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 /*
 576  * The main verify engine. Runs over the writes we previously submitted,
 577  * reads the blocks back in, and checks the crc/md5 of the data.
 578  */
 579 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 580 {
 581         struct fio_file *f;
 582         struct io_u *io_u;
 583         int ret, min_events;
 584         unsigned int i;
 585
 586         dprint(FD_VERIFY, "starting loop\n");
 587
 588         /*
 589          * sync io first and invalidate cache, to make sure we really
 590          * read from disk.
 591          */
 592         for_each_file(td, f, i) {
 593                 if (!fio_file_open(f))
 594                         continue;
 595                 if (fio_io_sync(td, f))
 596                         break;
 597                 if (file_invalidate_cache(td, f))
 598                         break;
 599         }
 600
 601         check_update_rusage(td);
 602
 603         if (td->error)
 604                 return;
 605
 606         /*
 607          * verify_state needs to be reset before verification
 608          * proceeds so that expected random seeds match actual
 609          * random seeds in headers. The main loop will reset
 610          * all random number generators if randrepeat is set.
 611          */
 612         if (!td->o.rand_repeatable)
 613                 td_fill_verify_state_seed(td);
 614
 615         td_set_runstate(td, TD_VERIFYING);
 616
 617         io_u = NULL;
 618         while (!td->terminate) {
 619                 enum fio_ddir ddir;
 620                 int full;
 621
 622                 update_tv_cache(td);
 623                 check_update_rusage(td);
 624
 625                 if (runtime_exceeded(td, &td->tv_cache)) {
 626                         __update_tv_cache(td);
 627                         if (runtime_exceeded(td, &td->tv_cache)) {
 628                                 fio_mark_td_terminate(td);
 629                                 break;
 630                         }
 631                 }
 632
 633                 if (flow_threshold_exceeded(td))
 634                         continue;
 635
 636                 if (!td->o.experimental_verify) {
 637                         io_u = __get_io_u(td);
 638                         if (!io_u)
 639                                 break;
 640
 641                         if (get_next_verify(td, io_u)) {
 642                                 put_io_u(td, io_u);
 643                                 break;
 644                         }
 645
 646                         if (td_io_prep(td, io_u)) {
 647                                 put_io_u(td, io_u);
 648                                 break;
 649                         }
 650                 } else {
 651                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 652                                 break;
 653
 654                         while ((io_u = get_io_u(td)) != NULL) {
 655                                 if (IS_ERR_OR_NULL(io_u)) {
 656                                         io_u = NULL;
 657                                         ret = FIO_Q_BUSY;
 658                                         goto reap;
 659                                 }
 660
 661                                 /*
 662                                  * We are only interested in the places where
 663                                  * we wrote or trimmed IOs. Turn those into
 664                                  * reads for verification purposes.
 665                                  */
 666                                 if (io_u->ddir == DDIR_READ) {
 667                                         /*
 668                                          * Pretend we issued it for rwmix
 669                                          * accounting
 670                                          */
 671                                         td->io_issues[DDIR_READ]++;
 672                                         put_io_u(td, io_u);
 673                                         continue;
 674                                 } else if (io_u->ddir == DDIR_TRIM) {
 675                                         io_u->ddir = DDIR_READ;
 676                                         io_u_set(io_u, IO_U_F_TRIMMED);
 677                                         break;
 678                                 } else if (io_u->ddir == DDIR_WRITE) {
 679                                         io_u->ddir = DDIR_READ;
 680                                         break;
 681                                 } else {
 682                                         put_io_u(td, io_u);
 683                                         continue;
 684                                 }
 685                         }
 686
 687                         if (!io_u)
 688                                 break;
 689                 }
 690
 691                 if (verify_state_should_stop(td, io_u)) {
 692                         put_io_u(td, io_u);
 693                         break;
 694                 }
 695
 696                 if (td->o.verify_async)
 697                         io_u->end_io = verify_io_u_async;
 698                 else
 699                         io_u->end_io = verify_io_u;
 700
 701                 ddir = io_u->ddir;
 702                 if (!td->o.disable_slat)
 703                         fio_gettime(&io_u->start_time, NULL);
 704
 705                 ret = td_io_queue(td, io_u);
 706
 707                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 708                         break;
 709
 710                 /*
 711                  * if we can queue more, do so. but check if there are
 712                  * completed io_u's first. Note that we can get BUSY even
 713                  * without IO queued, if the system is resource starved.
 714                  */
 715 reap:
 716                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 717                 if (full || io_in_polling(td))
 718                         ret = wait_for_completions(td, NULL);
 719
 720                 if (ret < 0)
 721                         break;
 722         }
 723
 724         check_update_rusage(td);
 725
 726         if (!td->error) {
 727                 min_events = td->cur_depth;
 728
 729                 if (min_events)
 730                         ret = io_u_queued_complete(td, min_events);
 731         } else
 732                 cleanup_pending_aio(td);
 733
 734         td_set_runstate(td, TD_RUNNING);
 735
 736         dprint(FD_VERIFY, "exiting loop\n");
 737 }
 738
 739 static bool exceeds_number_ios(struct thread_data *td)
 740 {
 741         unsigned long long number_ios;
 742
 743         if (!td->o.number_ios)
 744                 return false;
 745
 746         number_ios = ddir_rw_sum(td->io_blocks);
 747         number_ios += td->io_u_queued + td->io_u_in_flight;
 748
 749         return number_ios >= (td->o.number_ios * td->loops);
 750 }
 751
 752 static bool io_issue_bytes_exceeded(struct thread_data *td)
 753 {
 754         unsigned long long bytes, limit;
 755
 756         if (td_rw(td))
 757                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 758         else if (td_write(td))
 759                 bytes = td->io_issue_bytes[DDIR_WRITE];
 760         else if (td_read(td))
 761                 bytes = td->io_issue_bytes[DDIR_READ];
 762         else
 763                 bytes = td->io_issue_bytes[DDIR_TRIM];
 764
 765         if (td->o.io_limit)
 766                 limit = td->o.io_limit;
 767         else
 768                 limit = td->o.size;
 769
 770         limit *= td->loops;
 771         return bytes >= limit || exceeds_number_ios(td);
 772 }
 773
 774 static bool io_complete_bytes_exceeded(struct thread_data *td)
 775 {
 776         unsigned long long bytes, limit;
 777
 778         if (td_rw(td))
 779                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 780         else if (td_write(td))
 781                 bytes = td->this_io_bytes[DDIR_WRITE];
 782         else if (td_read(td))
 783                 bytes = td->this_io_bytes[DDIR_READ];
 784         else
 785                 bytes = td->this_io_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 /*
 797  * used to calculate the next io time for rate control
 798  *
 799  */
 800 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 801 {
 802         uint64_t secs, remainder, bps, bytes, iops;
 803
 804         assert(!(td->flags & TD_F_CHILD));
 805         bytes = td->rate_io_issue_bytes[ddir];
 806         bps = td->rate_bps[ddir];
 807
 808         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 809                 uint64_t val;
 810                 iops = bps / td->o.bs[ddir];
 811                 val = (int64_t) (1000000 / iops) *
 812                                 -logf(__rand_0_1(&td->poisson_state));
 813                 if (val) {
 814                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 815                                         (unsigned long long) 1000000 / val);
 816                 }
 817                 td->last_usec += val;
 818                 return td->last_usec;
 819         } else if (bps) {
 820                 secs = bytes / bps;
 821                 remainder = bytes % bps;
 822                 return remainder * 1000000 / bps + secs * 1000000;
 823         }
 824
 825         return 0;
 826 }
 827
 828 /*
 829  * Main IO worker function. It retrieves io_u's to process and queues
 830  * and reaps them, checking for rate and errors along the way.
