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         /*
 655          * If verify_backlog is enabled, we'll run the verify in this
 656          * handler as well. For that case, we may need up to twice the
 657          * amount of bytes.
 658          */
 659         total_bytes = td->o.size;
 660         if (td->o.verify != VERIFY_NONE &&
 661            (td_write(td) && td->o.verify_backlog))
 662                 total_bytes += td->o.size;
 663
 664         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 665                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 666                 td->o.time_based) {
 667                 struct timeval comp_time;
 668                 int min_evts = 0;
 669                 struct io_u *io_u;
 670                 int ret2, full;
 671                 enum fio_ddir ddir;
 672
 673                 check_update_rusage(td);
 674
 675                 if (td->terminate || td->done)
 676                         break;
 677
 678                 update_tv_cache(td);
 679
 680                 if (runtime_exceeded(td, &td->tv_cache)) {
 681                         __update_tv_cache(td);
 682                         if (runtime_exceeded(td, &td->tv_cache)) {
 683                                 td->terminate = 1;
 684                                 break;
 685                         }
 686                 }
 687
 688                 if (flow_threshold_exceeded(td))
 689                         continue;
 690
 691                 if (bytes_issued >= total_bytes)
 692                         break;
 693
 694                 io_u = get_io_u(td);
 695                 if (!io_u) {
 696                         if (td->o.latency_target)
 697                                 goto reap;
 698                         break;
 699                 }
 700
 701                 ddir = io_u->ddir;
 702
 703                 /*
 704                  * Add verification end_io handler if:
 705                  *      - Asked to verify (!td_rw(td))
 706                  *      - Or the io_u is from our verify list (mixed write/ver)
 707                  */
 708                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 709                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 710
 711                         if (!td->o.verify_pattern_bytes) {
 712                                 io_u->rand_seed = __rand(&td->__verify_state);
 713                                 if (sizeof(int) != sizeof(long *))
 714                                         io_u->rand_seed *= __rand(&td->__verify_state);
 715                         }
 716
 717                         if (td->o.verify_async)
 718                                 io_u->end_io = verify_io_u_async;
 719                         else
 720                                 io_u->end_io = verify_io_u;
 721                         td_set_runstate(td, TD_VERIFYING);
 722                 } else if (in_ramp_time(td))
 723                         td_set_runstate(td, TD_RAMP);
 724                 else
 725                         td_set_runstate(td, TD_RUNNING);
 726
 727                 /*
 728                  * Always log IO before it's issued, so we know the specific
 729                  * order of it. The logged unit will track when the IO has
 730                  * completed.
 731                  */
 732                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 733                     td->o.do_verify &&
 734                     td->o.verify != VERIFY_NONE &&
 735                     !td->o.experimental_verify)
 736                         log_io_piece(td, io_u);
 737
 738                 ret = td_io_queue(td, io_u);
 739                 switch (ret) {
 740                 case FIO_Q_COMPLETED:
 741                         if (io_u->error) {
 742                                 ret = -io_u->error;
 743                                 clear_io_u(td, io_u);
 744                         } else if (io_u->resid) {
 745                                 int bytes = io_u->xfer_buflen - io_u->resid;
 746                                 struct fio_file *f = io_u->file;
 747
 748                                 bytes_issued += bytes;
 749                                 /*
 750                                  * zero read, fail
 751                                  */
 752                                 if (!bytes) {
 753                                         td_verror(td, EIO, "full resid");
 754                                         put_io_u(td, io_u);
 755                                         break;
 756                                 }
 757
 758                                 io_u->xfer_buflen = io_u->resid;
 759                                 io_u->xfer_buf += bytes;
 760                                 io_u->offset += bytes;
 761
 762                                 if (ddir_rw(io_u->ddir))
 763                                         td->ts.short_io_u[io_u->ddir]++;
 764
 765                                 if (io_u->offset == f->real_file_size)
 766                                         goto sync_done;
 767
 768                                 requeue_io_u(td, &io_u);
 769                         } else {
 770 sync_done:
 771                                 if (__should_check_rate(td, DDIR_READ) ||
 772                                     __should_check_rate(td, DDIR_WRITE) ||
 773                                     __should_check_rate(td, DDIR_TRIM))
 774                                         fio_gettime(&comp_time, NULL);
 775
 776                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 777                                 if (ret < 0)
 778                                         break;
 779                                 bytes_issued += io_u->xfer_buflen;
 780                         }
 781                         break;
 782                 case FIO_Q_QUEUED:
 783                         /*
 784                          * if the engine doesn't have a commit hook,
 785                          * the io_u is really queued. if it does have such
 786                          * a hook, it has to call io_u_queued() itself.
 787                          */
 788                         if (td->io_ops->commit == NULL)
 789                                 io_u_queued(td, io_u);
 790                         bytes_issued += io_u->xfer_buflen;
 791                         break;
 792                 case FIO_Q_BUSY:
 793                         requeue_io_u(td, &io_u);
 794                         ret2 = td_io_commit(td);
 795                         if (ret2 < 0)
 796                                 ret = ret2;
 797                         break;
 798                 default:
 799                         assert(ret < 0);
 800                         put_io_u(td, io_u);
 801                         break;
 802                 }
 803
 804                 if (break_on_this_error(td, ddir, &ret))
 805                         break;
 806
 807                 /*
 808                  * See if we need to complete some commands. Note that we
 809                  * can get BUSY even without IO queued, if the system is
 810                  * resource starved.
