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                 /*
 528                  * when doing I/O (not when verifying),
 529                  * check for any errors that are to be ignored
 530                  */
 531                 if (!from_verify)
 532                         break;
 533
 534                 return 0;
 535         case FIO_Q_QUEUED:
 536                 /*
 537                  * if the engine doesn't have a commit hook,
 538                  * the io_u is really queued. if it does have such
 539                  * a hook, it has to call io_u_queued() itself.
 540                  */
 541                 if (td->io_ops->commit == NULL)
 542                         io_u_queued(td, io_u);
 543                 if (bytes_issued)
 544                         *bytes_issued += io_u->xfer_buflen;
 545                 break;
 546         case FIO_Q_BUSY:
 547                 if (!from_verify)
 548                         unlog_io_piece(td, io_u);
 549                 requeue_io_u(td, &io_u);
 550                 ret2 = td_io_commit(td);
 551                 if (ret2 < 0)
 552                         *ret = ret2;
 553                 break;
 554         default:
 555                 assert(*ret < 0);
 556                 td_verror(td, -(*ret), "td_io_queue");
 557                 break;
 558         }
 559
 560         if (break_on_this_error(td, ddir, ret))
 561                 return 1;
 562
 563         return 0;
 564 }
 565
 566 static inline bool io_in_polling(struct thread_data *td)
 567 {
 568         return !td->o.iodepth_batch_complete_min &&
 569                    !td->o.iodepth_batch_complete_max;
 570 }
 571
 572 /*
 573  * The main verify engine. Runs over the writes we previously submitted,
 574  * reads the blocks back in, and checks the crc/md5 of the data.
 575  */
 576 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 577 {
 578         struct fio_file *f;
 579         struct io_u *io_u;
 580         int ret, min_events;
 581         unsigned int i;
 582
 583         dprint(FD_VERIFY, "starting loop\n");
 584
 585         /*
 586          * sync io first and invalidate cache, to make sure we really
 587          * read from disk.
 588          */
 589         for_each_file(td, f, i) {
 590                 if (!fio_file_open(f))
 591                         continue;
 592                 if (fio_io_sync(td, f))
 593                         break;
 594                 if (file_invalidate_cache(td, f))
 595                         break;
 596         }
 597
 598         check_update_rusage(td);
 599
 600         if (td->error)
 601                 return;
 602
 603         td_set_runstate(td, TD_VERIFYING);
 604
 605         io_u = NULL;
 606         while (!td->terminate) {
 607                 enum fio_ddir ddir;
 608                 int full;
 609
 610                 update_tv_cache(td);
 611                 check_update_rusage(td);
 612
 613                 if (runtime_exceeded(td, &td->tv_cache)) {
 614                         __update_tv_cache(td);
 615                         if (runtime_exceeded(td, &td->tv_cache)) {
 616                                 fio_mark_td_terminate(td);
 617                                 break;
 618                         }
 619                 }
 620
 621                 if (flow_threshold_exceeded(td))
 622                         continue;
 623
 624                 if (!td->o.experimental_verify) {
 625                         io_u = __get_io_u(td);
 626                         if (!io_u)
 627                                 break;
 628
 629                         if (get_next_verify(td, io_u)) {
 630                                 put_io_u(td, io_u);
 631                                 break;
 632                         }
 633
 634                         if (td_io_prep(td, io_u)) {
 635                                 put_io_u(td, io_u);
 636                                 break;
 637                         }
 638                 } else {
 639                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 640                                 break;
 641
 642                         while ((io_u = get_io_u(td)) != NULL) {
 643                                 if (IS_ERR(io_u)) {
 644                                         io_u = NULL;
 645                                         ret = FIO_Q_BUSY;
 646                                         goto reap;
 647                                 }
 648
 649                                 /*
 650                                  * We are only interested in the places where
 651                                  * we wrote or trimmed IOs. Turn those into
 652                                  * reads for verification purposes.
 653                                  */
 654                                 if (io_u->ddir == DDIR_READ) {
 655                                         /*
 656                                          * Pretend we issued it for rwmix
 657                                          * accounting
 658                                          */
 659                                         td->io_issues[DDIR_READ]++;
 660                                         put_io_u(td, io_u);
 661                                         continue;
 662                                 } else if (io_u->ddir == DDIR_TRIM) {
 663                                         io_u->ddir = DDIR_READ;
 664                                         io_u_set(io_u, IO_U_F_TRIMMED);
 665                                         break;
 666                                 } else if (io_u->ddir == DDIR_WRITE) {
 667                                         io_u->ddir = DDIR_READ;
 668                                         break;
 669                                 } else {
 670                                         put_io_u(td, io_u);
 671                                         continue;
 672                                 }
 673                         }
 674
 675                         if (!io_u)
 676                                 break;
 677                 }
 678
 679                 if (verify_state_should_stop(td, io_u)) {
 680                         put_io_u(td, io_u);
 681                         break;
 682                 }
 683
 684                 if (td->o.verify_async)
 685                         io_u->end_io = verify_io_u_async;
 686                 else
 687                         io_u->end_io = verify_io_u;
 688
 689                 ddir = io_u->ddir;
 690                 if (!td->o.disable_slat)
 691                         fio_gettime(&io_u->start_time, NULL);
 692
 693                 ret = td_io_queue(td, io_u);
 694
 695                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 696                         break;
 697
 698                 /*
 699                  * if we can queue more, do so. but check if there are
 700                  * completed io_u's first. Note that we can get BUSY even
 701                  * without IO queued, if the system is resource starved.
