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 stat_number = 0;
  66 int shm_id = 0;
  67 int temp_stall_ts;
  68 unsigned long done_secs = 0;
  69 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  70
  71 #define JOB_START_TIMEOUT       (5 * 1000)
  72
  73 static void sig_int(int sig)
  74 {
  75         if (threads) {
  76                 if (is_backend)
  77                         fio_server_got_signal(sig);
  78                 else {
  79                         log_info("\nfio: terminating on signal %d\n", sig);
  80                         log_info_flush();
  81                         exit_value = 128;
  82                 }
  83
  84                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  85         }
  86 }
  87
  88 void sig_show_status(int sig)
  89 {
  90         show_running_run_stats();
  91 }
  92
  93 static void set_sig_handlers(void)
  94 {
  95         struct sigaction act;
  96
  97         memset(&act, 0, sizeof(act));
  98         act.sa_handler = sig_int;
  99         act.sa_flags = SA_RESTART;
 100         sigaction(SIGINT, &act, NULL);
 101
 102         memset(&act, 0, sizeof(act));
 103         act.sa_handler = sig_int;
 104         act.sa_flags = SA_RESTART;
 105         sigaction(SIGTERM, &act, NULL);
 106
 107 /* Windows uses SIGBREAK as a quit signal from other applications */
 108 #ifdef WIN32
 109         memset(&act, 0, sizeof(act));
 110         act.sa_handler = sig_int;
 111         act.sa_flags = SA_RESTART;
 112         sigaction(SIGBREAK, &act, NULL);
 113 #endif
 114
 115         memset(&act, 0, sizeof(act));
 116         act.sa_handler = sig_show_status;
 117         act.sa_flags = SA_RESTART;
 118         sigaction(SIGUSR1, &act, NULL);
 119
 120         if (is_backend) {
 121                 memset(&act, 0, sizeof(act));
 122                 act.sa_handler = sig_int;
 123                 act.sa_flags = SA_RESTART;
 124                 sigaction(SIGPIPE, &act, NULL);
 125         }
 126 }
 127
 128 /*
 129  * Check if we are above the minimum rate given.
 130  */
 131 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 132                              enum fio_ddir ddir)
 133 {
 134         unsigned long long bytes = 0;
 135         unsigned long iops = 0;
 136         unsigned long spent;
 137         unsigned long rate;
 138         unsigned int ratemin = 0;
 139         unsigned int rate_iops = 0;
 140         unsigned int rate_iops_min = 0;
 141
 142         assert(ddir_rw(ddir));
 143
 144         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 145                 return false;
 146
 147         /*
 148          * allow a 2 second settle period in the beginning
 149          */
 150         if (mtime_since(&td->start, now) < 2000)
 151                 return false;
 152
 153         iops += td->this_io_blocks[ddir];
 154         bytes += td->this_io_bytes[ddir];
 155         ratemin += td->o.ratemin[ddir];
 156         rate_iops += td->o.rate_iops[ddir];
 157         rate_iops_min += td->o.rate_iops_min[ddir];
 158
 159         /*
 160          * if rate blocks is set, sample is running
 161          */
 162         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 163                 spent = mtime_since(&td->lastrate[ddir], now);
 164                 if (spent < td->o.ratecycle)
 165                         return false;
 166
 167                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 168                         /*
 169                          * check bandwidth specified rate
 170                          */
 171                         if (bytes < td->rate_bytes[ddir]) {
 172                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 173                                         td->o.name, ratemin, bytes);
 174                                 return true;
 175                         } else {
 176                                 if (spent)
 177                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 178                                 else
 179                                         rate = 0;
 180
 181                                 if (rate < ratemin ||
 182                                     bytes < td->rate_bytes[ddir]) {
 183                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 184                                                 td->o.name, ratemin, rate);
 185                                         return true;
 186                                 }
 187                         }
 188                 } else {
 189                         /*
 190                          * checks iops specified rate
 191                          */
 192                         if (iops < rate_iops) {
 193                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 194                                                 td->o.name, rate_iops, iops);
 195                                 return true;
 196                         } else {
 197                                 if (spent)
 198                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 199                                 else
 200                                         rate = 0;
 201
 202                                 if (rate < rate_iops_min ||
 203                                     iops < td->rate_blocks[ddir]) {
 204                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 205                                                 td->o.name, rate_iops_min, rate);
 206                                         return true;
 207                                 }
 208                         }
 209                 }
 210         }
 211
 212         td->rate_bytes[ddir] = bytes;
 213         td->rate_blocks[ddir] = iops;
 214         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 215         return false;
 216 }
 217
 218 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 219 {
 220         bool ret = false;
 221
 222         if (td->bytes_done[DDIR_READ])
 223                 ret |= __check_min_rate(td, now, DDIR_READ);
 224         if (td->bytes_done[DDIR_WRITE])
 225                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 226         if (td->bytes_done[DDIR_TRIM])
 227                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 228
 229         return ret;
 230 }
 231
 232 /*
 233  * When job exits, we can cancel the in-flight IO if we are using async
 234  * io. Attempt to do so.
 235  */
 236 static void cleanup_pending_aio(struct thread_data *td)
 237 {
 238         int r;
 239
 240         /*
 241          * get immediately available events, if any
 242          */
 243         r = io_u_queued_complete(td, 0);
 244
 245         /*
 246          * now cancel remaining active events
 247          */
 248         if (td->io_ops->cancel) {
 249                 struct io_u *io_u;
 250                 int i;
 251
 252                 io_u_qiter(&td->io_u_all, io_u, i) {
 253                         if (io_u->flags & IO_U_F_FLIGHT) {
 254                                 r = td->io_ops->cancel(td, io_u);
 255                                 if (!r)
 256                                         put_io_u(td, io_u);
 257                         }
 258                 }
 259         }
 260
 261         if (td->cur_depth)
 262                 r = io_u_queued_complete(td, td->cur_depth);
 263 }
 264
 265 /*
 266  * Helper to handle the final sync of a file. Works just like the normal
 267  * io path, just does everything sync.
 268  */
 269 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 270 {
 271         struct io_u *io_u = __get_io_u(td);
 272         enum fio_q_status ret;
 273
 274         if (!io_u)
 275                 return true;
 276
 277         io_u->ddir = DDIR_SYNC;
 278         io_u->file = f;
 279         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 280
 281         if (td_io_prep(td, io_u)) {
 282                 put_io_u(td, io_u);
 283                 return true;
 284         }
 285
 286 requeue:
 287         ret = td_io_queue(td, io_u);
 288         switch (ret) {
 289         case FIO_Q_QUEUED:
 290                 td_io_commit(td);
 291                 if (io_u_queued_complete(td, 1) < 0)
 292                         return true;
 293                 break;
 294         case FIO_Q_COMPLETED:
 295                 if (io_u->error) {
 296                         td_verror(td, io_u->error, "td_io_queue");
 297                         return true;
 298                 }
 299
 300                 if (io_u_sync_complete(td, io_u) < 0)
 301                         return true;
 302                 break;
 303         case FIO_Q_BUSY:
 304                 td_io_commit(td);
 305                 goto requeue;
 306         }
 307
 308         return false;
 309 }
 310
 311 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 312 {
 313         int ret, ret2;
 314
 315         if (fio_file_open(f))
 316                 return fio_io_sync(td, f);
 317
 318         if (td_io_open_file(td, f))
 319                 return 1;
 320
 321         ret = fio_io_sync(td, f);
 322         ret2 = 0;
 323         if (fio_file_open(f))
 324                 ret2 = td_io_close_file(td, f);
 325         return (ret || ret2);
 326 }
 327
 328 static inline void __update_ts_cache(struct thread_data *td)
 329 {
 330         fio_gettime(&td->ts_cache, NULL);
 331 }
 332
 333 static inline void update_ts_cache(struct thread_data *td)
 334 {
 335         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 336                 __update_ts_cache(td);
 337 }
 338
 339 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 340 {
 341         if (in_ramp_time(td))
 342                 return false;
 343         if (!td->o.timeout)
 344                 return false;
 345         if (utime_since(&td->epoch, t) >= td->o.timeout)
 346                 return true;
 347
 348         return false;
 349 }
 350
 351 /*
 352  * We need to update the runtime consistently in ms, but keep a running
 353  * tally of the current elapsed time in microseconds for sub millisecond
 354  * updates.
 355  */
 356 static inline void update_runtime(struct thread_data *td,
 357                                   unsigned long long *elapsed_us,
 358                                   const enum fio_ddir ddir)
 359 {
 360         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 361                 return;
 362
 363         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 364         elapsed_us[ddir] += utime_since_now(&td->start);
 365         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 366 }
 367
 368 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 369                                 int *retptr)
 370 {
 371         int ret = *retptr;
 372
 373         if (ret < 0 || td->error) {
 374                 int err = td->error;
 375                 enum error_type_bit eb;
 376
 377                 if (ret < 0)
 378                         err = -ret;
 379
 380                 eb = td_error_type(ddir, err);
 381                 if (!(td->o.continue_on_error & (1 << eb)))
 382                         return true;
 383
 384                 if (td_non_fatal_error(td, eb, err)) {
 385                         /*
 386                          * Continue with the I/Os in case of
 387                          * a non fatal error.
