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 #ifdef WIN32
  94 static void sig_break(int sig)
  95 {
  96         int i;
  97
  98         sig_int(sig);
  99
 100         /**
 101          * Windows terminates all job processes on SIGBREAK after the handler
 102          * returns, so give them time to wrap-up and give stats
 103          */
 104         for_each_td(td, i) {
 105                 while (td->runstate < TD_EXITED)
 106                         sleep(1);
 107         }
 108 }
 109 #endif
 110
 111 void sig_show_status(int sig)
 112 {
 113         show_running_run_stats();
 114 }
 115
 116 static void set_sig_handlers(void)
 117 {
 118         struct sigaction act;
 119
 120         memset(&act, 0, sizeof(act));
 121         act.sa_handler = sig_int;
 122         act.sa_flags = SA_RESTART;
 123         sigaction(SIGINT, &act, NULL);
 124
 125         memset(&act, 0, sizeof(act));
 126         act.sa_handler = sig_int;
 127         act.sa_flags = SA_RESTART;
 128         sigaction(SIGTERM, &act, NULL);
 129
 130 /* Windows uses SIGBREAK as a quit signal from other applications */
 131 #ifdef WIN32
 132         memset(&act, 0, sizeof(act));
 133         act.sa_handler = sig_break;
 134         act.sa_flags = SA_RESTART;
 135         sigaction(SIGBREAK, &act, NULL);
 136 #endif
 137
 138         memset(&act, 0, sizeof(act));
 139         act.sa_handler = sig_show_status;
 140         act.sa_flags = SA_RESTART;
 141         sigaction(SIGUSR1, &act, NULL);
 142
 143         if (is_backend) {
 144                 memset(&act, 0, sizeof(act));
 145                 act.sa_handler = sig_int;
 146                 act.sa_flags = SA_RESTART;
 147                 sigaction(SIGPIPE, &act, NULL);
 148         }
 149 }
 150
 151 /*
 152  * Check if we are above the minimum rate given.
 153  */
 154 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 155                              enum fio_ddir ddir)
 156 {
 157         unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
 158         unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
 159         unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
 160         unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
 161
 162         assert(ddir_rw(ddir));
 163
 164         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 165                 return false;
 166
 167         /*
 168          * allow a 2 second settle period in the beginning
 169          */
 170         if (mtime_since(&td->start, now) < 2000)
 171                 return false;
 172
 173         /*
 174          * if last_rate_check_blocks or last_rate_check_bytes is set,
 175          * we can compute a rate per ratecycle
 176          */
 177         if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
 178                 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
 179                 if (spent < td->o.ratecycle || spent==0)
 180                         return false;
 181
 182                 if (td->o.ratemin[ddir]) {
 183                         /*
 184                          * check bandwidth specified rate
 185                          */
 186                         unsigned long long current_rate_bytes =
 187                                 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
 188                         if (current_rate_bytes < option_rate_bytes_min) {
 189                                 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 190                                         td->o.name, option_rate_bytes_min, current_rate_bytes);
 191                                 return true;
 192                         }
 193                 } else {
 194                         /*
 195                          * checks iops specified rate
 196                          */
 197                         unsigned long long current_rate_iops =
 198                                 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
 199
 200                         if (current_rate_iops < option_rate_iops_min) {
 201                                 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 202                                         td->o.name, option_rate_iops_min, current_rate_iops);
 203                                 return true;
 204                         }
 205                 }
 206         }
 207
 208         td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
 209         td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
 210         memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
 211         return false;
 212 }
 213
 214 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 215 {
 216         bool ret = false;
 217
 218         for_each_rw_ddir(ddir) {
 219                 if (td->bytes_done[ddir])
 220                         ret |= __check_min_rate(td, now, ddir);
 221         }
 222
 223         return ret;
 224 }
 225
 226 /*
 227  * When job exits, we can cancel the in-flight IO if we are using async
 228  * io. Attempt to do so.
 229  */
 230 static void cleanup_pending_aio(struct thread_data *td)
 231 {
 232         int r;
 233
 234         /*
 235          * get immediately available events, if any
 236          */
 237         r = io_u_queued_complete(td, 0);
 238
 239         /*
 240          * now cancel remaining active events
 241          */
 242         if (td->io_ops->cancel) {
 243                 struct io_u *io_u;
 244                 int i;
 245
 246                 io_u_qiter(&td->io_u_all, io_u, i) {
 247                         if (io_u->flags & IO_U_F_FLIGHT) {
 248                                 r = td->io_ops->cancel(td, io_u);
 249                                 if (!r)
 250                                         put_io_u(td, io_u);
 251                         }
 252                 }
 253         }
 254
 255         if (td->cur_depth)
 256                 r = io_u_queued_complete(td, td->cur_depth);
 257 }
 258
 259 /*
 260  * Helper to handle the final sync of a file. Works just like the normal
 261  * io path, just does everything sync.
 262  */
 263 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 264 {
 265         struct io_u *io_u = __get_io_u(td);
 266         enum fio_q_status ret;
 267
 268         if (!io_u)
 269                 return true;
 270
 271         io_u->ddir = DDIR_SYNC;
 272         io_u->file = f;
 273         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 274
 275         if (td_io_prep(td, io_u)) {
 276                 put_io_u(td, io_u);
 277                 return true;
 278         }
 279
 280 requeue:
 281         ret = td_io_queue(td, io_u);
 282         switch (ret) {
 283         case FIO_Q_QUEUED:
 284                 td_io_commit(td);
 285                 if (io_u_queued_complete(td, 1) < 0)
 286                         return true;
 287                 break;
 288         case FIO_Q_COMPLETED:
 289                 if (io_u->error) {
 290                         td_verror(td, io_u->error, "td_io_queue");
 291                         return true;
 292                 }
 293
 294                 if (io_u_sync_complete(td, io_u) < 0)
 295                         return true;
 296                 break;
 297         case FIO_Q_BUSY:
 298                 td_io_commit(td);
 299                 goto requeue;
 300         }
 301
 302         return false;
 303 }
 304
 305 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 306 {
 307         int ret, ret2;
 308
 309         if (fio_file_open(f))
 310                 return fio_io_sync(td, f);
 311
 312         if (td_io_open_file(td, f))
 313                 return 1;
 314
 315         ret = fio_io_sync(td, f);
 316         ret2 = 0;
 317         if (fio_file_open(f))
 318                 ret2 = td_io_close_file(td, f);
 319         return (ret || ret2);
 320 }
 321
 322 static inline void __update_ts_cache(struct thread_data *td)
 323 {
 324         fio_gettime(&td->ts_cache, NULL);
 325 }
 326
 327 static inline void update_ts_cache(struct thread_data *td)
 328 {
 329         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 330                 __update_ts_cache(td);
 331 }
 332
 333 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 334 {
 335         if (in_ramp_time(td))
 336                 return false;
 337         if (!td->o.timeout)
 338                 return false;
 339         if (utime_since(&td->epoch, t) >= td->o.timeout)
 340                 return true;
 341
 342         return false;
 343 }
 344
 345 /*
 346  * We need to update the runtime consistently in ms, but keep a running
 347  * tally of the current elapsed time in microseconds for sub millisecond
 348  * updates.
 349  */
 350 static inline void update_runtime(struct thread_data *td,
 351                                   unsigned long long *elapsed_us,
 352                                   const enum fio_ddir ddir)
 353 {
 354         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 355                 return;
 356
 357         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 358         elapsed_us[ddir] += utime_since_now(&td->start);
 359         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 360 }
 361
 362 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 363                                 int *retptr)
 364 {
 365         int ret = *retptr;
 366
 367         if (ret < 0 || td->error) {
 368                 int err = td->error;
 369                 enum error_type_bit eb;
 370
 371                 if (ret < 0)
 372                         err = -ret;
 373
 374                 eb = td_error_type(ddir, err);
 375                 if (!(td->o.continue_on_error & (1 << eb)))
 376                         return true;
 377
 378                 if (td_non_fatal_error(td, eb, err)) {
 379                         /*
 380                          * Continue with the I/Os in case of
 381                          * a non fatal error.
 382                          */
 383                         update_error_count(td, err);
 384                         td_clear_error(td);
 385                         *retptr = 0;
 386                         return false;
 387                 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
 388                         /*
 389                          * We expect to hit this error if
 390                          * fill_device option is set.
 391                          */
 392                         td_clear_error(td);
 393                         fio_mark_td_terminate(td);
 394                         return true;
 395                 } else {
 396                         /*
 397                          * Stop the I/O in case of a fatal
 398                          * error.
