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