 702                  */
 703 reap:
 704                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 705                 if (full || io_in_polling(td))
 706                         ret = wait_for_completions(td, NULL);
 707
 708                 if (ret < 0)
 709                         break;
 710         }
 711
 712         check_update_rusage(td);
 713
 714         if (!td->error) {
 715                 min_events = td->cur_depth;
 716
 717                 if (min_events)
 718                         ret = io_u_queued_complete(td, min_events);
 719         } else
 720                 cleanup_pending_aio(td);
 721
 722         td_set_runstate(td, TD_RUNNING);
 723
 724         dprint(FD_VERIFY, "exiting loop\n");
 725 }
 726
 727 static bool exceeds_number_ios(struct thread_data *td)
 728 {
 729         unsigned long long number_ios;
 730
 731         if (!td->o.number_ios)
 732                 return false;
 733
 734         number_ios = ddir_rw_sum(td->io_blocks);
 735         number_ios += td->io_u_queued + td->io_u_in_flight;
 736
 737         return number_ios >= (td->o.number_ios * td->loops);
 738 }
 739
 740 static bool io_issue_bytes_exceeded(struct thread_data *td)
 741 {
 742         unsigned long long bytes, limit;
 743
 744         if (td_rw(td))
 745                 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
 746         else if (td_write(td))
 747                 bytes = td->io_issue_bytes[DDIR_WRITE];
 748         else if (td_read(td))
 749                 bytes = td->io_issue_bytes[DDIR_READ];
 750         else
 751                 bytes = td->io_issue_bytes[DDIR_TRIM];
 752
 753         if (td->o.io_limit)
 754                 limit = td->o.io_limit;
 755         else
 756                 limit = td->o.size;
 757
 758         limit *= td->loops;
 759         return bytes >= limit || exceeds_number_ios(td);
 760 }
 761
 762 static bool io_complete_bytes_exceeded(struct thread_data *td)
 763 {
 764         unsigned long long bytes, limit;
 765
 766         if (td_rw(td))
 767                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 768         else if (td_write(td))
 769                 bytes = td->this_io_bytes[DDIR_WRITE];
 770         else if (td_read(td))
 771                 bytes = td->this_io_bytes[DDIR_READ];
 772         else
 773                 bytes = td->this_io_bytes[DDIR_TRIM];
 774
 775         if (td->o.io_limit)
 776                 limit = td->o.io_limit;
 777         else
 778                 limit = td->o.size;
 779
 780         limit *= td->loops;
 781         return bytes >= limit || exceeds_number_ios(td);
 782 }
 783
 784 /*
 785  * used to calculate the next io time for rate control
 786  *
 787  */
 788 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 789 {
 790         uint64_t secs, remainder, bps, bytes, iops;
 791
 792         assert(!(td->flags & TD_F_CHILD));
 793         bytes = td->rate_io_issue_bytes[ddir];
 794         bps = td->rate_bps[ddir];
 795
 796         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 797                 uint64_t val;
 798                 iops = bps / td->o.bs[ddir];
 799                 val = (int64_t) (1000000 / iops) *
 800                                 -logf(__rand_0_1(&td->poisson_state));
 801                 if (val) {
 802                         dprint(FD_RATE, "poisson rate iops=%llu\n",
 803                                         (unsigned long long) 1000000 / val);
 804                 }
 805                 td->last_usec += val;
 806                 return td->last_usec;
 807         } else if (bps) {
 808                 secs = bytes / bps;
 809                 remainder = bytes % bps;
 810                 return remainder * 1000000 / bps + secs * 1000000;
 811         }
 812
 813         return 0;
 814 }
 815
 816 /*
 817  * Main IO worker function. It retrieves io_u's to process and queues
 818  * and reaps them, checking for rate and errors along the way.
 819  *
 820  * Returns number of bytes written and trimmed.
 821  */
 822 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 823 {
 824         unsigned int i;
 825         int ret = 0;
 826         uint64_t total_bytes, bytes_issued = 0;
 827
 828         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 829                 bytes_done[i] = td->bytes_done[i];
 830
 831         if (in_ramp_time(td))
 832                 td_set_runstate(td, TD_RAMP);
 833         else
 834                 td_set_runstate(td, TD_RUNNING);
 835
 836         lat_target_init(td);
 837
 838         total_bytes = td->o.size;
 839         /*
 840         * Allow random overwrite workloads to write up to io_limit
 841         * before starting verification phase as 'size' doesn't apply.
 842         */
 843         if (td_write(td) && td_random(td) && td->o.norandommap)
 844                 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
 845         /*
 846          * If verify_backlog is enabled, we'll run the verify in this
 847          * handler as well. For that case, we may need up to twice the
 848          * amount of bytes.
 849          */
 850         if (td->o.verify != VERIFY_NONE &&
 851            (td_write(td) && td->o.verify_backlog))
 852                 total_bytes += td->o.size;
 853
 854         /* In trimwrite mode, each byte is trimmed and then written, so
 855          * allow total_bytes to be twice as big */
 856         if (td_trimwrite(td))
 857                 total_bytes += td->total_io_size;
 858
 859         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 860                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 861                 td->o.time_based) {
 862                 struct timeval comp_time;
 863                 struct io_u *io_u;
 864                 int full;
 865                 enum fio_ddir ddir;
 866
 867                 check_update_rusage(td);
 868
 869                 if (td->terminate || td->done)
 870                         break;
 871
 872                 update_tv_cache(td);
 873
 874                 if (runtime_exceeded(td, &td->tv_cache)) {
 875                         __update_tv_cache(td);
 876                         if (runtime_exceeded(td, &td->tv_cache)) {
 877                                 fio_mark_td_terminate(td);
 878                                 break;
 879                         }
 880                 }
 881
 882                 if (flow_threshold_exceeded(td))
 883                         continue;
 884
 885                 /*
 886                  * Break if we exceeded the bytes. The exception is time
 887                  * based runs, but we still need to break out of the loop
 888                  * for those to run verification, if enabled.
 889                  */
 890                 if (bytes_issued >= total_bytes &&
 891                     (!td->o.time_based ||
 892                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 893                         break;
 894
 895                 io_u = get_io_u(td);
 896                 if (IS_ERR_OR_NULL(io_u)) {
 897                         int err = PTR_ERR(io_u);
 898
 899                         io_u = NULL;
 900                         if (err == -EBUSY) {
 901                                 ret = FIO_Q_BUSY;
 902                                 goto reap;
 903                         }
 904                         if (td->o.latency_target)
 905                                 goto reap;
 906                         break;
 907                 }
 908
 909                 ddir = io_u->ddir;
 910
 911                 /*
 912                  * Add verification end_io handler if:
 913                  *      - Asked to verify (!td_rw(td))
 914                  *      - Or the io_u is from our verify list (mixed write/ver)
 915                  */
 916                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 917                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 918
 919                         if (!td->o.verify_pattern_bytes) {
 920                                 io_u->rand_seed = __rand(&td->verify_state);
 921                                 if (sizeof(int) != sizeof(long *))
 922                                         io_u->rand_seed *= __rand(&td->verify_state);
 923                         }
 924
 925                         if (verify_state_should_stop(td, io_u)) {
 926                                 put_io_u(td, io_u);
 927                                 break;
 928                         }
 929
 930                         if (td->o.verify_async)
 931                                 io_u->end_io = verify_io_u_async;
 932                         else
 933                                 io_u->end_io = verify_io_u;
 934                         td_set_runstate(td, TD_VERIFYING);
 935                 } else if (in_ramp_time(td))
 936                         td_set_runstate(td, TD_RAMP);
 937                 else
 938                         td_set_runstate(td, TD_RUNNING);
 939
 940                 /*
 941                  * Always log IO before it's issued, so we know the specific
 942                  * order of it. The logged unit will track when the IO has
 943                  * completed.
 944                  */
 945                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 946                     td->o.do_verify &&
 947                     td->o.verify != VERIFY_NONE &&
 948                     !td->o.experimental_verify)
 949                         log_io_piece(td, io_u);
 950
 951                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
 952                         const unsigned long blen = io_u->xfer_buflen;
 953                         const enum fio_ddir ddir = acct_ddir(io_u);
 954
 955                         if (td->error)
 956                                 break;
 957
 958                         workqueue_enqueue(&td->io_wq, &io_u->work);
 959                         ret = FIO_Q_QUEUED;
 960
 961                         if (ddir_rw(ddir)) {
 962                                 td->io_issues[ddir]++;
 963                                 td->io_issue_bytes[ddir] += blen;
 964                                 td->rate_io_issue_bytes[ddir] += blen;
 965                         }
 966
 967                         if (should_check_rate(td))
 968                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 969
 970                 } else {
 971                         ret = td_io_queue(td, io_u);
 972
 973                         if (should_check_rate(td))
 974                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 975
 976                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
 977                                 break;
 978
 979                         /*
 980                          * See if we need to complete some commands. Note that
 981                          * we can get BUSY even without IO queued, if the
 982                          * system is resource starved.