 811                  */
 812 reap:
 813                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 814                 if (full || !td->o.iodepth_batch_complete) {
 815                         min_evts = min(td->o.iodepth_batch_complete,
 816                                         td->cur_depth);
 817                         /*
 818                          * if the queue is full, we MUST reap at least 1 event
 819                          */
 820                         if (full && !min_evts)
 821                                 min_evts = 1;
 822
 823                         if (__should_check_rate(td, DDIR_READ) ||
 824                             __should_check_rate(td, DDIR_WRITE) ||
 825                             __should_check_rate(td, DDIR_TRIM))
 826                                 fio_gettime(&comp_time, NULL);
 827
 828                         do {
 829                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 830                                 if (ret < 0)
 831                                         break;
 832
 833                         } while (full && (td->cur_depth > td->o.iodepth_low));
 834                 }
 835
 836                 if (ret < 0)
 837                         break;
 838                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 839                         continue;
 840
 841                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 842                         if (check_min_rate(td, &comp_time, bytes_done)) {
 843                                 if (exitall_on_terminate)
 844                                         fio_terminate_threads(td->groupid);
 845                                 td_verror(td, EIO, "check_min_rate");
 846                                 break;
 847                         }
 848                 }
 849                 if (!in_ramp_time(td) && td->o.latency_target)
 850                         lat_target_check(td);
 851
 852                 if (td->o.thinktime) {
 853                         unsigned long long b;
 854
 855                         b = ddir_rw_sum(td->io_blocks);
 856                         if (!(b % td->o.thinktime_blocks)) {
 857                                 int left;
 858
 859                                 io_u_quiesce(td);
 860
 861                                 if (td->o.thinktime_spin)
 862                                         usec_spin(td->o.thinktime_spin);
 863
 864                                 left = td->o.thinktime - td->o.thinktime_spin;
 865                                 if (left)
 866                                         usec_sleep(td, left);
 867                         }
 868                 }
 869         }
 870
 871         check_update_rusage(td);
 872
 873         if (td->trim_entries)
 874                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 875
 876         if (td->o.fill_device && td->error == ENOSPC) {
 877                 td->error = 0;
 878                 td->terminate = 1;
 879         }
 880         if (!td->error) {
 881                 struct fio_file *f;
 882
 883                 i = td->cur_depth;
 884                 if (i) {
 885                         ret = io_u_queued_complete(td, i, bytes_done);
 886                         if (td->o.fill_device && td->error == ENOSPC)
 887                                 td->error = 0;
 888                 }
 889
 890                 if (should_fsync(td) && td->o.end_fsync) {
 891                         td_set_runstate(td, TD_FSYNCING);
 892
 893                         for_each_file(td, f, i) {
 894                                 if (!fio_file_fsync(td, f))
 895                                         continue;
 896
 897                                 log_err("fio: end_fsync failed for file %s\n",
 898                                                                 f->file_name);
 899                         }
 900                 }
 901         } else
 902                 cleanup_pending_aio(td);
 903
 904         /*
 905          * stop job if we failed doing any IO
 906          */
 907         if (!ddir_rw_sum(td->this_io_bytes))
 908                 td->done = 1;
 909
 910         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 911 }
 912
 913 static void cleanup_io_u(struct thread_data *td)
 914 {
 915         struct io_u *io_u;
 916
 917         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 918
 919                 if (td->io_ops->io_u_free)
 920                         td->io_ops->io_u_free(td, io_u);
 921
 922                 fio_memfree(io_u, sizeof(*io_u));
 923         }
 924
 925         free_io_mem(td);
 926
 927         io_u_rexit(&td->io_u_requeues);
 928         io_u_qexit(&td->io_u_freelist);
 929         io_u_qexit(&td->io_u_all);
 930 }
 931
 932 static int init_io_u(struct thread_data *td)
 933 {
 934         struct io_u *io_u;
 935         unsigned int max_bs, min_write;
 936         int cl_align, i, max_units;
 937         int data_xfer = 1, err;
 938         char *p;
 939
 940         max_units = td->o.iodepth;
 941         max_bs = td_max_bs(td);
 942         min_write = td->o.min_bs[DDIR_WRITE];
 943         td->orig_buffer_size = (unsigned long long) max_bs
 944                                         * (unsigned long long) max_units;
 945
 946         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 947                 data_xfer = 0;
 948
 949         err = 0;
 950         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 951         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 952         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 953
 954         if (err) {
 955                 log_err("fio: failed setting up IO queues\n");
 956                 return 1;
 957         }
 958
 959         /*
 960          * if we may later need to do address alignment, then add any
 961          * possible adjustment here so that we don't cause a buffer
 962          * overflow later. this adjustment may be too much if we get
 963          * lucky and the allocator gives us an aligned address.
