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., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38 #include <math.h>
  39
  40 #include "fio.h"
  41 #ifndef FIO_NO_HAVE_SHM_H
  42 #include <sys/shm.h>
  43 #endif
  44 #include "hash.h"
  45 #include "smalloc.h"
  46 #include "verify.h"
  47 #include "trim.h"
  48 #include "diskutil.h"
  49 #include "cgroup.h"
  50 #include "profile.h"
  51 #include "lib/rand.h"
  52 #include "lib/memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59 #include "rate-submit.h"
  60 #include "helper_thread.h"
  61
  62 static struct fio_mutex *startup_mutex;
  63 static struct flist_head *cgroup_list;
  64 static char *cgroup_mnt;
  65 static int exit_value;
  66 static volatile int fio_abort;
  67 static unsigned int nr_process = 0;
  68 static unsigned int nr_thread = 0;
  69
  70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  71
  72 int groupid = 0;
  73 unsigned int thread_number = 0;
  74 unsigned int stat_number = 0;
  75 int shm_id = 0;
  76 int temp_stall_ts;
  77 unsigned long done_secs = 0;
  78
  79 #define PAGE_ALIGN(buf) \
  80         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  81
  82 #define JOB_START_TIMEOUT       (5 * 1000)
  83
  84 static void sig_int(int sig)
  85 {
  86         if (threads) {
  87                 if (is_backend)
  88                         fio_server_got_signal(sig);
  89                 else {
  90                         log_info("\nfio: terminating on signal %d\n", sig);
  91                         log_info_flush();
  92                         exit_value = 128;
  93                 }
  94
  95                 fio_terminate_threads(TERMINATE_ALL);
  96         }
  97 }
  98
  99 void sig_show_status(int sig)
 100 {
 101         show_running_run_stats();
 102 }
 103
 104 static void set_sig_handlers(void)
 105 {
 106         struct sigaction act;
 107
 108         memset(&act, 0, sizeof(act));
 109         act.sa_handler = sig_int;
 110         act.sa_flags = SA_RESTART;
 111         sigaction(SIGINT, &act, NULL);
 112
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGTERM, &act, NULL);
 117
 118 /* Windows uses SIGBREAK as a quit signal from other applications */
 119 #ifdef WIN32
 120         memset(&act, 0, sizeof(act));
 121         act.sa_handler = sig_int;
 122         act.sa_flags = SA_RESTART;
 123         sigaction(SIGBREAK, &act, NULL);
 124 #endif
 125
 126         memset(&act, 0, sizeof(act));
 127         act.sa_handler = sig_show_status;
 128         act.sa_flags = SA_RESTART;
 129         sigaction(SIGUSR1, &act, NULL);
 130
 131         if (is_backend) {
 132                 memset(&act, 0, sizeof(act));
 133                 act.sa_handler = sig_int;
 134                 act.sa_flags = SA_RESTART;
 135                 sigaction(SIGPIPE, &act, NULL);
 136         }
 137 }
 138
 139 /*
 140  * Check if we are above the minimum rate given.
 141  */
 142 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 143                              enum fio_ddir ddir)
 144 {
 145         unsigned long long bytes = 0;
 146         unsigned long iops = 0;
 147         unsigned long spent;
 148         unsigned long rate;
 149         unsigned int ratemin = 0;
 150         unsigned int rate_iops = 0;
 151         unsigned int rate_iops_min = 0;
 152
 153         assert(ddir_rw(ddir));
 154
 155         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 156                 return false;
 157
 158         /*
 159          * allow a 2 second settle period in the beginning
 160          */
 161         if (mtime_since(&td->start, now) < 2000)
 162                 return false;
 163
 164         iops += td->this_io_blocks[ddir];
 165         bytes += td->this_io_bytes[ddir];
 166         ratemin += td->o.ratemin[ddir];
 167         rate_iops += td->o.rate_iops[ddir];
 168         rate_iops_min += td->o.rate_iops_min[ddir];
 169
 170         /*
 171          * if rate blocks is set, sample is running
 172          */
 173         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 174                 spent = mtime_since(&td->lastrate[ddir], now);
 175                 if (spent < td->o.ratecycle)
 176                         return false;
 177
 178                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 179                         /*
 180                          * check bandwidth specified rate
 181                          */
 182                         if (bytes < td->rate_bytes[ddir]) {
 183                                 log_err("%s: min rate %u not met\n", td->o.name,
 184                                                                 ratemin);
 185                                 return true;
 186                         } else {
 187                                 if (spent)
 188                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 189                                 else
 190                                         rate = 0;
 191
 192                                 if (rate < ratemin ||
 193                                     bytes < td->rate_bytes[ddir]) {
 194                                         log_err("%s: min rate %u not met, got"
 195                                                 " %luKB/sec\n", td->o.name,
 196                                                         ratemin, rate);
 197                                         return true;
 198                                 }
 199                         }
 200                 } else {
 201                         /*
 202                          * checks iops specified rate
 203                          */
 204                         if (iops < rate_iops) {
 205                                 log_err("%s: min iops rate %u not met\n",
 206                                                 td->o.name, rate_iops);
 207                                 return true;
 208                         } else {
 209                                 if (spent)
 210                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 211                                 else
 212                                         rate = 0;
 213
 214                                 if (rate < rate_iops_min ||
 215                                     iops < td->rate_blocks[ddir]) {
 216                                         log_err("%s: min iops rate %u not met,"
 217                                                 " got %lu\n", td->o.name,
 218                                                         rate_iops_min, rate);
 219                                         return true;
 220                                 }
 221                         }
 222                 }
 223         }
 224
 225         td->rate_bytes[ddir] = bytes;
 226         td->rate_blocks[ddir] = iops;
 227         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 228         return false;
 229 }
 230
 231 static bool check_min_rate(struct thread_data *td, struct timeval *now)
 232 {
 233         bool ret = false;
 234
 235         if (td->bytes_done[DDIR_READ])
 236                 ret |= __check_min_rate(td, now, DDIR_READ);
 237         if (td->bytes_done[DDIR_WRITE])
 238                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 239         if (td->bytes_done[DDIR_TRIM])
 240                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 241
 242         return ret;
 243 }
 244
 245 /*
 246  * When job exits, we can cancel the in-flight IO if we are using async
 247  * io. Attempt to do so.
 248  */
 249 static void cleanup_pending_aio(struct thread_data *td)
 250 {
 251         int r;
 252
 253         /*
 254          * get immediately available events, if any
 255          */
 256         r = io_u_queued_complete(td, 0);
 257         if (r < 0)
 258                 return;
 259
 260         /*
 261          * now cancel remaining active events
 262          */
 263         if (td->io_ops->cancel) {
 264                 struct io_u *io_u;
 265                 int i;
 266
 267                 io_u_qiter(&td->io_u_all, io_u, i) {
 268                         if (io_u->flags & IO_U_F_FLIGHT) {
 269                                 r = td->io_ops->cancel(td, io_u);
 270                                 if (!r)
 271                                         put_io_u(td, io_u);
 272                         }
 273                 }
 274         }
 275
 276         if (td->cur_depth)
 277                 r = io_u_queued_complete(td, td->cur_depth);
 278 }
 279
 280 /*
 281  * Helper to handle the final sync of a file. Works just like the normal
 282  * io path, just does everything sync.
 283  */
 284 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 285 {
 286         struct io_u *io_u = __get_io_u(td);
 287         int ret;
 288
 289         if (!io_u)
 290                 return true;
 291
 292         io_u->ddir = DDIR_SYNC;
 293         io_u->file = f;
 294
 295         if (td_io_prep(td, io_u)) {
 296                 put_io_u(td, io_u);
 297                 return true;
 298         }
 299
 300 requeue:
 301         ret = td_io_queue(td, io_u);
 302         if (ret < 0) {
 303                 td_verror(td, io_u->error, "td_io_queue");
 304                 put_io_u(td, io_u);
 305                 return true;
 306         } else if (ret == FIO_Q_QUEUED) {
 307                 if (td_io_commit(td))
 308                         return true;
 309                 if (io_u_queued_complete(td, 1) < 0)
 310                         return true;
 311         } else if (ret == FIO_Q_COMPLETED) {
 312                 if (io_u->error) {
 313                         td_verror(td, io_u->error, "td_io_queue");
 314                         return true;
 315                 }
 316
 317                 if (io_u_sync_complete(td, io_u) < 0)
 318                         return true;
 319         } else if (ret == FIO_Q_BUSY) {
 320                 if (td_io_commit(td))
 321                         return true;
 322                 goto requeue;
 323         }
 324
 325         return false;
 326 }
 327
 328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 329 {
 330         int ret;
 331
 332         if (fio_file_open(f))
 333                 return fio_io_sync(td, f);
 334
 335         if (td_io_open_file(td, f))
 336                 return 1;
 337
 338         ret = fio_io_sync(td, f);
 339         td_io_close_file(td, f);
 340         return ret;
 341 }
 342
 343 static inline void __update_tv_cache(struct thread_data *td)
 344 {
 345         fio_gettime(&td->tv_cache, NULL);
 346 }
 347
 348 static inline void update_tv_cache(struct thread_data *td)
 349 {
 350         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 351                 __update_tv_cache(td);
 352 }
 353
 354 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 355 {
 356         if (in_ramp_time(td))
 357                 return false;
 358         if (!td->o.timeout)
 359                 return false;
 360         if (utime_since(&td->epoch, t) >= td->o.timeout)
 361                 return true;
 362
 363         return false;
 364 }
 365
 366 /*
 367  * We need to update the runtime consistently in ms, but keep a running
 368  * tally of the current elapsed time in microseconds for sub millisecond
 369  * updates.