 983                          */
 984 reap:
 985                         full = queue_full(td) ||
 986                                 (ret == FIO_Q_BUSY && td->cur_depth);
 987                         if (full || io_in_polling(td))
 988                                 ret = wait_for_completions(td, &comp_time);
 989                 }
 990                 if (ret < 0)
 991                         break;
 992                 if (!ddir_rw_sum(td->bytes_done) &&
 993                     !(td->io_ops->flags & FIO_NOIO))
 994                         continue;
 995
 996                 if (!in_ramp_time(td) && should_check_rate(td)) {
 997                         if (check_min_rate(td, &comp_time)) {
 998                                 if (exitall_on_terminate || td->o.exitall_error)
 999                                         fio_terminate_threads(td->groupid);
1000                                 td_verror(td, EIO, "check_min_rate");
1001                                 break;
1002                         }
1003                 }
1004                 if (!in_ramp_time(td) && td->o.latency_target)
1005                         lat_target_check(td);
1006
1007                 if (td->o.thinktime) {
1008                         unsigned long long b;
1009
1010                         b = ddir_rw_sum(td->io_blocks);
1011                         if (!(b % td->o.thinktime_blocks)) {
1012                                 int left;
1013
1014                                 io_u_quiesce(td);
1015
1016                                 if (td->o.thinktime_spin)
1017                                         usec_spin(td->o.thinktime_spin);
1018
1019                                 left = td->o.thinktime - td->o.thinktime_spin;
1020                                 if (left)
1021                                         usec_sleep(td, left);
1022                         }
1023                 }
1024         }
1025
1026         check_update_rusage(td);
1027
1028         if (td->trim_entries)
1029                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1030
1031         if (td->o.fill_device && td->error == ENOSPC) {
1032                 td->error = 0;
1033                 fio_mark_td_terminate(td);
1034         }
1035         if (!td->error) {
1036                 struct fio_file *f;
1037
1038                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1039                         workqueue_flush(&td->io_wq);
1040                         i = 0;
1041                 } else
1042                         i = td->cur_depth;
1043
1044                 if (i) {
1045                         ret = io_u_queued_complete(td, i);
1046                         if (td->o.fill_device && td->error == ENOSPC)
1047                                 td->error = 0;
1048                 }
1049
1050                 if (should_fsync(td) && td->o.end_fsync) {
1051                         td_set_runstate(td, TD_FSYNCING);
1052
1053                         for_each_file(td, f, i) {
1054                                 if (!fio_file_fsync(td, f))
1055                                         continue;
1056
1057                                 log_err("fio: end_fsync failed for file %s\n",
1058                                                                 f->file_name);
1059                         }
1060                 }
1061         } else
1062                 cleanup_pending_aio(td);
1063
1064         /*
1065          * stop job if we failed doing any IO
1066          */
1067         if (!ddir_rw_sum(td->this_io_bytes))
1068                 td->done = 1;
1069
1070         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1071                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1072 }
1073
1074 static void free_file_completion_logging(struct thread_data *td)
1075 {
1076         struct fio_file *f;
1077         unsigned int i;
1078
1079         for_each_file(td, f, i) {
1080                 if (!f->last_write_comp)
1081                         break;
1082                 sfree(f->last_write_comp);
1083         }
1084 }
1085
1086 static int init_file_completion_logging(struct thread_data *td,
1087                                         unsigned int depth)
1088 {
1089         struct fio_file *f;
1090         unsigned int i;
1091
1092         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1093                 return 0;
1094
1095         for_each_file(td, f, i) {
1096                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1097                 if (!f->last_write_comp)
1098                         goto cleanup;
1099         }
1100
1101         return 0;
1102
1103 cleanup:
1104         free_file_completion_logging(td);
1105         log_err("fio: failed to alloc write comp data\n");
1106         return 1;
1107 }
1108
1109 static void cleanup_io_u(struct thread_data *td)
1110 {
1111         struct io_u *io_u;
1112
1113         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1114
1115                 if (td->io_ops->io_u_free)
1116                         td->io_ops->io_u_free(td, io_u);
1117
1118                 fio_memfree(io_u, sizeof(*io_u));
1119         }
1120
1121         free_io_mem(td);
1122
1123         io_u_rexit(&td->io_u_requeues);
1124         io_u_qexit(&td->io_u_freelist);
1125         io_u_qexit(&td->io_u_all);
1126
1127         free_file_completion_logging(td);
1128 }
1129
1130 static int init_io_u(struct thread_data *td)
1131 {
1132         struct io_u *io_u;
1133         unsigned int max_bs, min_write;
1134         int cl_align, i, max_units;
1135         int data_xfer = 1, err;
1136         char *p;
1137
1138         max_units = td->o.iodepth;
1139         max_bs = td_max_bs(td);
1140         min_write = td->o.min_bs[DDIR_WRITE];
1141         td->orig_buffer_size = (unsigned long long) max_bs
1142                                         * (unsigned long long) max_units;
1143
1144         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1145                 data_xfer = 0;
1146
1147         err = 0;
1148         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1149         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1150         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1151
1152         if (err) {
1153                 log_err("fio: failed setting up IO queues\n");
1154                 return 1;
1155         }
1156
1157         /*
1158          * if we may later need to do address alignment, then add any
1159          * possible adjustment here so that we don't cause a buffer
1160          * overflow later. this adjustment may be too much if we get
1161          * lucky and the allocator gives us an aligned address.
1162          */
1163         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1164             (td->io_ops->flags & FIO_RAWIO))
1165                 td->orig_buffer_size += page_mask + td->o.mem_align;
1166
1167         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1168                 unsigned long bs;
1169
1170                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1171                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1172         }
1173
1174         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1175                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1176                 return 1;
1177         }
1178
1179         if (data_xfer && allocate_io_mem(td))
1180                 return 1;
1181
1182         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1183             (td->io_ops->flags & FIO_RAWIO))
1184                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1185         else
1186                 p = td->orig_buffer;
1187
1188         cl_align = os_cache_line_size();
1189
1190         for (i = 0; i < max_units; i++) {
1191                 void *ptr;
1192
1193                 if (td->terminate)
1194                         return 1;
1195
1196                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1197                 if (!ptr) {
1198                         log_err("fio: unable to allocate aligned memory\n");
1199                         break;
1200                 }
1201
1202                 io_u = ptr;
1203                 memset(io_u, 0, sizeof(*io_u));
1204                 INIT_FLIST_HEAD(&io_u->verify_list);
1205                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1206
1207                 if (data_xfer) {
1208                         io_u->buf = p;
1209                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1210
1211                         if (td_write(td))
1212                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1213                         if (td_write(td) && td->o.verify_pattern_bytes) {
1214                                 /*
1215                                  * Fill the buffer with the pattern if we are
1216                                  * going to be doing writes.
