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         lat_target_init(td);
 653
 654         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 655                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 656                 td->o.time_based) {
 657                 struct timeval comp_time;
 658                 int min_evts = 0;
 659                 struct io_u *io_u;
 660                 int ret2, full;
 661                 enum fio_ddir ddir;
 662
 663                 check_update_rusage(td);
 664
 665                 if (td->terminate || td->done)
 666                         break;
 667
 668                 update_tv_cache(td);
 669
 670                 if (runtime_exceeded(td, &td->tv_cache)) {
 671                         __update_tv_cache(td);
 672                         if (runtime_exceeded(td, &td->tv_cache)) {
 673                                 td->terminate = 1;
 674                                 break;
 675                         }
 676                 }
 677
 678                 if (flow_threshold_exceeded(td))
 679                         continue;
 680
 681                 if (bytes_issued >= (uint64_t) td->o.size)
 682                         break;
 683
 684                 io_u = get_io_u(td);
 685                 if (!io_u) {
 686                         if (td->o.latency_target)
 687                                 goto reap;
 688                         break;
 689                 }
 690
 691                 ddir = io_u->ddir;
 692
 693                 /*
 694                  * Add verification end_io handler if:
 695                  *      - Asked to verify (!td_rw(td))
 696                  *      - Or the io_u is from our verify list (mixed write/ver)
 697                  */
 698                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 699                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 700
 701                         if (!td->o.verify_pattern_bytes) {
 702                                 io_u->rand_seed = __rand(&td->__verify_state);
 703                                 if (sizeof(int) != sizeof(long *))
 704                                         io_u->rand_seed *= __rand(&td->__verify_state);
 705                         }
 706
 707                         if (td->o.verify_async)
 708                                 io_u->end_io = verify_io_u_async;
 709                         else
 710                                 io_u->end_io = verify_io_u;
 711                         td_set_runstate(td, TD_VERIFYING);
 712                 } else if (in_ramp_time(td))
 713                         td_set_runstate(td, TD_RAMP);
 714                 else
 715                         td_set_runstate(td, TD_RUNNING);
 716
 717                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 718                     td->o.do_verify &&
 719                     td->o.verify != VERIFY_NONE &&
 720                     !td->o.experimental_verify)
 721                         log_io_piece(td, io_u);
 722
 723                 ret = td_io_queue(td, io_u);
 724                 switch (ret) {
 725                 case FIO_Q_COMPLETED:
 726                         if (io_u->error) {
 727                                 ret = -io_u->error;
 728                                 clear_io_u(td, io_u);
 729                         } else if (io_u->resid) {
 730                                 int bytes = io_u->xfer_buflen - io_u->resid;
 731                                 struct fio_file *f = io_u->file;
 732
 733                                 bytes_issued += bytes;
 734                                 /*
 735                                  * zero read, fail
 736                                  */
 737                                 if (!bytes) {
 738                                         td_verror(td, EIO, "full resid");
 739                                         put_io_u(td, io_u);
 740                                         break;
 741                                 }
 742
 743                                 io_u->xfer_buflen = io_u->resid;
 744                                 io_u->xfer_buf += bytes;
 745                                 io_u->offset += bytes;
 746
 747                                 if (ddir_rw(io_u->ddir))
 748                                         td->ts.short_io_u[io_u->ddir]++;
 749
 750                                 if (io_u->offset == f->real_file_size)
 751                                         goto sync_done;
 752
 753                                 requeue_io_u(td, &io_u);
 754                         } else {
 755 sync_done:
 756                                 if (__should_check_rate(td, DDIR_READ) ||
 757                                     __should_check_rate(td, DDIR_WRITE) ||
 758                                     __should_check_rate(td, DDIR_TRIM))
 759                                         fio_gettime(&comp_time, NULL);
 760
 761                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 762                                 if (ret < 0)
 763                                         break;
 764                                 bytes_issued += io_u->xfer_buflen;
 765                         }
 766                         break;
 767                 case FIO_Q_QUEUED:
 768                         /*
 769                          * if the engine doesn't have a commit hook,
 770                          * the io_u is really queued. if it does have such
 771                          * a hook, it has to call io_u_queued() itself.
