backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32 #include <pthread.h>
  33
  34 #include "fio.h"
  35 #include "smalloc.h"
  36 #include "verify.h"
  37 #include "diskutil.h"
  38 #include "cgroup.h"
  39 #include "profile.h"
  40 #include "lib/rand.h"
  41 #include "lib/memalign.h"
  42 #include "server.h"
  43 #include "lib/getrusage.h"
  44 #include "idletime.h"
  45 #include "err.h"
  46 #include "workqueue.h"
  47 #include "lib/mountcheck.h"
  48 #include "rate-submit.h"
  49 #include "helper_thread.h"
  50 #include "pshared.h"
  51 #include "zone-dist.h"
  52
  53 static struct fio_sem *startup_sem;
  54 static struct flist_head *cgroup_list;
  55 static struct cgroup_mnt *cgroup_mnt;
  56 static int exit_value;
  57 static volatile bool fio_abort;
  58 static unsigned int nr_process = 0;
  59 static unsigned int nr_thread = 0;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int stat_number = 0;
  66 int shm_id = 0;
  67 int temp_stall_ts;
  68 unsigned long done_secs = 0;
  69 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  70
  71 #define JOB_START_TIMEOUT       (5 * 1000)
  72
  73 static void sig_int(int sig)
  74 {
  75         if (threads) {
  76                 if (is_backend)
  77                         fio_server_got_signal(sig);
  78                 else {
  79                         log_info("\nfio: terminating on signal %d\n", sig);
  80                         log_info_flush();
  81                         exit_value = 128;
  82                 }
  83
  84                 fio_terminate_threads(TERMINATE_ALL);
  85         }
  86 }
  87
  88 void sig_show_status(int sig)
  89 {
  90         show_running_run_stats();
  91 }
  92
  93 static void set_sig_handlers(void)
  94 {
  95         struct sigaction act;
  96
  97         memset(&act, 0, sizeof(act));
  98         act.sa_handler = sig_int;
  99         act.sa_flags = SA_RESTART;
 100         sigaction(SIGINT, &act, NULL);
 101
 102         memset(&act, 0, sizeof(act));
 103         act.sa_handler = sig_int;
 104         act.sa_flags = SA_RESTART;
 105         sigaction(SIGTERM, &act, NULL);
 106
 107 /* Windows uses SIGBREAK as a quit signal from other applications */
 108 #ifdef WIN32
 109         memset(&act, 0, sizeof(act));
 110         act.sa_handler = sig_int;
 111         act.sa_flags = SA_RESTART;
 112         sigaction(SIGBREAK, &act, NULL);
 113 #endif
 114
 115         memset(&act, 0, sizeof(act));
 116         act.sa_handler = sig_show_status;
 117         act.sa_flags = SA_RESTART;
 118         sigaction(SIGUSR1, &act, NULL);
 119
 120         if (is_backend) {
 121                 memset(&act, 0, sizeof(act));
 122                 act.sa_handler = sig_int;
 123                 act.sa_flags = SA_RESTART;
 124                 sigaction(SIGPIPE, &act, NULL);
 125         }
 126 }
 127
 128 /*
 129  * Check if we are above the minimum rate given.
 130  */
 131 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 132                              enum fio_ddir ddir)
 133 {
 134         unsigned long long bytes = 0;
 135         unsigned long iops = 0;
 136         unsigned long spent;
 137         unsigned long rate;
 138         unsigned int ratemin = 0;
 139         unsigned int rate_iops = 0;
 140         unsigned int rate_iops_min = 0;
 141
 142         assert(ddir_rw(ddir));
 143
 144         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 145                 return false;
 146
 147         /*
 148          * allow a 2 second settle period in the beginning
 149          */
 150         if (mtime_since(&td->start, now) < 2000)
 151                 return false;
 152
 153         iops += td->this_io_blocks[ddir];
 154         bytes += td->this_io_bytes[ddir];
 155         ratemin += td->o.ratemin[ddir];
 156         rate_iops += td->o.rate_iops[ddir];
 157         rate_iops_min += td->o.rate_iops_min[ddir];
 158
 159         /*
 160          * if rate blocks is set, sample is running
 161          */
 162         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 163                 spent = mtime_since(&td->lastrate[ddir], now);
 164                 if (spent < td->o.ratecycle)
 165                         return false;
 166
 167                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 168                         /*
 169                          * check bandwidth specified rate
 170                          */
 171                         if (bytes < td->rate_bytes[ddir]) {
 172                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 173                                         td->o.name, ratemin, bytes);
 174                                 return true;
 175                         } else {
 176                                 if (spent)
 177                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 178                                 else
 179                                         rate = 0;
 180
 181                                 if (rate < ratemin ||
 182                                     bytes < td->rate_bytes[ddir]) {
 183                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 184                                                 td->o.name, ratemin, rate);
 185                                         return true;
 186                                 }
 187                         }
 188                 } else {
 189                         /*
 190                          * checks iops specified rate
 191                          */
 192                         if (iops < rate_iops) {
 193                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 194                                                 td->o.name, rate_iops, iops);
 195                                 return true;
 196                         } else {
 197                                 if (spent)
 198                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 199                                 else
 200                                         rate = 0;
 201
 202                                 if (rate < rate_iops_min ||
 203                                     iops < td->rate_blocks[ddir]) {
 204                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 205                                                 td->o.name, rate_iops_min, rate);
 206                                         return true;
 207                                 }
 208                         }
 209                 }
 210         }
 211
 212         td->rate_bytes[ddir] = bytes;
 213         td->rate_blocks[ddir] = iops;
 214         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 215         return false;
 216 }
 217
 218 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 219 {
 220         bool ret = false;
 221
 222         if (td->bytes_done[DDIR_READ])
 223                 ret |= __check_min_rate(td, now, DDIR_READ);
 224         if (td->bytes_done[DDIR_WRITE])
 225                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 226         if (td->bytes_done[DDIR_TRIM])
 227                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 228
 229         return ret;
 230 }
 231
 232 /*
 233  * When job exits, we can cancel the in-flight IO if we are using async
 234  * io. Attempt to do so.
 235  */
 236 static void cleanup_pending_aio(struct thread_data *td)
 237 {
 238         int r;
 239
 240         if (td->error)
 241                 return;
 242
 243         /*
 244          * get immediately available events, if any
 245          */
 246         r = io_u_queued_complete(td, 0);
 247         if (r < 0)
 248                 return;
 249
 250         /*
 251          * now cancel remaining active events
 252          */
 253         if (td->io_ops->cancel) {
 254                 struct io_u *io_u;
 255                 int i;
 256
 257                 io_u_qiter(&td->io_u_all, io_u, i) {
 258                         if (io_u->flags & IO_U_F_FLIGHT) {
 259                                 r = td->io_ops->cancel(td, io_u);
 260                                 if (!r)
 261                                         put_io_u(td, io_u);
 262                         }
 263                 }
 264         }
 265
 266         if (td->cur_depth)
 267                 r = io_u_queued_complete(td, td->cur_depth);
 268 }
 269
 270 /*
 271  * Helper to handle the final sync of a file. Works just like the normal
 272  * io path, just does everything sync.
 273  */
 274 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 275 {
 276         struct io_u *io_u = __get_io_u(td);
 277         enum fio_q_status ret;
 278
 279         if (!io_u)
 280                 return true;
 281
 282         io_u->ddir = DDIR_SYNC;
 283         io_u->file = f;
 284         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 285
 286         if (td_io_prep(td, io_u)) {
 287                 put_io_u(td, io_u);
 288                 return true;
 289         }
 290
 291 requeue:
 292         ret = td_io_queue(td, io_u);
 293         switch (ret) {
 294         case FIO_Q_QUEUED:
 295                 td_io_commit(td);
 296                 if (io_u_queued_complete(td, 1) < 0)
 297                         return true;
 298                 break;
 299         case FIO_Q_COMPLETED:
 300                 if (io_u->error) {
 301                         td_verror(td, io_u->error, "td_io_queue");
 302                         return true;
 303                 }
 304
 305                 if (io_u_sync_complete(td, io_u) < 0)
 306                         return true;
 307                 break;
 308         case FIO_Q_BUSY:
 309                 td_io_commit(td);
 310                 goto requeue;
 311         }
 312
 313         return false;
 314 }
 315
 316 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 317 {
 318         int ret, ret2;
 319
 320         if (fio_file_open(f))
 321                 return fio_io_sync(td, f);
 322
 323         if (td_io_open_file(td, f))
 324                 return 1;
 325
 326         ret = fio_io_sync(td, f);
 327         ret2 = 0;
 328         if (fio_file_open(f))
 329                 ret2 = td_io_close_file(td, f);
 330         return (ret || ret2);
 331 }
 332
 333 static inline void __update_ts_cache(struct thread_data *td)
 334 {
 335         fio_gettime(&td->ts_cache, NULL);
 336 }
 337
 338 static inline void update_ts_cache(struct thread_data *td)
 339 {
 340         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 341                 __update_ts_cache(td);
 342 }
 343
 344 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 345 {
 346         if (in_ramp_time(td))
 347                 return false;
 348         if (!td->o.timeout)
 349                 return false;
 350         if (utime_since(&td->epoch, t) >= td->o.timeout)
 351                 return true;
 352
 353         return false;
 354 }
 355
 356 /*
 357  * We need to update the runtime consistently in ms, but keep a running
 358  * tally of the current elapsed time in microseconds for sub millisecond
 359  * updates.
