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
  33 #include "fio.h"
  34 #include "smalloc.h"
  35 #include "verify.h"
  36 #include "diskutil.h"
  37 #include "cgroup.h"
  38 #include "profile.h"
  39 #include "lib/rand.h"
  40 #include "lib/memalign.h"
  41 #include "server.h"
  42 #include "lib/getrusage.h"
  43 #include "idletime.h"
  44 #include "err.h"
  45 #include "workqueue.h"
  46 #include "lib/mountcheck.h"
  47 #include "rate-submit.h"
  48 #include "helper_thread.h"
  49 #include "pshared.h"
  50
  51 static struct fio_sem *startup_sem;
  52 static struct flist_head *cgroup_list;
  53 static char *cgroup_mnt;
  54 static int exit_value;
  55 static volatile int fio_abort;
  56 static unsigned int nr_process = 0;
  57 static unsigned int nr_thread = 0;
  58
  59 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  60
  61 int groupid = 0;
  62 unsigned int thread_number = 0;
  63 unsigned int stat_number = 0;
  64 int shm_id = 0;
  65 int temp_stall_ts;
  66 unsigned long done_secs = 0;
  67
  68 #define JOB_START_TIMEOUT       (5 * 1000)
  69
  70 static void sig_int(int sig)
  71 {
  72         if (threads) {
  73                 if (is_backend)
  74                         fio_server_got_signal(sig);
  75                 else {
  76                         log_info("\nfio: terminating on signal %d\n", sig);
  77                         log_info_flush();
  78                         exit_value = 128;
  79                 }
  80
  81                 fio_terminate_threads(TERMINATE_ALL);
  82         }
  83 }
  84
  85 void sig_show_status(int sig)
  86 {
  87         show_running_run_stats();
  88 }
  89
  90 static void set_sig_handlers(void)
  91 {
  92         struct sigaction act;
  93
  94         memset(&act, 0, sizeof(act));
  95         act.sa_handler = sig_int;
  96         act.sa_flags = SA_RESTART;
  97         sigaction(SIGINT, &act, NULL);
  98
  99         memset(&act, 0, sizeof(act));
 100         act.sa_handler = sig_int;
 101         act.sa_flags = SA_RESTART;
 102         sigaction(SIGTERM, &act, NULL);
 103
 104 /* Windows uses SIGBREAK as a quit signal from other applications */
 105 #ifdef WIN32
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGBREAK, &act, NULL);
 110 #endif
 111
 112         memset(&act, 0, sizeof(act));
 113         act.sa_handler = sig_show_status;
 114         act.sa_flags = SA_RESTART;
 115         sigaction(SIGUSR1, &act, NULL);
 116
 117         if (is_backend) {
 118                 memset(&act, 0, sizeof(act));
 119                 act.sa_handler = sig_int;
 120                 act.sa_flags = SA_RESTART;
 121                 sigaction(SIGPIPE, &act, NULL);
 122         }
 123 }
 124
 125 /*
 126  * Check if we are above the minimum rate given.
 127  */
 128 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 129                              enum fio_ddir ddir)
 130 {
 131         unsigned long long bytes = 0;
 132         unsigned long iops = 0;
 133         unsigned long spent;
 134         unsigned long rate;
 135         unsigned int ratemin = 0;
 136         unsigned int rate_iops = 0;
 137         unsigned int rate_iops_min = 0;
 138
 139         assert(ddir_rw(ddir));
 140
 141         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 142                 return false;
 143
 144         /*
 145          * allow a 2 second settle period in the beginning
 146          */
 147         if (mtime_since(&td->start, now) < 2000)
 148                 return false;
 149
 150         iops += td->this_io_blocks[ddir];
 151         bytes += td->this_io_bytes[ddir];
 152         ratemin += td->o.ratemin[ddir];
 153         rate_iops += td->o.rate_iops[ddir];
 154         rate_iops_min += td->o.rate_iops_min[ddir];
 155
 156         /*
 157          * if rate blocks is set, sample is running
 158          */
 159         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 160                 spent = mtime_since(&td->lastrate[ddir], now);
 161                 if (spent < td->o.ratecycle)
 162                         return false;
 163
 164                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 165                         /*
 166                          * check bandwidth specified rate
 167                          */
 168                         if (bytes < td->rate_bytes[ddir]) {
 169                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 170                                         td->o.name, ratemin, bytes);
 171                                 return true;
 172                         } else {
 173                                 if (spent)
 174                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 175                                 else
 176                                         rate = 0;
 177
 178                                 if (rate < ratemin ||
 179                                     bytes < td->rate_bytes[ddir]) {
 180                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 181                                                 td->o.name, ratemin, rate);
 182                                         return true;
 183                                 }
 184                         }
 185                 } else {
 186                         /*
 187                          * checks iops specified rate
 188                          */
 189                         if (iops < rate_iops) {
 190                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 191                                                 td->o.name, rate_iops, iops);
 192                                 return true;
 193                         } else {
 194                                 if (spent)
 195                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 196                                 else
 197                                         rate = 0;
 198
 199                                 if (rate < rate_iops_min ||
 200                                     iops < td->rate_blocks[ddir]) {
 201                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 202                                                 td->o.name, rate_iops_min, rate);
 203                                         return true;
 204                                 }
 205                         }
 206                 }
 207         }
 208
 209         td->rate_bytes[ddir] = bytes;
 210         td->rate_blocks[ddir] = iops;
 211         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 212         return false;
 213 }
 214
 215 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 216 {
 217         bool ret = false;
 218
 219         if (td->bytes_done[DDIR_READ])
 220                 ret |= __check_min_rate(td, now, DDIR_READ);
 221         if (td->bytes_done[DDIR_WRITE])
 222                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 223         if (td->bytes_done[DDIR_TRIM])
 224                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 225
 226         return ret;
 227 }
 228
 229 /*
 230  * When job exits, we can cancel the in-flight IO if we are using async
 231  * io. Attempt to do so.
 232  */
 233 static void cleanup_pending_aio(struct thread_data *td)
 234 {
 235         int r;
 236
 237         /*
 238          * get immediately available events, if any
 239          */
 240         r = io_u_queued_complete(td, 0);
 241         if (r < 0)
 242                 return;
 243
 244         /*
 245          * now cancel remaining active events
 246          */
 247         if (td->io_ops->cancel) {
 248                 struct io_u *io_u;
 249                 int i;
 250
 251                 io_u_qiter(&td->io_u_all, io_u, i) {
 252                         if (io_u->flags & IO_U_F_FLIGHT) {
 253                                 r = td->io_ops->cancel(td, io_u);
 254                                 if (!r)
 255                                         put_io_u(td, io_u);
 256                         }
 257                 }
 258         }
 259
 260         if (td->cur_depth)
 261                 r = io_u_queued_complete(td, td->cur_depth);
 262 }
 263
 264 /*
 265  * Helper to handle the final sync of a file. Works just like the normal
 266  * io path, just does everything sync.
 267  */
 268 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 269 {
 270         struct io_u *io_u = __get_io_u(td);
 271         int ret;
 272
 273         if (!io_u)
 274                 return true;
 275
 276         io_u->ddir = DDIR_SYNC;
 277         io_u->file = f;
 278
 279         if (td_io_prep(td, io_u)) {
 280                 put_io_u(td, io_u);
 281                 return true;
 282         }
 283
 284 requeue:
 285         ret = td_io_queue(td, io_u);
 286         if (ret < 0) {
 287                 td_verror(td, io_u->error, "td_io_queue");
 288                 put_io_u(td, io_u);
 289                 return true;
 290         } else if (ret == FIO_Q_QUEUED) {
 291                 if (td_io_commit(td))
 292                         return true;
 293                 if (io_u_queued_complete(td, 1) < 0)
 294                         return true;
 295         } else if (ret == FIO_Q_COMPLETED) {
 296                 if (io_u->error) {
 297                         td_verror(td, io_u->error, "td_io_queue");
 298                         return true;
 299                 }
 300
 301                 if (io_u_sync_complete(td, io_u) < 0)
 302                         return true;
 303         } else if (ret == FIO_Q_BUSY) {
 304                 if (td_io_commit(td))
 305                         return true;
 306                 goto requeue;
 307         }
 308
 309         return false;
 310 }
 311
 312 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 313 {
 314         int ret;
 315
 316         if (fio_file_open(f))
 317                 return fio_io_sync(td, f);
 318
 319         if (td_io_open_file(td, f))
 320                 return 1;
 321
 322         ret = fio_io_sync(td, f);
 323         td_io_close_file(td, f);
 324         return ret;
 325 }
 326
 327 static inline void __update_ts_cache(struct thread_data *td)
 328 {
 329         fio_gettime(&td->ts_cache, NULL);
 330 }
 331
 332 static inline void update_ts_cache(struct thread_data *td)
 333 {
 334         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 335                 __update_ts_cache(td);
 336 }
 337
 338 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 339 {
 340         if (in_ramp_time(td))
 341                 return false;
 342         if (!td->o.timeout)
 343                 return false;
 344         if (utime_since(&td->epoch, t) >= td->o.timeout)
 345                 return true;
 346
 347         return false;
 348 }
 349
 350 /*
 351  * We need to update the runtime consistently in ms, but keep a running
 352  * tally of the current elapsed time in microseconds for sub millisecond
 353  * updates.