 831  *
 832  * Returns number of bytes written and trimmed.
 833  */
 834 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 835 {
 836         unsigned int i;
 837         int ret = 0;
 838         uint64_t total_bytes, bytes_issued = 0;
 839
 840         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 841                 bytes_done[i] = td->bytes_done[i];
 842
 843         if (in_ramp_time(td))
 844                 td_set_runstate(td, TD_RAMP);
 845         else
 846                 td_set_runstate(td, TD_RUNNING);
 847
 848         lat_target_init(td);
 849
 850         total_bytes = td->o.size;
 851         /*
 852         * Allow random overwrite workloads to write up to io_limit
 853         * before starting verification phase as 'size' doesn't apply.
 854         */
 855         if (td_write(td) && td_random(td) && td->o.norandommap)
 856                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 857         /*
 858          * If verify_backlog is enabled, we'll run the verify in this
 859          * handler as well. For that case, we may need up to twice the
 860          * amount of bytes.
 861          */
 862         if (td->o.verify != VERIFY_NONE &&
 863            (td_write(td) && td->o.verify_backlog))
 864                 total_bytes += td->o.size;
 865
 866         /* In trimwrite mode, each byte is trimmed and then written, so
 867          * allow total_bytes to be twice as big */
 868         if (td_trimwrite(td))
 869                 total_bytes += td->total_io_size;
 870
 871         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 872                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 873                 td->o.time_based) {
 874                 struct timeval comp_time;
 875                 struct io_u *io_u;
 876                 int full;
 877                 enum fio_ddir ddir;
 878
 879                 check_update_rusage(td);
 880
 881                 if (td->terminate || td->done)
 882                         break;
 883
 884                 update_tv_cache(td);
 885
 886                 if (runtime_exceeded(td, &td->tv_cache)) {
 887                         __update_tv_cache(td);
 888                         if (runtime_exceeded(td, &td->tv_cache)) {
 889                                 fio_mark_td_terminate(td);
 890                                 break;
 891                         }
 892                 }
 893
 894                 if (flow_threshold_exceeded(td))
 895                         continue;
 896
 897                 /*
 898                  * Break if we exceeded the bytes. The exception is time
 899                  * based runs, but we still need to break out of the loop
 900                  * for those to run verification, if enabled.
 901                  */
 902                 if (bytes_issued >= total_bytes &&
 903                     (!td->o.time_based ||
 904                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 905                         break;
 906
 907                 io_u = get_io_u(td);
 908                 if (IS_ERR_OR_NULL(io_u)) {
 909                         int err = PTR_ERR(io_u);
 910
 911                         io_u = NULL;
 912                         if (err == -EBUSY) {
 913                                 ret = FIO_Q_BUSY;
 914                                 goto reap;
 915                         }
 916                         if (td->o.latency_target)
 917                                 goto reap;
 918                         break;
 919                 }
 920
 921                 ddir = io_u->ddir;
 922
 923                 /*
 924                  * Add verification end_io handler if:
 925                  *      - Asked to verify (!td_rw(td))
 926                  *      - Or the io_u is from our verify list (mixed write/ver)
 927                  */
 928                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 929                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 930
 931                         if (!td->o.verify_pattern_bytes) {
 932                                 io_u->rand_seed = __rand(&td->verify_state);
 933                                 if (sizeof(int) != sizeof(long *))
 934                                         io_u->rand_seed *= __rand(&td->verify_state);
 935                         }
 936
 937                         if (verify_state_should_stop(td, io_u)) {
 938                                 put_io_u(td, io_u);
 939                                 break;
 940                         }
 941
 942                         if (td->o.verify_async)
 943                                 io_u->end_io = verify_io_u_async;
 944                         else
 945                                 io_u->end_io = verify_io_u;
 946                         td_set_runstate(td, TD_VERIFYING);
 947                 } else if (in_ramp_time(td))
 948                         td_set_runstate(td, TD_RAMP);
 949                 else
 950                         td_set_runstate(td, TD_RUNNING);
 951
 952                 /*
 953                  * Always log IO before it's issued, so we know the specific
 954                  * order of it. The logged unit will track when the IO has
 955                  * completed.
 956                  */
 957                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 958                     td->o.do_verify &&
 959                     td->o.verify != VERIFY_NONE &&
 960                     !td->o.experimental_verify)
 961                         log_io_piece(td, io_u);
 962
 963                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 964                         const unsigned long blen = io_u->xfer_buflen;
 965                         const enum fio_ddir ddir = acct_ddir(io_u);
 966
 967                         if (td->error)
 968                                 break;
 969
 970                         workqueue_enqueue(&td->io_wq, &io_u->work);
 971                         ret = FIO_Q_QUEUED;
 972
 973                         if (ddir_rw(ddir)) {
 974                                 td->io_issues[ddir]++;
 975                                 td->io_issue_bytes[ddir] += blen;
 976                                 td->rate_io_issue_bytes[ddir] += blen;
 977                         }
 978
 979                         if (should_check_rate(td))
 980                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 981
 982                 } else {
 983                         ret = td_io_queue(td, io_u);
 984
 985                         if (should_check_rate(td))
 986                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 987
 988                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 989                                 break;
 990
 991                         /*
 992                          * See if we need to complete some commands. Note that
 993                          * we can get BUSY even without IO queued, if the
 994                          * system is resource starved.
