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
  61 static struct fio_mutex *startup_mutex;
  62 static struct flist_head *cgroup_list;
  63 static char *cgroup_mnt;
  64 static int exit_value;
  65 static volatile int fio_abort;
  66 static unsigned int nr_process = 0;
  67 static unsigned int nr_thread = 0;
  68
  69 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  70
  71 int groupid = 0;
  72 unsigned int thread_number = 0;
  73 unsigned int stat_number = 0;
  74 int shm_id = 0;
  75 int temp_stall_ts;
  76 unsigned long done_secs = 0;
  77
  78 static struct helper_data {
  79         volatile int exit;
  80         volatile int reset;
  81         volatile int do_stat;
  82         struct sk_out *sk_out;
  83         pthread_t thread;
  84         pthread_mutex_t lock;
  85         pthread_cond_t cond;
  86 } *helper_data;
  87
  88 #define PAGE_ALIGN(buf) \
  89         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  90
  91 #define JOB_START_TIMEOUT       (5 * 1000)
  92
  93 static void sig_int(int sig)
  94 {
  95         if (threads) {
  96                 if (is_backend)
  97                         fio_server_got_signal(sig);
  98                 else {
  99                         log_info("\nfio: terminating on signal %d\n", sig);
 100                         log_info_flush();
 101                         exit_value = 128;
 102                 }
 103
 104                 fio_terminate_threads(TERMINATE_ALL);
 105         }
 106 }
 107
 108 void sig_show_status(int sig)
 109 {
 110         show_running_run_stats();
 111 }
 112
 113 static void set_sig_handlers(void)
 114 {
 115         struct sigaction act;
 116
 117         memset(&act, 0, sizeof(act));
 118         act.sa_handler = sig_int;
 119         act.sa_flags = SA_RESTART;
 120         sigaction(SIGINT, &act, NULL);
 121
 122         memset(&act, 0, sizeof(act));
 123         act.sa_handler = sig_int;
 124         act.sa_flags = SA_RESTART;
 125         sigaction(SIGTERM, &act, NULL);
 126
 127 /* Windows uses SIGBREAK as a quit signal from other applications */
 128 #ifdef WIN32
 129         memset(&act, 0, sizeof(act));
 130         act.sa_handler = sig_int;
 131         act.sa_flags = SA_RESTART;
 132         sigaction(SIGBREAK, &act, NULL);
 133 #endif
 134
 135         memset(&act, 0, sizeof(act));
 136         act.sa_handler = sig_show_status;
 137         act.sa_flags = SA_RESTART;
 138         sigaction(SIGUSR1, &act, NULL);
 139
 140         if (is_backend) {
 141                 memset(&act, 0, sizeof(act));
 142                 act.sa_handler = sig_int;
 143                 act.sa_flags = SA_RESTART;
 144                 sigaction(SIGPIPE, &act, NULL);
 145         }
 146 }
 147
 148 /*
 149  * Check if we are above the minimum rate given.
 150  */
 151 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 152                              enum fio_ddir ddir)
 153 {
 154         unsigned long long bytes = 0;
 155         unsigned long iops = 0;
 156         unsigned long spent;
 157         unsigned long rate;
 158         unsigned int ratemin = 0;
 159         unsigned int rate_iops = 0;
 160         unsigned int rate_iops_min = 0;
 161
 162         assert(ddir_rw(ddir));
 163
 164         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 165                 return false;
 166
 167         /*
 168          * allow a 2 second settle period in the beginning
 169          */
 170         if (mtime_since(&td->start, now) < 2000)
 171                 return false;
 172
 173         iops += td->this_io_blocks[ddir];
 174         bytes += td->this_io_bytes[ddir];
 175         ratemin += td->o.ratemin[ddir];
 176         rate_iops += td->o.rate_iops[ddir];
 177         rate_iops_min += td->o.rate_iops_min[ddir];
 178
 179         /*
 180          * if rate blocks is set, sample is running
 181          */
 182         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 183                 spent = mtime_since(&td->lastrate[ddir], now);
 184                 if (spent < td->o.ratecycle)
 185                         return false;
 186
 187                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 188                         /*
 189                          * check bandwidth specified rate
 190                          */
 191                         if (bytes < td->rate_bytes[ddir]) {
 192                                 log_err("%s: min rate %u not met\n", td->o.name,
 193                                                                 ratemin);
 194                                 return true;
 195                         } else {
 196                                 if (spent)
 197                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 198                                 else
 199                                         rate = 0;
 200
 201                                 if (rate < ratemin ||
 202                                     bytes < td->rate_bytes[ddir]) {
 203                                         log_err("%s: min rate %u not met, got"
 204                                                 " %luKB/sec\n", td->o.name,
 205                                                         ratemin, rate);
 206                                         return true;
 207                                 }
 208                         }
 209                 } else {
 210                         /*
 211                          * checks iops specified rate
 212                          */
 213                         if (iops < rate_iops) {
 214                                 log_err("%s: min iops rate %u not met\n",
 215                                                 td->o.name, rate_iops);
 216                                 return true;
 217                         } else {
 218                                 if (spent)
 219                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 220                                 else
 221                                         rate = 0;
 222
 223                                 if (rate < rate_iops_min ||
 224                                     iops < td->rate_blocks[ddir]) {
 225                                         log_err("%s: min iops rate %u not met,"
 226                                                 " got %lu\n", td->o.name,
 227                                                         rate_iops_min, rate);
 228                                         return true;
 229                                 }
 230                         }
 231                 }
 232         }
 233
 234         td->rate_bytes[ddir] = bytes;
 235         td->rate_blocks[ddir] = iops;
 236         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 237         return false;
 238 }
 239
 240 static bool check_min_rate(struct thread_data *td, struct timeval *now)
 241 {
 242         bool ret = false;
 243
 244         if (td->bytes_done[DDIR_READ])
 245                 ret |= __check_min_rate(td, now, DDIR_READ);
 246         if (td->bytes_done[DDIR_WRITE])
 247                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 248         if (td->bytes_done[DDIR_TRIM])
 249                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 250
 251         return ret;
 252 }
 253
 254 /*
 255  * When job exits, we can cancel the in-flight IO if we are using async
 256  * io. Attempt to do so.
 257  */
 258 static void cleanup_pending_aio(struct thread_data *td)
 259 {
 260         int r;
 261
 262         /*
 263          * get immediately available events, if any
 264          */
 265         r = io_u_queued_complete(td, 0);
 266         if (r < 0)
 267                 return;
 268
 269         /*
 270          * now cancel remaining active events
 271          */
 272         if (td->io_ops->cancel) {
 273                 struct io_u *io_u;
 274                 int i;
 275
 276                 io_u_qiter(&td->io_u_all, io_u, i) {
 277                         if (io_u->flags & IO_U_F_FLIGHT) {
 278                                 r = td->io_ops->cancel(td, io_u);
 279                                 if (!r)
 280                                         put_io_u(td, io_u);
 281                         }
 282                 }
 283         }
 284
 285         if (td->cur_depth)
 286                 r = io_u_queued_complete(td, td->cur_depth);
 287 }
 288
 289 /*
 290  * Helper to handle the final sync of a file. Works just like the normal
 291  * io path, just does everything sync.
 292  */
 293 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 294 {
 295         struct io_u *io_u = __get_io_u(td);
 296         int ret;
 297
 298         if (!io_u)
 299                 return true;
 300
 301         io_u->ddir = DDIR_SYNC;
 302         io_u->file = f;
 303
 304         if (td_io_prep(td, io_u)) {
 305                 put_io_u(td, io_u);
 306                 return true;
 307         }
 308
 309 requeue:
 310         ret = td_io_queue(td, io_u);
 311         if (ret < 0) {
 312                 td_verror(td, io_u->error, "td_io_queue");
 313                 put_io_u(td, io_u);
 314                 return true;
 315         } else if (ret == FIO_Q_QUEUED) {
 316                 if (td_io_commit(td))
 317                         return true;
 318                 if (io_u_queued_complete(td, 1) < 0)
 319                         return true;
 320         } else if (ret == FIO_Q_COMPLETED) {
 321                 if (io_u->error) {
 322                         td_verror(td, io_u->error, "td_io_queue");
 323                         return true;
 324                 }
 325
 326                 if (io_u_sync_complete(td, io_u) < 0)
 327                         return true;
 328         } else if (ret == FIO_Q_BUSY) {
 329                 if (td_io_commit(td))
 330                         return true;
 331                 goto requeue;
 332         }
 333
 334         return false;
 335 }
 336
 337 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 338 {
 339         int ret;
 340
 341         if (fio_file_open(f))
 342                 return fio_io_sync(td, f);
 343
 344         if (td_io_open_file(td, f))
 345                 return 1;
 346
 347         ret = fio_io_sync(td, f);
 348         td_io_close_file(td, f);
 349         return ret;
 350 }
 351
 352 static inline void __update_tv_cache(struct thread_data *td)
 353 {
 354         fio_gettime(&td->tv_cache, NULL);
 355 }
 356
 357 static inline void update_tv_cache(struct thread_data *td)
 358 {
 359         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 360                 __update_tv_cache(td);
 361 }
 362
 363 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 364 {
 365         if (in_ramp_time(td))
 366                 return false;
 367         if (!td->o.timeout)
 368                 return false;
 369         if (utime_since(&td->epoch, t) >= td->o.timeout)
 370                 return true;
 371
 372         return false;
 373 }
 374
 375 /*
 376  * We need to update the runtime consistently in ms, but keep a running
 377  * tally of the current elapsed time in microseconds for sub millisecond
 378  * updates.
