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 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
  70 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
  71 #else
  72 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  73 #endif
  74
  75 #define JOB_START_TIMEOUT       (5 * 1000)
  76
  77 static void sig_int(int sig)
  78 {
  79         if (threads) {
  80                 if (is_backend)
  81                         fio_server_got_signal(sig);
  82                 else {
  83                         log_info("\nfio: terminating on signal %d\n", sig);
  84                         log_info_flush();
  85                         exit_value = 128;
  86                 }
  87
  88                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  89         }
  90 }
  91
  92 void sig_show_status(int sig)
  93 {
  94         show_running_run_stats();
  95 }
  96
  97 static void set_sig_handlers(void)
  98 {
  99         struct sigaction act;
 100
 101         memset(&act, 0, sizeof(act));
 102         act.sa_handler = sig_int;
 103         act.sa_flags = SA_RESTART;
 104         sigaction(SIGINT, &act, NULL);
 105
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGTERM, &act, NULL);
 110
 111 /* Windows uses SIGBREAK as a quit signal from other applications */
 112 #ifdef WIN32
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGBREAK, &act, NULL);
 117 #endif
 118
 119         memset(&act, 0, sizeof(act));
 120         act.sa_handler = sig_show_status;
 121         act.sa_flags = SA_RESTART;
 122         sigaction(SIGUSR1, &act, NULL);
 123
 124         if (is_backend) {
 125                 memset(&act, 0, sizeof(act));
 126                 act.sa_handler = sig_int;
 127                 act.sa_flags = SA_RESTART;
 128                 sigaction(SIGPIPE, &act, NULL);
 129         }
 130 }
 131
 132 /*
 133  * Check if we are above the minimum rate given.
 134  */
 135 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 136                              enum fio_ddir ddir)
 137 {
 138         unsigned long long bytes = 0;
 139         unsigned long iops = 0;
 140         unsigned long spent;
 141         unsigned long long rate;
 142         unsigned long long ratemin = 0;
 143         unsigned int rate_iops = 0;
 144         unsigned int rate_iops_min = 0;
 145
 146         assert(ddir_rw(ddir));
 147
 148         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 149                 return false;
 150
 151         /*
 152          * allow a 2 second settle period in the beginning
 153          */
 154         if (mtime_since(&td->start, now) < 2000)
 155                 return false;
 156
 157         iops += td->this_io_blocks[ddir];
 158         bytes += td->this_io_bytes[ddir];
 159         ratemin += td->o.ratemin[ddir];
 160         rate_iops += td->o.rate_iops[ddir];
 161         rate_iops_min += td->o.rate_iops_min[ddir];
 162
 163         /*
 164          * if rate blocks is set, sample is running
 165          */
 166         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 167                 spent = mtime_since(&td->lastrate[ddir], now);
 168                 if (spent < td->o.ratecycle)
 169                         return false;
 170
 171                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 172                         /*
 173                          * check bandwidth specified rate
 174                          */
 175                         if (bytes < td->rate_bytes[ddir]) {
 176                                 log_err("%s: rate_min=%lluB/s not met, only transferred %lluB\n",
 177                                         td->o.name, ratemin, bytes);
 178                                 return true;
 179                         } else {
 180                                 if (spent)
 181                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 182                                 else
 183                                         rate = 0;
 184
 185                                 if (rate < ratemin ||
 186                                     bytes < td->rate_bytes[ddir]) {
 187                                         log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 188                                                 td->o.name, ratemin, rate);
 189                                         return true;
 190                                 }
 191                         }
 192                 } else {
 193                         /*
 194                          * checks iops specified rate
 195                          */
 196                         if (iops < rate_iops) {
 197                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 198                                                 td->o.name, rate_iops, iops);
 199                                 return true;
 200                         } else {
 201                                 if (spent)
 202                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 203                                 else
 204                                         rate = 0;
 205
 206                                 if (rate < rate_iops_min ||
 207                                     iops < td->rate_blocks[ddir]) {
 208                                         log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 209                                                 td->o.name, rate_iops_min, rate);
 210                                         return true;
 211                                 }
 212                         }
 213                 }
 214         }
 215
 216         td->rate_bytes[ddir] = bytes;
 217         td->rate_blocks[ddir] = iops;
 218         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 219         return false;
 220 }
 221
 222 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 223 {
 224         bool ret = false;
 225
 226         for_each_rw_ddir(ddir) {
 227                 if (td->bytes_done[ddir])
 228                         ret |= __check_min_rate(td, now, ddir);
 229         }
 230
 231         return ret;
 232 }
 233
 234 /*
 235  * When job exits, we can cancel the in-flight IO if we are using async
 236  * io. Attempt to do so.
 237  */
 238 static void cleanup_pending_aio(struct thread_data *td)
 239 {
 240         int r;
 241
 242         /*
 243          * get immediately available events, if any
 244          */
 245         r = io_u_queued_complete(td, 0);
 246
 247         /*
 248          * now cancel remaining active events
 249          */
 250         if (td->io_ops->cancel) {
 251                 struct io_u *io_u;
 252                 int i;
 253
 254                 io_u_qiter(&td->io_u_all, io_u, i) {
 255                         if (io_u->flags & IO_U_F_FLIGHT) {
 256                                 r = td->io_ops->cancel(td, io_u);
 257                                 if (!r)
 258                                         put_io_u(td, io_u);
 259                         }
 260                 }
 261         }
 262
 263         if (td->cur_depth)
 264                 r = io_u_queued_complete(td, td->cur_depth);
 265 }
 266
 267 /*
 268  * Helper to handle the final sync of a file. Works just like the normal
 269  * io path, just does everything sync.