 360  */
 361 static inline void update_runtime(struct thread_data *td,
 362                                   unsigned long long *elapsed_us,
 363                                   const enum fio_ddir ddir)
 364 {
 365         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 366                 return;
 367
 368         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 369         elapsed_us[ddir] += utime_since_now(&td->start);
 370         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 371 }
 372
 373 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 374                                 int *retptr)
 375 {
 376         int ret = *retptr;
 377
 378         if (ret < 0 || td->error) {
 379                 int err = td->error;
 380                 enum error_type_bit eb;
 381
 382                 if (ret < 0)
 383                         err = -ret;
 384
 385                 eb = td_error_type(ddir, err);
 386                 if (!(td->o.continue_on_error & (1 << eb)))
 387                         return true;
 388
 389                 if (td_non_fatal_error(td, eb, err)) {
 390                         /*
 391                          * Continue with the I/Os in case of
 392                          * a non fatal error.
 393                          */
 394                         update_error_count(td, err);
 395                         td_clear_error(td);
 396                         *retptr = 0;
 397                         return false;
 398                 } else if (td->o.fill_device && err == ENOSPC) {
 399                         /*
 400                          * We expect to hit this error if
 401                          * fill_device option is set.
 402                          */
 403                         td_clear_error(td);
 404                         fio_mark_td_terminate(td);
 405                         return true;
 406                 } else {
 407                         /*
 408                          * Stop the I/O in case of a fatal
 409                          * error.
 410                          */
 411                         update_error_count(td, err);
 412                         return true;
 413                 }
 414         }
 415
 416         return false;
 417 }
 418
 419 static void check_update_rusage(struct thread_data *td)
 420 {
 421         if (td->update_rusage) {
 422                 td->update_rusage = 0;
 423                 update_rusage_stat(td);
 424                 fio_sem_up(td->rusage_sem);
 425         }
 426 }
 427
 428 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 429 {
 430         const int full = queue_full(td);
 431         int min_evts = 0;
 432         int ret;
 433
 434         if (td->flags & TD_F_REGROW_LOGS)
 435                 return io_u_quiesce(td);
 436
 437         /*
 438          * if the queue is full, we MUST reap at least 1 event
 439          */
 440         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 441         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 442                 min_evts = 1;
 443
 444         if (time && __should_check_rate(td))
 445                 fio_gettime(time, NULL);
 446
 447         do {
 448                 ret = io_u_queued_complete(td, min_evts);
 449                 if (ret < 0)
 450                         break;
 451         } while (full && (td->cur_depth > td->o.iodepth_low));
 452
 453         return ret;
 454 }
 455
 456 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 457                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 458                    struct timespec *comp_time)
 459 {
 460         switch (*ret) {
 461         case FIO_Q_COMPLETED:
 462                 if (io_u->error) {
 463                         *ret = -io_u->error;
 464                         clear_io_u(td, io_u);
 465                 } else if (io_u->resid) {
 466                         long long bytes = io_u->xfer_buflen - io_u->resid;
 467                         struct fio_file *f = io_u->file;
 468
 469                         if (bytes_issued)
 470                                 *bytes_issued += bytes;
 471
 472                         if (!from_verify)
 473                                 trim_io_piece(io_u);
 474
 475                         /*
 476                          * zero read, fail
 477                          */
 478                         if (!bytes) {
 479                                 if (!from_verify)
 480                                         unlog_io_piece(td, io_u);
 481                                 td_verror(td, EIO, "full resid");
 482                                 put_io_u(td, io_u);
 483                                 break;
 484                         }
 485
 486                         io_u->xfer_buflen = io_u->resid;
 487                         io_u->xfer_buf += bytes;
 488                         io_u->offset += bytes;
 489
 490                         if (ddir_rw(io_u->ddir))
 491                                 td->ts.short_io_u[io_u->ddir]++;
 492
 493                         if (io_u->offset == f->real_file_size)
 494                                 goto sync_done;
 495
 496                         requeue_io_u(td, &io_u);
 497                 } else {
 498 sync_done:
 499                         if (comp_time && __should_check_rate(td))
 500                                 fio_gettime(comp_time, NULL);
 501
 502                         *ret = io_u_sync_complete(td, io_u);
 503                         if (*ret < 0)
 504                                 break;
 505                 }
 506
 507                 if (td->flags & TD_F_REGROW_LOGS)
 508                         regrow_logs(td);
 509
 510                 /*
 511                  * when doing I/O (not when verifying),
 512                  * check for any errors that are to be ignored
 513                  */
 514                 if (!from_verify)
 515                         break;
 516
 517                 return 0;
 518         case FIO_Q_QUEUED:
 519                 /*
 520                  * if the engine doesn't have a commit hook,
 521                  * the io_u is really queued. if it does have such
 522                  * a hook, it has to call io_u_queued() itself.
 523                  */
 524                 if (td->io_ops->commit == NULL)
 525                         io_u_queued(td, io_u);
 526                 if (bytes_issued)
 527                         *bytes_issued += io_u->xfer_buflen;
 528                 break;
 529         case FIO_Q_BUSY:
 530                 if (!from_verify)
 531                         unlog_io_piece(td, io_u);
 532                 requeue_io_u(td, &io_u);
 533                 td_io_commit(td);
 534                 break;
 535         default:
 536                 assert(*ret < 0);
 537                 td_verror(td, -(*ret), "td_io_queue");
 538                 break;
 539         }
 540
 541         if (break_on_this_error(td, ddir, ret))
 542                 return 1;
 543
 544         return 0;
 545 }
 546
 547 static inline bool io_in_polling(struct thread_data *td)
 548 {
 549         return !td->o.iodepth_batch_complete_min &&
 550                    !td->o.iodepth_batch_complete_max;
 551 }
 552 /*
 553  * Unlinks files from thread data fio_file structure
 554  */
 555 static int unlink_all_files(struct thread_data *td)
 556 {
 557         struct fio_file *f;
 558         unsigned int i;
 559         int ret = 0;
 560
 561         for_each_file(td, f, i) {
 562                 if (f->filetype != FIO_TYPE_FILE)
 563                         continue;
 564                 ret = td_io_unlink_file(td, f);
 565                 if (ret)
 566                         break;
 567         }
 568
 569         if (ret)
 570                 td_verror(td, ret, "unlink_all_files");
 571
 572         return ret;
 573 }
 574
 575 /*
 576  * Check if io_u will overlap an in-flight IO in the queue
 577  */
 578 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 579 {
 580         bool overlap;
 581         struct io_u *check_io_u;
 582         unsigned long long x1, x2, y1, y2;
 583         int i;
 584
 585         x1 = io_u->offset;
 586         x2 = io_u->offset + io_u->buflen;
 587         overlap = false;
 588         io_u_qiter(q, check_io_u, i) {
 589                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 590                         y1 = check_io_u->offset;
 591                         y2 = check_io_u->offset + check_io_u->buflen;
 592
 593                         if (x1 < y2 && y1 < x2) {
 594                                 overlap = true;
 595                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 596                                                 x1, io_u->buflen,
 597                                                 y1, check_io_u->buflen);
 598                                 break;
 599                         }
 600                 }
 601         }
 602
 603         return overlap;
 604 }
 605
 606 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 607 {
 608         /*
 609          * Check for overlap if the user asked us to, and we have
 610          * at least one IO in flight besides this one.
 611          */
 612         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 613             in_flight_overlap(&td->io_u_all, io_u))
 614                 return FIO_Q_BUSY;
 615
 616         return td_io_queue(td, io_u);
 617 }
 618
 619 /*
 620  * The main verify engine. Runs over the writes we previously submitted,
 621  * reads the blocks back in, and checks the crc/md5 of the data.
 622  */
 623 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 624 {
 625         struct fio_file *f;
 626         struct io_u *io_u;
 627         int ret, min_events;
 628         unsigned int i;
 629
 630         dprint(FD_VERIFY, "starting loop\n");
 631
 632         /*
 633          * sync io first and invalidate cache, to make sure we really
 634          * read from disk.
 635          */
 636         for_each_file(td, f, i) {
 637                 if (!fio_file_open(f))
 638                         continue;
 639                 if (fio_io_sync(td, f))
 640                         break;
 641                 if (file_invalidate_cache(td, f))
 642                         break;
 643         }
 644
 645         check_update_rusage(td);
 646
 647         if (td->error)
 648                 return;
 649
 650         /*
 651          * verify_state needs to be reset before verification
 652          * proceeds so that expected random seeds match actual
 653          * random seeds in headers. The main loop will reset
 654          * all random number generators if randrepeat is set.
