                 spent = utime_since_genesis();
2427
2428                                 if (td->o.start_delay > spent)
2429                                         continue;
2430                         }
2431
2432                         if (td->o.stonewall && (nr_started || nr_running)) {
2433                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2434                                                         td->o.name);
2435                                 break;
2436                         }
2437
2438                         if (waitee_running(td)) {
2439                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2440                                                 td->o.name, td->o.wait_for);
2441                                 continue;
2442                         }
2443
2444                         init_disk_util(td);
2445
2446                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2447                         td->update_rusage = 0;
2448
2449                         /*
2450                          * Set state to created. Thread will transition
2451                          * to TD_INITIALIZED when it's done setting up.
2452                          */
2453                         td_set_runstate(td, TD_CREATED);
2454                         map[this_jobs++] = td;
2455                         nr_started++;
2456
2457                         fd = calloc(1, sizeof(*fd));
2458                         fd->td = td;
2459                         fd->sk_out = sk_out;
2460
2461                         if (td->o.use_thread) {
2462                                 int ret;
2463
2464                                 dprint(FD_PROCESS, "will pthread_create\n");
2465                                 ret = pthread_create(&td->thread, NULL,
2466                                                         thread_main, fd);
2467                                 if (ret) {
2468                                         log_err("pthread_create: %s\n",
2469                                                         strerror(ret));
2470                                         free(fd);
2471                                         nr_started--;
2472                                         break;
2473                                 }
2474                                 fd = NULL;
2475                                 ret = pthread_detach(td->thread);
2476                                 if (ret)
2477                                         log_err("pthread_detach: %s",
2478                                                         strerror(ret));
2479                         } else {
2480                                 pid_t pid;
2481                                 void *eo;
2482                                 dprint(FD_PROCESS, "will fork\n");
2483                                 eo = td->eo;
2484                                 read_barrier();
2485                                 pid = fork();
2486                                 if (!pid) {
2487                                         int ret;
2488
2489                                         ret = (int)(uintptr_t)thread_main(fd);
2490                                         _exit(ret);
2491                                 } else if (i == fio_debug_jobno)
2492                                         *fio_debug_jobp = pid;
2493                                 free(eo);
2494                                 free(fd);
2495                                 fd = NULL;
2496                         }
2497                         dprint(FD_MUTEX, "wait on startup_sem\n");
2498                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2499                                 log_err("fio: job startup hung? exiting.\n");
2500                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2501                                 fio_abort = true;
2502                                 nr_started--;
2503                                 free(fd);
2504                                 break;
2505                         }
2506                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2507                 }
2508
2509                 /*
2510                  * Wait for the started threads to transition to
2511                  * TD_INITIALIZED.
2512                  */
2513                 fio_gettime(&this_start, NULL);
2514                 left = this_jobs;
2515                 while (left && !fio_abort) {
2516                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2517                                 break;
2518
2519                         do_usleep(100000);
2520
2521                         for (i = 0; i < this_jobs; i++) {
2522                                 td = map[i];
2523                                 if (!td)
2524                                         continue;
2525                                 if (td->runstate == TD_INITIALIZED) {
2526                                         map[i] = NULL;
2527                                         left--;
2528                                 } else if (td->runstate >= TD_EXITED) {
2529                                         map[i] = NULL;
2530                                         left--;
2531                                         todo--;
2532                                         nr_running++; /* work-around... */
2533                                 }
2534                         }
2535                 }
2536
2537                 if (left) {
2538                         log_err("fio: %d job%s failed to start\n", left,
2539                                         left > 1 ? "s" : "");
2540                         for (i = 0; i < this_jobs; i++) {
2541                                 td = map[i];
2542                                 if (!td)
2543                                         continue;
2544                                 kill(td->pid, SIGTERM);
2545                         }
2546                         break;
2547                 }
2548
2549                 /*
2550                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2551                  */
2552                 for_each_td(td, i) {
2553                         if (td->runstate != TD_INITIALIZED)
2554                                 continue;
2555
2556                         if (in_ramp_time(td))
2557                                 td_set_runstate(td, TD_RAMP);
2558                         else
2559                                 td_set_runstate(td, TD_RUNNING);
2560                         nr_running++;
2561                         nr_started--;
2562                         m_rate += ddir_rw_sum(td->o.ratemin);
2563                         t_rate += ddir_rw_sum(td->o.rate);
2564                         todo--;
2565                         fio_sem_up(td->sem);
2566                 }
2567
2568                 reap_threads(&nr_running, &t_rate, &m_rate);
2569
2570                 if (todo)
2571                         do_usleep(100000);
2572         }
2573
2574         while (nr_running) {
2575                 reap_threads(&nr_running, &t_rate, &m_rate);
2576                 do_usleep(10000);
2577         }
2578
2579         fio_idle_prof_stop();
2580
2581         update_io_ticks();
2582 }
2583
2584 static void free_disk_util(void)
2585 {
2586         disk_util_prune_entries();
2587         helper_thread_destroy();
2588 }
2589
2590 int fio_backend(struct sk_out *sk_out)
2591 {
2592         struct thread_data *td;
2593         int i;
2594
2595         if (exec_profile) {
2596                 if (load_profile(exec_profile))
2597                         return 1;
2598                 free(exec_profile);
2599                 exec_profile = NULL;
2600         }
2601         if (!thread_number)
2602                 return 0;
2603
2604         if (write_bw_log) {
2605                 struct log_params p = {
2606                         .log_type = IO_LOG_TYPE_BW,
2607                 };
2608
2609                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2610                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2611                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2612         }
2613
2614         if (init_global_dedupe_working_set_seeds()) {
2615                 log_err("fio: failed to initialize global dedupe working set\n");
2616                 return 1;
2617         }
2618
2619         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2620         if (!sk_out)
2621                 is_local_backend = true;
2622         if (startup_sem == NULL)
2623                 return 1;
2624
2625         set_genesis_time();
2626         stat_init();
2627         if (helper_thread_create(startup_sem, sk_out))
2628                 log_err("fio: failed to create helper thread\n");
2629
2630         cgroup_list = smalloc(sizeof(*cgroup_list));
2631         if (cgroup_list)
2632                 INIT_FLIST_HEAD(cgroup_list);
2633
2634         run_threads(sk_out);
2635
2636         helper_thread_exit();
2637
2638         if (!fio_abort) {
2639                 __show_run_stats();
2640                 if (write_bw_log) {
2641                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2642                                 struct io_log *log = agg_io_log[i];
2643
2644                                 flush_log(log, false);
2645                                 free_log(log);
2646                         }
2647                 }
2648         }
2649
2650         for_each_td(td, i) {
2651                 struct thread_stat *ts = &td->ts;
2652
2653                 free_clat_prio_stats(ts);
2654                 steadystate_free(td);
2655                 fio_options_free(td);
2656                 fio_dump_options_free(td);
2657                 if (td->rusage_sem) {
2658                         fio_sem_remove(td->rusage_sem);
2659                         td->rusage_sem = NULL;
2660                 }
2661                 fio_sem_remove(td->sem);
2662                 td->sem = NULL;
2663         }
2664
2665         free_disk_util();
2666         if (cgroup_list) {
2667                 cgroup_kill(cgroup_list);
2668                 sfree(cgroup_list);
2669         }
2670
2671         fio_sem_remove(startup_sem);
2672         stat_exit();
2673         return exit_value;
2674 }