backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38 #include <math.h>
  39
  40 #include "fio.h"
  41 #ifndef FIO_NO_HAVE_SHM_H
  42 #include <sys/shm.h>
  43 #endif
  44 #include "hash.h"
  45 #include "smalloc.h"
  46 #include "verify.h"
  47 #include "trim.h"
  48 #include "diskutil.h"
  49 #include "cgroup.h"
  50 #include "profile.h"
  51 #include "lib/rand.h"
  52 #include "lib/memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59 #include "rate-submit.h"
  60
  61 static pthread_t helper_thread;
  62 static pthread_mutex_t helper_lock;
  63 pthread_cond_t helper_cond;
  64 int helper_do_stat = 0;
  65
  66 static struct fio_mutex *startup_mutex;
  67 static struct flist_head *cgroup_list;
  68 static char *cgroup_mnt;
  69 static int exit_value;
  70 static volatile int fio_abort;
  71 static unsigned int nr_process = 0;
  72 static unsigned int nr_thread = 0;
  73
  74 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  75
  76 int groupid = 0;
  77 unsigned int thread_number = 0;
  78 unsigned int stat_number = 0;
  79 int shm_id = 0;
  80 int temp_stall_ts;
  81 unsigned long done_secs = 0;
  82 volatile int helper_exit = 0;
  83
  84 #define PAGE_ALIGN(buf) \
  85         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  86
  87 #define JOB_START_TIMEOUT       (5 * 1000)
  88
  89 static void sig_int(int sig)
  90 {
  91         if (threads) {
  92                 if (is_backend)
  93                         fio_server_got_signal(sig);
  94                 else {
  95                         log_info("\nfio: terminating on signal %d\n", sig);
  96                         log_info_flush();
  97                         exit_value = 128;
  98                 }
  99
 100                 fio_terminate_threads(TERMINATE_ALL);
 101         }
 102 }
 103
 104 void sig_show_status(int sig)
 105 {
 106         show_running_run_stats();
 107 }
 108
 109 static void set_sig_handlers(void)
 110 {
 111         struct sigaction act;
 112
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGINT, &act, NULL);
 117
 118         memset(&act, 0, sizeof(act));
 119         act.sa_handler = sig_int;
 120         act.sa_flags = SA_RESTART;
 121         sigaction(SIGTERM, &act, NULL);
 122
 123 /* Windows uses SIGBREAK as a quit signal from other applications */
 124 #ifdef WIN32
 125         memset(&act, 0, sizeof(act));
 126         act.sa_handler = sig_int;
 127         act.sa_flags = SA_RESTART;
 128         sigaction(SIGBREAK, &act, NULL);
 129 #endif
 130
 131         memset(&act, 0, sizeof(act));
 132         act.sa_handler = sig_show_status;
 133         act.sa_flags = SA_RESTART;
 134         sigaction(SIGUSR1, &act, NULL);
 135
 136         if (is_backend) {
 137                 memset(&act, 0, sizeof(act));
 138                 act.sa_handler = sig_int;
 139                 act.sa_flags = SA_RESTART;
 140                 sigaction(SIGPIPE, &act, NULL);
 141         }
 142 }
 143
 144 /*
 145  * Check if we are above the minimum rate given.
 146  */
 147 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 148                              enum fio_ddir ddir)
 149 {
 150         unsigned long long bytes = 0;
 151         unsigned long iops = 0;
 152         unsigned long spent;
 153         unsigned long rate;
 154         unsigned int ratemin = 0;
 155         unsigned int rate_iops = 0;
 156         unsigned int rate_iops_min = 0;
 157
 158         assert(ddir_rw(ddir));
 159
 160         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 161                 return false;
 162
 163         /*
 164          * allow a 2 second settle period in the beginning
 165          */
 166         if (mtime_since(&td->start, now) < 2000)
 167                 return false;
 168
 169         iops += td->this_io_blocks[ddir];
 170         bytes += td->this_io_bytes[ddir];
 171         ratemin += td->o.ratemin[ddir];
 172         rate_iops += td->o.rate_iops[ddir];
 173         rate_iops_min += td->o.rate_iops_min[ddir];
 174
 175         /*
 176          * if rate blocks is set, sample is running
 177          */
 178         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 179                 spent = mtime_since(&td->lastrate[ddir], now);
 180                 if (spent < td->o.ratecycle)
 181                         return false;
 182
 183                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 184                         /*
 185                          * check bandwidth specified rate
 186                          */
 187                         if (bytes < td->rate_bytes[ddir]) {
 188                                 log_err("%s: min rate %u not met\n", td->o.name,
 189                                                                 ratemin);
 190                                 return true;
 191                         } else {
 192                                 if (spent)
 193                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 194                                 else
 195                                         rate = 0;
 196
 197                                 if (rate < ratemin ||
 198                                     bytes < td->rate_bytes[ddir]) {
 199                                         log_err("%s: min rate %u not met, got"
 200                                                 " %luKB/sec\n", td->o.name,
 201                                                         ratemin, rate);
 202                                         return true;
 203                                 }
 204                         }
 205                 } else {
 206                         /*
 207                          * checks iops specified rate
 208                          */
 209                         if (iops < rate_iops) {
 210                                 log_err("%s: min iops rate %u not met\n",
 211                                                 td->o.name, rate_iops);
 212                                 return true;
 213                         } else {
 214                                 if (spent)
 215                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 216                                 else
 217                                         rate = 0;
 218
 219                                 if (rate < rate_iops_min ||
 220                                     iops < td->rate_blocks[ddir]) {
 221                                         log_err("%s: min iops rate %u not met,"
 222                                                 " got %lu\n", td->o.name,
 223                                                         rate_iops_min, rate);
 224                                         return true;
 225                                 }
 226                         }
 227                 }
 228         }
 229
 230         td->rate_bytes[ddir] = bytes;
 231         td->rate_blocks[ddir] = iops;
 232         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 233         return false;
 234 }
 235
 236 static bool check_min_rate(struct thread_data *td, struct timeval *now)
 237 {
 238         bool ret = false;
 239
 240         if (td->bytes_done[DDIR_READ])
 241                 ret |= __check_min_rate(td, now, DDIR_READ);
 242         if (td->bytes_done[DDIR_WRITE])
 243                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 244         if (td->bytes_done[DDIR_TRIM])
 245                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 246
 247         return ret;
 248 }
 249
 250 /*
 251  * When job exits, we can cancel the in-flight IO if we are using async
 252  * io. Attempt to do so.
 253  */
 254 static void cleanup_pending_aio(struct thread_data *td)
 255 {
 256         int r;
 257
 258         /*
 259          * get immediately available events, if any
 260          */
 261         r = io_u_queued_complete(td, 0);
 262         if (r < 0)
 263                 return;
 264
 265         /*
 266          * now cancel remaining active events
 267          */
 268         if (td->io_ops->cancel) {
 269                 struct io_u *io_u;
 270                 int i;
 271
 272                 io_u_qiter(&td->io_u_all, io_u, i) {
 273                         if (io_u->flags & IO_U_F_FLIGHT) {
 274                                 r = td->io_ops->cancel(td, io_u);
 275                                 if (!r)
 276                                         put_io_u(td, io_u);
 277                         }
 278                 }
 279         }
 280
 281         if (td->cur_depth)
 282                 r = io_u_queued_complete(td, td->cur_depth);
 283 }
 284
 285 /*
 286  * Helper to handle the final sync of a file. Works just like the normal
 287  * io path, just does everything sync.
 288  */
 289 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 290 {
 291         struct io_u *io_u = __get_io_u(td);
 292         int ret;
 293
 294         if (!io_u)
 295                 return true;
 296
 297         io_u->ddir = DDIR_SYNC;
 298         io_u->file = f;
 299
 300         if (td_io_prep(td, io_u)) {
 301                 put_io_u(td, io_u);
 302                 return true;
 303         }
 304
 305 requeue:
 306         ret = td_io_queue(td, io_u);
 307         if (ret < 0) {
 308                 td_verror(td, io_u->error, "td_io_queue");
 309                 put_io_u(td, io_u);
 310                 return true;
 311         } else if (ret == FIO_Q_QUEUED) {
 312                 if (td_io_commit(td))
 313                         return true;
 314                 if (io_u_queued_complete(td, 1) < 0)
 315                         return true;
 316         } else if (ret == FIO_Q_COMPLETED) {
 317                 if (io_u->error) {
 318                         td_verror(td, io_u->error, "td_io_queue");
 319                         return true;
 320                 }
 321
 322                 if (io_u_sync_complete(td, io_u) < 0)
 323                         return true;
 324         } else if (ret == FIO_Q_BUSY) {
 325                 if (td_io_commit(td))
 326                         return true;
 327                 goto requeue;
 328         }
 329
 330         return false;
 331 }
 332
 333 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 334 {
 335         int ret;
 336
 337         if (fio_file_open(f))
 338                 return fio_io_sync(td, f);
 339
 340         if (td_io_open_file(td, f))
 341                 return 1;
 342
 343         ret = fio_io_sync(td, f);
 344         td_io_close_file(td, f);
 345         return ret;
 346 }
 347
 348 static inline void __update_tv_cache(struct thread_data *td)
 349 {
 350         fio_gettime(&td->tv_cache, NULL);
 351 }
 352
 353 static inline void update_tv_cache(struct thread_data *td)
 354 {
 355         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 356                 __update_tv_cache(td);
 357 }
 358
 359 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 360 {
 361         if (in_ramp_time(td))
 362                 return false;
 363         if (!td->o.timeout)
 364                 return false;
 365         if (utime_since(&td->epoch, t) >= td->o.timeout)
 366                 return true;
 367
 368         return false;
 369 }
 370
 371 /*
 372  * We need to update the runtime consistently in ms, but keep a running
 373  * tally of the current elapsed time in microseconds for sub millisecond
 374  * updates.