 655          */
 656         if (!td->o.rand_repeatable)
 657                 td_fill_verify_state_seed(td);
 658
 659         td_set_runstate(td, TD_VERIFYING);
 660
 661         io_u = NULL;
 662         while (!td->terminate) {
 663                 enum fio_ddir ddir;
 664                 int full;
 665
 666                 update_ts_cache(td);
 667                 check_update_rusage(td);
 668
 669                 if (runtime_exceeded(td, &td->ts_cache)) {
 670                         __update_ts_cache(td);
 671                         if (runtime_exceeded(td, &td->ts_cache)) {
 672                                 fio_mark_td_terminate(td);
 673                                 break;
 674                         }
 675                 }
 676
 677                 if (flow_threshold_exceeded(td))
 678                         continue;
 679
 680                 if (!td->o.experimental_verify) {
 681                         io_u = __get_io_u(td);
 682                         if (!io_u)
 683                                 break;
 684
 685                         if (get_next_verify(td, io_u)) {
 686                                 put_io_u(td, io_u);
 687                                 break;
 688                         }
 689
 690                         if (td_io_prep(td, io_u)) {
 691                                 put_io_u(td, io_u);
 692                                 break;
 693                         }
 694                 } else {
 695                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 696                                 break;
 697
 698                         while ((io_u = get_io_u(td)) != NULL) {
 699                                 if (IS_ERR_OR_NULL(io_u)) {
 700                                         io_u = NULL;
 701                                         ret = FIO_Q_BUSY;
 702                                         goto reap;
 703                                 }
 704
 705                                 /*
 706                                  * We are only interested in the places where
 707                                  * we wrote or trimmed IOs. Turn those into
 708                                  * reads for verification purposes.
 709                                  */
 710                                 if (io_u->ddir == DDIR_READ) {
 711                                         /*
 712                                          * Pretend we issued it for rwmix
 713                                          * accounting
 714                                          */
 715                                         td->io_issues[DDIR_READ]++;
 716                                         put_io_u(td, io_u);
 717                                         continue;
 718                                 } else if (io_u->ddir == DDIR_TRIM) {
 719                                         io_u->ddir = DDIR_READ;
 720                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 721                                         break;
 722                                 } else if (io_u->ddir == DDIR_WRITE) {
 723                                         io_u->ddir = DDIR_READ;
 724                                         populate_verify_io_u(td, io_u);
 725                                         break;
 726                                 } else {
 727                                         put_io_u(td, io_u);
 728                                         continue;
 729                                 }
 730                         }
 731
 732                         if (!io_u)
 733                                 break;
 734                 }
 735
 736                 if (verify_state_should_stop(td, io_u)) {
 737                         put_io_u(td, io_u);
 738                         break;
 739                 }
 740
 741                 if (td->o.verify_async)
 742                         io_u->end_io = verify_io_u_async;
 743                 else
 744                         io_u->end_io = verify_io_u;
 745
 746                 ddir = io_u->ddir;
 747                 if (!td->o.disable_slat)
 748                         fio_gettime(&io_u->start_time, NULL);
 749
 750                 ret = io_u_submit(td, io_u);
 751
 752                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 753                         break;
 754
 755                 /*
 756                  * if we can queue more, do so. but check if there are
 757                  * completed io_u's first. Note that we can get BUSY even
 758                  * without IO queued, if the system is resource starved.
 759                  */
 760 reap:
 761                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 762                 if (full || io_in_polling(td))
 763                         ret = wait_for_completions(td, NULL);
 764
 765                 if (ret < 0)
 766                         break;
 767         }
 768
 769         check_update_rusage(td);
 770
 771         if (!td->error) {
 772                 min_events = td->cur_depth;
 773
 774                 if (min_events)
 775                         ret = io_u_queued_complete(td, min_events);
 776         } else
 777                 cleanup_pending_aio(td);
 778
 779         td_set_runstate(td, TD_RUNNING);
 780
 781         dprint(FD_VERIFY, "exiting loop\n");
 782 }
 783
 784 static bool exceeds_number_ios(struct thread_data *td)
 785 {
 786         unsigned long long number_ios;
 787
 788         if (!td->o.number_ios)
 789                 return false;
 790
 791         number_ios = ddir_rw_sum(td->io_blocks);
 792         number_ios += td->io_u_queued + td->io_u_in_flight;
 793
 794         return number_ios >= (td->o.number_ios * td->loops);
 795 }
 796
 797 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 798 {
 799         unsigned long long bytes, limit;
 800
 801         if (td_rw(td))
 802                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 803         else if (td_write(td))
 804                 bytes = this_bytes[DDIR_WRITE];
 805         else if (td_read(td))
 806                 bytes = this_bytes[DDIR_READ];
 807         else
 808                 bytes = this_bytes[DDIR_TRIM];
 809
 810         if (td->o.io_size)
 811                 limit = td->o.io_size;
 812         else
 813                 limit = td->o.size;
 814
 815         limit *= td->loops;
 816         return bytes >= limit || exceeds_number_ios(td);
 817 }
 818
 819 static bool io_issue_bytes_exceeded(struct thread_data *td)
 820 {
 821         return io_bytes_exceeded(td, td->io_issue_bytes);
 822 }
 823
 824 static bool io_complete_bytes_exceeded(struct thread_data *td)
 825 {
 826         return io_bytes_exceeded(td, td->this_io_bytes);
 827 }
 828
 829 /*
 830  * used to calculate the next io time for rate control
 831  *
 832  */
 833 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 834 {
 835         uint64_t bps = td->rate_bps[ddir];
 836
 837         assert(!(td->flags & TD_F_CHILD));
 838
 839         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 840                 uint64_t val, iops;
 841
 842                 iops = bps / td->o.bs[ddir];
 843                 val = (int64_t) (1000000 / iops) *
 844                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 845                 if (val) {
 846                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 847                                         (unsigned long long) 1000000 / val,
 848                                         ddir);
 849                 }
 850                 td->last_usec[ddir] += val;
 851                 return td->last_usec[ddir];
 852         } else if (bps) {
 853                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 854                 uint64_t secs = bytes / bps;
 855                 uint64_t remainder = bytes % bps;
 856
 857                 return remainder * 1000000 / bps + secs * 1000000;
 858         }
 859
 860         return 0;
 861 }
 862
 863 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 864 {
 865         unsigned long long b;
 866         uint64_t total;
 867         int left;
 868
 869         b = ddir_rw_sum(td->io_blocks);
 870         if (b % td->o.thinktime_blocks)
 871                 return;
 872
 873         io_u_quiesce(td);
 874
 875         total = 0;
 876         if (td->o.thinktime_spin)
 877                 total = usec_spin(td->o.thinktime_spin);
 878
 879         left = td->o.thinktime - total;
 880         if (left)
 881                 total += usec_sleep(td, left);
 882
 883         /*
 884          * If we're ignoring thinktime for the rate, add the number of bytes
 885          * we would have done while sleeping, minus one block to ensure we
 886          * start issuing immediately after the sleep.
 887          */
 888         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 889                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 890                 uint64_t bs = td->o.min_bs[ddir];
 891                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 892                 uint64_t over;
 893
 894                 if (usperop <= total)
 895                         over = bs;
 896                 else
 897                         over = (usperop - total) / usperop * -bs;
 898
 899                 td->rate_io_issue_bytes[ddir] += (missed - over);
 900                 /* adjust for rate_process=poisson */
 901                 td->last_usec[ddir] += total;
 902         }
 903 }
 904
 905 /*
 906  * Main IO worker function. It retrieves io_u's to process and queues
 907  * and reaps them, checking for rate and errors along the way.
 908  *
 909  * Returns number of bytes written and trimmed.
 910  */
 911 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 912 {
 913         unsigned int i;
 914         int ret = 0;
 915         uint64_t total_bytes, bytes_issued = 0;
 916
 917         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 918                 bytes_done[i] = td->bytes_done[i];
 919
 920         if (in_ramp_time(td))
 921                 td_set_runstate(td, TD_RAMP);
 922         else
 923                 td_set_runstate(td, TD_RUNNING);
 924
 925         lat_target_init(td);
 926
 927         total_bytes = td->o.size;
 928         /*
 929         * Allow random overwrite workloads to write up to io_size
 930         * before starting verification phase as 'size' doesn't apply.
 931         */
 932         if (td_write(td) && td_random(td) && td->o.norandommap)
 933                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 934         /*
 935          * If verify_backlog is enabled, we'll run the verify in this
 936          * handler as well. For that case, we may need up to twice the
 937          * amount of bytes.
 938          */
 939         if (td->o.verify != VERIFY_NONE &&
 940            (td_write(td) && td->o.verify_backlog))
 941                 total_bytes += td->o.size;
 942
 943         /* In trimwrite mode, each byte is trimmed and then written, so
 944          * allow total_bytes to be twice as big */
 945         if (td_trimwrite(td))
 946                 total_bytes += td->total_io_size;
 947
 948         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 949                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 950                 td->o.time_based) {
 951                 struct timespec comp_time;
 952                 struct io_u *io_u;
 953                 int full;
 954                 enum fio_ddir ddir;
 955
 956                 check_update_rusage(td);
 957
 958                 if (td->terminate || td->done)
 959                         break;
 960
 961                 update_ts_cache(td);
 962
 963                 if (runtime_exceeded(td, &td->ts_cache)) {
 964                         __update_ts_cache(td);
 965                         if (runtime_exceeded(td, &td->ts_cache)) {
 966                                 fio_mark_td_terminate(td);
 967                                 break;
 968                         }
 969                 }
 970
 971                 if (flow_threshold_exceeded(td))
 972                         continue;
 973
 974                 /*
 975                  * Break if we exceeded the bytes. The exception is time
 976                  * based runs, but we still need to break out of the loop
 977                  * for those to run verification, if enabled.
