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
  39 #include "fio.h"
  40 #ifndef FIO_NO_HAVE_SHM_H
  41 #include <sys/shm.h>
  42 #endif
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53 #include "lib/getrusage.h"
  54 #include "idletime.h"
  55 #include "err.h"
  56 #include "lib/tp.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59
  60 static pthread_t helper_thread;
  61 static pthread_mutex_t helper_lock;
  62 pthread_cond_t helper_cond;
  63 int helper_do_stat = 0;
  64
  65 static struct fio_mutex *startup_mutex;
  66 static struct flist_head *cgroup_list;
  67 static char *cgroup_mnt;
  68 static int exit_value;
  69 static volatile int fio_abort;
  70 static unsigned int nr_process = 0;
  71 static unsigned int nr_thread = 0;
  72
  73 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  74
  75 int groupid = 0;
  76 unsigned int thread_number = 0;
  77 unsigned int stat_number = 0;
  78 int shm_id = 0;
  79 int temp_stall_ts;
  80 unsigned long done_secs = 0;
  81 volatile int helper_exit = 0;
  82
  83 #define PAGE_ALIGN(buf) \
  84         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  85
  86 #define JOB_START_TIMEOUT       (5 * 1000)
  87
  88 static void sig_int(int sig)
  89 {
  90         if (threads) {
  91                 if (is_backend)
  92                         fio_server_got_signal(sig);
  93                 else {
  94                         log_info("\nfio: terminating on signal %d\n", sig);
  95                         log_info_flush();
  96                         exit_value = 128;
  97                 }
  98
  99                 fio_terminate_threads(TERMINATE_ALL);
 100         }
 101 }
 102
 103 static void sig_show_status(int sig)
 104 {
 105         show_running_run_stats();
 106 }
 107
 108 static void set_sig_handlers(void)
 109 {
 110         struct sigaction act;
 111
 112         memset(&act, 0, sizeof(act));
 113         act.sa_handler = sig_int;
 114         act.sa_flags = SA_RESTART;
 115         sigaction(SIGINT, &act, NULL);
 116
 117         memset(&act, 0, sizeof(act));
 118         act.sa_handler = sig_int;
 119         act.sa_flags = SA_RESTART;
 120         sigaction(SIGTERM, &act, NULL);
 121
 122 /* Windows uses SIGBREAK as a quit signal from other applications */
 123 #ifdef WIN32
 124         memset(&act, 0, sizeof(act));
 125         act.sa_handler = sig_int;
 126         act.sa_flags = SA_RESTART;
 127         sigaction(SIGBREAK, &act, NULL);
 128 #endif
 129
 130         memset(&act, 0, sizeof(act));
 131         act.sa_handler = sig_show_status;
 132         act.sa_flags = SA_RESTART;
 133         sigaction(SIGUSR1, &act, NULL);
 134
 135         if (is_backend) {
 136                 memset(&act, 0, sizeof(act));
 137                 act.sa_handler = sig_int;
 138                 act.sa_flags = SA_RESTART;
 139                 sigaction(SIGPIPE, &act, NULL);
 140         }
 141 }
 142
 143 /*
 144  * Check if we are above the minimum rate given.
 145  */
 146 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 147                             enum fio_ddir ddir)
 148 {
 149         unsigned long long bytes = 0;
 150         unsigned long iops = 0;
 151         unsigned long spent;
 152         unsigned long rate;
 153         unsigned int ratemin = 0;
 154         unsigned int rate_iops = 0;
 155         unsigned int rate_iops_min = 0;
 156
 157         assert(ddir_rw(ddir));
 158
 159         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 160                 return 0;
 161
 162         /*
 163          * allow a 2 second settle period in the beginning
 164          */
 165         if (mtime_since(&td->start, now) < 2000)
 166                 return 0;
 167
 168         iops += td->this_io_blocks[ddir];
 169         bytes += td->this_io_bytes[ddir];
 170         ratemin += td->o.ratemin[ddir];
 171         rate_iops += td->o.rate_iops[ddir];
 172         rate_iops_min += td->o.rate_iops_min[ddir];
 173
 174         /*
 175          * if rate blocks is set, sample is running
 176          */
 177         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 178                 spent = mtime_since(&td->lastrate[ddir], now);
 179                 if (spent < td->o.ratecycle)
 180                         return 0;
 181
 182                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 183                         /*
 184                          * check bandwidth specified rate
 185                          */
 186                         if (bytes < td->rate_bytes[ddir]) {
 187                                 log_err("%s: min rate %u not met\n", td->o.name,
 188                                                                 ratemin);
 189                                 return 1;
 190                         } else {
 191                                 if (spent)
 192                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 193                                 else
 194                                         rate = 0;
 195
 196                                 if (rate < ratemin ||
 197                                     bytes < td->rate_bytes[ddir]) {
 198                                         log_err("%s: min rate %u not met, got"
 199                                                 " %luKB/sec\n", td->o.name,
 200                                                         ratemin, rate);
 201                                         return 1;
 202                                 }
 203                         }
 204                 } else {
 205                         /*
 206                          * checks iops specified rate
 207                          */
 208                         if (iops < rate_iops) {
 209                                 log_err("%s: min iops rate %u not met\n",
 210                                                 td->o.name, rate_iops);
 211                                 return 1;
 212                         } else {
 213                                 if (spent)
 214                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 215                                 else
 216                                         rate = 0;
 217
 218                                 if (rate < rate_iops_min ||
 219                                     iops < td->rate_blocks[ddir]) {
 220                                         log_err("%s: min iops rate %u not met,"
 221                                                 " got %lu\n", td->o.name,
 222                                                         rate_iops_min, rate);
 223                                         return 1;
 224                                 }
 225                         }
 226                 }
 227         }
 228
 229         td->rate_bytes[ddir] = bytes;
 230         td->rate_blocks[ddir] = iops;
 231         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 232         return 0;
 233 }
 234
 235 static int check_min_rate(struct thread_data *td, struct timeval *now)
 236 {
 237         int ret = 0;
 238
 239         if (td->bytes_done[DDIR_READ])
 240                 ret |= __check_min_rate(td, now, DDIR_READ);
 241         if (td->bytes_done[DDIR_WRITE])
 242                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 243         if (td->bytes_done[DDIR_TRIM])
 244                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 245
 246         return ret;
 247 }
 248
 249 /*
 250  * When job exits, we can cancel the in-flight IO if we are using async
 251  * io. Attempt to do so.
 252  */
 253 static void cleanup_pending_aio(struct thread_data *td)
 254 {
 255         int r;
 256
 257         /*
 258          * get immediately available events, if any
 259          */
 260         r = io_u_queued_complete(td, 0);
 261         if (r < 0)
 262                 return;
 263
 264         /*
 265          * now cancel remaining active events
 266          */
 267         if (td->io_ops->cancel) {
 268                 struct io_u *io_u;
 269                 int i;
 270
 271                 io_u_qiter(&td->io_u_all, io_u, i) {
 272                         if (io_u->flags & IO_U_F_FLIGHT) {
 273                                 r = td->io_ops->cancel(td, io_u);
 274                                 if (!r)
 275                                         put_io_u(td, io_u);
 276                         }
 277                 }
 278         }
 279
 280         if (td->cur_depth)
 281                 r = io_u_queued_complete(td, td->cur_depth);
 282 }
 283
 284 /*
 285  * Helper to handle the final sync of a file. Works just like the normal
 286  * io path, just does everything sync.