 375  */
 376 static inline void update_runtime(struct thread_data *td,
 377                                   unsigned long long *elapsed_us,
 378                                   const enum fio_ddir ddir)
 379 {
 380         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 381                 return;
 382
 383         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 384         elapsed_us[ddir] += utime_since_now(&td->start);
 385         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 386 }
 387
 388 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 389                                 int *retptr)
 390 {
 391         int ret = *retptr;
 392
 393         if (ret < 0 || td->error) {
 394                 int err = td->error;
 395                 enum error_type_bit eb;
 396
 397                 if (ret < 0)
 398                         err = -ret;
 399
 400                 eb = td_error_type(ddir, err);
 401                 if (!(td->o.continue_on_error & (1 << eb)))
 402                         return true;
 403
 404                 if (td_non_fatal_error(td, eb, err)) {
 405                         /*
 406                          * Continue with the I/Os in case of
 407                          * a non fatal error.
 408                          */
 409                         update_error_count(td, err);
 410                         td_clear_error(td);
 411                         *retptr = 0;
 412                         return false;
 413                 } else if (td->o.fill_device && err == ENOSPC) {
 414                         /*
 415                          * We expect to hit this error if
 416                          * fill_device option is set.
 417                          */
 418                         td_clear_error(td);
 419                         fio_mark_td_terminate(td);
 420                         return true;
 421                 } else {
 422                         /*
 423                          * Stop the I/O in case of a fatal
 424                          * error.
 425                          */
 426                         update_error_count(td, err);
 427                         return true;
 428                 }
 429         }
 430
 431         return false;
 432 }
 433
 434 static void check_update_rusage(struct thread_data *td)
 435 {
 436         if (td->update_rusage) {
 437                 td->update_rusage = 0;
 438                 update_rusage_stat(td);
 439                 fio_mutex_up(td->rusage_sem);
 440         }
 441 }
 442
 443 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 444 {
 445         const int full = queue_full(td);
 446         int min_evts = 0;
 447         int ret;
 448
 449         /*
 450          * if the queue is full, we MUST reap at least 1 event
 451          */
 452         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 453         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 454                 min_evts = 1;
 455
 456         if (time && (__should_check_rate(td, DDIR_READ) ||
 457             __should_check_rate(td, DDIR_WRITE) ||
 458             __should_check_rate(td, DDIR_TRIM)))
 459                 fio_gettime(time, NULL);
 460
 461         do {
 462                 ret = io_u_queued_complete(td, min_evts);
 463                 if (ret < 0)
 464                         break;
 465         } while (full && (td->cur_depth > td->o.iodepth_low));
 466
 467         return ret;
 468 }
 469
 470 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 471                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 472                    struct timeval *comp_time)
 473 {
 474         int ret2;
 475
 476         switch (*ret) {
 477         case FIO_Q_COMPLETED:
 478                 if (io_u->error) {
 479                         *ret = -io_u->error;
 480                         clear_io_u(td, io_u);
 481                 } else if (io_u->resid) {
 482                         int bytes = io_u->xfer_buflen - io_u->resid;
 483                         struct fio_file *f = io_u->file;
 484
 485                         if (bytes_issued)
 486                                 *bytes_issued += bytes;
 487
 488                         if (!from_verify)
 489                                 trim_io_piece(td, io_u);
 490
 491                         /*
 492                          * zero read, fail
 493                          */
 494                         if (!bytes) {
 495                                 if (!from_verify)
 496                                         unlog_io_piece(td, io_u);
 497                                 td_verror(td, EIO, "full resid");
 498                                 put_io_u(td, io_u);
 499                                 break;
 500                         }
 501
 502                         io_u->xfer_buflen = io_u->resid;
 503                         io_u->xfer_buf += bytes;
 504                         io_u->offset += bytes;
 505
 506                         if (ddir_rw(io_u->ddir))
 507                                 td->ts.short_io_u[io_u->ddir]++;
 508
 509                         f = io_u->file;
 510                         if (io_u->offset == f->real_file_size)
 511                                 goto sync_done;
 512
 513                         requeue_io_u(td, &io_u);
 514                 } else {
 515 sync_done:
 516                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 517                             __should_check_rate(td, DDIR_WRITE) ||
 518                             __should_check_rate(td, DDIR_TRIM)))
 519                                 fio_gettime(comp_time, NULL);
 520
 521                         *ret = io_u_sync_complete(td, io_u);
 522                         if (*ret < 0)
 523                                 break;
 524                 }
 525
 526                 /*
 527                  * when doing I/O (not when verifying),
 528                  * check for any errors that are to be ignored
 529                  */
 530                 if (!from_verify)
 531                         break;
 532
 533                 return 0;
 534         case FIO_Q_QUEUED:
 535                 /*
 536                  * if the engine doesn't have a commit hook,
 537                  * the io_u is really queued. if it does have such
 538                  * a hook, it has to call io_u_queued() itself.
 539                  */
 540                 if (td->io_ops->commit == NULL)
 541                         io_u_queued(td, io_u);
 542                 if (bytes_issued)
 543                         *bytes_issued += io_u->xfer_buflen;
 544                 break;
 545         case FIO_Q_BUSY:
 546                 if (!from_verify)
 547                         unlog_io_piece(td, io_u);
 548                 requeue_io_u(td, &io_u);
 549                 ret2 = td_io_commit(td);
 550                 if (ret2 < 0)
 551                         *ret = ret2;
 552                 break;
 553         default:
 554                 assert(*ret < 0);
 555                 td_verror(td, -(*ret), "td_io_queue");
 556                 break;
 557         }
 558
 559         if (break_on_this_error(td, ddir, ret))
 560                 return 1;
 561
 562         return 0;
 563 }
 564
 565 static inline bool io_in_polling(struct thread_data *td)
 566 {
 567         return !td->o.iodepth_batch_complete_min &&
 568                    !td->o.iodepth_batch_complete_max;
 569 }
 570
 571 /*
 572  * The main verify engine. Runs over the writes we previously submitted,
 573  * reads the blocks back in, and checks the crc/md5 of the data.
 574  */
 575 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 576 {
 577         struct fio_file *f;
 578         struct io_u *io_u;
 579         int ret, min_events;
 580         unsigned int i;
 581
 582         dprint(FD_VERIFY, "starting loop\n");
 583
 584         /*
 585          * sync io first and invalidate cache, to make sure we really
 586          * read from disk.
 587          */
 588         for_each_file(td, f, i) {
 589                 if (!fio_file_open(f))
 590                         continue;
 591                 if (fio_io_sync(td, f))
 592                         break;
 593                 if (file_invalidate_cache(td, f))
 594                         break;
 595         }
 596
 597         check_update_rusage(td);
 598
 599         if (td->error)
 600                 return;
 601
 602         td_set_runstate(td, TD_VERIFYING);
 603
 604         io_u = NULL;
 605         while (!td->terminate) {
 606                 enum fio_ddir ddir;
 607                 int full;
 608
 609                 update_tv_cache(td);
 610                 check_update_rusage(td);
 611
 612                 if (runtime_exceeded(td, &td->tv_cache)) {
 613                         __update_tv_cache(td);
 614                         if (runtime_exceeded(td, &td->tv_cache)) {
 615                                 fio_mark_td_terminate(td);
 616                                 break;
 617                         }
 618                 }
 619
 620                 if (flow_threshold_exceeded(td))
 621                         continue;
 622
 623                 if (!td->o.experimental_verify) {
 624                         io_u = __get_io_u(td);
 625                         if (!io_u)
 626                                 break;
 627
 628                         if (get_next_verify(td, io_u)) {
 629                                 put_io_u(td, io_u);
 630                                 break;
 631                         }
 632
 633                         if (td_io_prep(td, io_u)) {
 634                                 put_io_u(td, io_u);
 635                                 break;
 636                         }
 637                 } else {
 638                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 639                                 break;
 640
 641                         while ((io_u = get_io_u(td)) != NULL) {
 642                                 if (IS_ERR(io_u)) {
 643                                         io_u = NULL;
 644                                         ret = FIO_Q_BUSY;
 645                                         goto reap;
 646                                 }
 647
 648                                 /*
 649                                  * We are only interested in the places where
 650                                  * we wrote or trimmed IOs. Turn those into
 651                                  * reads for verification purposes.