 964          */
 965         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 966             (td->io_ops->flags & FIO_RAWIO))
 967                 td->orig_buffer_size += page_mask + td->o.mem_align;
 968
 969         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 970                 unsigned long bs;
 971
 972                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 973                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 974         }
 975
 976         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 977                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 978                 return 1;
 979         }
 980
 981         if (data_xfer && allocate_io_mem(td))
 982                 return 1;
 983
 984         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 985             (td->io_ops->flags & FIO_RAWIO))
 986                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 987         else
 988                 p = td->orig_buffer;
 989
 990         cl_align = os_cache_line_size();
 991
 992         for (i = 0; i < max_units; i++) {
 993                 void *ptr;
 994
 995                 if (td->terminate)
 996                         return 1;
 997
 998                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 999                 if (!ptr) {
1000                         log_err("fio: unable to allocate aligned memory\n");
1001                         break;
1002                 }
1003
1004                 io_u = ptr;
1005                 memset(io_u, 0, sizeof(*io_u));
1006                 INIT_FLIST_HEAD(&io_u->verify_list);
1007                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1008
1009                 if (data_xfer) {
1010                         io_u->buf = p;
1011                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1012
1013                         if (td_write(td))
1014                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1015                         if (td_write(td) && td->o.verify_pattern_bytes) {
1016                                 /*
1017                                  * Fill the buffer with the pattern if we are
1018                                  * going to be doing writes.
1019                                  */
1020                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1021                         }
1022                 }
1023
1024                 io_u->index = i;
1025                 io_u->flags = IO_U_F_FREE;
1026                 io_u_qpush(&td->io_u_freelist, io_u);
1027
1028                 /*
1029                  * io_u never leaves this stack, used for iteration of all
1030                  * io_u buffers.
1031                  */
1032                 io_u_qpush(&td->io_u_all, io_u);
1033
1034                 if (td->io_ops->io_u_init) {
1035                         int ret = td->io_ops->io_u_init(td, io_u);
1036
1037                         if (ret) {
1038                                 log_err("fio: failed to init engine data: %d\n", ret);
1039                                 return 1;
1040                         }
1041                 }
1042
1043                 p += max_bs;
1044         }
1045
1046         return 0;
1047 }
1048
1049 static int switch_ioscheduler(struct thread_data *td)
1050 {
1051         char tmp[256], tmp2[128];
1052         FILE *f;
1053         int ret;
1054
1055         if (td->io_ops->flags & FIO_DISKLESSIO)
1056                 return 0;
1057
1058         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1059
1060         f = fopen(tmp, "r+");
1061         if (!f) {
1062                 if (errno == ENOENT) {
1063                         log_err("fio: os or kernel doesn't support IO scheduler"
1064                                 " switching\n");
1065                         return 0;
1066                 }
1067                 td_verror(td, errno, "fopen iosched");
1068                 return 1;
1069         }
1070
1071         /*
1072          * Set io scheduler.
1073          */
1074         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1075         if (ferror(f) || ret != 1) {
1076                 td_verror(td, errno, "fwrite");
1077                 fclose(f);
1078                 return 1;
1079         }
1080
1081         rewind(f);
1082
1083         /*
1084          * Read back and check that the selected scheduler is now the default.
1085          */
1086         ret = fread(tmp, 1, sizeof(tmp), f);
1087         if (ferror(f) || ret < 0) {
1088                 td_verror(td, errno, "fread");
1089                 fclose(f);
1090                 return 1;
1091         }
1092
1093         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1094         if (!strstr(tmp, tmp2)) {
1095                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1096                 td_verror(td, EINVAL, "iosched_switch");
1097                 fclose(f);
1098                 return 1;
1099         }
1100
1101         fclose(f);
1102         return 0;
1103 }
1104
1105 static int keep_running(struct thread_data *td)
1106 {
1107         if (td->done)
1108                 return 0;
1109         if (td->o.time_based)
1110                 return 1;
1111         if (td->o.loops) {
1112                 td->o.loops--;
1113                 return 1;
1114         }
1115
1116         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1117                 uint64_t diff;
1118
1119                 /*
1120                  * If the difference is less than the minimum IO size, we
1121                  * are done.
1122                  */
1123                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1124                 if (diff < td_max_bs(td))
1125                         return 0;
1126
1127                 return 1;
1128         }
1129
1130         return 0;
1131 }
1132
1133 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1134 {
1135         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1136         char *str;
1137
1138         str = malloc(newlen);
1139         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1140
1141         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1142         ret = system(str);
1143         if (ret == -1)
1144                 log_err("fio: exec of cmd <%s> failed\n", str);
1145
1146         free(str);
1147         return ret;
1148 }
1149
1150 /*
1151  * Dry run to compute correct state of numberio for verification.
