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                                 clear_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         /*
 901          * usec_spin() might run for slightly longer than intended in a VM
 902          * where the vCPU could get descheduled or the hypervisor could steal
 903          * CPU time. Ensure "left" doesn't become negative.
 904          */
 905         if (total < td->o.thinktime)
 906                 left = td->o.thinktime - total;
 907         else
 908                 left = 0;
 909
 910         if (td->o.timeout) {
 911                 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
 912                 if (runtime_left < (unsigned long long)left)
 913                         left = runtime_left;
 914         }
 915
 916         if (left)
 917                 total += usec_sleep(td, left);
 918
 919         /*
 920          * If we're ignoring thinktime for the rate, add the number of bytes
 921          * we would have done while sleeping, minus one block to ensure we
 922          * start issuing immediately after the sleep.
 923          */
 924         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 925                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 926                 uint64_t bs = td->o.min_bs[ddir];
 927                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 928                 uint64_t over;
 929
 930                 if (usperop <= total)
 931                         over = bs;
 932                 else
 933                         over = (usperop - total) / usperop * -bs;
 934
 935                 td->rate_io_issue_bytes[ddir] += (missed - over);
 936                 /* adjust for rate_process=poisson */
 937                 td->last_usec[ddir] += total;
 938         }
 939
 940         if (time && should_check_rate(td))
 941                 fio_gettime(time, NULL);
 942
 943         td->last_thinktime_blocks = b;
 944         if (td->o.thinktime_iotime) {
 945                 fio_gettime(&now, NULL);
 946                 td->last_thinktime = now;
 947         }
 948 }
 949
 950 /*
 951  * Main IO worker function. It retrieves io_u's to process and queues
 952  * and reaps them, checking for rate and errors along the way.
 953  *
 954  * Returns number of bytes written and trimmed.
 955  */
 956 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 957 {
 958         unsigned int i;
 959         int ret = 0;
 960         uint64_t total_bytes, bytes_issued = 0;
 961
 962         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 963                 bytes_done[i] = td->bytes_done[i];
 964
 965         if (in_ramp_time(td))
 966                 td_set_runstate(td, TD_RAMP);
 967         else
 968                 td_set_runstate(td, TD_RUNNING);
 969
 970         lat_target_init(td);
 971
 972         total_bytes = td->o.size;
 973         /*
 974         * Allow random overwrite workloads to write up to io_size
 975         * before starting verification phase as 'size' doesn't apply.
 976         */
 977         if (td_write(td) && td_random(td) && td->o.norandommap)
 978                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 979         /*
 980          * If verify_backlog is enabled, we'll run the verify in this
 981          * handler as well. For that case, we may need up to twice the
 982          * amount of bytes.
 983          */
 984         if (td->o.verify != VERIFY_NONE &&
 985            (td_write(td) && td->o.verify_backlog))
 986                 total_bytes += td->o.size;
 987
 988         /* In trimwrite mode, each byte is trimmed and then written, so
 989          * allow total_bytes or number of ios to be twice as big */
 990         if (td_trimwrite(td)) {
 991                 total_bytes += td->total_io_size;
 992                 td->o.number_ios *= 2;
 993         }
 994
 995         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 996                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 997                 td->o.time_based) {
 998                 struct timespec comp_time;
 999                 struct io_u *io_u;
1000                 int full;
1001                 enum fio_ddir ddir;
1002
1003                 check_update_rusage(td);
1004
1005                 if (td->terminate || td->done)
1006                         break;
1007
1008                 update_ts_cache(td);
1009
1010                 if (runtime_exceeded(td, &td->ts_cache)) {
1011                         __update_ts_cache(td);
1012                         if (runtime_exceeded(td, &td->ts_cache)) {
1013                                 fio_mark_td_terminate(td);
1014                                 break;
1015                         }
1016                 }
1017
1018                 if (flow_threshold_exceeded(td))
1019                         continue;
1020
1021                 /*
1022                  * Break if we exceeded the bytes. The exception is time
1023                  * based runs, but we still need to break out of the loop
1024                  * for those to run verification, if enabled.
1025                  * Jobs read from iolog do not use this stop condition.
1026                  */
1027                 if (bytes_issued >= total_bytes &&
1028                     !td->o.read_iolog_file &&
1029                     (!td->o.time_based ||
1030                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
1031                         break;
1032
1033                 io_u = get_io_u(td);
1034                 if (IS_ERR_OR_NULL(io_u)) {
1035                         int err = PTR_ERR(io_u);
1036
1037                         io_u = NULL;
1038                         ddir = DDIR_INVAL;
1039                         if (err == -EBUSY) {
1040                                 ret = FIO_Q_BUSY;
1041                                 goto reap;
1042                         }
1043                         if (td->o.latency_target)
1044                                 goto reap;
1045                         break;
1046                 }
1047
1048                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
1049                         if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
1050                                 io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
1051                                 io_u->numberio = td->io_issues[io_u->ddir];
1052                                 populate_verify_io_u(td, io_u);
1053                         }
1054                 }
1055
1056                 ddir = io_u->ddir;
1057
1058                 /*
1059                  * Add verification end_io handler if:
1060                  *      - Asked to verify (!td_rw(td))
1061                  *      - Or the io_u is from our verify list (mixed write/ver)
1062                  */
1063                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1064                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1065
1066                         if (verify_state_should_stop(td, io_u)) {
1067                                 put_io_u(td, io_u);
1068                                 break;
1069                         }
1070
1071                         if (td->o.verify_async)
1072                                 io_u->end_io = verify_io_u_async;
1073                         else
1074                                 io_u->end_io = verify_io_u;
1075                         td_set_runstate(td, TD_VERIFYING);
1076                 } else if (in_ramp_time(td))
1077                         td_set_runstate(td, TD_RAMP);
1078                 else
1079                         td_set_runstate(td, TD_RUNNING);
1080
1081                 /*
1082                  * Always log IO before it's issued, so we know the specific
1083                  * order of it. The logged unit will track when the IO has
1084                  * completed.
1085                  */
1086                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1087                     td->o.do_verify &&
1088                     td->o.verify != VERIFY_NONE &&
1089                     !td->o.experimental_verify)
1090                         log_io_piece(td, io_u);
1091
1092                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1093                         const unsigned long long blen = io_u->xfer_buflen;
1094                         const enum fio_ddir __ddir = acct_ddir(io_u);
1095
1096                         if (td->error)
1097                                 break;
1098
1099                         workqueue_enqueue(&td->io_wq, &io_u->work);
1100                         ret = FIO_Q_QUEUED;
1101
1102                         if (ddir_rw(__ddir)) {
1103                                 td->io_issues[__ddir]++;
1104                                 td->io_issue_bytes[__ddir] += blen;
1105                                 td->rate_io_issue_bytes[__ddir] += blen;
1106                         }
1107
1108                         if (should_check_rate(td)) {
1109                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1110                                 fio_gettime(&comp_time, NULL);
1111                         }
1112
1113                 } else {
1114                         ret = io_u_submit(td, io_u);
1115
1116                         if (should_check_rate(td))
1117                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1118
1119                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1120                                 break;
1121
1122                         /*
1123                          * See if we need to complete some commands. Note that
1124                          * we can get BUSY even without IO queued, if the
1125                          * system is resource starved.