 995                          */
 996 reap:
 997                         full = queue_full(td) ||
 998                                 (ret == FIO_Q_BUSY && td->cur_depth);
 999                         if (full || io_in_polling(td))
1000                                 ret = wait_for_completions(td, &comp_time);
1001                 }
1002                 if (ret < 0)
1003                         break;
1004                 if (!ddir_rw_sum(td->bytes_done) &&
1005                     !(td->io_ops->flags & FIO_NOIO))
1006                         continue;
1007
1008                 if (!in_ramp_time(td) && should_check_rate(td)) {
1009                         if (check_min_rate(td, &comp_time)) {
1010                                 if (exitall_on_terminate || td->o.exitall_error)
1011                                         fio_terminate_threads(td->groupid);
1012                                 td_verror(td, EIO, "check_min_rate");
1013                                 break;
1014                         }
1015                 }
1016                 if (!in_ramp_time(td) && td->o.latency_target)
1017                         lat_target_check(td);
1018
1019                 if (td->o.thinktime) {
1020                         unsigned long long b;
1021
1022                         b = ddir_rw_sum(td->io_blocks);
1023                         if (!(b % td->o.thinktime_blocks)) {
1024                                 int left;
1025
1026                                 io_u_quiesce(td);
1027
1028                                 if (td->o.thinktime_spin)
1029                                         usec_spin(td->o.thinktime_spin);
1030
1031                                 left = td->o.thinktime - td->o.thinktime_spin;
1032                                 if (left)
1033                                         usec_sleep(td, left);
1034                         }
1035                 }
1036         }
1037
1038         check_update_rusage(td);
1039
1040         if (td->trim_entries)
1041                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1042
1043         if (td->o.fill_device && td->error == ENOSPC) {
1044                 td->error = 0;
1045                 fio_mark_td_terminate(td);
1046         }
1047         if (!td->error) {
1048                 struct fio_file *f;
1049
1050                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1051                         workqueue_flush(&td->io_wq);
1052                         i = 0;
1053                 } else
1054                         i = td->cur_depth;
1055
1056                 if (i) {
1057                         ret = io_u_queued_complete(td, i);
1058                         if (td->o.fill_device && td->error == ENOSPC)
1059                                 td->error = 0;
1060                 }
1061
1062                 if (should_fsync(td) && td->o.end_fsync) {
1063                         td_set_runstate(td, TD_FSYNCING);
1064
1065                         for_each_file(td, f, i) {
1066                                 if (!fio_file_fsync(td, f))
1067                                         continue;
1068
1069                                 log_err("fio: end_fsync failed for file %s\n",
1070                                                                 f->file_name);
1071                         }
1072                 }
1073         } else
1074                 cleanup_pending_aio(td);
1075
1076         /*
1077          * stop job if we failed doing any IO
1078          */
1079         if (!ddir_rw_sum(td->this_io_bytes))
1080                 td->done = 1;
1081
1082         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1083                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1084 }
1085
1086 static void free_file_completion_logging(struct thread_data *td)
1087 {
1088         struct fio_file *f;
1089         unsigned int i;
1090
1091         for_each_file(td, f, i) {
1092                 if (!f->last_write_comp)
1093                         break;
1094                 sfree(f->last_write_comp);
1095         }
1096 }
1097
1098 static int init_file_completion_logging(struct thread_data *td,
1099                                         unsigned int depth)
1100 {
1101         struct fio_file *f;
1102         unsigned int i;
1103
1104         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1105                 return 0;
1106
1107         for_each_file(td, f, i) {
1108                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1109                 if (!f->last_write_comp)
1110                         goto cleanup;
1111         }
1112
1113         return 0;
1114
1115 cleanup:
1116         free_file_completion_logging(td);
1117         log_err("fio: failed to alloc write comp data\n");
1118         return 1;
1119 }
1120
1121 static void cleanup_io_u(struct thread_data *td)
1122 {
1123         struct io_u *io_u;
1124
1125         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1126
1127                 if (td->io_ops->io_u_free)
1128                         td->io_ops->io_u_free(td, io_u);
1129
1130                 fio_memfree(io_u, sizeof(*io_u));
1131         }
1132
1133         free_io_mem(td);
1134
1135         io_u_rexit(&td->io_u_requeues);
1136         io_u_qexit(&td->io_u_freelist);
1137         io_u_qexit(&td->io_u_all);
1138
1139         free_file_completion_logging(td);
1140 }
1141
1142 static int init_io_u(struct thread_data *td)
1143 {
1144         struct io_u *io_u;
1145         unsigned int max_bs, min_write;
1146         int cl_align, i, max_units;
1147         int data_xfer = 1, err;
1148         char *p;
1149
1150         max_units = td->o.iodepth;
1151         max_bs = td_max_bs(td);
1152         min_write = td->o.min_bs[DDIR_WRITE];
1153         td->orig_buffer_size = (unsigned long long) max_bs
1154                                         * (unsigned long long) max_units;
1155
1156         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1157                 data_xfer = 0;
1158
1159         err = 0;
1160         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1161         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1162         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1163
1164         if (err) {
1165                 log_err("fio: failed setting up IO queues\n");
1166                 return 1;
1167         }
1168
1169         /*
1170          * if we may later need to do address alignment, then add any
1171          * possible adjustment here so that we don't cause a buffer
1172          * overflow later. this adjustment may be too much if we get
1173          * lucky and the allocator gives us an aligned address.
1174          */
1175         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1176             (td->io_ops->flags & FIO_RAWIO))
1177                 td->orig_buffer_size += page_mask + td->o.mem_align;
1178
1179         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1180                 unsigned long bs;
1181
1182                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1183                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1184         }
1185
1186         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1187                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1188                 return 1;
1189         }
1190
1191         if (data_xfer && allocate_io_mem(td))
1192                 return 1;
1193
1194         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1195             (td->io_ops->flags & FIO_RAWIO))
1196                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1197         else
1198                 p = td->orig_buffer;
1199
1200         cl_align = os_cache_line_size();
1201
1202         for (i = 0; i < max_units; i++) {
1203                 void *ptr;
1204
1205                 if (td->terminate)
1206                         return 1;
1207
1208                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1209                 if (!ptr) {
1210                         log_err("fio: unable to allocate aligned memory\n");
1211                         break;
1212                 }
1213
1214                 io_u = ptr;
1215                 memset(io_u, 0, sizeof(*io_u));
1216                 INIT_FLIST_HEAD(&io_u->verify_list);
1217                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1218
1219                 if (data_xfer) {
1220                         io_u->buf = p;
1221                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1222
1223                         if (td_write(td))
1224                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1225                         if (td_write(td) && td->o.verify_pattern_bytes) {
1226                                 /*
1227                                  * Fill the buffer with the pattern if we are
1228                                  * going to be doing writes.
1229                                  */
1230                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1231                         }
1232                 }
1233
1234                 io_u->index = i;
1235                 io_u->flags = IO_U_F_FREE;
1236                 io_u_qpush(&td->io_u_freelist, io_u);
1237
1238                 /*
1239                  * io_u never leaves this stack, used for iteration of all
1240                  * io_u buffers.
