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 <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32 #include <pthread.h>
  33
  34 #include "fio.h"
  35 #include "smalloc.h"
  36 #include "verify.h"
  37 #include "diskutil.h"
  38 #include "cgroup.h"
  39 #include "profile.h"
  40 #include "lib/rand.h"
  41 #include "lib/memalign.h"
  42 #include "server.h"
  43 #include "lib/getrusage.h"
  44 #include "idletime.h"
  45 #include "err.h"
  46 #include "workqueue.h"
  47 #include "lib/mountcheck.h"
  48 #include "rate-submit.h"
  49 #include "helper_thread.h"
  50 #include "pshared.h"
  51 #include "zone-dist.h"
  52
  53 static struct fio_sem *startup_sem;
  54 static struct flist_head *cgroup_list;
  55 static struct cgroup_mnt *cgroup_mnt;
  56 static int exit_value;
  57 static volatile bool fio_abort;
  58 static unsigned int nr_process = 0;
  59 static unsigned int nr_thread = 0;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int nr_segments = 0;
  66 unsigned int cur_segment = 0;
  67 unsigned int stat_number = 0;
  68 int temp_stall_ts;
  69 unsigned long done_secs = 0;
  70 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
  71 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
  72 #else
  73 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  74 #endif
  75
  76 #define JOB_START_TIMEOUT       (5 * 1000)
  77
  78 static void sig_int(int sig)
  79 {
  80         if (nr_segments) {
  81                 if (is_backend)
  82                         fio_server_got_signal(sig);
  83                 else {
  84                         log_info("\nfio: terminating on signal %d\n", sig);
  85                         log_info_flush();
  86                         exit_value = 128;
  87                 }
  88
  89                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  90         }
  91 }
  92
  93 void sig_show_status(int sig)
  94 {
  95         show_running_run_stats();
  96 }
  97
  98 static void set_sig_handlers(void)
  99 {
 100         struct sigaction act;
 101
 102         memset(&act, 0, sizeof(act));
 103         act.sa_handler = sig_int;
 104         act.sa_flags = SA_RESTART;
 105         sigaction(SIGINT, &act, NULL);
 106
 107         memset(&act, 0, sizeof(act));
 108         act.sa_handler = sig_int;
 109         act.sa_flags = SA_RESTART;
 110         sigaction(SIGTERM, &act, NULL);
 111
 112 /* Windows uses SIGBREAK as a quit signal from other applications */
 113 #ifdef WIN32
 114         memset(&act, 0, sizeof(act));
 115         act.sa_handler = sig_int;
 116         act.sa_flags = SA_RESTART;
 117         sigaction(SIGBREAK, &act, NULL);
 118 #endif
 119
 120         memset(&act, 0, sizeof(act));
 121         act.sa_handler = sig_show_status;
 122         act.sa_flags = SA_RESTART;
 123         sigaction(SIGUSR1, &act, NULL);
 124
 125         if (is_backend) {
 126                 memset(&act, 0, sizeof(act));
 127                 act.sa_handler = sig_int;
 128                 act.sa_flags = SA_RESTART;
 129                 sigaction(SIGPIPE, &act, NULL);
 130         }
 131 }
 132
 133 /*
 134  * Check if we are above the minimum rate given.
 135  */
 136 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 137                              enum fio_ddir ddir)
 138 {
 139         unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
 140         unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
 141         unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
 142         unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
 143
 144         assert(ddir_rw(ddir));
 145
 146         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 147                 return false;
 148
 149         /*
 150          * allow a 2 second settle period in the beginning
 151          */
 152         if (mtime_since(&td->start, now) < 2000)
 153                 return false;
 154
 155         /*
 156          * if last_rate_check_blocks or last_rate_check_bytes is set,
 157          * we can compute a rate per ratecycle
 158          */
 159         if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
 160                 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
 161                 if (spent < td->o.ratecycle || spent==0)
 162                         return false;
 163
 164                 if (td->o.ratemin[ddir]) {
 165                         /*
 166                          * check bandwidth specified rate
 167                          */
 168                         unsigned long long current_rate_bytes =
 169                                 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
 170                         if (current_rate_bytes < option_rate_bytes_min) {
 171                                 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 172                                         td->o.name, option_rate_bytes_min, current_rate_bytes);
 173                                 return true;
 174                         }
 175                 } else {
 176                         /*
 177                          * checks iops specified rate
 178                          */
 179                         unsigned long long current_rate_iops =
 180                                 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
 181
 182                         if (current_rate_iops < option_rate_iops_min) {
 183                                 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 184                                         td->o.name, option_rate_iops_min, current_rate_iops);
 185                                 return true;
 186                         }
 187                 }
 188         }
 189
 190         td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
 191         td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
 192         memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
 193         return false;
 194 }
 195
 196 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 197 {
 198         bool ret = false;
 199
 200         for_each_rw_ddir(ddir) {
 201                 if (td->bytes_done[ddir])
 202                         ret |= __check_min_rate(td, now, ddir);
 203         }
 204
 205         return ret;
 206 }
 207
 208 /*
 209  * When job exits, we can cancel the in-flight IO if we are using async
 210  * io. Attempt to do so.
 211  */
 212 static void cleanup_pending_aio(struct thread_data *td)
 213 {
 214         int r;
 215
 216         /*
 217          * get immediately available events, if any
 218          */
 219         r = io_u_queued_complete(td, 0);
 220
 221         /*
 222          * now cancel remaining active events
 223          */
 224         if (td->io_ops->cancel) {
 225                 struct io_u *io_u;
 226                 int i;
 227
 228                 io_u_qiter(&td->io_u_all, io_u, i) {
 229                         if (io_u->flags & IO_U_F_FLIGHT) {
 230                                 r = td->io_ops->cancel(td, io_u);
 231                                 if (!r)
 232                                         put_io_u(td, io_u);
 233                         }
 234                 }
 235         }
 236
 237         if (td->cur_depth)
 238                 r = io_u_queued_complete(td, td->cur_depth);
 239 }
 240
 241 /*
 242  * Helper to handle the final sync of a file. Works just like the normal
 243  * io path, just does everything sync.
 244  */
 245 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 246 {
 247         struct io_u *io_u = __get_io_u(td);
 248         enum fio_q_status ret;
 249
 250         if (!io_u)
 251                 return true;
 252
 253         io_u->ddir = DDIR_SYNC;
 254         io_u->file = f;
 255         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 256
 257         if (td_io_prep(td, io_u)) {
 258                 put_io_u(td, io_u);
 259                 return true;
 260         }
 261
 262 requeue:
 263         ret = td_io_queue(td, io_u);
 264         switch (ret) {
 265         case FIO_Q_QUEUED:
 266                 td_io_commit(td);
 267                 if (io_u_queued_complete(td, 1) < 0)
 268                         return true;
 269                 break;
 270         case FIO_Q_COMPLETED:
 271                 if (io_u->error) {
 272                         td_verror(td, io_u->error, "td_io_queue");
 273                         return true;
 274                 }
 275
 276                 if (io_u_sync_complete(td, io_u) < 0)
 277                         return true;
 278                 break;
 279         case FIO_Q_BUSY:
 280                 td_io_commit(td);
 281                 goto requeue;
 282         }
 283
 284         return false;
 285 }
 286
 287 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 288 {
 289         int ret, ret2;
 290
 291         if (fio_file_open(f))
 292                 return fio_io_sync(td, f);
 293
 294         if (td_io_open_file(td, f))
 295                 return 1;
 296
 297         ret = fio_io_sync(td, f);
 298         ret2 = 0;
 299         if (fio_file_open(f))
 300                 ret2 = td_io_close_file(td, f);
 301         return (ret || ret2);
 302 }
 303
 304 static inline void __update_ts_cache(struct thread_data *td)
 305 {
 306         fio_gettime(&td->ts_cache, NULL);
 307 }
 308
 309 static inline void update_ts_cache(struct thread_data *td)
 310 {
 311         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 312                 __update_ts_cache(td);
 313 }
 314
 315 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 316 {
 317         if (in_ramp_time(td))
 318                 return false;
 319         if (!td->o.timeout)
 320                 return false;
 321         if (utime_since(&td->epoch, t) >= td->o.timeout)
 322                 return true;
 323
 324         return false;
 325 }
 326
 327 /*
 328  * We need to update the runtime consistently in ms, but keep a running
 329  * tally of the current elapsed time in microseconds for sub millisecond
 330  * updates.
 331  */
 332 static inline void update_runtime(struct thread_data *td,
 333                                   unsigned long long *elapsed_us,
 334                                   const enum fio_ddir ddir)
 335 {
 336         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 337                 return;
 338
 339         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 340         elapsed_us[ddir] += utime_since_now(&td->start);
 341         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 342 }
 343
 344 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 345                                 int *retptr)
 346 {
 347         int ret = *retptr;
 348
 349         if (ret < 0 || td->error) {
 350                 int err = td->error;
 351                 enum error_type_bit eb;
 352
 353                 if (ret < 0)
 354                         err = -ret;
 355
 356                 eb = td_error_type(ddir, err);
 357                 if (!(td->o.continue_on_error & (1 << eb)))
 358                         return true;
 359
 360                 if (td_non_fatal_error(td, eb, err)) {
 361                         /*
 362                          * Continue with the I/Os in case of
 363                          * a non fatal error.
 364                          */
 365                         update_error_count(td, err);
 366                         td_clear_error(td);
 367                         *retptr = 0;
 368                         return false;
 369                 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
 370                         /*
 371                          * We expect to hit this error if
 372                          * fill_device option is set.