 370  */
 371 static inline void update_runtime(struct thread_data *td,
 372                                   unsigned long long *elapsed_us,
 373                                   const enum fio_ddir ddir)
 374 {
 375         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 376                 return;
 377
 378         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 379         elapsed_us[ddir] += utime_since_now(&td->start);
 380         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 381 }
 382
 383 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 384                                 int *retptr)
 385 {
 386         int ret = *retptr;
 387
 388         if (ret < 0 || td->error) {
 389                 int err = td->error;
 390                 enum error_type_bit eb;
 391
 392                 if (ret < 0)
 393                         err = -ret;
 394
 395                 eb = td_error_type(ddir, err);
 396                 if (!(td->o.continue_on_error & (1 << eb)))
 397                         return true;
 398
 399                 if (td_non_fatal_error(td, eb, err)) {
 400                         /*
 401                          * Continue with the I/Os in case of
 402                          * a non fatal error.
 403                          */
 404                         update_error_count(td, err);
 405                         td_clear_error(td);
 406                         *retptr = 0;
 407                         return false;
 408                 } else if (td->o.fill_device && err == ENOSPC) {
 409                         /*
 410                          * We expect to hit this error if
 411                          * fill_device option is set.
 412                          */
 413                         td_clear_error(td);
 414                         fio_mark_td_terminate(td);
 415                         return true;
 416                 } else {
 417                         /*
 418                          * Stop the I/O in case of a fatal
 419                          * error.
 420                          */
 421                         update_error_count(td, err);
 422                         return true;
 423                 }
 424         }
 425
 426         return false;
 427 }
 428
 429 static void check_update_rusage(struct thread_data *td)
 430 {
 431         if (td->update_rusage) {
 432                 td->update_rusage = 0;
 433                 update_rusage_stat(td);
 434                 fio_mutex_up(td->rusage_sem);
 435         }
 436 }
 437
 438 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 439 {
 440         const int full = queue_full(td);
 441         int min_evts = 0;
 442         int ret;
 443
 444         if (td->flags & TD_F_REGROW_LOGS) {
 445                 ret = io_u_quiesce(td);
 446                 regrow_logs(td);
 447                 return ret;
 448         }
 449
 450         /*
 451          * if the queue is full, we MUST reap at least 1 event
 452          */
 453         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 454         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 455                 min_evts = 1;
 456
 457         if (time && (__should_check_rate(td, DDIR_READ) ||
 458             __should_check_rate(td, DDIR_WRITE) ||
 459             __should_check_rate(td, DDIR_TRIM)))
 460                 fio_gettime(time, NULL);
 461
 462         do {
 463                 ret = io_u_queued_complete(td, min_evts);
 464                 if (ret < 0)
 465                         break;
 466         } while (full && (td->cur_depth > td->o.iodepth_low));
 467
 468         return ret;
 469 }
 470
 471 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 472                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 473                    struct timeval *comp_time)
 474 {
 475         int ret2;
 476
 477         switch (*ret) {
 478         case FIO_Q_COMPLETED:
 479                 if (io_u->error) {
 480                         *ret = -io_u->error;
 481                         clear_io_u(td, io_u);
 482                 } else if (io_u->resid) {
 483                         int bytes = io_u->xfer_buflen - io_u->resid;
 484                         struct fio_file *f = io_u->file;
 485
 486                         if (bytes_issued)
 487                                 *bytes_issued += bytes;
 488
 489                         if (!from_verify)
 490                                 trim_io_piece(td, io_u);
 491
 492                         /*
 493                          * zero read, fail
 494                          */
 495                         if (!bytes) {
 496                                 if (!from_verify)
 497                                         unlog_io_piece(td, io_u);
 498                                 td_verror(td, EIO, "full resid");
 499                                 put_io_u(td, io_u);
 500                                 break;
 501                         }
 502
 503                         io_u->xfer_buflen = io_u->resid;
 504                         io_u->xfer_buf += bytes;
 505                         io_u->offset += bytes;
 506
 507                         if (ddir_rw(io_u->ddir))
 508                                 td->ts.short_io_u[io_u->ddir]++;
 509
 510                         f = io_u->file;
 511                         if (io_u->offset == f->real_file_size)
 512                                 goto sync_done;
 513
 514                         requeue_io_u(td, &io_u);
 515                 } else {
 516 sync_done:
 517                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 518                             __should_check_rate(td, DDIR_WRITE) ||
 519                             __should_check_rate(td, DDIR_TRIM)))
 520                                 fio_gettime(comp_time, NULL);
 521
 522                         *ret = io_u_sync_complete(td, io_u);
 523                         if (*ret < 0)
 524                                 break;
 525                 }
 526
 527                 if (td->flags & TD_F_REGROW_LOGS)
 528                         regrow_logs(td);
 529
 530                 /*
 531                  * when doing I/O (not when verifying),
 532                  * check for any errors that are to be ignored
 533                  */
 534                 if (!from_verify)
 535                         break;
 536
 537                 return 0;
 538         case FIO_Q_QUEUED:
 539                 /*
 540                  * if the engine doesn't have a commit hook,
 541                  * the io_u is really queued. if it does have such
 542                  * a hook, it has to call io_u_queued() itself.
 543                  */
 544                 if (td->io_ops->commit == NULL)
 545                         io_u_queued(td, io_u);
 546                 if (bytes_issued)
 547                         *bytes_issued += io_u->xfer_buflen;
 548                 break;
 549         case FIO_Q_BUSY:
 550                 if (!from_verify)
 551                         unlog_io_piece(td, io_u);
 552                 requeue_io_u(td, &io_u);
 553                 ret2 = td_io_commit(td);
 554                 if (ret2 < 0)
 555                         *ret = ret2;
 556                 break;
 557         default:
 558                 assert(*ret < 0);
 559                 td_verror(td, -(*ret), "td_io_queue");
 560                 break;
 561         }
 562
 563         if (break_on_this_error(td, ddir, ret))
 564                 return 1;
 565
 566         return 0;
 567 }
 568
 569 static inline bool io_in_polling(struct thread_data *td)
 570 {
 571         return !td->o.iodepth_batch_complete_min &&
 572                    !td->o.iodepth_batch_complete_max;
 573 }
 574 /*
 575  * Unlinks files from thread data fio_file structure
 576  */
 577 static int unlink_all_files(struct thread_data *td)
 578 {
 579         struct fio_file *f;
 580         unsigned int i;
 581         int ret = 0;
 582
 583         for_each_file(td, f, i) {
 584                 if (f->filetype != FIO_TYPE_FILE)
 585                         continue;
 586                 ret = td_io_unlink_file(td, f);
 587                 if (ret)
 588                         break;
 589         }
 590
 591         if (ret)
 592                 td_verror(td, ret, "unlink_all_files");
 593
 594         return ret;
 595 }
 596
 597 /*
 598  * The main verify engine. Runs over the writes we previously submitted,
 599  * reads the blocks back in, and checks the crc/md5 of the data.
 600  */
 601 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 602 {
 603         struct fio_file *f;
 604         struct io_u *io_u;
 605         int ret, min_events;
 606         unsigned int i;
 607
 608         dprint(FD_VERIFY, "starting loop\n");
 609
 610         /*
 611          * sync io first and invalidate cache, to make sure we really
 612          * read from disk.
 613          */
 614         for_each_file(td, f, i) {
 615                 if (!fio_file_open(f))
 616                         continue;
 617                 if (fio_io_sync(td, f))
 618                         break;
 619                 if (file_invalidate_cache(td, f))
 620                         break;
 621         }
 622
 623         check_update_rusage(td);
 624
 625         if (td->error)
 626                 return;
 627
 628         /*
 629          * verify_state needs to be reset before verification
 630          * proceeds so that expected random seeds match actual
 631          * random seeds in headers. The main loop will reset
 632          * all random number generators if randrepeat is set.