1126                          */
1127 reap:
1128                         full = queue_full(td) ||
1129                                 (ret == FIO_Q_BUSY && td->cur_depth);
1130                         if (full || io_in_polling(td))
1131                                 ret = wait_for_completions(td, &comp_time);
1132                 }
1133                 if (ret < 0)
1134                         break;
1135
1136                 if (ddir_rw(ddir) && td->o.thinktime)
1137                         handle_thinktime(td, ddir, &comp_time);
1138
1139                 if (!ddir_rw_sum(td->bytes_done) &&
1140                     !td_ioengine_flagged(td, FIO_NOIO))
1141                         continue;
1142
1143                 if (!in_ramp_time(td) && should_check_rate(td)) {
1144                         if (check_min_rate(td, &comp_time)) {
1145                                 if (exitall_on_terminate || td->o.exitall_error)
1146                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1147                                 td_verror(td, EIO, "check_min_rate");
1148                                 break;
1149                         }
1150                 }
1151                 if (!in_ramp_time(td) && td->o.latency_target)
1152                         lat_target_check(td);
1153         }
1154
1155         check_update_rusage(td);
1156
1157         if (td->trim_entries)
1158                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1159
1160         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1161                 td->error = 0;
1162                 fio_mark_td_terminate(td);
1163         }
1164         if (!td->error) {
1165                 struct fio_file *f;
1166
1167                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1168                         workqueue_flush(&td->io_wq);
1169                         i = 0;
1170                 } else
1171                         i = td->cur_depth;
1172
1173                 if (i) {
1174                         ret = io_u_queued_complete(td, i);
1175                         if (td->o.fill_device &&
1176                             (td->error == ENOSPC || td->error == EDQUOT))
1177                                 td->error = 0;
1178                 }
1179
1180                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1181                         td_set_runstate(td, TD_FSYNCING);
1182
1183                         for_each_file(td, f, i) {
1184                                 if (!fio_file_fsync(td, f))
1185                                         continue;
1186
1187                                 log_err("fio: end_fsync failed for file %s\n",
1188                                                                 f->file_name);
1189                         }
1190                 }
1191         } else {
1192                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1193                         workqueue_flush(&td->io_wq);
1194                 cleanup_pending_aio(td);
1195         }
1196
1197         /*
1198          * stop job if we failed doing any IO
1199          */
1200         if (!ddir_rw_sum(td->this_io_bytes))
1201                 td->done = 1;
1202
1203         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1204                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1205 }
1206
1207 static void free_file_completion_logging(struct thread_data *td)
1208 {
1209         struct fio_file *f;
1210         unsigned int i;
1211
1212         for_each_file(td, f, i) {
1213                 if (!f->last_write_comp)
1214                         break;
1215                 sfree(f->last_write_comp);
1216         }
1217 }
1218
1219 static int init_file_completion_logging(struct thread_data *td,
1220                                         unsigned int depth)
1221 {
1222         struct fio_file *f;
1223         unsigned int i;
1224
1225         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1226                 return 0;
1227
1228         for_each_file(td, f, i) {
1229                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1230                 if (!f->last_write_comp)
1231                         goto cleanup;
1232         }
1233
1234         return 0;
1235
1236 cleanup:
1237         free_file_completion_logging(td);
1238         log_err("fio: failed to alloc write comp data\n");
1239         return 1;
1240 }
1241
1242 static void cleanup_io_u(struct thread_data *td)
1243 {
1244         struct io_u *io_u;
1245
1246         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1247
1248                 if (td->io_ops->io_u_free)
1249                         td->io_ops->io_u_free(td, io_u);
1250
1251                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1252         }
1253
1254         free_io_mem(td);
1255
1256         io_u_rexit(&td->io_u_requeues);
1257         io_u_qexit(&td->io_u_freelist, false);
1258         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1259
1260         free_file_completion_logging(td);
1261 }
1262
1263 static int init_io_u(struct thread_data *td)
1264 {
1265         struct io_u *io_u;
1266         int cl_align, i, max_units;
1267         int err;
1268
1269         max_units = td->o.iodepth;
1270
1271         err = 0;
1272         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1273         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1274         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1275
1276         if (err) {
1277                 log_err("fio: failed setting up IO queues\n");
1278                 return 1;
1279         }
1280
1281         cl_align = os_cache_line_size();
1282
1283         for (i = 0; i < max_units; i++) {
1284                 void *ptr;
1285
1286                 if (td->terminate)
1287                         return 1;
1288
1289                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1290                 if (!ptr) {
1291                         log_err("fio: unable to allocate aligned memory\n");
1292                         return 1;
1293                 }
1294
1295                 io_u = ptr;
1296                 memset(io_u, 0, sizeof(*io_u));
1297                 INIT_FLIST_HEAD(&io_u->verify_list);
1298                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1299
1300                 io_u->index = i;
1301                 io_u->flags = IO_U_F_FREE;
1302                 io_u_qpush(&td->io_u_freelist, io_u);
1303
1304                 /*
1305                  * io_u never leaves this stack, used for iteration of all
1306                  * io_u buffers.
1307                  */
1308                 io_u_qpush(&td->io_u_all, io_u);
1309
1310                 if (td->io_ops->io_u_init) {
1311                         int ret = td->io_ops->io_u_init(td, io_u);
1312
1313                         if (ret) {
1314                                 log_err("fio: failed to init engine data: %d\n", ret);
1315                                 return 1;
1316                         }
1317                 }
1318         }
1319
1320         if (init_io_u_buffers(td))
1321                 return 1;
1322
1323         if (init_file_completion_logging(td, max_units))
1324                 return 1;
1325
1326         return 0;
1327 }
1328
1329 int init_io_u_buffers(struct thread_data *td)
1330 {
1331         struct io_u *io_u;
1332         unsigned long long max_bs, min_write;
1333         int i, max_units;
1334         int data_xfer = 1;
1335         char *p;
1336
1337         max_units = td->o.iodepth;
1338         max_bs = td_max_bs(td);
1339         min_write = td->o.min_bs[DDIR_WRITE];
1340         td->orig_buffer_size = (unsigned long long) max_bs
1341                                         * (unsigned long long) max_units;
1342
1343         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1344                 data_xfer = 0;
1345
1346         /*
1347          * if we may later need to do address alignment, then add any
1348          * possible adjustment here so that we don't cause a buffer
1349          * overflow later. this adjustment may be too much if we get
1350          * lucky and the allocator gives us an aligned address.
