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 pthread_t helper_thread;
  62 static pthread_mutex_t helper_lock;
  63 pthread_cond_t helper_cond;
  64 int helper_do_stat = 0;
  65
  66 static struct fio_mutex *startup_mutex;
  67 static struct flist_head *cgroup_list;
  68 static char *cgroup_mnt;
  69 static int exit_value;
  70 static volatile int fio_abort;
  71 static unsigned int nr_process = 0;
  72 static unsigned int nr_thread = 0;
  73
  74 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  75
  76 int groupid = 0;
  77 unsigned int thread_number = 0;
  78 unsigned int stat_number = 0;
  79 int shm_id = 0;
  80 int temp_stall_ts;
  81 unsigned long done_secs = 0;
  82 volatile int helper_exit = 0;
  83
  84 #define PAGE_ALIGN(buf) \
  85         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  86
  87 #define JOB_START_TIMEOUT       (5 * 1000)
  88
  89 static void sig_int(int sig)
  90 {
  91         if (threads) {
  92                 if (is_backend)
  93                         fio_server_got_signal(sig);
  94                 else {
  95                         log_info("\nfio: terminating on signal %d\n", sig);
  96                         log_info_flush();
  97                         exit_value = 128;
  98                 }
  99
 100                 fio_terminate_threads(TERMINATE_ALL);
 101         }
 102 }
 103
 104 void sig_show_status(int sig)
 105 {
 106         show_running_run_stats();
 107 }
 108
 109 static void set_sig_handlers(void)
 110 {
 111         struct sigaction act;
 112
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGINT, &act, NULL);
 117
 118         memset(&act, 0, sizeof(act));
 119         act.sa_handler = sig_int;
 120         act.sa_flags = SA_RESTART;
 121         sigaction(SIGTERM, &act, NULL);
 122
 123 /* Windows uses SIGBREAK as a quit signal from other applications */
 124 #ifdef WIN32
 125         memset(&act, 0, sizeof(act));
 126         act.sa_handler = sig_int;
 127         act.sa_flags = SA_RESTART;
 128         sigaction(SIGBREAK, &act, NULL);
 129 #endif
 130
 131         memset(&act, 0, sizeof(act));
 132         act.sa_handler = sig_show_status;
 133         act.sa_flags = SA_RESTART;
 134         sigaction(SIGUSR1, &act, NULL);
 135
 136         if (is_backend) {
 137                 memset(&act, 0, sizeof(act));
 138                 act.sa_handler = sig_int;
 139                 act.sa_flags = SA_RESTART;
 140                 sigaction(SIGPIPE, &act, NULL);
 141         }
 142 }
 143
 144 /*
 145  * Check if we are above the minimum rate given.
 146  */
 147 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 148                              enum fio_ddir ddir)
 149 {
 150         unsigned long long bytes = 0;
 151         unsigned long iops = 0;
 152         unsigned long spent;
 153         unsigned long rate;
 154         unsigned int ratemin = 0;
 155         unsigned int rate_iops = 0;
 156         unsigned int rate_iops_min = 0;
 157
 158         assert(ddir_rw(ddir));
 159
 160         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 161                 return false;
 162
 163         /*
 164          * allow a 2 second settle period in the beginning
 165          */
 166         if (mtime_since(&td->start, now) < 2000)
 167                 return false;
 168
 169         iops += td->this_io_blocks[ddir];
 170         bytes += td->this_io_bytes[ddir];
 171         ratemin += td->o.ratemin[ddir];
 172         rate_iops += td->o.rate_iops[ddir];
 173         rate_iops_min += td->o.rate_iops_min[ddir];
 174
 175         /*
 176          * if rate blocks is set, sample is running
 177          */
 178         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 179                 spent = mtime_since(&td->lastrate[ddir], now);
 180                 if (spent < td->o.ratecycle)
 181                         return false;
 182
 183                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 184                         /*
 185                          * check bandwidth specified rate
 186                          */
 187                         if (bytes < td->rate_bytes[ddir]) {
 188                                 log_err("%s: min rate %u not met\n", td->o.name,
 189                                                                 ratemin);
 190                                 return true;
 191                         } else {
 192                                 if (spent)
 193                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 194                                 else
 195                                         rate = 0;
 196
 197                                 if (rate < ratemin ||
 198                                     bytes < td->rate_bytes[ddir]) {
 199                                         log_err("%s: min rate %u not met, got"
 200                                                 " %luKB/sec\n", td->o.name,
 201                                                         ratemin, rate);
 202                                         return true;
 203                                 }
 204                         }
 205                 } else {
 206                         /*
 207                          * checks iops specified rate
 208                          */
 209                         if (iops < rate_iops) {
 210                                 log_err("%s: min iops rate %u not met\n",
 211                                                 td->o.name, rate_iops);
 212                                 return true;
 213                         } else {
 214                                 if (spent)
 215                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 216                                 else
 217                                         rate = 0;
 218
 219                                 if (rate < rate_iops_min ||
 220                                     iops < td->rate_blocks[ddir]) {
 221                                         log_err("%s: min iops rate %u not met,"
 222                                                 " got %lu\n", td->o.name,
 223                                                         rate_iops_min, rate);
 224                                         return true;
 225                                 }
 226                         }
 227                 }
 228         }
 229
 230         td->rate_bytes[ddir] = bytes;
 231         td->rate_blocks[ddir] = iops;
 232         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 233         return false;
 234 }
 235
 236 static bool check_min_rate(struct thread_data *td, struct timeval *now)
 237 {
 238         bool ret = false;
 239
 240         if (td->bytes_done[DDIR_READ])
 241                 ret |= __check_min_rate(td, now, DDIR_READ);
 242         if (td->bytes_done[DDIR_WRITE])
 243                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 244         if (td->bytes_done[DDIR_TRIM])
 245                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 246
 247         return ret;
 248 }
 249
 250 /*
 251  * When job exits, we can cancel the in-flight IO if we are using async
 252  * io. Attempt to do so.
 253  */
 254 static void cleanup_pending_aio(struct thread_data *td)
 255 {
 256         int r;
 257
 258         /*
 259          * get immediately available events, if any
 260          */
 261         r = io_u_queued_complete(td, 0);
 262         if (r < 0)
 263                 return;
 264
 265         /*
 266          * now cancel remaining active events
 267          */
 268         if (td->io_ops->cancel) {
 269                 struct io_u *io_u;
 270                 int i;
 271
 272                 io_u_qiter(&td->io_u_all, io_u, i) {
 273                         if (io_u->flags & IO_U_F_FLIGHT) {
 274                                 r = td->io_ops->cancel(td, io_u);
 275                                 if (!r)
 276                                         put_io_u(td, io_u);
 277                         }
 278                 }
 279         }
 280
 281         if (td->cur_depth)
 282                 r = io_u_queued_complete(td, td->cur_depth);
 283 }
 284
 285 /*
 286  * Helper to handle the final sync of a file. Works just like the normal
 287  * io path, just does everything sync.
 288  */
 289 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 290 {
 291         struct io_u *io_u = __get_io_u(td);
 292         int ret;
 293
 294         if (!io_u)
 295                 return true;
 296
 297         io_u->ddir = DDIR_SYNC;
 298         io_u->file = f;
 299
 300         if (td_io_prep(td, io_u)) {
 301                 put_io_u(td, io_u);
 302                 return true;
 303         }
 304
 305 requeue:
 306         ret = td_io_queue(td, io_u);
 307         if (ret < 0) {
 308                 td_verror(td, io_u->error, "td_io_queue");
 309                 put_io_u(td, io_u);
 310                 return true;
 311         } else if (ret == FIO_Q_QUEUED) {
 312                 if (io_u_queued_complete(td, 1) < 0)
 313                         return true;
 314         } else if (ret == FIO_Q_COMPLETED) {
 315                 if (io_u->error) {
 316                         td_verror(td, io_u->error, "td_io_queue");
 317                         return true;
 318                 }
 319
 320                 if (io_u_sync_complete(td, io_u) < 0)
 321                         return true;
 322         } else if (ret == FIO_Q_BUSY) {
 323                 if (td_io_commit(td))
 324                         return true;
 325                 goto requeue;
 326         }
 327
 328         return false;
 329 }
 330
 331 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 332 {
 333         int ret;
 334
 335         if (fio_file_open(f))
 336                 return fio_io_sync(td, f);
 337
 338         if (td_io_open_file(td, f))
 339                 return 1;
 340
 341         ret = fio_io_sync(td, f);
 342         td_io_close_file(td, f);
 343         return ret;
 344 }
 345
 346 static inline void __update_tv_cache(struct thread_data *td)
 347 {
 348         fio_gettime(&td->tv_cache, NULL);
 349 }
 350
 351 static inline void update_tv_cache(struct thread_data *td)
 352 {
 353         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 354                 __update_tv_cache(td);
 355 }
 356
 357 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 358 {
 359         if (in_ramp_time(td))
 360                 return false;
 361         if (!td->o.timeout)
 362                 return false;
 363         if (utime_since(&td->epoch, t) >= td->o.timeout)
 364                 return true;
 365
 366         return false;
 367 }
 368
 369 /*
 370  * We need to update the runtime consistently in ms, but keep a running
 371  * tally of the current elapsed time in microseconds for sub millisecond
 372  * updates.
