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