1241                  */
1242                 io_u_qpush(&td->io_u_all, io_u);
1243
1244                 if (td->io_ops->io_u_init) {
1245                         int ret = td->io_ops->io_u_init(td, io_u);
1246
1247                         if (ret) {
1248                                 log_err("fio: failed to init engine data: %d\n", ret);
1249                                 return 1;
1250                         }
1251                 }
1252
1253                 p += max_bs;
1254         }
1255
1256         if (init_file_completion_logging(td, max_units))
1257                 return 1;
1258
1259         return 0;
1260 }
1261
1262 static int switch_ioscheduler(struct thread_data *td)
1263 {
1264 #ifdef FIO_HAVE_IOSCHED_SWITCH
1265         char tmp[256], tmp2[128];
1266         FILE *f;
1267         int ret;
1268
1269         if (td->io_ops->flags & FIO_DISKLESSIO)
1270                 return 0;
1271
1272         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1273
1274         f = fopen(tmp, "r+");
1275         if (!f) {
1276                 if (errno == ENOENT) {
1277                         log_err("fio: os or kernel doesn't support IO scheduler"
1278                                 " switching\n");
1279                         return 0;
1280                 }
1281                 td_verror(td, errno, "fopen iosched");
1282                 return 1;
1283         }
1284
1285         /*
1286          * Set io scheduler.
1287          */
1288         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1289         if (ferror(f) || ret != 1) {
1290                 td_verror(td, errno, "fwrite");
1291                 fclose(f);
1292                 return 1;
1293         }
1294
1295         rewind(f);
1296
1297         /*
1298          * Read back and check that the selected scheduler is now the default.
1299          */
1300         memset(tmp, 0, sizeof(tmp));
1301         ret = fread(tmp, sizeof(tmp), 1, f);
1302         if (ferror(f) || ret < 0) {
1303                 td_verror(td, errno, "fread");
1304                 fclose(f);
1305                 return 1;
1306         }
1307         /*
1308          * either a list of io schedulers or "none\n" is expected.
1309          */
1310         tmp[strlen(tmp) - 1] = '\0';
1311
1312
1313         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1314         if (!strstr(tmp, tmp2)) {
1315                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1316                 td_verror(td, EINVAL, "iosched_switch");
1317                 fclose(f);
1318                 return 1;
1319         }
1320
1321         fclose(f);
1322         return 0;
1323 #else
1324         return 0;
1325 #endif
1326 }
1327
1328 static bool keep_running(struct thread_data *td)
1329 {
1330         unsigned long long limit;
1331
1332         if (td->done)
1333                 return false;
1334         if (td->o.time_based)
1335                 return true;
1336         if (td->o.loops) {
1337                 td->o.loops--;
1338                 return true;
1339         }
1340         if (exceeds_number_ios(td))
1341                 return false;
1342
1343         if (td->o.io_limit)
1344                 limit = td->o.io_limit;
1345         else
1346                 limit = td->o.size;
1347
1348         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1349                 uint64_t diff;
1350
1351                 /*
1352                  * If the difference is less than the minimum IO size, we
1353                  * are done.
1354                  */
1355                 diff = limit - ddir_rw_sum(td->io_bytes);
1356                 if (diff < td_max_bs(td))
1357                         return false;
1358
1359                 if (fio_files_done(td) && !td->o.io_limit)
1360                         return false;
1361
1362                 return true;
1363         }
1364
1365         return false;
1366 }
1367
1368 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1369 {
1370         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1371         int ret;
1372         char *str;
1373
1374         str = malloc(newlen);
1375         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1376
1377         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1378         ret = system(str);
1379         if (ret == -1)
1380                 log_err("fio: exec of cmd <%s> failed\n", str);
1381
1382         free(str);
1383         return ret;
1384 }
1385
1386 /*
1387  * Dry run to compute correct state of numberio for verification.
1388  */
1389 static uint64_t do_dry_run(struct thread_data *td)
1390 {
1391         td_set_runstate(td, TD_RUNNING);
1392
1393         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1394                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1395                 struct io_u *io_u;
1396                 int ret;
1397
1398                 if (td->terminate || td->done)
1399                         break;
1400
1401                 io_u = get_io_u(td);
1402                 if (IS_ERR_OR_NULL(io_u))
1403                         break;
1404
1405                 io_u_set(io_u, IO_U_F_FLIGHT);
1406                 io_u->error = 0;
1407                 io_u->resid = 0;
1408                 if (ddir_rw(acct_ddir(io_u)))
1409                         td->io_issues[acct_ddir(io_u)]++;
1410                 if (ddir_rw(io_u->ddir)) {
1411                         io_u_mark_depth(td, 1);
1412                         td->ts.total_io_u[io_u->ddir]++;
1413                 }
1414
1415                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1416                     td->o.do_verify &&
1417                     td->o.verify != VERIFY_NONE &&
1418                     !td->o.experimental_verify)
1419                         log_io_piece(td, io_u);
1420
1421                 ret = io_u_sync_complete(td, io_u);
1422                 (void) ret;
1423         }
1424
1425         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1426 }
1427
1428 struct fork_data {
1429         struct thread_data *td;
1430         struct sk_out *sk_out;
1431 };
1432
1433 /*
1434  * Entry point for the thread based jobs. The process based jobs end up
1435  * here as well, after a little setup.
1436  */
1437 static void *thread_main(void *data)
1438 {
1439         struct fork_data *fd = data;
1440         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1441         struct thread_data *td = fd->td;
1442         struct thread_options *o = &td->o;
1443         struct sk_out *sk_out = fd->sk_out;
1444         int clear_state;
1445         int ret;
1446
1447         sk_out_assign(sk_out);
1448         free(fd);
1449
1450         if (!o->use_thread) {
1451                 setsid();
1452                 td->pid = getpid();
1453         } else
1454                 td->pid = gettid();
1455
1456         fio_local_clock_init(o->use_thread);
1457
1458         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1459
1460         if (is_backend)
1461                 fio_server_send_start(td);
1462
1463         INIT_FLIST_HEAD(&td->io_log_list);
1464         INIT_FLIST_HEAD(&td->io_hist_list);
1465         INIT_FLIST_HEAD(&td->verify_list);
1466         INIT_FLIST_HEAD(&td->trim_list);
1467         INIT_FLIST_HEAD(&td->next_rand_list);
1468         td->io_hist_tree = RB_ROOT;
1469
1470         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1471         if (ret) {
1472                 td_verror(td, ret, "mutex_cond_init_pshared");
1473                 goto err;
1474         }
1475         ret = cond_init_pshared(&td->verify_cond);
1476         if (ret) {
1477                 td_verror(td, ret, "mutex_cond_pshared");
1478                 goto err;
1479         }
1480
1481         td_set_runstate(td, TD_INITIALIZED);
1482         dprint(FD_MUTEX, "up startup_mutex\n");
1483         fio_mutex_up(startup_mutex);
1484         dprint(FD_MUTEX, "wait on td->mutex\n");
1485         fio_mutex_down(td->mutex);
1486         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1487
1488         /*
1489          * A new gid requires privilege, so we need to do this before setting
1490          * the uid.