1351          */
1352         if (td->o.odirect || td->o.mem_align ||
1353             td_ioengine_flagged(td, FIO_RAWIO))
1354                 td->orig_buffer_size += page_mask + td->o.mem_align;
1355
1356         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1357                 unsigned long long bs;
1358
1359                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1360                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1361         }
1362
1363         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1364                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1365                 return 1;
1366         }
1367
1368         if (data_xfer && allocate_io_mem(td))
1369                 return 1;
1370
1371         if (td->o.odirect || td->o.mem_align ||
1372             td_ioengine_flagged(td, FIO_RAWIO))
1373                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1374         else
1375                 p = td->orig_buffer;
1376
1377         for (i = 0; i < max_units; i++) {
1378                 io_u = td->io_u_all.io_us[i];
1379                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1380
1381                 if (data_xfer) {
1382                         io_u->buf = p;
1383                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1384
1385                         if (td_write(td))
1386                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1387                         if (td_write(td) && td->o.verify_pattern_bytes) {
1388                                 /*
1389                                  * Fill the buffer with the pattern if we are
1390                                  * going to be doing writes.
1391                                  */
1392                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1393                         }
1394                 }
1395                 p += max_bs;
1396         }
1397
1398         return 0;
1399 }
1400
1401 #ifdef FIO_HAVE_IOSCHED_SWITCH
1402 /*
1403  * These functions are Linux specific.
1404  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1405  */
1406 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1407 {
1408         char tmp[256], tmp2[128], *p;
1409         FILE *f;
1410         int ret;
1411
1412         assert(file->du && file->du->sysfs_root);
1413         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1414
1415         f = fopen(tmp, "r+");
1416         if (!f) {
1417                 if (errno == ENOENT) {
1418                         log_err("fio: os or kernel doesn't support IO scheduler"
1419                                 " switching\n");
1420                         return 0;
1421                 }
1422                 td_verror(td, errno, "fopen iosched");
1423                 return 1;
1424         }
1425
1426         /*
1427          * Set io scheduler.
1428          */
1429         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1430         if (ferror(f) || ret != 1) {
1431                 td_verror(td, errno, "fwrite");
1432                 fclose(f);
1433                 return 1;
1434         }
1435
1436         rewind(f);
1437
1438         /*
1439          * Read back and check that the selected scheduler is now the default.
1440          */
1441         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1442         if (ferror(f) || ret < 0) {
1443                 td_verror(td, errno, "fread");
1444                 fclose(f);
1445                 return 1;
1446         }
1447         tmp[ret] = '\0';
1448         /*
1449          * either a list of io schedulers or "none\n" is expected. Strip the
1450          * trailing newline.
1451          */
1452         p = tmp;
1453         strsep(&p, "\n");
1454
1455         /*
1456          * Write to "none" entry doesn't fail, so check the result here.
1457          */
1458         if (!strcmp(tmp, "none")) {
1459                 log_err("fio: io scheduler is not tunable\n");
1460                 fclose(f);
1461                 return 0;
1462         }
1463
1464         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1465         if (!strstr(tmp, tmp2)) {
1466                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1467                 td_verror(td, EINVAL, "iosched_switch");
1468                 fclose(f);
1469                 return 1;
1470         }
1471
1472         fclose(f);
1473         return 0;
1474 }
1475
1476 static int switch_ioscheduler(struct thread_data *td)
1477 {
1478         struct fio_file *f;
1479         unsigned int i;
1480         int ret = 0;
1481
1482         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1483                 return 0;
1484
1485         assert(td->files && td->files[0]);
1486
1487         for_each_file(td, f, i) {
1488
1489                 /* Only consider regular files and block device files */
1490                 switch (f->filetype) {
1491                 case FIO_TYPE_FILE:
1492                 case FIO_TYPE_BLOCK:
1493                         /*
1494                          * Make sure that the device hosting the file could
1495                          * be determined.
1496                          */
1497                         if (!f->du)
1498                                 continue;
1499                         break;
1500                 case FIO_TYPE_CHAR:
1501                 case FIO_TYPE_PIPE:
1502                 default:
1503                         continue;
1504                 }
1505
1506                 ret = set_ioscheduler(td, f);
1507                 if (ret)
1508                         return ret;
1509         }
1510
1511         return 0;
1512 }
1513
1514 #else
1515
1516 static int switch_ioscheduler(struct thread_data *td)
1517 {
1518         return 0;
1519 }
1520
1521 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1522
1523 static bool keep_running(struct thread_data *td)
1524 {
1525         unsigned long long limit;
1526
1527         if (td->done)
1528                 return false;
1529         if (td->terminate)
1530                 return false;
1531         if (td->o.time_based)
1532                 return true;
1533         if (td->o.loops) {
1534                 td->o.loops--;
1535                 return true;
1536         }
1537         if (exceeds_number_ios(td))
1538                 return false;
1539
1540         if (td->o.io_size)
1541                 limit = td->o.io_size;
1542         else
1543                 limit = td->o.size;
1544
1545         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1546                 uint64_t diff;
1547
1548                 /*
1549                  * If the difference is less than the maximum IO size, we
1550                  * are done.
1551                  */
1552                 diff = limit - ddir_rw_sum(td->io_bytes);
1553                 if (diff < td_max_bs(td))
1554                         return false;
1555
1556                 if (fio_files_done(td) && !td->o.io_size)
1557                         return false;
1558
1559                 return true;
1560         }
1561
1562         return false;
1563 }
1564
1565 static int exec_string(struct thread_options *o, const char *string,
1566                        const char *mode)
1567 {
1568         int ret;
1569         char *str;
1570
1571         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1572                 return -1;
1573
1574         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1575                  o->name, mode);
1576         ret = system(str);
1577         if (ret == -1)
1578                 log_err("fio: exec of cmd <%s> failed\n", str);
1579
1580         free(str);
1581         return ret;
1582 }
1583
1584 /*
1585  * Dry run to compute correct state of numberio for verification.