 388                          */
 389                         update_error_count(td, err);
 390                         td_clear_error(td);
 391                         *retptr = 0;
 392                         return false;
 393                 } else if (td->o.fill_device && err == ENOSPC) {
 394                         /*
 395                          * We expect to hit this error if
 396                          * fill_device option is set.
 397                          */
 398                         td_clear_error(td);
 399                         fio_mark_td_terminate(td);
 400                         return true;
 401                 } else {
 402                         /*
 403                          * Stop the I/O in case of a fatal
 404                          * error.
 405                          */
 406                         update_error_count(td, err);
 407                         return true;
 408                 }
 409         }
 410
 411         return false;
 412 }
 413
 414 static void check_update_rusage(struct thread_data *td)
 415 {
 416         if (td->update_rusage) {
 417                 td->update_rusage = 0;
 418                 update_rusage_stat(td);
 419                 fio_sem_up(td->rusage_sem);
 420         }
 421 }
 422
 423 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 424 {
 425         const int full = queue_full(td);
 426         int min_evts = 0;
 427         int ret;
 428
 429         if (td->flags & TD_F_REGROW_LOGS)
 430                 return io_u_quiesce(td);
 431
 432         /*
 433          * if the queue is full, we MUST reap at least 1 event
 434          */
 435         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 436         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 437                 min_evts = 1;
 438
 439         if (time && __should_check_rate(td))
 440                 fio_gettime(time, NULL);
 441
 442         do {
 443                 ret = io_u_queued_complete(td, min_evts);
 444                 if (ret < 0)
 445                         break;
 446         } while (full && (td->cur_depth > td->o.iodepth_low));
 447
 448         return ret;
 449 }
 450
 451 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 452                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 453                    struct timespec *comp_time)
 454 {
 455         switch (*ret) {
 456         case FIO_Q_COMPLETED:
 457                 if (io_u->error) {
 458                         *ret = -io_u->error;
 459                         clear_io_u(td, io_u);
 460                 } else if (io_u->resid) {
 461                         long long bytes = io_u->xfer_buflen - io_u->resid;
 462                         struct fio_file *f = io_u->file;
 463
 464                         if (bytes_issued)
 465                                 *bytes_issued += bytes;
 466
 467                         if (!from_verify)
 468                                 trim_io_piece(io_u);
 469
 470                         /*
 471                          * zero read, fail
 472                          */
 473                         if (!bytes) {
 474                                 if (!from_verify)
 475                                         unlog_io_piece(td, io_u);
 476                                 td_verror(td, EIO, "full resid");
 477                                 put_io_u(td, io_u);
 478                                 break;
 479                         }
 480
 481                         io_u->xfer_buflen = io_u->resid;
 482                         io_u->xfer_buf += bytes;
 483                         io_u->offset += bytes;
 484
 485                         if (ddir_rw(io_u->ddir))
 486                                 td->ts.short_io_u[io_u->ddir]++;
 487
 488                         if (io_u->offset == f->real_file_size)
 489                                 goto sync_done;
 490
 491                         requeue_io_u(td, &io_u);
 492                 } else {
 493 sync_done:
 494                         if (comp_time && __should_check_rate(td))
 495                                 fio_gettime(comp_time, NULL);
 496
 497                         *ret = io_u_sync_complete(td, io_u);
 498                         if (*ret < 0)
 499                                 break;
 500                 }
 501
 502                 if (td->flags & TD_F_REGROW_LOGS)
 503                         regrow_logs(td);
 504
 505                 /*
 506                  * when doing I/O (not when verifying),
 507                  * check for any errors that are to be ignored
 508                  */
 509                 if (!from_verify)
 510                         break;
 511
 512                 return 0;
 513         case FIO_Q_QUEUED:
 514                 /*
 515                  * if the engine doesn't have a commit hook,
 516                  * the io_u is really queued. if it does have such
 517                  * a hook, it has to call io_u_queued() itself.
 518                  */
 519                 if (td->io_ops->commit == NULL)
 520                         io_u_queued(td, io_u);
 521                 if (bytes_issued)
 522                         *bytes_issued += io_u->xfer_buflen;
 523                 break;
 524         case FIO_Q_BUSY:
 525                 if (!from_verify)
 526                         unlog_io_piece(td, io_u);
 527                 requeue_io_u(td, &io_u);
 528                 td_io_commit(td);
 529                 break;
 530         default:
 531                 assert(*ret < 0);
 532                 td_verror(td, -(*ret), "td_io_queue");
 533                 break;
 534         }
 535
 536         if (break_on_this_error(td, ddir, ret))
 537                 return 1;
 538
 539         return 0;
 540 }
 541
 542 static inline bool io_in_polling(struct thread_data *td)
 543 {
 544         return !td->o.iodepth_batch_complete_min &&
 545                    !td->o.iodepth_batch_complete_max;
 546 }
 547 /*
 548  * Unlinks files from thread data fio_file structure
 549  */
 550 static int unlink_all_files(struct thread_data *td)
 551 {
 552         struct fio_file *f;
 553         unsigned int i;
 554         int ret = 0;
 555
 556         for_each_file(td, f, i) {
 557                 if (f->filetype != FIO_TYPE_FILE)
 558                         continue;
 559                 ret = td_io_unlink_file(td, f);
 560                 if (ret)
 561                         break;
 562         }
 563
 564         if (ret)
 565                 td_verror(td, ret, "unlink_all_files");
 566
 567         return ret;
 568 }
 569
 570 /*
 571  * Check if io_u will overlap an in-flight IO in the queue
 572  */
 573 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 574 {
 575         bool overlap;
 576         struct io_u *check_io_u;
 577         unsigned long long x1, x2, y1, y2;
 578         int i;
 579
 580         x1 = io_u->offset;
 581         x2 = io_u->offset + io_u->buflen;
 582         overlap = false;
 583         io_u_qiter(q, check_io_u, i) {
 584                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 585                         y1 = check_io_u->offset;
 586                         y2 = check_io_u->offset + check_io_u->buflen;
 587
 588                         if (x1 < y2 && y1 < x2) {
 589                                 overlap = true;
 590                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 591                                                 x1, io_u->buflen,
 592                                                 y1, check_io_u->buflen);
 593                                 break;
 594                         }
 595                 }
 596         }
 597
 598         return overlap;
 599 }
 600
 601 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 602 {
 603         /*
 604          * Check for overlap if the user asked us to, and we have
 605          * at least one IO in flight besides this one.
 606          */
 607         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 608             in_flight_overlap(&td->io_u_all, io_u))
 609                 return FIO_Q_BUSY;
 610
 611         return td_io_queue(td, io_u);
 612 }
 613
 614 /*
 615  * The main verify engine. Runs over the writes we previously submitted,
 616  * reads the blocks back in, and checks the crc/md5 of the data.
 617  */
 618 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 619 {
 620         struct fio_file *f;
 621         struct io_u *io_u;
 622         int ret, min_events;
 623         unsigned int i;
 624
 625         dprint(FD_VERIFY, "starting loop\n");
 626
 627         /*
 628          * sync io first and invalidate cache, to make sure we really
 629          * read from disk.
 630          */
 631         for_each_file(td, f, i) {
 632                 if (!fio_file_open(f))
 633                         continue;
 634                 if (fio_io_sync(td, f))
 635                         break;
 636                 if (file_invalidate_cache(td, f))
 637                         break;
 638         }
 639
 640         check_update_rusage(td);
 641
 642         if (td->error)
 643                 return;
 644
 645         /*
 646          * verify_state needs to be reset before verification
 647          * proceeds so that expected random seeds match actual
 648          * random seeds in headers. The main loop will reset
 649          * all random number generators if randrepeat is set.
 650          */
 651         if (!td->o.rand_repeatable)
 652                 td_fill_verify_state_seed(td);
 653
 654         td_set_runstate(td, TD_VERIFYING);
 655
 656         io_u = NULL;
 657         while (!td->terminate) {
 658                 enum fio_ddir ddir;
 659                 int full;
 660
 661                 update_ts_cache(td);
 662                 check_update_rusage(td);
 663
 664                 if (runtime_exceeded(td, &td->ts_cache)) {
 665                         __update_ts_cache(td);
 666                         if (runtime_exceeded(td, &td->ts_cache)) {
 667                                 fio_mark_td_terminate(td);
 668                                 break;
 669                         }
 670                 }
 671
 672                 if (flow_threshold_exceeded(td))
 673                         continue;
 674
 675                 if (!td->o.experimental_verify) {
 676                         io_u = __get_io_u(td);
 677                         if (!io_u)
 678                                 break;
 679
 680                         if (get_next_verify(td, io_u)) {
 681                                 put_io_u(td, io_u);
 682                                 break;
 683                         }
 684
 685                         if (td_io_prep(td, io_u)) {
 686                                 put_io_u(td, io_u);
 687                                 break;
 688                         }
 689                 } else {
 690                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 691                                 break;
 692
 693                         while ((io_u = get_io_u(td)) != NULL) {
 694                                 if (IS_ERR_OR_NULL(io_u)) {
 695                                         io_u = NULL;
 696                                         ret = FIO_Q_BUSY;
 697                                         goto reap;
 698                                 }
 699
 700                                 /*
 701                                  * We are only interested in the places where
 702                                  * we wrote or trimmed IOs. Turn those into
 703                                  * reads for verification purposes.
