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) {
 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 {
 863         unsigned long long b;
 864         uint64_t total;
 865         int left;
 866
 867         b = ddir_rw_sum(td->io_blocks);
 868         if (b % td->o.thinktime_blocks)
 869                 return;
 870
 871         io_u_quiesce(td);
 872
 873         total = 0;
 874         if (td->o.thinktime_spin)
 875                 total = usec_spin(td->o.thinktime_spin);
 876
 877         left = td->o.thinktime - total;
 878         if (left)
 879                 total += usec_sleep(td, left);
 880
 881         /*
 882          * If we're ignoring thinktime for the rate, add the number of bytes
 883          * we would have done while sleeping, minus one block to ensure we
 884          * start issuing immediately after the sleep.
 885          */
 886         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 887                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 888                 uint64_t bs = td->o.min_bs[ddir];
 889                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 890                 uint64_t over;
 891
 892                 if (usperop <= total)
 893                         over = bs;
 894                 else
 895                         over = (usperop - total) / usperop * -bs;
 896
 897                 td->rate_io_issue_bytes[ddir] += (missed - over);
 898                 /* adjust for rate_process=poisson */
 899                 td->last_usec[ddir] += total;
 900         }
 901 }
 902
 903 /*
 904  * Main IO worker function. It retrieves io_u's to process and queues
 905  * and reaps them, checking for rate and errors along the way.
 906  *
 907  * Returns number of bytes written and trimmed.
 908  */
 909 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 910 {
 911         unsigned int i;
 912         int ret = 0;
 913         uint64_t total_bytes, bytes_issued = 0;
 914
 915         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 916                 bytes_done[i] = td->bytes_done[i];
 917
 918         if (in_ramp_time(td))
 919                 td_set_runstate(td, TD_RAMP);
 920         else
 921                 td_set_runstate(td, TD_RUNNING);
 922
 923         lat_target_init(td);
 924
 925         total_bytes = td->o.size;
 926         /*
 927         * Allow random overwrite workloads to write up to io_size
 928         * before starting verification phase as 'size' doesn't apply.
 929         */
 930         if (td_write(td) && td_random(td) && td->o.norandommap)
 931                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 932         /*
 933          * If verify_backlog is enabled, we'll run the verify in this
 934          * handler as well. For that case, we may need up to twice the
 935          * amount of bytes.
 936          */
 937         if (td->o.verify != VERIFY_NONE &&
 938            (td_write(td) && td->o.verify_backlog))
 939                 total_bytes += td->o.size;
 940
 941         /* In trimwrite mode, each byte is trimmed and then written, so
 942          * allow total_bytes to be twice as big */
 943         if (td_trimwrite(td))
 944                 total_bytes += td->total_io_size;
 945
 946         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 947                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 948                 td->o.time_based) {
 949                 struct timespec comp_time;
 950                 struct io_u *io_u;
 951                 int full;
 952                 enum fio_ddir ddir;
 953
 954                 check_update_rusage(td);
 955
 956                 if (td->terminate || td->done)
 957                         break;
 958
 959                 update_ts_cache(td);
 960
 961                 if (runtime_exceeded(td, &td->ts_cache)) {
 962                         __update_ts_cache(td);
 963                         if (runtime_exceeded(td, &td->ts_cache)) {
 964                                 fio_mark_td_terminate(td);
 965                                 break;
 966                         }
 967                 }
 968
 969                 if (flow_threshold_exceeded(td))
 970                         continue;
 971
 972                 /*
 973                  * Break if we exceeded the bytes. The exception is time
 974                  * based runs, but we still need to break out of the loop
 975                  * for those to run verification, if enabled.
 976                  * Jobs read from iolog do not use this stop condition.
 977                  */
 978                 if (bytes_issued >= total_bytes &&
 979                     !td->o.read_iolog_file &&
 980                     (!td->o.time_based ||
 981                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 982                         break;
 983
 984                 io_u = get_io_u(td);
 985                 if (IS_ERR_OR_NULL(io_u)) {
 986                         int err = PTR_ERR(io_u);
 987
 988                         io_u = NULL;
 989                         ddir = DDIR_INVAL;
 990                         if (err == -EBUSY) {
 991                                 ret = FIO_Q_BUSY;
 992                                 goto reap;
 993                         }
 994                         if (td->o.latency_target)
 995                                 goto reap;
 996                         break;
 997                 }
 998
 999                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1000                         populate_verify_io_u(td, io_u);
1001
1002                 ddir = io_u->ddir;
1003
1004                 /*
1005                  * Add verification end_io handler if:
1006                  *      - Asked to verify (!td_rw(td))
1007                  *      - Or the io_u is from our verify list (mixed write/ver)
1008                  */
1009                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1010                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1011
1012                         if (verify_state_should_stop(td, io_u)) {
1013                                 put_io_u(td, io_u);
1014                                 break;
1015                         }
1016
1017                         if (td->o.verify_async)
1018                                 io_u->end_io = verify_io_u_async;
1019                         else
1020                                 io_u->end_io = verify_io_u;
1021                         td_set_runstate(td, TD_VERIFYING);
1022                 } else if (in_ramp_time(td))
1023                         td_set_runstate(td, TD_RAMP);
1024                 else
1025                         td_set_runstate(td, TD_RUNNING);
1026
1027                 /*
1028                  * Always log IO before it's issued, so we know the specific
1029                  * order of it. The logged unit will track when the IO has
1030                  * completed.
1031                  */
1032                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1033                     td->o.do_verify &&
1034                     td->o.verify != VERIFY_NONE &&
1035                     !td->o.experimental_verify)
1036                         log_io_piece(td, io_u);
1037
1038                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1039                         const unsigned long long blen = io_u->xfer_buflen;
1040                         const enum fio_ddir __ddir = acct_ddir(io_u);
1041
1042                         if (td->error)
1043                                 break;
1044
1045                         workqueue_enqueue(&td->io_wq, &io_u->work);
1046                         ret = FIO_Q_QUEUED;
1047
1048                         if (ddir_rw(__ddir)) {
1049                                 td->io_issues[__ddir]++;
1050                                 td->io_issue_bytes[__ddir] += blen;
1051                                 td->rate_io_issue_bytes[__ddir] += blen;
1052                         }
1053
1054                         if (should_check_rate(td))
1055                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1056
1057                 } else {
1058                         ret = io_u_submit(td, io_u);
1059
1060                         if (should_check_rate(td))
1061                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1062
1063                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1064                                 break;
1065
1066                         /*
1067                          * See if we need to complete some commands. Note that
1068                          * we can get BUSY even without IO queued, if the
1069                          * system is resource starved.