 379  */
 380 static inline void update_runtime(struct thread_data *td,
 381                                   unsigned long long *elapsed_us,
 382                                   const enum fio_ddir ddir)
 383 {
 384         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 385                 return;
 386
 387         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 388         elapsed_us[ddir] += utime_since_now(&td->start);
 389         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 390 }
 391
 392 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 393                                 int *retptr)
 394 {
 395         int ret = *retptr;
 396
 397         if (ret < 0 || td->error) {
 398                 int err = td->error;
 399                 enum error_type_bit eb;
 400
 401                 if (ret < 0)
 402                         err = -ret;
 403
 404                 eb = td_error_type(ddir, err);
 405                 if (!(td->o.continue_on_error & (1 << eb)))
 406                         return true;
 407
 408                 if (td_non_fatal_error(td, eb, err)) {
 409                         /*
 410                          * Continue with the I/Os in case of
 411                          * a non fatal error.
 412                          */
 413                         update_error_count(td, err);
 414                         td_clear_error(td);
 415                         *retptr = 0;
 416                         return false;
 417                 } else if (td->o.fill_device && err == ENOSPC) {
 418                         /*
 419                          * We expect to hit this error if
 420                          * fill_device option is set.
 421                          */
 422                         td_clear_error(td);
 423                         fio_mark_td_terminate(td);
 424                         return true;
 425                 } else {
 426                         /*
 427                          * Stop the I/O in case of a fatal
 428                          * error.
 429                          */
 430                         update_error_count(td, err);
 431                         return true;
 432                 }
 433         }
 434
 435         return false;
 436 }
 437
 438 static void check_update_rusage(struct thread_data *td)
 439 {
 440         if (td->update_rusage) {
 441                 td->update_rusage = 0;
 442                 update_rusage_stat(td);
 443                 fio_mutex_up(td->rusage_sem);
 444         }
 445 }
 446
 447 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 448 {
 449         const int full = queue_full(td);
 450         int min_evts = 0;
 451         int ret;
 452
 453         /*
 454          * if the queue is full, we MUST reap at least 1 event
 455          */
 456         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 457         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 458                 min_evts = 1;
 459
 460         if (time && (__should_check_rate(td, DDIR_READ) ||
 461             __should_check_rate(td, DDIR_WRITE) ||
 462             __should_check_rate(td, DDIR_TRIM)))
 463                 fio_gettime(time, NULL);
 464
 465         do {
 466                 ret = io_u_queued_complete(td, min_evts);
 467                 if (ret < 0)
 468                         break;
 469         } while (full && (td->cur_depth > td->o.iodepth_low));
 470
 471         return ret;
 472 }
 473
 474 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 475                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 476                    struct timeval *comp_time)
 477 {
 478         int ret2;
 479
 480         switch (*ret) {
 481         case FIO_Q_COMPLETED:
 482                 if (io_u->error) {
 483                         *ret = -io_u->error;
 484                         clear_io_u(td, io_u);
 485                 } else if (io_u->resid) {
 486                         int bytes = io_u->xfer_buflen - io_u->resid;
 487                         struct fio_file *f = io_u->file;
 488
 489                         if (bytes_issued)
 490                                 *bytes_issued += bytes;
 491
 492                         if (!from_verify)
 493                                 trim_io_piece(td, io_u);
 494
 495                         /*
 496                          * zero read, fail
 497                          */
 498                         if (!bytes) {
 499                                 if (!from_verify)
 500                                         unlog_io_piece(td, io_u);
 501                                 td_verror(td, EIO, "full resid");
 502                                 put_io_u(td, io_u);
 503                                 break;
 504                         }
 505
 506                         io_u->xfer_buflen = io_u->resid;
 507                         io_u->xfer_buf += bytes;
 508                         io_u->offset += bytes;
 509
 510                         if (ddir_rw(io_u->ddir))
 511                                 td->ts.short_io_u[io_u->ddir]++;
 512
 513                         f = io_u->file;
 514                         if (io_u->offset == f->real_file_size)
 515                                 goto sync_done;
 516
 517                         requeue_io_u(td, &io_u);
 518                 } else {
 519 sync_done:
 520                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 521                             __should_check_rate(td, DDIR_WRITE) ||
 522                             __should_check_rate(td, DDIR_TRIM)))
 523                                 fio_gettime(comp_time, NULL);
 524
 525                         *ret = io_u_sync_complete(td, io_u);
 526                         if (*ret < 0)
 527                                 break;
 528                 }
 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         td_set_runstate(td, TD_VERIFYING);
 607
 608         io_u = NULL;
 609         while (!td->terminate) {
 610                 enum fio_ddir ddir;
 611                 int full;
 612
 613                 update_tv_cache(td);
 614                 check_update_rusage(td);
 615
 616                 if (runtime_exceeded(td, &td->tv_cache)) {
 617                         __update_tv_cache(td);
 618                         if (runtime_exceeded(td, &td->tv_cache)) {
 619                                 fio_mark_td_terminate(td);
 620                                 break;
 621                         }
 622                 }
 623
 624                 if (flow_threshold_exceeded(td))
 625                         continue;
 626
 627                 if (!td->o.experimental_verify) {
 628                         io_u = __get_io_u(td);
 629                         if (!io_u)
 630                                 break;
 631
 632                         if (get_next_verify(td, io_u)) {
 633                                 put_io_u(td, io_u);
 634                                 break;
 635                         }
 636
 637                         if (td_io_prep(td, io_u)) {
 638                                 put_io_u(td, io_u);
 639                                 break;
 640                         }
 641                 } else {
 642                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 643                                 break;
 644
 645                         while ((io_u = get_io_u(td)) != NULL) {
 646                                 if (IS_ERR(io_u)) {
 647                                         io_u = NULL;
 648                                         ret = FIO_Q_BUSY;
 649                                         goto reap;
 650                                 }
 651
 652                                 /*
 653                                  * We are only interested in the places where
 654                                  * we wrote or trimmed IOs. Turn those into
 655                                  * reads for verification purposes.
 656                                  */
 657                                 if (io_u->ddir == DDIR_READ) {
 658                                         /*
 659                                          * Pretend we issued it for rwmix
 660                                          * accounting
 661                                          */
 662                                         td->io_issues[DDIR_READ]++;
 663                                         put_io_u(td, io_u);
 664                                         continue;
 665                                 } else if (io_u->ddir == DDIR_TRIM) {
 666                                         io_u->ddir = DDIR_READ;
 667                                         io_u_set(io_u, IO_U_F_TRIMMED);
 668                                         break;
 669                                 } else if (io_u->ddir == DDIR_WRITE) {
 670                                         io_u->ddir = DDIR_READ;
 671                                         break;
 672                                 } else {
 673                                         put_io_u(td, io_u);
 674                                         continue;
 675                                 }
 676                         }
 677
 678                         if (!io_u)
 679                                 break;
 680                 }
 681
 682                 if (verify_state_should_stop(td, io_u)) {
 683                         put_io_u(td, io_u);
 684                         break;
 685                 }
 686
 687                 if (td->o.verify_async)
 688                         io_u->end_io = verify_io_u_async;
 689                 else
 690                         io_u->end_io = verify_io_u;
 691
 692                 ddir = io_u->ddir;
 693                 if (!td->o.disable_slat)
 694                         fio_gettime(&io_u->start_time, NULL);
 695
 696                 ret = td_io_queue(td, io_u);
 697
 698                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 699                         break;
 700
 701                 /*
 702                  * if we can queue more, do so. but check if there are
 703                  * completed io_u's first. Note that we can get BUSY even
 704                  * without IO queued, if the system is resource starved.
