backend.c

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