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., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38 #include <math.h>
  39
  40 #include "fio.h"
  41 #ifndef FIO_NO_HAVE_SHM_H
  42 #include <sys/shm.h>
  43 #endif
  44 #include "hash.h"
  45 #include "smalloc.h"
  46 #include "verify.h"
  47 #include "trim.h"
  48 #include "diskutil.h"
  49 #include "cgroup.h"
  50 #include "profile.h"
  51 #include "lib/rand.h"
  52 #include "lib/memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59 #include "rate-submit.h"
  60 #include "helper_thread.h"
  61
  62 static struct fio_mutex *startup_mutex;
  63 static struct flist_head *cgroup_list;
  64 static char *cgroup_mnt;
  65 static int exit_value;
  66 static volatile int fio_abort;
  67 static unsigned int nr_process = 0;
  68 static unsigned int nr_thread = 0;
  69
  70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  71
  72 int groupid = 0;
  73 unsigned int thread_number = 0;
  74 unsigned int stat_number = 0;
  75 int shm_id = 0;
  76 int temp_stall_ts;
  77 unsigned long done_secs = 0;
  78
  79 #define PAGE_ALIGN(buf) \
  80         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  81
  82 #define JOB_START_TIMEOUT       (5 * 1000)
  83
  84 static void sig_int(int sig)
  85 {
  86         if (threads) {
  87                 if (is_backend)
  88                         fio_server_got_signal(sig);
  89                 else {
  90                         log_info("\nfio: terminating on signal %d\n", sig);
  91                         log_info_flush();
  92                         exit_value = 128;
  93                 }
  94
  95                 fio_terminate_threads(TERMINATE_ALL);
  96         }
  97 }
  98
  99 void sig_show_status(int sig)
 100 {
 101         show_running_run_stats();
 102 }
 103
 104 static void set_sig_handlers(void)
 105 {
 106         struct sigaction act;
 107
 108         memset(&act, 0, sizeof(act));
 109         act.sa_handler = sig_int;
 110         act.sa_flags = SA_RESTART;
 111         sigaction(SIGINT, &act, NULL);
 112
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGTERM, &act, NULL);
 117
 118 /* Windows uses SIGBREAK as a quit signal from other applications */
 119 #ifdef WIN32
 120         memset(&act, 0, sizeof(act));
 121         act.sa_handler = sig_int;
 122         act.sa_flags = SA_RESTART;
 123         sigaction(SIGBREAK, &act, NULL);
 124 #endif
 125
 126         memset(&act, 0, sizeof(act));
 127         act.sa_handler = sig_show_status;
 128         act.sa_flags = SA_RESTART;
 129         sigaction(SIGUSR1, &act, NULL);
 130
 131         if (is_backend) {
 132                 memset(&act, 0, sizeof(act));
 133                 act.sa_handler = sig_int;
 134                 act.sa_flags = SA_RESTART;
 135                 sigaction(SIGPIPE, &act, NULL);
 136         }
 137 }
 138
 139 /*
 140  * Check if we are above the minimum rate given.
 141  */
 142 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 143                              enum fio_ddir ddir)
 144 {
 145         unsigned long long bytes = 0;
 146         unsigned long iops = 0;
 147         unsigned long spent;
 148         unsigned long rate;
 149         unsigned int ratemin = 0;
 150         unsigned int rate_iops = 0;
 151         unsigned int rate_iops_min = 0;
 152
 153         assert(ddir_rw(ddir));
 154
 155         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 156                 return false;
 157
 158         /*
 159          * allow a 2 second settle period in the beginning
 160          */
 161         if (mtime_since(&td->start, now) < 2000)
 162                 return false;
 163
 164         iops += td->this_io_blocks[ddir];
 165         bytes += td->this_io_bytes[ddir];
 166         ratemin += td->o.ratemin[ddir];
 167         rate_iops += td->o.rate_iops[ddir];
 168         rate_iops_min += td->o.rate_iops_min[ddir];
 169
 170         /*
 171          * if rate blocks is set, sample is running
 172          */
 173         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 174                 spent = mtime_since(&td->lastrate[ddir], now);
 175                 if (spent < td->o.ratecycle)
 176                         return false;
 177
 178                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 179                         /*
 180                          * check bandwidth specified rate
 181                          */
 182                         if (bytes < td->rate_bytes[ddir]) {
 183                                 log_err("%s: min rate %u not met\n", td->o.name,
 184                                                                 ratemin);
 185                                 return true;
 186                         } else {
 187                                 if (spent)
 188                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 189                                 else
 190                                         rate = 0;
 191
 192                                 if (rate < ratemin ||
 193                                     bytes < td->rate_bytes[ddir]) {
 194                                         log_err("%s: min rate %u not met, got"
 195                                                 " %luKB/sec\n", td->o.name,
 196                                                         ratemin, rate);
 197                                         return true;
 198                                 }
 199                         }
 200                 } else {
 201                         /*
 202                          * checks iops specified rate
 203                          */
 204                         if (iops < rate_iops) {
 205                                 log_err("%s: min iops rate %u not met\n",
 206                                                 td->o.name, rate_iops);
 207                                 return true;
 208                         } else {
 209                                 if (spent)
 210                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 211                                 else
 212                                         rate = 0;
 213
 214                                 if (rate < rate_iops_min ||
 215                                     iops < td->rate_blocks[ddir]) {
 216                                         log_err("%s: min iops rate %u not met,"
 217                                                 " got %lu\n", td->o.name,
 218                                                         rate_iops_min, rate);
 219                                         return true;
 220                                 }
 221                         }
 222                 }
 223         }
 224
 225         td->rate_bytes[ddir] = bytes;
 226         td->rate_blocks[ddir] = iops;
 227         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 228         return false;
 229 }
 230
 231 static bool check_min_rate(struct thread_data *td, struct timeval *now)
 232 {
 233         bool ret = false;
 234
 235         if (td->bytes_done[DDIR_READ])
 236                 ret |= __check_min_rate(td, now, DDIR_READ);
 237         if (td->bytes_done[DDIR_WRITE])
 238                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 239         if (td->bytes_done[DDIR_TRIM])
 240                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 241
 242         return ret;
 243 }
 244
 245 /*
 246  * When job exits, we can cancel the in-flight IO if we are using async
 247  * io. Attempt to do so.
 248  */
 249 static void cleanup_pending_aio(struct thread_data *td)
 250 {
 251         int r;
 252
 253         /*
 254          * get immediately available events, if any
 255          */
 256         r = io_u_queued_complete(td, 0);
 257         if (r < 0)
 258                 return;
 259
 260         /*
 261          * now cancel remaining active events
 262          */
 263         if (td->io_ops->cancel) {
 264                 struct io_u *io_u;
 265                 int i;
 266
 267                 io_u_qiter(&td->io_u_all, io_u, i) {
 268                         if (io_u->flags & IO_U_F_FLIGHT) {
 269                                 r = td->io_ops->cancel(td, io_u);
 270                                 if (!r)
 271                                         put_io_u(td, io_u);
 272                         }
 273                 }
 274         }
 275
 276         if (td->cur_depth)
 277                 r = io_u_queued_complete(td, td->cur_depth);
 278 }
 279
 280 /*
 281  * Helper to handle the final sync of a file. Works just like the normal
 282  * io path, just does everything sync.
 283  */
 284 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 285 {
 286         struct io_u *io_u = __get_io_u(td);
 287         int ret;
 288
 289         if (!io_u)
 290                 return true;
 291
 292         io_u->ddir = DDIR_SYNC;
 293         io_u->file = f;
 294
 295         if (td_io_prep(td, io_u)) {
 296                 put_io_u(td, io_u);
 297                 return true;
 298         }
 299
 300 requeue:
 301         ret = td_io_queue(td, io_u);
 302         if (ret < 0) {
 303                 td_verror(td, io_u->error, "td_io_queue");
 304                 put_io_u(td, io_u);
 305                 return true;
 306         } else if (ret == FIO_Q_QUEUED) {
 307                 if (td_io_commit(td))
 308                         return true;
 309                 if (io_u_queued_complete(td, 1) < 0)
 310                         return true;
 311         } else if (ret == FIO_Q_COMPLETED) {
 312                 if (io_u->error) {
 313                         td_verror(td, io_u->error, "td_io_queue");
 314                         return true;
 315                 }
 316
 317                 if (io_u_sync_complete(td, io_u) < 0)
 318                         return true;
 319         } else if (ret == FIO_Q_BUSY) {
 320                 if (td_io_commit(td))
 321                         return true;
 322                 goto requeue;
 323         }
 324
 325         return false;
 326 }
 327
 328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 329 {
 330         int ret;
 331
 332         if (fio_file_open(f))
 333                 return fio_io_sync(td, f);
 334
 335         if (td_io_open_file(td, f))
 336                 return 1;
 337
 338         ret = fio_io_sync(td, f);
 339         td_io_close_file(td, f);
 340         return ret;
 341 }
 342
 343 static inline void __update_tv_cache(struct thread_data *td)
 344 {
 345         fio_gettime(&td->tv_cache, NULL);
 346 }
 347
 348 static inline void update_tv_cache(struct thread_data *td)
 349 {
 350         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 351                 __update_tv_cache(td);
 352 }
 353
 354 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
 355 {
 356         if (in_ramp_time(td))
 357                 return false;
 358         if (!td->o.timeout)
 359                 return false;
 360         if (utime_since(&td->epoch, t) >= td->o.timeout)
 361                 return true;
 362
 363         return false;
 364 }
 365
 366 /*
 367  * We need to update the runtime consistently in ms, but keep a running
 368  * tally of the current elapsed time in microseconds for sub millisecond
 369  * updates.