 399                          */
 400                         update_error_count(td, err);
 401                         return true;
 402                 }
 403         }
 404
 405         return false;
 406 }
 407
 408 static void check_update_rusage(struct thread_data *td)
 409 {
 410         if (td->update_rusage) {
 411                 td->update_rusage = 0;
 412                 update_rusage_stat(td);
 413                 fio_sem_up(td->rusage_sem);
 414         }
 415 }
 416
 417 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 418 {
 419         const int full = queue_full(td);
 420         int min_evts = 0;
 421         int ret;
 422
 423         if (td->flags & TD_F_REGROW_LOGS)
 424                 return io_u_quiesce(td);
 425
 426         /*
 427          * if the queue is full, we MUST reap at least 1 event
 428          */
 429         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 430         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 431                 min_evts = 1;
 432
 433         if (time && should_check_rate(td))
 434                 fio_gettime(time, NULL);
 435
 436         do {
 437                 ret = io_u_queued_complete(td, min_evts);
 438                 if (ret < 0)
 439                         break;
 440         } while (full && (td->cur_depth > td->o.iodepth_low));
 441
 442         return ret;
 443 }
 444
 445 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 446                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 447                    struct timespec *comp_time)
 448 {
 449         switch (*ret) {
 450         case FIO_Q_COMPLETED:
 451                 if (io_u->error) {
 452                         *ret = -io_u->error;
 453                         clear_io_u(td, io_u);
 454                 } else if (io_u->resid) {
 455                         long long bytes = io_u->xfer_buflen - io_u->resid;
 456                         struct fio_file *f = io_u->file;
 457
 458                         if (bytes_issued)
 459                                 *bytes_issued += bytes;
 460
 461                         if (!from_verify)
 462                                 trim_io_piece(io_u);
 463
 464                         /*
 465                          * zero read, fail
 466                          */
 467                         if (!bytes) {
 468                                 if (!from_verify)
 469                                         unlog_io_piece(td, io_u);
 470                                 td_verror(td, EIO, "full resid");
 471                                 put_io_u(td, io_u);
 472                                 break;
 473                         }
 474
 475                         io_u->xfer_buflen = io_u->resid;
 476                         io_u->xfer_buf += bytes;
 477                         io_u->offset += bytes;
 478
 479                         if (ddir_rw(io_u->ddir))
 480                                 td->ts.short_io_u[io_u->ddir]++;
 481
 482                         if (io_u->offset == f->real_file_size)
 483                                 goto sync_done;
 484
 485                         requeue_io_u(td, &io_u);
 486                 } else {
 487 sync_done:
 488                         if (comp_time && should_check_rate(td))
 489                                 fio_gettime(comp_time, NULL);
 490
 491                         *ret = io_u_sync_complete(td, io_u);
 492                         if (*ret < 0)
 493                                 break;
 494                 }
 495
 496                 if (td->flags & TD_F_REGROW_LOGS)
 497                         regrow_logs(td);
 498
 499                 /*
 500                  * when doing I/O (not when verifying),
 501                  * check for any errors that are to be ignored
 502                  */
 503                 if (!from_verify)
 504                         break;
 505
 506                 return 0;
 507         case FIO_Q_QUEUED:
 508                 /*
 509                  * if the engine doesn't have a commit hook,
 510                  * the io_u is really queued. if it does have such
 511                  * a hook, it has to call io_u_queued() itself.
 512                  */
 513                 if (td->io_ops->commit == NULL)
 514                         io_u_queued(td, io_u);
 515                 if (bytes_issued)
 516                         *bytes_issued += io_u->xfer_buflen;
 517                 break;
 518         case FIO_Q_BUSY:
 519                 if (!from_verify)
 520                         unlog_io_piece(td, io_u);
 521                 requeue_io_u(td, &io_u);
 522                 td_io_commit(td);
 523                 break;
 524         default:
 525                 assert(*ret < 0);
 526                 td_verror(td, -(*ret), "td_io_queue");
 527                 break;
 528         }
 529
 530         if (break_on_this_error(td, ddir, ret))
 531                 return 1;
 532
 533         return 0;
 534 }
 535
 536 static inline bool io_in_polling(struct thread_data *td)
 537 {
 538         return !td->o.iodepth_batch_complete_min &&
 539                    !td->o.iodepth_batch_complete_max;
 540 }
 541 /*
 542  * Unlinks files from thread data fio_file structure
 543  */
 544 static int unlink_all_files(struct thread_data *td)
 545 {
 546         struct fio_file *f;
 547         unsigned int i;
 548         int ret = 0;
 549
 550         for_each_file(td, f, i) {
 551                 if (f->filetype != FIO_TYPE_FILE)
 552                         continue;
 553                 ret = td_io_unlink_file(td, f);
 554                 if (ret)
 555                         break;
 556         }
 557
 558         if (ret)
 559                 td_verror(td, ret, "unlink_all_files");
 560
 561         return ret;
 562 }
 563
 564 /*
 565  * Check if io_u will overlap an in-flight IO in the queue
 566  */
 567 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 568 {
 569         bool overlap;
 570         struct io_u *check_io_u;
 571         unsigned long long x1, x2, y1, y2;
 572         int i;
 573
 574         x1 = io_u->offset;
 575         x2 = io_u->offset + io_u->buflen;
 576         overlap = false;
 577         io_u_qiter(q, check_io_u, i) {
 578                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 579                         y1 = check_io_u->offset;
 580                         y2 = check_io_u->offset + check_io_u->buflen;
 581
 582                         if (x1 < y2 && y1 < x2) {
 583                                 overlap = true;
 584                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 585                                                 x1, io_u->buflen,
 586                                                 y1, check_io_u->buflen);
 587                                 break;
 588                         }
 589                 }
 590         }
 591
 592         return overlap;
 593 }
 594
 595 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 596 {
 597         /*
 598          * Check for overlap if the user asked us to, and we have
 599          * at least one IO in flight besides this one.
 600          */
 601         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 602             in_flight_overlap(&td->io_u_all, io_u))
 603                 return FIO_Q_BUSY;
 604
 605         return td_io_queue(td, io_u);
 606 }
 607
 608 /*
 609  * The main verify engine. Runs over the writes we previously submitted,
 610  * reads the blocks back in, and checks the crc/md5 of the data.
 611  */
 612 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 613 {
 614         struct fio_file *f;
 615         struct io_u *io_u;
 616         int ret, min_events;
 617         unsigned int i;
 618
 619         dprint(FD_VERIFY, "starting loop\n");
 620
 621         /*
 622          * sync io first and invalidate cache, to make sure we really
 623          * read from disk.
 624          */
 625         for_each_file(td, f, i) {
 626                 if (!fio_file_open(f))
 627                         continue;
 628                 if (fio_io_sync(td, f))
 629                         break;
 630                 if (file_invalidate_cache(td, f))
 631                         break;
 632         }
 633
 634         check_update_rusage(td);
 635
 636         if (td->error)
 637                 return;
 638
 639         td_set_runstate(td, TD_VERIFYING);
 640
 641         io_u = NULL;
 642         while (!td->terminate) {
 643                 enum fio_ddir ddir;
 644                 int full;
 645
 646                 update_ts_cache(td);
 647                 check_update_rusage(td);
 648
 649                 if (runtime_exceeded(td, &td->ts_cache)) {
 650                         __update_ts_cache(td);
 651                         if (runtime_exceeded(td, &td->ts_cache)) {
 652                                 fio_mark_td_terminate(td);
 653                                 break;
 654                         }
 655                 }
 656
 657                 if (flow_threshold_exceeded(td))
 658                         continue;
 659
 660                 if (!td->o.experimental_verify) {
 661                         io_u = __get_io_u(td);
 662                         if (!io_u)
 663                                 break;
 664
 665                         if (get_next_verify(td, io_u)) {
 666                                 put_io_u(td, io_u);
 667                                 break;
 668                         }
 669
 670                         if (td_io_prep(td, io_u)) {
 671                                 put_io_u(td, io_u);
 672                                 break;
 673                         }
 674                 } else {
 675                         if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
 676                                 break;
 677
 678                         while ((io_u = get_io_u(td)) != NULL) {
 679                                 if (IS_ERR_OR_NULL(io_u)) {
 680                                         io_u = NULL;
 681                                         ret = FIO_Q_BUSY;
 682                                         goto reap;
 683                                 }
 684
 685                                 /*
 686                                  * We are only interested in the places where
 687                                  * we wrote or trimmed IOs. Turn those into
 688                                  * reads for verification purposes.
 689                                  */
 690                                 if (io_u->ddir == DDIR_READ) {
 691                                         /*
 692                                          * Pretend we issued it for rwmix
 693                                          * accounting
 694                                          */
 695                                         td->io_issues[DDIR_READ]++;
 696                                         put_io_u(td, io_u);
 697                                         continue;
 698                                 } else if (io_u->ddir == DDIR_TRIM) {
 699                                         io_u->ddir = DDIR_READ;
 700                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 701                                         break;
 702                                 } else if (io_u->ddir == DDIR_WRITE) {
 703                                         io_u->ddir = DDIR_READ;
 704                                         io_u->numberio = td->verify_read_issues;
 705                                         td->verify_read_issues++;
 706                                         populate_verify_io_u(td, io_u);
 707                                         break;
 708                                 } else {
 709                                         put_io_u(td, io_u);
 710                                         continue;
 711                                 }
 712                         }
 713
 714                         if (!io_u)
 715                                 break;
 716                 }
 717
 718                 if (verify_state_should_stop(td, io_u)) {
 719                         put_io_u(td, io_u);
 720                         break;
 721                 }
 722
 723                 if (td->o.verify_async)
 724                         io_u->end_io = verify_io_u_async;
 725                 else
 726                         io_u->end_io = verify_io_u;
 727
 728                 ddir = io_u->ddir;
 729                 if (!td->o.disable_slat)
 730                         fio_gettime(&io_u->start_time, NULL);
 731
 732                 ret = io_u_submit(td, io_u);
 733
 734                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 735                         break;
 736
 737                 /*
 738                  * if we can queue more, do so. but check if there are
 739                  * completed io_u's first. Note that we can get BUSY even
 740                  * without IO queued, if the system is resource starved.