 354  */
 355 static inline void update_runtime(struct thread_data *td,
 356                                   unsigned long long *elapsed_us,
 357                                   const enum fio_ddir ddir)
 358 {
 359         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 360                 return;
 361
 362         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 363         elapsed_us[ddir] += utime_since_now(&td->start);
 364         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 365 }
 366
 367 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 368                                 int *retptr)
 369 {
 370         int ret = *retptr;
 371
 372         if (ret < 0 || td->error) {
 373                 int err = td->error;
 374                 enum error_type_bit eb;
 375
 376                 if (ret < 0)
 377                         err = -ret;
 378
 379                 eb = td_error_type(ddir, err);
 380                 if (!(td->o.continue_on_error & (1 << eb)))
 381                         return true;
 382
 383                 if (td_non_fatal_error(td, eb, err)) {
 384                         /*
 385                          * Continue with the I/Os in case of
 386                          * a non fatal error.
 387                          */
 388                         update_error_count(td, err);
 389                         td_clear_error(td);
 390                         *retptr = 0;
 391                         return false;
 392                 } else if (td->o.fill_device && err == ENOSPC) {
 393                         /*
 394                          * We expect to hit this error if
 395                          * fill_device option is set.
 396                          */
 397                         td_clear_error(td);
 398                         fio_mark_td_terminate(td);
 399                         return true;
 400                 } else {
 401                         /*
 402                          * Stop the I/O in case of a fatal
 403                          * error.
 404                          */
 405                         update_error_count(td, err);
 406                         return true;
 407                 }
 408         }
 409
 410         return false;
 411 }
 412
 413 static void check_update_rusage(struct thread_data *td)
 414 {
 415         if (td->update_rusage) {
 416                 td->update_rusage = 0;
 417                 update_rusage_stat(td);
 418                 fio_sem_up(td->rusage_sem);
 419         }
 420 }
 421
 422 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 423 {
 424         const int full = queue_full(td);
 425         int min_evts = 0;
 426         int ret;
 427
 428         if (td->flags & TD_F_REGROW_LOGS)
 429                 return io_u_quiesce(td);
 430
 431         /*
 432          * if the queue is full, we MUST reap at least 1 event
 433          */
 434         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 435         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 436                 min_evts = 1;
 437
 438         if (time && (__should_check_rate(td, DDIR_READ) ||
 439             __should_check_rate(td, DDIR_WRITE) ||
 440             __should_check_rate(td, DDIR_TRIM)))
 441                 fio_gettime(time, NULL);
 442
 443         do {
 444                 ret = io_u_queued_complete(td, min_evts);
 445                 if (ret < 0)
 446                         break;
 447         } while (full && (td->cur_depth > td->o.iodepth_low));
 448
 449         return ret;
 450 }
 451
 452 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 453                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 454                    struct timespec *comp_time)
 455 {
 456         int ret2;
 457
 458         switch (*ret) {
 459         case FIO_Q_COMPLETED:
 460                 if (io_u->error) {
 461                         *ret = -io_u->error;
 462                         clear_io_u(td, io_u);
 463                 } else if (io_u->resid) {
 464                         int bytes = io_u->xfer_buflen - io_u->resid;
 465                         struct fio_file *f = io_u->file;
 466
 467                         if (bytes_issued)
 468                                 *bytes_issued += bytes;
 469
 470                         if (!from_verify)
 471                                 trim_io_piece(td, io_u);
 472
 473                         /*
 474                          * zero read, fail
 475                          */
 476                         if (!bytes) {
 477                                 if (!from_verify)
 478                                         unlog_io_piece(td, io_u);
 479                                 td_verror(td, EIO, "full resid");
 480                                 put_io_u(td, io_u);
 481                                 break;
 482                         }
 483
 484                         io_u->xfer_buflen = io_u->resid;
 485                         io_u->xfer_buf += bytes;
 486                         io_u->offset += bytes;
 487
 488                         if (ddir_rw(io_u->ddir))
 489                                 td->ts.short_io_u[io_u->ddir]++;
 490
 491                         if (io_u->offset == f->real_file_size)
 492                                 goto sync_done;
 493
 494                         requeue_io_u(td, &io_u);
 495                 } else {
 496 sync_done:
 497                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 498                             __should_check_rate(td, DDIR_WRITE) ||
 499                             __should_check_rate(td, DDIR_TRIM)))
 500                                 fio_gettime(comp_time, NULL);
 501
 502                         *ret = io_u_sync_complete(td, io_u);
 503                         if (*ret < 0)
 504                                 break;
 505                 }
 506
 507                 if (td->flags & TD_F_REGROW_LOGS)
 508                         regrow_logs(td);
 509
 510                 /*
 511                  * when doing I/O (not when verifying),
 512                  * check for any errors that are to be ignored
 513                  */
 514                 if (!from_verify)
 515                         break;
 516
 517                 return 0;
 518         case FIO_Q_QUEUED:
 519                 /*
 520                  * if the engine doesn't have a commit hook,
 521                  * the io_u is really queued. if it does have such
 522                  * a hook, it has to call io_u_queued() itself.
 523                  */
 524                 if (td->io_ops->commit == NULL)
 525                         io_u_queued(td, io_u);
 526                 if (bytes_issued)
 527                         *bytes_issued += io_u->xfer_buflen;
 528                 break;
 529         case FIO_Q_BUSY:
 530                 if (!from_verify)
 531                         unlog_io_piece(td, io_u);
 532                 requeue_io_u(td, &io_u);
 533                 ret2 = td_io_commit(td);
 534                 if (ret2 < 0)
 535                         *ret = ret2;
 536                 break;
 537         default:
 538                 assert(*ret < 0);
 539                 td_verror(td, -(*ret), "td_io_queue");
 540                 break;
 541         }
 542
 543         if (break_on_this_error(td, ddir, ret))
 544                 return 1;
 545
 546         return 0;
 547 }
 548
 549 static inline bool io_in_polling(struct thread_data *td)
 550 {
 551         return !td->o.iodepth_batch_complete_min &&
 552                    !td->o.iodepth_batch_complete_max;
 553 }
 554 /*
 555  * Unlinks files from thread data fio_file structure
 556  */
 557 static int unlink_all_files(struct thread_data *td)
 558 {
 559         struct fio_file *f;
 560         unsigned int i;
 561         int ret = 0;
 562
 563         for_each_file(td, f, i) {
 564                 if (f->filetype != FIO_TYPE_FILE)
 565                         continue;
 566                 ret = td_io_unlink_file(td, f);
 567                 if (ret)
 568                         break;
 569         }
 570
 571         if (ret)
 572                 td_verror(td, ret, "unlink_all_files");
 573
 574         return ret;
 575 }
 576
 577 /*
 578  * Check if io_u will overlap an in-flight IO in the queue
 579  */
 580 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 581 {
 582         bool overlap;
 583         struct io_u *check_io_u;
 584         unsigned long long x1, x2, y1, y2;
 585         int i;
 586
 587         x1 = io_u->offset;
 588         x2 = io_u->offset + io_u->buflen;
 589         overlap = false;
 590         io_u_qiter(q, check_io_u, i) {
 591                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 592                         y1 = check_io_u->offset;
 593                         y2 = check_io_u->offset + check_io_u->buflen;
 594
 595                         if (x1 < y2 && y1 < x2) {
 596                                 overlap = true;
 597                                 dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
 598                                                 x1, io_u->buflen,
 599                                                 y1, check_io_u->buflen);
 600                                 break;
 601                         }
 602                 }
 603         }
 604
 605         return overlap;
 606 }
 607
 608 static int io_u_submit(struct thread_data *td, struct io_u *io_u)
 609 {
 610         /*
 611          * Check for overlap if the user asked us to, and we have
 612          * at least one IO in flight besides this one.
 613          */
 614         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 615             in_flight_overlap(&td->io_u_all, io_u))
 616                 return FIO_Q_BUSY;
 617
 618         return td_io_queue(td, io_u);
 619 }
 620
 621 /*
 622  * The main verify engine. Runs over the writes we previously submitted,
 623  * reads the blocks back in, and checks the crc/md5 of the data.
 624  */
 625 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 626 {
 627         struct fio_file *f;
 628         struct io_u *io_u;
 629         int ret, min_events;
 630         unsigned int i;
 631
 632         dprint(FD_VERIFY, "starting loop\n");
 633
 634         /*
 635          * sync io first and invalidate cache, to make sure we really
 636          * read from disk.
 637          */
 638         for_each_file(td, f, i) {
 639                 if (!fio_file_open(f))
 640                         continue;
 641                 if (fio_io_sync(td, f))
 642                         break;
 643                 if (file_invalidate_cache(td, f))
 644                         break;
 645         }
 646
 647         check_update_rusage(td);
 648
 649         if (td->error)
 650                 return;
 651
 652         /*
 653          * verify_state needs to be reset before verification
 654          * proceeds so that expected random seeds match actual
 655          * random seeds in headers. The main loop will reset
 656          * all random number generators if randrepeat is set.