 652                                  */
 653                                 if (io_u->ddir == DDIR_READ) {
 654                                         /*
 655                                          * Pretend we issued it for rwmix
 656                                          * accounting
 657                                          */
 658                                         td->io_issues[DDIR_READ]++;
 659                                         put_io_u(td, io_u);
 660                                         continue;
 661                                 } else if (io_u->ddir == DDIR_TRIM) {
 662                                         io_u->ddir = DDIR_READ;
 663                                         io_u_set(io_u, IO_U_F_TRIMMED);
 664                                         break;
 665                                 } else if (io_u->ddir == DDIR_WRITE) {
 666                                         io_u->ddir = DDIR_READ;
 667                                         break;
 668                                 } else {
 669                                         put_io_u(td, io_u);
 670                                         continue;
 671                                 }
 672                         }
 673
 674                         if (!io_u)
 675                                 break;
 676                 }
 677
 678                 if (verify_state_should_stop(td, io_u)) {
 679                         put_io_u(td, io_u);
 680                         break;
 681                 }
 682
 683                 if (td->o.verify_async)
 684                         io_u->end_io = verify_io_u_async;
 685                 else
 686                         io_u->end_io = verify_io_u;
 687
 688                 ddir = io_u->ddir;
 689                 if (!td->o.disable_slat)
 690                         fio_gettime(&io_u->start_time, NULL);
 691
 692                 ret = td_io_queue(td, io_u);
 693
 694                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 695                         break;
 696
 697                 /*
 698                  * if we can queue more, do so. but check if there are
 699                  * completed io_u's first. Note that we can get BUSY even
 700                  * without IO queued, if the system is resource starved.
 701                  */
 702 reap:
 703                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 704                 if (full || io_in_polling(td))
 705                         ret = wait_for_completions(td, NULL);
 706
 707                 if (ret < 0)
 708                         break;
 709         }
 710
 711         check_update_rusage(td);
 712
 713         if (!td->error) {
 714                 min_events = td->cur_depth;
 715
 716                 if (min_events)
 717                         ret = io_u_queued_complete(td, min_events);
 718         } else
 719                 cleanup_pending_aio(td);
 720
 721         td_set_runstate(td, TD_RUNNING);
 722
 723         dprint(FD_VERIFY, "exiting loop\n");
 724 }
 725
 726 static bool exceeds_number_ios(struct thread_data *td)
 727 {
 728         unsigned long long number_ios;
 729
 730         if (!td->o.number_ios)
 731                 return false;
 732
 733         number_ios = ddir_rw_sum(td->io_blocks);
 734         number_ios += td->io_u_queued + td->io_u_in_flight;
 735
 736         return number_ios >= (td->o.number_ios * td->loops);
 737 }
 738
 739 static bool io_issue_bytes_exceeded(struct thread_data *td)
 740 {
 741         unsigned long long bytes, limit;
 742
 743         if (td_rw(td))
 744                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 745         else if (td_write(td))
 746                 bytes = td->io_issue_bytes[DDIR_WRITE];
 747         else if (td_read(td))
 748                 bytes = td->io_issue_bytes[DDIR_READ];
 749         else
 750                 bytes = td->io_issue_bytes[DDIR_TRIM];
 751
 752         if (td->o.io_limit)
 753                 limit = td->o.io_limit;
 754         else
 755                 limit = td->o.size;
 756
 757         limit *= td->loops;
 758         return bytes >= limit || exceeds_number_ios(td);
 759 }
 760
 761 static bool io_complete_bytes_exceeded(struct thread_data *td)
 762 {
 763         unsigned long long bytes, limit;
 764
 765         if (td_rw(td))
 766                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 767         else if (td_write(td))
 768                 bytes = td->this_io_bytes[DDIR_WRITE];
 769         else if (td_read(td))
 770                 bytes = td->this_io_bytes[DDIR_READ];
 771         else
 772                 bytes = td->this_io_bytes[DDIR_TRIM];
 773
 774         if (td->o.io_limit)
 775                 limit = td->o.io_limit;
 776         else
 777                 limit = td->o.size;
 778
 779         limit *= td->loops;
 780         return bytes >= limit || exceeds_number_ios(td);
 781 }
 782
 783 /*
 784  * used to calculate the next io time for rate control
 785  *
 786  */
 787 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 788 {
 789         uint64_t secs, remainder, bps, bytes, iops;
 790
 791         assert(!(td->flags & TD_F_CHILD));
 792         bytes = td->rate_io_issue_bytes[ddir];
 793         bps = td->rate_bps[ddir];
 794
 795         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 796                 uint64_t val;
 797                 iops = bps / td->o.bs[ddir];
 798                 val = (int64_t) (1000000 / iops) *
 799                                 -logf(__rand_0_1(&td->poisson_state));
 800                 if (val) {
 801                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 802                                         (unsigned long long) 1000000 / val);
 803                 }
 804                 td->last_usec += val;
 805                 return td->last_usec;
 806         } else if (bps) {
 807                 secs = bytes / bps;
 808                 remainder = bytes % bps;
 809                 return remainder * 1000000 / bps + secs * 1000000;
 810         }
 811
 812         return 0;
 813 }
 814
 815 /*
 816  * Main IO worker function. It retrieves io_u's to process and queues
 817  * and reaps them, checking for rate and errors along the way.
 818  *
 819  * Returns number of bytes written and trimmed.
 820  */
 821 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 822 {
 823         unsigned int i;
 824         int ret = 0;
 825         uint64_t total_bytes, bytes_issued = 0;
 826
 827         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 828                 bytes_done[i] = td->bytes_done[i];
 829
 830         if (in_ramp_time(td))
 831                 td_set_runstate(td, TD_RAMP);
 832         else
 833                 td_set_runstate(td, TD_RUNNING);
 834
 835         lat_target_init(td);
 836
 837         total_bytes = td->o.size;
 838         /*
 839         * Allow random overwrite workloads to write up to io_limit
 840         * before starting verification phase as 'size' doesn't apply.
 841         */
 842         if (td_write(td) && td_random(td) && td->o.norandommap)
 843                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 844         /*
 845          * If verify_backlog is enabled, we'll run the verify in this
 846          * handler as well. For that case, we may need up to twice the
 847          * amount of bytes.
 848          */
 849         if (td->o.verify != VERIFY_NONE &&
 850            (td_write(td) && td->o.verify_backlog))
 851                 total_bytes += td->o.size;
 852
 853         /* In trimwrite mode, each byte is trimmed and then written, so
 854          * allow total_bytes to be twice as big */
 855         if (td_trimwrite(td))
 856                 total_bytes += td->total_io_size;
 857
 858         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 859                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 860                 td->o.time_based) {
 861                 struct timeval comp_time;
 862                 struct io_u *io_u;
 863                 int full;
 864                 enum fio_ddir ddir;
 865
 866                 check_update_rusage(td);
 867
 868                 if (td->terminate || td->done)
 869                         break;
 870
 871                 update_tv_cache(td);
 872
 873                 if (runtime_exceeded(td, &td->tv_cache)) {
 874                         __update_tv_cache(td);
 875                         if (runtime_exceeded(td, &td->tv_cache)) {
 876                                 fio_mark_td_terminate(td);
 877                                 break;
 878                         }
 879                 }
 880
 881                 if (flow_threshold_exceeded(td))
 882                         continue;
 883
 884                 /*
 885                  * Break if we exceeded the bytes. The exception is time
 886                  * based runs, but we still need to break out of the loop
 887                  * for those to run verification, if enabled.
 888                  */
 889                 if (bytes_issued >= total_bytes &&
 890                     (!td->o.time_based ||
 891                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 892                         break;
 893
 894                 io_u = get_io_u(td);
 895                 if (IS_ERR_OR_NULL(io_u)) {
 896                         int err = PTR_ERR(io_u);
 897
 898                         io_u = NULL;
 899                         if (err == -EBUSY) {
 900                                 ret = FIO_Q_BUSY;
 901                                 goto reap;
 902                         }
 903                         if (td->o.latency_target)
 904                                 goto reap;
 905                         break;
 906                 }
 907
 908                 ddir = io_u->ddir;
 909
 910                 /*
 911                  * Add verification end_io handler if:
 912                  *      - Asked to verify (!td_rw(td))
 913                  *      - Or the io_u is from our verify list (mixed write/ver)
 914                  */
 915                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 916                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 917
 918                         if (!td->o.verify_pattern_bytes) {
 919                                 io_u->rand_seed = __rand(&td->verify_state);
 920                                 if (sizeof(int) != sizeof(long *))
 921                                         io_u->rand_seed *= __rand(&td->verify_state);
 922                         }
 923
 924                         if (verify_state_should_stop(td, io_u)) {
 925                                 put_io_u(td, io_u);
 926                                 break;
 927                         }
 928
 929                         if (td->o.verify_async)
 930                                 io_u->end_io = verify_io_u_async;
 931                         else
 932                                 io_u->end_io = verify_io_u;
 933                         td_set_runstate(td, TD_VERIFYING);
 934                 } else if (in_ramp_time(td))
 935                         td_set_runstate(td, TD_RAMP);
 936                 else
 937                         td_set_runstate(td, TD_RUNNING);
 938
 939                 /*
 940                  * Always log IO before it's issued, so we know the specific
 941                  * order of it. The logged unit will track when the IO has
 942                  * completed.