 370  */
 371 static inline void update_runtime(struct thread_data *td,
 372                                   unsigned long long *elapsed_us,
 373                                   const enum fio_ddir ddir)
 374 {
 375         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 376                 return;
 377
 378         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 379         elapsed_us[ddir] += utime_since_now(&td->start);
 380         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 381 }
 382
 383 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 384                                 int *retptr)
 385 {
 386         int ret = *retptr;
 387
 388         if (ret < 0 || td->error) {
 389                 int err = td->error;
 390                 enum error_type_bit eb;
 391
 392                 if (ret < 0)
 393                         err = -ret;
 394
 395                 eb = td_error_type(ddir, err);
 396                 if (!(td->o.continue_on_error & (1 << eb)))
 397                         return true;
 398
 399                 if (td_non_fatal_error(td, eb, err)) {
 400                         /*
 401                          * Continue with the I/Os in case of
 402                          * a non fatal error.
 403                          */
 404                         update_error_count(td, err);
 405                         td_clear_error(td);
 406                         *retptr = 0;
 407                         return false;
 408                 } else if (td->o.fill_device && err == ENOSPC) {
 409                         /*
 410                          * We expect to hit this error if
 411                          * fill_device option is set.
 412                          */
 413                         td_clear_error(td);
 414                         fio_mark_td_terminate(td);
 415                         return true;
 416                 } else {
 417                         /*
 418                          * Stop the I/O in case of a fatal
 419                          * error.
 420                          */
 421                         update_error_count(td, err);
 422                         return true;
 423                 }
 424         }
 425
 426         return false;
 427 }
 428
 429 static void check_update_rusage(struct thread_data *td)
 430 {
 431         if (td->update_rusage) {
 432                 td->update_rusage = 0;
 433                 update_rusage_stat(td);
 434                 fio_mutex_up(td->rusage_sem);
 435         }
 436 }
 437
 438 static int wait_for_completions(struct thread_data *td, struct timeval *time)
 439 {
 440         const int full = queue_full(td);
 441         int min_evts = 0;
 442         int ret;
 443
 444         if (td->flags & TD_F_REGROW_LOGS) {
 445                 ret = io_u_quiesce(td);
 446                 regrow_logs(td);
 447                 return ret;
 448         }
 449
 450         /*
 451          * if the queue is full, we MUST reap at least 1 event
 452          */
 453         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 454         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 455                 min_evts = 1;
 456
 457         if (time && (__should_check_rate(td, DDIR_READ) ||
 458             __should_check_rate(td, DDIR_WRITE) ||
 459             __should_check_rate(td, DDIR_TRIM)))
 460                 fio_gettime(time, NULL);
 461
 462         do {
 463                 ret = io_u_queued_complete(td, min_evts);
 464                 if (ret < 0)
 465                         break;
 466         } while (full && (td->cur_depth > td->o.iodepth_low));
 467
 468         return ret;
 469 }
 470
 471 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 472                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 473                    struct timeval *comp_time)
 474 {
 475         int ret2;
 476
 477         switch (*ret) {
 478         case FIO_Q_COMPLETED:
 479                 if (io_u->error) {
 480                         *ret = -io_u->error;
 481                         clear_io_u(td, io_u);
 482                 } else if (io_u->resid) {
 483                         int bytes = io_u->xfer_buflen - io_u->resid;
 484                         struct fio_file *f = io_u->file;
 485
 486                         if (bytes_issued)
 487                                 *bytes_issued += bytes;
 488
 489                         if (!from_verify)
 490                                 trim_io_piece(td, io_u);
 491
 492                         /*
 493                          * zero read, fail
 494                          */
 495                         if (!bytes) {
 496                                 if (!from_verify)
 497                                         unlog_io_piece(td, io_u);
 498                                 td_verror(td, EIO, "full resid");
 499                                 put_io_u(td, io_u);
 500                                 break;
 501                         }
 502
 503                         io_u->xfer_buflen = io_u->resid;
 504                         io_u->xfer_buf += bytes;
 505                         io_u->offset += bytes;
 506
 507                         if (ddir_rw(io_u->ddir))
 508                                 td->ts.short_io_u[io_u->ddir]++;
 509
 510                         f = io_u->file;
 511                         if (io_u->offset == f->real_file_size)
 512                                 goto sync_done;
 513
 514                         requeue_io_u(td, &io_u);
 515                 } else {
 516 sync_done:
 517                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 518                             __should_check_rate(td, DDIR_WRITE) ||
 519                             __should_check_rate(td, DDIR_TRIM)))
 520                                 fio_gettime(comp_time, NULL);
 521
 522                         *ret = io_u_sync_complete(td, io_u);
 523                         if (*ret < 0)
 524                                 break;
 525                 }
 526
 527                 if (td->flags & TD_F_REGROW_LOGS)
 528                         regrow_logs(td);
 529
 530                 /*
 531                  * when doing I/O (not when verifying),
 532                  * check for any errors that are to be ignored
 533                  */
 534                 if (!from_verify)
 535                         break;
 536
 537                 return 0;
 538         case FIO_Q_QUEUED:
 539                 /*
 540                  * if the engine doesn't have a commit hook,
 541                  * the io_u is really queued. if it does have such
 542                  * a hook, it has to call io_u_queued() itself.
 543                  */
 544                 if (td->io_ops->commit == NULL)
 545                         io_u_queued(td, io_u);
 546                 if (bytes_issued)
 547                         *bytes_issued += io_u->xfer_buflen;
 548                 break;
 549         case FIO_Q_BUSY:
 550                 if (!from_verify)
 551                         unlog_io_piece(td, io_u);
 552                 requeue_io_u(td, &io_u);
 553                 ret2 = td_io_commit(td);
 554                 if (ret2 < 0)
 555                         *ret = ret2;
 556                 break;
 557         default:
 558                 assert(*ret < 0);
 559                 td_verror(td, -(*ret), "td_io_queue");
 560                 break;
 561         }
 562
 563         if (break_on_this_error(td, ddir, ret))
 564                 return 1;
 565
 566         return 0;
 567 }
 568
 569 static inline bool io_in_polling(struct thread_data *td)
 570 {
 571         return !td->o.iodepth_batch_complete_min &&
 572                    !td->o.iodepth_batch_complete_max;
 573 }
 574 /*
 575  * Unlinks files from thread data fio_file structure
 576  */
 577 static int unlink_all_files(struct thread_data *td)
 578 {
 579         struct fio_file *f;
 580         unsigned int i;
 581         int ret = 0;
 582
 583         for_each_file(td, f, i) {
 584                 if (f->filetype != FIO_TYPE_FILE)
 585                         continue;
 586                 ret = td_io_unlink_file(td, f);
 587                 if (ret)
 588                         break;
 589         }
 590
 591         if (ret)
 592                 td_verror(td, ret, "unlink_all_files");
 593
 594         return ret;
 595 }
 596
 597 /*
 598  * The main verify engine. Runs over the writes we previously submitted,
 599  * reads the blocks back in, and checks the crc/md5 of the data.
 600  */
 601 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 602 {
 603         struct fio_file *f;
 604         struct io_u *io_u;
 605         int ret, min_events;
 606         unsigned int i;
 607
 608         dprint(FD_VERIFY, "starting loop\n");
 609
 610         /*
 611          * sync io first and invalidate cache, to make sure we really
 612          * read from disk.
 613          */
 614         for_each_file(td, f, i) {
 615                 if (!fio_file_open(f))
 616                         continue;
 617                 if (fio_io_sync(td, f))
 618                         break;
 619                 if (file_invalidate_cache(td, f))
 620                         break;
 621         }
 622
 623         check_update_rusage(td);
 624
 625         if (td->error)
 626                 return;
 627
 628         /*
 629          * verify_state needs to be reset before verification
 630          * proceeds so that expected random seeds match actual
 631          * random seeds in headers. The main loop will reset
 632          * all random number generators if randrepeat is set.
