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