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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38 #include <math.h>
  39
  40 #include "fio.h"
  41 #ifndef FIO_NO_HAVE_SHM_H
  42 #include <sys/shm.h>
  43 #endif
  44 #include "hash.h"
  45 #include "smalloc.h"
  46 #include "verify.h"
  47 #include "trim.h"
  48 #include "diskutil.h"
  49 #include "cgroup.h"
  50 #include "profile.h"
  51 #include "lib/rand.h"
  52 #include "lib/memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59 #include "rate-submit.h"
  60 #include "helper_thread.h"
  61
  62 static struct fio_mutex *startup_mutex;
  63 static struct flist_head *cgroup_list;
  64 static char *cgroup_mnt;
  65 static int exit_value;
  66 static volatile int fio_abort;
  67 static unsigned int nr_process = 0;
  68 static unsigned int nr_thread = 0;
  69
  70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  71
  72 int groupid = 0;
  73 unsigned int thread_number = 0;
  74 unsigned int stat_number = 0;
  75 int shm_id = 0;
  76 int temp_stall_ts;
  77 unsigned long done_secs = 0;
  78
  79 #define JOB_START_TIMEOUT       (5 * 1000)
  80
  81 static void sig_int(int sig)
  82 {
  83         if (threads) {
  84                 if (is_backend)
  85                         fio_server_got_signal(sig);
  86                 else {
  87                         log_info("\nfio: terminating on signal %d\n", sig);
  88                         log_info_flush();
  89                         exit_value = 128;
  90                 }
  91
  92                 fio_terminate_threads(TERMINATE_ALL);
  93         }
  94 }
  95
  96 void sig_show_status(int sig)
  97 {
  98         show_running_run_stats();
  99 }
 100
 101 static void set_sig_handlers(void)
 102 {
 103         struct sigaction act;
 104
 105         memset(&act, 0, sizeof(act));
 106         act.sa_handler = sig_int;
 107         act.sa_flags = SA_RESTART;
 108         sigaction(SIGINT, &act, NULL);
 109
 110         memset(&act, 0, sizeof(act));
 111         act.sa_handler = sig_int;
 112         act.sa_flags = SA_RESTART;
 113         sigaction(SIGTERM, &act, NULL);
 114
 115 /* Windows uses SIGBREAK as a quit signal from other applications */
 116 #ifdef WIN32
 117         memset(&act, 0, sizeof(act));
 118         act.sa_handler = sig_int;
 119         act.sa_flags = SA_RESTART;
 120         sigaction(SIGBREAK, &act, NULL);
 121 #endif
 122
 123         memset(&act, 0, sizeof(act));
 124         act.sa_handler = sig_show_status;
 125         act.sa_flags = SA_RESTART;
 126         sigaction(SIGUSR1, &act, NULL);
 127
 128         if (is_backend) {
 129                 memset(&act, 0, sizeof(act));
 130                 act.sa_handler = sig_int;
 131                 act.sa_flags = SA_RESTART;
 132                 sigaction(SIGPIPE, &act, NULL);
 133         }
 134 }
 135
 136 /*
 137  * Check if we are above the minimum rate given.
 138  */
 139 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 140                              enum fio_ddir ddir)
 141 {
 142         unsigned long long bytes = 0;
 143         unsigned long iops = 0;
 144         unsigned long spent;
 145         unsigned long rate;
 146         unsigned int ratemin = 0;
 147         unsigned int rate_iops = 0;
 148         unsigned int rate_iops_min = 0;
 149
 150         assert(ddir_rw(ddir));
 151
 152         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 153                 return false;
 154
 155         /*
 156          * allow a 2 second settle period in the beginning
 157          */
 158         if (mtime_since(&td->start, now) < 2000)
 159                 return false;
 160
 161         iops += td->this_io_blocks[ddir];
 162         bytes += td->this_io_bytes[ddir];
 163         ratemin += td->o.ratemin[ddir];
 164         rate_iops += td->o.rate_iops[ddir];
 165         rate_iops_min += td->o.rate_iops_min[ddir];
 166
 167         /*
 168          * if rate blocks is set, sample is running
 169          */
 170         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 171                 spent = mtime_since(&td->lastrate[ddir], now);
 172                 if (spent < td->o.ratecycle)
 173                         return false;
 174
 175                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 176                         /*
 177                          * check bandwidth specified rate
 178                          */
 179                         if (bytes < td->rate_bytes[ddir]) {
 180                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 181                                         td->o.name, ratemin, bytes);
 182                                 return true;
 183                         } else {
 184                                 if (spent)
 185                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 186                                 else
 187                                         rate = 0;
 188
 189                                 if (rate < ratemin ||
 190                                     bytes < td->rate_bytes[ddir]) {
 191                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 192                                                 td->o.name, ratemin, rate);
 193                                         return true;
 194                                 }
 195                         }
 196                 } else {
 197                         /*
 198                          * checks iops specified rate
 199                          */
 200                         if (iops < rate_iops) {
 201                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 202                                                 td->o.name, rate_iops, iops);
 203                                 return true;
 204                         } else {
 205                                 if (spent)
 206                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 207                                 else
 208                                         rate = 0;
 209
 210                                 if (rate < rate_iops_min ||
 211                                     iops < td->rate_blocks[ddir]) {
 212                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 213                                                 td->o.name, rate_iops_min, rate);
 214                                         return true;
 215                                 }
 216                         }
 217                 }
 218         }
 219
 220         td->rate_bytes[ddir] = bytes;
 221         td->rate_blocks[ddir] = iops;
 222         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 223         return false;
 224 }
 225
 226 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 227 {
 228         bool ret = false;
 229
 230         if (td->bytes_done[DDIR_READ])
 231                 ret |= __check_min_rate(td, now, DDIR_READ);
 232         if (td->bytes_done[DDIR_WRITE])
 233                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 234         if (td->bytes_done[DDIR_TRIM])
 235                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 236
 237         return ret;
 238 }
 239
 240 /*
 241  * When job exits, we can cancel the in-flight IO if we are using async
 242  * io. Attempt to do so.
 243  */
 244 static void cleanup_pending_aio(struct thread_data *td)
 245 {
 246         int r;
 247
 248         /*
 249          * get immediately available events, if any
 250          */
 251         r = io_u_queued_complete(td, 0);
 252         if (r < 0)
 253                 return;
 254
 255         /*
 256          * now cancel remaining active events
 257          */
 258         if (td->io_ops->cancel) {
 259                 struct io_u *io_u;
 260                 int i;
 261
 262                 io_u_qiter(&td->io_u_all, io_u, i) {
 263                         if (io_u->flags & IO_U_F_FLIGHT) {
 264                                 r = td->io_ops->cancel(td, io_u);
 265                                 if (!r)
 266                                         put_io_u(td, io_u);
 267                         }
 268                 }
 269         }
 270
 271         if (td->cur_depth)
 272                 r = io_u_queued_complete(td, td->cur_depth);
 273 }
 274
 275 /*
 276  * Helper to handle the final sync of a file. Works just like the normal
 277  * io path, just does everything sync.
 278  */
 279 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 280 {
 281         struct io_u *io_u = __get_io_u(td);
 282         int ret;
 283
 284         if (!io_u)
 285                 return true;
 286
 287         io_u->ddir = DDIR_SYNC;
 288         io_u->file = f;
 289
 290         if (td_io_prep(td, io_u)) {
 291                 put_io_u(td, io_u);
 292                 return true;
 293         }
 294
 295 requeue:
 296         ret = td_io_queue(td, io_u);
 297         if (ret < 0) {
 298                 td_verror(td, io_u->error, "td_io_queue");
 299                 put_io_u(td, io_u);
 300                 return true;
 301         } else if (ret == FIO_Q_QUEUED) {
 302                 if (td_io_commit(td))
 303                         return true;
 304                 if (io_u_queued_complete(td, 1) < 0)
 305                         return true;
 306         } else if (ret == FIO_Q_COMPLETED) {
 307                 if (io_u->error) {
 308                         td_verror(td, io_u->error, "td_io_queue");
 309                         return true;
 310                 }
 311
 312                 if (io_u_sync_complete(td, io_u) < 0)
 313                         return true;
 314         } else if (ret == FIO_Q_BUSY) {
 315                 if (td_io_commit(td))
 316                         return true;
 317                 goto requeue;
 318         }
 319
 320         return false;
 321 }
 322
 323 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 324 {
 325         int ret;
 326
 327         if (fio_file_open(f))
 328                 return fio_io_sync(td, f);
 329
 330         if (td_io_open_file(td, f))
 331                 return 1;
 332
 333         ret = fio_io_sync(td, f);
 334         td_io_close_file(td, f);
 335         return ret;
 336 }
 337
 338 static inline void __update_ts_cache(struct thread_data *td)
 339 {
 340         fio_gettime(&td->ts_cache, NULL);
 341 }
 342
 343 static inline void update_ts_cache(struct thread_data *td)
 344 {
 345         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 346                 __update_ts_cache(td);
 347 }
 348
 349 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 350 {
 351         if (in_ramp_time(td))
 352                 return false;
 353         if (!td->o.timeout)
 354                 return false;
 355         if (utime_since(&td->epoch, t) >= td->o.timeout)
 356                 return true;
 357
 358         return false;
 359 }
 360
 361 /*
 362  * We need to update the runtime consistently in ms, but keep a running
 363  * tally of the current elapsed time in microseconds for sub millisecond
 364  * updates.