 373  */
 374 static inline void update_runtime(struct thread_data *td,
 375                                   unsigned long long *elapsed_us,
 376                                   const enum fio_ddir ddir)
 377 {
 378         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 379                 return;
 380
 381         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 382         elapsed_us[ddir] += utime_since_now(&td->start);
 383         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 384 }
 385
 386 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 387                                 int *retptr)
 388 {
 389         int ret = *retptr;
 390
 391         if (ret < 0 || td->error) {
 392                 int err = td->error;
 393                 enum error_type_bit eb;
 394
 395                 if (ret < 0)
 396                         err = -ret;
 397
 398                 eb = td_error_type(ddir, err);
 399                 if (!(td->o.continue_on_error & (1 << eb)))
 400                         return true;
 401
 402                 if (td_non_fatal_error(td, eb, err)) {
 403                         /*
 404                          * Continue with the I/Os in case of
 405                          * a non fatal error.
 406                          */
 407                         update_error_count(td, err);
 408                         td_clear_error(td);
 409                         *retptr = 0;
 410                         return false;
 411                 } else if (td->o.fill_device && err == ENOSPC) {
 412                         /*
 413                          * We expect to hit this error if
 414                          * fill_device option is set.
 415                          */
 416                         td_clear_error(td);
 417                         fio_mark_td_terminate(td);
 418                         return true;
 419                 } else {
 420                         /*
 421                          * Stop the I/O in case of a fatal
 422                          * error.
 423                          */
 424                         update_error_count(td, err);
 425                         return true;
 426                 }
 427         }
 428
 429         return false;
 430 }
 431
 432 static void check_update_rusage(struct thread_data *td)
 433 {
 434         if (td->update_rusage) {
 435                 td->update_rusage = 0;
 436                 update_rusage_stat(td);
 437                 fio_mutex_up(td->rusage_sem);
 438         }
 439 }
 440
 441 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 442 {
 443         const int full = queue_full(td);
 444         int min_evts = 0;
 445         int ret;
 446
 447         /*
 448          * if the queue is full, we MUST reap at least 1 event
 449          */
 450         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 451         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 452                 min_evts = 1;
 453
 454         if (time && (__should_check_rate(td, DDIR_READ) ||
 455             __should_check_rate(td, DDIR_WRITE) ||
 456             __should_check_rate(td, DDIR_TRIM)))
 457                 fio_gettime(time, NULL);
 458
 459         do {
 460                 ret = io_u_queued_complete(td, min_evts);
 461                 if (ret < 0)
 462                         break;
 463         } while (full && (td->cur_depth > td->o.iodepth_low));
 464
 465         return ret;
 466 }
 467
 468 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 469                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 470                    struct timeval *comp_time)
 471 {
 472         int ret2;
 473
 474         switch (*ret) {
 475         case FIO_Q_COMPLETED:
 476                 if (io_u->error) {
 477                         *ret = -io_u->error;
 478                         clear_io_u(td, io_u);
 479                 } else if (io_u->resid) {
 480                         int bytes = io_u->xfer_buflen - io_u->resid;
 481                         struct fio_file *f = io_u->file;
 482
 483                         if (bytes_issued)
 484                                 *bytes_issued += bytes;
 485
 486                         if (!from_verify)
 487                                 trim_io_piece(td, io_u);
 488
 489                         /*
 490                          * zero read, fail
 491                          */
 492                         if (!bytes) {
 493                                 if (!from_verify)
 494                                         unlog_io_piece(td, io_u);
 495                                 td_verror(td, EIO, "full resid");
 496                                 put_io_u(td, io_u);
 497                                 break;
 498                         }
 499
 500                         io_u->xfer_buflen = io_u->resid;
 501                         io_u->xfer_buf += bytes;
 502                         io_u->offset += bytes;
 503
 504                         if (ddir_rw(io_u->ddir))
 505                                 td->ts.short_io_u[io_u->ddir]++;
 506
 507                         f = io_u->file;
 508                         if (io_u->offset == f->real_file_size)
 509                                 goto sync_done;
 510
 511                         requeue_io_u(td, &io_u);
 512                 } else {
 513 sync_done:
 514                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 515                             __should_check_rate(td, DDIR_WRITE) ||
 516                             __should_check_rate(td, DDIR_TRIM)))
 517                                 fio_gettime(comp_time, NULL);
 518
 519                         *ret = io_u_sync_complete(td, io_u);
 520                         if (*ret < 0)
 521                                 break;
 522                 }
 523                 return 0;
 524         case FIO_Q_QUEUED:
 525                 /*
 526                  * if the engine doesn't have a commit hook,
 527                  * the io_u is really queued. if it does have such
 528                  * a hook, it has to call io_u_queued() itself.
 529                  */
 530                 if (td->io_ops->commit == NULL)
 531                         io_u_queued(td, io_u);
 532                 if (bytes_issued)
 533                         *bytes_issued += io_u->xfer_buflen;
 534                 break;
 535         case FIO_Q_BUSY:
 536                 if (!from_verify)
 537                         unlog_io_piece(td, io_u);
 538                 requeue_io_u(td, &io_u);
 539                 ret2 = td_io_commit(td);
 540                 if (ret2 < 0)
 541                         *ret = ret2;
 542                 break;
 543         default:
 544                 assert(*ret < 0);
 545                 td_verror(td, -(*ret), "td_io_queue");
 546                 break;
 547         }
 548
 549         if (break_on_this_error(td, ddir, ret))
 550                 return 1;
 551
 552         return 0;
 553 }
 554
 555 static inline bool io_in_polling(struct thread_data *td)
 556 {
 557         return !td->o.iodepth_batch_complete_min &&
 558                    !td->o.iodepth_batch_complete_max;
 559 }
 560
 561 /*
 562  * The main verify engine. Runs over the writes we previously submitted,
 563  * reads the blocks back in, and checks the crc/md5 of the data.
 564  */
 565 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 566 {
 567         struct fio_file *f;
 568         struct io_u *io_u;
 569         int ret, min_events;
 570         unsigned int i;
 571
 572         dprint(FD_VERIFY, "starting loop\n");
 573
 574         /*
 575          * sync io first and invalidate cache, to make sure we really
 576          * read from disk.
 577          */
 578         for_each_file(td, f, i) {
 579                 if (!fio_file_open(f))
 580                         continue;
 581                 if (fio_io_sync(td, f))
 582                         break;
 583                 if (file_invalidate_cache(td, f))
 584                         break;
 585         }
 586
 587         check_update_rusage(td);
 588
 589         if (td->error)
 590                 return;
 591
 592         td_set_runstate(td, TD_VERIFYING);
 593
 594         io_u = NULL;
 595         while (!td->terminate) {
 596                 enum fio_ddir ddir;
 597                 int full;
 598
 599                 update_tv_cache(td);
 600                 check_update_rusage(td);
 601
 602                 if (runtime_exceeded(td, &td->tv_cache)) {
 603                         __update_tv_cache(td);
 604                         if (runtime_exceeded(td, &td->tv_cache)) {
 605                                 fio_mark_td_terminate(td);
 606                                 break;
 607                         }
 608                 }
 609
 610                 if (flow_threshold_exceeded(td))
 611                         continue;
 612
 613                 if (!td->o.experimental_verify) {
 614                         io_u = __get_io_u(td);
 615                         if (!io_u)
 616                                 break;
 617
 618                         if (get_next_verify(td, io_u)) {
 619                                 put_io_u(td, io_u);
 620                                 break;
 621                         }
 622
 623                         if (td_io_prep(td, io_u)) {
 624                                 put_io_u(td, io_u);
 625                                 break;
 626                         }
 627                 } else {
 628                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 629                                 break;
 630
 631                         while ((io_u = get_io_u(td)) != NULL) {
 632                                 if (IS_ERR(io_u)) {
 633                                         io_u = NULL;
 634                                         ret = FIO_Q_BUSY;
 635                                         goto reap;
 636                                 }
 637
 638                                 /*
 639                                  * We are only interested in the places where
 640                                  * we wrote or trimmed IOs. Turn those into
 641                                  * reads for verification purposes.
