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 "memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "lib/tp.h"
  58 #include "workqueue.h"
  59 #include "lib/mountcheck.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 int __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 0;
 162
 163         /*
 164          * allow a 2 second settle period in the beginning
 165          */
 166         if (mtime_since(&td->start, now) < 2000)
 167                 return 0;
 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 0;
 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 1;
 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 1;
 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 1;
 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 1;
 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 0;
 234 }
 235
 236 static int check_min_rate(struct thread_data *td, struct timeval *now)
 237 {
 238         int ret = 0;
 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 int 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 1;
 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 1;
 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 1;
 311         } else if (ret == FIO_Q_QUEUED) {
 312                 if (io_u_queued_complete(td, 1) < 0)
 313                         return 1;
 314         } else if (ret == FIO_Q_COMPLETED) {
 315                 if (io_u->error) {
 316                         td_verror(td, io_u->error, "td_io_queue");
 317                         return 1;
 318                 }
 319
 320                 if (io_u_sync_complete(td, io_u) < 0)
 321                         return 1;
 322         } else if (ret == FIO_Q_BUSY) {
 323                 if (td_io_commit(td))
 324                         return 1;
 325                 goto requeue;
 326         }
 327
 328         return 0;
 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 int runtime_exceeded(struct thread_data *td, struct timeval *t)
 358 {
 359         if (in_ramp_time(td))
 360                 return 0;
 361         if (!td->o.timeout)
 362                 return 0;
 363         if (utime_since(&td->epoch, t) >= td->o.timeout)
 364                 return 1;
 365
 366         return 0;
 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 int 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 1;
 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 0;
 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 1;
 419                 } else {
 420                         /*
 421                          * Stop the I/O in case of a fatal
 422                          * error.
 423                          */
 424                         update_error_count(td, err);
 425                         return 1;
 426                 }
 427         }
 428
 429         return 0;
 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 int 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 unsigned int exceeds_number_ios(struct thread_data *td)
 717 {
 718         unsigned long long number_ios;
 719
 720         if (!td->o.number_ios)
 721                 return 0;
 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 int 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 int 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.poisson_rate) {
 786                 iops = bps / td->o.bs[ddir];
 787                 td->last_usec += (int64_t) (1000000 / iops) *
 788                                         -logf(__rand_0_1(&td->poisson_state));
 789                 return td->last_usec;
 790         } else if (bps) {
 791                 secs = bytes / bps;
 792                 remainder = bytes % bps;
 793                 return remainder * 1000000 / bps + secs * 1000000;
 794         }
 795
 796         return 0;
 797 }
 798
 799 /*
 800  * Main IO worker function. It retrieves io_u's to process and queues
 801  * and reaps them, checking for rate and errors along the way.
 802  *
 803  * Returns number of bytes written and trimmed.
 804  */
 805 static uint64_t do_io(struct thread_data *td)
 806 {
 807         unsigned int i;
 808         int ret = 0;
 809         uint64_t total_bytes, bytes_issued = 0;
 810
 811         if (in_ramp_time(td))
 812                 td_set_runstate(td, TD_RAMP);
 813         else
 814                 td_set_runstate(td, TD_RUNNING);
 815
 816         lat_target_init(td);
 817
 818         total_bytes = td->o.size;
 819         /*
 820         * Allow random overwrite workloads to write up to io_limit
 821         * before starting verification phase as 'size' doesn't apply.
 822         */
 823         if (td_write(td) && td_random(td) && td->o.norandommap)
 824                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 825         /*
 826          * If verify_backlog is enabled, we'll run the verify in this
 827          * handler as well. For that case, we may need up to twice the
 828          * amount of bytes.
 829          */
 830         if (td->o.verify != VERIFY_NONE &&
 831            (td_write(td) && td->o.verify_backlog))
 832                 total_bytes += td->o.size;
 833
 834         /* In trimwrite mode, each byte is trimmed and then written, so
 835          * allow total_bytes to be twice as big */
 836         if (td_trimwrite(td))
 837                 total_bytes += td->total_io_size;
 838
 839         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 840                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 841                 td->o.time_based) {
 842                 struct timeval comp_time;
 843                 struct io_u *io_u;
 844                 int full;
 845                 enum fio_ddir ddir;
 846
 847                 check_update_rusage(td);
 848
 849                 if (td->terminate || td->done)
 850                         break;
 851
 852                 update_tv_cache(td);
 853
 854                 if (runtime_exceeded(td, &td->tv_cache)) {
 855                         __update_tv_cache(td);
 856                         if (runtime_exceeded(td, &td->tv_cache)) {
 857                                 fio_mark_td_terminate(td);
 858                                 break;
 859                         }
 860                 }
 861
 862                 if (flow_threshold_exceeded(td))
 863                         continue;
 864
 865                 if (bytes_issued >= total_bytes)
 866                         break;
 867
 868                 io_u = get_io_u(td);
 869                 if (IS_ERR_OR_NULL(io_u)) {
 870                         int err = PTR_ERR(io_u);
 871
 872                         io_u = NULL;
 873                         if (err == -EBUSY) {
 874                                 ret = FIO_Q_BUSY;
 875                                 goto reap;
 876                         }
 877                         if (td->o.latency_target)
 878                                 goto reap;
 879                         break;
 880                 }
 881
 882                 ddir = io_u->ddir;
 883
 884                 /*
 885                  * Add verification end_io handler if:
 886                  *      - Asked to verify (!td_rw(td))
 887                  *      - Or the io_u is from our verify list (mixed write/ver)
 888                  */
 889                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 890                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 891
 892                         if (!td->o.verify_pattern_bytes) {
 893                                 io_u->rand_seed = __rand(&td->verify_state);
 894                                 if (sizeof(int) != sizeof(long *))
 895                                         io_u->rand_seed *= __rand(&td->verify_state);
 896                         }
 897
 898                         if (verify_state_should_stop(td, io_u)) {
 899                                 put_io_u(td, io_u);
 900                                 break;
 901                         }
 902
 903                         if (td->o.verify_async)
 904                                 io_u->end_io = verify_io_u_async;
 905                         else
 906                                 io_u->end_io = verify_io_u;
 907                         td_set_runstate(td, TD_VERIFYING);
 908                 } else if (in_ramp_time(td))
 909                         td_set_runstate(td, TD_RAMP);
 910                 else
 911                         td_set_runstate(td, TD_RUNNING);
 912
 913                 /*
 914                  * Always log IO before it's issued, so we know the specific
 915                  * order of it. The logged unit will track when the IO has
 916                  * completed.