 705                  */
 706 reap:
 707                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 708                 if (full || io_in_polling(td))
 709                         ret = wait_for_completions(td, NULL);
 710
 711                 if (ret < 0)
 712                         break;
 713         }
 714
 715         check_update_rusage(td);
 716
 717         if (!td->error) {
 718                 min_events = td->cur_depth;
 719
 720                 if (min_events)
 721                         ret = io_u_queued_complete(td, min_events);
 722         } else
 723                 cleanup_pending_aio(td);
 724
 725         td_set_runstate(td, TD_RUNNING);
 726
 727         dprint(FD_VERIFY, "exiting loop\n");
 728 }
 729
 730 static bool exceeds_number_ios(struct thread_data *td)
 731 {
 732         unsigned long long number_ios;
 733
 734         if (!td->o.number_ios)
 735                 return false;
 736
 737         number_ios = ddir_rw_sum(td->io_blocks);
 738         number_ios += td->io_u_queued + td->io_u_in_flight;
 739
 740         return number_ios >= (td->o.number_ios * td->loops);
 741 }
 742
 743 static bool io_issue_bytes_exceeded(struct thread_data *td)
 744 {
 745         unsigned long long bytes, limit;
 746
 747         if (td_rw(td))
 748                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 749         else if (td_write(td))
 750                 bytes = td->io_issue_bytes[DDIR_WRITE];
 751         else if (td_read(td))
 752                 bytes = td->io_issue_bytes[DDIR_READ];
 753         else
 754                 bytes = td->io_issue_bytes[DDIR_TRIM];
 755
 756         if (td->o.io_limit)
 757                 limit = td->o.io_limit;
 758         else
 759                 limit = td->o.size;
 760
 761         limit *= td->loops;
 762         return bytes >= limit || exceeds_number_ios(td);
 763 }
 764
 765 static bool io_complete_bytes_exceeded(struct thread_data *td)
 766 {
 767         unsigned long long bytes, limit;
 768
 769         if (td_rw(td))
 770                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 771         else if (td_write(td))
 772                 bytes = td->this_io_bytes[DDIR_WRITE];
 773         else if (td_read(td))
 774                 bytes = td->this_io_bytes[DDIR_READ];
 775         else
 776                 bytes = td->this_io_bytes[DDIR_TRIM];
 777
 778         if (td->o.io_limit)
 779                 limit = td->o.io_limit;
 780         else
 781                 limit = td->o.size;
 782
 783         limit *= td->loops;
 784         return bytes >= limit || exceeds_number_ios(td);
 785 }
 786
 787 /*
 788  * used to calculate the next io time for rate control
 789  *
 790  */
 791 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 792 {
 793         uint64_t secs, remainder, bps, bytes, iops;
 794
 795         assert(!(td->flags & TD_F_CHILD));
 796         bytes = td->rate_io_issue_bytes[ddir];
 797         bps = td->rate_bps[ddir];
 798
 799         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 800                 uint64_t val;
 801                 iops = bps / td->o.bs[ddir];
 802                 val = (int64_t) (1000000 / iops) *
 803                                 -logf(__rand_0_1(&td->poisson_state));
 804                 if (val) {
 805                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 806                                         (unsigned long long) 1000000 / val);
 807                 }
 808                 td->last_usec += val;
 809                 return td->last_usec;
 810         } else if (bps) {
 811                 secs = bytes / bps;
 812                 remainder = bytes % bps;
 813                 return remainder * 1000000 / bps + secs * 1000000;
 814         }
 815
 816         return 0;
 817 }
 818
 819 /*
 820  * Main IO worker function. It retrieves io_u's to process and queues
 821  * and reaps them, checking for rate and errors along the way.
 822  *
 823  * Returns number of bytes written and trimmed.
 824  */
 825 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 826 {
 827         unsigned int i;
 828         int ret = 0;
 829         uint64_t total_bytes, bytes_issued = 0;
 830
 831         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 832                 bytes_done[i] = td->bytes_done[i];
 833
 834         if (in_ramp_time(td))
 835                 td_set_runstate(td, TD_RAMP);
 836         else
 837                 td_set_runstate(td, TD_RUNNING);
 838
 839         lat_target_init(td);
 840
 841         total_bytes = td->o.size;
 842         /*
 843         * Allow random overwrite workloads to write up to io_limit
 844         * before starting verification phase as 'size' doesn't apply.
 845         */
 846         if (td_write(td) && td_random(td) && td->o.norandommap)
 847                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 848         /*
 849          * If verify_backlog is enabled, we'll run the verify in this
 850          * handler as well. For that case, we may need up to twice the
 851          * amount of bytes.
 852          */
 853         if (td->o.verify != VERIFY_NONE &&
 854            (td_write(td) && td->o.verify_backlog))
 855                 total_bytes += td->o.size;
 856
 857         /* In trimwrite mode, each byte is trimmed and then written, so
 858          * allow total_bytes to be twice as big */
 859         if (td_trimwrite(td))
 860                 total_bytes += td->total_io_size;
 861
 862         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 863                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 864                 td->o.time_based) {
 865                 struct timeval comp_time;
 866                 struct io_u *io_u;
 867                 int full;
 868                 enum fio_ddir ddir;
 869
 870                 check_update_rusage(td);
 871
 872                 if (td->terminate || td->done)
 873                         break;
 874
 875                 update_tv_cache(td);
 876
 877                 if (runtime_exceeded(td, &td->tv_cache)) {
 878                         __update_tv_cache(td);
 879                         if (runtime_exceeded(td, &td->tv_cache)) {
 880                                 fio_mark_td_terminate(td);
 881                                 break;
 882                         }
 883                 }
 884
 885                 if (flow_threshold_exceeded(td))
 886                         continue;
 887
 888                 /*
 889                  * Break if we exceeded the bytes. The exception is time
 890                  * based runs, but we still need to break out of the loop
 891                  * for those to run verification, if enabled.
 892                  */
 893                 if (bytes_issued >= total_bytes &&
 894                     (!td->o.time_based ||
 895                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 896                         break;
 897
 898                 io_u = get_io_u(td);
 899                 if (IS_ERR_OR_NULL(io_u)) {
 900                         int err = PTR_ERR(io_u);
 901
 902                         io_u = NULL;
 903                         if (err == -EBUSY) {
 904                                 ret = FIO_Q_BUSY;
 905                                 goto reap;
 906                         }
 907                         if (td->o.latency_target)
 908                                 goto reap;
 909                         break;
 910                 }
 911
 912                 ddir = io_u->ddir;
 913
 914                 /*
 915                  * Add verification end_io handler if:
 916                  *      - Asked to verify (!td_rw(td))
 917                  *      - Or the io_u is from our verify list (mixed write/ver)
 918                  */
 919                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 920                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 921
 922                         if (!td->o.verify_pattern_bytes) {
 923                                 io_u->rand_seed = __rand(&td->verify_state);
 924                                 if (sizeof(int) != sizeof(long *))
 925                                         io_u->rand_seed *= __rand(&td->verify_state);
 926                         }
 927
 928                         if (verify_state_should_stop(td, io_u)) {
 929                                 put_io_u(td, io_u);
 930                                 break;
 931                         }
 932
 933                         if (td->o.verify_async)
 934                                 io_u->end_io = verify_io_u_async;
 935                         else
 936                                 io_u->end_io = verify_io_u;
 937                         td_set_runstate(td, TD_VERIFYING);
 938                 } else if (in_ramp_time(td))
 939                         td_set_runstate(td, TD_RAMP);
 940                 else
 941                         td_set_runstate(td, TD_RUNNING);
 942
 943                 /*
 944                  * Always log IO before it's issued, so we know the specific
 945                  * order of it. The logged unit will track when the IO has
 946                  * completed.
 947                  */
 948                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 949                     td->o.do_verify &&
 950                     td->o.verify != VERIFY_NONE &&
 951                     !td->o.experimental_verify)
 952                         log_io_piece(td, io_u);
 953
 954                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 955                         const unsigned long blen = io_u->xfer_buflen;
 956                         const enum fio_ddir ddir = acct_ddir(io_u);
 957
 958                         if (td->error)
 959                                 break;
 960
 961                         workqueue_enqueue(&td->io_wq, &io_u->work);
 962                         ret = FIO_Q_QUEUED;
 963
 964                         if (ddir_rw(ddir)) {
 965                                 td->io_issues[ddir]++;
 966                                 td->io_issue_bytes[ddir] += blen;
 967                                 td->rate_io_issue_bytes[ddir] += blen;
 968                         }
 969
 970                         if (should_check_rate(td))
 971                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 972
 973                 } else {
 974                         ret = td_io_queue(td, io_u);
 975
 976                         if (should_check_rate(td))
 977                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 978
 979                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 980                                 break;
 981
 982                         /*
 983                          * See if we need to complete some commands. Note that
 984                          * we can get BUSY even without IO queued, if the
 985                          * system is resource starved.