 373                          */
 374                         td_clear_error(td);
 375                         fio_mark_td_terminate(td);
 376                         return true;
 377                 } else {
 378                         /*
 379                          * Stop the I/O in case of a fatal
 380                          * error.
 381                          */
 382                         update_error_count(td, err);
 383                         return true;
 384                 }
 385         }
 386
 387         return false;
 388 }
 389
 390 static void check_update_rusage(struct thread_data *td)
 391 {
 392         if (td->update_rusage) {
 393                 td->update_rusage = 0;
 394                 update_rusage_stat(td);
 395                 fio_sem_up(td->rusage_sem);
 396         }
 397 }
 398
 399 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 400 {
 401         const int full = queue_full(td);
 402         int min_evts = 0;
 403         int ret;
 404
 405         if (td->flags & TD_F_REGROW_LOGS)
 406                 return io_u_quiesce(td);
 407
 408         /*
 409          * if the queue is full, we MUST reap at least 1 event
 410          */
 411         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 412         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 413                 min_evts = 1;
 414
 415         if (time && should_check_rate(td))
 416                 fio_gettime(time, NULL);
 417
 418         do {
 419                 ret = io_u_queued_complete(td, min_evts);
 420                 if (ret < 0)
 421                         break;
 422         } while (full && (td->cur_depth > td->o.iodepth_low));
 423
 424         return ret;
 425 }
 426
 427 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 428                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 429                    struct timespec *comp_time)
 430 {
 431         switch (*ret) {
 432         case FIO_Q_COMPLETED:
 433                 if (io_u->error) {
 434                         *ret = -io_u->error;
 435                         clear_io_u(td, io_u);
 436                 } else if (io_u->resid) {
 437                         long long bytes = io_u->xfer_buflen - io_u->resid;
 438                         struct fio_file *f = io_u->file;
 439
 440                         if (bytes_issued)
 441                                 *bytes_issued += bytes;
 442
 443                         if (!from_verify)
 444                                 trim_io_piece(io_u);
 445
 446                         /*
 447                          * zero read, fail
 448                          */
 449                         if (!bytes) {
 450                                 if (!from_verify)
 451                                         unlog_io_piece(td, io_u);
 452                                 td_verror(td, EIO, "full resid");
 453                                 put_io_u(td, io_u);
 454                                 break;
 455                         }
 456
 457                         io_u->xfer_buflen = io_u->resid;
 458                         io_u->xfer_buf += bytes;
 459                         io_u->offset += bytes;
 460
 461                         if (ddir_rw(io_u->ddir))
 462                                 td->ts.short_io_u[io_u->ddir]++;
 463
 464                         if (io_u->offset == f->real_file_size)
 465                                 goto sync_done;
 466
 467                         requeue_io_u(td, &io_u);
 468                 } else {
 469 sync_done:
 470                         if (comp_time && should_check_rate(td))
 471                                 fio_gettime(comp_time, NULL);
 472
 473                         *ret = io_u_sync_complete(td, io_u);
 474                         if (*ret < 0)
 475                                 break;
 476                 }
 477
 478                 if (td->flags & TD_F_REGROW_LOGS)
 479                         regrow_logs(td);
 480
 481                 /*
 482                  * when doing I/O (not when verifying),
 483                  * check for any errors that are to be ignored
 484                  */
 485                 if (!from_verify)
 486                         break;
 487
 488                 return 0;
 489         case FIO_Q_QUEUED:
 490                 /*
 491                  * if the engine doesn't have a commit hook,
 492                  * the io_u is really queued. if it does have such
 493                  * a hook, it has to call io_u_queued() itself.
 494                  */
 495                 if (td->io_ops->commit == NULL)
 496                         io_u_queued(td, io_u);
 497                 if (bytes_issued)
 498                         *bytes_issued += io_u->xfer_buflen;
 499                 break;
 500         case FIO_Q_BUSY:
 501                 if (!from_verify)
 502                         unlog_io_piece(td, io_u);
 503                 requeue_io_u(td, &io_u);
 504                 td_io_commit(td);
 505                 break;
 506         default:
 507                 assert(*ret < 0);
 508                 td_verror(td, -(*ret), "td_io_queue");
 509                 break;
 510         }
 511
 512         if (break_on_this_error(td, ddir, ret))
 513                 return 1;
 514
 515         return 0;
 516 }
 517
 518 static inline bool io_in_polling(struct thread_data *td)
 519 {
 520         return !td->o.iodepth_batch_complete_min &&
 521                    !td->o.iodepth_batch_complete_max;
 522 }
 523 /*
 524  * Unlinks files from thread data fio_file structure
 525  */
 526 static int unlink_all_files(struct thread_data *td)
 527 {
 528         struct fio_file *f;
 529         unsigned int i;
 530         int ret = 0;
 531
 532         for_each_file(td, f, i) {
 533                 if (f->filetype != FIO_TYPE_FILE)
 534                         continue;
 535                 ret = td_io_unlink_file(td, f);
 536                 if (ret)
 537                         break;
 538         }
 539
 540         if (ret)
 541                 td_verror(td, ret, "unlink_all_files");
 542
 543         return ret;
 544 }
 545
 546 /*
 547  * Check if io_u will overlap an in-flight IO in the queue
 548  */
 549 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 550 {
 551         bool overlap;
 552         struct io_u *check_io_u;
 553         unsigned long long x1, x2, y1, y2;
 554         int i;
 555
 556         x1 = io_u->offset;
 557         x2 = io_u->offset + io_u->buflen;
 558         overlap = false;
 559         io_u_qiter(q, check_io_u, i) {
 560                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 561                         y1 = check_io_u->offset;
 562                         y2 = check_io_u->offset + check_io_u->buflen;
 563
 564                         if (x1 < y2 && y1 < x2) {
 565                                 overlap = true;
 566                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 567                                                 x1, io_u->buflen,
 568                                                 y1, check_io_u->buflen);
 569                                 break;
 570                         }
 571                 }
 572         }
 573
 574         return overlap;
 575 }
 576
 577 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 578 {
 579         /*
 580          * Check for overlap if the user asked us to, and we have
 581          * at least one IO in flight besides this one.
 582          */
 583         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 584             in_flight_overlap(&td->io_u_all, io_u))
 585                 return FIO_Q_BUSY;
 586
 587         return td_io_queue(td, io_u);
 588 }
 589
 590 /*
 591  * The main verify engine. Runs over the writes we previously submitted,
 592  * reads the blocks back in, and checks the crc/md5 of the data.
 593  */
 594 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 595 {
 596         struct fio_file *f;
 597         struct io_u *io_u;
 598         int ret, min_events;
 599         unsigned int i;
 600
 601         dprint(FD_VERIFY, "starting loop\n");
 602
 603         /*
 604          * sync io first and invalidate cache, to make sure we really
 605          * read from disk.
 606          */
 607         for_each_file(td, f, i) {
 608                 if (!fio_file_open(f))
 609                         continue;
 610                 if (fio_io_sync(td, f))
 611                         break;
 612                 if (file_invalidate_cache(td, f))
 613                         break;
 614         }
 615
 616         check_update_rusage(td);
 617
 618         if (td->error)
 619                 return;
 620
 621         /*
 622          * verify_state needs to be reset before verification
 623          * proceeds so that expected random seeds match actual
 624          * random seeds in headers. The main loop will reset
 625          * all random number generators if randrepeat is set.
 626          */
 627         if (!td->o.rand_repeatable)
 628                 td_fill_verify_state_seed(td);
 629
 630         td_set_runstate(td, TD_VERIFYING);
 631
 632         io_u = NULL;
 633         while (!td->terminate) {
 634                 enum fio_ddir ddir;
 635                 int full;
 636
 637                 update_ts_cache(td);
 638                 check_update_rusage(td);
 639
 640                 if (runtime_exceeded(td, &td->ts_cache)) {
 641                         __update_ts_cache(td);
 642                         if (runtime_exceeded(td, &td->ts_cache)) {
 643                                 fio_mark_td_terminate(td);
 644                                 break;
 645                         }
 646                 }
 647
 648                 if (flow_threshold_exceeded(td))
 649                         continue;
 650
 651                 if (!td->o.experimental_verify) {
 652                         io_u = __get_io_u(td);
 653                         if (!io_u)
 654                                 break;
 655
 656                         if (get_next_verify(td, io_u)) {
 657                                 put_io_u(td, io_u);
 658                                 break;
 659                         }
 660
 661                         if (td_io_prep(td, io_u)) {
 662                                 put_io_u(td, io_u);
 663                                 break;
 664                         }
 665                 } else {
 666                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 667                                 break;
 668
 669                         while ((io_u = get_io_u(td)) != NULL) {
 670                                 if (IS_ERR_OR_NULL(io_u)) {
 671                                         io_u = NULL;
 672                                         ret = FIO_Q_BUSY;
 673                                         goto reap;
 674                                 }
 675
 676                                 /*
 677                                  * We are only interested in the places where
 678                                  * we wrote or trimmed IOs. Turn those into
 679                                  * reads for verification purposes.