 642                                  */
 643                                 if (io_u->ddir == DDIR_READ) {
 644                                         /*
 645                                          * Pretend we issued it for rwmix
 646                                          * accounting
 647                                          */
 648                                         td->io_issues[DDIR_READ]++;
 649                                         put_io_u(td, io_u);
 650                                         continue;
 651                                 } else if (io_u->ddir == DDIR_TRIM) {
 652                                         io_u->ddir = DDIR_READ;
 653                                         io_u_set(io_u, IO_U_F_TRIMMED);
 654                                         break;
 655                                 } else if (io_u->ddir == DDIR_WRITE) {
 656                                         io_u->ddir = DDIR_READ;
 657                                         break;
 658                                 } else {
 659                                         put_io_u(td, io_u);
 660                                         continue;
 661                                 }
 662                         }
 663
 664                         if (!io_u)
 665                                 break;
 666                 }
 667
 668                 if (verify_state_should_stop(td, io_u)) {
 669                         put_io_u(td, io_u);
 670                         break;
 671                 }
 672
 673                 if (td->o.verify_async)
 674                         io_u->end_io = verify_io_u_async;
 675                 else
 676                         io_u->end_io = verify_io_u;
 677
 678                 ddir = io_u->ddir;
 679                 if (!td->o.disable_slat)
 680                         fio_gettime(&io_u->start_time, NULL);
 681
 682                 ret = td_io_queue(td, io_u);
 683
 684                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 685                         break;
 686
 687                 /*
 688                  * if we can queue more, do so. but check if there are
 689                  * completed io_u's first. Note that we can get BUSY even
 690                  * without IO queued, if the system is resource starved.
 691                  */
 692 reap:
 693                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 694                 if (full || io_in_polling(td))
 695                         ret = wait_for_completions(td, NULL);
 696
 697                 if (ret < 0)
 698                         break;
 699         }
 700
 701         check_update_rusage(td);
 702
 703         if (!td->error) {
 704                 min_events = td->cur_depth;
 705
 706                 if (min_events)
 707                         ret = io_u_queued_complete(td, min_events);
 708         } else
 709                 cleanup_pending_aio(td);
 710
 711         td_set_runstate(td, TD_RUNNING);
 712
 713         dprint(FD_VERIFY, "exiting loop\n");
 714 }
 715
 716 static bool exceeds_number_ios(struct thread_data *td)
 717 {
 718         unsigned long long number_ios;
 719
 720         if (!td->o.number_ios)
 721                 return false;
 722
 723         number_ios = ddir_rw_sum(td->io_blocks);
 724         number_ios += td->io_u_queued + td->io_u_in_flight;
 725
 726         return number_ios >= (td->o.number_ios * td->loops);
 727 }
 728
 729 static bool io_issue_bytes_exceeded(struct thread_data *td)
 730 {
 731         unsigned long long bytes, limit;
 732
 733         if (td_rw(td))
 734                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 735         else if (td_write(td))
 736                 bytes = td->io_issue_bytes[DDIR_WRITE];
 737         else if (td_read(td))
 738                 bytes = td->io_issue_bytes[DDIR_READ];
 739         else
 740                 bytes = td->io_issue_bytes[DDIR_TRIM];
 741
 742         if (td->o.io_limit)
 743                 limit = td->o.io_limit;
 744         else
 745                 limit = td->o.size;
 746
 747         limit *= td->loops;
 748         return bytes >= limit || exceeds_number_ios(td);
 749 }
 750
 751 static bool io_complete_bytes_exceeded(struct thread_data *td)
 752 {
 753         unsigned long long bytes, limit;
 754
 755         if (td_rw(td))
 756                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 757         else if (td_write(td))
 758                 bytes = td->this_io_bytes[DDIR_WRITE];
 759         else if (td_read(td))
 760                 bytes = td->this_io_bytes[DDIR_READ];
 761         else
 762                 bytes = td->this_io_bytes[DDIR_TRIM];
 763
 764         if (td->o.io_limit)
 765                 limit = td->o.io_limit;
 766         else
 767                 limit = td->o.size;
 768
 769         limit *= td->loops;
 770         return bytes >= limit || exceeds_number_ios(td);
 771 }
 772
 773 /*
 774  * used to calculate the next io time for rate control
 775  *
 776  */
 777 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 778 {
 779         uint64_t secs, remainder, bps, bytes, iops;
 780
 781         assert(!(td->flags & TD_F_CHILD));
 782         bytes = td->rate_io_issue_bytes[ddir];
 783         bps = td->rate_bps[ddir];
 784
 785         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 786                 uint64_t val;
 787                 iops = bps / td->o.bs[ddir];
 788                 val = (int64_t) (1000000 / iops) *
 789                                 -logf(__rand_0_1(&td->poisson_state));
 790                 if (val) {
 791                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 792                                         (unsigned long long) 1000000 / val);
 793                 }
 794                 td->last_usec += val;
 795                 return td->last_usec;
 796         } else if (bps) {
 797                 secs = bytes / bps;
 798                 remainder = bytes % bps;
 799                 return remainder * 1000000 / bps + secs * 1000000;
 800         }
 801
 802         return 0;
 803 }
 804
 805 /*
 806  * Main IO worker function. It retrieves io_u's to process and queues
 807  * and reaps them, checking for rate and errors along the way.
 808  *
 809  * Returns number of bytes written and trimmed.
 810  */
 811 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 812 {
 813         unsigned int i;
 814         int ret = 0;
 815         uint64_t total_bytes, bytes_issued = 0;
 816
 817         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 818                 bytes_done[i] = td->bytes_done[i];
 819
 820         if (in_ramp_time(td))
 821                 td_set_runstate(td, TD_RAMP);
 822         else
 823                 td_set_runstate(td, TD_RUNNING);
 824
 825         lat_target_init(td);
 826
 827         total_bytes = td->o.size;
 828         /*
 829         * Allow random overwrite workloads to write up to io_limit
 830         * before starting verification phase as 'size' doesn't apply.
 831         */
 832         if (td_write(td) && td_random(td) && td->o.norandommap)
 833                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 834         /*
 835          * If verify_backlog is enabled, we'll run the verify in this
 836          * handler as well. For that case, we may need up to twice the
 837          * amount of bytes.
 838          */
 839         if (td->o.verify != VERIFY_NONE &&
 840            (td_write(td) && td->o.verify_backlog))
 841                 total_bytes += td->o.size;
 842
 843         /* In trimwrite mode, each byte is trimmed and then written, so
 844          * allow total_bytes to be twice as big */
 845         if (td_trimwrite(td))
 846                 total_bytes += td->total_io_size;
 847
 848         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 849                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 850                 td->o.time_based) {
 851                 struct timeval comp_time;
 852                 struct io_u *io_u;
 853                 int full;
 854                 enum fio_ddir ddir;
 855
 856                 check_update_rusage(td);
 857
 858                 if (td->terminate || td->done)
 859                         break;
 860
 861                 update_tv_cache(td);
 862
 863                 if (runtime_exceeded(td, &td->tv_cache)) {
 864                         __update_tv_cache(td);
 865                         if (runtime_exceeded(td, &td->tv_cache)) {
 866                                 fio_mark_td_terminate(td);
 867                                 break;
 868                         }
 869                 }
 870
 871                 if (flow_threshold_exceeded(td))
 872                         continue;
 873
 874                 if (!td->o.time_based && bytes_issued >= total_bytes)
 875                         break;
 876
 877                 io_u = get_io_u(td);
 878                 if (IS_ERR_OR_NULL(io_u)) {
 879                         int err = PTR_ERR(io_u);
 880
 881                         io_u = NULL;
 882                         if (err == -EBUSY) {
 883                                 ret = FIO_Q_BUSY;
 884                                 goto reap;
 885                         }
 886                         if (td->o.latency_target)
 887                                 goto reap;
 888                         break;
 889                 }
 890
 891                 ddir = io_u->ddir;
 892
 893                 /*
 894                  * Add verification end_io handler if:
 895                  *      - Asked to verify (!td_rw(td))
 896                  *      - Or the io_u is from our verify list (mixed write/ver)
 897                  */
 898                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 899                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 900
 901                         if (!td->o.verify_pattern_bytes) {
 902                                 io_u->rand_seed = __rand(&td->verify_state);
 903                                 if (sizeof(int) != sizeof(long *))
 904                                         io_u->rand_seed *= __rand(&td->verify_state);
 905                         }
 906
 907                         if (verify_state_should_stop(td, io_u)) {
 908                                 put_io_u(td, io_u);
 909                                 break;
 910                         }
 911
 912                         if (td->o.verify_async)
 913                                 io_u->end_io = verify_io_u_async;
 914                         else
 915                                 io_u->end_io = verify_io_u;
 916                         td_set_runstate(td, TD_VERIFYING);
 917                 } else if (in_ramp_time(td))
 918                         td_set_runstate(td, TD_RAMP);
 919                 else
 920                         td_set_runstate(td, TD_RUNNING);
 921
 922                 /*
 923                  * Always log IO before it's issued, so we know the specific
 924                  * order of it. The logged unit will track when the IO has
 925                  * completed.