 633          */
 634         if (!td->o.rand_repeatable)
 635                 td_fill_verify_state_seed(td);
 636
 637         td_set_runstate(td, TD_VERIFYING);
 638
 639         io_u = NULL;
 640         while (!td->terminate) {
 641                 enum fio_ddir ddir;
 642                 int full;
 643
 644                 update_tv_cache(td);
 645                 check_update_rusage(td);
 646
 647                 if (runtime_exceeded(td, &td->tv_cache)) {
 648                         __update_tv_cache(td);
 649                         if (runtime_exceeded(td, &td->tv_cache)) {
 650                                 fio_mark_td_terminate(td);
 651                                 break;
 652                         }
 653                 }
 654
 655                 if (flow_threshold_exceeded(td))
 656                         continue;
 657
 658                 if (!td->o.experimental_verify) {
 659                         io_u = __get_io_u(td);
 660                         if (!io_u)
 661                                 break;
 662
 663                         if (get_next_verify(td, io_u)) {
 664                                 put_io_u(td, io_u);
 665                                 break;
 666                         }
 667
 668                         if (td_io_prep(td, io_u)) {
 669                                 put_io_u(td, io_u);
 670                                 break;
 671                         }
 672                 } else {
 673                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 674                                 break;
 675
 676                         while ((io_u = get_io_u(td)) != NULL) {
 677                                 if (IS_ERR_OR_NULL(io_u)) {
 678                                         io_u = NULL;
 679                                         ret = FIO_Q_BUSY;
 680                                         goto reap;
 681                                 }
 682
 683                                 /*
 684                                  * We are only interested in the places where
 685                                  * we wrote or trimmed IOs. Turn those into
 686                                  * reads for verification purposes.
 687                                  */
 688                                 if (io_u->ddir == DDIR_READ) {
 689                                         /*
 690                                          * Pretend we issued it for rwmix
 691                                          * accounting
 692                                          */
 693                                         td->io_issues[DDIR_READ]++;
 694                                         put_io_u(td, io_u);
 695                                         continue;
 696                                 } else if (io_u->ddir == DDIR_TRIM) {
 697                                         io_u->ddir = DDIR_READ;
 698                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 699                                         break;
 700                                 } else if (io_u->ddir == DDIR_WRITE) {
 701                                         io_u->ddir = DDIR_READ;
 702                                         break;
 703                                 } else {
 704                                         put_io_u(td, io_u);
 705                                         continue;
 706                                 }
 707                         }
 708
 709                         if (!io_u)
 710                                 break;
 711                 }
 712
 713                 if (verify_state_should_stop(td, io_u)) {
 714                         put_io_u(td, io_u);
 715                         break;
 716                 }
 717
 718                 if (td->o.verify_async)
 719                         io_u->end_io = verify_io_u_async;
 720                 else
 721                         io_u->end_io = verify_io_u;
 722
 723                 ddir = io_u->ddir;
 724                 if (!td->o.disable_slat)
 725                         fio_gettime(&io_u->start_time, NULL);
 726
 727                 ret = td_io_queue(td, io_u);
 728
 729                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 730                         break;
 731
 732                 /*
 733                  * if we can queue more, do so. but check if there are
 734                  * completed io_u's first. Note that we can get BUSY even
 735                  * without IO queued, if the system is resource starved.
 736                  */
 737 reap:
 738                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 739                 if (full || io_in_polling(td))
 740                         ret = wait_for_completions(td, NULL);
 741
 742                 if (ret < 0)
 743                         break;
 744         }
 745
 746         check_update_rusage(td);
 747
 748         if (!td->error) {
 749                 min_events = td->cur_depth;
 750
 751                 if (min_events)
 752                         ret = io_u_queued_complete(td, min_events);
 753         } else
 754                 cleanup_pending_aio(td);
 755
 756         td_set_runstate(td, TD_RUNNING);
 757
 758         dprint(FD_VERIFY, "exiting loop\n");
 759 }
 760
 761 static bool exceeds_number_ios(struct thread_data *td)
 762 {
 763         unsigned long long number_ios;
 764
 765         if (!td->o.number_ios)
 766                 return false;
 767
 768         number_ios = ddir_rw_sum(td->io_blocks);
 769         number_ios += td->io_u_queued + td->io_u_in_flight;
 770
 771         return number_ios >= (td->o.number_ios * td->loops);
 772 }
 773
 774 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 775 {
 776         unsigned long long bytes, limit;
 777
 778         if (td_rw(td))
 779                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 780         else if (td_write(td))
 781                 bytes = this_bytes[DDIR_WRITE];
 782         else if (td_read(td))
 783                 bytes = this_bytes[DDIR_READ];
 784         else
 785                 bytes = this_bytes[DDIR_TRIM];
 786
 787         if (td->o.io_limit)
 788                 limit = td->o.io_limit;
 789         else
 790                 limit = td->o.size;
 791
 792         limit *= td->loops;
 793         return bytes >= limit || exceeds_number_ios(td);
 794 }
 795
 796 static bool io_issue_bytes_exceeded(struct thread_data *td)
 797 {
 798         return io_bytes_exceeded(td, td->io_issue_bytes);
 799 }
 800
 801 static bool io_complete_bytes_exceeded(struct thread_data *td)
 802 {
 803         return io_bytes_exceeded(td, td->this_io_bytes);
 804 }
 805
 806 /*
 807  * used to calculate the next io time for rate control
 808  *
 809  */
 810 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 811 {
 812         uint64_t secs, remainder, bps, bytes, iops;
 813
 814         assert(!(td->flags & TD_F_CHILD));
 815         bytes = td->rate_io_issue_bytes[ddir];
 816         bps = td->rate_bps[ddir];
 817
 818         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 819                 uint64_t val;
 820                 iops = bps / td->o.bs[ddir];
 821                 val = (int64_t) (1000000 / iops) *
 822                                 -logf(__rand_0_1(&td->poisson_state));
 823                 if (val) {
 824                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 825                                         (unsigned long long) 1000000 / val);
 826                 }
 827                 td->last_usec += val;
 828                 return td->last_usec;
 829         } else if (bps) {
 830                 secs = bytes / bps;
 831                 remainder = bytes % bps;
 832                 return remainder * 1000000 / bps + secs * 1000000;
 833         }
 834
 835         return 0;
 836 }
 837
 838 /*
 839  * Main IO worker function. It retrieves io_u's to process and queues
 840  * and reaps them, checking for rate and errors along the way.
 841  *
 842  * Returns number of bytes written and trimmed.
 843  */
 844 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 845 {
 846         unsigned int i;
 847         int ret = 0;
 848         uint64_t total_bytes, bytes_issued = 0;
 849
 850         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 851                 bytes_done[i] = td->bytes_done[i];
 852
 853         if (in_ramp_time(td))
 854                 td_set_runstate(td, TD_RAMP);
 855         else
 856                 td_set_runstate(td, TD_RUNNING);
 857
 858         lat_target_init(td);
 859
 860         total_bytes = td->o.size;
 861         /*
 862         * Allow random overwrite workloads to write up to io_limit
 863         * before starting verification phase as 'size' doesn't apply.
 864         */
 865         if (td_write(td) && td_random(td) && td->o.norandommap)
 866                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 867         /*
 868          * If verify_backlog is enabled, we'll run the verify in this
 869          * handler as well. For that case, we may need up to twice the
 870          * amount of bytes.
 871          */
 872         if (td->o.verify != VERIFY_NONE &&
 873            (td_write(td) && td->o.verify_backlog))
 874                 total_bytes += td->o.size;
 875
 876         /* In trimwrite mode, each byte is trimmed and then written, so
 877          * allow total_bytes to be twice as big */
 878         if (td_trimwrite(td))
 879                 total_bytes += td->total_io_size;
 880
 881         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 882                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 883                 td->o.time_based) {
 884                 struct timeval comp_time;
 885                 struct io_u *io_u;
 886                 int full;
 887                 enum fio_ddir ddir;
 888
 889                 check_update_rusage(td);
 890
 891                 if (td->terminate || td->done)
 892                         break;
 893
 894                 update_tv_cache(td);
 895
 896                 if (runtime_exceeded(td, &td->tv_cache)) {
 897                         __update_tv_cache(td);
 898                         if (runtime_exceeded(td, &td->tv_cache)) {
 899                                 fio_mark_td_terminate(td);
 900                                 break;
 901                         }
 902                 }
 903
 904                 if (flow_threshold_exceeded(td))
 905                         continue;
 906
 907                 /*
 908                  * Break if we exceeded the bytes. The exception is time
 909                  * based runs, but we still need to break out of the loop
 910                  * for those to run verification, if enabled.