 365  */
 366 static inline void update_runtime(struct thread_data *td,
 367                                   unsigned long long *elapsed_us,
 368                                   const enum fio_ddir ddir)
 369 {
 370         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 371                 return;
 372
 373         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 374         elapsed_us[ddir] += utime_since_now(&td->start);
 375         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 376 }
 377
 378 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 379                                 int *retptr)
 380 {
 381         int ret = *retptr;
 382
 383         if (ret < 0 || td->error) {
 384                 int err = td->error;
 385                 enum error_type_bit eb;
 386
 387                 if (ret < 0)
 388                         err = -ret;
 389
 390                 eb = td_error_type(ddir, err);
 391                 if (!(td->o.continue_on_error & (1 << eb)))
 392                         return true;
 393
 394                 if (td_non_fatal_error(td, eb, err)) {
 395                         /*
 396                          * Continue with the I/Os in case of
 397                          * a non fatal error.
 398                          */
 399                         update_error_count(td, err);
 400                         td_clear_error(td);
 401                         *retptr = 0;
 402                         return false;
 403                 } else if (td->o.fill_device && err == ENOSPC) {
 404                         /*
 405                          * We expect to hit this error if
 406                          * fill_device option is set.
 407                          */
 408                         td_clear_error(td);
 409                         fio_mark_td_terminate(td);
 410                         return true;
 411                 } else {
 412                         /*
 413                          * Stop the I/O in case of a fatal
 414                          * error.
 415                          */
 416                         update_error_count(td, err);
 417                         return true;
 418                 }
 419         }
 420
 421         return false;
 422 }
 423
 424 static void check_update_rusage(struct thread_data *td)
 425 {
 426         if (td->update_rusage) {
 427                 td->update_rusage = 0;
 428                 update_rusage_stat(td);
 429                 fio_mutex_up(td->rusage_sem);
 430         }
 431 }
 432
 433 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 434 {
 435         const int full = queue_full(td);
 436         int min_evts = 0;
 437         int ret;
 438
 439         if (td->flags & TD_F_REGROW_LOGS)
 440                 return io_u_quiesce(td);
 441
 442         /*
 443          * if the queue is full, we MUST reap at least 1 event
 444          */
 445         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 446         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 447                 min_evts = 1;
 448
 449         if (time && (__should_check_rate(td, DDIR_READ) ||
 450             __should_check_rate(td, DDIR_WRITE) ||
 451             __should_check_rate(td, DDIR_TRIM)))
 452                 fio_gettime(time, NULL);
 453
 454         do {
 455                 ret = io_u_queued_complete(td, min_evts);
 456                 if (ret < 0)
 457                         break;
 458         } while (full && (td->cur_depth > td->o.iodepth_low));
 459
 460         return ret;
 461 }
 462
 463 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 464                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 465                    struct timespec *comp_time)
 466 {
 467         int ret2;
 468
 469         switch (*ret) {
 470         case FIO_Q_COMPLETED:
 471                 if (io_u->error) {
 472                         *ret = -io_u->error;
 473                         clear_io_u(td, io_u);
 474                 } else if (io_u->resid) {
 475                         int bytes = io_u->xfer_buflen - io_u->resid;
 476                         struct fio_file *f = io_u->file;
 477
 478                         if (bytes_issued)
 479                                 *bytes_issued += bytes;
 480
 481                         if (!from_verify)
 482                                 trim_io_piece(td, io_u);
 483
 484                         /*
 485                          * zero read, fail
 486                          */
 487                         if (!bytes) {
 488                                 if (!from_verify)
 489                                         unlog_io_piece(td, io_u);
 490                                 td_verror(td, EIO, "full resid");
 491                                 put_io_u(td, io_u);
 492                                 break;
 493                         }
 494
 495                         io_u->xfer_buflen = io_u->resid;
 496                         io_u->xfer_buf += bytes;
 497                         io_u->offset += bytes;
 498
 499                         if (ddir_rw(io_u->ddir))
 500                                 td->ts.short_io_u[io_u->ddir]++;
 501
 502                         if (io_u->offset == f->real_file_size)
 503                                 goto sync_done;
 504
 505                         requeue_io_u(td, &io_u);
 506                 } else {
 507 sync_done:
 508                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 509                             __should_check_rate(td, DDIR_WRITE) ||
 510                             __should_check_rate(td, DDIR_TRIM)))
 511                                 fio_gettime(comp_time, NULL);
 512
 513                         *ret = io_u_sync_complete(td, io_u);
 514                         if (*ret < 0)
 515                                 break;
 516                 }
 517
 518                 if (td->flags & TD_F_REGROW_LOGS)
 519                         regrow_logs(td);
 520
 521                 /*
 522                  * when doing I/O (not when verifying),
 523                  * check for any errors that are to be ignored
 524                  */
 525                 if (!from_verify)
 526                         break;
 527
 528                 return 0;
 529         case FIO_Q_QUEUED:
 530                 /*
 531                  * if the engine doesn't have a commit hook,
 532                  * the io_u is really queued. if it does have such
 533                  * a hook, it has to call io_u_queued() itself.
 534                  */
 535                 if (td->io_ops->commit == NULL)
 536                         io_u_queued(td, io_u);
 537                 if (bytes_issued)
 538                         *bytes_issued += io_u->xfer_buflen;
 539                 break;
 540         case FIO_Q_BUSY:
 541                 if (!from_verify)
 542                         unlog_io_piece(td, io_u);
 543                 requeue_io_u(td, &io_u);
 544                 ret2 = td_io_commit(td);
 545                 if (ret2 < 0)
 546                         *ret = ret2;
 547                 break;
 548         default:
 549                 assert(*ret < 0);
 550                 td_verror(td, -(*ret), "td_io_queue");
 551                 break;
 552         }
 553
 554         if (break_on_this_error(td, ddir, ret))
 555                 return 1;
 556
 557         return 0;
 558 }
 559
 560 static inline bool io_in_polling(struct thread_data *td)
 561 {
 562         return !td->o.iodepth_batch_complete_min &&
 563                    !td->o.iodepth_batch_complete_max;
 564 }
 565 /*
 566  * Unlinks files from thread data fio_file structure
 567  */
 568 static int unlink_all_files(struct thread_data *td)
 569 {
 570         struct fio_file *f;
 571         unsigned int i;
 572         int ret = 0;
 573
 574         for_each_file(td, f, i) {
 575                 if (f->filetype != FIO_TYPE_FILE)
 576                         continue;
 577                 ret = td_io_unlink_file(td, f);
 578                 if (ret)
 579                         break;
 580         }
 581
 582         if (ret)
 583                 td_verror(td, ret, "unlink_all_files");
 584
 585         return ret;
 586 }
 587
 588 /*
 589  * Check if io_u will overlap an in-flight IO in the queue
 590  */
 591 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 592 {
 593         bool overlap;
 594         struct io_u *check_io_u;
 595         unsigned long long x1, x2, y1, y2;
 596         int i;
 597
 598         x1 = io_u->offset;
 599         x2 = io_u->offset + io_u->buflen;
 600         overlap = false;
 601         io_u_qiter(q, check_io_u, i) {
 602                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 603                         y1 = check_io_u->offset;
 604                         y2 = check_io_u->offset + check_io_u->buflen;
 605
 606                         if (x1 < y2 && y1 < x2) {
 607                                 overlap = true;
 608                                 dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
 609                                                 x1, io_u->buflen,
 610                                                 y1, check_io_u->buflen);
 611                                 break;
 612                         }
 613                 }
 614         }
 615
 616         return overlap;
 617 }
 618
 619 static int io_u_submit(struct thread_data *td, struct io_u *io_u)
 620 {
 621         /*
 622          * Check for overlap if the user asked us to, and we have
 623          * at least one IO in flight besides this one.
 624          */
 625         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 626             in_flight_overlap(&td->io_u_all, io_u))
 627                 return FIO_Q_BUSY;
 628
 629         return td_io_queue(td, io_u);
 630 }
 631
 632 /*
 633  * The main verify engine. Runs over the writes we previously submitted,
 634  * reads the blocks back in, and checks the crc/md5 of the data.
 635  */
 636 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 637 {
 638         struct fio_file *f;
 639         struct io_u *io_u;
 640         int ret, min_events;
 641         unsigned int i;
 642
 643         dprint(FD_VERIFY, "starting loop\n");
 644
 645         /*
 646          * sync io first and invalidate cache, to make sure we really
 647          * read from disk.
 648          */
 649         for_each_file(td, f, i) {
 650                 if (!fio_file_open(f))
 651                         continue;
 652                 if (fio_io_sync(td, f))
 653                         break;
 654                 if (file_invalidate_cache(td, f))
 655                         break;
 656         }
 657
 658         check_update_rusage(td);
 659
 660         if (td->error)
 661                 return;
 662
 663         /*
 664          * verify_state needs to be reset before verification
 665          * proceeds so that expected random seeds match actual
 666          * random seeds in headers. The main loop will reset
 667          * all random number generators if randrepeat is set.