 633          */
 634         if (!td->o.rand_repeatable)
 635                 td_fill_verify_state_seed(td);
 636
 637         td_set_runstate(td, TD_VERIFYING);
 638
 639         io_u = NULL;
 640         while (!td->terminate) {
 641                 enum fio_ddir ddir;
 642                 int full;
 643
 644                 update_tv_cache(td);
 645                 check_update_rusage(td);
 646
 647                 if (runtime_exceeded(td, &td->tv_cache)) {
 648                         __update_tv_cache(td);
 649                         if (runtime_exceeded(td, &td->tv_cache)) {
 650                                 fio_mark_td_terminate(td);
 651                                 break;
 652                         }
 653                 }
 654
 655                 if (flow_threshold_exceeded(td))
 656                         continue;
 657
 658                 if (!td->o.experimental_verify) {
 659                         io_u = __get_io_u(td);
 660                         if (!io_u)
 661                                 break;
 662
 663                         if (get_next_verify(td, io_u)) {
 664                                 put_io_u(td, io_u);
 665                                 break;
 666                         }
 667
 668                         if (td_io_prep(td, io_u)) {
 669                                 put_io_u(td, io_u);
 670                                 break;
 671                         }
 672                 } else {
 673                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 674                                 break;
 675
 676                         while ((io_u = get_io_u(td)) != NULL) {
 677                                 if (IS_ERR_OR_NULL(io_u)) {
 678                                         io_u = NULL;
 679                                         ret = FIO_Q_BUSY;
 680                                         goto reap;
 681                                 }
 682
 683                                 /*
 684                                  * We are only interested in the places where
 685                                  * we wrote or trimmed IOs. Turn those into
 686                                  * reads for verification purposes.
 687                                  */
 688                                 if (io_u->ddir == DDIR_READ) {
 689                                         /*
 690                                          * Pretend we issued it for rwmix
 691                                          * accounting
 692                                          */
 693                                         td->io_issues[DDIR_READ]++;
 694                                         put_io_u(td, io_u);
 695                                         continue;
 696                                 } else if (io_u->ddir == DDIR_TRIM) {
 697                                         io_u->ddir = DDIR_READ;
 698                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 699                                         break;
 700                                 } else if (io_u->ddir == DDIR_WRITE) {
 701                                         io_u->ddir = DDIR_READ;
 702                                         break;
 703                                 } else {
 704                                         put_io_u(td, io_u);
 705                                         continue;
 706                                 }
 707                         }
 708
 709                         if (!io_u)
 710                                 break;
 711                 }
 712
 713                 if (verify_state_should_stop(td, io_u)) {
 714                         put_io_u(td, io_u);
 715                         break;
 716                 }
 717
 718                 if (td->o.verify_async)
 719                         io_u->end_io = verify_io_u_async;
 720                 else
 721                         io_u->end_io = verify_io_u;
 722
 723                 ddir = io_u->ddir;
 724                 if (!td->o.disable_slat)
 725                         fio_gettime(&io_u->start_time, NULL);
 726
 727                 ret = td_io_queue(td, io_u);
 728
 729                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 730                         break;
 731
 732                 /*
 733                  * if we can queue more, do so. but check if there are
 734                  * completed io_u's first. Note that we can get BUSY even
 735                  * without IO queued, if the system is resource starved.
 736                  */
 737 reap:
 738                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 739                 if (full || io_in_polling(td))
 740                         ret = wait_for_completions(td, NULL);
 741
 742                 if (ret < 0)
 743                         break;
 744         }
 745
 746         check_update_rusage(td);
 747
 748         if (!td->error) {
 749                 min_events = td->cur_depth;
 750
 751                 if (min_events)
 752                         ret = io_u_queued_complete(td, min_events);
 753         } else
 754                 cleanup_pending_aio(td);
 755
 756         td_set_runstate(td, TD_RUNNING);
 757
 758         dprint(FD_VERIFY, "exiting loop\n");
 759 }
 760
 761 static bool exceeds_number_ios(struct thread_data *td)
 762 {
 763         unsigned long long number_ios;
 764
 765         if (!td->o.number_ios)
 766                 return false;
 767
 768         number_ios = ddir_rw_sum(td->io_blocks);
 769         number_ios += td->io_u_queued + td->io_u_in_flight;
 770
 771         return number_ios >= (td->o.number_ios * td->loops);
 772 }
 773
 774 static bool io_issue_bytes_exceeded(struct thread_data *td)
 775 {
 776         unsigned long long bytes, limit;
 777
 778         if (td_rw(td))
 779                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 780         else if (td_write(td))
 781                 bytes = td->io_issue_bytes[DDIR_WRITE];
 782         else if (td_read(td))
 783                 bytes = td->io_issue_bytes[DDIR_READ];
 784         else
 785                 bytes = td->io_issue_bytes[DDIR_TRIM];
 786
 787         if (td->o.io_limit)
 788                 limit = td->o.io_limit;
 789         else
 790                 limit = td->o.size;
 791
 792         limit *= td->loops;
 793         return bytes >= limit || exceeds_number_ios(td);
 794 }
 795
 796 static bool io_complete_bytes_exceeded(struct thread_data *td)
 797 {
 798         unsigned long long bytes, limit;
 799
 800         if (td_rw(td))
 801                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 802         else if (td_write(td))
 803                 bytes = td->this_io_bytes[DDIR_WRITE];
 804         else if (td_read(td))
 805                 bytes = td->this_io_bytes[DDIR_READ];
 806         else
 807                 bytes = td->this_io_bytes[DDIR_TRIM];
 808
 809         if (td->o.io_limit)
 810                 limit = td->o.io_limit;
 811         else
 812                 limit = td->o.size;
 813
 814         limit *= td->loops;
 815         return bytes >= limit || exceeds_number_ios(td);
 816 }
 817
 818 /*
 819  * used to calculate the next io time for rate control
 820  *
 821  */
 822 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 823 {
 824         uint64_t secs, remainder, bps, bytes, iops;
 825
 826         assert(!(td->flags & TD_F_CHILD));
 827         bytes = td->rate_io_issue_bytes[ddir];
 828         bps = td->rate_bps[ddir];
 829
 830         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 831                 uint64_t val;
 832                 iops = bps / td->o.bs[ddir];
 833                 val = (int64_t) (1000000 / iops) *
 834                                 -logf(__rand_0_1(&td->poisson_state));
 835                 if (val) {
 836                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 837                                         (unsigned long long) 1000000 / val);
 838                 }
 839                 td->last_usec += val;
 840                 return td->last_usec;
 841         } else if (bps) {
 842                 secs = bytes / bps;
 843                 remainder = bytes % bps;
 844                 return remainder * 1000000 / bps + secs * 1000000;
 845         }
 846
 847         return 0;
 848 }
 849
 850 /*
 851  * Main IO worker function. It retrieves io_u's to process and queues
 852  * and reaps them, checking for rate and errors along the way.
 853  *
 854  * Returns number of bytes written and trimmed.
 855  */
 856 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 857 {
 858         unsigned int i;
 859         int ret = 0;
 860         uint64_t total_bytes, bytes_issued = 0;
 861
 862         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 863                 bytes_done[i] = td->bytes_done[i];
 864
 865         if (in_ramp_time(td))
 866                 td_set_runstate(td, TD_RAMP);
 867         else
 868                 td_set_runstate(td, TD_RUNNING);
 869
 870         lat_target_init(td);
 871
 872         total_bytes = td->o.size;
 873         /*
 874         * Allow random overwrite workloads to write up to io_limit
 875         * before starting verification phase as 'size' doesn't apply.
 876         */
 877         if (td_write(td) && td_random(td) && td->o.norandommap)
 878                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 879         /*
 880          * If verify_backlog is enabled, we'll run the verify in this
 881          * handler as well. For that case, we may need up to twice the
 882          * amount of bytes.
 883          */
 884         if (td->o.verify != VERIFY_NONE &&
 885            (td_write(td) && td->o.verify_backlog))
 886                 total_bytes += td->o.size;
 887
 888         /* In trimwrite mode, each byte is trimmed and then written, so
 889          * allow total_bytes to be twice as big */
 890         if (td_trimwrite(td))
 891                 total_bytes += td->total_io_size;
 892
 893         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 894                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 895                 td->o.time_based) {
 896                 struct timeval comp_time;
 897                 struct io_u *io_u;
 898                 int full;
 899                 enum fio_ddir ddir;
 900
 901                 check_update_rusage(td);
 902
 903                 if (td->terminate || td->done)
 904                         break;
 905
 906                 update_tv_cache(td);
 907
 908                 if (runtime_exceeded(td, &td->tv_cache)) {
 909                         __update_tv_cache(td);
 910                         if (runtime_exceeded(td, &td->tv_cache)) {
 911                                 fio_mark_td_terminate(td);
 912                                 break;
 913                         }
 914                 }
 915
 916                 if (flow_threshold_exceeded(td))
 917                         continue;
 918
 919                 /*
 920                  * Break if we exceeded the bytes. The exception is time
 921                  * based runs, but we still need to break out of the loop
 922                  * for those to run verification, if enabled.