 772                          */
 773                         if (td->io_ops->commit == NULL)
 774                                 io_u_queued(td, io_u);
 775                         bytes_issued += io_u->xfer_buflen;
 776                         break;
 777                 case FIO_Q_BUSY:
 778                         requeue_io_u(td, &io_u);
 779                         ret2 = td_io_commit(td);
 780                         if (ret2 < 0)
 781                                 ret = ret2;
 782                         break;
 783                 default:
 784                         assert(ret < 0);
 785                         put_io_u(td, io_u);
 786                         break;
 787                 }
 788
 789                 if (break_on_this_error(td, ddir, &ret))
 790                         break;
 791
 792                 /*
 793                  * See if we need to complete some commands. Note that we
 794                  * can get BUSY even without IO queued, if the system is
 795                  * resource starved.
 796                  */
 797 reap:
 798                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 799                 if (full || !td->o.iodepth_batch_complete) {
 800                         min_evts = min(td->o.iodepth_batch_complete,
 801                                         td->cur_depth);
 802                         /*
 803                          * if the queue is full, we MUST reap at least 1 event
 804                          */
 805                         if (full && !min_evts)
 806                                 min_evts = 1;
 807
 808                         if (__should_check_rate(td, DDIR_READ) ||
 809                             __should_check_rate(td, DDIR_WRITE) ||
 810                             __should_check_rate(td, DDIR_TRIM))
 811                                 fio_gettime(&comp_time, NULL);
 812
 813                         do {
 814                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 815                                 if (ret < 0)
 816                                         break;
 817
 818                         } while (full && (td->cur_depth > td->o.iodepth_low));
 819                 }
 820
 821                 if (ret < 0)
 822                         break;
 823                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 824                         continue;
 825
 826                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 827                         if (check_min_rate(td, &comp_time, bytes_done)) {
 828                                 if (exitall_on_terminate)
 829                                         fio_terminate_threads(td->groupid);
 830                                 td_verror(td, EIO, "check_min_rate");
 831                                 break;
 832                         }
 833                 }
 834                 if (!in_ramp_time(td) && td->o.latency_target)
 835                         lat_target_check(td);
 836
 837                 if (td->o.thinktime) {
 838                         unsigned long long b;
 839
 840                         b = ddir_rw_sum(td->io_blocks);
 841                         if (!(b % td->o.thinktime_blocks)) {
 842                                 int left;
 843
 844                                 io_u_quiesce(td);
 845
 846                                 if (td->o.thinktime_spin)
 847                                         usec_spin(td->o.thinktime_spin);
 848
 849                                 left = td->o.thinktime - td->o.thinktime_spin;
 850                                 if (left)
 851                                         usec_sleep(td, left);
 852                         }
 853                 }
 854         }
 855
 856         check_update_rusage(td);
 857
 858         if (td->trim_entries)
 859                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 860
 861         if (td->o.fill_device && td->error == ENOSPC) {
 862                 td->error = 0;
 863                 td->terminate = 1;
 864         }
 865         if (!td->error) {
 866                 struct fio_file *f;
 867
 868                 i = td->cur_depth;
 869                 if (i) {
 870                         ret = io_u_queued_complete(td, i, bytes_done);
 871                         if (td->o.fill_device && td->error == ENOSPC)
 872                                 td->error = 0;
 873                 }
 874
 875                 if (should_fsync(td) && td->o.end_fsync) {
 876                         td_set_runstate(td, TD_FSYNCING);
 877
 878                         for_each_file(td, f, i) {
 879                                 if (!fio_file_fsync(td, f))
 880                                         continue;
 881
 882                                 log_err("fio: end_fsync failed for file %s\n",
 883                                                                 f->file_name);
 884                         }
 885                 }
 886         } else
 887                 cleanup_pending_aio(td);
 888
 889         /*
 890          * stop job if we failed doing any IO
 891          */
 892         if (!ddir_rw_sum(td->this_io_bytes))
 893                 td->done = 1;
 894
 895         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 896 }
 897
 898 static void cleanup_io_u(struct thread_data *td)
 899 {
 900         struct io_u *io_u;
 901
 902         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 903
 904                 if (td->io_ops->io_u_free)
 905                         td->io_ops->io_u_free(td, io_u);
 906