 668          */
 669         if (!td->o.rand_repeatable)
 670                 td_fill_verify_state_seed(td);
 671
 672         td_set_runstate(td, TD_VERIFYING);
 673
 674         io_u = NULL;
 675         while (!td->terminate) {
 676                 enum fio_ddir ddir;
 677                 int full;
 678
 679                 update_ts_cache(td);
 680                 check_update_rusage(td);
 681
 682                 if (runtime_exceeded(td, &td->ts_cache)) {
 683                         __update_ts_cache(td);
 684                         if (runtime_exceeded(td, &td->ts_cache)) {
 685                                 fio_mark_td_terminate(td);
 686                                 break;
 687                         }
 688                 }
 689
 690                 if (flow_threshold_exceeded(td))
 691                         continue;
 692
 693                 if (!td->o.experimental_verify) {
 694                         io_u = __get_io_u(td);
 695                         if (!io_u)
 696                                 break;
 697
 698                         if (get_next_verify(td, io_u)) {
 699                                 put_io_u(td, io_u);
 700                                 break;
 701                         }
 702
 703                         if (td_io_prep(td, io_u)) {
 704                                 put_io_u(td, io_u);
 705                                 break;
 706                         }
 707                 } else {
 708                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 709                                 break;
 710
 711                         while ((io_u = get_io_u(td)) != NULL) {
 712                                 if (IS_ERR_OR_NULL(io_u)) {
 713                                         io_u = NULL;
 714                                         ret = FIO_Q_BUSY;
 715                                         goto reap;
 716                                 }
 717
 718                                 /*
 719                                  * We are only interested in the places where
 720                                  * we wrote or trimmed IOs. Turn those into
 721                                  * reads for verification purposes.
 722                                  */
 723                                 if (io_u->ddir == DDIR_READ) {
 724                                         /*
 725                                          * Pretend we issued it for rwmix
 726                                          * accounting
 727                                          */
 728                                         td->io_issues[DDIR_READ]++;
 729                                         put_io_u(td, io_u);
 730                                         continue;
 731                                 } else if (io_u->ddir == DDIR_TRIM) {
 732                                         io_u->ddir = DDIR_READ;
 733                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 734                                         break;
 735                                 } else if (io_u->ddir == DDIR_WRITE) {
 736                                         io_u->ddir = DDIR_READ;
 737                                         break;
 738                                 } else {
 739                                         put_io_u(td, io_u);
 740                                         continue;
 741                                 }
 742                         }
 743
 744                         if (!io_u)
 745                                 break;
 746                 }
 747
 748                 if (verify_state_should_stop(td, io_u)) {
 749                         put_io_u(td, io_u);
 750                         break;
 751                 }
 752
 753                 if (td->o.verify_async)
 754                         io_u->end_io = verify_io_u_async;
 755                 else
 756                         io_u->end_io = verify_io_u;
 757
 758                 ddir = io_u->ddir;
 759                 if (!td->o.disable_slat)
 760                         fio_gettime(&io_u->start_time, NULL);
 761
 762                 ret = io_u_submit(td, io_u);
 763
 764                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 765                         break;
 766
 767                 /*
 768                  * if we can queue more, do so. but check if there are
 769                  * completed io_u's first. Note that we can get BUSY even
 770                  * without IO queued, if the system is resource starved.
 771                  */
 772 reap:
 773                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 774                 if (full || io_in_polling(td))
 775                         ret = wait_for_completions(td, NULL);
 776
 777                 if (ret < 0)
 778                         break;
 779         }
 780
 781         check_update_rusage(td);
 782
 783         if (!td->error) {
 784                 min_events = td->cur_depth;
 785
 786                 if (min_events)
 787                         ret = io_u_queued_complete(td, min_events);
 788         } else
 789                 cleanup_pending_aio(td);
 790
 791         td_set_runstate(td, TD_RUNNING);
 792
 793         dprint(FD_VERIFY, "exiting loop\n");
 794 }
 795
 796 static bool exceeds_number_ios(struct thread_data *td)
 797 {
 798         unsigned long long number_ios;
 799
 800         if (!td->o.number_ios)
 801                 return false;
 802
 803         number_ios = ddir_rw_sum(td->io_blocks);
 804         number_ios += td->io_u_queued + td->io_u_in_flight;
 805
 806         return number_ios >= (td->o.number_ios * td->loops);
 807 }
 808
 809 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 810 {
 811         unsigned long long bytes, limit;
 812
 813         if (td_rw(td))
 814                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 815         else if (td_write(td))
 816                 bytes = this_bytes[DDIR_WRITE];
 817         else if (td_read(td))
 818                 bytes = this_bytes[DDIR_READ];
 819         else
 820                 bytes = this_bytes[DDIR_TRIM];
 821
 822         if (td->o.io_size)
 823                 limit = td->o.io_size;
 824         else
 825                 limit = td->o.size;
 826
 827         limit *= td->loops;
 828         return bytes >= limit || exceeds_number_ios(td);
 829 }
 830
 831 static bool io_issue_bytes_exceeded(struct thread_data *td)
 832 {
 833         return io_bytes_exceeded(td, td->io_issue_bytes);
 834 }
 835
 836 static bool io_complete_bytes_exceeded(struct thread_data *td)
 837 {
 838         return io_bytes_exceeded(td, td->this_io_bytes);
 839 }
 840
 841 /*
 842  * used to calculate the next io time for rate control
 843  *
 844  */
 845 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 846 {
 847         uint64_t bps = td->rate_bps[ddir];
 848
 849         assert(!(td->flags & TD_F_CHILD));
 850
 851         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 852                 uint64_t val, iops;
 853
 854                 iops = bps / td->o.bs[ddir];
 855                 val = (int64_t) (1000000 / iops) *
 856                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 857                 if (val) {
 858                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 859                                         (unsigned long long) 1000000 / val,
 860                                         ddir);
 861                 }
 862                 td->last_usec[ddir] += val;
 863                 return td->last_usec[ddir];
 864         } else if (bps) {
 865                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 866                 uint64_t secs = bytes / bps;
 867                 uint64_t remainder = bytes % bps;
 868
 869                 return remainder * 1000000 / bps + secs * 1000000;
 870         }
 871
 872         return 0;
 873 }
 874
 875 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 876 {
 877         unsigned long long b;
 878         uint64_t total;
 879         int left;
 880
 881         b = ddir_rw_sum(td->io_blocks);
 882         if (b % td->o.thinktime_blocks)
 883                 return;
 884
 885         io_u_quiesce(td);
 886
 887         total = 0;
 888         if (td->o.thinktime_spin)
 889                 total = usec_spin(td->o.thinktime_spin);
 890
 891         left = td->o.thinktime - total;
 892         if (left)
 893                 total += usec_sleep(td, left);
 894
 895         /*
 896          * If we're ignoring thinktime for the rate, add the number of bytes
 897          * we would have done while sleeping, minus one block to ensure we
 898          * start issuing immediately after the sleep.
 899          */
 900         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 901                 td->rate_io_issue_bytes[ddir] += (td->rate_bps[ddir] * 1000000) / total;
 902                 td->rate_io_issue_bytes[ddir] -= td->o.min_bs[ddir];
 903         }
 904 }
 905
 906 /*
 907  * Main IO worker function. It retrieves io_u's to process and queues
 908  * and reaps them, checking for rate and errors along the way.
 909  *
 910  * Returns number of bytes written and trimmed.
 911  */
 912 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 913 {
 914         unsigned int i;
 915         int ret = 0;
 916         uint64_t total_bytes, bytes_issued = 0;
 917
 918         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 919                 bytes_done[i] = td->bytes_done[i];
 920
 921         if (in_ramp_time(td))
 922                 td_set_runstate(td, TD_RAMP);
 923         else
 924                 td_set_runstate(td, TD_RUNNING);
 925
 926         lat_target_init(td);
 927
 928         total_bytes = td->o.size;
 929         /*
 930         * Allow random overwrite workloads to write up to io_size
 931         * before starting verification phase as 'size' doesn't apply.
 932         */
 933         if (td_write(td) && td_random(td) && td->o.norandommap)
 934                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 935         /*
 936          * If verify_backlog is enabled, we'll run the verify in this
 937          * handler as well. For that case, we may need up to twice the
 938          * amount of bytes.
 939          */
 940         if (td->o.verify != VERIFY_NONE &&
 941            (td_write(td) && td->o.verify_backlog))
 942                 total_bytes += td->o.size;
 943
 944         /* In trimwrite mode, each byte is trimmed and then written, so
 945          * allow total_bytes to be twice as big */
 946         if (td_trimwrite(td))
 947                 total_bytes += td->total_io_size;
 948
 949         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 950                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 951                 td->o.time_based) {
 952                 struct timespec comp_time;
 953                 struct io_u *io_u;
 954                 int full;
 955                 enum fio_ddir ddir;
 956
 957                 check_update_rusage(td);
 958
 959                 if (td->terminate || td->done)
 960                         break;
 961
 962                 update_ts_cache(td);
 963
 964                 if (runtime_exceeded(td, &td->ts_cache)) {
 965                         __update_ts_cache(td);
 966                         if (runtime_exceeded(td, &td->ts_cache)) {
 967                                 fio_mark_td_terminate(td);
 968                                 break;
 969                         }
 970                 }
 971
 972                 if (flow_threshold_exceeded(td))
 973                         continue;
 974
 975                 /*
 976                  * Break if we exceeded the bytes. The exception is time
 977                  * based runs, but we still need to break out of the loop
 978                  * for those to run verification, if enabled.