 704                                  */
 705                                 if (io_u->ddir == DDIR_READ) {
 706                                         /*
 707                                          * Pretend we issued it for rwmix
 708                                          * accounting
 709                                          */
 710                                         td->io_issues[DDIR_READ]++;
 711                                         put_io_u(td, io_u);
 712                                         continue;
 713                                 } else if (io_u->ddir == DDIR_TRIM) {
 714                                         io_u->ddir = DDIR_READ;
 715                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 716                                         break;
 717                                 } else if (io_u->ddir == DDIR_WRITE) {
 718                                         io_u->ddir = DDIR_READ;
 719                                         populate_verify_io_u(td, io_u);
 720                                         break;
 721                                 } else {
 722                                         put_io_u(td, io_u);
 723                                         continue;
 724                                 }
 725                         }
 726
 727                         if (!io_u)
 728                                 break;
 729                 }
 730
 731                 if (verify_state_should_stop(td, io_u)) {
 732                         put_io_u(td, io_u);
 733                         break;
 734                 }
 735
 736                 if (td->o.verify_async)
 737                         io_u->end_io = verify_io_u_async;
 738                 else
 739                         io_u->end_io = verify_io_u;
 740
 741                 ddir = io_u->ddir;
 742                 if (!td->o.disable_slat)
 743                         fio_gettime(&io_u->start_time, NULL);
 744
 745                 ret = io_u_submit(td, io_u);
 746
 747                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 748                         break;
 749
 750                 /*
 751                  * if we can queue more, do so. but check if there are
 752                  * completed io_u's first. Note that we can get BUSY even
 753                  * without IO queued, if the system is resource starved.
 754                  */
 755 reap:
 756                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 757                 if (full || io_in_polling(td))
 758                         ret = wait_for_completions(td, NULL);
 759
 760                 if (ret < 0)
 761                         break;
 762         }
 763
 764         check_update_rusage(td);
 765
 766         if (!td->error) {
 767                 min_events = td->cur_depth;
 768
 769                 if (min_events)
 770                         ret = io_u_queued_complete(td, min_events);
 771         } else
 772                 cleanup_pending_aio(td);
 773
 774         td_set_runstate(td, TD_RUNNING);
 775
 776         dprint(FD_VERIFY, "exiting loop\n");
 777 }
 778
 779 static bool exceeds_number_ios(struct thread_data *td)
 780 {
 781         unsigned long long number_ios;
 782
 783         if (!td->o.number_ios)
 784                 return false;
 785
 786         number_ios = ddir_rw_sum(td->io_blocks);
 787         number_ios += td->io_u_queued + td->io_u_in_flight;
 788
 789         return number_ios >= (td->o.number_ios * td->loops);
 790 }
 791
 792 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 793 {
 794         unsigned long long bytes, limit;
 795
 796         if (td_rw(td))
 797                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 798         else if (td_write(td))
 799                 bytes = this_bytes[DDIR_WRITE];
 800         else if (td_read(td))
 801                 bytes = this_bytes[DDIR_READ];
 802         else
 803                 bytes = this_bytes[DDIR_TRIM];
 804
 805         if (td->o.io_size)
 806                 limit = td->o.io_size;
 807         else
 808                 limit = td->o.size;
 809
 810         limit *= td->loops;
 811         return bytes >= limit || exceeds_number_ios(td);
 812 }
 813
 814 static bool io_issue_bytes_exceeded(struct thread_data *td)
 815 {
 816         return io_bytes_exceeded(td, td->io_issue_bytes);
 817 }
 818
 819 static bool io_complete_bytes_exceeded(struct thread_data *td)
 820 {
 821         return io_bytes_exceeded(td, td->this_io_bytes);
 822 }
 823
 824 /*
 825  * used to calculate the next io time for rate control
 826  *
 827  */
 828 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 829 {
 830         uint64_t bps = td->rate_bps[ddir];
 831
 832         assert(!(td->flags & TD_F_CHILD));
 833
 834         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 835                 uint64_t val, iops;
 836
 837                 iops = bps / td->o.bs[ddir];
 838                 val = (int64_t) (1000000 / iops) *
 839                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 840                 if (val) {
 841                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 842                                         (unsigned long long) 1000000 / val,
 843                                         ddir);
 844                 }
 845                 td->last_usec[ddir] += val;
 846                 return td->last_usec[ddir];
 847         } else if (bps) {
 848                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 849                 uint64_t secs = bytes / bps;
 850                 uint64_t remainder = bytes % bps;
 851
 852                 return remainder * 1000000 / bps + secs * 1000000;
 853         }
 854
 855         return 0;
 856 }
 857
 858 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 859 {
 860         unsigned long long b;
 861         uint64_t total;
 862         int left;
 863
 864         b = ddir_rw_sum(td->io_blocks);
 865         if (b % td->o.thinktime_blocks)
 866                 return;
 867
 868         io_u_quiesce(td);
 869
 870         total = 0;
 871         if (td->o.thinktime_spin)
 872                 total = usec_spin(td->o.thinktime_spin);
 873
 874         left = td->o.thinktime - total;
 875         if (left)
 876                 total += usec_sleep(td, left);
 877
 878         /*
 879          * If we're ignoring thinktime for the rate, add the number of bytes
 880          * we would have done while sleeping, minus one block to ensure we
 881          * start issuing immediately after the sleep.
 882          */
 883         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 884                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 885                 uint64_t bs = td->o.min_bs[ddir];
 886                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 887                 uint64_t over;
 888
 889                 if (usperop <= total)
 890                         over = bs;
 891                 else
 892                         over = (usperop - total) / usperop * -bs;
 893
 894                 td->rate_io_issue_bytes[ddir] += (missed - over);
 895                 /* adjust for rate_process=poisson */
 896                 td->last_usec[ddir] += total;
 897         }
 898 }
 899
 900 /*
 901  * Main IO worker function. It retrieves io_u's to process and queues
 902  * and reaps them, checking for rate and errors along the way.
 903  *
 904  * Returns number of bytes written and trimmed.
 905  */
 906 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 907 {
 908         unsigned int i;
 909         int ret = 0;
 910         uint64_t total_bytes, bytes_issued = 0;
 911
 912         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 913                 bytes_done[i] = td->bytes_done[i];
 914
 915         if (in_ramp_time(td))
 916                 td_set_runstate(td, TD_RAMP);
 917         else
 918                 td_set_runstate(td, TD_RUNNING);
 919
 920         lat_target_init(td);
 921
 922         total_bytes = td->o.size;
 923         /*
 924         * Allow random overwrite workloads to write up to io_size
 925         * before starting verification phase as 'size' doesn't apply.
 926         */
 927         if (td_write(td) && td_random(td) && td->o.norandommap)
 928                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 929         /*
 930          * If verify_backlog is enabled, we'll run the verify in this
 931          * handler as well. For that case, we may need up to twice the
 932          * amount of bytes.
 933          */
 934         if (td->o.verify != VERIFY_NONE &&
 935            (td_write(td) && td->o.verify_backlog))
 936                 total_bytes += td->o.size;
 937
 938         /* In trimwrite mode, each byte is trimmed and then written, so
 939          * allow total_bytes to be twice as big */
 940         if (td_trimwrite(td))
 941                 total_bytes += td->total_io_size;
 942
 943         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 944                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 945                 td->o.time_based) {
 946                 struct timespec comp_time;
 947                 struct io_u *io_u;
 948                 int full;
 949                 enum fio_ddir ddir;
 950
 951                 check_update_rusage(td);
 952
 953                 if (td->terminate || td->done)
 954                         break;
 955
 956                 update_ts_cache(td);
 957
 958                 if (runtime_exceeded(td, &td->ts_cache)) {
 959                         __update_ts_cache(td);
 960                         if (runtime_exceeded(td, &td->ts_cache)) {
 961                                 fio_mark_td_terminate(td);
 962                                 break;
 963                         }
 964                 }
 965
 966                 if (flow_threshold_exceeded(td))
 967                         continue;
 968
 969                 /*
 970                  * Break if we exceeded the bytes. The exception is time
 971                  * based runs, but we still need to break out of the loop
 972                  * for those to run verification, if enabled.
 973                  * Jobs read from iolog do not use this stop condition.