 287  */
 288 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 289 {
 290         struct io_u *io_u = __get_io_u(td);
 291         int ret;
 292
 293         if (!io_u)
 294                 return 1;
 295
 296         io_u->ddir = DDIR_SYNC;
 297         io_u->file = f;
 298
 299         if (td_io_prep(td, io_u)) {
 300                 put_io_u(td, io_u);
 301                 return 1;
 302         }
 303
 304 requeue:
 305         ret = td_io_queue(td, io_u);
 306         if (ret < 0) {
 307                 td_verror(td, io_u->error, "td_io_queue");
 308                 put_io_u(td, io_u);
 309                 return 1;
 310         } else if (ret == FIO_Q_QUEUED) {
 311                 if (io_u_queued_complete(td, 1) < 0)
 312                         return 1;
 313         } else if (ret == FIO_Q_COMPLETED) {
 314                 if (io_u->error) {
 315                         td_verror(td, io_u->error, "td_io_queue");
 316                         return 1;
 317                 }
 318
 319                 if (io_u_sync_complete(td, io_u) < 0)
 320                         return 1;
 321         } else if (ret == FIO_Q_BUSY) {
 322                 if (td_io_commit(td))
 323                         return 1;
 324                 goto requeue;
 325         }
 326
 327         return 0;
 328 }
 329
 330 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 331 {
 332         int ret;
 333
 334         if (fio_file_open(f))
 335                 return fio_io_sync(td, f);
 336
 337         if (td_io_open_file(td, f))
 338                 return 1;
 339
 340         ret = fio_io_sync(td, f);
 341         td_io_close_file(td, f);
 342         return ret;
 343 }
 344
 345 static inline void __update_tv_cache(struct thread_data *td)
 346 {
 347         fio_gettime(&td->tv_cache, NULL);
 348 }
 349
 350 static inline void update_tv_cache(struct thread_data *td)
 351 {
 352         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 353                 __update_tv_cache(td);
 354 }
 355
 356 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 357 {
 358         if (in_ramp_time(td))
 359                 return 0;
 360         if (!td->o.timeout)
 361                 return 0;
 362         if (utime_since(&td->epoch, t) >= td->o.timeout)
 363                 return 1;
 364
 365         return 0;
 366 }
 367
 368 /*
 369  * We need to update the runtime consistently in ms, but keep a running
 370  * tally of the current elapsed time in microseconds for sub millisecond
 371  * updates.
 372  */
 373 static inline void update_runtime(struct thread_data *td,
 374                                   unsigned long long *elapsed_us,
 375                                   const enum fio_ddir ddir)
 376 {
 377         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 378                 return;
 379
 380         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 381         elapsed_us[ddir] += utime_since_now(&td->start);
 382         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 383 }
 384
 385 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 386                                int *retptr)
 387 {
 388         int ret = *retptr;
 389
 390         if (ret < 0 || td->error) {
 391                 int err = td->error;
 392                 enum error_type_bit eb;
 393
 394                 if (ret < 0)
 395                         err = -ret;
 396
 397                 eb = td_error_type(ddir, err);
 398                 if (!(td->o.continue_on_error & (1 << eb)))
 399                         return 1;
 400
 401                 if (td_non_fatal_error(td, eb, err)) {
 402                         /*
 403                          * Continue with the I/Os in case of
 404                          * a non fatal error.
 405                          */
 406                         update_error_count(td, err);
 407                         td_clear_error(td);
 408                         *retptr = 0;
 409                         return 0;
 410                 } else if (td->o.fill_device && err == ENOSPC) {
 411                         /*
 412                          * We expect to hit this error if
 413                          * fill_device option is set.
 414                          */
 415                         td_clear_error(td);
 416                         fio_mark_td_terminate(td);
 417                         return 1;
 418                 } else {
 419                         /*
 420                          * Stop the I/O in case of a fatal
 421                          * error.
 422                          */
 423                         update_error_count(td, err);
 424                         return 1;
 425                 }
 426         }
 427
 428         return 0;
 429 }
 430
 431 static void check_update_rusage(struct thread_data *td)
 432 {
 433         if (td->update_rusage) {
 434                 td->update_rusage = 0;
 435                 update_rusage_stat(td);
 436                 fio_mutex_up(td->rusage_sem);
 437         }
 438 }
 439
 440 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 441 {
 442         const int full = queue_full(td);
 443         int min_evts = 0;
 444         int ret;
 445
 446         /*
 447          * if the queue is full, we MUST reap at least 1 event
 448          */
 449         min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
 450     if ((full && !min_evts) || !td->o.iodepth_batch_complete)
 451                 min_evts = 1;
 452
 453         if (time && (__should_check_rate(td, DDIR_READ) ||
 454             __should_check_rate(td, DDIR_WRITE) ||
 455             __should_check_rate(td, DDIR_TRIM)))
 456                 fio_gettime(time, NULL);
 457
 458         do {
 459                 ret = io_u_queued_complete(td, min_evts);
 460                 if (ret < 0)
 461                         break;
 462         } while (full && (td->cur_depth > td->o.iodepth_low));
 463
 464         return ret;
 465 }
 466
 467 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 468                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 469                    struct timeval *comp_time)
 470 {
 471         int ret2;
 472
 473         switch (*ret) {
 474         case FIO_Q_COMPLETED:
 475                 if (io_u->error) {
 476                         *ret = -io_u->error;
 477                         clear_io_u(td, io_u);
 478                 } else if (io_u->resid) {
 479                         int bytes = io_u->xfer_buflen - io_u->resid;
 480                         struct fio_file *f = io_u->file;
 481
 482                         if (bytes_issued)
 483                                 *bytes_issued += bytes;
 484
 485                         if (!from_verify)
 486                                 trim_io_piece(td, io_u);
 487
 488                         /*
 489                          * zero read, fail
 490                          */
 491                         if (!bytes) {
 492                                 if (!from_verify)
 493                                         unlog_io_piece(td, io_u);
 494                                 td_verror(td, EIO, "full resid");
 495                                 put_io_u(td, io_u);
 496                                 break;
 497                         }
 498
 499                         io_u->xfer_buflen = io_u->resid;
 500                         io_u->xfer_buf += bytes;
 501                         io_u->offset += bytes;
 502
 503                         if (ddir_rw(io_u->ddir))
 504                                 td->ts.short_io_u[io_u->ddir]++;
 505
 506                         f = io_u->file;
 507                         if (io_u->offset == f->real_file_size)
 508                                 goto sync_done;
 509
 510                         requeue_io_u(td, &io_u);
 511                 } else {
 512 sync_done:
 513                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 514                             __should_check_rate(td, DDIR_WRITE) ||
 515                             __should_check_rate(td, DDIR_TRIM)))
 516                                 fio_gettime(comp_time, NULL);
 517
 518                         *ret = io_u_sync_complete(td, io_u);
 519                         if (*ret < 0)
 520                                 break;
 521                 }
 522                 return 0;
 523         case FIO_Q_QUEUED:
 524                 /*
 525                  * if the engine doesn't have a commit hook,
 526                  * the io_u is really queued. if it does have such
 527                  * a hook, it has to call io_u_queued() itself.
 528                  */
 529                 if (td->io_ops->commit == NULL)
 530                         io_u_queued(td, io_u);
 531                 if (bytes_issued)
 532                         *bytes_issued += io_u->xfer_buflen;
 533                 break;
 534         case FIO_Q_BUSY:
 535                 if (!from_verify)
 536                         unlog_io_piece(td, io_u);
 537                 requeue_io_u(td, &io_u);
 538                 ret2 = td_io_commit(td);
 539                 if (ret2 < 0)
 540                         *ret = ret2;
 541                 break;
 542         default:
 543                 assert(ret < 0);
 544                 td_verror(td, -(*ret), "td_io_queue");
 545                 break;
 546         }
 547
 548         if (break_on_this_error(td, ddir, ret))
 549                 return 1;
 550
 551         return 0;
 552 }
 553
 554 /*
 555  * The main verify engine. Runs over the writes we previously submitted,
 556  * reads the blocks back in, and checks the crc/md5 of the data.
 557  */
 558 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 559 {
 560         struct fio_file *f;
 561         struct io_u *io_u;
 562         int ret, min_events;
 563         unsigned int i;
 564
 565         dprint(FD_VERIFY, "starting loop\n");
 566
 567         /*
 568          * sync io first and invalidate cache, to make sure we really
 569          * read from disk.
 570          */
 571         for_each_file(td, f, i) {
 572                 if (!fio_file_open(f))
 573                         continue;
 574                 if (fio_io_sync(td, f))
 575                         break;
 576                 if (file_invalidate_cache(td, f))
 577                         break;
 578         }
 579
 580         check_update_rusage(td);
 581
 582         if (td->error)
 583                 return;
 584
 585         td_set_runstate(td, TD_VERIFYING);
 586
 587         io_u = NULL;
 588         while (!td->terminate) {
 589                 enum fio_ddir ddir;
 590                 int full;
 591
 592                 update_tv_cache(td);
 593                 check_update_rusage(td);
 594
 595                 if (runtime_exceeded(td, &td->tv_cache)) {
 596                         __update_tv_cache(td);
 597                         if (runtime_exceeded(td, &td->tv_cache)) {
 598                                 fio_mark_td_terminate(td);
 599                                 break;
 600                         }
 601                 }
 602
 603                 if (flow_threshold_exceeded(td))
 604                         continue;
 605
 606                 if (!td->o.experimental_verify) {
 607                         io_u = __get_io_u(td);
 608                         if (!io_u)
 609                                 break;
 610
 611                         if (get_next_verify(td, io_u)) {
 612                                 put_io_u(td, io_u);
 613                                 break;
 614                         }
 615
 616                         if (td_io_prep(td, io_u)) {
 617                                 put_io_u(td, io_u);
 618                                 break;
 619                         }
 620                 } else {
 621                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 622                                 break;
 623
 624                         while ((io_u = get_io_u(td)) != NULL) {
 625                                 if (IS_ERR(io_u)) {
 626                                         io_u = NULL;
 627                                         ret = FIO_Q_BUSY;
 628                                         goto reap;
 629                                 }
 630
 631                                 /*
 632                                  * We are only interested in the places where
 633                                  * we wrote or trimmed IOs. Turn those into
 634                                  * reads for verification purposes.