1152  */
1153 static uint64_t do_dry_run(struct thread_data *td)
1154 {
1155         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1156
1157         td_set_runstate(td, TD_RUNNING);
1158
1159         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1160                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1161                 struct io_u *io_u;
1162                 int ret;
1163
1164                 if (td->terminate || td->done)
1165                         break;
1166
1167                 io_u = get_io_u(td);
1168                 if (!io_u)
1169                         break;
1170
1171                 io_u->flags |= IO_U_F_FLIGHT;
1172                 io_u->error = 0;
1173                 io_u->resid = 0;
1174                 if (ddir_rw(acct_ddir(io_u)))
1175                         td->io_issues[acct_ddir(io_u)]++;
1176                 if (ddir_rw(io_u->ddir)) {
1177                         io_u_mark_depth(td, 1);
1178                         td->ts.total_io_u[io_u->ddir]++;
1179                 }
1180
1181                 ret = io_u_sync_complete(td, io_u, bytes_done);
1182                 (void) ret;
1183         }
1184
1185         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1186 }
1187
1188 /*
1189  * Entry point for the thread based jobs. The process based jobs end up
1190  * here as well, after a little setup.
1191  */
1192 static void *thread_main(void *data)
1193 {
1194         unsigned long long elapsed;
1195         struct thread_data *td = data;
1196         struct thread_options *o = &td->o;
1197         pthread_condattr_t attr;
1198         int clear_state;
1199         int ret;
1200
1201         if (!o->use_thread) {
1202                 setsid();
1203                 td->pid = getpid();
1204         } else
1205                 td->pid = gettid();
1206
1207         /*
1208          * fio_time_init() may not have been called yet if running as a server
1209          */
1210         fio_time_init();
1211
1212         fio_local_clock_init(o->use_thread);
1213
1214         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1215
1216         if (is_backend)
1217                 fio_server_send_start(td);
1218
1219         INIT_FLIST_HEAD(&td->io_log_list);
1220         INIT_FLIST_HEAD(&td->io_hist_list);
1221         INIT_FLIST_HEAD(&td->verify_list);
1222         INIT_FLIST_HEAD(&td->trim_list);
1223         INIT_FLIST_HEAD(&td->next_rand_list);
1224         pthread_mutex_init(&td->io_u_lock, NULL);
1225         td->io_hist_tree = RB_ROOT;
1226
1227         pthread_condattr_init(&attr);
1228         pthread_cond_init(&td->verify_cond, &attr);
1229         pthread_cond_init(&td->free_cond, &attr);
1230
1231         td_set_runstate(td, TD_INITIALIZED);
1232         dprint(FD_MUTEX, "up startup_mutex\n");
1233         fio_mutex_up(startup_mutex);
1234         dprint(FD_MUTEX, "wait on td->mutex\n");
1235         fio_mutex_down(td->mutex);
1236         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1237
1238         /*
1239          * the ->mutex mutex is now no longer used, close it to avoid
1240          * eating a file descriptor
1241          */
1242         fio_mutex_remove(td->mutex);
1243         td->mutex = NULL;
1244
1245         /*
1246          * A new gid requires privilege, so we need to do this before setting
1247          * the uid.
1248          */
1249         if (o->gid != -1U && setgid(o->gid)) {
1250                 td_verror(td, errno, "setgid");
1251                 goto err;
1252         }
1253         if (o->uid != -1U && setuid(o->uid)) {
1254                 td_verror(td, errno, "setuid");
1255                 goto err;
1256         }
1257
1258         /*
1259          * If we have a gettimeofday() thread, make sure we exclude that
1260          * thread from this job
1261          */
1262         if (o->gtod_cpu)
1263                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1264
1265         /*
1266          * Set affinity first, in case it has an impact on the memory
1267          * allocations.
1268          */
1269         if (o->cpumask_set) {
1270                 ret = fio_setaffinity(td->pid, o->cpumask);
1271                 if (ret == -1) {
1272                         td_verror(td, errno, "cpu_set_affinity");
1273                         goto err;
1274                 }
1275         }
1276
1277 #ifdef CONFIG_LIBNUMA
1278         /* numa node setup */
1279         if (o->numa_cpumask_set || o->numa_memmask_set) {
1280                 int ret;
1281
1282                 if (numa_available() < 0) {
1283                         td_verror(td, errno, "Does not support NUMA API\n");
1284                         goto err;
1285                 }
1286
1287                 if (o->numa_cpumask_set) {
1288                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1289                         if (ret == -1) {
1290                                 td_verror(td, errno, \
1291                                         "numa_run_on_node_mask failed\n");
1292                                 goto err;
1293                         }
1294                 }
1295
1296                 if (o->numa_memmask_set) {
1297
1298                         switch (o->numa_mem_mode) {
1299                         case MPOL_INTERLEAVE:
1300                                 numa_set_interleave_mask(o->numa_memnodesmask);
1301                                 break;
1302                         case MPOL_BIND:
1303                                 numa_set_membind(o->numa_memnodesmask);
1304                                 break;
1305                         case MPOL_LOCAL:
1306                                 numa_set_localalloc();
1307                                 break;
1308                         case MPOL_PREFERRED:
1309                                 numa_set_preferred(o->numa_mem_prefer_node);
1310                                 break;
1311                         case MPOL_DEFAULT:
1312                         default:
1313                                 break;
1314                         }
1315
1316                 }
1317         }
1318 #endif
1319
1320         if (fio_pin_memory(td))
1321                 goto err;
1322
1323         /*
1324          * May alter parameters that init_io_u() will use, so we need to
1325          * do this first.