1586  */
1587 static uint64_t do_dry_run(struct thread_data *td)
1588 {
1589         td_set_runstate(td, TD_RUNNING);
1590
1591         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1592                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1593                 struct io_u *io_u;
1594                 int ret;
1595
1596                 if (td->terminate || td->done)
1597                         break;
1598
1599                 io_u = get_io_u(td);
1600                 if (IS_ERR_OR_NULL(io_u))
1601                         break;
1602
1603                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1604                 io_u->error = 0;
1605                 io_u->resid = 0;
1606                 if (ddir_rw(acct_ddir(io_u)))
1607                         td->io_issues[acct_ddir(io_u)]++;
1608                 if (ddir_rw(io_u->ddir)) {
1609                         io_u_mark_depth(td, 1);
1610                         td->ts.total_io_u[io_u->ddir]++;
1611                 }
1612
1613                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1614                     td->o.do_verify &&
1615                     td->o.verify != VERIFY_NONE &&
1616                     !td->o.experimental_verify)
1617                         log_io_piece(td, io_u);
1618
1619                 ret = io_u_sync_complete(td, io_u);
1620                 (void) ret;
1621         }
1622
1623         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1624 }
1625
1626 struct fork_data {
1627         struct thread_data *td;
1628         struct sk_out *sk_out;
1629 };
1630
1631 /*
1632  * Entry point for the thread based jobs. The process based jobs end up
1633  * here as well, after a little setup.
1634  */
1635 static void *thread_main(void *data)
1636 {
1637         struct fork_data *fd = data;
1638         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1639         struct thread_data *td = fd->td;
1640         struct thread_options *o = &td->o;
1641         struct sk_out *sk_out = fd->sk_out;
1642         uint64_t bytes_done[DDIR_RWDIR_CNT];
1643         int deadlock_loop_cnt;
1644         bool clear_state;
1645         int ret;
1646
1647         sk_out_assign(sk_out);
1648         free(fd);
1649
1650         if (!o->use_thread) {
1651                 setsid();
1652                 td->pid = getpid();
1653         } else
1654                 td->pid = gettid();
1655
1656         fio_local_clock_init();
1657
1658         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1659
1660         if (is_backend)
1661                 fio_server_send_start(td);
1662
1663         INIT_FLIST_HEAD(&td->io_log_list);
1664         INIT_FLIST_HEAD(&td->io_hist_list);
1665         INIT_FLIST_HEAD(&td->verify_list);
1666         INIT_FLIST_HEAD(&td->trim_list);
1667         td->io_hist_tree = RB_ROOT;
1668
1669         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1670         if (ret) {
1671                 td_verror(td, ret, "mutex_cond_init_pshared");
1672                 goto err;
1673         }
1674         ret = cond_init_pshared(&td->verify_cond);
1675         if (ret) {
1676                 td_verror(td, ret, "mutex_cond_pshared");
1677                 goto err;
1678         }
1679
1680         td_set_runstate(td, TD_INITIALIZED);
1681         dprint(FD_MUTEX, "up startup_sem\n");
1682         fio_sem_up(startup_sem);
1683         dprint(FD_MUTEX, "wait on td->sem\n");
1684         fio_sem_down(td->sem);
1685         dprint(FD_MUTEX, "done waiting on td->sem\n");
1686
1687         /*
1688          * A new gid requires privilege, so we need to do this before setting
1689          * the uid.
1690          */
1691         if (o->gid != -1U && setgid(o->gid)) {
1692                 td_verror(td, errno, "setgid");
1693                 goto err;
1694         }
1695         if (o->uid != -1U && setuid(o->uid)) {
1696                 td_verror(td, errno, "setuid");
1697                 goto err;
1698         }
1699
1700         td_zone_gen_index(td);
1701
1702         /*
1703          * Do this early, we don't want the compress threads to be limited
1704          * to the same CPUs as the IO workers. So do this before we set
1705          * any potential CPU affinity
1706          */
1707         if (iolog_compress_init(td, sk_out))
1708                 goto err;
1709
1710         /*
1711          * If we have a gettimeofday() thread, make sure we exclude that
1712          * thread from this job
1713          */
1714         if (o->gtod_cpu)
1715                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1716
1717         /*
1718          * Set affinity first, in case it has an impact on the memory
1719          * allocations.
1720          */
1721         if (fio_option_is_set(o, cpumask)) {
1722                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1723                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1724                         if (!ret) {
1725                                 log_err("fio: no CPUs set\n");
1726                                 log_err("fio: Try increasing number of available CPUs\n");
1727                                 td_verror(td, EINVAL, "cpus_split");
1728                                 goto err;
1729                         }
1730                 }
1731                 ret = fio_setaffinity(td->pid, o->cpumask);
1732                 if (ret == -1) {
1733                         td_verror(td, errno, "cpu_set_affinity");
1734                         goto err;
1735                 }
1736         }
1737
1738 #ifdef CONFIG_LIBNUMA
1739         /* numa node setup */
1740         if (fio_option_is_set(o, numa_cpunodes) ||
1741             fio_option_is_set(o, numa_memnodes)) {
1742                 struct bitmask *mask;
1743
1744                 if (numa_available() < 0) {
1745                         td_verror(td, errno, "Does not support NUMA API\n");
1746                         goto err;
1747                 }
1748
1749                 if (fio_option_is_set(o, numa_cpunodes)) {
1750                         mask = numa_parse_nodestring(o->numa_cpunodes);
1751                         ret = numa_run_on_node_mask(mask);
1752                         numa_free_nodemask(mask);
1753                         if (ret == -1) {
1754                                 td_verror(td, errno, \
1755                                         "numa_run_on_node_mask failed\n");
1756                                 goto err;
1757                         }
1758                 }
1759
1760                 if (fio_option_is_set(o, numa_memnodes)) {
1761                         mask = NULL;
1762                         if (o->numa_memnodes)
1763                                 mask = numa_parse_nodestring(o->numa_memnodes);
1764
1765                         switch (o->numa_mem_mode) {
1766                         case MPOL_INTERLEAVE:
1767                                 numa_set_interleave_mask(mask);
1768                                 break;
1769                         case MPOL_BIND:
1770                                 numa_set_membind(mask);
1771                                 break;
1772                         case MPOL_LOCAL:
1773                                 numa_set_localalloc();
1774                                 break;
1775                         case MPOL_PREFERRED:
1776                                 numa_set_preferred(o->numa_mem_prefer_node);
1777                                 break;
1778                         case MPOL_DEFAULT:
1779                         default:
1780                                 break;
1781                         }
1782
1783                         if (mask)
1784                                 numa_free_nodemask(mask);
1785
1786                 }
1787         }
1788 #endif
1789
1790         if (fio_pin_memory(td))
1791                 goto err;
1792
1793         /*
1794          * May alter parameters that init_io_u() will use, so we need to
1795          * do this first.