 270  */
 271 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 272 {
 273         struct io_u *io_u = __get_io_u(td);
 274         enum fio_q_status ret;
 275
 276         if (!io_u)
 277                 return true;
 278
 279         io_u->ddir = DDIR_SYNC;
 280         io_u->file = f;
 281         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 282
 283         if (td_io_prep(td, io_u)) {
 284                 put_io_u(td, io_u);
 285                 return true;
 286         }
 287
 288 requeue:
 289         ret = td_io_queue(td, io_u);
 290         switch (ret) {
 291         case FIO_Q_QUEUED:
 292                 td_io_commit(td);
 293                 if (io_u_queued_complete(td, 1) < 0)
 294                         return true;
 295                 break;
 296         case FIO_Q_COMPLETED:
 297                 if (io_u->error) {
 298                         td_verror(td, io_u->error, "td_io_queue");
 299                         return true;
 300                 }
 301
 302                 if (io_u_sync_complete(td, io_u) < 0)
 303                         return true;
 304                 break;
 305         case FIO_Q_BUSY:
 306                 td_io_commit(td);
 307                 goto requeue;
 308         }
 309
 310         return false;
 311 }
 312
 313 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 314 {
 315         int ret, ret2;
 316
 317         if (fio_file_open(f))
 318                 return fio_io_sync(td, f);
 319
 320         if (td_io_open_file(td, f))
 321                 return 1;
 322
 323         ret = fio_io_sync(td, f);
 324         ret2 = 0;
 325         if (fio_file_open(f))
 326                 ret2 = td_io_close_file(td, f);
 327         return (ret || ret2);
 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 (verify_state_should_stop(td, io_u)) {
1012                                 put_io_u(td, io_u);
1013                                 break;
1014                         }
1015
1016                         if (td->o.verify_async)
1017                                 io_u->end_io = verify_io_u_async;
1018                         else
1019                                 io_u->end_io = verify_io_u;
1020                         td_set_runstate(td, TD_VERIFYING);
1021                 } else if (in_ramp_time(td))
1022                         td_set_runstate(td, TD_RAMP);
1023                 else
1024                         td_set_runstate(td, TD_RUNNING);
1025
1026                 /*
1027                  * Always log IO before it's issued, so we know the specific
1028                  * order of it. The logged unit will track when the IO has
1029                  * completed.
1030                  */
1031                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1032                     td->o.do_verify &&
1033                     td->o.verify != VERIFY_NONE &&
1034                     !td->o.experimental_verify)
1035                         log_io_piece(td, io_u);
1036
1037                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1038                         const unsigned long long blen = io_u->xfer_buflen;
1039                         const enum fio_ddir __ddir = acct_ddir(io_u);
1040
1041                         if (td->error)
1042                                 break;
1043
1044                         workqueue_enqueue(&td->io_wq, &io_u->work);
1045                         ret = FIO_Q_QUEUED;
1046
1047                         if (ddir_rw(__ddir)) {
1048                                 td->io_issues[__ddir]++;
1049                                 td->io_issue_bytes[__ddir] += blen;
1050                                 td->rate_io_issue_bytes[__ddir] += blen;
1051                         }
1052
1053                         if (should_check_rate(td))
1054                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1055
1056                 } else {
1057                         ret = io_u_submit(td, io_u);
1058
1059                         if (should_check_rate(td))
1060                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1061
1062                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1063                                 break;
1064
1065                         /*
1066                          * See if we need to complete some commands. Note that
1067                          * we can get BUSY even without IO queued, if the
1068                          * system is resource starved.
1069                          */
1070 reap:
1071                         full = queue_full(td) ||
1072                                 (ret == FIO_Q_BUSY && td->cur_depth);
1073                         if (full || io_in_polling(td))
1074                                 ret = wait_for_completions(td, &comp_time);
1075                 }
1076                 if (ret < 0)
1077                         break;
1078                 if (!ddir_rw_sum(td->bytes_done) &&
1079                     !td_ioengine_flagged(td, FIO_NOIO))
1080                         continue;
1081
1082                 if (!in_ramp_time(td) && should_check_rate(td)) {
1083                         if (check_min_rate(td, &comp_time)) {
1084                                 if (exitall_on_terminate || td->o.exitall_error)
1085                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1086                                 td_verror(td, EIO, "check_min_rate");
1087                                 break;
1088                         }
1089                 }
1090                 if (!in_ramp_time(td) && td->o.latency_target)
1091                         lat_target_check(td);
1092
1093                 if (ddir_rw(ddir) && td->o.thinktime)
1094                         handle_thinktime(td, ddir);
1095         }
1096
1097         check_update_rusage(td);
1098
1099         if (td->trim_entries)
1100                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1101
1102         if (td->o.fill_device && td->error == ENOSPC) {
1103                 td->error = 0;
1104                 fio_mark_td_terminate(td);
1105         }
1106         if (!td->error) {
1107                 struct fio_file *f;
1108
1109                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1110                         workqueue_flush(&td->io_wq);
1111                         i = 0;
1112                 } else
1113                         i = td->cur_depth;
1114
1115                 if (i) {
1116                         ret = io_u_queued_complete(td, i);
1117                         if (td->o.fill_device && td->error == ENOSPC)
1118                                 td->error = 0;
1119                 }
1120
1121                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1122                         td_set_runstate(td, TD_FSYNCING);
1123
1124                         for_each_file(td, f, i) {
1125                                 if (!fio_file_fsync(td, f))
1126                                         continue;
1127
1128                                 log_err("fio: end_fsync failed for file %s\n",
1129                                                                 f->file_name);
1130                         }
1131                 }
1132         } else
1133                 cleanup_pending_aio(td);
1134
1135         /*
1136          * stop job if we failed doing any IO
1137          */
1138         if (!ddir_rw_sum(td->this_io_bytes))
1139                 td->done = 1;
1140
1141         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1142                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1143 }
1144
1145 static void free_file_completion_logging(struct thread_data *td)
1146 {
1147         struct fio_file *f;
1148         unsigned int i;
1149
1150         for_each_file(td, f, i) {
1151                 if (!f->last_write_comp)
1152                         break;
1153                 sfree(f->last_write_comp);
1154         }
1155 }
1156
1157 static int init_file_completion_logging(struct thread_data *td,
1158                                         unsigned int depth)
1159 {
1160         struct fio_file *f;
1161         unsigned int i;
1162
1163         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1164                 return 0;
1165
1166         for_each_file(td, f, i) {
1167                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1168                 if (!f->last_write_comp)
1169                         goto cleanup;
1170         }
1171
1172         return 0;
1173
1174 cleanup:
1175         free_file_completion_logging(td);
1176         log_err("fio: failed to alloc write comp data\n");
1177         return 1;
1178 }
1179
1180 static void cleanup_io_u(struct thread_data *td)
1181 {
1182         struct io_u *io_u;
1183
1184         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1185
1186                 if (td->io_ops->io_u_free)
1187                         td->io_ops->io_u_free(td, io_u);
1188
1189                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1190         }
1191
1192         free_io_mem(td);
1193
1194         io_u_rexit(&td->io_u_requeues);
1195         io_u_qexit(&td->io_u_freelist, false);
1196         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1197
1198         free_file_completion_logging(td);
1199 }
1200
1201 static int init_io_u(struct thread_data *td)
1202 {
1203         struct io_u *io_u;
1204         int cl_align, i, max_units;
1205         int err;
1206
1207         max_units = td->o.iodepth;
1208
1209         err = 0;
1210         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1211         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1212         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1213
1214         if (err) {
1215                 log_err("fio: failed setting up IO queues\n");
1216                 return 1;
1217         }
1218
1219         cl_align = os_cache_line_size();
1220
1221         for (i = 0; i < max_units; i++) {
1222                 void *ptr;
1223
1224                 if (td->terminate)
1225                         return 1;
1226
1227                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1228                 if (!ptr) {
1229                         log_err("fio: unable to allocate aligned memory\n");
1230                         return 1;
1231                 }
1232
1233                 io_u = ptr;
1234                 memset(io_u, 0, sizeof(*io_u));
1235                 INIT_FLIST_HEAD(&io_u->verify_list);
1236                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1237
1238                 io_u->index = i;
1239                 io_u->flags = IO_U_F_FREE;
1240                 io_u_qpush(&td->io_u_freelist, io_u);
1241
1242                 /*
1243                  * io_u never leaves this stack, used for iteration of all
1244                  * io_u buffers.