 917                  */
 918                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 919                     td->o.do_verify &&
 920                     td->o.verify != VERIFY_NONE &&
 921                     !td->o.experimental_verify)
 922                         log_io_piece(td, io_u);
 923
 924                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 925                         if (td->error)
 926                                 break;
 927                         ret = workqueue_enqueue(&td->io_wq, io_u);
 928
 929                         if (should_check_rate(td))
 930                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 931
 932                 } else {
 933                         ret = td_io_queue(td, io_u);
 934
 935                         if (should_check_rate(td))
 936                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 937
 938                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 939                                 break;
 940
 941                         /*
 942                          * See if we need to complete some commands. Note that
 943                          * we can get BUSY even without IO queued, if the
 944                          * system is resource starved.
 945                          */
 946 reap:
 947                         full = queue_full(td) ||
 948                                 (ret == FIO_Q_BUSY && td->cur_depth);
 949                         if (full || io_in_polling(td))
 950                                 ret = wait_for_completions(td, &comp_time);
 951                 }
 952                 if (ret < 0)
 953                         break;
 954                 if (!ddir_rw_sum(td->bytes_done) &&
 955                     !(td->io_ops->flags & FIO_NOIO))
 956                         continue;
 957
 958                 if (!in_ramp_time(td) && should_check_rate(td)) {
 959                         if (check_min_rate(td, &comp_time)) {
 960                                 if (exitall_on_terminate)
 961                                         fio_terminate_threads(td->groupid);
 962                                 td_verror(td, EIO, "check_min_rate");
 963                                 break;
 964                         }
 965                 }
 966                 if (!in_ramp_time(td) && td->o.latency_target)
 967                         lat_target_check(td);
 968
 969                 if (td->o.thinktime) {
 970                         unsigned long long b;
 971
 972                         b = ddir_rw_sum(td->io_blocks);
 973                         if (!(b % td->o.thinktime_blocks)) {
 974                                 int left;
 975
 976                                 io_u_quiesce(td);
 977
 978                                 if (td->o.thinktime_spin)
 979                                         usec_spin(td->o.thinktime_spin);
 980
 981                                 left = td->o.thinktime - td->o.thinktime_spin;
 982                                 if (left)
 983                                         usec_sleep(td, left);
 984                         }
 985                 }
 986         }
 987
 988         check_update_rusage(td);
 989
 990         if (td->trim_entries)
 991                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 992
 993         if (td->o.fill_device && td->error == ENOSPC) {
 994                 td->error = 0;
 995                 fio_mark_td_terminate(td);
 996         }
 997         if (!td->error) {
 998                 struct fio_file *f;
 999
1000                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1001                         workqueue_flush(&td->io_wq);
1002                         i = 0;
1003                 } else
1004                         i = td->cur_depth;
1005
1006                 if (i) {
1007                         ret = io_u_queued_complete(td, i);
1008                         if (td->o.fill_device && td->error == ENOSPC)
1009                                 td->error = 0;
1010                 }
1011
1012                 if (should_fsync(td) && td->o.end_fsync) {
1013                         td_set_runstate(td, TD_FSYNCING);
1014
1015                         for_each_file(td, f, i) {
1016                                 if (!fio_file_fsync(td, f))
1017                                         continue;
1018
1019                                 log_err("fio: end_fsync failed for file %s\n",
1020                                                                 f->file_name);
1021                         }
1022                 }
1023         } else
1024                 cleanup_pending_aio(td);
1025
1026         /*
1027          * stop job if we failed doing any IO
1028          */
1029         if (!ddir_rw_sum(td->this_io_bytes))
1030                 td->done = 1;
1031
1032         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1033 }
1034
1035 static void cleanup_io_u(struct thread_data *td)
1036 {
1037         struct io_u *io_u;
1038
1039         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1040
1041                 if (td->io_ops->io_u_free)
1042                         td->io_ops->io_u_free(td, io_u);
1043
1044                 fio_memfree(io_u, sizeof(*io_u));
1045         }
1046
1047         free_io_mem(td);
1048
1049         io_u_rexit(&td->io_u_requeues);
1050         io_u_qexit(&td->io_u_freelist);
1051         io_u_qexit(&td->io_u_all);
1052
1053         if (td->last_write_comp)
1054                 sfree(td->last_write_comp);
1055 }
1056
1057 static int init_io_u(struct thread_data *td)
1058 {
1059         struct io_u *io_u;
1060         unsigned int max_bs, min_write;
1061         int cl_align, i, max_units;
1062         int data_xfer = 1, err;
1063         char *p;
1064
1065         max_units = td->o.iodepth;
1066         max_bs = td_max_bs(td);
1067         min_write = td->o.min_bs[DDIR_WRITE];
1068         td->orig_buffer_size = (unsigned long long) max_bs
1069                                         * (unsigned long long) max_units;
1070
1071         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1072                 data_xfer = 0;
1073
1074         err = 0;
1075         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1076         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1077         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1078
1079         if (err) {
1080                 log_err("fio: failed setting up IO queues\n");
1081                 return 1;
1082         }
1083
1084         /*
1085          * if we may later need to do address alignment, then add any
1086          * possible adjustment here so that we don't cause a buffer
1087          * overflow later. this adjustment may be too much if we get
1088          * lucky and the allocator gives us an aligned address.
