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
  39 #include "fio.h"
  40 #ifndef FIO_NO_HAVE_SHM_H
  41 #include <sys/shm.h>
  42 #endif
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53 #include "lib/getrusage.h"
  54 #include "idletime.h"
  55
  56 static pthread_t disk_util_thread;
  57 static struct fio_mutex *disk_thread_mutex;
  58 static struct fio_mutex *startup_mutex;
  59 static struct fio_mutex *writeout_mutex;
  60 static struct flist_head *cgroup_list;
  61 static char *cgroup_mnt;
  62 static int exit_value;
  63 static volatile int fio_abort;
  64 static unsigned int nr_process = 0;
  65 static unsigned int nr_thread = 0;
  66
  67 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  68
  69 int groupid = 0;
  70 unsigned int thread_number = 0;
  71 unsigned int stat_number = 0;
  72 int shm_id = 0;
  73 int temp_stall_ts;
  74 unsigned long done_secs = 0;
  75 volatile int disk_util_exit = 0;
  76
  77 #define PAGE_ALIGN(buf) \
  78         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  79
  80 #define JOB_START_TIMEOUT       (5 * 1000)
  81
  82 static void sig_int(int sig)
  83 {
  84         if (threads) {
  85                 if (is_backend)
  86                         fio_server_got_signal(sig);
  87                 else {
  88                         log_info("\nfio: terminating on signal %d\n", sig);
  89                         fflush(stdout);
  90                         exit_value = 128;
  91                 }
  92
  93                 fio_terminate_threads(TERMINATE_ALL);
  94         }
  95 }
  96
  97 static void sig_show_status(int sig)
  98 {
  99         show_running_run_stats();
 100 }
 101
 102 static void set_sig_handlers(void)
 103 {
 104         struct sigaction act;
 105
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGINT, &act, NULL);
 110
 111         memset(&act, 0, sizeof(act));
 112         act.sa_handler = sig_int;
 113         act.sa_flags = SA_RESTART;
 114         sigaction(SIGTERM, &act, NULL);
 115
 116 /* Windows uses SIGBREAK as a quit signal from other applications */
 117 #ifdef WIN32
 118         memset(&act, 0, sizeof(act));
 119         act.sa_handler = sig_int;
 120         act.sa_flags = SA_RESTART;
 121         sigaction(SIGBREAK, &act, NULL);
 122 #endif
 123
 124         memset(&act, 0, sizeof(act));
 125         act.sa_handler = sig_show_status;
 126         act.sa_flags = SA_RESTART;
 127         sigaction(SIGUSR1, &act, NULL);
 128
 129         if (is_backend) {
 130                 memset(&act, 0, sizeof(act));
 131                 act.sa_handler = sig_int;
 132                 act.sa_flags = SA_RESTART;
 133                 sigaction(SIGPIPE, &act, NULL);
 134         }
 135 }
 136
 137 /*
 138  * Check if we are above the minimum rate given.
 139  */
 140 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 141                             enum fio_ddir ddir)
 142 {
 143         unsigned long long bytes = 0;
 144         unsigned long iops = 0;
 145         unsigned long spent;
 146         unsigned long rate;
 147         unsigned int ratemin = 0;
 148         unsigned int rate_iops = 0;
 149         unsigned int rate_iops_min = 0;
 150
 151         assert(ddir_rw(ddir));
 152
 153         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 154                 return 0;
 155
 156         /*
 157          * allow a 2 second settle period in the beginning
 158          */
 159         if (mtime_since(&td->start, now) < 2000)
 160                 return 0;
 161
 162         iops += td->this_io_blocks[ddir];
 163         bytes += td->this_io_bytes[ddir];
 164         ratemin += td->o.ratemin[ddir];
 165         rate_iops += td->o.rate_iops[ddir];
 166         rate_iops_min += td->o.rate_iops_min[ddir];
 167
 168         /*
 169          * if rate blocks is set, sample is running
 170          */
 171         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 172                 spent = mtime_since(&td->lastrate[ddir], now);
 173                 if (spent < td->o.ratecycle)
 174                         return 0;
 175
 176                 if (td->o.rate[ddir]) {
 177                         /*
 178                          * check bandwidth specified rate
 179                          */
 180                         if (bytes < td->rate_bytes[ddir]) {
 181                                 log_err("%s: min rate %u not met\n", td->o.name,
 182                                                                 ratemin);
 183                                 return 1;
 184                         } else {
 185                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 186                                 if (rate < ratemin ||
 187                                     bytes < td->rate_bytes[ddir]) {
 188                                         log_err("%s: min rate %u not met, got"
 189                                                 " %luKB/sec\n", td->o.name,
 190                                                         ratemin, rate);
 191                                         return 1;
 192                                 }
 193                         }
 194                 } else {
 195                         /*
 196                          * checks iops specified rate
 197                          */
 198                         if (iops < rate_iops) {
 199                                 log_err("%s: min iops rate %u not met\n",
 200                                                 td->o.name, rate_iops);
 201                                 return 1;
 202                         } else {
 203                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 204                                 if (rate < rate_iops_min ||
 205                                     iops < td->rate_blocks[ddir]) {
 206                                         log_err("%s: min iops rate %u not met,"
 207                                                 " got %lu\n", td->o.name,
 208                                                         rate_iops_min, rate);
 209                                 }
 210                         }
 211                 }
 212         }
 213
 214         td->rate_bytes[ddir] = bytes;
 215         td->rate_blocks[ddir] = iops;
 216         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 217         return 0;
 218 }
 219
 220 static int check_min_rate(struct thread_data *td, struct timeval *now,
 221                           uint64_t *bytes_done)
 222 {
 223         int ret = 0;
 224
 225         if (bytes_done[DDIR_READ])
 226                 ret |= __check_min_rate(td, now, DDIR_READ);
 227         if (bytes_done[DDIR_WRITE])
 228                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 229         if (bytes_done[DDIR_TRIM])
 230                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 231
 232         return ret;
 233 }
 234
 235 /*
 236  * When job exits, we can cancel the in-flight IO if we are using async
 237  * io. Attempt to do so.
 238  */
 239 static void cleanup_pending_aio(struct thread_data *td)
 240 {
 241         int r;
 242
 243         /*
 244          * get immediately available events, if any
 245          */
 246         r = io_u_queued_complete(td, 0, NULL);
 247         if (r < 0)
 248                 return;
 249
 250         /*
 251          * now cancel remaining active events
 252          */
 253         if (td->io_ops->cancel) {
 254                 struct io_u *io_u;
 255                 int i;
 256
 257                 io_u_qiter(&td->io_u_all, io_u, i) {
 258                         if (io_u->flags & IO_U_F_FLIGHT) {
 259                                 r = td->io_ops->cancel(td, io_u);
 260                                 if (!r)
 261                                         put_io_u(td, io_u);
 262                         }
 263                 }
 264         }
 265
 266         if (td->cur_depth)
 267                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 268 }
 269
 270 /*
 271  * Helper to handle the final sync of a file. Works just like the normal
 272  * io path, just does everything sync.