 926                  */
 927                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 928                     td->o.do_verify &&
 929                     td->o.verify != VERIFY_NONE &&
 930                     !td->o.experimental_verify)
 931                         log_io_piece(td, io_u);
 932
 933                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 934                         const unsigned long blen = io_u->xfer_buflen;
 935                         const enum fio_ddir ddir = acct_ddir(io_u);
 936
 937                         if (td->error)
 938                                 break;
 939
 940                         workqueue_enqueue(&td->io_wq, &io_u->work);
 941                         ret = FIO_Q_QUEUED;
 942
 943                         if (ddir_rw(ddir)) {
 944                                 td->io_issues[ddir]++;
 945                                 td->io_issue_bytes[ddir] += blen;
 946                                 td->rate_io_issue_bytes[ddir] += blen;
 947                         }
 948
 949                         if (should_check_rate(td))
 950                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 951
 952                 } else {
 953                         ret = td_io_queue(td, io_u);
 954
 955                         if (should_check_rate(td))
 956                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 957
 958                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 959                                 break;
 960
 961                         /*
 962                          * See if we need to complete some commands. Note that
 963                          * we can get BUSY even without IO queued, if the
 964                          * system is resource starved.
 965                          */
 966 reap:
 967                         full = queue_full(td) ||
 968                                 (ret == FIO_Q_BUSY && td->cur_depth);
 969                         if (full || io_in_polling(td))
 970                                 ret = wait_for_completions(td, &comp_time);
 971                 }
 972                 if (ret < 0)
 973                         break;
 974                 if (!ddir_rw_sum(td->bytes_done) &&
 975                     !(td->io_ops->flags & FIO_NOIO))
 976                         continue;
 977
 978                 if (!in_ramp_time(td) && should_check_rate(td)) {
 979                         if (check_min_rate(td, &comp_time)) {
 980                                 if (exitall_on_terminate || td->o.exitall_error)
 981                                         fio_terminate_threads(td->groupid);
 982                                 td_verror(td, EIO, "check_min_rate");
 983                                 break;
 984                         }
 985                 }
 986                 if (!in_ramp_time(td) && td->o.latency_target)
 987                         lat_target_check(td);
 988
 989                 if (td->o.thinktime) {
 990                         unsigned long long b;
 991
 992                         b = ddir_rw_sum(td->io_blocks);
 993                         if (!(b % td->o.thinktime_blocks)) {
 994                                 int left;
 995
 996                                 io_u_quiesce(td);
 997
 998                                 if (td->o.thinktime_spin)
 999                                         usec_spin(td->o.thinktime_spin);
1000
1001                                 left = td->o.thinktime - td->o.thinktime_spin;
1002                                 if (left)
1003                                         usec_sleep(td, left);
1004                         }
1005                 }
1006         }
1007
1008         check_update_rusage(td);
1009
1010         if (td->trim_entries)
1011                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1012
1013         if (td->o.fill_device && td->error == ENOSPC) {
1014                 td->error = 0;
1015                 fio_mark_td_terminate(td);
1016         }
1017         if (!td->error) {
1018                 struct fio_file *f;
1019
1020                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1021                         workqueue_flush(&td->io_wq);
1022                         i = 0;
1023                 } else
1024                         i = td->cur_depth;
1025
1026                 if (i) {
1027                         ret = io_u_queued_complete(td, i);
1028                         if (td->o.fill_device && td->error == ENOSPC)
1029                                 td->error = 0;
1030                 }
1031
1032                 if (should_fsync(td) && td->o.end_fsync) {
1033                         td_set_runstate(td, TD_FSYNCING);
1034
1035                         for_each_file(td, f, i) {
1036                                 if (!fio_file_fsync(td, f))
1037                                         continue;
1038
1039                                 log_err("fio: end_fsync failed for file %s\n",
1040                                                                 f->file_name);
1041                         }
1042                 }
1043         } else
1044                 cleanup_pending_aio(td);
1045
1046         /*
1047          * stop job if we failed doing any IO
1048          */
1049         if (!ddir_rw_sum(td->this_io_bytes))
1050                 td->done = 1;
1051
1052         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1053                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1054 }
1055
1056 static void cleanup_io_u(struct thread_data *td)
1057 {
1058         struct io_u *io_u;
1059
1060         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1061
1062                 if (td->io_ops->io_u_free)
1063                         td->io_ops->io_u_free(td, io_u);
1064
1065                 fio_memfree(io_u, sizeof(*io_u));
1066         }
1067
1068         free_io_mem(td);
1069
1070         io_u_rexit(&td->io_u_requeues);
1071         io_u_qexit(&td->io_u_freelist);
1072         io_u_qexit(&td->io_u_all);
1073
1074         if (td->last_write_comp)
1075                 sfree(td->last_write_comp);
1076 }
1077
1078 static int init_io_u(struct thread_data *td)
1079 {
1080         struct io_u *io_u;
1081         unsigned int max_bs, min_write;
1082         int cl_align, i, max_units;
1083         int data_xfer = 1, err;
1084         char *p;
1085
1086         max_units = td->o.iodepth;
1087         max_bs = td_max_bs(td);
1088         min_write = td->o.min_bs[DDIR_WRITE];
1089         td->orig_buffer_size = (unsigned long long) max_bs
1090                                         * (unsigned long long) max_units;
1091
1092         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1093                 data_xfer = 0;
1094
1095         err = 0;
1096         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1097         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1098         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1099
1100         if (err) {
1101                 log_err("fio: failed setting up IO queues\n");
1102                 return 1;
1103         }
1104
1105         /*
1106          * if we may later need to do address alignment, then add any
1107          * possible adjustment here so that we don't cause a buffer
1108          * overflow later. this adjustment may be too much if we get
1109          * lucky and the allocator gives us an aligned address.
1110          */
1111         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1112             (td->io_ops->flags & FIO_RAWIO))
1113                 td->orig_buffer_size += page_mask + td->o.mem_align;
1114
1115         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1116                 unsigned long bs;
1117
1118                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1119                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1120         }
1121
1122         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1123                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1124                 return 1;
1125         }
1126
1127         if (data_xfer && allocate_io_mem(td))
1128                 return 1;
1129
1130         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1131             (td->io_ops->flags & FIO_RAWIO))
1132                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1133         else
1134                 p = td->orig_buffer;
1135
1136         cl_align = os_cache_line_size();
1137
1138         for (i = 0; i < max_units; i++) {
1139                 void *ptr;
1140
1141                 if (td->terminate)
1142                         return 1;
1143
1144                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1145                 if (!ptr) {
1146                         log_err("fio: unable to allocate aligned memory\n");
1147                         break;
1148                 }
1149
1150                 io_u = ptr;
1151                 memset(io_u, 0, sizeof(*io_u));
1152                 INIT_FLIST_HEAD(&io_u->verify_list);
1153                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1154
1155                 if (data_xfer) {
1156                         io_u->buf = p;
1157                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1158
1159                         if (td_write(td))
1160                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1161                         if (td_write(td) && td->o.verify_pattern_bytes) {
1162                                 /*
1163                                  * Fill the buffer with the pattern if we are
1164                                  * going to be doing writes.
1165                                  */
1166                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1167                         }
1168                 }
1169
1170                 io_u->index = i;
1171                 io_u->flags = IO_U_F_FREE;
1172                 io_u_qpush(&td->io_u_freelist, io_u);
1173
1174                 /*
1175                  * io_u never leaves this stack, used for iteration of all
1176                  * io_u buffers.
1177                  */
1178                 io_u_qpush(&td->io_u_all, io_u);
1179
1180                 if (td->io_ops->io_u_init) {
1181                         int ret = td->io_ops->io_u_init(td, io_u);
1182
1183                         if (ret) {
1184                                 log_err("fio: failed to init engine data: %d\n", ret);
1185                                 return 1;
1186                         }
1187                 }
1188
1189                 p += max_bs;
1190         }
1191
1192         if (td->o.verify != VERIFY_NONE) {
1193                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1194                 if (!td->last_write_comp) {
1195                         log_err("fio: failed to alloc write comp data\n");
1196                         return 1;
1197                 }
1198         }
1199
1200         return 0;
1201 }
1202
1203 static int switch_ioscheduler(struct thread_data *td)
1204 {
1205         char tmp[256], tmp2[128];
1206         FILE *f;
1207         int ret;
1208
1209         if (td->io_ops->flags & FIO_DISKLESSIO)
1210                 return 0;
1211
1212         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1213
1214         f = fopen(tmp, "r+");
1215         if (!f) {
1216                 if (errno == ENOENT) {
1217                         log_err("fio: os or kernel doesn't support IO scheduler"
1218                                 " switching\n");
1219                         return 0;
1220                 }
1221                 td_verror(td, errno, "fopen iosched");
1222                 return 1;
1223         }
1224
1225         /*
1226          * Set io scheduler.
