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 "lib/tp.h"
  58 #include "workqueue.h"
  59 #include "lib/mountcheck.h"
  60
  61 static pthread_t helper_thread;
  62 static pthread_mutex_t helper_lock;
  63 pthread_cond_t helper_cond;
  64 int helper_do_stat = 0;
  65
  66 static struct fio_mutex *startup_mutex;
  67 static struct flist_head *cgroup_list;
  68 static char *cgroup_mnt;
  69 static int exit_value;
  70 static volatile int fio_abort;
  71 static unsigned int nr_process = 0;
  72 static unsigned int nr_thread = 0;
  73
  74 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  75
  76 int groupid = 0;
  77 unsigned int thread_number = 0;
  78 unsigned int stat_number = 0;
  79 int shm_id = 0;
  80 int temp_stall_ts;
  81 unsigned long done_secs = 0;
  82 volatile int helper_exit = 0;
  83
  84 #define PAGE_ALIGN(buf) \
  85         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  86
  87 #define JOB_START_TIMEOUT       (5 * 1000)
  88
  89 static void sig_int(int sig)
  90 {
  91         if (threads) {
  92                 if (is_backend)
  93                         fio_server_got_signal(sig);
  94                 else {
  95                         log_info("\nfio: terminating on signal %d\n", sig);
  96                         log_info_flush();
  97                         exit_value = 128;
  98                 }
  99
 100                 fio_terminate_threads(TERMINATE_ALL);
 101         }
 102 }
 103
 104 void sig_show_status(int sig)
 105 {
 106         show_running_run_stats();
 107 }
 108
 109 static void set_sig_handlers(void)
 110 {
 111         struct sigaction act;
 112
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGINT, &act, NULL);
 117
 118         memset(&act, 0, sizeof(act));
 119         act.sa_handler = sig_int;
 120         act.sa_flags = SA_RESTART;
 121         sigaction(SIGTERM, &act, NULL);
 122
 123 /* Windows uses SIGBREAK as a quit signal from other applications */
 124 #ifdef WIN32
 125         memset(&act, 0, sizeof(act));
 126         act.sa_handler = sig_int;
 127         act.sa_flags = SA_RESTART;
 128         sigaction(SIGBREAK, &act, NULL);
 129 #endif
 130
 131         memset(&act, 0, sizeof(act));
 132         act.sa_handler = sig_show_status;
 133         act.sa_flags = SA_RESTART;
 134         sigaction(SIGUSR1, &act, NULL);
 135
 136         if (is_backend) {
 137                 memset(&act, 0, sizeof(act));
 138                 act.sa_handler = sig_int;
 139                 act.sa_flags = SA_RESTART;
 140                 sigaction(SIGPIPE, &act, NULL);
 141         }
 142 }
 143
 144 /*
 145  * Check if we are above the minimum rate given.
 146  */
 147 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 148                              enum fio_ddir ddir)
 149 {
 150         unsigned long long bytes = 0;
 151         unsigned long iops = 0;
 152         unsigned long spent;
 153         unsigned long rate;
 154         unsigned int ratemin = 0;
 155         unsigned int rate_iops = 0;
 156         unsigned int rate_iops_min = 0;
 157
 158         assert(ddir_rw(ddir));
 159
 160         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 161                 return false;
 162
 163         /*
 164          * allow a 2 second settle period in the beginning
 165          */
 166         if (mtime_since(&td->start, now) < 2000)
 167                 return false;
 168
 169         iops += td->this_io_blocks[ddir];
 170         bytes += td->this_io_bytes[ddir];
 171         ratemin += td->o.ratemin[ddir];
 172         rate_iops += td->o.rate_iops[ddir];
 173         rate_iops_min += td->o.rate_iops_min[ddir];
 174
 175         /*
 176          * if rate blocks is set, sample is running
 177          */
 178         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 179                 spent = mtime_since(&td->lastrate[ddir], now);
 180                 if (spent < td->o.ratecycle)
 181                         return false;
 182
 183                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 184                         /*
 185                          * check bandwidth specified rate
 186                          */
 187                         if (bytes < td->rate_bytes[ddir]) {
 188                                 log_err("%s: min rate %u not met\n", td->o.name,
 189                                                                 ratemin);
 190                                 return true;
 191                         } else {
 192                                 if (spent)
 193                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 194                                 else
 195                                         rate = 0;
 196
 197                                 if (rate < ratemin ||
 198                                     bytes < td->rate_bytes[ddir]) {
 199                                         log_err("%s: min rate %u not met, got"
 200                                                 " %luKB/sec\n", td->o.name,
 201                                                         ratemin, rate);
 202                                         return true;
 203                                 }
 204                         }
 205                 } else {
 206                         /*
 207                          * checks iops specified rate
 208                          */
 209                         if (iops < rate_iops) {
 210                                 log_err("%s: min iops rate %u not met\n",
 211                                                 td->o.name, rate_iops);
 212                                 return true;
 213                         } else {
 214                                 if (spent)
 215                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 216                                 else
 217                                         rate = 0;
 218
 219                                 if (rate < rate_iops_min ||
 220                                     iops < td->rate_blocks[ddir]) {
 221                                         log_err("%s: min iops rate %u not met,"
 222                                                 " got %lu\n", td->o.name,
 223                                                         rate_iops_min, rate);
 224                                         return true;
 225                                 }
 226                         }
 227                 }
 228         }
 229
 230         td->rate_bytes[ddir] = bytes;
 231         td->rate_blocks[ddir] = iops;
 232         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 233         return false;
 234 }
 235
 236 static bool check_min_rate(struct thread_data *td, struct timeval *now)
 237 {
 238         bool ret = false;
 239
 240         if (td->bytes_done[DDIR_READ])
 241                 ret |= __check_min_rate(td, now, DDIR_READ);
 242         if (td->bytes_done[DDIR_WRITE])
 243                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 244         if (td->bytes_done[DDIR_TRIM])
 245                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 246
 247         return ret;
 248 }
 249
 250 /*
 251  * When job exits, we can cancel the in-flight IO if we are using async
 252  * io. Attempt to do so.
 253  */
 254 static void cleanup_pending_aio(struct thread_data *td)
 255 {
 256         int r;
 257
 258         /*
 259          * get immediately available events, if any
 260          */
 261         r = io_u_queued_complete(td, 0);
 262         if (r < 0)
 263                 return;
 264
 265         /*
 266          * now cancel remaining active events
 267          */
 268         if (td->io_ops->cancel) {
 269                 struct io_u *io_u;
 270                 int i;
 271
 272                 io_u_qiter(&td->io_u_all, io_u, i) {
 273                         if (io_u->flags & IO_U_F_FLIGHT) {
 274                                 r = td->io_ops->cancel(td, io_u);
 275                                 if (!r)
 276                                         put_io_u(td, io_u);
 277                         }
 278                 }
 279         }
 280
 281         if (td->cur_depth)
 282                 r = io_u_queued_complete(td, td->cur_depth);
 283 }
 284
 285 /*
 286  * Helper to handle the final sync of a file. Works just like the normal
 287  * io path, just does everything sync.
 288  */
 289 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 290 {
 291         struct io_u *io_u = __get_io_u(td);
 292         int ret;
 293
 294         if (!io_u)
 295                 return true;
 296
 297         io_u->ddir = DDIR_SYNC;
 298         io_u->file = f;
 299
 300         if (td_io_prep(td, io_u)) {
 301                 put_io_u(td, io_u);
 302                 return true;
 303         }
 304
 305 requeue:
 306         ret = td_io_queue(td, io_u);
 307         if (ret < 0) {
 308                 td_verror(td, io_u->error, "td_io_queue");
 309                 put_io_u(td, io_u);
 310                 return true;
 311         } else if (ret == FIO_Q_QUEUED) {
 312                 if (io_u_queued_complete(td, 1) < 0)
 313                         return true;
 314         } else if (ret == FIO_Q_COMPLETED) {
 315                 if (io_u->error) {
 316                         td_verror(td, io_u->error, "td_io_queue");
 317                         return true;
 318                 }
 319
 320                 if (io_u_sync_complete(td, io_u) < 0)
 321                         return true;
 322         } else if (ret == FIO_Q_BUSY) {
 323                 if (td_io_commit(td))
 324                         return true;
 325                 goto requeue;
 326         }
 327
 328         return false;
 329 }
 330
 331 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 332 {
 333         int ret;
 334
 335         if (fio_file_open(f))
 336                 return fio_io_sync(td, f);
 337
 338         if (td_io_open_file(td, f))
 339                 return 1;
 340
 341         ret = fio_io_sync(td, f);
 342         td_io_close_file(td, f);
 343         return ret;
 344 }
 345
 346 static inline void __update_tv_cache(struct thread_data *td)
 347 {
 348         fio_gettime(&td->tv_cache, NULL);
 349 }
 350
 351 static inline void update_tv_cache(struct thread_data *td)
 352 {
 353         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 354                 __update_tv_cache(td);
 355 }
 356
 357 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 358 {
 359         if (in_ramp_time(td))
 360                 return false;
 361         if (!td->o.timeout)
 362                 return false;
 363         if (utime_since(&td->epoch, t) >= td->o.timeout)
 364                 return true;
 365
 366         return false;
 367 }
 368
 369 /*
 370  * We need to update the runtime consistently in ms, but keep a running
 371  * tally of the current elapsed time in microseconds for sub millisecond
 372  * updates.