1245                  */
1246                 io_u_qpush(&td->io_u_all, io_u);
1247
1248                 if (td->io_ops->io_u_init) {
1249                         int ret = td->io_ops->io_u_init(td, io_u);
1250
1251                         if (ret) {
1252                                 log_err("fio: failed to init engine data: %d\n", ret);
1253                                 return 1;
1254                         }
1255                 }
1256         }
1257
1258         init_io_u_buffers(td);
1259
1260         if (init_file_completion_logging(td, max_units))
1261                 return 1;
1262
1263         return 0;
1264 }
1265
1266 int init_io_u_buffers(struct thread_data *td)
1267 {
1268         struct io_u *io_u;
1269         unsigned long long max_bs, min_write;
1270         int i, max_units;
1271         int data_xfer = 1;
1272         char *p;
1273
1274         max_units = td->o.iodepth;
1275         max_bs = td_max_bs(td);
1276         min_write = td->o.min_bs[DDIR_WRITE];
1277         td->orig_buffer_size = (unsigned long long) max_bs
1278                                         * (unsigned long long) max_units;
1279
1280         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1281                 data_xfer = 0;
1282
1283         /*
1284          * if we may later need to do address alignment, then add any
1285          * possible adjustment here so that we don't cause a buffer
1286          * overflow later. this adjustment may be too much if we get
1287          * lucky and the allocator gives us an aligned address.
1288          */
1289         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1290             td_ioengine_flagged(td, FIO_RAWIO))
1291                 td->orig_buffer_size += page_mask + td->o.mem_align;
1292
1293         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1294                 unsigned long long bs;
1295
1296                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1297                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1298         }
1299
1300         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1301                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1302                 return 1;
1303         }
1304
1305         if (data_xfer && allocate_io_mem(td))
1306                 return 1;
1307
1308         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1309             td_ioengine_flagged(td, FIO_RAWIO))
1310                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1311         else
1312                 p = td->orig_buffer;
1313
1314         for (i = 0; i < max_units; i++) {
1315                 io_u = td->io_u_all.io_us[i];
1316                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1317
1318                 if (data_xfer) {
1319                         io_u->buf = p;
1320                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1321
1322                         if (td_write(td))
1323                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1324                         if (td_write(td) && td->o.verify_pattern_bytes) {
1325                                 /*
1326                                  * Fill the buffer with the pattern if we are
1327                                  * going to be doing writes.
1328                                  */
1329                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1330                         }
1331                 }
1332                 p += max_bs;
1333         }
1334
1335         return 0;
1336 }
1337
1338 /*
1339  * This function is Linux specific.
1340  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1341  */
1342 static int switch_ioscheduler(struct thread_data *td)
1343 {
1344 #ifdef FIO_HAVE_IOSCHED_SWITCH
1345         char tmp[256], tmp2[128], *p;
1346         FILE *f;
1347         int ret;
1348
1349         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1350                 return 0;
1351
1352         assert(td->files && td->files[0]);
1353         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1354
1355         f = fopen(tmp, "r+");
1356         if (!f) {
1357                 if (errno == ENOENT) {
1358                         log_err("fio: os or kernel doesn't support IO scheduler"
1359                                 " switching\n");
1360                         return 0;
1361                 }
1362                 td_verror(td, errno, "fopen iosched");
1363                 return 1;
1364         }
1365
1366         /*
1367          * Set io scheduler.
1368          */
1369         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1370         if (ferror(f) || ret != 1) {
1371                 td_verror(td, errno, "fwrite");
1372                 fclose(f);
1373                 return 1;
1374         }
1375
1376         rewind(f);
1377
1378         /*
1379          * Read back and check that the selected scheduler is now the default.
1380          */
1381         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1382         if (ferror(f) || ret < 0) {
1383                 td_verror(td, errno, "fread");
1384                 fclose(f);
1385                 return 1;
1386         }
1387         tmp[ret] = '\0';
1388         /*
1389          * either a list of io schedulers or "none\n" is expected. Strip the
1390          * trailing newline.
1391          */
1392         p = tmp;
1393         strsep(&p, "\n");
1394
1395         /*
1396          * Write to "none" entry doesn't fail, so check the result here.
1397          */
1398         if (!strcmp(tmp, "none")) {
1399                 log_err("fio: io scheduler is not tunable\n");
1400                 fclose(f);
1401                 return 0;
1402         }
1403
1404         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1405         if (!strstr(tmp, tmp2)) {
1406                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1407                 td_verror(td, EINVAL, "iosched_switch");
1408                 fclose(f);
1409                 return 1;
1410         }
1411
1412         fclose(f);
1413         return 0;
1414 #else
1415         return 0;
1416 #endif
1417 }
1418
1419 static bool keep_running(struct thread_data *td)
1420 {
1421         unsigned long long limit;
1422
1423         if (td->done)
1424                 return false;
1425         if (td->terminate)
1426                 return false;
1427         if (td->o.time_based)
1428                 return true;
1429         if (td->o.loops) {
1430                 td->o.loops--;
1431                 return true;
1432         }
1433         if (exceeds_number_ios(td))
1434                 return false;
1435
1436         if (td->o.io_size)
1437                 limit = td->o.io_size;
1438         else
1439                 limit = td->o.size;
1440
1441         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1442                 uint64_t diff;
1443
1444                 /*
1445                  * If the difference is less than the maximum IO size, we
1446                  * are done.