 923                  */
 924                 if (bytes_issued >= total_bytes &&
 925                     (!td->o.time_based ||
 926                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 927                         break;
 928
 929                 io_u = get_io_u(td);
 930                 if (IS_ERR_OR_NULL(io_u)) {
 931                         int err = PTR_ERR(io_u);
 932
 933                         io_u = NULL;
 934                         if (err == -EBUSY) {
 935                                 ret = FIO_Q_BUSY;
 936                                 goto reap;
 937                         }
 938                         if (td->o.latency_target)
 939                                 goto reap;
 940                         break;
 941                 }
 942
 943                 ddir = io_u->ddir;
 944
 945                 /*
 946                  * Add verification end_io handler if:
 947                  *      - Asked to verify (!td_rw(td))
 948                  *      - Or the io_u is from our verify list (mixed write/ver)
 949                  */
 950                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 951                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 952
 953                         if (!td->o.verify_pattern_bytes) {
 954                                 io_u->rand_seed = __rand(&td->verify_state);
 955                                 if (sizeof(int) != sizeof(long *))
 956                                         io_u->rand_seed *= __rand(&td->verify_state);
 957                         }
 958
 959                         if (verify_state_should_stop(td, io_u)) {
 960                                 put_io_u(td, io_u);
 961                                 break;
 962                         }
 963
 964                         if (td->o.verify_async)
 965                                 io_u->end_io = verify_io_u_async;
 966                         else
 967                                 io_u->end_io = verify_io_u;
 968                         td_set_runstate(td, TD_VERIFYING);
 969                 } else if (in_ramp_time(td))
 970                         td_set_runstate(td, TD_RAMP);
 971                 else
 972                         td_set_runstate(td, TD_RUNNING);
 973
 974                 /*
 975                  * Always log IO before it's issued, so we know the specific
 976                  * order of it. The logged unit will track when the IO has
 977                  * completed.
 978                  */
 979                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 980                     td->o.do_verify &&
 981                     td->o.verify != VERIFY_NONE &&
 982                     !td->o.experimental_verify)
 983                         log_io_piece(td, io_u);
 984
 985                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 986                         const unsigned long blen = io_u->xfer_buflen;
 987                         const enum fio_ddir ddir = acct_ddir(io_u);
 988
 989                         if (td->error)
 990                                 break;
 991
 992                         workqueue_enqueue(&td->io_wq, &io_u->work);
 993                         ret = FIO_Q_QUEUED;
 994
 995                         if (ddir_rw(ddir)) {
 996                                 td->io_issues[ddir]++;
 997                                 td->io_issue_bytes[ddir] += blen;
 998                                 td->rate_io_issue_bytes[ddir] += blen;
 999                         }
1000
1001                         if (should_check_rate(td))
1002                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1003
1004                 } else {
1005                         ret = td_io_queue(td, io_u);
1006
1007                         if (should_check_rate(td))
1008                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1009
1010                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1011                                 break;
1012
1013                         /*
1014                          * See if we need to complete some commands. Note that
1015                          * we can get BUSY even without IO queued, if the
1016                          * system is resource starved.
1017                          */
1018 reap:
1019                         full = queue_full(td) ||
1020                                 (ret == FIO_Q_BUSY && td->cur_depth);
1021                         if (full || io_in_polling(td))
1022                                 ret = wait_for_completions(td, &comp_time);
1023                 }
1024                 if (ret < 0)
1025                         break;
1026                 if (!ddir_rw_sum(td->bytes_done) &&
1027                     !td_ioengine_flagged(td, FIO_NOIO))
1028                         continue;
1029
1030                 if (!in_ramp_time(td) && should_check_rate(td)) {
1031                         if (check_min_rate(td, &comp_time)) {
1032                                 if (exitall_on_terminate || td->o.exitall_error)
1033                                         fio_terminate_threads(td->groupid);
1034                                 td_verror(td, EIO, "check_min_rate");
1035                                 break;
1036                         }
1037                 }
1038                 if (!in_ramp_time(td) && td->o.latency_target)
1039                         lat_target_check(td);
1040
1041                 if (td->o.thinktime) {
1042                         unsigned long long b;
1043
1044                         b = ddir_rw_sum(td->io_blocks);
1045                         if (!(b % td->o.thinktime_blocks)) {
1046                                 int left;
1047
1048                                 io_u_quiesce(td);
1049
1050                                 if (td->o.thinktime_spin)
1051                                         usec_spin(td->o.thinktime_spin);
1052
1053                                 left = td->o.thinktime - td->o.thinktime_spin;
1054                                 if (left)
1055                                         usec_sleep(td, left);
1056                         }
1057                 }
1058         }
1059
1060         check_update_rusage(td);
1061
1062         if (td->trim_entries)
1063                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1064
1065         if (td->o.fill_device && td->error == ENOSPC) {
1066                 td->error = 0;
1067                 fio_mark_td_terminate(td);
1068         }
1069         if (!td->error) {
1070                 struct fio_file *f;
1071
1072                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1073                         workqueue_flush(&td->io_wq);
1074                         i = 0;
1075                 } else
1076                         i = td->cur_depth;
1077
1078                 if (i) {
1079                         ret = io_u_queued_complete(td, i);
1080                         if (td->o.fill_device && td->error == ENOSPC)
1081                                 td->error = 0;
1082                 }
1083
1084                 if (should_fsync(td) && td->o.end_fsync) {
1085                         td_set_runstate(td, TD_FSYNCING);
1086
1087                         for_each_file(td, f, i) {
1088                                 if (!fio_file_fsync(td, f))
1089                                         continue;
1090
1091                                 log_err("fio: end_fsync failed for file %s\n",
1092                                                                 f->file_name);
1093                         }
1094                 }
1095         } else
1096                 cleanup_pending_aio(td);
1097
1098         /*
1099          * stop job if we failed doing any IO
1100          */
1101         if (!ddir_rw_sum(td->this_io_bytes))
1102                 td->done = 1;
1103
1104         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1105                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1106 }
1107
1108 static void free_file_completion_logging(struct thread_data *td)
1109 {
1110         struct fio_file *f;
1111         unsigned int i;
1112
1113         for_each_file(td, f, i) {
1114                 if (!f->last_write_comp)
1115                         break;
1116                 sfree(f->last_write_comp);
1117         }
1118 }
1119
1120 static int init_file_completion_logging(struct thread_data *td,
1121                                         unsigned int depth)
1122 {
1123         struct fio_file *f;
1124         unsigned int i;
1125
1126         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1127                 return 0;
1128
1129         for_each_file(td, f, i) {
1130                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1131                 if (!f->last_write_comp)
1132                         goto cleanup;
1133         }
1134
1135         return 0;
1136
1137 cleanup:
1138         free_file_completion_logging(td);
1139         log_err("fio: failed to alloc write comp data\n");
1140         return 1;
1141 }
1142
1143 static void cleanup_io_u(struct thread_data *td)
1144 {
1145         struct io_u *io_u;
1146
1147         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1148
1149                 if (td->io_ops->io_u_free)
1150                         td->io_ops->io_u_free(td, io_u);
1151
1152                 fio_memfree(io_u, sizeof(*io_u));
1153         }
1154
1155         free_io_mem(td);
1156
1157         io_u_rexit(&td->io_u_requeues);
1158         io_u_qexit(&td->io_u_freelist);
1159         io_u_qexit(&td->io_u_all);
1160
1161         free_file_completion_logging(td);
1162 }
1163
1164 static int init_io_u(struct thread_data *td)
1165 {
1166         struct io_u *io_u;
1167         unsigned int max_bs, min_write;
1168         int cl_align, i, max_units;
1169         int data_xfer = 1, err;
1170         char *p;
1171
1172         max_units = td->o.iodepth;
1173         max_bs = td_max_bs(td);
1174         min_write = td->o.min_bs[DDIR_WRITE];
1175         td->orig_buffer_size = (unsigned long long) max_bs
1176                                         * (unsigned long long) max_units;
1177
1178         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1179                 data_xfer = 0;
1180
1181         err = 0;
1182         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1183         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1184         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1185
1186         if (err) {
1187                 log_err("fio: failed setting up IO queues\n");
1188                 return 1;
1189         }
1190
1191         /*
1192          * if we may later need to do address alignment, then add any
1193          * possible adjustment here so that we don't cause a buffer
1194          * overflow later. this adjustment may be too much if we get
1195          * lucky and the allocator gives us an aligned address.