 373  */
 374 static inline void update_runtime(struct thread_data *td,
 375                                   unsigned long long *elapsed_us,
 376                                   const enum fio_ddir ddir)
 377 {
 378         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 379                 return;
 380
 381         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 382         elapsed_us[ddir] += utime_since_now(&td->start);
 383         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 384 }
 385
 386 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 387                                 int *retptr)
 388 {
 389         int ret = *retptr;
 390
 391         if (ret < 0 || td->error) {
 392                 int err = td->error;
 393                 enum error_type_bit eb;
 394
 395                 if (ret < 0)
 396                         err = -ret;
 397
 398                 eb = td_error_type(ddir, err);
 399                 if (!(td->o.continue_on_error & (1 << eb)))
 400                         return true;
 401
 402                 if (td_non_fatal_error(td, eb, err)) {
 403                         /*
 404                          * Continue with the I/Os in case of
 405                          * a non fatal error.
 406                          */
 407                         update_error_count(td, err);
 408                         td_clear_error(td);
 409                         *retptr = 0;
 410                         return false;
 411                 } else if (td->o.fill_device && err == ENOSPC) {
 412                         /*
 413                          * We expect to hit this error if
 414                          * fill_device option is set.
 415                          */
 416                         td_clear_error(td);
 417                         fio_mark_td_terminate(td);
 418                         return true;
 419                 } else {
 420                         /*
 421                          * Stop the I/O in case of a fatal
 422                          * error.
 423                          */
 424                         update_error_count(td, err);
 425                         return true;
 426                 }
 427         }
 428
 429         return false;
 430 }
 431
 432 static void check_update_rusage(struct thread_data *td)
 433 {
 434         if (td->update_rusage) {
 435                 td->update_rusage = 0;
 436                 update_rusage_stat(td);
 437                 fio_mutex_up(td->rusage_sem);
 438         }
 439 }
 440
 441 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 442 {
 443         const int full = queue_full(td);
 444         int min_evts = 0;
 445         int ret;
 446
 447         /*
 448          * if the queue is full, we MUST reap at least 1 event
 449          */
 450         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 451         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 452                 min_evts = 1;
 453
 454         if (time && (__should_check_rate(td, DDIR_READ) ||
 455             __should_check_rate(td, DDIR_WRITE) ||
 456             __should_check_rate(td, DDIR_TRIM)))
 457                 fio_gettime(time, NULL);
 458
 459         do {
 460                 ret = io_u_queued_complete(td, min_evts);
 461                 if (ret < 0)
 462                         break;
 463         } while (full && (td->cur_depth > td->o.iodepth_low));
 464
 465         return ret;
 466 }
 467
 468 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 469                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 470                    struct timeval *comp_time)
 471 {
 472         int ret2;
 473
 474         switch (*ret) {
 475         case FIO_Q_COMPLETED:
 476                 if (io_u->error) {
 477                         *ret = -io_u->error;
 478                         clear_io_u(td, io_u);
 479                 } else if (io_u->resid) {
 480                         int bytes = io_u->xfer_buflen - io_u->resid;
 481                         struct fio_file *f = io_u->file;
 482
 483                         if (bytes_issued)
 484                                 *bytes_issued += bytes;
 485
 486                         if (!from_verify)
 487                                 trim_io_piece(td, io_u);
 488
 489                         /*
 490                          * zero read, fail
 491                          */
 492                         if (!bytes) {
 493                                 if (!from_verify)
 494                                         unlog_io_piece(td, io_u);
 495                                 td_verror(td, EIO, "full resid");
 496                                 put_io_u(td, io_u);
 497                                 break;
 498                         }
 499
 500                         io_u->xfer_buflen = io_u->resid;
 501                         io_u->xfer_buf += bytes;
 502                         io_u->offset += bytes;
 503
 504                         if (ddir_rw(io_u->ddir))
 505                                 td->ts.short_io_u[io_u->ddir]++;
 506
 507                         f = io_u->file;
 508                         if (io_u->offset == f->real_file_size)
 509                                 goto sync_done;
 510
 511                         requeue_io_u(td, &io_u);
 512                 } else {
 513 sync_done:
 514                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 515                             __should_check_rate(td, DDIR_WRITE) ||
 516                             __should_check_rate(td, DDIR_TRIM)))
 517                                 fio_gettime(comp_time, NULL);
 518
 519                         *ret = io_u_sync_complete(td, io_u);
 520                         if (*ret < 0)
 521                                 break;
 522                 }
 523                 return 0;
 524         case FIO_Q_QUEUED:
 525                 /*
 526                  * if the engine doesn't have a commit hook,
 527                  * the io_u is really queued. if it does have such
 528                  * a hook, it has to call io_u_queued() itself.
 529                  */
 530                 if (td->io_ops->commit == NULL)
 531                         io_u_queued(td, io_u);
 532                 if (bytes_issued)
 533                         *bytes_issued += io_u->xfer_buflen;
 534                 break;
 535         case FIO_Q_BUSY:
 536                 if (!from_verify)
 537                         unlog_io_piece(td, io_u);
 538                 requeue_io_u(td, &io_u);
 539                 ret2 = td_io_commit(td);
 540                 if (ret2 < 0)
 541                         *ret = ret2;
 542                 break;
 543         default:
 544                 assert(*ret < 0);
 545                 td_verror(td, -(*ret), "td_io_queue");
 546                 break;
 547         }
 548
 549         if (break_on_this_error(td, ddir, ret))
 550                 return 1;
 551
 552         return 0;
 553 }
 554
 555 static inline bool io_in_polling(struct thread_data *td)
 556 {
 557         return !td->o.iodepth_batch_complete_min &&
 558                    !td->o.iodepth_batch_complete_max;
 559 }
 560
 561 /*
 562  * The main verify engine. Runs over the writes we previously submitted,
 563  * reads the blocks back in, and checks the crc/md5 of the data.
 564  */
 565 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 566 {
 567         struct fio_file *f;
 568         struct io_u *io_u;
 569         int ret, min_events;
 570         unsigned int i;
 571
 572         dprint(FD_VERIFY, "starting loop\n");
 573
 574         /*
 575          * sync io first and invalidate cache, to make sure we really
 576          * read from disk.
 577          */
 578         for_each_file(td, f, i) {
 579                 if (!fio_file_open(f))
 580                         continue;
 581                 if (fio_io_sync(td, f))
 582                         break;
 583                 if (file_invalidate_cache(td, f))
 584                         break;
 585         }
 586
 587         check_update_rusage(td);
 588
 589         if (td->error)
 590                 return;
 591
 592         td_set_runstate(td, TD_VERIFYING);
 593
 594         io_u = NULL;
 595         while (!td->terminate) {
 596                 enum fio_ddir ddir;
 597                 int full;
 598
 599                 update_tv_cache(td);
 600                 check_update_rusage(td);
 601
 602                 if (runtime_exceeded(td, &td->tv_cache)) {
 603                         __update_tv_cache(td);
 604                         if (runtime_exceeded(td, &td->tv_cache)) {
 605                                 fio_mark_td_terminate(td);
 606                                 break;
 607                         }
 608                 }
 609
 610                 if (flow_threshold_exceeded(td))
 611                         continue;
 612
 613                 if (!td->o.experimental_verify) {
 614                         io_u = __get_io_u(td);
 615                         if (!io_u)
 616                                 break;
 617
 618                         if (get_next_verify(td, io_u)) {
 619                                 put_io_u(td, io_u);
 620                                 break;
 621                         }
 622
 623                         if (td_io_prep(td, io_u)) {
 624                                 put_io_u(td, io_u);
 625                                 break;
 626                         }
 627                 } else {
 628                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 629                                 break;
 630
 631                         while ((io_u = get_io_u(td)) != NULL) {
 632                                 if (IS_ERR(io_u)) {
 633                                         io_u = NULL;
 634                                         ret = FIO_Q_BUSY;
 635                                         goto reap;
 636                                 }
 637
 638                                 /*
 639                                  * We are only interested in the places where
 640                                  * we wrote or trimmed IOs. Turn those into
 641                                  * reads for verification purposes.
 642                                  */
 643                                 if (io_u->ddir == DDIR_READ) {
 644                                         /*
 645                                          * Pretend we issued it for rwmix
 646                                          * accounting
 647                                          */
 648                                         td->io_issues[DDIR_READ]++;
 649                                         put_io_u(td, io_u);
 650                                         continue;
 651                                 } else if (io_u->ddir == DDIR_TRIM) {
 652                                         io_u->ddir = DDIR_READ;
 653                                         io_u_set(io_u, IO_U_F_TRIMMED);
 654                                         break;
 655                                 } else if (io_u->ddir == DDIR_WRITE) {
 656                                         io_u->ddir = DDIR_READ;
 657                                         break;
 658                                 } else {
 659                                         put_io_u(td, io_u);
 660                                         continue;
 661                                 }
 662                         }
 663
 664                         if (!io_u)
 665                                 break;
 666                 }
 667
 668                 if (verify_state_should_stop(td, io_u)) {
 669                         put_io_u(td, io_u);
 670                         break;
 671                 }
 672
 673                 if (td->o.verify_async)
 674                         io_u->end_io = verify_io_u_async;
 675                 else
 676                         io_u->end_io = verify_io_u;
 677
 678                 ddir = io_u->ddir;
 679                 if (!td->o.disable_slat)
 680                         fio_gettime(&io_u->start_time, NULL);
 681
 682                 ret = td_io_queue(td, io_u);
 683
 684                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 685                         break;
 686
 687                 /*
 688                  * if we can queue more, do so. but check if there are
 689                  * completed io_u's first. Note that we can get BUSY even
 690                  * without IO queued, if the system is resource starved.