1227          */
1228         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1229         if (ferror(f) || ret != 1) {
1230                 td_verror(td, errno, "fwrite");
1231                 fclose(f);
1232                 return 1;
1233         }
1234
1235         rewind(f);
1236
1237         /*
1238          * Read back and check that the selected scheduler is now the default.
1239          */
1240         memset(tmp, 0, sizeof(tmp));
1241         ret = fread(tmp, sizeof(tmp), 1, f);
1242         if (ferror(f) || ret < 0) {
1243                 td_verror(td, errno, "fread");
1244                 fclose(f);
1245                 return 1;
1246         }
1247         /*
1248          * either a list of io schedulers or "none\n" is expected.
1249          */
1250         tmp[strlen(tmp) - 1] = '\0';
1251
1252
1253         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1254         if (!strstr(tmp, tmp2)) {
1255                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1256                 td_verror(td, EINVAL, "iosched_switch");
1257                 fclose(f);
1258                 return 1;
1259         }
1260
1261         fclose(f);
1262         return 0;
1263 }
1264
1265 static bool keep_running(struct thread_data *td)
1266 {
1267         unsigned long long limit;
1268
1269         if (td->done)
1270                 return false;
1271         if (td->o.time_based)
1272                 return true;
1273         if (td->o.loops) {
1274                 td->o.loops--;
1275                 return true;
1276         }
1277         if (exceeds_number_ios(td))
1278                 return false;
1279
1280         if (td->o.io_limit)
1281                 limit = td->o.io_limit;
1282         else
1283                 limit = td->o.size;
1284
1285         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1286                 uint64_t diff;
1287
1288                 /*
1289                  * If the difference is less than the minimum IO size, we
1290                  * are done.
1291                  */
1292                 diff = limit - ddir_rw_sum(td->io_bytes);
1293                 if (diff < td_max_bs(td))
1294                         return false;
1295
1296                 if (fio_files_done(td) && !td->o.io_limit)
1297                         return false;
1298
1299                 return true;
1300         }
1301
1302         return false;
1303 }
1304
1305 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1306 {
1307         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1308         int ret;
1309         char *str;
1310
1311         str = malloc(newlen);
1312         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1313
1314         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1315         ret = system(str);
1316         if (ret == -1)
1317                 log_err("fio: exec of cmd <%s> failed\n", str);
1318
1319         free(str);
1320         return ret;
1321 }
1322
1323 /*
1324  * Dry run to compute correct state of numberio for verification.
1325  */
1326 static uint64_t do_dry_run(struct thread_data *td)
1327 {
1328         td_set_runstate(td, TD_RUNNING);
1329
1330         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1331                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1332                 struct io_u *io_u;
1333                 int ret;
1334
1335                 if (td->terminate || td->done)
1336                         break;
1337
1338                 io_u = get_io_u(td);
1339                 if (!io_u)
1340                         break;
1341
1342                 io_u_set(io_u, IO_U_F_FLIGHT);
1343                 io_u->error = 0;
1344                 io_u->resid = 0;
1345                 if (ddir_rw(acct_ddir(io_u)))
1346                         td->io_issues[acct_ddir(io_u)]++;
1347                 if (ddir_rw(io_u->ddir)) {
1348                         io_u_mark_depth(td, 1);
1349                         td->ts.total_io_u[io_u->ddir]++;
1350                 }
1351
1352                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1353                     td->o.do_verify &&
1354                     td->o.verify != VERIFY_NONE &&
1355                     !td->o.experimental_verify)
1356                         log_io_piece(td, io_u);
1357
1358                 ret = io_u_sync_complete(td, io_u);
1359                 (void) ret;
1360         }
1361
1362         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1363 }
1364
1365 struct fork_data {
1366         struct thread_data *td;
1367         struct sk_out *sk_out;
1368 };
1369
1370 /*
1371  * Entry point for the thread based jobs. The process based jobs end up
1372  * here as well, after a little setup.
1373  */
1374 static void *thread_main(void *data)
1375 {
1376         struct fork_data *fd = data;
1377         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1378         struct thread_data *td = fd->td;
1379         struct thread_options *o = &td->o;
1380         struct sk_out *sk_out = fd->sk_out;
1381         pthread_condattr_t attr;
1382         int clear_state;
1383         int ret;
1384
1385         sk_out_assign(sk_out);
1386         free(fd);
1387
1388         if (!o->use_thread) {
1389                 setsid();
1390                 td->pid = getpid();
1391         } else
1392                 td->pid = gettid();
1393
1394         fio_local_clock_init(o->use_thread);
1395
1396         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1397
1398         if (is_backend)
1399                 fio_server_send_start(td);
1400
1401         INIT_FLIST_HEAD(&td->io_log_list);
1402         INIT_FLIST_HEAD(&td->io_hist_list);
1403         INIT_FLIST_HEAD(&td->verify_list);
1404         INIT_FLIST_HEAD(&td->trim_list);
1405         INIT_FLIST_HEAD(&td->next_rand_list);
1406         pthread_mutex_init(&td->io_u_lock, NULL);
1407         td->io_hist_tree = RB_ROOT;
1408
1409         pthread_condattr_init(&attr);
1410         pthread_cond_init(&td->verify_cond, &attr);
1411         pthread_cond_init(&td->free_cond, &attr);
1412
1413         td_set_runstate(td, TD_INITIALIZED);
1414         dprint(FD_MUTEX, "up startup_mutex\n");
1415         fio_mutex_up(startup_mutex);
1416         dprint(FD_MUTEX, "wait on td->mutex\n");
1417         fio_mutex_down(td->mutex);
1418         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1419
1420         /*
1421          * A new gid requires privilege, so we need to do this before setting
1422          * the uid.
1423          */
1424         if (o->gid != -1U && setgid(o->gid)) {
1425                 td_verror(td, errno, "setgid");
1426                 goto err;
1427         }
1428         if (o->uid != -1U && setuid(o->uid)) {
1429                 td_verror(td, errno, "setuid");
1430                 goto err;
1431         }
1432
1433         /*
1434          * If we have a gettimeofday() thread, make sure we exclude that
1435          * thread from this job
1436          */
1437         if (o->gtod_cpu)
1438                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1439
1440         /*
1441          * Set affinity first, in case it has an impact on the memory
1442          * allocations.
1443          */
1444         if (fio_option_is_set(o, cpumask)) {
1445                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1446                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1447                         if (!ret) {
1448                                 log_err("fio: no CPUs set\n");
1449                                 log_err("fio: Try increasing number of available CPUs\n");
1450                                 td_verror(td, EINVAL, "cpus_split");
1451                                 goto err;
1452                         }
1453                 }
1454                 ret = fio_setaffinity(td->pid, o->cpumask);
1455                 if (ret == -1) {
1456                         td_verror(td, errno, "cpu_set_affinity");
1457                         goto err;
1458                 }
1459         }
1460
1461 #ifdef CONFIG_LIBNUMA
1462         /* numa node setup */
1463         if (fio_option_is_set(o, numa_cpunodes) ||
1464             fio_option_is_set(o, numa_memnodes)) {
1465                 struct bitmask *mask;
1466
1467                 if (numa_available() < 0) {
1468                         td_verror(td, errno, "Does not support NUMA API\n");
1469                         goto err;
1470                 }
1471
1472                 if (fio_option_is_set(o, numa_cpunodes)) {
1473                         mask = numa_parse_nodestring(o->numa_cpunodes);
1474                         ret = numa_run_on_node_mask(mask);
1475                         numa_free_nodemask(mask);
1476                         if (ret == -1) {
1477                                 td_verror(td, errno, \
1478                                         "numa_run_on_node_mask failed\n");
1479                                 goto err;
1480                         }
1481                 }
1482
1483                 if (fio_option_is_set(o, numa_memnodes)) {
1484                         mask = NULL;
1485                         if (o->numa_memnodes)
1486                                 mask = numa_parse_nodestring(o->numa_memnodes);
1487
1488                         switch (o->numa_mem_mode) {
1489                         case MPOL_INTERLEAVE:
1490                                 numa_set_interleave_mask(mask);
1491                                 break;
1492                         case MPOL_BIND:
1493                                 numa_set_membind(mask);
1494                                 break;
1495                         case MPOL_LOCAL:
1496                                 numa_set_localalloc();
1497                                 break;
1498                         case MPOL_PREFERRED:
1499                                 numa_set_preferred(o->numa_mem_prefer_node);
1500                                 break;
1501                         case MPOL_DEFAULT:
1502                         default:
1503                                 break;
1504                         }
1505
1506                         if (mask)
1507                                 numa_free_nodemask(mask);
1508
1509                 }
1510         }
1511 #endif
1512
1513         if (fio_pin_memory(td))
1514                 goto err;
1515
1516         /*
1517          * May alter parameters that init_io_u() will use, so we need to
1518          * do this first.