1089          */
1090         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1091             (td->io_ops->flags & FIO_RAWIO))
1092                 td->orig_buffer_size += page_mask + td->o.mem_align;
1093
1094         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1095                 unsigned long bs;
1096
1097                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1098                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1099         }
1100
1101         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1102                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1103                 return 1;
1104         }
1105
1106         if (data_xfer && allocate_io_mem(td))
1107                 return 1;
1108
1109         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1110             (td->io_ops->flags & FIO_RAWIO))
1111                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1112         else
1113                 p = td->orig_buffer;
1114
1115         cl_align = os_cache_line_size();
1116
1117         for (i = 0; i < max_units; i++) {
1118                 void *ptr;
1119
1120                 if (td->terminate)
1121                         return 1;
1122
1123                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1124                 if (!ptr) {
1125                         log_err("fio: unable to allocate aligned memory\n");
1126                         break;
1127                 }
1128
1129                 io_u = ptr;
1130                 memset(io_u, 0, sizeof(*io_u));
1131                 INIT_FLIST_HEAD(&io_u->verify_list);
1132                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1133
1134                 if (data_xfer) {
1135                         io_u->buf = p;
1136                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1137
1138                         if (td_write(td))
1139                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1140                         if (td_write(td) && td->o.verify_pattern_bytes) {
1141                                 /*
1142                                  * Fill the buffer with the pattern if we are
1143                                  * going to be doing writes.
1144                                  */
1145                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1146                         }
1147                 }
1148
1149                 io_u->index = i;
1150                 io_u->flags = IO_U_F_FREE;
1151                 io_u_qpush(&td->io_u_freelist, io_u);
1152
1153                 /*
1154                  * io_u never leaves this stack, used for iteration of all
1155                  * io_u buffers.
1156                  */
1157                 io_u_qpush(&td->io_u_all, io_u);
1158
1159                 if (td->io_ops->io_u_init) {
1160                         int ret = td->io_ops->io_u_init(td, io_u);
1161
1162                         if (ret) {
1163                                 log_err("fio: failed to init engine data: %d\n", ret);
1164                                 return 1;
1165                         }
1166                 }
1167
1168                 p += max_bs;
1169         }
1170
1171         if (td->o.verify != VERIFY_NONE) {
1172                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1173                 if (!td->last_write_comp) {
1174                         log_err("fio: failed to alloc write comp data\n");
1175                         return 1;
1176                 }
1177         }
1178
1179         return 0;
1180 }
1181
1182 static int switch_ioscheduler(struct thread_data *td)
1183 {
1184         char tmp[256], tmp2[128];
1185         FILE *f;
1186         int ret;
1187
1188         if (td->io_ops->flags & FIO_DISKLESSIO)
1189                 return 0;
1190
1191         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1192
1193         f = fopen(tmp, "r+");
1194         if (!f) {
1195                 if (errno == ENOENT) {
1196                         log_err("fio: os or kernel doesn't support IO scheduler"
1197                                 " switching\n");
1198                         return 0;
1199                 }
1200                 td_verror(td, errno, "fopen iosched");
1201                 return 1;
1202         }
1203
1204         /*
1205          * Set io scheduler.
1206          */
1207         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1208         if (ferror(f) || ret != 1) {
1209                 td_verror(td, errno, "fwrite");
1210                 fclose(f);
1211                 return 1;
1212         }
1213
1214         rewind(f);
1215
1216         /*
1217          * Read back and check that the selected scheduler is now the default.
1218          */
1219         memset(tmp, 0, sizeof(tmp));
1220         ret = fread(tmp, sizeof(tmp), 1, f);
1221         if (ferror(f) || ret < 0) {
1222                 td_verror(td, errno, "fread");
1223                 fclose(f);
1224                 return 1;
1225         }
1226         /*
1227          * either a list of io schedulers or "none\n" is expected.
1228          */
1229         tmp[strlen(tmp) - 1] = '\0';
1230
1231
1232         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1233         if (!strstr(tmp, tmp2)) {
1234                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1235                 td_verror(td, EINVAL, "iosched_switch");
1236                 fclose(f);
1237                 return 1;
1238         }
1239
1240         fclose(f);
1241         return 0;
1242 }
1243
1244 static int keep_running(struct thread_data *td)
1245 {
1246         unsigned long long limit;
1247
1248         if (td->done)
1249                 return 0;
1250         if (td->o.time_based)
1251                 return 1;
1252         if (td->o.loops) {
1253                 td->o.loops--;
1254                 return 1;
1255         }
1256         if (exceeds_number_ios(td))
1257                 return 0;
1258
1259         if (td->o.io_limit)
1260                 limit = td->o.io_limit;
1261         else
1262                 limit = td->o.size;
1263
1264         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1265                 uint64_t diff;
1266
1267                 /*
1268                  * If the difference is less than the minimum IO size, we
1269                  * are done.
1270                  */
1271                 diff = limit - ddir_rw_sum(td->io_bytes);
1272                 if (diff < td_max_bs(td))
1273                         return 0;
1274
1275                 if (fio_files_done(td))
1276                         return 0;
1277
1278                 return 1;
1279         }
1280
1281         return 0;
1282 }
1283
1284 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1285 {
1286         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1287         int ret;
1288         char *str;
1289
1290         str = malloc(newlen);
1291         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1292
1293         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1294         ret = system(str);
1295         if (ret == -1)
1296                 log_err("fio: exec of cmd <%s> failed\n", str);
1297
1298         free(str);
1299         return ret;
1300 }
1301
1302 /*
1303  * Dry run to compute correct state of numberio for verification.