 680                                  */
 681                                 if (io_u->ddir == DDIR_READ) {
 682                                         /*
 683                                          * Pretend we issued it for rwmix
 684                                          * accounting
 685                                          */
 686                                         td->io_issues[DDIR_READ]++;
 687                                         put_io_u(td, io_u);
 688                                         continue;
 689                                 } else if (io_u->ddir == DDIR_TRIM) {
 690                                         io_u->ddir = DDIR_READ;
 691                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 692                                         break;
 693                                 } else if (io_u->ddir == DDIR_WRITE) {
 694                                         io_u->ddir = DDIR_READ;
 695                                         populate_verify_io_u(td, io_u);
 696                                         break;
 697                                 } else {
 698                                         put_io_u(td, io_u);
 699                                         continue;
 700                                 }
 701                         }
 702
 703                         if (!io_u)
 704                                 break;
 705                 }
 706
 707                 if (verify_state_should_stop(td, io_u)) {
 708                         put_io_u(td, io_u);
 709                         break;
 710                 }
 711
 712                 if (td->o.verify_async)
 713                         io_u->end_io = verify_io_u_async;
 714                 else
 715                         io_u->end_io = verify_io_u;
 716
 717                 ddir = io_u->ddir;
 718                 if (!td->o.disable_slat)
 719                         fio_gettime(&io_u->start_time, NULL);
 720
 721                 ret = io_u_submit(td, io_u);
 722
 723                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 724                         break;
 725
 726                 /*
 727                  * if we can queue more, do so. but check if there are
 728                  * completed io_u's first. Note that we can get BUSY even
 729                  * without IO queued, if the system is resource starved.
 730                  */
 731 reap:
 732                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 733                 if (full || io_in_polling(td))
 734                         ret = wait_for_completions(td, NULL);
 735
 736                 if (ret < 0)
 737                         break;
 738         }
 739
 740         check_update_rusage(td);
 741
 742         if (!td->error) {
 743                 min_events = td->cur_depth;
 744
 745                 if (min_events)
 746                         ret = io_u_queued_complete(td, min_events);
 747         } else
 748                 cleanup_pending_aio(td);
 749
 750         td_set_runstate(td, TD_RUNNING);
 751
 752         dprint(FD_VERIFY, "exiting loop\n");
 753 }
 754
 755 static bool exceeds_number_ios(struct thread_data *td)
 756 {
 757         unsigned long long number_ios;
 758
 759         if (!td->o.number_ios)
 760                 return false;
 761
 762         number_ios = ddir_rw_sum(td->io_blocks);
 763         number_ios += td->io_u_queued + td->io_u_in_flight;
 764
 765         return number_ios >= (td->o.number_ios * td->loops);
 766 }
 767
 768 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 769 {
 770         unsigned long long bytes, limit;
 771
 772         if (td_rw(td))
 773                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 774         else if (td_write(td))
 775                 bytes = this_bytes[DDIR_WRITE];
 776         else if (td_read(td))
 777                 bytes = this_bytes[DDIR_READ];
 778         else
 779                 bytes = this_bytes[DDIR_TRIM];
 780
 781         if (td->o.io_size)
 782                 limit = td->o.io_size;
 783         else
 784                 limit = td->o.size;
 785
 786         limit *= td->loops;
 787         return bytes >= limit || exceeds_number_ios(td);
 788 }
 789
 790 static bool io_issue_bytes_exceeded(struct thread_data *td)
 791 {
 792         return io_bytes_exceeded(td, td->io_issue_bytes);
 793 }
 794
 795 static bool io_complete_bytes_exceeded(struct thread_data *td)
 796 {
 797         return io_bytes_exceeded(td, td->this_io_bytes);
 798 }
 799
 800 /*
 801  * used to calculate the next io time for rate control
 802  *
 803  */
 804 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 805 {
 806         uint64_t bps = td->rate_bps[ddir];
 807
 808         assert(!(td->flags & TD_F_CHILD));
 809
 810         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 811                 uint64_t val, iops;
 812
 813                 iops = bps / td->o.min_bs[ddir];
 814                 val = (int64_t) (1000000 / iops) *
 815                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 816                 if (val) {
 817                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 818                                         (unsigned long long) 1000000 / val,
 819                                         ddir);
 820                 }
 821                 td->last_usec[ddir] += val;
 822                 return td->last_usec[ddir];
 823         } else if (bps) {
 824                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 825                 uint64_t secs = bytes / bps;
 826                 uint64_t remainder = bytes % bps;
 827
 828                 return remainder * 1000000 / bps + secs * 1000000;
 829         }
 830
 831         return 0;
 832 }
 833
 834 static void init_thinktime(struct thread_data *td)
 835 {
 836         if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
 837                 td->thinktime_blocks_counter = td->io_blocks;
 838         else
 839                 td->thinktime_blocks_counter = td->io_issues;
 840         td->last_thinktime = td->epoch;
 841         td->last_thinktime_blocks = 0;
 842 }
 843
 844 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
 845                              struct timespec *time)
 846 {
 847         unsigned long long b;
 848         uint64_t total;
 849         int left;
 850         struct timespec now;
 851         bool stall = false;
 852
 853         if (td->o.thinktime_iotime) {
 854                 fio_gettime(&now, NULL);
 855                 if (utime_since(&td->last_thinktime, &now)
 856                     >= td->o.thinktime_iotime + td->o.thinktime) {
 857                         stall = true;
 858                 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
 859                         /*
 860                          * When thinktime_iotime is set and thinktime_blocks is
 861                          * not set, skip the thinktime_blocks check, since
 862                          * thinktime_blocks default value 1 does not work
 863                          * together with thinktime_iotime.
 864                          */
 865                         return;
 866                 }
 867
 868         }
 869
 870         b = ddir_rw_sum(td->thinktime_blocks_counter);
 871         if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
 872                 stall = true;
 873
 874         if (!stall)
 875                 return;
 876
 877         io_u_quiesce(td);
 878
 879         total = 0;
 880         if (td->o.thinktime_spin)
 881                 total = usec_spin(td->o.thinktime_spin);
 882
 883         left = td->o.thinktime - total;
 884         if (left)
 885                 total += usec_sleep(td, left);
 886
 887         /*
 888          * If we're ignoring thinktime for the rate, add the number of bytes
 889          * we would have done while sleeping, minus one block to ensure we
 890          * start issuing immediately after the sleep.
 891          */
 892         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 893                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 894                 uint64_t bs = td->o.min_bs[ddir];
 895                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 896                 uint64_t over;
 897
 898                 if (usperop <= total)
 899                         over = bs;
 900                 else
 901                         over = (usperop - total) / usperop * -bs;
 902
 903                 td->rate_io_issue_bytes[ddir] += (missed - over);
 904                 /* adjust for rate_process=poisson */
 905                 td->last_usec[ddir] += total;
 906         }
 907
 908         if (time && should_check_rate(td))
 909                 fio_gettime(time, NULL);
 910
 911         td->last_thinktime_blocks = b;
 912         if (td->o.thinktime_iotime)
 913                 td->last_thinktime = now;
 914 }
 915
 916 /*
 917  * Main IO worker function. It retrieves io_u's to process and queues
 918  * and reaps them, checking for rate and errors along the way.
 919  *
 920  * Returns number of bytes written and trimmed.
 921  */
 922 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 923 {
 924         unsigned int i;
 925         int ret = 0;
 926         uint64_t total_bytes, bytes_issued = 0;
 927
 928         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 929                 bytes_done[i] = td->bytes_done[i];
 930
 931         if (in_ramp_time(td))
 932                 td_set_runstate(td, TD_RAMP);
 933         else
 934                 td_set_runstate(td, TD_RUNNING);
 935
 936         lat_target_init(td);
 937
 938         total_bytes = td->o.size;
 939         /*
 940         * Allow random overwrite workloads to write up to io_size
 941         * before starting verification phase as 'size' doesn't apply.
 942         */
 943         if (td_write(td) && td_random(td) && td->o.norandommap)
 944                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 945         /*
 946          * If verify_backlog is enabled, we'll run the verify in this
 947          * handler as well. For that case, we may need up to twice the
 948          * amount of bytes.
 949          */
 950         if (td->o.verify != VERIFY_NONE &&
 951            (td_write(td) && td->o.verify_backlog))
 952                 total_bytes += td->o.size;
 953
 954         /* In trimwrite mode, each byte is trimmed and then written, so
 955          * allow total_bytes to be twice as big */
 956         if (td_trimwrite(td))
 957                 total_bytes += td->total_io_size;
 958
 959         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 960                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 961                 td->o.time_based) {
 962                 struct timespec comp_time;
 963                 struct io_u *io_u;
 964                 int full;
 965                 enum fio_ddir ddir;
 966
 967                 check_update_rusage(td);
 968
 969                 if (td->terminate || td->done)
 970                         break;
 971
 972                 update_ts_cache(td);
 973
 974                 if (runtime_exceeded(td, &td->ts_cache)) {
 975                         __update_ts_cache(td);
 976                         if (runtime_exceeded(td, &td->ts_cache)) {
 977                                 fio_mark_td_terminate(td);
 978                                 break;
 979                         }
 980                 }
 981
 982                 if (flow_threshold_exceeded(td))
 983                         continue;
 984
 985                 /*
 986                  * Break if we exceeded the bytes. The exception is time
 987                  * based runs, but we still need to break out of the loop
 988                  * for those to run verification, if enabled.
 989                  * Jobs read from iolog do not use this stop condition.