 907                 fio_memfree(io_u, sizeof(*io_u));
 908         }
 909
 910         free_io_mem(td);
 911
 912         io_u_rexit(&td->io_u_requeues);
 913         io_u_qexit(&td->io_u_freelist);
 914         io_u_qexit(&td->io_u_all);
 915 }
 916
 917 static int init_io_u(struct thread_data *td)
 918 {
 919         struct io_u *io_u;
 920         unsigned int max_bs, min_write;
 921         int cl_align, i, max_units;
 922         int data_xfer = 1, err;
 923         char *p;
 924
 925         max_units = td->o.iodepth;
 926         max_bs = td_max_bs(td);
 927         min_write = td->o.min_bs[DDIR_WRITE];
 928         td->orig_buffer_size = (unsigned long long) max_bs
 929                                         * (unsigned long long) max_units;
 930
 931         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 932                 data_xfer = 0;
 933
 934         err = 0;
 935         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 936         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 937         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 938
 939         if (err) {
 940                 log_err("fio: failed setting up IO queues\n");
 941                 return 1;
 942         }
 943
 944         /*
 945          * if we may later need to do address alignment, then add any
 946          * possible adjustment here so that we don't cause a buffer
 947          * overflow later. this adjustment may be too much if we get
 948          * lucky and the allocator gives us an aligned address.
 949          */
 950         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 951             (td->io_ops->flags & FIO_RAWIO))
 952                 td->orig_buffer_size += page_mask + td->o.mem_align;
 953
 954         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 955                 unsigned long bs;
 956
 957                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 958                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 959         }
 960
 961         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 962                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 963                 return 1;
 964         }
 965
 966         if (data_xfer && allocate_io_mem(td))
 967                 return 1;
 968
 969         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 970             (td->io_ops->flags & FIO_RAWIO))
 971                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 972         else
 973                 p = td->orig_buffer;
 974
 975         cl_align = os_cache_line_size();
 976
 977         for (i = 0; i < max_units; i++) {
 978                 void *ptr;
 979
 980                 if (td->terminate)
 981                         return 1;
 982
 983                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 984                 if (!ptr) {
 985                         log_err("fio: unable to allocate aligned memory\n");
 986                         break;
 987                 }
 988
 989                 io_u = ptr;
 990                 memset(io_u, 0, sizeof(*io_u));
 991                 INIT_FLIST_HEAD(&io_u->verify_list);
 992                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 993
 994                 if (data_xfer) {
 995                         io_u->buf = p;
 996                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 997
 998                         if (td_write(td))
 999                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1000                         if (td_write(td) && td->o.verify_pattern_bytes) {
1001                                 /*
1002                                  * Fill the buffer with the pattern if we are
1003                                  * going to be doing writes.
1004                                  */
1005                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1006                         }
1007                 }
1008
1009                 io_u->index = i;
1010                 io_u->flags = IO_U_F_FREE;
1011                 io_u_qpush(&td->io_u_freelist, io_u);
1012
1013                 /*
1014                  * io_u never leaves this stack, used for iteration of all
1015                  * io_u buffers.
1016                  */
1017                 io_u_qpush(&td->io_u_all, io_u);
1018
1019                 if (td->io_ops->io_u_init) {
1020                         int ret = td->io_ops->io_u_init(td, io_u);
1021
1022                         if (ret) {
1023                                 log_err("fio: failed to init engine data: %d\n", ret);
1024                                 return 1;
1025                         }
1026                 }
1027
1028                 p += max_bs;
1029         }
1030
1031         return 0;
1032 }
1033
1034 static int switch_ioscheduler(struct thread_data *td)
1035 {
1036         char tmp[256], tmp2[128];
1037         FILE *f;
1038         int ret;
1039
1040         if (td->io_ops->flags & FIO_DISKLESSIO)
1041                 return 0;
1042
1043         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1044
1045         f = fopen(tmp, "r+");
1046         if (!f) {
1047                 if (errno == ENOENT) {
1048                         log_err("fio: os or kernel doesn't support IO scheduler"
1049                                 " switching\n");
1050                         return 0;
1051                 }
1052                 td_verror(td, errno, "fopen iosched");
1053                 return 1;
1054         }
1055
1056         /*
1057          * Set io scheduler.