1491          */
1492         if (o->gid != -1U && setgid(o->gid)) {
1493                 td_verror(td, errno, "setgid");
1494                 goto err;
1495         }
1496         if (o->uid != -1U && setuid(o->uid)) {
1497                 td_verror(td, errno, "setuid");
1498                 goto err;
1499         }
1500
1501         /*
1502          * Do this early, we don't want the compress threads to be limited
1503          * to the same CPUs as the IO workers. So do this before we set
1504          * any potential CPU affinity
1505          */
1506         if (iolog_compress_init(td, sk_out))
1507                 goto err;
1508
1509         /*
1510          * If we have a gettimeofday() thread, make sure we exclude that
1511          * thread from this job
1512          */
1513         if (o->gtod_cpu)
1514                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1515
1516         /*
1517          * Set affinity first, in case it has an impact on the memory
1518          * allocations.
1519          */
1520         if (fio_option_is_set(o, cpumask)) {
1521                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1522                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1523                         if (!ret) {
1524                                 log_err("fio: no CPUs set\n");
1525                                 log_err("fio: Try increasing number of available CPUs\n");
1526                                 td_verror(td, EINVAL, "cpus_split");
1527                                 goto err;
1528                         }
1529                 }
1530                 ret = fio_setaffinity(td->pid, o->cpumask);
1531                 if (ret == -1) {
1532                         td_verror(td, errno, "cpu_set_affinity");
1533                         goto err;
1534                 }
1535         }
1536
1537 #ifdef CONFIG_LIBNUMA
1538         /* numa node setup */
1539         if (fio_option_is_set(o, numa_cpunodes) ||
1540             fio_option_is_set(o, numa_memnodes)) {
1541                 struct bitmask *mask;
1542
1543                 if (numa_available() < 0) {
1544                         td_verror(td, errno, "Does not support NUMA API\n");
1545                         goto err;
1546                 }
1547
1548                 if (fio_option_is_set(o, numa_cpunodes)) {
1549                         mask = numa_parse_nodestring(o->numa_cpunodes);
1550                         ret = numa_run_on_node_mask(mask);
1551                         numa_free_nodemask(mask);
1552                         if (ret == -1) {
1553                                 td_verror(td, errno, \
1554                                         "numa_run_on_node_mask failed\n");
1555                                 goto err;
1556                         }
1557                 }
1558
1559                 if (fio_option_is_set(o, numa_memnodes)) {
1560                         mask = NULL;
1561                         if (o->numa_memnodes)
1562                                 mask = numa_parse_nodestring(o->numa_memnodes);
1563
1564                         switch (o->numa_mem_mode) {
1565                         case MPOL_INTERLEAVE:
1566                                 numa_set_interleave_mask(mask);
1567                                 break;
1568                         case MPOL_BIND:
1569                                 numa_set_membind(mask);
1570                                 break;
1571                         case MPOL_LOCAL:
1572                                 numa_set_localalloc();
1573                                 break;
1574                         case MPOL_PREFERRED:
1575                                 numa_set_preferred(o->numa_mem_prefer_node);
1576                                 break;
1577                         case MPOL_DEFAULT:
1578                         default:
1579                                 break;
1580                         }
1581
1582                         if (mask)
1583                                 numa_free_nodemask(mask);
1584
1585                 }
1586         }
1587 #endif
1588
1589         if (fio_pin_memory(td))
1590                 goto err;
1591
1592         /*
1593          * May alter parameters that init_io_u() will use, so we need to
1594          * do this first.
1595          */
1596         if (init_iolog(td))
1597                 goto err;
1598
1599         if (init_io_u(td))
1600                 goto err;
1601
1602         if (o->verify_async && verify_async_init(td))
1603                 goto err;
1604
1605         if (fio_option_is_set(o, ioprio) ||
1606             fio_option_is_set(o, ioprio_class)) {
1607                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1608                 if (ret == -1) {
1609                         td_verror(td, errno, "ioprio_set");
1610                         goto err;
1611                 }
1612         }
1613
1614         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1615                 goto err;
1616
1617         errno = 0;
1618         if (nice(o->nice) == -1 && errno != 0) {
1619                 td_verror(td, errno, "nice");
1620                 goto err;
1621         }
1622
1623         if (o->ioscheduler && switch_ioscheduler(td))
1624                 goto err;
1625
1626         if (!o->create_serialize && setup_files(td))
1627                 goto err;
1628
1629         if (td_io_init(td))
1630                 goto err;
1631
1632         if (init_random_map(td))
1633                 goto err;
1634
1635         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1636                 goto err;
1637
1638         if (o->pre_read) {
1639                 if (pre_read_files(td) < 0)
1640                         goto err;
1641         }
1642
1643         fio_verify_init(td);
1644
1645         if (rate_submit_init(td, sk_out))
1646                 goto err;
1647
1648         fio_gettime(&td->epoch, NULL);
1649         fio_getrusage(&td->ru_start);
1650         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1651         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1652
1653         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1654                         o->ratemin[DDIR_TRIM]) {
1655                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1656                                         sizeof(td->bw_sample_time));
1657                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1658                                         sizeof(td->bw_sample_time));
1659                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1660                                         sizeof(td->bw_sample_time));
1661         }
1662
1663         clear_state = 0;
1664         while (keep_running(td)) {
1665                 uint64_t verify_bytes;
1666
1667                 fio_gettime(&td->start, NULL);
1668                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1669
1670                 if (clear_state)
1671                         clear_io_state(td, 0);
1672
1673                 prune_io_piece_log(td);
1674
1675                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1676                         verify_bytes = do_dry_run(td);
1677                 else {
1678                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1679
1680                         do_io(td, bytes_done);
1681
1682                         if (!ddir_rw_sum(bytes_done)) {
1683                                 fio_mark_td_terminate(td);
1684                                 verify_bytes = 0;
1685                         } else {
1686                                 verify_bytes = bytes_done[DDIR_WRITE] +
1687                                                 bytes_done[DDIR_TRIM];
1688                         }
1689                 }
1690
1691                 clear_state = 1;
1692
1693                 /*
1694                  * Make sure we've successfully updated the rusage stats
1695                  * before waiting on the stat mutex. Otherwise we could have
1696                  * the stat thread holding stat mutex and waiting for
1697                  * the rusage_sem, which would never get upped because
1698                  * this thread is waiting for the stat mutex.