1447                  */
1448                 diff = limit - ddir_rw_sum(td->io_bytes);
1449                 if (diff < td_max_bs(td))
1450                         return false;
1451
1452                 if (fio_files_done(td) && !td->o.io_size)
1453                         return false;
1454
1455                 return true;
1456         }
1457
1458         return false;
1459 }
1460
1461 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1462 {
1463         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 13 + 1;
1464         int ret;
1465         char *str;
1466
1467         str = malloc(newlen);
1468         sprintf(str, "%s > %s.%s.txt 2>&1", string, o->name, mode);
1469
1470         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1471         ret = system(str);
1472         if (ret == -1)
1473                 log_err("fio: exec of cmd <%s> failed\n", str);
1474
1475         free(str);
1476         return ret;
1477 }
1478
1479 /*
1480  * Dry run to compute correct state of numberio for verification.
1481  */
1482 static uint64_t do_dry_run(struct thread_data *td)
1483 {
1484         td_set_runstate(td, TD_RUNNING);
1485
1486         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1487                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1488                 struct io_u *io_u;
1489                 int ret;
1490
1491                 if (td->terminate || td->done)
1492                         break;
1493
1494                 io_u = get_io_u(td);
1495                 if (IS_ERR_OR_NULL(io_u))
1496                         break;
1497
1498                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1499                 io_u->error = 0;
1500                 io_u->resid = 0;
1501                 if (ddir_rw(acct_ddir(io_u)))
1502                         td->io_issues[acct_ddir(io_u)]++;
1503                 if (ddir_rw(io_u->ddir)) {
1504                         io_u_mark_depth(td, 1);
1505                         td->ts.total_io_u[io_u->ddir]++;
1506                 }
1507
1508                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1509                     td->o.do_verify &&
1510                     td->o.verify != VERIFY_NONE &&
1511                     !td->o.experimental_verify)
1512                         log_io_piece(td, io_u);
1513
1514                 ret = io_u_sync_complete(td, io_u);
1515                 (void) ret;
1516         }
1517
1518         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1519 }
1520
1521 struct fork_data {
1522         struct thread_data *td;
1523         struct sk_out *sk_out;
1524 };
1525
1526 /*
1527  * Entry point for the thread based jobs. The process based jobs end up
1528  * here as well, after a little setup.
1529  */
1530 static void *thread_main(void *data)
1531 {
1532         struct fork_data *fd = data;
1533         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1534         struct thread_data *td = fd->td;
1535         struct thread_options *o = &td->o;
1536         struct sk_out *sk_out = fd->sk_out;
1537         uint64_t bytes_done[DDIR_RWDIR_CNT];
1538         int deadlock_loop_cnt;
1539         bool clear_state;
1540         int res, ret;
1541
1542         sk_out_assign(sk_out);
1543         free(fd);
1544
1545         if (!o->use_thread) {
1546                 setsid();
1547                 td->pid = getpid();
1548         } else
1549                 td->pid = gettid();
1550
1551         fio_local_clock_init();
1552
1553         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1554
1555         if (is_backend)
1556                 fio_server_send_start(td);
1557
1558         INIT_FLIST_HEAD(&td->io_log_list);
1559         INIT_FLIST_HEAD(&td->io_hist_list);
1560         INIT_FLIST_HEAD(&td->verify_list);
1561         INIT_FLIST_HEAD(&td->trim_list);
1562         td->io_hist_tree = RB_ROOT;
1563
1564         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1565         if (ret) {
1566                 td_verror(td, ret, "mutex_cond_init_pshared");
1567                 goto err;
1568         }
1569         ret = cond_init_pshared(&td->verify_cond);
1570         if (ret) {
1571                 td_verror(td, ret, "mutex_cond_pshared");
1572                 goto err;
1573         }
1574
1575         td_set_runstate(td, TD_INITIALIZED);
1576         dprint(FD_MUTEX, "up startup_sem\n");
1577         fio_sem_up(startup_sem);
1578         dprint(FD_MUTEX, "wait on td->sem\n");
1579         fio_sem_down(td->sem);
1580         dprint(FD_MUTEX, "done waiting on td->sem\n");
1581
1582         /*
1583          * A new gid requires privilege, so we need to do this before setting
1584          * the uid.
1585          */
1586         if (o->gid != -1U && setgid(o->gid)) {
1587                 td_verror(td, errno, "setgid");
1588                 goto err;
1589         }
1590         if (o->uid != -1U && setuid(o->uid)) {
1591                 td_verror(td, errno, "setuid");
1592                 goto err;
1593         }
1594
1595         td_zone_gen_index(td);
1596
1597         /*
1598          * Do this early, we don't want the compress threads to be limited
1599          * to the same CPUs as the IO workers. So do this before we set
1600          * any potential CPU affinity
1601          */
1602         if (iolog_compress_init(td, sk_out))
1603                 goto err;
1604
1605         /*
1606          * If we have a gettimeofday() thread, make sure we exclude that
1607          * thread from this job
1608          */
1609         if (o->gtod_cpu)
1610                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1611
1612         /*
1613          * Set affinity first, in case it has an impact on the memory
1614          * allocations.
1615          */
1616         if (fio_option_is_set(o, cpumask)) {
1617                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1618                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1619                         if (!ret) {
1620                                 log_err("fio: no CPUs set\n");
1621                                 log_err("fio: Try increasing number of available CPUs\n");
1622                                 td_verror(td, EINVAL, "cpus_split");
1623                                 goto err;
1624                         }
1625                 }
1626                 ret = fio_setaffinity(td->pid, o->cpumask);
1627                 if (ret == -1) {
1628                         td_verror(td, errno, "cpu_set_affinity");
1629                         goto err;
1630                 }
1631         }
1632
1633 #ifdef CONFIG_LIBNUMA
1634         /* numa node setup */
1635         if (fio_option_is_set(o, numa_cpunodes) ||
1636             fio_option_is_set(o, numa_memnodes)) {
1637                 struct bitmask *mask;
1638
1639                 if (numa_available() < 0) {
1640                         td_verror(td, errno, "Does not support NUMA API\n");
1641                         goto err;
1642                 }
1643
1644                 if (fio_option_is_set(o, numa_cpunodes)) {
1645                         mask = numa_parse_nodestring(o->numa_cpunodes);
1646                         ret = numa_run_on_node_mask(mask);
1647                         numa_free_nodemask(mask);
1648                         if (ret == -1) {
1649                                 td_verror(td, errno, \
1650                                         "numa_run_on_node_mask failed\n");
1651                                 goto err;
1652                         }
1653                 }
1654
1655                 if (fio_option_is_set(o, numa_memnodes)) {
1656                         mask = NULL;
1657                         if (o->numa_memnodes)
1658                                 mask = numa_parse_nodestring(o->numa_memnodes);
1659
1660                         switch (o->numa_mem_mode) {
1661                         case MPOL_INTERLEAVE:
1662                                 numa_set_interleave_mask(mask);
1663                                 break;
1664                         case MPOL_BIND:
1665                                 numa_set_membind(mask);
1666                                 break;
1667                         case MPOL_LOCAL:
1668                                 numa_set_localalloc();
1669                                 break;
1670                         case MPOL_PREFERRED:
1671                                 numa_set_preferred(o->numa_mem_prefer_node);
1672                                 break;
1673                         case MPOL_DEFAULT:
1674                         default:
1675                                 break;
1676                         }
1677
1678                         if (mask)
1679                                 numa_free_nodemask(mask);
1680
1681                 }
1682         }
1683 #endif
1684
1685         if (fio_pin_memory(td))
1686                 goto err;
1687
1688         /*
1689          * May alter parameters that init_io_u() will use, so we need to
1690          * do this first.