1058          */
1059         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1060         if (ferror(f) || ret != 1) {
1061                 td_verror(td, errno, "fwrite");
1062                 fclose(f);
1063                 return 1;
1064         }
1065
1066         rewind(f);
1067
1068         /*
1069          * Read back and check that the selected scheduler is now the default.
1070          */
1071         ret = fread(tmp, 1, sizeof(tmp), f);
1072         if (ferror(f) || ret < 0) {
1073                 td_verror(td, errno, "fread");
1074                 fclose(f);
1075                 return 1;
1076         }
1077
1078         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1079         if (!strstr(tmp, tmp2)) {
1080                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1081                 td_verror(td, EINVAL, "iosched_switch");
1082                 fclose(f);
1083                 return 1;
1084         }
1085
1086         fclose(f);
1087         return 0;
1088 }
1089
1090 static int keep_running(struct thread_data *td)
1091 {
1092         if (td->done)
1093                 return 0;
1094         if (td->o.time_based)
1095                 return 1;
1096         if (td->o.loops) {
1097                 td->o.loops--;
1098                 return 1;
1099         }
1100
1101         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1102                 uint64_t diff;
1103
1104                 /*
1105                  * If the difference is less than the minimum IO size, we
1106                  * are done.
1107                  */
1108                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1109                 if (diff < td_max_bs(td))
1110                         return 0;
1111
1112                 return 1;
1113         }
1114
1115         return 0;
1116 }
1117
1118 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1119 {
1120         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1121         char *str;
1122
1123         str = malloc(newlen);
1124         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1125
1126         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1127         ret = system(str);
1128         if (ret == -1)
1129                 log_err("fio: exec of cmd <%s> failed\n", str);
1130
1131         free(str);
1132         return ret;
1133 }
1134
1135 /*
1136  * Dry run to compute correct state of numberio for verification.
1137  */
1138 static uint64_t do_dry_run(struct thread_data *td)
1139 {
1140         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1141
1142         td_set_runstate(td, TD_RUNNING);
1143
1144         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1145                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1146                 struct io_u *io_u;
1147                 int ret;
1148
1149                 if (td->terminate || td->done)
1150                         break;
1151
1152                 io_u = get_io_u(td);
1153                 if (!io_u)
1154                         break;
1155
1156                 io_u->flags |= IO_U_F_FLIGHT;
1157                 io_u->error = 0;
1158                 io_u->resid = 0;
1159                 if (ddir_rw(acct_ddir(io_u)))
1160                         td->io_issues[acct_ddir(io_u)]++;
1161                 if (ddir_rw(io_u->ddir)) {
1162                         io_u_mark_depth(td, 1);
1163                         td->ts.total_io_u[io_u->ddir]++;
1164                 }
1165
1166                 ret = io_u_sync_complete(td, io_u, bytes_done);
1167                 (void) ret;
1168         }
1169
1170         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1171 }
1172
1173 /*
1174  * Entry point for the thread based jobs. The process based jobs end up
1175  * here as well, after a little setup.