 978                  * Jobs read from iolog do not use this stop condition.
 979                  */
 980                 if (bytes_issued >= total_bytes &&
 981                     !td->o.read_iolog_file &&
 982                     (!td->o.time_based ||
 983                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 984                         break;
 985
 986                 io_u = get_io_u(td);
 987                 if (IS_ERR_OR_NULL(io_u)) {
 988                         int err = PTR_ERR(io_u);
 989
 990                         io_u = NULL;
 991                         ddir = DDIR_INVAL;
 992                         if (err == -EBUSY) {
 993                                 ret = FIO_Q_BUSY;
 994                                 goto reap;
 995                         }
 996                         if (td->o.latency_target)
 997                                 goto reap;
 998                         break;
 999                 }
1000
1001                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1002                         populate_verify_io_u(td, io_u);
1003
1004                 ddir = io_u->ddir;
1005
1006                 /*
1007                  * Add verification end_io handler if:
1008                  *      - Asked to verify (!td_rw(td))
1009                  *      - Or the io_u is from our verify list (mixed write/ver)
1010                  */
1011                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1012                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1013
1014                         if (!td->o.verify_pattern_bytes) {
1015                                 io_u->rand_seed = __rand(&td->verify_state);
1016                                 if (sizeof(int) != sizeof(long *))
1017                                         io_u->rand_seed *= __rand(&td->verify_state);
1018                         }
1019
1020                         if (verify_state_should_stop(td, io_u)) {
1021                                 put_io_u(td, io_u);
1022                                 break;
1023                         }
1024
1025                         if (td->o.verify_async)
1026                                 io_u->end_io = verify_io_u_async;
1027                         else
1028                                 io_u->end_io = verify_io_u;
1029                         td_set_runstate(td, TD_VERIFYING);
1030                 } else if (in_ramp_time(td))
1031                         td_set_runstate(td, TD_RAMP);
1032                 else
1033                         td_set_runstate(td, TD_RUNNING);
1034
1035                 /*
1036                  * Always log IO before it's issued, so we know the specific
1037                  * order of it. The logged unit will track when the IO has
1038                  * completed.
1039                  */
1040                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1041                     td->o.do_verify &&
1042                     td->o.verify != VERIFY_NONE &&
1043                     !td->o.experimental_verify)
1044                         log_io_piece(td, io_u);
1045
1046                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1047                         const unsigned long long blen = io_u->xfer_buflen;
1048                         const enum fio_ddir __ddir = acct_ddir(io_u);
1049
1050                         if (td->error)
1051                                 break;
1052
1053                         workqueue_enqueue(&td->io_wq, &io_u->work);
1054                         ret = FIO_Q_QUEUED;
1055
1056                         if (ddir_rw(__ddir)) {
1057                                 td->io_issues[__ddir]++;
1058                                 td->io_issue_bytes[__ddir] += blen;
1059                                 td->rate_io_issue_bytes[__ddir] += blen;
1060                         }
1061
1062                         if (should_check_rate(td))
1063                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1064
1065                 } else {
1066                         ret = io_u_submit(td, io_u);
1067
1068                         if (should_check_rate(td))
1069                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1070
1071                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1072                                 break;
1073
1074                         /*
1075                          * See if we need to complete some commands. Note that
1076                          * we can get BUSY even without IO queued, if the
1077                          * system is resource starved.
1078                          */
1079 reap:
1080                         full = queue_full(td) ||
1081                                 (ret == FIO_Q_BUSY && td->cur_depth);
1082                         if (full || io_in_polling(td))
1083                                 ret = wait_for_completions(td, &comp_time);
1084                 }
1085                 if (ret < 0)
1086                         break;
1087                 if (!ddir_rw_sum(td->bytes_done) &&
1088                     !td_ioengine_flagged(td, FIO_NOIO))
1089                         continue;
1090
1091                 if (!in_ramp_time(td) && should_check_rate(td)) {
1092                         if (check_min_rate(td, &comp_time)) {
1093                                 if (exitall_on_terminate || td->o.exitall_error)
1094                                         fio_terminate_threads(td->groupid);
1095                                 td_verror(td, EIO, "check_min_rate");
1096                                 break;
1097                         }
1098                 }
1099                 if (!in_ramp_time(td) && td->o.latency_target)
1100                         lat_target_check(td);
1101
1102                 if (ddir_rw(ddir) && td->o.thinktime)
1103                         handle_thinktime(td, ddir);
1104         }
1105
1106         check_update_rusage(td);
1107
1108         if (td->trim_entries)
1109                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1110
1111         if (td->o.fill_device && td->error == ENOSPC) {
1112                 td->error = 0;
1113                 fio_mark_td_terminate(td);
1114         }
1115         if (!td->error) {
1116                 struct fio_file *f;
1117
1118                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1119                         workqueue_flush(&td->io_wq);
1120                         i = 0;
1121                 } else
1122                         i = td->cur_depth;
1123
1124                 if (i) {
1125                         ret = io_u_queued_complete(td, i);
1126                         if (td->o.fill_device && td->error == ENOSPC)
1127                                 td->error = 0;
1128                 }
1129
1130                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1131                         td_set_runstate(td, TD_FSYNCING);
1132
1133                         for_each_file(td, f, i) {
1134                                 if (!fio_file_fsync(td, f))
1135                                         continue;
1136
1137                                 log_err("fio: end_fsync failed for file %s\n",
1138                                                                 f->file_name);
1139                         }
1140                 }
1141         } else
1142                 cleanup_pending_aio(td);
1143
1144         /*
1145          * stop job if we failed doing any IO
1146          */
1147         if (!ddir_rw_sum(td->this_io_bytes))
1148                 td->done = 1;
1149
1150         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1151                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1152 }
1153
1154 static void free_file_completion_logging(struct thread_data *td)
1155 {
1156         struct fio_file *f;
1157         unsigned int i;
1158
1159         for_each_file(td, f, i) {
1160                 if (!f->last_write_comp)
1161                         break;
1162                 sfree(f->last_write_comp);
1163         }
1164 }
1165
1166 static int init_file_completion_logging(struct thread_data *td,
1167                                         unsigned int depth)
1168 {
1169         struct fio_file *f;
1170         unsigned int i;
1171
1172         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1173                 return 0;
1174
1175         for_each_file(td, f, i) {
1176                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1177                 if (!f->last_write_comp)
1178                         goto cleanup;
1179         }
1180
1181         return 0;
1182
1183 cleanup:
1184         free_file_completion_logging(td);
1185         log_err("fio: failed to alloc write comp data\n");
1186         return 1;
1187 }
1188
1189 static void cleanup_io_u(struct thread_data *td)
1190 {
1191         struct io_u *io_u;
1192
1193         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1194
1195                 if (td->io_ops->io_u_free)
1196                         td->io_ops->io_u_free(td, io_u);
1197
1198                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1199         }
1200
1201         free_io_mem(td);
1202
1203         io_u_rexit(&td->io_u_requeues);
1204         io_u_qexit(&td->io_u_freelist, false);
1205         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1206
1207         free_file_completion_logging(td);
1208 }
1209
1210 static int init_io_u(struct thread_data *td)
1211 {
1212         struct io_u *io_u;
1213         int cl_align, i, max_units;
1214         int err;
1215
1216         max_units = td->o.iodepth;
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, false);
1221         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1222
1223         if (err) {
1224                 log_err("fio: failed setting up IO queues\n");
1225                 return 1;
1226         }
1227
1228         cl_align = os_cache_line_size();
1229
1230         for (i = 0; i < max_units; i++) {
1231                 void *ptr;
1232
1233                 if (td->terminate)
1234                         return 1;
1235
1236                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1237                 if (!ptr) {
1238                         log_err("fio: unable to allocate aligned memory\n");
1239                         break;
1240                 }
1241
1242                 io_u = ptr;
1243                 memset(io_u, 0, sizeof(*io_u));
1244                 INIT_FLIST_HEAD(&io_u->verify_list);
1245                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1246
1247                 io_u->index = i;
1248                 io_u->flags = IO_U_F_FREE;
1249                 io_u_qpush(&td->io_u_freelist, io_u);
1250
1251                 /*
1252                  * io_u never leaves this stack, used for iteration of all
1253                  * io_u buffers.
1254                  */
1255                 io_u_qpush(&td->io_u_all, io_u);
1256
1257                 if (td->io_ops->io_u_init) {
1258                         int ret = td->io_ops->io_u_init(td, io_u);
1259
1260                         if (ret) {
1261                                 log_err("fio: failed to init engine data: %d\n", ret);
1262                                 return 1;
1263                         }
1264                 }
1265         }
1266
1267         init_io_u_buffers(td);
1268
1269         if (init_file_completion_logging(td, max_units))
1270                 return 1;
1271
1272         return 0;
1273 }
1274
1275 int init_io_u_buffers(struct thread_data *td)
1276 {
1277         struct io_u *io_u;
1278         unsigned long long max_bs, min_write;
1279         int i, max_units;
1280         int data_xfer = 1;
1281         char *p;
1282
1283         max_units = td->o.iodepth;
1284         max_bs = td_max_bs(td);
1285         min_write = td->o.min_bs[DDIR_WRITE];
1286         td->orig_buffer_size = (unsigned long long) max_bs
1287                                         * (unsigned long long) max_units;
1288
1289         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1290                 data_xfer = 0;
1291
1292         /*
1293          * if we may later need to do address alignment, then add any
1294          * possible adjustment here so that we don't cause a buffer
1295          * overflow later. this adjustment may be too much if we get
1296          * lucky and the allocator gives us an aligned address.
