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