1196          */
1197         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1198             td_ioengine_flagged(td, FIO_RAWIO))
1199                 td->orig_buffer_size += page_mask + td->o.mem_align;
1200
1201         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1202                 unsigned long bs;
1203
1204                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1205                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1206         }
1207
1208         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1209                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1210                 return 1;
1211         }
1212
1213         if (data_xfer && allocate_io_mem(td))
1214                 return 1;
1215
1216         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1217             td_ioengine_flagged(td, FIO_RAWIO))
1218                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1219         else
1220                 p = td->orig_buffer;
1221
1222         cl_align = os_cache_line_size();
1223
1224         for (i = 0; i < max_units; i++) {
1225                 void *ptr;
1226
1227                 if (td->terminate)
1228                         return 1;
1229
1230                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1231                 if (!ptr) {
1232                         log_err("fio: unable to allocate aligned memory\n");
1233                         break;
1234                 }
1235
1236                 io_u = ptr;
1237                 memset(io_u, 0, sizeof(*io_u));
1238                 INIT_FLIST_HEAD(&io_u->verify_list);
1239                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1240
1241                 if (data_xfer) {
1242                         io_u->buf = p;
1243                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1244
1245                         if (td_write(td))
1246                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1247                         if (td_write(td) && td->o.verify_pattern_bytes) {
1248                                 /*
1249                                  * Fill the buffer with the pattern if we are
1250                                  * going to be doing writes.
1251                                  */
1252                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1253                         }
1254                 }
1255
1256                 io_u->index = i;
1257                 io_u->flags = IO_U_F_FREE;
1258                 io_u_qpush(&td->io_u_freelist, io_u);
1259
1260                 /*
1261                  * io_u never leaves this stack, used for iteration of all
1262                  * io_u buffers.
1263                  */
1264                 io_u_qpush(&td->io_u_all, io_u);
1265
1266                 if (td->io_ops->io_u_init) {
1267                         int ret = td->io_ops->io_u_init(td, io_u);
1268
1269                         if (ret) {
1270                                 log_err("fio: failed to init engine data: %d\n", ret);
1271                                 return 1;
1272                         }
1273                 }
1274
1275                 p += max_bs;
1276         }
1277
1278         if (init_file_completion_logging(td, max_units))
1279                 return 1;
1280
1281         return 0;
1282 }
1283
1284 static int switch_ioscheduler(struct thread_data *td)
1285 {
1286 #ifdef FIO_HAVE_IOSCHED_SWITCH
1287         char tmp[256], tmp2[128];
1288         FILE *f;
1289         int ret;
1290
1291         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1292                 return 0;
1293
1294         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1295
1296         f = fopen(tmp, "r+");
1297         if (!f) {
1298                 if (errno == ENOENT) {
1299                         log_err("fio: os or kernel doesn't support IO scheduler"
1300                                 " switching\n");
1301                         return 0;
1302                 }
1303                 td_verror(td, errno, "fopen iosched");
1304                 return 1;
1305         }
1306
1307         /*
1308          * Set io scheduler.
1309          */
1310         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1311         if (ferror(f) || ret != 1) {
1312                 td_verror(td, errno, "fwrite");
1313                 fclose(f);
1314                 return 1;
1315         }
1316
1317         rewind(f);
1318
1319         /*
1320          * Read back and check that the selected scheduler is now the default.
1321          */
1322         memset(tmp, 0, sizeof(tmp));
1323         ret = fread(tmp, sizeof(tmp), 1, f);
1324         if (ferror(f) || ret < 0) {
1325                 td_verror(td, errno, "fread");
1326                 fclose(f);
1327                 return 1;
1328         }
1329         /*
1330          * either a list of io schedulers or "none\n" is expected.
1331          */
1332         tmp[strlen(tmp) - 1] = '\0';
1333
1334         /*
1335          * Write to "none" entry doesn't fail, so check the result here.
1336          */
1337         if (!strcmp(tmp, "none")) {
1338                 log_err("fio: io scheduler is not tunable\n");
1339                 fclose(f);
1340                 return 0;
1341         }
1342
1343         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1344         if (!strstr(tmp, tmp2)) {
1345                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1346                 td_verror(td, EINVAL, "iosched_switch");
1347                 fclose(f);
1348                 return 1;
1349         }
1350
1351         fclose(f);
1352         return 0;
1353 #else
1354         return 0;
1355 #endif
1356 }
1357
1358 static bool keep_running(struct thread_data *td)
1359 {
1360         unsigned long long limit;
1361
1362         if (td->done)
1363                 return false;
1364         if (td->o.time_based)
1365                 return true;
1366         if (td->o.loops) {
1367                 td->o.loops--;
1368                 return true;
1369         }
1370         if (exceeds_number_ios(td))
1371                 return false;
1372
1373         if (td->o.io_limit)
1374                 limit = td->o.io_limit;
1375         else
1376                 limit = td->o.size;
1377
1378         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1379                 uint64_t diff;
1380
1381                 /*
1382                  * If the difference is less than the minimum IO size, we
1383                  * are done.
1384                  */
1385                 diff = limit - ddir_rw_sum(td->io_bytes);
1386                 if (diff < td_max_bs(td))
1387                         return false;
1388
1389                 if (fio_files_done(td) && !td->o.io_limit)
1390                         return false;
1391
1392                 return true;
1393         }
1394
1395         return false;
1396 }
1397
1398 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1399 {
1400         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1401         int ret;
1402         char *str;
1403
1404         str = malloc(newlen);
1405         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1406
1407         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1408         ret = system(str);
1409         if (ret == -1)
1410                 log_err("fio: exec of cmd <%s> failed\n", str);
1411
1412         free(str);
1413         return ret;
1414 }
1415
1416 /*
1417  * Dry run to compute correct state of numberio for verification.
1418  */
1419 static uint64_t do_dry_run(struct thread_data *td)
1420 {
1421         td_set_runstate(td, TD_RUNNING);
1422
1423         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1424                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1425                 struct io_u *io_u;
1426                 int ret;
1427
1428                 if (td->terminate || td->done)
1429                         break;
1430
1431                 io_u = get_io_u(td);
1432                 if (IS_ERR_OR_NULL(io_u))
1433                         break;
1434
1435                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1436                 io_u->error = 0;
1437                 io_u->resid = 0;
1438                 if (ddir_rw(acct_ddir(io_u)))
1439                         td->io_issues[acct_ddir(io_u)]++;
1440                 if (ddir_rw(io_u->ddir)) {
1441                         io_u_mark_depth(td, 1);
1442                         td->ts.total_io_u[io_u->ddir]++;
1443                 }
1444
1445                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1446                     td->o.do_verify &&
1447                     td->o.verify != VERIFY_NONE &&
1448                     !td->o.experimental_verify)
1449                         log_io_piece(td, io_u);
1450
1451                 ret = io_u_sync_complete(td, io_u);
1452                 (void) ret;
1453         }
1454
1455         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1456 }
1457
1458 struct fork_data {
1459         struct thread_data *td;
1460         struct sk_out *sk_out;
1461 };
1462
1463 /*
1464  * Entry point for the thread based jobs. The process based jobs end up
1465  * here as well, after a little setup.
1466  */
1467 static void *thread_main(void *data)
1468 {
1469         struct fork_data *fd = data;
1470         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1471         struct thread_data *td = fd->td;
1472         struct thread_options *o = &td->o;
1473         struct sk_out *sk_out = fd->sk_out;
1474         int deadlock_loop_cnt;
1475         int clear_state;
1476         int ret;
1477
1478         sk_out_assign(sk_out);
1479         free(fd);
1480
1481         if (!o->use_thread) {
1482                 setsid();
1483                 td->pid = getpid();
1484         } else
1485                 td->pid = gettid();
1486
1487         fio_local_clock_init(o->use_thread);
1488
1489         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1490
1491         if (is_backend)
1492                 fio_server_send_start(td);
1493
1494         INIT_FLIST_HEAD(&td->io_log_list);
1495         INIT_FLIST_HEAD(&td->io_hist_list);
1496         INIT_FLIST_HEAD(&td->verify_list);
1497         INIT_FLIST_HEAD(&td->trim_list);
1498         INIT_FLIST_HEAD(&td->next_rand_list);
1499         td->io_hist_tree = RB_ROOT;
1500
1501         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1502         if (ret) {
1503                 td_verror(td, ret, "mutex_cond_init_pshared");
1504                 goto err;
1505         }
1506         ret = cond_init_pshared(&td->verify_cond);
1507         if (ret) {
1508                 td_verror(td, ret, "mutex_cond_pshared");
1509                 goto err;
1510         }
1511
1512         td_set_runstate(td, TD_INITIALIZED);
1513         dprint(FD_MUTEX, "up startup_mutex\n");
1514         fio_mutex_up(startup_mutex);
1515         dprint(FD_MUTEX, "wait on td->mutex\n");
1516         fio_mutex_down(td->mutex);
1517         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1518
1519         /*
1520          * A new gid requires privilege, so we need to do this before setting
1521          * the uid.