 986                          */
 987 reap:
 988                         full = queue_full(td) ||
 989                                 (ret == FIO_Q_BUSY && td->cur_depth);
 990                         if (full || io_in_polling(td))
 991                                 ret = wait_for_completions(td, &comp_time);
 992                 }
 993                 if (ret < 0)
 994                         break;
 995                 if (!ddir_rw_sum(td->bytes_done) &&
 996                     !(td->io_ops->flags & FIO_NOIO))
 997                         continue;
 998
 999                 if (!in_ramp_time(td) && should_check_rate(td)) {
1000                         if (check_min_rate(td, &comp_time)) {
1001                                 if (exitall_on_terminate || td->o.exitall_error)
1002                                         fio_terminate_threads(td->groupid);
1003                                 td_verror(td, EIO, "check_min_rate");
1004                                 break;
1005                         }
1006                 }
1007                 if (!in_ramp_time(td) && td->o.latency_target)
1008                         lat_target_check(td);
1009
1010                 if (td->o.thinktime) {
1011                         unsigned long long b;
1012
1013                         b = ddir_rw_sum(td->io_blocks);
1014                         if (!(b % td->o.thinktime_blocks)) {
1015                                 int left;
1016
1017                                 io_u_quiesce(td);
1018
1019                                 if (td->o.thinktime_spin)
1020                                         usec_spin(td->o.thinktime_spin);
1021
1022                                 left = td->o.thinktime - td->o.thinktime_spin;
1023                                 if (left)
1024                                         usec_sleep(td, left);
1025                         }
1026                 }
1027         }
1028
1029         check_update_rusage(td);
1030
1031         if (td->trim_entries)
1032                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1033
1034         if (td->o.fill_device && td->error == ENOSPC) {
1035                 td->error = 0;
1036                 fio_mark_td_terminate(td);
1037         }
1038         if (!td->error) {
1039                 struct fio_file *f;
1040
1041                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1042                         workqueue_flush(&td->io_wq);
1043                         i = 0;
1044                 } else
1045                         i = td->cur_depth;
1046
1047                 if (i) {
1048                         ret = io_u_queued_complete(td, i);
1049                         if (td->o.fill_device && td->error == ENOSPC)
1050                                 td->error = 0;
1051                 }
1052
1053                 if (should_fsync(td) && td->o.end_fsync) {
1054                         td_set_runstate(td, TD_FSYNCING);
1055
1056                         for_each_file(td, f, i) {
1057                                 if (!fio_file_fsync(td, f))
1058                                         continue;
1059
1060                                 log_err("fio: end_fsync failed for file %s\n",
1061                                                                 f->file_name);
1062                         }
1063                 }
1064         } else
1065                 cleanup_pending_aio(td);
1066
1067         /*
1068          * stop job if we failed doing any IO
1069          */
1070         if (!ddir_rw_sum(td->this_io_bytes))
1071                 td->done = 1;
1072
1073         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1074                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1075 }
1076
1077 static void free_file_completion_logging(struct thread_data *td)
1078 {
1079         struct fio_file *f;
1080         unsigned int i;
1081
1082         for_each_file(td, f, i) {
1083                 if (!f->last_write_comp)
1084                         break;
1085                 sfree(f->last_write_comp);
1086         }
1087 }
1088
1089 static int init_file_completion_logging(struct thread_data *td,
1090                                         unsigned int depth)
1091 {
1092         struct fio_file *f;
1093         unsigned int i;
1094
1095         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1096                 return 0;
1097
1098         for_each_file(td, f, i) {
1099                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1100                 if (!f->last_write_comp)
1101                         goto cleanup;
1102         }
1103
1104         return 0;
1105
1106 cleanup:
1107         free_file_completion_logging(td);
1108         log_err("fio: failed to alloc write comp data\n");
1109         return 1;
1110 }
1111
1112 static void cleanup_io_u(struct thread_data *td)
1113 {
1114         struct io_u *io_u;
1115
1116         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1117
1118                 if (td->io_ops->io_u_free)
1119                         td->io_ops->io_u_free(td, io_u);
1120
1121                 fio_memfree(io_u, sizeof(*io_u));
1122         }
1123
1124         free_io_mem(td);
1125
1126         io_u_rexit(&td->io_u_requeues);
1127         io_u_qexit(&td->io_u_freelist);
1128         io_u_qexit(&td->io_u_all);
1129
1130         free_file_completion_logging(td);
1131 }
1132
1133 static int init_io_u(struct thread_data *td)
1134 {
1135         struct io_u *io_u;
1136         unsigned int max_bs, min_write;
1137         int cl_align, i, max_units;
1138         int data_xfer = 1, err;
1139         char *p;
1140
1141         max_units = td->o.iodepth;
1142         max_bs = td_max_bs(td);
1143         min_write = td->o.min_bs[DDIR_WRITE];
1144         td->orig_buffer_size = (unsigned long long) max_bs
1145                                         * (unsigned long long) max_units;
1146
1147         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1148                 data_xfer = 0;
1149
1150         err = 0;
1151         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1152         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1153         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1154
1155         if (err) {
1156                 log_err("fio: failed setting up IO queues\n");
1157                 return 1;
1158         }
1159
1160         /*
1161          * if we may later need to do address alignment, then add any
1162          * possible adjustment here so that we don't cause a buffer
1163          * overflow later. this adjustment may be too much if we get
1164          * lucky and the allocator gives us an aligned address.
1165          */
1166         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1167             (td->io_ops->flags & FIO_RAWIO))
1168                 td->orig_buffer_size += page_mask + td->o.mem_align;
1169
1170         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1171                 unsigned long bs;
1172
1173                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1174                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1175         }
1176
1177         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1178                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1179                 return 1;
1180         }
1181
1182         if (data_xfer && allocate_io_mem(td))
1183                 return 1;
1184
1185         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1186             (td->io_ops->flags & FIO_RAWIO))
1187                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1188         else
1189                 p = td->orig_buffer;
1190
1191         cl_align = os_cache_line_size();
1192
1193         for (i = 0; i < max_units; i++) {
1194                 void *ptr;
1195
1196                 if (td->terminate)
1197                         return 1;
1198
1199                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1200                 if (!ptr) {
1201                         log_err("fio: unable to allocate aligned memory\n");
1202                         break;
1203                 }
1204
1205                 io_u = ptr;
1206                 memset(io_u, 0, sizeof(*io_u));
1207                 INIT_FLIST_HEAD(&io_u->verify_list);
1208                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1209
1210                 if (data_xfer) {
1211                         io_u->buf = p;
1212                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1213
1214                         if (td_write(td))
1215                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1216                         if (td_write(td) && td->o.verify_pattern_bytes) {
1217                                 /*
1218                                  * Fill the buffer with the pattern if we are
1219                                  * going to be doing writes.
1220                                  */
1221                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1222                         }
1223                 }
1224
1225                 io_u->index = i;
1226                 io_u->flags = IO_U_F_FREE;
1227                 io_u_qpush(&td->io_u_freelist, io_u);
1228
1229                 /*
1230                  * io_u never leaves this stack, used for iteration of all
1231                  * io_u buffers.
1232                  */
1233                 io_u_qpush(&td->io_u_all, io_u);
1234
1235                 if (td->io_ops->io_u_init) {
1236                         int ret = td->io_ops->io_u_init(td, io_u);
1237
1238                         if (ret) {
1239                                 log_err("fio: failed to init engine data: %d\n", ret);
1240                                 return 1;
1241                         }
1242                 }
1243
1244                 p += max_bs;
1245         }
1246
1247         if (init_file_completion_logging(td, max_units))
1248                 return 1;
1249
1250         return 0;
1251 }
1252
1253 static int switch_ioscheduler(struct thread_data *td)
1254 {
1255         char tmp[256], tmp2[128];
1256         FILE *f;
1257         int ret;
1258
1259         if (td->io_ops->flags & FIO_DISKLESSIO)
1260                 return 0;
1261
1262         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1263
1264         f = fopen(tmp, "r+");
1265         if (!f) {
1266                 if (errno == ENOENT) {
1267                         log_err("fio: os or kernel doesn't support IO scheduler"
1268                                 " switching\n");
1269                         return 0;
1270                 }
1271                 td_verror(td, errno, "fopen iosched");
1272                 return 1;
1273         }
1274
1275         /*
1276          * Set io scheduler.
1277          */
1278         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1279         if (ferror(f) || ret != 1) {
1280                 td_verror(td, errno, "fwrite");
1281                 fclose(f);
1282                 return 1;
1283         }
1284
1285         rewind(f);
1286
1287         /*
1288          * Read back and check that the selected scheduler is now the default.
1289          */
1290         memset(tmp, 0, sizeof(tmp));
1291         ret = fread(tmp, sizeof(tmp), 1, f);
1292         if (ferror(f) || ret < 0) {
1293                 td_verror(td, errno, "fread");
1294                 fclose(f);
1295                 return 1;
1296         }
1297         /*
1298          * either a list of io schedulers or "none\n" is expected.
1299          */
1300         tmp[strlen(tmp) - 1] = '\0';
1301
1302
1303         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1304         if (!strstr(tmp, tmp2)) {
1305                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1306                 td_verror(td, EINVAL, "iosched_switch");
1307                 fclose(f);
1308                 return 1;
1309         }
1310
1311         fclose(f);
1312         return 0;
1313 }
1314
1315 static bool keep_running(struct thread_data *td)
1316 {
1317         unsigned long long limit;
1318
1319         if (td->done)
1320                 return false;
1321         if (td->o.time_based)
1322                 return true;
1323         if (td->o.loops) {
1324                 td->o.loops--;
1325                 return true;
1326         }
1327         if (exceeds_number_ios(td))
1328                 return false;
1329
1330         if (td->o.io_limit)
1331                 limit = td->o.io_limit;
1332         else
1333                 limit = td->o.size;
1334
1335         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1336                 uint64_t diff;
1337
1338                 /*
1339                  * If the difference is less than the minimum IO size, we
1340                  * are done.