1217                                  */
1218                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1219                         }
1220                 }
1221
1222                 io_u->index = i;
1223                 io_u->flags = IO_U_F_FREE;
1224                 io_u_qpush(&td->io_u_freelist, io_u);
1225
1226                 /*
1227                  * io_u never leaves this stack, used for iteration of all
1228                  * io_u buffers.
1229                  */
1230                 io_u_qpush(&td->io_u_all, io_u);
1231
1232                 if (td->io_ops->io_u_init) {
1233                         int ret = td->io_ops->io_u_init(td, io_u);
1234
1235                         if (ret) {
1236                                 log_err("fio: failed to init engine data: %d\n", ret);
1237                                 return 1;
1238                         }
1239                 }
1240
1241                 p += max_bs;
1242         }
1243
1244         if (init_file_completion_logging(td, max_units))
1245                 return 1;
1246
1247         return 0;
1248 }
1249
1250 static int switch_ioscheduler(struct thread_data *td)
1251 {
1252         char tmp[256], tmp2[128];
1253         FILE *f;
1254         int ret;
1255
1256         if (td->io_ops->flags & FIO_DISKLESSIO)
1257                 return 0;
1258
1259         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1260
1261         f = fopen(tmp, "r+");
1262         if (!f) {
1263                 if (errno == ENOENT) {
1264                         log_err("fio: os or kernel doesn't support IO scheduler"
1265                                 " switching\n");
1266                         return 0;
1267                 }
1268                 td_verror(td, errno, "fopen iosched");
1269                 return 1;
1270         }
1271
1272         /*
1273          * Set io scheduler.
1274          */
1275         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1276         if (ferror(f) || ret != 1) {
1277                 td_verror(td, errno, "fwrite");
1278                 fclose(f);
1279                 return 1;
1280         }
1281
1282         rewind(f);
1283
1284         /*
1285          * Read back and check that the selected scheduler is now the default.
1286          */
1287         memset(tmp, 0, sizeof(tmp));
1288         ret = fread(tmp, sizeof(tmp), 1, f);
1289         if (ferror(f) || ret < 0) {
1290                 td_verror(td, errno, "fread");
1291                 fclose(f);
1292                 return 1;
1293         }
1294         /*
1295          * either a list of io schedulers or "none\n" is expected.
1296          */
1297         tmp[strlen(tmp) - 1] = '\0';
1298
1299
1300         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1301         if (!strstr(tmp, tmp2)) {
1302                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1303                 td_verror(td, EINVAL, "iosched_switch");
1304                 fclose(f);
1305                 return 1;
1306         }
1307
1308         fclose(f);
1309         return 0;
1310 }
1311
1312 static bool keep_running(struct thread_data *td)
1313 {
1314         unsigned long long limit;
1315
1316         if (td->done)
1317                 return false;
1318         if (td->o.time_based)
1319                 return true;
1320         if (td->o.loops) {
1321                 td->o.loops--;
1322                 return true;
1323         }
1324         if (exceeds_number_ios(td))
1325                 return false;
1326
1327         if (td->o.io_limit)
1328                 limit = td->o.io_limit;
1329         else
1330                 limit = td->o.size;
1331
1332         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1333                 uint64_t diff;
1334
1335                 /*
1336                  * If the difference is less than the minimum IO size, we
1337                  * are done.
1338                  */
1339                 diff = limit - ddir_rw_sum(td->io_bytes);
1340                 if (diff < td_max_bs(td))
1341                         return false;
1342
1343                 if (fio_files_done(td) && !td->o.io_limit)
1344                         return false;
1345
1346                 return true;
1347         }
1348
1349         return false;
1350 }
1351
1352 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1353 {
1354         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1355         int ret;
1356         char *str;
1357
1358         str = malloc(newlen);
1359         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1360
1361         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1362         ret = system(str);
1363         if (ret == -1)
1364                 log_err("fio: exec of cmd <%s> failed\n", str);
1365
1366         free(str);
1367         return ret;
1368 }
1369
1370 /*
1371  * Dry run to compute correct state of numberio for verification.
1372  */
1373 static uint64_t do_dry_run(struct thread_data *td)
1374 {
1375         td_set_runstate(td, TD_RUNNING);
1376
1377         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1378                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1379                 struct io_u *io_u;
1380                 int ret;
1381
1382                 if (td->terminate || td->done)
1383                         break;
1384
1385                 io_u = get_io_u(td);
1386                 if (!io_u)
1387                         break;
1388
1389                 io_u_set(io_u, IO_U_F_FLIGHT);
1390                 io_u->error = 0;
1391                 io_u->resid = 0;
1392                 if (ddir_rw(acct_ddir(io_u)))
1393                         td->io_issues[acct_ddir(io_u)]++;
1394                 if (ddir_rw(io_u->ddir)) {
1395                         io_u_mark_depth(td, 1);
1396                         td->ts.total_io_u[io_u->ddir]++;
1397                 }
1398
1399                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1400                     td->o.do_verify &&
1401                     td->o.verify != VERIFY_NONE &&
1402                     !td->o.experimental_verify)
1403                         log_io_piece(td, io_u);
1404
1405                 ret = io_u_sync_complete(td, io_u);
1406                 (void) ret;
1407         }
1408
1409         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1410 }
1411
1412 struct fork_data {
1413         struct thread_data *td;
1414         struct sk_out *sk_out;
1415 };
1416
1417 /*
1418  * Entry point for the thread based jobs. The process based jobs end up
1419  * here as well, after a little setup.
1420  */
1421 static void *thread_main(void *data)
1422 {
1423         struct fork_data *fd = data;
1424         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1425         struct thread_data *td = fd->td;
1426         struct thread_options *o = &td->o;
1427         struct sk_out *sk_out = fd->sk_out;
1428         pthread_condattr_t attr;
1429         int clear_state;
1430         int ret;
1431
1432         sk_out_assign(sk_out);
1433         free(fd);
1434
1435         if (!o->use_thread) {
1436                 setsid();
1437                 td->pid = getpid();
1438         } else
1439                 td->pid = gettid();
1440
1441         fio_local_clock_init(o->use_thread);
1442
1443         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1444
1445         if (is_backend)
1446                 fio_server_send_start(td);
1447
1448         INIT_FLIST_HEAD(&td->io_log_list);
1449         INIT_FLIST_HEAD(&td->io_hist_list);
1450         INIT_FLIST_HEAD(&td->verify_list);
1451         INIT_FLIST_HEAD(&td->trim_list);
1452         INIT_FLIST_HEAD(&td->next_rand_list);
1453         pthread_mutex_init(&td->io_u_lock, NULL);
1454         td->io_hist_tree = RB_ROOT;
1455
1456         pthread_condattr_init(&attr);
1457         pthread_cond_init(&td->verify_cond, &attr);
1458         pthread_cond_init(&td->free_cond, &attr);
1459
1460         td_set_runstate(td, TD_INITIALIZED);
1461         dprint(FD_MUTEX, "up startup_mutex\n");
1462         fio_mutex_up(startup_mutex);
1463         dprint(FD_MUTEX, "wait on td->mutex\n");
1464         fio_mutex_down(td->mutex);
1465         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1466
1467         /*
1468          * A new gid requires privilege, so we need to do this before setting
1469          * the uid.
