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