1519          */
1520         if (init_iolog(td))
1521                 goto err;
1522
1523         if (init_io_u(td))
1524                 goto err;
1525
1526         if (o->verify_async && verify_async_init(td))
1527                 goto err;
1528
1529         if (fio_option_is_set(o, ioprio) ||
1530             fio_option_is_set(o, ioprio_class)) {
1531                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1532                 if (ret == -1) {
1533                         td_verror(td, errno, "ioprio_set");
1534                         goto err;
1535                 }
1536         }
1537
1538         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1539                 goto err;
1540
1541         errno = 0;
1542         if (nice(o->nice) == -1 && errno != 0) {
1543                 td_verror(td, errno, "nice");
1544                 goto err;
1545         }
1546
1547         if (o->ioscheduler && switch_ioscheduler(td))
1548                 goto err;
1549
1550         if (!o->create_serialize && setup_files(td))
1551                 goto err;
1552
1553         if (td_io_init(td))
1554                 goto err;
1555
1556         if (init_random_map(td))
1557                 goto err;
1558
1559         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1560                 goto err;
1561
1562         if (o->pre_read) {
1563                 if (pre_read_files(td) < 0)
1564                         goto err;
1565         }
1566
1567         if (iolog_compress_init(td, sk_out))
1568                 goto err;
1569
1570         fio_verify_init(td);
1571
1572         if (rate_submit_init(td, sk_out))
1573                 goto err;
1574
1575         fio_gettime(&td->epoch, NULL);
1576         fio_getrusage(&td->ru_start);
1577         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1578         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1579
1580         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1581                         o->ratemin[DDIR_TRIM]) {
1582                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1583                                         sizeof(td->bw_sample_time));
1584                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1585                                         sizeof(td->bw_sample_time));
1586                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1587                                         sizeof(td->bw_sample_time));
1588         }
1589
1590         clear_state = 0;
1591         while (keep_running(td)) {
1592                 uint64_t verify_bytes;
1593
1594                 fio_gettime(&td->start, NULL);
1595                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1596
1597                 if (clear_state)
1598                         clear_io_state(td, 0);
1599
1600                 prune_io_piece_log(td);
1601
1602                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1603                         verify_bytes = do_dry_run(td);
1604                 else {
1605                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1606
1607                         do_io(td, bytes_done);
1608
1609                         if (!ddir_rw_sum(bytes_done)) {
1610                                 fio_mark_td_terminate(td);
1611                                 verify_bytes = 0;
1612                         } else {
1613                                 verify_bytes = bytes_done[DDIR_WRITE] +
1614                                                 bytes_done[DDIR_TRIM];
1615                         }
1616                 }
1617
1618                 clear_state = 1;
1619
1620                 /*
1621                  * Make sure we've successfully updated the rusage stats
1622                  * before waiting on the stat mutex. Otherwise we could have
1623                  * the stat thread holding stat mutex and waiting for
1624                  * the rusage_sem, which would never get upped because
1625                  * this thread is waiting for the stat mutex.
1626                  */
1627                 check_update_rusage(td);
1628
1629                 fio_mutex_down(stat_mutex);
1630                 if (td_read(td) && td->io_bytes[DDIR_READ])
1631                         update_runtime(td, elapsed_us, DDIR_READ);
1632                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1633                         update_runtime(td, elapsed_us, DDIR_WRITE);
1634                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1635                         update_runtime(td, elapsed_us, DDIR_TRIM);
1636                 fio_gettime(&td->start, NULL);
1637                 fio_mutex_up(stat_mutex);
1638
1639                 if (td->error || td->terminate)
1640                         break;
1641
1642                 if (!o->do_verify ||
1643                     o->verify == VERIFY_NONE ||
1644                     (td->io_ops->flags & FIO_UNIDIR))
1645                         continue;
1646
1647                 clear_io_state(td, 0);
1648
1649                 fio_gettime(&td->start, NULL);
1650
1651                 do_verify(td, verify_bytes);
1652
1653                 /*
1654                  * See comment further up for why this is done here.
1655                  */
1656                 check_update_rusage(td);
1657
1658                 fio_mutex_down(stat_mutex);
1659                 update_runtime(td, elapsed_us, DDIR_READ);
1660                 fio_gettime(&td->start, NULL);
1661                 fio_mutex_up(stat_mutex);
1662
1663                 if (td->error || td->terminate)
1664                         break;
1665         }
1666
1667         update_rusage_stat(td);
1668         td->ts.total_run_time = mtime_since_now(&td->epoch);
1669         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1670         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1671         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1672
1673         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1674             (td->o.verify != VERIFY_NONE && td_write(td)))
1675                 verify_save_state(td->thread_number);
1676
1677         fio_unpin_memory(td);
1678
1679         fio_writeout_logs(td);
1680
1681         iolog_compress_exit(td);
1682         rate_submit_exit(td);
1683
1684         if (o->exec_postrun)
1685                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1686
1687         if (exitall_on_terminate || (o->exitall_error && td->error))
1688                 fio_terminate_threads(td->groupid);
1689
1690 err:
1691         if (td->error)
1692                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1693                                                         td->verror);
1694
1695         if (o->verify_async)
1696                 verify_async_exit(td);
1697
1698         close_and_free_files(td);
1699         cleanup_io_u(td);
1700         close_ioengine(td);
1701         cgroup_shutdown(td, &cgroup_mnt);
1702         verify_free_state(td);
1703
1704         if (fio_option_is_set(o, cpumask)) {
1705                 ret = fio_cpuset_exit(&o->cpumask);
1706                 if (ret)
1707                         td_verror(td, ret, "fio_cpuset_exit");
1708         }
1709
1710         /*
1711          * do this very late, it will log file closing as well
1712          */
1713         if (o->write_iolog_file)
1714                 write_iolog_close(td);
1715
1716         fio_mutex_remove(td->mutex);
1717         td->mutex = NULL;
1718
1719         td_set_runstate(td, TD_EXITED);
1720
1721         /*
1722          * Do this last after setting our runstate to exited, so we
1723          * know that the stat thread is signaled.
1724          */
1725         check_update_rusage(td);
1726
1727         sk_out_drop();
1728         return (void *) (uintptr_t) td->error;
1729 }
1730
1731
1732 /*
1733  * We cannot pass the td data into a forked process, so attach the td and
1734  * pass it to the thread worker.
1735  */
1736 static int fork_main(struct sk_out *sk_out, int shmid, int offset)
1737 {
1738         struct fork_data *fd;
1739         void *data, *ret;
1740
1741 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1742         data = shmat(shmid, NULL, 0);
1743         if (data == (void *) -1) {
1744                 int __err = errno;
1745
1746                 perror("shmat");
1747                 return __err;
1748         }
1749 #else
1750         /*
1751          * HP-UX inherits shm mappings?
1752          */
1753         data = threads;
1754 #endif
1755
1756         fd = calloc(1, sizeof(*fd));
1757         fd->td = data + offset * sizeof(struct thread_data);
1758         fd->sk_out = sk_out;
1759         ret = thread_main(fd);
1760         shmdt(data);
1761         return (int) (uintptr_t) ret;
1762 }
1763
1764 static void dump_td_info(struct thread_data *td)
1765 {
1766         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1767                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1768                         (unsigned long) time_since_now(&td->terminate_time));
1769 }
1770
1771 /*
1772  * Run over the job map and reap the threads that have exited, if any.