 979                  */
 980                 if (bytes_issued >= total_bytes &&
 981                     (!td->o.time_based ||
 982                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 983                         break;
 984
 985                 io_u = get_io_u(td);
 986                 if (IS_ERR_OR_NULL(io_u)) {
 987                         int err = PTR_ERR(io_u);
 988
 989                         io_u = NULL;
 990                         ddir = DDIR_INVAL;
 991                         if (err == -EBUSY) {
 992                                 ret = FIO_Q_BUSY;
 993                                 goto reap;
 994                         }
 995                         if (td->o.latency_target)
 996                                 goto reap;
 997                         break;
 998                 }
 999
1000                 ddir = io_u->ddir;
1001
1002                 /*
1003                  * Add verification end_io handler if:
1004                  *      - Asked to verify (!td_rw(td))
1005                  *      - Or the io_u is from our verify list (mixed write/ver)
1006                  */
1007                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1008                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1009
1010                         if (!td->o.verify_pattern_bytes) {
1011                                 io_u->rand_seed = __rand(&td->verify_state);
1012                                 if (sizeof(int) != sizeof(long *))
1013                                         io_u->rand_seed *= __rand(&td->verify_state);
1014                         }
1015
1016                         if (verify_state_should_stop(td, io_u)) {
1017                                 put_io_u(td, io_u);
1018                                 break;
1019                         }
1020
1021                         if (td->o.verify_async)
1022                                 io_u->end_io = verify_io_u_async;
1023                         else
1024                                 io_u->end_io = verify_io_u;
1025                         td_set_runstate(td, TD_VERIFYING);
1026                 } else if (in_ramp_time(td))
1027                         td_set_runstate(td, TD_RAMP);
1028                 else
1029                         td_set_runstate(td, TD_RUNNING);
1030
1031                 /*
1032                  * Always log IO before it's issued, so we know the specific
1033                  * order of it. The logged unit will track when the IO has
1034                  * completed.
1035                  */
1036                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1037                     td->o.do_verify &&
1038                     td->o.verify != VERIFY_NONE &&
1039                     !td->o.experimental_verify)
1040                         log_io_piece(td, io_u);
1041
1042                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1043                         const unsigned long blen = io_u->xfer_buflen;
1044                         const enum fio_ddir ddir = acct_ddir(io_u);
1045
1046                         if (td->error)
1047                                 break;
1048
1049                         workqueue_enqueue(&td->io_wq, &io_u->work);
1050                         ret = FIO_Q_QUEUED;
1051
1052                         if (ddir_rw(ddir)) {
1053                                 td->io_issues[ddir]++;
1054                                 td->io_issue_bytes[ddir] += blen;
1055                                 td->rate_io_issue_bytes[ddir] += blen;
1056                         }
1057
1058                         if (should_check_rate(td))
1059                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1060
1061                 } else {
1062                         ret = io_u_submit(td, io_u);
1063
1064                         if (should_check_rate(td))
1065                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1066
1067                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1068                                 break;
1069
1070                         /*
1071                          * See if we need to complete some commands. Note that
1072                          * we can get BUSY even without IO queued, if the
1073                          * system is resource starved.
1074                          */
1075 reap:
1076                         full = queue_full(td) ||
1077                                 (ret == FIO_Q_BUSY && td->cur_depth);
1078                         if (full || io_in_polling(td))
1079                                 ret = wait_for_completions(td, &comp_time);
1080                 }
1081                 if (ret < 0)
1082                         break;
1083                 if (!ddir_rw_sum(td->bytes_done) &&
1084                     !td_ioengine_flagged(td, FIO_NOIO))
1085                         continue;
1086
1087                 if (!in_ramp_time(td) && should_check_rate(td)) {
1088                         if (check_min_rate(td, &comp_time)) {
1089                                 if (exitall_on_terminate || td->o.exitall_error)
1090                                         fio_terminate_threads(td->groupid);
1091                                 td_verror(td, EIO, "check_min_rate");
1092                                 break;
1093                         }
1094                 }
1095                 if (!in_ramp_time(td) && td->o.latency_target)
1096                         lat_target_check(td);
1097
1098                 if (ddir_rw(ddir) && td->o.thinktime)
1099                         handle_thinktime(td, ddir);
1100         }
1101
1102         check_update_rusage(td);
1103
1104         if (td->trim_entries)
1105                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1106
1107         if (td->o.fill_device && td->error == ENOSPC) {
1108                 td->error = 0;
1109                 fio_mark_td_terminate(td);
1110         }
1111         if (!td->error) {
1112                 struct fio_file *f;
1113
1114                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1115                         workqueue_flush(&td->io_wq);
1116                         i = 0;
1117                 } else
1118                         i = td->cur_depth;
1119
1120                 if (i) {
1121                         ret = io_u_queued_complete(td, i);
1122                         if (td->o.fill_device && td->error == ENOSPC)
1123                                 td->error = 0;
1124                 }
1125
1126                 if (should_fsync(td) && td->o.end_fsync) {
1127                         td_set_runstate(td, TD_FSYNCING);
1128
1129                         for_each_file(td, f, i) {
1130                                 if (!fio_file_fsync(td, f))
1131                                         continue;
1132
1133                                 log_err("fio: end_fsync failed for file %s\n",
1134                                                                 f->file_name);
1135                         }
1136                 }
1137         } else
1138                 cleanup_pending_aio(td);
1139
1140         /*
1141          * stop job if we failed doing any IO
1142          */
1143         if (!ddir_rw_sum(td->this_io_bytes))
1144                 td->done = 1;
1145
1146         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1147                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1148 }
1149
1150 static void free_file_completion_logging(struct thread_data *td)
1151 {
1152         struct fio_file *f;
1153         unsigned int i;
1154
1155         for_each_file(td, f, i) {
1156                 if (!f->last_write_comp)
1157                         break;
1158                 sfree(f->last_write_comp);
1159         }
1160 }
1161
1162 static int init_file_completion_logging(struct thread_data *td,
1163                                         unsigned int depth)
1164 {
1165         struct fio_file *f;
1166         unsigned int i;
1167
1168         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1169                 return 0;
1170
1171         for_each_file(td, f, i) {
1172                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1173                 if (!f->last_write_comp)
1174                         goto cleanup;
1175         }
1176
1177         return 0;
1178
1179 cleanup:
1180         free_file_completion_logging(td);
1181         log_err("fio: failed to alloc write comp data\n");
1182         return 1;
1183 }
1184
1185 static void cleanup_io_u(struct thread_data *td)
1186 {
1187         struct io_u *io_u;
1188
1189         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1190
1191                 if (td->io_ops->io_u_free)
1192                         td->io_ops->io_u_free(td, io_u);
1193
1194                 fio_memfree(io_u, sizeof(*io_u));
1195         }
1196
1197         free_io_mem(td);
1198
1199         io_u_rexit(&td->io_u_requeues);
1200         io_u_qexit(&td->io_u_freelist);
1201         io_u_qexit(&td->io_u_all);
1202
1203         free_file_completion_logging(td);
1204 }
1205
1206 static int init_io_u(struct thread_data *td)
1207 {
1208         struct io_u *io_u;
1209         unsigned int max_bs, min_write;
1210         int cl_align, i, max_units;
1211         int data_xfer = 1, err;
1212         char *p;
1213
1214         max_units = td->o.iodepth;
1215         max_bs = td_max_bs(td);
1216         min_write = td->o.min_bs[DDIR_WRITE];
1217         td->orig_buffer_size = (unsigned long long) max_bs
1218                                         * (unsigned long long) max_units;
1219
1220         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1221                 data_xfer = 0;
1222
1223         err = 0;
1224         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1225         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1226         err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1227
1228         if (err) {
1229                 log_err("fio: failed setting up IO queues\n");
1230                 return 1;
1231         }
1232
1233         /*
1234          * if we may later need to do address alignment, then add any
1235          * possible adjustment here so that we don't cause a buffer
1236          * overflow later. this adjustment may be too much if we get
1237          * lucky and the allocator gives us an aligned address.
1238          */
1239         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1240             td_ioengine_flagged(td, FIO_RAWIO))
1241                 td->orig_buffer_size += page_mask + td->o.mem_align;
1242
1243         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1244                 unsigned long bs;
1245
1246                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1247                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1248         }
1249
1250         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1251                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1252                 return 1;
1253         }
1254
1255         if (data_xfer && allocate_io_mem(td))
1256                 return 1;
1257
1258         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1259             td_ioengine_flagged(td, FIO_RAWIO))
1260                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1261         else
1262                 p = td->orig_buffer;
1263
1264         cl_align = os_cache_line_size();
1265
1266         for (i = 0; i < max_units; i++) {
1267                 void *ptr;
1268
1269                 if (td->terminate)
1270                         return 1;
1271
1272                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1273                 if (!ptr) {
1274                         log_err("fio: unable to allocate aligned memory\n");
1275                         break;
1276                 }
1277
1278                 io_u = ptr;
1279                 memset(io_u, 0, sizeof(*io_u));
1280                 INIT_FLIST_HEAD(&io_u->verify_list);
1281                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1282
1283                 if (data_xfer) {
1284                         io_u->buf = p;
1285                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1286
1287                         if (td_write(td))
1288                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1289                         if (td_write(td) && td->o.verify_pattern_bytes) {
1290                                 /*
1291                                  * Fill the buffer with the pattern if we are
1292                                  * going to be doing writes.
