backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32 #include <pthread.h>
  33
  34 #include "fio.h"
  35 #include "smalloc.h"
  36 #include "verify.h"
  37 #include "diskutil.h"
  38 #include "cgroup.h"
  39 #include "profile.h"
  40 #include "lib/rand.h"
  41 #include "lib/memalign.h"
  42 #include "server.h"
  43 #include "lib/getrusage.h"
  44 #include "idletime.h"
  45 #include "err.h"
  46 #include "workqueue.h"
  47 #include "lib/mountcheck.h"
  48 #include "rate-submit.h"
  49 #include "helper_thread.h"
  50 #include "pshared.h"
  51 #include "zone-dist.h"
  52
  53 static struct fio_sem *startup_sem;
  54 static struct flist_head *cgroup_list;
  55 static struct cgroup_mnt *cgroup_mnt;
  56 static int exit_value;
  57 static volatile bool fio_abort;
  58 static unsigned int nr_process = 0;
  59 static unsigned int nr_thread = 0;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int nr_segments = 0;
  66 unsigned int cur_segment = 0;
  67 unsigned int stat_number = 0;
  68 int temp_stall_ts;
  69 unsigned long done_secs = 0;
  70 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
  71 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
  72 #else
  73 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  74 #endif
  75
  76 #define JOB_START_TIMEOUT       (5 * 1000)
  77
  78 static void sig_int(int sig)
  79 {
  80         if (nr_segments) {
  81                 if (is_backend)
  82                         fio_server_got_signal(sig);
  83                 else {
  84                         log_info("\nfio: terminating on signal %d\n", sig);
  85                         log_info_flush();
  86                         exit_value = 128;
  87                 }
  88
  89                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  90         }
  91 }
  92
  93 #ifdef WIN32
  94 static void sig_break(int sig)
  95 {
  96         struct thread_data *td;
  97         int i;
  98
  99         sig_int(sig);
 100
 101         /**
 102          * Windows terminates all job processes on SIGBREAK after the handler
 103          * returns, so give them time to wrap-up and give stats
 104          */
 105         for_each_td(td, i) {
 106                 while (td->runstate < TD_EXITED)
 107                         sleep(1);
 108         }
 109 }
 110 #endif
 111
 112 void sig_show_status(int sig)
 113 {
 114         show_running_run_stats();
 115 }
 116
 117 static void set_sig_handlers(void)
 118 {
 119         struct sigaction act;
 120
 121         memset(&act, 0, sizeof(act));
 122         act.sa_handler = sig_int;
 123         act.sa_flags = SA_RESTART;
 124         sigaction(SIGINT, &act, NULL);
 125
 126         memset(&act, 0, sizeof(act));
 127         act.sa_handler = sig_int;
 128         act.sa_flags = SA_RESTART;
 129         sigaction(SIGTERM, &act, NULL);
 130
 131 /* Windows uses SIGBREAK as a quit signal from other applications */
 132 #ifdef WIN32
 133         memset(&act, 0, sizeof(act));
 134         act.sa_handler = sig_break;
 135         act.sa_flags = SA_RESTART;
 136         sigaction(SIGBREAK, &act, NULL);
 137 #endif
 138
 139         memset(&act, 0, sizeof(act));
 140         act.sa_handler = sig_show_status;
 141         act.sa_flags = SA_RESTART;
 142         sigaction(SIGUSR1, &act, NULL);
 143
 144         if (is_backend) {
 145                 memset(&act, 0, sizeof(act));
 146                 act.sa_handler = sig_int;
 147                 act.sa_flags = SA_RESTART;
 148                 sigaction(SIGPIPE, &act, NULL);
 149         }
 150 }
 151
 152 /*
 153  * Check if we are above the minimum rate given.
 154  */
 155 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 156                              enum fio_ddir ddir)
 157 {
 158         unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
 159         unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
 160         unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
 161         unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
 162
 163         assert(ddir_rw(ddir));
 164
 165         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 166                 return false;
 167
 168         /*
 169          * allow a 2 second settle period in the beginning
 170          */
 171         if (mtime_since(&td->start, now) < 2000)
 172                 return false;
 173
 174         /*
 175          * if last_rate_check_blocks or last_rate_check_bytes is set,
 176          * we can compute a rate per ratecycle
 177          */
 178         if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
 179                 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
 180                 if (spent < td->o.ratecycle || spent==0)
 181                         return false;
 182
 183                 if (td->o.ratemin[ddir]) {
 184                         /*
 185                          * check bandwidth specified rate
 186                          */
 187                         unsigned long long current_rate_bytes =
 188                                 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
 189                         if (current_rate_bytes < option_rate_bytes_min) {
 190                                 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 191                                         td->o.name, option_rate_bytes_min, current_rate_bytes);
 192                                 return true;
 193                         }
 194                 } else {
 195                         /*
 196                          * checks iops specified rate
 197                          */
 198                         unsigned long long current_rate_iops =
 199                                 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
 200
 201                         if (current_rate_iops < option_rate_iops_min) {
 202                                 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 203                                         td->o.name, option_rate_iops_min, current_rate_iops);
 204                                 return true;
 205                         }
 206                 }
 207         }
 208
 209         td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
 210         td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
 211         memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
 212         return false;
 213 }
 214
 215 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 216 {
 217         bool ret = false;
 218
 219         for_each_rw_ddir(ddir) {
 220                 if (td->bytes_done[ddir])
 221                         ret |= __check_min_rate(td, now, ddir);
 222         }
 223
 224         return ret;
 225 }
 226
 227 /*
 228  * When job exits, we can cancel the in-flight IO if we are using async
 229  * io. Attempt to do so.
 230  */
 231 static void cleanup_pending_aio(struct thread_data *td)
 232 {
 233         int r;
 234
 235         /*
 236          * get immediately available events, if any
 237          */
 238         r = io_u_queued_complete(td, 0);
 239
 240         /*
 241          * now cancel remaining active events
 242          */
 243         if (td->io_ops->cancel) {
 244                 struct io_u *io_u;
 245                 int i;
 246
 247                 io_u_qiter(&td->io_u_all, io_u, i) {
 248                         if (io_u->flags & IO_U_F_FLIGHT) {
 249                                 r = td->io_ops->cancel(td, io_u);
 250                                 if (!r)
 251                                         put_io_u(td, io_u);
 252                         }
 253                 }
 254         }
 255
 256         if (td->cur_depth)
 257                 r = io_u_queued_complete(td, td->cur_depth);
 258 }
 259
 260 /*
 261  * Helper to handle the final sync of a file. Works just like the normal
 262  * io path, just does everything sync.
 263  */
 264 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 265 {
 266         struct io_u *io_u = __get_io_u(td);
 267         enum fio_q_status ret;
 268
 269         if (!io_u)
 270                 return true;
 271
 272         io_u->ddir = DDIR_SYNC;
 273         io_u->file = f;
 274         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 275
 276         if (td_io_prep(td, io_u)) {
 277                 put_io_u(td, io_u);
 278                 return true;
 279         }
 280
 281 requeue:
 282         ret = td_io_queue(td, io_u);
 283         switch (ret) {
 284         case FIO_Q_QUEUED:
 285                 td_io_commit(td);
 286                 if (io_u_queued_complete(td, 1) < 0)
 287                         return true;
 288                 break;
 289         case FIO_Q_COMPLETED:
 290                 if (io_u->error) {
 291                         td_verror(td, io_u->error, "td_io_queue");
 292                         return true;
 293                 }
 294
 295                 if (io_u_sync_complete(td, io_u) < 0)
 296                         return true;
 297                 break;
 298         case FIO_Q_BUSY:
 299                 td_io_commit(td);
 300                 goto requeue;
 301         }
 302
 303         return false;
 304 }
 305
 306 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 307 {
 308         int ret, ret2;
 309
 310         if (fio_file_open(f))
 311                 return fio_io_sync(td, f);
 312
 313         if (td_io_open_file(td, f))
 314                 return 1;
 315
 316         ret = fio_io_sync(td, f);
 317         ret2 = 0;
 318         if (fio_file_open(f))
 319                 ret2 = td_io_close_file(td, f);
 320         return (ret || ret2);
 321 }
 322
 323 static inline void __update_ts_cache(struct thread_data *td)
 324 {
 325         fio_gettime(&td->ts_cache, NULL);
 326 }
 327
 328 static inline void update_ts_cache(struct thread_data *td)
 329 {
 330         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 331                 __update_ts_cache(td);
 332 }
 333
 334 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 335 {
 336         if (in_ramp_time(td))
 337                 return false;
 338         if (!td->o.timeout)
 339                 return false;
 340         if (utime_since(&td->epoch, t) >= td->o.timeout)
 341                 return true;
 342
 343         return false;
 344 }
 345
 346 /*
 347  * We need to update the runtime consistently in ms, but keep a running
 348  * tally of the current elapsed time in microseconds for sub millisecond
 349  * updates.
 350  */
 351 static inline void update_runtime(struct thread_data *td,
 352                                   unsigned long long *elapsed_us,
 353                                   const enum fio_ddir ddir)
 354 {
 355         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 356                 return;
 357
 358         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 359         elapsed_us[ddir] += utime_since_now(&td->start);
 360         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 361 }
 362
 363 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 364                                 int *retptr)
 365 {
 366         int ret = *retptr;
 367
 368         if (ret < 0 || td->error) {
 369                 int err = td->error;
 370                 enum error_type_bit eb;
 371
 372                 if (ret < 0)
 373                         err = -ret;
 374
 375                 eb = td_error_type(ddir, err);
 376                 if (!(td->o.continue_on_error & (1 << eb)))
 377                         return true;
 378
 379                 if (td_non_fatal_error(td, eb, err)) {
 380                         /*
 381                          * Continue with the I/Os in case of
 382                          * a non fatal error.