 635                                  */
 636                                 if (io_u->ddir == DDIR_READ) {
 637                                         /*
 638                                          * Pretend we issued it for rwmix
 639                                          * accounting
 640                                          */
 641                                         td->io_issues[DDIR_READ]++;
 642                                         put_io_u(td, io_u);
 643                                         continue;
 644                                 } else if (io_u->ddir == DDIR_TRIM) {
 645                                         io_u->ddir = DDIR_READ;
 646                                         io_u_set(io_u, IO_U_F_TRIMMED);
 647                                         break;
 648                                 } else if (io_u->ddir == DDIR_WRITE) {
 649                                         io_u->ddir = DDIR_READ;
 650                                         break;
 651                                 } else {
 652                                         put_io_u(td, io_u);
 653                                         continue;
 654                                 }
 655                         }
 656
 657                         if (!io_u)
 658                                 break;
 659                 }
 660
 661                 if (verify_state_should_stop(td, io_u)) {
 662                         put_io_u(td, io_u);
 663                         break;
 664                 }
 665
 666                 if (td->o.verify_async)
 667                         io_u->end_io = verify_io_u_async;
 668                 else
 669                         io_u->end_io = verify_io_u;
 670
 671                 ddir = io_u->ddir;
 672                 if (!td->o.disable_slat)
 673                         fio_gettime(&io_u->start_time, NULL);
 674
 675                 ret = td_io_queue(td, io_u);
 676
 677                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 678                         break;
 679
 680                 /*
 681                  * if we can queue more, do so. but check if there are
 682                  * completed io_u's first. Note that we can get BUSY even
 683                  * without IO queued, if the system is resource starved.
 684                  */
 685 reap:
 686                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 687                 if (full || !td->o.iodepth_batch_complete)
 688                         ret = wait_for_completions(td, NULL);
 689
 690                 if (ret < 0)
 691                         break;
 692         }
 693
 694         check_update_rusage(td);
 695
 696         if (!td->error) {
 697                 min_events = td->cur_depth;
 698
 699                 if (min_events)
 700                         ret = io_u_queued_complete(td, min_events);
 701         } else
 702                 cleanup_pending_aio(td);
 703
 704         td_set_runstate(td, TD_RUNNING);
 705
 706         dprint(FD_VERIFY, "exiting loop\n");
 707 }
 708
 709 static unsigned int exceeds_number_ios(struct thread_data *td)
 710 {
 711         unsigned long long number_ios;
 712
 713         if (!td->o.number_ios)
 714                 return 0;
 715
 716         number_ios = ddir_rw_sum(td->io_blocks);
 717         number_ios += td->io_u_queued + td->io_u_in_flight;
 718
 719         return number_ios >= (td->o.number_ios * td->loops);
 720 }
 721
 722 static int io_issue_bytes_exceeded(struct thread_data *td)
 723 {
 724         unsigned long long bytes, limit;
 725
 726         if (td_rw(td))
 727                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 728         else if (td_write(td))
 729                 bytes = td->io_issue_bytes[DDIR_WRITE];
 730         else if (td_read(td))
 731                 bytes = td->io_issue_bytes[DDIR_READ];
 732         else
 733                 bytes = td->io_issue_bytes[DDIR_TRIM];
 734
 735         if (td->o.io_limit)
 736                 limit = td->o.io_limit;
 737         else
 738                 limit = td->o.size;
 739
 740         limit *= td->loops;
 741         return bytes >= limit || exceeds_number_ios(td);
 742 }
 743
 744 static int io_complete_bytes_exceeded(struct thread_data *td)
 745 {
 746         unsigned long long bytes, limit;
 747
 748         if (td_rw(td))
 749                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 750         else if (td_write(td))
 751                 bytes = td->this_io_bytes[DDIR_WRITE];
 752         else if (td_read(td))
 753                 bytes = td->this_io_bytes[DDIR_READ];
 754         else
 755                 bytes = td->this_io_bytes[DDIR_TRIM];
 756
 757         if (td->o.io_limit)
 758                 limit = td->o.io_limit;
 759         else
 760                 limit = td->o.size;
 761
 762         limit *= td->loops;
 763         return bytes >= limit || exceeds_number_ios(td);
 764 }
 765
 766 /*
 767  * used to calculate the next io time for rate control
 768  *
 769  */
 770 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 771 {
 772         uint64_t secs, remainder, bps, bytes;
 773
 774         assert(!(td->flags & TD_F_CHILD));
 775         bytes = td->rate_io_issue_bytes[ddir];
 776         bps = td->rate_bps[ddir];
 777         if (bps) {
 778                 secs = bytes / bps;
 779                 remainder = bytes % bps;
 780                 return remainder * 1000000 / bps + secs * 1000000;
 781         } else
 782                 return 0;
 783 }
 784
 785 /*
 786  * Main IO worker function. It retrieves io_u's to process and queues
 787  * and reaps them, checking for rate and errors along the way.
 788  *
 789  * Returns number of bytes written and trimmed.
 790  */
 791 static uint64_t do_io(struct thread_data *td)
 792 {
 793         unsigned int i;
 794         int ret = 0;
 795         uint64_t total_bytes, bytes_issued = 0;
 796
 797         if (in_ramp_time(td))
 798                 td_set_runstate(td, TD_RAMP);
 799         else
 800                 td_set_runstate(td, TD_RUNNING);
 801
 802         lat_target_init(td);
 803
 804         total_bytes = td->o.size;
 805         /*
 806         * Allow random overwrite workloads to write up to io_limit
 807         * before starting verification phase as 'size' doesn't apply.
 808         */
 809         if (td_write(td) && td_random(td) && td->o.norandommap)
 810                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 811         /*
 812          * If verify_backlog is enabled, we'll run the verify in this
 813          * handler as well. For that case, we may need up to twice the
 814          * amount of bytes.
 815          */
 816         if (td->o.verify != VERIFY_NONE &&
 817            (td_write(td) && td->o.verify_backlog))
 818                 total_bytes += td->o.size;
 819
 820         /* In trimwrite mode, each byte is trimmed and then written, so
 821          * allow total_bytes to be twice as big */
 822         if (td_trimwrite(td))
 823                 total_bytes += td->total_io_size;
 824
 825         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 826                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 827                 td->o.time_based) {
 828                 struct timeval comp_time;
 829                 struct io_u *io_u;
 830                 int full;
 831                 enum fio_ddir ddir;
 832
 833                 check_update_rusage(td);
 834
 835                 if (td->terminate || td->done)
 836                         break;
 837
 838                 update_tv_cache(td);
 839
 840                 if (runtime_exceeded(td, &td->tv_cache)) {
 841                         __update_tv_cache(td);
 842                         if (runtime_exceeded(td, &td->tv_cache)) {
 843                                 fio_mark_td_terminate(td);
 844                                 break;
 845                         }
 846                 }
 847
 848                 if (flow_threshold_exceeded(td))
 849                         continue;
 850
 851                 if (bytes_issued >= total_bytes)
 852                         break;
 853
 854                 io_u = get_io_u(td);
 855                 if (IS_ERR_OR_NULL(io_u)) {
 856                         int err = PTR_ERR(io_u);
 857
 858                         io_u = NULL;
 859                         if (err == -EBUSY) {
 860                                 ret = FIO_Q_BUSY;
 861                                 goto reap;
 862                         }
 863                         if (td->o.latency_target)
 864                                 goto reap;
 865                         break;
 866                 }
 867
 868                 ddir = io_u->ddir;
 869
 870                 /*
 871                  * Add verification end_io handler if:
 872                  *      - Asked to verify (!td_rw(td))
 873                  *      - Or the io_u is from our verify list (mixed write/ver)
 874                  */
 875                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 876                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 877
 878                         if (!td->o.verify_pattern_bytes) {
 879                                 io_u->rand_seed = __rand(&td->verify_state);
 880                                 if (sizeof(int) != sizeof(long *))
 881                                         io_u->rand_seed *= __rand(&td->verify_state);
 882                         }
 883
 884                         if (verify_state_should_stop(td, io_u)) {
 885                                 put_io_u(td, io_u);
 886                                 break;
 887                         }
 888
 889                         if (td->o.verify_async)
 890                                 io_u->end_io = verify_io_u_async;
 891                         else
 892                                 io_u->end_io = verify_io_u;
 893                         td_set_runstate(td, TD_VERIFYING);
 894                 } else if (in_ramp_time(td))
 895                         td_set_runstate(td, TD_RAMP);
 896                 else
 897                         td_set_runstate(td, TD_RUNNING);
 898
 899                 /*
 900                  * Always log IO before it's issued, so we know the specific
 901                  * order of it. The logged unit will track when the IO has
 902                  * completed.