1691          */
1692         if (!init_iolog(td))
1693                 goto err;
1694
1695         if (td_io_init(td))
1696                 goto err;
1697
1698         if (init_io_u(td))
1699                 goto err;
1700
1701         if (td->io_ops->post_init && td->io_ops->post_init(td))
1702                 goto err;
1703
1704         if (o->verify_async && verify_async_init(td))
1705                 goto err;
1706
1707         if (fio_option_is_set(o, ioprio) ||
1708             fio_option_is_set(o, ioprio_class)) {
1709                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1710                 if (ret == -1) {
1711                         td_verror(td, errno, "ioprio_set");
1712                         goto err;
1713                 }
1714         }
1715
1716         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1717                 goto err;
1718
1719         errno = 0;
1720         if (nice(o->nice) == -1 && errno != 0) {
1721                 td_verror(td, errno, "nice");
1722                 goto err;
1723         }
1724
1725         if (o->ioscheduler && switch_ioscheduler(td))
1726                 goto err;
1727
1728         if (!o->create_serialize && setup_files(td))
1729                 goto err;
1730
1731         if (!init_random_map(td))
1732                 goto err;
1733
1734         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1735                 goto err;
1736
1737         if (o->pre_read && !pre_read_files(td))
1738                 goto err;
1739
1740         fio_verify_init(td);
1741
1742         if (rate_submit_init(td, sk_out))
1743                 goto err;
1744
1745         set_epoch_time(td, o->log_unix_epoch);
1746         fio_getrusage(&td->ru_start);
1747         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1748         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1749         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1750
1751         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1752                         o->ratemin[DDIR_TRIM]) {
1753                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1754                                         sizeof(td->bw_sample_time));
1755                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1756                                         sizeof(td->bw_sample_time));
1757                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1758                                         sizeof(td->bw_sample_time));
1759         }
1760
1761         memset(bytes_done, 0, sizeof(bytes_done));
1762         clear_state = false;
1763
1764         while (keep_running(td)) {
1765                 uint64_t verify_bytes;
1766
1767                 fio_gettime(&td->start, NULL);
1768                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1769
1770                 if (clear_state) {
1771                         clear_io_state(td, 0);
1772
1773                         if (o->unlink_each_loop && unlink_all_files(td))
1774                                 break;
1775                 }
1776
1777                 prune_io_piece_log(td);
1778
1779                 if (td->o.verify_only && td_write(td))
1780                         verify_bytes = do_dry_run(td);
1781                 else {
1782                         do_io(td, bytes_done);
1783
1784                         if (!ddir_rw_sum(bytes_done)) {
1785                                 fio_mark_td_terminate(td);
1786                                 verify_bytes = 0;
1787                         } else {
1788                                 verify_bytes = bytes_done[DDIR_WRITE] +
1789                                                 bytes_done[DDIR_TRIM];
1790                         }
1791                 }
1792
1793                 /*
1794                  * If we took too long to shut down, the main thread could
1795                  * already consider us reaped/exited. If that happens, break
1796                  * out and clean up.
1797                  */
1798                 if (td->runstate >= TD_EXITED)
1799                         break;
1800
1801                 clear_state = true;
1802
1803                 /*
1804                  * Make sure we've successfully updated the rusage stats
1805                  * before waiting on the stat mutex. Otherwise we could have
1806                  * the stat thread holding stat mutex and waiting for
1807                  * the rusage_sem, which would never get upped because
1808                  * this thread is waiting for the stat mutex.
1809                  */
1810                 deadlock_loop_cnt = 0;
1811                 do {
1812                         check_update_rusage(td);
1813                         if (!fio_sem_down_trylock(stat_sem))
1814                                 break;
1815                         usleep(1000);
1816                         if (deadlock_loop_cnt++ > 5000) {
1817                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1818                                 td->error = EDEADLK;
1819                                 goto err;
1820                         }
1821                 } while (1);
1822
1823                 if (td_read(td) && td->io_bytes[DDIR_READ])
1824                         update_runtime(td, elapsed_us, DDIR_READ);
1825                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1826                         update_runtime(td, elapsed_us, DDIR_WRITE);
1827                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1828                         update_runtime(td, elapsed_us, DDIR_TRIM);
1829                 fio_gettime(&td->start, NULL);
1830                 fio_sem_up(stat_sem);
1831
1832                 if (td->error || td->terminate)
1833                         break;
1834
1835                 if (!o->do_verify ||
1836                     o->verify == VERIFY_NONE ||
1837                     td_ioengine_flagged(td, FIO_UNIDIR))
1838                         continue;
1839
1840                 clear_io_state(td, 0);
1841
1842                 fio_gettime(&td->start, NULL);
1843
1844                 do_verify(td, verify_bytes);
1845
1846                 /*
1847                  * See comment further up for why this is done here.