 943                  */
 944                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 945                     td->o.do_verify &&
 946                     td->o.verify != VERIFY_NONE &&
 947                     !td->o.experimental_verify)
 948                         log_io_piece(td, io_u);
 949
 950                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 951                         const unsigned long blen = io_u->xfer_buflen;
 952                         const enum fio_ddir ddir = acct_ddir(io_u);
 953
 954                         if (td->error)
 955                                 break;
 956
 957                         workqueue_enqueue(&td->io_wq, &io_u->work);
 958                         ret = FIO_Q_QUEUED;
 959
 960                         if (ddir_rw(ddir)) {
 961                                 td->io_issues[ddir]++;
 962                                 td->io_issue_bytes[ddir] += blen;
 963                                 td->rate_io_issue_bytes[ddir] += blen;
 964                         }
 965
 966                         if (should_check_rate(td))
 967                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 968
 969                 } else {
 970                         ret = td_io_queue(td, io_u);
 971
 972                         if (should_check_rate(td))
 973                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 974
 975                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 976                                 break;
 977
 978                         /*
 979                          * See if we need to complete some commands. Note that
 980                          * we can get BUSY even without IO queued, if the
 981                          * system is resource starved.
 982                          */
 983 reap:
 984                         full = queue_full(td) ||
 985                                 (ret == FIO_Q_BUSY && td->cur_depth);
 986                         if (full || io_in_polling(td))
 987                                 ret = wait_for_completions(td, &comp_time);
 988                 }
 989                 if (ret < 0)
 990                         break;
 991                 if (!ddir_rw_sum(td->bytes_done) &&
 992                     !(td->io_ops->flags & FIO_NOIO))
 993                         continue;
 994
 995                 if (!in_ramp_time(td) && should_check_rate(td)) {
 996                         if (check_min_rate(td, &comp_time)) {
 997                                 if (exitall_on_terminate || td->o.exitall_error)
 998                                         fio_terminate_threads(td->groupid);
 999                                 td_verror(td, EIO, "check_min_rate");
1000                                 break;
1001                         }
1002                 }
1003                 if (!in_ramp_time(td) && td->o.latency_target)
1004                         lat_target_check(td);
1005
1006                 if (td->o.thinktime) {
1007                         unsigned long long b;
1008
1009                         b = ddir_rw_sum(td->io_blocks);
1010                         if (!(b % td->o.thinktime_blocks)) {
1011                                 int left;
1012
1013                                 io_u_quiesce(td);
1014
1015                                 if (td->o.thinktime_spin)
1016                                         usec_spin(td->o.thinktime_spin);
1017
1018                                 left = td->o.thinktime - td->o.thinktime_spin;
1019                                 if (left)
1020                                         usec_sleep(td, left);
1021                         }
1022                 }
1023         }
1024
1025         check_update_rusage(td);
1026
1027         if (td->trim_entries)
1028                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1029
1030         if (td->o.fill_device && td->error == ENOSPC) {
1031                 td->error = 0;
1032                 fio_mark_td_terminate(td);
1033         }
1034         if (!td->error) {
1035                 struct fio_file *f;
1036
1037                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1038                         workqueue_flush(&td->io_wq);
1039                         i = 0;
1040                 } else
1041                         i = td->cur_depth;
1042
1043                 if (i) {
1044                         ret = io_u_queued_complete(td, i);
1045                         if (td->o.fill_device && td->error == ENOSPC)
1046                                 td->error = 0;
1047                 }
1048
1049                 if (should_fsync(td) && td->o.end_fsync) {
1050                         td_set_runstate(td, TD_FSYNCING);
1051
1052                         for_each_file(td, f, i) {
1053                                 if (!fio_file_fsync(td, f))
1054                                         continue;
1055
1056                                 log_err("fio: end_fsync failed for file %s\n",
1057                                                                 f->file_name);
1058                         }
1059                 }
1060         } else
1061                 cleanup_pending_aio(td);
1062
1063         /*
1064          * stop job if we failed doing any IO
1065          */
1066         if (!ddir_rw_sum(td->this_io_bytes))
1067                 td->done = 1;
1068
1069         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1070                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1071 }
1072
1073 static void free_file_completion_logging(struct thread_data *td)
1074 {
1075         struct fio_file *f;
1076         unsigned int i;
1077
1078         for_each_file(td, f, i) {
1079                 if (!f->last_write_comp)
1080                         break;
1081                 sfree(f->last_write_comp);
1082         }
1083 }
1084
1085 static int init_file_completion_logging(struct thread_data *td,
1086                                         unsigned int depth)
1087 {
1088         struct fio_file *f;
1089         unsigned int i;
1090
1091         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1092                 return 0;
1093
1094         for_each_file(td, f, i) {
1095                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1096                 if (!f->last_write_comp)
1097                         goto cleanup;
1098         }
1099
1100         return 0;
1101
1102 cleanup:
1103         free_file_completion_logging(td);
1104         log_err("fio: failed to alloc write comp data\n");
1105         return 1;
1106 }
1107
1108 static void cleanup_io_u(struct thread_data *td)
1109 {
1110         struct io_u *io_u;
1111
1112         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1113
1114                 if (td->io_ops->io_u_free)
1115                         td->io_ops->io_u_free(td, io_u);
1116
1117                 fio_memfree(io_u, sizeof(*io_u));
1118         }
1119
1120         free_io_mem(td);
1121
1122         io_u_rexit(&td->io_u_requeues);
1123         io_u_qexit(&td->io_u_freelist);
1124         io_u_qexit(&td->io_u_all);
1125
1126         free_file_completion_logging(td);
1127 }
1128
1129 static int init_io_u(struct thread_data *td)
1130 {
1131         struct io_u *io_u;
1132         unsigned int max_bs, min_write;
1133         int cl_align, i, max_units;
1134         int data_xfer = 1, err;
1135         char *p;
1136
1137         max_units = td->o.iodepth;
1138         max_bs = td_max_bs(td);
1139         min_write = td->o.min_bs[DDIR_WRITE];
1140         td->orig_buffer_size = (unsigned long long) max_bs
1141                                         * (unsigned long long) max_units;
1142
1143         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1144                 data_xfer = 0;
1145
1146         err = 0;
1147         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1148         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1149         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1150
1151         if (err) {
1152                 log_err("fio: failed setting up IO queues\n");
1153                 return 1;
1154         }
1155
1156         /*
1157          * if we may later need to do address alignment, then add any
1158          * possible adjustment here so that we don't cause a buffer
1159          * overflow later. this adjustment may be too much if we get
1160          * lucky and the allocator gives us an aligned address.
1161          */
1162         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1163             (td->io_ops->flags & FIO_RAWIO))
1164                 td->orig_buffer_size += page_mask + td->o.mem_align;
1165
1166         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1167                 unsigned long bs;
1168
1169                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1170                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1171         }
1172
1173         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1174                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1175                 return 1;
1176         }
1177
1178         if (data_xfer && allocate_io_mem(td))
1179                 return 1;
1180
1181         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1182             (td->io_ops->flags & FIO_RAWIO))
1183                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1184         else
1185                 p = td->orig_buffer;
1186
1187         cl_align = os_cache_line_size();
1188
1189         for (i = 0; i < max_units; i++) {
1190                 void *ptr;
1191
1192                 if (td->terminate)
1193                         return 1;
1194
1195                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1196                 if (!ptr) {
1197                         log_err("fio: unable to allocate aligned memory\n");
1198                         break;
1199                 }
1200
1201                 io_u = ptr;
1202                 memset(io_u, 0, sizeof(*io_u));
1203                 INIT_FLIST_HEAD(&io_u->verify_list);
1204                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1205
1206                 if (data_xfer) {
1207                         io_u->buf = p;
1208                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1209
1210                         if (td_write(td))
1211                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1212                         if (td_write(td) && td->o.verify_pattern_bytes) {
1213                                 /*
1214                                  * Fill the buffer with the pattern if we are
1215                                  * going to be doing writes.
1216                                  */
1217                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1218                         }
1219                 }
1220
1221                 io_u->index = i;
1222                 io_u->flags = IO_U_F_FREE;
1223                 io_u_qpush(&td->io_u_freelist, io_u);
1224
1225                 /*
1226                  * io_u never leaves this stack, used for iteration of all
1227                  * io_u buffers.