1773  */
1774 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1775                          unsigned int *m_rate)
1776 {
1777         struct thread_data *td;
1778         unsigned int cputhreads, realthreads, pending;
1779         int i, status, ret;
1780
1781         /*
1782          * reap exited threads (TD_EXITED -> TD_REAPED)
1783          */
1784         realthreads = pending = cputhreads = 0;
1785         for_each_td(td, i) {
1786                 int flags = 0;
1787
1788                 /*
1789                  * ->io_ops is NULL for a thread that has closed its
1790                  * io engine
1791                  */
1792                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1793                         cputhreads++;
1794                 else
1795                         realthreads++;
1796
1797                 if (!td->pid) {
1798                         pending++;
1799                         continue;
1800                 }
1801                 if (td->runstate == TD_REAPED)
1802                         continue;
1803                 if (td->o.use_thread) {
1804                         if (td->runstate == TD_EXITED) {
1805                                 td_set_runstate(td, TD_REAPED);
1806                                 goto reaped;
1807                         }
1808                         continue;
1809                 }
1810
1811                 flags = WNOHANG;
1812                 if (td->runstate == TD_EXITED)
1813                         flags = 0;
1814
1815                 /*
1816                  * check if someone quit or got killed in an unusual way
1817                  */
1818                 ret = waitpid(td->pid, &status, flags);
1819                 if (ret < 0) {
1820                         if (errno == ECHILD) {
1821                                 log_err("fio: pid=%d disappeared %d\n",
1822                                                 (int) td->pid, td->runstate);
1823                                 td->sig = ECHILD;
1824                                 td_set_runstate(td, TD_REAPED);
1825                                 goto reaped;
1826                         }
1827                         perror("waitpid");
1828                 } else if (ret == td->pid) {
1829                         if (WIFSIGNALED(status)) {
1830                                 int sig = WTERMSIG(status);
1831
1832                                 if (sig != SIGTERM && sig != SIGUSR2)
1833                                         log_err("fio: pid=%d, got signal=%d\n",
1834                                                         (int) td->pid, sig);
1835                                 td->sig = sig;
1836                                 td_set_runstate(td, TD_REAPED);
1837                                 goto reaped;
1838                         }
1839                         if (WIFEXITED(status)) {
1840                                 if (WEXITSTATUS(status) && !td->error)
1841                                         td->error = WEXITSTATUS(status);
1842
1843                                 td_set_runstate(td, TD_REAPED);
1844                                 goto reaped;
1845                         }
1846                 }
1847
1848                 /*
1849                  * If the job is stuck, do a forceful timeout of it and
1850                  * move on.
1851                  */
1852                 if (td->terminate &&
1853                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1854                         dump_td_info(td);
1855                         td_set_runstate(td, TD_REAPED);
1856                         goto reaped;
1857                 }
1858
1859                 /*
1860                  * thread is not dead, continue
1861                  */
1862                 pending++;
1863                 continue;
1864 reaped:
1865                 (*nr_running)--;
1866                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1867                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1868                 if (!td->pid)
1869                         pending--;
1870
1871                 if (td->error)
1872                         exit_value++;
1873
1874                 done_secs += mtime_since_now(&td->epoch) / 1000;
1875                 profile_td_exit(td);
1876         }
1877
1878         if (*nr_running == cputhreads && !pending && realthreads)
1879                 fio_terminate_threads(TERMINATE_ALL);
1880 }
1881
1882 static bool __check_trigger_file(void)
1883 {
1884         struct stat sb;
1885
1886         if (!trigger_file)
1887                 return false;
1888
1889         if (stat(trigger_file, &sb))
1890                 return false;
1891
1892         if (unlink(trigger_file) < 0)
1893                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1894                                                         strerror(errno));
1895
1896         return true;
1897 }
1898
1899 static bool trigger_timedout(void)
1900 {
1901         if (trigger_timeout)
1902                 return time_since_genesis() >= trigger_timeout;
1903
1904         return false;
1905 }
1906
1907 void exec_trigger(const char *cmd)
1908 {
1909         int ret;
1910
1911         if (!cmd)
1912                 return;
1913
1914         ret = system(cmd);
1915         if (ret == -1)
1916                 log_err("fio: failed executing %s trigger\n", cmd);
1917 }
1918
1919 void check_trigger_file(void)
1920 {
1921         if (__check_trigger_file() || trigger_timedout()) {
1922                 if (nr_clients)
1923                         fio_clients_send_trigger(trigger_remote_cmd);
1924                 else {
1925                         verify_save_state(IO_LIST_ALL);
1926                         fio_terminate_threads(TERMINATE_ALL);
1927                         exec_trigger(trigger_cmd);
1928                 }
1929         }
1930 }
1931
1932 static int fio_verify_load_state(struct thread_data *td)
1933 {
1934         int ret;
1935
1936         if (!td->o.verify_state)
1937                 return 0;
1938
1939         if (is_backend) {
1940                 void *data;
1941                 int ver;
1942
1943                 ret = fio_server_get_verify_state(td->o.name,
1944                                         td->thread_number - 1, &data, &ver);
1945                 if (!ret)
1946                         verify_convert_assign_state(td, data, ver);
1947         } else
1948                 ret = verify_load_state(td, "local");
1949
1950         return ret;
1951 }
1952
1953 static void do_usleep(unsigned int usecs)
1954 {
1955         check_for_running_stats();
1956         check_trigger_file();
1957         usleep(usecs);
1958 }
1959
1960 static bool check_mount_writes(struct thread_data *td)
1961 {
1962         struct fio_file *f;
1963         unsigned int i;
1964
1965         if (!td_write(td) || td->o.allow_mounted_write)
1966                 return false;
1967
1968         for_each_file(td, f, i) {
1969                 if (f->filetype != FIO_TYPE_BD)
1970                         continue;
1971                 if (device_is_mounted(f->file_name))
1972                         goto mounted;
1973         }
1974
1975         return false;
1976 mounted:
1977         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1978         return true;
1979 }
1980
1981 static bool waitee_running(struct thread_data *me)
1982 {
1983         const char *waitee = me->o.wait_for;
1984         const char *self = me->o.name;
1985         struct thread_data *td;
1986         int i;
1987
1988         if (!waitee)
1989                 return false;
1990
1991         for_each_td(td, i) {
1992                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
1993                         continue;
1994
1995                 if (td->runstate < TD_EXITED) {
1996                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
1997                                         self, td->o.name,
1998                                         runstate_to_name(td->runstate));
1999                         return true;
2000                 }
2001         }
2002
2003         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2004         return false;
2005 }
2006
2007 /*
2008  * Main function for kicking off and reaping jobs, as needed.
2009  */
2010 static void run_threads(struct sk_out *sk_out)
2011 {
2012         struct thread_data *td;
2013         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2014         uint64_t spent;
2015
2016         if (fio_gtod_offload && fio_start_gtod_thread())
2017                 return;
2018
2019         fio_idle_prof_init();
2020
2021         set_sig_handlers();
2022
2023         nr_thread = nr_process = 0;
2024         for_each_td(td, i) {
2025                 if (check_mount_writes(td))
2026                         return;
2027                 if (td->o.use_thread)
2028                         nr_thread++;
2029                 else
2030                         nr_process++;
2031         }
2032
2033         if (output_format & FIO_OUTPUT_NORMAL) {
2034                 log_info("Starting ");
2035                 if (nr_thread)
2036                         log_info("%d thread%s", nr_thread,
2037                                                 nr_thread > 1 ? "s" : "");
2038                 if (nr_process) {
2039                         if (nr_thread)
2040                                 log_info(" and ");
2041                         log_info("%d process%s", nr_process,
2042                                                 nr_process > 1 ? "es" : "");
2043                 }
2044                 log_info("\n");
2045                 log_info_flush();
2046         }
2047
2048         todo = thread_number;
2049         nr_running = 0;
2050         nr_started = 0;
2051         m_rate = t_rate = 0;
2052
2053         for_each_td(td, i) {
2054                 print_status_init(td->thread_number - 1);
2055
2056                 if (!td->o.create_serialize)
2057                         continue;
2058
2059                 if (fio_verify_load_state(td))
2060                         goto reap;
2061
2062                 /*
2063                  * do file setup here so it happens sequentially,
2064                  * we don't want X number of threads getting their
2065                  * client data interspersed on disk
2066                  */
2067                 if (setup_files(td)) {
2068 reap:
2069                         exit_value++;
2070                         if (td->error)
2071                                 log_err("fio: pid=%d, err=%d/%s\n",
2072                                         (int) td->pid, td->error, td->verror);
2073                         td_set_runstate(td, TD_REAPED);
2074                         todo--;
2075                 } else {
2076                         struct fio_file *f;
2077                         unsigned int j;
2078
2079                         /*
2080                          * for sharing to work, each job must always open
2081                          * its own files. so close them, if we opened them
2082                          * for creation
2083                          */
2084                         for_each_file(td, f, j) {
2085                                 if (fio_file_open(f))
2086                                         td_io_close_file(td, f);
2087                         }
2088                 }
2089         }
2090
2091         /* start idle threads before io threads start to run */
2092         fio_idle_prof_start();
2093
2094         set_genesis_time();
2095
2096         while (todo) {
2097                 struct thread_data *map[REAL_MAX_JOBS];
2098                 struct timeval this_start;
2099                 int this_jobs = 0, left;
2100
2101                 /*
2102                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2103                  */
2104                 for_each_td(td, i) {
2105                         if (td->runstate != TD_NOT_CREATED)
2106                                 continue;
2107
2108                         /*
2109                          * never got a chance to start, killed by other
2110                          * thread for some reason
2111                          */
2112                         if (td->terminate) {
2113                                 todo--;
2114                                 continue;
2115                         }
2116
2117                         if (td->o.start_delay) {
2118                                 spent = utime_since_genesis();
2119
2120                                 if (td->o.start_delay > spent)
2121                                         continue;
2122                         }
2123
2124                         if (td->o.stonewall && (nr_started || nr_running)) {
2125                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2126                                                         td->o.name);
2127                                 break;
2128                         }
2129
2130                         if (waitee_running(td)) {
2131                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2132                                                 td->o.name, td->o.wait_for);
2133                                 continue;
2134                         }
2135
2136                         init_disk_util(td);
2137
2138                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2139                         td->update_rusage = 0;
2140
2141                         /*
2142                          * Set state to created. Thread will transition
2143                          * to TD_INITIALIZED when it's done setting up.