 911                  */
 912                 if (bytes_issued >= total_bytes &&
 913                     (!td->o.time_based ||
 914                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 915                         break;
 916
 917                 io_u = get_io_u(td);
 918                 if (IS_ERR_OR_NULL(io_u)) {
 919                         int err = PTR_ERR(io_u);
 920
 921                         io_u = NULL;
 922                         if (err == -EBUSY) {
 923                                 ret = FIO_Q_BUSY;
 924                                 goto reap;
 925                         }
 926                         if (td->o.latency_target)
 927                                 goto reap;
 928                         break;
 929                 }
 930
 931                 ddir = io_u->ddir;
 932
 933                 /*
 934                  * Add verification end_io handler if:
 935                  *      - Asked to verify (!td_rw(td))
 936                  *      - Or the io_u is from our verify list (mixed write/ver)
 937                  */
 938                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 939                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 940
 941                         if (!td->o.verify_pattern_bytes) {
 942                                 io_u->rand_seed = __rand(&td->verify_state);
 943                                 if (sizeof(int) != sizeof(long *))
 944                                         io_u->rand_seed *= __rand(&td->verify_state);
 945                         }
 946
 947                         if (verify_state_should_stop(td, io_u)) {
 948                                 put_io_u(td, io_u);
 949                                 break;
 950                         }
 951
 952                         if (td->o.verify_async)
 953                                 io_u->end_io = verify_io_u_async;
 954                         else
 955                                 io_u->end_io = verify_io_u;
 956                         td_set_runstate(td, TD_VERIFYING);
 957                 } else if (in_ramp_time(td))
 958                         td_set_runstate(td, TD_RAMP);
 959                 else
 960                         td_set_runstate(td, TD_RUNNING);
 961
 962                 /*
 963                  * Always log IO before it's issued, so we know the specific
 964                  * order of it. The logged unit will track when the IO has
 965                  * completed.
 966                  */
 967                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 968                     td->o.do_verify &&
 969                     td->o.verify != VERIFY_NONE &&
 970                     !td->o.experimental_verify)
 971                         log_io_piece(td, io_u);
 972
 973                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 974                         const unsigned long blen = io_u->xfer_buflen;
 975                         const enum fio_ddir ddir = acct_ddir(io_u);
 976
 977                         if (td->error)
 978                                 break;
 979
 980                         workqueue_enqueue(&td->io_wq, &io_u->work);
 981                         ret = FIO_Q_QUEUED;
 982
 983                         if (ddir_rw(ddir)) {
 984                                 td->io_issues[ddir]++;
 985                                 td->io_issue_bytes[ddir] += blen;
 986                                 td->rate_io_issue_bytes[ddir] += blen;
 987                         }
 988
 989                         if (should_check_rate(td))
 990                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 991
 992                 } else {
 993                         ret = td_io_queue(td, io_u);
 994
 995                         if (should_check_rate(td))
 996                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 997
 998                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 999                                 break;
1000
1001                         /*
1002                          * See if we need to complete some commands. Note that
1003                          * we can get BUSY even without IO queued, if the
1004                          * system is resource starved.
1005                          */
1006 reap:
1007                         full = queue_full(td) ||
1008                                 (ret == FIO_Q_BUSY && td->cur_depth);
1009                         if (full || io_in_polling(td))
1010                                 ret = wait_for_completions(td, &comp_time);
1011                 }
1012                 if (ret < 0)
1013                         break;
1014                 if (!ddir_rw_sum(td->bytes_done) &&
1015                     !td_ioengine_flagged(td, FIO_NOIO))
1016                         continue;
1017
1018                 if (!in_ramp_time(td) && should_check_rate(td)) {
1019                         if (check_min_rate(td, &comp_time)) {
1020                                 if (exitall_on_terminate || td->o.exitall_error)
1021                                         fio_terminate_threads(td->groupid);
1022                                 td_verror(td, EIO, "check_min_rate");
1023                                 break;
1024                         }
1025                 }
1026                 if (!in_ramp_time(td) && td->o.latency_target)
1027                         lat_target_check(td);
1028
1029                 if (td->o.thinktime) {
1030                         unsigned long long b;
1031
1032                         b = ddir_rw_sum(td->io_blocks);
1033                         if (!(b % td->o.thinktime_blocks)) {
1034                                 int left;
1035
1036                                 io_u_quiesce(td);
1037
1038                                 if (td->o.thinktime_spin)
1039                                         usec_spin(td->o.thinktime_spin);
1040
1041                                 left = td->o.thinktime - td->o.thinktime_spin;
1042                                 if (left)
1043                                         usec_sleep(td, left);
1044                         }
1045                 }
1046         }
1047
1048         check_update_rusage(td);
1049
1050         if (td->trim_entries)
1051                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1052
1053         if (td->o.fill_device && td->error == ENOSPC) {
1054                 td->error = 0;
1055                 fio_mark_td_terminate(td);
1056         }
1057         if (!td->error) {
1058                 struct fio_file *f;
1059
1060                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1061                         workqueue_flush(&td->io_wq);
1062                         i = 0;
1063                 } else
1064                         i = td->cur_depth;
1065
1066                 if (i) {
1067                         ret = io_u_queued_complete(td, i);
1068                         if (td->o.fill_device && td->error == ENOSPC)
1069                                 td->error = 0;
1070                 }
1071
1072                 if (should_fsync(td) && td->o.end_fsync) {
1073                         td_set_runstate(td, TD_FSYNCING);
1074
1075                         for_each_file(td, f, i) {
1076                                 if (!fio_file_fsync(td, f))
1077                                         continue;
1078
1079                                 log_err("fio: end_fsync failed for file %s\n",
1080                                                                 f->file_name);
1081                         }
1082                 }
1083         } else
1084                 cleanup_pending_aio(td);
1085
1086         /*
1087          * stop job if we failed doing any IO
1088          */
1089         if (!ddir_rw_sum(td->this_io_bytes))
1090                 td->done = 1;
1091
1092         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1093                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1094 }
1095
1096 static void free_file_completion_logging(struct thread_data *td)
1097 {
1098         struct fio_file *f;
1099         unsigned int i;
1100
1101         for_each_file(td, f, i) {
1102                 if (!f->last_write_comp)
1103                         break;
1104                 sfree(f->last_write_comp);
1105         }
1106 }
1107
1108 static int init_file_completion_logging(struct thread_data *td,
1109                                         unsigned int depth)
1110 {
1111         struct fio_file *f;
1112         unsigned int i;
1113
1114         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1115                 return 0;
1116
1117         for_each_file(td, f, i) {
1118                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1119                 if (!f->last_write_comp)
1120                         goto cleanup;
1121         }
1122
1123         return 0;
1124
1125 cleanup:
1126         free_file_completion_logging(td);
1127         log_err("fio: failed to alloc write comp data\n");
1128         return 1;
1129 }
1130
1131 static void cleanup_io_u(struct thread_data *td)
1132 {
1133         struct io_u *io_u;
1134
1135         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1136
1137                 if (td->io_ops->io_u_free)
1138                         td->io_ops->io_u_free(td, io_u);
1139
1140                 fio_memfree(io_u, sizeof(*io_u));
1141         }
1142
1143         free_io_mem(td);
1144
1145         io_u_rexit(&td->io_u_requeues);
1146         io_u_qexit(&td->io_u_freelist);
1147         io_u_qexit(&td->io_u_all);
1148
1149         free_file_completion_logging(td);
1150 }
1151
1152 static int init_io_u(struct thread_data *td)
1153 {
1154         struct io_u *io_u;
1155         unsigned int max_bs, min_write;
1156         int cl_align, i, max_units;
1157         int data_xfer = 1, err;
1158         char *p;
1159
1160         max_units = td->o.iodepth;
1161         max_bs = td_max_bs(td);
1162         min_write = td->o.min_bs[DDIR_WRITE];
1163         td->orig_buffer_size = (unsigned long long) max_bs
1164                                         * (unsigned long long) max_units;
1165
1166         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1167                 data_xfer = 0;
1168
1169         err = 0;
1170         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1171         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1172         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1173
1174         if (err) {
1175                 log_err("fio: failed setting up IO queues\n");
1176                 return 1;
1177         }
1178
1179         /*
1180          * if we may later need to do address alignment, then add any
1181          * possible adjustment here so that we don't cause a buffer
1182          * overflow later. this adjustment may be too much if we get
1183          * lucky and the allocator gives us an aligned address.
1184          */
1185         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1186             td_ioengine_flagged(td, FIO_RAWIO))
1187                 td->orig_buffer_size += page_mask + td->o.mem_align;
1188
1189         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1190                 unsigned long bs;
1191
1192                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1193                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1194         }
1195
1196         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1197                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1198                 return 1;
1199         }
1200
1201         if (data_xfer && allocate_io_mem(td))
1202                 return 1;
1203
1204         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1205             td_ioengine_flagged(td, FIO_RAWIO))
1206                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1207         else
1208                 p = td->orig_buffer;
1209
1210         cl_align = os_cache_line_size();
1211
1212         for (i = 0; i < max_units; i++) {
1213                 void *ptr;
1214
1215                 if (td->terminate)
1216                         return 1;
1217
1218                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1219                 if (!ptr) {
1220                         log_err("fio: unable to allocate aligned memory\n");
1221                         break;
1222                 }
1223
1224                 io_u = ptr;
1225                 memset(io_u, 0, sizeof(*io_u));
1226                 INIT_FLIST_HEAD(&io_u->verify_list);
1227                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1228
1229                 if (data_xfer) {
1230                         io_u->buf = p;
1231                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1232
1233                         if (td_write(td))
1234                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1235                         if (td_write(td) && td->o.verify_pattern_bytes) {
1236                                 /*
1237                                  * Fill the buffer with the pattern if we are
1238                                  * going to be doing writes.