 657          */
 658         if (!td->o.rand_repeatable)
 659                 td_fill_verify_state_seed(td);
 660
 661         td_set_runstate(td, TD_VERIFYING);
 662
 663         io_u = NULL;
 664         while (!td->terminate) {
 665                 enum fio_ddir ddir;
 666                 int full;
 667
 668                 update_ts_cache(td);
 669                 check_update_rusage(td);
 670
 671                 if (runtime_exceeded(td, &td->ts_cache)) {
 672                         __update_ts_cache(td);
 673                         if (runtime_exceeded(td, &td->ts_cache)) {
 674                                 fio_mark_td_terminate(td);
 675                                 break;
 676                         }
 677                 }
 678
 679                 if (flow_threshold_exceeded(td))
 680                         continue;
 681
 682                 if (!td->o.experimental_verify) {
 683                         io_u = __get_io_u(td);
 684                         if (!io_u)
 685                                 break;
 686
 687                         if (get_next_verify(td, io_u)) {
 688                                 put_io_u(td, io_u);
 689                                 break;
 690                         }
 691
 692                         if (td_io_prep(td, io_u)) {
 693                                 put_io_u(td, io_u);
 694                                 break;
 695                         }
 696                 } else {
 697                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 698                                 break;
 699
 700                         while ((io_u = get_io_u(td)) != NULL) {
 701                                 if (IS_ERR_OR_NULL(io_u)) {
 702                                         io_u = NULL;
 703                                         ret = FIO_Q_BUSY;
 704                                         goto reap;
 705                                 }
 706
 707                                 /*
 708                                  * We are only interested in the places where
 709                                  * we wrote or trimmed IOs. Turn those into
 710                                  * reads for verification purposes.
 711                                  */
 712                                 if (io_u->ddir == DDIR_READ) {
 713                                         /*
 714                                          * Pretend we issued it for rwmix
 715                                          * accounting
 716                                          */
 717                                         td->io_issues[DDIR_READ]++;
 718                                         put_io_u(td, io_u);
 719                                         continue;
 720                                 } else if (io_u->ddir == DDIR_TRIM) {
 721                                         io_u->ddir = DDIR_READ;
 722                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 723                                         break;
 724                                 } else if (io_u->ddir == DDIR_WRITE) {
 725                                         io_u->ddir = DDIR_READ;
 726                                         populate_verify_io_u(td, io_u);
 727                                         break;
 728                                 } else {
 729                                         put_io_u(td, io_u);
 730                                         continue;
 731                                 }
 732                         }
 733
 734                         if (!io_u)
 735                                 break;
 736                 }
 737
 738                 if (verify_state_should_stop(td, io_u)) {
 739                         put_io_u(td, io_u);
 740                         break;
 741                 }
 742
 743                 if (td->o.verify_async)
 744                         io_u->end_io = verify_io_u_async;
 745                 else
 746                         io_u->end_io = verify_io_u;
 747
 748                 ddir = io_u->ddir;
 749                 if (!td->o.disable_slat)
 750                         fio_gettime(&io_u->start_time, NULL);
 751
 752                 ret = io_u_submit(td, io_u);
 753
 754                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 755                         break;
 756
 757                 /*
 758                  * if we can queue more, do so. but check if there are
 759                  * completed io_u's first. Note that we can get BUSY even
 760                  * without IO queued, if the system is resource starved.
 761                  */
 762 reap:
 763                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 764                 if (full || io_in_polling(td))
 765                         ret = wait_for_completions(td, NULL);
 766
 767                 if (ret < 0)
 768                         break;
 769         }
 770
 771         check_update_rusage(td);
 772
 773         if (!td->error) {
 774                 min_events = td->cur_depth;
 775
 776                 if (min_events)
 777                         ret = io_u_queued_complete(td, min_events);
 778         } else
 779                 cleanup_pending_aio(td);
 780
 781         td_set_runstate(td, TD_RUNNING);
 782
 783         dprint(FD_VERIFY, "exiting loop\n");
 784 }
 785
 786 static bool exceeds_number_ios(struct thread_data *td)
 787 {
 788         unsigned long long number_ios;
 789
 790         if (!td->o.number_ios)
 791                 return false;
 792
 793         number_ios = ddir_rw_sum(td->io_blocks);
 794         number_ios += td->io_u_queued + td->io_u_in_flight;
 795
 796         return number_ios >= (td->o.number_ios * td->loops);
 797 }
 798
 799 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 800 {
 801         unsigned long long bytes, limit;
 802
 803         if (td_rw(td))
 804                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 805         else if (td_write(td))
 806                 bytes = this_bytes[DDIR_WRITE];
 807         else if (td_read(td))
 808                 bytes = this_bytes[DDIR_READ];
 809         else
 810                 bytes = this_bytes[DDIR_TRIM];
 811
 812         if (td->o.io_size)
 813                 limit = td->o.io_size;
 814         else
 815                 limit = td->o.size;
 816
 817         limit *= td->loops;
 818         return bytes >= limit || exceeds_number_ios(td);
 819 }
 820
 821 static bool io_issue_bytes_exceeded(struct thread_data *td)
 822 {
 823         return io_bytes_exceeded(td, td->io_issue_bytes);
 824 }
 825
 826 static bool io_complete_bytes_exceeded(struct thread_data *td)
 827 {
 828         return io_bytes_exceeded(td, td->this_io_bytes);
 829 }
 830
 831 /*
 832  * used to calculate the next io time for rate control
 833  *
 834  */
 835 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 836 {
 837         uint64_t bps = td->rate_bps[ddir];
 838
 839         assert(!(td->flags & TD_F_CHILD));
 840
 841         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 842                 uint64_t val, iops;
 843
 844                 iops = bps / td->o.bs[ddir];
 845                 val = (int64_t) (1000000 / iops) *
 846                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 847                 if (val) {
 848                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 849                                         (unsigned long long) 1000000 / val,
 850                                         ddir);
 851                 }
 852                 td->last_usec[ddir] += val;
 853                 return td->last_usec[ddir];
 854         } else if (bps) {
 855                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 856                 uint64_t secs = bytes / bps;
 857                 uint64_t remainder = bytes % bps;
 858
 859                 return remainder * 1000000 / bps + secs * 1000000;
 860         }
 861
 862         return 0;
 863 }
 864
 865 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 866 {
 867         unsigned long long b;
 868         uint64_t total;
 869         int left;
 870
 871         b = ddir_rw_sum(td->io_blocks);
 872         if (b % td->o.thinktime_blocks)
 873                 return;
 874
 875         io_u_quiesce(td);
 876
 877         total = 0;
 878         if (td->o.thinktime_spin)
 879                 total = usec_spin(td->o.thinktime_spin);
 880
 881         left = td->o.thinktime - total;
 882         if (left)
 883                 total += usec_sleep(td, left);
 884
 885         /*
 886          * If we're ignoring thinktime for the rate, add the number of bytes
 887          * we would have done while sleeping, minus one block to ensure we
 888          * start issuing immediately after the sleep.
 889          */
 890         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 891                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 892                 uint64_t bs = td->o.min_bs[ddir];
 893                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 894                 uint64_t over;
 895
 896                 if (usperop <= total)
 897                         over = bs;
 898                 else
 899                         over = (usperop - total) / usperop * -bs;
 900
 901                 td->rate_io_issue_bytes[ddir] += (missed - over);
 902         }
 903 }
 904
 905 /*
 906  * Main IO worker function. It retrieves io_u's to process and queues
 907  * and reaps them, checking for rate and errors along the way.
 908  *
 909  * Returns number of bytes written and trimmed.
 910  */
 911 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 912 {
 913         unsigned int i;
 914         int ret = 0;
 915         uint64_t total_bytes, bytes_issued = 0;
 916
 917         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 918                 bytes_done[i] = td->bytes_done[i];
 919
 920         if (in_ramp_time(td))
 921                 td_set_runstate(td, TD_RAMP);
 922         else
 923                 td_set_runstate(td, TD_RUNNING);
 924
 925         lat_target_init(td);
 926
 927         total_bytes = td->o.size;
 928         /*
 929         * Allow random overwrite workloads to write up to io_size
 930         * before starting verification phase as 'size' doesn't apply.
 931         */
 932         if (td_write(td) && td_random(td) && td->o.norandommap)
 933                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 934         /*
 935          * If verify_backlog is enabled, we'll run the verify in this
 936          * handler as well. For that case, we may need up to twice the
 937          * amount of bytes.
 938          */
 939         if (td->o.verify != VERIFY_NONE &&
 940            (td_write(td) && td->o.verify_backlog))
 941                 total_bytes += td->o.size;
 942
 943         /* In trimwrite mode, each byte is trimmed and then written, so
 944          * allow total_bytes to be twice as big */
 945         if (td_trimwrite(td))
 946                 total_bytes += td->total_io_size;
 947
 948         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 949                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 950                 td->o.time_based) {
 951                 struct timespec comp_time;
 952                 struct io_u *io_u;
 953                 int full;
 954                 enum fio_ddir ddir;
 955
 956                 check_update_rusage(td);
 957
 958                 if (td->terminate || td->done)
 959                         break;
 960
 961                 update_ts_cache(td);
 962
 963                 if (runtime_exceeded(td, &td->ts_cache)) {
 964                         __update_ts_cache(td);
 965                         if (runtime_exceeded(td, &td->ts_cache)) {
 966                                 fio_mark_td_terminate(td);
 967                                 break;
 968                         }
 969                 }
 970
 971                 if (flow_threshold_exceeded(td))
 972                         continue;
 973
 974                 /*
 975                  * Break if we exceeded the bytes. The exception is time
 976                  * based runs, but we still need to break out of the loop
 977                  * for those to run verification, if enabled.