1070                          */
1071 reap:
1072                         full = queue_full(td) ||
1073                                 (ret == FIO_Q_BUSY && td->cur_depth);
1074                         if (full || io_in_polling(td))
1075                                 ret = wait_for_completions(td, &comp_time);
1076                 }
1077                 if (ret < 0)
1078                         break;
1079                 if (!ddir_rw_sum(td->bytes_done) &&
1080                     !td_ioengine_flagged(td, FIO_NOIO))
1081                         continue;
1082
1083                 if (!in_ramp_time(td) && should_check_rate(td)) {
1084                         if (check_min_rate(td, &comp_time)) {
1085                                 if (exitall_on_terminate || td->o.exitall_error)
1086                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1087                                 td_verror(td, EIO, "check_min_rate");
1088                                 break;
1089                         }
1090                 }
1091                 if (!in_ramp_time(td) && td->o.latency_target)
1092                         lat_target_check(td);
1093
1094                 if (ddir_rw(ddir) && td->o.thinktime)
1095                         handle_thinktime(td, ddir);
1096         }
1097
1098         check_update_rusage(td);
1099
1100         if (td->trim_entries)
1101                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1102
1103         if (td->o.fill_device && td->error == ENOSPC) {
1104                 td->error = 0;
1105                 fio_mark_td_terminate(td);
1106         }
1107         if (!td->error) {
1108                 struct fio_file *f;
1109
1110                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1111                         workqueue_flush(&td->io_wq);
1112                         i = 0;
1113                 } else
1114                         i = td->cur_depth;
1115
1116                 if (i) {
1117                         ret = io_u_queued_complete(td, i);
1118                         if (td->o.fill_device && td->error == ENOSPC)
1119                                 td->error = 0;
1120                 }
1121
1122                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1123                         td_set_runstate(td, TD_FSYNCING);
1124
1125                         for_each_file(td, f, i) {
1126                                 if (!fio_file_fsync(td, f))
1127                                         continue;
1128
1129                                 log_err("fio: end_fsync failed for file %s\n",
1130                                                                 f->file_name);
1131                         }
1132                 }
1133         } else
1134                 cleanup_pending_aio(td);
1135
1136         /*
1137          * stop job if we failed doing any IO
1138          */
1139         if (!ddir_rw_sum(td->this_io_bytes))
1140                 td->done = 1;
1141
1142         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1143                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1144 }
1145
1146 static void free_file_completion_logging(struct thread_data *td)
1147 {
1148         struct fio_file *f;
1149         unsigned int i;
1150
1151         for_each_file(td, f, i) {
1152                 if (!f->last_write_comp)
1153                         break;
1154                 sfree(f->last_write_comp);
1155         }
1156 }
1157
1158 static int init_file_completion_logging(struct thread_data *td,
1159                                         unsigned int depth)
1160 {
1161         struct fio_file *f;
1162         unsigned int i;
1163
1164         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1165                 return 0;
1166
1167         for_each_file(td, f, i) {
1168                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1169                 if (!f->last_write_comp)
1170                         goto cleanup;
1171         }
1172
1173         return 0;
1174
1175 cleanup:
1176         free_file_completion_logging(td);
1177         log_err("fio: failed to alloc write comp data\n");
1178         return 1;
1179 }
1180
1181 static void cleanup_io_u(struct thread_data *td)
1182 {
1183         struct io_u *io_u;
1184
1185         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1186
1187                 if (td->io_ops->io_u_free)
1188                         td->io_ops->io_u_free(td, io_u);
1189
1190                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1191         }
1192
1193         free_io_mem(td);
1194
1195         io_u_rexit(&td->io_u_requeues);
1196         io_u_qexit(&td->io_u_freelist, false);
1197         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1198
1199         free_file_completion_logging(td);
1200 }
1201
1202 static int init_io_u(struct thread_data *td)
1203 {
1204         struct io_u *io_u;
1205         int cl_align, i, max_units;
1206         int err;
1207
1208         max_units = td->o.iodepth;
1209
1210         err = 0;
1211         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1212         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1213         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1214
1215         if (err) {
1216                 log_err("fio: failed setting up IO queues\n");
1217                 return 1;
1218         }
1219
1220         cl_align = os_cache_line_size();
1221
1222         for (i = 0; i < max_units; i++) {
1223                 void *ptr;
1224
1225                 if (td->terminate)
1226                         return 1;
1227
1228                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1229                 if (!ptr) {
1230                         log_err("fio: unable to allocate aligned memory\n");
1231                         return 1;
1232                 }
1233
1234                 io_u = ptr;
1235                 memset(io_u, 0, sizeof(*io_u));
1236                 INIT_FLIST_HEAD(&io_u->verify_list);
1237                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1238
1239                 io_u->index = i;
1240                 io_u->flags = IO_U_F_FREE;
1241                 io_u_qpush(&td->io_u_freelist, io_u);
1242
1243                 /*
1244                  * io_u never leaves this stack, used for iteration of all
1245                  * io_u buffers.
1246                  */
1247                 io_u_qpush(&td->io_u_all, io_u);
1248
1249                 if (td->io_ops->io_u_init) {
1250                         int ret = td->io_ops->io_u_init(td, io_u);
1251
1252                         if (ret) {
1253                                 log_err("fio: failed to init engine data: %d\n", ret);
1254                                 return 1;
1255                         }
1256                 }
1257         }
1258
1259         init_io_u_buffers(td);
1260
1261         if (init_file_completion_logging(td, max_units))
1262                 return 1;
1263
1264         return 0;
1265 }
1266
1267 int init_io_u_buffers(struct thread_data *td)
1268 {
1269         struct io_u *io_u;
1270         unsigned long long max_bs, min_write;
1271         int i, max_units;
1272         int data_xfer = 1;
1273         char *p;
1274
1275         max_units = td->o.iodepth;
1276         max_bs = td_max_bs(td);
1277         min_write = td->o.min_bs[DDIR_WRITE];
1278         td->orig_buffer_size = (unsigned long long) max_bs
1279                                         * (unsigned long long) max_units;
1280
1281         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1282                 data_xfer = 0;
1283
1284         /*
1285          * if we may later need to do address alignment, then add any
1286          * possible adjustment here so that we don't cause a buffer
1287          * overflow later. this adjustment may be too much if we get
1288          * lucky and the allocator gives us an aligned address.