2144                          */
2145                         td_set_runstate(td, TD_CREATED);
2146                         map[this_jobs++] = td;
2147                         nr_started++;
2148
2149                         if (td->o.use_thread) {
2150                                 struct fork_data *fd;
2151                                 int ret;
2152
2153                                 fd = calloc(1, sizeof(*fd));
2154                                 fd->td = td;
2155                                 fd->sk_out = sk_out;
2156
2157                                 dprint(FD_PROCESS, "will pthread_create\n");
2158                                 ret = pthread_create(&td->thread, NULL,
2159                                                         thread_main, fd);
2160                                 if (ret) {
2161                                         log_err("pthread_create: %s\n",
2162                                                         strerror(ret));
2163                                         free(fd);
2164                                         nr_started--;
2165                                         break;
2166                                 }
2167                                 ret = pthread_detach(td->thread);
2168                                 if (ret)
2169                                         log_err("pthread_detach: %s",
2170                                                         strerror(ret));
2171                         } else {
2172                                 pid_t pid;
2173                                 dprint(FD_PROCESS, "will fork\n");
2174                                 pid = fork();
2175                                 if (!pid) {
2176                                         int ret = fork_main(sk_out, shm_id, i);
2177
2178                                         _exit(ret);
2179                                 } else if (i == fio_debug_jobno)
2180                                         *fio_debug_jobp = pid;
2181                         }
2182                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2183                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2184                                 log_err("fio: job startup hung? exiting.\n");
2185                                 fio_terminate_threads(TERMINATE_ALL);
2186                                 fio_abort = 1;
2187                                 nr_started--;
2188                                 break;
2189                         }
2190                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2191                 }
2192
2193                 /*
2194                  * Wait for the started threads to transition to
2195                  * TD_INITIALIZED.
2196                  */
2197                 fio_gettime(&this_start, NULL);
2198                 left = this_jobs;
2199                 while (left && !fio_abort) {
2200                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2201                                 break;
2202
2203                         do_usleep(100000);
2204
2205                         for (i = 0; i < this_jobs; i++) {
2206                                 td = map[i];
2207                                 if (!td)
2208                                         continue;
2209                                 if (td->runstate == TD_INITIALIZED) {
2210                                         map[i] = NULL;
2211                                         left--;
2212                                 } else if (td->runstate >= TD_EXITED) {
2213                                         map[i] = NULL;
2214                                         left--;
2215                                         todo--;
2216                                         nr_running++; /* work-around... */
2217                                 }
2218                         }
2219                 }
2220
2221                 if (left) {
2222                         log_err("fio: %d job%s failed to start\n", left,
2223                                         left > 1 ? "s" : "");
2224                         for (i = 0; i < this_jobs; i++) {
2225                                 td = map[i];
2226                                 if (!td)
2227                                         continue;
2228                                 kill(td->pid, SIGTERM);
2229                         }
2230                         break;
2231                 }
2232
2233                 /*
2234                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2235                  */
2236                 for_each_td(td, i) {
2237                         if (td->runstate != TD_INITIALIZED)
2238                                 continue;
2239
2240                         if (in_ramp_time(td))
2241                                 td_set_runstate(td, TD_RAMP);
2242                         else
2243                                 td_set_runstate(td, TD_RUNNING);
2244                         nr_running++;
2245                         nr_started--;
2246                         m_rate += ddir_rw_sum(td->o.ratemin);
2247                         t_rate += ddir_rw_sum(td->o.rate);
2248                         todo--;
2249                         fio_mutex_up(td->mutex);
2250                 }
2251
2252                 reap_threads(&nr_running, &t_rate, &m_rate);
2253
2254                 if (todo)
2255                         do_usleep(100000);
2256         }
2257
2258         while (nr_running) {
2259                 reap_threads(&nr_running, &t_rate, &m_rate);
2260                 do_usleep(10000);
2261         }
2262
2263         fio_idle_prof_stop();
2264
2265         update_io_ticks();
2266 }
2267
2268 static void wait_for_helper_thread_exit(void)
2269 {
2270         void *ret;
2271
2272         helper_exit = 1;
2273         pthread_cond_signal(&helper_cond);
2274         pthread_join(helper_thread, &ret);
2275 }
2276
2277 static void free_disk_util(void)
2278 {
2279         disk_util_prune_entries();
2280
2281         pthread_cond_destroy(&helper_cond);
2282 }
2283
2284 static void *helper_thread_main(void *data)
2285 {
2286         struct sk_out *sk_out = data;
2287         int ret = 0;
2288
2289         sk_out_assign(sk_out);
2290
2291         fio_mutex_up(startup_mutex);
2292
2293         while (!ret) {
2294                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2295                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2296                 struct timespec ts;
2297                 struct timeval tv;
2298
2299                 gettimeofday(&tv, NULL);
2300                 ts.tv_sec = tv.tv_sec + sec;
2301                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2302
2303                 if (ts.tv_nsec >= 1000000000ULL) {
2304                         ts.tv_nsec -= 1000000000ULL;
2305                         ts.tv_sec++;
2306                 }
2307
2308                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2309
2310                 ret = update_io_ticks();
2311
2312                 if (helper_do_stat) {
2313                         helper_do_stat = 0;
2314                         __show_running_run_stats();
2315                 }
2316
2317                 if (!is_backend)
2318                         print_thread_status();
2319         }
2320
2321         sk_out_drop();
2322         return NULL;
2323 }
2324
2325 static int create_helper_thread(struct sk_out *sk_out)
2326 {
2327         int ret;
2328
2329         setup_disk_util();
2330
2331         pthread_cond_init(&helper_cond, NULL);
2332         pthread_mutex_init(&helper_lock, NULL);
2333
2334         ret = pthread_create(&helper_thread, NULL, helper_thread_main, sk_out);
2335         if (ret) {
2336                 log_err("Can't create helper thread: %s\n", strerror(ret));
2337                 return 1;
2338         }
2339
2340         dprint(FD_MUTEX, "wait on startup_mutex\n");
2341         fio_mutex_down(startup_mutex);
2342         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2343         return 0;
2344 }
2345
2346 int fio_backend(struct sk_out *sk_out)
2347 {
2348         struct thread_data *td;
2349         int i;
2350
2351         if (exec_profile) {
2352                 if (load_profile(exec_profile))
2353                         return 1;
2354                 free(exec_profile);
2355                 exec_profile = NULL;
2356         }
2357         if (!thread_number)
2358                 return 0;
2359
2360         if (write_bw_log) {
2361                 struct log_params p = {
2362                         .log_type = IO_LOG_TYPE_BW,
2363                 };
2364
2365                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2366                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2367                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2368         }
2369
2370         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2371         if (startup_mutex == NULL)
2372                 return 1;
2373
2374         set_genesis_time();
2375         stat_init();
2376         create_helper_thread(sk_out);
2377
2378         cgroup_list = smalloc(sizeof(*cgroup_list));
2379         INIT_FLIST_HEAD(cgroup_list);
2380
2381         run_threads(sk_out);
2382
2383         wait_for_helper_thread_exit();
2384
2385         if (!fio_abort) {
2386                 __show_run_stats();
2387                 if (write_bw_log) {
2388                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2389                                 struct io_log *log = agg_io_log[i];
2390
2391                                 flush_log(log, 0);
2392                                 free_log(log);
2393                         }
2394                 }
2395         }
2396
2397         for_each_td(td, i) {
2398                 fio_options_free(td);
2399                 if (td->rusage_sem) {
2400                         fio_mutex_remove(td->rusage_sem);
2401                         td->rusage_sem = NULL;
2402                 }
2403         }
2404
2405         free_disk_util();
2406         cgroup_kill(cgroup_list);
2407         sfree(cgroup_list);
2408         sfree(cgroup_mnt);
2409
2410         fio_mutex_remove(startup_mutex);
2411         stat_exit();
2412         return exit_value;
2413 }