1470          */
1471         if (o->gid != -1U && setgid(o->gid)) {
1472                 td_verror(td, errno, "setgid");
1473                 goto err;
1474         }
1475         if (o->uid != -1U && setuid(o->uid)) {
1476                 td_verror(td, errno, "setuid");
1477                 goto err;
1478         }
1479
1480         /*
1481          * Do this early, we don't want the compress threads to be limited
1482          * to the same CPUs as the IO workers. So do this before we set
1483          * any potential CPU affinity
1484          */
1485         if (iolog_compress_init(td, sk_out))
1486                 goto err;
1487
1488         /*
1489          * If we have a gettimeofday() thread, make sure we exclude that
1490          * thread from this job
1491          */
1492         if (o->gtod_cpu)
1493                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1494
1495         /*
1496          * Set affinity first, in case it has an impact on the memory
1497          * allocations.
1498          */
1499         if (fio_option_is_set(o, cpumask)) {
1500                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1501                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1502                         if (!ret) {
1503                                 log_err("fio: no CPUs set\n");
1504                                 log_err("fio: Try increasing number of available CPUs\n");
1505                                 td_verror(td, EINVAL, "cpus_split");
1506                                 goto err;
1507                         }
1508                 }
1509                 ret = fio_setaffinity(td->pid, o->cpumask);
1510                 if (ret == -1) {
1511                         td_verror(td, errno, "cpu_set_affinity");
1512                         goto err;
1513                 }
1514         }
1515
1516 #ifdef CONFIG_LIBNUMA
1517         /* numa node setup */
1518         if (fio_option_is_set(o, numa_cpunodes) ||
1519             fio_option_is_set(o, numa_memnodes)) {
1520                 struct bitmask *mask;
1521
1522                 if (numa_available() < 0) {
1523                         td_verror(td, errno, "Does not support NUMA API\n");
1524                         goto err;
1525                 }
1526
1527                 if (fio_option_is_set(o, numa_cpunodes)) {
1528                         mask = numa_parse_nodestring(o->numa_cpunodes);
1529                         ret = numa_run_on_node_mask(mask);
1530                         numa_free_nodemask(mask);
1531                         if (ret == -1) {
1532                                 td_verror(td, errno, \
1533                                         "numa_run_on_node_mask failed\n");
1534                                 goto err;
1535                         }
1536                 }
1537
1538                 if (fio_option_is_set(o, numa_memnodes)) {
1539                         mask = NULL;
1540                         if (o->numa_memnodes)
1541                                 mask = numa_parse_nodestring(o->numa_memnodes);
1542
1543                         switch (o->numa_mem_mode) {
1544                         case MPOL_INTERLEAVE:
1545                                 numa_set_interleave_mask(mask);
1546                                 break;
1547                         case MPOL_BIND:
1548                                 numa_set_membind(mask);
1549                                 break;
1550                         case MPOL_LOCAL:
1551                                 numa_set_localalloc();
1552                                 break;
1553                         case MPOL_PREFERRED:
1554                                 numa_set_preferred(o->numa_mem_prefer_node);
1555                                 break;
1556                         case MPOL_DEFAULT:
1557                         default:
1558                                 break;
1559                         }
1560
1561                         if (mask)
1562                                 numa_free_nodemask(mask);
1563
1564                 }
1565         }
1566 #endif
1567
1568         if (fio_pin_memory(td))
1569                 goto err;
1570
1571         /*
1572          * May alter parameters that init_io_u() will use, so we need to
1573          * do this first.
1574          */
1575         if (init_iolog(td))
1576                 goto err;
1577
1578         if (init_io_u(td))
1579                 goto err;
1580
1581         if (o->verify_async && verify_async_init(td))
1582                 goto err;
1583
1584         if (fio_option_is_set(o, ioprio) ||
1585             fio_option_is_set(o, ioprio_class)) {
1586                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1587                 if (ret == -1) {
1588                         td_verror(td, errno, "ioprio_set");
1589                         goto err;
1590                 }
1591         }
1592
1593         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1594                 goto err;
1595
1596         errno = 0;
1597         if (nice(o->nice) == -1 && errno != 0) {
1598                 td_verror(td, errno, "nice");
1599                 goto err;
1600         }
1601
1602         if (o->ioscheduler && switch_ioscheduler(td))
1603                 goto err;
1604
1605         if (!o->create_serialize && setup_files(td))
1606                 goto err;
1607
1608         if (td_io_init(td))
1609                 goto err;
1610
1611         if (init_random_map(td))
1612                 goto err;
1613
1614         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1615                 goto err;
1616
1617         if (o->pre_read) {
1618                 if (pre_read_files(td) < 0)
1619                         goto err;
1620         }
1621
1622         fio_verify_init(td);
1623
1624         if (rate_submit_init(td, sk_out))
1625                 goto err;
1626
1627         fio_gettime(&td->epoch, NULL);
1628         fio_getrusage(&td->ru_start);
1629         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1630         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1631
1632         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1633                         o->ratemin[DDIR_TRIM]) {
1634                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1635                                         sizeof(td->bw_sample_time));
1636                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1637                                         sizeof(td->bw_sample_time));
1638                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1639                                         sizeof(td->bw_sample_time));
1640         }
1641
1642         clear_state = 0;
1643         while (keep_running(td)) {
1644                 uint64_t verify_bytes;
1645
1646                 fio_gettime(&td->start, NULL);
1647                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1648
1649                 if (clear_state)
1650                         clear_io_state(td, 0);
1651
1652                 prune_io_piece_log(td);
1653
1654                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1655                         verify_bytes = do_dry_run(td);
1656                 else {
1657                         uint64_t bytes_done[DDIR_RWDIR_CNT];
1658
1659                         do_io(td, bytes_done);
1660
1661                         if (!ddir_rw_sum(bytes_done)) {
1662                                 fio_mark_td_terminate(td);
1663                                 verify_bytes = 0;
1664                         } else {
1665                                 verify_bytes = bytes_done[DDIR_WRITE] +
1666                                                 bytes_done[DDIR_TRIM];
1667                         }
1668                 }
1669
1670                 clear_state = 1;
1671
1672                 /*
1673                  * Make sure we've successfully updated the rusage stats
1674                  * before waiting on the stat mutex. Otherwise we could have
1675                  * the stat thread holding stat mutex and waiting for
1676                  * the rusage_sem, which would never get upped because
1677                  * this thread is waiting for the stat mutex.