1289          */
1290         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1291             td_ioengine_flagged(td, FIO_RAWIO))
1292                 td->orig_buffer_size += page_mask + td->o.mem_align;
1293
1294         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1295                 unsigned long long bs;
1296
1297                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1298                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1299         }
1300
1301         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1302                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1303                 return 1;
1304         }
1305
1306         if (data_xfer && allocate_io_mem(td))
1307                 return 1;
1308
1309         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1310             td_ioengine_flagged(td, FIO_RAWIO))
1311                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1312         else
1313                 p = td->orig_buffer;
1314
1315         for (i = 0; i < max_units; i++) {
1316                 io_u = td->io_u_all.io_us[i];
1317                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1318
1319                 if (data_xfer) {
1320                         io_u->buf = p;
1321                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1322
1323                         if (td_write(td))
1324                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1325                         if (td_write(td) && td->o.verify_pattern_bytes) {
1326                                 /*
1327                                  * Fill the buffer with the pattern if we are
1328                                  * going to be doing writes.
1329                                  */
1330                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1331                         }
1332                 }
1333                 p += max_bs;
1334         }
1335
1336         return 0;
1337 }
1338
1339 /*
1340  * This function is Linux specific.
1341  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1342  */
1343 static int switch_ioscheduler(struct thread_data *td)
1344 {
1345 #ifdef FIO_HAVE_IOSCHED_SWITCH
1346         char tmp[256], tmp2[128], *p;
1347         FILE *f;
1348         int ret;
1349
1350         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1351                 return 0;
1352
1353         assert(td->files && td->files[0]);
1354         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1355
1356         f = fopen(tmp, "r+");
1357         if (!f) {
1358                 if (errno == ENOENT) {
1359                         log_err("fio: os or kernel doesn't support IO scheduler"
1360                                 " switching\n");
1361                         return 0;
1362                 }
1363                 td_verror(td, errno, "fopen iosched");
1364                 return 1;
1365         }
1366
1367         /*
1368          * Set io scheduler.
1369          */
1370         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1371         if (ferror(f) || ret != 1) {
1372                 td_verror(td, errno, "fwrite");
1373                 fclose(f);
1374                 return 1;
1375         }
1376
1377         rewind(f);
1378
1379         /*
1380          * Read back and check that the selected scheduler is now the default.
1381          */
1382         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1383         if (ferror(f) || ret < 0) {
1384                 td_verror(td, errno, "fread");
1385                 fclose(f);
1386                 return 1;
1387         }
1388         tmp[ret] = '\0';
1389         /*
1390          * either a list of io schedulers or "none\n" is expected. Strip the
1391          * trailing newline.
1392          */
1393         p = tmp;
1394         strsep(&p, "\n");
1395
1396         /*
1397          * Write to "none" entry doesn't fail, so check the result here.
1398          */
1399         if (!strcmp(tmp, "none")) {
1400                 log_err("fio: io scheduler is not tunable\n");
1401                 fclose(f);
1402                 return 0;
1403         }
1404
1405         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1406         if (!strstr(tmp, tmp2)) {
1407                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1408                 td_verror(td, EINVAL, "iosched_switch");
1409                 fclose(f);
1410                 return 1;
1411         }
1412
1413         fclose(f);
1414         return 0;
1415 #else
1416         return 0;
1417 #endif
1418 }
1419
1420 static bool keep_running(struct thread_data *td)
1421 {
1422         unsigned long long limit;
1423
1424         if (td->done)
1425                 return false;
1426         if (td->terminate)
1427                 return false;
1428         if (td->o.time_based)
1429                 return true;
1430         if (td->o.loops) {
1431                 td->o.loops--;
1432                 return true;
1433         }
1434         if (exceeds_number_ios(td))
1435                 return false;
1436
1437         if (td->o.io_size)
1438                 limit = td->o.io_size;
1439         else
1440                 limit = td->o.size;
1441
1442         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1443                 uint64_t diff;
1444
1445                 /*
1446                  * If the difference is less than the maximum IO size, we
1447                  * are done.
1448                  */
1449                 diff = limit - ddir_rw_sum(td->io_bytes);
1450                 if (diff < td_max_bs(td))
1451                         return false;
1452
1453                 if (fio_files_done(td) && !td->o.io_size)
1454                         return false;
1455
1456                 return true;
1457         }
1458
1459         return false;
1460 }
1461
1462 static int exec_string(struct thread_options *o, const char *string,
1463                        const char *mode)
1464 {
1465         int ret;
1466         char *str;
1467
1468         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1469                 return -1;
1470
1471         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1472                  o->name, mode);
1473         ret = system(str);
1474         if (ret == -1)
1475                 log_err("fio: exec of cmd <%s> failed\n", str);
1476
1477         free(str);
1478         return ret;
1479 }
1480
1481 /*
1482  * Dry run to compute correct state of numberio for verification.
1483  */
1484 static uint64_t do_dry_run(struct thread_data *td)
1485 {
1486         td_set_runstate(td, TD_RUNNING);
1487
1488         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1489                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1490                 struct io_u *io_u;
1491                 int ret;
1492
1493                 if (td->terminate || td->done)
1494                         break;
1495
1496                 io_u = get_io_u(td);
1497                 if (IS_ERR_OR_NULL(io_u))
1498                         break;
1499
1500                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1501                 io_u->error = 0;
1502                 io_u->resid = 0;
1503                 if (ddir_rw(acct_ddir(io_u)))
1504                         td->io_issues[acct_ddir(io_u)]++;
1505                 if (ddir_rw(io_u->ddir)) {
1506                         io_u_mark_depth(td, 1);
1507                         td->ts.total_io_u[io_u->ddir]++;
1508                 }
1509
1510                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1511                     td->o.do_verify &&
1512                     td->o.verify != VERIFY_NONE &&
1513                     !td->o.experimental_verify)
1514                         log_io_piece(td, io_u);
1515
1516                 ret = io_u_sync_complete(td, io_u);
1517                 (void) ret;
1518         }
1519
1520         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1521 }
1522
1523 struct fork_data {
1524         struct thread_data *td;
1525         struct sk_out *sk_out;
1526 };
1527
1528 /*
1529  * Entry point for the thread based jobs. The process based jobs end up
1530  * here as well, after a little setup.