 273  */
 274 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 275 {
 276         struct io_u *io_u = __get_io_u(td);
 277         int ret;
 278
 279         if (!io_u)
 280                 return 1;
 281
 282         io_u->ddir = DDIR_SYNC;
 283         io_u->file = f;
 284
 285         if (td_io_prep(td, io_u)) {
 286                 put_io_u(td, io_u);
 287                 return 1;
 288         }
 289
 290 requeue:
 291         ret = td_io_queue(td, io_u);
 292         if (ret < 0) {
 293                 td_verror(td, io_u->error, "td_io_queue");
 294                 put_io_u(td, io_u);
 295                 return 1;
 296         } else if (ret == FIO_Q_QUEUED) {
 297                 if (io_u_queued_complete(td, 1, NULL) < 0)
 298                         return 1;
 299         } else if (ret == FIO_Q_COMPLETED) {
 300                 if (io_u->error) {
 301                         td_verror(td, io_u->error, "td_io_queue");
 302                         return 1;
 303                 }
 304
 305                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 306                         return 1;
 307         } else if (ret == FIO_Q_BUSY) {
 308                 if (td_io_commit(td))
 309                         return 1;
 310                 goto requeue;
 311         }
 312
 313         return 0;
 314 }
 315
 316 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 317 {
 318         int ret;
 319
 320         if (fio_file_open(f))
 321                 return fio_io_sync(td, f);
 322
 323         if (td_io_open_file(td, f))
 324                 return 1;
 325
 326         ret = fio_io_sync(td, f);
 327         td_io_close_file(td, f);
 328         return ret;
 329 }
 330
 331 static inline void __update_tv_cache(struct thread_data *td)
 332 {
 333         fio_gettime(&td->tv_cache, NULL);
 334 }
 335
 336 static inline void update_tv_cache(struct thread_data *td)
 337 {
 338         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 339                 __update_tv_cache(td);
 340 }
 341
 342 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 343 {
 344         if (in_ramp_time(td))
 345                 return 0;
 346         if (!td->o.timeout)
 347                 return 0;
 348         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 349                 return 1;
 350
 351         return 0;
 352 }
 353
 354 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 355                                int *retptr)
 356 {
 357         int ret = *retptr;
 358
 359         if (ret < 0 || td->error) {
 360                 int err = td->error;
 361                 enum error_type_bit eb;
 362
 363                 if (ret < 0)
 364                         err = -ret;
 365
 366                 eb = td_error_type(ddir, err);
 367                 if (!(td->o.continue_on_error & (1 << eb)))
 368                         return 1;
 369
 370                 if (td_non_fatal_error(td, eb, err)) {
 371                         /*
 372                          * Continue with the I/Os in case of
 373                          * a non fatal error.
 374                          */
 375                         update_error_count(td, err);
 376                         td_clear_error(td);
 377                         *retptr = 0;
 378                         return 0;
 379                 } else if (td->o.fill_device && err == ENOSPC) {
 380                         /*
 381                          * We expect to hit this error if
 382                          * fill_device option is set.
 383                          */
 384                         td_clear_error(td);
 385                         td->terminate = 1;
 386                         return 1;
 387                 } else {
 388                         /*
 389                          * Stop the I/O in case of a fatal
 390                          * error.
 391                          */
 392                         update_error_count(td, err);
 393                         return 1;
 394                 }
 395         }
 396
 397         return 0;
 398 }
 399
 400 static void check_update_rusage(struct thread_data *td)
 401 {
 402         if (td->update_rusage) {
 403                 td->update_rusage = 0;
 404                 update_rusage_stat(td);
 405                 fio_mutex_up(td->rusage_sem);
 406         }
 407 }
 408
 409 /*
 410  * The main verify engine. Runs over the writes we previously submitted,
 411  * reads the blocks back in, and checks the crc/md5 of the data.
 412  */
 413 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 414 {
 415         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 416         struct fio_file *f;
 417         struct io_u *io_u;
 418         int ret, min_events;
 419         unsigned int i;
 420
 421         dprint(FD_VERIFY, "starting loop\n");
 422
 423         /*
 424          * sync io first and invalidate cache, to make sure we really
 425          * read from disk.
 426          */
 427         for_each_file(td, f, i) {
 428                 if (!fio_file_open(f))
 429                         continue;
 430                 if (fio_io_sync(td, f))
 431                         break;
 432                 if (file_invalidate_cache(td, f))
 433                         break;
 434         }
 435
 436         check_update_rusage(td);
 437
 438         if (td->error)
 439                 return;
 440
 441         td_set_runstate(td, TD_VERIFYING);
 442
 443         io_u = NULL;
 444         while (!td->terminate) {
 445                 enum fio_ddir ddir;
 446                 int ret2, full;
 447
 448                 update_tv_cache(td);
 449                 check_update_rusage(td);
 450
 451                 if (runtime_exceeded(td, &td->tv_cache)) {
 452                         __update_tv_cache(td);
 453                         if (runtime_exceeded(td, &td->tv_cache)) {
 454                                 td->terminate = 1;
 455                                 break;
 456                         }
 457                 }
 458
 459                 if (flow_threshold_exceeded(td))
 460                         continue;
 461
 462                 if (!td->o.experimental_verify) {
 463                         io_u = __get_io_u(td);
 464                         if (!io_u)
 465                                 break;
 466
 467                         if (get_next_verify(td, io_u)) {
 468                                 put_io_u(td, io_u);
 469                                 break;
 470                         }
 471
 472                         if (td_io_prep(td, io_u)) {
 473                                 put_io_u(td, io_u);
 474                                 break;
 475                         }
 476                 } else {
 477                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 478                                 break;
 479
 480                         while ((io_u = get_io_u(td)) != NULL) {
 481                                 /*
 482                                  * We are only interested in the places where
 483                                  * we wrote or trimmed IOs. Turn those into
 484                                  * reads for verification purposes.
 485                                  */
 486                                 if (io_u->ddir == DDIR_READ) {
 487                                         /*
 488                                          * Pretend we issued it for rwmix
 489                                          * accounting
 490                                          */
 491                                         td->io_issues[DDIR_READ]++;
 492                                         put_io_u(td, io_u);
 493                                         continue;
 494                                 } else if (io_u->ddir == DDIR_TRIM) {
 495                                         io_u->ddir = DDIR_READ;
 496                                         io_u->flags |= IO_U_F_TRIMMED;
 497                                         break;
 498                                 } else if (io_u->ddir == DDIR_WRITE) {
 499                                         io_u->ddir = DDIR_READ;
 500                                         break;
 501                                 } else {
 502                                         put_io_u(td, io_u);
 503                                         continue;
 504                                 }
 505                         }
 506
 507                         if (!io_u)
 508                                 break;
 509                 }
 510
 511                 if (td->o.verify_async)
 512                         io_u->end_io = verify_io_u_async;
 513                 else
 514                         io_u->end_io = verify_io_u;
 515
 516                 ddir = io_u->ddir;
 517
 518                 ret = td_io_queue(td, io_u);
 519                 switch (ret) {
 520                 case FIO_Q_COMPLETED:
 521                         if (io_u->error) {
 522                                 ret = -io_u->error;
 523                                 clear_io_u(td, io_u);
 524                         } else if (io_u->resid) {
 525                                 int bytes = io_u->xfer_buflen - io_u->resid;
 526
 527                                 /*
 528                                  * zero read, fail
 529                                  */
 530                                 if (!bytes) {
 531                                         td_verror(td, EIO, "full resid");
 532                                         put_io_u(td, io_u);
 533                                         break;
 534                                 }
 535
 536                                 io_u->xfer_buflen = io_u->resid;
 537                                 io_u->xfer_buf += bytes;
 538                                 io_u->offset += bytes;
 539
 540                                 if (ddir_rw(io_u->ddir))
 541                                         td->ts.short_io_u[io_u->ddir]++;
 542
 543                                 f = io_u->file;
 544                                 if (io_u->offset == f->real_file_size)
 545                                         goto sync_done;
 546
 547                                 requeue_io_u(td, &io_u);
 548                         } else {
 549 sync_done:
 550                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 551                                 if (ret < 0)
 552                                         break;
 553                         }
 554                         continue;
 555                 case FIO_Q_QUEUED:
 556                         break;
 557                 case FIO_Q_BUSY:
 558                         requeue_io_u(td, &io_u);
 559                         ret2 = td_io_commit(td);
 560                         if (ret2 < 0)
 561                                 ret = ret2;
 562                         break;
 563                 default:
 564                         assert(ret < 0);
 565                         td_verror(td, -ret, "td_io_queue");
 566                         break;
 567                 }
 568
 569                 if (break_on_this_error(td, ddir, &ret))
 570                         break;
 571
 572                 /*
 573                  * if we can queue more, do so. but check if there are
 574                  * completed io_u's first. Note that we can get BUSY even
 575                  * without IO queued, if the system is resource starved.