1228                  */
1229                 io_u_qpush(&td->io_u_all, io_u);
1230
1231                 if (td->io_ops->io_u_init) {
1232                         int ret = td->io_ops->io_u_init(td, io_u);
1233
1234                         if (ret) {
1235                                 log_err("fio: failed to init engine data: %d\n", ret);
1236                                 return 1;
1237                         }
1238                 }
1239
1240                 p += max_bs;
1241         }
1242
1243         if (init_file_completion_logging(td, max_units))
1244                 return 1;
1245
1246         return 0;
1247 }
1248
1249 static int switch_ioscheduler(struct thread_data *td)
1250 {
1251         char tmp[256], tmp2[128];
1252         FILE *f;
1253         int ret;
1254
1255         if (td->io_ops->flags & FIO_DISKLESSIO)
1256                 return 0;
1257
1258         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1259
1260         f = fopen(tmp, "r+");
1261         if (!f) {
1262                 if (errno == ENOENT) {
1263                         log_err("fio: os or kernel doesn't support IO scheduler"
1264                                 " switching\n");
1265                         return 0;
1266                 }
1267                 td_verror(td, errno, "fopen iosched");
1268                 return 1;
1269         }
1270
1271         /*
1272          * Set io scheduler.
1273          */
1274         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1275         if (ferror(f) || ret != 1) {
1276                 td_verror(td, errno, "fwrite");
1277                 fclose(f);
1278                 return 1;
1279         }
1280
1281         rewind(f);
1282
1283         /*
1284          * Read back and check that the selected scheduler is now the default.
1285          */
1286         memset(tmp, 0, sizeof(tmp));
1287         ret = fread(tmp, sizeof(tmp), 1, f);
1288         if (ferror(f) || ret < 0) {
1289                 td_verror(td, errno, "fread");
1290                 fclose(f);
1291                 return 1;
1292         }
1293         /*
1294          * either a list of io schedulers or "none\n" is expected.
1295          */
1296         tmp[strlen(tmp) - 1] = '\0';
1297
1298
1299         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1300         if (!strstr(tmp, tmp2)) {
1301                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1302                 td_verror(td, EINVAL, "iosched_switch");
1303                 fclose(f);
1304                 return 1;
1305         }
1306
1307         fclose(f);
1308         return 0;
1309 }
1310
1311 static bool keep_running(struct thread_data *td)
1312 {
1313         unsigned long long limit;
1314
1315         if (td->done)
1316                 return false;
1317         if (td->o.time_based)
1318                 return true;
1319         if (td->o.loops) {
1320                 td->o.loops--;
1321                 return true;
1322         }
1323         if (exceeds_number_ios(td))
1324                 return false;
1325
1326         if (td->o.io_limit)
1327                 limit = td->o.io_limit;
1328         else
1329                 limit = td->o.size;
1330
1331         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1332                 uint64_t diff;
1333
1334                 /*
1335                  * If the difference is less than the minimum IO size, we
1336                  * are done.
1337                  */
1338                 diff = limit - ddir_rw_sum(td->io_bytes);
1339                 if (diff < td_max_bs(td))
1340                         return false;
1341
1342                 if (fio_files_done(td) && !td->o.io_limit)
1343                         return false;
1344
1345                 return true;
1346         }
1347
1348         return false;
1349 }
1350
1351 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1352 {
1353         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1354         int ret;
1355         char *str;
1356
1357         str = malloc(newlen);
1358         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1359
1360         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1361         ret = system(str);
1362         if (ret == -1)
1363                 log_err("fio: exec of cmd <%s> failed\n", str);
1364
1365         free(str);
1366         return ret;
1367 }
1368
1369 /*
1370  * Dry run to compute correct state of numberio for verification.
1371  */
1372 static uint64_t do_dry_run(struct thread_data *td)
1373 {
1374         td_set_runstate(td, TD_RUNNING);
1375
1376         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1377                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1378                 struct io_u *io_u;
1379                 int ret;
1380
1381                 if (td->terminate || td->done)
1382                         break;
1383
1384                 io_u = get_io_u(td);
1385                 if (!io_u)
1386                         break;
1387
1388                 io_u_set(io_u, IO_U_F_FLIGHT);
1389                 io_u->error = 0;
1390                 io_u->resid = 0;
1391                 if (ddir_rw(acct_ddir(io_u)))
1392                         td->io_issues[acct_ddir(io_u)]++;
1393                 if (ddir_rw(io_u->ddir)) {
1394                         io_u_mark_depth(td, 1);
1395                         td->ts.total_io_u[io_u->ddir]++;
1396                 }
1397
1398                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1399                     td->o.do_verify &&
1400                     td->o.verify != VERIFY_NONE &&
1401                     !td->o.experimental_verify)
1402                         log_io_piece(td, io_u);
1403
1404                 ret = io_u_sync_complete(td, io_u);
1405                 (void) ret;
1406         }
1407
1408         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1409 }
1410
1411 struct fork_data {
1412         struct thread_data *td;
1413         struct sk_out *sk_out;
1414 };
1415
1416 /*
1417  * Entry point for the thread based jobs. The process based jobs end up
1418  * here as well, after a little setup.
1419  */
1420 static void *thread_main(void *data)
1421 {
1422         struct fork_data *fd = data;
1423         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1424         struct thread_data *td = fd->td;
1425         struct thread_options *o = &td->o;
1426         struct sk_out *sk_out = fd->sk_out;
1427         pthread_condattr_t attr;
1428         int clear_state;
1429         int ret;
1430
1431         sk_out_assign(sk_out);
1432         free(fd);
1433
1434         if (!o->use_thread) {
1435                 setsid();
1436                 td->pid = getpid();
1437         } else
1438                 td->pid = gettid();
1439
1440         fio_local_clock_init(o->use_thread);
1441
1442         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1443
1444         if (is_backend)
1445                 fio_server_send_start(td);
1446
1447         INIT_FLIST_HEAD(&td->io_log_list);
1448         INIT_FLIST_HEAD(&td->io_hist_list);
1449         INIT_FLIST_HEAD(&td->verify_list);
1450         INIT_FLIST_HEAD(&td->trim_list);
1451         INIT_FLIST_HEAD(&td->next_rand_list);
1452         pthread_mutex_init(&td->io_u_lock, NULL);
1453         td->io_hist_tree = RB_ROOT;
1454
1455         pthread_condattr_init(&attr);
1456         pthread_cond_init(&td->verify_cond, &attr);
1457         pthread_cond_init(&td->free_cond, &attr);
1458
1459         td_set_runstate(td, TD_INITIALIZED);
1460         dprint(FD_MUTEX, "up startup_mutex\n");
1461         fio_mutex_up(startup_mutex);
1462         dprint(FD_MUTEX, "wait on td->mutex\n");
1463         fio_mutex_down(td->mutex);
1464         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1465
1466         /*
1467          * A new gid requires privilege, so we need to do this before setting
1468          * the uid.
1469          */
1470         if (o->gid != -1U && setgid(o->gid)) {
1471                 td_verror(td, errno, "setgid");
1472                 goto err;
1473         }
1474         if (o->uid != -1U && setuid(o->uid)) {
1475                 td_verror(td, errno, "setuid");
1476                 goto err;
1477         }
1478
1479         /*
1480          * If we have a gettimeofday() thread, make sure we exclude that
1481          * thread from this job
1482          */
1483         if (o->gtod_cpu)
1484                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1485
1486         /*
1487          * Set affinity first, in case it has an impact on the memory
1488          * allocations.
1489          */
1490         if (fio_option_is_set(o, cpumask)) {
1491                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1492                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1493                         if (!ret) {
1494                                 log_err("fio: no CPUs set\n");
1495                                 log_err("fio: Try increasing number of available CPUs\n");
1496                                 td_verror(td, EINVAL, "cpus_split");
1497                                 goto err;
1498                         }
1499                 }
1500                 ret = fio_setaffinity(td->pid, o->cpumask);
1501                 if (ret == -1) {
1502                         td_verror(td, errno, "cpu_set_affinity");
1503                         goto err;
1504                 }
1505         }
1506
1507 #ifdef CONFIG_LIBNUMA
1508         /* numa node setup */
1509         if (fio_option_is_set(o, numa_cpunodes) ||
1510             fio_option_is_set(o, numa_memnodes)) {
1511                 struct bitmask *mask;
1512
1513                 if (numa_available() < 0) {
1514                         td_verror(td, errno, "Does not support NUMA API\n");
1515                         goto err;
1516                 }
1517
1518                 if (fio_option_is_set(o, numa_cpunodes)) {
1519                         mask = numa_parse_nodestring(o->numa_cpunodes);
1520                         ret = numa_run_on_node_mask(mask);
1521                         numa_free_nodemask(mask);
1522                         if (ret == -1) {
1523                                 td_verror(td, errno, \
1524                                         "numa_run_on_node_mask failed\n");
1525                                 goto err;
1526                         }
1527                 }
1528
1529                 if (fio_option_is_set(o, numa_memnodes)) {
1530                         mask = NULL;
1531                         if (o->numa_memnodes)
1532                                 mask = numa_parse_nodestring(o->numa_memnodes);
1533
1534                         switch (o->numa_mem_mode) {
1535                         case MPOL_INTERLEAVE:
1536                                 numa_set_interleave_mask(mask);
1537                                 break;
1538                         case MPOL_BIND:
1539                                 numa_set_membind(mask);
1540                                 break;
1541                         case MPOL_LOCAL:
1542                                 numa_set_localalloc();
1543                                 break;
1544                         case MPOL_PREFERRED:
1545                                 numa_set_preferred(o->numa_mem_prefer_node);
1546                                 break;
1547                         case MPOL_DEFAULT:
1548                         default:
1549                                 break;
1550                         }
1551
1552                         if (mask)
1553                                 numa_free_nodemask(mask);
1554
1555                 }
1556         }
1557 #endif
1558
1559         if (fio_pin_memory(td))
1560                 goto err;
1561
1562         /*
1563          * May alter parameters that init_io_u() will use, so we need to
1564          * do this first.