1699                  */
1700                 check_update_rusage(td);
1701
1702                 fio_mutex_down(stat_mutex);
1703                 if (td_read(td) && td->io_bytes[DDIR_READ])
1704                         update_runtime(td, elapsed_us, DDIR_READ);
1705                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1706                         update_runtime(td, elapsed_us, DDIR_WRITE);
1707                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1708                         update_runtime(td, elapsed_us, DDIR_TRIM);
1709                 fio_gettime(&td->start, NULL);
1710                 fio_mutex_up(stat_mutex);
1711
1712                 if (td->error || td->terminate)
1713                         break;
1714
1715                 if (!o->do_verify ||
1716                     o->verify == VERIFY_NONE ||
1717                     (td->io_ops->flags & FIO_UNIDIR))
1718                         continue;
1719
1720                 clear_io_state(td, 0);
1721
1722                 fio_gettime(&td->start, NULL);
1723
1724                 do_verify(td, verify_bytes);
1725
1726                 /*
1727                  * See comment further up for why this is done here.
1728                  */
1729                 check_update_rusage(td);
1730
1731                 fio_mutex_down(stat_mutex);
1732                 update_runtime(td, elapsed_us, DDIR_READ);
1733                 fio_gettime(&td->start, NULL);
1734                 fio_mutex_up(stat_mutex);
1735
1736                 if (td->error || td->terminate)
1737                         break;
1738         }
1739
1740         td_set_runstate(td, TD_FINISHING);
1741
1742         update_rusage_stat(td);
1743         td->ts.total_run_time = mtime_since_now(&td->epoch);
1744         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1745         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1746         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1747
1748         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1749             (td->o.verify != VERIFY_NONE && td_write(td)))
1750                 verify_save_state(td->thread_number);
1751
1752         fio_unpin_memory(td);
1753
1754         td_writeout_logs(td, true);
1755
1756         iolog_compress_exit(td);
1757         rate_submit_exit(td);
1758
1759         if (o->exec_postrun)
1760                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1761
1762         if (exitall_on_terminate || (o->exitall_error && td->error))
1763                 fio_terminate_threads(td->groupid);
1764
1765 err:
1766         if (td->error)
1767                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1768                                                         td->verror);
1769
1770         if (o->verify_async)
1771                 verify_async_exit(td);
1772
1773         close_and_free_files(td);
1774         cleanup_io_u(td);
1775         close_ioengine(td);
1776         cgroup_shutdown(td, &cgroup_mnt);
1777         verify_free_state(td);
1778
1779         if (td->zone_state_index) {
1780                 int i;
1781
1782                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1783                         free(td->zone_state_index[i]);
1784                 free(td->zone_state_index);
1785                 td->zone_state_index = NULL;
1786         }
1787
1788         if (fio_option_is_set(o, cpumask)) {
1789                 ret = fio_cpuset_exit(&o->cpumask);
1790                 if (ret)
1791                         td_verror(td, ret, "fio_cpuset_exit");
1792         }
1793
1794         /*
1795          * do this very late, it will log file closing as well
1796          */
1797         if (o->write_iolog_file)
1798                 write_iolog_close(td);
1799
1800         fio_mutex_remove(td->mutex);
1801         td->mutex = NULL;
1802
1803         td_set_runstate(td, TD_EXITED);
1804
1805         /*
1806          * Do this last after setting our runstate to exited, so we
1807          * know that the stat thread is signaled.
1808          */
1809         check_update_rusage(td);
1810
1811         sk_out_drop();
1812         return (void *) (uintptr_t) td->error;
1813 }
1814
1815 static void dump_td_info(struct thread_data *td)
1816 {
1817         log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
1818                 "appears to be stuck. Doing forceful exit of this job.\n",
1819                         td->o.name, td->runstate,
1820                         (unsigned long) time_since_now(&td->terminate_time));
1821 }
1822
1823 /*
1824  * Run over the job map and reap the threads that have exited, if any.
1825  */
1826 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1827                          unsigned int *m_rate)
1828 {
1829         struct thread_data *td;
1830         unsigned int cputhreads, realthreads, pending;
1831         int i, status, ret;
1832
1833         /*
1834          * reap exited threads (TD_EXITED -> TD_REAPED)
1835          */
1836         realthreads = pending = cputhreads = 0;
1837         for_each_td(td, i) {
1838                 int flags = 0;
1839
1840                 /*
1841                  * ->io_ops is NULL for a thread that has closed its
1842                  * io engine
1843                  */
1844                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1845                         cputhreads++;
1846                 else
1847                         realthreads++;
1848
1849                 if (!td->pid) {
1850                         pending++;
1851                         continue;
1852                 }
1853                 if (td->runstate == TD_REAPED)
1854                         continue;
1855                 if (td->o.use_thread) {
1856                         if (td->runstate == TD_EXITED) {
1857                                 td_set_runstate(td, TD_REAPED);
1858                                 goto reaped;
1859                         }
1860                         continue;
1861                 }
1862
1863                 flags = WNOHANG;
1864                 if (td->runstate == TD_EXITED)
1865                         flags = 0;
1866
1867                 /*
1868                  * check if someone quit or got killed in an unusual way
1869                  */
1870                 ret = waitpid(td->pid, &status, flags);
1871                 if (ret < 0) {
1872                         if (errno == ECHILD) {
1873                                 log_err("fio: pid=%d disappeared %d\n",
1874                                                 (int) td->pid, td->runstate);
1875                                 td->sig = ECHILD;
1876                                 td_set_runstate(td, TD_REAPED);
1877                                 goto reaped;
1878                         }
1879                         perror("waitpid");
1880                 } else if (ret == td->pid) {
1881                         if (WIFSIGNALED(status)) {
1882                                 int sig = WTERMSIG(status);
1883
1884                                 if (sig != SIGTERM && sig != SIGUSR2)
1885                                         log_err("fio: pid=%d, got signal=%d\n",
1886                                                         (int) td->pid, sig);
1887                                 td->sig = sig;
1888                                 td_set_runstate(td, TD_REAPED);
1889                                 goto reaped;
1890                         }
1891                         if (WIFEXITED(status)) {
1892                                 if (WEXITSTATUS(status) && !td->error)
1893                                         td->error = WEXITSTATUS(status);
1894
1895                                 td_set_runstate(td, TD_REAPED);
1896                                 goto reaped;
1897                         }
1898                 }
1899
1900                 /*
1901                  * If the job is stuck, do a forceful timeout of it and
1902                  * move on.
