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