 741                  */
 742 reap:
 743                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 744                 if (full || io_in_polling(td))
 745                         ret = wait_for_completions(td, NULL);
 746
 747                 if (ret < 0)
 748                         break;
 749         }
 750
 751         check_update_rusage(td);
 752
 753         if (!td->error) {
 754                 min_events = td->cur_depth;
 755
 756                 if (min_events)
 757                         ret = io_u_queued_complete(td, min_events);
 758         } else
 759                 cleanup_pending_aio(td);
 760
 761         td_set_runstate(td, TD_RUNNING);
 762
 763         dprint(FD_VERIFY, "exiting loop\n");
 764 }
 765
 766 static bool exceeds_number_ios(struct thread_data *td)
 767 {
 768         unsigned long long number_ios;
 769
 770         if (!td->o.number_ios)
 771                 return false;
 772
 773         number_ios = ddir_rw_sum(td->io_blocks);
 774         number_ios += td->io_u_queued + td->io_u_in_flight;
 775
 776         return number_ios >= (td->o.number_ios * td->loops);
 777 }
 778
 779 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 780 {
 781         unsigned long long bytes, limit;
 782
 783         if (td_rw(td))
 784                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 785         else if (td_write(td))
 786                 bytes = this_bytes[DDIR_WRITE];
 787         else if (td_read(td))
 788                 bytes = this_bytes[DDIR_READ];
 789         else
 790                 bytes = this_bytes[DDIR_TRIM];
 791
 792         if (td->o.io_size)
 793                 limit = td->o.io_size;
 794         else
 795                 limit = td->o.size;
 796
 797         limit *= td->loops;
 798         return bytes >= limit || exceeds_number_ios(td);
 799 }
 800
 801 static bool io_issue_bytes_exceeded(struct thread_data *td)
 802 {
 803         return io_bytes_exceeded(td, td->io_issue_bytes);
 804 }
 805
 806 static bool io_complete_bytes_exceeded(struct thread_data *td)
 807 {
 808         return io_bytes_exceeded(td, td->this_io_bytes);
 809 }
 810
 811 /*
 812  * used to calculate the next io time for rate control
 813  *
 814  */
 815 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 816 {
 817         uint64_t bps = td->rate_bps[ddir];
 818
 819         assert(!(td->flags & TD_F_CHILD));
 820
 821         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 822                 uint64_t val, iops;
 823
 824                 iops = bps / td->o.min_bs[ddir];
 825                 val = (int64_t) (1000000 / iops) *
 826                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 827                 if (val) {
 828                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 829                                         (unsigned long long) 1000000 / val,
 830                                         ddir);
 831                 }
 832                 td->last_usec[ddir] += val;
 833                 return td->last_usec[ddir];
 834         } else if (bps) {
 835                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 836                 uint64_t secs = bytes / bps;
 837                 uint64_t remainder = bytes % bps;
 838
 839                 return remainder * 1000000 / bps + secs * 1000000;
 840         }
 841
 842         return 0;
 843 }
 844
 845 static void init_thinktime(struct thread_data *td)
 846 {
 847         if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
 848                 td->thinktime_blocks_counter = td->io_blocks;
 849         else
 850                 td->thinktime_blocks_counter = td->io_issues;
 851         td->last_thinktime = td->epoch;
 852         td->last_thinktime_blocks = 0;
 853 }
 854
 855 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
 856                              struct timespec *time)
 857 {
 858         unsigned long long b;
 859         unsigned long long runtime_left;
 860         uint64_t total;
 861         int left;
 862         struct timespec now;
 863         bool stall = false;
 864
 865         if (td->o.thinktime_iotime) {
 866                 fio_gettime(&now, NULL);
 867                 if (utime_since(&td->last_thinktime, &now)
 868                     >= td->o.thinktime_iotime) {
 869                         stall = true;
 870                 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
 871                         /*
 872                          * When thinktime_iotime is set and thinktime_blocks is
 873                          * not set, skip the thinktime_blocks check, since
 874                          * thinktime_blocks default value 1 does not work
 875                          * together with thinktime_iotime.
 876                          */
 877                         return;
 878                 }
 879
 880         }
 881
 882         b = ddir_rw_sum(td->thinktime_blocks_counter);
 883         if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
 884                 stall = true;
 885
 886         if (!stall)
 887                 return;
 888
 889         io_u_quiesce(td);
 890
 891         left = td->o.thinktime_spin;
 892         if (td->o.timeout) {
 893                 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
 894                 if (runtime_left < (unsigned long long)left)
 895                         left = runtime_left;
 896         }
 897
 898         total = 0;
 899         if (left)
 900                 total = usec_spin(left);
 901
 902         left = td->o.thinktime - total;
 903         if (td->o.timeout) {
 904                 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
 905                 if (runtime_left < (unsigned long long)left)
 906                         left = runtime_left;
 907         }
 908
 909         if (left)
 910                 total += usec_sleep(td, left);
 911
 912         /*
 913          * If we're ignoring thinktime for the rate, add the number of bytes
 914          * we would have done while sleeping, minus one block to ensure we
 915          * start issuing immediately after the sleep.
 916          */
 917         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 918                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 919                 uint64_t bs = td->o.min_bs[ddir];
 920                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 921                 uint64_t over;
 922
 923                 if (usperop <= total)
 924                         over = bs;
 925                 else
 926                         over = (usperop - total) / usperop * -bs;
 927
 928                 td->rate_io_issue_bytes[ddir] += (missed - over);
 929                 /* adjust for rate_process=poisson */
 930                 td->last_usec[ddir] += total;
 931         }
 932
 933         if (time && should_check_rate(td))
 934                 fio_gettime(time, NULL);
 935
 936         td->last_thinktime_blocks = b;
 937         if (td->o.thinktime_iotime) {
 938                 fio_gettime(&now, NULL);
 939                 td->last_thinktime = now;
 940         }
 941 }
 942
 943 /*
 944  * Main IO worker function. It retrieves io_u's to process and queues
 945  * and reaps them, checking for rate and errors along the way.
 946  *
 947  * Returns number of bytes written and trimmed.
 948  */
 949 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 950 {
 951         unsigned int i;
 952         int ret = 0;
 953         uint64_t total_bytes, bytes_issued = 0;
 954
 955         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 956                 bytes_done[i] = td->bytes_done[i];
 957
 958         if (in_ramp_time(td))
 959                 td_set_runstate(td, TD_RAMP);
 960         else
 961                 td_set_runstate(td, TD_RUNNING);
 962
 963         lat_target_init(td);
 964
 965         total_bytes = td->o.size;
 966         /*
 967         * Allow random overwrite workloads to write up to io_size
 968         * before starting verification phase as 'size' doesn't apply.
 969         */
 970         if (td_write(td) && td_random(td) && td->o.norandommap)
 971                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 972         /*
 973          * If verify_backlog is enabled, we'll run the verify in this
 974          * handler as well. For that case, we may need up to twice the
 975          * amount of bytes.
 976          */
 977         if (td->o.verify != VERIFY_NONE &&
 978            (td_write(td) && td->o.verify_backlog))
 979                 total_bytes += td->o.size;
 980
 981         /* In trimwrite mode, each byte is trimmed and then written, so
 982          * allow total_bytes or number of ios to be twice as big */
 983         if (td_trimwrite(td)) {
 984                 total_bytes += td->total_io_size;
 985                 td->o.number_ios *= 2;
 986         }
 987
 988         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 989                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 990                 td->o.time_based) {
 991                 struct timespec comp_time;
 992                 struct io_u *io_u;
 993                 int full;
 994                 enum fio_ddir ddir;
 995
 996                 check_update_rusage(td);
 997
 998                 if (td->terminate || td->done)
 999                         break;
1000
1001                 update_ts_cache(td);
1002
1003                 if (runtime_exceeded(td, &td->ts_cache)) {
1004                         __update_ts_cache(td);
1005                         if (runtime_exceeded(td, &td->ts_cache)) {
1006                                 fio_mark_td_terminate(td);
1007                                 break;
1008                         }
1009                 }
1010
1011                 if (flow_threshold_exceeded(td))
1012                         continue;
1013
1014                 /*
1015                  * Break if we exceeded the bytes. The exception is time
1016                  * based runs, but we still need to break out of the loop
1017                  * for those to run verification, if enabled.
1018                  * Jobs read from iolog do not use this stop condition.
1019                  */
1020                 if (bytes_issued >= total_bytes &&
1021                     !td->o.read_iolog_file &&
1022                     (!td->o.time_based ||
1023                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
1024                         break;
1025
1026                 io_u = get_io_u(td);
1027                 if (IS_ERR_OR_NULL(io_u)) {
1028                         int err = PTR_ERR(io_u);
1029
1030                         io_u = NULL;
1031                         ddir = DDIR_INVAL;
1032                         if (err == -EBUSY) {
1033                                 ret = FIO_Q_BUSY;
1034                                 goto reap;
1035                         }
1036                         if (td->o.latency_target)
1037                                 goto reap;
1038                         break;
1039                 }
1040
1041                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
1042                         if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
1043                                 io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
1044                                 io_u->numberio = td->io_issues[io_u->ddir];
1045                                 populate_verify_io_u(td, io_u);
1046                         }
1047                 }
1048
1049                 ddir = io_u->ddir;
1050
1051                 /*
1052                  * Add verification end_io handler if:
1053                  *      - Asked to verify (!td_rw(td))
1054                  *      - Or the io_u is from our verify list (mixed write/ver)
1055                  */
1056                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1057                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1058
1059                         if (verify_state_should_stop(td, io_u)) {
1060                                 put_io_u(td, io_u);
1061                                 break;
1062                         }
1063
1064                         if (td->o.verify_async)
1065                                 io_u->end_io = verify_io_u_async;
1066                         else
1067                                 io_u->end_io = verify_io_u;
1068                         td_set_runstate(td, TD_VERIFYING);
1069                 } else if (in_ramp_time(td))
1070                         td_set_runstate(td, TD_RAMP);
1071                 else
1072                         td_set_runstate(td, TD_RUNNING);
1073
1074                 /*
1075                  * Always log IO before it's issued, so we know the specific
1076                  * order of it. The logged unit will track when the IO has
1077                  * completed.