1176  */
1177 static void *thread_main(void *data)
1178 {
1179         unsigned long long elapsed;
1180         struct thread_data *td = data;
1181         struct thread_options *o = &td->o;
1182         pthread_condattr_t attr;
1183         int clear_state;
1184         int ret;
1185
1186         if (!o->use_thread) {
1187                 setsid();
1188                 td->pid = getpid();
1189         } else
1190                 td->pid = gettid();
1191
1192         /*
1193          * fio_time_init() may not have been called yet if running as a server
1194          */
1195         fio_time_init();
1196
1197         fio_local_clock_init(o->use_thread);
1198
1199         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1200
1201         if (is_backend)
1202                 fio_server_send_start(td);
1203
1204         INIT_FLIST_HEAD(&td->io_log_list);
1205         INIT_FLIST_HEAD(&td->io_hist_list);
1206         INIT_FLIST_HEAD(&td->verify_list);
1207         INIT_FLIST_HEAD(&td->trim_list);
1208         INIT_FLIST_HEAD(&td->next_rand_list);
1209         pthread_mutex_init(&td->io_u_lock, NULL);
1210         td->io_hist_tree = RB_ROOT;
1211
1212         pthread_condattr_init(&attr);
1213         pthread_cond_init(&td->verify_cond, &attr);
1214         pthread_cond_init(&td->free_cond, &attr);
1215
1216         td_set_runstate(td, TD_INITIALIZED);
1217         dprint(FD_MUTEX, "up startup_mutex\n");
1218         fio_mutex_up(startup_mutex);
1219         dprint(FD_MUTEX, "wait on td->mutex\n");
1220         fio_mutex_down(td->mutex);
1221         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1222
1223         /*
1224          * the ->mutex mutex is now no longer used, close it to avoid
1225          * eating a file descriptor
1226          */
1227         fio_mutex_remove(td->mutex);
1228         td->mutex = NULL;
1229
1230         /*
1231          * A new gid requires privilege, so we need to do this before setting
1232          * the uid.
1233          */
1234         if (o->gid != -1U && setgid(o->gid)) {
1235                 td_verror(td, errno, "setgid");
1236                 goto err;
1237         }
1238         if (o->uid != -1U && setuid(o->uid)) {
1239                 td_verror(td, errno, "setuid");
1240                 goto err;
1241         }
1242
1243         /*
1244          * If we have a gettimeofday() thread, make sure we exclude that
1245          * thread from this job
1246          */
1247         if (o->gtod_cpu)
1248                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1249
1250         /*
1251          * Set affinity first, in case it has an impact on the memory
1252          * allocations.
1253          */
1254         if (o->cpumask_set) {
1255                 ret = fio_setaffinity(td->pid, o->cpumask);
1256                 if (ret == -1) {
1257                         td_verror(td, errno, "cpu_set_affinity");
1258                         goto err;
1259                 }
1260         }
1261
1262 #ifdef CONFIG_LIBNUMA
1263         /* numa node setup */
1264         if (o->numa_cpumask_set || o->numa_memmask_set) {
1265                 int ret;
1266
1267                 if (numa_available() < 0) {
1268                         td_verror(td, errno, "Does not support NUMA API\n");
1269                         goto err;
1270                 }
1271
1272                 if (o->numa_cpumask_set) {
1273                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1274                         if (ret == -1) {
1275                                 td_verror(td, errno, \
1276                                         "numa_run_on_node_mask failed\n");
1277                                 goto err;
1278                         }
1279                 }
1280
1281                 if (o->numa_memmask_set) {
1282
1283                         switch (o->numa_mem_mode) {
1284                         case MPOL_INTERLEAVE:
1285                                 numa_set_interleave_mask(o->numa_memnodesmask);
1286                                 break;
1287                         case MPOL_BIND:
1288                                 numa_set_membind(o->numa_memnodesmask);
1289                                 break;
1290                         case MPOL_LOCAL:
1291                                 numa_set_localalloc();
1292                                 break;
1293                         case MPOL_PREFERRED:
1294                                 numa_set_preferred(o->numa_mem_prefer_node);
1295                                 break;
1296                         case MPOL_DEFAULT:
1297                         default:
1298                                 break;
1299                         }
1300
1301                 }
1302         }
1303 #endif
1304
1305         if (fio_pin_memory(td))
1306                 goto err;
1307
1308         /*
1309          * May alter parameters that init_io_u() will use, so we need to
1310          * do this first.