 978                  */
 979                 if (bytes_issued >= total_bytes &&
 980                     (!td->o.time_based ||
 981                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 982                         break;
 983
 984                 io_u = get_io_u(td);
 985                 if (IS_ERR_OR_NULL(io_u)) {
 986                         int err = PTR_ERR(io_u);
 987
 988                         io_u = NULL;
 989                         ddir = DDIR_INVAL;
 990                         if (err == -EBUSY) {
 991                                 ret = FIO_Q_BUSY;
 992                                 goto reap;
 993                         }
 994                         if (td->o.latency_target)
 995                                 goto reap;
 996                         break;
 997                 }
 998
 999                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1000                         populate_verify_io_u(td, io_u);
1001
1002                 ddir = io_u->ddir;
1003
1004                 /*
1005                  * Add verification end_io handler if:
1006                  *      - Asked to verify (!td_rw(td))
1007                  *      - Or the io_u is from our verify list (mixed write/ver)
1008                  */
1009                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1010                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1011
1012                         if (!td->o.verify_pattern_bytes) {
1013                                 io_u->rand_seed = __rand(&td->verify_state);
1014                                 if (sizeof(int) != sizeof(long *))
1015                                         io_u->rand_seed *= __rand(&td->verify_state);
1016                         }
1017
1018                         if (verify_state_should_stop(td, io_u)) {
1019                                 put_io_u(td, io_u);
1020                                 break;
1021                         }
1022
1023                         if (td->o.verify_async)
1024                                 io_u->end_io = verify_io_u_async;
1025                         else
1026                                 io_u->end_io = verify_io_u;
1027                         td_set_runstate(td, TD_VERIFYING);
1028                 } else if (in_ramp_time(td))
1029                         td_set_runstate(td, TD_RAMP);
1030                 else
1031                         td_set_runstate(td, TD_RUNNING);
1032
1033                 /*
1034                  * Always log IO before it's issued, so we know the specific
1035                  * order of it. The logged unit will track when the IO has
1036                  * completed.
1037                  */
1038                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1039                     td->o.do_verify &&
1040                     td->o.verify != VERIFY_NONE &&
1041                     !td->o.experimental_verify)
1042                         log_io_piece(td, io_u);
1043
1044                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1045                         const unsigned long blen = io_u->xfer_buflen;
1046                         const enum fio_ddir ddir = acct_ddir(io_u);
1047
1048                         if (td->error)
1049                                 break;
1050
1051                         workqueue_enqueue(&td->io_wq, &io_u->work);
1052                         ret = FIO_Q_QUEUED;
1053
1054                         if (ddir_rw(ddir)) {
1055                                 td->io_issues[ddir]++;
1056                                 td->io_issue_bytes[ddir] += blen;
1057                                 td->rate_io_issue_bytes[ddir] += blen;
1058                         }
1059
1060                         if (should_check_rate(td))
1061                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1062
1063                 } else {
1064                         ret = io_u_submit(td, io_u);
1065
1066                         if (should_check_rate(td))
1067                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1068
1069                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1070                                 break;
1071
1072                         /*
1073                          * See if we need to complete some commands. Note that
1074                          * we can get BUSY even without IO queued, if the
1075                          * system is resource starved.
1076                          */
1077 reap:
1078                         full = queue_full(td) ||
1079                                 (ret == FIO_Q_BUSY && td->cur_depth);
1080                         if (full || io_in_polling(td))
1081                                 ret = wait_for_completions(td, &comp_time);
1082                 }
1083                 if (ret < 0)
1084                         break;
1085                 if (!ddir_rw_sum(td->bytes_done) &&
1086                     !td_ioengine_flagged(td, FIO_NOIO))
1087                         continue;
1088
1089                 if (!in_ramp_time(td) && should_check_rate(td)) {
1090                         if (check_min_rate(td, &comp_time)) {
1091                                 if (exitall_on_terminate || td->o.exitall_error)
1092                                         fio_terminate_threads(td->groupid);
1093                                 td_verror(td, EIO, "check_min_rate");
1094                                 break;
1095                         }
1096                 }
1097                 if (!in_ramp_time(td) && td->o.latency_target)
1098                         lat_target_check(td);
1099
1100                 if (ddir_rw(ddir) && td->o.thinktime)
1101                         handle_thinktime(td, ddir);
1102         }
1103
1104         check_update_rusage(td);
1105
1106         if (td->trim_entries)
1107                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1108
1109         if (td->o.fill_device && td->error == ENOSPC) {
1110                 td->error = 0;
1111                 fio_mark_td_terminate(td);
1112         }
1113         if (!td->error) {
1114                 struct fio_file *f;
1115
1116                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1117                         workqueue_flush(&td->io_wq);
1118                         i = 0;
1119                 } else
1120                         i = td->cur_depth;
1121
1122                 if (i) {
1123                         ret = io_u_queued_complete(td, i);
1124                         if (td->o.fill_device && td->error == ENOSPC)
1125                                 td->error = 0;
1126                 }
1127
1128                 if (should_fsync(td) && td->o.end_fsync) {
1129                         td_set_runstate(td, TD_FSYNCING);
1130
1131                         for_each_file(td, f, i) {
1132                                 if (!fio_file_fsync(td, f))
1133                                         continue;
1134
1135                                 log_err("fio: end_fsync failed for file %s\n",
1136                                                                 f->file_name);
1137                         }
1138                 }
1139         } else
1140                 cleanup_pending_aio(td);
1141
1142         /*
1143          * stop job if we failed doing any IO
1144          */
1145         if (!ddir_rw_sum(td->this_io_bytes))
1146                 td->done = 1;
1147
1148         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1149                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1150 }
1151
1152 static void free_file_completion_logging(struct thread_data *td)
1153 {
1154         struct fio_file *f;
1155         unsigned int i;
1156
1157         for_each_file(td, f, i) {
1158                 if (!f->last_write_comp)
1159                         break;
1160                 sfree(f->last_write_comp);
1161         }
1162 }
1163
1164 static int init_file_completion_logging(struct thread_data *td,
1165                                         unsigned int depth)
1166 {
1167         struct fio_file *f;
1168         unsigned int i;
1169
1170         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1171                 return 0;
1172
1173         for_each_file(td, f, i) {
1174                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1175                 if (!f->last_write_comp)
1176                         goto cleanup;
1177         }
1178
1179         return 0;
1180
1181 cleanup:
1182         free_file_completion_logging(td);
1183         log_err("fio: failed to alloc write comp data\n");
1184         return 1;
1185 }
1186
1187 static void cleanup_io_u(struct thread_data *td)
1188 {
1189         struct io_u *io_u;
1190
1191         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1192
1193                 if (td->io_ops->io_u_free)
1194                         td->io_ops->io_u_free(td, io_u);
1195
1196                 fio_memfree(io_u, sizeof(*io_u));
1197         }
1198
1199         free_io_mem(td);
1200
1201         io_u_rexit(&td->io_u_requeues);
1202         io_u_qexit(&td->io_u_freelist);
1203         io_u_qexit(&td->io_u_all);
1204
1205         free_file_completion_logging(td);
1206 }
1207
1208 static int init_io_u(struct thread_data *td)
1209 {
1210         struct io_u *io_u;
1211         unsigned int max_bs, min_write;
1212         int cl_align, i, max_units;
1213         int data_xfer = 1, err;
1214         char *p;
1215
1216         max_units = td->o.iodepth;
1217         max_bs = td_max_bs(td);
1218         min_write = td->o.min_bs[DDIR_WRITE];
1219         td->orig_buffer_size = (unsigned long long) max_bs
1220                                         * (unsigned long long) max_units;
1221
1222         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1223                 data_xfer = 0;
1224
1225         err = 0;
1226         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1227         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1228         err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1229
1230         if (err) {
1231                 log_err("fio: failed setting up IO queues\n");
1232                 return 1;
1233         }
1234
1235         /*
1236          * if we may later need to do address alignment, then add any
1237          * possible adjustment here so that we don't cause a buffer
1238          * overflow later. this adjustment may be too much if we get
1239          * lucky and the allocator gives us an aligned address.
1240          */
1241         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1242             td_ioengine_flagged(td, FIO_RAWIO))
1243                 td->orig_buffer_size += page_mask + td->o.mem_align;
1244
1245         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1246                 unsigned long bs;
1247
1248                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1249                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1250         }
1251
1252         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1253                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1254                 return 1;
1255         }
1256
1257         if (data_xfer && allocate_io_mem(td))
1258                 return 1;
1259
1260         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1261             td_ioengine_flagged(td, FIO_RAWIO))
1262                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1263         else
1264                 p = td->orig_buffer;
1265
1266         cl_align = os_cache_line_size();
1267
1268         for (i = 0; i < max_units; i++) {
1269                 void *ptr;
1270
1271                 if (td->terminate)
1272                         return 1;
1273
1274                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1275                 if (!ptr) {
1276                         log_err("fio: unable to allocate aligned memory\n");
1277                         break;
1278                 }
1279
1280                 io_u = ptr;
1281                 memset(io_u, 0, sizeof(*io_u));
1282                 INIT_FLIST_HEAD(&io_u->verify_list);
1283                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1284
1285                 if (data_xfer) {
1286                         io_u->buf = p;
1287                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1288
1289                         if (td_write(td))
1290                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1291                         if (td_write(td) && td->o.verify_pattern_bytes) {
1292                                 /*
1293                                  * Fill the buffer with the pattern if we are
1294                                  * going to be doing writes.