1304  */
1305 static uint64_t do_dry_run(struct thread_data *td)
1306 {
1307         td_set_runstate(td, TD_RUNNING);
1308
1309         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1310                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1311                 struct io_u *io_u;
1312                 int ret;
1313
1314                 if (td->terminate || td->done)
1315                         break;
1316
1317                 io_u = get_io_u(td);
1318                 if (!io_u)
1319                         break;
1320
1321                 io_u_set(io_u, IO_U_F_FLIGHT);
1322                 io_u->error = 0;
1323                 io_u->resid = 0;
1324                 if (ddir_rw(acct_ddir(io_u)))
1325                         td->io_issues[acct_ddir(io_u)]++;
1326                 if (ddir_rw(io_u->ddir)) {
1327                         io_u_mark_depth(td, 1);
1328                         td->ts.total_io_u[io_u->ddir]++;
1329                 }
1330
1331                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1332                     td->o.do_verify &&
1333                     td->o.verify != VERIFY_NONE &&
1334                     !td->o.experimental_verify)
1335                         log_io_piece(td, io_u);
1336
1337                 ret = io_u_sync_complete(td, io_u);
1338                 (void) ret;
1339         }
1340
1341         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1342 }
1343
1344 static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1345 {
1346         const enum fio_ddir ddir = io_u->ddir;
1347         int ret;
1348
1349         dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1350
1351         io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1352
1353         td->cur_depth++;
1354
1355         ret = td_io_queue(td, io_u);
1356
1357         dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1358
1359         io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1360
1361         if (ret == FIO_Q_QUEUED)
1362                 ret = io_u_queued_complete(td, 1);
1363
1364         td->cur_depth--;
1365 }
1366
1367 /*
1368  * Entry point for the thread based jobs. The process based jobs end up
1369  * here as well, after a little setup.
1370  */
1371 static void *thread_main(void *data)
1372 {
1373         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1374         struct thread_data *td = data;
1375         struct thread_options *o = &td->o;
1376         pthread_condattr_t attr;
1377         int clear_state;
1378         int ret;
1379
1380         if (!o->use_thread) {
1381                 setsid();
1382                 td->pid = getpid();
1383         } else
1384                 td->pid = gettid();
1385
1386         fio_local_clock_init(o->use_thread);
1387
1388         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1389
1390         if (is_backend)
1391                 fio_server_send_start(td);
1392
1393         INIT_FLIST_HEAD(&td->io_log_list);
1394         INIT_FLIST_HEAD(&td->io_hist_list);
1395         INIT_FLIST_HEAD(&td->verify_list);
1396         INIT_FLIST_HEAD(&td->trim_list);
1397         INIT_FLIST_HEAD(&td->next_rand_list);
1398         pthread_mutex_init(&td->io_u_lock, NULL);
1399         td->io_hist_tree = RB_ROOT;
1400
1401         pthread_condattr_init(&attr);
1402         pthread_cond_init(&td->verify_cond, &attr);
1403         pthread_cond_init(&td->free_cond, &attr);
1404
1405         td_set_runstate(td, TD_INITIALIZED);
1406         dprint(FD_MUTEX, "up startup_mutex\n");
1407         fio_mutex_up(startup_mutex);
1408         dprint(FD_MUTEX, "wait on td->mutex\n");
1409         fio_mutex_down(td->mutex);
1410         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1411
1412         /*
1413          * A new gid requires privilege, so we need to do this before setting
1414          * the uid.
1415          */
1416         if (o->gid != -1U && setgid(o->gid)) {
1417                 td_verror(td, errno, "setgid");
1418                 goto err;
1419         }
1420         if (o->uid != -1U && setuid(o->uid)) {
1421                 td_verror(td, errno, "setuid");
1422                 goto err;
1423         }
1424
1425         /*
1426          * If we have a gettimeofday() thread, make sure we exclude that
1427          * thread from this job
1428          */
1429         if (o->gtod_cpu)
1430                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1431
1432         /*
1433          * Set affinity first, in case it has an impact on the memory
1434          * allocations.
1435          */
1436         if (fio_option_is_set(o, cpumask)) {
1437                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1438                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1439                         if (!ret) {
1440                                 log_err("fio: no CPUs set\n");
1441                                 log_err("fio: Try increasing number of available CPUs\n");
1442                                 td_verror(td, EINVAL, "cpus_split");
1443                                 goto err;
1444                         }
1445                 }
1446                 ret = fio_setaffinity(td->pid, o->cpumask);
1447                 if (ret == -1) {
1448                         td_verror(td, errno, "cpu_set_affinity");
1449                         goto err;
1450                 }
1451         }
1452
1453 #ifdef CONFIG_LIBNUMA
1454         /* numa node setup */
1455         if (fio_option_is_set(o, numa_cpunodes) ||
1456             fio_option_is_set(o, numa_memnodes)) {
1457                 struct bitmask *mask;
1458
1459                 if (numa_available() < 0) {
1460                         td_verror(td, errno, "Does not support NUMA API\n");
1461                         goto err;
1462                 }
1463
1464                 if (fio_option_is_set(o, numa_cpunodes)) {
1465                         mask = numa_parse_nodestring(o->numa_cpunodes);
1466                         ret = numa_run_on_node_mask(mask);
1467                         numa_free_nodemask(mask);
1468                         if (ret == -1) {
1469                                 td_verror(td, errno, \
1470                                         "numa_run_on_node_mask failed\n");
1471                                 goto err;
1472                         }
1473                 }
1474
1475                 if (fio_option_is_set(o, numa_memnodes)) {
1476                         mask = NULL;
1477                         if (o->numa_memnodes)
1478                                 mask = numa_parse_nodestring(o->numa_memnodes);
1479
1480                         switch (o->numa_mem_mode) {
1481                         case MPOL_INTERLEAVE:
1482                                 numa_set_interleave_mask(mask);
1483                                 break;
1484                         case MPOL_BIND:
1485                                 numa_set_membind(mask);
1486                                 break;
1487                         case MPOL_LOCAL:
1488                                 numa_set_localalloc();
1489                                 break;
1490                         case MPOL_PREFERRED:
1491                                 numa_set_preferred(o->numa_mem_prefer_node);
1492                                 break;
1493                         case MPOL_DEFAULT:
1494                         default:
1495                                 break;
1496                         }
1497
1498                         if (mask)
1499                                 numa_free_nodemask(mask);
1500
1501                 }
1502         }
1503 #endif
1504
1505         if (fio_pin_memory(td))
1506                 goto err;
1507
1508         /*
1509          * May alter parameters that init_io_u() will use, so we need to
1510          * do this first.