 903                  */
 904                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 905                     td->o.do_verify &&
 906                     td->o.verify != VERIFY_NONE &&
 907                     !td->o.experimental_verify)
 908                         log_io_piece(td, io_u);
 909
 910                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 911                         if (td->error)
 912                                 break;
 913                         ret = workqueue_enqueue(&td->io_wq, io_u);
 914
 915                         if (should_check_rate(td))
 916                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 917
 918                 } else {
 919                         ret = td_io_queue(td, io_u);
 920
 921                         if (should_check_rate(td))
 922                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 923
 924                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 925                                 break;
 926
 927                         /*
 928                          * See if we need to complete some commands. Note that
 929                          * we can get BUSY even without IO queued, if the
 930                          * system is resource starved.
 931                          */
 932 reap:
 933                         full = queue_full(td) ||
 934                                 (ret == FIO_Q_BUSY && td->cur_depth);
 935                         if (full || !td->o.iodepth_batch_complete)
 936                                 ret = wait_for_completions(td, &comp_time);
 937                 }
 938                 if (ret < 0)
 939                         break;
 940                 if (!ddir_rw_sum(td->bytes_done) &&
 941                     !(td->io_ops->flags & FIO_NOIO))
 942                         continue;
 943
 944                 if (!in_ramp_time(td) && should_check_rate(td)) {
 945                         if (check_min_rate(td, &comp_time)) {
 946                                 if (exitall_on_terminate)
 947                                         fio_terminate_threads(td->groupid);
 948                                 td_verror(td, EIO, "check_min_rate");
 949                                 break;
 950                         }
 951                 }
 952                 if (!in_ramp_time(td) && td->o.latency_target)
 953                         lat_target_check(td);
 954
 955                 if (td->o.thinktime) {
 956                         unsigned long long b;
 957
 958                         b = ddir_rw_sum(td->io_blocks);
 959                         if (!(b % td->o.thinktime_blocks)) {
 960                                 int left;
 961
 962                                 io_u_quiesce(td);
 963
 964                                 if (td->o.thinktime_spin)
 965                                         usec_spin(td->o.thinktime_spin);
 966
 967                                 left = td->o.thinktime - td->o.thinktime_spin;
 968                                 if (left)
 969                                         usec_sleep(td, left);
 970                         }
 971                 }
 972         }
 973
 974         check_update_rusage(td);
 975
 976         if (td->trim_entries)
 977                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 978
 979         if (td->o.fill_device && td->error == ENOSPC) {
 980                 td->error = 0;
 981                 fio_mark_td_terminate(td);
 982         }
 983         if (!td->error) {
 984                 struct fio_file *f;
 985
 986                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 987                         workqueue_flush(&td->io_wq);
 988                         i = 0;
 989                 } else
 990                         i = td->cur_depth;
 991
 992                 if (i) {
 993                         ret = io_u_queued_complete(td, i);
 994                         if (td->o.fill_device && td->error == ENOSPC)
 995                                 td->error = 0;
 996                 }
 997
 998                 if (should_fsync(td) && td->o.end_fsync) {
 999                         td_set_runstate(td, TD_FSYNCING);
1000
1001                         for_each_file(td, f, i) {
1002                                 if (!fio_file_fsync(td, f))
1003                                         continue;
1004
1005                                 log_err("fio: end_fsync failed for file %s\n",
1006                                                                 f->file_name);
1007                         }
1008                 }
1009         } else
1010                 cleanup_pending_aio(td);
1011
1012         /*
1013          * stop job if we failed doing any IO
1014          */
1015         if (!ddir_rw_sum(td->this_io_bytes))
1016                 td->done = 1;
1017
1018         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1019 }
1020
1021 static void cleanup_io_u(struct thread_data *td)
1022 {
1023         struct io_u *io_u;
1024
1025         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1026
1027                 if (td->io_ops->io_u_free)
1028                         td->io_ops->io_u_free(td, io_u);
1029
1030                 fio_memfree(io_u, sizeof(*io_u));
1031         }
1032
1033         free_io_mem(td);
1034
1035         io_u_rexit(&td->io_u_requeues);
1036         io_u_qexit(&td->io_u_freelist);
1037         io_u_qexit(&td->io_u_all);
1038
1039         if (td->last_write_comp)
1040                 sfree(td->last_write_comp);
1041 }
1042
1043 static int init_io_u(struct thread_data *td)
1044 {
1045         struct io_u *io_u;
1046         unsigned int max_bs, min_write;
1047         int cl_align, i, max_units;
1048         int data_xfer = 1, err;
1049         char *p;
1050
1051         max_units = td->o.iodepth;
1052         max_bs = td_max_bs(td);
1053         min_write = td->o.min_bs[DDIR_WRITE];
1054         td->orig_buffer_size = (unsigned long long) max_bs
1055                                         * (unsigned long long) max_units;
1056
1057         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1058                 data_xfer = 0;
1059
1060         err = 0;
1061         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1062         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1063         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1064
1065         if (err) {
1066                 log_err("fio: failed setting up IO queues\n");
1067                 return 1;
1068         }
1069
1070         /*
1071          * if we may later need to do address alignment, then add any
1072          * possible adjustment here so that we don't cause a buffer
1073          * overflow later. this adjustment may be too much if we get
1074          * lucky and the allocator gives us an aligned address.
1075          */
1076         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1077             (td->io_ops->flags & FIO_RAWIO))
1078                 td->orig_buffer_size += page_mask + td->o.mem_align;
1079
1080         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1081                 unsigned long bs;
1082
1083                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1084                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1085         }
1086
1087         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1088                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1089                 return 1;
1090         }
1091
1092         if (data_xfer && allocate_io_mem(td))
1093                 return 1;
1094
1095         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1096             (td->io_ops->flags & FIO_RAWIO))
1097                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1098         else
1099                 p = td->orig_buffer;
1100
1101         cl_align = os_cache_line_size();
1102
1103         for (i = 0; i < max_units; i++) {
1104                 void *ptr;
1105
1106                 if (td->terminate)
1107                         return 1;
1108
1109                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1110                 if (!ptr) {
1111                         log_err("fio: unable to allocate aligned memory\n");
1112                         break;
1113                 }
1114
1115                 io_u = ptr;
1116                 memset(io_u, 0, sizeof(*io_u));
1117                 INIT_FLIST_HEAD(&io_u->verify_list);
1118                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1119
1120                 if (data_xfer) {
1121                         io_u->buf = p;
1122                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1123
1124                         if (td_write(td))
1125                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1126                         if (td_write(td) && td->o.verify_pattern_bytes) {
1127                                 /*
1128                                  * Fill the buffer with the pattern if we are
1129                                  * going to be doing writes.
1130                                  */
1131                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1132                         }
1133                 }
1134
1135                 io_u->index = i;
1136                 io_u->flags = IO_U_F_FREE;
1137                 io_u_qpush(&td->io_u_freelist, io_u);
1138
1139                 /*
1140                  * io_u never leaves this stack, used for iteration of all
1141                  * io_u buffers.