 990                  */
 991                 if (bytes_issued >= total_bytes &&
 992                     !td->o.read_iolog_file &&
 993                     (!td->o.time_based ||
 994                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 995                         break;
 996
 997                 io_u = get_io_u(td);
 998                 if (IS_ERR_OR_NULL(io_u)) {
 999                         int err = PTR_ERR(io_u);
1000
1001                         io_u = NULL;
1002                         ddir = DDIR_INVAL;
1003                         if (err == -EBUSY) {
1004                                 ret = FIO_Q_BUSY;
1005                                 goto reap;
1006                         }
1007                         if (td->o.latency_target)
1008                                 goto reap;
1009                         break;
1010                 }
1011
1012                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1013                         populate_verify_io_u(td, io_u);
1014
1015                 ddir = io_u->ddir;
1016
1017                 /*
1018                  * Add verification end_io handler if:
1019                  *      - Asked to verify (!td_rw(td))
1020                  *      - Or the io_u is from our verify list (mixed write/ver)
1021                  */
1022                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1023                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1024
1025                         if (verify_state_should_stop(td, io_u)) {
1026                                 put_io_u(td, io_u);
1027                                 break;
1028                         }
1029
1030                         if (td->o.verify_async)
1031                                 io_u->end_io = verify_io_u_async;
1032                         else
1033                                 io_u->end_io = verify_io_u;
1034                         td_set_runstate(td, TD_VERIFYING);
1035                 } else if (in_ramp_time(td))
1036                         td_set_runstate(td, TD_RAMP);
1037                 else
1038                         td_set_runstate(td, TD_RUNNING);
1039
1040                 /*
1041                  * Always log IO before it's issued, so we know the specific
1042                  * order of it. The logged unit will track when the IO has
1043                  * completed.
1044                  */
1045                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1046                     td->o.do_verify &&
1047                     td->o.verify != VERIFY_NONE &&
1048                     !td->o.experimental_verify)
1049                         log_io_piece(td, io_u);
1050
1051                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1052                         const unsigned long long blen = io_u->xfer_buflen;
1053                         const enum fio_ddir __ddir = acct_ddir(io_u);
1054
1055                         if (td->error)
1056                                 break;
1057
1058                         workqueue_enqueue(&td->io_wq, &io_u->work);
1059                         ret = FIO_Q_QUEUED;
1060
1061                         if (ddir_rw(__ddir)) {
1062                                 td->io_issues[__ddir]++;
1063                                 td->io_issue_bytes[__ddir] += blen;
1064                                 td->rate_io_issue_bytes[__ddir] += blen;
1065                         }
1066
1067                         if (should_check_rate(td)) {
1068                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1069                                 fio_gettime(&comp_time, NULL);
1070                         }
1071
1072                 } else {
1073                         ret = io_u_submit(td, io_u);
1074
1075                         if (should_check_rate(td))
1076                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1077
1078                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1079                                 break;
1080
1081                         /*
1082                          * See if we need to complete some commands. Note that
1083                          * we can get BUSY even without IO queued, if the
1084                          * system is resource starved.
1085                          */
1086 reap:
1087                         full = queue_full(td) ||
1088                                 (ret == FIO_Q_BUSY && td->cur_depth);
1089                         if (full || io_in_polling(td))
1090                                 ret = wait_for_completions(td, &comp_time);
1091                 }
1092                 if (ret < 0)
1093                         break;
1094
1095                 if (ddir_rw(ddir) && td->o.thinktime)
1096                         handle_thinktime(td, ddir, &comp_time);
1097
1098                 if (!ddir_rw_sum(td->bytes_done) &&
1099                     !td_ioengine_flagged(td, FIO_NOIO))
1100                         continue;
1101
1102                 if (!in_ramp_time(td) && should_check_rate(td)) {
1103                         if (check_min_rate(td, &comp_time)) {
1104                                 if (exitall_on_terminate || td->o.exitall_error)
1105                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1106                                 td_verror(td, EIO, "check_min_rate");
1107                                 break;
1108                         }
1109                 }
1110                 if (!in_ramp_time(td) && td->o.latency_target)
1111                         lat_target_check(td);
1112         }
1113
1114         check_update_rusage(td);
1115
1116         if (td->trim_entries)
1117                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1118
1119         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1120                 td->error = 0;
1121                 fio_mark_td_terminate(td);
1122         }
1123         if (!td->error) {
1124                 struct fio_file *f;
1125
1126                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1127                         workqueue_flush(&td->io_wq);
1128                         i = 0;
1129                 } else
1130                         i = td->cur_depth;
1131
1132                 if (i) {
1133                         ret = io_u_queued_complete(td, i);
1134                         if (td->o.fill_device &&
1135                             (td->error == ENOSPC || td->error == EDQUOT))
1136                                 td->error = 0;
1137                 }
1138
1139                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1140                         td_set_runstate(td, TD_FSYNCING);
1141
1142                         for_each_file(td, f, i) {
1143                                 if (!fio_file_fsync(td, f))
1144                                         continue;
1145
1146                                 log_err("fio: end_fsync failed for file %s\n",
1147                                                                 f->file_name);
1148                         }
1149                 }
1150         } else {
1151                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1152                         workqueue_flush(&td->io_wq);
1153                 cleanup_pending_aio(td);
1154         }
1155
1156         /*
1157          * stop job if we failed doing any IO
1158          */
1159         if (!ddir_rw_sum(td->this_io_bytes))
1160                 td->done = 1;
1161
1162         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1163                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1164 }
1165
1166 static void free_file_completion_logging(struct thread_data *td)
1167 {
1168         struct fio_file *f;
1169         unsigned int i;
1170
1171         for_each_file(td, f, i) {
1172                 if (!f->last_write_comp)
1173                         break;
1174                 sfree(f->last_write_comp);
1175         }
1176 }
1177
1178 static int init_file_completion_logging(struct thread_data *td,
1179                                         unsigned int depth)
1180 {
1181         struct fio_file *f;
1182         unsigned int i;
1183
1184         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1185                 return 0;
1186
1187         for_each_file(td, f, i) {
1188                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1189                 if (!f->last_write_comp)
1190                         goto cleanup;
1191         }
1192
1193         return 0;
1194
1195 cleanup:
1196         free_file_completion_logging(td);
1197         log_err("fio: failed to alloc write comp data\n");
1198         return 1;
1199 }
1200
1201 static void cleanup_io_u(struct thread_data *td)
1202 {
1203         struct io_u *io_u;
1204
1205         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1206
1207                 if (td->io_ops->io_u_free)
1208                         td->io_ops->io_u_free(td, io_u);
1209
1210                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1211         }
1212
1213         free_io_mem(td);
1214
1215         io_u_rexit(&td->io_u_requeues);
1216         io_u_qexit(&td->io_u_freelist, false);
1217         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1218
1219         free_file_completion_logging(td);
1220 }
1221
1222 static int init_io_u(struct thread_data *td)
1223 {
1224         struct io_u *io_u;
1225         int cl_align, i, max_units;
1226         int err;
1227
1228         max_units = td->o.iodepth;
1229
1230         err = 0;
1231         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1232         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1233         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1234
1235         if (err) {
1236                 log_err("fio: failed setting up IO queues\n");
1237                 return 1;
1238         }
1239
1240         cl_align = os_cache_line_size();
1241
1242         for (i = 0; i < max_units; i++) {
1243                 void *ptr;
1244
1245                 if (td->terminate)
1246                         return 1;
1247
1248                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1249                 if (!ptr) {
1250                         log_err("fio: unable to allocate aligned memory\n");
1251                         return 1;
1252                 }
1253
1254                 io_u = ptr;
1255                 memset(io_u, 0, sizeof(*io_u));
1256                 INIT_FLIST_HEAD(&io_u->verify_list);
1257                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1258
1259                 io_u->index = i;
1260                 io_u->flags = IO_U_F_FREE;
1261                 io_u_qpush(&td->io_u_freelist, io_u);
1262
1263                 /*
1264                  * io_u never leaves this stack, used for iteration of all
1265                  * io_u buffers.
1266                  */
1267                 io_u_qpush(&td->io_u_all, io_u);
1268
1269                 if (td->io_ops->io_u_init) {
1270                         int ret = td->io_ops->io_u_init(td, io_u);
1271
1272                         if (ret) {
1273                                 log_err("fio: failed to init engine data: %d\n", ret);
1274                                 return 1;
1275                         }
1276                 }
1277         }
1278
1279         init_io_u_buffers(td);
1280
1281         if (init_file_completion_logging(td, max_units))
1282                 return 1;
1283
1284         return 0;
1285 }
1286
1287 int init_io_u_buffers(struct thread_data *td)
1288 {
1289         struct io_u *io_u;
1290         unsigned long long max_bs, min_write;
1291         int i, max_units;
1292         int data_xfer = 1;
1293         char *p;
1294
1295         max_units = td->o.iodepth;
1296         max_bs = td_max_bs(td);
1297         min_write = td->o.min_bs[DDIR_WRITE];
1298         td->orig_buffer_size = (unsigned long long) max_bs
1299                                         * (unsigned long long) max_units;
1300
1301         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1302                 data_xfer = 0;
1303
1304         /*
1305          * if we may later need to do address alignment, then add any
1306          * possible adjustment here so that we don't cause a buffer
1307          * overflow later. this adjustment may be too much if we get
1308          * lucky and the allocator gives us an aligned address.
1309          */
1310         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1311             td_ioengine_flagged(td, FIO_RAWIO))
1312                 td->orig_buffer_size += page_mask + td->o.mem_align;
1313
1314         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1315                 unsigned long long bs;
1316
1317                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1318                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1319         }
1320
1321         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1322                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1323                 return 1;
1324         }
1325
1326         if (data_xfer && allocate_io_mem(td))
1327                 return 1;
1328
1329         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1330             td_ioengine_flagged(td, FIO_RAWIO))
1331                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1332         else
1333                 p = td->orig_buffer;
1334
1335         for (i = 0; i < max_units; i++) {
1336                 io_u = td->io_u_all.io_us[i];
1337                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1338
1339                 if (data_xfer) {
1340                         io_u->buf = p;
1341                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1342
1343                         if (td_write(td))
1344                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1345                         if (td_write(td) && td->o.verify_pattern_bytes) {
1346                                 /*
1347                                  * Fill the buffer with the pattern if we are
1348                                  * going to be doing writes.