1239                                  */
1240                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1241                         }
1242                 }
1243
1244                 io_u->index = i;
1245                 io_u->flags = IO_U_F_FREE;
1246                 io_u_qpush(&td->io_u_freelist, io_u);
1247
1248                 /*
1249                  * io_u never leaves this stack, used for iteration of all
1250                  * io_u buffers.
1251                  */
1252                 io_u_qpush(&td->io_u_all, io_u);
1253
1254                 if (td->io_ops->io_u_init) {
1255                         int ret = td->io_ops->io_u_init(td, io_u);
1256
1257                         if (ret) {
1258                                 log_err("fio: failed to init engine data: %d\n", ret);
1259                                 return 1;
1260                         }
1261                 }
1262
1263                 p += max_bs;
1264         }
1265
1266         if (init_file_completion_logging(td, max_units))
1267                 return 1;
1268
1269         return 0;
1270 }
1271
1272 static int switch_ioscheduler(struct thread_data *td)
1273 {
1274 #ifdef FIO_HAVE_IOSCHED_SWITCH
1275         char tmp[256], tmp2[128];
1276         FILE *f;
1277         int ret;
1278
1279         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1280                 return 0;
1281
1282         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1283
1284         f = fopen(tmp, "r+");
1285         if (!f) {
1286                 if (errno == ENOENT) {
1287                         log_err("fio: os or kernel doesn't support IO scheduler"
1288                                 " switching\n");
1289                         return 0;
1290                 }
1291                 td_verror(td, errno, "fopen iosched");
1292                 return 1;
1293         }
1294
1295         /*
1296          * Set io scheduler.
1297          */
1298         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1299         if (ferror(f) || ret != 1) {
1300                 td_verror(td, errno, "fwrite");
1301                 fclose(f);
1302                 return 1;
1303         }
1304
1305         rewind(f);
1306
1307         /*
1308          * Read back and check that the selected scheduler is now the default.
1309          */
1310         memset(tmp, 0, sizeof(tmp));
1311         ret = fread(tmp, sizeof(tmp), 1, f);
1312         if (ferror(f) || ret < 0) {
1313                 td_verror(td, errno, "fread");
1314                 fclose(f);
1315                 return 1;
1316         }
1317         /*
1318          * either a list of io schedulers or "none\n" is expected.
1319          */
1320         tmp[strlen(tmp) - 1] = '\0';
1321
1322         /*
1323          * Write to "none" entry doesn't fail, so check the result here.
1324          */
1325         if (!strcmp(tmp, "none")) {
1326                 log_err("fio: io scheduler is not tunable\n");
1327                 fclose(f);
1328                 return 0;
1329         }
1330
1331         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1332         if (!strstr(tmp, tmp2)) {
1333                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1334                 td_verror(td, EINVAL, "iosched_switch");
1335                 fclose(f);
1336                 return 1;
1337         }
1338
1339         fclose(f);
1340         return 0;
1341 #else
1342         return 0;
1343 #endif
1344 }
1345
1346 static bool keep_running(struct thread_data *td)
1347 {
1348         unsigned long long limit;
1349
1350         if (td->done)
1351                 return false;
1352         if (td->o.time_based)
1353                 return true;
1354         if (td->o.loops) {
1355                 td->o.loops--;
1356                 return true;
1357         }
1358         if (exceeds_number_ios(td))
1359                 return false;
1360
1361         if (td->o.io_limit)
1362                 limit = td->o.io_limit;
1363         else
1364                 limit = td->o.size;
1365
1366         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1367                 uint64_t diff;
1368
1369                 /*
1370                  * If the difference is less than the minimum IO size, we
1371                  * are done.
1372                  */
1373                 diff = limit - ddir_rw_sum(td->io_bytes);
1374                 if (diff < td_max_bs(td))
1375                         return false;
1376
1377                 if (fio_files_done(td) && !td->o.io_limit)
1378                         return false;
1379
1380                 return true;
1381         }
1382
1383         return false;
1384 }
1385
1386 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1387 {
1388         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1389         int ret;
1390         char *str;
1391
1392         str = malloc(newlen);
1393         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1394
1395         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1396         ret = system(str);
1397         if (ret == -1)
1398                 log_err("fio: exec of cmd <%s> failed\n", str);
1399
1400         free(str);
1401         return ret;
1402 }
1403
1404 /*
1405  * Dry run to compute correct state of numberio for verification.
1406  */
1407 static uint64_t do_dry_run(struct thread_data *td)
1408 {
1409         td_set_runstate(td, TD_RUNNING);
1410
1411         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1412                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1413                 struct io_u *io_u;
1414                 int ret;
1415
1416                 if (td->terminate || td->done)
1417                         break;
1418
1419                 io_u = get_io_u(td);
1420                 if (IS_ERR_OR_NULL(io_u))
1421                         break;
1422
1423                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1424                 io_u->error = 0;
1425                 io_u->resid = 0;
1426                 if (ddir_rw(acct_ddir(io_u)))
1427                         td->io_issues[acct_ddir(io_u)]++;
1428                 if (ddir_rw(io_u->ddir)) {
1429                         io_u_mark_depth(td, 1);
1430                         td->ts.total_io_u[io_u->ddir]++;
1431                 }
1432
1433                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1434                     td->o.do_verify &&
1435                     td->o.verify != VERIFY_NONE &&
1436                     !td->o.experimental_verify)
1437                         log_io_piece(td, io_u);
1438
1439                 ret = io_u_sync_complete(td, io_u);
1440                 (void) ret;
1441         }
1442
1443         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1444 }
1445
1446 struct fork_data {
1447         struct thread_data *td;
1448         struct sk_out *sk_out;
1449 };
1450
1451 /*
1452  * Entry point for the thread based jobs. The process based jobs end up
1453  * here as well, after a little setup.
1454  */
1455 static void *thread_main(void *data)
1456 {
1457         struct fork_data *fd = data;
1458         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1459         struct thread_data *td = fd->td;
1460         struct thread_options *o = &td->o;
1461         struct sk_out *sk_out = fd->sk_out;
1462         int deadlock_loop_cnt;
1463         int clear_state;
1464         int ret;
1465
1466         sk_out_assign(sk_out);
1467         free(fd);
1468
1469         if (!o->use_thread) {
1470                 setsid();
1471                 td->pid = getpid();
1472         } else
1473                 td->pid = gettid();
1474
1475         fio_local_clock_init(o->use_thread);
1476
1477         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1478
1479         if (is_backend)
1480                 fio_server_send_start(td);
1481
1482         INIT_FLIST_HEAD(&td->io_log_list);
1483         INIT_FLIST_HEAD(&td->io_hist_list);
1484         INIT_FLIST_HEAD(&td->verify_list);
1485         INIT_FLIST_HEAD(&td->trim_list);
1486         INIT_FLIST_HEAD(&td->next_rand_list);
1487         td->io_hist_tree = RB_ROOT;
1488
1489         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1490         if (ret) {
1491                 td_verror(td, ret, "mutex_cond_init_pshared");
1492                 goto err;
1493         }
1494         ret = cond_init_pshared(&td->verify_cond);
1495         if (ret) {
1496                 td_verror(td, ret, "mutex_cond_pshared");
1497                 goto err;
1498         }
1499
1500         td_set_runstate(td, TD_INITIALIZED);
1501         dprint(FD_MUTEX, "up startup_mutex\n");
1502         fio_mutex_up(startup_mutex);
1503         dprint(FD_MUTEX, "wait on td->mutex\n");
1504         fio_mutex_down(td->mutex);
1505         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1506
1507         /*
1508          * A new gid requires privilege, so we need to do this before setting
1509          * the uid.
1510          */
1511         if (o->gid != -1U && setgid(o->gid)) {
1512                 td_verror(td, errno, "setgid");
1513                 goto err;
1514         }
1515         if (o->uid != -1U && setuid(o->uid)) {
1516                 td_verror(td, errno, "setuid");
1517                 goto err;
1518         }
1519
1520         /*
1521          * Do this early, we don't want the compress threads to be limited
1522          * to the same CPUs as the IO workers. So do this before we set
1523          * any potential CPU affinity
1524          */
1525         if (iolog_compress_init(td, sk_out))
1526                 goto err;
1527
1528         /*
1529          * If we have a gettimeofday() thread, make sure we exclude that
1530          * thread from this job
1531          */
1532         if (o->gtod_cpu)
1533                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1534
1535         /*
1536          * Set affinity first, in case it has an impact on the memory
1537          * allocations.