1078                  */
1079                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1080                     td->o.do_verify &&
1081                     td->o.verify != VERIFY_NONE &&
1082                     !td->o.experimental_verify)
1083                         log_io_piece(td, io_u);
1084
1085                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1086                         const unsigned long long blen = io_u->xfer_buflen;
1087                         const enum fio_ddir __ddir = acct_ddir(io_u);
1088
1089                         if (td->error)
1090                                 break;
1091
1092                         workqueue_enqueue(&td->io_wq, &io_u->work);
1093                         ret = FIO_Q_QUEUED;
1094
1095                         if (ddir_rw(__ddir)) {
1096                                 td->io_issues[__ddir]++;
1097                                 td->io_issue_bytes[__ddir] += blen;
1098                                 td->rate_io_issue_bytes[__ddir] += blen;
1099                         }
1100
1101                         if (should_check_rate(td)) {
1102                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1103                                 fio_gettime(&comp_time, NULL);
1104                         }
1105
1106                 } else {
1107                         ret = io_u_submit(td, io_u);
1108
1109                         if (should_check_rate(td))
1110                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1111
1112                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1113                                 break;
1114
1115                         /*
1116                          * See if we need to complete some commands. Note that
1117                          * we can get BUSY even without IO queued, if the
1118                          * system is resource starved.
1119                          */
1120 reap:
1121                         full = queue_full(td) ||
1122                                 (ret == FIO_Q_BUSY && td->cur_depth);
1123                         if (full || io_in_polling(td))
1124                                 ret = wait_for_completions(td, &comp_time);
1125                 }
1126                 if (ret < 0)
1127                         break;
1128
1129                 if (ddir_rw(ddir) && td->o.thinktime)
1130                         handle_thinktime(td, ddir, &comp_time);
1131
1132                 if (!ddir_rw_sum(td->bytes_done) &&
1133                     !td_ioengine_flagged(td, FIO_NOIO))
1134                         continue;
1135
1136                 if (!in_ramp_time(td) && should_check_rate(td)) {
1137                         if (check_min_rate(td, &comp_time)) {
1138                                 if (exitall_on_terminate || td->o.exitall_error)
1139                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1140                                 td_verror(td, EIO, "check_min_rate");
1141                                 break;
1142                         }
1143                 }
1144                 if (!in_ramp_time(td) && td->o.latency_target)
1145                         lat_target_check(td);
1146         }
1147
1148         check_update_rusage(td);
1149
1150         if (td->trim_entries)
1151                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1152
1153         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1154                 td->error = 0;
1155                 fio_mark_td_terminate(td);
1156         }
1157         if (!td->error) {
1158                 struct fio_file *f;
1159
1160                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1161                         workqueue_flush(&td->io_wq);
1162                         i = 0;
1163                 } else
1164                         i = td->cur_depth;
1165
1166                 if (i) {
1167                         ret = io_u_queued_complete(td, i);
1168                         if (td->o.fill_device &&
1169                             (td->error == ENOSPC || td->error == EDQUOT))
1170                                 td->error = 0;
1171                 }
1172
1173                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1174                         td_set_runstate(td, TD_FSYNCING);
1175
1176                         for_each_file(td, f, i) {
1177                                 if (!fio_file_fsync(td, f))
1178                                         continue;
1179
1180                                 log_err("fio: end_fsync failed for file %s\n",
1181                                                                 f->file_name);
1182                         }
1183                 }
1184         } else {
1185                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1186                         workqueue_flush(&td->io_wq);
1187                 cleanup_pending_aio(td);
1188         }
1189
1190         /*
1191          * stop job if we failed doing any IO
1192          */
1193         if (!ddir_rw_sum(td->this_io_bytes))
1194                 td->done = 1;
1195
1196         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1197                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1198 }
1199
1200 static void free_file_completion_logging(struct thread_data *td)
1201 {
1202         struct fio_file *f;
1203         unsigned int i;
1204
1205         for_each_file(td, f, i) {
1206                 if (!f->last_write_comp)
1207                         break;
1208                 sfree(f->last_write_comp);
1209         }
1210 }
1211
1212 static int init_file_completion_logging(struct thread_data *td,
1213                                         unsigned int depth)
1214 {
1215         struct fio_file *f;
1216         unsigned int i;
1217
1218         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1219                 return 0;
1220
1221         for_each_file(td, f, i) {
1222                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1223                 if (!f->last_write_comp)
1224                         goto cleanup;
1225         }
1226
1227         return 0;
1228
1229 cleanup:
1230         free_file_completion_logging(td);
1231         log_err("fio: failed to alloc write comp data\n");
1232         return 1;
1233 }
1234
1235 static void cleanup_io_u(struct thread_data *td)
1236 {
1237         struct io_u *io_u;
1238
1239         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1240
1241                 if (td->io_ops->io_u_free)
1242                         td->io_ops->io_u_free(td, io_u);
1243
1244                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1245         }
1246
1247         free_io_mem(td);
1248
1249         io_u_rexit(&td->io_u_requeues);
1250         io_u_qexit(&td->io_u_freelist, false);
1251         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1252
1253         free_file_completion_logging(td);
1254 }
1255
1256 static int init_io_u(struct thread_data *td)
1257 {
1258         struct io_u *io_u;
1259         int cl_align, i, max_units;
1260         int err;
1261
1262         max_units = td->o.iodepth;
1263
1264         err = 0;
1265         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1266         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1267         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1268
1269         if (err) {
1270                 log_err("fio: failed setting up IO queues\n");
1271                 return 1;
1272         }
1273
1274         cl_align = os_cache_line_size();
1275
1276         for (i = 0; i < max_units; i++) {
1277                 void *ptr;
1278
1279                 if (td->terminate)
1280                         return 1;
1281
1282                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1283                 if (!ptr) {
1284                         log_err("fio: unable to allocate aligned memory\n");
1285                         return 1;
1286                 }
1287
1288                 io_u = ptr;
1289                 memset(io_u, 0, sizeof(*io_u));
1290                 INIT_FLIST_HEAD(&io_u->verify_list);
1291                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1292
1293                 io_u->index = i;
1294                 io_u->flags = IO_U_F_FREE;
1295                 io_u_qpush(&td->io_u_freelist, io_u);
1296
1297                 /*
1298                  * io_u never leaves this stack, used for iteration of all
1299                  * io_u buffers.
1300                  */
1301                 io_u_qpush(&td->io_u_all, io_u);
1302
1303                 if (td->io_ops->io_u_init) {
1304                         int ret = td->io_ops->io_u_init(td, io_u);
1305
1306                         if (ret) {
1307                                 log_err("fio: failed to init engine data: %d\n", ret);
1308                                 return 1;
1309                         }
1310                 }
1311         }
1312
1313         if (init_io_u_buffers(td))
1314                 return 1;
1315
1316         if (init_file_completion_logging(td, max_units))
1317                 return 1;
1318
1319         return 0;
1320 }
1321
1322 int init_io_u_buffers(struct thread_data *td)
1323 {
1324         struct io_u *io_u;
1325         unsigned long long max_bs, min_write;
1326         int i, max_units;
1327         int data_xfer = 1;
1328         char *p;
1329
1330         max_units = td->o.iodepth;
1331         max_bs = td_max_bs(td);
1332         min_write = td->o.min_bs[DDIR_WRITE];
1333         td->orig_buffer_size = (unsigned long long) max_bs
1334                                         * (unsigned long long) max_units;
1335
1336         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1337                 data_xfer = 0;
1338
1339         /*
1340          * if we may later need to do address alignment, then add any
1341          * possible adjustment here so that we don't cause a buffer
1342          * overflow later. this adjustment may be too much if we get
1343          * lucky and the allocator gives us an aligned address.
1344          */
1345         if (td->o.odirect || td->o.mem_align ||
1346             td_ioengine_flagged(td, FIO_RAWIO))
1347                 td->orig_buffer_size += page_mask + td->o.mem_align;
1348
1349         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1350                 unsigned long long bs;
1351
1352                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1353                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1354         }
1355
1356         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1357                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1358                 return 1;
1359         }
1360
1361         if (data_xfer && allocate_io_mem(td))
1362                 return 1;
1363
1364         if (td->o.odirect || td->o.mem_align ||
1365             td_ioengine_flagged(td, FIO_RAWIO))
1366                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1367         else
1368                 p = td->orig_buffer;
1369
1370         for (i = 0; i < max_units; i++) {
1371                 io_u = td->io_u_all.io_us[i];
1372                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1373
1374                 if (data_xfer) {
1375                         io_u->buf = p;
1376                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1377
1378                         if (td_write(td))
1379                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1380                         if (td_write(td) && td->o.verify_pattern_bytes) {
1381                                 /*
1382                                  * Fill the buffer with the pattern if we are
1383                                  * going to be doing writes.