 383                          */
 384                         update_error_count(td, err);
 385                         td_clear_error(td);
 386                         *retptr = 0;
 387                         return false;
 388                 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
 389                         /*
 390                          * We expect to hit this error if
 391                          * fill_device option is set.
 392                          */
 393                         td_clear_error(td);
 394                         fio_mark_td_terminate(td);
 395                         return true;
 396                 } else {
 397                         /*
 398                          * Stop the I/O in case of a fatal
 399                          * error.
 400                          */
 401                         update_error_count(td, err);
 402                         return true;
 403                 }
 404         }
 405
 406         return false;
 407 }
 408
 409 static void check_update_rusage(struct thread_data *td)
 410 {
 411         if (td->update_rusage) {
 412                 td->update_rusage = 0;
 413                 update_rusage_stat(td);
 414                 fio_sem_up(td->rusage_sem);
 415         }
 416 }
 417
 418 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 419 {
 420         const int full = queue_full(td);
 421         int min_evts = 0;
 422         int ret;
 423
 424         if (td->flags & TD_F_REGROW_LOGS)
 425                 return io_u_quiesce(td);
 426
 427         /*
 428          * if the queue is full, we MUST reap at least 1 event
 429          */
 430         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 431         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 432                 min_evts = 1;
 433
 434         if (time && should_check_rate(td))
 435                 fio_gettime(time, NULL);
 436
 437         do {
 438                 ret = io_u_queued_complete(td, min_evts);
 439                 if (ret < 0)
 440                         break;
 441         } while (full && (td->cur_depth > td->o.iodepth_low));
 442
 443         return ret;
 444 }
 445
 446 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 447                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 448                    struct timespec *comp_time)
 449 {
 450         switch (*ret) {
 451         case FIO_Q_COMPLETED:
 452                 if (io_u->error) {
 453                         *ret = -io_u->error;
 454                         clear_io_u(td, io_u);
 455                 } else if (io_u->resid) {
 456                         long long bytes = io_u->xfer_buflen - io_u->resid;
 457                         struct fio_file *f = io_u->file;
 458
 459                         if (bytes_issued)
 460                                 *bytes_issued += bytes;
 461
 462                         if (!from_verify)
 463                                 trim_io_piece(io_u);
 464
 465                         /*
 466                          * zero read, fail
 467                          */
 468                         if (!bytes) {
 469                                 if (!from_verify)
 470                                         unlog_io_piece(td, io_u);
 471                                 td_verror(td, EIO, "full resid");
 472                                 put_io_u(td, io_u);
 473                                 break;
 474                         }
 475
 476                         io_u->xfer_buflen = io_u->resid;
 477                         io_u->xfer_buf += bytes;
 478                         io_u->offset += bytes;
 479
 480                         if (ddir_rw(io_u->ddir))
 481                                 td->ts.short_io_u[io_u->ddir]++;
 482
 483                         if (io_u->offset == f->real_file_size)
 484                                 goto sync_done;
 485
 486                         requeue_io_u(td, &io_u);
 487                 } else {
 488 sync_done:
 489                         if (comp_time && should_check_rate(td))
 490                                 fio_gettime(comp_time, NULL);
 491
 492                         *ret = io_u_sync_complete(td, io_u);
 493                         if (*ret < 0)
 494                                 break;
 495                 }
 496
 497                 if (td->flags & TD_F_REGROW_LOGS)
 498                         regrow_logs(td);
 499
 500                 /*
 501                  * when doing I/O (not when verifying),
 502                  * check for any errors that are to be ignored
 503                  */
 504                 if (!from_verify)
 505                         break;
 506
 507                 return 0;
 508         case FIO_Q_QUEUED:
 509                 /*
 510                  * if the engine doesn't have a commit hook,
 511                  * the io_u is really queued. if it does have such
 512                  * a hook, it has to call io_u_queued() itself.
 513                  */
 514                 if (td->io_ops->commit == NULL)
 515                         io_u_queued(td, io_u);
 516                 if (bytes_issued)
 517                         *bytes_issued += io_u->xfer_buflen;
 518                 break;
 519         case FIO_Q_BUSY:
 520                 if (!from_verify)
 521                         unlog_io_piece(td, io_u);
 522                 requeue_io_u(td, &io_u);
 523                 td_io_commit(td);
 524                 break;
 525         default:
 526                 assert(*ret < 0);
 527                 td_verror(td, -(*ret), "td_io_queue");
 528                 break;
 529         }
 530
 531         if (break_on_this_error(td, ddir, ret))
 532                 return 1;
 533
 534         return 0;
 535 }
 536
 537 static inline bool io_in_polling(struct thread_data *td)
 538 {
 539         return !td->o.iodepth_batch_complete_min &&
 540                    !td->o.iodepth_batch_complete_max;
 541 }
 542 /*
 543  * Unlinks files from thread data fio_file structure
 544  */
 545 static int unlink_all_files(struct thread_data *td)
 546 {
 547         struct fio_file *f;
 548         unsigned int i;
 549         int ret = 0;
 550
 551         for_each_file(td, f, i) {
 552                 if (f->filetype != FIO_TYPE_FILE)
 553                         continue;
 554                 ret = td_io_unlink_file(td, f);
 555                 if (ret)
 556                         break;
 557         }
 558
 559         if (ret)
 560                 td_verror(td, ret, "unlink_all_files");
 561
 562         return ret;
 563 }
 564
 565 /*
 566  * Check if io_u will overlap an in-flight IO in the queue
 567  */
 568 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 569 {
 570         bool overlap;
 571         struct io_u *check_io_u;
 572         unsigned long long x1, x2, y1, y2;
 573         int i;
 574
 575         x1 = io_u->offset;
 576         x2 = io_u->offset + io_u->buflen;
 577         overlap = false;
 578         io_u_qiter(q, check_io_u, i) {
 579                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 580                         y1 = check_io_u->offset;
 581                         y2 = check_io_u->offset + check_io_u->buflen;
 582
 583                         if (x1 < y2 && y1 < x2) {
 584                                 overlap = true;
 585                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 586                                                 x1, io_u->buflen,
 587                                                 y1, check_io_u->buflen);
 588                                 break;
 589                         }
 590                 }
 591         }
 592
 593         return overlap;
 594 }
 595
 596 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 597 {
 598         /*
 599          * Check for overlap if the user asked us to, and we have
 600          * at least one IO in flight besides this one.
 601          */
 602         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 603             in_flight_overlap(&td->io_u_all, io_u))
 604                 return FIO_Q_BUSY;
 605
 606         return td_io_queue(td, io_u);
 607 }
 608
 609 /*
 610  * The main verify engine. Runs over the writes we previously submitted,
 611  * reads the blocks back in, and checks the crc/md5 of the data.
 612  */
 613 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 614 {
 615         struct fio_file *f;
 616         struct io_u *io_u;
 617         int ret, min_events;
 618         unsigned int i;
 619
 620         dprint(FD_VERIFY, "starting loop\n");
 621
 622         /*
 623          * sync io first and invalidate cache, to make sure we really
 624          * read from disk.
 625          */
 626         for_each_file(td, f, i) {
 627                 if (!fio_file_open(f))
 628                         continue;
 629                 if (fio_io_sync(td, f))
 630                         break;
 631                 if (file_invalidate_cache(td, f))
 632                         break;
 633         }
 634
 635         check_update_rusage(td);
 636
 637         if (td->error)
 638                 return;
 639
 640         td_set_runstate(td, TD_VERIFYING);
 641
 642         io_u = NULL;
 643         while (!td->terminate) {
 644                 enum fio_ddir ddir;
 645                 int full;
 646
 647                 update_ts_cache(td);
 648                 check_update_rusage(td);
 649
 650                 if (runtime_exceeded(td, &td->ts_cache)) {
 651                         __update_ts_cache(td);
 652                         if (runtime_exceeded(td, &td->ts_cache)) {
 653                                 fio_mark_td_terminate(td);
 654                                 break;
 655                         }
 656                 }
 657
 658                 if (flow_threshold_exceeded(td))
 659                         continue;
 660
 661                 if (!td->o.experimental_verify) {
 662                         io_u = __get_io_u(td);
 663                         if (!io_u)
 664                                 break;
 665
 666                         if (get_next_verify(td, io_u)) {
 667                                 put_io_u(td, io_u);
 668                                 break;
 669                         }
 670
 671                         if (td_io_prep(td, io_u)) {
 672                                 put_io_u(td, io_u);
 673                                 break;
 674                         }
 675                 } else {
 676                         if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
 677                                 break;
 678
 679                         while ((io_u = get_io_u(td)) != NULL) {
 680                                 if (IS_ERR_OR_NULL(io_u)) {
 681                                         io_u = NULL;
 682                                         ret = FIO_Q_BUSY;
 683                                         goto reap;
 684                                 }
 685
 686                                 /*
 687                                  * We are only interested in the places where
 688                                  * we wrote or trimmed IOs. Turn those into
 689                                  * reads for verification purposes.