1297          */
1298         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1299             td_ioengine_flagged(td, FIO_RAWIO))
1300                 td->orig_buffer_size += page_mask + td->o.mem_align;
1301
1302         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1303                 unsigned long long bs;
1304
1305                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1306                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1307         }
1308
1309         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1310                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1311                 return 1;
1312         }
1313
1314         if (data_xfer && allocate_io_mem(td))
1315                 return 1;
1316
1317         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1318             td_ioengine_flagged(td, FIO_RAWIO))
1319                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1320         else
1321                 p = td->orig_buffer;
1322
1323         for (i = 0; i < max_units; i++) {
1324                 io_u = td->io_u_all.io_us[i];
1325                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1326
1327                 if (data_xfer) {
1328                         io_u->buf = p;
1329                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1330
1331                         if (td_write(td))
1332                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1333                         if (td_write(td) && td->o.verify_pattern_bytes) {
1334                                 /*
1335                                  * Fill the buffer with the pattern if we are
1336                                  * going to be doing writes.
1337                                  */
1338                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1339                         }
1340                 }
1341                 p += max_bs;
1342         }
1343
1344         return 0;
1345 }
1346
1347 /*
1348  * This function is Linux specific.
1349  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1350  */
1351 static int switch_ioscheduler(struct thread_data *td)
1352 {
1353 #ifdef FIO_HAVE_IOSCHED_SWITCH
1354         char tmp[256], tmp2[128], *p;
1355         FILE *f;
1356         int ret;
1357
1358         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1359                 return 0;
1360
1361         assert(td->files && td->files[0]);
1362         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1363
1364         f = fopen(tmp, "r+");
1365         if (!f) {
1366                 if (errno == ENOENT) {
1367                         log_err("fio: os or kernel doesn't support IO scheduler"
1368                                 " switching\n");
1369                         return 0;
1370                 }
1371                 td_verror(td, errno, "fopen iosched");
1372                 return 1;
1373         }
1374
1375         /*
1376          * Set io scheduler.
1377          */
1378         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1379         if (ferror(f) || ret != 1) {
1380                 td_verror(td, errno, "fwrite");
1381                 fclose(f);
1382                 return 1;
1383         }
1384
1385         rewind(f);
1386
1387         /*
1388          * Read back and check that the selected scheduler is now the default.
1389          */
1390         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1391         if (ferror(f) || ret < 0) {
1392                 td_verror(td, errno, "fread");
1393                 fclose(f);
1394                 return 1;
1395         }
1396         tmp[ret] = '\0';
1397         /*
1398          * either a list of io schedulers or "none\n" is expected. Strip the
1399          * trailing newline.
1400          */
1401         p = tmp;
1402         strsep(&p, "\n");
1403
1404         /*
1405          * Write to "none" entry doesn't fail, so check the result here.
1406          */
1407         if (!strcmp(tmp, "none")) {
1408                 log_err("fio: io scheduler is not tunable\n");
1409                 fclose(f);
1410                 return 0;
1411         }
1412
1413         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1414         if (!strstr(tmp, tmp2)) {
1415                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1416                 td_verror(td, EINVAL, "iosched_switch");
1417                 fclose(f);
1418                 return 1;
1419         }
1420
1421         fclose(f);
1422         return 0;
1423 #else
1424         return 0;
1425 #endif
1426 }
1427
1428 static bool keep_running(struct thread_data *td)
1429 {
1430         unsigned long long limit;
1431
1432         if (td->done)
1433                 return false;
1434         if (td->terminate)
1435                 return false;
1436         if (td->o.time_based)
1437                 return true;
1438         if (td->o.loops) {
1439                 td->o.loops--;
1440                 return true;
1441         }
1442         if (exceeds_number_ios(td))
1443                 return false;
1444
1445         if (td->o.io_size)
1446                 limit = td->o.io_size;
1447         else
1448                 limit = td->o.size;
1449
1450         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1451                 uint64_t diff;
1452
1453                 /*
1454                  * If the difference is less than the maximum IO size, we
1455                  * are done.
1456                  */
1457                 diff = limit - ddir_rw_sum(td->io_bytes);
1458                 if (diff < td_max_bs(td))
1459                         return false;
1460
1461                 if (fio_files_done(td) && !td->o.io_size)
1462                         return false;
1463
1464                 return true;
1465         }
1466
1467         return false;
1468 }
1469
1470 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1471 {
1472         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 13 + 1;
1473         int ret;
1474         char *str;
1475
1476         str = malloc(newlen);
1477         sprintf(str, "%s > %s.%s.txt 2>&1", string, o->name, mode);
1478
1479         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1480         ret = system(str);
1481         if (ret == -1)
1482                 log_err("fio: exec of cmd <%s> failed\n", str);
1483
1484         free(str);
1485         return ret;
1486 }
1487
1488 /*
1489  * Dry run to compute correct state of numberio for verification.
1490  */
1491 static uint64_t do_dry_run(struct thread_data *td)
1492 {
1493         td_set_runstate(td, TD_RUNNING);
1494
1495         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1496                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1497                 struct io_u *io_u;
1498                 int ret;
1499
1500                 if (td->terminate || td->done)
1501                         break;
1502
1503                 io_u = get_io_u(td);
1504                 if (IS_ERR_OR_NULL(io_u))
1505                         break;
1506
1507                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1508                 io_u->error = 0;
1509                 io_u->resid = 0;
1510                 if (ddir_rw(acct_ddir(io_u)))
1511                         td->io_issues[acct_ddir(io_u)]++;
1512                 if (ddir_rw(io_u->ddir)) {
1513                         io_u_mark_depth(td, 1);
1514                         td->ts.total_io_u[io_u->ddir]++;
1515                 }
1516
1517                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1518                     td->o.do_verify &&
1519                     td->o.verify != VERIFY_NONE &&
1520                     !td->o.experimental_verify)
1521                         log_io_piece(td, io_u);
1522
1523                 ret = io_u_sync_complete(td, io_u);
1524                 (void) ret;
1525         }
1526
1527         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1528 }
1529
1530 struct fork_data {
1531         struct thread_data *td;
1532         struct sk_out *sk_out;
1533 };
1534
1535 /*
1536  * Entry point for the thread based jobs. The process based jobs end up
1537  * here as well, after a little setup.
1538  */
1539 static void *thread_main(void *data)
1540 {
1541         struct fork_data *fd = data;
1542         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1543         struct thread_data *td = fd->td;
1544         struct thread_options *o = &td->o;
1545         struct sk_out *sk_out = fd->sk_out;
1546         uint64_t bytes_done[DDIR_RWDIR_CNT];
1547         int deadlock_loop_cnt;
1548         bool clear_state;
1549         int ret;
1550
1551         sk_out_assign(sk_out);
1552         free(fd);
1553
1554         if (!o->use_thread) {
1555                 setsid();
1556                 td->pid = getpid();
1557         } else
1558                 td->pid = gettid();
1559
1560         fio_local_clock_init();
1561
1562         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1563
1564         if (is_backend)
1565                 fio_server_send_start(td);
1566
1567         INIT_FLIST_HEAD(&td->io_log_list);
1568         INIT_FLIST_HEAD(&td->io_hist_list);
1569         INIT_FLIST_HEAD(&td->verify_list);
1570         INIT_FLIST_HEAD(&td->trim_list);
1571         td->io_hist_tree = RB_ROOT;
1572
1573         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1574         if (ret) {
1575                 td_verror(td, ret, "mutex_cond_init_pshared");
1576                 goto err;
1577         }
1578         ret = cond_init_pshared(&td->verify_cond);
1579         if (ret) {
1580                 td_verror(td, ret, "mutex_cond_pshared");
1581                 goto err;
1582         }
1583
1584         td_set_runstate(td, TD_INITIALIZED);
1585         dprint(FD_MUTEX, "up startup_sem\n");
1586         fio_sem_up(startup_sem);
1587         dprint(FD_MUTEX, "wait on td->sem\n");
1588         fio_sem_down(td->sem);
1589         dprint(FD_MUTEX, "done waiting on td->sem\n");
1590
1591         /*
1592          * A new gid requires privilege, so we need to do this before setting
1593          * the uid.
1594          */
1595         if (o->gid != -1U && setgid(o->gid)) {
1596                 td_verror(td, errno, "setgid");
1597                 goto err;
1598         }
1599         if (o->uid != -1U && setuid(o->uid)) {
1600                 td_verror(td, errno, "setuid");
1601                 goto err;
1602         }
1603
1604         td_zone_gen_index(td);
1605
1606         /*
1607          * Do this early, we don't want the compress threads to be limited
1608          * to the same CPUs as the IO workers. So do this before we set
1609          * any potential CPU affinity
1610          */
1611         if (iolog_compress_init(td, sk_out))
1612                 goto err;
1613
1614         /*
1615          * If we have a gettimeofday() thread, make sure we exclude that
1616          * thread from this job
1617          */
1618         if (o->gtod_cpu)
1619                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1620
1621         /*
1622          * Set affinity first, in case it has an impact on the memory
1623          * allocations.