1511          */
1512         if (init_iolog(td))
1513                 goto err;
1514
1515         if (init_io_u(td))
1516                 goto err;
1517
1518         if (o->verify_async && verify_async_init(td))
1519                 goto err;
1520
1521         if (fio_option_is_set(o, ioprio) ||
1522             fio_option_is_set(o, ioprio_class)) {
1523                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1524                 if (ret == -1) {
1525                         td_verror(td, errno, "ioprio_set");
1526                         goto err;
1527                 }
1528         }
1529
1530         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1531                 goto err;
1532
1533         errno = 0;
1534         if (nice(o->nice) == -1 && errno != 0) {
1535                 td_verror(td, errno, "nice");
1536                 goto err;
1537         }
1538
1539         if (o->ioscheduler && switch_ioscheduler(td))
1540                 goto err;
1541
1542         if (!o->create_serialize && setup_files(td))
1543                 goto err;
1544
1545         if (td_io_init(td))
1546                 goto err;
1547
1548         if (init_random_map(td))
1549                 goto err;
1550
1551         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1552                 goto err;
1553
1554         if (o->pre_read) {
1555                 if (pre_read_files(td) < 0)
1556                         goto err;
1557         }
1558
1559         if (td->flags & TD_F_COMPRESS_LOG)
1560                 tp_init(&td->tp_data);
1561
1562         fio_verify_init(td);
1563
1564         if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1565             workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1566                 goto err;
1567
1568         fio_gettime(&td->epoch, NULL);
1569         fio_getrusage(&td->ru_start);
1570         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1571         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1572
1573         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1574                         o->ratemin[DDIR_TRIM]) {
1575                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1576                                         sizeof(td->bw_sample_time));
1577                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1578                                         sizeof(td->bw_sample_time));
1579                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1580                                         sizeof(td->bw_sample_time));
1581         }
1582
1583         clear_state = 0;
1584         while (keep_running(td)) {
1585                 uint64_t verify_bytes;
1586
1587                 fio_gettime(&td->start, NULL);
1588                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1589
1590                 if (clear_state)
1591                         clear_io_state(td, 0);
1592
1593                 prune_io_piece_log(td);
1594
1595                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1596                         verify_bytes = do_dry_run(td);
1597                 else
1598                         verify_bytes = do_io(td);
1599
1600                 clear_state = 1;
1601
1602                 /*
1603                  * Make sure we've successfully updated the rusage stats
1604                  * before waiting on the stat mutex. Otherwise we could have
1605                  * the stat thread holding stat mutex and waiting for
1606                  * the rusage_sem, which would never get upped because
1607                  * this thread is waiting for the stat mutex.
1608                  */
1609                 check_update_rusage(td);
1610
1611                 fio_mutex_down(stat_mutex);
1612                 if (td_read(td) && td->io_bytes[DDIR_READ])
1613                         update_runtime(td, elapsed_us, DDIR_READ);
1614                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1615                         update_runtime(td, elapsed_us, DDIR_WRITE);
1616                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1617                         update_runtime(td, elapsed_us, DDIR_TRIM);
1618                 fio_gettime(&td->start, NULL);
1619                 fio_mutex_up(stat_mutex);
1620
1621                 if (td->error || td->terminate)
1622                         break;
1623
1624                 if (!o->do_verify ||
1625                     o->verify == VERIFY_NONE ||
1626                     (td->io_ops->flags & FIO_UNIDIR))
1627                         continue;
1628
1629                 clear_io_state(td, 0);
1630
1631                 fio_gettime(&td->start, NULL);
1632
1633                 do_verify(td, verify_bytes);
1634
1635                 /*
1636                  * See comment further up for why this is done here.
1637                  */
1638                 check_update_rusage(td);
1639
1640                 fio_mutex_down(stat_mutex);
1641                 update_runtime(td, elapsed_us, DDIR_READ);
1642                 fio_gettime(&td->start, NULL);
1643                 fio_mutex_up(stat_mutex);
1644
1645                 if (td->error || td->terminate)
1646                         break;
1647         }
1648
1649         update_rusage_stat(td);
1650         td->ts.total_run_time = mtime_since_now(&td->epoch);
1651         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1652         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1653         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1654
1655         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1656             (td->o.verify != VERIFY_NONE && td_write(td)))
1657                 verify_save_state(td->thread_number);
1658
1659         fio_unpin_memory(td);
1660
1661         fio_writeout_logs(td);
1662
1663         if (o->io_submit_mode == IO_MODE_OFFLOAD)
1664                 workqueue_exit(&td->io_wq);
1665
1666         if (td->flags & TD_F_COMPRESS_LOG)
1667                 tp_exit(&td->tp_data);
1668
1669         if (o->exec_postrun)
1670                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1671
1672         if (exitall_on_terminate)
1673                 fio_terminate_threads(td->groupid);
1674
1675 err:
1676         if (td->error)
1677                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1678                                                         td->verror);
1679
1680         if (o->verify_async)
1681                 verify_async_exit(td);
1682
1683         close_and_free_files(td);
1684         cleanup_io_u(td);
1685         close_ioengine(td);
1686         cgroup_shutdown(td, &cgroup_mnt);
1687         verify_free_state(td);
1688
1689         if (fio_option_is_set(o, cpumask)) {
1690                 ret = fio_cpuset_exit(&o->cpumask);
1691                 if (ret)
1692                         td_verror(td, ret, "fio_cpuset_exit");
1693         }
1694
1695         /*
1696          * do this very late, it will log file closing as well
1697          */
1698         if (o->write_iolog_file)
1699                 write_iolog_close(td);
1700
1701         fio_mutex_remove(td->mutex);
1702         td->mutex = NULL;
1703
1704         td_set_runstate(td, TD_EXITED);
1705
1706         /*
1707          * Do this last after setting our runstate to exited, so we
1708          * know that the stat thread is signaled.
1709          */
1710         check_update_rusage(td);
1711
1712         return (void *) (uintptr_t) td->error;
1713 }
1714
1715
1716 /*
1717  * We cannot pass the td data into a forked process, so attach the td and
1718  * pass it to the thread worker.