1796          */
1797         if (!init_iolog(td))
1798                 goto err;
1799
1800         /* ioprio_set() has to be done before td_io_init() */
1801         if (fio_option_is_set(o, ioprio) ||
1802             fio_option_is_set(o, ioprio_class) ||
1803             fio_option_is_set(o, ioprio_hint)) {
1804                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class,
1805                                  o->ioprio, o->ioprio_hint);
1806                 if (ret == -1) {
1807                         td_verror(td, errno, "ioprio_set");
1808                         goto err;
1809                 }
1810                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio,
1811                                           o->ioprio_hint);
1812                 td->ts.ioprio = td->ioprio;
1813         }
1814
1815         if (td_io_init(td))
1816                 goto err;
1817
1818         if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
1819                 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1820                          "are selected, queue depth will be capped at 1\n");
1821         }
1822
1823         if (init_io_u(td))
1824                 goto err;
1825
1826         if (td->io_ops->post_init && td->io_ops->post_init(td))
1827                 goto err;
1828
1829         if (o->verify_async && verify_async_init(td))
1830                 goto err;
1831
1832         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1833                 goto err;
1834
1835         errno = 0;
1836         if (nice(o->nice) == -1 && errno != 0) {
1837                 td_verror(td, errno, "nice");
1838                 goto err;
1839         }
1840
1841         if (o->ioscheduler && switch_ioscheduler(td))
1842                 goto err;
1843
1844         if (!o->create_serialize && setup_files(td))
1845                 goto err;
1846
1847         if (!init_random_map(td))
1848                 goto err;
1849
1850         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1851                 goto err;
1852
1853         if (o->pre_read && !pre_read_files(td))
1854                 goto err;
1855
1856         fio_verify_init(td);
1857
1858         if (rate_submit_init(td, sk_out))
1859                 goto err;
1860
1861         set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1862         fio_getrusage(&td->ru_start);
1863         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1864         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1865         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1866
1867         init_thinktime(td);
1868
1869         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1870                         o->ratemin[DDIR_TRIM]) {
1871                 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1872                                         sizeof(td->bw_sample_time));
1873                 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1874                                         sizeof(td->bw_sample_time));
1875                 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1876                                         sizeof(td->bw_sample_time));
1877         }
1878
1879         memset(bytes_done, 0, sizeof(bytes_done));
1880         clear_state = false;
1881
1882         while (keep_running(td)) {
1883                 uint64_t verify_bytes;
1884
1885                 fio_gettime(&td->start, NULL);
1886                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1887
1888                 if (clear_state) {
1889                         clear_io_state(td, 0);
1890
1891                         if (o->unlink_each_loop && unlink_all_files(td))
1892                                 break;
1893                 }
1894
1895                 prune_io_piece_log(td);
1896
1897                 if (td->o.verify_only && td_write(td))
1898                         verify_bytes = do_dry_run(td);
1899                 else {
1900                         if (!td->o.rand_repeatable)
1901                                 /* save verify rand state to replay hdr seeds later at verify */
1902                                 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
1903                         do_io(td, bytes_done);
1904                         if (!td->o.rand_repeatable)
1905                                 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
1906                         if (!ddir_rw_sum(bytes_done)) {
1907                                 fio_mark_td_terminate(td);
1908                                 verify_bytes = 0;
1909                         } else {
1910                                 verify_bytes = bytes_done[DDIR_WRITE] +
1911                                                 bytes_done[DDIR_TRIM];
1912                         }
1913                 }
1914
1915                 /*
1916                  * If we took too long to shut down, the main thread could
1917                  * already consider us reaped/exited. If that happens, break
1918                  * out and clean up.
1919                  */
1920                 if (td->runstate >= TD_EXITED)
1921                         break;
1922
1923                 clear_state = true;
1924
1925                 /*
1926                  * Make sure we've successfully updated the rusage stats
1927                  * before waiting on the stat mutex. Otherwise we could have
1928                  * the stat thread holding stat mutex and waiting for
1929                  * the rusage_sem, which would never get upped because
1930                  * this thread is waiting for the stat mutex.
1931                  */
1932                 deadlock_loop_cnt = 0;
1933                 do {
1934                         check_update_rusage(td);
1935                         if (!fio_sem_down_trylock(stat_sem))
1936                                 break;
1937                         usleep(1000);
1938                         if (deadlock_loop_cnt++ > 5000) {
1939                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1940                                 td->error = EDEADLK;
1941                                 goto err;
1942                         }
1943                 } while (1);
1944
1945                 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1946                         ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
1947                         update_runtime(td, elapsed_us, DDIR_READ);
1948                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1949                         update_runtime(td, elapsed_us, DDIR_WRITE);
1950                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1951                         update_runtime(td, elapsed_us, DDIR_TRIM);
1952                 fio_gettime(&td->start, NULL);
1953                 fio_sem_up(stat_sem);
1954
1955                 if (td->error || td->terminate)
1956                         break;
1957
1958                 if (!o->do_verify ||
1959                     o->verify == VERIFY_NONE ||
1960                     td_ioengine_flagged(td, FIO_UNIDIR))
1961                         continue;
1962
1963                 clear_io_state(td, 0);
1964
1965                 fio_gettime(&td->start, NULL);
1966
1967                 do_verify(td, verify_bytes);
1968
1969                 /*
1970                  * See comment further up for why this is done here.