1531  */
1532 static void *thread_main(void *data)
1533 {
1534         struct fork_data *fd = data;
1535         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1536         struct thread_data *td = fd->td;
1537         struct thread_options *o = &td->o;
1538         struct sk_out *sk_out = fd->sk_out;
1539         uint64_t bytes_done[DDIR_RWDIR_CNT];
1540         int deadlock_loop_cnt;
1541         bool clear_state;
1542         int res, ret;
1543
1544         sk_out_assign(sk_out);
1545         free(fd);
1546
1547         if (!o->use_thread) {
1548                 setsid();
1549                 td->pid = getpid();
1550         } else
1551                 td->pid = gettid();
1552
1553         fio_local_clock_init();
1554
1555         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1556
1557         if (is_backend)
1558                 fio_server_send_start(td);
1559
1560         INIT_FLIST_HEAD(&td->io_log_list);
1561         INIT_FLIST_HEAD(&td->io_hist_list);
1562         INIT_FLIST_HEAD(&td->verify_list);
1563         INIT_FLIST_HEAD(&td->trim_list);
1564         td->io_hist_tree = RB_ROOT;
1565
1566         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1567         if (ret) {
1568                 td_verror(td, ret, "mutex_cond_init_pshared");
1569                 goto err;
1570         }
1571         ret = cond_init_pshared(&td->verify_cond);
1572         if (ret) {
1573                 td_verror(td, ret, "mutex_cond_pshared");
1574                 goto err;
1575         }
1576
1577         td_set_runstate(td, TD_INITIALIZED);
1578         dprint(FD_MUTEX, "up startup_sem\n");
1579         fio_sem_up(startup_sem);
1580         dprint(FD_MUTEX, "wait on td->sem\n");
1581         fio_sem_down(td->sem);
1582         dprint(FD_MUTEX, "done waiting on td->sem\n");
1583
1584         /*
1585          * A new gid requires privilege, so we need to do this before setting
1586          * the uid.
1587          */
1588         if (o->gid != -1U && setgid(o->gid)) {
1589                 td_verror(td, errno, "setgid");
1590                 goto err;
1591         }
1592         if (o->uid != -1U && setuid(o->uid)) {
1593                 td_verror(td, errno, "setuid");
1594                 goto err;
1595         }
1596
1597         td_zone_gen_index(td);
1598
1599         /*
1600          * Do this early, we don't want the compress threads to be limited
1601          * to the same CPUs as the IO workers. So do this before we set
1602          * any potential CPU affinity
1603          */
1604         if (iolog_compress_init(td, sk_out))
1605                 goto err;
1606
1607         /*
1608          * If we have a gettimeofday() thread, make sure we exclude that
1609          * thread from this job
1610          */
1611         if (o->gtod_cpu)
1612                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1613
1614         /*
1615          * Set affinity first, in case it has an impact on the memory
1616          * allocations.
1617          */
1618         if (fio_option_is_set(o, cpumask)) {
1619                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1620                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1621                         if (!ret) {
1622                                 log_err("fio: no CPUs set\n");
1623                                 log_err("fio: Try increasing number of available CPUs\n");
1624                                 td_verror(td, EINVAL, "cpus_split");
1625                                 goto err;
1626                         }
1627                 }
1628                 ret = fio_setaffinity(td->pid, o->cpumask);
1629                 if (ret == -1) {
1630                         td_verror(td, errno, "cpu_set_affinity");
1631                         goto err;
1632                 }
1633         }
1634
1635 #ifdef CONFIG_LIBNUMA
1636         /* numa node setup */
1637         if (fio_option_is_set(o, numa_cpunodes) ||
1638             fio_option_is_set(o, numa_memnodes)) {
1639                 struct bitmask *mask;
1640
1641                 if (numa_available() < 0) {
1642                         td_verror(td, errno, "Does not support NUMA API\n");
1643                         goto err;
1644                 }
1645
1646                 if (fio_option_is_set(o, numa_cpunodes)) {
1647                         mask = numa_parse_nodestring(o->numa_cpunodes);
1648                         ret = numa_run_on_node_mask(mask);
1649                         numa_free_nodemask(mask);
1650                         if (ret == -1) {
1651                                 td_verror(td, errno, \
1652                                         "numa_run_on_node_mask failed\n");
1653                                 goto err;
1654                         }
1655                 }
1656
1657                 if (fio_option_is_set(o, numa_memnodes)) {
1658                         mask = NULL;
1659                         if (o->numa_memnodes)
1660                                 mask = numa_parse_nodestring(o->numa_memnodes);
1661
1662                         switch (o->numa_mem_mode) {
1663                         case MPOL_INTERLEAVE:
1664                                 numa_set_interleave_mask(mask);
1665                                 break;
1666                         case MPOL_BIND:
1667                                 numa_set_membind(mask);
1668                                 break;
1669                         case MPOL_LOCAL:
1670                                 numa_set_localalloc();
1671                                 break;
1672                         case MPOL_PREFERRED:
1673                                 numa_set_preferred(o->numa_mem_prefer_node);
1674                                 break;
1675                         case MPOL_DEFAULT:
1676                         default:
1677                                 break;
1678                         }
1679
1680                         if (mask)
1681                                 numa_free_nodemask(mask);
1682
1683                 }
1684         }
1685 #endif
1686
1687         if (fio_pin_memory(td))
1688                 goto err;
1689
1690         /*
1691          * May alter parameters that init_io_u() will use, so we need to
1692          * do this first.