1326          */
1327         if (init_iolog(td))
1328                 goto err;
1329
1330         if (init_io_u(td))
1331                 goto err;
1332
1333         if (o->verify_async && verify_async_init(td))
1334                 goto err;
1335
1336         if (o->ioprio) {
1337                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1338                 if (ret == -1) {
1339                         td_verror(td, errno, "ioprio_set");
1340                         goto err;
1341                 }
1342         }
1343
1344         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1345                 goto err;
1346
1347         errno = 0;
1348         if (nice(o->nice) == -1 && errno != 0) {
1349                 td_verror(td, errno, "nice");
1350                 goto err;
1351         }
1352
1353         if (o->ioscheduler && switch_ioscheduler(td))
1354                 goto err;
1355
1356         if (!o->create_serialize && setup_files(td))
1357                 goto err;
1358
1359         if (td_io_init(td))
1360                 goto err;
1361
1362         if (init_random_map(td))
1363                 goto err;
1364
1365         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1366                 goto err;
1367
1368         if (o->pre_read) {
1369                 if (pre_read_files(td) < 0)
1370                         goto err;
1371         }
1372
1373         fio_verify_init(td);
1374
1375         fio_gettime(&td->epoch, NULL);
1376         fio_getrusage(&td->ru_start);
1377         clear_state = 0;
1378         while (keep_running(td)) {
1379                 uint64_t verify_bytes;
1380
1381                 fio_gettime(&td->start, NULL);
1382                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1383                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1384                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1385
1386                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1387                                 o->ratemin[DDIR_TRIM]) {
1388                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1389                                                 sizeof(td->bw_sample_time));
1390                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1391                                                 sizeof(td->bw_sample_time));
1392                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1393                                                 sizeof(td->bw_sample_time));
1394                 }
1395
1396                 if (clear_state)
1397                         clear_io_state(td);
1398
1399                 prune_io_piece_log(td);
1400
1401                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1402                         verify_bytes = do_dry_run(td);
1403                 else
1404                         verify_bytes = do_io(td);
1405
1406                 clear_state = 1;
1407
1408                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1409                         elapsed = utime_since_now(&td->start);
1410                         td->ts.runtime[DDIR_READ] += elapsed;
1411                 }
1412                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1413                         elapsed = utime_since_now(&td->start);
1414                         td->ts.runtime[DDIR_WRITE] += elapsed;
1415                 }
1416                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1417                         elapsed = utime_since_now(&td->start);
1418                         td->ts.runtime[DDIR_TRIM] += elapsed;
1419                 }
1420
1421                 if (td->error || td->terminate)
1422                         break;
1423
1424                 if (!o->do_verify ||
1425                     o->verify == VERIFY_NONE ||
1426                     (td->io_ops->flags & FIO_UNIDIR))
1427                         continue;
1428
1429                 clear_io_state(td);
1430
1431                 fio_gettime(&td->start, NULL);
1432
1433                 do_verify(td, verify_bytes);
1434
1435                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1436
1437                 if (td->error || td->terminate)
1438                         break;
1439         }
1440
1441         update_rusage_stat(td);
1442         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1443         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1444         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1445         td->ts.total_run_time = mtime_since_now(&td->epoch);
1446         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1447         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1448         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1449
1450         fio_unpin_memory(td);
1451
1452         fio_mutex_down(writeout_mutex);
1453         if (td->bw_log) {
1454                 if (o->bw_log_file) {
1455                         finish_log_named(td, td->bw_log,
1456                                                 o->bw_log_file, "bw");
1457                 } else
1458                         finish_log(td, td->bw_log, "bw");
1459         }
1460         if (td->lat_log) {
1461                 if (o->lat_log_file) {
1462                         finish_log_named(td, td->lat_log,
1463                                                 o->lat_log_file, "lat");
1464                 } else
1465                         finish_log(td, td->lat_log, "lat");
1466         }
1467         if (td->slat_log) {
1468                 if (o->lat_log_file) {
1469                         finish_log_named(td, td->slat_log,
1470                                                 o->lat_log_file, "slat");
1471                 } else
1472                         finish_log(td, td->slat_log, "slat");
1473         }
1474         if (td->clat_log) {
1475                 if (o->lat_log_file) {
1476                         finish_log_named(td, td->clat_log,
1477                                                 o->lat_log_file, "clat");
1478                 } else
1479                         finish_log(td, td->clat_log, "clat");
1480         }
1481         if (td->iops_log) {
1482                 if (o->iops_log_file) {
1483                         finish_log_named(td, td->iops_log,
1484                                                 o->iops_log_file, "iops");
1485                 } else
1486                         finish_log(td, td->iops_log, "iops");
1487         }
1488
1489         fio_mutex_up(writeout_mutex);
1490         if (o->exec_postrun)
1491                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1492
1493         if (exitall_on_terminate)
1494                 fio_terminate_threads(td->groupid);
1495
1496 err:
1497         if (td->error)
1498                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1499                                                         td->verror);
1500
1501         if (o->verify_async)
1502                 verify_async_exit(td);
1503
1504         close_and_free_files(td);
1505         cleanup_io_u(td);
1506         close_ioengine(td);
1507         cgroup_shutdown(td, &cgroup_mnt);
1508
1509         if (o->cpumask_set) {
1510                 int ret = fio_cpuset_exit(&o->cpumask);
1511
1512                 td_verror(td, ret, "fio_cpuset_exit");
1513         }
1514
1515         /*
1516          * do this very late, it will log file closing as well
1517          */
1518         if (o->write_iolog_file)
1519                 write_iolog_close(td);
1520
1521         fio_mutex_remove(td->rusage_sem);
1522         td->rusage_sem = NULL;
1523
1524         td_set_runstate(td, TD_EXITED);
1525         return (void *) (uintptr_t) td->error;
1526 }
1527
1528
1529 /*
1530  * We cannot pass the td data into a forked process, so attach the td and
1531  * pass it to the thread worker.