1719  */
1720 static int fork_main(int shmid, int offset)
1721 {
1722         struct thread_data *td;
1723         void *data, *ret;
1724
1725 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1726         data = shmat(shmid, NULL, 0);
1727         if (data == (void *) -1) {
1728                 int __err = errno;
1729
1730                 perror("shmat");
1731                 return __err;
1732         }
1733 #else
1734         /*
1735          * HP-UX inherits shm mappings?
1736          */
1737         data = threads;
1738 #endif
1739
1740         td = data + offset * sizeof(struct thread_data);
1741         ret = thread_main(td);
1742         shmdt(data);
1743         return (int) (uintptr_t) ret;
1744 }
1745
1746 static void dump_td_info(struct thread_data *td)
1747 {
1748         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1749                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1750                         (unsigned long) time_since_now(&td->terminate_time));
1751 }
1752
1753 /*
1754  * Run over the job map and reap the threads that have exited, if any.
1755  */
1756 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1757                          unsigned int *m_rate)
1758 {
1759         struct thread_data *td;
1760         unsigned int cputhreads, realthreads, pending;
1761         int i, status, ret;
1762
1763         /*
1764          * reap exited threads (TD_EXITED -> TD_REAPED)
1765          */
1766         realthreads = pending = cputhreads = 0;
1767         for_each_td(td, i) {
1768                 int flags = 0;
1769
1770                 /*
1771                  * ->io_ops is NULL for a thread that has closed its
1772                  * io engine
1773                  */
1774                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1775                         cputhreads++;
1776                 else
1777                         realthreads++;
1778
1779                 if (!td->pid) {
1780                         pending++;
1781                         continue;
1782                 }
1783                 if (td->runstate == TD_REAPED)
1784                         continue;
1785                 if (td->o.use_thread) {
1786                         if (td->runstate == TD_EXITED) {
1787                                 td_set_runstate(td, TD_REAPED);
1788                                 goto reaped;
1789                         }
1790                         continue;
1791                 }
1792
1793                 flags = WNOHANG;
1794                 if (td->runstate == TD_EXITED)
1795                         flags = 0;
1796
1797                 /*
1798                  * check if someone quit or got killed in an unusual way
1799                  */
1800                 ret = waitpid(td->pid, &status, flags);
1801                 if (ret < 0) {
1802                         if (errno == ECHILD) {
1803                                 log_err("fio: pid=%d disappeared %d\n",
1804                                                 (int) td->pid, td->runstate);
1805                                 td->sig = ECHILD;
1806                                 td_set_runstate(td, TD_REAPED);
1807                                 goto reaped;
1808                         }
1809                         perror("waitpid");
1810                 } else if (ret == td->pid) {
1811                         if (WIFSIGNALED(status)) {
1812                                 int sig = WTERMSIG(status);
1813
1814                                 if (sig != SIGTERM && sig != SIGUSR2)
1815                                         log_err("fio: pid=%d, got signal=%d\n",
1816                                                         (int) td->pid, sig);
1817                                 td->sig = sig;
1818                                 td_set_runstate(td, TD_REAPED);
1819                                 goto reaped;
1820                         }
1821                         if (WIFEXITED(status)) {
1822                                 if (WEXITSTATUS(status) && !td->error)
1823                                         td->error = WEXITSTATUS(status);
1824
1825                                 td_set_runstate(td, TD_REAPED);
1826                                 goto reaped;
1827                         }
1828                 }
1829
1830                 /*
1831                  * If the job is stuck, do a forceful timeout of it and
1832                  * move on.
1833                  */
1834                 if (td->terminate &&
1835                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1836                         dump_td_info(td);
1837                         td_set_runstate(td, TD_REAPED);
1838                         goto reaped;
1839                 }
1840
1841                 /*
1842                  * thread is not dead, continue
1843                  */
1844                 pending++;
1845                 continue;
1846 reaped:
1847                 (*nr_running)--;
1848                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1849                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1850                 if (!td->pid)
1851                         pending--;
1852
1853                 if (td->error)
1854                         exit_value++;
1855
1856                 done_secs += mtime_since_now(&td->epoch) / 1000;
1857                 profile_td_exit(td);
1858         }
1859
1860         if (*nr_running == cputhreads && !pending && realthreads)
1861                 fio_terminate_threads(TERMINATE_ALL);
1862 }
1863
1864 static int __check_trigger_file(void)
1865 {
1866         struct stat sb;
1867
1868         if (!trigger_file)
1869                 return 0;
1870
1871         if (stat(trigger_file, &sb))
1872                 return 0;
1873
1874         if (unlink(trigger_file) < 0)
1875                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1876                                                         strerror(errno));
1877
1878         return 1;
1879 }
1880
1881 static int trigger_timedout(void)
1882 {
1883         if (trigger_timeout)
1884                 return time_since_genesis() >= trigger_timeout;
1885
1886         return 0;
1887 }
1888
1889 void exec_trigger(const char *cmd)
1890 {
1891         int ret;
1892
1893         if (!cmd)
1894                 return;
1895
1896         ret = system(cmd);
1897         if (ret == -1)
1898                 log_err("fio: failed executing %s trigger\n", cmd);
1899 }
1900
1901 void check_trigger_file(void)
1902 {
1903         if (__check_trigger_file() || trigger_timedout()) {
1904                 if (nr_clients)
1905                         fio_clients_send_trigger(trigger_remote_cmd);
1906                 else {
1907                         verify_save_state(IO_LIST_ALL);
1908                         fio_terminate_threads(TERMINATE_ALL);
1909                         exec_trigger(trigger_cmd);
1910                 }
1911         }
1912 }
1913
1914 static int fio_verify_load_state(struct thread_data *td)
1915 {
1916         int ret;
1917
1918         if (!td->o.verify_state)
1919                 return 0;
1920
1921         if (is_backend) {
1922                 void *data;
1923                 int ver;
1924
1925                 ret = fio_server_get_verify_state(td->o.name,
1926                                         td->thread_number - 1, &data, &ver);
1927                 if (!ret)
1928                         verify_convert_assign_state(td, data, ver);
1929         } else
1930                 ret = verify_load_state(td, "local");
1931
1932         return ret;
1933 }
1934
1935 static void do_usleep(unsigned int usecs)
1936 {
1937         check_for_running_stats();
1938         check_trigger_file();
1939         usleep(usecs);
1940 }
1941
1942 static int check_mount_writes(struct thread_data *td)
1943 {
1944         struct fio_file *f;
1945         unsigned int i;
1946
1947         if (!td_write(td) || td->o.allow_mounted_write)
1948                 return 0;
1949
1950         for_each_file(td, f, i) {
1951                 if (f->filetype != FIO_TYPE_BD)
1952                         continue;
1953                 if (device_is_mounted(f->file_name))
1954                         goto mounted;
1955         }
1956
1957         return 0;
1958 mounted:
1959         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1960         return 1;
1961 }
1962
1963 /*
1964  * Main function for kicking off and reaping jobs, as needed.