1971                  */
1972                 check_update_rusage(td);
1973
1974                 fio_sem_down(stat_sem);
1975                 update_runtime(td, elapsed_us, DDIR_READ);
1976                 fio_gettime(&td->start, NULL);
1977                 fio_sem_up(stat_sem);
1978
1979                 if (td->error || td->terminate)
1980                         break;
1981         }
1982
1983         /*
1984          * Acquire this lock if we were doing overlap checking in
1985          * offload mode so that we don't clean up this job while
1986          * another thread is checking its io_u's for overlap
1987          */
1988         if (td_offload_overlap(td)) {
1989                 int res;
1990
1991                 res = pthread_mutex_lock(&overlap_check);
1992                 if (res) {
1993                         td->error = errno;
1994                         goto err;
1995                 }
1996         }
1997         td_set_runstate(td, TD_FINISHING);
1998         if (td_offload_overlap(td)) {
1999                 int res;
2000
2001                 res = pthread_mutex_unlock(&overlap_check);
2002                 if (res) {
2003                         td->error = errno;
2004                         goto err;
2005                 }
2006         }
2007
2008         update_rusage_stat(td);
2009         td->ts.total_run_time = mtime_since_now(&td->epoch);
2010         for_each_rw_ddir(ddir) {
2011                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2012         }
2013
2014         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
2015             (td->o.verify != VERIFY_NONE && td_write(td)))
2016                 verify_save_state(td->thread_number);
2017
2018         fio_unpin_memory(td);
2019
2020         td_writeout_logs(td, true);
2021
2022         iolog_compress_exit(td);
2023         rate_submit_exit(td);
2024
2025         if (o->exec_postrun)
2026                 exec_string(o, o->exec_postrun, "postrun");
2027
2028         if (exitall_on_terminate || (o->exitall_error && td->error))
2029                 fio_terminate_threads(td->groupid, td->o.exit_what);
2030
2031 err:
2032         if (td->error)
2033                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
2034                                                         td->verror);
2035
2036         if (o->verify_async)
2037                 verify_async_exit(td);
2038
2039         close_and_free_files(td);
2040         cleanup_io_u(td);
2041         close_ioengine(td);
2042         cgroup_shutdown(td, cgroup_mnt);
2043         verify_free_state(td);
2044         td_zone_free_index(td);
2045
2046         if (fio_option_is_set(o, cpumask)) {
2047                 ret = fio_cpuset_exit(&o->cpumask);
2048                 if (ret)
2049                         td_verror(td, ret, "fio_cpuset_exit");
2050         }
2051
2052         /*
2053          * do this very late, it will log file closing as well
2054          */
2055         if (o->write_iolog_file)
2056                 write_iolog_close(td);
2057         if (td->io_log_rfile)
2058                 fclose(td->io_log_rfile);
2059
2060         td_set_runstate(td, TD_EXITED);
2061
2062         /*
2063          * Do this last after setting our runstate to exited, so we
2064          * know that the stat thread is signaled.
2065          */
2066         check_update_rusage(td);
2067
2068         sk_out_drop();
2069         return (void *) (uintptr_t) td->error;
2070 }
2071
2072 /*
2073  * Run over the job map and reap the threads that have exited, if any.
2074  */
2075 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2076                          uint64_t *m_rate)
2077 {
2078         unsigned int cputhreads, realthreads, pending;
2079         int status, ret;
2080
2081         /*
2082          * reap exited threads (TD_EXITED -> TD_REAPED)
2083          */
2084         realthreads = pending = cputhreads = 0;
2085         for_each_td(td) {
2086                 int flags = 0;
2087
2088                 if (!strcmp(td->o.ioengine, "cpuio"))
2089                         cputhreads++;
2090                 else
2091                         realthreads++;
2092
2093                 if (!td->pid) {
2094                         pending++;
2095                         continue;
2096                 }
2097                 if (td->runstate == TD_REAPED)
2098                         continue;
2099                 if (td->o.use_thread) {
2100                         if (td->runstate == TD_EXITED) {
2101                                 td_set_runstate(td, TD_REAPED);
2102                                 goto reaped;
2103                         }
2104                         continue;
2105                 }
2106
2107                 flags = WNOHANG;
2108                 if (td->runstate == TD_EXITED)
2109                         flags = 0;
2110
2111                 /*
2112                  * check if someone quit or got killed in an unusual way
2113                  */
2114                 ret = waitpid(td->pid, &status, flags);
2115                 if (ret < 0) {
2116                         if (errno == ECHILD) {
2117                                 log_err("fio: pid=%d disappeared %d\n",
2118                                                 (int) td->pid, td->runstate);
2119                                 td->sig = ECHILD;
2120                                 td_set_runstate(td, TD_REAPED);
2121                                 goto reaped;
2122                         }
2123                         perror("waitpid");
2124                 } else if (ret == td->pid) {
2125                         if (WIFSIGNALED(status)) {
2126                                 int sig = WTERMSIG(status);
2127
2128                                 if (sig != SIGTERM && sig != SIGUSR2)
2129                                         log_err("fio: pid=%d, got signal=%d\n",
2130                                                         (int) td->pid, sig);
2131                                 td->sig = sig;
2132                                 td_set_runstate(td, TD_REAPED);
2133                                 goto reaped;
2134                         }
2135                         if (WIFEXITED(status)) {
2136                                 if (WEXITSTATUS(status) && !td->error)
2137                                         td->error = WEXITSTATUS(status);
2138
2139                                 td_set_runstate(td, TD_REAPED);
2140                                 goto reaped;
2141                         }
2142                 }
2143
2144                 /*
2145                  * If the job is stuck, do a forceful timeout of it and
2146                  * move on.
2147                  */
2148                 if (td->terminate &&
2149                     td->runstate < TD_FSYNCING &&
2150                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2151                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2152                                 "%lu seconds, it appears to be stuck. Doing "
2153                                 "forceful exit of this job.\n",
2154                                 td->o.name, td->runstate,
2155                                 (unsigned long) time_since_now(&td->terminate_time));
2156                         td_set_runstate(td, TD_REAPED);
2157                         goto reaped;
2158                 }
2159
2160                 /*
2161                  * thread is not dead, continue
2162                  */
2163                 pending++;
2164                 continue;
2165 reaped:
2166                 (*nr_running)--;
2167                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2168                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2169                 if (!td->pid)
2170                         pending--;
2171
2172                 if (td->error)
2173                         exit_value++;
2174
2175                 done_secs += mtime_since_now(&td->epoch) / 1000;
2176                 profile_td_exit(td);
2177                 flow_exit_job(td);
2178         } end_for_each();
2179
2180         if (*nr_running == cputhreads && !pending && realthreads)
2181                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2182 }
2183
2184 static bool __check_trigger_file(void)
2185 {
2186         struct stat sb;
2187
2188         if (!trigger_file)
2189                 return false;
2190
2191         if (stat(trigger_file, &sb))
2192                 return false;
2193
2194         if (unlink(trigger_file) < 0)
2195                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2196                                                         strerror(errno));
2197
2198         return true;
2199 }
2200
2201 static bool trigger_timedout(void)
2202 {
2203         if (trigger_timeout)
2204                 if (time_since_genesis() >= trigger_timeout) {
2205                         trigger_timeout = 0;
2206                         return true;
2207                 }
2208
2209         return false;
2210 }
2211
2212 void exec_trigger(const char *cmd)
2213 {
2214         int ret;
2215
2216         if (!cmd || cmd[0] == '\0')
2217                 return;
2218
2219         ret = system(cmd);
2220         if (ret == -1)
2221                 log_err("fio: failed executing %s trigger\n", cmd);
2222 }
2223
2224 void check_trigger_file(void)
2225 {
2226         if (__check_trigger_file() || trigger_timedout()) {
2227                 if (nr_clients)
2228                         fio_clients_send_trigger(trigger_remote_cmd);
2229                 else {
2230                         verify_save_state(IO_LIST_ALL);
2231                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2232                         exec_trigger(trigger_cmd);
2233                 }
2234         }
2235 }
2236
2237 static int fio_verify_load_state(struct thread_data *td)
2238 {
2239         int ret;
2240
2241         if (!td->o.verify_state)
2242                 return 0;
2243
2244         if (is_backend) {
2245                 void *data;
2246
2247                 ret = fio_server_get_verify_state(td->o.name,
2248                                         td->thread_number - 1, &data);
2249                 if (!ret)
2250                         verify_assign_state(td, data);
2251         } else {
2252                 char prefix[PATH_MAX];
2253
2254                 if (aux_path)
2255                         sprintf(prefix, "%s%clocal", aux_path,
2256                                         FIO_OS_PATH_SEPARATOR);
2257                 else
2258                         strcpy(prefix, "local");
2259                 ret = verify_load_state(td, prefix);
2260         }
2261
2262         return ret;
2263 }
2264
2265 static void do_usleep(unsigned int usecs)
2266 {
2267         check_for_running_stats();
2268         check_trigger_file();
2269         usleep(usecs);
2270 }
2271
2272 static bool check_mount_writes(struct thread_data *td)
2273 {
2274         struct fio_file *f;
2275         unsigned int i;
2276
2277         if (!td_write(td) || td->o.allow_mounted_write)
2278                 return false;
2279
2280         /*
2281          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2282          * are mkfs'd and mounted.