 974                  */
 975                 if (bytes_issued >= total_bytes &&
 976                     !td->o.read_iolog_file &&
 977                     (!td->o.time_based ||
 978                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 979                         break;
 980
 981                 io_u = get_io_u(td);
 982                 if (IS_ERR_OR_NULL(io_u)) {
 983                         int err = PTR_ERR(io_u);
 984
 985                         io_u = NULL;
 986                         ddir = DDIR_INVAL;
 987                         if (err == -EBUSY) {
 988                                 ret = FIO_Q_BUSY;
 989                                 goto reap;
 990                         }
 991                         if (td->o.latency_target)
 992                                 goto reap;
 993                         break;
 994                 }
 995
 996                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
 997                         populate_verify_io_u(td, io_u);
 998
 999                 ddir = io_u->ddir;
1000
1001                 /*
1002                  * Add verification end_io handler if:
1003                  *      - Asked to verify (!td_rw(td))
1004                  *      - Or the io_u is from our verify list (mixed write/ver)
1005                  */
1006                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1007                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1008
1009                         if (!td->o.verify_pattern_bytes) {
1010                                 io_u->rand_seed = __rand(&td->verify_state);
1011                                 if (sizeof(int) != sizeof(long *))
1012                                         io_u->rand_seed *= __rand(&td->verify_state);
1013                         }
1014
1015                         if (verify_state_should_stop(td, io_u)) {
1016                                 put_io_u(td, io_u);
1017                                 break;
1018                         }
1019
1020                         if (td->o.verify_async)
1021                                 io_u->end_io = verify_io_u_async;
1022                         else
1023                                 io_u->end_io = verify_io_u;
1024                         td_set_runstate(td, TD_VERIFYING);
1025                 } else if (in_ramp_time(td))
1026                         td_set_runstate(td, TD_RAMP);
1027                 else
1028                         td_set_runstate(td, TD_RUNNING);
1029
1030                 /*
1031                  * Always log IO before it's issued, so we know the specific
1032                  * order of it. The logged unit will track when the IO has
1033                  * completed.
1034                  */
1035                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1036                     td->o.do_verify &&
1037                     td->o.verify != VERIFY_NONE &&
1038                     !td->o.experimental_verify)
1039                         log_io_piece(td, io_u);
1040
1041                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1042                         const unsigned long long blen = io_u->xfer_buflen;
1043                         const enum fio_ddir __ddir = acct_ddir(io_u);
1044
1045                         if (td->error)
1046                                 break;
1047
1048                         workqueue_enqueue(&td->io_wq, &io_u->work);
1049                         ret = FIO_Q_QUEUED;
1050
1051                         if (ddir_rw(__ddir)) {
1052                                 td->io_issues[__ddir]++;
1053                                 td->io_issue_bytes[__ddir] += blen;
1054                                 td->rate_io_issue_bytes[__ddir] += blen;
1055                         }
1056
1057                         if (should_check_rate(td))
1058                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1059
1060                 } else {
1061                         ret = io_u_submit(td, io_u);
1062
1063                         if (should_check_rate(td))
1064                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1065
1066                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1067                                 break;
1068
1069                         /*
1070                          * See if we need to complete some commands. Note that
1071                          * we can get BUSY even without IO queued, if the
1072                          * system is resource starved.
1073                          */
1074 reap:
1075                         full = queue_full(td) ||
1076                                 (ret == FIO_Q_BUSY && td->cur_depth);
1077                         if (full || io_in_polling(td))
1078                                 ret = wait_for_completions(td, &comp_time);
1079                 }
1080                 if (ret < 0)
1081                         break;
1082                 if (!ddir_rw_sum(td->bytes_done) &&
1083                     !td_ioengine_flagged(td, FIO_NOIO))
1084                         continue;
1085
1086                 if (!in_ramp_time(td) && should_check_rate(td)) {
1087                         if (check_min_rate(td, &comp_time)) {
1088                                 if (exitall_on_terminate || td->o.exitall_error)
1089                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1090                                 td_verror(td, EIO, "check_min_rate");
1091                                 break;
1092                         }
1093                 }
1094                 if (!in_ramp_time(td) && td->o.latency_target)
1095                         lat_target_check(td);
1096
1097                 if (ddir_rw(ddir) && td->o.thinktime)
1098                         handle_thinktime(td, ddir);
1099         }
1100
1101         check_update_rusage(td);
1102
1103         if (td->trim_entries)
1104                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1105
1106         if (td->o.fill_device && td->error == ENOSPC) {
1107                 td->error = 0;
1108                 fio_mark_td_terminate(td);
1109         }
1110         if (!td->error) {
1111                 struct fio_file *f;
1112
1113                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1114                         workqueue_flush(&td->io_wq);
1115                         i = 0;
1116                 } else
1117                         i = td->cur_depth;
1118
1119                 if (i) {
1120                         ret = io_u_queued_complete(td, i);
1121                         if (td->o.fill_device && td->error == ENOSPC)
1122                                 td->error = 0;
1123                 }
1124
1125                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1126                         td_set_runstate(td, TD_FSYNCING);
1127
1128                         for_each_file(td, f, i) {
1129                                 if (!fio_file_fsync(td, f))
1130                                         continue;
1131
1132                                 log_err("fio: end_fsync failed for file %s\n",
1133                                                                 f->file_name);
1134                         }
1135                 }
1136         } else
1137                 cleanup_pending_aio(td);
1138
1139         /*
1140          * stop job if we failed doing any IO
1141          */
1142         if (!ddir_rw_sum(td->this_io_bytes))
1143                 td->done = 1;
1144
1145         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1146                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1147 }
1148
1149 static void free_file_completion_logging(struct thread_data *td)
1150 {
1151         struct fio_file *f;
1152         unsigned int i;
1153
1154         for_each_file(td, f, i) {
1155                 if (!f->last_write_comp)
1156                         break;
1157                 sfree(f->last_write_comp);
1158         }
1159 }
1160
1161 static int init_file_completion_logging(struct thread_data *td,
1162                                         unsigned int depth)
1163 {
1164         struct fio_file *f;
1165         unsigned int i;
1166
1167         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1168                 return 0;
1169
1170         for_each_file(td, f, i) {
1171                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1172                 if (!f->last_write_comp)
1173                         goto cleanup;
1174         }
1175
1176         return 0;
1177
1178 cleanup:
1179         free_file_completion_logging(td);
1180         log_err("fio: failed to alloc write comp data\n");
1181         return 1;
1182 }
1183
1184 static void cleanup_io_u(struct thread_data *td)
1185 {
1186         struct io_u *io_u;
1187
1188         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1189
1190                 if (td->io_ops->io_u_free)
1191                         td->io_ops->io_u_free(td, io_u);
1192
1193                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1194         }
1195
1196         free_io_mem(td);
1197
1198         io_u_rexit(&td->io_u_requeues);
1199         io_u_qexit(&td->io_u_freelist, false);
1200         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1201
1202         free_file_completion_logging(td);
1203 }
1204
1205 static int init_io_u(struct thread_data *td)
1206 {
1207         struct io_u *io_u;
1208         int cl_align, i, max_units;
1209         int err;
1210
1211         max_units = td->o.iodepth;
1212
1213         err = 0;
1214         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1215         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1216         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1217
1218         if (err) {
1219                 log_err("fio: failed setting up IO queues\n");
1220                 return 1;
1221         }
1222
1223         cl_align = os_cache_line_size();
1224
1225         for (i = 0; i < max_units; i++) {
1226                 void *ptr;
1227
1228                 if (td->terminate)
1229                         return 1;
1230
1231                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1232                 if (!ptr) {
1233                         log_err("fio: unable to allocate aligned memory\n");
1234                         return 1;
1235                 }
1236
1237                 io_u = ptr;
1238                 memset(io_u, 0, sizeof(*io_u));
1239                 INIT_FLIST_HEAD(&io_u->verify_list);
1240                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1241
1242                 io_u->index = i;
1243                 io_u->flags = IO_U_F_FREE;
1244                 io_u_qpush(&td->io_u_freelist, io_u);
1245
1246                 /*
1247                  * io_u never leaves this stack, used for iteration of all
1248                  * io_u buffers.
1249                  */
1250                 io_u_qpush(&td->io_u_all, io_u);
1251
1252                 if (td->io_ops->io_u_init) {
1253                         int ret = td->io_ops->io_u_init(td, io_u);
1254
1255                         if (ret) {
1256                                 log_err("fio: failed to init engine data: %d\n", ret);
1257                                 return 1;
1258                         }
1259                 }
1260         }
1261
1262         init_io_u_buffers(td);
1263
1264         if (init_file_completion_logging(td, max_units))
1265                 return 1;
1266
1267         return 0;
1268 }
1269
1270 int init_io_u_buffers(struct thread_data *td)
1271 {
1272         struct io_u *io_u;
1273         unsigned long long max_bs, min_write;
1274         int i, max_units;
1275         int data_xfer = 1;
1276         char *p;
1277
1278         max_units = td->o.iodepth;
1279         max_bs = td_max_bs(td);
1280         min_write = td->o.min_bs[DDIR_WRITE];
1281         td->orig_buffer_size = (unsigned long long) max_bs
1282                                         * (unsigned long long) max_units;
1283
1284         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1285                 data_xfer = 0;
1286
1287         /*
1288          * if we may later need to do address alignment, then add any
1289          * possible adjustment here so that we don't cause a buffer
1290          * overflow later. this adjustment may be too much if we get
1291          * lucky and the allocator gives us an aligned address.