1965  */
1966 static void run_threads(void)
1967 {
1968         struct thread_data *td;
1969         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1970         uint64_t spent;
1971
1972         if (fio_gtod_offload && fio_start_gtod_thread())
1973                 return;
1974
1975         fio_idle_prof_init();
1976
1977         set_sig_handlers();
1978
1979         nr_thread = nr_process = 0;
1980         for_each_td(td, i) {
1981                 if (check_mount_writes(td))
1982                         return;
1983                 if (td->o.use_thread)
1984                         nr_thread++;
1985                 else
1986                         nr_process++;
1987         }
1988
1989         if (output_format & FIO_OUTPUT_NORMAL) {
1990                 log_info("Starting ");
1991                 if (nr_thread)
1992                         log_info("%d thread%s", nr_thread,
1993                                                 nr_thread > 1 ? "s" : "");
1994                 if (nr_process) {
1995                         if (nr_thread)
1996                                 log_info(" and ");
1997                         log_info("%d process%s", nr_process,
1998                                                 nr_process > 1 ? "es" : "");
1999                 }
2000                 log_info("\n");
2001                 log_info_flush();
2002         }
2003
2004         todo = thread_number;
2005         nr_running = 0;
2006         nr_started = 0;
2007         m_rate = t_rate = 0;
2008
2009         for_each_td(td, i) {
2010                 print_status_init(td->thread_number - 1);
2011
2012                 if (!td->o.create_serialize)
2013                         continue;
2014
2015                 if (fio_verify_load_state(td))
2016                         goto reap;
2017
2018                 /*
2019                  * do file setup here so it happens sequentially,
2020                  * we don't want X number of threads getting their
2021                  * client data interspersed on disk
2022                  */
2023                 if (setup_files(td)) {
2024 reap:
2025                         exit_value++;
2026                         if (td->error)
2027                                 log_err("fio: pid=%d, err=%d/%s\n",
2028                                         (int) td->pid, td->error, td->verror);
2029                         td_set_runstate(td, TD_REAPED);
2030                         todo--;
2031                 } else {
2032                         struct fio_file *f;
2033                         unsigned int j;
2034
2035                         /*
2036                          * for sharing to work, each job must always open
2037                          * its own files. so close them, if we opened them
2038                          * for creation
2039                          */
2040                         for_each_file(td, f, j) {
2041                                 if (fio_file_open(f))
2042                                         td_io_close_file(td, f);
2043                         }
2044                 }
2045         }
2046
2047         /* start idle threads before io threads start to run */
2048         fio_idle_prof_start();
2049
2050         set_genesis_time();
2051
2052         while (todo) {
2053                 struct thread_data *map[REAL_MAX_JOBS];
2054                 struct timeval this_start;
2055                 int this_jobs = 0, left;
2056
2057                 /*
2058                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2059                  */
2060                 for_each_td(td, i) {
2061                         if (td->runstate != TD_NOT_CREATED)
2062                                 continue;
2063
2064                         /*
2065                          * never got a chance to start, killed by other
2066                          * thread for some reason
2067                          */
2068                         if (td->terminate) {
2069                                 todo--;
2070                                 continue;
2071                         }
2072
2073                         if (td->o.start_delay) {
2074                                 spent = utime_since_genesis();
2075
2076                                 if (td->o.start_delay > spent)
2077                                         continue;
2078                         }
2079
2080                         if (td->o.stonewall && (nr_started || nr_running)) {
2081                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2082                                                         td->o.name);
2083                                 break;
2084                         }
2085
2086                         init_disk_util(td);
2087
2088                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2089                         td->update_rusage = 0;
2090
2091                         /*
2092                          * Set state to created. Thread will transition
2093                          * to TD_INITIALIZED when it's done setting up.
2094                          */
2095                         td_set_runstate(td, TD_CREATED);
2096                         map[this_jobs++] = td;
2097                         nr_started++;
2098
2099                         if (td->o.use_thread) {
2100                                 int ret;
2101
2102                                 dprint(FD_PROCESS, "will pthread_create\n");
2103                                 ret = pthread_create(&td->thread, NULL,
2104                                                         thread_main, td);
2105                                 if (ret) {
2106                                         log_err("pthread_create: %s\n",
2107                                                         strerror(ret));
2108                                         nr_started--;
2109                                         break;
2110                                 }
2111                                 ret = pthread_detach(td->thread);
2112                                 if (ret)
2113                                         log_err("pthread_detach: %s",
2114                                                         strerror(ret));
2115                         } else {
2116                                 pid_t pid;
2117                                 dprint(FD_PROCESS, "will fork\n");
2118                                 pid = fork();
2119                                 if (!pid) {
2120                                         int ret = fork_main(shm_id, i);
2121
2122                                         _exit(ret);
2123                                 } else if (i == fio_debug_jobno)
2124                                         *fio_debug_jobp = pid;
2125                         }
2126                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2127                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2128                                 log_err("fio: job startup hung? exiting.\n");
2129                                 fio_terminate_threads(TERMINATE_ALL);
2130                                 fio_abort = 1;
2131                                 nr_started--;
2132                                 break;
2133                         }
2134                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2135                 }
2136
2137                 /*
2138                  * Wait for the started threads to transition to
2139                  * TD_INITIALIZED.