1624          */
1625         if (fio_option_is_set(o, cpumask)) {
1626                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1627                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1628                         if (!ret) {
1629                                 log_err("fio: no CPUs set\n");
1630                                 log_err("fio: Try increasing number of available CPUs\n");
1631                                 td_verror(td, EINVAL, "cpus_split");
1632                                 goto err;
1633                         }
1634                 }
1635                 ret = fio_setaffinity(td->pid, o->cpumask);
1636                 if (ret == -1) {
1637                         td_verror(td, errno, "cpu_set_affinity");
1638                         goto err;
1639                 }
1640         }
1641
1642 #ifdef CONFIG_LIBNUMA
1643         /* numa node setup */
1644         if (fio_option_is_set(o, numa_cpunodes) ||
1645             fio_option_is_set(o, numa_memnodes)) {
1646                 struct bitmask *mask;
1647
1648                 if (numa_available() < 0) {
1649                         td_verror(td, errno, "Does not support NUMA API\n");
1650                         goto err;
1651                 }
1652
1653                 if (fio_option_is_set(o, numa_cpunodes)) {
1654                         mask = numa_parse_nodestring(o->numa_cpunodes);
1655                         ret = numa_run_on_node_mask(mask);
1656                         numa_free_nodemask(mask);
1657                         if (ret == -1) {
1658                                 td_verror(td, errno, \
1659                                         "numa_run_on_node_mask failed\n");
1660                                 goto err;
1661                         }
1662                 }
1663
1664                 if (fio_option_is_set(o, numa_memnodes)) {
1665                         mask = NULL;
1666                         if (o->numa_memnodes)
1667                                 mask = numa_parse_nodestring(o->numa_memnodes);
1668
1669                         switch (o->numa_mem_mode) {
1670                         case MPOL_INTERLEAVE:
1671                                 numa_set_interleave_mask(mask);
1672                                 break;
1673                         case MPOL_BIND:
1674                                 numa_set_membind(mask);
1675                                 break;
1676                         case MPOL_LOCAL:
1677                                 numa_set_localalloc();
1678                                 break;
1679                         case MPOL_PREFERRED:
1680                                 numa_set_preferred(o->numa_mem_prefer_node);
1681                                 break;
1682                         case MPOL_DEFAULT:
1683                         default:
1684                                 break;
1685                         }
1686
1687                         if (mask)
1688                                 numa_free_nodemask(mask);
1689
1690                 }
1691         }
1692 #endif
1693
1694         if (fio_pin_memory(td))
1695                 goto err;
1696
1697         /*
1698          * May alter parameters that init_io_u() will use, so we need to
1699          * do this first.
1700          */
1701         if (!init_iolog(td))
1702                 goto err;
1703
1704         if (td_io_init(td))
1705                 goto err;
1706
1707         if (init_io_u(td))
1708                 goto err;
1709
1710         if (td->io_ops->post_init && td->io_ops->post_init(td))
1711                 goto err;
1712
1713         if (o->verify_async && verify_async_init(td))
1714                 goto err;
1715
1716         if (fio_option_is_set(o, ioprio) ||
1717             fio_option_is_set(o, ioprio_class)) {
1718                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1719                 if (ret == -1) {
1720                         td_verror(td, errno, "ioprio_set");
1721                         goto err;
1722                 }
1723         }
1724
1725         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1726                 goto err;
1727
1728         errno = 0;
1729         if (nice(o->nice) == -1 && errno != 0) {
1730                 td_verror(td, errno, "nice");
1731                 goto err;
1732         }
1733
1734         if (o->ioscheduler && switch_ioscheduler(td))
1735                 goto err;
1736
1737         if (!o->create_serialize && setup_files(td))
1738                 goto err;
1739
1740         if (!init_random_map(td))
1741                 goto err;
1742
1743         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1744                 goto err;
1745
1746         if (o->pre_read && !pre_read_files(td))
1747                 goto err;
1748
1749         fio_verify_init(td);
1750
1751         if (rate_submit_init(td, sk_out))
1752                 goto err;
1753
1754         set_epoch_time(td, o->log_unix_epoch);
1755         fio_getrusage(&td->ru_start);
1756         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1757         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1758         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1759
1760         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1761                         o->ratemin[DDIR_TRIM]) {
1762                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1763                                         sizeof(td->bw_sample_time));
1764                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1765                                         sizeof(td->bw_sample_time));
1766                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1767                                         sizeof(td->bw_sample_time));
1768         }
1769
1770         memset(bytes_done, 0, sizeof(bytes_done));
1771         clear_state = false;
1772
1773         while (keep_running(td)) {
1774                 uint64_t verify_bytes;
1775
1776                 fio_gettime(&td->start, NULL);
1777                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1778
1779                 if (clear_state) {
1780                         clear_io_state(td, 0);
1781
1782                         if (o->unlink_each_loop && unlink_all_files(td))
1783                                 break;
1784                 }
1785
1786                 prune_io_piece_log(td);
1787
1788                 if (td->o.verify_only && td_write(td))
1789                         verify_bytes = do_dry_run(td);
1790                 else {
1791                         do_io(td, bytes_done);
1792
1793                         if (!ddir_rw_sum(bytes_done)) {
1794                                 fio_mark_td_terminate(td);
1795                                 verify_bytes = 0;
1796                         } else {
1797                                 verify_bytes = bytes_done[DDIR_WRITE] +
1798                                                 bytes_done[DDIR_TRIM];
1799                         }
1800                 }
1801
1802                 /*
1803                  * If we took too long to shut down, the main thread could
1804                  * already consider us reaped/exited. If that happens, break
1805                  * out and clean up.
1806                  */
1807                 if (td->runstate >= TD_EXITED)
1808                         break;
1809
1810                 clear_state = true;
1811
1812                 /*
1813                  * Make sure we've successfully updated the rusage stats
1814                  * before waiting on the stat mutex. Otherwise we could have
1815                  * the stat thread holding stat mutex and waiting for
1816                  * the rusage_sem, which would never get upped because
1817                  * this thread is waiting for the stat mutex.
1818                  */
1819                 deadlock_loop_cnt = 0;
1820                 do {
1821                         check_update_rusage(td);
1822                         if (!fio_sem_down_trylock(stat_sem))
1823                                 break;
1824                         usleep(1000);
1825                         if (deadlock_loop_cnt++ > 5000) {
1826                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1827                                 td->error = EDEADLK;
1828                                 goto err;
1829                         }
1830                 } while (1);
1831
1832                 if (td_read(td) && td->io_bytes[DDIR_READ])
1833                         update_runtime(td, elapsed_us, DDIR_READ);
1834                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1835                         update_runtime(td, elapsed_us, DDIR_WRITE);
1836                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1837                         update_runtime(td, elapsed_us, DDIR_TRIM);
1838                 fio_gettime(&td->start, NULL);
1839                 fio_sem_up(stat_sem);
1840
1841                 if (td->error || td->terminate)
1842                         break;
1843
1844                 if (!o->do_verify ||
1845                     o->verify == VERIFY_NONE ||
1846                     td_ioengine_flagged(td, FIO_UNIDIR))
1847                         continue;
1848
1849                 clear_io_state(td, 0);
1850
1851                 fio_gettime(&td->start, NULL);
1852
1853                 do_verify(td, verify_bytes);
1854
1855                 /*
1856                  * See comment further up for why this is done here.
1857                  */
1858                 check_update_rusage(td);
1859
1860                 fio_sem_down(stat_sem);
1861                 update_runtime(td, elapsed_us, DDIR_READ);
1862                 fio_gettime(&td->start, NULL);
1863                 fio_sem_up(stat_sem);
1864
1865                 if (td->error || td->terminate)
1866                         break;
1867         }
1868
1869         /*
1870          * Acquire this lock if we were doing overlap checking in
1871          * offload mode so that we don't clean up this job while
1872          * another thread is checking its io_u's for overlap
1873          */
1874         if (td_offload_overlap(td))
1875                 pthread_mutex_lock(&overlap_check);
1876         td_set_runstate(td, TD_FINISHING);
1877         if (td_offload_overlap(td))
1878                 pthread_mutex_unlock(&overlap_check);
1879
1880         update_rusage_stat(td);
1881         td->ts.total_run_time = mtime_since_now(&td->epoch);
1882         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1883         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1884         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1885
1886         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1887             (td->o.verify != VERIFY_NONE && td_write(td)))
1888                 verify_save_state(td->thread_number);
1889
1890         fio_unpin_memory(td);
1891
1892         td_writeout_logs(td, true);
1893
1894         iolog_compress_exit(td);
1895         rate_submit_exit(td);
1896
1897         if (o->exec_postrun)
1898                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1899
1900         if (exitall_on_terminate || (o->exitall_error && td->error))
1901                 fio_terminate_threads(td->groupid);
1902
1903 err:
1904         if (td->error)
1905                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1906                                                         td->verror);
1907
1908         if (o->verify_async)
1909                 verify_async_exit(td);
1910
1911         close_and_free_files(td);
1912         cleanup_io_u(td);
1913         close_ioengine(td);
1914         cgroup_shutdown(td, cgroup_mnt);
1915         verify_free_state(td);
1916         td_zone_free_index(td);
1917
1918         if (fio_option_is_set(o, cpumask)) {
1919                 ret = fio_cpuset_exit(&o->cpumask);
1920                 if (ret)
1921                         td_verror(td, ret, "fio_cpuset_exit");
1922         }
1923
1924         /*
1925          * do this very late, it will log file closing as well
1926          */
1927         if (o->write_iolog_file)
1928                 write_iolog_close(td);
1929         if (td->io_log_rfile)
1930                 fclose(td->io_log_rfile);
1931
1932         td_set_runstate(td, TD_EXITED);
1933
1934         /*
1935          * Do this last after setting our runstate to exited, so we
1936          * know that the stat thread is signaled.