1349                                  */
1350                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1351                         }
1352                 }
1353                 p += max_bs;
1354         }
1355
1356         return 0;
1357 }
1358
1359 #ifdef FIO_HAVE_IOSCHED_SWITCH
1360 /*
1361  * These functions are Linux specific.
1362  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1363  */
1364 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1365 {
1366         char tmp[256], tmp2[128], *p;
1367         FILE *f;
1368         int ret;
1369
1370         assert(file->du && file->du->sysfs_root);
1371         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1372
1373         f = fopen(tmp, "r+");
1374         if (!f) {
1375                 if (errno == ENOENT) {
1376                         log_err("fio: os or kernel doesn't support IO scheduler"
1377                                 " switching\n");
1378                         return 0;
1379                 }
1380                 td_verror(td, errno, "fopen iosched");
1381                 return 1;
1382         }
1383
1384         /*
1385          * Set io scheduler.
1386          */
1387         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1388         if (ferror(f) || ret != 1) {
1389                 td_verror(td, errno, "fwrite");
1390                 fclose(f);
1391                 return 1;
1392         }
1393
1394         rewind(f);
1395
1396         /*
1397          * Read back and check that the selected scheduler is now the default.
1398          */
1399         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1400         if (ferror(f) || ret < 0) {
1401                 td_verror(td, errno, "fread");
1402                 fclose(f);
1403                 return 1;
1404         }
1405         tmp[ret] = '\0';
1406         /*
1407          * either a list of io schedulers or "none\n" is expected. Strip the
1408          * trailing newline.
1409          */
1410         p = tmp;
1411         strsep(&p, "\n");
1412
1413         /*
1414          * Write to "none" entry doesn't fail, so check the result here.
1415          */
1416         if (!strcmp(tmp, "none")) {
1417                 log_err("fio: io scheduler is not tunable\n");
1418                 fclose(f);
1419                 return 0;
1420         }
1421
1422         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1423         if (!strstr(tmp, tmp2)) {
1424                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1425                 td_verror(td, EINVAL, "iosched_switch");
1426                 fclose(f);
1427                 return 1;
1428         }
1429
1430         fclose(f);
1431         return 0;
1432 }
1433
1434 static int switch_ioscheduler(struct thread_data *td)
1435 {
1436         struct fio_file *f;
1437         unsigned int i;
1438         int ret = 0;
1439
1440         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1441                 return 0;
1442
1443         assert(td->files && td->files[0]);
1444
1445         for_each_file(td, f, i) {
1446
1447                 /* Only consider regular files and block device files */
1448                 switch (f->filetype) {
1449                 case FIO_TYPE_FILE:
1450                 case FIO_TYPE_BLOCK:
1451                         /*
1452                          * Make sure that the device hosting the file could
1453                          * be determined.
1454                          */
1455                         if (!f->du)
1456                                 continue;
1457                         break;
1458                 case FIO_TYPE_CHAR:
1459                 case FIO_TYPE_PIPE:
1460                 default:
1461                         continue;
1462                 }
1463
1464                 ret = set_ioscheduler(td, f);
1465                 if (ret)
1466                         return ret;
1467         }
1468
1469         return 0;
1470 }
1471
1472 #else
1473
1474 static int switch_ioscheduler(struct thread_data *td)
1475 {
1476         return 0;
1477 }
1478
1479 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1480
1481 static bool keep_running(struct thread_data *td)
1482 {
1483         unsigned long long limit;
1484
1485         if (td->done)
1486                 return false;
1487         if (td->terminate)
1488                 return false;
1489         if (td->o.time_based)
1490                 return true;
1491         if (td->o.loops) {
1492                 td->o.loops--;
1493                 return true;
1494         }
1495         if (exceeds_number_ios(td))
1496                 return false;
1497
1498         if (td->o.io_size)
1499                 limit = td->o.io_size;
1500         else
1501                 limit = td->o.size;
1502
1503         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1504                 uint64_t diff;
1505
1506                 /*
1507                  * If the difference is less than the maximum IO size, we
1508                  * are done.
1509                  */
1510                 diff = limit - ddir_rw_sum(td->io_bytes);
1511                 if (diff < td_max_bs(td))
1512                         return false;
1513
1514                 if (fio_files_done(td) && !td->o.io_size)
1515                         return false;
1516
1517                 return true;
1518         }
1519
1520         return false;
1521 }
1522
1523 static int exec_string(struct thread_options *o, const char *string,
1524                        const char *mode)
1525 {
1526         int ret;
1527         char *str;
1528
1529         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1530                 return -1;
1531
1532         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1533                  o->name, mode);
1534         ret = system(str);
1535         if (ret == -1)
1536                 log_err("fio: exec of cmd <%s> failed\n", str);
1537
1538         free(str);
1539         return ret;
1540 }
1541
1542 /*
1543  * Dry run to compute correct state of numberio for verification.
1544  */
1545 static uint64_t do_dry_run(struct thread_data *td)
1546 {
1547         td_set_runstate(td, TD_RUNNING);
1548
1549         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1550                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1551                 struct io_u *io_u;
1552                 int ret;
1553
1554                 if (td->terminate || td->done)
1555                         break;
1556
1557                 io_u = get_io_u(td);
1558                 if (IS_ERR_OR_NULL(io_u))
1559                         break;
1560
1561                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1562                 io_u->error = 0;
1563                 io_u->resid = 0;
1564                 if (ddir_rw(acct_ddir(io_u)))
1565                         td->io_issues[acct_ddir(io_u)]++;
1566                 if (ddir_rw(io_u->ddir)) {
1567                         io_u_mark_depth(td, 1);
1568                         td->ts.total_io_u[io_u->ddir]++;
1569                 }
1570
1571                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1572                     td->o.do_verify &&
1573                     td->o.verify != VERIFY_NONE &&
1574                     !td->o.experimental_verify)
1575                         log_io_piece(td, io_u);
1576
1577                 ret = io_u_sync_complete(td, io_u);
1578                 (void) ret;
1579         }
1580
1581         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1582 }
1583
1584 struct fork_data {
1585         struct thread_data *td;
1586         struct sk_out *sk_out;
1587 };
1588
1589 /*
1590  * Entry point for the thread based jobs. The process based jobs end up
1591  * here as well, after a little setup.
1592  */
1593 static void *thread_main(void *data)
1594 {
1595         struct fork_data *fd = data;
1596         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1597         struct thread_data *td = fd->td;
1598         struct thread_options *o = &td->o;
1599         struct sk_out *sk_out = fd->sk_out;
1600         uint64_t bytes_done[DDIR_RWDIR_CNT];
1601         int deadlock_loop_cnt;
1602         bool clear_state;
1603         int res, ret;
1604
1605         sk_out_assign(sk_out);
1606         free(fd);
1607
1608         if (!o->use_thread) {
1609                 setsid();
1610                 td->pid = getpid();
1611         } else
1612                 td->pid = gettid();
1613
1614         fio_local_clock_init();
1615
1616         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1617
1618         if (is_backend)
1619                 fio_server_send_start(td);
1620
1621         INIT_FLIST_HEAD(&td->io_log_list);
1622         INIT_FLIST_HEAD(&td->io_hist_list);
1623         INIT_FLIST_HEAD(&td->verify_list);
1624         INIT_FLIST_HEAD(&td->trim_list);
1625         td->io_hist_tree = RB_ROOT;
1626
1627         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1628         if (ret) {
1629                 td_verror(td, ret, "mutex_cond_init_pshared");
1630                 goto err;
1631         }
1632         ret = cond_init_pshared(&td->verify_cond);
1633         if (ret) {
1634                 td_verror(td, ret, "mutex_cond_pshared");
1635                 goto err;
1636         }
1637
1638         td_set_runstate(td, TD_INITIALIZED);
1639         dprint(FD_MUTEX, "up startup_sem\n");
1640         fio_sem_up(startup_sem);
1641         dprint(FD_MUTEX, "wait on td->sem\n");
1642         fio_sem_down(td->sem);
1643         dprint(FD_MUTEX, "done waiting on td->sem\n");
1644
1645         /*
1646          * A new gid requires privilege, so we need to do this before setting
1647          * the uid.
1648          */
1649         if (o->gid != -1U && setgid(o->gid)) {
1650                 td_verror(td, errno, "setgid");
1651                 goto err;
1652         }
1653         if (o->uid != -1U && setuid(o->uid)) {
1654                 td_verror(td, errno, "setuid");
1655                 goto err;
1656         }
1657
1658         td_zone_gen_index(td);
1659
1660         /*
1661          * Do this early, we don't want the compress threads to be limited
1662          * to the same CPUs as the IO workers. So do this before we set
1663          * any potential CPU affinity
1664          */
1665         if (iolog_compress_init(td, sk_out))
1666                 goto err;
1667
1668         /*
1669          * If we have a gettimeofday() thread, make sure we exclude that
1670          * thread from this job
1671          */
1672         if (o->gtod_cpu)
1673                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1674
1675         /*
1676          * Set affinity first, in case it has an impact on the memory
1677          * allocations.