1848                  */
1849                 check_update_rusage(td);
1850
1851                 fio_sem_down(stat_sem);
1852                 update_runtime(td, elapsed_us, DDIR_READ);
1853                 fio_gettime(&td->start, NULL);
1854                 fio_sem_up(stat_sem);
1855
1856                 if (td->error || td->terminate)
1857                         break;
1858         }
1859
1860         /*
1861          * Acquire this lock if we were doing overlap checking in
1862          * offload mode so that we don't clean up this job while
1863          * another thread is checking its io_u's for overlap
1864          */
1865         if (td_offload_overlap(td)) {
1866                 int res = pthread_mutex_lock(&overlap_check);
1867                 assert(res == 0);
1868         }
1869         td_set_runstate(td, TD_FINISHING);
1870         if (td_offload_overlap(td)) {
1871                 res = pthread_mutex_unlock(&overlap_check);
1872                 assert(res == 0);
1873         }
1874
1875         update_rusage_stat(td);
1876         td->ts.total_run_time = mtime_since_now(&td->epoch);
1877         for_each_rw_ddir(ddir) {
1878                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1879         }
1880
1881         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1882             (td->o.verify != VERIFY_NONE && td_write(td)))
1883                 verify_save_state(td->thread_number);
1884
1885         fio_unpin_memory(td);
1886
1887         td_writeout_logs(td, true);
1888
1889         iolog_compress_exit(td);
1890         rate_submit_exit(td);
1891
1892         if (o->exec_postrun)
1893                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1894
1895         if (exitall_on_terminate || (o->exitall_error && td->error))
1896                 fio_terminate_threads(td->groupid, td->o.exit_what);
1897
1898 err:
1899         if (td->error)
1900                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1901                                                         td->verror);
1902
1903         if (o->verify_async)
1904                 verify_async_exit(td);
1905
1906         close_and_free_files(td);
1907         cleanup_io_u(td);
1908         close_ioengine(td);
1909         cgroup_shutdown(td, cgroup_mnt);
1910         verify_free_state(td);
1911         td_zone_free_index(td);
1912
1913         if (fio_option_is_set(o, cpumask)) {
1914                 ret = fio_cpuset_exit(&o->cpumask);
1915                 if (ret)
1916                         td_verror(td, ret, "fio_cpuset_exit");
1917         }
1918
1919         /*
1920          * do this very late, it will log file closing as well
1921          */
1922         if (o->write_iolog_file)
1923                 write_iolog_close(td);
1924         if (td->io_log_rfile)
1925                 fclose(td->io_log_rfile);
1926
1927         td_set_runstate(td, TD_EXITED);
1928
1929         /*
1930          * Do this last after setting our runstate to exited, so we
1931          * know that the stat thread is signaled.
1932          */
1933         check_update_rusage(td);
1934
1935         sk_out_drop();
1936         return (void *) (uintptr_t) td->error;
1937 }
1938
1939 /*
1940  * Run over the job map and reap the threads that have exited, if any.
1941  */
1942 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1943                          uint64_t *m_rate)
1944 {
1945         struct thread_data *td;
1946         unsigned int cputhreads, realthreads, pending;
1947         int i, status, ret;
1948
1949         /*
1950          * reap exited threads (TD_EXITED -> TD_REAPED)
1951          */
1952         realthreads = pending = cputhreads = 0;
1953         for_each_td(td, i) {
1954                 int flags = 0;
1955
1956                  if (!strcmp(td->o.ioengine, "cpuio"))
1957                         cputhreads++;
1958                 else
1959                         realthreads++;
1960
1961                 if (!td->pid) {
1962                         pending++;
1963                         continue;
1964                 }
1965                 if (td->runstate == TD_REAPED)
1966                         continue;
1967                 if (td->o.use_thread) {
1968                         if (td->runstate == TD_EXITED) {
1969                                 td_set_runstate(td, TD_REAPED);
1970                                 goto reaped;
1971                         }
1972                         continue;
1973                 }
1974
1975                 flags = WNOHANG;
1976                 if (td->runstate == TD_EXITED)
1977                         flags = 0;
1978
1979                 /*
1980                  * check if someone quit or got killed in an unusual way
1981                  */
1982                 ret = waitpid(td->pid, &status, flags);
1983                 if (ret < 0) {
1984                         if (errno == ECHILD) {
1985                                 log_err("fio: pid=%d disappeared %d\n",
1986                                                 (int) td->pid, td->runstate);
1987                                 td->sig = ECHILD;
1988                                 td_set_runstate(td, TD_REAPED);
1989                                 goto reaped;
1990                         }
1991                         perror("waitpid");
1992                 } else if (ret == td->pid) {
1993                         if (WIFSIGNALED(status)) {
1994                                 int sig = WTERMSIG(status);
1995
1996                                 if (sig != SIGTERM && sig != SIGUSR2)
1997                                         log_err("fio: pid=%d, got signal=%d\n",
1998                                                         (int) td->pid, sig);
1999                                 td->sig = sig;
2000                                 td_set_runstate(td, TD_REAPED);
2001                                 goto reaped;
2002                         }
2003                         if (WIFEXITED(status)) {
2004                                 if (WEXITSTATUS(status) && !td->error)
2005                                         td->error = WEXITSTATUS(status);
2006
2007                                 td_set_runstate(td, TD_REAPED);
2008                                 goto reaped;
2009                         }
2010                 }
2011
2012                 /*
2013                  * If the job is stuck, do a forceful timeout of it and
2014                  * move on.
2015                  */
2016                 if (td->terminate &&
2017                     td->runstate < TD_FSYNCING &&
2018                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2019                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2020                                 "%lu seconds, it appears to be stuck. Doing "
2021                                 "forceful exit of this job.\n",
2022                                 td->o.name, td->runstate,
2023                                 (unsigned long) time_since_now(&td->terminate_time));
2024                         td_set_runstate(td, TD_REAPED);
2025                         goto reaped;
2026                 }
2027
2028                 /*
2029                  * thread is not dead, continue
2030                  */
2031                 pending++;
2032                 continue;
2033 reaped:
2034                 (*nr_running)--;
2035                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2036                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2037                 if (!td->pid)
2038                         pending--;
2039
2040                 if (td->error)
2041                         exit_value++;
2042
2043                 done_secs += mtime_since_now(&td->epoch) / 1000;
2044                 profile_td_exit(td);
2045                 flow_exit_job(td);
2046         }
2047
2048         if (*nr_running == cputhreads && !pending && realthreads)
2049                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2050 }
2051
2052 static bool __check_trigger_file(void)
2053 {
2054         struct stat sb;
2055
2056         if (!trigger_file)
2057                 return false;
2058
2059         if (stat(trigger_file, &sb))
2060                 return false;
2061
2062         if (unlink(trigger_file) < 0)
2063                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2064                                                         strerror(errno));
2065
2066         return true;
2067 }
2068
2069 static bool trigger_timedout(void)
2070 {
2071         if (trigger_timeout)
2072                 if (time_since_genesis() >= trigger_timeout) {
2073                         trigger_timeout = 0;
2074                         return true;
2075                 }
2076
2077         return false;
2078 }
2079
2080 void exec_trigger(const char *cmd)
2081 {
2082         int ret;
2083
2084         if (!cmd || cmd[0] == '\0')
2085                 return;
2086
2087         ret = system(cmd);
2088         if (ret == -1)
2089                 log_err("fio: failed executing %s trigger\n", cmd);
2090 }
2091
2092 void check_trigger_file(void)
2093 {
2094         if (__check_trigger_file() || trigger_timedout()) {
2095                 if (nr_clients)
2096                         fio_clients_send_trigger(trigger_remote_cmd);
2097                 else {
2098                         verify_save_state(IO_LIST_ALL);
2099                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2100                         exec_trigger(trigger_cmd);
2101                 }
2102         }
2103 }
2104
2105 static int fio_verify_load_state(struct thread_data *td)
2106 {
2107         int ret;
2108
2109         if (!td->o.verify_state)
2110                 return 0;
2111
2112         if (is_backend) {
2113                 void *data;
2114
2115                 ret = fio_server_get_verify_state(td->o.name,
2116                                         td->thread_number - 1, &data);
2117                 if (!ret)
2118                         verify_assign_state(td, data);
2119         } else {
2120                 char prefix[PATH_MAX];
2121
2122                 if (aux_path)
2123                         sprintf(prefix, "%s%clocal", aux_path,
2124                                         FIO_OS_PATH_SEPARATOR);
2125                 else
2126                         strcpy(prefix, "local");
2127                 ret = verify_load_state(td, prefix);
2128         }
2129
2130         return ret;
2131 }
2132
2133 static void do_usleep(unsigned int usecs)
2134 {
2135         check_for_running_stats();
2136         check_trigger_file();
2137         usleep(usecs);
2138 }
2139
2140 static bool check_mount_writes(struct thread_data *td)
2141 {
2142         struct fio_file *f;
2143         unsigned int i;
2144
2145         if (!td_write(td) || td->o.allow_mounted_write)
2146                 return false;
2147
2148         /*
2149          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2150          * are mkfs'd and mounted.