1522          */
1523         if (o->gid != -1U && setgid(o->gid)) {
1524                 td_verror(td, errno, "setgid");
1525                 goto err;
1526         }
1527         if (o->uid != -1U && setuid(o->uid)) {
1528                 td_verror(td, errno, "setuid");
1529                 goto err;
1530         }
1531
1532         /*
1533          * Do this early, we don't want the compress threads to be limited
1534          * to the same CPUs as the IO workers. So do this before we set
1535          * any potential CPU affinity
1536          */
1537         if (iolog_compress_init(td, sk_out))
1538                 goto err;
1539
1540         /*
1541          * If we have a gettimeofday() thread, make sure we exclude that
1542          * thread from this job
1543          */
1544         if (o->gtod_cpu)
1545                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1546
1547         /*
1548          * Set affinity first, in case it has an impact on the memory
1549          * allocations.
1550          */
1551         if (fio_option_is_set(o, cpumask)) {
1552                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1553                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1554                         if (!ret) {
1555                                 log_err("fio: no CPUs set\n");
1556                                 log_err("fio: Try increasing number of available CPUs\n");
1557                                 td_verror(td, EINVAL, "cpus_split");
1558                                 goto err;
1559                         }
1560                 }
1561                 ret = fio_setaffinity(td->pid, o->cpumask);
1562                 if (ret == -1) {
1563                         td_verror(td, errno, "cpu_set_affinity");
1564                         goto err;
1565                 }
1566         }
1567
1568 #ifdef CONFIG_LIBNUMA
1569         /* numa node setup */
1570         if (fio_option_is_set(o, numa_cpunodes) ||
1571             fio_option_is_set(o, numa_memnodes)) {
1572                 struct bitmask *mask;
1573
1574                 if (numa_available() < 0) {
1575                         td_verror(td, errno, "Does not support NUMA API\n");
1576                         goto err;
1577                 }
1578
1579                 if (fio_option_is_set(o, numa_cpunodes)) {
1580                         mask = numa_parse_nodestring(o->numa_cpunodes);
1581                         ret = numa_run_on_node_mask(mask);
1582                         numa_free_nodemask(mask);
1583                         if (ret == -1) {
1584                                 td_verror(td, errno, \
1585                                         "numa_run_on_node_mask failed\n");
1586                                 goto err;
1587                         }
1588                 }
1589
1590                 if (fio_option_is_set(o, numa_memnodes)) {
1591                         mask = NULL;
1592                         if (o->numa_memnodes)
1593                                 mask = numa_parse_nodestring(o->numa_memnodes);
1594
1595                         switch (o->numa_mem_mode) {
1596                         case MPOL_INTERLEAVE:
1597                                 numa_set_interleave_mask(mask);
1598                                 break;
1599                         case MPOL_BIND:
1600                                 numa_set_membind(mask);
1601                                 break;
1602                         case MPOL_LOCAL:
1603                                 numa_set_localalloc();
1604                                 break;
1605                         case MPOL_PREFERRED:
1606                                 numa_set_preferred(o->numa_mem_prefer_node);
1607                                 break;
1608                         case MPOL_DEFAULT:
1609                         default:
1610                                 break;
1611                         }
1612
1613                         if (mask)
1614                                 numa_free_nodemask(mask);
1615
1616                 }
1617         }
1618 #endif
1619
1620         if (fio_pin_memory(td))
1621                 goto err;
1622
1623         /*
1624          * May alter parameters that init_io_u() will use, so we need to
1625          * do this first.
1626          */
1627         if (init_iolog(td))
1628                 goto err;
1629
1630         if (init_io_u(td))
1631                 goto err;
1632
1633         if (o->verify_async && verify_async_init(td))
1634                 goto err;
1635
1636         if (fio_option_is_set(o, ioprio) ||
1637             fio_option_is_set(o, ioprio_class)) {
1638                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1639                 if (ret == -1) {
1640                         td_verror(td, errno, "ioprio_set");
1641                         goto err;
1642                 }
1643         }
1644
1645         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1646                 goto err;
1647
1648         errno = 0;
1649         if (nice(o->nice) == -1 && errno != 0) {
1650                 td_verror(td, errno, "nice");
1651                 goto err;
1652         }
1653
1654         if (o->ioscheduler && switch_ioscheduler(td))
1655                 goto err;
1656
1657         if (!o->create_serialize && setup_files(td))
1658                 goto err;
1659
1660         if (td_io_init(td))
1661                 goto err;
1662
1663         if (init_random_map(td))
1664                 goto err;
1665
1666         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1667                 goto err;
1668
1669         if (o->pre_read) {
1670                 if (pre_read_files(td) < 0)
1671                         goto err;
1672         }
1673
1674         fio_verify_init(td);
1675
1676         if (rate_submit_init(td, sk_out))
1677                 goto err;
1678
1679         set_epoch_time(td, o->log_unix_epoch);
1680         fio_getrusage(&td->ru_start);
1681         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1682         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1683
1684         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1685                         o->ratemin[DDIR_TRIM]) {
1686                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1687                                         sizeof(td->bw_sample_time));
1688                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1689                                         sizeof(td->bw_sample_time));
1690                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1691                                         sizeof(td->bw_sample_time));
1692         }
1693
1694         clear_state = 0;
1695         while (keep_running(td)) {
1696                 uint64_t verify_bytes;
1697
1698                 fio_gettime(&td->start, NULL);
1699                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1700
1701                 if (clear_state) {
1702                         clear_io_state(td, 0);
1703
1704                         if (o->unlink_each_loop && unlink_all_files(td))
1705                                 break;
1706                 }
1707
1708                 prune_io_piece_log(td);
1709
1710                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1711                         verify_bytes = do_dry_run(td);
1712                 else {
1713                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1714
1715                         do_io(td, bytes_done);
1716
1717                         if (!ddir_rw_sum(bytes_done)) {
1718                                 fio_mark_td_terminate(td);
1719                                 verify_bytes = 0;
1720                         } else {
1721                                 verify_bytes = bytes_done[DDIR_WRITE] +
1722                                                 bytes_done[DDIR_TRIM];
1723                         }
1724                 }
1725
1726                 clear_state = 1;
1727
1728                 /*
1729                  * Make sure we've successfully updated the rusage stats
1730                  * before waiting on the stat mutex. Otherwise we could have
1731                  * the stat thread holding stat mutex and waiting for
1732                  * the rusage_sem, which would never get upped because
1733                  * this thread is waiting for the stat mutex.
1734                  */
1735                 deadlock_loop_cnt = 0;
1736                 do {
1737                         check_update_rusage(td);
1738                         if (!fio_mutex_down_trylock(stat_mutex))
1739                                 break;
1740                         usleep(1000);
1741                         if (deadlock_loop_cnt++ > 5000) {
1742                                 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1743                                 td->error = EDEADLOCK;
1744                                 goto err;
1745                         }
1746                 } while (1);
1747
1748                 if (td_read(td) && td->io_bytes[DDIR_READ])
1749                         update_runtime(td, elapsed_us, DDIR_READ);
1750                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1751                         update_runtime(td, elapsed_us, DDIR_WRITE);
1752                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1753                         update_runtime(td, elapsed_us, DDIR_TRIM);
1754                 fio_gettime(&td->start, NULL);
1755                 fio_mutex_up(stat_mutex);
1756
1757                 if (td->error || td->terminate)
1758                         break;
1759
1760                 if (!o->do_verify ||
1761                     o->verify == VERIFY_NONE ||
1762                     td_ioengine_flagged(td, FIO_UNIDIR))
1763                         continue;
1764
1765                 clear_io_state(td, 0);
1766
1767                 fio_gettime(&td->start, NULL);
1768
1769                 do_verify(td, verify_bytes);
1770
1771                 /*
1772                  * See comment further up for why this is done here.
1773                  */
1774                 check_update_rusage(td);
1775
1776                 fio_mutex_down(stat_mutex);
1777                 update_runtime(td, elapsed_us, DDIR_READ);
1778                 fio_gettime(&td->start, NULL);
1779                 fio_mutex_up(stat_mutex);
1780
1781                 if (td->error || td->terminate)
1782                         break;
1783         }
1784
1785         td_set_runstate(td, TD_FINISHING);
1786
1787         update_rusage_stat(td);
1788         td->ts.total_run_time = mtime_since_now(&td->epoch);
1789         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1790         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1791         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1792
1793         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1794             (td->o.verify != VERIFY_NONE && td_write(td)))
1795                 verify_save_state(td->thread_number);
1796
1797         fio_unpin_memory(td);
1798
1799         td_writeout_logs(td, true);
1800
1801         iolog_compress_exit(td);
1802         rate_submit_exit(td);
1803
1804         if (o->exec_postrun)
1805                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1806
1807         if (exitall_on_terminate || (o->exitall_error && td->error))
1808                 fio_terminate_threads(td->groupid);
1809
1810 err:
1811         if (td->error)
1812                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1813                                                         td->verror);
1814
1815         if (o->verify_async)
1816                 verify_async_exit(td);
1817
1818         close_and_free_files(td);
1819         cleanup_io_u(td);
1820         close_ioengine(td);
1821         cgroup_shutdown(td, &cgroup_mnt);
1822         verify_free_state(td);
1823
1824         if (td->zone_state_index) {
1825                 int i;
1826
1827                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1828                         free(td->zone_state_index[i]);
1829                 free(td->zone_state_index);
1830                 td->zone_state_index = NULL;
1831         }
1832
1833         if (fio_option_is_set(o, cpumask)) {
1834                 ret = fio_cpuset_exit(&o->cpumask);
1835                 if (ret)
1836                         td_verror(td, ret, "fio_cpuset_exit");
1837         }
1838
1839         /*
1840          * do this very late, it will log file closing as well
1841          */
1842         if (o->write_iolog_file)
1843                 write_iolog_close(td);
1844
1845         fio_mutex_remove(td->mutex);
1846         td->mutex = NULL;
1847
1848         td_set_runstate(td, TD_EXITED);
1849
1850         /*
1851          * Do this last after setting our runstate to exited, so we
1852          * know that the stat thread is signaled.