 690                                  */
 691                                 if (io_u->ddir == DDIR_READ) {
 692                                         /*
 693                                          * Pretend we issued it for rwmix
 694                                          * accounting
 695                                          */
 696                                         td->io_issues[DDIR_READ]++;
 697                                         put_io_u(td, io_u);
 698                                         continue;
 699                                 } else if (io_u->ddir == DDIR_TRIM) {
 700                                         io_u->ddir = DDIR_READ;
 701                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 702                                         break;
 703                                 } else if (io_u->ddir == DDIR_WRITE) {
 704                                         io_u->ddir = DDIR_READ;
 705                                         io_u->numberio = td->verify_read_issues;
 706                                         td->verify_read_issues++;
 707                                         populate_verify_io_u(td, io_u);
 708                                         break;
 709                                 } else {
 710                                         put_io_u(td, io_u);
 711                                         continue;
 712                                 }
 713                         }
 714
 715                         if (!io_u)
 716                                 break;
 717                 }
 718
 719                 if (verify_state_should_stop(td, io_u)) {
 720                         put_io_u(td, io_u);
 721                         break;
 722                 }
 723
 724                 if (td->o.verify_async)
 725                         io_u->end_io = verify_io_u_async;
 726                 else
 727                         io_u->end_io = verify_io_u;
 728
 729                 ddir = io_u->ddir;
 730                 if (!td->o.disable_slat)
 731                         fio_gettime(&io_u->start_time, NULL);
 732
 733                 ret = io_u_submit(td, io_u);
 734
 735                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 736                         break;
 737
 738                 /*
 739                  * if we can queue more, do so. but check if there are
 740                  * completed io_u's first. Note that we can get BUSY even
 741                  * without IO queued, if the system is resource starved.
 742                  */
 743 reap:
 744                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 745                 if (full || io_in_polling(td))
 746                         ret = wait_for_completions(td, NULL);
 747
 748                 if (ret < 0)
 749                         break;
 750         }
 751
 752         check_update_rusage(td);
 753
 754         if (!td->error) {
 755                 min_events = td->cur_depth;
 756
 757                 if (min_events)
 758                         ret = io_u_queued_complete(td, min_events);
 759         } else
 760                 cleanup_pending_aio(td);
 761
 762         td_set_runstate(td, TD_RUNNING);
 763
 764         dprint(FD_VERIFY, "exiting loop\n");
 765 }
 766
 767 static bool exceeds_number_ios(struct thread_data *td)
 768 {
 769         unsigned long long number_ios;
 770
 771         if (!td->o.number_ios)
 772                 return false;
 773
 774         number_ios = ddir_rw_sum(td->io_blocks);
 775         number_ios += td->io_u_queued + td->io_u_in_flight;
 776
 777         return number_ios >= (td->o.number_ios * td->loops);
 778 }
 779
 780 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 781 {
 782         unsigned long long bytes, limit;
 783
 784         if (td_rw(td))
 785                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 786         else if (td_write(td))
 787                 bytes = this_bytes[DDIR_WRITE];
 788         else if (td_read(td))
 789                 bytes = this_bytes[DDIR_READ];
 790         else
 791                 bytes = this_bytes[DDIR_TRIM];
 792
 793         if (td->o.io_size)
 794                 limit = td->o.io_size;
 795         else
 796                 limit = td->o.size;
 797
 798         limit *= td->loops;
 799         return bytes >= limit || exceeds_number_ios(td);
 800 }
 801
 802 static bool io_issue_bytes_exceeded(struct thread_data *td)
 803 {
 804         return io_bytes_exceeded(td, td->io_issue_bytes);
 805 }
 806
 807 static bool io_complete_bytes_exceeded(struct thread_data *td)
 808 {
 809         return io_bytes_exceeded(td, td->this_io_bytes);
 810 }
 811
 812 /*
 813  * used to calculate the next io time for rate control
 814  *
 815  */
 816 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 817 {
 818         uint64_t bps = td->rate_bps[ddir];
 819
 820         assert(!(td->flags & TD_F_CHILD));
 821
 822         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 823                 uint64_t val, iops;
 824
 825                 iops = bps / td->o.min_bs[ddir];
 826                 val = (int64_t) (1000000 / iops) *
 827                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 828                 if (val) {
 829                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 830                                         (unsigned long long) 1000000 / val,
 831                                         ddir);
 832                 }
 833                 td->last_usec[ddir] += val;
 834                 return td->last_usec[ddir];
 835         } else if (bps) {
 836                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 837                 uint64_t secs = bytes / bps;
 838                 uint64_t remainder = bytes % bps;
 839
 840                 return remainder * 1000000 / bps + secs * 1000000;
 841         }
 842
 843         return 0;
 844 }
 845
 846 static void init_thinktime(struct thread_data *td)
 847 {
 848         if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
 849                 td->thinktime_blocks_counter = td->io_blocks;
 850         else
 851                 td->thinktime_blocks_counter = td->io_issues;
 852         td->last_thinktime = td->epoch;
 853         td->last_thinktime_blocks = 0;
 854 }
 855
 856 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
 857                              struct timespec *time)
 858 {
 859         unsigned long long b;
 860         uint64_t total;
 861         int left;
 862         struct timespec now;
 863         bool stall = false;
 864
 865         if (td->o.thinktime_iotime) {
 866                 fio_gettime(&now, NULL);
 867                 if (utime_since(&td->last_thinktime, &now)
 868                     >= td->o.thinktime_iotime + td->o.thinktime) {
 869                         stall = true;
 870                 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
 871                         /*
 872                          * When thinktime_iotime is set and thinktime_blocks is
 873                          * not set, skip the thinktime_blocks check, since
 874                          * thinktime_blocks default value 1 does not work
 875                          * together with thinktime_iotime.
 876                          */
 877                         return;
 878                 }
 879
 880         }
 881
 882         b = ddir_rw_sum(td->thinktime_blocks_counter);
 883         if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
 884                 stall = true;
 885
 886         if (!stall)
 887                 return;
 888
 889         io_u_quiesce(td);
 890
 891         total = 0;
 892         if (td->o.thinktime_spin)
 893                 total = usec_spin(td->o.thinktime_spin);
 894
 895         left = td->o.thinktime - total;
 896         if (left)
 897                 total += usec_sleep(td, left);
 898
 899         /*
 900          * If we're ignoring thinktime for the rate, add the number of bytes
 901          * we would have done while sleeping, minus one block to ensure we
 902          * start issuing immediately after the sleep.
 903          */
 904         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 905                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 906                 uint64_t bs = td->o.min_bs[ddir];
 907                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 908                 uint64_t over;
 909
 910                 if (usperop <= total)
 911                         over = bs;
 912                 else
 913                         over = (usperop - total) / usperop * -bs;
 914
 915                 td->rate_io_issue_bytes[ddir] += (missed - over);
 916                 /* adjust for rate_process=poisson */
 917                 td->last_usec[ddir] += total;
 918         }
 919
 920         if (time && should_check_rate(td))
 921                 fio_gettime(time, NULL);
 922
 923         td->last_thinktime_blocks = b;
 924         if (td->o.thinktime_iotime)
 925                 td->last_thinktime = now;
 926 }
 927
 928 /*
 929  * Main IO worker function. It retrieves io_u's to process and queues
 930  * and reaps them, checking for rate and errors along the way.
 931  *
 932  * Returns number of bytes written and trimmed.
 933  */
 934 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 935 {
 936         unsigned int i;
 937         int ret = 0;
 938         uint64_t total_bytes, bytes_issued = 0;
 939
 940         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 941                 bytes_done[i] = td->bytes_done[i];
 942
 943         if (in_ramp_time(td))
 944                 td_set_runstate(td, TD_RAMP);
 945         else
 946                 td_set_runstate(td, TD_RUNNING);
 947
 948         lat_target_init(td);
 949
 950         total_bytes = td->o.size;
 951         /*
 952         * Allow random overwrite workloads to write up to io_size
 953         * before starting verification phase as 'size' doesn't apply.
 954         */
 955         if (td_write(td) && td_random(td) && td->o.norandommap)
 956                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 957         /*
 958          * If verify_backlog is enabled, we'll run the verify in this
 959          * handler as well. For that case, we may need up to twice the
 960          * amount of bytes.
 961          */
 962         if (td->o.verify != VERIFY_NONE &&
 963            (td_write(td) && td->o.verify_backlog))
 964                 total_bytes += td->o.size;
 965
 966         /* In trimwrite mode, each byte is trimmed and then written, so
 967          * allow total_bytes or number of ios to be twice as big */
 968         if (td_trimwrite(td)) {
 969                 total_bytes += td->total_io_size;
 970                 td->o.number_ios *= 2;
 971         }
 972
 973         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 974                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 975                 td->o.time_based) {
 976                 struct timespec comp_time;
 977                 struct io_u *io_u;
 978                 int full;
 979                 enum fio_ddir ddir;
 980
 981                 check_update_rusage(td);
 982
 983                 if (td->terminate || td->done)
 984                         break;
 985
 986                 update_ts_cache(td);
 987
 988                 if (runtime_exceeded(td, &td->ts_cache)) {
 989                         __update_ts_cache(td);
 990                         if (runtime_exceeded(td, &td->ts_cache)) {
 991                                 fio_mark_td_terminate(td);
 992                                 break;
 993                         }
 994                 }
 995
 996                 if (flow_threshold_exceeded(td))
 997                         continue;
 998
 999                 /*
1000                  * Break if we exceeded the bytes. The exception is time
1001                  * based runs, but we still need to break out of the loop
1002                  * for those to run verification, if enabled.
1003                  * Jobs read from iolog do not use this stop condition.