1293                                  */
1294                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1295                         }
1296                 }
1297
1298                 io_u->index = i;
1299                 io_u->flags = IO_U_F_FREE;
1300                 io_u_qpush(&td->io_u_freelist, io_u);
1301
1302                 /*
1303                  * io_u never leaves this stack, used for iteration of all
1304                  * io_u buffers.
1305                  */
1306                 io_u_qpush(&td->io_u_all, io_u);
1307
1308                 if (td->io_ops->io_u_init) {
1309                         int ret = td->io_ops->io_u_init(td, io_u);
1310
1311                         if (ret) {
1312                                 log_err("fio: failed to init engine data: %d\n", ret);
1313                                 return 1;
1314                         }
1315                 }
1316
1317                 p += max_bs;
1318         }
1319
1320         if (init_file_completion_logging(td, max_units))
1321                 return 1;
1322
1323         return 0;
1324 }
1325
1326 /*
1327  * This function is Linux specific.
1328  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1329  */
1330 static int switch_ioscheduler(struct thread_data *td)
1331 {
1332 #ifdef FIO_HAVE_IOSCHED_SWITCH
1333         char tmp[256], tmp2[128];
1334         FILE *f;
1335         int ret;
1336
1337         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1338                 return 0;
1339
1340         assert(td->files && td->files[0]);
1341         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1342
1343         f = fopen(tmp, "r+");
1344         if (!f) {
1345                 if (errno == ENOENT) {
1346                         log_err("fio: os or kernel doesn't support IO scheduler"
1347                                 " switching\n");
1348                         return 0;
1349                 }
1350                 td_verror(td, errno, "fopen iosched");
1351                 return 1;
1352         }
1353
1354         /*
1355          * Set io scheduler.
1356          */
1357         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1358         if (ferror(f) || ret != 1) {
1359                 td_verror(td, errno, "fwrite");
1360                 fclose(f);
1361                 return 1;
1362         }
1363
1364         rewind(f);
1365
1366         /*
1367          * Read back and check that the selected scheduler is now the default.
1368          */
1369         memset(tmp, 0, sizeof(tmp));
1370         ret = fread(tmp, sizeof(tmp), 1, f);
1371         if (ferror(f) || ret < 0) {
1372                 td_verror(td, errno, "fread");
1373                 fclose(f);
1374                 return 1;
1375         }
1376         /*
1377          * either a list of io schedulers or "none\n" is expected.
1378          */
1379         tmp[strlen(tmp) - 1] = '\0';
1380
1381         /*
1382          * Write to "none" entry doesn't fail, so check the result here.
1383          */
1384         if (!strcmp(tmp, "none")) {
1385                 log_err("fio: io scheduler is not tunable\n");
1386                 fclose(f);
1387                 return 0;
1388         }
1389
1390         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1391         if (!strstr(tmp, tmp2)) {
1392                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1393                 td_verror(td, EINVAL, "iosched_switch");
1394                 fclose(f);
1395                 return 1;
1396         }
1397
1398         fclose(f);
1399         return 0;
1400 #else
1401         return 0;
1402 #endif
1403 }
1404
1405 static bool keep_running(struct thread_data *td)
1406 {
1407         unsigned long long limit;
1408
1409         if (td->done)
1410                 return false;
1411         if (td->terminate)
1412                 return false;
1413         if (td->o.time_based)
1414                 return true;
1415         if (td->o.loops) {
1416                 td->o.loops--;
1417                 return true;
1418         }
1419         if (exceeds_number_ios(td))
1420                 return false;
1421
1422         if (td->o.io_size)
1423                 limit = td->o.io_size;
1424         else
1425                 limit = td->o.size;
1426
1427         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1428                 uint64_t diff;
1429
1430                 /*
1431                  * If the difference is less than the maximum IO size, we
1432                  * are done.
1433                  */
1434                 diff = limit - ddir_rw_sum(td->io_bytes);
1435                 if (diff < td_max_bs(td))
1436                         return false;
1437
1438                 if (fio_files_done(td) && !td->o.io_size)
1439                         return false;
1440
1441                 return true;
1442         }
1443
1444         return false;
1445 }
1446
1447 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1448 {
1449         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1450         int ret;
1451         char *str;
1452
1453         str = malloc(newlen);
1454         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1455
1456         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1457         ret = system(str);
1458         if (ret == -1)
1459                 log_err("fio: exec of cmd <%s> failed\n", str);
1460
1461         free(str);
1462         return ret;
1463 }
1464
1465 /*
1466  * Dry run to compute correct state of numberio for verification.
1467  */
1468 static uint64_t do_dry_run(struct thread_data *td)
1469 {
1470         td_set_runstate(td, TD_RUNNING);
1471
1472         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1473                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1474                 struct io_u *io_u;
1475                 int ret;
1476
1477                 if (td->terminate || td->done)
1478                         break;
1479
1480                 io_u = get_io_u(td);
1481                 if (IS_ERR_OR_NULL(io_u))
1482                         break;
1483
1484                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1485                 io_u->error = 0;
1486                 io_u->resid = 0;
1487                 if (ddir_rw(acct_ddir(io_u)))
1488                         td->io_issues[acct_ddir(io_u)]++;
1489                 if (ddir_rw(io_u->ddir)) {
1490                         io_u_mark_depth(td, 1);
1491                         td->ts.total_io_u[io_u->ddir]++;
1492                 }
1493
1494                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1495                     td->o.do_verify &&
1496                     td->o.verify != VERIFY_NONE &&
1497                     !td->o.experimental_verify)
1498                         log_io_piece(td, io_u);
1499
1500                 ret = io_u_sync_complete(td, io_u);
1501                 (void) ret;
1502         }
1503
1504         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1505 }
1506
1507 struct fork_data {
1508         struct thread_data *td;
1509         struct sk_out *sk_out;
1510 };
1511
1512 /*
1513  * Entry point for the thread based jobs. The process based jobs end up
1514  * here as well, after a little setup.
1515  */
1516 static void *thread_main(void *data)
1517 {
1518         struct fork_data *fd = data;
1519         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1520         struct thread_data *td = fd->td;
1521         struct thread_options *o = &td->o;
1522         struct sk_out *sk_out = fd->sk_out;
1523         uint64_t bytes_done[DDIR_RWDIR_CNT];
1524         int deadlock_loop_cnt;
1525         bool clear_state, did_some_io;
1526         int ret;
1527
1528         sk_out_assign(sk_out);
1529         free(fd);
1530
1531         if (!o->use_thread) {
1532                 setsid();
1533                 td->pid = getpid();
1534         } else
1535                 td->pid = gettid();
1536
1537         fio_local_clock_init(o->use_thread);
1538
1539         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1540
1541         if (is_backend)
1542                 fio_server_send_start(td);
1543
1544         INIT_FLIST_HEAD(&td->io_log_list);
1545         INIT_FLIST_HEAD(&td->io_hist_list);
1546         INIT_FLIST_HEAD(&td->verify_list);
1547         INIT_FLIST_HEAD(&td->trim_list);
1548         INIT_FLIST_HEAD(&td->next_rand_list);
1549         td->io_hist_tree = RB_ROOT;
1550
1551         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1552         if (ret) {
1553                 td_verror(td, ret, "mutex_cond_init_pshared");
1554                 goto err;
1555         }
1556         ret = cond_init_pshared(&td->verify_cond);
1557         if (ret) {
1558                 td_verror(td, ret, "mutex_cond_pshared");
1559                 goto err;
1560         }
1561
1562         td_set_runstate(td, TD_INITIALIZED);
1563         dprint(FD_MUTEX, "up startup_mutex\n");
1564         fio_mutex_up(startup_mutex);
1565         dprint(FD_MUTEX, "wait on td->mutex\n");
1566         fio_mutex_down(td->mutex);
1567         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1568
1569         /*
1570          * A new gid requires privilege, so we need to do this before setting
1571          * the uid.
1572          */
1573         if (o->gid != -1U && setgid(o->gid)) {
1574                 td_verror(td, errno, "setgid");
1575                 goto err;
1576         }
1577         if (o->uid != -1U && setuid(o->uid)) {
1578                 td_verror(td, errno, "setuid");
1579                 goto err;
1580         }
1581
1582         /*
1583          * Do this early, we don't want the compress threads to be limited
1584          * to the same CPUs as the IO workers. So do this before we set
1585          * any potential CPU affinity
1586          */
1587         if (iolog_compress_init(td, sk_out))
1588                 goto err;
1589
1590         /*
1591          * If we have a gettimeofday() thread, make sure we exclude that
1592          * thread from this job
1593          */
1594         if (o->gtod_cpu)
1595                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1596
1597         /*
1598          * Set affinity first, in case it has an impact on the memory
1599          * allocations.