1538          */
1539         if (fio_option_is_set(o, cpumask)) {
1540                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1541                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1542                         if (!ret) {
1543                                 log_err("fio: no CPUs set\n");
1544                                 log_err("fio: Try increasing number of available CPUs\n");
1545                                 td_verror(td, EINVAL, "cpus_split");
1546                                 goto err;
1547                         }
1548                 }
1549                 ret = fio_setaffinity(td->pid, o->cpumask);
1550                 if (ret == -1) {
1551                         td_verror(td, errno, "cpu_set_affinity");
1552                         goto err;
1553                 }
1554         }
1555
1556 #ifdef CONFIG_LIBNUMA
1557         /* numa node setup */
1558         if (fio_option_is_set(o, numa_cpunodes) ||
1559             fio_option_is_set(o, numa_memnodes)) {
1560                 struct bitmask *mask;
1561
1562                 if (numa_available() < 0) {
1563                         td_verror(td, errno, "Does not support NUMA API\n");
1564                         goto err;
1565                 }
1566
1567                 if (fio_option_is_set(o, numa_cpunodes)) {
1568                         mask = numa_parse_nodestring(o->numa_cpunodes);
1569                         ret = numa_run_on_node_mask(mask);
1570                         numa_free_nodemask(mask);
1571                         if (ret == -1) {
1572                                 td_verror(td, errno, \
1573                                         "numa_run_on_node_mask failed\n");
1574                                 goto err;
1575                         }
1576                 }
1577
1578                 if (fio_option_is_set(o, numa_memnodes)) {
1579                         mask = NULL;
1580                         if (o->numa_memnodes)
1581                                 mask = numa_parse_nodestring(o->numa_memnodes);
1582
1583                         switch (o->numa_mem_mode) {
1584                         case MPOL_INTERLEAVE:
1585                                 numa_set_interleave_mask(mask);
1586                                 break;
1587                         case MPOL_BIND:
1588                                 numa_set_membind(mask);
1589                                 break;
1590                         case MPOL_LOCAL:
1591                                 numa_set_localalloc();
1592                                 break;
1593                         case MPOL_PREFERRED:
1594                                 numa_set_preferred(o->numa_mem_prefer_node);
1595                                 break;
1596                         case MPOL_DEFAULT:
1597                         default:
1598                                 break;
1599                         }
1600
1601                         if (mask)
1602                                 numa_free_nodemask(mask);
1603
1604                 }
1605         }
1606 #endif
1607
1608         if (fio_pin_memory(td))
1609                 goto err;
1610
1611         /*
1612          * May alter parameters that init_io_u() will use, so we need to
1613          * do this first.
1614          */
1615         if (init_iolog(td))
1616                 goto err;
1617
1618         if (init_io_u(td))
1619                 goto err;
1620
1621         if (o->verify_async && verify_async_init(td))
1622                 goto err;
1623
1624         if (fio_option_is_set(o, ioprio) ||
1625             fio_option_is_set(o, ioprio_class)) {
1626                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1627                 if (ret == -1) {
1628                         td_verror(td, errno, "ioprio_set");
1629                         goto err;
1630                 }
1631         }
1632
1633         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1634                 goto err;
1635
1636         errno = 0;
1637         if (nice(o->nice) == -1 && errno != 0) {
1638                 td_verror(td, errno, "nice");
1639                 goto err;
1640         }
1641
1642         if (o->ioscheduler && switch_ioscheduler(td))
1643                 goto err;
1644
1645         if (!o->create_serialize && setup_files(td))
1646                 goto err;
1647
1648         if (td_io_init(td))
1649                 goto err;
1650
1651         if (init_random_map(td))
1652                 goto err;
1653
1654         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1655                 goto err;
1656
1657         if (o->pre_read) {
1658                 if (pre_read_files(td) < 0)
1659                         goto err;
1660         }
1661
1662         fio_verify_init(td);
1663
1664         if (rate_submit_init(td, sk_out))
1665                 goto err;
1666
1667         set_epoch_time(td, o->log_unix_epoch);
1668         fio_getrusage(&td->ru_start);
1669         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1670         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1671
1672         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1673                         o->ratemin[DDIR_TRIM]) {
1674                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1675                                         sizeof(td->bw_sample_time));
1676                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1677                                         sizeof(td->bw_sample_time));
1678                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1679                                         sizeof(td->bw_sample_time));
1680         }
1681
1682         clear_state = 0;
1683         while (keep_running(td)) {
1684                 uint64_t verify_bytes;
1685
1686                 fio_gettime(&td->start, NULL);
1687                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1688
1689                 if (clear_state) {
1690                         clear_io_state(td, 0);
1691
1692                         if (o->unlink_each_loop && unlink_all_files(td))
1693                                 break;
1694                 }
1695
1696                 prune_io_piece_log(td);
1697
1698                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1699                         verify_bytes = do_dry_run(td);
1700                 else {
1701                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1702
1703                         do_io(td, bytes_done);
1704
1705                         if (!ddir_rw_sum(bytes_done)) {
1706                                 fio_mark_td_terminate(td);
1707                                 verify_bytes = 0;
1708                         } else {
1709                                 verify_bytes = bytes_done[DDIR_WRITE] +
1710                                                 bytes_done[DDIR_TRIM];
1711                         }
1712                 }
1713
1714                 /*
1715                  * If we took too long to shut down, the main thread could
1716                  * already consider us reaped/exited. If that happens, break
1717                  * out and clean up.
1718                  */
1719                 if (td->runstate >= TD_EXITED)
1720                         break;
1721
1722                 clear_state = 1;
1723
1724                 /*
1725                  * Make sure we've successfully updated the rusage stats
1726                  * before waiting on the stat mutex. Otherwise we could have
1727                  * the stat thread holding stat mutex and waiting for
1728                  * the rusage_sem, which would never get upped because
1729                  * this thread is waiting for the stat mutex.
1730                  */
1731                 deadlock_loop_cnt = 0;
1732                 do {
1733                         check_update_rusage(td);
1734                         if (!fio_mutex_down_trylock(stat_mutex))
1735                                 break;
1736                         usleep(1000);
1737                         if (deadlock_loop_cnt++ > 5000) {
1738                                 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1739                                 td->error = EDEADLK;
1740                                 goto err;
1741                         }
1742                 } while (1);
1743
1744                 if (td_read(td) && td->io_bytes[DDIR_READ])
1745                         update_runtime(td, elapsed_us, DDIR_READ);
1746                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1747                         update_runtime(td, elapsed_us, DDIR_WRITE);
1748                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1749                         update_runtime(td, elapsed_us, DDIR_TRIM);
1750                 fio_gettime(&td->start, NULL);
1751                 fio_mutex_up(stat_mutex);
1752
1753                 if (td->error || td->terminate)
1754                         break;
1755
1756                 if (!o->do_verify ||
1757                     o->verify == VERIFY_NONE ||
1758                     td_ioengine_flagged(td, FIO_UNIDIR))
1759                         continue;
1760
1761                 clear_io_state(td, 0);
1762
1763                 fio_gettime(&td->start, NULL);
1764
1765                 do_verify(td, verify_bytes);
1766
1767                 /*
1768                  * See comment further up for why this is done here.
1769                  */
1770                 check_update_rusage(td);
1771
1772                 fio_mutex_down(stat_mutex);
1773                 update_runtime(td, elapsed_us, DDIR_READ);
1774                 fio_gettime(&td->start, NULL);
1775                 fio_mutex_up(stat_mutex);
1776
1777                 if (td->error || td->terminate)
1778                         break;
1779         }
1780
1781         td_set_runstate(td, TD_FINISHING);
1782
1783         update_rusage_stat(td);
1784         td->ts.total_run_time = mtime_since_now(&td->epoch);
1785         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1786         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1787         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1788
1789         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1790             (td->o.verify != VERIFY_NONE && td_write(td)))
1791                 verify_save_state(td->thread_number);
1792
1793         fio_unpin_memory(td);
1794
1795         td_writeout_logs(td, true);
1796
1797         iolog_compress_exit(td);
1798         rate_submit_exit(td);
1799
1800         if (o->exec_postrun)
1801                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1802
1803         if (exitall_on_terminate || (o->exitall_error && td->error))
1804                 fio_terminate_threads(td->groupid);
1805
1806 err:
1807         if (td->error)
1808                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1809                                                         td->verror);
1810
1811         if (o->verify_async)
1812                 verify_async_exit(td);
1813
1814         close_and_free_files(td);
1815         cleanup_io_u(td);
1816         close_ioengine(td);
1817         cgroup_shutdown(td, &cgroup_mnt);
1818         verify_free_state(td);
1819
1820         if (td->zone_state_index) {
1821                 int i;
1822
1823                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1824                         free(td->zone_state_index[i]);
1825                 free(td->zone_state_index);
1826                 td->zone_state_index = NULL;
1827         }
1828
1829         if (fio_option_is_set(o, cpumask)) {
1830                 ret = fio_cpuset_exit(&o->cpumask);
1831                 if (ret)
1832                         td_verror(td, ret, "fio_cpuset_exit");
1833         }
1834
1835         /*
1836          * do this very late, it will log file closing as well
1837          */
1838         if (o->write_iolog_file)
1839                 write_iolog_close(td);
1840
1841         fio_mutex_remove(td->mutex);
1842         td->mutex = NULL;
1843
1844         td_set_runstate(td, TD_EXITED);
1845
1846         /*
1847          * Do this last after setting our runstate to exited, so we
1848          * know that the stat thread is signaled.