2283          */
2284         for_each_file(td, f, i) {
2285 #ifdef FIO_HAVE_CHARDEV_SIZE
2286                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2287 #else
2288                 if (f->filetype != FIO_TYPE_BLOCK)
2289 #endif
2290                         continue;
2291                 if (device_is_mounted(f->file_name))
2292                         goto mounted;
2293         }
2294
2295         return false;
2296 mounted:
2297         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2298         return true;
2299 }
2300
2301 static bool waitee_running(struct thread_data *me)
2302 {
2303         const char *waitee = me->o.wait_for;
2304         const char *self = me->o.name;
2305
2306         if (!waitee)
2307                 return false;
2308
2309         for_each_td(td) {
2310                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2311                         continue;
2312
2313                 if (td->runstate < TD_EXITED) {
2314                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2315                                         self, td->o.name,
2316                                         runstate_to_name(td->runstate));
2317                         return true;
2318                 }
2319         } end_for_each();
2320
2321         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2322         return false;
2323 }
2324
2325 /*
2326  * Main function for kicking off and reaping jobs, as needed.
2327  */
2328 static void run_threads(struct sk_out *sk_out)
2329 {
2330         struct thread_data *td;
2331         unsigned int i, todo, nr_running, nr_started;
2332         uint64_t m_rate, t_rate;
2333         uint64_t spent;
2334
2335         if (fio_gtod_offload && fio_start_gtod_thread())
2336                 return;
2337
2338         fio_idle_prof_init();
2339
2340         set_sig_handlers();
2341
2342         nr_thread = nr_process = 0;
2343         for_each_td(td) {
2344                 if (check_mount_writes(td))
2345                         return;
2346                 if (td->o.use_thread)
2347                         nr_thread++;
2348                 else
2349                         nr_process++;
2350         } end_for_each();
2351
2352         if (output_format & FIO_OUTPUT_NORMAL) {
2353                 struct buf_output out;
2354
2355                 buf_output_init(&out);
2356                 __log_buf(&out, "Starting ");
2357                 if (nr_thread)
2358                         __log_buf(&out, "%d thread%s", nr_thread,
2359                                                 nr_thread > 1 ? "s" : "");
2360                 if (nr_process) {
2361                         if (nr_thread)
2362                                 __log_buf(&out, " and ");
2363                         __log_buf(&out, "%d process%s", nr_process,
2364                                                 nr_process > 1 ? "es" : "");
2365                 }
2366                 __log_buf(&out, "\n");
2367                 log_info_buf(out.buf, out.buflen);
2368                 buf_output_free(&out);
2369         }
2370
2371         todo = thread_number;
2372         nr_running = 0;
2373         nr_started = 0;
2374         m_rate = t_rate = 0;
2375
2376         for_each_td(td) {
2377                 print_status_init(td->thread_number - 1);
2378
2379                 if (!td->o.create_serialize)
2380                         continue;
2381
2382                 if (fio_verify_load_state(td))
2383                         goto reap;
2384
2385                 /*
2386                  * do file setup here so it happens sequentially,
2387                  * we don't want X number of threads getting their
2388                  * client data interspersed on disk
2389                  */
2390                 if (setup_files(td)) {
2391 reap:
2392                         exit_value++;
2393                         if (td->error)
2394                                 log_err("fio: pid=%d, err=%d/%s\n",
2395                                         (int) td->pid, td->error, td->verror);
2396                         td_set_runstate(td, TD_REAPED);
2397                         todo--;
2398                 } else {
2399                         struct fio_file *f;
2400                         unsigned int j;
2401
2402                         /*
2403                          * for sharing to work, each job must always open
2404                          * its own files. so close them, if we opened them
2405                          * for creation
2406                          */
2407                         for_each_file(td, f, j) {
2408                                 if (fio_file_open(f))
2409                                         td_io_close_file(td, f);
2410                         }
2411                 }
2412         } end_for_each();
2413
2414         /* start idle threads before io threads start to run */
2415         fio_idle_prof_start();
2416
2417         set_genesis_time();
2418
2419         while (todo) {
2420                 struct thread_data *map[REAL_MAX_JOBS];
2421                 struct timespec this_start;
2422                 int this_jobs = 0, left;
2423                 struct fork_data *fd;
2424
2425                 /*
2426                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2427                  */
2428                 for_each_td(td) {
2429                         if (td->runstate != TD_NOT_CREATED)
2430                                 continue;
2431
2432                         /*
2433                          * never got a chance to start, killed by other
2434                          * thread for some reason
2435                          */
2436                         if (td->terminate) {
2437                                 todo--;
2438                                 continue;
2439                         }
2440
2441                         if (td->o.start_delay) {
2442                                 spent = utime_since_genesis();
2443
2444                                 if (td->o.start_delay > spent)
2445                                         continue;
2446                         }
2447
2448                         if (td->o.stonewall && (nr_started || nr_running)) {
2449                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2450                                                         td->o.name);
2451                                 break;
2452                         }
2453
2454                         if (waitee_running(td)) {
2455                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2456                                                 td->o.name, td->o.wait_for);
2457                                 continue;
2458                         }
2459
2460                         init_disk_util(td);
2461
2462                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2463                         td->update_rusage = 0;
2464
2465                         /*
2466                          * Set state to created. Thread will transition
2467                          * to TD_INITIALIZED when it's done setting up.