1600          */
1601         if (fio_option_is_set(o, cpumask)) {
1602                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1603                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1604                         if (!ret) {
1605                                 log_err("fio: no CPUs set\n");
1606                                 log_err("fio: Try increasing number of available CPUs\n");
1607                                 td_verror(td, EINVAL, "cpus_split");
1608                                 goto err;
1609                         }
1610                 }
1611                 ret = fio_setaffinity(td->pid, o->cpumask);
1612                 if (ret == -1) {
1613                         td_verror(td, errno, "cpu_set_affinity");
1614                         goto err;
1615                 }
1616         }
1617
1618 #ifdef CONFIG_LIBNUMA
1619         /* numa node setup */
1620         if (fio_option_is_set(o, numa_cpunodes) ||
1621             fio_option_is_set(o, numa_memnodes)) {
1622                 struct bitmask *mask;
1623
1624                 if (numa_available() < 0) {
1625                         td_verror(td, errno, "Does not support NUMA API\n");
1626                         goto err;
1627                 }
1628
1629                 if (fio_option_is_set(o, numa_cpunodes)) {
1630                         mask = numa_parse_nodestring(o->numa_cpunodes);
1631                         ret = numa_run_on_node_mask(mask);
1632                         numa_free_nodemask(mask);
1633                         if (ret == -1) {
1634                                 td_verror(td, errno, \
1635                                         "numa_run_on_node_mask failed\n");
1636                                 goto err;
1637                         }
1638                 }
1639
1640                 if (fio_option_is_set(o, numa_memnodes)) {
1641                         mask = NULL;
1642                         if (o->numa_memnodes)
1643                                 mask = numa_parse_nodestring(o->numa_memnodes);
1644
1645                         switch (o->numa_mem_mode) {
1646                         case MPOL_INTERLEAVE:
1647                                 numa_set_interleave_mask(mask);
1648                                 break;
1649                         case MPOL_BIND:
1650                                 numa_set_membind(mask);
1651                                 break;
1652                         case MPOL_LOCAL:
1653                                 numa_set_localalloc();
1654                                 break;
1655                         case MPOL_PREFERRED:
1656                                 numa_set_preferred(o->numa_mem_prefer_node);
1657                                 break;
1658                         case MPOL_DEFAULT:
1659                         default:
1660                                 break;
1661                         }
1662
1663                         if (mask)
1664                                 numa_free_nodemask(mask);
1665
1666                 }
1667         }
1668 #endif
1669
1670         if (fio_pin_memory(td))
1671                 goto err;
1672
1673         /*
1674          * May alter parameters that init_io_u() will use, so we need to
1675          * do this first.
1676          */
1677         if (init_iolog(td))
1678                 goto err;
1679
1680         if (init_io_u(td))
1681                 goto err;
1682
1683         if (o->verify_async && verify_async_init(td))
1684                 goto err;
1685
1686         if (fio_option_is_set(o, ioprio) ||
1687             fio_option_is_set(o, ioprio_class)) {
1688                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1689                 if (ret == -1) {
1690                         td_verror(td, errno, "ioprio_set");
1691                         goto err;
1692                 }
1693         }
1694
1695         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1696                 goto err;
1697
1698         errno = 0;
1699         if (nice(o->nice) == -1 && errno != 0) {
1700                 td_verror(td, errno, "nice");
1701                 goto err;
1702         }
1703
1704         if (o->ioscheduler && switch_ioscheduler(td))
1705                 goto err;
1706
1707         if (!o->create_serialize && setup_files(td))
1708                 goto err;
1709
1710         if (td_io_init(td))
1711                 goto err;
1712
1713         if (!init_random_map(td))
1714                 goto err;
1715
1716         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1717                 goto err;
1718
1719         if (o->pre_read && !pre_read_files(td))
1720                 goto err;
1721
1722         fio_verify_init(td);
1723
1724         if (rate_submit_init(td, sk_out))
1725                 goto err;
1726
1727         set_epoch_time(td, o->log_unix_epoch);
1728         fio_getrusage(&td->ru_start);
1729         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1730         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1731         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1732
1733         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1734                         o->ratemin[DDIR_TRIM]) {
1735                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1736                                         sizeof(td->bw_sample_time));
1737                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1738                                         sizeof(td->bw_sample_time));
1739                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1740                                         sizeof(td->bw_sample_time));
1741         }
1742
1743         memset(bytes_done, 0, sizeof(bytes_done));
1744         clear_state = false;
1745         did_some_io = false;
1746
1747         while (keep_running(td)) {
1748                 uint64_t verify_bytes;
1749
1750                 fio_gettime(&td->start, NULL);
1751                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1752
1753                 if (clear_state) {
1754                         clear_io_state(td, 0);
1755
1756                         if (o->unlink_each_loop && unlink_all_files(td))
1757                                 break;
1758                 }
1759
1760                 prune_io_piece_log(td);
1761
1762                 if (td->o.verify_only && td_write(td))
1763                         verify_bytes = do_dry_run(td);
1764                 else {
1765                         do_io(td, bytes_done);
1766
1767                         if (!ddir_rw_sum(bytes_done)) {
1768                                 fio_mark_td_terminate(td);
1769                                 verify_bytes = 0;
1770                         } else {
1771                                 verify_bytes = bytes_done[DDIR_WRITE] +
1772                                                 bytes_done[DDIR_TRIM];
1773                         }
1774                 }
1775
1776                 /*
1777                  * If we took too long to shut down, the main thread could
1778                  * already consider us reaped/exited. If that happens, break
1779                  * out and clean up.
1780                  */
1781                 if (td->runstate >= TD_EXITED)
1782                         break;
1783
1784                 clear_state = true;
1785
1786                 /*
1787                  * Make sure we've successfully updated the rusage stats
1788                  * before waiting on the stat mutex. Otherwise we could have
1789                  * the stat thread holding stat mutex and waiting for
1790                  * the rusage_sem, which would never get upped because
1791                  * this thread is waiting for the stat mutex.
1792                  */
1793                 deadlock_loop_cnt = 0;
1794                 do {
1795                         check_update_rusage(td);
1796                         if (!fio_mutex_down_trylock(stat_mutex))
1797                                 break;
1798                         usleep(1000);
1799                         if (deadlock_loop_cnt++ > 5000) {
1800                                 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1801                                 td->error = EDEADLK;
1802                                 goto err;
1803                         }
1804                 } while (1);
1805
1806                 if (td_read(td) && td->io_bytes[DDIR_READ])
1807                         update_runtime(td, elapsed_us, DDIR_READ);
1808                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1809                         update_runtime(td, elapsed_us, DDIR_WRITE);
1810                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1811                         update_runtime(td, elapsed_us, DDIR_TRIM);
1812                 fio_gettime(&td->start, NULL);
1813                 fio_mutex_up(stat_mutex);
1814
1815                 if (td->error || td->terminate)
1816                         break;
1817
1818                 if (!o->do_verify ||
1819                     o->verify == VERIFY_NONE ||
1820                     td_ioengine_flagged(td, FIO_UNIDIR))
1821                         continue;
1822
1823                 if (ddir_rw_sum(bytes_done))
1824                         did_some_io = true;
1825
1826                 clear_io_state(td, 0);
1827
1828                 fio_gettime(&td->start, NULL);
1829
1830                 do_verify(td, verify_bytes);
1831
1832                 /*
1833                  * See comment further up for why this is done here.
1834                  */
1835                 check_update_rusage(td);
1836
1837                 fio_mutex_down(stat_mutex);
1838                 update_runtime(td, elapsed_us, DDIR_READ);
1839                 fio_gettime(&td->start, NULL);
1840                 fio_mutex_up(stat_mutex);
1841
1842                 if (td->error || td->terminate)
1843                         break;
1844         }
1845
1846         /*
1847          * If td ended up with no I/O when it should have had,
1848          * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1849          * (Are we not missing other flags that can be ignored ?)
1850          */
1851         if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1852             !did_some_io && !td->o.create_only &&
1853             !(td_ioengine_flagged(td, FIO_NOIO) ||
1854               td_ioengine_flagged(td, FIO_DISKLESSIO)))
1855                 log_err("%s: No I/O performed by %s, "
1856                          "perhaps try --debug=io option for details?\n",
1857                          td->o.name, td->io_ops->name);
1858
1859         td_set_runstate(td, TD_FINISHING);
1860
1861         update_rusage_stat(td);
1862         td->ts.total_run_time = mtime_since_now(&td->epoch);
1863         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1864         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1865         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1866
1867         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1868             (td->o.verify != VERIFY_NONE && td_write(td)))
1869                 verify_save_state(td->thread_number);
1870
1871         fio_unpin_memory(td);
1872
1873         td_writeout_logs(td, true);
1874
1875         iolog_compress_exit(td);
1876         rate_submit_exit(td);
1877
1878         if (o->exec_postrun)
1879                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1880
1881         if (exitall_on_terminate || (o->exitall_error && td->error))
1882                 fio_terminate_threads(td->groupid);
1883
1884 err:
1885         if (td->error)
1886                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1887                                                         td->verror);
1888
1889         if (o->verify_async)
1890                 verify_async_exit(td);
1891
1892         close_and_free_files(td);
1893         cleanup_io_u(td);
1894         close_ioengine(td);
1895         cgroup_shutdown(td, &cgroup_mnt);
1896         verify_free_state(td);
1897
1898         if (td->zone_state_index) {
1899                 int i;
1900
1901                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1902                         free(td->zone_state_index[i]);
1903                 free(td->zone_state_index);
1904                 td->zone_state_index = NULL;
1905         }
1906
1907         if (fio_option_is_set(o, cpumask)) {
1908                 ret = fio_cpuset_exit(&o->cpumask);
1909                 if (ret)
1910                         td_verror(td, ret, "fio_cpuset_exit");
1911         }
1912
1913         /*
1914          * do this very late, it will log file closing as well
1915          */
1916         if (o->write_iolog_file)
1917                 write_iolog_close(td);
1918
1919         td_set_runstate(td, TD_EXITED);
1920
1921         /*
1922          * Do this last after setting our runstate to exited, so we
1923          * know that the stat thread is signaled.