 576                  */
 577                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 578                 if (full || !td->o.iodepth_batch_complete) {
 579                         min_events = min(td->o.iodepth_batch_complete,
 580                                          td->cur_depth);
 581                         /*
 582                          * if the queue is full, we MUST reap at least 1 event
 583                          */
 584                         if (full && !min_events)
 585                                 min_events = 1;
 586
 587                         do {
 588                                 /*
 589                                  * Reap required number of io units, if any,
 590                                  * and do the verification on them through
 591                                  * the callback handler
 592                                  */
 593                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 594                                         ret = -1;
 595                                         break;
 596                                 }
 597                         } while (full && (td->cur_depth > td->o.iodepth_low));
 598                 }
 599                 if (ret < 0)
 600                         break;
 601         }
 602
 603         check_update_rusage(td);
 604
 605         if (!td->error) {
 606                 min_events = td->cur_depth;
 607
 608                 if (min_events)
 609                         ret = io_u_queued_complete(td, min_events, NULL);
 610         } else
 611                 cleanup_pending_aio(td);
 612
 613         td_set_runstate(td, TD_RUNNING);
 614
 615         dprint(FD_VERIFY, "exiting loop\n");
 616 }
 617
 618 static int io_bytes_exceeded(struct thread_data *td)
 619 {
 620         unsigned long long bytes;
 621
 622         if (td_rw(td))
 623                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 624         else if (td_write(td))
 625                 bytes = td->this_io_bytes[DDIR_WRITE];
 626         else if (td_read(td))
 627                 bytes = td->this_io_bytes[DDIR_READ];
 628         else
 629                 bytes = td->this_io_bytes[DDIR_TRIM];
 630
 631         return bytes >= td->o.size;
 632 }
 633
 634 /*
 635  * Main IO worker function. It retrieves io_u's to process and queues
 636  * and reaps them, checking for rate and errors along the way.
 637  *
 638  * Returns number of bytes written and trimmed.
 639  */
 640 static uint64_t do_io(struct thread_data *td)
 641 {
 642         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 643         unsigned int i;
 644         int ret = 0;
 645         uint64_t total_bytes, bytes_issued = 0;
 646
 647         if (in_ramp_time(td))
 648                 td_set_runstate(td, TD_RAMP);
 649         else
 650                 td_set_runstate(td, TD_RUNNING);
 651
 652         lat_target_init(td);
 653
 654         total_bytes = td->o.size;
 655         if (td->o.verify != VERIFY_NONE && td_write(td))
 656                 total_bytes += td->o.size;
 657
 658         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 659                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 660                 td->o.time_based) {
 661                 struct timeval comp_time;
 662                 int min_evts = 0;
 663                 struct io_u *io_u;
 664                 int ret2, full;
 665                 enum fio_ddir ddir;
 666
 667                 check_update_rusage(td);
 668
 669                 if (td->terminate || td->done)
 670                         break;
 671
 672                 update_tv_cache(td);
 673
 674                 if (runtime_exceeded(td, &td->tv_cache)) {
 675                         __update_tv_cache(td);
 676                         if (runtime_exceeded(td, &td->tv_cache)) {
 677                                 td->terminate = 1;
 678                                 break;
 679                         }
 680                 }
 681
 682                 if (flow_threshold_exceeded(td))
 683                         continue;
 684
 685                 if (bytes_issued >= total_bytes)
 686                         break;
 687
 688                 io_u = get_io_u(td);
 689                 if (!io_u) {
 690                         if (td->o.latency_target)
 691                                 goto reap;
 692                         break;
 693                 }
 694
 695                 ddir = io_u->ddir;
 696
 697                 /*
 698                  * Add verification end_io handler if:
 699                  *      - Asked to verify (!td_rw(td))
 700                  *      - Or the io_u is from our verify list (mixed write/ver)
 701                  */
 702                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 703                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 704
 705                         if (!td->o.verify_pattern_bytes) {
 706                                 io_u->rand_seed = __rand(&td->__verify_state);
 707                                 if (sizeof(int) != sizeof(long *))
 708                                         io_u->rand_seed *= __rand(&td->__verify_state);
 709                         }
 710
 711                         if (td->o.verify_async)
 712                                 io_u->end_io = verify_io_u_async;
 713                         else
 714                                 io_u->end_io = verify_io_u;
 715                         td_set_runstate(td, TD_VERIFYING);
 716                 } else if (in_ramp_time(td))
 717                         td_set_runstate(td, TD_RAMP);
 718                 else
 719                         td_set_runstate(td, TD_RUNNING);
 720
 721                 /*
 722                  * Verify_backlog disabled: We need to log rand seed before the
 723                  * actual IO to be able to replay it correctly in the verify phase.
 724                  */
 725                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 726                     td->o.do_verify &&
 727                     td->o.verify != VERIFY_NONE &&
 728                     !td->o.experimental_verify &&
 729                     !(td->flags & TD_F_VER_BACKLOG))
 730                         log_io_piece(td, io_u);
 731
 732                 ret = td_io_queue(td, io_u);
 733                 switch (ret) {
 734                 case FIO_Q_COMPLETED:
 735                         if (io_u->error) {
 736                                 ret = -io_u->error;
 737                                 clear_io_u(td, io_u);
 738                         } else if (io_u->resid) {
 739                                 int bytes = io_u->xfer_buflen - io_u->resid;
 740                                 struct fio_file *f = io_u->file;
 741
 742                                 bytes_issued += bytes;
 743                                 /*
 744                                  * zero read, fail
 745                                  */
 746                                 if (!bytes) {
 747                                         td_verror(td, EIO, "full resid");
 748                                         put_io_u(td, io_u);
 749                                         break;
 750                                 }
 751
 752                                 io_u->xfer_buflen = io_u->resid;
 753                                 io_u->xfer_buf += bytes;
 754                                 io_u->offset += bytes;
 755
 756                                 if (ddir_rw(io_u->ddir))
 757                                         td->ts.short_io_u[io_u->ddir]++;
 758
 759                                 if (io_u->offset == f->real_file_size)
 760                                         goto sync_done;
 761
 762                                 requeue_io_u(td, &io_u);
 763                         } else {
 764 sync_done:
 765                                 if (__should_check_rate(td, DDIR_READ) ||
 766                                     __should_check_rate(td, DDIR_WRITE) ||
 767                                     __should_check_rate(td, DDIR_TRIM))
 768                                         fio_gettime(&comp_time, NULL);
 769
 770                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 771                                 if (ret < 0)
 772                                         break;
 773                                 bytes_issued += io_u->xfer_buflen;
 774                         }
 775                         break;
 776                 case FIO_Q_QUEUED:
 777                         /*
 778                          * if the engine doesn't have a commit hook,
 779                          * the io_u is really queued. if it does have such
 780                          * a hook, it has to call io_u_queued() itself.