1311          */
1312         if (init_iolog(td))
1313                 goto err;
1314
1315         if (init_io_u(td))
1316                 goto err;
1317
1318         if (o->verify_async && verify_async_init(td))
1319                 goto err;
1320
1321         if (o->ioprio) {
1322                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1323                 if (ret == -1) {
1324                         td_verror(td, errno, "ioprio_set");
1325                         goto err;
1326                 }
1327         }
1328
1329         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1330                 goto err;
1331
1332         errno = 0;
1333         if (nice(o->nice) == -1 && errno != 0) {
1334                 td_verror(td, errno, "nice");
1335                 goto err;
1336         }
1337
1338         if (o->ioscheduler && switch_ioscheduler(td))
1339                 goto err;
1340
1341         if (!o->create_serialize && setup_files(td))
1342                 goto err;
1343
1344         if (td_io_init(td))
1345                 goto err;
1346
1347         if (init_random_map(td))
1348                 goto err;
1349
1350         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1351                 goto err;
1352
1353         if (o->pre_read) {
1354                 if (pre_read_files(td) < 0)
1355                         goto err;
1356         }
1357
1358         fio_verify_init(td);
1359
1360         fio_gettime(&td->epoch, NULL);
1361         fio_getrusage(&td->ru_start);
1362         clear_state = 0;
1363         while (keep_running(td)) {
1364                 uint64_t verify_bytes;
1365
1366                 fio_gettime(&td->start, NULL);
1367                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1368                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1369                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1370
1371                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1372                                 o->ratemin[DDIR_TRIM]) {
1373                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1374                                                 sizeof(td->bw_sample_time));
1375                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1376                                                 sizeof(td->bw_sample_time));
1377                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1378                                                 sizeof(td->bw_sample_time));
1379                 }
1380
1381                 if (clear_state)
1382                         clear_io_state(td);
1383
1384                 prune_io_piece_log(td);
1385
1386                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1387                         verify_bytes = do_dry_run(td);
1388                 else
1389                         verify_bytes = do_io(td);
1390
1391                 clear_state = 1;
1392
1393                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1394                         elapsed = utime_since_now(&td->start);
1395                         td->ts.runtime[DDIR_READ] += elapsed;
1396                 }
1397                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1398                         elapsed = utime_since_now(&td->start);
1399                         td->ts.runtime[DDIR_WRITE] += elapsed;
1400                 }
1401                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1402                         elapsed = utime_since_now(&td->start);
1403                         td->ts.runtime[DDIR_TRIM] += elapsed;
1404                 }
1405
1406                 if (td->error || td->terminate)
1407                         break;
1408
1409                 if (!o->do_verify ||
1410                     o->verify == VERIFY_NONE ||
1411                     (td->io_ops->flags & FIO_UNIDIR))
1412                         continue;
1413
1414                 clear_io_state(td);
1415
1416                 fio_gettime(&td->start, NULL);
1417
1418                 do_verify(td, verify_bytes);
1419
1420                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1421
1422                 if (td->error || td->terminate)
1423                         break;
1424         }
1425
1426         update_rusage_stat(td);
1427         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1428         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1429         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1430         td->ts.total_run_time = mtime_since_now(&td->epoch);
1431         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1432         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1433         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1434
1435         fio_unpin_memory(td);
1436
1437         fio_mutex_down(writeout_mutex);
1438         if (td->bw_log) {
1439                 if (o->bw_log_file) {
1440                         finish_log_named(td, td->bw_log,
1441                                                 o->bw_log_file, "bw");
1442                 } else
1443                         finish_log(td, td->bw_log, "bw");
1444         }
1445         if (td->lat_log) {
1446                 if (o->lat_log_file) {
1447                         finish_log_named(td, td->lat_log,
1448                                                 o->lat_log_file, "lat");
1449                 } else
1450                         finish_log(td, td->lat_log, "lat");
1451         }