1341                  */
1342                 diff = limit - ddir_rw_sum(td->io_bytes);
1343                 if (diff < td_max_bs(td))
1344                         return false;
1345
1346                 if (fio_files_done(td) && !td->o.io_limit)
1347                         return false;
1348
1349                 return true;
1350         }
1351
1352         return false;
1353 }
1354
1355 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1356 {
1357         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1358         int ret;
1359         char *str;
1360
1361         str = malloc(newlen);
1362         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1363
1364         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1365         ret = system(str);
1366         if (ret == -1)
1367                 log_err("fio: exec of cmd <%s> failed\n", str);
1368
1369         free(str);
1370         return ret;
1371 }
1372
1373 /*
1374  * Dry run to compute correct state of numberio for verification.
1375  */
1376 static uint64_t do_dry_run(struct thread_data *td)
1377 {
1378         td_set_runstate(td, TD_RUNNING);
1379
1380         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1381                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1382                 struct io_u *io_u;
1383                 int ret;
1384
1385                 if (td->terminate || td->done)
1386                         break;
1387
1388                 io_u = get_io_u(td);
1389                 if (!io_u)
1390                         break;
1391
1392                 io_u_set(io_u, IO_U_F_FLIGHT);
1393                 io_u->error = 0;
1394                 io_u->resid = 0;
1395                 if (ddir_rw(acct_ddir(io_u)))
1396                         td->io_issues[acct_ddir(io_u)]++;
1397                 if (ddir_rw(io_u->ddir)) {
1398                         io_u_mark_depth(td, 1);
1399                         td->ts.total_io_u[io_u->ddir]++;
1400                 }
1401
1402                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1403                     td->o.do_verify &&
1404                     td->o.verify != VERIFY_NONE &&
1405                     !td->o.experimental_verify)
1406                         log_io_piece(td, io_u);
1407
1408                 ret = io_u_sync_complete(td, io_u);
1409                 (void) ret;
1410         }
1411
1412         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1413 }
1414
1415 struct fork_data {
1416         struct thread_data *td;
1417         struct sk_out *sk_out;
1418 };
1419
1420 /*
1421  * Entry point for the thread based jobs. The process based jobs end up
1422  * here as well, after a little setup.
1423  */
1424 static void *thread_main(void *data)
1425 {
1426         struct fork_data *fd = data;
1427         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1428         struct thread_data *td = fd->td;
1429         struct thread_options *o = &td->o;
1430         struct sk_out *sk_out = fd->sk_out;
1431         pthread_condattr_t attr;
1432         int clear_state;
1433         int ret;
1434
1435         sk_out_assign(sk_out);
1436         free(fd);
1437
1438         if (!o->use_thread) {
1439                 setsid();
1440                 td->pid = getpid();
1441         } else
1442                 td->pid = gettid();
1443
1444         fio_local_clock_init(o->use_thread);
1445
1446         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1447
1448         if (is_backend)
1449                 fio_server_send_start(td);
1450
1451         INIT_FLIST_HEAD(&td->io_log_list);
1452         INIT_FLIST_HEAD(&td->io_hist_list);
1453         INIT_FLIST_HEAD(&td->verify_list);
1454         INIT_FLIST_HEAD(&td->trim_list);
1455         INIT_FLIST_HEAD(&td->next_rand_list);
1456         pthread_mutex_init(&td->io_u_lock, NULL);
1457         td->io_hist_tree = RB_ROOT;
1458
1459         pthread_condattr_init(&attr);
1460         pthread_cond_init(&td->verify_cond, &attr);
1461         pthread_cond_init(&td->free_cond, &attr);
1462
1463         td_set_runstate(td, TD_INITIALIZED);
1464         dprint(FD_MUTEX, "up startup_mutex\n");
1465         fio_mutex_up(startup_mutex);
1466         dprint(FD_MUTEX, "wait on td->mutex\n");
1467         fio_mutex_down(td->mutex);
1468         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1469
1470         /*
1471          * A new gid requires privilege, so we need to do this before setting
1472          * the uid.
1473          */
1474         if (o->gid != -1U && setgid(o->gid)) {
1475                 td_verror(td, errno, "setgid");
1476                 goto err;
1477         }
1478         if (o->uid != -1U && setuid(o->uid)) {
1479                 td_verror(td, errno, "setuid");
1480                 goto err;
1481         }
1482
1483         /*
1484          * If we have a gettimeofday() thread, make sure we exclude that
1485          * thread from this job
1486          */
1487         if (o->gtod_cpu)
1488                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1489
1490         /*
1491          * Set affinity first, in case it has an impact on the memory
1492          * allocations.
1493          */
1494         if (fio_option_is_set(o, cpumask)) {
1495                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1496                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1497                         if (!ret) {
1498                                 log_err("fio: no CPUs set\n");
1499                                 log_err("fio: Try increasing number of available CPUs\n");
1500                                 td_verror(td, EINVAL, "cpus_split");
1501                                 goto err;
1502                         }
1503                 }
1504                 ret = fio_setaffinity(td->pid, o->cpumask);
1505                 if (ret == -1) {
1506                         td_verror(td, errno, "cpu_set_affinity");
1507                         goto err;
1508                 }
1509         }
1510
1511 #ifdef CONFIG_LIBNUMA
1512         /* numa node setup */
1513         if (fio_option_is_set(o, numa_cpunodes) ||
1514             fio_option_is_set(o, numa_memnodes)) {
1515                 struct bitmask *mask;
1516
1517                 if (numa_available() < 0) {
1518                         td_verror(td, errno, "Does not support NUMA API\n");
1519                         goto err;
1520                 }
1521
1522                 if (fio_option_is_set(o, numa_cpunodes)) {
1523                         mask = numa_parse_nodestring(o->numa_cpunodes);
1524                         ret = numa_run_on_node_mask(mask);
1525                         numa_free_nodemask(mask);
1526                         if (ret == -1) {
1527                                 td_verror(td, errno, \
1528                                         "numa_run_on_node_mask failed\n");
1529                                 goto err;
1530                         }
1531                 }
1532
1533                 if (fio_option_is_set(o, numa_memnodes)) {
1534                         mask = NULL;
1535                         if (o->numa_memnodes)
1536                                 mask = numa_parse_nodestring(o->numa_memnodes);
1537
1538                         switch (o->numa_mem_mode) {
1539                         case MPOL_INTERLEAVE:
1540                                 numa_set_interleave_mask(mask);
1541                                 break;
1542                         case MPOL_BIND:
1543                                 numa_set_membind(mask);
1544                                 break;
1545                         case MPOL_LOCAL:
1546                                 numa_set_localalloc();
1547                                 break;
1548                         case MPOL_PREFERRED:
1549                                 numa_set_preferred(o->numa_mem_prefer_node);
1550                                 break;
1551                         case MPOL_DEFAULT:
1552                         default:
1553                                 break;
1554                         }
1555
1556                         if (mask)
1557                                 numa_free_nodemask(mask);
1558
1559                 }
1560         }
1561 #endif
1562
1563         if (fio_pin_memory(td))
1564                 goto err;
1565
1566         /*
1567          * May alter parameters that init_io_u() will use, so we need to
1568          * do this first.
1569          */
1570         if (init_iolog(td))
1571                 goto err;
1572
1573         if (init_io_u(td))
1574                 goto err;
1575
1576         if (o->verify_async && verify_async_init(td))
1577                 goto err;
1578
1579         if (fio_option_is_set(o, ioprio) ||
1580             fio_option_is_set(o, ioprio_class)) {
1581                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1582                 if (ret == -1) {
1583                         td_verror(td, errno, "ioprio_set");
1584                         goto err;
1585                 }
1586         }
1587
1588         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1589                 goto err;
1590
1591         errno = 0;
1592         if (nice(o->nice) == -1 && errno != 0) {
1593                 td_verror(td, errno, "nice");
1594                 goto err;
1595         }
1596
1597         if (o->ioscheduler && switch_ioscheduler(td))
1598                 goto err;
1599
1600         if (!o->create_serialize && setup_files(td))
1601                 goto err;
1602
1603         if (td_io_init(td))
1604                 goto err;
1605
1606         if (init_random_map(td))
1607                 goto err;
1608
1609         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1610                 goto err;
1611
1612         if (o->pre_read) {
1613                 if (pre_read_files(td) < 0)
1614                         goto err;
1615         }
1616
1617         if (iolog_compress_init(td, sk_out))
1618                 goto err;
1619
1620         fio_verify_init(td);
1621
1622         if (rate_submit_init(td, sk_out))
1623                 goto err;
1624
1625         fio_gettime(&td->epoch, NULL);
1626         fio_getrusage(&td->ru_start);
1627         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1628         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1629
1630         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1631                         o->ratemin[DDIR_TRIM]) {
1632                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1633                                         sizeof(td->bw_sample_time));
1634                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1635                                         sizeof(td->bw_sample_time));
1636                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1637                                         sizeof(td->bw_sample_time));
1638         }
1639
1640         clear_state = 0;
1641         while (keep_running(td)) {
1642                 uint64_t verify_bytes;
1643
1644                 fio_gettime(&td->start, NULL);
1645                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1646
1647                 if (clear_state)
1648                         clear_io_state(td, 0);
1649
1650                 prune_io_piece_log(td);
1651
1652                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1653                         verify_bytes = do_dry_run(td);
1654                 else {
1655                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1656
1657                         do_io(td, bytes_done);
1658
1659                         if (!ddir_rw_sum(bytes_done)) {
1660                                 fio_mark_td_terminate(td);
1661                                 verify_bytes = 0;
1662                         } else {
1663                                 verify_bytes = bytes_done[DDIR_WRITE] +
1664                                                 bytes_done[DDIR_TRIM];
1665                         }
1666                 }
1667
1668                 clear_state = 1;
1669
1670                 /*
1671                  * Make sure we've successfully updated the rusage stats
1672                  * before waiting on the stat mutex. Otherwise we could have
1673                  * the stat thread holding stat mutex and waiting for
1674                  * the rusage_sem, which would never get upped because
1675                  * this thread is waiting for the stat mutex.