2151          */
2152         for_each_file(td, f, i) {
2153 #ifdef FIO_HAVE_CHARDEV_SIZE
2154                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2155 #else
2156                 if (f->filetype != FIO_TYPE_BLOCK)
2157 #endif
2158                         continue;
2159                 if (device_is_mounted(f->file_name))
2160                         goto mounted;
2161         }
2162
2163         return false;
2164 mounted:
2165         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2166         return true;
2167 }
2168
2169 static bool waitee_running(struct thread_data *me)
2170 {
2171         const char *waitee = me->o.wait_for;
2172         const char *self = me->o.name;
2173         struct thread_data *td;
2174         int i;
2175
2176         if (!waitee)
2177                 return false;
2178
2179         for_each_td(td, i) {
2180                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2181                         continue;
2182
2183                 if (td->runstate < TD_EXITED) {
2184                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2185                                         self, td->o.name,
2186                                         runstate_to_name(td->runstate));
2187                         return true;
2188                 }
2189         }
2190
2191         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2192         return false;
2193 }
2194
2195 /*
2196  * Main function for kicking off and reaping jobs, as needed.
2197  */
2198 static void run_threads(struct sk_out *sk_out)
2199 {
2200         struct thread_data *td;
2201         unsigned int i, todo, nr_running, nr_started;
2202         uint64_t m_rate, t_rate;
2203         uint64_t spent;
2204
2205         if (fio_gtod_offload && fio_start_gtod_thread())
2206                 return;
2207
2208         fio_idle_prof_init();
2209
2210         set_sig_handlers();
2211
2212         nr_thread = nr_process = 0;
2213         for_each_td(td, i) {
2214                 if (check_mount_writes(td))
2215                         return;
2216                 if (td->o.use_thread)
2217                         nr_thread++;
2218                 else
2219                         nr_process++;
2220         }
2221
2222         if (output_format & FIO_OUTPUT_NORMAL) {
2223                 struct buf_output out;
2224
2225                 buf_output_init(&out);
2226                 __log_buf(&out, "Starting ");
2227                 if (nr_thread)
2228                         __log_buf(&out, "%d thread%s", nr_thread,
2229                                                 nr_thread > 1 ? "s" : "");
2230                 if (nr_process) {
2231                         if (nr_thread)
2232                                 __log_buf(&out, " and ");
2233                         __log_buf(&out, "%d process%s", nr_process,
2234                                                 nr_process > 1 ? "es" : "");
2235                 }
2236                 __log_buf(&out, "\n");
2237                 log_info_buf(out.buf, out.buflen);
2238                 buf_output_free(&out);
2239         }
2240
2241         todo = thread_number;
2242         nr_running = 0;
2243         nr_started = 0;
2244         m_rate = t_rate = 0;
2245
2246         for_each_td(td, i) {
2247                 print_status_init(td->thread_number - 1);
2248
2249                 if (!td->o.create_serialize)
2250                         continue;
2251
2252                 if (fio_verify_load_state(td))
2253                         goto reap;
2254
2255                 /*
2256                  * do file setup here so it happens sequentially,
2257                  * we don't want X number of threads getting their
2258                  * client data interspersed on disk
2259                  */
2260                 if (setup_files(td)) {
2261 reap:
2262                         exit_value++;
2263                         if (td->error)
2264                                 log_err("fio: pid=%d, err=%d/%s\n",
2265                                         (int) td->pid, td->error, td->verror);
2266                         td_set_runstate(td, TD_REAPED);
2267                         todo--;
2268                 } else {
2269                         struct fio_file *f;
2270                         unsigned int j;
2271
2272                         /*
2273                          * for sharing to work, each job must always open
2274                          * its own files. so close them, if we opened them
2275                          * for creation
2276                          */
2277                         for_each_file(td, f, j) {
2278                                 if (fio_file_open(f))
2279                                         td_io_close_file(td, f);
2280                         }
2281                 }
2282         }
2283
2284         /* start idle threads before io threads start to run */
2285         fio_idle_prof_start();
2286
2287         set_genesis_time();
2288
2289         while (todo) {
2290                 struct thread_data *map[REAL_MAX_JOBS];
2291                 struct timespec this_start;
2292                 int this_jobs = 0, left;
2293                 struct fork_data *fd;
2294
2295                 /*
2296                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2297                  */
2298                 for_each_td(td, i) {
2299                         if (td->runstate != TD_NOT_CREATED)
2300                                 continue;
2301
2302                         /*
2303                          * never got a chance to start, killed by other
2304                          * thread for some reason
2305                          */
2306                         if (td->terminate) {
2307                                 todo--;
2308                                 continue;
2309                         }
2310
2311                         if (td->o.start_delay) {
2312                                 spent = utime_since_genesis();
2313
2314                                 if (td->o.start_delay > spent)
2315                                         continue;
2316                         }
2317
2318                         if (td->o.stonewall && (nr_started || nr_running)) {
2319                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2320                                                         td->o.name);
2321                                 break;
2322                         }
2323
2324                         if (waitee_running(td)) {
2325                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2326                                                 td->o.name, td->o.wait_for);
2327                                 continue;
2328                         }
2329
2330                         init_disk_util(td);
2331
2332                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2333                         td->update_rusage = 0;
2334
2335                         /*
2336                          * Set state to created. Thread will transition
2337                          * to TD_INITIALIZED when it's done setting up.