 691                  */
 692 reap:
 693                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 694                 if (full || io_in_polling(td))
 695                         ret = wait_for_completions(td, NULL);
 696
 697                 if (ret < 0)
 698                         break;
 699         }
 700
 701         check_update_rusage(td);
 702
 703         if (!td->error) {
 704                 min_events = td->cur_depth;
 705
 706                 if (min_events)
 707                         ret = io_u_queued_complete(td, min_events);
 708         } else
 709                 cleanup_pending_aio(td);
 710
 711         td_set_runstate(td, TD_RUNNING);
 712
 713         dprint(FD_VERIFY, "exiting loop\n");
 714 }
 715
 716 static bool exceeds_number_ios(struct thread_data *td)
 717 {
 718         unsigned long long number_ios;
 719
 720         if (!td->o.number_ios)
 721                 return false;
 722
 723         number_ios = ddir_rw_sum(td->io_blocks);
 724         number_ios += td->io_u_queued + td->io_u_in_flight;
 725
 726         return number_ios >= (td->o.number_ios * td->loops);
 727 }
 728
 729 static bool io_issue_bytes_exceeded(struct thread_data *td)
 730 {
 731         unsigned long long bytes, limit;
 732
 733         if (td_rw(td))
 734                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 735         else if (td_write(td))
 736                 bytes = td->io_issue_bytes[DDIR_WRITE];
 737         else if (td_read(td))
 738                 bytes = td->io_issue_bytes[DDIR_READ];
 739         else
 740                 bytes = td->io_issue_bytes[DDIR_TRIM];
 741
 742         if (td->o.io_limit)
 743                 limit = td->o.io_limit;
 744         else
 745                 limit = td->o.size;
 746
 747         limit *= td->loops;
 748         return bytes >= limit || exceeds_number_ios(td);
 749 }
 750
 751 static bool io_complete_bytes_exceeded(struct thread_data *td)
 752 {
 753         unsigned long long bytes, limit;
 754
 755         if (td_rw(td))
 756                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 757         else if (td_write(td))
 758                 bytes = td->this_io_bytes[DDIR_WRITE];
 759         else if (td_read(td))
 760                 bytes = td->this_io_bytes[DDIR_READ];
 761         else
 762                 bytes = td->this_io_bytes[DDIR_TRIM];
 763
 764         if (td->o.io_limit)
 765                 limit = td->o.io_limit;
 766         else
 767                 limit = td->o.size;
 768
 769         limit *= td->loops;
 770         return bytes >= limit || exceeds_number_ios(td);
 771 }
 772
 773 /*
 774  * used to calculate the next io time for rate control
 775  *
 776  */
 777 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 778 {
 779         uint64_t secs, remainder, bps, bytes, iops;
 780
 781         assert(!(td->flags & TD_F_CHILD));
 782         bytes = td->rate_io_issue_bytes[ddir];
 783         bps = td->rate_bps[ddir];
 784
 785         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 786                 uint64_t val;
 787                 iops = bps / td->o.bs[ddir];
 788                 val = (int64_t) (1000000 / iops) *
 789                                 -logf(__rand_0_1(&td->poisson_state));
 790                 if (val) {
 791                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 792                                         (unsigned long long) 1000000 / val);
 793                 }
 794                 td->last_usec += val;
 795                 return td->last_usec;
 796         } else if (bps) {
 797                 secs = bytes / bps;
 798                 remainder = bytes % bps;
 799                 return remainder * 1000000 / bps + secs * 1000000;
 800         }
 801
 802         return 0;
 803 }
 804
 805 /*
 806  * Main IO worker function. It retrieves io_u's to process and queues
 807  * and reaps them, checking for rate and errors along the way.
 808  *
 809  * Returns number of bytes written and trimmed.
 810  */
 811 static uint64_t do_io(struct thread_data *td)
 812 {
 813         unsigned int i;
 814         int ret = 0;
 815         uint64_t total_bytes, bytes_issued = 0;
 816
 817         if (in_ramp_time(td))
 818                 td_set_runstate(td, TD_RAMP);
 819         else
 820                 td_set_runstate(td, TD_RUNNING);
 821
 822         lat_target_init(td);
 823
 824         total_bytes = td->o.size;
 825         /*
 826         * Allow random overwrite workloads to write up to io_limit
 827         * before starting verification phase as 'size' doesn't apply.
 828         */
 829         if (td_write(td) && td_random(td) && td->o.norandommap)
 830                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 831         /*
 832          * If verify_backlog is enabled, we'll run the verify in this
 833          * handler as well. For that case, we may need up to twice the
 834          * amount of bytes.
 835          */
 836         if (td->o.verify != VERIFY_NONE &&
 837            (td_write(td) && td->o.verify_backlog))
 838                 total_bytes += td->o.size;
 839
 840         /* In trimwrite mode, each byte is trimmed and then written, so
 841          * allow total_bytes to be twice as big */
 842         if (td_trimwrite(td))
 843                 total_bytes += td->total_io_size;
 844
 845         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 846                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 847                 td->o.time_based) {
 848                 struct timeval comp_time;
 849                 struct io_u *io_u;
 850                 int full;
 851                 enum fio_ddir ddir;
 852
 853                 check_update_rusage(td);
 854
 855                 if (td->terminate || td->done)
 856                         break;
 857
 858                 update_tv_cache(td);
 859
 860                 if (runtime_exceeded(td, &td->tv_cache)) {
 861                         __update_tv_cache(td);
 862                         if (runtime_exceeded(td, &td->tv_cache)) {
 863                                 fio_mark_td_terminate(td);
 864                                 break;
 865                         }
 866                 }
 867
 868                 if (flow_threshold_exceeded(td))
 869                         continue;
 870
 871                 if (!td->o.time_based && bytes_issued >= total_bytes)
 872                         break;
 873
 874                 io_u = get_io_u(td);
 875                 if (IS_ERR_OR_NULL(io_u)) {
 876                         int err = PTR_ERR(io_u);
 877
 878                         io_u = NULL;
 879                         if (err == -EBUSY) {
 880                                 ret = FIO_Q_BUSY;
 881                                 goto reap;
 882                         }
 883                         if (td->o.latency_target)
 884                                 goto reap;
 885                         break;
 886                 }
 887
 888                 ddir = io_u->ddir;
 889
 890                 /*
 891                  * Add verification end_io handler if:
 892                  *      - Asked to verify (!td_rw(td))
 893                  *      - Or the io_u is from our verify list (mixed write/ver)
 894                  */
 895                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 896                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 897
 898                         if (!td->o.verify_pattern_bytes) {
 899                                 io_u->rand_seed = __rand(&td->verify_state);
 900                                 if (sizeof(int) != sizeof(long *))
 901                                         io_u->rand_seed *= __rand(&td->verify_state);
 902                         }
 903
 904                         if (verify_state_should_stop(td, io_u)) {
 905                                 put_io_u(td, io_u);
 906                                 break;
 907                         }
 908
 909                         if (td->o.verify_async)
 910                                 io_u->end_io = verify_io_u_async;
 911                         else
 912                                 io_u->end_io = verify_io_u;
 913                         td_set_runstate(td, TD_VERIFYING);
 914                 } else if (in_ramp_time(td))
 915                         td_set_runstate(td, TD_RAMP);
 916                 else
 917                         td_set_runstate(td, TD_RUNNING);
 918
 919                 /*
 920                  * Always log IO before it's issued, so we know the specific
 921                  * order of it. The logged unit will track when the IO has
 922                  * completed.
 923                  */
 924                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 925                     td->o.do_verify &&
 926                     td->o.verify != VERIFY_NONE &&
 927                     !td->o.experimental_verify)
 928                         log_io_piece(td, io_u);
 929
 930                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 931                         const unsigned long blen = io_u->xfer_buflen;
 932                         const enum fio_ddir ddir = acct_ddir(io_u);
 933
 934                         if (td->error)
 935                                 break;
 936
 937                         ret = workqueue_enqueue(&td->io_wq, &io_u->work);
 938                         if (ret)
 939                                 ret = FIO_Q_QUEUED;
 940                         else
 941                                 ret = FIO_Q_BUSY;
 942
 943                         if (ret == FIO_Q_QUEUED && ddir_rw(ddir)) {
 944                                 td->io_issues[ddir]++;
 945                                 td->io_issue_bytes[ddir] += blen;
 946                                 td->rate_io_issue_bytes[ddir] += blen;
 947                         }
 948
 949                         if (should_check_rate(td))
 950                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 951
 952                 } else {
 953                         ret = td_io_queue(td, io_u);
 954
 955                         if (should_check_rate(td))
 956                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 957
 958                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 959                                 break;
 960
 961                         /*
 962                          * See if we need to complete some commands. Note that
 963                          * we can get BUSY even without IO queued, if the
 964                          * system is resource starved.