1849          */
1850         check_update_rusage(td);
1851
1852         sk_out_drop();
1853         return (void *) (uintptr_t) td->error;
1854 }
1855
1856 static void dump_td_info(struct thread_data *td)
1857 {
1858         log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
1859                 "appears to be stuck. Doing forceful exit of this job.\n",
1860                         td->o.name, td->runstate,
1861                         (unsigned long) time_since_now(&td->terminate_time));
1862 }
1863
1864 /*
1865  * Run over the job map and reap the threads that have exited, if any.
1866  */
1867 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1868                          unsigned int *m_rate)
1869 {
1870         struct thread_data *td;
1871         unsigned int cputhreads, realthreads, pending;
1872         int i, status, ret;
1873
1874         /*
1875          * reap exited threads (TD_EXITED -> TD_REAPED)
1876          */
1877         realthreads = pending = cputhreads = 0;
1878         for_each_td(td, i) {
1879                 int flags = 0;
1880
1881                 /*
1882                  * ->io_ops is NULL for a thread that has closed its
1883                  * io engine
1884                  */
1885                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1886                         cputhreads++;
1887                 else
1888                         realthreads++;
1889
1890                 if (!td->pid) {
1891                         pending++;
1892                         continue;
1893                 }
1894                 if (td->runstate == TD_REAPED)
1895                         continue;
1896                 if (td->o.use_thread) {
1897                         if (td->runstate == TD_EXITED) {
1898                                 td_set_runstate(td, TD_REAPED);
1899                                 goto reaped;
1900                         }
1901                         continue;
1902                 }
1903
1904                 flags = WNOHANG;
1905                 if (td->runstate == TD_EXITED)
1906                         flags = 0;
1907
1908                 /*
1909                  * check if someone quit or got killed in an unusual way
1910                  */
1911                 ret = waitpid(td->pid, &status, flags);
1912                 if (ret < 0) {
1913                         if (errno == ECHILD) {
1914                                 log_err("fio: pid=%d disappeared %d\n",
1915                                                 (int) td->pid, td->runstate);
1916                                 td->sig = ECHILD;
1917                                 td_set_runstate(td, TD_REAPED);
1918                                 goto reaped;
1919                         }
1920                         perror("waitpid");
1921                 } else if (ret == td->pid) {
1922                         if (WIFSIGNALED(status)) {
1923                                 int sig = WTERMSIG(status);
1924
1925                                 if (sig != SIGTERM && sig != SIGUSR2)
1926                                         log_err("fio: pid=%d, got signal=%d\n",
1927                                                         (int) td->pid, sig);
1928                                 td->sig = sig;
1929                                 td_set_runstate(td, TD_REAPED);
1930                                 goto reaped;
1931                         }
1932                         if (WIFEXITED(status)) {
1933                                 if (WEXITSTATUS(status) && !td->error)
1934                                         td->error = WEXITSTATUS(status);
1935
1936                                 td_set_runstate(td, TD_REAPED);
1937                                 goto reaped;
1938                         }
1939                 }
1940
1941                 /*
1942                  * If the job is stuck, do a forceful timeout of it and
1943                  * move on.
1944                  */
1945                 if (td->terminate &&
1946                     td->runstate < TD_FSYNCING &&
1947                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1948                         dump_td_info(td);
1949                         td_set_runstate(td, TD_REAPED);
1950                         goto reaped;
1951                 }
1952
1953                 /*
1954                  * thread is not dead, continue
1955                  */
1956                 pending++;
1957                 continue;
1958 reaped:
1959                 (*nr_running)--;
1960                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1961                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1962                 if (!td->pid)
1963                         pending--;
1964
1965                 if (td->error)
1966                         exit_value++;
1967
1968                 done_secs += mtime_since_now(&td->epoch) / 1000;
1969                 profile_td_exit(td);
1970         }
1971
1972         if (*nr_running == cputhreads && !pending && realthreads)
1973                 fio_terminate_threads(TERMINATE_ALL);
1974 }
1975
1976 static bool __check_trigger_file(void)
1977 {
1978         struct stat sb;
1979
1980         if (!trigger_file)
1981                 return false;
1982
1983         if (stat(trigger_file, &sb))
1984                 return false;
1985
1986         if (unlink(trigger_file) < 0)
1987                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1988                                                         strerror(errno));
1989
1990         return true;
1991 }
1992
1993 static bool trigger_timedout(void)
1994 {
1995         if (trigger_timeout)
1996                 return time_since_genesis() >= trigger_timeout;
1997
1998         return false;
1999 }
2000
2001 void exec_trigger(const char *cmd)
2002 {
2003         int ret;
2004
2005         if (!cmd)
2006                 return;
2007
2008         ret = system(cmd);
2009         if (ret == -1)
2010                 log_err("fio: failed executing %s trigger\n", cmd);
2011 }
2012
2013 void check_trigger_file(void)
2014 {
2015         if (__check_trigger_file() || trigger_timedout()) {
2016                 if (nr_clients)
2017                         fio_clients_send_trigger(trigger_remote_cmd);
2018                 else {
2019                         verify_save_state(IO_LIST_ALL);
2020                         fio_terminate_threads(TERMINATE_ALL);
2021                         exec_trigger(trigger_cmd);
2022                 }
2023         }
2024 }
2025
2026 static int fio_verify_load_state(struct thread_data *td)
2027 {
2028         int ret;
2029
2030         if (!td->o.verify_state)
2031                 return 0;
2032
2033         if (is_backend) {
2034                 void *data;
2035
2036                 ret = fio_server_get_verify_state(td->o.name,
2037                                         td->thread_number - 1, &data);
2038                 if (!ret)
2039                         verify_assign_state(td, data);
2040         } else
2041                 ret = verify_load_state(td, "local");
2042
2043         return ret;
2044 }
2045
2046 static void do_usleep(unsigned int usecs)
2047 {
2048         check_for_running_stats();
2049         check_trigger_file();
2050         usleep(usecs);
2051 }
2052
2053 static bool check_mount_writes(struct thread_data *td)
2054 {
2055         struct fio_file *f;
2056         unsigned int i;
2057
2058         if (!td_write(td) || td->o.allow_mounted_write)
2059                 return false;
2060
2061         for_each_file(td, f, i) {
2062                 if (f->filetype != FIO_TYPE_BD)
2063                         continue;
2064                 if (device_is_mounted(f->file_name))
2065                         goto mounted;
2066         }
2067
2068         return false;
2069 mounted:
2070         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2071         return true;
2072 }
2073
2074 static bool waitee_running(struct thread_data *me)
2075 {
2076         const char *waitee = me->o.wait_for;
2077         const char *self = me->o.name;
2078         struct thread_data *td;
2079         int i;
2080
2081         if (!waitee)
2082                 return false;
2083
2084         for_each_td(td, i) {
2085                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2086                         continue;
2087
2088                 if (td->runstate < TD_EXITED) {
2089                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2090                                         self, td->o.name,
2091                                         runstate_to_name(td->runstate));
2092                         return true;
2093                 }
2094         }
2095
2096         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2097         return false;
2098 }
2099
2100 /*
2101  * Main function for kicking off and reaping jobs, as needed.