1292          */
1293         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1294             td_ioengine_flagged(td, FIO_RAWIO))
1295                 td->orig_buffer_size += page_mask + td->o.mem_align;
1296
1297         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1298                 unsigned long long bs;
1299
1300                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1301                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1302         }
1303
1304         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1305                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1306                 return 1;
1307         }
1308
1309         if (data_xfer && allocate_io_mem(td))
1310                 return 1;
1311
1312         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1313             td_ioengine_flagged(td, FIO_RAWIO))
1314                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1315         else
1316                 p = td->orig_buffer;
1317
1318         for (i = 0; i < max_units; i++) {
1319                 io_u = td->io_u_all.io_us[i];
1320                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1321
1322                 if (data_xfer) {
1323                         io_u->buf = p;
1324                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1325
1326                         if (td_write(td))
1327                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1328                         if (td_write(td) && td->o.verify_pattern_bytes) {
1329                                 /*
1330                                  * Fill the buffer with the pattern if we are
1331                                  * going to be doing writes.
1332                                  */
1333                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1334                         }
1335                 }
1336                 p += max_bs;
1337         }
1338
1339         return 0;
1340 }
1341
1342 /*
1343  * This function is Linux specific.
1344  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1345  */
1346 static int switch_ioscheduler(struct thread_data *td)
1347 {
1348 #ifdef FIO_HAVE_IOSCHED_SWITCH
1349         char tmp[256], tmp2[128], *p;
1350         FILE *f;
1351         int ret;
1352
1353         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1354                 return 0;
1355
1356         assert(td->files && td->files[0]);
1357         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1358
1359         f = fopen(tmp, "r+");
1360         if (!f) {
1361                 if (errno == ENOENT) {
1362                         log_err("fio: os or kernel doesn't support IO scheduler"
1363                                 " switching\n");
1364                         return 0;
1365                 }
1366                 td_verror(td, errno, "fopen iosched");
1367                 return 1;
1368         }
1369
1370         /*
1371          * Set io scheduler.
1372          */
1373         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1374         if (ferror(f) || ret != 1) {
1375                 td_verror(td, errno, "fwrite");
1376                 fclose(f);
1377                 return 1;
1378         }
1379
1380         rewind(f);
1381
1382         /*
1383          * Read back and check that the selected scheduler is now the default.
1384          */
1385         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1386         if (ferror(f) || ret < 0) {
1387                 td_verror(td, errno, "fread");
1388                 fclose(f);
1389                 return 1;
1390         }
1391         tmp[ret] = '\0';
1392         /*
1393          * either a list of io schedulers or "none\n" is expected. Strip the
1394          * trailing newline.
1395          */
1396         p = tmp;
1397         strsep(&p, "\n");
1398
1399         /*
1400          * Write to "none" entry doesn't fail, so check the result here.
1401          */
1402         if (!strcmp(tmp, "none")) {
1403                 log_err("fio: io scheduler is not tunable\n");
1404                 fclose(f);
1405                 return 0;
1406         }
1407
1408         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1409         if (!strstr(tmp, tmp2)) {
1410                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1411                 td_verror(td, EINVAL, "iosched_switch");
1412                 fclose(f);
1413                 return 1;
1414         }
1415
1416         fclose(f);
1417         return 0;
1418 #else
1419         return 0;
1420 #endif
1421 }
1422
1423 static bool keep_running(struct thread_data *td)
1424 {
1425         unsigned long long limit;
1426
1427         if (td->done)
1428                 return false;
1429         if (td->terminate)
1430                 return false;
1431         if (td->o.time_based)
1432                 return true;
1433         if (td->o.loops) {
1434                 td->o.loops--;
1435                 return true;
1436         }
1437         if (exceeds_number_ios(td))
1438                 return false;
1439
1440         if (td->o.io_size)
1441                 limit = td->o.io_size;
1442         else
1443                 limit = td->o.size;
1444
1445         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1446                 uint64_t diff;
1447
1448                 /*
1449                  * If the difference is less than the maximum IO size, we
1450                  * are done.
1451                  */
1452                 diff = limit - ddir_rw_sum(td->io_bytes);
1453                 if (diff < td_max_bs(td))
1454                         return false;
1455
1456                 if (fio_files_done(td) && !td->o.io_size)
1457                         return false;
1458
1459                 return true;
1460         }
1461
1462         return false;
1463 }
1464
1465 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1466 {
1467         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 13 + 1;
1468         int ret;
1469         char *str;
1470
1471         str = malloc(newlen);
1472         sprintf(str, "%s > %s.%s.txt 2>&1", string, o->name, mode);
1473
1474         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1475         ret = system(str);
1476         if (ret == -1)
1477                 log_err("fio: exec of cmd <%s> failed\n", str);
1478
1479         free(str);
1480         return ret;
1481 }
1482
1483 /*
1484  * Dry run to compute correct state of numberio for verification.
1485  */
1486 static uint64_t do_dry_run(struct thread_data *td)
1487 {
1488         td_set_runstate(td, TD_RUNNING);
1489
1490         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1491                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1492                 struct io_u *io_u;
1493                 int ret;
1494
1495                 if (td->terminate || td->done)
1496                         break;
1497
1498                 io_u = get_io_u(td);
1499                 if (IS_ERR_OR_NULL(io_u))
1500                         break;
1501
1502                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1503                 io_u->error = 0;
1504                 io_u->resid = 0;
1505                 if (ddir_rw(acct_ddir(io_u)))
1506                         td->io_issues[acct_ddir(io_u)]++;
1507                 if (ddir_rw(io_u->ddir)) {
1508                         io_u_mark_depth(td, 1);
1509                         td->ts.total_io_u[io_u->ddir]++;
1510                 }
1511
1512                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1513                     td->o.do_verify &&
1514                     td->o.verify != VERIFY_NONE &&
1515                     !td->o.experimental_verify)
1516                         log_io_piece(td, io_u);
1517
1518                 ret = io_u_sync_complete(td, io_u);
1519                 (void) ret;
1520         }
1521
1522         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1523 }
1524
1525 struct fork_data {
1526         struct thread_data *td;
1527         struct sk_out *sk_out;
1528 };
1529
1530 /*
1531  * Entry point for the thread based jobs. The process based jobs end up
1532  * here as well, after a little setup.
1533  */
1534 static void *thread_main(void *data)
1535 {
1536         struct fork_data *fd = data;
1537         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1538         struct thread_data *td = fd->td;
1539         struct thread_options *o = &td->o;
1540         struct sk_out *sk_out = fd->sk_out;
1541         uint64_t bytes_done[DDIR_RWDIR_CNT];
1542         int deadlock_loop_cnt;
1543         bool clear_state;
1544         int ret;
1545
1546         sk_out_assign(sk_out);
1547         free(fd);
1548
1549         if (!o->use_thread) {
1550                 setsid();
1551                 td->pid = getpid();
1552         } else
1553                 td->pid = gettid();
1554
1555         fio_local_clock_init();
1556
1557         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1558
1559         if (is_backend)
1560                 fio_server_send_start(td);
1561
1562         INIT_FLIST_HEAD(&td->io_log_list);
1563         INIT_FLIST_HEAD(&td->io_hist_list);
1564         INIT_FLIST_HEAD(&td->verify_list);
1565         INIT_FLIST_HEAD(&td->trim_list);
1566         td->io_hist_tree = RB_ROOT;
1567
1568         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1569         if (ret) {
1570                 td_verror(td, ret, "mutex_cond_init_pshared");
1571                 goto err;
1572         }
1573         ret = cond_init_pshared(&td->verify_cond);
1574         if (ret) {
1575                 td_verror(td, ret, "mutex_cond_pshared");
1576                 goto err;
1577         }
1578
1579         td_set_runstate(td, TD_INITIALIZED);
1580         dprint(FD_MUTEX, "up startup_sem\n");
1581         fio_sem_up(startup_sem);
1582         dprint(FD_MUTEX, "wait on td->sem\n");
1583         fio_sem_down(td->sem);
1584         dprint(FD_MUTEX, "done waiting on td->sem\n");
1585
1586         /*
1587          * A new gid requires privilege, so we need to do this before setting
1588          * the uid.
1589          */
1590         if (o->gid != -1U && setgid(o->gid)) {
1591                 td_verror(td, errno, "setgid");
1592                 goto err;
1593         }
1594         if (o->uid != -1U && setuid(o->uid)) {
1595                 td_verror(td, errno, "setuid");
1596                 goto err;
1597         }
1598
1599         td_zone_gen_index(td);
1600
1601         /*
1602          * Do this early, we don't want the compress threads to be limited
1603          * to the same CPUs as the IO workers. So do this before we set
1604          * any potential CPU affinity
1605          */
1606         if (iolog_compress_init(td, sk_out))
1607                 goto err;
1608
1609         /*
1610          * If we have a gettimeofday() thread, make sure we exclude that
1611          * thread from this job
1612          */
1613         if (o->gtod_cpu)
1614                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1615
1616         /*
1617          * Set affinity first, in case it has an impact on the memory
1618          * allocations.