1693          */
1694         if (!init_iolog(td))
1695                 goto err;
1696
1697         if (td_io_init(td))
1698                 goto err;
1699
1700         if (init_io_u(td))
1701                 goto err;
1702
1703         if (td->io_ops->post_init && td->io_ops->post_init(td))
1704                 goto err;
1705
1706         if (o->verify_async && verify_async_init(td))
1707                 goto err;
1708
1709         if (fio_option_is_set(o, ioprio) ||
1710             fio_option_is_set(o, ioprio_class)) {
1711                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1712                 if (ret == -1) {
1713                         td_verror(td, errno, "ioprio_set");
1714                         goto err;
1715                 }
1716         }
1717
1718         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1719                 goto err;
1720
1721         errno = 0;
1722         if (nice(o->nice) == -1 && errno != 0) {
1723                 td_verror(td, errno, "nice");
1724                 goto err;
1725         }
1726
1727         if (o->ioscheduler && switch_ioscheduler(td))
1728                 goto err;
1729
1730         if (!o->create_serialize && setup_files(td))
1731                 goto err;
1732
1733         if (!init_random_map(td))
1734                 goto err;
1735
1736         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1737                 goto err;
1738
1739         if (o->pre_read && !pre_read_files(td))
1740                 goto err;
1741
1742         fio_verify_init(td);
1743
1744         if (rate_submit_init(td, sk_out))
1745                 goto err;
1746
1747         set_epoch_time(td, o->log_unix_epoch);
1748         fio_getrusage(&td->ru_start);
1749         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1750         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1751         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1752
1753         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1754                         o->ratemin[DDIR_TRIM]) {
1755                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1756                                         sizeof(td->bw_sample_time));
1757                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1758                                         sizeof(td->bw_sample_time));
1759                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1760                                         sizeof(td->bw_sample_time));
1761         }
1762
1763         memset(bytes_done, 0, sizeof(bytes_done));
1764         clear_state = false;
1765
1766         while (keep_running(td)) {
1767                 uint64_t verify_bytes;
1768
1769                 fio_gettime(&td->start, NULL);
1770                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1771
1772                 if (clear_state) {
1773                         clear_io_state(td, 0);
1774
1775                         if (o->unlink_each_loop && unlink_all_files(td))
1776                                 break;
1777                 }
1778
1779                 prune_io_piece_log(td);
1780
1781                 if (td->o.verify_only && td_write(td))
1782                         verify_bytes = do_dry_run(td);
1783                 else {
1784                         do_io(td, bytes_done);
1785
1786                         if (!ddir_rw_sum(bytes_done)) {
1787                                 fio_mark_td_terminate(td);
1788                                 verify_bytes = 0;
1789                         } else {
1790                                 verify_bytes = bytes_done[DDIR_WRITE] +
1791                                                 bytes_done[DDIR_TRIM];
1792                         }
1793                 }
1794
1795                 /*
1796                  * If we took too long to shut down, the main thread could
1797                  * already consider us reaped/exited. If that happens, break
1798                  * out and clean up.
1799                  */
1800                 if (td->runstate >= TD_EXITED)
1801                         break;
1802
1803                 clear_state = true;
1804
1805                 /*
1806                  * Make sure we've successfully updated the rusage stats
1807                  * before waiting on the stat mutex. Otherwise we could have
1808                  * the stat thread holding stat mutex and waiting for
1809                  * the rusage_sem, which would never get upped because
1810                  * this thread is waiting for the stat mutex.
1811                  */
1812                 deadlock_loop_cnt = 0;
1813                 do {
1814                         check_update_rusage(td);
1815                         if (!fio_sem_down_trylock(stat_sem))
1816                                 break;
1817                         usleep(1000);
1818                         if (deadlock_loop_cnt++ > 5000) {
1819                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1820                                 td->error = EDEADLK;
1821                                 goto err;
1822                         }
1823                 } while (1);
1824
1825                 if (td_read(td) && td->io_bytes[DDIR_READ])
1826                         update_runtime(td, elapsed_us, DDIR_READ);
1827                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1828                         update_runtime(td, elapsed_us, DDIR_WRITE);
1829                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1830                         update_runtime(td, elapsed_us, DDIR_TRIM);
1831                 fio_gettime(&td->start, NULL);
1832                 fio_sem_up(stat_sem);
1833
1834                 if (td->error || td->terminate)
1835                         break;
1836
1837                 if (!o->do_verify ||
1838                     o->verify == VERIFY_NONE ||
1839                     td_ioengine_flagged(td, FIO_UNIDIR))
1840                         continue;
1841
1842                 clear_io_state(td, 0);
1843
1844                 fio_gettime(&td->start, NULL);
1845
1846                 do_verify(td, verify_bytes);
1847
1848                 /*
1849                  * See comment further up for why this is done here.
1850                  */
1851                 check_update_rusage(td);
1852
1853                 fio_sem_down(stat_sem);
1854                 update_runtime(td, elapsed_us, DDIR_READ);
1855                 fio_gettime(&td->start, NULL);
1856                 fio_sem_up(stat_sem);
1857
1858                 if (td->error || td->terminate)
1859                         break;
1860         }
1861
1862         /*
1863          * Acquire this lock if we were doing overlap checking in
1864          * offload mode so that we don't clean up this job while
1865          * another thread is checking its io_u's for overlap
1866          */
1867         if (td_offload_overlap(td)) {
1868                 int res = pthread_mutex_lock(&overlap_check);
1869                 assert(res == 0);
1870         }
1871         td_set_runstate(td, TD_FINISHING);
1872         if (td_offload_overlap(td)) {
1873                 res = pthread_mutex_unlock(&overlap_check);
1874                 assert(res == 0);
1875         }
1876
1877         update_rusage_stat(td);
1878         td->ts.total_run_time = mtime_since_now(&td->epoch);
1879         for_each_rw_ddir(ddir) {
1880                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1881         }
1882
1883         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1884             (td->o.verify != VERIFY_NONE && td_write(td)))
1885                 verify_save_state(td->thread_number);
1886
1887         fio_unpin_memory(td);
1888
1889         td_writeout_logs(td, true);
1890
1891         iolog_compress_exit(td);
1892         rate_submit_exit(td);
1893
1894         if (o->exec_postrun)
1895                 exec_string(o, o->exec_postrun, "postrun");
1896
1897         if (exitall_on_terminate || (o->exitall_error && td->error))
1898                 fio_terminate_threads(td->groupid, td->o.exit_what);
1899
1900 err:
1901         if (td->error)
1902                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1903                                                         td->verror);
1904
1905         if (o->verify_async)
1906                 verify_async_exit(td);
1907
1908         close_and_free_files(td);
1909         cleanup_io_u(td);
1910         close_ioengine(td);
1911         cgroup_shutdown(td, cgroup_mnt);
1912         verify_free_state(td);
1913         td_zone_free_index(td);
1914
1915         if (fio_option_is_set(o, cpumask)) {
1916                 ret = fio_cpuset_exit(&o->cpumask);
1917                 if (ret)
1918                         td_verror(td, ret, "fio_cpuset_exit");
1919         }
1920
1921         /*
1922          * do this very late, it will log file closing as well
1923          */
1924         if (o->write_iolog_file)
1925                 write_iolog_close(td);
1926         if (td->io_log_rfile)
1927                 fclose(td->io_log_rfile);
1928
1929         td_set_runstate(td, TD_EXITED);
1930
1931         /*
1932          * Do this last after setting our runstate to exited, so we
1933          * know that the stat thread is signaled.