1678                  */
1679                 check_update_rusage(td);
1680
1681                 fio_mutex_down(stat_mutex);
1682                 if (td_read(td) && td->io_bytes[DDIR_READ])
1683                         update_runtime(td, elapsed_us, DDIR_READ);
1684                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1685                         update_runtime(td, elapsed_us, DDIR_WRITE);
1686                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1687                         update_runtime(td, elapsed_us, DDIR_TRIM);
1688                 fio_gettime(&td->start, NULL);
1689                 fio_mutex_up(stat_mutex);
1690
1691                 if (td->error || td->terminate)
1692                         break;
1693
1694                 if (!o->do_verify ||
1695                     o->verify == VERIFY_NONE ||
1696                     (td->io_ops->flags & FIO_UNIDIR))
1697                         continue;
1698
1699                 clear_io_state(td, 0);
1700
1701                 fio_gettime(&td->start, NULL);
1702
1703                 do_verify(td, verify_bytes);
1704
1705                 /*
1706                  * See comment further up for why this is done here.
1707                  */
1708                 check_update_rusage(td);
1709
1710                 fio_mutex_down(stat_mutex);
1711                 update_runtime(td, elapsed_us, DDIR_READ);
1712                 fio_gettime(&td->start, NULL);
1713                 fio_mutex_up(stat_mutex);
1714
1715                 if (td->error || td->terminate)
1716                         break;
1717         }
1718
1719         td_set_runstate(td, TD_FINISHING);
1720
1721         update_rusage_stat(td);
1722         td->ts.total_run_time = mtime_since_now(&td->epoch);
1723         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1724         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1725         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1726
1727         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1728             (td->o.verify != VERIFY_NONE && td_write(td)))
1729                 verify_save_state(td->thread_number);
1730
1731         fio_unpin_memory(td);
1732
1733         td_writeout_logs(td, true);
1734
1735         iolog_compress_exit(td);
1736         rate_submit_exit(td);
1737
1738         if (o->exec_postrun)
1739                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1740
1741         if (exitall_on_terminate || (o->exitall_error && td->error))
1742                 fio_terminate_threads(td->groupid);
1743
1744 err:
1745         if (td->error)
1746                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1747                                                         td->verror);
1748
1749         if (o->verify_async)
1750                 verify_async_exit(td);
1751
1752         close_and_free_files(td);
1753         cleanup_io_u(td);
1754         close_ioengine(td);
1755         cgroup_shutdown(td, &cgroup_mnt);
1756         verify_free_state(td);
1757
1758         if (td->zone_state_index) {
1759                 int i;
1760
1761                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1762                         free(td->zone_state_index[i]);
1763                 free(td->zone_state_index);
1764                 td->zone_state_index = NULL;
1765         }
1766
1767         if (fio_option_is_set(o, cpumask)) {
1768                 ret = fio_cpuset_exit(&o->cpumask);
1769                 if (ret)
1770                         td_verror(td, ret, "fio_cpuset_exit");
1771         }
1772
1773         /*
1774          * do this very late, it will log file closing as well
1775          */
1776         if (o->write_iolog_file)
1777                 write_iolog_close(td);
1778
1779         fio_mutex_remove(td->mutex);
1780         td->mutex = NULL;
1781
1782         td_set_runstate(td, TD_EXITED);
1783
1784         /*
1785          * Do this last after setting our runstate to exited, so we
1786          * know that the stat thread is signaled.
1787          */
1788         check_update_rusage(td);
1789
1790         sk_out_drop();
1791         return (void *) (uintptr_t) td->error;
1792 }
1793
1794
1795 /*
1796  * We cannot pass the td data into a forked process, so attach the td and
1797  * pass it to the thread worker.
1798  */
1799 static int fork_main(struct sk_out *sk_out, int shmid, int offset)
1800 {
1801         struct fork_data *fd;
1802         void *data, *ret;
1803
1804 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1805         data = shmat(shmid, NULL, 0);
1806         if (data == (void *) -1) {
1807                 int __err = errno;
1808
1809                 perror("shmat");
1810                 return __err;
1811         }
1812 #else
1813         /*
1814          * HP-UX inherits shm mappings?
1815          */
1816         data = threads;
1817 #endif
1818
1819         fd = calloc(1, sizeof(*fd));
1820         fd->td = data + offset * sizeof(struct thread_data);
1821         fd->sk_out = sk_out;
1822         ret = thread_main(fd);
1823         shmdt(data);
1824         return (int) (uintptr_t) ret;
1825 }
1826
1827 static void dump_td_info(struct thread_data *td)
1828 {
1829         log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
1830                 "appears to be stuck. Doing forceful exit of this job.\n",
1831                         td->o.name, td->runstate,
1832                         (unsigned long) time_since_now(&td->terminate_time));
1833 }
1834
1835 /*
1836  * Run over the job map and reap the threads that have exited, if any.
1837  */
1838 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1839                          unsigned int *m_rate)
1840 {
1841         struct thread_data *td;
1842         unsigned int cputhreads, realthreads, pending;
1843         int i, status, ret;
1844
1845         /*
1846          * reap exited threads (TD_EXITED -> TD_REAPED)
1847          */
1848         realthreads = pending = cputhreads = 0;
1849         for_each_td(td, i) {
1850                 int flags = 0;
1851
1852                 /*
1853                  * ->io_ops is NULL for a thread that has closed its
1854                  * io engine
1855                  */
1856                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1857                         cputhreads++;
1858                 else
1859                         realthreads++;
1860
1861                 if (!td->pid) {
1862                         pending++;
1863                         continue;
1864                 }
1865                 if (td->runstate == TD_REAPED)
1866                         continue;
1867                 if (td->o.use_thread) {
1868                         if (td->runstate == TD_EXITED) {
1869                                 td_set_runstate(td, TD_REAPED);
1870                                 goto reaped;
1871                         }
1872                         continue;
1873                 }
1874
1875                 flags = WNOHANG;
1876                 if (td->runstate == TD_EXITED)
1877                         flags = 0;
1878
1879                 /*
1880                  * check if someone quit or got killed in an unusual way
1881                  */
1882                 ret = waitpid(td->pid, &status, flags);
1883                 if (ret < 0) {
1884                         if (errno == ECHILD) {
1885                                 log_err("fio: pid=%d disappeared %d\n",
1886                                                 (int) td->pid, td->runstate);
1887                                 td->sig = ECHILD;
1888                                 td_set_runstate(td, TD_REAPED);
1889                                 goto reaped;
1890                         }
1891                         perror("waitpid");
1892                 } else if (ret == td->pid) {
1893                         if (WIFSIGNALED(status)) {
1894                                 int sig = WTERMSIG(status);
1895
1896                                 if (sig != SIGTERM && sig != SIGUSR2)
1897                                         log_err("fio: pid=%d, got signal=%d\n",
1898                                                         (int) td->pid, sig);
1899                                 td->sig = sig;
1900                                 td_set_runstate(td, TD_REAPED);
1901                                 goto reaped;
1902                         }
1903                         if (WIFEXITED(status)) {
1904                                 if (WEXITSTATUS(status) && !td->error)
1905                                         td->error = WEXITSTATUS(status);
1906
1907                                 td_set_runstate(td, TD_REAPED);
1908                                 goto reaped;
1909                         }
1910                 }
1911
1912                 /*
1913                  * If the job is stuck, do a forceful timeout of it and
1914                  * move on.