1678          */
1679         if (fio_option_is_set(o, cpumask)) {
1680                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1681                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1682                         if (!ret) {
1683                                 log_err("fio: no CPUs set\n");
1684                                 log_err("fio: Try increasing number of available CPUs\n");
1685                                 td_verror(td, EINVAL, "cpus_split");
1686                                 goto err;
1687                         }
1688                 }
1689                 ret = fio_setaffinity(td->pid, o->cpumask);
1690                 if (ret == -1) {
1691                         td_verror(td, errno, "cpu_set_affinity");
1692                         goto err;
1693                 }
1694         }
1695
1696 #ifdef CONFIG_LIBNUMA
1697         /* numa node setup */
1698         if (fio_option_is_set(o, numa_cpunodes) ||
1699             fio_option_is_set(o, numa_memnodes)) {
1700                 struct bitmask *mask;
1701
1702                 if (numa_available() < 0) {
1703                         td_verror(td, errno, "Does not support NUMA API\n");
1704                         goto err;
1705                 }
1706
1707                 if (fio_option_is_set(o, numa_cpunodes)) {
1708                         mask = numa_parse_nodestring(o->numa_cpunodes);
1709                         ret = numa_run_on_node_mask(mask);
1710                         numa_free_nodemask(mask);
1711                         if (ret == -1) {
1712                                 td_verror(td, errno, \
1713                                         "numa_run_on_node_mask failed\n");
1714                                 goto err;
1715                         }
1716                 }
1717
1718                 if (fio_option_is_set(o, numa_memnodes)) {
1719                         mask = NULL;
1720                         if (o->numa_memnodes)
1721                                 mask = numa_parse_nodestring(o->numa_memnodes);
1722
1723                         switch (o->numa_mem_mode) {
1724                         case MPOL_INTERLEAVE:
1725                                 numa_set_interleave_mask(mask);
1726                                 break;
1727                         case MPOL_BIND:
1728                                 numa_set_membind(mask);
1729                                 break;
1730                         case MPOL_LOCAL:
1731                                 numa_set_localalloc();
1732                                 break;
1733                         case MPOL_PREFERRED:
1734                                 numa_set_preferred(o->numa_mem_prefer_node);
1735                                 break;
1736                         case MPOL_DEFAULT:
1737                         default:
1738                                 break;
1739                         }
1740
1741                         if (mask)
1742                                 numa_free_nodemask(mask);
1743
1744                 }
1745         }
1746 #endif
1747
1748         if (fio_pin_memory(td))
1749                 goto err;
1750
1751         /*
1752          * May alter parameters that init_io_u() will use, so we need to
1753          * do this first.
1754          */
1755         if (!init_iolog(td))
1756                 goto err;
1757
1758         /* ioprio_set() has to be done before td_io_init() */
1759         if (fio_option_is_set(o, ioprio) ||
1760             fio_option_is_set(o, ioprio_class)) {
1761                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1762                 if (ret == -1) {
1763                         td_verror(td, errno, "ioprio_set");
1764                         goto err;
1765                 }
1766                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1767                 td->ts.ioprio = td->ioprio;
1768         }
1769
1770         if (td_io_init(td))
1771                 goto err;
1772
1773         if (init_io_u(td))
1774                 goto err;
1775
1776         if (td->io_ops->post_init && td->io_ops->post_init(td))
1777                 goto err;
1778
1779         if (o->verify_async && verify_async_init(td))
1780                 goto err;
1781
1782         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1783                 goto err;
1784
1785         errno = 0;
1786         if (nice(o->nice) == -1 && errno != 0) {
1787                 td_verror(td, errno, "nice");
1788                 goto err;
1789         }
1790
1791         if (o->ioscheduler && switch_ioscheduler(td))
1792                 goto err;
1793
1794         if (!o->create_serialize && setup_files(td))
1795                 goto err;
1796
1797         if (!init_random_map(td))
1798                 goto err;
1799
1800         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1801                 goto err;
1802
1803         if (o->pre_read && !pre_read_files(td))
1804                 goto err;
1805
1806         fio_verify_init(td);
1807
1808         if (rate_submit_init(td, sk_out))
1809                 goto err;
1810
1811         set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1812         fio_getrusage(&td->ru_start);
1813         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1814         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1815         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1816
1817         init_thinktime(td);
1818
1819         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1820                         o->ratemin[DDIR_TRIM]) {
1821                 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1822                                         sizeof(td->bw_sample_time));
1823                 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1824                                         sizeof(td->bw_sample_time));
1825                 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1826                                         sizeof(td->bw_sample_time));
1827         }
1828
1829         memset(bytes_done, 0, sizeof(bytes_done));
1830         clear_state = false;
1831
1832         while (keep_running(td)) {
1833                 uint64_t verify_bytes;
1834
1835                 fio_gettime(&td->start, NULL);
1836                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1837
1838                 if (clear_state) {
1839                         clear_io_state(td, 0);
1840
1841                         if (o->unlink_each_loop && unlink_all_files(td))
1842                                 break;
1843                 }
1844
1845                 prune_io_piece_log(td);
1846
1847                 if (td->o.verify_only && td_write(td))
1848                         verify_bytes = do_dry_run(td);
1849                 else {
1850                         do_io(td, bytes_done);
1851
1852                         if (!ddir_rw_sum(bytes_done)) {
1853                                 fio_mark_td_terminate(td);
1854                                 verify_bytes = 0;
1855                         } else {
1856                                 verify_bytes = bytes_done[DDIR_WRITE] +
1857                                                 bytes_done[DDIR_TRIM];
1858                         }
1859                 }
1860
1861                 /*
1862                  * If we took too long to shut down, the main thread could
1863                  * already consider us reaped/exited. If that happens, break
1864                  * out and clean up.
1865                  */
1866                 if (td->runstate >= TD_EXITED)
1867                         break;
1868
1869                 clear_state = true;
1870
1871                 /*
1872                  * Make sure we've successfully updated the rusage stats
1873                  * before waiting on the stat mutex. Otherwise we could have
1874                  * the stat thread holding stat mutex and waiting for
1875                  * the rusage_sem, which would never get upped because
1876                  * this thread is waiting for the stat mutex.
1877                  */
1878                 deadlock_loop_cnt = 0;
1879                 do {
1880                         check_update_rusage(td);
1881                         if (!fio_sem_down_trylock(stat_sem))
1882                                 break;
1883                         usleep(1000);
1884                         if (deadlock_loop_cnt++ > 5000) {
1885                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1886                                 td->error = EDEADLK;
1887                                 goto err;
1888                         }
1889                 } while (1);
1890
1891                 if (td_read(td) && td->io_bytes[DDIR_READ])
1892                         update_runtime(td, elapsed_us, DDIR_READ);
1893                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1894                         update_runtime(td, elapsed_us, DDIR_WRITE);
1895                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1896                         update_runtime(td, elapsed_us, DDIR_TRIM);
1897                 fio_gettime(&td->start, NULL);
1898                 fio_sem_up(stat_sem);
1899
1900                 if (td->error || td->terminate)
1901                         break;
1902
1903                 if (!o->do_verify ||
1904                     o->verify == VERIFY_NONE ||
1905                     td_ioengine_flagged(td, FIO_UNIDIR))
1906                         continue;
1907
1908                 clear_io_state(td, 0);
1909
1910                 fio_gettime(&td->start, NULL);
1911
1912                 do_verify(td, verify_bytes);
1913
1914                 /*
1915                  * See comment further up for why this is done here.
1916                  */
1917                 check_update_rusage(td);
1918
1919                 fio_sem_down(stat_sem);
1920                 update_runtime(td, elapsed_us, DDIR_READ);
1921                 fio_gettime(&td->start, NULL);
1922                 fio_sem_up(stat_sem);
1923
1924                 if (td->error || td->terminate)
1925                         break;
1926         }
1927
1928         /*
1929          * Acquire this lock if we were doing overlap checking in
1930          * offload mode so that we don't clean up this job while
1931          * another thread is checking its io_u's for overlap
1932          */
1933         if (td_offload_overlap(td)) {
1934                 int res = pthread_mutex_lock(&overlap_check);
1935                 assert(res == 0);
1936         }
1937         td_set_runstate(td, TD_FINISHING);
1938         if (td_offload_overlap(td)) {
1939                 res = pthread_mutex_unlock(&overlap_check);
1940                 assert(res == 0);
1941         }
1942
1943         update_rusage_stat(td);
1944         td->ts.total_run_time = mtime_since_now(&td->epoch);
1945         for_each_rw_ddir(ddir) {
1946                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1947         }
1948
1949         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1950             (td->o.verify != VERIFY_NONE && td_write(td)))
1951                 verify_save_state(td->thread_number);
1952
1953         fio_unpin_memory(td);
1954
1955         td_writeout_logs(td, true);
1956
1957         iolog_compress_exit(td);
1958         rate_submit_exit(td);
1959
1960         if (o->exec_postrun)
1961                 exec_string(o, o->exec_postrun, "postrun");
1962
1963         if (exitall_on_terminate || (o->exitall_error && td->error))
1964                 fio_terminate_threads(td->groupid, td->o.exit_what);
1965
1966 err:
1967         if (td->error)
1968                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1969                                                         td->verror);
1970
1971         if (o->verify_async)
1972                 verify_async_exit(td);
1973
1974         close_and_free_files(td);
1975         cleanup_io_u(td);
1976         close_ioengine(td);
1977         cgroup_shutdown(td, cgroup_mnt);
1978         verify_free_state(td);
1979         td_zone_free_index(td);
1980
1981         if (fio_option_is_set(o, cpumask)) {
1982                 ret = fio_cpuset_exit(&o->cpumask);
1983                 if (ret)
1984                         td_verror(td, ret, "fio_cpuset_exit");
1985         }
1986
1987         /*
1988          * do this very late, it will log file closing as well
1989          */
1990         if (o->write_iolog_file)
1991                 write_iolog_close(td);
1992         if (td->io_log_rfile)
1993                 fclose(td->io_log_rfile);
1994
1995         td_set_runstate(td, TD_EXITED);
1996
1997         /*
1998          * Do this last after setting our runstate to exited, so we
1999          * know that the stat thread is signaled.