1924          */
1925         check_update_rusage(td);
1926
1927         sk_out_drop();
1928         return (void *) (uintptr_t) td->error;
1929 }
1930
1931 /*
1932  * Run over the job map and reap the threads that have exited, if any.
1933  */
1934 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1935                          uint64_t *m_rate)
1936 {
1937         struct thread_data *td;
1938         unsigned int cputhreads, realthreads, pending;
1939         int i, status, ret;
1940
1941         /*
1942          * reap exited threads (TD_EXITED -> TD_REAPED)
1943          */
1944         realthreads = pending = cputhreads = 0;
1945         for_each_td(td, i) {
1946                 int flags = 0;
1947
1948                  if (!strcmp(td->o.ioengine, "cpuio"))
1949                         cputhreads++;
1950                 else
1951                         realthreads++;
1952
1953                 if (!td->pid) {
1954                         pending++;
1955                         continue;
1956                 }
1957                 if (td->runstate == TD_REAPED)
1958                         continue;
1959                 if (td->o.use_thread) {
1960                         if (td->runstate == TD_EXITED) {
1961                                 td_set_runstate(td, TD_REAPED);
1962                                 goto reaped;
1963                         }
1964                         continue;
1965                 }
1966
1967                 flags = WNOHANG;
1968                 if (td->runstate == TD_EXITED)
1969                         flags = 0;
1970
1971                 /*
1972                  * check if someone quit or got killed in an unusual way
1973                  */
1974                 ret = waitpid(td->pid, &status, flags);
1975                 if (ret < 0) {
1976                         if (errno == ECHILD) {
1977                                 log_err("fio: pid=%d disappeared %d\n",
1978                                                 (int) td->pid, td->runstate);
1979                                 td->sig = ECHILD;
1980                                 td_set_runstate(td, TD_REAPED);
1981                                 goto reaped;
1982                         }
1983                         perror("waitpid");
1984                 } else if (ret == td->pid) {
1985                         if (WIFSIGNALED(status)) {
1986                                 int sig = WTERMSIG(status);
1987
1988                                 if (sig != SIGTERM && sig != SIGUSR2)
1989                                         log_err("fio: pid=%d, got signal=%d\n",
1990                                                         (int) td->pid, sig);
1991                                 td->sig = sig;
1992                                 td_set_runstate(td, TD_REAPED);
1993                                 goto reaped;
1994                         }
1995                         if (WIFEXITED(status)) {
1996                                 if (WEXITSTATUS(status) && !td->error)
1997                                         td->error = WEXITSTATUS(status);
1998
1999                                 td_set_runstate(td, TD_REAPED);
2000                                 goto reaped;
2001                         }
2002                 }
2003
2004                 /*
2005                  * If the job is stuck, do a forceful timeout of it and
2006                  * move on.
2007                  */
2008                 if (td->terminate &&
2009                     td->runstate < TD_FSYNCING &&
2010                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2011                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2012                                 "%lu seconds, it appears to be stuck. Doing "
2013                                 "forceful exit of this job.\n",
2014                                 td->o.name, td->runstate,
2015                                 (unsigned long) time_since_now(&td->terminate_time));
2016                         td_set_runstate(td, TD_REAPED);
2017                         goto reaped;
2018                 }
2019
2020                 /*
2021                  * thread is not dead, continue
2022                  */
2023                 pending++;
2024                 continue;
2025 reaped:
2026                 (*nr_running)--;
2027                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2028                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2029                 if (!td->pid)
2030                         pending--;
2031
2032                 if (td->error)
2033                         exit_value++;
2034
2035                 done_secs += mtime_since_now(&td->epoch) / 1000;
2036                 profile_td_exit(td);
2037         }
2038
2039         if (*nr_running == cputhreads && !pending && realthreads)
2040                 fio_terminate_threads(TERMINATE_ALL);
2041 }
2042
2043 static bool __check_trigger_file(void)
2044 {
2045         struct stat sb;
2046
2047         if (!trigger_file)
2048                 return false;
2049
2050         if (stat(trigger_file, &sb))
2051                 return false;
2052
2053         if (unlink(trigger_file) < 0)
2054                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2055                                                         strerror(errno));
2056
2057         return true;
2058 }
2059
2060 static bool trigger_timedout(void)
2061 {
2062         if (trigger_timeout)
2063                 if (time_since_genesis() >= trigger_timeout) {
2064                         trigger_timeout = 0;
2065                         return true;
2066                 }
2067
2068         return false;
2069 }
2070
2071 void exec_trigger(const char *cmd)
2072 {
2073         int ret;
2074
2075         if (!cmd || cmd[0] == '\0')
2076                 return;
2077
2078         ret = system(cmd);
2079         if (ret == -1)
2080                 log_err("fio: failed executing %s trigger\n", cmd);
2081 }
2082
2083 void check_trigger_file(void)
2084 {
2085         if (__check_trigger_file() || trigger_timedout()) {
2086                 if (nr_clients)
2087                         fio_clients_send_trigger(trigger_remote_cmd);
2088                 else {
2089                         verify_save_state(IO_LIST_ALL);
2090                         fio_terminate_threads(TERMINATE_ALL);
2091                         exec_trigger(trigger_cmd);
2092                 }
2093         }
2094 }
2095
2096 static int fio_verify_load_state(struct thread_data *td)
2097 {
2098         int ret;
2099
2100         if (!td->o.verify_state)
2101                 return 0;
2102
2103         if (is_backend) {
2104                 void *data;
2105
2106                 ret = fio_server_get_verify_state(td->o.name,
2107                                         td->thread_number - 1, &data);
2108                 if (!ret)
2109                         verify_assign_state(td, data);
2110         } else
2111                 ret = verify_load_state(td, "local");
2112
2113         return ret;
2114 }
2115
2116 static void do_usleep(unsigned int usecs)
2117 {
2118         check_for_running_stats();
2119         check_trigger_file();
2120         usleep(usecs);
2121 }
2122
2123 static bool check_mount_writes(struct thread_data *td)
2124 {
2125         struct fio_file *f;
2126         unsigned int i;
2127
2128         if (!td_write(td) || td->o.allow_mounted_write)
2129                 return false;
2130
2131         /*
2132          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2133          * are mkfs'd and mounted.
2134          */
2135         for_each_file(td, f, i) {
2136 #ifdef FIO_HAVE_CHARDEV_SIZE
2137                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2138 #else
2139                 if (f->filetype != FIO_TYPE_BLOCK)
2140 #endif
2141                         continue;
2142                 if (device_is_mounted(f->file_name))
2143                         goto mounted;
2144         }
2145
2146         return false;
2147 mounted:
2148         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2149         return true;
2150 }
2151
2152 static bool waitee_running(struct thread_data *me)
2153 {
2154         const char *waitee = me->o.wait_for;
2155         const char *self = me->o.name;
2156         struct thread_data *td;
2157         int i;
2158
2159         if (!waitee)
2160                 return false;
2161
2162         for_each_td(td, i) {
2163                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2164                         continue;
2165
2166                 if (td->runstate < TD_EXITED) {
2167                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2168                                         self, td->o.name,
2169                                         runstate_to_name(td->runstate));
2170                         return true;
2171                 }
2172         }
2173
2174         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2175         return false;
2176 }
2177
2178 /*
2179  * Main function for kicking off and reaping jobs, as needed.