1142                  */
1143                 io_u_qpush(&td->io_u_all, io_u);
1144
1145                 if (td->io_ops->io_u_init) {
1146                         int ret = td->io_ops->io_u_init(td, io_u);
1147
1148                         if (ret) {
1149                                 log_err("fio: failed to init engine data: %d\n", ret);
1150                                 return 1;
1151                         }
1152                 }
1153
1154                 p += max_bs;
1155         }
1156
1157         if (td->o.verify != VERIFY_NONE) {
1158                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1159                 if (!td->last_write_comp) {
1160                         log_err("fio: failed to alloc write comp data\n");
1161                         return 1;
1162                 }
1163         }
1164
1165         return 0;
1166 }
1167
1168 static int switch_ioscheduler(struct thread_data *td)
1169 {
1170         char tmp[256], tmp2[128];
1171         FILE *f;
1172         int ret;
1173
1174         if (td->io_ops->flags & FIO_DISKLESSIO)
1175                 return 0;
1176
1177         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1178
1179         f = fopen(tmp, "r+");
1180         if (!f) {
1181                 if (errno == ENOENT) {
1182                         log_err("fio: os or kernel doesn't support IO scheduler"
1183                                 " switching\n");
1184                         return 0;
1185                 }
1186                 td_verror(td, errno, "fopen iosched");
1187                 return 1;
1188         }
1189
1190         /*
1191          * Set io scheduler.
1192          */
1193         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1194         if (ferror(f) || ret != 1) {
1195                 td_verror(td, errno, "fwrite");
1196                 fclose(f);
1197                 return 1;
1198         }
1199
1200         rewind(f);
1201
1202         /*
1203          * Read back and check that the selected scheduler is now the default.
1204          */
1205         memset(tmp, 0, sizeof(tmp));
1206         ret = fread(tmp, sizeof(tmp), 1, f);
1207         if (ferror(f) || ret < 0) {
1208                 td_verror(td, errno, "fread");
1209                 fclose(f);
1210                 return 1;
1211         }
1212         /*
1213          * either a list of io schedulers or "none\n" is expected.
1214          */
1215         tmp[strlen(tmp) - 1] = '\0';
1216
1217
1218         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1219         if (!strstr(tmp, tmp2)) {
1220                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1221                 td_verror(td, EINVAL, "iosched_switch");
1222                 fclose(f);
1223                 return 1;
1224         }
1225
1226         fclose(f);
1227         return 0;
1228 }
1229
1230 static int keep_running(struct thread_data *td)
1231 {
1232         unsigned long long limit;
1233
1234         if (td->done)
1235                 return 0;
1236         if (td->o.time_based)
1237                 return 1;
1238         if (td->o.loops) {
1239                 td->o.loops--;
1240                 return 1;
1241         }
1242         if (exceeds_number_ios(td))
1243                 return 0;
1244
1245         if (td->o.io_limit)
1246                 limit = td->o.io_limit;
1247         else
1248                 limit = td->o.size;
1249
1250         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1251                 uint64_t diff;
1252
1253                 /*
1254                  * If the difference is less than the minimum IO size, we
1255                  * are done.
1256                  */
1257                 diff = limit - ddir_rw_sum(td->io_bytes);
1258                 if (diff < td_max_bs(td))
1259                         return 0;
1260
1261                 if (fio_files_done(td))
1262                         return 0;
1263
1264                 return 1;
1265         }
1266
1267         return 0;
1268 }
1269
1270 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1271 {
1272         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1273         int ret;
1274         char *str;
1275
1276         str = malloc(newlen);
1277         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1278
1279         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1280         ret = system(str);
1281         if (ret == -1)
1282                 log_err("fio: exec of cmd <%s> failed\n", str);
1283
1284         free(str);
1285         return ret;
1286 }
1287
1288 /*
1289  * Dry run to compute correct state of numberio for verification.
1290  */
1291 static uint64_t do_dry_run(struct thread_data *td)
1292 {
1293         td_set_runstate(td, TD_RUNNING);
1294
1295         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1296                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1297                 struct io_u *io_u;
1298                 int ret;
1299
1300                 if (td->terminate || td->done)
1301                         break;
1302
1303                 io_u = get_io_u(td);
1304                 if (!io_u)
1305                         break;
1306
1307                 io_u_set(io_u, IO_U_F_FLIGHT);
1308                 io_u->error = 0;
1309                 io_u->resid = 0;
1310                 if (ddir_rw(acct_ddir(io_u)))
1311                         td->io_issues[acct_ddir(io_u)]++;
1312                 if (ddir_rw(io_u->ddir)) {
1313                         io_u_mark_depth(td, 1);
1314                         td->ts.total_io_u[io_u->ddir]++;
1315                 }
1316
1317                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1318                     td->o.do_verify &&
1319                     td->o.verify != VERIFY_NONE &&
1320                     !td->o.experimental_verify)
1321                         log_io_piece(td, io_u);
1322
1323                 ret = io_u_sync_complete(td, io_u);
1324                 (void) ret;
1325         }
1326
1327         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1328 }
1329
1330 static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1331 {
1332         const enum fio_ddir ddir = io_u->ddir;
1333         int ret;
1334
1335         dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1336
1337         io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1338
1339         td->cur_depth++;
1340
1341         ret = td_io_queue(td, io_u);
1342
1343         dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1344
1345         io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1346
1347         if (ret == FIO_Q_QUEUED)
1348                 ret = io_u_queued_complete(td, 1);
1349
1350         td->cur_depth--;
1351 }
1352
1353 /*
1354  * Entry point for the thread based jobs. The process based jobs end up
1355  * here as well, after a little setup.
1356  */
1357 static void *thread_main(void *data)
1358 {
1359         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1360         struct thread_data *td = data;
1361         struct thread_options *o = &td->o;
1362         pthread_condattr_t attr;
1363         int clear_state;
1364         int ret;
1365
1366         if (!o->use_thread) {
1367                 setsid();
1368                 td->pid = getpid();
1369         } else
1370                 td->pid = gettid();
1371
1372         fio_local_clock_init(o->use_thread);
1373
1374         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1375
1376         if (is_backend)
1377                 fio_server_send_start(td);
1378
1379         INIT_FLIST_HEAD(&td->io_log_list);
1380         INIT_FLIST_HEAD(&td->io_hist_list);
1381         INIT_FLIST_HEAD(&td->verify_list);
1382         INIT_FLIST_HEAD(&td->trim_list);
1383         INIT_FLIST_HEAD(&td->next_rand_list);
1384         pthread_mutex_init(&td->io_u_lock, NULL);
1385         td->io_hist_tree = RB_ROOT;
1386
1387         pthread_condattr_init(&attr);
1388         pthread_cond_init(&td->verify_cond, &attr);
1389         pthread_cond_init(&td->free_cond, &attr);
1390
1391         td_set_runstate(td, TD_INITIALIZED);
1392         dprint(FD_MUTEX, "up startup_mutex\n");
1393         fio_mutex_up(startup_mutex);
1394         dprint(FD_MUTEX, "wait on td->mutex\n");
1395         fio_mutex_down(td->mutex);
1396         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1397
1398         /*
1399          * A new gid requires privilege, so we need to do this before setting
1400          * the uid.
1401          */
1402         if (o->gid != -1U && setgid(o->gid)) {
1403                 td_verror(td, errno, "setgid");
1404                 goto err;
1405         }
1406         if (o->uid != -1U && setuid(o->uid)) {
1407                 td_verror(td, errno, "setuid");
1408                 goto err;
1409         }
1410
1411         /*
1412          * If we have a gettimeofday() thread, make sure we exclude that
1413          * thread from this job
1414          */
1415         if (o->gtod_cpu)
1416                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1417
1418         /*
1419          * Set affinity first, in case it has an impact on the memory
1420          * allocations.
1421          */
1422         if (fio_option_is_set(o, cpumask)) {
1423                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1424                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1425                         if (!ret) {
1426                                 log_err("fio: no CPUs set\n");
1427                                 log_err("fio: Try increasing number of available CPUs\n");
1428                                 td_verror(td, EINVAL, "cpus_split");
1429                                 goto err;
1430                         }
1431                 }
1432                 ret = fio_setaffinity(td->pid, o->cpumask);
1433                 if (ret == -1) {
1434                         td_verror(td, errno, "cpu_set_affinity");
1435                         goto err;
1436                 }
1437         }
1438
1439 #ifdef CONFIG_LIBNUMA
1440         /* numa node setup */
1441         if (fio_option_is_set(o, numa_cpunodes) ||
1442             fio_option_is_set(o, numa_memnodes)) {
1443                 struct bitmask *mask;
1444
1445                 if (numa_available() < 0) {
1446                         td_verror(td, errno, "Does not support NUMA API\n");
1447                         goto err;
1448                 }
1449
1450                 if (fio_option_is_set(o, numa_cpunodes)) {
1451                         mask = numa_parse_nodestring(o->numa_cpunodes);
1452                         ret = numa_run_on_node_mask(mask);
1453                         numa_free_nodemask(mask);
1454                         if (ret == -1) {
1455                                 td_verror(td, errno, \
1456                                         "numa_run_on_node_mask failed\n");
1457                                 goto err;
1458                         }
1459                 }
1460
1461                 if (fio_option_is_set(o, numa_memnodes)) {
1462                         mask = NULL;
1463                         if (o->numa_memnodes)
1464                                 mask = numa_parse_nodestring(o->numa_memnodes);
1465
1466                         switch (o->numa_mem_mode) {
1467                         case MPOL_INTERLEAVE:
1468                                 numa_set_interleave_mask(mask);
1469                                 break;
1470                         case MPOL_BIND:
1471                                 numa_set_membind(mask);
1472                                 break;
1473                         case MPOL_LOCAL:
1474                                 numa_set_localalloc();
1475                                 break;
1476                         case MPOL_PREFERRED:
1477                                 numa_set_preferred(o->numa_mem_prefer_node);
1478                                 break;
1479                         case MPOL_DEFAULT:
1480                         default:
1481                                 break;
1482                         }
1483
1484                         if (mask)
1485                                 numa_free_nodemask(mask);
1486
1487                 }
1488         }
1489 #endif
1490
1491         if (fio_pin_memory(td))
1492                 goto err;
1493
1494         /*
1495          * May alter parameters that init_io_u() will use, so we need to
1496          * do this first.