1903                  */
1904                 if (td->terminate &&
1905                     td->runstate < TD_FSYNCING &&
1906                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1907                         dump_td_info(td);
1908                         td_set_runstate(td, TD_REAPED);
1909                         goto reaped;
1910                 }
1911
1912                 /*
1913                  * thread is not dead, continue
1914                  */
1915                 pending++;
1916                 continue;
1917 reaped:
1918                 (*nr_running)--;
1919                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1920                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1921                 if (!td->pid)
1922                         pending--;
1923
1924                 if (td->error)
1925                         exit_value++;
1926
1927                 done_secs += mtime_since_now(&td->epoch) / 1000;
1928                 profile_td_exit(td);
1929         }
1930
1931         if (*nr_running == cputhreads && !pending && realthreads)
1932                 fio_terminate_threads(TERMINATE_ALL);
1933 }
1934
1935 static bool __check_trigger_file(void)
1936 {
1937         struct stat sb;
1938
1939         if (!trigger_file)
1940                 return false;
1941
1942         if (stat(trigger_file, &sb))
1943                 return false;
1944
1945         if (unlink(trigger_file) < 0)
1946                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1947                                                         strerror(errno));
1948
1949         return true;
1950 }
1951
1952 static bool trigger_timedout(void)
1953 {
1954         if (trigger_timeout)
1955                 return time_since_genesis() >= trigger_timeout;
1956
1957         return false;
1958 }
1959
1960 void exec_trigger(const char *cmd)
1961 {
1962         int ret;
1963
1964         if (!cmd)
1965                 return;
1966
1967         ret = system(cmd);
1968         if (ret == -1)
1969                 log_err("fio: failed executing %s trigger\n", cmd);
1970 }
1971
1972 void check_trigger_file(void)
1973 {
1974         if (__check_trigger_file() || trigger_timedout()) {
1975                 if (nr_clients)
1976                         fio_clients_send_trigger(trigger_remote_cmd);
1977                 else {
1978                         verify_save_state(IO_LIST_ALL);
1979                         fio_terminate_threads(TERMINATE_ALL);
1980                         exec_trigger(trigger_cmd);
1981                 }
1982         }
1983 }
1984
1985 static int fio_verify_load_state(struct thread_data *td)
1986 {
1987         int ret;
1988
1989         if (!td->o.verify_state)
1990                 return 0;
1991
1992         if (is_backend) {
1993                 void *data;
1994
1995                 ret = fio_server_get_verify_state(td->o.name,
1996                                         td->thread_number - 1, &data);
1997                 if (!ret)
1998                         verify_assign_state(td, data);
1999         } else
2000                 ret = verify_load_state(td, "local");
2001
2002         return ret;
2003 }
2004
2005 static void do_usleep(unsigned int usecs)
2006 {
2007         check_for_running_stats();
2008         check_trigger_file();
2009         usleep(usecs);
2010 }
2011
2012 static bool check_mount_writes(struct thread_data *td)
2013 {
2014         struct fio_file *f;
2015         unsigned int i;
2016
2017         if (!td_write(td) || td->o.allow_mounted_write)
2018                 return false;
2019
2020         for_each_file(td, f, i) {
2021                 if (f->filetype != FIO_TYPE_BD)
2022                         continue;
2023                 if (device_is_mounted(f->file_name))
2024                         goto mounted;
2025         }
2026
2027         return false;
2028 mounted:
2029         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2030         return true;
2031 }
2032
2033 static bool waitee_running(struct thread_data *me)
2034 {
2035         const char *waitee = me->o.wait_for;
2036         const char *self = me->o.name;
2037         struct thread_data *td;
2038         int i;
2039
2040         if (!waitee)
2041                 return false;
2042
2043         for_each_td(td, i) {
2044                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2045                         continue;
2046
2047                 if (td->runstate < TD_EXITED) {
2048                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2049                                         self, td->o.name,
2050                                         runstate_to_name(td->runstate));
2051                         return true;
2052                 }
2053         }
2054
2055         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2056         return false;
2057 }
2058
2059 /*
2060  * Main function for kicking off and reaping jobs, as needed.
2061  */
2062 static void run_threads(struct sk_out *sk_out)
2063 {
2064         struct thread_data *td;
2065         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2066         uint64_t spent;
2067
2068         if (fio_gtod_offload && fio_start_gtod_thread())
2069                 return;
2070
2071         fio_idle_prof_init();
2072
2073         set_sig_handlers();
2074
2075         nr_thread = nr_process = 0;
2076         for_each_td(td, i) {
2077                 if (check_mount_writes(td))
2078                         return;
2079                 if (td->o.use_thread)
2080                         nr_thread++;
2081                 else
2082                         nr_process++;
2083         }
2084
2085         if (output_format & FIO_OUTPUT_NORMAL) {
2086                 log_info("Starting ");
2087                 if (nr_thread)
2088                         log_info("%d thread%s", nr_thread,
2089                                                 nr_thread > 1 ? "s" : "");
2090                 if (nr_process) {
2091                         if (nr_thread)
2092                                 log_info(" and ");
2093                         log_info("%d process%s", nr_process,
2094                                                 nr_process > 1 ? "es" : "");
2095                 }
2096                 log_info("\n");
2097                 log_info_flush();
2098         }
2099
2100         todo = thread_number;
2101         nr_running = 0;
2102         nr_started = 0;
2103         m_rate = t_rate = 0;
2104
2105         for_each_td(td, i) {
2106                 print_status_init(td->thread_number - 1);
2107
2108                 if (!td->o.create_serialize)
2109                         continue;
2110
2111                 if (fio_verify_load_state(td))
2112                         goto reap;
2113
2114                 /*
2115                  * do file setup here so it happens sequentially,
2116                  * we don't want X number of threads getting their
2117                  * client data interspersed on disk
2118                  */
2119                 if (setup_files(td)) {
2120 reap:
2121                         exit_value++;
2122                         if (td->error)
2123                                 log_err("fio: pid=%d, err=%d/%s\n",
2124                                         (int) td->pid, td->error, td->verror);
2125                         td_set_runstate(td, TD_REAPED);
2126                         todo--;
2127                 } else {
2128                         struct fio_file *f;
2129                         unsigned int j;
2130
2131                         /*
2132                          * for sharing to work, each job must always open
2133                          * its own files. so close them, if we opened them
2134                          * for creation
2135                          */
2136                         for_each_file(td, f, j) {
2137                                 if (fio_file_open(f))
2138                                         td_io_close_file(td, f);
2139                         }
2140                 }
2141         }
2142
2143         /* start idle threads before io threads start to run */
2144         fio_idle_prof_start();
2145
2146         set_genesis_time();
2147
2148         while (todo) {
2149                 struct thread_data *map[REAL_MAX_JOBS];
2150                 struct timeval this_start;
2151                 int this_jobs = 0, left;
2152                 struct fork_data *fd;
2153
2154                 /*
2155                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2156                  */
2157                 for_each_td(td, i) {
2158                         if (td->runstate != TD_NOT_CREATED)
2159                                 continue;
2160
2161                         /*
2162                          * never got a chance to start, killed by other
2163                          * thread for some reason
2164                          */
2165                         if (td->terminate) {
2166                                 todo--;
2167                                 continue;
2168                         }
2169
2170                         if (td->o.start_delay) {
2171                                 spent = utime_since_genesis();
2172
2173                                 if (td->o.start_delay > spent)
2174                                         continue;
2175                         }
2176
2177                         if (td->o.stonewall && (nr_started || nr_running)) {
2178                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2179                                                         td->o.name);
2180                                 break;
2181                         }
2182
2183                         if (waitee_running(td)) {
2184                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2185                                                 td->o.name, td->o.wait_for);
2186                                 continue;
2187                         }
2188
2189                         init_disk_util(td);
2190
2191                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2192                         td->update_rusage = 0;
2193
2194                         /*
2195                          * Set state to created. Thread will transition
2196                          * to TD_INITIALIZED when it's done setting up.