1532  */
1533 static int fork_main(int shmid, int offset)
1534 {
1535         struct thread_data *td;
1536         void *data, *ret;
1537
1538 #ifndef __hpux
1539         data = shmat(shmid, NULL, 0);
1540         if (data == (void *) -1) {
1541                 int __err = errno;
1542
1543                 perror("shmat");
1544                 return __err;
1545         }
1546 #else
1547         /*
1548          * HP-UX inherits shm mappings?
1549          */
1550         data = threads;
1551 #endif
1552
1553         td = data + offset * sizeof(struct thread_data);
1554         ret = thread_main(td);
1555         shmdt(data);
1556         return (int) (uintptr_t) ret;
1557 }
1558
1559 /*
1560  * Run over the job map and reap the threads that have exited, if any.
1561  */
1562 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1563                          unsigned int *m_rate)
1564 {
1565         struct thread_data *td;
1566         unsigned int cputhreads, realthreads, pending;
1567         int i, status, ret;
1568
1569         /*
1570          * reap exited threads (TD_EXITED -> TD_REAPED)
1571          */
1572         realthreads = pending = cputhreads = 0;
1573         for_each_td(td, i) {
1574                 int flags = 0;
1575
1576                 /*
1577                  * ->io_ops is NULL for a thread that has closed its
1578                  * io engine
1579                  */
1580                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1581                         cputhreads++;
1582                 else
1583                         realthreads++;
1584
1585                 if (!td->pid) {
1586                         pending++;
1587                         continue;
1588                 }
1589                 if (td->runstate == TD_REAPED)
1590                         continue;
1591                 if (td->o.use_thread) {
1592                         if (td->runstate == TD_EXITED) {
1593                                 td_set_runstate(td, TD_REAPED);
1594                                 goto reaped;
1595                         }
1596                         continue;
1597                 }
1598
1599                 flags = WNOHANG;
1600                 if (td->runstate == TD_EXITED)
1601                         flags = 0;
1602
1603                 /*
1604                  * check if someone quit or got killed in an unusual way
1605                  */
1606                 ret = waitpid(td->pid, &status, flags);
1607                 if (ret < 0) {
1608                         if (errno == ECHILD) {
1609                                 log_err("fio: pid=%d disappeared %d\n",
1610                                                 (int) td->pid, td->runstate);
1611                                 td->sig = ECHILD;
1612                                 td_set_runstate(td, TD_REAPED);
1613                                 goto reaped;
1614                         }
1615                         perror("waitpid");
1616                 } else if (ret == td->pid) {
1617                         if (WIFSIGNALED(status)) {
1618                                 int sig = WTERMSIG(status);
1619
1620                                 if (sig != SIGTERM && sig != SIGUSR2)
1621                                         log_err("fio: pid=%d, got signal=%d\n",
1622                                                         (int) td->pid, sig);
1623                                 td->sig = sig;
1624                                 td_set_runstate(td, TD_REAPED);
1625                                 goto reaped;
1626                         }
1627                         if (WIFEXITED(status)) {
1628                                 if (WEXITSTATUS(status) && !td->error)
1629                                         td->error = WEXITSTATUS(status);
1630
1631                                 td_set_runstate(td, TD_REAPED);
1632                                 goto reaped;
1633                         }
1634                 }
1635
1636                 /*
1637                  * thread is not dead, continue
1638                  */
1639                 pending++;
1640                 continue;
1641 reaped:
1642                 (*nr_running)--;
1643                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1644                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1645                 if (!td->pid)
1646                         pending--;
1647
1648                 if (td->error)
1649                         exit_value++;
1650
1651                 done_secs += mtime_since_now(&td->epoch) / 1000;
1652                 profile_td_exit(td);
1653         }
1654
1655         if (*nr_running == cputhreads && !pending && realthreads)
1656                 fio_terminate_threads(TERMINATE_ALL);
1657 }
1658
1659 static void do_usleep(unsigned int usecs)
1660 {
1661         check_for_running_stats();
1662         usleep(usecs);
1663 }
1664
1665 /*
1666  * Main function for kicking off and reaping jobs, as needed.