1565          */
1566         if (init_iolog(td))
1567                 goto err;
1568
1569         if (init_io_u(td))
1570                 goto err;
1571
1572         if (o->verify_async && verify_async_init(td))
1573                 goto err;
1574
1575         if (fio_option_is_set(o, ioprio) ||
1576             fio_option_is_set(o, ioprio_class)) {
1577                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1578                 if (ret == -1) {
1579                         td_verror(td, errno, "ioprio_set");
1580                         goto err;
1581                 }
1582         }
1583
1584         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1585                 goto err;
1586
1587         errno = 0;
1588         if (nice(o->nice) == -1 && errno != 0) {
1589                 td_verror(td, errno, "nice");
1590                 goto err;
1591         }
1592
1593         if (o->ioscheduler && switch_ioscheduler(td))
1594                 goto err;
1595
1596         if (!o->create_serialize && setup_files(td))
1597                 goto err;
1598
1599         if (td_io_init(td))
1600                 goto err;
1601
1602         if (init_random_map(td))
1603                 goto err;
1604
1605         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1606                 goto err;
1607
1608         if (o->pre_read) {
1609                 if (pre_read_files(td) < 0)
1610                         goto err;
1611         }
1612
1613         if (iolog_compress_init(td, sk_out))
1614                 goto err;
1615
1616         fio_verify_init(td);
1617
1618         if (rate_submit_init(td, sk_out))
1619                 goto err;
1620
1621         fio_gettime(&td->epoch, NULL);
1622         fio_getrusage(&td->ru_start);
1623         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1624         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1625
1626         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1627                         o->ratemin[DDIR_TRIM]) {
1628                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1629                                         sizeof(td->bw_sample_time));
1630                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1631                                         sizeof(td->bw_sample_time));
1632                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1633                                         sizeof(td->bw_sample_time));
1634         }
1635
1636         clear_state = 0;
1637         while (keep_running(td)) {
1638                 uint64_t verify_bytes;
1639
1640                 fio_gettime(&td->start, NULL);
1641                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1642
1643                 if (clear_state)
1644                         clear_io_state(td, 0);
1645
1646                 prune_io_piece_log(td);
1647
1648                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1649                         verify_bytes = do_dry_run(td);
1650                 else {
1651                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1652
1653                         do_io(td, bytes_done);
1654
1655                         if (!ddir_rw_sum(bytes_done)) {
1656                                 fio_mark_td_terminate(td);
1657                                 verify_bytes = 0;
1658                         } else {
1659                                 verify_bytes = bytes_done[DDIR_WRITE] +
1660                                                 bytes_done[DDIR_TRIM];
1661                         }
1662                 }
1663
1664                 clear_state = 1;
1665
1666                 /*
1667                  * Make sure we've successfully updated the rusage stats
1668                  * before waiting on the stat mutex. Otherwise we could have
1669                  * the stat thread holding stat mutex and waiting for
1670                  * the rusage_sem, which would never get upped because
1671                  * this thread is waiting for the stat mutex.
1672                  */
1673                 check_update_rusage(td);
1674
1675                 fio_mutex_down(stat_mutex);
1676                 if (td_read(td) && td->io_bytes[DDIR_READ])
1677                         update_runtime(td, elapsed_us, DDIR_READ);
1678                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1679                         update_runtime(td, elapsed_us, DDIR_WRITE);
1680                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1681                         update_runtime(td, elapsed_us, DDIR_TRIM);
1682                 fio_gettime(&td->start, NULL);
1683                 fio_mutex_up(stat_mutex);
1684
1685                 if (td->error || td->terminate)
1686                         break;
1687
1688                 if (!o->do_verify ||
1689                     o->verify == VERIFY_NONE ||
1690                     (td->io_ops->flags & FIO_UNIDIR))
1691                         continue;
1692
1693                 clear_io_state(td, 0);
1694
1695                 fio_gettime(&td->start, NULL);
1696
1697                 do_verify(td, verify_bytes);
1698
1699                 /*
1700                  * See comment further up for why this is done here.
1701                  */
1702                 check_update_rusage(td);
1703
1704                 fio_mutex_down(stat_mutex);
1705                 update_runtime(td, elapsed_us, DDIR_READ);
1706                 fio_gettime(&td->start, NULL);
1707                 fio_mutex_up(stat_mutex);
1708
1709                 if (td->error || td->terminate)
1710                         break;
1711         }
1712
1713         update_rusage_stat(td);
1714         td->ts.total_run_time = mtime_since_now(&td->epoch);
1715         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1716         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1717         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1718
1719         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1720             (td->o.verify != VERIFY_NONE && td_write(td)))
1721                 verify_save_state(td->thread_number);
1722
1723         fio_unpin_memory(td);
1724
1725         fio_writeout_logs(td);
1726
1727         iolog_compress_exit(td);
1728         rate_submit_exit(td);
1729
1730         if (o->exec_postrun)
1731                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1732
1733         if (exitall_on_terminate || (o->exitall_error && td->error))
1734                 fio_terminate_threads(td->groupid);
1735
1736 err:
1737         if (td->error)
1738                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1739                                                         td->verror);
1740
1741         if (o->verify_async)
1742                 verify_async_exit(td);
1743
1744         close_and_free_files(td);
1745         cleanup_io_u(td);
1746         close_ioengine(td);
1747         cgroup_shutdown(td, &cgroup_mnt);
1748         verify_free_state(td);
1749
1750         if (td->zone_state_index) {
1751                 int i;
1752
1753                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1754                         free(td->zone_state_index[i]);
1755                 free(td->zone_state_index);
1756                 td->zone_state_index = NULL;
1757         }
1758
1759         if (fio_option_is_set(o, cpumask)) {
1760                 ret = fio_cpuset_exit(&o->cpumask);
1761                 if (ret)
1762                         td_verror(td, ret, "fio_cpuset_exit");
1763         }
1764
1765         /*
1766          * do this very late, it will log file closing as well
1767          */
1768         if (o->write_iolog_file)
1769                 write_iolog_close(td);
1770
1771         fio_mutex_remove(td->mutex);
1772         td->mutex = NULL;
1773
1774         td_set_runstate(td, TD_EXITED);
1775
1776         /*
1777          * Do this last after setting our runstate to exited, so we
1778          * know that the stat thread is signaled.
1779          */
1780         check_update_rusage(td);
1781
1782         sk_out_drop();
1783         return (void *) (uintptr_t) td->error;
1784 }
1785
1786
1787 /*
1788  * We cannot pass the td data into a forked process, so attach the td and
1789  * pass it to the thread worker.
1790  */
1791 static int fork_main(struct sk_out *sk_out, int shmid, int offset)
1792 {
1793         struct fork_data *fd;
1794         void *data, *ret;
1795
1796 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1797         data = shmat(shmid, NULL, 0);
1798         if (data == (void *) -1) {
1799                 int __err = errno;
1800
1801                 perror("shmat");
1802                 return __err;
1803         }
1804 #else
1805         /*
1806          * HP-UX inherits shm mappings?
1807          */
1808         data = threads;
1809 #endif
1810
1811         fd = calloc(1, sizeof(*fd));
1812         fd->td = data + offset * sizeof(struct thread_data);
1813         fd->sk_out = sk_out;
1814         ret = thread_main(fd);
1815         shmdt(data);
1816         return (int) (uintptr_t) ret;
1817 }
1818
1819 static void dump_td_info(struct thread_data *td)
1820 {
1821         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1822                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1823                         (unsigned long) time_since_now(&td->terminate_time));
1824 }
1825
1826 /*
1827  * Run over the job map and reap the threads that have exited, if any.