2102  */
2103 static void run_threads(struct sk_out *sk_out)
2104 {
2105         struct thread_data *td;
2106         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2107         uint64_t spent;
2108
2109         if (fio_gtod_offload && fio_start_gtod_thread())
2110                 return;
2111
2112         fio_idle_prof_init();
2113
2114         set_sig_handlers();
2115
2116         nr_thread = nr_process = 0;
2117         for_each_td(td, i) {
2118                 if (check_mount_writes(td))
2119                         return;
2120                 if (td->o.use_thread)
2121                         nr_thread++;
2122                 else
2123                         nr_process++;
2124         }
2125
2126         if (output_format & FIO_OUTPUT_NORMAL) {
2127                 log_info("Starting ");
2128                 if (nr_thread)
2129                         log_info("%d thread%s", nr_thread,
2130                                                 nr_thread > 1 ? "s" : "");
2131                 if (nr_process) {
2132                         if (nr_thread)
2133                                 log_info(" and ");
2134                         log_info("%d process%s", nr_process,
2135                                                 nr_process > 1 ? "es" : "");
2136                 }
2137                 log_info("\n");
2138                 log_info_flush();
2139         }
2140
2141         todo = thread_number;
2142         nr_running = 0;
2143         nr_started = 0;
2144         m_rate = t_rate = 0;
2145
2146         for_each_td(td, i) {
2147                 print_status_init(td->thread_number - 1);
2148
2149                 if (!td->o.create_serialize)
2150                         continue;
2151
2152                 if (fio_verify_load_state(td))
2153                         goto reap;
2154
2155                 /*
2156                  * do file setup here so it happens sequentially,
2157                  * we don't want X number of threads getting their
2158                  * client data interspersed on disk
2159                  */
2160                 if (setup_files(td)) {
2161 reap:
2162                         exit_value++;
2163                         if (td->error)
2164                                 log_err("fio: pid=%d, err=%d/%s\n",
2165                                         (int) td->pid, td->error, td->verror);
2166                         td_set_runstate(td, TD_REAPED);
2167                         todo--;
2168                 } else {
2169                         struct fio_file *f;
2170                         unsigned int j;
2171
2172                         /*
2173                          * for sharing to work, each job must always open
2174                          * its own files. so close them, if we opened them
2175                          * for creation
2176                          */
2177                         for_each_file(td, f, j) {
2178                                 if (fio_file_open(f))
2179                                         td_io_close_file(td, f);
2180                         }
2181                 }
2182         }
2183
2184         /* start idle threads before io threads start to run */
2185         fio_idle_prof_start();
2186
2187         set_genesis_time();
2188
2189         while (todo) {
2190                 struct thread_data *map[REAL_MAX_JOBS];
2191                 struct timeval this_start;
2192                 int this_jobs = 0, left;
2193                 struct fork_data *fd;
2194
2195                 /*
2196                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2197                  */
2198                 for_each_td(td, i) {
2199                         if (td->runstate != TD_NOT_CREATED)
2200                                 continue;
2201
2202                         /*
2203                          * never got a chance to start, killed by other
2204                          * thread for some reason
2205                          */
2206                         if (td->terminate) {
2207                                 todo--;
2208                                 continue;
2209                         }
2210
2211                         if (td->o.start_delay) {
2212                                 spent = utime_since_genesis();
2213
2214                                 if (td->o.start_delay > spent)
2215                                         continue;
2216                         }
2217
2218                         if (td->o.stonewall && (nr_started || nr_running)) {
2219                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2220                                                         td->o.name);
2221                                 break;
2222                         }
2223
2224                         if (waitee_running(td)) {
2225                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2226                                                 td->o.name, td->o.wait_for);
2227                                 continue;
2228                         }
2229
2230                         init_disk_util(td);
2231
2232                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2233                         td->update_rusage = 0;
2234
2235                         /*
2236                          * Set state to created. Thread will transition
2237                          * to TD_INITIALIZED when it's done setting up.
2238                          */
2239                         td_set_runstate(td, TD_CREATED);
2240                         map[this_jobs++] = td;
2241                         nr_started++;
2242
2243                         fd = calloc(1, sizeof(*fd));
2244                         fd->td = td;
2245                         fd->sk_out = sk_out;
2246
2247                         if (td->o.use_thread) {
2248                                 int ret;
2249
2250                                 dprint(FD_PROCESS, "will pthread_create\n");
2251                                 ret = pthread_create(&td->thread, NULL,
2252                                                         thread_main, fd);
2253                                 if (ret) {
2254                                         log_err("pthread_create: %s\n",
2255                                                         strerror(ret));
2256                                         free(fd);
2257                                         nr_started--;
2258                                         break;
2259                                 }
2260                                 ret = pthread_detach(td->thread);
2261                                 if (ret)
2262                                         log_err("pthread_detach: %s",
2263                                                         strerror(ret));
2264                         } else {
2265                                 pid_t pid;
2266                                 dprint(FD_PROCESS, "will fork\n");
2267                                 pid = fork();
2268                                 if (!pid) {
2269                                         int ret;
2270
2271                                         ret = (int)(uintptr_t)thread_main(fd);
2272                                         _exit(ret);
2273                                 } else if (i == fio_debug_jobno)
2274                                         *fio_debug_jobp = pid;
2275                         }
2276                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2277                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2278                                 log_err("fio: job startup hung? exiting.\n");
2279                                 fio_terminate_threads(TERMINATE_ALL);
2280                                 fio_abort = 1;
2281                                 nr_started--;
2282                                 break;
2283                         }
2284                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2285                 }
2286
2287                 /*
2288                  * Wait for the started threads to transition to
2289                  * TD_INITIALIZED.
2290                  */
2291                 fio_gettime(&this_start, NULL);
2292                 left = this_jobs;
2293                 while (left && !fio_abort) {
2294                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2295                                 break;
2296
2297                         do_usleep(100000);
2298
2299                         for (i = 0; i < this_jobs; i++) {
2300                                 td = map[i];
2301                                 if (!td)
2302                                         continue;
2303                                 if (td->runstate == TD_INITIALIZED) {
2304                                         map[i] = NULL;
2305                                         left--;
2306                                 } else if (td->runstate >= TD_EXITED) {
2307                                         map[i] = NULL;
2308                                         left--;
2309                                         todo--;
2310                                         nr_running++; /* work-around... */
2311                                 }
2312                         }
2313                 }
2314
2315                 if (left) {
2316                         log_err("fio: %d job%s failed to start\n", left,
2317                                         left > 1 ? "s" : "");
2318                         for (i = 0; i < this_jobs; i++) {
2319                                 td = map[i];
2320                                 if (!td)
2321                                         continue;
2322                                 kill(td->pid, SIGTERM);
2323                         }
2324                         break;
2325                 }
2326
2327                 /*
2328                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2329                  */
2330                 for_each_td(td, i) {
2331                         if (td->runstate != TD_INITIALIZED)
2332                                 continue;
2333
2334                         if (in_ramp_time(td))
2335                                 td_set_runstate(td, TD_RAMP);
2336                         else
2337                                 td_set_runstate(td, TD_RUNNING);
2338                         nr_running++;
2339                         nr_started--;
2340                         m_rate += ddir_rw_sum(td->o.ratemin);
2341                         t_rate += ddir_rw_sum(td->o.rate);
2342                         todo--;
2343                         fio_mutex_up(td->mutex);
2344                 }
2345
2346                 reap_threads(&nr_running, &t_rate, &m_rate);
2347
2348                 if (todo)
2349                         do_usleep(100000);
2350         }
2351
2352         while (nr_running) {
2353                 reap_threads(&nr_running, &t_rate, &m_rate);
2354                 do_usleep(10000);
2355         }
2356
2357         fio_idle_prof_stop();
2358
2359         update_io_ticks();
2360 }
2361
2362 static void free_disk_util(void)
2363 {
2364         disk_util_prune_entries();
2365         helper_thread_destroy();
2366 }
2367
2368 int fio_backend(struct sk_out *sk_out)
2369 {
2370         struct thread_data *td;
2371         int i;
2372
2373         if (exec_profile) {
2374                 if (load_profile(exec_profile))
2375                         return 1;
2376                 free(exec_profile);
2377                 exec_profile = NULL;
2378         }
2379         if (!thread_number)
2380                 return 0;
2381
2382         if (write_bw_log) {
2383                 struct log_params p = {
2384                         .log_type = IO_LOG_TYPE_BW,
2385                 };
2386
2387                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2388                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2389                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2390         }
2391
2392         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2393         if (startup_mutex == NULL)
2394                 return 1;
2395
2396         set_genesis_time();
2397         stat_init();
2398         helper_thread_create(startup_mutex, sk_out);
2399
2400         cgroup_list = smalloc(sizeof(*cgroup_list));
2401         INIT_FLIST_HEAD(cgroup_list);
2402
2403         run_threads(sk_out);
2404
2405         helper_thread_exit();
2406
2407         if (!fio_abort) {
2408                 __show_run_stats();
2409                 if (write_bw_log) {
2410                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2411                                 struct io_log *log = agg_io_log[i];
2412
2413                                 flush_log(log, false);
2414                                 free_log(log);
2415                         }
2416                 }
2417         }
2418
2419         for_each_td(td, i) {
2420                 fio_options_free(td);
2421                 if (td->rusage_sem) {
2422                         fio_mutex_remove(td->rusage_sem);
2423                         td->rusage_sem = NULL;
2424                 }
2425         }
2426
2427         free_disk_util();
2428         cgroup_kill(cgroup_list);
2429         sfree(cgroup_list);
2430         sfree(cgroup_mnt);
2431
2432         fio_mutex_remove(startup_mutex);
2433         stat_exit();
2434         return exit_value;
2435 }