 965                          */
 966 reap:
 967                         full = queue_full(td) ||
 968                                 (ret == FIO_Q_BUSY && td->cur_depth);
 969                         if (full || io_in_polling(td))
 970                                 ret = wait_for_completions(td, &comp_time);
 971                 }
 972                 if (ret < 0)
 973                         break;
 974                 if (!ddir_rw_sum(td->bytes_done) &&
 975                     !(td->io_ops->flags & FIO_NOIO))
 976                         continue;
 977
 978                 if (!in_ramp_time(td) && should_check_rate(td)) {
 979                         if (check_min_rate(td, &comp_time)) {
 980                                 if (exitall_on_terminate)
 981                                         fio_terminate_threads(td->groupid);
 982                                 td_verror(td, EIO, "check_min_rate");
 983                                 break;
 984                         }
 985                 }
 986                 if (!in_ramp_time(td) && td->o.latency_target)
 987                         lat_target_check(td);
 988
 989                 if (td->o.thinktime) {
 990                         unsigned long long b;
 991
 992                         b = ddir_rw_sum(td->io_blocks);
 993                         if (!(b % td->o.thinktime_blocks)) {
 994                                 int left;
 995
 996                                 io_u_quiesce(td);
 997
 998                                 if (td->o.thinktime_spin)
 999                                         usec_spin(td->o.thinktime_spin);
1000
1001                                 left = td->o.thinktime - td->o.thinktime_spin;
1002                                 if (left)
1003                                         usec_sleep(td, left);
1004                         }
1005                 }
1006         }
1007
1008         check_update_rusage(td);
1009
1010         if (td->trim_entries)
1011                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1012
1013         if (td->o.fill_device && td->error == ENOSPC) {
1014                 td->error = 0;
1015                 fio_mark_td_terminate(td);
1016         }
1017         if (!td->error) {
1018                 struct fio_file *f;
1019
1020                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1021                         workqueue_flush(&td->io_wq);
1022                         i = 0;
1023                 } else
1024                         i = td->cur_depth;
1025
1026                 if (i) {
1027                         ret = io_u_queued_complete(td, i);
1028                         if (td->o.fill_device && td->error == ENOSPC)
1029                                 td->error = 0;
1030                 }
1031
1032                 if (should_fsync(td) && td->o.end_fsync) {
1033                         td_set_runstate(td, TD_FSYNCING);
1034
1035                         for_each_file(td, f, i) {
1036                                 if (!fio_file_fsync(td, f))
1037                                         continue;
1038
1039                                 log_err("fio: end_fsync failed for file %s\n",
1040                                                                 f->file_name);
1041                         }
1042                 }
1043         } else
1044                 cleanup_pending_aio(td);
1045
1046         /*
1047          * stop job if we failed doing any IO
1048          */
1049         if (!ddir_rw_sum(td->this_io_bytes))
1050                 td->done = 1;
1051
1052         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1053 }
1054
1055 static void cleanup_io_u(struct thread_data *td)
1056 {
1057         struct io_u *io_u;
1058
1059         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1060
1061                 if (td->io_ops->io_u_free)
1062                         td->io_ops->io_u_free(td, io_u);
1063
1064                 fio_memfree(io_u, sizeof(*io_u));
1065         }
1066
1067         free_io_mem(td);
1068
1069         io_u_rexit(&td->io_u_requeues);
1070         io_u_qexit(&td->io_u_freelist);
1071         io_u_qexit(&td->io_u_all);
1072
1073         if (td->last_write_comp)
1074                 sfree(td->last_write_comp);
1075 }
1076
1077 static int init_io_u(struct thread_data *td)
1078 {
1079         struct io_u *io_u;
1080         unsigned int max_bs, min_write;
1081         int cl_align, i, max_units;
1082         int data_xfer = 1, err;
1083         char *p;
1084
1085         max_units = td->o.iodepth;
1086         max_bs = td_max_bs(td);
1087         min_write = td->o.min_bs[DDIR_WRITE];
1088         td->orig_buffer_size = (unsigned long long) max_bs
1089                                         * (unsigned long long) max_units;
1090
1091         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1092                 data_xfer = 0;
1093
1094         err = 0;
1095         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1096         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1097         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1098
1099         if (err) {
1100                 log_err("fio: failed setting up IO queues\n");
1101                 return 1;
1102         }
1103
1104         /*
1105          * if we may later need to do address alignment, then add any
1106          * possible adjustment here so that we don't cause a buffer
1107          * overflow later. this adjustment may be too much if we get
1108          * lucky and the allocator gives us an aligned address.
1109          */
1110         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1111             (td->io_ops->flags & FIO_RAWIO))
1112                 td->orig_buffer_size += page_mask + td->o.mem_align;
1113
1114         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1115                 unsigned long bs;
1116
1117                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1118                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1119         }
1120
1121         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1122                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1123                 return 1;
1124         }
1125
1126         if (data_xfer && allocate_io_mem(td))
1127                 return 1;
1128
1129         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1130             (td->io_ops->flags & FIO_RAWIO))
1131                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1132         else
1133                 p = td->orig_buffer;
1134
1135         cl_align = os_cache_line_size();
1136
1137         for (i = 0; i < max_units; i++) {
1138                 void *ptr;
1139
1140                 if (td->terminate)
1141                         return 1;
1142
1143                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1144                 if (!ptr) {
1145                         log_err("fio: unable to allocate aligned memory\n");
1146                         break;
1147                 }
1148
1149                 io_u = ptr;
1150                 memset(io_u, 0, sizeof(*io_u));
1151                 INIT_FLIST_HEAD(&io_u->verify_list);
1152                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1153
1154                 if (data_xfer) {
1155                         io_u->buf = p;
1156                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1157
1158                         if (td_write(td))
1159                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1160                         if (td_write(td) && td->o.verify_pattern_bytes) {
1161                                 /*
1162                                  * Fill the buffer with the pattern if we are
1163                                  * going to be doing writes.
1164                                  */
1165                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1166                         }
1167                 }
1168
1169                 io_u->index = i;
1170                 io_u->flags = IO_U_F_FREE;
1171                 io_u_qpush(&td->io_u_freelist, io_u);
1172
1173                 /*
1174                  * io_u never leaves this stack, used for iteration of all
1175                  * io_u buffers.
1176                  */
1177                 io_u_qpush(&td->io_u_all, io_u);
1178
1179                 if (td->io_ops->io_u_init) {
1180                         int ret = td->io_ops->io_u_init(td, io_u);
1181
1182                         if (ret) {
1183                                 log_err("fio: failed to init engine data: %d\n", ret);
1184                                 return 1;
1185                         }
1186                 }
1187
1188                 p += max_bs;
1189         }
1190
1191         if (td->o.verify != VERIFY_NONE) {
1192                 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1193                 if (!td->last_write_comp) {
1194                         log_err("fio: failed to alloc write comp data\n");
1195                         return 1;
1196                 }
1197         }
1198
1199         return 0;
1200 }
1201
1202 static int switch_ioscheduler(struct thread_data *td)
1203 {
1204         char tmp[256], tmp2[128];
1205         FILE *f;
1206         int ret;
1207
1208         if (td->io_ops->flags & FIO_DISKLESSIO)
1209                 return 0;
1210
1211         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1212
1213         f = fopen(tmp, "r+");
1214         if (!f) {
1215                 if (errno == ENOENT) {
1216                         log_err("fio: os or kernel doesn't support IO scheduler"
1217                                 " switching\n");
1218                         return 0;
1219                 }
1220                 td_verror(td, errno, "fopen iosched");
1221                 return 1;
1222         }
1223
1224         /*
1225          * Set io scheduler.
1226          */
1227         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1228         if (ferror(f) || ret != 1) {
1229                 td_verror(td, errno, "fwrite");
1230                 fclose(f);
1231                 return 1;
1232         }
1233
1234         rewind(f);
1235
1236         /*
1237          * Read back and check that the selected scheduler is now the default.
1238          */
1239         memset(tmp, 0, sizeof(tmp));
1240         ret = fread(tmp, sizeof(tmp), 1, f);
1241         if (ferror(f) || ret < 0) {
1242                 td_verror(td, errno, "fread");
1243                 fclose(f);
1244                 return 1;
1245         }
1246         /*
1247          * either a list of io schedulers or "none\n" is expected.
1248          */
1249         tmp[strlen(tmp) - 1] = '\0';
1250
1251
1252         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1253         if (!strstr(tmp, tmp2)) {
1254                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1255                 td_verror(td, EINVAL, "iosched_switch");
1256                 fclose(f);
1257                 return 1;
1258         }
1259
1260         fclose(f);
1261         return 0;
1262 }
1263
1264 static bool keep_running(struct thread_data *td)
1265 {
1266         unsigned long long limit;
1267
1268         if (td->done)
1269                 return false;
1270         if (td->o.time_based)
1271                 return true;
1272         if (td->o.loops) {
1273                 td->o.loops--;
1274                 return true;
1275         }
1276         if (exceeds_number_ios(td))
1277                 return false;
1278
1279         if (td->o.io_limit)
1280                 limit = td->o.io_limit;
1281         else
1282                 limit = td->o.size;
1283
1284         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1285                 uint64_t diff;
1286
1287                 /*
1288                  * If the difference is less than the minimum IO size, we
1289                  * are done.