1828  */
1829 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1830                          unsigned int *m_rate)
1831 {
1832         struct thread_data *td;
1833         unsigned int cputhreads, realthreads, pending;
1834         int i, status, ret;
1835
1836         /*
1837          * reap exited threads (TD_EXITED -> TD_REAPED)
1838          */
1839         realthreads = pending = cputhreads = 0;
1840         for_each_td(td, i) {
1841                 int flags = 0;
1842
1843                 /*
1844                  * ->io_ops is NULL for a thread that has closed its
1845                  * io engine
1846                  */
1847                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1848                         cputhreads++;
1849                 else
1850                         realthreads++;
1851
1852                 if (!td->pid) {
1853                         pending++;
1854                         continue;
1855                 }
1856                 if (td->runstate == TD_REAPED)
1857                         continue;
1858                 if (td->o.use_thread) {
1859                         if (td->runstate == TD_EXITED) {
1860                                 td_set_runstate(td, TD_REAPED);
1861                                 goto reaped;
1862                         }
1863                         continue;
1864                 }
1865
1866                 flags = WNOHANG;
1867                 if (td->runstate == TD_EXITED)
1868                         flags = 0;
1869
1870                 /*
1871                  * check if someone quit or got killed in an unusual way
1872                  */
1873                 ret = waitpid(td->pid, &status, flags);
1874                 if (ret < 0) {
1875                         if (errno == ECHILD) {
1876                                 log_err("fio: pid=%d disappeared %d\n",
1877                                                 (int) td->pid, td->runstate);
1878                                 td->sig = ECHILD;
1879                                 td_set_runstate(td, TD_REAPED);
1880                                 goto reaped;
1881                         }
1882                         perror("waitpid");
1883                 } else if (ret == td->pid) {
1884                         if (WIFSIGNALED(status)) {
1885                                 int sig = WTERMSIG(status);
1886
1887                                 if (sig != SIGTERM && sig != SIGUSR2)
1888                                         log_err("fio: pid=%d, got signal=%d\n",
1889                                                         (int) td->pid, sig);
1890                                 td->sig = sig;
1891                                 td_set_runstate(td, TD_REAPED);
1892                                 goto reaped;
1893                         }
1894                         if (WIFEXITED(status)) {
1895                                 if (WEXITSTATUS(status) && !td->error)
1896                                         td->error = WEXITSTATUS(status);
1897
1898                                 td_set_runstate(td, TD_REAPED);
1899                                 goto reaped;
1900                         }
1901                 }
1902
1903                 /*
1904                  * If the job is stuck, do a forceful timeout of it and
1905                  * move on.
1906                  */
1907                 if (td->terminate &&
1908                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1909                         dump_td_info(td);
1910                         td_set_runstate(td, TD_REAPED);
1911                         goto reaped;
1912                 }
1913
1914                 /*
1915                  * thread is not dead, continue
1916                  */
1917                 pending++;
1918                 continue;
1919 reaped:
1920                 (*nr_running)--;
1921                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1922                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1923                 if (!td->pid)
1924                         pending--;
1925
1926                 if (td->error)
1927                         exit_value++;
1928
1929                 done_secs += mtime_since_now(&td->epoch) / 1000;
1930                 profile_td_exit(td);
1931         }
1932
1933         if (*nr_running == cputhreads && !pending && realthreads)
1934                 fio_terminate_threads(TERMINATE_ALL);
1935 }
1936
1937 static bool __check_trigger_file(void)
1938 {
1939         struct stat sb;
1940
1941         if (!trigger_file)
1942                 return false;
1943
1944         if (stat(trigger_file, &sb))
1945                 return false;
1946
1947         if (unlink(trigger_file) < 0)
1948                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1949                                                         strerror(errno));
1950
1951         return true;
1952 }
1953
1954 static bool trigger_timedout(void)
1955 {
1956         if (trigger_timeout)
1957                 return time_since_genesis() >= trigger_timeout;
1958
1959         return false;
1960 }
1961
1962 void exec_trigger(const char *cmd)
1963 {
1964         int ret;
1965
1966         if (!cmd)
1967                 return;
1968
1969         ret = system(cmd);
1970         if (ret == -1)
1971                 log_err("fio: failed executing %s trigger\n", cmd);
1972 }
1973
1974 void check_trigger_file(void)
1975 {
1976         if (__check_trigger_file() || trigger_timedout()) {
1977                 if (nr_clients)
1978                         fio_clients_send_trigger(trigger_remote_cmd);
1979                 else {
1980                         verify_save_state(IO_LIST_ALL);
1981                         fio_terminate_threads(TERMINATE_ALL);
1982                         exec_trigger(trigger_cmd);
1983                 }
1984         }
1985 }
1986
1987 static int fio_verify_load_state(struct thread_data *td)
1988 {
1989         int ret;
1990
1991         if (!td->o.verify_state)
1992                 return 0;
1993
1994         if (is_backend) {
1995                 void *data;
1996
1997                 ret = fio_server_get_verify_state(td->o.name,
1998                                         td->thread_number - 1, &data);
1999                 if (!ret)
2000                         verify_assign_state(td, data);
2001         } else
2002                 ret = verify_load_state(td, "local");
2003
2004         return ret;
2005 }
2006
2007 static void do_usleep(unsigned int usecs)
2008 {
2009         check_for_running_stats();
2010         check_trigger_file();
2011         usleep(usecs);
2012 }
2013
2014 static bool check_mount_writes(struct thread_data *td)
2015 {
2016         struct fio_file *f;
2017         unsigned int i;
2018
2019         if (!td_write(td) || td->o.allow_mounted_write)
2020                 return false;
2021
2022         for_each_file(td, f, i) {
2023                 if (f->filetype != FIO_TYPE_BD)
2024                         continue;
2025                 if (device_is_mounted(f->file_name))
2026                         goto mounted;
2027         }
2028
2029         return false;
2030 mounted:
2031         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2032         return true;
2033 }
2034
2035 static bool waitee_running(struct thread_data *me)
2036 {
2037         const char *waitee = me->o.wait_for;
2038         const char *self = me->o.name;
2039         struct thread_data *td;
2040         int i;
2041
2042         if (!waitee)
2043                 return false;
2044
2045         for_each_td(td, i) {
2046                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2047                         continue;
2048
2049                 if (td->runstate < TD_EXITED) {
2050                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2051                                         self, td->o.name,
2052                                         runstate_to_name(td->runstate));
2053                         return true;
2054                 }
2055         }
2056
2057         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2058         return false;
2059 }
2060
2061 /*
2062  * Main function for kicking off and reaping jobs, as needed.
2063  */
2064 static void run_threads(struct sk_out *sk_out)
2065 {
2066         struct thread_data *td;
2067         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2068         uint64_t spent;
2069
2070         if (fio_gtod_offload && fio_start_gtod_thread())
2071                 return;
2072
2073         fio_idle_prof_init();
2074
2075         set_sig_handlers();
2076
2077         nr_thread = nr_process = 0;
2078         for_each_td(td, i) {
2079                 if (check_mount_writes(td))
2080                         return;
2081                 if (td->o.use_thread)
2082                         nr_thread++;
2083                 else
2084                         nr_process++;
2085         }
2086
2087         if (output_format & FIO_OUTPUT_NORMAL) {
2088                 log_info("Starting ");
2089                 if (nr_thread)
2090                         log_info("%d thread%s", nr_thread,
2091                                                 nr_thread > 1 ? "s" : "");
2092                 if (nr_process) {
2093                         if (nr_thread)
2094                                 log_info(" and ");
2095                         log_info("%d process%s", nr_process,
2096                                                 nr_process > 1 ? "es" : "");
2097                 }
2098                 log_info("\n");
2099                 log_info_flush();
2100         }
2101
2102         todo = thread_number;
2103         nr_running = 0;
2104         nr_started = 0;
2105         m_rate = t_rate = 0;
2106
2107         for_each_td(td, i) {
2108                 print_status_init(td->thread_number - 1);
2109
2110                 if (!td->o.create_serialize)
2111                         continue;
2112
2113                 if (fio_verify_load_state(td))
2114                         goto reap;
2115
2116                 /*
2117                  * do file setup here so it happens sequentially,
2118                  * we don't want X number of threads getting their
2119                  * client data interspersed on disk
2120                  */
2121                 if (setup_files(td)) {
2122 reap:
2123                         exit_value++;
2124                         if (td->error)
2125                                 log_err("fio: pid=%d, err=%d/%s\n",
2126                                         (int) td->pid, td->error, td->verror);
2127                         td_set_runstate(td, TD_REAPED);
2128                         todo--;
2129                 } else {
2130                         struct fio_file *f;
2131                         unsigned int j;
2132
2133                         /*
2134                          * for sharing to work, each job must always open
2135                          * its own files. so close them, if we opened them
2136                          * for creation
2137                          */
2138                         for_each_file(td, f, j) {
2139                                 if (fio_file_open(f))
2140                                         td_io_close_file(td, f);
2141                         }
2142                 }
2143         }
2144
2145         /* start idle threads before io threads start to run */
2146         fio_idle_prof_start();
2147
2148         set_genesis_time();
2149
2150         while (todo) {
2151                 struct thread_data *map[REAL_MAX_JOBS];
2152                 struct timeval this_start;
2153                 int this_jobs = 0, left;
2154
2155                 /*
2156                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2157                  */
2158                 for_each_td(td, i) {
2159                         if (td->runstate != TD_NOT_CREATED)
2160                                 continue;
2161
2162                         /*
2163                          * never got a chance to start, killed by other
2164                          * thread for some reason
2165                          */
2166                         if (td->terminate) {
2167                                 todo--;
2168                                 continue;
2169                         }
2170
2171                         if (td->o.start_delay) {
2172                                 spent = utime_since_genesis();
2173
2174                                 if (td->o.start_delay > spent)
2175                                         continue;
2176                         }
2177
2178                         if (td->o.stonewall && (nr_started || nr_running)) {
2179                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2180                                                         td->o.name);
2181                                 break;
2182                         }
2183
2184                         if (waitee_running(td)) {
2185                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2186                                                 td->o.name, td->o.wait_for);
2187                                 continue;
2188                         }
2189
2190                         init_disk_util(td);
2191
2192                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2193                         td->update_rusage = 0;
2194
2195                         /*
2196                          * Set state to created. Thread will transition
2197                          * to TD_INITIALIZED when it's done setting up.