1295                                  */
1296                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1297                         }
1298                 }
1299
1300                 io_u->index = i;
1301                 io_u->flags = IO_U_F_FREE;
1302                 io_u_qpush(&td->io_u_freelist, io_u);
1303
1304                 /*
1305                  * io_u never leaves this stack, used for iteration of all
1306                  * io_u buffers.
1307                  */
1308                 io_u_qpush(&td->io_u_all, io_u);
1309
1310                 if (td->io_ops->io_u_init) {
1311                         int ret = td->io_ops->io_u_init(td, io_u);
1312
1313                         if (ret) {
1314                                 log_err("fio: failed to init engine data: %d\n", ret);
1315                                 return 1;
1316                         }
1317                 }
1318
1319                 p += max_bs;
1320         }
1321
1322         if (init_file_completion_logging(td, max_units))
1323                 return 1;
1324
1325         return 0;
1326 }
1327
1328 /*
1329  * This function is Linux specific.
1330  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1331  */
1332 static int switch_ioscheduler(struct thread_data *td)
1333 {
1334 #ifdef FIO_HAVE_IOSCHED_SWITCH
1335         char tmp[256], tmp2[128], *p;
1336         FILE *f;
1337         int ret;
1338
1339         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1340                 return 0;
1341
1342         assert(td->files && td->files[0]);
1343         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1344
1345         f = fopen(tmp, "r+");
1346         if (!f) {
1347                 if (errno == ENOENT) {
1348                         log_err("fio: os or kernel doesn't support IO scheduler"
1349                                 " switching\n");
1350                         return 0;
1351                 }
1352                 td_verror(td, errno, "fopen iosched");
1353                 return 1;
1354         }
1355
1356         /*
1357          * Set io scheduler.
1358          */
1359         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1360         if (ferror(f) || ret != 1) {
1361                 td_verror(td, errno, "fwrite");
1362                 fclose(f);
1363                 return 1;
1364         }
1365
1366         rewind(f);
1367
1368         /*
1369          * Read back and check that the selected scheduler is now the default.
1370          */
1371         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1372         if (ferror(f) || ret < 0) {
1373                 td_verror(td, errno, "fread");
1374                 fclose(f);
1375                 return 1;
1376         }
1377         tmp[ret] = '\0';
1378         /*
1379          * either a list of io schedulers or "none\n" is expected. Strip the
1380          * trailing newline.
1381          */
1382         p = tmp;
1383         strsep(&p, "\n");
1384
1385         /*
1386          * Write to "none" entry doesn't fail, so check the result here.
1387          */
1388         if (!strcmp(tmp, "none")) {
1389                 log_err("fio: io scheduler is not tunable\n");
1390                 fclose(f);
1391                 return 0;
1392         }
1393
1394         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1395         if (!strstr(tmp, tmp2)) {
1396                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1397                 td_verror(td, EINVAL, "iosched_switch");
1398                 fclose(f);
1399                 return 1;
1400         }
1401
1402         fclose(f);
1403         return 0;
1404 #else
1405         return 0;
1406 #endif
1407 }
1408
1409 static bool keep_running(struct thread_data *td)
1410 {
1411         unsigned long long limit;
1412
1413         if (td->done)
1414                 return false;
1415         if (td->terminate)
1416                 return false;
1417         if (td->o.time_based)
1418                 return true;
1419         if (td->o.loops) {
1420                 td->o.loops--;
1421                 return true;
1422         }
1423         if (exceeds_number_ios(td))
1424                 return false;
1425
1426         if (td->o.io_size)
1427                 limit = td->o.io_size;
1428         else
1429                 limit = td->o.size;
1430
1431         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1432                 uint64_t diff;
1433
1434                 /*
1435                  * If the difference is less than the maximum IO size, we
1436                  * are done.
1437                  */
1438                 diff = limit - ddir_rw_sum(td->io_bytes);
1439                 if (diff < td_max_bs(td))
1440                         return false;
1441
1442                 if (fio_files_done(td) && !td->o.io_size)
1443                         return false;
1444
1445                 return true;
1446         }
1447
1448         return false;
1449 }
1450
1451 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1452 {
1453         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1454         int ret;
1455         char *str;
1456
1457         str = malloc(newlen);
1458         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1459
1460         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1461         ret = system(str);
1462         if (ret == -1)
1463                 log_err("fio: exec of cmd <%s> failed\n", str);
1464
1465         free(str);
1466         return ret;
1467 }
1468
1469 /*
1470  * Dry run to compute correct state of numberio for verification.
1471  */
1472 static uint64_t do_dry_run(struct thread_data *td)
1473 {
1474         td_set_runstate(td, TD_RUNNING);
1475
1476         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1477                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1478                 struct io_u *io_u;
1479                 int ret;
1480
1481                 if (td->terminate || td->done)
1482                         break;
1483
1484                 io_u = get_io_u(td);
1485                 if (IS_ERR_OR_NULL(io_u))
1486                         break;
1487
1488                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1489                 io_u->error = 0;
1490                 io_u->resid = 0;
1491                 if (ddir_rw(acct_ddir(io_u)))
1492                         td->io_issues[acct_ddir(io_u)]++;
1493                 if (ddir_rw(io_u->ddir)) {
1494                         io_u_mark_depth(td, 1);
1495                         td->ts.total_io_u[io_u->ddir]++;
1496                 }
1497
1498                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1499                     td->o.do_verify &&
1500                     td->o.verify != VERIFY_NONE &&
1501                     !td->o.experimental_verify)
1502                         log_io_piece(td, io_u);
1503
1504                 ret = io_u_sync_complete(td, io_u);
1505                 (void) ret;
1506         }
1507
1508         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1509 }
1510
1511 struct fork_data {
1512         struct thread_data *td;
1513         struct sk_out *sk_out;
1514 };
1515
1516 /*
1517  * Entry point for the thread based jobs. The process based jobs end up
1518  * here as well, after a little setup.
1519  */
1520 static void *thread_main(void *data)
1521 {
1522         struct fork_data *fd = data;
1523         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1524         struct thread_data *td = fd->td;
1525         struct thread_options *o = &td->o;
1526         struct sk_out *sk_out = fd->sk_out;
1527         uint64_t bytes_done[DDIR_RWDIR_CNT];
1528         int deadlock_loop_cnt;
1529         bool clear_state, did_some_io;
1530         int ret;
1531
1532         sk_out_assign(sk_out);
1533         free(fd);
1534
1535         if (!o->use_thread) {
1536                 setsid();
1537                 td->pid = getpid();
1538         } else
1539                 td->pid = gettid();
1540
1541         fio_local_clock_init(o->use_thread);
1542
1543         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1544
1545         if (is_backend)
1546                 fio_server_send_start(td);
1547
1548         INIT_FLIST_HEAD(&td->io_log_list);
1549         INIT_FLIST_HEAD(&td->io_hist_list);
1550         INIT_FLIST_HEAD(&td->verify_list);
1551         INIT_FLIST_HEAD(&td->trim_list);
1552         td->io_hist_tree = RB_ROOT;
1553
1554         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1555         if (ret) {
1556                 td_verror(td, ret, "mutex_cond_init_pshared");
1557                 goto err;
1558         }
1559         ret = cond_init_pshared(&td->verify_cond);
1560         if (ret) {
1561                 td_verror(td, ret, "mutex_cond_pshared");
1562                 goto err;
1563         }
1564
1565         td_set_runstate(td, TD_INITIALIZED);
1566         dprint(FD_MUTEX, "up startup_sem\n");
1567         fio_sem_up(startup_sem);
1568         dprint(FD_MUTEX, "wait on td->sem\n");
1569         fio_sem_down(td->sem);
1570         dprint(FD_MUTEX, "done waiting on td->sem\n");
1571
1572         /*
1573          * A new gid requires privilege, so we need to do this before setting
1574          * the uid.
1575          */
1576         if (o->gid != -1U && setgid(o->gid)) {
1577                 td_verror(td, errno, "setgid");
1578                 goto err;
1579         }
1580         if (o->uid != -1U && setuid(o->uid)) {
1581                 td_verror(td, errno, "setuid");
1582                 goto err;
1583         }
1584
1585         /*
1586          * Do this early, we don't want the compress threads to be limited
1587          * to the same CPUs as the IO workers. So do this before we set
1588          * any potential CPU affinity
1589          */
1590         if (iolog_compress_init(td, sk_out))
1591                 goto err;
1592
1593         /*
1594          * If we have a gettimeofday() thread, make sure we exclude that
1595          * thread from this job
1596          */
1597         if (o->gtod_cpu)
1598                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1599
1600         /*
1601          * Set affinity first, in case it has an impact on the memory
1602          * allocations.