1290                  */
1291                 diff = limit - ddir_rw_sum(td->io_bytes);
1292                 if (diff < td_max_bs(td))
1293                         return false;
1294
1295                 if (fio_files_done(td))
1296                         return false;
1297
1298                 return true;
1299         }
1300
1301         return false;
1302 }
1303
1304 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1305 {
1306         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1307         int ret;
1308         char *str;
1309
1310         str = malloc(newlen);
1311         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1312
1313         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1314         ret = system(str);
1315         if (ret == -1)
1316                 log_err("fio: exec of cmd <%s> failed\n", str);
1317
1318         free(str);
1319         return ret;
1320 }
1321
1322 /*
1323  * Dry run to compute correct state of numberio for verification.
1324  */
1325 static uint64_t do_dry_run(struct thread_data *td)
1326 {
1327         td_set_runstate(td, TD_RUNNING);
1328
1329         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1330                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1331                 struct io_u *io_u;
1332                 int ret;
1333
1334                 if (td->terminate || td->done)
1335                         break;
1336
1337                 io_u = get_io_u(td);
1338                 if (!io_u)
1339                         break;
1340
1341                 io_u_set(io_u, IO_U_F_FLIGHT);
1342                 io_u->error = 0;
1343                 io_u->resid = 0;
1344                 if (ddir_rw(acct_ddir(io_u)))
1345                         td->io_issues[acct_ddir(io_u)]++;
1346                 if (ddir_rw(io_u->ddir)) {
1347                         io_u_mark_depth(td, 1);
1348                         td->ts.total_io_u[io_u->ddir]++;
1349                 }
1350
1351                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1352                     td->o.do_verify &&
1353                     td->o.verify != VERIFY_NONE &&
1354                     !td->o.experimental_verify)
1355                         log_io_piece(td, io_u);
1356
1357                 ret = io_u_sync_complete(td, io_u);
1358                 (void) ret;
1359         }
1360
1361         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1362 }
1363
1364 static void io_workqueue_fn(struct submit_worker *sw,
1365                             struct workqueue_work *work)
1366 {
1367         struct io_u *io_u = container_of(work, struct io_u, work);
1368         const enum fio_ddir ddir = io_u->ddir;
1369         struct thread_data *td = sw->private;
1370         int ret;
1371
1372         dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1373
1374         io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1375
1376         td->cur_depth++;
1377
1378         do {
1379                 ret = td_io_queue(td, io_u);
1380                 if (ret != FIO_Q_BUSY)
1381                         break;
1382                 ret = io_u_queued_complete(td, 1);
1383                 if (ret > 0)
1384                         td->cur_depth -= ret;
1385                 io_u_clear(io_u, IO_U_F_FLIGHT);
1386         } while (1);
1387
1388         dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1389
1390         io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1391
1392         if (ret == FIO_Q_COMPLETED)
1393                 td->cur_depth--;
1394         else if (ret == FIO_Q_QUEUED) {
1395                 unsigned int min_evts;
1396
1397                 if (td->o.iodepth == 1)
1398                         min_evts = 1;
1399                 else
1400                         min_evts = 0;
1401
1402                 ret = io_u_queued_complete(td, min_evts);
1403                 if (ret > 0)
1404                         td->cur_depth -= ret;
1405         } else if (ret == FIO_Q_BUSY) {
1406                 ret = io_u_queued_complete(td, td->cur_depth);
1407                 if (ret > 0)
1408                         td->cur_depth -= ret;
1409         }
1410 }
1411
1412 static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
1413 {
1414         struct thread_data *td = sw->private;
1415
1416         if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
1417                 return true;
1418
1419         return false;
1420 }
1421
1422 static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
1423 {
1424         struct thread_data *td = sw->private;
1425         int ret;
1426
1427         ret = io_u_quiesce(td);
1428         if (ret > 0)
1429                 td->cur_depth -= ret;
1430 }
1431
1432 static int io_workqueue_alloc_fn(struct submit_worker *sw)
1433 {
1434         struct thread_data *td;
1435
1436         td = calloc(1, sizeof(*td));
1437         sw->private = td;
1438         return 0;
1439 }
1440
1441 static void io_workqueue_free_fn(struct submit_worker *sw)
1442 {
1443         free(sw->private);
1444         sw->private = NULL;
1445 }
1446
1447 static int io_workqueue_init_worker_fn(struct submit_worker *sw)
1448 {
1449         struct thread_data *parent = sw->wq->td;
1450         struct thread_data *td = sw->private;
1451         int fio_unused ret;
1452
1453         memcpy(&td->o, &parent->o, sizeof(td->o));
1454         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
1455         td->o.uid = td->o.gid = -1U;
1456         dup_files(td, parent);
1457         td->eo = parent->eo;
1458         fio_options_mem_dupe(td);
1459
1460         if (ioengine_load(td))
1461                 goto err;
1462
1463         if (td->o.odirect)
1464                 td->io_ops->flags |= FIO_RAWIO;
1465
1466         td->pid = gettid();
1467
1468         INIT_FLIST_HEAD(&td->io_log_list);
1469         INIT_FLIST_HEAD(&td->io_hist_list);
1470         INIT_FLIST_HEAD(&td->verify_list);
1471         INIT_FLIST_HEAD(&td->trim_list);
1472         INIT_FLIST_HEAD(&td->next_rand_list);
1473         td->io_hist_tree = RB_ROOT;
1474
1475         td->o.iodepth = 1;
1476         if (td_io_init(td))
1477                 goto err_io_init;
1478
1479         fio_gettime(&td->epoch, NULL);
1480         fio_getrusage(&td->ru_start);
1481         clear_io_state(td, 1);
1482
1483         td_set_runstate(td, TD_RUNNING);
1484         td->flags |= TD_F_CHILD;
1485         td->parent = parent;
1486         return 0;
1487
1488 err_io_init:
1489         close_ioengine(td);
1490 err:
1491         return 1;
1492
1493 }
1494
1495 static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
1496                                         unsigned int *sum_cnt)
1497 {
1498         struct thread_data *td = sw->private;
1499
1500         (*sum_cnt)++;
1501         sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
1502
1503         fio_options_free(td);
1504         close_and_free_files(td);
1505         if (td->io_ops)
1506                 close_ioengine(td);
1507         td_set_runstate(td, TD_EXITED);
1508 }
1509
1510 #ifdef CONFIG_SFAA
1511 static void sum_val(uint64_t *dst, uint64_t *src)
1512 {
1513         if (*src) {
1514                 __sync_fetch_and_add(dst, *src);
1515                 *src = 0;
1516         }
1517 }
1518 #else
1519 static void sum_val(uint64_t *dst, uint64_t *src)
1520 {
1521         if (*src) {
1522                 *dst += *src;
1523                 *src = 0;
1524         }
1525 }
1526 #endif
1527
1528 static void pthread_double_unlock(pthread_mutex_t *lock1,
1529                                   pthread_mutex_t *lock2)
1530 {
1531 #ifndef CONFIG_SFAA
1532         pthread_mutex_unlock(lock1);
1533         pthread_mutex_unlock(lock2);
1534 #endif
1535 }
1536
1537 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
1538 {
1539 #ifndef CONFIG_SFAA
1540         if (lock1 < lock2) {
1541                 pthread_mutex_lock(lock1);
1542                 pthread_mutex_lock(lock2);
1543         } else {
1544                 pthread_mutex_lock(lock2);
1545                 pthread_mutex_lock(lock1);
1546         }
1547 #endif
1548 }
1549
1550 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
1551                      enum fio_ddir ddir)
1552 {
1553         pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
1554
1555         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
1556         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
1557         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
1558         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
1559         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
1560
1561         pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
1562 }
1563
1564 static void io_workqueue_update_acct_fn(struct submit_worker *sw)
1565 {
1566         struct thread_data *src = sw->private;
1567         struct thread_data *dst = sw->wq->td;
1568
1569         if (td_read(src))
1570                 sum_ddir(dst, src, DDIR_READ);
1571         if (td_write(src))
1572                 sum_ddir(dst, src, DDIR_WRITE);
1573         if (td_trim(src))
1574                 sum_ddir(dst, src, DDIR_TRIM);
1575
1576 }
1577
1578 struct workqueue_ops rated_wq_ops = {
1579         .fn                     = io_workqueue_fn,
1580         .pre_sleep_flush_fn     = io_workqueue_pre_sleep_flush_fn,
1581         .pre_sleep_fn           = io_workqueue_pre_sleep_fn,
1582         .update_acct_fn         = io_workqueue_update_acct_fn,
1583         .alloc_worker_fn        = io_workqueue_alloc_fn,
1584         .free_worker_fn         = io_workqueue_free_fn,
1585         .init_worker_fn         = io_workqueue_init_worker_fn,
1586         .exit_worker_fn         = io_workqueue_exit_worker_fn,
1587 };
1588
1589 /*
1590  * Entry point for the thread based jobs. The process based jobs end up
1591  * here as well, after a little setup.
