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