1603          */
1604         if (fio_option_is_set(o, cpumask)) {
1605                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1606                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1607                         if (!ret) {
1608                                 log_err("fio: no CPUs set\n");
1609                                 log_err("fio: Try increasing number of available CPUs\n");
1610                                 td_verror(td, EINVAL, "cpus_split");
1611                                 goto err;
1612                         }
1613                 }
1614                 ret = fio_setaffinity(td->pid, o->cpumask);
1615                 if (ret == -1) {
1616                         td_verror(td, errno, "cpu_set_affinity");
1617                         goto err;
1618                 }
1619         }
1620
1621 #ifdef CONFIG_LIBNUMA
1622         /* numa node setup */
1623         if (fio_option_is_set(o, numa_cpunodes) ||
1624             fio_option_is_set(o, numa_memnodes)) {
1625                 struct bitmask *mask;
1626
1627                 if (numa_available() < 0) {
1628                         td_verror(td, errno, "Does not support NUMA API\n");
1629                         goto err;
1630                 }
1631
1632                 if (fio_option_is_set(o, numa_cpunodes)) {
1633                         mask = numa_parse_nodestring(o->numa_cpunodes);
1634                         ret = numa_run_on_node_mask(mask);
1635                         numa_free_nodemask(mask);
1636                         if (ret == -1) {
1637                                 td_verror(td, errno, \
1638                                         "numa_run_on_node_mask failed\n");
1639                                 goto err;
1640                         }
1641                 }
1642
1643                 if (fio_option_is_set(o, numa_memnodes)) {
1644                         mask = NULL;
1645                         if (o->numa_memnodes)
1646                                 mask = numa_parse_nodestring(o->numa_memnodes);
1647
1648                         switch (o->numa_mem_mode) {
1649                         case MPOL_INTERLEAVE:
1650                                 numa_set_interleave_mask(mask);
1651                                 break;
1652                         case MPOL_BIND:
1653                                 numa_set_membind(mask);
1654                                 break;
1655                         case MPOL_LOCAL:
1656                                 numa_set_localalloc();
1657                                 break;
1658                         case MPOL_PREFERRED:
1659                                 numa_set_preferred(o->numa_mem_prefer_node);
1660                                 break;
1661                         case MPOL_DEFAULT:
1662                         default:
1663                                 break;
1664                         }
1665
1666                         if (mask)
1667                                 numa_free_nodemask(mask);
1668
1669                 }
1670         }
1671 #endif
1672
1673         if (fio_pin_memory(td))
1674                 goto err;
1675
1676         /*
1677          * May alter parameters that init_io_u() will use, so we need to
1678          * do this first.
1679          */
1680         if (init_iolog(td))
1681                 goto err;
1682
1683         if (init_io_u(td))
1684                 goto err;
1685
1686         if (o->verify_async && verify_async_init(td))
1687                 goto err;
1688
1689         if (fio_option_is_set(o, ioprio) ||
1690             fio_option_is_set(o, ioprio_class)) {
1691                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1692                 if (ret == -1) {
1693                         td_verror(td, errno, "ioprio_set");
1694                         goto err;
1695                 }
1696         }
1697
1698         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1699                 goto err;
1700
1701         errno = 0;
1702         if (nice(o->nice) == -1 && errno != 0) {
1703                 td_verror(td, errno, "nice");
1704                 goto err;
1705         }
1706
1707         if (o->ioscheduler && switch_ioscheduler(td))
1708                 goto err;
1709
1710         if (!o->create_serialize && setup_files(td))
1711                 goto err;
1712
1713         if (td_io_init(td))
1714                 goto err;
1715
1716         if (!init_random_map(td))
1717                 goto err;
1718
1719         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1720                 goto err;
1721
1722         if (o->pre_read && !pre_read_files(td))
1723                 goto err;
1724
1725         fio_verify_init(td);
1726
1727         if (rate_submit_init(td, sk_out))
1728                 goto err;
1729
1730         set_epoch_time(td, o->log_unix_epoch);
1731         fio_getrusage(&td->ru_start);
1732         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1733         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1734         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1735
1736         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1737                         o->ratemin[DDIR_TRIM]) {
1738                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1739                                         sizeof(td->bw_sample_time));
1740                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1741                                         sizeof(td->bw_sample_time));
1742                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1743                                         sizeof(td->bw_sample_time));
1744         }
1745
1746         memset(bytes_done, 0, sizeof(bytes_done));
1747         clear_state = false;
1748         did_some_io = false;
1749
1750         while (keep_running(td)) {
1751                 uint64_t verify_bytes;
1752
1753                 fio_gettime(&td->start, NULL);
1754                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1755
1756                 if (clear_state) {
1757                         clear_io_state(td, 0);
1758
1759                         if (o->unlink_each_loop && unlink_all_files(td))
1760                                 break;
1761                 }
1762
1763                 prune_io_piece_log(td);
1764
1765                 if (td->o.verify_only && td_write(td))
1766                         verify_bytes = do_dry_run(td);
1767                 else {
1768                         do_io(td, bytes_done);
1769
1770                         if (!ddir_rw_sum(bytes_done)) {
1771                                 fio_mark_td_terminate(td);
1772                                 verify_bytes = 0;
1773                         } else {
1774                                 verify_bytes = bytes_done[DDIR_WRITE] +
1775                                                 bytes_done[DDIR_TRIM];
1776                         }
1777                 }
1778
1779                 /*
1780                  * If we took too long to shut down, the main thread could
1781                  * already consider us reaped/exited. If that happens, break
1782                  * out and clean up.
1783                  */
1784                 if (td->runstate >= TD_EXITED)
1785                         break;
1786
1787                 clear_state = true;
1788
1789                 /*
1790                  * Make sure we've successfully updated the rusage stats
1791                  * before waiting on the stat mutex. Otherwise we could have
1792                  * the stat thread holding stat mutex and waiting for
1793                  * the rusage_sem, which would never get upped because
1794                  * this thread is waiting for the stat mutex.
1795                  */
1796                 deadlock_loop_cnt = 0;
1797                 do {
1798                         check_update_rusage(td);
1799                         if (!fio_sem_down_trylock(stat_sem))
1800                                 break;
1801                         usleep(1000);
1802                         if (deadlock_loop_cnt++ > 5000) {
1803                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1804                                 td->error = EDEADLK;
1805                                 goto err;
1806                         }
1807                 } while (1);
1808
1809                 if (td_read(td) && td->io_bytes[DDIR_READ])
1810                         update_runtime(td, elapsed_us, DDIR_READ);
1811                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1812                         update_runtime(td, elapsed_us, DDIR_WRITE);
1813                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1814                         update_runtime(td, elapsed_us, DDIR_TRIM);
1815                 fio_gettime(&td->start, NULL);
1816                 fio_sem_up(stat_sem);
1817
1818                 if (td->error || td->terminate)
1819                         break;
1820
1821                 if (!o->do_verify ||
1822                     o->verify == VERIFY_NONE ||
1823                     td_ioengine_flagged(td, FIO_UNIDIR))
1824                         continue;
1825
1826                 if (ddir_rw_sum(bytes_done))
1827                         did_some_io = true;
1828
1829                 clear_io_state(td, 0);
1830
1831                 fio_gettime(&td->start, NULL);
1832
1833                 do_verify(td, verify_bytes);
1834
1835                 /*
1836                  * See comment further up for why this is done here.
1837                  */
1838                 check_update_rusage(td);
1839
1840                 fio_sem_down(stat_sem);
1841                 update_runtime(td, elapsed_us, DDIR_READ);
1842                 fio_gettime(&td->start, NULL);
1843                 fio_sem_up(stat_sem);
1844
1845                 if (td->error || td->terminate)
1846                         break;
1847         }
1848
1849         /*
1850          * If td ended up with no I/O when it should have had,
1851          * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1852          * (Are we not missing other flags that can be ignored ?)
1853          */
1854         if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1855             !did_some_io && !td->o.create_only &&
1856             !(td_ioengine_flagged(td, FIO_NOIO) ||
1857               td_ioengine_flagged(td, FIO_DISKLESSIO)))
1858                 log_err("%s: No I/O performed by %s, "
1859                          "perhaps try --debug=io option for details?\n",
1860                          td->o.name, td->io_ops->name);
1861
1862         td_set_runstate(td, TD_FINISHING);
1863
1864         update_rusage_stat(td);
1865         td->ts.total_run_time = mtime_since_now(&td->epoch);
1866         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1867         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1868         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1869
1870         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1871             (td->o.verify != VERIFY_NONE && td_write(td)))
1872                 verify_save_state(td->thread_number);
1873
1874         fio_unpin_memory(td);
1875
1876         td_writeout_logs(td, true);
1877
1878         iolog_compress_exit(td);
1879         rate_submit_exit(td);
1880
1881         if (o->exec_postrun)
1882                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1883
1884         if (exitall_on_terminate || (o->exitall_error && td->error))
1885                 fio_terminate_threads(td->groupid);
1886
1887 err:
1888         if (td->error)
1889                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1890                                                         td->verror);
1891
1892         if (o->verify_async)
1893                 verify_async_exit(td);
1894
1895         close_and_free_files(td);
1896         cleanup_io_u(td);
1897         close_ioengine(td);
1898         cgroup_shutdown(td, &cgroup_mnt);
1899         verify_free_state(td);
1900
1901         if (td->zone_state_index) {
1902                 int i;
1903
1904                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1905                         free(td->zone_state_index[i]);
1906                 free(td->zone_state_index);
1907                 td->zone_state_index = NULL;
1908         }
1909
1910         if (fio_option_is_set(o, cpumask)) {
1911                 ret = fio_cpuset_exit(&o->cpumask);
1912                 if (ret)
1913                         td_verror(td, ret, "fio_cpuset_exit");
1914         }
1915
1916         /*
1917          * do this very late, it will log file closing as well
1918          */
1919         if (o->write_iolog_file)
1920                 write_iolog_close(td);
1921
1922         td_set_runstate(td, TD_EXITED);
1923
1924         /*
1925          * Do this last after setting our runstate to exited, so we
1926          * know that the stat thread is signaled.