1619          */
1620         if (fio_option_is_set(o, cpumask)) {
1621                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1622                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1623                         if (!ret) {
1624                                 log_err("fio: no CPUs set\n");
1625                                 log_err("fio: Try increasing number of available CPUs\n");
1626                                 td_verror(td, EINVAL, "cpus_split");
1627                                 goto err;
1628                         }
1629                 }
1630                 ret = fio_setaffinity(td->pid, o->cpumask);
1631                 if (ret == -1) {
1632                         td_verror(td, errno, "cpu_set_affinity");
1633                         goto err;
1634                 }
1635         }
1636
1637 #ifdef CONFIG_LIBNUMA
1638         /* numa node setup */
1639         if (fio_option_is_set(o, numa_cpunodes) ||
1640             fio_option_is_set(o, numa_memnodes)) {
1641                 struct bitmask *mask;
1642
1643                 if (numa_available() < 0) {
1644                         td_verror(td, errno, "Does not support NUMA API\n");
1645                         goto err;
1646                 }
1647
1648                 if (fio_option_is_set(o, numa_cpunodes)) {
1649                         mask = numa_parse_nodestring(o->numa_cpunodes);
1650                         ret = numa_run_on_node_mask(mask);
1651                         numa_free_nodemask(mask);
1652                         if (ret == -1) {
1653                                 td_verror(td, errno, \
1654                                         "numa_run_on_node_mask failed\n");
1655                                 goto err;
1656                         }
1657                 }
1658
1659                 if (fio_option_is_set(o, numa_memnodes)) {
1660                         mask = NULL;
1661                         if (o->numa_memnodes)
1662                                 mask = numa_parse_nodestring(o->numa_memnodes);
1663
1664                         switch (o->numa_mem_mode) {
1665                         case MPOL_INTERLEAVE:
1666                                 numa_set_interleave_mask(mask);
1667                                 break;
1668                         case MPOL_BIND:
1669                                 numa_set_membind(mask);
1670                                 break;
1671                         case MPOL_LOCAL:
1672                                 numa_set_localalloc();
1673                                 break;
1674                         case MPOL_PREFERRED:
1675                                 numa_set_preferred(o->numa_mem_prefer_node);
1676                                 break;
1677                         case MPOL_DEFAULT:
1678                         default:
1679                                 break;
1680                         }
1681
1682                         if (mask)
1683                                 numa_free_nodemask(mask);
1684
1685                 }
1686         }
1687 #endif
1688
1689         if (fio_pin_memory(td))
1690                 goto err;
1691
1692         /*
1693          * May alter parameters that init_io_u() will use, so we need to
1694          * do this first.
1695          */
1696         if (!init_iolog(td))
1697                 goto err;
1698
1699         if (td_io_init(td))
1700                 goto err;
1701
1702         if (init_io_u(td))
1703                 goto err;
1704
1705         if (td->io_ops->post_init && td->io_ops->post_init(td))
1706                 goto err;
1707
1708         if (o->verify_async && verify_async_init(td))
1709                 goto err;
1710
1711         if (fio_option_is_set(o, ioprio) ||
1712             fio_option_is_set(o, ioprio_class)) {
1713                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1714                 if (ret == -1) {
1715                         td_verror(td, errno, "ioprio_set");
1716                         goto err;
1717                 }
1718         }
1719
1720         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1721                 goto err;
1722
1723         errno = 0;
1724         if (nice(o->nice) == -1 && errno != 0) {
1725                 td_verror(td, errno, "nice");
1726                 goto err;
1727         }
1728
1729         if (o->ioscheduler && switch_ioscheduler(td))
1730                 goto err;
1731
1732         if (!o->create_serialize && setup_files(td))
1733                 goto err;
1734
1735         if (!init_random_map(td))
1736                 goto err;
1737
1738         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1739                 goto err;
1740
1741         if (o->pre_read && !pre_read_files(td))
1742                 goto err;
1743
1744         fio_verify_init(td);
1745
1746         if (rate_submit_init(td, sk_out))
1747                 goto err;
1748
1749         set_epoch_time(td, o->log_unix_epoch);
1750         fio_getrusage(&td->ru_start);
1751         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1752         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1753         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1754
1755         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1756                         o->ratemin[DDIR_TRIM]) {
1757                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1758                                         sizeof(td->bw_sample_time));
1759                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1760                                         sizeof(td->bw_sample_time));
1761                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1762                                         sizeof(td->bw_sample_time));
1763         }
1764
1765         memset(bytes_done, 0, sizeof(bytes_done));
1766         clear_state = false;
1767
1768         while (keep_running(td)) {
1769                 uint64_t verify_bytes;
1770
1771                 fio_gettime(&td->start, NULL);
1772                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1773
1774                 if (clear_state) {
1775                         clear_io_state(td, 0);
1776
1777                         if (o->unlink_each_loop && unlink_all_files(td))
1778                                 break;
1779                 }
1780
1781                 prune_io_piece_log(td);
1782
1783                 if (td->o.verify_only && td_write(td))
1784                         verify_bytes = do_dry_run(td);
1785                 else {
1786                         do_io(td, bytes_done);
1787
1788                         if (!ddir_rw_sum(bytes_done)) {
1789                                 fio_mark_td_terminate(td);
1790                                 verify_bytes = 0;
1791                         } else {
1792                                 verify_bytes = bytes_done[DDIR_WRITE] +
1793                                                 bytes_done[DDIR_TRIM];
1794                         }
1795                 }
1796
1797                 /*
1798                  * If we took too long to shut down, the main thread could
1799                  * already consider us reaped/exited. If that happens, break
1800                  * out and clean up.
1801                  */
1802                 if (td->runstate >= TD_EXITED)
1803                         break;
1804
1805                 clear_state = true;
1806
1807                 /*
1808                  * Make sure we've successfully updated the rusage stats
1809                  * before waiting on the stat mutex. Otherwise we could have
1810                  * the stat thread holding stat mutex and waiting for
1811                  * the rusage_sem, which would never get upped because
1812                  * this thread is waiting for the stat mutex.
1813                  */
1814                 deadlock_loop_cnt = 0;
1815                 do {
1816                         check_update_rusage(td);
1817                         if (!fio_sem_down_trylock(stat_sem))
1818                                 break;
1819                         usleep(1000);
1820                         if (deadlock_loop_cnt++ > 5000) {
1821                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1822                                 td->error = EDEADLK;
1823                                 goto err;
1824                         }
1825                 } while (1);
1826
1827                 if (td_read(td) && td->io_bytes[DDIR_READ])
1828                         update_runtime(td, elapsed_us, DDIR_READ);
1829                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1830                         update_runtime(td, elapsed_us, DDIR_WRITE);
1831                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1832                         update_runtime(td, elapsed_us, DDIR_TRIM);
1833                 fio_gettime(&td->start, NULL);
1834                 fio_sem_up(stat_sem);
1835
1836                 if (td->error || td->terminate)
1837                         break;
1838
1839                 if (!o->do_verify ||
1840                     o->verify == VERIFY_NONE ||
1841                     td_ioengine_flagged(td, FIO_UNIDIR))
1842                         continue;
1843
1844                 clear_io_state(td, 0);
1845
1846                 fio_gettime(&td->start, NULL);
1847
1848                 do_verify(td, verify_bytes);
1849
1850                 /*
1851                  * See comment further up for why this is done here.
1852                  */
1853                 check_update_rusage(td);
1854
1855                 fio_sem_down(stat_sem);
1856                 update_runtime(td, elapsed_us, DDIR_READ);
1857                 fio_gettime(&td->start, NULL);
1858                 fio_sem_up(stat_sem);
1859
1860                 if (td->error || td->terminate)
1861                         break;
1862         }
1863
1864         /*
1865          * Acquire this lock if we were doing overlap checking in
1866          * offload mode so that we don't clean up this job while
1867          * another thread is checking its io_u's for overlap
1868          */
1869         if (td_offload_overlap(td))
1870                 pthread_mutex_lock(&overlap_check);
1871         td_set_runstate(td, TD_FINISHING);
1872         if (td_offload_overlap(td))
1873                 pthread_mutex_unlock(&overlap_check);
1874
1875         update_rusage_stat(td);
1876         td->ts.total_run_time = mtime_since_now(&td->epoch);
1877         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1878         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1879         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1880
1881         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1882             (td->o.verify != VERIFY_NONE && td_write(td)))
1883                 verify_save_state(td->thread_number);
1884
1885         fio_unpin_memory(td);
1886
1887         td_writeout_logs(td, true);
1888
1889         iolog_compress_exit(td);
1890         rate_submit_exit(td);
1891
1892         if (o->exec_postrun)
1893                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1894
1895         if (exitall_on_terminate || (o->exitall_error && td->error))
1896                 fio_terminate_threads(td->groupid, td->o.exit_what);
1897
1898 err:
1899         if (td->error)
1900                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1901                                                         td->verror);
1902
1903         if (o->verify_async)
1904                 verify_async_exit(td);
1905
1906         close_and_free_files(td);
1907         cleanup_io_u(td);
1908         close_ioengine(td);
1909         cgroup_shutdown(td, cgroup_mnt);
1910         verify_free_state(td);
1911         td_zone_free_index(td);
1912
1913         if (fio_option_is_set(o, cpumask)) {
1914                 ret = fio_cpuset_exit(&o->cpumask);
1915                 if (ret)
1916                         td_verror(td, ret, "fio_cpuset_exit");
1917         }
1918
1919         /*
1920          * do this very late, it will log file closing as well
1921          */
1922         if (o->write_iolog_file)
1923                 write_iolog_close(td);
1924         if (td->io_log_rfile)
1925                 fclose(td->io_log_rfile);
1926
1927         td_set_runstate(td, TD_EXITED);
1928
1929         /*
1930          * Do this last after setting our runstate to exited, so we
1931          * know that the stat thread is signaled.