1004                  */
1005                 if (bytes_issued >= total_bytes &&
1006                     !td->o.read_iolog_file &&
1007                     (!td->o.time_based ||
1008                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
1009                         break;
1010
1011                 io_u = get_io_u(td);
1012                 if (IS_ERR_OR_NULL(io_u)) {
1013                         int err = PTR_ERR(io_u);
1014
1015                         io_u = NULL;
1016                         ddir = DDIR_INVAL;
1017                         if (err == -EBUSY) {
1018                                 ret = FIO_Q_BUSY;
1019                                 goto reap;
1020                         }
1021                         if (td->o.latency_target)
1022                                 goto reap;
1023                         break;
1024                 }
1025
1026                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
1027                         io_u->numberio = td->io_issues[io_u->ddir];
1028                         populate_verify_io_u(td, io_u);
1029                 }
1030
1031                 ddir = io_u->ddir;
1032
1033                 /*
1034                  * Add verification end_io handler if:
1035                  *      - Asked to verify (!td_rw(td))
1036                  *      - Or the io_u is from our verify list (mixed write/ver)
1037                  */
1038                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1039                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1040
1041                         if (verify_state_should_stop(td, io_u)) {
1042                                 put_io_u(td, io_u);
1043                                 break;
1044                         }
1045
1046                         if (td->o.verify_async)
1047                                 io_u->end_io = verify_io_u_async;
1048                         else
1049                                 io_u->end_io = verify_io_u;
1050                         td_set_runstate(td, TD_VERIFYING);
1051                 } else if (in_ramp_time(td))
1052                         td_set_runstate(td, TD_RAMP);
1053                 else
1054                         td_set_runstate(td, TD_RUNNING);
1055
1056                 /*
1057                  * Always log IO before it's issued, so we know the specific
1058                  * order of it. The logged unit will track when the IO has
1059                  * completed.
1060                  */
1061                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1062                     td->o.do_verify &&
1063                     td->o.verify != VERIFY_NONE &&
1064                     !td->o.experimental_verify)
1065                         log_io_piece(td, io_u);
1066
1067                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1068                         const unsigned long long blen = io_u->xfer_buflen;
1069                         const enum fio_ddir __ddir = acct_ddir(io_u);
1070
1071                         if (td->error)
1072                                 break;
1073
1074                         workqueue_enqueue(&td->io_wq, &io_u->work);
1075                         ret = FIO_Q_QUEUED;
1076
1077                         if (ddir_rw(__ddir)) {
1078                                 td->io_issues[__ddir]++;
1079                                 td->io_issue_bytes[__ddir] += blen;
1080                                 td->rate_io_issue_bytes[__ddir] += blen;
1081                         }
1082
1083                         if (should_check_rate(td)) {
1084                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1085                                 fio_gettime(&comp_time, NULL);
1086                         }
1087
1088                 } else {
1089                         ret = io_u_submit(td, io_u);
1090
1091                         if (should_check_rate(td))
1092                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1093
1094                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1095                                 break;
1096
1097                         /*
1098                          * See if we need to complete some commands. Note that
1099                          * we can get BUSY even without IO queued, if the
1100                          * system is resource starved.
1101                          */
1102 reap:
1103                         full = queue_full(td) ||
1104                                 (ret == FIO_Q_BUSY && td->cur_depth);
1105                         if (full || io_in_polling(td))
1106                                 ret = wait_for_completions(td, &comp_time);
1107                 }
1108                 if (ret < 0)
1109                         break;
1110
1111                 if (ddir_rw(ddir) && td->o.thinktime)
1112                         handle_thinktime(td, ddir, &comp_time);
1113
1114                 if (!ddir_rw_sum(td->bytes_done) &&
1115                     !td_ioengine_flagged(td, FIO_NOIO))
1116                         continue;
1117
1118                 if (!in_ramp_time(td) && should_check_rate(td)) {
1119                         if (check_min_rate(td, &comp_time)) {
1120                                 if (exitall_on_terminate || td->o.exitall_error)
1121                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1122                                 td_verror(td, EIO, "check_min_rate");
1123                                 break;
1124                         }
1125                 }
1126                 if (!in_ramp_time(td) && td->o.latency_target)
1127                         lat_target_check(td);
1128         }
1129
1130         check_update_rusage(td);
1131
1132         if (td->trim_entries)
1133                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1134
1135         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1136                 td->error = 0;
1137                 fio_mark_td_terminate(td);
1138         }
1139         if (!td->error) {
1140                 struct fio_file *f;
1141
1142                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1143                         workqueue_flush(&td->io_wq);
1144                         i = 0;
1145                 } else
1146                         i = td->cur_depth;
1147
1148                 if (i) {
1149                         ret = io_u_queued_complete(td, i);
1150                         if (td->o.fill_device &&
1151                             (td->error == ENOSPC || td->error == EDQUOT))
1152                                 td->error = 0;
1153                 }
1154
1155                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1156                         td_set_runstate(td, TD_FSYNCING);
1157
1158                         for_each_file(td, f, i) {
1159                                 if (!fio_file_fsync(td, f))
1160                                         continue;
1161
1162                                 log_err("fio: end_fsync failed for file %s\n",
1163                                                                 f->file_name);
1164                         }
1165                 }
1166         } else {
1167                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1168                         workqueue_flush(&td->io_wq);
1169                 cleanup_pending_aio(td);
1170         }
1171
1172         /*
1173          * stop job if we failed doing any IO
1174          */
1175         if (!ddir_rw_sum(td->this_io_bytes))
1176                 td->done = 1;
1177
1178         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1179                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1180 }
1181
1182 static void free_file_completion_logging(struct thread_data *td)
1183 {
1184         struct fio_file *f;
1185         unsigned int i;
1186
1187         for_each_file(td, f, i) {
1188                 if (!f->last_write_comp)
1189                         break;
1190                 sfree(f->last_write_comp);
1191         }
1192 }
1193
1194 static int init_file_completion_logging(struct thread_data *td,
1195                                         unsigned int depth)
1196 {
1197         struct fio_file *f;
1198         unsigned int i;
1199
1200         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1201                 return 0;
1202
1203         for_each_file(td, f, i) {
1204                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1205                 if (!f->last_write_comp)
1206                         goto cleanup;
1207         }
1208
1209         return 0;
1210
1211 cleanup:
1212         free_file_completion_logging(td);
1213         log_err("fio: failed to alloc write comp data\n");
1214         return 1;
1215 }
1216
1217 static void cleanup_io_u(struct thread_data *td)
1218 {
1219         struct io_u *io_u;
1220
1221         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1222
1223                 if (td->io_ops->io_u_free)
1224                         td->io_ops->io_u_free(td, io_u);
1225
1226                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1227         }
1228
1229         free_io_mem(td);
1230
1231         io_u_rexit(&td->io_u_requeues);
1232         io_u_qexit(&td->io_u_freelist, false);
1233         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1234
1235         free_file_completion_logging(td);
1236 }
1237
1238 static int init_io_u(struct thread_data *td)
1239 {
1240         struct io_u *io_u;
1241         int cl_align, i, max_units;
1242         int err;
1243
1244         max_units = td->o.iodepth;
1245
1246         err = 0;
1247         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1248         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1249         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1250
1251         if (err) {
1252                 log_err("fio: failed setting up IO queues\n");
1253                 return 1;
1254         }
1255
1256         cl_align = os_cache_line_size();
1257
1258         for (i = 0; i < max_units; i++) {
1259                 void *ptr;
1260
1261                 if (td->terminate)
1262                         return 1;
1263
1264                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1265                 if (!ptr) {
1266                         log_err("fio: unable to allocate aligned memory\n");
1267                         return 1;
1268                 }
1269
1270                 io_u = ptr;
1271                 memset(io_u, 0, sizeof(*io_u));
1272                 INIT_FLIST_HEAD(&io_u->verify_list);
1273                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1274
1275                 io_u->index = i;
1276                 io_u->flags = IO_U_F_FREE;
1277                 io_u_qpush(&td->io_u_freelist, io_u);
1278
1279                 /*
1280                  * io_u never leaves this stack, used for iteration of all
1281                  * io_u buffers.
1282                  */
1283                 io_u_qpush(&td->io_u_all, io_u);
1284
1285                 if (td->io_ops->io_u_init) {
1286                         int ret = td->io_ops->io_u_init(td, io_u);
1287
1288                         if (ret) {
1289                                 log_err("fio: failed to init engine data: %d\n", ret);
1290                                 return 1;
1291                         }
1292                 }
1293         }
1294
1295         if (init_io_u_buffers(td))
1296                 return 1;
1297
1298         if (init_file_completion_logging(td, max_units))
1299                 return 1;
1300
1301         return 0;
1302 }
1303
1304 int init_io_u_buffers(struct thread_data *td)
1305 {
1306         struct io_u *io_u;
1307         unsigned long long max_bs, min_write;
1308         int i, max_units;
1309         int data_xfer = 1;
1310         char *p;
1311
1312         max_units = td->o.iodepth;
1313         max_bs = td_max_bs(td);
1314         min_write = td->o.min_bs[DDIR_WRITE];
1315         td->orig_buffer_size = (unsigned long long) max_bs
1316                                         * (unsigned long long) max_units;
1317
1318         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1319                 data_xfer = 0;
1320
1321         /*
1322          * if we may later need to do address alignment, then add any
1323          * possible adjustment here so that we don't cause a buffer
1324          * overflow later. this adjustment may be too much if we get
1325          * lucky and the allocator gives us an aligned address.