2338                          */
2339                         td_set_runstate(td, TD_CREATED);
2340                         map[this_jobs++] = td;
2341                         nr_started++;
2342
2343                         fd = calloc(1, sizeof(*fd));
2344                         fd->td = td;
2345                         fd->sk_out = sk_out;
2346
2347                         if (td->o.use_thread) {
2348                                 int ret;
2349
2350                                 dprint(FD_PROCESS, "will pthread_create\n");
2351                                 ret = pthread_create(&td->thread, NULL,
2352                                                         thread_main, fd);
2353                                 if (ret) {
2354                                         log_err("pthread_create: %s\n",
2355                                                         strerror(ret));
2356                                         free(fd);
2357                                         nr_started--;
2358                                         break;
2359                                 }
2360                                 fd = NULL;
2361                                 ret = pthread_detach(td->thread);
2362                                 if (ret)
2363                                         log_err("pthread_detach: %s",
2364                                                         strerror(ret));
2365                         } else {
2366                                 pid_t pid;
2367                                 dprint(FD_PROCESS, "will fork\n");
2368                                 pid = fork();
2369                                 if (!pid) {
2370                                         int ret;
2371
2372                                         ret = (int)(uintptr_t)thread_main(fd);
2373                                         _exit(ret);
2374                                 } else if (i == fio_debug_jobno)
2375                                         *fio_debug_jobp = pid;
2376                         }
2377                         dprint(FD_MUTEX, "wait on startup_sem\n");
2378                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2379                                 log_err("fio: job startup hung? exiting.\n");
2380                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2381                                 fio_abort = true;
2382                                 nr_started--;
2383                                 free(fd);
2384                                 break;
2385                         }
2386                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2387                 }
2388
2389                 /*
2390                  * Wait for the started threads to transition to
2391                  * TD_INITIALIZED.
2392                  */
2393                 fio_gettime(&this_start, NULL);
2394                 left = this_jobs;
2395                 while (left && !fio_abort) {
2396                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2397                                 break;
2398
2399                         do_usleep(100000);
2400
2401                         for (i = 0; i < this_jobs; i++) {
2402                                 td = map[i];
2403                                 if (!td)
2404                                         continue;
2405                                 if (td->runstate == TD_INITIALIZED) {
2406                                         map[i] = NULL;
2407                                         left--;
2408                                 } else if (td->runstate >= TD_EXITED) {
2409                                         map[i] = NULL;
2410                                         left--;
2411                                         todo--;
2412                                         nr_running++; /* work-around... */
2413                                 }
2414                         }
2415                 }
2416
2417                 if (left) {
2418                         log_err("fio: %d job%s failed to start\n", left,
2419                                         left > 1 ? "s" : "");
2420                         for (i = 0; i < this_jobs; i++) {
2421                                 td = map[i];
2422                                 if (!td)
2423                                         continue;
2424                                 kill(td->pid, SIGTERM);
2425                         }
2426                         break;
2427                 }
2428
2429                 /*
2430                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2431                  */
2432                 for_each_td(td, i) {
2433                         if (td->runstate != TD_INITIALIZED)
2434                                 continue;
2435
2436                         if (in_ramp_time(td))
2437                                 td_set_runstate(td, TD_RAMP);
2438                         else
2439                                 td_set_runstate(td, TD_RUNNING);
2440                         nr_running++;
2441                         nr_started--;
2442                         m_rate += ddir_rw_sum(td->o.ratemin);
2443                         t_rate += ddir_rw_sum(td->o.rate);
2444                         todo--;
2445                         fio_sem_up(td->sem);
2446                 }
2447
2448                 reap_threads(&nr_running, &t_rate, &m_rate);
2449
2450                 if (todo)
2451                         do_usleep(100000);
2452         }
2453
2454         while (nr_running) {
2455                 reap_threads(&nr_running, &t_rate, &m_rate);
2456                 do_usleep(10000);
2457         }
2458
2459         fio_idle_prof_stop();
2460
2461         update_io_ticks();
2462 }
2463
2464 static void free_disk_util(void)
2465 {
2466         disk_util_prune_entries();
2467         helper_thread_destroy();
2468 }
2469
2470 int fio_backend(struct sk_out *sk_out)
2471 {
2472         struct thread_data *td;
2473         int i;
2474
2475         if (exec_profile) {
2476                 if (load_profile(exec_profile))
2477                         return 1;
2478                 free(exec_profile);
2479                 exec_profile = NULL;
2480         }
2481         if (!thread_number)
2482                 return 0;
2483
2484         if (write_bw_log) {
2485                 struct log_params p = {
2486                         .log_type = IO_LOG_TYPE_BW,
2487                 };
2488
2489                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2490                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2491                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2492         }
2493
2494         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2495         if (!sk_out)
2496                 is_local_backend = true;
2497         if (startup_sem == NULL)
2498                 return 1;
2499
2500         set_genesis_time();
2501         stat_init();
2502         if (helper_thread_create(startup_sem, sk_out))
2503                 log_err("fio: failed to create helper thread\n");
2504
2505         cgroup_list = smalloc(sizeof(*cgroup_list));
2506         if (cgroup_list)
2507                 INIT_FLIST_HEAD(cgroup_list);
2508
2509         run_threads(sk_out);
2510
2511         helper_thread_exit();
2512
2513         if (!fio_abort) {
2514                 __show_run_stats();
2515                 if (write_bw_log) {
2516                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2517                                 struct io_log *log = agg_io_log[i];
2518
2519                                 flush_log(log, false);
2520                                 free_log(log);
2521                         }
2522                 }
2523         }
2524
2525         for_each_td(td, i) {
2526                 steadystate_free(td);
2527                 fio_options_free(td);
2528                 if (td->rusage_sem) {
2529                         fio_sem_remove(td->rusage_sem);
2530                         td->rusage_sem = NULL;
2531                 }
2532                 fio_sem_remove(td->sem);
2533                 td->sem = NULL;
2534         }
2535
2536         free_disk_util();
2537         if (cgroup_list) {
2538                 cgroup_kill(cgroup_list);
2539                 sfree(cgroup_list);
2540         }
2541
2542         fio_sem_remove(startup_sem);
2543         stat_exit();
2544         return exit_value;
2545 }