 781                          */
 782                         if (td->io_ops->commit == NULL)
 783                                 io_u_queued(td, io_u);
 784                         bytes_issued += io_u->xfer_buflen;
 785                         break;
 786                 case FIO_Q_BUSY:
 787                         requeue_io_u(td, &io_u);
 788                         ret2 = td_io_commit(td);
 789                         if (ret2 < 0)
 790                                 ret = ret2;
 791                         break;
 792                 default:
 793                         assert(ret < 0);
 794                         put_io_u(td, io_u);
 795                         break;
 796                 }
 797
 798                 if (break_on_this_error(td, ddir, &ret))
 799                         break;
 800
 801                 /*
 802                  * See if we need to complete some commands. Note that we
 803                  * can get BUSY even without IO queued, if the system is
 804                  * resource starved.
 805                  */
 806 reap:
 807                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 808                 if (full || !td->o.iodepth_batch_complete) {
 809                         min_evts = min(td->o.iodepth_batch_complete,
 810                                         td->cur_depth);
 811                         /*
 812                          * if the queue is full, we MUST reap at least 1 event
 813                          */
 814                         if (full && !min_evts)
 815                                 min_evts = 1;
 816
 817                         if (__should_check_rate(td, DDIR_READ) ||
 818                             __should_check_rate(td, DDIR_WRITE) ||
 819                             __should_check_rate(td, DDIR_TRIM))
 820                                 fio_gettime(&comp_time, NULL);
 821
 822                         do {
 823                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 824                                 if (ret < 0)
 825                                         break;
 826
 827                         } while (full && (td->cur_depth > td->o.iodepth_low));
 828                 }
 829
 830                 if (ret < 0)
 831                         break;
 832                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 833                         continue;
 834
 835                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 836                         if (check_min_rate(td, &comp_time, bytes_done)) {
 837                                 if (exitall_on_terminate)
 838                                         fio_terminate_threads(td->groupid);
 839                                 td_verror(td, EIO, "check_min_rate");
 840                                 break;
 841                         }
 842                 }
 843                 if (!in_ramp_time(td) && td->o.latency_target)
 844                         lat_target_check(td);
 845
 846                 if (td->o.thinktime) {
 847                         unsigned long long b;
 848
 849                         b = ddir_rw_sum(td->io_blocks);
 850                         if (!(b % td->o.thinktime_blocks)) {
 851                                 int left;
 852
 853                                 io_u_quiesce(td);
 854
 855                                 if (td->o.thinktime_spin)
 856                                         usec_spin(td->o.thinktime_spin);
 857
 858                                 left = td->o.thinktime - td->o.thinktime_spin;
 859                                 if (left)
 860                                         usec_sleep(td, left);
 861                         }
 862                 }
 863         }
 864
 865         check_update_rusage(td);
 866
 867         if (td->trim_entries)
 868                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 869
 870         if (td->o.fill_device && td->error == ENOSPC) {
 871                 td->error = 0;
 872                 td->terminate = 1;
 873         }
 874         if (!td->error) {
 875                 struct fio_file *f;
 876
 877                 i = td->cur_depth;
 878                 if (i) {
 879                         ret = io_u_queued_complete(td, i, bytes_done);
 880                         if (td->o.fill_device && td->error == ENOSPC)
 881                                 td->error = 0;
 882                 }
 883
 884                 if (should_fsync(td) && td->o.end_fsync) {
 885                         td_set_runstate(td, TD_FSYNCING);
 886
 887                         for_each_file(td, f, i) {
 888                                 if (!fio_file_fsync(td, f))
 889                                         continue;
 890
 891                                 log_err("fio: end_fsync failed for file %s\n",
 892                                                                 f->file_name);
 893                         }
 894                 }
 895         } else
 896                 cleanup_pending_aio(td);
 897
 898         /*
 899          * stop job if we failed doing any IO
 900          */
 901         if (!ddir_rw_sum(td->this_io_bytes))
 902                 td->done = 1;
 903
 904         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 905 }
 906
 907 static void cleanup_io_u(struct thread_data *td)
 908 {
 909         struct io_u *io_u;
 910
 911         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 912
 913                 if (td->io_ops->io_u_free)
 914                         td->io_ops->io_u_free(td, io_u);
 915
 916                 fio_memfree(io_u, sizeof(*io_u));
 917         }
 918
 919         free_io_mem(td);
 920
 921         io_u_rexit(&td->io_u_requeues);
 922         io_u_qexit(&td->io_u_freelist);
 923         io_u_qexit(&td->io_u_all);
 924 }
 925
 926 static int init_io_u(struct thread_data *td)
 927 {
 928         struct io_u *io_u;
 929         unsigned int max_bs, min_write;
 930         int cl_align, i, max_units;
 931         int data_xfer = 1, err;
 932         char *p;
 933
 934         max_units = td->o.iodepth;
 935         max_bs = td_max_bs(td);
 936         min_write = td->o.min_bs[DDIR_WRITE];
 937         td->orig_buffer_size = (unsigned long long) max_bs
 938                                         * (unsigned long long) max_units;
 939
 940         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 941                 data_xfer = 0;
 942
 943         err = 0;
 944         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 945         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 946         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 947
 948         if (err) {
 949                 log_err("fio: failed setting up IO queues\n");
 950                 return 1;
 951         }
 952
 953         /*
 954          * if we may later need to do address alignment, then add any
 955          * possible adjustment here so that we don't cause a buffer
 956          * overflow later. this adjustment may be too much if we get
 957          * lucky and the allocator gives us an aligned address.
 958          */
 959         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 960             (td->io_ops->flags & FIO_RAWIO))
 961                 td->orig_buffer_size += page_mask + td->o.mem_align;
 962
 963         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 964                 unsigned long bs;
 965
 966                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 967                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 968         }
 969
 970         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 971                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 972                 return 1;
 973         }
 974
 975         if (data_xfer && allocate_io_mem(td))
 976                 return 1;
 977
 978         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 979             (td->io_ops->flags & FIO_RAWIO))
 980                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 981         else
 982                 p = td->orig_buffer;
 983
 984         cl_align = os_cache_line_size();
 985
 986         for (i = 0; i < max_units; i++) {
 987                 void *ptr;
 988
 989                 if (td->terminate)
 990                         return 1;
 991
 992                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 993                 if (!ptr) {
 994                         log_err("fio: unable to allocate aligned memory\n");
 995                         break;
 996                 }
 997
 998                 io_u = ptr;
 999                 memset(io_u, 0, sizeof(*io_u));
1000                 INIT_FLIST_HEAD(&io_u->verify_list);
1001                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1002
1003                 if (data_xfer) {
1004                         io_u->buf = p;
1005                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1006
1007                         if (td_write(td))
1008                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1009                         if (td_write(td) && td->o.verify_pattern_bytes) {
1010                                 /*
1011                                  * Fill the buffer with the pattern if we are
1012                                  * going to be doing writes.