1592  */
1593 static void *thread_main(void *data)
1594 {
1595         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1596         struct thread_data *td = data;
1597         struct thread_options *o = &td->o;
1598         pthread_condattr_t attr;
1599         int clear_state;
1600         int ret;
1601
1602         if (!o->use_thread) {
1603                 setsid();
1604                 td->pid = getpid();
1605         } else
1606                 td->pid = gettid();
1607
1608         fio_local_clock_init(o->use_thread);
1609
1610         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1611
1612         if (is_backend)
1613                 fio_server_send_start(td);
1614
1615         INIT_FLIST_HEAD(&td->io_log_list);
1616         INIT_FLIST_HEAD(&td->io_hist_list);
1617         INIT_FLIST_HEAD(&td->verify_list);
1618         INIT_FLIST_HEAD(&td->trim_list);
1619         INIT_FLIST_HEAD(&td->next_rand_list);
1620         pthread_mutex_init(&td->io_u_lock, NULL);
1621         td->io_hist_tree = RB_ROOT;
1622
1623         pthread_condattr_init(&attr);
1624         pthread_cond_init(&td->verify_cond, &attr);
1625         pthread_cond_init(&td->free_cond, &attr);
1626
1627         td_set_runstate(td, TD_INITIALIZED);
1628         dprint(FD_MUTEX, "up startup_mutex\n");
1629         fio_mutex_up(startup_mutex);
1630         dprint(FD_MUTEX, "wait on td->mutex\n");
1631         fio_mutex_down(td->mutex);
1632         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1633
1634         /*
1635          * A new gid requires privilege, so we need to do this before setting
1636          * the uid.
1637          */
1638         if (o->gid != -1U && setgid(o->gid)) {
1639                 td_verror(td, errno, "setgid");
1640                 goto err;
1641         }
1642         if (o->uid != -1U && setuid(o->uid)) {
1643                 td_verror(td, errno, "setuid");
1644                 goto err;
1645         }
1646
1647         /*
1648          * If we have a gettimeofday() thread, make sure we exclude that
1649          * thread from this job
1650          */
1651         if (o->gtod_cpu)
1652                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1653
1654         /*
1655          * Set affinity first, in case it has an impact on the memory
1656          * allocations.
1657          */
1658         if (fio_option_is_set(o, cpumask)) {
1659                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1660                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1661                         if (!ret) {
1662                                 log_err("fio: no CPUs set\n");
1663                                 log_err("fio: Try increasing number of available CPUs\n");
1664                                 td_verror(td, EINVAL, "cpus_split");
1665                                 goto err;
1666                         }
1667                 }
1668                 ret = fio_setaffinity(td->pid, o->cpumask);
1669                 if (ret == -1) {
1670                         td_verror(td, errno, "cpu_set_affinity");
1671                         goto err;
1672                 }
1673         }
1674
1675 #ifdef CONFIG_LIBNUMA
1676         /* numa node setup */
1677         if (fio_option_is_set(o, numa_cpunodes) ||
1678             fio_option_is_set(o, numa_memnodes)) {
1679                 struct bitmask *mask;
1680
1681                 if (numa_available() < 0) {
1682                         td_verror(td, errno, "Does not support NUMA API\n");
1683                         goto err;
1684                 }
1685
1686                 if (fio_option_is_set(o, numa_cpunodes)) {
1687                         mask = numa_parse_nodestring(o->numa_cpunodes);
1688                         ret = numa_run_on_node_mask(mask);
1689                         numa_free_nodemask(mask);
1690                         if (ret == -1) {
1691                                 td_verror(td, errno, \
1692                                         "numa_run_on_node_mask failed\n");
1693                                 goto err;
1694                         }
1695                 }
1696
1697                 if (fio_option_is_set(o, numa_memnodes)) {
1698                         mask = NULL;
1699                         if (o->numa_memnodes)
1700                                 mask = numa_parse_nodestring(o->numa_memnodes);
1701
1702                         switch (o->numa_mem_mode) {
1703                         case MPOL_INTERLEAVE:
1704                                 numa_set_interleave_mask(mask);
1705                                 break;
1706                         case MPOL_BIND:
1707                                 numa_set_membind(mask);
1708                                 break;
1709                         case MPOL_LOCAL:
1710                                 numa_set_localalloc();
1711                                 break;
1712                         case MPOL_PREFERRED:
1713                                 numa_set_preferred(o->numa_mem_prefer_node);
1714                                 break;
1715                         case MPOL_DEFAULT:
1716                         default:
1717                                 break;
1718                         }
1719
1720                         if (mask)
1721                                 numa_free_nodemask(mask);
1722
1723                 }
1724         }
1725 #endif
1726
1727         if (fio_pin_memory(td))
1728                 goto err;
1729
1730         /*
1731          * May alter parameters that init_io_u() will use, so we need to
1732          * do this first.
1733          */
1734         if (init_iolog(td))
1735                 goto err;
1736
1737         if (init_io_u(td))
1738                 goto err;
1739
1740         if (o->verify_async && verify_async_init(td))
1741                 goto err;
1742
1743         if (fio_option_is_set(o, ioprio) ||
1744             fio_option_is_set(o, ioprio_class)) {
1745                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1746                 if (ret == -1) {
1747                         td_verror(td, errno, "ioprio_set");
1748                         goto err;
1749                 }
1750         }
1751
1752         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1753                 goto err;
1754
1755         errno = 0;
1756         if (nice(o->nice) == -1 && errno != 0) {
1757                 td_verror(td, errno, "nice");
1758                 goto err;
1759         }
1760
1761         if (o->ioscheduler && switch_ioscheduler(td))
1762                 goto err;
1763
1764         if (!o->create_serialize && setup_files(td))
1765                 goto err;
1766
1767         if (td_io_init(td))
1768                 goto err;
1769
1770         if (init_random_map(td))
1771                 goto err;
1772
1773         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1774                 goto err;
1775
1776         if (o->pre_read) {
1777                 if (pre_read_files(td) < 0)
1778                         goto err;
1779         }
1780
1781         if (td->flags & TD_F_COMPRESS_LOG)
1782                 tp_init(&td->tp_data);
1783
1784         fio_verify_init(td);
1785
1786         if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1787             workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth))
1788                 goto err;
1789
1790         fio_gettime(&td->epoch, NULL);
1791         fio_getrusage(&td->ru_start);
1792         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1793         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1794
1795         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1796                         o->ratemin[DDIR_TRIM]) {
1797                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1798                                         sizeof(td->bw_sample_time));
1799                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1800                                         sizeof(td->bw_sample_time));
1801                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1802                                         sizeof(td->bw_sample_time));
1803         }
1804
1805         clear_state = 0;
1806         while (keep_running(td)) {
1807                 uint64_t verify_bytes;
1808
1809                 fio_gettime(&td->start, NULL);
1810                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1811
1812                 if (clear_state)
1813                         clear_io_state(td, 0);
1814
1815                 prune_io_piece_log(td);
1816
1817                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1818                         verify_bytes = do_dry_run(td);
1819                 else
1820                         verify_bytes = do_io(td);
1821
1822                 clear_state = 1;
1823
1824                 /*
1825                  * Make sure we've successfully updated the rusage stats
1826                  * before waiting on the stat mutex. Otherwise we could have
1827                  * the stat thread holding stat mutex and waiting for
1828                  * the rusage_sem, which would never get upped because
1829                  * this thread is waiting for the stat mutex.
1830                  */
1831                 check_update_rusage(td);
1832
1833                 fio_mutex_down(stat_mutex);
1834                 if (td_read(td) && td->io_bytes[DDIR_READ])
1835                         update_runtime(td, elapsed_us, DDIR_READ);
1836                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1837                         update_runtime(td, elapsed_us, DDIR_WRITE);
1838                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1839                         update_runtime(td, elapsed_us, DDIR_TRIM);
1840                 fio_gettime(&td->start, NULL);
1841                 fio_mutex_up(stat_mutex);
1842
1843                 if (td->error || td->terminate)
1844                         break;
1845
1846                 if (!o->do_verify ||
1847                     o->verify == VERIFY_NONE ||
1848                     (td->io_ops->flags & FIO_UNIDIR))
1849                         continue;
1850
1851                 clear_io_state(td, 0);
1852
1853                 fio_gettime(&td->start, NULL);
1854
1855                 do_verify(td, verify_bytes);
1856
1857                 /*
1858                  * See comment further up for why this is done here.
1859                  */
1860                 check_update_rusage(td);
1861
1862                 fio_mutex_down(stat_mutex);
1863                 update_runtime(td, elapsed_us, DDIR_READ);
1864                 fio_gettime(&td->start, NULL);
1865                 fio_mutex_up(stat_mutex);
1866
1867                 if (td->error || td->terminate)
1868                         break;
1869         }
1870
1871         update_rusage_stat(td);
1872         td->ts.total_run_time = mtime_since_now(&td->epoch);
1873         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1874         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1875         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1876
1877         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1878             (td->o.verify != VERIFY_NONE && td_write(td)))
1879                 verify_save_state(td->thread_number);
1880
1881         fio_unpin_memory(td);
1882
1883         fio_writeout_logs(td);
1884
1885         if (o->io_submit_mode == IO_MODE_OFFLOAD)
1886                 workqueue_exit(&td->io_wq);
1887
1888         if (td->flags & TD_F_COMPRESS_LOG)
1889                 tp_exit(&td->tp_data);
1890
1891         if (o->exec_postrun)
1892                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1893
1894         if (exitall_on_terminate)
1895                 fio_terminate_threads(td->groupid);
1896
1897 err:
1898         if (td->error)
1899                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1900                                                         td->verror);
1901
1902         if (o->verify_async)
1903                 verify_async_exit(td);
1904
1905         close_and_free_files(td);
1906         cleanup_io_u(td);
1907         close_ioengine(td);
1908         cgroup_shutdown(td, &cgroup_mnt);
1909         verify_free_state(td);
1910
1911         if (fio_option_is_set(o, cpumask)) {
1912                 ret = fio_cpuset_exit(&o->cpumask);
1913                 if (ret)
1914                         td_verror(td, ret, "fio_cpuset_exit");
1915         }
1916
1917         /*
1918          * do this very late, it will log file closing as well
1919          */
1920         if (o->write_iolog_file)
1921                 write_iolog_close(td);
1922
1923         fio_mutex_remove(td->mutex);
1924         td->mutex = NULL;
1925
1926         td_set_runstate(td, TD_EXITED);
1927
1928         /*
1929          * Do this last after setting our runstate to exited, so we
1930          * know that the stat thread is signaled.