2000          */
2001         check_update_rusage(td);
2002
2003         sk_out_drop();
2004         return (void *) (uintptr_t) td->error;
2005 }
2006
2007 /*
2008  * Run over the job map and reap the threads that have exited, if any.
2009  */
2010 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2011                          uint64_t *m_rate)
2012 {
2013         struct thread_data *td;
2014         unsigned int cputhreads, realthreads, pending;
2015         int i, status, ret;
2016
2017         /*
2018          * reap exited threads (TD_EXITED -> TD_REAPED)
2019          */
2020         realthreads = pending = cputhreads = 0;
2021         for_each_td(td, i) {
2022                 int flags = 0;
2023
2024                 if (!strcmp(td->o.ioengine, "cpuio"))
2025                         cputhreads++;
2026                 else
2027                         realthreads++;
2028
2029                 if (!td->pid) {
2030                         pending++;
2031                         continue;
2032                 }
2033                 if (td->runstate == TD_REAPED)
2034                         continue;
2035                 if (td->o.use_thread) {
2036                         if (td->runstate == TD_EXITED) {
2037                                 td_set_runstate(td, TD_REAPED);
2038                                 goto reaped;
2039                         }
2040                         continue;
2041                 }
2042
2043                 flags = WNOHANG;
2044                 if (td->runstate == TD_EXITED)
2045                         flags = 0;
2046
2047                 /*
2048                  * check if someone quit or got killed in an unusual way
2049                  */
2050                 ret = waitpid(td->pid, &status, flags);
2051                 if (ret < 0) {
2052                         if (errno == ECHILD) {
2053                                 log_err("fio: pid=%d disappeared %d\n",
2054                                                 (int) td->pid, td->runstate);
2055                                 td->sig = ECHILD;
2056                                 td_set_runstate(td, TD_REAPED);
2057                                 goto reaped;
2058                         }
2059                         perror("waitpid");
2060                 } else if (ret == td->pid) {
2061                         if (WIFSIGNALED(status)) {
2062                                 int sig = WTERMSIG(status);
2063
2064                                 if (sig != SIGTERM && sig != SIGUSR2)
2065                                         log_err("fio: pid=%d, got signal=%d\n",
2066                                                         (int) td->pid, sig);
2067                                 td->sig = sig;
2068                                 td_set_runstate(td, TD_REAPED);
2069                                 goto reaped;
2070                         }
2071                         if (WIFEXITED(status)) {
2072                                 if (WEXITSTATUS(status) && !td->error)
2073                                         td->error = WEXITSTATUS(status);
2074
2075                                 td_set_runstate(td, TD_REAPED);
2076                                 goto reaped;
2077                         }
2078                 }
2079
2080                 /*
2081                  * If the job is stuck, do a forceful timeout of it and
2082                  * move on.
2083                  */
2084                 if (td->terminate &&
2085                     td->runstate < TD_FSYNCING &&
2086                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2087                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2088                                 "%lu seconds, it appears to be stuck. Doing "
2089                                 "forceful exit of this job.\n",
2090                                 td->o.name, td->runstate,
2091                                 (unsigned long) time_since_now(&td->terminate_time));
2092                         td_set_runstate(td, TD_REAPED);
2093                         goto reaped;
2094                 }
2095
2096                 /*
2097                  * thread is not dead, continue
2098                  */
2099                 pending++;
2100                 continue;
2101 reaped:
2102                 (*nr_running)--;
2103                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2104                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2105                 if (!td->pid)
2106                         pending--;
2107
2108                 if (td->error)
2109                         exit_value++;
2110
2111                 done_secs += mtime_since_now(&td->epoch) / 1000;
2112                 profile_td_exit(td);
2113                 flow_exit_job(td);
2114         }
2115
2116         if (*nr_running == cputhreads && !pending && realthreads)
2117                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2118 }
2119
2120 static bool __check_trigger_file(void)
2121 {
2122         struct stat sb;
2123
2124         if (!trigger_file)
2125                 return false;
2126
2127         if (stat(trigger_file, &sb))
2128                 return false;
2129
2130         if (unlink(trigger_file) < 0)
2131                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2132                                                         strerror(errno));
2133
2134         return true;
2135 }
2136
2137 static bool trigger_timedout(void)
2138 {
2139         if (trigger_timeout)
2140                 if (time_since_genesis() >= trigger_timeout) {
2141                         trigger_timeout = 0;
2142                         return true;
2143                 }
2144
2145         return false;
2146 }
2147
2148 void exec_trigger(const char *cmd)
2149 {
2150         int ret;
2151
2152         if (!cmd || cmd[0] == '\0')
2153                 return;
2154
2155         ret = system(cmd);
2156         if (ret == -1)
2157                 log_err("fio: failed executing %s trigger\n", cmd);
2158 }
2159
2160 void check_trigger_file(void)
2161 {
2162         if (__check_trigger_file() || trigger_timedout()) {
2163                 if (nr_clients)
2164                         fio_clients_send_trigger(trigger_remote_cmd);
2165                 else {
2166                         verify_save_state(IO_LIST_ALL);
2167                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2168                         exec_trigger(trigger_cmd);
2169                 }
2170         }
2171 }
2172
2173 static int fio_verify_load_state(struct thread_data *td)
2174 {
2175         int ret;
2176
2177         if (!td->o.verify_state)
2178                 return 0;
2179
2180         if (is_backend) {
2181                 void *data;
2182
2183                 ret = fio_server_get_verify_state(td->o.name,
2184                                         td->thread_number - 1, &data);
2185                 if (!ret)
2186                         verify_assign_state(td, data);
2187         } else {
2188                 char prefix[PATH_MAX];
2189
2190                 if (aux_path)
2191                         sprintf(prefix, "%s%clocal", aux_path,
2192                                         FIO_OS_PATH_SEPARATOR);
2193                 else
2194                         strcpy(prefix, "local");
2195                 ret = verify_load_state(td, prefix);
2196         }
2197
2198         return ret;
2199 }
2200
2201 static void do_usleep(unsigned int usecs)
2202 {
2203         check_for_running_stats();
2204         check_trigger_file();
2205         usleep(usecs);
2206 }
2207
2208 static bool check_mount_writes(struct thread_data *td)
2209 {
2210         struct fio_file *f;
2211         unsigned int i;
2212
2213         if (!td_write(td) || td->o.allow_mounted_write)
2214                 return false;
2215
2216         /*
2217          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2218          * are mkfs'd and mounted.
2219          */
2220         for_each_file(td, f, i) {
2221 #ifdef FIO_HAVE_CHARDEV_SIZE
2222                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2223 #else
2224                 if (f->filetype != FIO_TYPE_BLOCK)
2225 #endif
2226                         continue;
2227                 if (device_is_mounted(f->file_name))
2228                         goto mounted;
2229         }
2230
2231         return false;
2232 mounted:
2233         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2234         return true;
2235 }
2236
2237 static bool waitee_running(struct thread_data *me)
2238 {
2239         const char *waitee = me->o.wait_for;
2240         const char *self = me->o.name;
2241         struct thread_data *td;
2242         int i;
2243
2244         if (!waitee)
2245                 return false;
2246
2247         for_each_td(td, i) {
2248                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2249                         continue;
2250
2251                 if (td->runstate < TD_EXITED) {
2252                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2253                                         self, td->o.name,
2254                                         runstate_to_name(td->runstate));
2255                         return true;
2256                 }
2257         }
2258
2259         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2260         return false;
2261 }
2262
2263 /*
2264  * Main function for kicking off and reaping jobs, as needed.