1013                                  */
1014                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1015                         }
1016                 }
1017
1018                 io_u->index = i;
1019                 io_u->flags = IO_U_F_FREE;
1020                 io_u_qpush(&td->io_u_freelist, io_u);
1021
1022                 /*
1023                  * io_u never leaves this stack, used for iteration of all
1024                  * io_u buffers.
1025                  */
1026                 io_u_qpush(&td->io_u_all, io_u);
1027
1028                 if (td->io_ops->io_u_init) {
1029                         int ret = td->io_ops->io_u_init(td, io_u);
1030
1031                         if (ret) {
1032                                 log_err("fio: failed to init engine data: %d\n", ret);
1033                                 return 1;
1034                         }
1035                 }
1036
1037                 p += max_bs;
1038         }
1039
1040         return 0;
1041 }
1042
1043 static int switch_ioscheduler(struct thread_data *td)
1044 {
1045         char tmp[256], tmp2[128];
1046         FILE *f;
1047         int ret;
1048
1049         if (td->io_ops->flags & FIO_DISKLESSIO)
1050                 return 0;
1051
1052         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1053
1054         f = fopen(tmp, "r+");
1055         if (!f) {
1056                 if (errno == ENOENT) {
1057                         log_err("fio: os or kernel doesn't support IO scheduler"
1058                                 " switching\n");
1059                         return 0;
1060                 }
1061                 td_verror(td, errno, "fopen iosched");
1062                 return 1;
1063         }
1064
1065         /*
1066          * Set io scheduler.
1067          */
1068         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1069         if (ferror(f) || ret != 1) {
1070                 td_verror(td, errno, "fwrite");
1071                 fclose(f);
1072                 return 1;
1073         }
1074
1075         rewind(f);
1076
1077         /*
1078          * Read back and check that the selected scheduler is now the default.
1079          */
1080         ret = fread(tmp, 1, sizeof(tmp), f);
1081         if (ferror(f) || ret < 0) {
1082                 td_verror(td, errno, "fread");
1083                 fclose(f);
1084                 return 1;
1085         }
1086
1087         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1088         if (!strstr(tmp, tmp2)) {
1089                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1090                 td_verror(td, EINVAL, "iosched_switch");
1091                 fclose(f);
1092                 return 1;
1093         }
1094
1095         fclose(f);
1096         return 0;
1097 }
1098
1099 static int keep_running(struct thread_data *td)
1100 {
1101         if (td->done)
1102                 return 0;
1103         if (td->o.time_based)
1104                 return 1;
1105         if (td->o.loops) {
1106                 td->o.loops--;
1107                 return 1;
1108         }
1109
1110         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1111                 uint64_t diff;
1112
1113                 /*
1114                  * If the difference is less than the minimum IO size, we
1115                  * are done.
1116                  */
1117                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1118                 if (diff < td_max_bs(td))
1119                         return 0;
1120
1121                 return 1;
1122         }
1123
1124         return 0;
1125 }
1126
1127 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1128 {
1129         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1130         char *str;
1131
1132         str = malloc(newlen);
1133         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1134
1135         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1136         ret = system(str);
1137         if (ret == -1)
1138                 log_err("fio: exec of cmd <%s> failed\n", str);
1139
1140         free(str);
1141         return ret;
1142 }
1143
1144 /*
1145  * Dry run to compute correct state of numberio for verification.
1146  */
1147 static uint64_t do_dry_run(struct thread_data *td)
1148 {
1149         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1150
1151         td_set_runstate(td, TD_RUNNING);
1152
1153         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1154                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1155                 struct io_u *io_u;
1156                 int ret;
1157
1158                 if (td->terminate || td->done)
1159                         break;
1160
1161                 io_u = get_io_u(td);
1162                 if (!io_u)
1163                         break;
1164
1165                 io_u->flags |= IO_U_F_FLIGHT;
1166                 io_u->error = 0;
1167                 io_u->resid = 0;
1168                 if (ddir_rw(acct_ddir(io_u)))
1169                         td->io_issues[acct_ddir(io_u)]++;
1170                 if (ddir_rw(io_u->ddir)) {
1171                         io_u_mark_depth(td, 1);
1172                         td->ts.total_io_u[io_u->ddir]++;
1173                 }
1174
1175                 ret = io_u_sync_complete(td, io_u, bytes_done);
1176                 (void) ret;
1177         }
1178
1179         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1180 }
1181
1182 /*
1183  * Entry point for the thread based jobs. The process based jobs end up
1184  * here as well, after a little setup.
1185  */
1186 static void *thread_main(void *data)
1187 {
1188         unsigned long long elapsed;
1189         struct thread_data *td = data;
1190         struct thread_options *o = &td->o;
1191         pthread_condattr_t attr;
1192         int clear_state;
1193         int ret;
1194
1195         if (!o->use_thread) {
1196                 setsid();
1197                 td->pid = getpid();
1198         } else
1199                 td->pid = gettid();
1200
1201         /*
1202          * fio_time_init() may not have been called yet if running as a server
1203          */
1204         fio_time_init();
1205
1206         fio_local_clock_init(o->use_thread);
1207
1208         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1209
1210         if (is_backend)
1211                 fio_server_send_start(td);
1212
1213         INIT_FLIST_HEAD(&td->io_log_list);
1214         INIT_FLIST_HEAD(&td->io_hist_list);
1215         INIT_FLIST_HEAD(&td->verify_list);
1216         INIT_FLIST_HEAD(&td->trim_list);
1217         INIT_FLIST_HEAD(&td->next_rand_list);
1218         pthread_mutex_init(&td->io_u_lock, NULL);
1219         td->io_hist_tree = RB_ROOT;
1220
1221         pthread_condattr_init(&attr);
1222         pthread_cond_init(&td->verify_cond, &attr);
1223         pthread_cond_init(&td->free_cond, &attr);
1224
1225         td_set_runstate(td, TD_INITIALIZED);
1226         dprint(FD_MUTEX, "up startup_mutex\n");
1227         fio_mutex_up(startup_mutex);
1228         dprint(FD_MUTEX, "wait on td->mutex\n");
1229         fio_mutex_down(td->mutex);
1230         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1231
1232         /*
1233          * the ->mutex mutex is now no longer used, close it to avoid
1234          * eating a file descriptor
1235          */
1236         fio_mutex_remove(td->mutex);
1237         td->mutex = NULL;
1238
1239         /*
1240          * A new gid requires privilege, so we need to do this before setting
1241          * the uid.
1242          */
1243         if (o->gid != -1U && setgid(o->gid)) {
1244                 td_verror(td, errno, "setgid");
1245                 goto err;
1246         }
1247         if (o->uid != -1U && setuid(o->uid)) {
1248                 td_verror(td, errno, "setuid");
1249                 goto err;
1250         }
1251
1252         /*
1253          * If we have a gettimeofday() thread, make sure we exclude that
1254          * thread from this job
1255          */
1256         if (o->gtod_cpu)
1257                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1258
1259         /*
1260          * Set affinity first, in case it has an impact on the memory
1261          * allocations.