2140                  */
2141                 fio_gettime(&this_start, NULL);
2142                 left = this_jobs;
2143                 while (left && !fio_abort) {
2144                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2145                                 break;
2146
2147                         do_usleep(100000);
2148
2149                         for (i = 0; i < this_jobs; i++) {
2150                                 td = map[i];
2151                                 if (!td)
2152                                         continue;
2153                                 if (td->runstate == TD_INITIALIZED) {
2154                                         map[i] = NULL;
2155                                         left--;
2156                                 } else if (td->runstate >= TD_EXITED) {
2157                                         map[i] = NULL;
2158                                         left--;
2159                                         todo--;
2160                                         nr_running++; /* work-around... */
2161                                 }
2162                         }
2163                 }
2164
2165                 if (left) {
2166                         log_err("fio: %d job%s failed to start\n", left,
2167                                         left > 1 ? "s" : "");
2168                         for (i = 0; i < this_jobs; i++) {
2169                                 td = map[i];
2170                                 if (!td)
2171                                         continue;
2172                                 kill(td->pid, SIGTERM);
2173                         }
2174                         break;
2175                 }
2176
2177                 /*
2178                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2179                  */
2180                 for_each_td(td, i) {
2181                         if (td->runstate != TD_INITIALIZED)
2182                                 continue;
2183
2184                         if (in_ramp_time(td))
2185                                 td_set_runstate(td, TD_RAMP);
2186                         else
2187                                 td_set_runstate(td, TD_RUNNING);
2188                         nr_running++;
2189                         nr_started--;
2190                         m_rate += ddir_rw_sum(td->o.ratemin);
2191                         t_rate += ddir_rw_sum(td->o.rate);
2192                         todo--;
2193                         fio_mutex_up(td->mutex);
2194                 }
2195
2196                 reap_threads(&nr_running, &t_rate, &m_rate);
2197
2198                 if (todo)
2199                         do_usleep(100000);
2200         }
2201
2202         while (nr_running) {
2203                 reap_threads(&nr_running, &t_rate, &m_rate);
2204                 do_usleep(10000);
2205         }
2206
2207         fio_idle_prof_stop();
2208
2209         update_io_ticks();
2210 }
2211
2212 static void wait_for_helper_thread_exit(void)
2213 {
2214         void *ret;
2215
2216         helper_exit = 1;
2217         pthread_cond_signal(&helper_cond);
2218         pthread_join(helper_thread, &ret);
2219 }
2220
2221 static void free_disk_util(void)
2222 {
2223         disk_util_prune_entries();
2224
2225         pthread_cond_destroy(&helper_cond);
2226 }
2227
2228 static void *helper_thread_main(void *data)
2229 {
2230         int ret = 0;
2231
2232         fio_mutex_up(startup_mutex);
2233
2234         while (!ret) {
2235                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2236                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2237                 struct timespec ts;
2238                 struct timeval tv;
2239
2240                 gettimeofday(&tv, NULL);
2241                 ts.tv_sec = tv.tv_sec + sec;
2242                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2243
2244                 if (ts.tv_nsec >= 1000000000ULL) {
2245                         ts.tv_nsec -= 1000000000ULL;
2246                         ts.tv_sec++;
2247                 }
2248
2249                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2250
2251                 ret = update_io_ticks();
2252
2253                 if (helper_do_stat) {
2254                         helper_do_stat = 0;
2255                         __show_running_run_stats();
2256                 }
2257
2258                 if (!is_backend)
2259                         print_thread_status();
2260         }
2261
2262         return NULL;
2263 }
2264
2265 static int create_helper_thread(void)
2266 {
2267         int ret;
2268
2269         setup_disk_util();
2270
2271         pthread_cond_init(&helper_cond, NULL);
2272         pthread_mutex_init(&helper_lock, NULL);
2273
2274         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2275         if (ret) {
2276                 log_err("Can't create helper thread: %s\n", strerror(ret));
2277                 return 1;
2278         }
2279
2280         dprint(FD_MUTEX, "wait on startup_mutex\n");
2281         fio_mutex_down(startup_mutex);
2282         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2283         return 0;
2284 }
2285
2286 int fio_backend(void)
2287 {
2288         struct thread_data *td;
2289         int i;
2290
2291         if (exec_profile) {
2292                 if (load_profile(exec_profile))
2293                         return 1;
2294                 free(exec_profile);
2295                 exec_profile = NULL;
2296         }
2297         if (!thread_number)
2298                 return 0;
2299
2300         if (write_bw_log) {
2301                 struct log_params p = {
2302                         .log_type = IO_LOG_TYPE_BW,
2303                 };
2304
2305                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2306                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2307                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2308         }
2309
2310         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2311         if (startup_mutex == NULL)
2312                 return 1;
2313
2314         set_genesis_time();
2315         stat_init();
2316         create_helper_thread();
2317
2318         cgroup_list = smalloc(sizeof(*cgroup_list));
2319         INIT_FLIST_HEAD(cgroup_list);
2320
2321         run_threads();
2322
2323         wait_for_helper_thread_exit();
2324
2325         if (!fio_abort) {
2326                 __show_run_stats();
2327                 if (write_bw_log) {
2328                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2329                                 struct io_log *log = agg_io_log[i];
2330
2331                                 flush_log(log, 0);
2332                                 free_log(log);
2333                         }
2334                 }
2335         }
2336
2337         for_each_td(td, i) {
2338                 fio_options_free(td);
2339                 if (td->rusage_sem) {
2340                         fio_mutex_remove(td->rusage_sem);
2341                         td->rusage_sem = NULL;
2342                 }
2343         }
2344
2345         free_disk_util();
2346         cgroup_kill(cgroup_list);
2347         sfree(cgroup_list);
2348         sfree(cgroup_mnt);
2349
2350         fio_mutex_remove(startup_mutex);
2351         stat_exit();
2352         return exit_value;
2353 }