1932          */
1933         check_update_rusage(td);
1934
1935         sk_out_drop();
1936         return (void *) (uintptr_t) td->error;
1937 }
1938
1939 /*
1940  * Run over the job map and reap the threads that have exited, if any.
1941  */
1942 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1943                          uint64_t *m_rate)
1944 {
1945         struct thread_data *td;
1946         unsigned int cputhreads, realthreads, pending;
1947         int i, status, ret;
1948
1949         /*
1950          * reap exited threads (TD_EXITED -> TD_REAPED)
1951          */
1952         realthreads = pending = cputhreads = 0;
1953         for_each_td(td, i) {
1954                 int flags = 0;
1955
1956                  if (!strcmp(td->o.ioengine, "cpuio"))
1957                         cputhreads++;
1958                 else
1959                         realthreads++;
1960
1961                 if (!td->pid) {
1962                         pending++;
1963                         continue;
1964                 }
1965                 if (td->runstate == TD_REAPED)
1966                         continue;
1967                 if (td->o.use_thread) {
1968                         if (td->runstate == TD_EXITED) {
1969                                 td_set_runstate(td, TD_REAPED);
1970                                 goto reaped;
1971                         }
1972                         continue;
1973                 }
1974
1975                 flags = WNOHANG;
1976                 if (td->runstate == TD_EXITED)
1977                         flags = 0;
1978
1979                 /*
1980                  * check if someone quit or got killed in an unusual way
1981                  */
1982                 ret = waitpid(td->pid, &status, flags);
1983                 if (ret < 0) {
1984                         if (errno == ECHILD) {
1985                                 log_err("fio: pid=%d disappeared %d\n",
1986                                                 (int) td->pid, td->runstate);
1987                                 td->sig = ECHILD;
1988                                 td_set_runstate(td, TD_REAPED);
1989                                 goto reaped;
1990                         }
1991                         perror("waitpid");
1992                 } else if (ret == td->pid) {
1993                         if (WIFSIGNALED(status)) {
1994                                 int sig = WTERMSIG(status);
1995
1996                                 if (sig != SIGTERM && sig != SIGUSR2)
1997                                         log_err("fio: pid=%d, got signal=%d\n",
1998                                                         (int) td->pid, sig);
1999                                 td->sig = sig;
2000                                 td_set_runstate(td, TD_REAPED);
2001                                 goto reaped;
2002                         }
2003                         if (WIFEXITED(status)) {
2004                                 if (WEXITSTATUS(status) && !td->error)
2005                                         td->error = WEXITSTATUS(status);
2006
2007                                 td_set_runstate(td, TD_REAPED);
2008                                 goto reaped;
2009                         }
2010                 }
2011
2012                 /*
2013                  * If the job is stuck, do a forceful timeout of it and
2014                  * move on.
2015                  */
2016                 if (td->terminate &&
2017                     td->runstate < TD_FSYNCING &&
2018                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2019                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2020                                 "%lu seconds, it appears to be stuck. Doing "
2021                                 "forceful exit of this job.\n",
2022                                 td->o.name, td->runstate,
2023                                 (unsigned long) time_since_now(&td->terminate_time));
2024                         td_set_runstate(td, TD_REAPED);
2025                         goto reaped;
2026                 }
2027
2028                 /*
2029                  * thread is not dead, continue
2030                  */
2031                 pending++;
2032                 continue;
2033 reaped:
2034                 (*nr_running)--;
2035                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2036                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2037                 if (!td->pid)
2038                         pending--;
2039
2040                 if (td->error)
2041                         exit_value++;
2042
2043                 done_secs += mtime_since_now(&td->epoch) / 1000;
2044                 profile_td_exit(td);
2045         }
2046
2047         if (*nr_running == cputhreads && !pending && realthreads)
2048                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2049 }
2050
2051 static bool __check_trigger_file(void)
2052 {
2053         struct stat sb;
2054
2055         if (!trigger_file)
2056                 return false;
2057
2058         if (stat(trigger_file, &sb))
2059                 return false;
2060
2061         if (unlink(trigger_file) < 0)
2062                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2063                                                         strerror(errno));
2064
2065         return true;
2066 }
2067
2068 static bool trigger_timedout(void)
2069 {
2070         if (trigger_timeout)
2071                 if (time_since_genesis() >= trigger_timeout) {
2072                         trigger_timeout = 0;
2073                         return true;
2074                 }
2075
2076         return false;
2077 }
2078
2079 void exec_trigger(const char *cmd)
2080 {
2081         int ret;
2082
2083         if (!cmd || cmd[0] == '\0')
2084                 return;
2085
2086         ret = system(cmd);
2087         if (ret == -1)
2088                 log_err("fio: failed executing %s trigger\n", cmd);
2089 }
2090
2091 void check_trigger_file(void)
2092 {
2093         if (__check_trigger_file() || trigger_timedout()) {
2094                 if (nr_clients)
2095                         fio_clients_send_trigger(trigger_remote_cmd);
2096                 else {
2097                         verify_save_state(IO_LIST_ALL);
2098                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2099                         exec_trigger(trigger_cmd);
2100                 }
2101         }
2102 }
2103
2104 static int fio_verify_load_state(struct thread_data *td)
2105 {
2106         int ret;
2107
2108         if (!td->o.verify_state)
2109                 return 0;
2110
2111         if (is_backend) {
2112                 void *data;
2113
2114                 ret = fio_server_get_verify_state(td->o.name,
2115                                         td->thread_number - 1, &data);
2116                 if (!ret)
2117                         verify_assign_state(td, data);
2118         } else {
2119                 char prefix[PATH_MAX];
2120
2121                 if (aux_path)
2122                         sprintf(prefix, "%s%clocal", aux_path,
2123                                         FIO_OS_PATH_SEPARATOR);
2124                 else
2125                         strcpy(prefix, "local");
2126                 ret = verify_load_state(td, prefix);
2127         }
2128
2129         return ret;
2130 }
2131
2132 static void do_usleep(unsigned int usecs)
2133 {
2134         check_for_running_stats();
2135         check_trigger_file();
2136         usleep(usecs);
2137 }
2138
2139 static bool check_mount_writes(struct thread_data *td)
2140 {
2141         struct fio_file *f;
2142         unsigned int i;
2143
2144         if (!td_write(td) || td->o.allow_mounted_write)
2145                 return false;
2146
2147         /*
2148          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2149          * are mkfs'd and mounted.
2150          */
2151         for_each_file(td, f, i) {
2152 #ifdef FIO_HAVE_CHARDEV_SIZE
2153                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2154 #else
2155                 if (f->filetype != FIO_TYPE_BLOCK)
2156 #endif
2157                         continue;
2158                 if (device_is_mounted(f->file_name))
2159                         goto mounted;
2160         }
2161
2162         return false;
2163 mounted:
2164         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2165         return true;
2166 }
2167
2168 static bool waitee_running(struct thread_data *me)
2169 {
2170         const char *waitee = me->o.wait_for;
2171         const char *self = me->o.name;
2172         struct thread_data *td;
2173         int i;
2174
2175         if (!waitee)
2176                 return false;
2177
2178         for_each_td(td, i) {
2179                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2180                         continue;
2181
2182                 if (td->runstate < TD_EXITED) {
2183                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2184                                         self, td->o.name,
2185                                         runstate_to_name(td->runstate));
2186                         return true;
2187                 }
2188         }
2189
2190         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2191         return false;
2192 }
2193
2194 /*
2195  * Main function for kicking off and reaping jobs, as needed.