1452         if (td->slat_log) {
1453                 if (o->lat_log_file) {
1454                         finish_log_named(td, td->slat_log,
1455                                                 o->lat_log_file, "slat");
1456                 } else
1457                         finish_log(td, td->slat_log, "slat");
1458         }
1459         if (td->clat_log) {
1460                 if (o->lat_log_file) {
1461                         finish_log_named(td, td->clat_log,
1462                                                 o->lat_log_file, "clat");
1463                 } else
1464                         finish_log(td, td->clat_log, "clat");
1465         }
1466         if (td->iops_log) {
1467                 if (o->iops_log_file) {
1468                         finish_log_named(td, td->iops_log,
1469                                                 o->iops_log_file, "iops");
1470                 } else
1471                         finish_log(td, td->iops_log, "iops");
1472         }
1473
1474         fio_mutex_up(writeout_mutex);
1475         if (o->exec_postrun)
1476                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1477
1478         if (exitall_on_terminate)
1479                 fio_terminate_threads(td->groupid);
1480
1481 err:
1482         if (td->error)
1483                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1484                                                         td->verror);
1485
1486         if (o->verify_async)
1487                 verify_async_exit(td);
1488
1489         close_and_free_files(td);
1490         cleanup_io_u(td);
1491         close_ioengine(td);
1492         cgroup_shutdown(td, &cgroup_mnt);
1493
1494         if (o->cpumask_set) {
1495                 int ret = fio_cpuset_exit(&o->cpumask);
1496
1497                 td_verror(td, ret, "fio_cpuset_exit");
1498         }
1499
1500         /*
1501          * do this very late, it will log file closing as well
1502          */
1503         if (o->write_iolog_file)
1504                 write_iolog_close(td);
1505
1506         fio_mutex_remove(td->rusage_sem);
1507         td->rusage_sem = NULL;
1508
1509         td_set_runstate(td, TD_EXITED);
1510         return (void *) (uintptr_t) td->error;
1511 }
1512
1513
1514 /*
1515  * We cannot pass the td data into a forked process, so attach the td and
1516  * pass it to the thread worker.
1517  */
1518 static int fork_main(int shmid, int offset)
1519 {
1520         struct thread_data *td;
1521         void *data, *ret;
1522
1523 #ifndef __hpux
1524         data = shmat(shmid, NULL, 0);
1525         if (data == (void *) -1) {
1526                 int __err = errno;
1527
1528                 perror("shmat");
1529                 return __err;
1530         }
1531 #else
1532         /*
1533          * HP-UX inherits shm mappings?
1534          */
1535         data = threads;
1536 #endif
1537
1538         td = data + offset * sizeof(struct thread_data);
1539         ret = thread_main(td);
1540         shmdt(data);
1541         return (int) (uintptr_t) ret;
1542 }
1543
1544 /*
1545  * Run over the job map and reap the threads that have exited, if any.
1546  */
1547 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1548                          unsigned int *m_rate)
1549 {
1550         struct thread_data *td;
1551         unsigned int cputhreads, realthreads, pending;
1552         int i, status, ret;
1553
1554         /*
1555          * reap exited threads (TD_EXITED -> TD_REAPED)
1556          */
1557         realthreads = pending = cputhreads = 0;
1558         for_each_td(td, i) {
1559                 int flags = 0;
1560
1561                 /*
1562                  * ->io_ops is NULL for a thread that has closed its
1563                  * io engine
1564                  */
1565                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1566                         cputhreads++;
1567                 else
1568                         realthreads++;
1569
1570                 if (!td->pid) {
1571                         pending++;
1572                         continue;
1573                 }
1574                 if (td->runstate == TD_REAPED)
1575                         continue;
1576                 if (td->o.use_thread) {
1577                         if (td->runstate == TD_EXITED) {
1578                                 td_set_runstate(td, TD_REAPED);
1579                                 goto reaped;
1580                         }
1581                         continue;
1582                 }
1583
1584                 flags = WNOHANG;
1585                 if (td->runstate == TD_EXITED)
1586                         flags = 0;
1587
1588                 /*
1589                  * check if someone quit or got killed in an unusual way
1590                  */
1591                 ret = waitpid(td->pid, &status, flags);
1592                 if (ret < 0) {
1593                         if (errno == ECHILD) {
1594                                 log_err("fio: pid=%d disappeared %d\n",
1595                                                 (int) td->pid, td->runstate);
1596                                 td->sig = ECHILD;
1597                                 td_set_runstate(td, TD_REAPED);
1598                                 goto reaped;
1599                         }
1600                         perror("waitpid");
1601                 } else if (ret == td->pid) {
1602                         if (WIFSIGNALED(status)) {
1603                                 int sig = WTERMSIG(status);