1497          */
1498         if (init_iolog(td))
1499                 goto err;
1500
1501         if (init_io_u(td))
1502                 goto err;
1503
1504         if (o->verify_async && verify_async_init(td))
1505                 goto err;
1506
1507         if (fio_option_is_set(o, ioprio) ||
1508             fio_option_is_set(o, ioprio_class)) {
1509                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1510                 if (ret == -1) {
1511                         td_verror(td, errno, "ioprio_set");
1512                         goto err;
1513                 }
1514         }
1515
1516         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1517                 goto err;
1518
1519         errno = 0;
1520         if (nice(o->nice) == -1 && errno != 0) {
1521                 td_verror(td, errno, "nice");
1522                 goto err;
1523         }
1524
1525         if (o->ioscheduler && switch_ioscheduler(td))
1526                 goto err;
1527
1528         if (!o->create_serialize && setup_files(td))
1529                 goto err;
1530
1531         if (td_io_init(td))
1532                 goto err;
1533
1534         if (init_random_map(td))
1535                 goto err;
1536
1537         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1538                 goto err;
1539
1540         if (o->pre_read) {
1541                 if (pre_read_files(td) < 0)
1542                         goto err;
1543         }
1544
1545         if (td->flags & TD_F_COMPRESS_LOG)
1546                 tp_init(&td->tp_data);
1547
1548         fio_verify_init(td);
1549
1550         if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1551             workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1552                 goto err;
1553
1554         fio_gettime(&td->epoch, NULL);
1555         fio_getrusage(&td->ru_start);
1556         clear_state = 0;
1557         while (keep_running(td)) {
1558                 uint64_t verify_bytes;
1559
1560                 fio_gettime(&td->start, NULL);
1561                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1562                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1563                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1564
1565                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1566                                 o->ratemin[DDIR_TRIM]) {
1567                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1568                                                 sizeof(td->bw_sample_time));
1569                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1570                                                 sizeof(td->bw_sample_time));
1571                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1572                                                 sizeof(td->bw_sample_time));
1573                 }
1574
1575                 if (clear_state)
1576                         clear_io_state(td);
1577
1578                 prune_io_piece_log(td);
1579
1580                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1581                         verify_bytes = do_dry_run(td);
1582                 else
1583                         verify_bytes = do_io(td);
1584
1585                 clear_state = 1;
1586
1587                 /*
1588                  * Make sure we've successfully updated the rusage stats
1589                  * before waiting on the stat mutex. Otherwise we could have
1590                  * the stat thread holding stat mutex and waiting for
1591                  * the rusage_sem, which would never get upped because
1592                  * this thread is waiting for the stat mutex.
1593                  */
1594                 check_update_rusage(td);
1595
1596                 fio_mutex_down(stat_mutex);
1597                 if (td_read(td) && td->io_bytes[DDIR_READ])
1598                         update_runtime(td, elapsed_us, DDIR_READ);
1599                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1600                         update_runtime(td, elapsed_us, DDIR_WRITE);
1601                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1602                         update_runtime(td, elapsed_us, DDIR_TRIM);
1603                 fio_gettime(&td->start, NULL);
1604                 fio_mutex_up(stat_mutex);
1605
1606                 if (td->error || td->terminate)
1607                         break;
1608
1609                 if (!o->do_verify ||
1610                     o->verify == VERIFY_NONE ||
1611                     (td->io_ops->flags & FIO_UNIDIR))
1612                         continue;
1613
1614                 clear_io_state(td);
1615
1616                 fio_gettime(&td->start, NULL);
1617
1618                 do_verify(td, verify_bytes);
1619
1620                 /*
1621                  * See comment further up for why this is done here.
1622                  */
1623                 check_update_rusage(td);
1624
1625                 fio_mutex_down(stat_mutex);
1626                 update_runtime(td, elapsed_us, DDIR_READ);
1627                 fio_gettime(&td->start, NULL);
1628                 fio_mutex_up(stat_mutex);
1629
1630                 if (td->error || td->terminate)
1631                         break;
1632         }
1633
1634         update_rusage_stat(td);
1635         td->ts.total_run_time = mtime_since_now(&td->epoch);
1636         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1637         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1638         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1639
1640         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1641             (td->o.verify != VERIFY_NONE && td_write(td))) {
1642                 struct all_io_list *state;
1643                 size_t sz;
1644
1645                 state = get_all_io_list(td->thread_number, &sz);
1646                 if (state) {
1647                         __verify_save_state(state, "local");
1648                         free(state);
1649                 }
1650         }
1651
1652         fio_unpin_memory(td);
1653
1654         fio_writeout_logs(td);
1655
1656         if (o->io_submit_mode == IO_MODE_OFFLOAD)
1657                 workqueue_exit(&td->io_wq);
1658
1659         if (td->flags & TD_F_COMPRESS_LOG)
1660                 tp_exit(&td->tp_data);
1661
1662         if (o->exec_postrun)
1663                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1664
1665         if (exitall_on_terminate)
1666                 fio_terminate_threads(td->groupid);
1667
1668 err:
1669         if (td->error)
1670                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1671                                                         td->verror);
1672
1673         if (o->verify_async)
1674                 verify_async_exit(td);
1675
1676         close_and_free_files(td);
1677         cleanup_io_u(td);
1678         close_ioengine(td);
1679         cgroup_shutdown(td, &cgroup_mnt);
1680         verify_free_state(td);
1681
1682         if (fio_option_is_set(o, cpumask)) {
1683                 ret = fio_cpuset_exit(&o->cpumask);
1684                 if (ret)
1685                         td_verror(td, ret, "fio_cpuset_exit");
1686         }
1687
1688         /*
1689          * do this very late, it will log file closing as well
1690          */
1691         if (o->write_iolog_file)
1692                 write_iolog_close(td);
1693
1694         fio_mutex_remove(td->mutex);
1695         td->mutex = NULL;
1696
1697         td_set_runstate(td, TD_EXITED);
1698
1699         /*
1700          * Do this last after setting our runstate to exited, so we
1701          * know that the stat thread is signaled.
1702          */
1703         check_update_rusage(td);
1704
1705         return (void *) (uintptr_t) td->error;
1706 }
1707
1708
1709 /*
1710  * We cannot pass the td data into a forked process, so attach the td and
1711  * pass it to the thread worker.
1712  */
1713 static int fork_main(int shmid, int offset)
1714 {
1715         struct thread_data *td;
1716         void *data, *ret;
1717
1718 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1719         data = shmat(shmid, NULL, 0);
1720         if (data == (void *) -1) {
1721                 int __err = errno;
1722
1723                 perror("shmat");
1724                 return __err;
1725         }
1726 #else
1727         /*
1728          * HP-UX inherits shm mappings?
1729          */
1730         data = threads;
1731 #endif
1732
1733         td = data + offset * sizeof(struct thread_data);
1734         ret = thread_main(td);
1735         shmdt(data);
1736         return (int) (uintptr_t) ret;
1737 }
1738
1739 static void dump_td_info(struct thread_data *td)
1740 {
1741         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1742                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1743                         (unsigned long) time_since_now(&td->terminate_time));
1744 }
1745
1746 /*
1747  * Run over the job map and reap the threads that have exited, if any.