1931          */
1932         check_update_rusage(td);
1933
1934         return (void *) (uintptr_t) td->error;
1935 }
1936
1937
1938 /*
1939  * We cannot pass the td data into a forked process, so attach the td and
1940  * pass it to the thread worker.
1941  */
1942 static int fork_main(int shmid, int offset)
1943 {
1944         struct thread_data *td;
1945         void *data, *ret;
1946
1947 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1948         data = shmat(shmid, NULL, 0);
1949         if (data == (void *) -1) {
1950                 int __err = errno;
1951
1952                 perror("shmat");
1953                 return __err;
1954         }
1955 #else
1956         /*
1957          * HP-UX inherits shm mappings?
1958          */
1959         data = threads;
1960 #endif
1961
1962         td = data + offset * sizeof(struct thread_data);
1963         ret = thread_main(td);
1964         shmdt(data);
1965         return (int) (uintptr_t) ret;
1966 }
1967
1968 static void dump_td_info(struct thread_data *td)
1969 {
1970         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1971                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1972                         (unsigned long) time_since_now(&td->terminate_time));
1973 }
1974
1975 /*
1976  * Run over the job map and reap the threads that have exited, if any.
1977  */
1978 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1979                          unsigned int *m_rate)
1980 {
1981         struct thread_data *td;
1982         unsigned int cputhreads, realthreads, pending;
1983         int i, status, ret;
1984
1985         /*
1986          * reap exited threads (TD_EXITED -> TD_REAPED)
1987          */
1988         realthreads = pending = cputhreads = 0;
1989         for_each_td(td, i) {
1990                 int flags = 0;
1991
1992                 /*
1993                  * ->io_ops is NULL for a thread that has closed its
1994                  * io engine
1995                  */
1996                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1997                         cputhreads++;
1998                 else
1999                         realthreads++;
2000
2001                 if (!td->pid) {
2002                         pending++;
2003                         continue;
2004                 }
2005                 if (td->runstate == TD_REAPED)
2006                         continue;
2007                 if (td->o.use_thread) {
2008                         if (td->runstate == TD_EXITED) {
2009                                 td_set_runstate(td, TD_REAPED);
2010                                 goto reaped;
2011                         }
2012                         continue;
2013                 }
2014
2015                 flags = WNOHANG;
2016                 if (td->runstate == TD_EXITED)
2017                         flags = 0;
2018
2019                 /*
2020                  * check if someone quit or got killed in an unusual way
2021                  */
2022                 ret = waitpid(td->pid, &status, flags);
2023                 if (ret < 0) {
2024                         if (errno == ECHILD) {
2025                                 log_err("fio: pid=%d disappeared %d\n",
2026                                                 (int) td->pid, td->runstate);
2027                                 td->sig = ECHILD;
2028                                 td_set_runstate(td, TD_REAPED);
2029                                 goto reaped;
2030                         }
2031                         perror("waitpid");
2032                 } else if (ret == td->pid) {
2033                         if (WIFSIGNALED(status)) {
2034                                 int sig = WTERMSIG(status);
2035
2036                                 if (sig != SIGTERM && sig != SIGUSR2)
2037                                         log_err("fio: pid=%d, got signal=%d\n",
2038                                                         (int) td->pid, sig);
2039                                 td->sig = sig;
2040                                 td_set_runstate(td, TD_REAPED);
2041                                 goto reaped;
2042                         }
2043                         if (WIFEXITED(status)) {
2044                                 if (WEXITSTATUS(status) && !td->error)
2045                                         td->error = WEXITSTATUS(status);
2046
2047                                 td_set_runstate(td, TD_REAPED);
2048                                 goto reaped;
2049                         }
2050                 }
2051
2052                 /*
2053                  * If the job is stuck, do a forceful timeout of it and
2054                  * move on.
2055                  */
2056                 if (td->terminate &&
2057                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2058                         dump_td_info(td);
2059                         td_set_runstate(td, TD_REAPED);
2060                         goto reaped;
2061                 }
2062
2063                 /*
2064                  * thread is not dead, continue
2065                  */
2066                 pending++;
2067                 continue;
2068 reaped:
2069                 (*nr_running)--;
2070                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2071                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2072                 if (!td->pid)
2073                         pending--;
2074
2075                 if (td->error)
2076                         exit_value++;
2077
2078                 done_secs += mtime_since_now(&td->epoch) / 1000;
2079                 profile_td_exit(td);
2080         }
2081
2082         if (*nr_running == cputhreads && !pending && realthreads)
2083                 fio_terminate_threads(TERMINATE_ALL);
2084 }
2085
2086 static bool __check_trigger_file(void)
2087 {
2088         struct stat sb;
2089
2090         if (!trigger_file)
2091                 return false;
2092
2093         if (stat(trigger_file, &sb))
2094                 return false;
2095
2096         if (unlink(trigger_file) < 0)
2097                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2098                                                         strerror(errno));
2099
2100         return true;
2101 }
2102
2103 static bool trigger_timedout(void)
2104 {
2105         if (trigger_timeout)
2106                 return time_since_genesis() >= trigger_timeout;
2107
2108         return false;
2109 }
2110
2111 void exec_trigger(const char *cmd)
2112 {
2113         int ret;
2114
2115         if (!cmd)
2116                 return;
2117
2118         ret = system(cmd);
2119         if (ret == -1)
2120                 log_err("fio: failed executing %s trigger\n", cmd);
2121 }
2122
2123 void check_trigger_file(void)
2124 {
2125         if (__check_trigger_file() || trigger_timedout()) {
2126                 if (nr_clients)
2127                         fio_clients_send_trigger(trigger_remote_cmd);
2128                 else {
2129                         verify_save_state(IO_LIST_ALL);
2130                         fio_terminate_threads(TERMINATE_ALL);
2131                         exec_trigger(trigger_cmd);
2132                 }
2133         }
2134 }
2135
2136 static int fio_verify_load_state(struct thread_data *td)
2137 {
2138         int ret;
2139
2140         if (!td->o.verify_state)
2141                 return 0;
2142
2143         if (is_backend) {
2144                 void *data;
2145                 int ver;
2146
2147                 ret = fio_server_get_verify_state(td->o.name,
2148                                         td->thread_number - 1, &data, &ver);
2149                 if (!ret)
2150                         verify_convert_assign_state(td, data, ver);
2151         } else
2152                 ret = verify_load_state(td, "local");
2153
2154         return ret;
2155 }
2156
2157 static void do_usleep(unsigned int usecs)
2158 {
2159         check_for_running_stats();
2160         check_trigger_file();
2161         usleep(usecs);
2162 }
2163
2164 static bool check_mount_writes(struct thread_data *td)
2165 {
2166         struct fio_file *f;
2167         unsigned int i;
2168
2169         if (!td_write(td) || td->o.allow_mounted_write)
2170                 return false;
2171
2172         for_each_file(td, f, i) {
2173                 if (f->filetype != FIO_TYPE_BD)
2174                         continue;
2175                 if (device_is_mounted(f->file_name))
2176                         goto mounted;
2177         }
2178
2179         return false;
2180 mounted:
2181         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2182         return true;
2183 }
2184
2185 /*
2186  * Main function for kicking off and reaping jobs, as needed.