1676                  */
1677                 check_update_rusage(td);
1678
1679                 fio_mutex_down(stat_mutex);
1680                 if (td_read(td) && td->io_bytes[DDIR_READ])
1681                         update_runtime(td, elapsed_us, DDIR_READ);
1682                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1683                         update_runtime(td, elapsed_us, DDIR_WRITE);
1684                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1685                         update_runtime(td, elapsed_us, DDIR_TRIM);
1686                 fio_gettime(&td->start, NULL);
1687                 fio_mutex_up(stat_mutex);
1688
1689                 if (td->error || td->terminate)
1690                         break;
1691
1692                 if (!o->do_verify ||
1693                     o->verify == VERIFY_NONE ||
1694                     (td->io_ops->flags & FIO_UNIDIR))
1695                         continue;
1696
1697                 clear_io_state(td, 0);
1698
1699                 fio_gettime(&td->start, NULL);
1700
1701                 do_verify(td, verify_bytes);
1702
1703                 /*
1704                  * See comment further up for why this is done here.
1705                  */
1706                 check_update_rusage(td);
1707
1708                 fio_mutex_down(stat_mutex);
1709                 update_runtime(td, elapsed_us, DDIR_READ);
1710                 fio_gettime(&td->start, NULL);
1711                 fio_mutex_up(stat_mutex);
1712
1713                 if (td->error || td->terminate)
1714                         break;
1715         }
1716
1717         update_rusage_stat(td);
1718         td->ts.total_run_time = mtime_since_now(&td->epoch);
1719         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1720         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1721         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1722
1723         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1724             (td->o.verify != VERIFY_NONE && td_write(td)))
1725                 verify_save_state(td->thread_number);
1726
1727         fio_unpin_memory(td);
1728
1729         td_writeout_logs(td, true);
1730
1731         iolog_compress_exit(td);
1732         rate_submit_exit(td);
1733
1734         if (o->exec_postrun)
1735                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1736
1737         if (exitall_on_terminate || (o->exitall_error && td->error))
1738                 fio_terminate_threads(td->groupid);
1739
1740 err:
1741         if (td->error)
1742                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1743                                                         td->verror);
1744
1745         if (o->verify_async)
1746                 verify_async_exit(td);
1747
1748         close_and_free_files(td);
1749         cleanup_io_u(td);
1750         close_ioengine(td);
1751         cgroup_shutdown(td, &cgroup_mnt);
1752         verify_free_state(td);
1753
1754         if (td->zone_state_index) {
1755                 int i;
1756
1757                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1758                         free(td->zone_state_index[i]);
1759                 free(td->zone_state_index);
1760                 td->zone_state_index = NULL;
1761         }
1762
1763         if (fio_option_is_set(o, cpumask)) {
1764                 ret = fio_cpuset_exit(&o->cpumask);
1765                 if (ret)
1766                         td_verror(td, ret, "fio_cpuset_exit");
1767         }
1768
1769         /*
1770          * do this very late, it will log file closing as well
1771          */
1772         if (o->write_iolog_file)
1773                 write_iolog_close(td);
1774
1775         fio_mutex_remove(td->mutex);
1776         td->mutex = NULL;
1777
1778         td_set_runstate(td, TD_EXITED);
1779
1780         /*
1781          * Do this last after setting our runstate to exited, so we
1782          * know that the stat thread is signaled.
1783          */
1784         check_update_rusage(td);
1785
1786         sk_out_drop();
1787         return (void *) (uintptr_t) td->error;
1788 }
1789
1790
1791 /*
1792  * We cannot pass the td data into a forked process, so attach the td and
1793  * pass it to the thread worker.
1794  */
1795 static int fork_main(struct sk_out *sk_out, int shmid, int offset)
1796 {
1797         struct fork_data *fd;
1798         void *data, *ret;
1799
1800 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1801         data = shmat(shmid, NULL, 0);
1802         if (data == (void *) -1) {
1803                 int __err = errno;
1804
1805                 perror("shmat");
1806                 return __err;
1807         }
1808 #else
1809         /*
1810          * HP-UX inherits shm mappings?
1811          */
1812         data = threads;
1813 #endif
1814
1815         fd = calloc(1, sizeof(*fd));
1816         fd->td = data + offset * sizeof(struct thread_data);
1817         fd->sk_out = sk_out;
1818         ret = thread_main(fd);
1819         shmdt(data);
1820         return (int) (uintptr_t) ret;
1821 }
1822
1823 static void dump_td_info(struct thread_data *td)
1824 {
1825         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1826                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1827                         (unsigned long) time_since_now(&td->terminate_time));
1828 }
1829
1830 /*
1831  * Run over the job map and reap the threads that have exited, if any.
1832  */
1833 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1834                          unsigned int *m_rate)
1835 {
1836         struct thread_data *td;
1837         unsigned int cputhreads, realthreads, pending;
1838         int i, status, ret;
1839
1840         /*
1841          * reap exited threads (TD_EXITED -> TD_REAPED)
1842          */
1843         realthreads = pending = cputhreads = 0;
1844         for_each_td(td, i) {
1845                 int flags = 0;
1846
1847                 /*
1848                  * ->io_ops is NULL for a thread that has closed its
1849                  * io engine
1850                  */
1851                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1852                         cputhreads++;
1853                 else
1854                         realthreads++;
1855
1856                 if (!td->pid) {
1857                         pending++;
1858                         continue;
1859                 }
1860                 if (td->runstate == TD_REAPED)
1861                         continue;
1862                 if (td->o.use_thread) {
1863                         if (td->runstate == TD_EXITED) {
1864                                 td_set_runstate(td, TD_REAPED);
1865                                 goto reaped;
1866                         }
1867                         continue;
1868                 }
1869
1870                 flags = WNOHANG;
1871                 if (td->runstate == TD_EXITED)
1872                         flags = 0;
1873
1874                 /*
1875                  * check if someone quit or got killed in an unusual way
1876                  */
1877                 ret = waitpid(td->pid, &status, flags);
1878                 if (ret < 0) {
1879                         if (errno == ECHILD) {
1880                                 log_err("fio: pid=%d disappeared %d\n",
1881                                                 (int) td->pid, td->runstate);
1882                                 td->sig = ECHILD;
1883                                 td_set_runstate(td, TD_REAPED);
1884                                 goto reaped;
1885                         }
1886                         perror("waitpid");
1887                 } else if (ret == td->pid) {
1888                         if (WIFSIGNALED(status)) {
1889                                 int sig = WTERMSIG(status);
1890
1891                                 if (sig != SIGTERM && sig != SIGUSR2)
1892                                         log_err("fio: pid=%d, got signal=%d\n",
1893                                                         (int) td->pid, sig);
1894                                 td->sig = sig;
1895                                 td_set_runstate(td, TD_REAPED);
1896                                 goto reaped;
1897                         }
1898                         if (WIFEXITED(status)) {
1899                                 if (WEXITSTATUS(status) && !td->error)
1900                                         td->error = WEXITSTATUS(status);
1901
1902                                 td_set_runstate(td, TD_REAPED);
1903                                 goto reaped;
1904                         }
1905                 }
1906
1907                 /*
1908                  * If the job is stuck, do a forceful timeout of it and
1909                  * move on.