1262          */
1263         if (o->cpumask_set) {
1264                 ret = fio_setaffinity(td->pid, o->cpumask);
1265                 if (ret == -1) {
1266                         td_verror(td, errno, "cpu_set_affinity");
1267                         goto err;
1268                 }
1269         }
1270
1271 #ifdef CONFIG_LIBNUMA
1272         /* numa node setup */
1273         if (o->numa_cpumask_set || o->numa_memmask_set) {
1274                 int ret;
1275
1276                 if (numa_available() < 0) {
1277                         td_verror(td, errno, "Does not support NUMA API\n");
1278                         goto err;
1279                 }
1280
1281                 if (o->numa_cpumask_set) {
1282                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1283                         if (ret == -1) {
1284                                 td_verror(td, errno, \
1285                                         "numa_run_on_node_mask failed\n");
1286                                 goto err;
1287                         }
1288                 }
1289
1290                 if (o->numa_memmask_set) {
1291
1292                         switch (o->numa_mem_mode) {
1293                         case MPOL_INTERLEAVE:
1294                                 numa_set_interleave_mask(o->numa_memnodesmask);
1295                                 break;
1296                         case MPOL_BIND:
1297                                 numa_set_membind(o->numa_memnodesmask);
1298                                 break;
1299                         case MPOL_LOCAL:
1300                                 numa_set_localalloc();
1301                                 break;
1302                         case MPOL_PREFERRED:
1303                                 numa_set_preferred(o->numa_mem_prefer_node);
1304                                 break;
1305                         case MPOL_DEFAULT:
1306                         default:
1307                                 break;
1308                         }
1309
1310                 }
1311         }
1312 #endif
1313
1314         if (fio_pin_memory(td))
1315                 goto err;
1316
1317         /*
1318          * May alter parameters that init_io_u() will use, so we need to
1319          * do this first.
1320          */
1321         if (init_iolog(td))
1322                 goto err;
1323
1324         if (init_io_u(td))
1325                 goto err;
1326
1327         if (o->verify_async && verify_async_init(td))
1328                 goto err;
1329
1330         if (o->ioprio) {
1331                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1332                 if (ret == -1) {
1333                         td_verror(td, errno, "ioprio_set");
1334                         goto err;
1335                 }
1336         }
1337
1338         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1339                 goto err;
1340
1341         errno = 0;
1342         if (nice(o->nice) == -1 && errno != 0) {
1343                 td_verror(td, errno, "nice");
1344                 goto err;
1345         }
1346
1347         if (o->ioscheduler && switch_ioscheduler(td))
1348                 goto err;
1349
1350         if (!o->create_serialize && setup_files(td))
1351                 goto err;
1352
1353         if (td_io_init(td))
1354                 goto err;
1355
1356         if (init_random_map(td))
1357                 goto err;
1358
1359         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1360                 goto err;
1361
1362         if (o->pre_read) {
1363                 if (pre_read_files(td) < 0)
1364                         goto err;
1365         }
1366
1367         fio_verify_init(td);
1368
1369         fio_gettime(&td->epoch, NULL);
1370         fio_getrusage(&td->ru_start);
1371         clear_state = 0;
1372         while (keep_running(td)) {
1373                 uint64_t verify_bytes;
1374
1375                 fio_gettime(&td->start, NULL);
1376                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1377                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1378                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1379
1380                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1381                                 o->ratemin[DDIR_TRIM]) {
1382                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1383                                                 sizeof(td->bw_sample_time));
1384                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1385                                                 sizeof(td->bw_sample_time));
1386                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1387                                                 sizeof(td->bw_sample_time));
1388                 }
1389
1390                 if (clear_state)
1391                         clear_io_state(td);
1392
1393                 prune_io_piece_log(td);
1394
1395                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1396                         verify_bytes = do_dry_run(td);
1397                 else
1398                         verify_bytes = do_io(td);
1399
1400                 clear_state = 1;
1401
1402                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1403                         elapsed = utime_since_now(&td->start);
1404                         td->ts.runtime[DDIR_READ] += elapsed;
1405                 }
1406                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1407                         elapsed = utime_since_now(&td->start);
1408                         td->ts.runtime[DDIR_WRITE] += elapsed;
1409                 }
1410                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1411                         elapsed = utime_since_now(&td->start);
1412                         td->ts.runtime[DDIR_TRIM] += elapsed;
1413                 }
1414
1415                 if (td->error || td->terminate)
1416                         break;
1417
1418                 if (!o->do_verify ||
1419                     o->verify == VERIFY_NONE ||
1420                     (td->io_ops->flags & FIO_UNIDIR))
1421                         continue;
1422
1423                 clear_io_state(td);
1424
1425                 fio_gettime(&td->start, NULL);
1426
1427                 do_verify(td, verify_bytes);
1428
1429                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1430
1431                 if (td->error || td->terminate)
1432                         break;
1433         }
1434
1435         update_rusage_stat(td);
1436         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1437         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1438         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1439         td->ts.total_run_time = mtime_since_now(&td->epoch);
1440         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1441         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1442         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1443
1444         fio_unpin_memory(td);
1445
1446         fio_mutex_down(writeout_mutex);
1447         if (td->bw_log) {
1448                 if (o->bw_log_file) {
1449                         finish_log_named(td, td->bw_log,
1450                                                 o->bw_log_file, "bw");
1451                 } else
1452                         finish_log(td, td->bw_log, "bw");
1453         }
1454         if (td->lat_log) {
1455                 if (o->lat_log_file) {
1456                         finish_log_named(td, td->lat_log,
1457                                                 o->lat_log_file, "lat");
1458                 } else
1459                         finish_log(td, td->lat_log, "lat");
1460         }
1461         if (td->slat_log) {
1462                 if (o->lat_log_file) {
1463                         finish_log_named(td, td->slat_log,
1464                                                 o->lat_log_file, "slat");
1465                 } else
1466                         finish_log(td, td->slat_log, "slat");
1467         }
1468         if (td->clat_log) {
1469                 if (o->lat_log_file) {
1470                         finish_log_named(td, td->clat_log,
1471                                                 o->lat_log_file, "clat");
1472                 } else
1473                         finish_log(td, td->clat_log, "clat");
1474         }
1475         if (td->iops_log) {
1476                 if (o->iops_log_file) {
1477                         finish_log_named(td, td->iops_log,
1478                                                 o->iops_log_file, "iops");
1479                 } else
1480                         finish_log(td, td->iops_log, "iops");
1481         }
1482
1483         fio_mutex_up(writeout_mutex);
1484         if (o->exec_postrun)
1485                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1486
1487         if (exitall_on_terminate)
1488                 fio_terminate_threads(td->groupid);
1489
1490 err:
1491         if (td->error)
1492                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1493                                                         td->verror);
1494
1495         if (o->verify_async)
1496                 verify_async_exit(td);
1497
1498         close_and_free_files(td);
1499         cleanup_io_u(td);
1500         close_ioengine(td);
1501         cgroup_shutdown(td, &cgroup_mnt);
1502
1503         if (o->cpumask_set) {
1504                 int ret = fio_cpuset_exit(&o->cpumask);
1505
1506                 td_verror(td, ret, "fio_cpuset_exit");
1507         }
1508
1509         /*
1510          * do this very late, it will log file closing as well
1511          */
1512         if (o->write_iolog_file)
1513                 write_iolog_close(td);
1514
1515         fio_mutex_remove(td->rusage_sem);
1516         td->rusage_sem = NULL;
1517
1518         td_set_runstate(td, TD_EXITED);
1519         return (void *) (uintptr_t) td->error;
1520 }
1521
1522
1523 /*
1524  * We cannot pass the td data into a forked process, so attach the td and
1525  * pass it to the thread worker.