1853          */
1854         check_update_rusage(td);
1855
1856         sk_out_drop();
1857         return (void *) (uintptr_t) td->error;
1858 }
1859
1860 static void dump_td_info(struct thread_data *td)
1861 {
1862         log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
1863                 "appears to be stuck. Doing forceful exit of this job.\n",
1864                         td->o.name, td->runstate,
1865                         (unsigned long) time_since_now(&td->terminate_time));
1866 }
1867
1868 /*
1869  * Run over the job map and reap the threads that have exited, if any.
1870  */
1871 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1872                          unsigned int *m_rate)
1873 {
1874         struct thread_data *td;
1875         unsigned int cputhreads, realthreads, pending;
1876         int i, status, ret;
1877
1878         /*
1879          * reap exited threads (TD_EXITED -> TD_REAPED)
1880          */
1881         realthreads = pending = cputhreads = 0;
1882         for_each_td(td, i) {
1883                 int flags = 0;
1884
1885                 /*
1886                  * ->io_ops is NULL for a thread that has closed its
1887                  * io engine
1888                  */
1889                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1890                         cputhreads++;
1891                 else
1892                         realthreads++;
1893
1894                 if (!td->pid) {
1895                         pending++;
1896                         continue;
1897                 }
1898                 if (td->runstate == TD_REAPED)
1899                         continue;
1900                 if (td->o.use_thread) {
1901                         if (td->runstate == TD_EXITED) {
1902                                 td_set_runstate(td, TD_REAPED);
1903                                 goto reaped;
1904                         }
1905                         continue;
1906                 }
1907
1908                 flags = WNOHANG;
1909                 if (td->runstate == TD_EXITED)
1910                         flags = 0;
1911
1912                 /*
1913                  * check if someone quit or got killed in an unusual way
1914                  */
1915                 ret = waitpid(td->pid, &status, flags);
1916                 if (ret < 0) {
1917                         if (errno == ECHILD) {
1918                                 log_err("fio: pid=%d disappeared %d\n",
1919                                                 (int) td->pid, td->runstate);
1920                                 td->sig = ECHILD;
1921                                 td_set_runstate(td, TD_REAPED);
1922                                 goto reaped;
1923                         }
1924                         perror("waitpid");
1925                 } else if (ret == td->pid) {
1926                         if (WIFSIGNALED(status)) {
1927                                 int sig = WTERMSIG(status);
1928
1929                                 if (sig != SIGTERM && sig != SIGUSR2)
1930                                         log_err("fio: pid=%d, got signal=%d\n",
1931                                                         (int) td->pid, sig);
1932                                 td->sig = sig;
1933                                 td_set_runstate(td, TD_REAPED);
1934                                 goto reaped;
1935                         }
1936                         if (WIFEXITED(status)) {
1937                                 if (WEXITSTATUS(status) && !td->error)
1938                                         td->error = WEXITSTATUS(status);
1939
1940                                 td_set_runstate(td, TD_REAPED);
1941                                 goto reaped;
1942                         }
1943                 }
1944
1945                 /*
1946                  * If the job is stuck, do a forceful timeout of it and
1947                  * move on.
1948                  */
1949                 if (td->terminate &&
1950                     td->runstate < TD_FSYNCING &&
1951                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1952                         dump_td_info(td);
1953                         td_set_runstate(td, TD_REAPED);
1954                         goto reaped;
1955                 }
1956
1957                 /*
1958                  * thread is not dead, continue
1959                  */
1960                 pending++;
1961                 continue;
1962 reaped:
1963                 (*nr_running)--;
1964                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1965                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1966                 if (!td->pid)
1967                         pending--;
1968
1969                 if (td->error)
1970                         exit_value++;
1971
1972                 done_secs += mtime_since_now(&td->epoch) / 1000;
1973                 profile_td_exit(td);
1974         }
1975
1976         if (*nr_running == cputhreads && !pending && realthreads)
1977                 fio_terminate_threads(TERMINATE_ALL);
1978 }
1979
1980 static bool __check_trigger_file(void)
1981 {
1982         struct stat sb;
1983
1984         if (!trigger_file)
1985                 return false;
1986
1987         if (stat(trigger_file, &sb))
1988                 return false;
1989
1990         if (unlink(trigger_file) < 0)
1991                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1992                                                         strerror(errno));
1993
1994         return true;
1995 }
1996
1997 static bool trigger_timedout(void)
1998 {
1999         if (trigger_timeout)
2000                 return time_since_genesis() >= trigger_timeout;
2001
2002         return false;
2003 }
2004
2005 void exec_trigger(const char *cmd)
2006 {
2007         int ret;
2008
2009         if (!cmd)
2010                 return;
2011
2012         ret = system(cmd);
2013         if (ret == -1)
2014                 log_err("fio: failed executing %s trigger\n", cmd);
2015 }
2016
2017 void check_trigger_file(void)
2018 {
2019         if (__check_trigger_file() || trigger_timedout()) {
2020                 if (nr_clients)
2021                         fio_clients_send_trigger(trigger_remote_cmd);
2022                 else {
2023                         verify_save_state(IO_LIST_ALL);
2024                         fio_terminate_threads(TERMINATE_ALL);
2025                         exec_trigger(trigger_cmd);
2026                 }
2027         }
2028 }
2029
2030 static int fio_verify_load_state(struct thread_data *td)
2031 {
2032         int ret;
2033
2034         if (!td->o.verify_state)
2035                 return 0;
2036
2037         if (is_backend) {
2038                 void *data;
2039
2040                 ret = fio_server_get_verify_state(td->o.name,
2041                                         td->thread_number - 1, &data);
2042                 if (!ret)
2043                         verify_assign_state(td, data);
2044         } else
2045                 ret = verify_load_state(td, "local");
2046
2047         return ret;
2048 }
2049
2050 static void do_usleep(unsigned int usecs)
2051 {
2052         check_for_running_stats();
2053         check_trigger_file();
2054         usleep(usecs);
2055 }
2056
2057 static bool check_mount_writes(struct thread_data *td)
2058 {
2059         struct fio_file *f;
2060         unsigned int i;
2061
2062         if (!td_write(td) || td->o.allow_mounted_write)
2063                 return false;
2064
2065         for_each_file(td, f, i) {
2066                 if (f->filetype != FIO_TYPE_BD)
2067                         continue;
2068                 if (device_is_mounted(f->file_name))
2069                         goto mounted;
2070         }
2071
2072         return false;
2073 mounted:
2074         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2075         return true;
2076 }
2077
2078 static bool waitee_running(struct thread_data *me)
2079 {
2080         const char *waitee = me->o.wait_for;
2081         const char *self = me->o.name;
2082         struct thread_data *td;
2083         int i;
2084
2085         if (!waitee)
2086                 return false;
2087
2088         for_each_td(td, i) {
2089                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2090                         continue;
2091
2092                 if (td->runstate < TD_EXITED) {
2093                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2094                                         self, td->o.name,
2095                                         runstate_to_name(td->runstate));
2096                         return true;
2097                 }
2098         }
2099
2100         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2101         return false;
2102 }
2103
2104 /*
2105  * Main function for kicking off and reaping jobs, as needed.