1934          */
1935         check_update_rusage(td);
1936
1937         sk_out_drop();
1938         return (void *) (uintptr_t) td->error;
1939 }
1940
1941 /*
1942  * Run over the job map and reap the threads that have exited, if any.
1943  */
1944 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1945                          uint64_t *m_rate)
1946 {
1947         struct thread_data *td;
1948         unsigned int cputhreads, realthreads, pending;
1949         int i, status, ret;
1950
1951         /*
1952          * reap exited threads (TD_EXITED -> TD_REAPED)
1953          */
1954         realthreads = pending = cputhreads = 0;
1955         for_each_td(td, i) {
1956                 int flags = 0;
1957
1958                  if (!strcmp(td->o.ioengine, "cpuio"))
1959                         cputhreads++;
1960                 else
1961                         realthreads++;
1962
1963                 if (!td->pid) {
1964                         pending++;
1965                         continue;
1966                 }
1967                 if (td->runstate == TD_REAPED)
1968                         continue;
1969                 if (td->o.use_thread) {
1970                         if (td->runstate == TD_EXITED) {
1971                                 td_set_runstate(td, TD_REAPED);
1972                                 goto reaped;
1973                         }
1974                         continue;
1975                 }
1976
1977                 flags = WNOHANG;
1978                 if (td->runstate == TD_EXITED)
1979                         flags = 0;
1980
1981                 /*
1982                  * check if someone quit or got killed in an unusual way
1983                  */
1984                 ret = waitpid(td->pid, &status, flags);
1985                 if (ret < 0) {
1986                         if (errno == ECHILD) {
1987                                 log_err("fio: pid=%d disappeared %d\n",
1988                                                 (int) td->pid, td->runstate);
1989                                 td->sig = ECHILD;
1990                                 td_set_runstate(td, TD_REAPED);
1991                                 goto reaped;
1992                         }
1993                         perror("waitpid");
1994                 } else if (ret == td->pid) {
1995                         if (WIFSIGNALED(status)) {
1996                                 int sig = WTERMSIG(status);
1997
1998                                 if (sig != SIGTERM && sig != SIGUSR2)
1999                                         log_err("fio: pid=%d, got signal=%d\n",
2000                                                         (int) td->pid, sig);
2001                                 td->sig = sig;
2002                                 td_set_runstate(td, TD_REAPED);
2003                                 goto reaped;
2004                         }
2005                         if (WIFEXITED(status)) {
2006                                 if (WEXITSTATUS(status) && !td->error)
2007                                         td->error = WEXITSTATUS(status);
2008
2009                                 td_set_runstate(td, TD_REAPED);
2010                                 goto reaped;
2011                         }
2012                 }
2013
2014                 /*
2015                  * If the job is stuck, do a forceful timeout of it and
2016                  * move on.
2017                  */
2018                 if (td->terminate &&
2019                     td->runstate < TD_FSYNCING &&
2020                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2021                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2022                                 "%lu seconds, it appears to be stuck. Doing "
2023                                 "forceful exit of this job.\n",
2024                                 td->o.name, td->runstate,
2025                                 (unsigned long) time_since_now(&td->terminate_time));
2026                         td_set_runstate(td, TD_REAPED);
2027                         goto reaped;
2028                 }
2029
2030                 /*
2031                  * thread is not dead, continue
2032                  */
2033                 pending++;
2034                 continue;
2035 reaped:
2036                 (*nr_running)--;
2037                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2038                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2039                 if (!td->pid)
2040                         pending--;
2041
2042                 if (td->error)
2043                         exit_value++;
2044
2045                 done_secs += mtime_since_now(&td->epoch) / 1000;
2046                 profile_td_exit(td);
2047                 flow_exit_job(td);
2048         }
2049
2050         if (*nr_running == cputhreads && !pending && realthreads)
2051                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2052 }
2053
2054 static bool __check_trigger_file(void)
2055 {
2056         struct stat sb;
2057
2058         if (!trigger_file)
2059                 return false;
2060
2061         if (stat(trigger_file, &sb))
2062                 return false;
2063
2064         if (unlink(trigger_file) < 0)
2065                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2066                                                         strerror(errno));
2067
2068         return true;
2069 }
2070
2071 static bool trigger_timedout(void)
2072 {
2073         if (trigger_timeout)
2074                 if (time_since_genesis() >= trigger_timeout) {
2075                         trigger_timeout = 0;
2076                         return true;
2077                 }
2078
2079         return false;
2080 }
2081
2082 void exec_trigger(const char *cmd)
2083 {
2084         int ret;
2085
2086         if (!cmd || cmd[0] == '\0')
2087                 return;
2088
2089         ret = system(cmd);
2090         if (ret == -1)
2091                 log_err("fio: failed executing %s trigger\n", cmd);
2092 }
2093
2094 void check_trigger_file(void)
2095 {
2096         if (__check_trigger_file() || trigger_timedout()) {
2097                 if (nr_clients)
2098                         fio_clients_send_trigger(trigger_remote_cmd);
2099                 else {
2100                         verify_save_state(IO_LIST_ALL);
2101                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2102                         exec_trigger(trigger_cmd);
2103                 }
2104         }
2105 }
2106
2107 static int fio_verify_load_state(struct thread_data *td)
2108 {
2109         int ret;
2110
2111         if (!td->o.verify_state)
2112                 return 0;
2113
2114         if (is_backend) {
2115                 void *data;
2116
2117                 ret = fio_server_get_verify_state(td->o.name,
2118                                         td->thread_number - 1, &data);
2119                 if (!ret)
2120                         verify_assign_state(td, data);
2121         } else {
2122                 char prefix[PATH_MAX];
2123
2124                 if (aux_path)
2125                         sprintf(prefix, "%s%clocal", aux_path,
2126                                         FIO_OS_PATH_SEPARATOR);
2127                 else
2128                         strcpy(prefix, "local");
2129                 ret = verify_load_state(td, prefix);
2130         }
2131
2132         return ret;
2133 }
2134
2135 static void do_usleep(unsigned int usecs)
2136 {
2137         check_for_running_stats();
2138         check_trigger_file();
2139         usleep(usecs);
2140 }
2141
2142 static bool check_mount_writes(struct thread_data *td)
2143 {
2144         struct fio_file *f;
2145         unsigned int i;
2146
2147         if (!td_write(td) || td->o.allow_mounted_write)
2148                 return false;
2149
2150         /*
2151          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2152          * are mkfs'd and mounted.