1526  */
1527 static int fork_main(int shmid, int offset)
1528 {
1529         struct thread_data *td;
1530         void *data, *ret;
1531
1532 #ifndef __hpux
1533         data = shmat(shmid, NULL, 0);
1534         if (data == (void *) -1) {
1535                 int __err = errno;
1536
1537                 perror("shmat");
1538                 return __err;
1539         }
1540 #else
1541         /*
1542          * HP-UX inherits shm mappings?
1543          */
1544         data = threads;
1545 #endif
1546
1547         td = data + offset * sizeof(struct thread_data);
1548         ret = thread_main(td);
1549         shmdt(data);
1550         return (int) (uintptr_t) ret;
1551 }
1552
1553 /*
1554  * Run over the job map and reap the threads that have exited, if any.
1555  */
1556 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1557                          unsigned int *m_rate)
1558 {
1559         struct thread_data *td;
1560         unsigned int cputhreads, realthreads, pending;
1561         int i, status, ret;
1562
1563         /*
1564          * reap exited threads (TD_EXITED -> TD_REAPED)
1565          */
1566         realthreads = pending = cputhreads = 0;
1567         for_each_td(td, i) {
1568                 int flags = 0;
1569
1570                 /*
1571                  * ->io_ops is NULL for a thread that has closed its
1572                  * io engine
1573                  */
1574                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1575                         cputhreads++;
1576                 else
1577                         realthreads++;
1578
1579                 if (!td->pid) {
1580                         pending++;
1581                         continue;
1582                 }
1583                 if (td->runstate == TD_REAPED)
1584                         continue;
1585                 if (td->o.use_thread) {
1586                         if (td->runstate == TD_EXITED) {
1587                                 td_set_runstate(td, TD_REAPED);
1588                                 goto reaped;
1589                         }
1590                         continue;
1591                 }
1592
1593                 flags = WNOHANG;
1594                 if (td->runstate == TD_EXITED)
1595                         flags = 0;
1596
1597                 /*
1598                  * check if someone quit or got killed in an unusual way
1599                  */
1600                 ret = waitpid(td->pid, &status, flags);
1601                 if (ret < 0) {
1602                         if (errno == ECHILD) {
1603                                 log_err("fio: pid=%d disappeared %d\n",
1604                                                 (int) td->pid, td->runstate);
1605                                 td->sig = ECHILD;
1606                                 td_set_runstate(td, TD_REAPED);
1607                                 goto reaped;
1608                         }
1609                         perror("waitpid");
1610                 } else if (ret == td->pid) {
1611                         if (WIFSIGNALED(status)) {
1612                                 int sig = WTERMSIG(status);
1613
1614                                 if (sig != SIGTERM && sig != SIGUSR2)
1615                                         log_err("fio: pid=%d, got signal=%d\n",
1616                                                         (int) td->pid, sig);
1617                                 td->sig = sig;
1618                                 td_set_runstate(td, TD_REAPED);
1619                                 goto reaped;
1620                         }
1621                         if (WIFEXITED(status)) {
1622                                 if (WEXITSTATUS(status) && !td->error)
1623                                         td->error = WEXITSTATUS(status);
1624
1625                                 td_set_runstate(td, TD_REAPED);
1626                                 goto reaped;
1627                         }
1628                 }
1629
1630                 /*
1631                  * thread is not dead, continue
1632                  */
1633                 pending++;
1634                 continue;
1635 reaped:
1636                 (*nr_running)--;
1637                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1638                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1639                 if (!td->pid)
1640                         pending--;
1641
1642                 if (td->error)
1643                         exit_value++;
1644
1645                 done_secs += mtime_since_now(&td->epoch) / 1000;
1646                 profile_td_exit(td);
1647         }
1648
1649         if (*nr_running == cputhreads && !pending && realthreads)
1650                 fio_terminate_threads(TERMINATE_ALL);
1651 }
1652
1653 static void do_usleep(unsigned int usecs)
1654 {
1655         check_for_running_stats();
1656         usleep(usecs);
1657 }
1658
1659 /*
1660  * Main function for kicking off and reaping jobs, as needed.
1661  */
1662 static void run_threads(void)
1663 {
1664         struct thread_data *td;
1665         unsigned long spent;
1666         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1667
1668         if (fio_gtod_offload && fio_start_gtod_thread())
1669                 return;
1670
1671         fio_idle_prof_init();
1672
1673         set_sig_handlers();
1674
1675         nr_thread = nr_process = 0;
1676         for_each_td(td, i) {
1677                 if (td->o.use_thread)
1678                         nr_thread++;
1679                 else
1680                         nr_process++;
1681         }
1682
1683         if (output_format == FIO_OUTPUT_NORMAL) {
1684                 log_info("Starting ");
1685                 if (nr_thread)
1686                         log_info("%d thread%s", nr_thread,
1687                                                 nr_thread > 1 ? "s" : "");
1688                 if (nr_process) {
1689                         if (nr_thread)
1690                                 log_info(" and ");
1691                         log_info("%d process%s", nr_process,
1692                                                 nr_process > 1 ? "es" : "");
1693                 }
1694                 log_info("\n");
1695                 fflush(stdout);
1696         }
1697
1698         todo = thread_number;
1699         nr_running = 0;
1700         nr_started = 0;
1701         m_rate = t_rate = 0;
1702
1703         for_each_td(td, i) {
1704                 print_status_init(td->thread_number - 1);
1705
1706                 if (!td->o.create_serialize)
1707                         continue;
1708
1709                 /*
1710                  * do file setup here so it happens sequentially,
1711                  * we don't want X number of threads getting their
1712                  * client data interspersed on disk
1713                  */
1714                 if (setup_files(td)) {
1715                         exit_value++;
1716                         if (td->error)
1717                                 log_err("fio: pid=%d, err=%d/%s\n",
1718                                         (int) td->pid, td->error, td->verror);
1719                         td_set_runstate(td, TD_REAPED);
1720                         todo--;
1721                 } else {
1722                         struct fio_file *f;
1723                         unsigned int j;
1724
1725                         /*
1726                          * for sharing to work, each job must always open
1727                          * its own files. so close them, if we opened them
1728                          * for creation
1729                          */
1730                         for_each_file(td, f, j) {
1731                                 if (fio_file_open(f))
1732                                         td_io_close_file(td, f);
1733                         }
1734                 }
1735         }
1736
1737         /* start idle threads before io threads start to run */
1738         fio_idle_prof_start();
1739
1740         set_genesis_time();
1741
1742         while (todo) {
1743                 struct thread_data *map[REAL_MAX_JOBS];
1744                 struct timeval this_start;
1745                 int this_jobs = 0, left;
1746
1747                 /*
1748                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1749                  */
1750                 for_each_td(td, i) {
1751                         if (td->runstate != TD_NOT_CREATED)
1752                                 continue;
1753
1754                         /*
1755                          * never got a chance to start, killed by other
1756                          * thread for some reason
1757                          */
1758                         if (td->terminate) {
1759                                 todo--;
1760                                 continue;
1761                         }
1762
1763                         if (td->o.start_delay) {
1764                                 spent = mtime_since_genesis();
1765
1766                                 if (td->o.start_delay * 1000 > spent)
1767                                         continue;
1768                         }
1769
1770                         if (td->o.stonewall && (nr_started || nr_running)) {
1771                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1772                                                         td->o.name);
1773                                 break;
1774                         }
1775
1776                         init_disk_util(td);
1777
1778                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1779                         td->update_rusage = 0;
1780
1781                         /*
1782                          * Set state to created. Thread will transition
1783                          * to TD_INITIALIZED when it's done setting up.