1748  */
1749 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1750                          unsigned int *m_rate)
1751 {
1752         struct thread_data *td;
1753         unsigned int cputhreads, realthreads, pending;
1754         int i, status, ret;
1755
1756         /*
1757          * reap exited threads (TD_EXITED -> TD_REAPED)
1758          */
1759         realthreads = pending = cputhreads = 0;
1760         for_each_td(td, i) {
1761                 int flags = 0;
1762
1763                 /*
1764                  * ->io_ops is NULL for a thread that has closed its
1765                  * io engine
1766                  */
1767                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1768                         cputhreads++;
1769                 else
1770                         realthreads++;
1771
1772                 if (!td->pid) {
1773                         pending++;
1774                         continue;
1775                 }
1776                 if (td->runstate == TD_REAPED)
1777                         continue;
1778                 if (td->o.use_thread) {
1779                         if (td->runstate == TD_EXITED) {
1780                                 td_set_runstate(td, TD_REAPED);
1781                                 goto reaped;
1782                         }
1783                         continue;
1784                 }
1785
1786                 flags = WNOHANG;
1787                 if (td->runstate == TD_EXITED)
1788                         flags = 0;
1789
1790                 /*
1791                  * check if someone quit or got killed in an unusual way
1792                  */
1793                 ret = waitpid(td->pid, &status, flags);
1794                 if (ret < 0) {
1795                         if (errno == ECHILD) {
1796                                 log_err("fio: pid=%d disappeared %d\n",
1797                                                 (int) td->pid, td->runstate);
1798                                 td->sig = ECHILD;
1799                                 td_set_runstate(td, TD_REAPED);
1800                                 goto reaped;
1801                         }
1802                         perror("waitpid");
1803                 } else if (ret == td->pid) {
1804                         if (WIFSIGNALED(status)) {
1805                                 int sig = WTERMSIG(status);
1806
1807                                 if (sig != SIGTERM && sig != SIGUSR2)
1808                                         log_err("fio: pid=%d, got signal=%d\n",
1809                                                         (int) td->pid, sig);
1810                                 td->sig = sig;
1811                                 td_set_runstate(td, TD_REAPED);
1812                                 goto reaped;
1813                         }
1814                         if (WIFEXITED(status)) {
1815                                 if (WEXITSTATUS(status) && !td->error)
1816                                         td->error = WEXITSTATUS(status);
1817
1818                                 td_set_runstate(td, TD_REAPED);
1819                                 goto reaped;
1820                         }
1821                 }
1822
1823                 /*
1824                  * If the job is stuck, do a forceful timeout of it and
1825                  * move on.
1826                  */
1827                 if (td->terminate &&
1828                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1829                         dump_td_info(td);
1830                         td_set_runstate(td, TD_REAPED);
1831                         goto reaped;
1832                 }
1833
1834                 /*
1835                  * thread is not dead, continue
1836                  */
1837                 pending++;
1838                 continue;
1839 reaped:
1840                 (*nr_running)--;
1841                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1842                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1843                 if (!td->pid)
1844                         pending--;
1845
1846                 if (td->error)
1847                         exit_value++;
1848
1849                 done_secs += mtime_since_now(&td->epoch) / 1000;
1850                 profile_td_exit(td);
1851         }
1852
1853         if (*nr_running == cputhreads && !pending && realthreads)
1854                 fio_terminate_threads(TERMINATE_ALL);
1855 }
1856
1857 static int __check_trigger_file(void)
1858 {
1859         struct stat sb;
1860
1861         if (!trigger_file)
1862                 return 0;
1863
1864         if (stat(trigger_file, &sb))
1865                 return 0;
1866
1867         if (unlink(trigger_file) < 0)
1868                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1869                                                         strerror(errno));
1870
1871         return 1;
1872 }
1873
1874 static int trigger_timedout(void)
1875 {
1876         if (trigger_timeout)
1877                 return time_since_genesis() >= trigger_timeout;
1878
1879         return 0;
1880 }
1881
1882 void exec_trigger(const char *cmd)
1883 {
1884         int ret;
1885
1886         if (!cmd)
1887                 return;
1888
1889         ret = system(cmd);
1890         if (ret == -1)
1891                 log_err("fio: failed executing %s trigger\n", cmd);
1892 }
1893
1894 void check_trigger_file(void)
1895 {
1896         if (__check_trigger_file() || trigger_timedout()) {
1897                 if (nr_clients)
1898                         fio_clients_send_trigger(trigger_remote_cmd);
1899                 else {
1900                         verify_save_state();
1901                         fio_terminate_threads(TERMINATE_ALL);
1902                         exec_trigger(trigger_cmd);
1903                 }
1904         }
1905 }
1906
1907 static int fio_verify_load_state(struct thread_data *td)
1908 {
1909         int ret;
1910
1911         if (!td->o.verify_state)
1912                 return 0;
1913
1914         if (is_backend) {
1915                 void *data;
1916                 int ver;
1917
1918                 ret = fio_server_get_verify_state(td->o.name,
1919                                         td->thread_number - 1, &data, &ver);
1920                 if (!ret)
1921                         verify_convert_assign_state(td, data, ver);
1922         } else
1923                 ret = verify_load_state(td, "local");
1924
1925         return ret;
1926 }
1927
1928 static void do_usleep(unsigned int usecs)
1929 {
1930         check_for_running_stats();
1931         check_trigger_file();
1932         usleep(usecs);
1933 }
1934
1935 static int check_mount_writes(struct thread_data *td)
1936 {
1937         struct fio_file *f;
1938         unsigned int i;
1939
1940         if (!td_write(td) || td->o.allow_mounted_write)
1941                 return 0;
1942
1943         for_each_file(td, f, i) {
1944                 if (f->filetype != FIO_TYPE_BD)
1945                         continue;
1946                 if (device_is_mounted(f->file_name))
1947                         goto mounted;
1948         }
1949
1950         return 0;
1951 mounted:
1952         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1953         return 1;
1954 }
1955
1956 /*
1957  * Main function for kicking off and reaping jobs, as needed.
1958  */
1959 static void run_threads(void)
1960 {
1961         struct thread_data *td;
1962         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1963         uint64_t spent;
1964
1965         if (fio_gtod_offload && fio_start_gtod_thread())
1966                 return;
1967
1968         fio_idle_prof_init();
1969
1970         set_sig_handlers();
1971
1972         nr_thread = nr_process = 0;
1973         for_each_td(td, i) {
1974                 if (check_mount_writes(td))
1975                         return;
1976                 if (td->o.use_thread)
1977                         nr_thread++;
1978                 else
1979                         nr_process++;
1980         }
1981
1982         if (output_format == FIO_OUTPUT_NORMAL) {
1983                 log_info("Starting ");
1984                 if (nr_thread)
1985                         log_info("%d thread%s", nr_thread,
1986                                                 nr_thread > 1 ? "s" : "");
1987                 if (nr_process) {
1988                         if (nr_thread)
1989                                 log_info(" and ");
1990                         log_info("%d process%s", nr_process,
1991                                                 nr_process > 1 ? "es" : "");
1992                 }
1993                 log_info("\n");
1994                 log_info_flush();
1995         }
1996
1997         todo = thread_number;
1998         nr_running = 0;
1999         nr_started = 0;
2000         m_rate = t_rate = 0;
2001
2002         for_each_td(td, i) {
2003                 print_status_init(td->thread_number - 1);
2004
2005                 if (!td->o.create_serialize)
2006                         continue;
2007
2008                 if (fio_verify_load_state(td))
2009                         goto reap;
2010
2011                 /*
2012                  * do file setup here so it happens sequentially,
2013                  * we don't want X number of threads getting their
2014                  * client data interspersed on disk
2015                  */
2016                 if (setup_files(td)) {
2017 reap:
2018                         exit_value++;
2019                         if (td->error)
2020                                 log_err("fio: pid=%d, err=%d/%s\n",
2021                                         (int) td->pid, td->error, td->verror);
2022                         td_set_runstate(td, TD_REAPED);
2023                         todo--;
2024                 } else {
2025                         struct fio_file *f;
2026                         unsigned int j;
2027
2028                         /*
2029                          * for sharing to work, each job must always open
2030                          * its own files. so close them, if we opened them
2031                          * for creation
2032                          */
2033                         for_each_file(td, f, j) {
2034                                 if (fio_file_open(f))
2035                                         td_io_close_file(td, f);
2036                         }
2037                 }
2038         }
2039
2040         /* start idle threads before io threads start to run */
2041         fio_idle_prof_start();
2042
2043         set_genesis_time();
2044
2045         while (todo) {
2046                 struct thread_data *map[REAL_MAX_JOBS];
2047                 struct timeval this_start;
2048                 int this_jobs = 0, left;
2049
2050                 /*
2051                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2052                  */
2053                 for_each_td(td, i) {
2054                         if (td->runstate != TD_NOT_CREATED)
2055                                 continue;
2056
2057                         /*
2058                          * never got a chance to start, killed by other
2059                          * thread for some reason
2060                          */
2061                         if (td->terminate) {
2062                                 todo--;
2063                                 continue;
2064                         }
2065
2066                         if (td->o.start_delay) {
2067                                 spent = utime_since_genesis();
2068
2069                                 if (td->o.start_delay > spent)
2070                                         continue;
2071                         }
2072
2073                         if (td->o.stonewall && (nr_started || nr_running)) {
2074                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2075                                                         td->o.name);
2076                                 break;
2077                         }
2078
2079                         init_disk_util(td);
2080
2081                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2082                         td->update_rusage = 0;
2083
2084                         /*
2085                          * Set state to created. Thread will transition
2086                          * to TD_INITIALIZED when it's done setting up.