1667  */
1668 static void run_threads(void)
1669 {
1670         struct thread_data *td;
1671         unsigned long spent;
1672         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1673
1674         if (fio_gtod_offload && fio_start_gtod_thread())
1675                 return;
1676
1677         fio_idle_prof_init();
1678
1679         set_sig_handlers();
1680
1681         nr_thread = nr_process = 0;
1682         for_each_td(td, i) {
1683                 if (td->o.use_thread)
1684                         nr_thread++;
1685                 else
1686                         nr_process++;
1687         }
1688
1689         if (output_format == FIO_OUTPUT_NORMAL) {
1690                 log_info("Starting ");
1691                 if (nr_thread)
1692                         log_info("%d thread%s", nr_thread,
1693                                                 nr_thread > 1 ? "s" : "");
1694                 if (nr_process) {
1695                         if (nr_thread)
1696                                 log_info(" and ");
1697                         log_info("%d process%s", nr_process,
1698                                                 nr_process > 1 ? "es" : "");
1699                 }
1700                 log_info("\n");
1701                 fflush(stdout);
1702         }
1703
1704         todo = thread_number;
1705         nr_running = 0;
1706         nr_started = 0;
1707         m_rate = t_rate = 0;
1708
1709         for_each_td(td, i) {
1710                 print_status_init(td->thread_number - 1);
1711
1712                 if (!td->o.create_serialize)
1713                         continue;
1714
1715                 /*
1716                  * do file setup here so it happens sequentially,
1717                  * we don't want X number of threads getting their
1718                  * client data interspersed on disk
1719                  */
1720                 if (setup_files(td)) {
1721                         exit_value++;
1722                         if (td->error)
1723                                 log_err("fio: pid=%d, err=%d/%s\n",
1724                                         (int) td->pid, td->error, td->verror);
1725                         td_set_runstate(td, TD_REAPED);
1726                         todo--;
1727                 } else {
1728                         struct fio_file *f;
1729                         unsigned int j;
1730
1731                         /*
1732                          * for sharing to work, each job must always open
1733                          * its own files. so close them, if we opened them
1734                          * for creation
1735                          */
1736                         for_each_file(td, f, j) {
1737                                 if (fio_file_open(f))
1738                                         td_io_close_file(td, f);
1739                         }
1740                 }
1741         }
1742
1743         /* start idle threads before io threads start to run */
1744         fio_idle_prof_start();
1745
1746         set_genesis_time();
1747
1748         while (todo) {
1749                 struct thread_data *map[REAL_MAX_JOBS];
1750                 struct timeval this_start;
1751                 int this_jobs = 0, left;
1752
1753                 /*
1754                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1755                  */
1756                 for_each_td(td, i) {
1757                         if (td->runstate != TD_NOT_CREATED)
1758                                 continue;
1759
1760                         /*
1761                          * never got a chance to start, killed by other
1762                          * thread for some reason
1763                          */
1764                         if (td->terminate) {
1765                                 todo--;
1766                                 continue;
1767                         }
1768
1769                         if (td->o.start_delay) {
1770                                 spent = mtime_since_genesis();
1771
1772                                 if (td->o.start_delay * 1000 > spent)
1773                                         continue;
1774                         }
1775
1776                         if (td->o.stonewall && (nr_started || nr_running)) {
1777                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1778                                                         td->o.name);
1779                                 break;
1780                         }
1781
1782                         init_disk_util(td);
1783
1784                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1785                         td->update_rusage = 0;
1786
1787                         /*
1788                          * Set state to created. Thread will transition
1789                          * to TD_INITIALIZED when it's done setting up.
1790                          */
1791                         td_set_runstate(td, TD_CREATED);
1792                         map[this_jobs++] = td;
1793                         nr_started++;
1794
1795                         if (td->o.use_thread) {
1796                                 int ret;
1797
1798                                 dprint(FD_PROCESS, "will pthread_create\n");
1799                                 ret = pthread_create(&td->thread, NULL,
1800                                                         thread_main, td);
1801                                 if (ret) {
1802                                         log_err("pthread_create: %s\n",
1803                                                         strerror(ret));
1804                                         nr_started--;
1805                                         break;
1806                                 }
1807                                 ret = pthread_detach(td->thread);
1808                                 if (ret)
1809                                         log_err("pthread_detach: %s",
1810                                                         strerror(ret));
1811                         } else {
1812                                 pid_t pid;
1813                                 dprint(FD_PROCESS, "will fork\n");
1814                                 pid = fork();
1815                                 if (!pid) {
1816                                         int ret = fork_main(shm_id, i);
1817
1818                                         _exit(ret);
1819                                 } else if (i == fio_debug_jobno)
1820                                         *fio_debug_jobp = pid;
1821                         }
1822                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1823                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1824                                 log_err("fio: job startup hung? exiting.\n");
1825                                 fio_terminate_threads(TERMINATE_ALL);
1826                                 fio_abort = 1;
1827                                 nr_started--;
1828                                 break;
1829                         }
1830                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1831                 }
1832
1833                 /*
1834                  * Wait for the started threads to transition to
1835                  * TD_INITIALIZED.