2198                          */
2199                         td_set_runstate(td, TD_CREATED);
2200                         map[this_jobs++] = td;
2201                         nr_started++;
2202
2203                         if (td->o.use_thread) {
2204                                 struct fork_data *fd;
2205                                 int ret;
2206
2207                                 fd = calloc(1, sizeof(*fd));
2208                                 fd->td = td;
2209                                 fd->sk_out = sk_out;
2210
2211                                 dprint(FD_PROCESS, "will pthread_create\n");
2212                                 ret = pthread_create(&td->thread, NULL,
2213                                                         thread_main, fd);
2214                                 if (ret) {
2215                                         log_err("pthread_create: %s\n",
2216                                                         strerror(ret));
2217                                         free(fd);
2218                                         nr_started--;
2219                                         break;
2220                                 }
2221                                 ret = pthread_detach(td->thread);
2222                                 if (ret)
2223                                         log_err("pthread_detach: %s",
2224                                                         strerror(ret));
2225                         } else {
2226                                 pid_t pid;
2227                                 dprint(FD_PROCESS, "will fork\n");
2228                                 pid = fork();
2229                                 if (!pid) {
2230                                         int ret = fork_main(sk_out, shm_id, i);
2231
2232                                         _exit(ret);
2233                                 } else if (i == fio_debug_jobno)
2234                                         *fio_debug_jobp = pid;
2235                         }
2236                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2237                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2238                                 log_err("fio: job startup hung? exiting.\n");
2239                                 fio_terminate_threads(TERMINATE_ALL);
2240                                 fio_abort = 1;
2241                                 nr_started--;
2242                                 break;
2243                         }
2244                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2245                 }
2246
2247                 /*
2248                  * Wait for the started threads to transition to
2249                  * TD_INITIALIZED.
2250                  */
2251                 fio_gettime(&this_start, NULL);
2252                 left = this_jobs;
2253                 while (left && !fio_abort) {
2254                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2255                                 break;
2256
2257                         do_usleep(100000);
2258
2259                         for (i = 0; i < this_jobs; i++) {
2260                                 td = map[i];
2261                                 if (!td)
2262                                         continue;
2263                                 if (td->runstate == TD_INITIALIZED) {
2264                                         map[i] = NULL;
2265                                         left--;
2266                                 } else if (td->runstate >= TD_EXITED) {
2267                                         map[i] = NULL;
2268                                         left--;
2269                                         todo--;
2270                                         nr_running++; /* work-around... */
2271                                 }
2272                         }
2273                 }
2274
2275                 if (left) {
2276                         log_err("fio: %d job%s failed to start\n", left,
2277                                         left > 1 ? "s" : "");
2278                         for (i = 0; i < this_jobs; i++) {
2279                                 td = map[i];
2280                                 if (!td)
2281                                         continue;
2282                                 kill(td->pid, SIGTERM);
2283                         }
2284                         break;
2285                 }
2286
2287                 /*
2288                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2289                  */
2290                 for_each_td(td, i) {
2291                         if (td->runstate != TD_INITIALIZED)
2292                                 continue;
2293
2294                         if (in_ramp_time(td))
2295                                 td_set_runstate(td, TD_RAMP);
2296                         else
2297                                 td_set_runstate(td, TD_RUNNING);
2298                         nr_running++;
2299                         nr_started--;
2300                         m_rate += ddir_rw_sum(td->o.ratemin);
2301                         t_rate += ddir_rw_sum(td->o.rate);
2302                         todo--;
2303                         fio_mutex_up(td->mutex);
2304                 }
2305
2306                 reap_threads(&nr_running, &t_rate, &m_rate);
2307
2308                 if (todo)
2309                         do_usleep(100000);
2310         }
2311
2312         while (nr_running) {
2313                 reap_threads(&nr_running, &t_rate, &m_rate);
2314                 do_usleep(10000);
2315         }
2316
2317         fio_idle_prof_stop();
2318
2319         update_io_ticks();
2320 }
2321
2322 static void wait_for_helper_thread_exit(void)
2323 {
2324         void *ret;
2325
2326         helper_exit = 1;
2327         pthread_cond_signal(&helper_cond);
2328         pthread_join(helper_thread, &ret);
2329 }
2330
2331 static void free_disk_util(void)
2332 {
2333         disk_util_prune_entries();
2334
2335         pthread_cond_destroy(&helper_cond);
2336 }
2337
2338 static void *helper_thread_main(void *data)
2339 {
2340         struct sk_out *sk_out = data;
2341         int ret = 0;
2342
2343         sk_out_assign(sk_out);
2344
2345         fio_mutex_up(startup_mutex);
2346
2347         while (!ret) {
2348                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2349                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2350                 struct timespec ts;
2351                 struct timeval tv;
2352
2353                 gettimeofday(&tv, NULL);
2354                 ts.tv_sec = tv.tv_sec + sec;
2355                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2356
2357                 if (ts.tv_nsec >= 1000000000ULL) {
2358                         ts.tv_nsec -= 1000000000ULL;
2359                         ts.tv_sec++;
2360                 }
2361
2362                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2363
2364                 ret = update_io_ticks();
2365
2366                 if (helper_do_stat) {
2367                         helper_do_stat = 0;
2368                         __show_running_run_stats();
2369                 }
2370
2371                 if (!is_backend)
2372                         print_thread_status();
2373         }
2374
2375         sk_out_drop();
2376         return NULL;
2377 }
2378
2379 static int create_helper_thread(struct sk_out *sk_out)
2380 {
2381         int ret;
2382
2383         setup_disk_util();
2384
2385         pthread_cond_init(&helper_cond, NULL);
2386         pthread_mutex_init(&helper_lock, NULL);
2387
2388         ret = pthread_create(&helper_thread, NULL, helper_thread_main, sk_out);
2389         if (ret) {
2390                 log_err("Can't create helper thread: %s\n", strerror(ret));
2391                 return 1;
2392         }
2393
2394         dprint(FD_MUTEX, "wait on startup_mutex\n");
2395         fio_mutex_down(startup_mutex);
2396         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2397         return 0;
2398 }
2399
2400 int fio_backend(struct sk_out *sk_out)
2401 {
2402         struct thread_data *td;
2403         int i;
2404
2405         if (exec_profile) {
2406                 if (load_profile(exec_profile))
2407                         return 1;
2408                 free(exec_profile);
2409                 exec_profile = NULL;
2410         }
2411         if (!thread_number)
2412                 return 0;
2413
2414         if (write_bw_log) {
2415                 struct log_params p = {
2416                         .log_type = IO_LOG_TYPE_BW,
2417                 };
2418
2419                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2420                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2421                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2422         }
2423
2424         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2425         if (startup_mutex == NULL)
2426                 return 1;
2427
2428         set_genesis_time();
2429         stat_init();
2430         create_helper_thread(sk_out);
2431
2432         cgroup_list = smalloc(sizeof(*cgroup_list));
2433         INIT_FLIST_HEAD(cgroup_list);
2434
2435         run_threads(sk_out);
2436
2437         wait_for_helper_thread_exit();
2438
2439         if (!fio_abort) {
2440                 __show_run_stats();
2441                 if (write_bw_log) {
2442                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2443                                 struct io_log *log = agg_io_log[i];
2444
2445                                 flush_log(log, 0);
2446                                 free_log(log);
2447                         }
2448                 }
2449         }
2450
2451         for_each_td(td, i) {
2452                 fio_options_free(td);
2453                 if (td->rusage_sem) {
2454                         fio_mutex_remove(td->rusage_sem);
2455                         td->rusage_sem = NULL;
2456                 }
2457         }
2458
2459         free_disk_util();
2460         cgroup_kill(cgroup_list);
2461         sfree(cgroup_list);
2462         sfree(cgroup_mnt);
2463
2464         fio_mutex_remove(startup_mutex);
2465         stat_exit();
2466         return exit_value;
2467 }