1326          */
1327         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1328             td_ioengine_flagged(td, FIO_RAWIO))
1329                 td->orig_buffer_size += page_mask + td->o.mem_align;
1330
1331         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1332                 unsigned long long bs;
1333
1334                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1335                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1336         }
1337
1338         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1339                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1340                 return 1;
1341         }
1342
1343         if (data_xfer && allocate_io_mem(td))
1344                 return 1;
1345
1346         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1347             td_ioengine_flagged(td, FIO_RAWIO))
1348                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1349         else
1350                 p = td->orig_buffer;
1351
1352         for (i = 0; i < max_units; i++) {
1353                 io_u = td->io_u_all.io_us[i];
1354                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1355
1356                 if (data_xfer) {
1357                         io_u->buf = p;
1358                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1359
1360                         if (td_write(td))
1361                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1362                         if (td_write(td) && td->o.verify_pattern_bytes) {
1363                                 /*
1364                                  * Fill the buffer with the pattern if we are
1365                                  * going to be doing writes.
1366                                  */
1367                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1368                         }
1369                 }
1370                 p += max_bs;
1371         }
1372
1373         return 0;
1374 }
1375
1376 #ifdef FIO_HAVE_IOSCHED_SWITCH
1377 /*
1378  * These functions are Linux specific.
1379  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1380  */
1381 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1382 {
1383         char tmp[256], tmp2[128], *p;
1384         FILE *f;
1385         int ret;
1386
1387         assert(file->du && file->du->sysfs_root);
1388         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1389
1390         f = fopen(tmp, "r+");
1391         if (!f) {
1392                 if (errno == ENOENT) {
1393                         log_err("fio: os or kernel doesn't support IO scheduler"
1394                                 " switching\n");
1395                         return 0;
1396                 }
1397                 td_verror(td, errno, "fopen iosched");
1398                 return 1;
1399         }
1400
1401         /*
1402          * Set io scheduler.
1403          */
1404         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1405         if (ferror(f) || ret != 1) {
1406                 td_verror(td, errno, "fwrite");
1407                 fclose(f);
1408                 return 1;
1409         }
1410
1411         rewind(f);
1412
1413         /*
1414          * Read back and check that the selected scheduler is now the default.
1415          */
1416         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1417         if (ferror(f) || ret < 0) {
1418                 td_verror(td, errno, "fread");
1419                 fclose(f);
1420                 return 1;
1421         }
1422         tmp[ret] = '\0';
1423         /*
1424          * either a list of io schedulers or "none\n" is expected. Strip the
1425          * trailing newline.
1426          */
1427         p = tmp;
1428         strsep(&p, "\n");
1429
1430         /*
1431          * Write to "none" entry doesn't fail, so check the result here.
1432          */
1433         if (!strcmp(tmp, "none")) {
1434                 log_err("fio: io scheduler is not tunable\n");
1435                 fclose(f);
1436                 return 0;
1437         }
1438
1439         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1440         if (!strstr(tmp, tmp2)) {
1441                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1442                 td_verror(td, EINVAL, "iosched_switch");
1443                 fclose(f);
1444                 return 1;
1445         }
1446
1447         fclose(f);
1448         return 0;
1449 }
1450
1451 static int switch_ioscheduler(struct thread_data *td)
1452 {
1453         struct fio_file *f;
1454         unsigned int i;
1455         int ret = 0;
1456
1457         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1458                 return 0;
1459
1460         assert(td->files && td->files[0]);
1461
1462         for_each_file(td, f, i) {
1463
1464                 /* Only consider regular files and block device files */
1465                 switch (f->filetype) {
1466                 case FIO_TYPE_FILE:
1467                 case FIO_TYPE_BLOCK:
1468                         /*
1469                          * Make sure that the device hosting the file could
1470                          * be determined.
1471                          */
1472                         if (!f->du)
1473                                 continue;
1474                         break;
1475                 case FIO_TYPE_CHAR:
1476                 case FIO_TYPE_PIPE:
1477                 default:
1478                         continue;
1479                 }
1480
1481                 ret = set_ioscheduler(td, f);
1482                 if (ret)
1483                         return ret;
1484         }
1485
1486         return 0;
1487 }
1488
1489 #else
1490
1491 static int switch_ioscheduler(struct thread_data *td)
1492 {
1493         return 0;
1494 }
1495
1496 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1497
1498 static bool keep_running(struct thread_data *td)
1499 {
1500         unsigned long long limit;
1501
1502         if (td->done)
1503                 return false;
1504         if (td->terminate)
1505                 return false;
1506         if (td->o.time_based)
1507                 return true;
1508         if (td->o.loops) {
1509                 td->o.loops--;
1510                 return true;
1511         }
1512         if (exceeds_number_ios(td))
1513                 return false;
1514
1515         if (td->o.io_size)
1516                 limit = td->o.io_size;
1517         else
1518                 limit = td->o.size;
1519
1520         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1521                 uint64_t diff;
1522
1523                 /*
1524                  * If the difference is less than the maximum IO size, we
1525                  * are done.
1526                  */
1527                 diff = limit - ddir_rw_sum(td->io_bytes);
1528                 if (diff < td_max_bs(td))
1529                         return false;
1530
1531                 if (fio_files_done(td) && !td->o.io_size)
1532                         return false;
1533
1534                 return true;
1535         }
1536
1537         return false;
1538 }
1539
1540 static int exec_string(struct thread_options *o, const char *string,
1541                        const char *mode)
1542 {
1543         int ret;
1544         char *str;
1545
1546         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1547                 return -1;
1548
1549         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1550                  o->name, mode);
1551         ret = system(str);
1552         if (ret == -1)
1553                 log_err("fio: exec of cmd <%s> failed\n", str);
1554
1555         free(str);
1556         return ret;
1557 }
1558
1559 /*
1560  * Dry run to compute correct state of numberio for verification.
1561  */
1562 static uint64_t do_dry_run(struct thread_data *td)
1563 {
1564         td_set_runstate(td, TD_RUNNING);
1565
1566         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1567                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1568                 struct io_u *io_u;
1569                 int ret;
1570
1571                 if (td->terminate || td->done)
1572                         break;
1573
1574                 io_u = get_io_u(td);
1575                 if (IS_ERR_OR_NULL(io_u))
1576                         break;
1577
1578                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1579                 io_u->error = 0;
1580                 io_u->resid = 0;
1581                 if (ddir_rw(acct_ddir(io_u)))
1582                         td->io_issues[acct_ddir(io_u)]++;
1583                 if (ddir_rw(io_u->ddir)) {
1584                         io_u_mark_depth(td, 1);
1585                         td->ts.total_io_u[io_u->ddir]++;
1586                 }
1587
1588                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1589                     td->o.do_verify &&
1590                     td->o.verify != VERIFY_NONE &&
1591                     !td->o.experimental_verify)
1592                         log_io_piece(td, io_u);
1593
1594                 ret = io_u_sync_complete(td, io_u);
1595                 (void) ret;
1596         }
1597
1598         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1599 }
1600
1601 struct fork_data {
1602         struct thread_data *td;
1603         struct sk_out *sk_out;
1604 };
1605
1606 /*
1607  * Entry point for the thread based jobs. The process based jobs end up
1608  * here as well, after a little setup.
1609  */
1610 static void *thread_main(void *data)
1611 {
1612         struct fork_data *fd = data;
1613         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1614         struct thread_data *td = fd->td;
1615         struct thread_options *o = &td->o;
1616         struct sk_out *sk_out = fd->sk_out;
1617         uint64_t bytes_done[DDIR_RWDIR_CNT];
1618         int deadlock_loop_cnt;
1619         bool clear_state;
1620         int res, ret;
1621
1622         sk_out_assign(sk_out);
1623         free(fd);
1624
1625         if (!o->use_thread) {
1626                 setsid();
1627                 td->pid = getpid();
1628         } else
1629                 td->pid = gettid();
1630
1631         fio_local_clock_init();
1632
1633         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1634
1635         if (is_backend)
1636                 fio_server_send_start(td);
1637
1638         INIT_FLIST_HEAD(&td->io_log_list);
1639         INIT_FLIST_HEAD(&td->io_hist_list);
1640         INIT_FLIST_HEAD(&td->verify_list);
1641         INIT_FLIST_HEAD(&td->trim_list);
1642         td->io_hist_tree = RB_ROOT;
1643
1644         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1645         if (ret) {
1646                 td_verror(td, ret, "mutex_cond_init_pshared");
1647                 goto err;
1648         }
1649         ret = cond_init_pshared(&td->verify_cond);
1650         if (ret) {
1651                 td_verror(td, ret, "mutex_cond_pshared");
1652                 goto err;
1653         }
1654
1655         td_set_runstate(td, TD_INITIALIZED);
1656         dprint(FD_MUTEX, "up startup_sem\n");
1657         fio_sem_up(startup_sem);
1658         dprint(FD_MUTEX, "wait on td->sem\n");
1659         fio_sem_down(td->sem);
1660         dprint(FD_MUTEX, "done waiting on td->sem\n");
1661
1662         /*
1663          * A new gid requires privilege, so we need to do this before setting
1664          * the uid.
1665          */
1666         if (o->gid != -1U && setgid(o->gid)) {
1667                 td_verror(td, errno, "setgid");
1668                 goto err;
1669         }
1670         if (o->uid != -1U && setuid(o->uid)) {
1671                 td_verror(td, errno, "setuid");
1672                 goto err;
1673         }
1674
1675         td_zone_gen_index(td);
1676
1677         /*
1678          * Do this early, we don't want the compress threads to be limited
1679          * to the same CPUs as the IO workers. So do this before we set
1680          * any potential CPU affinity
1681          */
1682         if (iolog_compress_init(td, sk_out))
1683                 goto err;
1684
1685         /*
1686          * If we have a gettimeofday() thread, make sure we exclude that
1687          * thread from this job
1688          */
1689         if (o->gtod_cpu)
1690                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1691
1692         /*
1693          * Set affinity first, in case it has an impact on the memory
1694          * allocations.