1384                                  */
1385                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1386                         }
1387                 }
1388                 p += max_bs;
1389         }
1390
1391         return 0;
1392 }
1393
1394 #ifdef FIO_HAVE_IOSCHED_SWITCH
1395 /*
1396  * These functions are Linux specific.
1397  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1398  */
1399 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1400 {
1401         char tmp[256], tmp2[128], *p;
1402         FILE *f;
1403         int ret;
1404
1405         assert(file->du && file->du->sysfs_root);
1406         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1407
1408         f = fopen(tmp, "r+");
1409         if (!f) {
1410                 if (errno == ENOENT) {
1411                         log_err("fio: os or kernel doesn't support IO scheduler"
1412                                 " switching\n");
1413                         return 0;
1414                 }
1415                 td_verror(td, errno, "fopen iosched");
1416                 return 1;
1417         }
1418
1419         /*
1420          * Set io scheduler.
1421          */
1422         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1423         if (ferror(f) || ret != 1) {
1424                 td_verror(td, errno, "fwrite");
1425                 fclose(f);
1426                 return 1;
1427         }
1428
1429         rewind(f);
1430
1431         /*
1432          * Read back and check that the selected scheduler is now the default.
1433          */
1434         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1435         if (ferror(f) || ret < 0) {
1436                 td_verror(td, errno, "fread");
1437                 fclose(f);
1438                 return 1;
1439         }
1440         tmp[ret] = '\0';
1441         /*
1442          * either a list of io schedulers or "none\n" is expected. Strip the
1443          * trailing newline.
1444          */
1445         p = tmp;
1446         strsep(&p, "\n");
1447
1448         /*
1449          * Write to "none" entry doesn't fail, so check the result here.
1450          */
1451         if (!strcmp(tmp, "none")) {
1452                 log_err("fio: io scheduler is not tunable\n");
1453                 fclose(f);
1454                 return 0;
1455         }
1456
1457         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1458         if (!strstr(tmp, tmp2)) {
1459                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1460                 td_verror(td, EINVAL, "iosched_switch");
1461                 fclose(f);
1462                 return 1;
1463         }
1464
1465         fclose(f);
1466         return 0;
1467 }
1468
1469 static int switch_ioscheduler(struct thread_data *td)
1470 {
1471         struct fio_file *f;
1472         unsigned int i;
1473         int ret = 0;
1474
1475         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1476                 return 0;
1477
1478         assert(td->files && td->files[0]);
1479
1480         for_each_file(td, f, i) {
1481
1482                 /* Only consider regular files and block device files */
1483                 switch (f->filetype) {
1484                 case FIO_TYPE_FILE:
1485                 case FIO_TYPE_BLOCK:
1486                         /*
1487                          * Make sure that the device hosting the file could
1488                          * be determined.
1489                          */
1490                         if (!f->du)
1491                                 continue;
1492                         break;
1493                 case FIO_TYPE_CHAR:
1494                 case FIO_TYPE_PIPE:
1495                 default:
1496                         continue;
1497                 }
1498
1499                 ret = set_ioscheduler(td, f);
1500                 if (ret)
1501                         return ret;
1502         }
1503
1504         return 0;
1505 }
1506
1507 #else
1508
1509 static int switch_ioscheduler(struct thread_data *td)
1510 {
1511         return 0;
1512 }
1513
1514 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1515
1516 static bool keep_running(struct thread_data *td)
1517 {
1518         unsigned long long limit;
1519
1520         if (td->done)
1521                 return false;
1522         if (td->terminate)
1523                 return false;
1524         if (td->o.time_based)
1525                 return true;
1526         if (td->o.loops) {
1527                 td->o.loops--;
1528                 return true;
1529         }
1530         if (exceeds_number_ios(td))
1531                 return false;
1532
1533         if (td->o.io_size)
1534                 limit = td->o.io_size;
1535         else
1536                 limit = td->o.size;
1537
1538         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1539                 uint64_t diff;
1540
1541                 /*
1542                  * If the difference is less than the maximum IO size, we
1543                  * are done.
1544                  */
1545                 diff = limit - ddir_rw_sum(td->io_bytes);
1546                 if (diff < td_max_bs(td))
1547                         return false;
1548
1549                 if (fio_files_done(td) && !td->o.io_size)
1550                         return false;
1551
1552                 return true;
1553         }
1554
1555         return false;
1556 }
1557
1558 static int exec_string(struct thread_options *o, const char *string,
1559                        const char *mode)
1560 {
1561         int ret;
1562         char *str;
1563
1564         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1565                 return -1;
1566
1567         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1568                  o->name, mode);
1569         ret = system(str);
1570         if (ret == -1)
1571                 log_err("fio: exec of cmd <%s> failed\n", str);
1572
1573         free(str);
1574         return ret;
1575 }
1576
1577 /*
1578  * Dry run to compute correct state of numberio for verification.
1579  */
1580 static uint64_t do_dry_run(struct thread_data *td)
1581 {
1582         td_set_runstate(td, TD_RUNNING);
1583
1584         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1585                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1586                 struct io_u *io_u;
1587                 int ret;
1588
1589                 if (td->terminate || td->done)
1590                         break;
1591
1592                 io_u = get_io_u(td);
1593                 if (IS_ERR_OR_NULL(io_u))
1594                         break;
1595
1596                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1597                 io_u->error = 0;
1598                 io_u->resid = 0;
1599                 if (ddir_rw(acct_ddir(io_u)))
1600                         td->io_issues[acct_ddir(io_u)]++;
1601                 if (ddir_rw(io_u->ddir)) {
1602                         io_u_mark_depth(td, 1);
1603                         td->ts.total_io_u[io_u->ddir]++;
1604                 }
1605
1606                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1607                     td->o.do_verify &&
1608                     td->o.verify != VERIFY_NONE &&
1609                     !td->o.experimental_verify)
1610                         log_io_piece(td, io_u);
1611
1612                 ret = io_u_sync_complete(td, io_u);
1613                 (void) ret;
1614         }
1615
1616         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1617 }
1618
1619 struct fork_data {
1620         struct thread_data *td;
1621         struct sk_out *sk_out;
1622 };
1623
1624 /*
1625  * Entry point for the thread based jobs. The process based jobs end up
1626  * here as well, after a little setup.
1627  */
1628 static void *thread_main(void *data)
1629 {
1630         struct fork_data *fd = data;
1631         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1632         struct thread_data *td = fd->td;
1633         struct thread_options *o = &td->o;
1634         struct sk_out *sk_out = fd->sk_out;
1635         uint64_t bytes_done[DDIR_RWDIR_CNT];
1636         int deadlock_loop_cnt;
1637         bool clear_state;
1638         int res, ret;
1639
1640         sk_out_assign(sk_out);
1641         free(fd);
1642
1643         if (!o->use_thread) {
1644                 setsid();
1645                 td->pid = getpid();
1646         } else
1647                 td->pid = gettid();
1648
1649         fio_local_clock_init();
1650
1651         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1652
1653         if (is_backend)
1654                 fio_server_send_start(td);
1655
1656         INIT_FLIST_HEAD(&td->io_log_list);
1657         INIT_FLIST_HEAD(&td->io_hist_list);
1658         INIT_FLIST_HEAD(&td->verify_list);
1659         INIT_FLIST_HEAD(&td->trim_list);
1660         td->io_hist_tree = RB_ROOT;
1661
1662         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1663         if (ret) {
1664                 td_verror(td, ret, "mutex_cond_init_pshared");
1665                 goto err;
1666         }
1667         ret = cond_init_pshared(&td->verify_cond);
1668         if (ret) {
1669                 td_verror(td, ret, "mutex_cond_pshared");
1670                 goto err;
1671         }
1672
1673         td_set_runstate(td, TD_INITIALIZED);
1674         dprint(FD_MUTEX, "up startup_sem\n");
1675         fio_sem_up(startup_sem);
1676         dprint(FD_MUTEX, "wait on td->sem\n");
1677         fio_sem_down(td->sem);
1678         dprint(FD_MUTEX, "done waiting on td->sem\n");
1679
1680         /*
1681          * A new gid requires privilege, so we need to do this before setting
1682          * the uid.
1683          */
1684         if (o->gid != -1U && setgid(o->gid)) {
1685                 td_verror(td, errno, "setgid");
1686                 goto err;
1687         }
1688         if (o->uid != -1U && setuid(o->uid)) {
1689                 td_verror(td, errno, "setuid");
1690                 goto err;
1691         }
1692
1693         td_zone_gen_index(td);
1694
1695         /*
1696          * Do this early, we don't want the compress threads to be limited
1697          * to the same CPUs as the IO workers. So do this before we set
1698          * any potential CPU affinity
1699          */
1700         if (iolog_compress_init(td, sk_out))
1701                 goto err;
1702
1703         /*
1704          * If we have a gettimeofday() thread, make sure we exclude that
1705          * thread from this job
1706          */
1707         if (o->gtod_cpu)
1708                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1709
1710         /*
1711          * Set affinity first, in case it has an impact on the memory
1712          * allocations.