1836                  */
1837                 fio_gettime(&this_start, NULL);
1838                 left = this_jobs;
1839                 while (left && !fio_abort) {
1840                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1841                                 break;
1842
1843                         do_usleep(100000);
1844
1845                         for (i = 0; i < this_jobs; i++) {
1846                                 td = map[i];
1847                                 if (!td)
1848                                         continue;
1849                                 if (td->runstate == TD_INITIALIZED) {
1850                                         map[i] = NULL;
1851                                         left--;
1852                                 } else if (td->runstate >= TD_EXITED) {
1853                                         map[i] = NULL;
1854                                         left--;
1855                                         todo--;
1856                                         nr_running++; /* work-around... */
1857                                 }
1858                         }
1859                 }
1860
1861                 if (left) {
1862                         log_err("fio: %d job%s failed to start\n", left,
1863                                         left > 1 ? "s" : "");
1864                         for (i = 0; i < this_jobs; i++) {
1865                                 td = map[i];
1866                                 if (!td)
1867                                         continue;
1868                                 kill(td->pid, SIGTERM);
1869                         }
1870                         break;
1871                 }
1872
1873                 /*
1874                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1875                  */
1876                 for_each_td(td, i) {
1877                         if (td->runstate != TD_INITIALIZED)
1878                                 continue;
1879
1880                         if (in_ramp_time(td))
1881                                 td_set_runstate(td, TD_RAMP);
1882                         else
1883                                 td_set_runstate(td, TD_RUNNING);
1884                         nr_running++;
1885                         nr_started--;
1886                         m_rate += ddir_rw_sum(td->o.ratemin);
1887                         t_rate += ddir_rw_sum(td->o.rate);
1888                         todo--;
1889                         fio_mutex_up(td->mutex);
1890                 }
1891
1892                 reap_threads(&nr_running, &t_rate, &m_rate);
1893
1894                 if (todo)
1895                         do_usleep(100000);
1896         }
1897
1898         while (nr_running) {
1899                 reap_threads(&nr_running, &t_rate, &m_rate);
1900                 do_usleep(10000);
1901         }
1902
1903         fio_idle_prof_stop();
1904
1905         update_io_ticks();
1906 }
1907
1908 void wait_for_disk_thread_exit(void)
1909 {
1910         fio_mutex_down(disk_thread_mutex);
1911 }
1912
1913 static void free_disk_util(void)
1914 {
1915         disk_util_start_exit();
1916         wait_for_disk_thread_exit();
1917         disk_util_prune_entries();
1918 }
1919
1920 static void *disk_thread_main(void *data)
1921 {
1922         int ret = 0;
1923
1924         fio_mutex_up(startup_mutex);
1925
1926         while (threads && !ret) {
1927                 usleep(DISK_UTIL_MSEC * 1000);
1928                 if (!threads)
1929                         break;
1930                 ret = update_io_ticks();
1931
1932                 if (!is_backend)
1933                         print_thread_status();
1934         }
1935
1936         fio_mutex_up(disk_thread_mutex);
1937         return NULL;
1938 }
1939
1940 static int create_disk_util_thread(void)
1941 {
1942         int ret;
1943
1944         setup_disk_util();
1945
1946         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1947
1948         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1949         if (ret) {
1950                 fio_mutex_remove(disk_thread_mutex);
1951                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1952                 return 1;
1953         }
1954
1955         ret = pthread_detach(disk_util_thread);
1956         if (ret) {
1957                 fio_mutex_remove(disk_thread_mutex);
1958                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1959                 return 1;
1960         }
1961
1962         dprint(FD_MUTEX, "wait on startup_mutex\n");
1963         fio_mutex_down(startup_mutex);
1964         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1965         return 0;
1966 }
1967
1968 int fio_backend(void)
1969 {
1970         struct thread_data *td;
1971         int i;
1972
1973         if (exec_profile) {
1974                 if (load_profile(exec_profile))
1975                         return 1;
1976                 free(exec_profile);
1977                 exec_profile = NULL;
1978         }
1979         if (!thread_number)
1980                 return 0;
1981
1982         if (write_bw_log) {
1983                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1984                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1985                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1986         }
1987
1988         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1989         if (startup_mutex == NULL)
1990                 return 1;
1991         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1992         if (writeout_mutex == NULL)
1993                 return 1;
1994
1995         set_genesis_time();
1996         stat_init();
1997         create_disk_util_thread();
1998
1999         cgroup_list = smalloc(sizeof(*cgroup_list));
2000         INIT_FLIST_HEAD(cgroup_list);
2001
2002         run_threads();
2003
2004         if (!fio_abort) {
2005                 show_run_stats();
2006                 if (write_bw_log) {
2007                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2008                         __finish_log(agg_io_log[DDIR_WRITE],
2009                                         "agg-write_bw.log");
2010                         __finish_log(agg_io_log[DDIR_TRIM],
2011                                         "agg-write_bw.log");
2012                 }
2013         }
2014
2015         for_each_td(td, i)
2016                 fio_options_free(td);
2017
2018         free_disk_util();
2019         cgroup_kill(cgroup_list);
2020         sfree(cgroup_list);
2021         sfree(cgroup_mnt);
2022
2023         fio_mutex_remove(startup_mutex);
2024         fio_mutex_remove(writeout_mutex);
2025         fio_mutex_remove(disk_thread_mutex);
2026         stat_exit();
2027         return exit_value;
2028 }