1927          */
1928         check_update_rusage(td);
1929
1930         sk_out_drop();
1931         return (void *) (uintptr_t) td->error;
1932 }
1933
1934 /*
1935  * Run over the job map and reap the threads that have exited, if any.
1936  */
1937 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1938                          uint64_t *m_rate)
1939 {
1940         struct thread_data *td;
1941         unsigned int cputhreads, realthreads, pending;
1942         int i, status, ret;
1943
1944         /*
1945          * reap exited threads (TD_EXITED -> TD_REAPED)
1946          */
1947         realthreads = pending = cputhreads = 0;
1948         for_each_td(td, i) {
1949                 int flags = 0;
1950
1951                  if (!strcmp(td->o.ioengine, "cpuio"))
1952                         cputhreads++;
1953                 else
1954                         realthreads++;
1955
1956                 if (!td->pid) {
1957                         pending++;
1958                         continue;
1959                 }
1960                 if (td->runstate == TD_REAPED)
1961                         continue;
1962                 if (td->o.use_thread) {
1963                         if (td->runstate == TD_EXITED) {
1964                                 td_set_runstate(td, TD_REAPED);
1965                                 goto reaped;
1966                         }
1967                         continue;
1968                 }
1969
1970                 flags = WNOHANG;
1971                 if (td->runstate == TD_EXITED)
1972                         flags = 0;
1973
1974                 /*
1975                  * check if someone quit or got killed in an unusual way
1976                  */
1977                 ret = waitpid(td->pid, &status, flags);
1978                 if (ret < 0) {
1979                         if (errno == ECHILD) {
1980                                 log_err("fio: pid=%d disappeared %d\n",
1981                                                 (int) td->pid, td->runstate);
1982                                 td->sig = ECHILD;
1983                                 td_set_runstate(td, TD_REAPED);
1984                                 goto reaped;
1985                         }
1986                         perror("waitpid");
1987                 } else if (ret == td->pid) {
1988                         if (WIFSIGNALED(status)) {
1989                                 int sig = WTERMSIG(status);
1990
1991                                 if (sig != SIGTERM && sig != SIGUSR2)
1992                                         log_err("fio: pid=%d, got signal=%d\n",
1993                                                         (int) td->pid, sig);
1994                                 td->sig = sig;
1995                                 td_set_runstate(td, TD_REAPED);
1996                                 goto reaped;
1997                         }
1998                         if (WIFEXITED(status)) {
1999                                 if (WEXITSTATUS(status) && !td->error)
2000                                         td->error = WEXITSTATUS(status);
2001
2002                                 td_set_runstate(td, TD_REAPED);
2003                                 goto reaped;
2004                         }
2005                 }
2006
2007                 /*
2008                  * If the job is stuck, do a forceful timeout of it and
2009                  * move on.
2010                  */
2011                 if (td->terminate &&
2012                     td->runstate < TD_FSYNCING &&
2013                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2014                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2015                                 "%lu seconds, it appears to be stuck. Doing "
2016                                 "forceful exit of this job.\n",
2017                                 td->o.name, td->runstate,
2018                                 (unsigned long) time_since_now(&td->terminate_time));
2019                         td_set_runstate(td, TD_REAPED);
2020                         goto reaped;
2021                 }
2022
2023                 /*
2024                  * thread is not dead, continue
2025                  */
2026                 pending++;
2027                 continue;
2028 reaped:
2029                 (*nr_running)--;
2030                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2031                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2032                 if (!td->pid)
2033                         pending--;
2034
2035                 if (td->error)
2036                         exit_value++;
2037
2038                 done_secs += mtime_since_now(&td->epoch) / 1000;
2039                 profile_td_exit(td);
2040         }
2041
2042         if (*nr_running == cputhreads && !pending && realthreads)
2043                 fio_terminate_threads(TERMINATE_ALL);
2044 }
2045
2046 static bool __check_trigger_file(void)
2047 {
2048         struct stat sb;
2049
2050         if (!trigger_file)
2051                 return false;
2052
2053         if (stat(trigger_file, &sb))
2054                 return false;
2055
2056         if (unlink(trigger_file) < 0)
2057                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2058                                                         strerror(errno));
2059
2060         return true;
2061 }
2062
2063 static bool trigger_timedout(void)
2064 {
2065         if (trigger_timeout)
2066                 if (time_since_genesis() >= trigger_timeout) {
2067                         trigger_timeout = 0;
2068                         return true;
2069                 }
2070
2071         return false;
2072 }
2073
2074 void exec_trigger(const char *cmd)
2075 {
2076         int ret;
2077
2078         if (!cmd || cmd[0] == '\0')
2079                 return;
2080
2081         ret = system(cmd);
2082         if (ret == -1)
2083                 log_err("fio: failed executing %s trigger\n", cmd);
2084 }
2085
2086 void check_trigger_file(void)
2087 {
2088         if (__check_trigger_file() || trigger_timedout()) {
2089                 if (nr_clients)
2090                         fio_clients_send_trigger(trigger_remote_cmd);
2091                 else {
2092                         verify_save_state(IO_LIST_ALL);
2093                         fio_terminate_threads(TERMINATE_ALL);
2094                         exec_trigger(trigger_cmd);
2095                 }
2096         }
2097 }
2098
2099 static int fio_verify_load_state(struct thread_data *td)
2100 {
2101         int ret;
2102
2103         if (!td->o.verify_state)
2104                 return 0;
2105
2106         if (is_backend) {
2107                 void *data;
2108
2109                 ret = fio_server_get_verify_state(td->o.name,
2110                                         td->thread_number - 1, &data);
2111                 if (!ret)
2112                         verify_assign_state(td, data);
2113         } else
2114                 ret = verify_load_state(td, "local");
2115
2116         return ret;
2117 }
2118
2119 static void do_usleep(unsigned int usecs)
2120 {
2121         check_for_running_stats();
2122         check_trigger_file();
2123         usleep(usecs);
2124 }
2125
2126 static bool check_mount_writes(struct thread_data *td)
2127 {
2128         struct fio_file *f;
2129         unsigned int i;
2130
2131         if (!td_write(td) || td->o.allow_mounted_write)
2132                 return false;
2133
2134         /*
2135          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2136          * are mkfs'd and mounted.
2137          */
2138         for_each_file(td, f, i) {
2139 #ifdef FIO_HAVE_CHARDEV_SIZE
2140                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2141 #else
2142                 if (f->filetype != FIO_TYPE_BLOCK)
2143 #endif
2144                         continue;
2145                 if (device_is_mounted(f->file_name))
2146                         goto mounted;
2147         }
2148
2149         return false;
2150 mounted:
2151         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2152         return true;
2153 }
2154
2155 static bool waitee_running(struct thread_data *me)
2156 {
2157         const char *waitee = me->o.wait_for;
2158         const char *self = me->o.name;
2159         struct thread_data *td;
2160         int i;
2161
2162         if (!waitee)
2163                 return false;
2164
2165         for_each_td(td, i) {
2166                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2167                         continue;
2168
2169                 if (td->runstate < TD_EXITED) {
2170                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2171                                         self, td->o.name,
2172                                         runstate_to_name(td->runstate));
2173                         return true;
2174                 }
2175         }
2176
2177         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2178         return false;
2179 }
2180
2181 /*
2182  * Main function for kicking off and reaping jobs, as needed.