1713          */
1714         if (fio_option_is_set(o, cpumask)) {
1715                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1716                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1717                         if (!ret) {
1718                                 log_err("fio: no CPUs set\n");
1719                                 log_err("fio: Try increasing number of available CPUs\n");
1720                                 td_verror(td, EINVAL, "cpus_split");
1721                                 goto err;
1722                         }
1723                 }
1724                 ret = fio_setaffinity(td->pid, o->cpumask);
1725                 if (ret == -1) {
1726                         td_verror(td, errno, "cpu_set_affinity");
1727                         goto err;
1728                 }
1729         }
1730
1731 #ifdef CONFIG_LIBNUMA
1732         /* numa node setup */
1733         if (fio_option_is_set(o, numa_cpunodes) ||
1734             fio_option_is_set(o, numa_memnodes)) {
1735                 struct bitmask *mask;
1736
1737                 if (numa_available() < 0) {
1738                         td_verror(td, errno, "Does not support NUMA API\n");
1739                         goto err;
1740                 }
1741
1742                 if (fio_option_is_set(o, numa_cpunodes)) {
1743                         mask = numa_parse_nodestring(o->numa_cpunodes);
1744                         ret = numa_run_on_node_mask(mask);
1745                         numa_free_nodemask(mask);
1746                         if (ret == -1) {
1747                                 td_verror(td, errno, \
1748                                         "numa_run_on_node_mask failed\n");
1749                                 goto err;
1750                         }
1751                 }
1752
1753                 if (fio_option_is_set(o, numa_memnodes)) {
1754                         mask = NULL;
1755                         if (o->numa_memnodes)
1756                                 mask = numa_parse_nodestring(o->numa_memnodes);
1757
1758                         switch (o->numa_mem_mode) {
1759                         case MPOL_INTERLEAVE:
1760                                 numa_set_interleave_mask(mask);
1761                                 break;
1762                         case MPOL_BIND:
1763                                 numa_set_membind(mask);
1764                                 break;
1765                         case MPOL_LOCAL:
1766                                 numa_set_localalloc();
1767                                 break;
1768                         case MPOL_PREFERRED:
1769                                 numa_set_preferred(o->numa_mem_prefer_node);
1770                                 break;
1771                         case MPOL_DEFAULT:
1772                         default:
1773                                 break;
1774                         }
1775
1776                         if (mask)
1777                                 numa_free_nodemask(mask);
1778
1779                 }
1780         }
1781 #endif
1782
1783         if (fio_pin_memory(td))
1784                 goto err;
1785
1786         /*
1787          * May alter parameters that init_io_u() will use, so we need to
1788          * do this first.
1789          */
1790         if (!init_iolog(td))
1791                 goto err;
1792
1793         /* ioprio_set() has to be done before td_io_init() */
1794         if (fio_option_is_set(o, ioprio) ||
1795             fio_option_is_set(o, ioprio_class)) {
1796                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1797                 if (ret == -1) {
1798                         td_verror(td, errno, "ioprio_set");
1799                         goto err;
1800                 }
1801                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1802                 td->ts.ioprio = td->ioprio;
1803         }
1804
1805         if (td_io_init(td))
1806                 goto err;
1807
1808         if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
1809                 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1810                          "are selected, queue depth will be capped at 1\n");
1811         }
1812
1813         if (init_io_u(td))
1814                 goto err;
1815
1816         if (td->io_ops->post_init && td->io_ops->post_init(td))
1817                 goto err;
1818
1819         if (o->verify_async && verify_async_init(td))
1820                 goto err;
1821
1822         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1823                 goto err;
1824
1825         errno = 0;
1826         if (nice(o->nice) == -1 && errno != 0) {
1827                 td_verror(td, errno, "nice");
1828                 goto err;
1829         }
1830
1831         if (o->ioscheduler && switch_ioscheduler(td))
1832                 goto err;
1833
1834         if (!o->create_serialize && setup_files(td))
1835                 goto err;
1836
1837         if (!init_random_map(td))
1838                 goto err;
1839
1840         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1841                 goto err;
1842
1843         if (o->pre_read && !pre_read_files(td))
1844                 goto err;
1845
1846         fio_verify_init(td);
1847
1848         if (rate_submit_init(td, sk_out))
1849                 goto err;
1850
1851         set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1852         fio_getrusage(&td->ru_start);
1853         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1854         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1855         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1856
1857         init_thinktime(td);
1858
1859         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1860                         o->ratemin[DDIR_TRIM]) {
1861                 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1862                                         sizeof(td->bw_sample_time));
1863                 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1864                                         sizeof(td->bw_sample_time));
1865                 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1866                                         sizeof(td->bw_sample_time));
1867         }
1868
1869         memset(bytes_done, 0, sizeof(bytes_done));
1870         clear_state = false;
1871
1872         while (keep_running(td)) {
1873                 uint64_t verify_bytes;
1874
1875                 fio_gettime(&td->start, NULL);
1876                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1877
1878                 if (clear_state) {
1879                         clear_io_state(td, 0);
1880
1881                         if (o->unlink_each_loop && unlink_all_files(td))
1882                                 break;
1883                 }
1884
1885                 prune_io_piece_log(td);
1886
1887                 if (td->o.verify_only && td_write(td))
1888                         verify_bytes = do_dry_run(td);
1889                 else {
1890                         if (!td->o.rand_repeatable)
1891                                 /* save verify rand state to replay hdr seeds later at verify */
1892                                 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
1893                         do_io(td, bytes_done);
1894                         if (!td->o.rand_repeatable)
1895                                 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
1896                         if (!ddir_rw_sum(bytes_done)) {
1897                                 fio_mark_td_terminate(td);
1898                                 verify_bytes = 0;
1899                         } else {
1900                                 verify_bytes = bytes_done[DDIR_WRITE] +
1901                                                 bytes_done[DDIR_TRIM];
1902                         }
1903                 }
1904
1905                 /*
1906                  * If we took too long to shut down, the main thread could
1907                  * already consider us reaped/exited. If that happens, break
1908                  * out and clean up.
1909                  */
1910                 if (td->runstate >= TD_EXITED)
1911                         break;
1912
1913                 clear_state = true;
1914
1915                 /*
1916                  * Make sure we've successfully updated the rusage stats
1917                  * before waiting on the stat mutex. Otherwise we could have
1918                  * the stat thread holding stat mutex and waiting for
1919                  * the rusage_sem, which would never get upped because
1920                  * this thread is waiting for the stat mutex.
1921                  */
1922                 deadlock_loop_cnt = 0;
1923                 do {
1924                         check_update_rusage(td);
1925                         if (!fio_sem_down_trylock(stat_sem))
1926                                 break;
1927                         usleep(1000);
1928                         if (deadlock_loop_cnt++ > 5000) {
1929                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1930                                 td->error = EDEADLK;
1931                                 goto err;
1932                         }
1933                 } while (1);
1934
1935                 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1936                         ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
1937                         update_runtime(td, elapsed_us, DDIR_READ);
1938                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1939                         update_runtime(td, elapsed_us, DDIR_WRITE);
1940                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1941                         update_runtime(td, elapsed_us, DDIR_TRIM);
1942                 fio_gettime(&td->start, NULL);
1943                 fio_sem_up(stat_sem);
1944
1945                 if (td->error || td->terminate)
1946                         break;
1947
1948                 if (!o->do_verify ||
1949                     o->verify == VERIFY_NONE ||
1950                     td_ioengine_flagged(td, FIO_UNIDIR))
1951                         continue;
1952
1953                 clear_io_state(td, 0);
1954
1955                 fio_gettime(&td->start, NULL);
1956
1957                 do_verify(td, verify_bytes);
1958
1959                 /*
1960                  * See comment further up for why this is done here.
1961                  */
1962                 check_update_rusage(td);
1963
1964                 fio_sem_down(stat_sem);
1965                 update_runtime(td, elapsed_us, DDIR_READ);
1966                 fio_gettime(&td->start, NULL);
1967                 fio_sem_up(stat_sem);
1968
1969                 if (td->error || td->terminate)
1970                         break;
1971         }
1972
1973         /*
1974          * Acquire this lock if we were doing overlap checking in
1975          * offload mode so that we don't clean up this job while
1976          * another thread is checking its io_u's for overlap
1977          */
1978         if (td_offload_overlap(td)) {
1979                 int res = pthread_mutex_lock(&overlap_check);
1980                 assert(res == 0);
1981         }
1982         td_set_runstate(td, TD_FINISHING);
1983         if (td_offload_overlap(td)) {
1984                 res = pthread_mutex_unlock(&overlap_check);
1985                 assert(res == 0);
1986         }
1987
1988         update_rusage_stat(td);
1989         td->ts.total_run_time = mtime_since_now(&td->epoch);
1990         for_each_rw_ddir(ddir) {
1991                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1992         }
1993
1994         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1995             (td->o.verify != VERIFY_NONE && td_write(td)))
1996                 verify_save_state(td->thread_number);
1997
1998         fio_unpin_memory(td);
1999
2000         td_writeout_logs(td, true);
2001
2002         iolog_compress_exit(td);
2003         rate_submit_exit(td);
2004
2005         if (o->exec_postrun)
2006                 exec_string(o, o->exec_postrun, "postrun");
2007
2008         if (exitall_on_terminate || (o->exitall_error && td->error))
2009                 fio_terminate_threads(td->groupid, td->o.exit_what);
2010
2011 err:
2012         if (td->error)
2013                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
2014                                                         td->verror);
2015
2016         if (o->verify_async)
2017                 verify_async_exit(td);
2018
2019         close_and_free_files(td);
2020         cleanup_io_u(td);
2021         close_ioengine(td);
2022         cgroup_shutdown(td, cgroup_mnt);
2023         verify_free_state(td);
2024         td_zone_free_index(td);
2025
2026         if (fio_option_is_set(o, cpumask)) {
2027                 ret = fio_cpuset_exit(&o->cpumask);
2028                 if (ret)
2029                         td_verror(td, ret, "fio_cpuset_exit");
2030         }
2031
2032         /*
2033          * do this very late, it will log file closing as well
2034          */
2035         if (o->write_iolog_file)
2036                 write_iolog_close(td);
2037         if (td->io_log_rfile)
2038                 fclose(td->io_log_rfile);
2039
2040         td_set_runstate(td, TD_EXITED);
2041
2042         /*
2043          * Do this last after setting our runstate to exited, so we
2044          * know that the stat thread is signaled.