2187  */
2188 static void run_threads(void)
2189 {
2190         struct thread_data *td;
2191         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2192         uint64_t spent;
2193
2194         if (fio_gtod_offload && fio_start_gtod_thread())
2195                 return;
2196
2197         fio_idle_prof_init();
2198
2199         set_sig_handlers();
2200
2201         nr_thread = nr_process = 0;
2202         for_each_td(td, i) {
2203                 if (check_mount_writes(td))
2204                         return;
2205                 if (td->o.use_thread)
2206                         nr_thread++;
2207                 else
2208                         nr_process++;
2209         }
2210
2211         if (output_format & FIO_OUTPUT_NORMAL) {
2212                 log_info("Starting ");
2213                 if (nr_thread)
2214                         log_info("%d thread%s", nr_thread,
2215                                                 nr_thread > 1 ? "s" : "");
2216                 if (nr_process) {
2217                         if (nr_thread)
2218                                 log_info(" and ");
2219                         log_info("%d process%s", nr_process,
2220                                                 nr_process > 1 ? "es" : "");
2221                 }
2222                 log_info("\n");
2223                 log_info_flush();
2224         }
2225
2226         todo = thread_number;
2227         nr_running = 0;
2228         nr_started = 0;
2229         m_rate = t_rate = 0;
2230
2231         for_each_td(td, i) {
2232                 print_status_init(td->thread_number - 1);
2233
2234                 if (!td->o.create_serialize)
2235                         continue;
2236
2237                 if (fio_verify_load_state(td))
2238                         goto reap;
2239
2240                 /*
2241                  * do file setup here so it happens sequentially,
2242                  * we don't want X number of threads getting their
2243                  * client data interspersed on disk
2244                  */
2245                 if (setup_files(td)) {
2246 reap:
2247                         exit_value++;
2248                         if (td->error)
2249                                 log_err("fio: pid=%d, err=%d/%s\n",
2250                                         (int) td->pid, td->error, td->verror);
2251                         td_set_runstate(td, TD_REAPED);
2252                         todo--;
2253                 } else {
2254                         struct fio_file *f;
2255                         unsigned int j;
2256
2257                         /*
2258                          * for sharing to work, each job must always open
2259                          * its own files. so close them, if we opened them
2260                          * for creation
2261                          */
2262                         for_each_file(td, f, j) {
2263                                 if (fio_file_open(f))
2264                                         td_io_close_file(td, f);
2265                         }
2266                 }
2267         }
2268
2269         /* start idle threads before io threads start to run */
2270         fio_idle_prof_start();
2271
2272         set_genesis_time();
2273
2274         while (todo) {
2275                 struct thread_data *map[REAL_MAX_JOBS];
2276                 struct timeval this_start;
2277                 int this_jobs = 0, left;
2278
2279                 /*
2280                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2281                  */
2282                 for_each_td(td, i) {
2283                         if (td->runstate != TD_NOT_CREATED)
2284                                 continue;
2285
2286                         /*
2287                          * never got a chance to start, killed by other
2288                          * thread for some reason
2289                          */
2290                         if (td->terminate) {
2291                                 todo--;
2292                                 continue;
2293                         }
2294
2295                         if (td->o.start_delay) {
2296                                 spent = utime_since_genesis();
2297
2298                                 if (td->o.start_delay > spent)
2299                                         continue;
2300                         }
2301
2302                         if (td->o.stonewall && (nr_started || nr_running)) {
2303                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2304                                                         td->o.name);
2305                                 break;
2306                         }
2307
2308                         init_disk_util(td);
2309
2310                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2311                         td->update_rusage = 0;
2312
2313                         /*
2314                          * Set state to created. Thread will transition
2315                          * to TD_INITIALIZED when it's done setting up.
2316                          */
2317                         td_set_runstate(td, TD_CREATED);
2318                         map[this_jobs++] = td;
2319                         nr_started++;
2320
2321                         if (td->o.use_thread) {
2322                                 int ret;
2323
2324                                 dprint(FD_PROCESS, "will pthread_create\n");
2325                                 ret = pthread_create(&td->thread, NULL,
2326                                                         thread_main, td);
2327                                 if (ret) {
2328                                         log_err("pthread_create: %s\n",
2329                                                         strerror(ret));
2330                                         nr_started--;
2331                                         break;
2332                                 }
2333                                 ret = pthread_detach(td->thread);
2334                                 if (ret)
2335                                         log_err("pthread_detach: %s",
2336                                                         strerror(ret));
2337                         } else {
2338                                 pid_t pid;
2339                                 dprint(FD_PROCESS, "will fork\n");
2340                                 pid = fork();
2341                                 if (!pid) {
2342                                         int ret = fork_main(shm_id, i);
2343
2344                                         _exit(ret);
2345                                 } else if (i == fio_debug_jobno)
2346                                         *fio_debug_jobp = pid;
2347                         }
2348                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2349                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
2350                                 log_err("fio: job startup hung? exiting.\n");
2351                                 fio_terminate_threads(TERMINATE_ALL);
2352                                 fio_abort = 1;
2353                                 nr_started--;
2354                                 break;
2355                         }
2356                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2357                 }
2358
2359                 /*
2360                  * Wait for the started threads to transition to
2361                  * TD_INITIALIZED.
2362                  */
2363                 fio_gettime(&this_start, NULL);
2364                 left = this_jobs;
2365                 while (left && !fio_abort) {
2366                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2367                                 break;
2368
2369                         do_usleep(100000);
2370
2371                         for (i = 0; i < this_jobs; i++) {
2372                                 td = map[i];
2373                                 if (!td)
2374                                         continue;
2375                                 if (td->runstate == TD_INITIALIZED) {
2376                                         map[i] = NULL;
2377                                         left--;
2378                                 } else if (td->runstate >= TD_EXITED) {
2379                                         map[i] = NULL;
2380                                         left--;
2381                                         todo--;
2382                                         nr_running++; /* work-around... */
2383                                 }
2384                         }
2385                 }
2386
2387                 if (left) {
2388                         log_err("fio: %d job%s failed to start\n", left,
2389                                         left > 1 ? "s" : "");
2390                         for (i = 0; i < this_jobs; i++) {
2391                                 td = map[i];
2392                                 if (!td)
2393                                         continue;
2394                                 kill(td->pid, SIGTERM);
2395                         }
2396                         break;
2397                 }
2398
2399                 /*
2400                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2401                  */
2402                 for_each_td(td, i) {
2403                         if (td->runstate != TD_INITIALIZED)
2404                                 continue;
2405
2406                         if (in_ramp_time(td))
2407                                 td_set_runstate(td, TD_RAMP);
2408                         else
2409                                 td_set_runstate(td, TD_RUNNING);
2410                         nr_running++;
2411                         nr_started--;
2412                         m_rate += ddir_rw_sum(td->o.ratemin);
2413                         t_rate += ddir_rw_sum(td->o.rate);
2414                         todo--;
2415                         fio_mutex_up(td->mutex);
2416                 }
2417
2418                 reap_threads(&nr_running, &t_rate, &m_rate);
2419
2420                 if (todo)
2421                         do_usleep(100000);
2422         }
2423
2424         while (nr_running) {
2425                 reap_threads(&nr_running, &t_rate, &m_rate);
2426                 do_usleep(10000);
2427         }
2428
2429         fio_idle_prof_stop();
2430
2431         update_io_ticks();
2432 }
2433
2434 static void wait_for_helper_thread_exit(void)
2435 {
2436         void *ret;
2437
2438         helper_exit = 1;
2439         pthread_cond_signal(&helper_cond);
2440         pthread_join(helper_thread, &ret);
2441 }
2442
2443 static void free_disk_util(void)
2444 {
2445         disk_util_prune_entries();
2446
2447         pthread_cond_destroy(&helper_cond);
2448 }
2449
2450 static void *helper_thread_main(void *data)
2451 {
2452         int ret = 0;
2453
2454         fio_mutex_up(startup_mutex);
2455
2456         while (!ret) {
2457                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2458                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2459                 struct timespec ts;
2460                 struct timeval tv;
2461
2462                 gettimeofday(&tv, NULL);
2463                 ts.tv_sec = tv.tv_sec + sec;
2464                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2465
2466                 if (ts.tv_nsec >= 1000000000ULL) {
2467                         ts.tv_nsec -= 1000000000ULL;
2468                         ts.tv_sec++;
2469                 }
2470
2471                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2472
2473                 ret = update_io_ticks();
2474
2475                 if (helper_do_stat) {
2476                         helper_do_stat = 0;
2477                         __show_running_run_stats();
2478                 }
2479
2480                 if (!is_backend)
2481                         print_thread_status();
2482         }
2483
2484         return NULL;
2485 }
2486
2487 static int create_helper_thread(void)
2488 {
2489         int ret;
2490
2491         setup_disk_util();
2492
2493         pthread_cond_init(&helper_cond, NULL);
2494         pthread_mutex_init(&helper_lock, NULL);
2495
2496         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2497         if (ret) {
2498                 log_err("Can't create helper thread: %s\n", strerror(ret));
2499                 return 1;
2500         }
2501
2502         dprint(FD_MUTEX, "wait on startup_mutex\n");
2503         fio_mutex_down(startup_mutex);
2504         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2505         return 0;
2506 }
2507
2508 int fio_backend(void)
2509 {
2510         struct thread_data *td;
2511         int i;
2512
2513         if (exec_profile) {
2514                 if (load_profile(exec_profile))
2515                         return 1;
2516                 free(exec_profile);
2517                 exec_profile = NULL;
2518         }
2519         if (!thread_number)
2520                 return 0;
2521
2522         if (write_bw_log) {
2523                 struct log_params p = {
2524                         .log_type = IO_LOG_TYPE_BW,
2525                 };
2526
2527                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2528                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2529                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2530         }
2531
2532         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2533         if (startup_mutex == NULL)
2534                 return 1;
2535
2536         set_genesis_time();
2537         stat_init();
2538         create_helper_thread();
2539
2540         cgroup_list = smalloc(sizeof(*cgroup_list));
2541         INIT_FLIST_HEAD(cgroup_list);
2542
2543         run_threads();
2544
2545         wait_for_helper_thread_exit();
2546
2547         if (!fio_abort) {
2548                 __show_run_stats();
2549                 if (write_bw_log) {
2550                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2551                                 struct io_log *log = agg_io_log[i];
2552
2553                                 flush_log(log, 0);
2554                                 free_log(log);
2555                         }
2556                 }
2557         }
2558
2559         for_each_td(td, i) {
2560                 fio_options_free(td);
2561                 if (td->rusage_sem) {
2562                         fio_mutex_remove(td->rusage_sem);
2563                         td->rusage_sem = NULL;
2564                 }
2565         }
2566
2567         free_disk_util();
2568         cgroup_kill(cgroup_list);
2569         sfree(cgroup_list);
2570         sfree(cgroup_mnt);
2571
2572         fio_mutex_remove(startup_mutex);
2573         stat_exit();
2574         return exit_value;
2575 }