1695          */
1696         if (fio_option_is_set(o, cpumask)) {
1697                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1698                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1699                         if (!ret) {
1700                                 log_err("fio: no CPUs set\n");
1701                                 log_err("fio: Try increasing number of available CPUs\n");
1702                                 td_verror(td, EINVAL, "cpus_split");
1703                                 goto err;
1704                         }
1705                 }
1706                 ret = fio_setaffinity(td->pid, o->cpumask);
1707                 if (ret == -1) {
1708                         td_verror(td, errno, "cpu_set_affinity");
1709                         goto err;
1710                 }
1711         }
1712
1713 #ifdef CONFIG_LIBNUMA
1714         /* numa node setup */
1715         if (fio_option_is_set(o, numa_cpunodes) ||
1716             fio_option_is_set(o, numa_memnodes)) {
1717                 struct bitmask *mask;
1718
1719                 if (numa_available() < 0) {
1720                         td_verror(td, errno, "Does not support NUMA API\n");
1721                         goto err;
1722                 }
1723
1724                 if (fio_option_is_set(o, numa_cpunodes)) {
1725                         mask = numa_parse_nodestring(o->numa_cpunodes);
1726                         ret = numa_run_on_node_mask(mask);
1727                         numa_free_nodemask(mask);
1728                         if (ret == -1) {
1729                                 td_verror(td, errno, \
1730                                         "numa_run_on_node_mask failed\n");
1731                                 goto err;
1732                         }
1733                 }
1734
1735                 if (fio_option_is_set(o, numa_memnodes)) {
1736                         mask = NULL;
1737                         if (o->numa_memnodes)
1738                                 mask = numa_parse_nodestring(o->numa_memnodes);
1739
1740                         switch (o->numa_mem_mode) {
1741                         case MPOL_INTERLEAVE:
1742                                 numa_set_interleave_mask(mask);
1743                                 break;
1744                         case MPOL_BIND:
1745                                 numa_set_membind(mask);
1746                                 break;
1747                         case MPOL_LOCAL:
1748                                 numa_set_localalloc();
1749                                 break;
1750                         case MPOL_PREFERRED:
1751                                 numa_set_preferred(o->numa_mem_prefer_node);
1752                                 break;
1753                         case MPOL_DEFAULT:
1754                         default:
1755                                 break;
1756                         }
1757
1758                         if (mask)
1759                                 numa_free_nodemask(mask);
1760
1761                 }
1762         }
1763 #endif
1764
1765         if (fio_pin_memory(td))
1766                 goto err;
1767
1768         /*
1769          * May alter parameters that init_io_u() will use, so we need to
1770          * do this first.
1771          */
1772         if (!init_iolog(td))
1773                 goto err;
1774
1775         /* ioprio_set() has to be done before td_io_init() */
1776         if (fio_option_is_set(o, ioprio) ||
1777             fio_option_is_set(o, ioprio_class)) {
1778                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1779                 if (ret == -1) {
1780                         td_verror(td, errno, "ioprio_set");
1781                         goto err;
1782                 }
1783                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1784                 td->ts.ioprio = td->ioprio;
1785         }
1786
1787         if (td_io_init(td))
1788                 goto err;
1789
1790         if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
1791                 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1792                          "are selected, queue depth will be capped at 1\n");
1793         }
1794
1795         if (init_io_u(td))
1796                 goto err;
1797
1798         if (td->io_ops->post_init && td->io_ops->post_init(td))
1799                 goto err;
1800
1801         if (o->verify_async && verify_async_init(td))
1802                 goto err;
1803
1804         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1805                 goto err;
1806
1807         errno = 0;
1808         if (nice(o->nice) == -1 && errno != 0) {
1809                 td_verror(td, errno, "nice");
1810                 goto err;
1811         }
1812
1813         if (o->ioscheduler && switch_ioscheduler(td))
1814                 goto err;
1815
1816         if (!o->create_serialize && setup_files(td))
1817                 goto err;
1818
1819         if (!init_random_map(td))
1820                 goto err;
1821
1822         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1823                 goto err;
1824
1825         if (o->pre_read && !pre_read_files(td))
1826                 goto err;
1827
1828         fio_verify_init(td);
1829
1830         if (rate_submit_init(td, sk_out))
1831                 goto err;
1832
1833         set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1834         fio_getrusage(&td->ru_start);
1835         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1836         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1837         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1838
1839         init_thinktime(td);
1840
1841         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1842                         o->ratemin[DDIR_TRIM]) {
1843                 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1844                                         sizeof(td->bw_sample_time));
1845                 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1846                                         sizeof(td->bw_sample_time));
1847                 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1848                                         sizeof(td->bw_sample_time));
1849         }
1850
1851         memset(bytes_done, 0, sizeof(bytes_done));
1852         clear_state = false;
1853
1854         while (keep_running(td)) {
1855                 uint64_t verify_bytes;
1856
1857                 fio_gettime(&td->start, NULL);
1858                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1859
1860                 if (clear_state) {
1861                         clear_io_state(td, 0);
1862
1863                         if (o->unlink_each_loop && unlink_all_files(td))
1864                                 break;
1865                 }
1866
1867                 prune_io_piece_log(td);
1868
1869                 if (td->o.verify_only && td_write(td))
1870                         verify_bytes = do_dry_run(td);
1871                 else {
1872                         if (!td->o.rand_repeatable)
1873                                 /* save verify rand state to replay hdr seeds later at verify */
1874                                 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
1875                         do_io(td, bytes_done);
1876                         if (!td->o.rand_repeatable)
1877                                 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
1878                         if (!ddir_rw_sum(bytes_done)) {
1879                                 fio_mark_td_terminate(td);
1880                                 verify_bytes = 0;
1881                         } else {
1882                                 verify_bytes = bytes_done[DDIR_WRITE] +
1883                                                 bytes_done[DDIR_TRIM];
1884                         }
1885                 }
1886
1887                 /*
1888                  * If we took too long to shut down, the main thread could
1889                  * already consider us reaped/exited. If that happens, break
1890                  * out and clean up.
1891                  */
1892                 if (td->runstate >= TD_EXITED)
1893                         break;
1894
1895                 clear_state = true;
1896
1897                 /*
1898                  * Make sure we've successfully updated the rusage stats
1899                  * before waiting on the stat mutex. Otherwise we could have
1900                  * the stat thread holding stat mutex and waiting for
1901                  * the rusage_sem, which would never get upped because
1902                  * this thread is waiting for the stat mutex.
1903                  */
1904                 deadlock_loop_cnt = 0;
1905                 do {
1906                         check_update_rusage(td);
1907                         if (!fio_sem_down_trylock(stat_sem))
1908                                 break;
1909                         usleep(1000);
1910                         if (deadlock_loop_cnt++ > 5000) {
1911                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1912                                 td->error = EDEADLK;
1913                                 goto err;
1914                         }
1915                 } while (1);
1916
1917                 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1918                         ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
1919                         update_runtime(td, elapsed_us, DDIR_READ);
1920                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1921                         update_runtime(td, elapsed_us, DDIR_WRITE);
1922                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1923                         update_runtime(td, elapsed_us, DDIR_TRIM);
1924                 fio_gettime(&td->start, NULL);
1925                 fio_sem_up(stat_sem);
1926
1927                 if (td->error || td->terminate)
1928                         break;
1929
1930                 if (!o->do_verify ||
1931                     o->verify == VERIFY_NONE ||
1932                     td_ioengine_flagged(td, FIO_UNIDIR))
1933                         continue;
1934
1935                 clear_io_state(td, 0);
1936
1937                 fio_gettime(&td->start, NULL);
1938
1939                 do_verify(td, verify_bytes);
1940
1941                 /*
1942                  * See comment further up for why this is done here.
1943                  */
1944                 check_update_rusage(td);
1945
1946                 fio_sem_down(stat_sem);
1947                 update_runtime(td, elapsed_us, DDIR_READ);
1948                 fio_gettime(&td->start, NULL);
1949                 fio_sem_up(stat_sem);
1950
1951                 if (td->error || td->terminate)
1952                         break;
1953         }
1954
1955         /*
1956          * Acquire this lock if we were doing overlap checking in
1957          * offload mode so that we don't clean up this job while
1958          * another thread is checking its io_u's for overlap
1959          */
1960         if (td_offload_overlap(td)) {
1961                 int res = pthread_mutex_lock(&overlap_check);
1962                 assert(res == 0);
1963         }
1964         td_set_runstate(td, TD_FINISHING);
1965         if (td_offload_overlap(td)) {
1966                 res = pthread_mutex_unlock(&overlap_check);
1967                 assert(res == 0);
1968         }
1969
1970         update_rusage_stat(td);
1971         td->ts.total_run_time = mtime_since_now(&td->epoch);
1972         for_each_rw_ddir(ddir) {
1973                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1974         }
1975
1976         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1977             (td->o.verify != VERIFY_NONE && td_write(td)))
1978                 verify_save_state(td->thread_number);
1979
1980         fio_unpin_memory(td);
1981
1982         td_writeout_logs(td, true);
1983
1984         iolog_compress_exit(td);
1985         rate_submit_exit(td);
1986
1987         if (o->exec_postrun)
1988                 exec_string(o, o->exec_postrun, "postrun");
1989
1990         if (exitall_on_terminate || (o->exitall_error && td->error))
1991                 fio_terminate_threads(td->groupid, td->o.exit_what);
1992
1993 err:
1994         if (td->error)
1995                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1996                                                         td->verror);
1997
1998         if (o->verify_async)
1999                 verify_async_exit(td);
2000
2001         close_and_free_files(td);
2002         cleanup_io_u(td);
2003         close_ioengine(td);
2004         cgroup_shutdown(td, cgroup_mnt);
2005         verify_free_state(td);
2006         td_zone_free_index(td);
2007
2008         if (fio_option_is_set(o, cpumask)) {
2009                 ret = fio_cpuset_exit(&o->cpumask);
2010                 if (ret)
2011                         td_verror(td, ret, "fio_cpuset_exit");
2012         }
2013
2014         /*
2015          * do this very late, it will log file closing as well
2016          */
2017         if (o->write_iolog_file)
2018                 write_iolog_close(td);
2019         if (td->io_log_rfile)
2020                 fclose(td->io_log_rfile);
2021
2022         td_set_runstate(td, TD_EXITED);
2023
2024         /*
2025          * Do this last after setting our runstate to exited, so we
2026          * know that the stat thread is signaled.