2045          */
2046         check_update_rusage(td);
2047
2048         sk_out_drop();
2049         return (void *) (uintptr_t) td->error;
2050 }
2051
2052 /*
2053  * Run over the job map and reap the threads that have exited, if any.
2054  */
2055 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2056                          uint64_t *m_rate)
2057 {
2058         unsigned int cputhreads, realthreads, pending;
2059         int i, status, ret;
2060
2061         /*
2062          * reap exited threads (TD_EXITED -> TD_REAPED)
2063          */
2064         realthreads = pending = cputhreads = 0;
2065         for_each_td(td, i) {
2066                 int flags = 0;
2067
2068                 if (!strcmp(td->o.ioengine, "cpuio"))
2069                         cputhreads++;
2070                 else
2071                         realthreads++;
2072
2073                 if (!td->pid) {
2074                         pending++;
2075                         continue;
2076                 }
2077                 if (td->runstate == TD_REAPED)
2078                         continue;
2079                 if (td->o.use_thread) {
2080                         if (td->runstate == TD_EXITED) {
2081                                 td_set_runstate(td, TD_REAPED);
2082                                 goto reaped;
2083                         }
2084                         continue;
2085                 }
2086
2087                 flags = WNOHANG;
2088                 if (td->runstate == TD_EXITED)
2089                         flags = 0;
2090
2091                 /*
2092                  * check if someone quit or got killed in an unusual way
2093                  */
2094                 ret = waitpid(td->pid, &status, flags);
2095                 if (ret < 0) {
2096                         if (errno == ECHILD) {
2097                                 log_err("fio: pid=%d disappeared %d\n",
2098                                                 (int) td->pid, td->runstate);
2099                                 td->sig = ECHILD;
2100                                 td_set_runstate(td, TD_REAPED);
2101                                 goto reaped;
2102                         }
2103                         perror("waitpid");
2104                 } else if (ret == td->pid) {
2105                         if (WIFSIGNALED(status)) {
2106                                 int sig = WTERMSIG(status);
2107
2108                                 if (sig != SIGTERM && sig != SIGUSR2)
2109                                         log_err("fio: pid=%d, got signal=%d\n",
2110                                                         (int) td->pid, sig);
2111                                 td->sig = sig;
2112                                 td_set_runstate(td, TD_REAPED);
2113                                 goto reaped;
2114                         }
2115                         if (WIFEXITED(status)) {
2116                                 if (WEXITSTATUS(status) && !td->error)
2117                                         td->error = WEXITSTATUS(status);
2118
2119                                 td_set_runstate(td, TD_REAPED);
2120                                 goto reaped;
2121                         }
2122                 }
2123
2124                 /*
2125                  * If the job is stuck, do a forceful timeout of it and
2126                  * move on.
2127                  */
2128                 if (td->terminate &&
2129                     td->runstate < TD_FSYNCING &&
2130                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2131                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2132                                 "%lu seconds, it appears to be stuck. Doing "
2133                                 "forceful exit of this job.\n",
2134                                 td->o.name, td->runstate,
2135                                 (unsigned long) time_since_now(&td->terminate_time));
2136                         td_set_runstate(td, TD_REAPED);
2137                         goto reaped;
2138                 }
2139
2140                 /*
2141                  * thread is not dead, continue
2142                  */
2143                 pending++;
2144                 continue;
2145 reaped:
2146                 (*nr_running)--;
2147                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2148                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2149                 if (!td->pid)
2150                         pending--;
2151
2152                 if (td->error)
2153                         exit_value++;
2154
2155                 done_secs += mtime_since_now(&td->epoch) / 1000;
2156                 profile_td_exit(td);
2157                 flow_exit_job(td);
2158         }
2159
2160         if (*nr_running == cputhreads && !pending && realthreads)
2161                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2162 }
2163
2164 static bool __check_trigger_file(void)
2165 {
2166         struct stat sb;
2167
2168         if (!trigger_file)
2169                 return false;
2170
2171         if (stat(trigger_file, &sb))
2172                 return false;
2173
2174         if (unlink(trigger_file) < 0)
2175                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2176                                                         strerror(errno));
2177
2178         return true;
2179 }
2180
2181 static bool trigger_timedout(void)
2182 {
2183         if (trigger_timeout)
2184                 if (time_since_genesis() >= trigger_timeout) {
2185                         trigger_timeout = 0;
2186                         return true;
2187                 }
2188
2189         return false;
2190 }
2191
2192 void exec_trigger(const char *cmd)
2193 {
2194         int ret;
2195
2196         if (!cmd || cmd[0] == '\0')
2197                 return;
2198
2199         ret = system(cmd);
2200         if (ret == -1)
2201                 log_err("fio: failed executing %s trigger\n", cmd);
2202 }
2203
2204 void check_trigger_file(void)
2205 {
2206         if (__check_trigger_file() || trigger_timedout()) {
2207                 if (nr_clients)
2208                         fio_clients_send_trigger(trigger_remote_cmd);
2209                 else {
2210                         verify_save_state(IO_LIST_ALL);
2211                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2212                         exec_trigger(trigger_cmd);
2213                 }
2214         }
2215 }
2216
2217 static int fio_verify_load_state(struct thread_data *td)
2218 {
2219         int ret;
2220
2221         if (!td->o.verify_state)
2222                 return 0;
2223
2224         if (is_backend) {
2225                 void *data;
2226
2227                 ret = fio_server_get_verify_state(td->o.name,
2228                                         td->thread_number - 1, &data);
2229                 if (!ret)
2230                         verify_assign_state(td, data);
2231         } else {
2232                 char prefix[PATH_MAX];
2233
2234                 if (aux_path)
2235                         sprintf(prefix, "%s%clocal", aux_path,
2236                                         FIO_OS_PATH_SEPARATOR);
2237                 else
2238                         strcpy(prefix, "local");
2239                 ret = verify_load_state(td, prefix);
2240         }
2241
2242         return ret;
2243 }
2244
2245 static void do_usleep(unsigned int usecs)
2246 {
2247         check_for_running_stats();
2248         check_trigger_file();
2249         usleep(usecs);
2250 }
2251
2252 static bool check_mount_writes(struct thread_data *td)
2253 {
2254         struct fio_file *f;
2255         unsigned int i;
2256
2257         if (!td_write(td) || td->o.allow_mounted_write)
2258                 return false;
2259
2260         /*
2261          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2262          * are mkfs'd and mounted.
2263          */
2264         for_each_file(td, f, i) {
2265 #ifdef FIO_HAVE_CHARDEV_SIZE
2266                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2267 #else
2268                 if (f->filetype != FIO_TYPE_BLOCK)
2269 #endif
2270                         continue;
2271                 if (device_is_mounted(f->file_name))
2272                         goto mounted;
2273         }
2274
2275         return false;
2276 mounted:
2277         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2278         return true;
2279 }
2280
2281 static bool waitee_running(struct thread_data *me)
2282 {
2283         const char *waitee = me->o.wait_for;
2284         const char *self = me->o.name;
2285         int i;
2286
2287         if (!waitee)
2288                 return false;
2289
2290         for_each_td(td, i) {
2291                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2292                         continue;
2293
2294                 if (td->runstate < TD_EXITED) {
2295                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2296                                         self, td->o.name,
2297                                         runstate_to_name(td->runstate));
2298                         return true;
2299                 }
2300         }
2301
2302         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2303         return false;
2304 }
2305
2306 /*
2307  * Main function for kicking off and reaping jobs, as needed.