2196  */
2197 static void run_threads(struct sk_out *sk_out)
2198 {
2199         struct thread_data *td;
2200         unsigned int i, todo, nr_running, nr_started;
2201         uint64_t m_rate, t_rate;
2202         uint64_t spent;
2203
2204         if (fio_gtod_offload && fio_start_gtod_thread())
2205                 return;
2206
2207         fio_idle_prof_init();
2208
2209         set_sig_handlers();
2210
2211         nr_thread = nr_process = 0;
2212         for_each_td(td, i) {
2213                 if (check_mount_writes(td))
2214                         return;
2215                 if (td->o.use_thread)
2216                         nr_thread++;
2217                 else
2218                         nr_process++;
2219         }
2220
2221         if (output_format & FIO_OUTPUT_NORMAL) {
2222                 struct buf_output out;
2223
2224                 buf_output_init(&out);
2225                 __log_buf(&out, "Starting ");
2226                 if (nr_thread)
2227                         __log_buf(&out, "%d thread%s", nr_thread,
2228                                                 nr_thread > 1 ? "s" : "");
2229                 if (nr_process) {
2230                         if (nr_thread)
2231                                 __log_buf(&out, " and ");
2232                         __log_buf(&out, "%d process%s", nr_process,
2233                                                 nr_process > 1 ? "es" : "");
2234                 }
2235                 __log_buf(&out, "\n");
2236                 log_info_buf(out.buf, out.buflen);
2237                 buf_output_free(&out);
2238         }
2239
2240         todo = thread_number;
2241         nr_running = 0;
2242         nr_started = 0;
2243         m_rate = t_rate = 0;
2244
2245         for_each_td(td, i) {
2246                 print_status_init(td->thread_number - 1);
2247
2248                 if (!td->o.create_serialize)
2249                         continue;
2250
2251                 if (fio_verify_load_state(td))
2252                         goto reap;
2253
2254                 /*
2255                  * do file setup here so it happens sequentially,
2256                  * we don't want X number of threads getting their
2257                  * client data interspersed on disk
2258                  */
2259                 if (setup_files(td)) {
2260 reap:
2261                         exit_value++;
2262                         if (td->error)
2263                                 log_err("fio: pid=%d, err=%d/%s\n",
2264                                         (int) td->pid, td->error, td->verror);
2265                         td_set_runstate(td, TD_REAPED);
2266                         todo--;
2267                 } else {
2268                         struct fio_file *f;
2269                         unsigned int j;
2270
2271                         /*
2272                          * for sharing to work, each job must always open
2273                          * its own files. so close them, if we opened them
2274                          * for creation
2275                          */
2276                         for_each_file(td, f, j) {
2277                                 if (fio_file_open(f))
2278                                         td_io_close_file(td, f);
2279                         }
2280                 }
2281         }
2282
2283         /* start idle threads before io threads start to run */
2284         fio_idle_prof_start();
2285
2286         set_genesis_time();
2287
2288         while (todo) {
2289                 struct thread_data *map[REAL_MAX_JOBS];
2290                 struct timespec this_start;
2291                 int this_jobs = 0, left;
2292                 struct fork_data *fd;
2293
2294                 /*
2295                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2296                  */
2297                 for_each_td(td, i) {
2298                         if (td->runstate != TD_NOT_CREATED)
2299                                 continue;
2300
2301                         /*
2302                          * never got a chance to start, killed by other
2303                          * thread for some reason
2304                          */
2305                         if (td->terminate) {
2306                                 todo--;
2307                                 continue;
2308                         }
2309
2310                         if (td->o.start_delay) {
2311                                 spent = utime_since_genesis();
2312
2313                                 if (td->o.start_delay > spent)
2314                                         continue;
2315                         }
2316
2317                         if (td->o.stonewall && (nr_started || nr_running)) {
2318                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2319                                                         td->o.name);
2320                                 break;
2321                         }
2322
2323                         if (waitee_running(td)) {
2324                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2325                                                 td->o.name, td->o.wait_for);
2326                                 continue;
2327                         }
2328
2329                         init_disk_util(td);
2330
2331                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2332                         td->update_rusage = 0;
2333
2334                         /*
2335                          * Set state to created. Thread will transition
2336                          * to TD_INITIALIZED when it's done setting up.
2337                          */
2338                         td_set_runstate(td, TD_CREATED);
2339                         map[this_jobs++] = td;
2340                         nr_started++;
2341
2342                         fd = calloc(1, sizeof(*fd));
2343                         fd->td = td;
2344                         fd->sk_out = sk_out;
2345
2346                         if (td->o.use_thread) {
2347                                 int ret;
2348
2349                                 dprint(FD_PROCESS, "will pthread_create\n");
2350                                 ret = pthread_create(&td->thread, NULL,
2351                                                         thread_main, fd);
2352                                 if (ret) {
2353                                         log_err("pthread_create: %s\n",
2354                                                         strerror(ret));
2355                                         free(fd);
2356                                         nr_started--;
2357                                         break;
2358                                 }
2359                                 fd = NULL;
2360                                 ret = pthread_detach(td->thread);
2361                                 if (ret)
2362                                         log_err("pthread_detach: %s",
2363                                                         strerror(ret));
2364                         } else {
2365                                 pid_t pid;
2366                                 dprint(FD_PROCESS, "will fork\n");
2367                                 pid = fork();
2368                                 if (!pid) {
2369                                         int ret;
2370
2371                                         ret = (int)(uintptr_t)thread_main(fd);
2372                                         _exit(ret);
2373                                 } else if (i == fio_debug_jobno)
2374                                         *fio_debug_jobp = pid;
2375                         }
2376                         dprint(FD_MUTEX, "wait on startup_sem\n");
2377                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2378                                 log_err("fio: job startup hung? exiting.\n");
2379                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2380                                 fio_abort = true;
2381                                 nr_started--;
2382                                 free(fd);
2383                                 break;
2384                         }
2385                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2386                 }
2387
2388                 /*
2389                  * Wait for the started threads to transition to
2390                  * TD_INITIALIZED.
2391                  */
2392                 fio_gettime(&this_start, NULL);
2393                 left = this_jobs;
2394                 while (left && !fio_abort) {
2395                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2396                                 break;
2397
2398                         do_usleep(100000);
2399
2400                         for (i = 0; i < this_jobs; i++) {
2401                                 td = map[i];
2402                                 if (!td)
2403                                         continue;
2404                                 if (td->runstate == TD_INITIALIZED) {
2405                                         map[i] = NULL;
2406                                         left--;
2407                                 } else if (td->runstate >= TD_EXITED) {
2408                                         map[i] = NULL;
2409                                         left--;
2410                                         todo--;
2411                                         nr_running++; /* work-around... */
2412                                 }
2413                         }
2414                 }
2415
2416                 if (left) {
2417                         log_err("fio: %d job%s failed to start\n", left,
2418                                         left > 1 ? "s" : "");
2419                         for (i = 0; i < this_jobs; i++) {
2420                                 td = map[i];
2421                                 if (!td)
2422                                         continue;
2423                                 kill(td->pid, SIGTERM);
2424                         }
2425                         break;
2426                 }
2427
2428                 /*
2429                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2430                  */
2431                 for_each_td(td, i) {
2432                         if (td->runstate != TD_INITIALIZED)
2433                                 continue;
2434
2435                         if (in_ramp_time(td))
2436                                 td_set_runstate(td, TD_RAMP);
2437                         else
2438                                 td_set_runstate(td, TD_RUNNING);
2439                         nr_running++;
2440                         nr_started--;
2441                         m_rate += ddir_rw_sum(td->o.ratemin);
2442                         t_rate += ddir_rw_sum(td->o.rate);
2443                         todo--;
2444                         fio_sem_up(td->sem);
2445                 }
2446
2447                 reap_threads(&nr_running, &t_rate, &m_rate);
2448
2449                 if (todo)
2450                         do_usleep(100000);
2451         }
2452
2453         while (nr_running) {
2454                 reap_threads(&nr_running, &t_rate, &m_rate);
2455                 do_usleep(10000);
2456         }
2457
2458         fio_idle_prof_stop();
2459
2460         update_io_ticks();
2461 }
2462
2463 static void free_disk_util(void)
2464 {
2465         disk_util_prune_entries();
2466         helper_thread_destroy();
2467 }
2468
2469 int fio_backend(struct sk_out *sk_out)
2470 {
2471         struct thread_data *td;
2472         int i;
2473
2474         if (exec_profile) {
2475                 if (load_profile(exec_profile))
2476                         return 1;
2477                 free(exec_profile);
2478                 exec_profile = NULL;
2479         }
2480         if (!thread_number)
2481                 return 0;
2482
2483         if (write_bw_log) {
2484                 struct log_params p = {
2485                         .log_type = IO_LOG_TYPE_BW,
2486                 };
2487
2488                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2489                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2490                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2491         }
2492
2493         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2494         if (!sk_out)
2495                 is_local_backend = true;
2496         if (startup_sem == NULL)
2497                 return 1;
2498
2499         set_genesis_time();
2500         stat_init();
2501         if (helper_thread_create(startup_sem, sk_out))
2502                 log_err("fio: failed to create helper thread\n");
2503
2504         cgroup_list = smalloc(sizeof(*cgroup_list));
2505         if (cgroup_list)
2506                 INIT_FLIST_HEAD(cgroup_list);
2507
2508         run_threads(sk_out);
2509
2510         helper_thread_exit();
2511
2512         if (!fio_abort) {
2513                 __show_run_stats();
2514                 if (write_bw_log) {
2515                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2516                                 struct io_log *log = agg_io_log[i];
2517
2518                                 flush_log(log, false);
2519                                 free_log(log);
2520                         }
2521                 }
2522         }
2523
2524         for_each_td(td, i) {
2525                 steadystate_free(td);
2526                 fio_options_free(td);
2527                 if (td->rusage_sem) {
2528                         fio_sem_remove(td->rusage_sem);
2529                         td->rusage_sem = NULL;
2530                 }
2531                 fio_sem_remove(td->sem);
2532                 td->sem = NULL;
2533         }
2534
2535         free_disk_util();
2536         if (cgroup_list) {
2537                 cgroup_kill(cgroup_list);
2538                 sfree(cgroup_list);
2539         }
2540
2541         fio_sem_remove(startup_sem);
2542         stat_exit();
2543         return exit_value;
2544 }