2027          */
2028         check_update_rusage(td);
2029
2030         sk_out_drop();
2031         return (void *) (uintptr_t) td->error;
2032 }
2033
2034 /*
2035  * Run over the job map and reap the threads that have exited, if any.
2036  */
2037 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2038                          uint64_t *m_rate)
2039 {
2040         struct thread_data *td;
2041         unsigned int cputhreads, realthreads, pending;
2042         int i, status, ret;
2043
2044         /*
2045          * reap exited threads (TD_EXITED -> TD_REAPED)
2046          */
2047         realthreads = pending = cputhreads = 0;
2048         for_each_td(td, i) {
2049                 int flags = 0;
2050
2051                 if (!strcmp(td->o.ioengine, "cpuio"))
2052                         cputhreads++;
2053                 else
2054                         realthreads++;
2055
2056                 if (!td->pid) {
2057                         pending++;
2058                         continue;
2059                 }
2060                 if (td->runstate == TD_REAPED)
2061                         continue;
2062                 if (td->o.use_thread) {
2063                         if (td->runstate == TD_EXITED) {
2064                                 td_set_runstate(td, TD_REAPED);
2065                                 goto reaped;
2066                         }
2067                         continue;
2068                 }
2069
2070                 flags = WNOHANG;
2071                 if (td->runstate == TD_EXITED)
2072                         flags = 0;
2073
2074                 /*
2075                  * check if someone quit or got killed in an unusual way
2076                  */
2077                 ret = waitpid(td->pid, &status, flags);
2078                 if (ret < 0) {
2079                         if (errno == ECHILD) {
2080                                 log_err("fio: pid=%d disappeared %d\n",
2081                                                 (int) td->pid, td->runstate);
2082                                 td->sig = ECHILD;
2083                                 td_set_runstate(td, TD_REAPED);
2084                                 goto reaped;
2085                         }
2086                         perror("waitpid");
2087                 } else if (ret == td->pid) {
2088                         if (WIFSIGNALED(status)) {
2089                                 int sig = WTERMSIG(status);
2090
2091                                 if (sig != SIGTERM && sig != SIGUSR2)
2092                                         log_err("fio: pid=%d, got signal=%d\n",
2093                                                         (int) td->pid, sig);
2094                                 td->sig = sig;
2095                                 td_set_runstate(td, TD_REAPED);
2096                                 goto reaped;
2097                         }
2098                         if (WIFEXITED(status)) {
2099                                 if (WEXITSTATUS(status) && !td->error)
2100                                         td->error = WEXITSTATUS(status);
2101
2102                                 td_set_runstate(td, TD_REAPED);
2103                                 goto reaped;
2104                         }
2105                 }
2106
2107                 /*
2108                  * If the job is stuck, do a forceful timeout of it and
2109                  * move on.
2110                  */
2111                 if (td->terminate &&
2112                     td->runstate < TD_FSYNCING &&
2113                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2114                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2115                                 "%lu seconds, it appears to be stuck. Doing "
2116                                 "forceful exit of this job.\n",
2117                                 td->o.name, td->runstate,
2118                                 (unsigned long) time_since_now(&td->terminate_time));
2119                         td_set_runstate(td, TD_REAPED);
2120                         goto reaped;
2121                 }
2122
2123                 /*
2124                  * thread is not dead, continue
2125                  */
2126                 pending++;
2127                 continue;
2128 reaped:
2129                 (*nr_running)--;
2130                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2131                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2132                 if (!td->pid)
2133                         pending--;
2134
2135                 if (td->error)
2136                         exit_value++;
2137
2138                 done_secs += mtime_since_now(&td->epoch) / 1000;
2139                 profile_td_exit(td);
2140                 flow_exit_job(td);
2141         }
2142
2143         if (*nr_running == cputhreads && !pending && realthreads)
2144                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2145 }
2146
2147 static bool __check_trigger_file(void)
2148 {
2149         struct stat sb;
2150
2151         if (!trigger_file)
2152                 return false;
2153
2154         if (stat(trigger_file, &sb))
2155                 return false;
2156
2157         if (unlink(trigger_file) < 0)
2158                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2159                                                         strerror(errno));
2160
2161         return true;
2162 }
2163
2164 static bool trigger_timedout(void)
2165 {
2166         if (trigger_timeout)
2167                 if (time_since_genesis() >= trigger_timeout) {
2168                         trigger_timeout = 0;
2169                         return true;
2170                 }
2171
2172         return false;
2173 }
2174
2175 void exec_trigger(const char *cmd)
2176 {
2177         int ret;
2178
2179         if (!cmd || cmd[0] == '\0')
2180                 return;
2181
2182         ret = system(cmd);
2183         if (ret == -1)
2184                 log_err("fio: failed executing %s trigger\n", cmd);
2185 }
2186
2187 void check_trigger_file(void)
2188 {
2189         if (__check_trigger_file() || trigger_timedout()) {
2190                 if (nr_clients)
2191                         fio_clients_send_trigger(trigger_remote_cmd);
2192                 else {
2193                         verify_save_state(IO_LIST_ALL);
2194                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2195                         exec_trigger(trigger_cmd);
2196                 }
2197         }
2198 }
2199
2200 static int fio_verify_load_state(struct thread_data *td)
2201 {
2202         int ret;
2203
2204         if (!td->o.verify_state)
2205                 return 0;
2206
2207         if (is_backend) {
2208                 void *data;
2209
2210                 ret = fio_server_get_verify_state(td->o.name,
2211                                         td->thread_number - 1, &data);
2212                 if (!ret)
2213                         verify_assign_state(td, data);
2214         } else {
2215                 char prefix[PATH_MAX];
2216
2217                 if (aux_path)
2218                         sprintf(prefix, "%s%clocal", aux_path,
2219                                         FIO_OS_PATH_SEPARATOR);
2220                 else
2221                         strcpy(prefix, "local");
2222                 ret = verify_load_state(td, prefix);
2223         }
2224
2225         return ret;
2226 }
2227
2228 static void do_usleep(unsigned int usecs)
2229 {
2230         check_for_running_stats();
2231         check_trigger_file();
2232         usleep(usecs);
2233 }
2234
2235 static bool check_mount_writes(struct thread_data *td)
2236 {
2237         struct fio_file *f;
2238         unsigned int i;
2239
2240         if (!td_write(td) || td->o.allow_mounted_write)
2241                 return false;
2242
2243         /*
2244          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2245          * are mkfs'd and mounted.
2246          */
2247         for_each_file(td, f, i) {
2248 #ifdef FIO_HAVE_CHARDEV_SIZE
2249                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2250 #else
2251                 if (f->filetype != FIO_TYPE_BLOCK)
2252 #endif
2253                         continue;
2254                 if (device_is_mounted(f->file_name))
2255                         goto mounted;
2256         }
2257
2258         return false;
2259 mounted:
2260         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2261         return true;
2262 }
2263
2264 static bool waitee_running(struct thread_data *me)
2265 {
2266         const char *waitee = me->o.wait_for;
2267         const char *self = me->o.name;
2268         struct thread_data *td;
2269         int i;
2270
2271         if (!waitee)
2272                 return false;
2273
2274         for_each_td(td, i) {
2275                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2276                         continue;
2277
2278                 if (td->runstate < TD_EXITED) {
2279                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2280                                         self, td->o.name,
2281                                         runstate_to_name(td->runstate));
2282                         return true;
2283                 }
2284         }
2285
2286         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2287         return false;
2288 }
2289
2290 /*
2291  * Main function for kicking off and reaping jobs, as needed.