1784                          */
1785                         td_set_runstate(td, TD_CREATED);
1786                         map[this_jobs++] = td;
1787                         nr_started++;
1788
1789                         if (td->o.use_thread) {
1790                                 int ret;
1791
1792                                 dprint(FD_PROCESS, "will pthread_create\n");
1793                                 ret = pthread_create(&td->thread, NULL,
1794                                                         thread_main, td);
1795                                 if (ret) {
1796                                         log_err("pthread_create: %s\n",
1797                                                         strerror(ret));
1798                                         nr_started--;
1799                                         break;
1800                                 }
1801                                 ret = pthread_detach(td->thread);
1802                                 if (ret)
1803                                         log_err("pthread_detach: %s",
1804                                                         strerror(ret));
1805                         } else {
1806                                 pid_t pid;
1807                                 dprint(FD_PROCESS, "will fork\n");
1808                                 pid = fork();
1809                                 if (!pid) {
1810                                         int ret = fork_main(shm_id, i);
1811
1812                                         _exit(ret);
1813                                 } else if (i == fio_debug_jobno)
1814                                         *fio_debug_jobp = pid;
1815                         }
1816                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1817                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1818                                 log_err("fio: job startup hung? exiting.\n");
1819                                 fio_terminate_threads(TERMINATE_ALL);
1820                                 fio_abort = 1;
1821                                 nr_started--;
1822                                 break;
1823                         }
1824                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1825                 }
1826
1827                 /*
1828                  * Wait for the started threads to transition to
1829                  * TD_INITIALIZED.
1830                  */
1831                 fio_gettime(&this_start, NULL);
1832                 left = this_jobs;
1833                 while (left && !fio_abort) {
1834                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1835                                 break;
1836
1837                         do_usleep(100000);
1838
1839                         for (i = 0; i < this_jobs; i++) {
1840                                 td = map[i];
1841                                 if (!td)
1842                                         continue;
1843                                 if (td->runstate == TD_INITIALIZED) {
1844                                         map[i] = NULL;
1845                                         left--;
1846                                 } else if (td->runstate >= TD_EXITED) {
1847                                         map[i] = NULL;
1848                                         left--;
1849                                         todo--;
1850                                         nr_running++; /* work-around... */
1851                                 }
1852                         }
1853                 }
1854
1855                 if (left) {
1856                         log_err("fio: %d job%s failed to start\n", left,
1857                                         left > 1 ? "s" : "");
1858                         for (i = 0; i < this_jobs; i++) {
1859                                 td = map[i];
1860                                 if (!td)
1861                                         continue;
1862                                 kill(td->pid, SIGTERM);
1863                         }
1864                         break;
1865                 }
1866
1867                 /*
1868                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1869                  */
1870                 for_each_td(td, i) {
1871                         if (td->runstate != TD_INITIALIZED)
1872                                 continue;
1873
1874                         if (in_ramp_time(td))
1875                                 td_set_runstate(td, TD_RAMP);
1876                         else
1877                                 td_set_runstate(td, TD_RUNNING);
1878                         nr_running++;
1879                         nr_started--;
1880                         m_rate += ddir_rw_sum(td->o.ratemin);
1881                         t_rate += ddir_rw_sum(td->o.rate);
1882                         todo--;
1883                         fio_mutex_up(td->mutex);
1884                 }
1885
1886                 reap_threads(&nr_running, &t_rate, &m_rate);
1887
1888                 if (todo)
1889                         do_usleep(100000);
1890         }
1891
1892         while (nr_running) {
1893                 reap_threads(&nr_running, &t_rate, &m_rate);
1894                 do_usleep(10000);
1895         }
1896
1897         fio_idle_prof_stop();
1898
1899         update_io_ticks();
1900 }
1901
1902 void wait_for_disk_thread_exit(void)
1903 {
1904         fio_mutex_down(disk_thread_mutex);
1905 }
1906
1907 static void free_disk_util(void)
1908 {
1909         disk_util_start_exit();
1910         wait_for_disk_thread_exit();
1911         disk_util_prune_entries();
1912 }
1913
1914 static void *disk_thread_main(void *data)
1915 {
1916         int ret = 0;
1917
1918         fio_mutex_up(startup_mutex);
1919
1920         while (threads && !ret) {
1921                 usleep(DISK_UTIL_MSEC * 1000);
1922                 if (!threads)
1923                         break;
1924                 ret = update_io_ticks();
1925
1926                 if (!is_backend)
1927                         print_thread_status();
1928         }
1929
1930         fio_mutex_up(disk_thread_mutex);
1931         return NULL;
1932 }
1933
1934 static int create_disk_util_thread(void)
1935 {
1936         int ret;
1937
1938         setup_disk_util();
1939
1940         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1941
1942         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1943         if (ret) {
1944                 fio_mutex_remove(disk_thread_mutex);
1945                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1946                 return 1;
1947         }
1948
1949         ret = pthread_detach(disk_util_thread);
1950         if (ret) {
1951                 fio_mutex_remove(disk_thread_mutex);
1952                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1953                 return 1;
1954         }
1955
1956         dprint(FD_MUTEX, "wait on startup_mutex\n");
1957         fio_mutex_down(startup_mutex);
1958         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1959         return 0;
1960 }
1961
1962 int fio_backend(void)
1963 {
1964         struct thread_data *td;
1965         int i;
1966
1967         if (exec_profile) {
1968                 if (load_profile(exec_profile))
1969                         return 1;
1970                 free(exec_profile);
1971                 exec_profile = NULL;
1972         }
1973         if (!thread_number)
1974                 return 0;
1975
1976         if (write_bw_log) {
1977                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1978                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1979                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1980         }
1981
1982         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1983         if (startup_mutex == NULL)
1984                 return 1;
1985         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1986         if (writeout_mutex == NULL)
1987                 return 1;
1988
1989         set_genesis_time();
1990         stat_init();
1991         create_disk_util_thread();
1992
1993         cgroup_list = smalloc(sizeof(*cgroup_list));
1994         INIT_FLIST_HEAD(cgroup_list);
1995
1996         run_threads();
1997
1998         if (!fio_abort) {
1999                 show_run_stats();
2000                 if (write_bw_log) {
2001                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2002                         __finish_log(agg_io_log[DDIR_WRITE],
2003                                         "agg-write_bw.log");
2004                         __finish_log(agg_io_log[DDIR_TRIM],
2005                                         "agg-write_bw.log");
2006                 }
2007         }
2008
2009         for_each_td(td, i)
2010                 fio_options_free(td);
2011
2012         free_disk_util();
2013         cgroup_kill(cgroup_list);
2014         sfree(cgroup_list);
2015         sfree(cgroup_mnt);
2016
2017         fio_mutex_remove(startup_mutex);
2018         fio_mutex_remove(writeout_mutex);
2019         fio_mutex_remove(disk_thread_mutex);
2020         stat_exit();
2021         return exit_value;
2022 }