2468                          */
2469                         td_set_runstate(td, TD_CREATED);
2470                         map[this_jobs++] = td;
2471                         nr_started++;
2472
2473                         fd = calloc(1, sizeof(*fd));
2474                         fd->td = td;
2475                         fd->sk_out = sk_out;
2476
2477                         if (td->o.use_thread) {
2478                                 int ret;
2479
2480                                 dprint(FD_PROCESS, "will pthread_create\n");
2481                                 ret = pthread_create(&td->thread, NULL,
2482                                                         thread_main, fd);
2483                                 if (ret) {
2484                                         log_err("pthread_create: %s\n",
2485                                                         strerror(ret));
2486                                         free(fd);
2487                                         nr_started--;
2488                                         break;
2489                                 }
2490                                 fd = NULL;
2491                                 ret = pthread_detach(td->thread);
2492                                 if (ret)
2493                                         log_err("pthread_detach: %s",
2494                                                         strerror(ret));
2495                         } else {
2496                                 pid_t pid;
2497                                 dprint(FD_PROCESS, "will fork\n");
2498                                 pid = fork();
2499                                 if (!pid) {
2500                                         int ret;
2501
2502                                         ret = (int)(uintptr_t)thread_main(fd);
2503                                         _exit(ret);
2504                                 } else if (__td_index == fio_debug_jobno)
2505                                         *fio_debug_jobp = pid;
2506                                 free(fd);
2507                                 fd = NULL;
2508                         }
2509                         dprint(FD_MUTEX, "wait on startup_sem\n");
2510                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2511                                 log_err("fio: job startup hung? exiting.\n");
2512                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2513                                 fio_abort = true;
2514                                 nr_started--;
2515                                 free(fd);
2516                                 break;
2517                         }
2518                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2519                 } end_for_each();
2520
2521                 /*
2522                  * Wait for the started threads to transition to
2523                  * TD_INITIALIZED.
2524                  */
2525                 fio_gettime(&this_start, NULL);
2526                 left = this_jobs;
2527                 while (left && !fio_abort) {
2528                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2529                                 break;
2530
2531                         do_usleep(100000);
2532
2533                         for (i = 0; i < this_jobs; i++) {
2534                                 td = map[i];
2535                                 if (!td)
2536                                         continue;
2537                                 if (td->runstate == TD_INITIALIZED) {
2538                                         map[i] = NULL;
2539                                         left--;
2540                                 } else if (td->runstate >= TD_EXITED) {
2541                                         map[i] = NULL;
2542                                         left--;
2543                                         todo--;
2544                                         nr_running++; /* work-around... */
2545                                 }
2546                         }
2547                 }
2548
2549                 if (left) {
2550                         log_err("fio: %d job%s failed to start\n", left,
2551                                         left > 1 ? "s" : "");
2552                         for (i = 0; i < this_jobs; i++) {
2553                                 td = map[i];
2554                                 if (!td)
2555                                         continue;
2556                                 kill(td->pid, SIGTERM);
2557                         }
2558                         break;
2559                 }
2560
2561                 /*
2562                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2563                  */
2564                 for_each_td(td) {
2565                         if (td->runstate != TD_INITIALIZED)
2566                                 continue;
2567
2568                         if (in_ramp_time(td))
2569                                 td_set_runstate(td, TD_RAMP);
2570                         else
2571                                 td_set_runstate(td, TD_RUNNING);
2572                         nr_running++;
2573                         nr_started--;
2574                         m_rate += ddir_rw_sum(td->o.ratemin);
2575                         t_rate += ddir_rw_sum(td->o.rate);
2576                         todo--;
2577                         fio_sem_up(td->sem);
2578                 } end_for_each();
2579
2580                 reap_threads(&nr_running, &t_rate, &m_rate);
2581
2582                 if (todo)
2583                         do_usleep(100000);
2584         }
2585
2586         while (nr_running) {
2587                 reap_threads(&nr_running, &t_rate, &m_rate);
2588                 do_usleep(10000);
2589         }
2590
2591         fio_idle_prof_stop();
2592
2593         update_io_ticks();
2594 }
2595
2596 static void free_disk_util(void)
2597 {
2598         disk_util_prune_entries();
2599         helper_thread_destroy();
2600 }
2601
2602 int fio_backend(struct sk_out *sk_out)
2603 {
2604         int i;
2605         if (exec_profile) {
2606                 if (load_profile(exec_profile))
2607                         return 1;
2608                 free(exec_profile);
2609                 exec_profile = NULL;
2610         }
2611         if (!thread_number)
2612                 return 0;
2613
2614         if (write_bw_log) {
2615                 struct log_params p = {
2616                         .log_type = IO_LOG_TYPE_BW,
2617                 };
2618
2619                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2620                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2621                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2622         }
2623
2624         if (init_global_dedupe_working_set_seeds()) {
2625                 log_err("fio: failed to initialize global dedupe working set\n");
2626                 return 1;
2627         }
2628
2629         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2630         if (!sk_out)
2631                 is_local_backend = true;
2632         if (startup_sem == NULL)
2633                 return 1;
2634
2635         set_genesis_time();
2636         stat_init();
2637         if (helper_thread_create(startup_sem, sk_out))
2638                 log_err("fio: failed to create helper thread\n");
2639
2640         cgroup_list = smalloc(sizeof(*cgroup_list));
2641         if (cgroup_list)
2642                 INIT_FLIST_HEAD(cgroup_list);
2643
2644         run_threads(sk_out);
2645
2646         helper_thread_exit();
2647
2648         if (!fio_abort) {
2649                 __show_run_stats();
2650                 if (write_bw_log) {
2651                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2652                                 struct io_log *log = agg_io_log[i];
2653
2654                                 flush_log(log, false);
2655                                 free_log(log);
2656                         }
2657                 }
2658         }
2659
2660         for_each_td(td) {
2661                 struct thread_stat *ts = &td->ts;
2662
2663                 free_clat_prio_stats(ts);
2664                 steadystate_free(td);
2665                 fio_options_free(td);
2666                 fio_dump_options_free(td);
2667                 if (td->rusage_sem) {
2668                         fio_sem_remove(td->rusage_sem);
2669                         td->rusage_sem = NULL;
2670                 }
2671                 fio_sem_remove(td->sem);
2672                 td->sem = NULL;
2673         } end_for_each();
2674
2675         free_disk_util();
2676         if (cgroup_list) {
2677                 cgroup_kill(cgroup_list);
2678                 sfree(cgroup_list);
2679         }
2680
2681         fio_sem_remove(startup_sem);
2682         stat_exit();
2683         return exit_value;
2684 }