2197                          */
2198                         td_set_runstate(td, TD_CREATED);
2199                         map[this_jobs++] = td;
2200                         nr_started++;
2201
2202                         fd = calloc(1, sizeof(*fd));
2203                         fd->td = td;
2204                         fd->sk_out = sk_out;
2205
2206                         if (td->o.use_thread) {
2207                                 int ret;
2208
2209                                 dprint(FD_PROCESS, "will pthread_create\n");
2210                                 ret = pthread_create(&td->thread, NULL,
2211                                                         thread_main, fd);
2212                                 if (ret) {
2213                                         log_err("pthread_create: %s\n",
2214                                                         strerror(ret));
2215                                         free(fd);
2216                                         nr_started--;
2217                                         break;
2218                                 }
2219                                 ret = pthread_detach(td->thread);
2220                                 if (ret)
2221                                         log_err("pthread_detach: %s",
2222                                                         strerror(ret));
2223                         } else {
2224                                 pid_t pid;
2225                                 dprint(FD_PROCESS, "will fork\n");
2226                                 pid = fork();
2227                                 if (!pid) {
2228                                         int ret;
2229
2230                                         ret = (int)(uintptr_t)thread_main(fd);
2231                                         _exit(ret);
2232                                 } else if (i == fio_debug_jobno)
2233                                         *fio_debug_jobp = pid;
2234                         }
2235                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2236                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2237                                 log_err("fio: job startup hung? exiting.\n");
2238                                 fio_terminate_threads(TERMINATE_ALL);
2239                                 fio_abort = 1;
2240                                 nr_started--;
2241                                 break;
2242                         }
2243                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2244                 }
2245
2246                 /*
2247                  * Wait for the started threads to transition to
2248                  * TD_INITIALIZED.
2249                  */
2250                 fio_gettime(&this_start, NULL);
2251                 left = this_jobs;
2252                 while (left && !fio_abort) {
2253                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2254                                 break;
2255
2256                         do_usleep(100000);
2257
2258                         for (i = 0; i < this_jobs; i++) {
2259                                 td = map[i];
2260                                 if (!td)
2261                                         continue;
2262                                 if (td->runstate == TD_INITIALIZED) {
2263                                         map[i] = NULL;
2264                                         left--;
2265                                 } else if (td->runstate >= TD_EXITED) {
2266                                         map[i] = NULL;
2267                                         left--;
2268                                         todo--;
2269                                         nr_running++; /* work-around... */
2270                                 }
2271                         }
2272                 }
2273
2274                 if (left) {
2275                         log_err("fio: %d job%s failed to start\n", left,
2276                                         left > 1 ? "s" : "");
2277                         for (i = 0; i < this_jobs; i++) {
2278                                 td = map[i];
2279                                 if (!td)
2280                                         continue;
2281                                 kill(td->pid, SIGTERM);
2282                         }
2283                         break;
2284                 }
2285
2286                 /*
2287                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2288                  */
2289                 for_each_td(td, i) {
2290                         if (td->runstate != TD_INITIALIZED)
2291                                 continue;
2292
2293                         if (in_ramp_time(td))
2294                                 td_set_runstate(td, TD_RAMP);
2295                         else
2296                                 td_set_runstate(td, TD_RUNNING);
2297                         nr_running++;
2298                         nr_started--;
2299                         m_rate += ddir_rw_sum(td->o.ratemin);
2300                         t_rate += ddir_rw_sum(td->o.rate);
2301                         todo--;
2302                         fio_mutex_up(td->mutex);
2303                 }
2304
2305                 reap_threads(&nr_running, &t_rate, &m_rate);
2306
2307                 if (todo)
2308                         do_usleep(100000);
2309         }
2310
2311         while (nr_running) {
2312                 reap_threads(&nr_running, &t_rate, &m_rate);
2313                 do_usleep(10000);
2314         }
2315
2316         fio_idle_prof_stop();
2317
2318         update_io_ticks();
2319 }
2320
2321 static void free_disk_util(void)
2322 {
2323         disk_util_prune_entries();
2324         helper_thread_destroy();
2325 }
2326
2327 int fio_backend(struct sk_out *sk_out)
2328 {
2329         struct thread_data *td;
2330         int i;
2331
2332         if (exec_profile) {
2333                 if (load_profile(exec_profile))
2334                         return 1;
2335                 free(exec_profile);
2336                 exec_profile = NULL;
2337         }
2338         if (!thread_number)
2339                 return 0;
2340
2341         if (write_bw_log) {
2342                 struct log_params p = {
2343                         .log_type = IO_LOG_TYPE_BW,
2344                 };
2345
2346                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2347                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2348                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2349         }
2350
2351         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2352         if (startup_mutex == NULL)
2353                 return 1;
2354
2355         set_genesis_time();
2356         stat_init();
2357         helper_thread_create(startup_mutex, sk_out);
2358
2359         cgroup_list = smalloc(sizeof(*cgroup_list));
2360         INIT_FLIST_HEAD(cgroup_list);
2361
2362         run_threads(sk_out);
2363
2364         helper_thread_exit();
2365
2366         if (!fio_abort) {
2367                 __show_run_stats();
2368                 if (write_bw_log) {
2369                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2370                                 struct io_log *log = agg_io_log[i];
2371
2372                                 flush_log(log, false);
2373                                 free_log(log);
2374                         }
2375                 }
2376         }
2377
2378         for_each_td(td, i) {
2379                 fio_options_free(td);
2380                 if (td->rusage_sem) {
2381                         fio_mutex_remove(td->rusage_sem);
2382                         td->rusage_sem = NULL;
2383                 }
2384         }
2385
2386         free_disk_util();
2387         cgroup_kill(cgroup_list);
2388         sfree(cgroup_list);
2389         sfree(cgroup_mnt);
2390
2391         fio_mutex_remove(startup_mutex);
2392         stat_exit();
2393         return exit_value;
2394 }