2180  */
2181 static void run_threads(struct sk_out *sk_out)
2182 {
2183         struct thread_data *td;
2184         unsigned int i, todo, nr_running, nr_started;
2185         uint64_t m_rate, t_rate;
2186         uint64_t spent;
2187
2188         if (fio_gtod_offload && fio_start_gtod_thread())
2189                 return;
2190
2191         fio_idle_prof_init();
2192
2193         set_sig_handlers();
2194
2195         nr_thread = nr_process = 0;
2196         for_each_td(td, i) {
2197                 if (check_mount_writes(td))
2198                         return;
2199                 if (td->o.use_thread)
2200                         nr_thread++;
2201                 else
2202                         nr_process++;
2203         }
2204
2205         if (output_format & FIO_OUTPUT_NORMAL) {
2206                 log_info("Starting ");
2207                 if (nr_thread)
2208                         log_info("%d thread%s", nr_thread,
2209                                                 nr_thread > 1 ? "s" : "");
2210                 if (nr_process) {
2211                         if (nr_thread)
2212                                 log_info(" and ");
2213                         log_info("%d process%s", nr_process,
2214                                                 nr_process > 1 ? "es" : "");
2215                 }
2216                 log_info("\n");
2217                 log_info_flush();
2218         }
2219
2220         todo = thread_number;
2221         nr_running = 0;
2222         nr_started = 0;
2223         m_rate = t_rate = 0;
2224
2225         for_each_td(td, i) {
2226                 print_status_init(td->thread_number - 1);
2227
2228                 if (!td->o.create_serialize)
2229                         continue;
2230
2231                 if (fio_verify_load_state(td))
2232                         goto reap;
2233
2234                 /*
2235                  * do file setup here so it happens sequentially,
2236                  * we don't want X number of threads getting their
2237                  * client data interspersed on disk
2238                  */
2239                 if (setup_files(td)) {
2240 reap:
2241                         exit_value++;
2242                         if (td->error)
2243                                 log_err("fio: pid=%d, err=%d/%s\n",
2244                                         (int) td->pid, td->error, td->verror);
2245                         td_set_runstate(td, TD_REAPED);
2246                         todo--;
2247                 } else {
2248                         struct fio_file *f;
2249                         unsigned int j;
2250
2251                         /*
2252                          * for sharing to work, each job must always open
2253                          * its own files. so close them, if we opened them
2254                          * for creation
2255                          */
2256                         for_each_file(td, f, j) {
2257                                 if (fio_file_open(f))
2258                                         td_io_close_file(td, f);
2259                         }
2260                 }
2261         }
2262
2263         /* start idle threads before io threads start to run */
2264         fio_idle_prof_start();
2265
2266         set_genesis_time();
2267
2268         while (todo) {
2269                 struct thread_data *map[REAL_MAX_JOBS];
2270                 struct timespec this_start;
2271                 int this_jobs = 0, left;
2272                 struct fork_data *fd;
2273
2274                 /*
2275                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2276                  */
2277                 for_each_td(td, i) {
2278                         if (td->runstate != TD_NOT_CREATED)
2279                                 continue;
2280
2281                         /*
2282                          * never got a chance to start, killed by other
2283                          * thread for some reason
2284                          */
2285                         if (td->terminate) {
2286                                 todo--;
2287                                 continue;
2288                         }
2289
2290                         if (td->o.start_delay) {
2291                                 spent = utime_since_genesis();
2292
2293                                 if (td->o.start_delay > spent)
2294                                         continue;
2295                         }
2296
2297                         if (td->o.stonewall && (nr_started || nr_running)) {
2298                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2299                                                         td->o.name);
2300                                 break;
2301                         }
2302
2303                         if (waitee_running(td)) {
2304                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2305                                                 td->o.name, td->o.wait_for);
2306                                 continue;
2307                         }
2308
2309                         init_disk_util(td);
2310
2311                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2312                         td->update_rusage = 0;
2313
2314                         /*
2315                          * Set state to created. Thread will transition
2316                          * to TD_INITIALIZED when it's done setting up.
2317                          */
2318                         td_set_runstate(td, TD_CREATED);
2319                         map[this_jobs++] = td;
2320                         nr_started++;
2321
2322                         fd = calloc(1, sizeof(*fd));
2323                         fd->td = td;
2324                         fd->sk_out = sk_out;
2325
2326                         if (td->o.use_thread) {
2327                                 int ret;
2328
2329                                 dprint(FD_PROCESS, "will pthread_create\n");
2330                                 ret = pthread_create(&td->thread, NULL,
2331                                                         thread_main, fd);
2332                                 if (ret) {
2333                                         log_err("pthread_create: %s\n",
2334                                                         strerror(ret));
2335                                         free(fd);
2336                                         nr_started--;
2337                                         break;
2338                                 }
2339                                 fd = NULL;
2340                                 ret = pthread_detach(td->thread);
2341                                 if (ret)
2342                                         log_err("pthread_detach: %s",
2343                                                         strerror(ret));
2344                         } else {
2345                                 pid_t pid;
2346                                 dprint(FD_PROCESS, "will fork\n");
2347                                 pid = fork();
2348                                 if (!pid) {
2349                                         int ret;
2350
2351                                         ret = (int)(uintptr_t)thread_main(fd);
2352                                         _exit(ret);
2353                                 } else if (i == fio_debug_jobno)
2354                                         *fio_debug_jobp = pid;
2355                         }
2356                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2357                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2358                                 log_err("fio: job startup hung? exiting.\n");
2359                                 fio_terminate_threads(TERMINATE_ALL);
2360                                 fio_abort = 1;
2361                                 nr_started--;
2362                                 free(fd);
2363                                 break;
2364                         }
2365                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2366                 }
2367
2368                 /*
2369                  * Wait for the started threads to transition to
2370                  * TD_INITIALIZED.
2371                  */
2372                 fio_gettime(&this_start, NULL);
2373                 left = this_jobs;
2374                 while (left && !fio_abort) {
2375                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2376                                 break;
2377
2378                         do_usleep(100000);
2379
2380                         for (i = 0; i < this_jobs; i++) {
2381                                 td = map[i];
2382                                 if (!td)
2383                                         continue;
2384                                 if (td->runstate == TD_INITIALIZED) {
2385                                         map[i] = NULL;
2386                                         left--;
2387                                 } else if (td->runstate >= TD_EXITED) {
2388                                         map[i] = NULL;
2389                                         left--;
2390                                         todo--;
2391                                         nr_running++; /* work-around... */
2392                                 }
2393                         }
2394                 }
2395
2396                 if (left) {
2397                         log_err("fio: %d job%s failed to start\n", left,
2398                                         left > 1 ? "s" : "");
2399                         for (i = 0; i < this_jobs; i++) {
2400                                 td = map[i];
2401                                 if (!td)
2402                                         continue;
2403                                 kill(td->pid, SIGTERM);
2404                         }
2405                         break;
2406                 }
2407
2408                 /*
2409                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2410                  */
2411                 for_each_td(td, i) {
2412                         if (td->runstate != TD_INITIALIZED)
2413                                 continue;
2414
2415                         if (in_ramp_time(td))
2416                                 td_set_runstate(td, TD_RAMP);
2417                         else
2418                                 td_set_runstate(td, TD_RUNNING);
2419                         nr_running++;
2420                         nr_started--;
2421                         m_rate += ddir_rw_sum(td->o.ratemin);
2422                         t_rate += ddir_rw_sum(td->o.rate);
2423                         todo--;
2424                         fio_mutex_up(td->mutex);
2425                 }
2426
2427                 reap_threads(&nr_running, &t_rate, &m_rate);
2428
2429                 if (todo)
2430                         do_usleep(100000);
2431         }
2432
2433         while (nr_running) {
2434                 reap_threads(&nr_running, &t_rate, &m_rate);
2435                 do_usleep(10000);
2436         }
2437
2438         fio_idle_prof_stop();
2439
2440         update_io_ticks();
2441 }
2442
2443 static void free_disk_util(void)
2444 {
2445         disk_util_prune_entries();
2446         helper_thread_destroy();
2447 }
2448
2449 int fio_backend(struct sk_out *sk_out)
2450 {
2451         struct thread_data *td;
2452         int i;
2453
2454         if (exec_profile) {
2455                 if (load_profile(exec_profile))
2456                         return 1;
2457                 free(exec_profile);
2458                 exec_profile = NULL;
2459         }
2460         if (!thread_number)
2461                 return 0;
2462
2463         if (write_bw_log) {
2464                 struct log_params p = {
2465                         .log_type = IO_LOG_TYPE_BW,
2466                 };
2467
2468                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2469                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2470                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2471         }
2472
2473         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2474         if (startup_mutex == NULL)
2475                 return 1;
2476
2477         set_genesis_time();
2478         stat_init();
2479         helper_thread_create(startup_mutex, sk_out);
2480
2481         cgroup_list = smalloc(sizeof(*cgroup_list));
2482         INIT_FLIST_HEAD(cgroup_list);
2483
2484         run_threads(sk_out);
2485
2486         helper_thread_exit();
2487
2488         if (!fio_abort) {
2489                 __show_run_stats();
2490                 if (write_bw_log) {
2491                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2492                                 struct io_log *log = agg_io_log[i];
2493
2494                                 flush_log(log, false);
2495                                 free_log(log);
2496                         }
2497                 }
2498         }
2499
2500         for_each_td(td, i) {
2501                 steadystate_free(td);
2502                 fio_options_free(td);
2503                 if (td->rusage_sem) {
2504                         fio_mutex_remove(td->rusage_sem);
2505                         td->rusage_sem = NULL;
2506                 }
2507                 fio_mutex_remove(td->mutex);
2508                 td->mutex = NULL;
2509         }
2510
2511         free_disk_util();
2512         cgroup_kill(cgroup_list);
2513         sfree(cgroup_list);
2514         sfree(cgroup_mnt);
2515
2516         fio_mutex_remove(startup_mutex);
2517         stat_exit();
2518         return exit_value;
2519 }