1915                  */
1916                 if (td->terminate &&
1917                     td->runstate < TD_FSYNCING &&
1918                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1919                         dump_td_info(td);
1920                         td_set_runstate(td, TD_REAPED);
1921                         goto reaped;
1922                 }
1923
1924                 /*
1925                  * thread is not dead, continue
1926                  */
1927                 pending++;
1928                 continue;
1929 reaped:
1930                 (*nr_running)--;
1931                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1932                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1933                 if (!td->pid)
1934                         pending--;
1935
1936                 if (td->error)
1937                         exit_value++;
1938
1939                 done_secs += mtime_since_now(&td->epoch) / 1000;
1940                 profile_td_exit(td);
1941         }
1942
1943         if (*nr_running == cputhreads && !pending && realthreads)
1944                 fio_terminate_threads(TERMINATE_ALL);
1945 }
1946
1947 static bool __check_trigger_file(void)
1948 {
1949         struct stat sb;
1950
1951         if (!trigger_file)
1952                 return false;
1953
1954         if (stat(trigger_file, &sb))
1955                 return false;
1956
1957         if (unlink(trigger_file) < 0)
1958                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1959                                                         strerror(errno));
1960
1961         return true;
1962 }
1963
1964 static bool trigger_timedout(void)
1965 {
1966         if (trigger_timeout)
1967                 return time_since_genesis() >= trigger_timeout;
1968
1969         return false;
1970 }
1971
1972 void exec_trigger(const char *cmd)
1973 {
1974         int ret;
1975
1976         if (!cmd)
1977                 return;
1978
1979         ret = system(cmd);
1980         if (ret == -1)
1981                 log_err("fio: failed executing %s trigger\n", cmd);
1982 }
1983
1984 void check_trigger_file(void)
1985 {
1986         if (__check_trigger_file() || trigger_timedout()) {
1987                 if (nr_clients)
1988                         fio_clients_send_trigger(trigger_remote_cmd);
1989                 else {
1990                         verify_save_state(IO_LIST_ALL);
1991                         fio_terminate_threads(TERMINATE_ALL);
1992                         exec_trigger(trigger_cmd);
1993                 }
1994         }
1995 }
1996
1997 static int fio_verify_load_state(struct thread_data *td)
1998 {
1999         int ret;
2000
2001         if (!td->o.verify_state)
2002                 return 0;
2003
2004         if (is_backend) {
2005                 void *data;
2006
2007                 ret = fio_server_get_verify_state(td->o.name,
2008                                         td->thread_number - 1, &data);
2009                 if (!ret)
2010                         verify_assign_state(td, data);
2011         } else
2012                 ret = verify_load_state(td, "local");
2013
2014         return ret;
2015 }
2016
2017 static void do_usleep(unsigned int usecs)
2018 {
2019         check_for_running_stats();
2020         check_trigger_file();
2021         usleep(usecs);
2022 }
2023
2024 static bool check_mount_writes(struct thread_data *td)
2025 {
2026         struct fio_file *f;
2027         unsigned int i;
2028
2029         if (!td_write(td) || td->o.allow_mounted_write)
2030                 return false;
2031
2032         for_each_file(td, f, i) {
2033                 if (f->filetype != FIO_TYPE_BD)
2034                         continue;
2035                 if (device_is_mounted(f->file_name))
2036                         goto mounted;
2037         }
2038
2039         return false;
2040 mounted:
2041         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2042         return true;
2043 }
2044
2045 static bool waitee_running(struct thread_data *me)
2046 {
2047         const char *waitee = me->o.wait_for;
2048         const char *self = me->o.name;
2049         struct thread_data *td;
2050         int i;
2051
2052         if (!waitee)
2053                 return false;
2054
2055         for_each_td(td, i) {
2056                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2057                         continue;
2058
2059                 if (td->runstate < TD_EXITED) {
2060                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2061                                         self, td->o.name,
2062                                         runstate_to_name(td->runstate));
2063                         return true;
2064                 }
2065         }
2066
2067         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2068         return false;
2069 }
2070
2071 /*
2072  * Main function for kicking off and reaping jobs, as needed.
2073  */
2074 static void run_threads(struct sk_out *sk_out)
2075 {
2076         struct thread_data *td;
2077         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2078         uint64_t spent;
2079
2080         if (fio_gtod_offload && fio_start_gtod_thread())
2081                 return;
2082
2083         fio_idle_prof_init();
2084
2085         set_sig_handlers();
2086
2087         nr_thread = nr_process = 0;
2088         for_each_td(td, i) {
2089                 if (check_mount_writes(td))
2090                         return;
2091                 if (td->o.use_thread)
2092                         nr_thread++;
2093                 else
2094                         nr_process++;
2095         }
2096
2097         if (output_format & FIO_OUTPUT_NORMAL) {
2098                 log_info("Starting ");
2099                 if (nr_thread)
2100                         log_info("%d thread%s", nr_thread,
2101                                                 nr_thread > 1 ? "s" : "");
2102                 if (nr_process) {
2103                         if (nr_thread)
2104                                 log_info(" and ");
2105                         log_info("%d process%s", nr_process,
2106                                                 nr_process > 1 ? "es" : "");
2107                 }
2108                 log_info("\n");
2109                 log_info_flush();
2110         }
2111
2112         todo = thread_number;
2113         nr_running = 0;
2114         nr_started = 0;
2115         m_rate = t_rate = 0;
2116
2117         for_each_td(td, i) {
2118                 print_status_init(td->thread_number - 1);
2119
2120                 if (!td->o.create_serialize)
2121                         continue;
2122
2123                 if (fio_verify_load_state(td))
2124                         goto reap;
2125
2126                 /*
2127                  * do file setup here so it happens sequentially,
2128                  * we don't want X number of threads getting their
2129                  * client data interspersed on disk
2130                  */
2131                 if (setup_files(td)) {
2132 reap:
2133                         exit_value++;
2134                         if (td->error)
2135                                 log_err("fio: pid=%d, err=%d/%s\n",
2136                                         (int) td->pid, td->error, td->verror);
2137                         td_set_runstate(td, TD_REAPED);
2138                         todo--;
2139                 } else {
2140                         struct fio_file *f;
2141                         unsigned int j;
2142
2143                         /*
2144                          * for sharing to work, each job must always open
2145                          * its own files. so close them, if we opened them
2146                          * for creation
2147                          */
2148                         for_each_file(td, f, j) {
2149                                 if (fio_file_open(f))
2150                                         td_io_close_file(td, f);
2151                         }
2152                 }
2153         }
2154
2155         /* start idle threads before io threads start to run */
2156         fio_idle_prof_start();
2157
2158         set_genesis_time();
2159
2160         while (todo) {
2161                 struct thread_data *map[REAL_MAX_JOBS];
2162                 struct timeval this_start;
2163                 int this_jobs = 0, left;
2164
2165                 /*
2166                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2167                  */
2168                 for_each_td(td, i) {
2169                         if (td->runstate != TD_NOT_CREATED)
2170                                 continue;
2171
2172                         /*
2173                          * never got a chance to start, killed by other
2174                          * thread for some reason
2175                          */
2176                         if (td->terminate) {
2177                                 todo--;
2178                                 continue;
2179                         }
2180
2181                         if (td->o.start_delay) {
2182                                 spent = utime_since_genesis();
2183
2184                                 if (td->o.start_delay > spent)
2185                                         continue;
2186                         }
2187
2188                         if (td->o.stonewall && (nr_started || nr_running)) {
2189                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2190                                                         td->o.name);
2191                                 break;
2192                         }
2193
2194                         if (waitee_running(td)) {
2195                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2196                                                 td->o.name, td->o.wait_for);
2197                                 continue;
2198                         }
2199
2200                         init_disk_util(td);
2201
2202                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2203                         td->update_rusage = 0;
2204
2205                         /*
2206                          * Set state to created. Thread will transition
2207                          * to TD_INITIALIZED when it's done setting up.