1910                  */
1911                 if (td->terminate &&
1912                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1913                         dump_td_info(td);
1914                         td_set_runstate(td, TD_REAPED);
1915                         goto reaped;
1916                 }
1917
1918                 /*
1919                  * thread is not dead, continue
1920                  */
1921                 pending++;
1922                 continue;
1923 reaped:
1924                 (*nr_running)--;
1925                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1926                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1927                 if (!td->pid)
1928                         pending--;
1929
1930                 if (td->error)
1931                         exit_value++;
1932
1933                 done_secs += mtime_since_now(&td->epoch) / 1000;
1934                 profile_td_exit(td);
1935         }
1936
1937         if (*nr_running == cputhreads && !pending && realthreads)
1938                 fio_terminate_threads(TERMINATE_ALL);
1939 }
1940
1941 static bool __check_trigger_file(void)
1942 {
1943         struct stat sb;
1944
1945         if (!trigger_file)
1946                 return false;
1947
1948         if (stat(trigger_file, &sb))
1949                 return false;
1950
1951         if (unlink(trigger_file) < 0)
1952                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1953                                                         strerror(errno));
1954
1955         return true;
1956 }
1957
1958 static bool trigger_timedout(void)
1959 {
1960         if (trigger_timeout)
1961                 return time_since_genesis() >= trigger_timeout;
1962
1963         return false;
1964 }
1965
1966 void exec_trigger(const char *cmd)
1967 {
1968         int ret;
1969
1970         if (!cmd)
1971                 return;
1972
1973         ret = system(cmd);
1974         if (ret == -1)
1975                 log_err("fio: failed executing %s trigger\n", cmd);
1976 }
1977
1978 void check_trigger_file(void)
1979 {
1980         if (__check_trigger_file() || trigger_timedout()) {
1981                 if (nr_clients)
1982                         fio_clients_send_trigger(trigger_remote_cmd);
1983                 else {
1984                         verify_save_state(IO_LIST_ALL);
1985                         fio_terminate_threads(TERMINATE_ALL);
1986                         exec_trigger(trigger_cmd);
1987                 }
1988         }
1989 }
1990
1991 static int fio_verify_load_state(struct thread_data *td)
1992 {
1993         int ret;
1994
1995         if (!td->o.verify_state)
1996                 return 0;
1997
1998         if (is_backend) {
1999                 void *data;
2000
2001                 ret = fio_server_get_verify_state(td->o.name,
2002                                         td->thread_number - 1, &data);
2003                 if (!ret)
2004                         verify_assign_state(td, data);
2005         } else
2006                 ret = verify_load_state(td, "local");
2007
2008         return ret;
2009 }
2010
2011 static void do_usleep(unsigned int usecs)
2012 {
2013         check_for_running_stats();
2014         check_trigger_file();
2015         usleep(usecs);
2016 }
2017
2018 static bool check_mount_writes(struct thread_data *td)
2019 {
2020         struct fio_file *f;
2021         unsigned int i;
2022
2023         if (!td_write(td) || td->o.allow_mounted_write)
2024                 return false;
2025
2026         for_each_file(td, f, i) {
2027                 if (f->filetype != FIO_TYPE_BD)
2028                         continue;
2029                 if (device_is_mounted(f->file_name))
2030                         goto mounted;
2031         }
2032
2033         return false;
2034 mounted:
2035         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2036         return true;
2037 }
2038
2039 static bool waitee_running(struct thread_data *me)
2040 {
2041         const char *waitee = me->o.wait_for;
2042         const char *self = me->o.name;
2043         struct thread_data *td;
2044         int i;
2045
2046         if (!waitee)
2047                 return false;
2048
2049         for_each_td(td, i) {
2050                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2051                         continue;
2052
2053                 if (td->runstate < TD_EXITED) {
2054                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2055                                         self, td->o.name,
2056                                         runstate_to_name(td->runstate));
2057                         return true;
2058                 }
2059         }
2060
2061         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2062         return false;
2063 }
2064
2065 /*
2066  * Main function for kicking off and reaping jobs, as needed.
2067  */
2068 static void run_threads(struct sk_out *sk_out)
2069 {
2070         struct thread_data *td;
2071         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2072         uint64_t spent;
2073
2074         if (fio_gtod_offload && fio_start_gtod_thread())
2075                 return;
2076
2077         fio_idle_prof_init();
2078
2079         set_sig_handlers();
2080
2081         nr_thread = nr_process = 0;
2082         for_each_td(td, i) {
2083                 if (check_mount_writes(td))
2084                         return;
2085                 if (td->o.use_thread)
2086                         nr_thread++;
2087                 else
2088                         nr_process++;
2089         }
2090
2091         if (output_format & FIO_OUTPUT_NORMAL) {
2092                 log_info("Starting ");
2093                 if (nr_thread)
2094                         log_info("%d thread%s", nr_thread,
2095                                                 nr_thread > 1 ? "s" : "");
2096                 if (nr_process) {
2097                         if (nr_thread)
2098                                 log_info(" and ");
2099                         log_info("%d process%s", nr_process,
2100                                                 nr_process > 1 ? "es" : "");
2101                 }
2102                 log_info("\n");
2103                 log_info_flush();
2104         }
2105
2106         todo = thread_number;
2107         nr_running = 0;
2108         nr_started = 0;
2109         m_rate = t_rate = 0;
2110
2111         for_each_td(td, i) {
2112                 print_status_init(td->thread_number - 1);
2113
2114                 if (!td->o.create_serialize)
2115                         continue;
2116
2117                 if (fio_verify_load_state(td))
2118                         goto reap;
2119
2120                 /*
2121                  * do file setup here so it happens sequentially,
2122                  * we don't want X number of threads getting their
2123                  * client data interspersed on disk
2124                  */
2125                 if (setup_files(td)) {
2126 reap:
2127                         exit_value++;
2128                         if (td->error)
2129                                 log_err("fio: pid=%d, err=%d/%s\n",
2130                                         (int) td->pid, td->error, td->verror);
2131                         td_set_runstate(td, TD_REAPED);
2132                         todo--;
2133                 } else {
2134                         struct fio_file *f;
2135                         unsigned int j;
2136
2137                         /*
2138                          * for sharing to work, each job must always open
2139                          * its own files. so close them, if we opened them
2140                          * for creation
2141                          */
2142                         for_each_file(td, f, j) {
2143                                 if (fio_file_open(f))
2144                                         td_io_close_file(td, f);
2145                         }
2146                 }
2147         }
2148
2149         /* start idle threads before io threads start to run */
2150         fio_idle_prof_start();
2151
2152         set_genesis_time();
2153
2154         while (todo) {
2155                 struct thread_data *map[REAL_MAX_JOBS];
2156                 struct timeval this_start;
2157                 int this_jobs = 0, left;
2158
2159                 /*
2160                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2161                  */
2162                 for_each_td(td, i) {
2163                         if (td->runstate != TD_NOT_CREATED)
2164                                 continue;
2165
2166                         /*
2167                          * never got a chance to start, killed by other
2168                          * thread for some reason
2169                          */
2170                         if (td->terminate) {
2171                                 todo--;
2172                                 continue;
2173                         }
2174
2175                         if (td->o.start_delay) {
2176                                 spent = utime_since_genesis();
2177
2178                                 if (td->o.start_delay > spent)
2179                                         continue;
2180                         }
2181
2182                         if (td->o.stonewall && (nr_started || nr_running)) {
2183                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2184                                                         td->o.name);
2185                                 break;
2186                         }
2187
2188                         if (waitee_running(td)) {
2189                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2190                                                 td->o.name, td->o.wait_for);
2191                                 continue;
2192                         }
2193
2194                         init_disk_util(td);
2195
2196                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2197                         td->update_rusage = 0;
2198
2199                         /*
2200                          * Set state to created. Thread will transition
2201                          * to TD_INITIALIZED when it's done setting up.
2202                          */
2203                         td_set_runstate(td, TD_CREATED);
2204                         map[this_jobs++] = td;
2205                         nr_started++;
2206
2207                         if (td->o.use_thread) {
2208                                 struct fork_data *fd;
2209                                 int ret;
2210
2211                                 fd = calloc(1, sizeof(*fd));
2212                                 fd->td = td;
2213                                 fd->sk_out = sk_out;
2214
2215                                 dprint(FD_PROCESS, "will pthread_create\n");
2216                                 ret = pthread_create(&td->thread, NULL,
2217                                                         thread_main, fd);
2218                                 if (ret) {
2219                                         log_err("pthread_create: %s\n",
2220                                                         strerror(ret));
2221                                         free(fd);
2222                                         nr_started--;
2223                                         break;
2224                                 }
2225                                 ret = pthread_detach(td->thread);
2226                                 if (ret)
2227                                         log_err("pthread_detach: %s",
2228                                                         strerror(ret));
2229                         } else {
2230                                 pid_t pid;
2231                                 dprint(FD_PROCESS, "will fork\n");
2232                                 pid = fork();
2233                                 if (!pid) {
2234                                         int ret = fork_main(sk_out, shm_id, i);
2235
2236                                         _exit(ret);
2237                                 } else if (i == fio_debug_jobno)
2238                                         *fio_debug_jobp = pid;
2239                         }
2240                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2241                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2242                                 log_err("fio: job startup hung? exiting.\n");
2243                                 fio_terminate_threads(TERMINATE_ALL);
2244                                 fio_abort = 1;
2245                                 nr_started--;
2246                                 break;
2247                         }
2248                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2249                 }
2250
2251                 /*
2252                  * Wait for the started threads to transition to
2253                  * TD_INITIALIZED.