2292  */
2293 static void run_threads(struct sk_out *sk_out)
2294 {
2295         struct thread_data *td;
2296         unsigned int i, todo, nr_running, nr_started;
2297         uint64_t m_rate, t_rate;
2298         uint64_t spent;
2299
2300         if (fio_gtod_offload && fio_start_gtod_thread())
2301                 return;
2302
2303         fio_idle_prof_init();
2304
2305         set_sig_handlers();
2306
2307         nr_thread = nr_process = 0;
2308         for_each_td(td, i) {
2309                 if (check_mount_writes(td))
2310                         return;
2311                 if (td->o.use_thread)
2312                         nr_thread++;
2313                 else
2314                         nr_process++;
2315         }
2316
2317         if (output_format & FIO_OUTPUT_NORMAL) {
2318                 struct buf_output out;
2319
2320                 buf_output_init(&out);
2321                 __log_buf(&out, "Starting ");
2322                 if (nr_thread)
2323                         __log_buf(&out, "%d thread%s", nr_thread,
2324                                                 nr_thread > 1 ? "s" : "");
2325                 if (nr_process) {
2326                         if (nr_thread)
2327                                 __log_buf(&out, " and ");
2328                         __log_buf(&out, "%d process%s", nr_process,
2329                                                 nr_process > 1 ? "es" : "");
2330                 }
2331                 __log_buf(&out, "\n");
2332                 log_info_buf(out.buf, out.buflen);
2333                 buf_output_free(&out);
2334         }
2335
2336         todo = thread_number;
2337         nr_running = 0;
2338         nr_started = 0;
2339         m_rate = t_rate = 0;
2340
2341         for_each_td(td, i) {
2342                 print_status_init(td->thread_number - 1);
2343
2344                 if (!td->o.create_serialize)
2345                         continue;
2346
2347                 if (fio_verify_load_state(td))
2348                         goto reap;
2349
2350                 /*
2351                  * do file setup here so it happens sequentially,
2352                  * we don't want X number of threads getting their
2353                  * client data interspersed on disk
2354                  */
2355                 if (setup_files(td)) {
2356 reap:
2357                         exit_value++;
2358                         if (td->error)
2359                                 log_err("fio: pid=%d, err=%d/%s\n",
2360                                         (int) td->pid, td->error, td->verror);
2361                         td_set_runstate(td, TD_REAPED);
2362                         todo--;
2363                 } else {
2364                         struct fio_file *f;
2365                         unsigned int j;
2366
2367                         /*
2368                          * for sharing to work, each job must always open
2369                          * its own files. so close them, if we opened them
2370                          * for creation
2371                          */
2372                         for_each_file(td, f, j) {
2373                                 if (fio_file_open(f))
2374                                         td_io_close_file(td, f);
2375                         }
2376                 }
2377         }
2378
2379         /* start idle threads before io threads start to run */
2380         fio_idle_prof_start();
2381
2382         set_genesis_time();
2383
2384         while (todo) {
2385                 struct thread_data *map[REAL_MAX_JOBS];
2386                 struct timespec this_start;
2387                 int this_jobs = 0, left;
2388                 struct fork_data *fd;
2389
2390                 /*
2391                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2392                  */
2393                 for_each_td(td, i) {
2394                         if (td->runstate != TD_NOT_CREATED)
2395                                 continue;
2396
2397                         /*
2398                          * never got a chance to start, killed by other
2399                          * thread for some reason
2400                          */
2401                         if (td->terminate) {
2402                                 todo--;
2403                                 continue;
2404                         }
2405
2406                         if (td->o.start_delay) {
2407                                 spent = utime_since_genesis();
2408
2409                                 if (td->o.start_delay > spent)
2410                                         continue;
2411                         }
2412
2413                         if (td->o.stonewall && (nr_started || nr_running)) {
2414                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2415                                                         td->o.name);
2416                                 break;
2417                         }
2418
2419                         if (waitee_running(td)) {
2420                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2421                                                 td->o.name, td->o.wait_for);
2422                                 continue;
2423                         }
2424
2425                         init_disk_util(td);
2426
2427                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2428                         td->update_rusage = 0;
2429
2430                         /*
2431                          * Set state to created. Thread will transition
2432                          * to TD_INITIALIZED when it's done setting up.
2433                          */
2434                         td_set_runstate(td, TD_CREATED);
2435                         map[this_jobs++] = td;
2436                         nr_started++;
2437
2438                         fd = calloc(1, sizeof(*fd));
2439                         fd->td = td;
2440                         fd->sk_out = sk_out;
2441
2442                         if (td->o.use_thread) {
2443                                 int ret;
2444
2445                                 dprint(FD_PROCESS, "will pthread_create\n");
2446                                 ret = pthread_create(&td->thread, NULL,
2447                                                         thread_main, fd);
2448                                 if (ret) {
2449                                         log_err("pthread_create: %s\n",
2450                                                         strerror(ret));
2451                                         free(fd);
2452                                         nr_started--;
2453                                         break;
2454                                 }
2455                                 fd = NULL;
2456                                 ret = pthread_detach(td->thread);
2457                                 if (ret)
2458                                         log_err("pthread_detach: %s",
2459                                                         strerror(ret));
2460                         } else {
2461                                 pid_t pid;
2462                                 void *eo;
2463                                 dprint(FD_PROCESS, "will fork\n");
2464                                 eo = td->eo;
2465                                 read_barrier();
2466                                 pid = fork();
2467                                 if (!pid) {
2468                                         int ret;
2469
2470                                         ret = (int)(uintptr_t)thread_main(fd);
2471                                         _exit(ret);
2472                                 } else if (i == fio_debug_jobno)
2473                                         *fio_debug_jobp = pid;
2474                                 free(eo);
2475                                 free(fd);
2476                                 fd = NULL;
2477                         }
2478                         dprint(FD_MUTEX, "wait on startup_sem\n");
2479                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2480                                 log_err("fio: job startup hung? exiting.\n");
2481                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2482                                 fio_abort = true;
2483                                 nr_started--;
2484                                 free(fd);
2485                                 break;
2486                         }
2487                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2488                 }
2489
2490                 /*
2491                  * Wait for the started threads to transition to
2492                  * TD_INITIALIZED.
2493                  */
2494                 fio_gettime(&this_start, NULL);
2495                 left = this_jobs;
2496                 while (left && !fio_abort) {
2497                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2498                                 break;
2499
2500                         do_usleep(100000);
2501
2502                         for (i = 0; i < this_jobs; i++) {
2503                                 td = map[i];
2504                                 if (!td)
2505                                         continue;
2506                                 if (td->runstate == TD_INITIALIZED) {
2507                                         map[i] = NULL;
2508                                         left--;
2509                                 } else if (td->runstate >= TD_EXITED) {
2510                                         map[i] = NULL;
2511                                         left--;
2512                                         todo--;
2513                                         nr_running++; /* work-around... */
2514                                 }
2515                         }
2516                 }
2517
2518                 if (left) {
2519                         log_err("fio: %d job%s failed to start\n", left,
2520                                         left > 1 ? "s" : "");
2521                         for (i = 0; i < this_jobs; i++) {
2522                                 td = map[i];
2523                                 if (!td)
2524                                         continue;
2525                                 kill(td->pid, SIGTERM);
2526                         }
2527                         break;
2528                 }
2529
2530                 /*
2531                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2532                  */
2533                 for_each_td(td, i) {
2534                         if (td->runstate != TD_INITIALIZED)
2535                                 continue;
2536
2537                         if (in_ramp_time(td))
2538                                 td_set_runstate(td, TD_RAMP);
2539                         else
2540                                 td_set_runstate(td, TD_RUNNING);
2541                         nr_running++;
2542                         nr_started--;
2543                         m_rate += ddir_rw_sum(td->o.ratemin);
2544                         t_rate += ddir_rw_sum(td->o.rate);
2545                         todo--;
2546                         fio_sem_up(td->sem);
2547                 }
2548
2549                 reap_threads(&nr_running, &t_rate, &m_rate);
2550
2551                 if (todo)
2552                         do_usleep(100000);
2553         }
2554
2555         while (nr_running) {
2556                 reap_threads(&nr_running, &t_rate, &m_rate);
2557                 do_usleep(10000);
2558         }
2559
2560         fio_idle_prof_stop();
2561
2562         update_io_ticks();
2563 }
2564
2565 static void free_disk_util(void)
2566 {
2567         disk_util_prune_entries();
2568         helper_thread_destroy();
2569 }
2570
2571 int fio_backend(struct sk_out *sk_out)
2572 {
2573         struct thread_data *td;
2574         int i;
2575
2576         if (exec_profile) {
2577                 if (load_profile(exec_profile))
2578                         return 1;
2579                 free(exec_profile);
2580                 exec_profile = NULL;
2581         }
2582         if (!thread_number)
2583                 return 0;
2584
2585         if (write_bw_log) {
2586                 struct log_params p = {
2587                         .log_type = IO_LOG_TYPE_BW,
2588                 };
2589
2590                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2591                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2592                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2593         }
2594
2595         if (init_global_dedupe_working_set_seeds()) {
2596                 log_err("fio: failed to initialize global dedupe working set\n");
2597                 return 1;
2598         }
2599
2600         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2601         if (!sk_out)
2602                 is_local_backend = true;
2603         if (startup_sem == NULL)
2604                 return 1;
2605
2606         set_genesis_time();
2607         stat_init();
2608         if (helper_thread_create(startup_sem, sk_out))
2609                 log_err("fio: failed to create helper thread\n");
2610
2611         cgroup_list = smalloc(sizeof(*cgroup_list));
2612         if (cgroup_list)
2613                 INIT_FLIST_HEAD(cgroup_list);
2614
2615         run_threads(sk_out);
2616
2617         helper_thread_exit();
2618
2619         if (!fio_abort) {
2620                 __show_run_stats();
2621                 if (write_bw_log) {
2622                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2623                                 struct io_log *log = agg_io_log[i];
2624
2625                                 flush_log(log, false);
2626                                 free_log(log);
2627                         }
2628                 }
2629         }
2630
2631         for_each_td(td, i) {
2632                 struct thread_stat *ts = &td->ts;
2633
2634                 free_clat_prio_stats(ts);
2635                 steadystate_free(td);
2636                 fio_options_free(td);
2637                 fio_dump_options_free(td);
2638                 if (td->rusage_sem) {
2639                         fio_sem_remove(td->rusage_sem);
2640                         td->rusage_sem = NULL;
2641                 }
2642                 fio_sem_remove(td->sem);
2643                 td->sem = NULL;
2644         }
2645
2646         free_disk_util();
2647         if (cgroup_list) {
2648                 cgroup_kill(cgroup_list);
2649                 sfree(cgroup_list);
2650         }
2651
2652         fio_sem_remove(startup_sem);
2653         stat_exit();
2654         return exit_value;
2655 }