2153          */
2154         for_each_file(td, f, i) {
2155 #ifdef FIO_HAVE_CHARDEV_SIZE
2156                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2157 #else
2158                 if (f->filetype != FIO_TYPE_BLOCK)
2159 #endif
2160                         continue;
2161                 if (device_is_mounted(f->file_name))
2162                         goto mounted;
2163         }
2164
2165         return false;
2166 mounted:
2167         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2168         return true;
2169 }
2170
2171 static bool waitee_running(struct thread_data *me)
2172 {
2173         const char *waitee = me->o.wait_for;
2174         const char *self = me->o.name;
2175         struct thread_data *td;
2176         int i;
2177
2178         if (!waitee)
2179                 return false;
2180
2181         for_each_td(td, i) {
2182                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2183                         continue;
2184
2185                 if (td->runstate < TD_EXITED) {
2186                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2187                                         self, td->o.name,
2188                                         runstate_to_name(td->runstate));
2189                         return true;
2190                 }
2191         }
2192
2193         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2194         return false;
2195 }
2196
2197 /*
2198  * Main function for kicking off and reaping jobs, as needed.
2199  */
2200 static void run_threads(struct sk_out *sk_out)
2201 {
2202         struct thread_data *td;
2203         unsigned int i, todo, nr_running, nr_started;
2204         uint64_t m_rate, t_rate;
2205         uint64_t spent;
2206
2207         if (fio_gtod_offload && fio_start_gtod_thread())
2208                 return;
2209
2210         fio_idle_prof_init();
2211
2212         set_sig_handlers();
2213
2214         nr_thread = nr_process = 0;
2215         for_each_td(td, i) {
2216                 if (check_mount_writes(td))
2217                         return;
2218                 if (td->o.use_thread)
2219                         nr_thread++;
2220                 else
2221                         nr_process++;
2222         }
2223
2224         if (output_format & FIO_OUTPUT_NORMAL) {
2225                 struct buf_output out;
2226
2227                 buf_output_init(&out);
2228                 __log_buf(&out, "Starting ");
2229                 if (nr_thread)
2230                         __log_buf(&out, "%d thread%s", nr_thread,
2231                                                 nr_thread > 1 ? "s" : "");
2232                 if (nr_process) {
2233                         if (nr_thread)
2234                                 __log_buf(&out, " and ");
2235                         __log_buf(&out, "%d process%s", nr_process,
2236                                                 nr_process > 1 ? "es" : "");
2237                 }
2238                 __log_buf(&out, "\n");
2239                 log_info_buf(out.buf, out.buflen);
2240                 buf_output_free(&out);
2241         }
2242
2243         todo = thread_number;
2244         nr_running = 0;
2245         nr_started = 0;
2246         m_rate = t_rate = 0;
2247
2248         for_each_td(td, i) {
2249                 print_status_init(td->thread_number - 1);
2250
2251                 if (!td->o.create_serialize)
2252                         continue;
2253
2254                 if (fio_verify_load_state(td))
2255                         goto reap;
2256
2257                 /*
2258                  * do file setup here so it happens sequentially,
2259                  * we don't want X number of threads getting their
2260                  * client data interspersed on disk
2261                  */
2262                 if (setup_files(td)) {
2263 reap:
2264                         exit_value++;
2265                         if (td->error)
2266                                 log_err("fio: pid=%d, err=%d/%s\n",
2267                                         (int) td->pid, td->error, td->verror);
2268                         td_set_runstate(td, TD_REAPED);
2269                         todo--;
2270                 } else {
2271                         struct fio_file *f;
2272                         unsigned int j;
2273
2274                         /*
2275                          * for sharing to work, each job must always open
2276                          * its own files. so close them, if we opened them
2277                          * for creation
2278                          */
2279                         for_each_file(td, f, j) {
2280                                 if (fio_file_open(f))
2281                                         td_io_close_file(td, f);
2282                         }
2283                 }
2284         }
2285
2286         /* start idle threads before io threads start to run */
2287         fio_idle_prof_start();
2288
2289         set_genesis_time();
2290
2291         while (todo) {
2292                 struct thread_data *map[REAL_MAX_JOBS];
2293                 struct timespec this_start;
2294                 int this_jobs = 0, left;
2295                 struct fork_data *fd;
2296
2297                 /*
2298                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2299                  */
2300                 for_each_td(td, i) {
2301                         if (td->runstate != TD_NOT_CREATED)
2302                                 continue;
2303
2304                         /*
2305                          * never got a chance to start, killed by other
2306                          * thread for some reason
2307                          */
2308                         if (td->terminate) {
2309                                 todo--;
2310                                 continue;
2311                         }
2312
2313                         if (td->o.start_delay) {
2314                                 spent = utime_since_genesis();
2315
2316                                 if (td->o.start_delay > spent)
2317                                         continue;
2318                         }
2319
2320                         if (td->o.stonewall && (nr_started || nr_running)) {
2321                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2322                                                         td->o.name);
2323                                 break;
2324                         }
2325
2326                         if (waitee_running(td)) {
2327                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2328                                                 td->o.name, td->o.wait_for);
2329                                 continue;
2330                         }
2331
2332                         init_disk_util(td);
2333
2334                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2335                         td->update_rusage = 0;
2336
2337                         /*
2338                          * Set state to created. Thread will transition
2339                          * to TD_INITIALIZED when it's done setting up.