2265  */
2266 static void run_threads(struct sk_out *sk_out)
2267 {
2268         struct thread_data *td;
2269         unsigned int i, todo, nr_running, nr_started;
2270         uint64_t m_rate, t_rate;
2271         uint64_t spent;
2272
2273         if (fio_gtod_offload && fio_start_gtod_thread())
2274                 return;
2275
2276         fio_idle_prof_init();
2277
2278         set_sig_handlers();
2279
2280         nr_thread = nr_process = 0;
2281         for_each_td(td, i) {
2282                 if (check_mount_writes(td))
2283                         return;
2284                 if (td->o.use_thread)
2285                         nr_thread++;
2286                 else
2287                         nr_process++;
2288         }
2289
2290         if (output_format & FIO_OUTPUT_NORMAL) {
2291                 struct buf_output out;
2292
2293                 buf_output_init(&out);
2294                 __log_buf(&out, "Starting ");
2295                 if (nr_thread)
2296                         __log_buf(&out, "%d thread%s", nr_thread,
2297                                                 nr_thread > 1 ? "s" : "");
2298                 if (nr_process) {
2299                         if (nr_thread)
2300                                 __log_buf(&out, " and ");
2301                         __log_buf(&out, "%d process%s", nr_process,
2302                                                 nr_process > 1 ? "es" : "");
2303                 }
2304                 __log_buf(&out, "\n");
2305                 log_info_buf(out.buf, out.buflen);
2306                 buf_output_free(&out);
2307         }
2308
2309         todo = thread_number;
2310         nr_running = 0;
2311         nr_started = 0;
2312         m_rate = t_rate = 0;
2313
2314         for_each_td(td, i) {
2315                 print_status_init(td->thread_number - 1);
2316
2317                 if (!td->o.create_serialize)
2318                         continue;
2319
2320                 if (fio_verify_load_state(td))
2321                         goto reap;
2322
2323                 /*
2324                  * do file setup here so it happens sequentially,
2325                  * we don't want X number of threads getting their
2326                  * client data interspersed on disk
2327                  */
2328                 if (setup_files(td)) {
2329 reap:
2330                         exit_value++;
2331                         if (td->error)
2332                                 log_err("fio: pid=%d, err=%d/%s\n",
2333                                         (int) td->pid, td->error, td->verror);
2334                         td_set_runstate(td, TD_REAPED);
2335                         todo--;
2336                 } else {
2337                         struct fio_file *f;
2338                         unsigned int j;
2339
2340                         /*
2341                          * for sharing to work, each job must always open
2342                          * its own files. so close them, if we opened them
2343                          * for creation
2344                          */
2345                         for_each_file(td, f, j) {
2346                                 if (fio_file_open(f))
2347                                         td_io_close_file(td, f);
2348                         }
2349                 }
2350         }
2351
2352         /* start idle threads before io threads start to run */
2353         fio_idle_prof_start();
2354
2355         set_genesis_time();
2356
2357         while (todo) {
2358                 struct thread_data *map[REAL_MAX_JOBS];
2359                 struct timespec this_start;
2360                 int this_jobs = 0, left;
2361                 struct fork_data *fd;
2362
2363                 /*
2364                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2365                  */
2366                 for_each_td(td, i) {
2367                         if (td->runstate != TD_NOT_CREATED)
2368                                 continue;
2369
2370                         /*
2371                          * never got a chance to start, killed by other
2372                          * thread for some reason
2373                          */
2374                         if (td->terminate) {
2375                                 todo--;
2376                                 continue;
2377                         }
2378
2379                         if (td->o.start_delay) {
2380                                 spent = utime_since_genesis();
2381
2382                                 if (td->o.start_delay > spent)
2383                                         continue;
2384                         }
2385
2386                         if (td->o.stonewall && (nr_started || nr_running)) {
2387                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2388                                                         td->o.name);
2389                                 break;
2390                         }
2391
2392                         if (waitee_running(td)) {
2393                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2394                                                 td->o.name, td->o.wait_for);
2395                                 continue;
2396                         }
2397
2398                         init_disk_util(td);
2399
2400                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2401                         td->update_rusage = 0;
2402
2403                         /*
2404                          * Set state to created. Thread will transition
2405                          * to TD_INITIALIZED when it's done setting up.
2406                          */
2407                         td_set_runstate(td, TD_CREATED);
2408                         map[this_jobs++] = td;
2409                         nr_started++;
2410
2411                         fd = calloc(1, sizeof(*fd));
2412                         fd->td = td;
2413                         fd->sk_out = sk_out;
2414
2415                         if (td->o.use_thread) {
2416                                 int ret;
2417
2418                                 dprint(FD_PROCESS, "will pthread_create\n");
2419                                 ret = pthread_create(&td->thread, NULL,
2420                                                         thread_main, fd);
2421                                 if (ret) {
2422                                         log_err("pthread_create: %s\n",
2423                                                         strerror(ret));
2424                                         free(fd);
2425                                         nr_started--;
2426                                         break;
2427                                 }
2428                                 fd = NULL;
2429                                 ret = pthread_detach(td->thread);
2430                                 if (ret)
2431                                         log_err("pthread_detach: %s",
2432                                                         strerror(ret));
2433                         } else {
2434                                 pid_t pid;
2435                                 struct fio_file **files;
2436                                 void *eo;
2437                                 dprint(FD_PROCESS, "will fork\n");
2438                                 files = td->files;
2439                                 eo = td->eo;
2440                                 read_barrier();
2441                                 pid = fork();
2442                                 if (!pid) {
2443                                         int ret;
2444
2445                                         ret = (int)(uintptr_t)thread_main(fd);
2446                                         _exit(ret);
2447                                 } else if (i == fio_debug_jobno)
2448                                         *fio_debug_jobp = pid;
2449                                 // freeing previously allocated memory for files
2450                                 // this memory freed MUST NOT be shared between processes, only the pointer itself may be shared within TD
2451                                 free(files);
2452                                 free(eo);
2453                                 free(fd);
2454                                 fd = NULL;
2455                         }
2456                         dprint(FD_MUTEX, "wait on startup_sem\n");
2457                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2458                                 log_err("fio: job startup hung? exiting.\n");
2459                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2460                                 fio_abort = true;
2461                                 nr_started--;
2462                                 free(fd);
2463                                 break;
2464                         }
2465                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2466                 }
2467
2468                 /*
2469                  * Wait for the started threads to transition to
2470                  * TD_INITIALIZED.
2471                  */
2472                 fio_gettime(&this_start, NULL);
2473                 left = this_jobs;
2474                 while (left && !fio_abort) {
2475                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2476                                 break;
2477
2478                         do_usleep(100000);
2479
2480                         for (i = 0; i < this_jobs; i++) {
2481                                 td = map[i];
2482                                 if (!td)
2483                                         continue;
2484                                 if (td->runstate == TD_INITIALIZED) {
2485                                         map[i] = NULL;
2486                                         left--;
2487                                 } else if (td->runstate >= TD_EXITED) {
2488                                         map[i] = NULL;
2489                                         left--;
2490                                         todo--;
2491                                         nr_running++; /* work-around... */
2492                                 }
2493                         }
2494                 }
2495
2496                 if (left) {
2497                         log_err("fio: %d job%s failed to start\n", left,
2498                                         left > 1 ? "s" : "");
2499                         for (i = 0; i < this_jobs; i++) {
2500                                 td = map[i];
2501                                 if (!td)
2502                                         continue;
2503                                 kill(td->pid, SIGTERM);
2504                         }
2505                         break;
2506                 }
2507
2508                 /*
2509                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2510                  */
2511                 for_each_td(td, i) {
2512                         if (td->runstate != TD_INITIALIZED)
2513                                 continue;
2514
2515                         if (in_ramp_time(td))
2516                                 td_set_runstate(td, TD_RAMP);
2517                         else
2518                                 td_set_runstate(td, TD_RUNNING);
2519                         nr_running++;
2520                         nr_started--;
2521                         m_rate += ddir_rw_sum(td->o.ratemin);
2522                         t_rate += ddir_rw_sum(td->o.rate);
2523                         todo--;
2524                         fio_sem_up(td->sem);
2525                 }
2526
2527                 reap_threads(&nr_running, &t_rate, &m_rate);
2528
2529                 if (todo)
2530                         do_usleep(100000);
2531         }
2532
2533         while (nr_running) {
2534                 reap_threads(&nr_running, &t_rate, &m_rate);
2535                 do_usleep(10000);
2536         }
2537
2538         fio_idle_prof_stop();
2539
2540         update_io_ticks();
2541 }
2542
2543 static void free_disk_util(void)
2544 {
2545         disk_util_prune_entries();
2546         helper_thread_destroy();
2547 }
2548
2549 int fio_backend(struct sk_out *sk_out)
2550 {
2551         struct thread_data *td;
2552         int i;
2553
2554         if (exec_profile) {
2555                 if (load_profile(exec_profile))
2556                         return 1;
2557                 free(exec_profile);
2558                 exec_profile = NULL;
2559         }
2560         if (!thread_number)
2561                 return 0;
2562
2563         if (write_bw_log) {
2564                 struct log_params p = {
2565                         .log_type = IO_LOG_TYPE_BW,
2566                 };
2567
2568                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2569                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2570                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2571         }
2572
2573         if (init_global_dedupe_working_set_seeds()) {
2574                 log_err("fio: failed to initialize global dedupe working set\n");
2575                 return 1;
2576         }
2577
2578         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2579         if (!sk_out)
2580                 is_local_backend = true;
2581         if (startup_sem == NULL)
2582                 return 1;
2583
2584         set_genesis_time();
2585         stat_init();
2586         if (helper_thread_create(startup_sem, sk_out))
2587                 log_err("fio: failed to create helper thread\n");
2588
2589         cgroup_list = smalloc(sizeof(*cgroup_list));
2590         if (cgroup_list)
2591                 INIT_FLIST_HEAD(cgroup_list);
2592
2593         run_threads(sk_out);
2594
2595         helper_thread_exit();
2596
2597         if (!fio_abort) {
2598                 __show_run_stats();
2599                 if (write_bw_log) {
2600                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2601                                 struct io_log *log = agg_io_log[i];
2602
2603                                 flush_log(log, false);
2604                                 free_log(log);
2605                         }
2606                 }
2607         }
2608
2609         for_each_td(td, i) {
2610                 struct thread_stat *ts = &td->ts;
2611
2612                 free_clat_prio_stats(ts);
2613                 steadystate_free(td);
2614                 fio_options_free(td);
2615                 fio_dump_options_free(td);
2616                 if (td->rusage_sem) {
2617                         fio_sem_remove(td->rusage_sem);
2618                         td->rusage_sem = NULL;
2619                 }
2620                 fio_sem_remove(td->sem);
2621                 td->sem = NULL;
2622         }
2623
2624         free_disk_util();
2625         if (cgroup_list) {
2626                 cgroup_kill(cgroup_list);
2627                 sfree(cgroup_list);
2628         }
2629
2630         fio_sem_remove(startup_sem);
2631         stat_exit();
2632         return exit_value;
2633 }