1937          */
1938         check_update_rusage(td);
1939
1940         sk_out_drop();
1941         return (void *) (uintptr_t) td->error;
1942 }
1943
1944 /*
1945  * Run over the job map and reap the threads that have exited, if any.
1946  */
1947 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1948                          uint64_t *m_rate)
1949 {
1950         struct thread_data *td;
1951         unsigned int cputhreads, realthreads, pending;
1952         int i, status, ret;
1953
1954         /*
1955          * reap exited threads (TD_EXITED -> TD_REAPED)
1956          */
1957         realthreads = pending = cputhreads = 0;
1958         for_each_td(td, i) {
1959                 int flags = 0;
1960
1961                  if (!strcmp(td->o.ioengine, "cpuio"))
1962                         cputhreads++;
1963                 else
1964                         realthreads++;
1965
1966                 if (!td->pid) {
1967                         pending++;
1968                         continue;
1969                 }
1970                 if (td->runstate == TD_REAPED)
1971                         continue;
1972                 if (td->o.use_thread) {
1973                         if (td->runstate == TD_EXITED) {
1974                                 td_set_runstate(td, TD_REAPED);
1975                                 goto reaped;
1976                         }
1977                         continue;
1978                 }
1979
1980                 flags = WNOHANG;
1981                 if (td->runstate == TD_EXITED)
1982                         flags = 0;
1983
1984                 /*
1985                  * check if someone quit or got killed in an unusual way
1986                  */
1987                 ret = waitpid(td->pid, &status, flags);
1988                 if (ret < 0) {
1989                         if (errno == ECHILD) {
1990                                 log_err("fio: pid=%d disappeared %d\n",
1991                                                 (int) td->pid, td->runstate);
1992                                 td->sig = ECHILD;
1993                                 td_set_runstate(td, TD_REAPED);
1994                                 goto reaped;
1995                         }
1996                         perror("waitpid");
1997                 } else if (ret == td->pid) {
1998                         if (WIFSIGNALED(status)) {
1999                                 int sig = WTERMSIG(status);
2000
2001                                 if (sig != SIGTERM && sig != SIGUSR2)
2002                                         log_err("fio: pid=%d, got signal=%d\n",
2003                                                         (int) td->pid, sig);
2004                                 td->sig = sig;
2005                                 td_set_runstate(td, TD_REAPED);
2006                                 goto reaped;
2007                         }
2008                         if (WIFEXITED(status)) {
2009                                 if (WEXITSTATUS(status) && !td->error)
2010                                         td->error = WEXITSTATUS(status);
2011
2012                                 td_set_runstate(td, TD_REAPED);
2013                                 goto reaped;
2014                         }
2015                 }
2016
2017                 /*
2018                  * If the job is stuck, do a forceful timeout of it and
2019                  * move on.
2020                  */
2021                 if (td->terminate &&
2022                     td->runstate < TD_FSYNCING &&
2023                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2024                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2025                                 "%lu seconds, it appears to be stuck. Doing "
2026                                 "forceful exit of this job.\n",
2027                                 td->o.name, td->runstate,
2028                                 (unsigned long) time_since_now(&td->terminate_time));
2029                         td_set_runstate(td, TD_REAPED);
2030                         goto reaped;
2031                 }
2032
2033                 /*
2034                  * thread is not dead, continue
2035                  */
2036                 pending++;
2037                 continue;
2038 reaped:
2039                 (*nr_running)--;
2040                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2041                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2042                 if (!td->pid)
2043                         pending--;
2044
2045                 if (td->error)
2046                         exit_value++;
2047
2048                 done_secs += mtime_since_now(&td->epoch) / 1000;
2049                 profile_td_exit(td);
2050         }
2051
2052         if (*nr_running == cputhreads && !pending && realthreads)
2053                 fio_terminate_threads(TERMINATE_ALL);
2054 }
2055
2056 static bool __check_trigger_file(void)
2057 {
2058         struct stat sb;
2059
2060         if (!trigger_file)
2061                 return false;
2062
2063         if (stat(trigger_file, &sb))
2064                 return false;
2065
2066         if (unlink(trigger_file) < 0)
2067                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2068                                                         strerror(errno));
2069
2070         return true;
2071 }
2072
2073 static bool trigger_timedout(void)
2074 {
2075         if (trigger_timeout)
2076                 if (time_since_genesis() >= trigger_timeout) {
2077                         trigger_timeout = 0;
2078                         return true;
2079                 }
2080
2081         return false;
2082 }
2083
2084 void exec_trigger(const char *cmd)
2085 {
2086         int ret;
2087
2088         if (!cmd || cmd[0] == '\0')
2089                 return;
2090
2091         ret = system(cmd);
2092         if (ret == -1)
2093                 log_err("fio: failed executing %s trigger\n", cmd);
2094 }
2095
2096 void check_trigger_file(void)
2097 {
2098         if (__check_trigger_file() || trigger_timedout()) {
2099                 if (nr_clients)
2100                         fio_clients_send_trigger(trigger_remote_cmd);
2101                 else {
2102                         verify_save_state(IO_LIST_ALL);
2103                         fio_terminate_threads(TERMINATE_ALL);
2104                         exec_trigger(trigger_cmd);
2105                 }
2106         }
2107 }
2108
2109 static int fio_verify_load_state(struct thread_data *td)
2110 {
2111         int ret;
2112
2113         if (!td->o.verify_state)
2114                 return 0;
2115
2116         if (is_backend) {
2117                 void *data;
2118
2119                 ret = fio_server_get_verify_state(td->o.name,
2120                                         td->thread_number - 1, &data);
2121                 if (!ret)
2122                         verify_assign_state(td, data);
2123         } else
2124                 ret = verify_load_state(td, "local");
2125
2126         return ret;
2127 }
2128
2129 static void do_usleep(unsigned int usecs)
2130 {
2131         check_for_running_stats();
2132         check_trigger_file();
2133         usleep(usecs);
2134 }
2135
2136 static bool check_mount_writes(struct thread_data *td)
2137 {
2138         struct fio_file *f;
2139         unsigned int i;
2140
2141         if (!td_write(td) || td->o.allow_mounted_write)
2142                 return false;
2143
2144         /*
2145          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2146          * are mkfs'd and mounted.
2147          */
2148         for_each_file(td, f, i) {
2149 #ifdef FIO_HAVE_CHARDEV_SIZE
2150                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2151 #else
2152                 if (f->filetype != FIO_TYPE_BLOCK)
2153 #endif
2154                         continue;
2155                 if (device_is_mounted(f->file_name))
2156                         goto mounted;
2157         }
2158
2159         return false;
2160 mounted:
2161         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2162         return true;
2163 }
2164
2165 static bool waitee_running(struct thread_data *me)
2166 {
2167         const char *waitee = me->o.wait_for;
2168         const char *self = me->o.name;
2169         struct thread_data *td;
2170         int i;
2171
2172         if (!waitee)
2173                 return false;
2174
2175         for_each_td(td, i) {
2176                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2177                         continue;
2178
2179                 if (td->runstate < TD_EXITED) {
2180                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2181                                         self, td->o.name,
2182                                         runstate_to_name(td->runstate));
2183                         return true;
2184                 }
2185         }
2186
2187         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2188         return false;
2189 }
2190
2191 /*
2192  * Main function for kicking off and reaping jobs, as needed.