2208                          */
2209                         td_set_runstate(td, TD_CREATED);
2210                         map[this_jobs++] = td;
2211                         nr_started++;
2212
2213                         if (td->o.use_thread) {
2214                                 struct fork_data *fd;
2215                                 int ret;
2216
2217                                 fd = calloc(1, sizeof(*fd));
2218                                 fd->td = td;
2219                                 fd->sk_out = sk_out;
2220
2221                                 dprint(FD_PROCESS, "will pthread_create\n");
2222                                 ret = pthread_create(&td->thread, NULL,
2223                                                         thread_main, fd);
2224                                 if (ret) {
2225                                         log_err("pthread_create: %s\n",
2226                                                         strerror(ret));
2227                                         free(fd);
2228                                         nr_started--;
2229                                         break;
2230                                 }
2231                                 ret = pthread_detach(td->thread);
2232                                 if (ret)
2233                                         log_err("pthread_detach: %s",
2234                                                         strerror(ret));
2235                         } else {
2236                                 pid_t pid;
2237                                 dprint(FD_PROCESS, "will fork\n");
2238                                 pid = fork();
2239                                 if (!pid) {
2240                                         int ret = fork_main(sk_out, shm_id, i);
2241
2242                                         _exit(ret);
2243                                 } else if (i == fio_debug_jobno)
2244                                         *fio_debug_jobp = pid;
2245                         }
2246                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2247                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2248                                 log_err("fio: job startup hung? exiting.\n");
2249                                 fio_terminate_threads(TERMINATE_ALL);
2250                                 fio_abort = 1;
2251                                 nr_started--;
2252                                 break;
2253                         }
2254                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2255                 }
2256
2257                 /*
2258                  * Wait for the started threads to transition to
2259                  * TD_INITIALIZED.
2260                  */
2261                 fio_gettime(&this_start, NULL);
2262                 left = this_jobs;
2263                 while (left && !fio_abort) {
2264                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2265                                 break;
2266
2267                         do_usleep(100000);
2268
2269                         for (i = 0; i < this_jobs; i++) {
2270                                 td = map[i];
2271                                 if (!td)
2272                                         continue;
2273                                 if (td->runstate == TD_INITIALIZED) {
2274                                         map[i] = NULL;
2275                                         left--;
2276                                 } else if (td->runstate >= TD_EXITED) {
2277                                         map[i] = NULL;
2278                                         left--;
2279                                         todo--;
2280                                         nr_running++; /* work-around... */
2281                                 }
2282                         }
2283                 }
2284
2285                 if (left) {
2286                         log_err("fio: %d job%s failed to start\n", left,
2287                                         left > 1 ? "s" : "");
2288                         for (i = 0; i < this_jobs; i++) {
2289                                 td = map[i];
2290                                 if (!td)
2291                                         continue;
2292                                 kill(td->pid, SIGTERM);
2293                         }
2294                         break;
2295                 }
2296
2297                 /*
2298                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2299                  */
2300                 for_each_td(td, i) {
2301                         if (td->runstate != TD_INITIALIZED)
2302                                 continue;
2303
2304                         if (in_ramp_time(td))
2305                                 td_set_runstate(td, TD_RAMP);
2306                         else
2307                                 td_set_runstate(td, TD_RUNNING);
2308                         nr_running++;
2309                         nr_started--;
2310                         m_rate += ddir_rw_sum(td->o.ratemin);
2311                         t_rate += ddir_rw_sum(td->o.rate);
2312                         todo--;
2313                         fio_mutex_up(td->mutex);
2314                 }
2315
2316                 reap_threads(&nr_running, &t_rate, &m_rate);
2317
2318                 if (todo)
2319                         do_usleep(100000);
2320         }
2321
2322         while (nr_running) {
2323                 reap_threads(&nr_running, &t_rate, &m_rate);
2324                 do_usleep(10000);
2325         }
2326
2327         fio_idle_prof_stop();
2328
2329         update_io_ticks();
2330 }
2331
2332 static void free_disk_util(void)
2333 {
2334         disk_util_prune_entries();
2335         helper_thread_destroy();
2336 }
2337
2338 int fio_backend(struct sk_out *sk_out)
2339 {
2340         struct thread_data *td;
2341         int i;
2342
2343         if (exec_profile) {
2344                 if (load_profile(exec_profile))
2345                         return 1;
2346                 free(exec_profile);
2347                 exec_profile = NULL;
2348         }
2349         if (!thread_number)
2350                 return 0;
2351
2352         if (write_bw_log) {
2353                 struct log_params p = {
2354                         .log_type = IO_LOG_TYPE_BW,
2355                 };
2356
2357                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2358                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2359                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2360         }
2361
2362         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2363         if (startup_mutex == NULL)
2364                 return 1;
2365
2366         set_genesis_time();
2367         stat_init();
2368         helper_thread_create(startup_mutex, sk_out);
2369
2370         cgroup_list = smalloc(sizeof(*cgroup_list));
2371         INIT_FLIST_HEAD(cgroup_list);
2372
2373         run_threads(sk_out);
2374
2375         helper_thread_exit();
2376
2377         if (!fio_abort) {
2378                 __show_run_stats();
2379                 if (write_bw_log) {
2380                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2381                                 struct io_log *log = agg_io_log[i];
2382
2383                                 flush_log(log, 0);
2384                                 free_log(log);
2385                         }
2386                 }
2387         }
2388
2389         for_each_td(td, i) {
2390                 fio_options_free(td);
2391                 if (td->rusage_sem) {
2392                         fio_mutex_remove(td->rusage_sem);
2393                         td->rusage_sem = NULL;
2394                 }
2395         }
2396
2397         free_disk_util();
2398         cgroup_kill(cgroup_list);
2399         sfree(cgroup_list);
2400         sfree(cgroup_mnt);
2401
2402         fio_mutex_remove(startup_mutex);
2403         stat_exit();
2404         return exit_value;
2405 }