2087                          */
2088                         td_set_runstate(td, TD_CREATED);
2089                         map[this_jobs++] = td;
2090                         nr_started++;
2091
2092                         if (td->o.use_thread) {
2093                                 int ret;
2094
2095                                 dprint(FD_PROCESS, "will pthread_create\n");
2096                                 ret = pthread_create(&td->thread, NULL,
2097                                                         thread_main, td);
2098                                 if (ret) {
2099                                         log_err("pthread_create: %s\n",
2100                                                         strerror(ret));
2101                                         nr_started--;
2102                                         break;
2103                                 }
2104                                 ret = pthread_detach(td->thread);
2105                                 if (ret)
2106                                         log_err("pthread_detach: %s",
2107                                                         strerror(ret));
2108                         } else {
2109                                 pid_t pid;
2110                                 dprint(FD_PROCESS, "will fork\n");
2111                                 pid = fork();
2112                                 if (!pid) {
2113                                         int ret = fork_main(shm_id, i);
2114
2115                                         _exit(ret);
2116                                 } else if (i == fio_debug_jobno)
2117                                         *fio_debug_jobp = pid;
2118                         }
2119                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2120                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2121                                 log_err("fio: job startup hung? exiting.\n");
2122                                 fio_terminate_threads(TERMINATE_ALL);
2123                                 fio_abort = 1;
2124                                 nr_started--;
2125                                 break;
2126                         }
2127                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2128                 }
2129
2130                 /*
2131                  * Wait for the started threads to transition to
2132                  * TD_INITIALIZED.
2133                  */
2134                 fio_gettime(&this_start, NULL);
2135                 left = this_jobs;
2136                 while (left && !fio_abort) {
2137                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2138                                 break;
2139
2140                         do_usleep(100000);
2141
2142                         for (i = 0; i < this_jobs; i++) {
2143                                 td = map[i];
2144                                 if (!td)
2145                                         continue;
2146                                 if (td->runstate == TD_INITIALIZED) {
2147                                         map[i] = NULL;
2148                                         left--;
2149                                 } else if (td->runstate >= TD_EXITED) {
2150                                         map[i] = NULL;
2151                                         left--;
2152                                         todo--;
2153                                         nr_running++; /* work-around... */
2154                                 }
2155                         }
2156                 }
2157
2158                 if (left) {
2159                         log_err("fio: %d job%s failed to start\n", left,
2160                                         left > 1 ? "s" : "");
2161                         for (i = 0; i < this_jobs; i++) {
2162                                 td = map[i];
2163                                 if (!td)
2164                                         continue;
2165                                 kill(td->pid, SIGTERM);
2166                         }
2167                         break;
2168                 }
2169
2170                 /*
2171                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2172                  */
2173                 for_each_td(td, i) {
2174                         if (td->runstate != TD_INITIALIZED)
2175                                 continue;
2176
2177                         if (in_ramp_time(td))
2178                                 td_set_runstate(td, TD_RAMP);
2179                         else
2180                                 td_set_runstate(td, TD_RUNNING);
2181                         nr_running++;
2182                         nr_started--;
2183                         m_rate += ddir_rw_sum(td->o.ratemin);
2184                         t_rate += ddir_rw_sum(td->o.rate);
2185                         todo--;
2186                         fio_mutex_up(td->mutex);
2187                 }
2188
2189                 reap_threads(&nr_running, &t_rate, &m_rate);
2190
2191                 if (todo)
2192                         do_usleep(100000);
2193         }
2194
2195         while (nr_running) {
2196                 reap_threads(&nr_running, &t_rate, &m_rate);
2197                 do_usleep(10000);
2198         }
2199
2200         fio_idle_prof_stop();
2201
2202         update_io_ticks();
2203 }
2204
2205 static void wait_for_helper_thread_exit(void)
2206 {
2207         void *ret;
2208
2209         helper_exit = 1;
2210         pthread_cond_signal(&helper_cond);
2211         pthread_join(helper_thread, &ret);
2212 }
2213
2214 static void free_disk_util(void)
2215 {
2216         disk_util_prune_entries();
2217
2218         pthread_cond_destroy(&helper_cond);
2219 }
2220
2221 static void *helper_thread_main(void *data)
2222 {
2223         int ret = 0;
2224
2225         fio_mutex_up(startup_mutex);
2226
2227         while (!ret) {
2228                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2229                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2230                 struct timespec ts;
2231                 struct timeval tv;
2232
2233                 gettimeofday(&tv, NULL);
2234                 ts.tv_sec = tv.tv_sec + sec;
2235                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2236
2237                 if (ts.tv_nsec >= 1000000000ULL) {
2238                         ts.tv_nsec -= 1000000000ULL;
2239                         ts.tv_sec++;
2240                 }
2241
2242                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2243
2244                 ret = update_io_ticks();
2245
2246                 if (helper_do_stat) {
2247                         helper_do_stat = 0;
2248                         __show_running_run_stats();
2249                 }
2250
2251                 if (!is_backend)
2252                         print_thread_status();
2253         }
2254
2255         return NULL;
2256 }
2257
2258 static int create_helper_thread(void)
2259 {
2260         int ret;
2261
2262         setup_disk_util();
2263
2264         pthread_cond_init(&helper_cond, NULL);
2265         pthread_mutex_init(&helper_lock, NULL);
2266
2267         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2268         if (ret) {
2269                 log_err("Can't create helper thread: %s\n", strerror(ret));
2270                 return 1;
2271         }
2272
2273         dprint(FD_MUTEX, "wait on startup_mutex\n");
2274         fio_mutex_down(startup_mutex);
2275         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2276         return 0;
2277 }
2278
2279 int fio_backend(void)
2280 {
2281         struct thread_data *td;
2282         int i;
2283
2284         if (exec_profile) {
2285                 if (load_profile(exec_profile))
2286                         return 1;
2287                 free(exec_profile);
2288                 exec_profile = NULL;
2289         }
2290         if (!thread_number)
2291                 return 0;
2292
2293         if (write_bw_log) {
2294                 struct log_params p = {
2295                         .log_type = IO_LOG_TYPE_BW,
2296                 };
2297
2298                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2299                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2300                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2301         }
2302
2303         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2304         if (startup_mutex == NULL)
2305                 return 1;
2306
2307         set_genesis_time();
2308         stat_init();
2309         create_helper_thread();
2310
2311         cgroup_list = smalloc(sizeof(*cgroup_list));
2312         INIT_FLIST_HEAD(cgroup_list);
2313
2314         run_threads();
2315
2316         wait_for_helper_thread_exit();
2317
2318         if (!fio_abort) {
2319                 __show_run_stats();
2320                 if (write_bw_log) {
2321                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2322                                 struct io_log *log = agg_io_log[i];
2323
2324                                 flush_log(log, 0);
2325                                 free_log(log);
2326                         }
2327                 }
2328         }
2329
2330         for_each_td(td, i) {
2331                 fio_options_free(td);
2332                 if (td->rusage_sem) {
2333                         fio_mutex_remove(td->rusage_sem);
2334                         td->rusage_sem = NULL;
2335                 }
2336         }
2337
2338         free_disk_util();
2339         cgroup_kill(cgroup_list);
2340         sfree(cgroup_list);
2341         sfree(cgroup_mnt);
2342
2343         fio_mutex_remove(startup_mutex);
2344         stat_exit();
2345         return exit_value;
2346 }