2254                  */
2255                 fio_gettime(&this_start, NULL);
2256                 left = this_jobs;
2257                 while (left && !fio_abort) {
2258                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2259                                 break;
2260
2261                         do_usleep(100000);
2262
2263                         for (i = 0; i < this_jobs; i++) {
2264                                 td = map[i];
2265                                 if (!td)
2266                                         continue;
2267                                 if (td->runstate == TD_INITIALIZED) {
2268                                         map[i] = NULL;
2269                                         left--;
2270                                 } else if (td->runstate >= TD_EXITED) {
2271                                         map[i] = NULL;
2272                                         left--;
2273                                         todo--;
2274                                         nr_running++; /* work-around... */
2275                                 }
2276                         }
2277                 }
2278
2279                 if (left) {
2280                         log_err("fio: %d job%s failed to start\n", left,
2281                                         left > 1 ? "s" : "");
2282                         for (i = 0; i < this_jobs; i++) {
2283                                 td = map[i];
2284                                 if (!td)
2285                                         continue;
2286                                 kill(td->pid, SIGTERM);
2287                         }
2288                         break;
2289                 }
2290
2291                 /*
2292                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2293                  */
2294                 for_each_td(td, i) {
2295                         if (td->runstate != TD_INITIALIZED)
2296                                 continue;
2297
2298                         if (in_ramp_time(td))
2299                                 td_set_runstate(td, TD_RAMP);
2300                         else
2301                                 td_set_runstate(td, TD_RUNNING);
2302                         nr_running++;
2303                         nr_started--;
2304                         m_rate += ddir_rw_sum(td->o.ratemin);
2305                         t_rate += ddir_rw_sum(td->o.rate);
2306                         todo--;
2307                         fio_mutex_up(td->mutex);
2308                 }
2309
2310                 reap_threads(&nr_running, &t_rate, &m_rate);
2311
2312                 if (todo)
2313                         do_usleep(100000);
2314         }
2315
2316         while (nr_running) {
2317                 reap_threads(&nr_running, &t_rate, &m_rate);
2318                 do_usleep(10000);
2319         }
2320
2321         fio_idle_prof_stop();
2322
2323         update_io_ticks();
2324 }
2325
2326 void helper_reset(void)
2327 {
2328         if (!helper_data)
2329                 return;
2330
2331         pthread_mutex_lock(&helper_data->lock);
2332
2333         if (!helper_data->reset) {
2334                 helper_data->reset = 1;
2335                 pthread_cond_signal(&helper_data->cond);
2336         }
2337
2338         pthread_mutex_unlock(&helper_data->lock);
2339 }
2340
2341 void helper_do_stat(void)
2342 {
2343         if (!helper_data)
2344                 return;
2345
2346         pthread_mutex_lock(&helper_data->lock);
2347         helper_data->do_stat = 1;
2348         pthread_cond_signal(&helper_data->cond);
2349         pthread_mutex_unlock(&helper_data->lock);
2350 }
2351
2352 bool helper_should_exit(void)
2353 {
2354         if (!helper_data)
2355                 return true;
2356
2357         return helper_data->exit;
2358 }
2359
2360 static void wait_for_helper_thread_exit(void)
2361 {
2362         void *ret;
2363
2364         pthread_mutex_lock(&helper_data->lock);
2365         helper_data->exit = 1;
2366         pthread_cond_signal(&helper_data->cond);
2367         pthread_mutex_unlock(&helper_data->lock);
2368
2369         pthread_join(helper_data->thread, &ret);
2370 }
2371
2372 static void free_disk_util(void)
2373 {
2374         disk_util_prune_entries();
2375
2376         pthread_cond_destroy(&helper_data->cond);
2377         pthread_mutex_destroy(&helper_data->lock);
2378         sfree(helper_data);
2379 }
2380
2381 static void *helper_thread_main(void *data)
2382 {
2383         struct helper_data *hd = data;
2384         unsigned int msec_to_next_event, next_log;
2385         struct timeval tv, last_du;
2386         int ret = 0;
2387
2388         sk_out_assign(hd->sk_out);
2389
2390         gettimeofday(&tv, NULL);
2391         memcpy(&last_du, &tv, sizeof(tv));
2392
2393         fio_mutex_up(startup_mutex);
2394
2395         msec_to_next_event = DISK_UTIL_MSEC;
2396         while (!ret && !hd->exit) {
2397                 struct timespec ts;
2398                 struct timeval now;
2399                 uint64_t since_du;
2400
2401                 timeval_add_msec(&tv, msec_to_next_event);
2402                 ts.tv_sec = tv.tv_sec;
2403                 ts.tv_nsec = tv.tv_usec * 1000;
2404
2405                 pthread_mutex_lock(&hd->lock);
2406                 pthread_cond_timedwait(&hd->cond, &hd->lock, &ts);
2407
2408                 gettimeofday(&now, NULL);
2409
2410                 if (hd->reset) {
2411                         memcpy(&tv, &now, sizeof(tv));
2412                         memcpy(&last_du, &now, sizeof(last_du));
2413                         hd->reset = 0;
2414                 }
2415
2416                 pthread_mutex_unlock(&hd->lock);
2417
2418                 since_du = mtime_since(&last_du, &now);
2419                 if (since_du >= DISK_UTIL_MSEC || DISK_UTIL_MSEC - since_du < 10) {
2420                         ret = update_io_ticks();
2421                         timeval_add_msec(&last_du, DISK_UTIL_MSEC);
2422                         msec_to_next_event = DISK_UTIL_MSEC;
2423                         if (since_du >= DISK_UTIL_MSEC)
2424                                 msec_to_next_event -= (since_du - DISK_UTIL_MSEC);
2425                 } else {
2426                         if (since_du >= DISK_UTIL_MSEC)
2427                                 msec_to_next_event = DISK_UTIL_MSEC - (DISK_UTIL_MSEC - since_du);
2428                         else
2429                                 msec_to_next_event = DISK_UTIL_MSEC;
2430                 }
2431
2432                 if (hd->do_stat) {
2433                         hd->do_stat = 0;
2434                         __show_running_run_stats();
2435                 }
2436
2437                 next_log = calc_log_samples();
2438                 if (!next_log)
2439                         next_log = DISK_UTIL_MSEC;
2440
2441                 msec_to_next_event = min(next_log, msec_to_next_event);
2442
2443                 if (!is_backend)
2444                         print_thread_status();
2445         }
2446
2447         fio_writeout_logs(false);
2448
2449         sk_out_drop();
2450         return NULL;
2451 }
2452
2453 static int create_helper_thread(struct sk_out *sk_out)
2454 {
2455         struct helper_data *hd;
2456         int ret;
2457
2458         hd = smalloc(sizeof(*hd));
2459
2460         setup_disk_util();
2461
2462         hd->sk_out = sk_out;
2463         pthread_cond_init(&hd->cond, NULL);
2464         pthread_mutex_init(&hd->lock, NULL);
2465
2466         ret = pthread_create(&hd->thread, NULL, helper_thread_main, hd);
2467         if (ret) {
2468                 log_err("Can't create helper thread: %s\n", strerror(ret));
2469                 return 1;
2470         }
2471
2472         helper_data = hd;
2473
2474         dprint(FD_MUTEX, "wait on startup_mutex\n");
2475         fio_mutex_down(startup_mutex);
2476         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2477         return 0;
2478 }
2479
2480 int fio_backend(struct sk_out *sk_out)
2481 {
2482         struct thread_data *td;
2483         int i;
2484
2485         if (exec_profile) {
2486                 if (load_profile(exec_profile))
2487                         return 1;
2488                 free(exec_profile);
2489                 exec_profile = NULL;
2490         }
2491         if (!thread_number)
2492                 return 0;
2493
2494         if (write_bw_log) {
2495                 struct log_params p = {
2496                         .log_type = IO_LOG_TYPE_BW,
2497                 };
2498
2499                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2500                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2501                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2502         }
2503
2504         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2505         if (startup_mutex == NULL)
2506                 return 1;
2507
2508         set_genesis_time();
2509         stat_init();
2510         create_helper_thread(sk_out);
2511
2512         cgroup_list = smalloc(sizeof(*cgroup_list));
2513         INIT_FLIST_HEAD(cgroup_list);
2514
2515         run_threads(sk_out);
2516
2517         wait_for_helper_thread_exit();
2518
2519         if (!fio_abort) {
2520                 __show_run_stats();
2521                 if (write_bw_log) {
2522                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2523                                 struct io_log *log = agg_io_log[i];
2524
2525                                 flush_log(log, 0);
2526                                 free_log(log);
2527                         }
2528                 }
2529         }
2530
2531         for_each_td(td, i) {
2532                 fio_options_free(td);
2533                 if (td->rusage_sem) {
2534                         fio_mutex_remove(td->rusage_sem);
2535                         td->rusage_sem = NULL;
2536                 }
2537         }
2538
2539         free_disk_util();
2540         cgroup_kill(cgroup_list);
2541         sfree(cgroup_list);
2542         sfree(cgroup_mnt);
2543
2544         fio_mutex_remove(startup_mutex);
2545         stat_exit();
2546         return exit_value;
2547 }