2340                          */
2341                         td_set_runstate(td, TD_CREATED);
2342                         map[this_jobs++] = td;
2343                         nr_started++;
2344
2345                         fd = calloc(1, sizeof(*fd));
2346                         fd->td = td;
2347                         fd->sk_out = sk_out;
2348
2349                         if (td->o.use_thread) {
2350                                 int ret;
2351
2352                                 dprint(FD_PROCESS, "will pthread_create\n");
2353                                 ret = pthread_create(&td->thread, NULL,
2354                                                         thread_main, fd);
2355                                 if (ret) {
2356                                         log_err("pthread_create: %s\n",
2357                                                         strerror(ret));
2358                                         free(fd);
2359                                         nr_started--;
2360                                         break;
2361                                 }
2362                                 fd = NULL;
2363                                 ret = pthread_detach(td->thread);
2364                                 if (ret)
2365                                         log_err("pthread_detach: %s",
2366                                                         strerror(ret));
2367                         } else {
2368                                 pid_t pid;
2369                                 dprint(FD_PROCESS, "will fork\n");
2370                                 pid = fork();
2371                                 if (!pid) {
2372                                         int ret;
2373
2374                                         ret = (int)(uintptr_t)thread_main(fd);
2375                                         _exit(ret);
2376                                 } else if (i == fio_debug_jobno)
2377                                         *fio_debug_jobp = pid;
2378                         }
2379                         dprint(FD_MUTEX, "wait on startup_sem\n");
2380                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2381                                 log_err("fio: job startup hung? exiting.\n");
2382                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2383                                 fio_abort = true;
2384                                 nr_started--;
2385                                 free(fd);
2386                                 break;
2387                         }
2388                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2389                 }
2390
2391                 /*
2392                  * Wait for the started threads to transition to
2393                  * TD_INITIALIZED.
2394                  */
2395                 fio_gettime(&this_start, NULL);
2396                 left = this_jobs;
2397                 while (left && !fio_abort) {
2398                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2399                                 break;
2400
2401                         do_usleep(100000);
2402
2403                         for (i = 0; i < this_jobs; i++) {
2404                                 td = map[i];
2405                                 if (!td)
2406                                         continue;
2407                                 if (td->runstate == TD_INITIALIZED) {
2408                                         map[i] = NULL;
2409                                         left--;
2410                                 } else if (td->runstate >= TD_EXITED) {
2411                                         map[i] = NULL;
2412                                         left--;
2413                                         todo--;
2414                                         nr_running++; /* work-around... */
2415                                 }
2416                         }
2417                 }
2418
2419                 if (left) {
2420                         log_err("fio: %d job%s failed to start\n", left,
2421                                         left > 1 ? "s" : "");
2422                         for (i = 0; i < this_jobs; i++) {
2423                                 td = map[i];
2424                                 if (!td)
2425                                         continue;
2426                                 kill(td->pid, SIGTERM);
2427                         }
2428                         break;
2429                 }
2430
2431                 /*
2432                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2433                  */
2434                 for_each_td(td, i) {
2435                         if (td->runstate != TD_INITIALIZED)
2436                                 continue;
2437
2438                         if (in_ramp_time(td))
2439                                 td_set_runstate(td, TD_RAMP);
2440                         else
2441                                 td_set_runstate(td, TD_RUNNING);
2442                         nr_running++;
2443                         nr_started--;
2444                         m_rate += ddir_rw_sum(td->o.ratemin);
2445                         t_rate += ddir_rw_sum(td->o.rate);
2446                         todo--;
2447                         fio_sem_up(td->sem);
2448                 }
2449
2450                 reap_threads(&nr_running, &t_rate, &m_rate);
2451
2452                 if (todo)
2453                         do_usleep(100000);
2454         }
2455
2456         while (nr_running) {
2457                 reap_threads(&nr_running, &t_rate, &m_rate);
2458                 do_usleep(10000);
2459         }
2460
2461         fio_idle_prof_stop();
2462
2463         update_io_ticks();
2464 }
2465
2466 static void free_disk_util(void)
2467 {
2468         disk_util_prune_entries();
2469         helper_thread_destroy();
2470 }
2471
2472 int fio_backend(struct sk_out *sk_out)
2473 {
2474         struct thread_data *td;
2475         int i;
2476
2477         if (exec_profile) {
2478                 if (load_profile(exec_profile))
2479                         return 1;
2480                 free(exec_profile);
2481                 exec_profile = NULL;
2482         }
2483         if (!thread_number)
2484                 return 0;
2485
2486         if (write_bw_log) {
2487                 struct log_params p = {
2488                         .log_type = IO_LOG_TYPE_BW,
2489                 };
2490
2491                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2492                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2493                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2494         }
2495
2496         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2497         if (!sk_out)
2498                 is_local_backend = true;
2499         if (startup_sem == NULL)
2500                 return 1;
2501
2502         set_genesis_time();
2503         stat_init();
2504         if (helper_thread_create(startup_sem, sk_out))
2505                 log_err("fio: failed to create helper thread\n");
2506
2507         cgroup_list = smalloc(sizeof(*cgroup_list));
2508         if (cgroup_list)
2509                 INIT_FLIST_HEAD(cgroup_list);
2510
2511         run_threads(sk_out);
2512
2513         helper_thread_exit();
2514
2515         if (!fio_abort) {
2516                 __show_run_stats();
2517                 if (write_bw_log) {
2518                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2519                                 struct io_log *log = agg_io_log[i];
2520
2521                                 flush_log(log, false);
2522                                 free_log(log);
2523                         }
2524                 }
2525         }
2526
2527         for_each_td(td, i) {
2528                 steadystate_free(td);
2529                 fio_options_free(td);
2530                 if (td->rusage_sem) {
2531                         fio_sem_remove(td->rusage_sem);
2532                         td->rusage_sem = NULL;
2533                 }
2534                 fio_sem_remove(td->sem);
2535                 td->sem = NULL;
2536         }
2537
2538         free_disk_util();
2539         if (cgroup_list) {
2540                 cgroup_kill(cgroup_list);
2541                 sfree(cgroup_list);
2542         }
2543
2544         fio_sem_remove(startup_sem);
2545         stat_exit();
2546         return exit_value;
2547 }