2106  */
2107 static void run_threads(struct sk_out *sk_out)
2108 {
2109         struct thread_data *td;
2110         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2111         uint64_t spent;
2112
2113         if (fio_gtod_offload && fio_start_gtod_thread())
2114                 return;
2115
2116         fio_idle_prof_init();
2117
2118         set_sig_handlers();
2119
2120         nr_thread = nr_process = 0;
2121         for_each_td(td, i) {
2122                 if (check_mount_writes(td))
2123                         return;
2124                 if (td->o.use_thread)
2125                         nr_thread++;
2126                 else
2127                         nr_process++;
2128         }
2129
2130         if (output_format & FIO_OUTPUT_NORMAL) {
2131                 log_info("Starting ");
2132                 if (nr_thread)
2133                         log_info("%d thread%s", nr_thread,
2134                                                 nr_thread > 1 ? "s" : "");
2135                 if (nr_process) {
2136                         if (nr_thread)
2137                                 log_info(" and ");
2138                         log_info("%d process%s", nr_process,
2139                                                 nr_process > 1 ? "es" : "");
2140                 }
2141                 log_info("\n");
2142                 log_info_flush();
2143         }
2144
2145         todo = thread_number;
2146         nr_running = 0;
2147         nr_started = 0;
2148         m_rate = t_rate = 0;
2149
2150         for_each_td(td, i) {
2151                 print_status_init(td->thread_number - 1);
2152
2153                 if (!td->o.create_serialize)
2154                         continue;
2155
2156                 if (fio_verify_load_state(td))
2157                         goto reap;
2158
2159                 /*
2160                  * do file setup here so it happens sequentially,
2161                  * we don't want X number of threads getting their
2162                  * client data interspersed on disk
2163                  */
2164                 if (setup_files(td)) {
2165 reap:
2166                         exit_value++;
2167                         if (td->error)
2168                                 log_err("fio: pid=%d, err=%d/%s\n",
2169                                         (int) td->pid, td->error, td->verror);
2170                         td_set_runstate(td, TD_REAPED);
2171                         todo--;
2172                 } else {
2173                         struct fio_file *f;
2174                         unsigned int j;
2175
2176                         /*
2177                          * for sharing to work, each job must always open
2178                          * its own files. so close them, if we opened them
2179                          * for creation
2180                          */
2181                         for_each_file(td, f, j) {
2182                                 if (fio_file_open(f))
2183                                         td_io_close_file(td, f);
2184                         }
2185                 }
2186         }
2187
2188         /* start idle threads before io threads start to run */
2189         fio_idle_prof_start();
2190
2191         set_genesis_time();
2192
2193         while (todo) {
2194                 struct thread_data *map[REAL_MAX_JOBS];
2195                 struct timeval this_start;
2196                 int this_jobs = 0, left;
2197                 struct fork_data *fd;
2198
2199                 /*
2200                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2201                  */
2202                 for_each_td(td, i) {
2203                         if (td->runstate != TD_NOT_CREATED)
2204                                 continue;
2205
2206                         /*
2207                          * never got a chance to start, killed by other
2208                          * thread for some reason
2209                          */
2210                         if (td->terminate) {
2211                                 todo--;
2212                                 continue;
2213                         }
2214
2215                         if (td->o.start_delay) {
2216                                 spent = utime_since_genesis();
2217
2218                                 if (td->o.start_delay > spent)
2219                                         continue;
2220                         }
2221
2222                         if (td->o.stonewall && (nr_started || nr_running)) {
2223                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2224                                                         td->o.name);
2225                                 break;
2226                         }
2227
2228                         if (waitee_running(td)) {
2229                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2230                                                 td->o.name, td->o.wait_for);
2231                                 continue;
2232                         }
2233
2234                         init_disk_util(td);
2235
2236                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2237                         td->update_rusage = 0;
2238
2239                         /*
2240                          * Set state to created. Thread will transition
2241                          * to TD_INITIALIZED when it's done setting up.
2242                          */
2243                         td_set_runstate(td, TD_CREATED);
2244                         map[this_jobs++] = td;
2245                         nr_started++;
2246
2247                         fd = calloc(1, sizeof(*fd));
2248                         fd->td = td;
2249                         fd->sk_out = sk_out;
2250
2251                         if (td->o.use_thread) {
2252                                 int ret;
2253
2254                                 dprint(FD_PROCESS, "will pthread_create\n");
2255                                 ret = pthread_create(&td->thread, NULL,
2256                                                         thread_main, fd);
2257                                 if (ret) {
2258                                         log_err("pthread_create: %s\n",
2259                                                         strerror(ret));
2260                                         free(fd);
2261                                         nr_started--;
2262                                         break;
2263                                 }
2264                                 ret = pthread_detach(td->thread);
2265                                 if (ret)
2266                                         log_err("pthread_detach: %s",
2267                                                         strerror(ret));
2268                         } else {
2269                                 pid_t pid;
2270                                 dprint(FD_PROCESS, "will fork\n");
2271                                 pid = fork();
2272                                 if (!pid) {
2273                                         int ret;
2274
2275                                         ret = (int)(uintptr_t)thread_main(fd);
2276                                         _exit(ret);
2277                                 } else if (i == fio_debug_jobno)
2278                                         *fio_debug_jobp = pid;
2279                         }
2280                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2281                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2282                                 log_err("fio: job startup hung? exiting.\n");
2283                                 fio_terminate_threads(TERMINATE_ALL);
2284                                 fio_abort = 1;
2285                                 nr_started--;
2286                                 break;
2287                         }
2288                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2289                 }
2290
2291                 /*
2292                  * Wait for the started threads to transition to
2293                  * TD_INITIALIZED.
2294                  */
2295                 fio_gettime(&this_start, NULL);
2296                 left = this_jobs;
2297                 while (left && !fio_abort) {
2298                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2299                                 break;
2300
2301                         do_usleep(100000);
2302
2303                         for (i = 0; i < this_jobs; i++) {
2304                                 td = map[i];
2305                                 if (!td)
2306                                         continue;
2307                                 if (td->runstate == TD_INITIALIZED) {
2308                                         map[i] = NULL;
2309                                         left--;
2310                                 } else if (td->runstate >= TD_EXITED) {
2311                                         map[i] = NULL;
2312                                         left--;
2313                                         todo--;
2314                                         nr_running++; /* work-around... */
2315                                 }
2316                         }
2317                 }
2318
2319                 if (left) {
2320                         log_err("fio: %d job%s failed to start\n", left,
2321                                         left > 1 ? "s" : "");
2322                         for (i = 0; i < this_jobs; i++) {
2323                                 td = map[i];
2324                                 if (!td)
2325                                         continue;
2326                                 kill(td->pid, SIGTERM);
2327                         }
2328                         break;
2329                 }
2330
2331                 /*
2332                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2333                  */
2334                 for_each_td(td, i) {
2335                         if (td->runstate != TD_INITIALIZED)
2336                                 continue;
2337
2338                         if (in_ramp_time(td))
2339                                 td_set_runstate(td, TD_RAMP);
2340                         else
2341                                 td_set_runstate(td, TD_RUNNING);
2342                         nr_running++;
2343                         nr_started--;
2344                         m_rate += ddir_rw_sum(td->o.ratemin);
2345                         t_rate += ddir_rw_sum(td->o.rate);
2346                         todo--;
2347                         fio_mutex_up(td->mutex);
2348                 }
2349
2350                 reap_threads(&nr_running, &t_rate, &m_rate);
2351
2352                 if (todo)
2353                         do_usleep(100000);
2354         }
2355
2356         while (nr_running) {
2357                 reap_threads(&nr_running, &t_rate, &m_rate);
2358                 do_usleep(10000);
2359         }
2360
2361         fio_idle_prof_stop();
2362
2363         update_io_ticks();
2364 }
2365
2366 static void free_disk_util(void)
2367 {
2368         disk_util_prune_entries();
2369         helper_thread_destroy();
2370 }
2371
2372 int fio_backend(struct sk_out *sk_out)
2373 {
2374         struct thread_data *td;
2375         int i;
2376
2377         if (exec_profile) {
2378                 if (load_profile(exec_profile))
2379                         return 1;
2380                 free(exec_profile);
2381                 exec_profile = NULL;
2382         }
2383         if (!thread_number)
2384                 return 0;
2385
2386         if (write_bw_log) {
2387                 struct log_params p = {
2388                         .log_type = IO_LOG_TYPE_BW,
2389                 };
2390
2391                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2392                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2393                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2394         }
2395
2396         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2397         if (startup_mutex == NULL)
2398                 return 1;
2399
2400         set_genesis_time();
2401         stat_init();
2402         helper_thread_create(startup_mutex, sk_out);
2403
2404         cgroup_list = smalloc(sizeof(*cgroup_list));
2405         INIT_FLIST_HEAD(cgroup_list);
2406
2407         run_threads(sk_out);
2408
2409         helper_thread_exit();
2410
2411         if (!fio_abort) {
2412                 __show_run_stats();
2413                 if (write_bw_log) {
2414                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2415                                 struct io_log *log = agg_io_log[i];
2416
2417                                 flush_log(log, false);
2418                                 free_log(log);
2419                         }
2420                 }
2421         }
2422
2423         for_each_td(td, i) {
2424                 fio_options_free(td);
2425                 if (td->rusage_sem) {
2426                         fio_mutex_remove(td->rusage_sem);
2427                         td->rusage_sem = NULL;
2428                 }
2429         }
2430
2431         free_disk_util();
2432         cgroup_kill(cgroup_list);
2433         sfree(cgroup_list);
2434         sfree(cgroup_mnt);
2435
2436         fio_mutex_remove(startup_mutex);
2437         stat_exit();
2438         return exit_value;
2439 }