2183  */
2184 static void run_threads(struct sk_out *sk_out)
2185 {
2186         struct thread_data *td;
2187         unsigned int i, todo, nr_running, nr_started;
2188         uint64_t m_rate, t_rate;
2189         uint64_t spent;
2190
2191         if (fio_gtod_offload && fio_start_gtod_thread())
2192                 return;
2193
2194         fio_idle_prof_init();
2195
2196         set_sig_handlers();
2197
2198         nr_thread = nr_process = 0;
2199         for_each_td(td, i) {
2200                 if (check_mount_writes(td))
2201                         return;
2202                 if (td->o.use_thread)
2203                         nr_thread++;
2204                 else
2205                         nr_process++;
2206         }
2207
2208         if (output_format & FIO_OUTPUT_NORMAL) {
2209                 log_info("Starting ");
2210                 if (nr_thread)
2211                         log_info("%d thread%s", nr_thread,
2212                                                 nr_thread > 1 ? "s" : "");
2213                 if (nr_process) {
2214                         if (nr_thread)
2215                                 log_info(" and ");
2216                         log_info("%d process%s", nr_process,
2217                                                 nr_process > 1 ? "es" : "");
2218                 }
2219                 log_info("\n");
2220                 log_info_flush();
2221         }
2222
2223         todo = thread_number;
2224         nr_running = 0;
2225         nr_started = 0;
2226         m_rate = t_rate = 0;
2227
2228         for_each_td(td, i) {
2229                 print_status_init(td->thread_number - 1);
2230
2231                 if (!td->o.create_serialize)
2232                         continue;
2233
2234                 if (fio_verify_load_state(td))
2235                         goto reap;
2236
2237                 /*
2238                  * do file setup here so it happens sequentially,
2239                  * we don't want X number of threads getting their
2240                  * client data interspersed on disk
2241                  */
2242                 if (setup_files(td)) {
2243 reap:
2244                         exit_value++;
2245                         if (td->error)
2246                                 log_err("fio: pid=%d, err=%d/%s\n",
2247                                         (int) td->pid, td->error, td->verror);
2248                         td_set_runstate(td, TD_REAPED);
2249                         todo--;
2250                 } else {
2251                         struct fio_file *f;
2252                         unsigned int j;
2253
2254                         /*
2255                          * for sharing to work, each job must always open
2256                          * its own files. so close them, if we opened them
2257                          * for creation
2258                          */
2259                         for_each_file(td, f, j) {
2260                                 if (fio_file_open(f))
2261                                         td_io_close_file(td, f);
2262                         }
2263                 }
2264         }
2265
2266         /* start idle threads before io threads start to run */
2267         fio_idle_prof_start();
2268
2269         set_genesis_time();
2270
2271         while (todo) {
2272                 struct thread_data *map[REAL_MAX_JOBS];
2273                 struct timespec this_start;
2274                 int this_jobs = 0, left;
2275                 struct fork_data *fd;
2276
2277                 /*
2278                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2279                  */
2280                 for_each_td(td, i) {
2281                         if (td->runstate != TD_NOT_CREATED)
2282                                 continue;
2283
2284                         /*
2285                          * never got a chance to start, killed by other
2286                          * thread for some reason
2287                          */
2288                         if (td->terminate) {
2289                                 todo--;
2290                                 continue;
2291                         }
2292
2293                         if (td->o.start_delay) {
2294                                 spent = utime_since_genesis();
2295
2296                                 if (td->o.start_delay > spent)
2297                                         continue;
2298                         }
2299
2300                         if (td->o.stonewall && (nr_started || nr_running)) {
2301                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2302                                                         td->o.name);
2303                                 break;
2304                         }
2305
2306                         if (waitee_running(td)) {
2307                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2308                                                 td->o.name, td->o.wait_for);
2309                                 continue;
2310                         }
2311
2312                         init_disk_util(td);
2313
2314                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2315                         td->update_rusage = 0;
2316
2317                         /*
2318                          * Set state to created. Thread will transition
2319                          * to TD_INITIALIZED when it's done setting up.
2320                          */
2321                         td_set_runstate(td, TD_CREATED);
2322                         map[this_jobs++] = td;
2323                         nr_started++;
2324
2325                         fd = calloc(1, sizeof(*fd));
2326                         fd->td = td;
2327                         fd->sk_out = sk_out;
2328
2329                         if (td->o.use_thread) {
2330                                 int ret;
2331
2332                                 dprint(FD_PROCESS, "will pthread_create\n");
2333                                 ret = pthread_create(&td->thread, NULL,
2334                                                         thread_main, fd);
2335                                 if (ret) {
2336                                         log_err("pthread_create: %s\n",
2337                                                         strerror(ret));
2338                                         free(fd);
2339                                         nr_started--;
2340                                         break;
2341                                 }
2342                                 fd = NULL;
2343                                 ret = pthread_detach(td->thread);
2344                                 if (ret)
2345                                         log_err("pthread_detach: %s",
2346                                                         strerror(ret));
2347                         } else {
2348                                 pid_t pid;
2349                                 dprint(FD_PROCESS, "will fork\n");
2350                                 pid = fork();
2351                                 if (!pid) {
2352                                         int ret;
2353
2354                                         ret = (int)(uintptr_t)thread_main(fd);
2355                                         _exit(ret);
2356                                 } else if (i == fio_debug_jobno)
2357                                         *fio_debug_jobp = pid;
2358                         }
2359                         dprint(FD_MUTEX, "wait on startup_sem\n");
2360                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2361                                 log_err("fio: job startup hung? exiting.\n");
2362                                 fio_terminate_threads(TERMINATE_ALL);
2363                                 fio_abort = 1;
2364                                 nr_started--;
2365                                 free(fd);
2366                                 break;
2367                         }
2368                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2369                 }
2370
2371                 /*
2372                  * Wait for the started threads to transition to
2373                  * TD_INITIALIZED.
2374                  */
2375                 fio_gettime(&this_start, NULL);
2376                 left = this_jobs;
2377                 while (left && !fio_abort) {
2378                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2379                                 break;
2380
2381                         do_usleep(100000);
2382
2383                         for (i = 0; i < this_jobs; i++) {
2384                                 td = map[i];
2385                                 if (!td)
2386                                         continue;
2387                                 if (td->runstate == TD_INITIALIZED) {
2388                                         map[i] = NULL;
2389                                         left--;
2390                                 } else if (td->runstate >= TD_EXITED) {
2391                                         map[i] = NULL;
2392                                         left--;
2393                                         todo--;
2394                                         nr_running++; /* work-around... */
2395                                 }
2396                         }
2397                 }
2398
2399                 if (left) {
2400                         log_err("fio: %d job%s failed to start\n", left,
2401                                         left > 1 ? "s" : "");
2402                         for (i = 0; i < this_jobs; i++) {
2403                                 td = map[i];
2404                                 if (!td)
2405                                         continue;
2406                                 kill(td->pid, SIGTERM);
2407                         }
2408                         break;
2409                 }
2410
2411                 /*
2412                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2413                  */
2414                 for_each_td(td, i) {
2415                         if (td->runstate != TD_INITIALIZED)
2416                                 continue;
2417
2418                         if (in_ramp_time(td))
2419                                 td_set_runstate(td, TD_RAMP);
2420                         else
2421                                 td_set_runstate(td, TD_RUNNING);
2422                         nr_running++;
2423                         nr_started--;
2424                         m_rate += ddir_rw_sum(td->o.ratemin);
2425                         t_rate += ddir_rw_sum(td->o.rate);
2426                         todo--;
2427                         fio_sem_up(td->sem);
2428                 }
2429
2430                 reap_threads(&nr_running, &t_rate, &m_rate);
2431
2432                 if (todo)
2433                         do_usleep(100000);
2434         }
2435
2436         while (nr_running) {
2437                 reap_threads(&nr_running, &t_rate, &m_rate);
2438                 do_usleep(10000);
2439         }
2440
2441         fio_idle_prof_stop();
2442
2443         update_io_ticks();
2444 }
2445
2446 static void free_disk_util(void)
2447 {
2448         disk_util_prune_entries();
2449         helper_thread_destroy();
2450 }
2451
2452 int fio_backend(struct sk_out *sk_out)
2453 {
2454         struct thread_data *td;
2455         int i;
2456
2457         if (exec_profile) {
2458                 if (load_profile(exec_profile))
2459                         return 1;
2460                 free(exec_profile);
2461                 exec_profile = NULL;
2462         }
2463         if (!thread_number)
2464                 return 0;
2465
2466         if (write_bw_log) {
2467                 struct log_params p = {
2468                         .log_type = IO_LOG_TYPE_BW,
2469                 };
2470
2471                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2472                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2473                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2474         }
2475
2476         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2477         if (startup_sem == NULL)
2478                 return 1;
2479
2480         set_genesis_time();
2481         stat_init();
2482         helper_thread_create(startup_sem, sk_out);
2483
2484         cgroup_list = smalloc(sizeof(*cgroup_list));
2485         if (cgroup_list)
2486                 INIT_FLIST_HEAD(cgroup_list);
2487
2488         run_threads(sk_out);
2489
2490         helper_thread_exit();
2491
2492         if (!fio_abort) {
2493                 __show_run_stats();
2494                 if (write_bw_log) {
2495                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2496                                 struct io_log *log = agg_io_log[i];
2497
2498                                 flush_log(log, false);
2499                                 free_log(log);
2500                         }
2501                 }
2502         }
2503
2504         for_each_td(td, i) {
2505                 steadystate_free(td);
2506                 fio_options_free(td);
2507                 if (td->rusage_sem) {
2508                         fio_sem_remove(td->rusage_sem);
2509                         td->rusage_sem = NULL;
2510                 }
2511                 fio_sem_remove(td->sem);
2512                 td->sem = NULL;
2513         }
2514
2515         free_disk_util();
2516         if (cgroup_list) {
2517                 cgroup_kill(cgroup_list);
2518                 sfree(cgroup_list);
2519         }
2520         sfree(cgroup_mnt);
2521
2522         fio_sem_remove(startup_sem);
2523         stat_exit();
2524         return exit_value;
2525 }