2193  */
2194 static void run_threads(struct sk_out *sk_out)
2195 {
2196         struct thread_data *td;
2197         unsigned int i, todo, nr_running, nr_started;
2198         uint64_t m_rate, t_rate;
2199         uint64_t spent;
2200
2201         if (fio_gtod_offload && fio_start_gtod_thread())
2202                 return;
2203
2204         fio_idle_prof_init();
2205
2206         set_sig_handlers();
2207
2208         nr_thread = nr_process = 0;
2209         for_each_td(td, i) {
2210                 if (check_mount_writes(td))
2211                         return;
2212                 if (td->o.use_thread)
2213                         nr_thread++;
2214                 else
2215                         nr_process++;
2216         }
2217
2218         if (output_format & FIO_OUTPUT_NORMAL) {
2219                 struct buf_output out;
2220
2221                 buf_output_init(&out);
2222                 __log_buf(&out, "Starting ");
2223                 if (nr_thread)
2224                         __log_buf(&out, "%d thread%s", nr_thread,
2225                                                 nr_thread > 1 ? "s" : "");
2226                 if (nr_process) {
2227                         if (nr_thread)
2228                                 __log_buf(&out, " and ");
2229                         __log_buf(&out, "%d process%s", nr_process,
2230                                                 nr_process > 1 ? "es" : "");
2231                 }
2232                 __log_buf(&out, "\n");
2233                 log_info_buf(out.buf, out.buflen);
2234                 buf_output_free(&out);
2235         }
2236
2237         todo = thread_number;
2238         nr_running = 0;
2239         nr_started = 0;
2240         m_rate = t_rate = 0;
2241
2242         for_each_td(td, i) {
2243                 print_status_init(td->thread_number - 1);
2244
2245                 if (!td->o.create_serialize)
2246                         continue;
2247
2248                 if (fio_verify_load_state(td))
2249                         goto reap;
2250
2251                 /*
2252                  * do file setup here so it happens sequentially,
2253                  * we don't want X number of threads getting their
2254                  * client data interspersed on disk
2255                  */
2256                 if (setup_files(td)) {
2257 reap:
2258                         exit_value++;
2259                         if (td->error)
2260                                 log_err("fio: pid=%d, err=%d/%s\n",
2261                                         (int) td->pid, td->error, td->verror);
2262                         td_set_runstate(td, TD_REAPED);
2263                         todo--;
2264                 } else {
2265                         struct fio_file *f;
2266                         unsigned int j;
2267
2268                         /*
2269                          * for sharing to work, each job must always open
2270                          * its own files. so close them, if we opened them
2271                          * for creation
2272                          */
2273                         for_each_file(td, f, j) {
2274                                 if (fio_file_open(f))
2275                                         td_io_close_file(td, f);
2276                         }
2277                 }
2278         }
2279
2280         /* start idle threads before io threads start to run */
2281         fio_idle_prof_start();
2282
2283         set_genesis_time();
2284
2285         while (todo) {
2286                 struct thread_data *map[REAL_MAX_JOBS];
2287                 struct timespec this_start;
2288                 int this_jobs = 0, left;
2289                 struct fork_data *fd;
2290
2291                 /*
2292                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2293                  */
2294                 for_each_td(td, i) {
2295                         if (td->runstate != TD_NOT_CREATED)
2296                                 continue;
2297
2298                         /*
2299                          * never got a chance to start, killed by other
2300                          * thread for some reason
2301                          */
2302                         if (td->terminate) {
2303                                 todo--;
2304                                 continue;
2305                         }
2306
2307                         if (td->o.start_delay) {
2308                                 spent = utime_since_genesis();
2309
2310                                 if (td->o.start_delay > spent)
2311                                         continue;
2312                         }
2313
2314                         if (td->o.stonewall && (nr_started || nr_running)) {
2315                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2316                                                         td->o.name);
2317                                 break;
2318                         }
2319
2320                         if (waitee_running(td)) {
2321                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2322                                                 td->o.name, td->o.wait_for);
2323                                 continue;
2324                         }
2325
2326                         init_disk_util(td);
2327
2328                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2329                         td->update_rusage = 0;
2330
2331                         /*
2332                          * Set state to created. Thread will transition
2333                          * to TD_INITIALIZED when it's done setting up.
2334                          */
2335                         td_set_runstate(td, TD_CREATED);
2336                         map[this_jobs++] = td;
2337                         nr_started++;
2338
2339                         fd = calloc(1, sizeof(*fd));
2340                         fd->td = td;
2341                         fd->sk_out = sk_out;
2342
2343                         if (td->o.use_thread) {
2344                                 int ret;
2345
2346                                 dprint(FD_PROCESS, "will pthread_create\n");
2347                                 ret = pthread_create(&td->thread, NULL,
2348                                                         thread_main, fd);
2349                                 if (ret) {
2350                                         log_err("pthread_create: %s\n",
2351                                                         strerror(ret));
2352                                         free(fd);
2353                                         nr_started--;
2354                                         break;
2355                                 }
2356                                 fd = NULL;
2357                                 ret = pthread_detach(td->thread);
2358                                 if (ret)
2359                                         log_err("pthread_detach: %s",
2360                                                         strerror(ret));
2361                         } else {
2362                                 pid_t pid;
2363                                 dprint(FD_PROCESS, "will fork\n");
2364                                 pid = fork();
2365                                 if (!pid) {
2366                                         int ret;
2367
2368                                         ret = (int)(uintptr_t)thread_main(fd);
2369                                         _exit(ret);
2370                                 } else if (i == fio_debug_jobno)
2371                                         *fio_debug_jobp = pid;
2372                         }
2373                         dprint(FD_MUTEX, "wait on startup_sem\n");
2374                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2375                                 log_err("fio: job startup hung? exiting.\n");
2376                                 fio_terminate_threads(TERMINATE_ALL);
2377                                 fio_abort = true;
2378                                 nr_started--;
2379                                 free(fd);
2380                                 break;
2381                         }
2382                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2383                 }
2384
2385                 /*
2386                  * Wait for the started threads to transition to
2387                  * TD_INITIALIZED.
2388                  */
2389                 fio_gettime(&this_start, NULL);
2390                 left = this_jobs;
2391                 while (left && !fio_abort) {
2392                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2393                                 break;
2394
2395                         do_usleep(100000);
2396
2397                         for (i = 0; i < this_jobs; i++) {
2398                                 td = map[i];
2399                                 if (!td)
2400                                         continue;
2401                                 if (td->runstate == TD_INITIALIZED) {
2402                                         map[i] = NULL;
2403                                         left--;
2404                                 } else if (td->runstate >= TD_EXITED) {
2405                                         map[i] = NULL;
2406                                         left--;
2407                                         todo--;
2408                                         nr_running++; /* work-around... */
2409                                 }
2410                         }
2411                 }
2412
2413                 if (left) {
2414                         log_err("fio: %d job%s failed to start\n", left,
2415                                         left > 1 ? "s" : "");
2416                         for (i = 0; i < this_jobs; i++) {
2417                                 td = map[i];
2418                                 if (!td)
2419                                         continue;
2420                                 kill(td->pid, SIGTERM);
2421                         }
2422                         break;
2423                 }
2424
2425                 /*
2426                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2427                  */
2428                 for_each_td(td, i) {
2429                         if (td->runstate != TD_INITIALIZED)
2430                                 continue;
2431
2432                         if (in_ramp_time(td))
2433                                 td_set_runstate(td, TD_RAMP);
2434                         else
2435                                 td_set_runstate(td, TD_RUNNING);
2436                         nr_running++;
2437                         nr_started--;
2438                         m_rate += ddir_rw_sum(td->o.ratemin);
2439                         t_rate += ddir_rw_sum(td->o.rate);
2440                         todo--;
2441                         fio_sem_up(td->sem);
2442                 }
2443
2444                 reap_threads(&nr_running, &t_rate, &m_rate);
2445
2446                 if (todo)
2447                         do_usleep(100000);
2448         }
2449
2450         while (nr_running) {
2451                 reap_threads(&nr_running, &t_rate, &m_rate);
2452                 do_usleep(10000);
2453         }
2454
2455         fio_idle_prof_stop();
2456
2457         update_io_ticks();
2458 }
2459
2460 static void free_disk_util(void)
2461 {
2462         disk_util_prune_entries();
2463         helper_thread_destroy();
2464 }
2465
2466 int fio_backend(struct sk_out *sk_out)
2467 {
2468         struct thread_data *td;
2469         int i;
2470
2471         if (exec_profile) {
2472                 if (load_profile(exec_profile))
2473                         return 1;
2474                 free(exec_profile);
2475                 exec_profile = NULL;
2476         }
2477         if (!thread_number)
2478                 return 0;
2479
2480         if (write_bw_log) {
2481                 struct log_params p = {
2482                         .log_type = IO_LOG_TYPE_BW,
2483                 };
2484
2485                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2486                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2487                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2488         }
2489
2490         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2491         if (!sk_out)
2492                 is_local_backend = true;
2493         if (startup_sem == NULL)
2494                 return 1;
2495
2496         set_genesis_time();
2497         stat_init();
2498         helper_thread_create(startup_sem, sk_out);
2499
2500         cgroup_list = smalloc(sizeof(*cgroup_list));
2501         if (cgroup_list)
2502                 INIT_FLIST_HEAD(cgroup_list);
2503
2504         run_threads(sk_out);
2505
2506         helper_thread_exit();
2507
2508         if (!fio_abort) {
2509                 __show_run_stats();
2510                 if (write_bw_log) {
2511                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2512                                 struct io_log *log = agg_io_log[i];
2513
2514                                 flush_log(log, false);
2515                                 free_log(log);
2516                         }
2517                 }
2518         }
2519
2520         for_each_td(td, i) {
2521                 steadystate_free(td);
2522                 fio_options_free(td);
2523                 if (td->rusage_sem) {
2524                         fio_sem_remove(td->rusage_sem);
2525                         td->rusage_sem = NULL;
2526                 }
2527                 fio_sem_remove(td->sem);
2528                 td->sem = NULL;
2529         }
2530
2531         free_disk_util();
2532         if (cgroup_list) {
2533                 cgroup_kill(cgroup_list);
2534                 sfree(cgroup_list);
2535         }
2536
2537         fio_sem_remove(startup_sem);
2538         stat_exit();
2539         return exit_value;
2540 }