2308  */
2309 static void run_threads(struct sk_out *sk_out)
2310 {
2311         unsigned int i, todo, nr_running, nr_started;
2312         uint64_t m_rate, t_rate;
2313         uint64_t spent;
2314
2315         if (fio_gtod_offload && fio_start_gtod_thread())
2316                 return;
2317
2318         fio_idle_prof_init();
2319
2320         set_sig_handlers();
2321
2322         nr_thread = nr_process = 0;
2323         for_each_td(td, i) {
2324                 if (check_mount_writes(td))
2325                         return;
2326                 if (td->o.use_thread)
2327                         nr_thread++;
2328                 else
2329                         nr_process++;
2330         }
2331
2332         if (output_format & FIO_OUTPUT_NORMAL) {
2333                 struct buf_output out;
2334
2335                 buf_output_init(&out);
2336                 __log_buf(&out, "Starting ");
2337                 if (nr_thread)
2338                         __log_buf(&out, "%d thread%s", nr_thread,
2339                                                 nr_thread > 1 ? "s" : "");
2340                 if (nr_process) {
2341                         if (nr_thread)
2342                                 __log_buf(&out, " and ");
2343                         __log_buf(&out, "%d process%s", nr_process,
2344                                                 nr_process > 1 ? "es" : "");
2345                 }
2346                 __log_buf(&out, "\n");
2347                 log_info_buf(out.buf, out.buflen);
2348                 buf_output_free(&out);
2349         }
2350
2351         todo = thread_number;
2352         nr_running = 0;
2353         nr_started = 0;
2354         m_rate = t_rate = 0;
2355
2356         for_each_td(td, i) {
2357                 print_status_init(td->thread_number - 1);
2358
2359                 if (!td->o.create_serialize)
2360                         continue;
2361
2362                 if (fio_verify_load_state(td))
2363                         goto reap;
2364
2365                 /*
2366                  * do file setup here so it happens sequentially,
2367                  * we don't want X number of threads getting their
2368                  * client data interspersed on disk
2369                  */
2370                 if (setup_files(td)) {
2371 reap:
2372                         exit_value++;
2373                         if (td->error)
2374                                 log_err("fio: pid=%d, err=%d/%s\n",
2375                                         (int) td->pid, td->error, td->verror);
2376                         td_set_runstate(td, TD_REAPED);
2377                         todo--;
2378                 } else {
2379                         struct fio_file *f;
2380                         unsigned int j;
2381
2382                         /*
2383                          * for sharing to work, each job must always open
2384                          * its own files. so close them, if we opened them
2385                          * for creation
2386                          */
2387                         for_each_file(td, f, j) {
2388                                 if (fio_file_open(f))
2389                                         td_io_close_file(td, f);
2390                         }
2391                 }
2392         }
2393
2394         /* start idle threads before io threads start to run */
2395         fio_idle_prof_start();
2396
2397         set_genesis_time();
2398
2399         while (todo) {
2400                 struct thread_data *map[REAL_MAX_JOBS];
2401                 struct timespec this_start;
2402                 int this_jobs = 0, left;
2403                 struct fork_data *fd;
2404
2405                 /*
2406                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2407                  */
2408                 for_each_td(td, i) {
2409                         if (td->runstate != TD_NOT_CREATED)
2410                                 continue;
2411
2412                         /*
2413                          * never got a chance to start, killed by other
2414                          * thread for some reason
2415                          */
2416                         if (td->terminate) {
2417                                 todo--;
2418                                 continue;
2419                         }
2420
2421                         if (td->o.start_delay) {
2422                                 spent = utime_since_genesis();
2423
2424                                 if (td->o.start_delay > spent)
2425                                         continue;
2426                         }
2427
2428                         if (td->o.stonewall && (nr_started || nr_running)) {
2429                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2430                                                         td->o.name);
2431                                 break;
2432                         }
2433
2434                         if (waitee_running(td)) {
2435                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2436                                                 td->o.name, td->o.wait_for);
2437                                 continue;
2438                         }
2439
2440                         init_disk_util(td);
2441
2442                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2443                         td->update_rusage = 0;
2444
2445                         /*
2446                          * Set state to created. Thread will transition
2447                          * to TD_INITIALIZED when it's done setting up.
2448                          */
2449                         td_set_runstate(td, TD_CREATED);
2450                         map[this_jobs++] = td;
2451                         nr_started++;
2452
2453                         fd = calloc(1, sizeof(*fd));
2454                         fd->td = td;
2455                         fd->sk_out = sk_out;
2456
2457                         if (td->o.use_thread) {
2458                                 int ret;
2459
2460                                 dprint(FD_PROCESS, "will pthread_create\n");
2461                                 ret = pthread_create(&td->thread, NULL,
2462                                                         thread_main, fd);
2463                                 if (ret) {
2464                                         log_err("pthread_create: %s\n",
2465                                                         strerror(ret));
2466                                         free(fd);
2467                                         nr_started--;
2468                                         break;
2469                                 }
2470                                 fd = NULL;
2471                                 ret = pthread_detach(td->thread);
2472                                 if (ret)
2473                                         log_err("pthread_detach: %s",
2474                                                         strerror(ret));
2475                         } else {
2476                                 pid_t pid;
2477                                 void *eo;
2478                                 dprint(FD_PROCESS, "will fork\n");
2479                                 eo = td->eo;
2480                                 read_barrier();
2481                                 pid = fork();
2482                                 if (!pid) {
2483                                         int ret;
2484
2485                                         ret = (int)(uintptr_t)thread_main(fd);
2486                                         _exit(ret);
2487                                 } else if (i == fio_debug_jobno)
2488                                         *fio_debug_jobp = pid;
2489                                 free(eo);
2490                                 free(fd);
2491                                 fd = NULL;
2492                         }
2493                         dprint(FD_MUTEX, "wait on startup_sem\n");
2494                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2495                                 log_err("fio: job startup hung? exiting.\n");
2496                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2497                                 fio_abort = true;
2498                                 nr_started--;
2499                                 free(fd);
2500                                 break;
2501                         }
2502                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2503                 }
2504
2505                 /*
2506                  * Wait for the started threads to transition to
2507                  * TD_INITIALIZED.
2508                  */
2509                 fio_gettime(&this_start, NULL);
2510                 left = this_jobs;
2511                 while (left && !fio_abort) {
2512                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2513                                 break;
2514
2515                         do_usleep(100000);
2516
2517                         for (i = 0; i < this_jobs; i++) {
2518                                 struct thread_data *td = map[i];
2519                                 if (!td)
2520                                         continue;
2521                                 if (td->runstate == TD_INITIALIZED) {
2522                                         map[i] = NULL;
2523                                         left--;
2524                                 } else if (td->runstate >= TD_EXITED) {
2525                                         map[i] = NULL;
2526                                         left--;
2527                                         todo--;
2528                                         nr_running++; /* work-around... */
2529                                 }
2530                         }
2531                 }
2532
2533                 if (left) {
2534                         log_err("fio: %d job%s failed to start\n", left,
2535                                         left > 1 ? "s" : "");
2536                         for (i = 0; i < this_jobs; i++) {
2537                                 struct thread_data *td = map[i];
2538                                 if (!td)
2539                                         continue;
2540                                 kill(td->pid, SIGTERM);
2541                         }
2542                         break;
2543                 }
2544
2545                 /*
2546                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2547                  */
2548                 for_each_td(td, i) {
2549                         if (td->runstate != TD_INITIALIZED)
2550                                 continue;
2551
2552                         if (in_ramp_time(td))
2553                                 td_set_runstate(td, TD_RAMP);
2554                         else
2555                                 td_set_runstate(td, TD_RUNNING);
2556                         nr_running++;
2557                         nr_started--;
2558                         m_rate += ddir_rw_sum(td->o.ratemin);
2559                         t_rate += ddir_rw_sum(td->o.rate);
2560                         todo--;
2561                         fio_sem_up(td->sem);
2562                 }
2563
2564                 reap_threads(&nr_running, &t_rate, &m_rate);
2565
2566                 if (todo)
2567                         do_usleep(100000);
2568         }
2569
2570         while (nr_running) {
2571                 reap_threads(&nr_running, &t_rate, &m_rate);
2572                 do_usleep(10000);
2573         }
2574
2575         fio_idle_prof_stop();
2576
2577         update_io_ticks();
2578 }
2579
2580 static void free_disk_util(void)
2581 {
2582         disk_util_prune_entries();
2583         helper_thread_destroy();
2584 }
2585
2586 int fio_backend(struct sk_out *sk_out)
2587 {
2588         int i;
2589
2590         if (exec_profile) {
2591                 if (load_profile(exec_profile))
2592                         return 1;
2593                 free(exec_profile);
2594                 exec_profile = NULL;
2595         }
2596         if (!thread_number)
2597                 return 0;
2598
2599         if (write_bw_log) {
2600                 struct log_params p = {
2601                         .log_type = IO_LOG_TYPE_BW,
2602                 };
2603
2604                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2605                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2606                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2607         }
2608
2609         if (init_global_dedupe_working_set_seeds()) {
2610                 log_err("fio: failed to initialize global dedupe working set\n");
2611                 return 1;
2612         }
2613
2614         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2615         if (!sk_out)
2616                 is_local_backend = true;
2617         if (startup_sem == NULL)
2618                 return 1;
2619
2620         set_genesis_time();
2621         stat_init();
2622         if (helper_thread_create(startup_sem, sk_out))
2623                 log_err("fio: failed to create helper thread\n");
2624
2625         cgroup_list = smalloc(sizeof(*cgroup_list));
2626         if (cgroup_list)
2627                 INIT_FLIST_HEAD(cgroup_list);
2628
2629         run_threads(sk_out);
2630
2631         helper_thread_exit();
2632
2633         if (!fio_abort) {
2634                 __show_run_stats();
2635                 if (write_bw_log) {
2636                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2637                                 struct io_log *log = agg_io_log[i];
2638
2639                                 flush_log(log, false);
2640                                 free_log(log);
2641                         }
2642                 }
2643         }
2644
2645         for_each_td(td, i) {
2646                 struct thread_stat *ts = &td->ts;
2647
2648                 free_clat_prio_stats(ts);
2649                 steadystate_free(td);
2650                 fio_options_free(td);
2651                 fio_dump_options_free(td);
2652                 if (td->rusage_sem) {
2653                         fio_sem_remove(td->rusage_sem);
2654                         td->rusage_sem = NULL;
2655                 }
2656                 fio_sem_remove(td->sem);
2657                 td->sem = NULL;
2658         }
2659
2660         free_disk_util();
2661         if (cgroup_list) {
2662                 cgroup_kill(cgroup_list);
2663                 sfree(cgroup_list);
2664         }
2665
2666         fio_sem_remove(startup_sem);
2667         stat_exit();
2668         return exit_value;
2669 }