1604
1605                                 if (sig != SIGTERM && sig != SIGUSR2)
1606                                         log_err("fio: pid=%d, got signal=%d\n",
1607                                                         (int) td->pid, sig);
1608                                 td->sig = sig;
1609                                 td_set_runstate(td, TD_REAPED);
1610                                 goto reaped;
1611                         }
1612                         if (WIFEXITED(status)) {
1613                                 if (WEXITSTATUS(status) && !td->error)
1614                                         td->error = WEXITSTATUS(status);
1615
1616                                 td_set_runstate(td, TD_REAPED);
1617                                 goto reaped;
1618                         }
1619                 }
1620
1621                 /*
1622                  * thread is not dead, continue
1623                  */
1624                 pending++;
1625                 continue;
1626 reaped:
1627                 (*nr_running)--;
1628                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1629                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1630                 if (!td->pid)
1631                         pending--;
1632
1633                 if (td->error)
1634                         exit_value++;
1635
1636                 done_secs += mtime_since_now(&td->epoch) / 1000;
1637                 profile_td_exit(td);
1638         }
1639
1640         if (*nr_running == cputhreads && !pending && realthreads)
1641                 fio_terminate_threads(TERMINATE_ALL);
1642 }
1643
1644 static void do_usleep(unsigned int usecs)
1645 {
1646         check_for_running_stats();
1647         usleep(usecs);
1648 }
1649
1650 /*
1651  * Main function for kicking off and reaping jobs, as needed.
1652  */
1653 static void run_threads(void)
1654 {
1655         struct thread_data *td;
1656         unsigned long spent;
1657         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1658
1659         if (fio_gtod_offload && fio_start_gtod_thread())
1660                 return;
1661
1662         fio_idle_prof_init();
1663
1664         set_sig_handlers();
1665
1666         nr_thread = nr_process = 0;
1667         for_each_td(td, i) {
1668                 if (td->o.use_thread)
1669                         nr_thread++;
1670                 else
1671                         nr_process++;
1672         }
1673
1674         if (output_format == FIO_OUTPUT_NORMAL) {
1675                 log_info("Starting ");
1676                 if (nr_thread)
1677                         log_info("%d thread%s", nr_thread,
1678                                                 nr_thread > 1 ? "s" : "");
1679                 if (nr_process) {
1680                         if (nr_thread)
1681                                 log_info(" and ");
1682                         log_info("%d process%s", nr_process,
1683                                                 nr_process > 1 ? "es" : "");
1684                 }
1685                 log_info("\n");
1686                 fflush(stdout);
1687         }
1688
1689         todo = thread_number;
1690         nr_running = 0;
1691         nr_started = 0;
1692         m_rate = t_rate = 0;
1693
1694         for_each_td(td, i) {
1695                 print_status_init(td->thread_number - 1);
1696
1697                 if (!td->o.create_serialize)
1698                         continue;
1699
1700                 /*
1701                  * do file setup here so it happens sequentially,
1702                  * we don't want X number of threads getting their
1703                  * client data interspersed on disk
1704                  */
1705                 if (setup_files(td)) {
1706                         exit_value++;
1707                         if (td->error)
1708                                 log_err("fio: pid=%d, err=%d/%s\n",
1709                                         (int) td->pid, td->error, td->verror);
1710                         td_set_runstate(td, TD_REAPED);
1711                         todo--;
1712                 } else {
1713                         struct fio_file *f;
1714                         unsigned int j;
1715
1716                         /*
1717                          * for sharing to work, each job must always open
1718                          * its own files. so close them, if we opened them
1719                          * for creation
1720                          */
1721                         for_each_file(td, f, j) {
1722                                 if (fio_file_open(f))
1723                                         td_io_close_file(td, f);
1724                         }
1725                 }
1726         }
1727
1728         /* start idle threads before io threads start to run */
1729         fio_idle_prof_start();
1730
1731         set_genesis_time();
1732
1733         while (todo) {
1734                 struct thread_data *map[REAL_MAX_JOBS];
1735                 struct timeval this_start;
1736                 int this_jobs = 0, left;
1737
1738                 /*
1739                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1740                  */
1741                 for_each_td(td, i) {
1742                         if (td->runstate != TD_NOT_CREATED)
1743                                 continue;
1744
1745                         /*
1746                          * never got a chance to start, killed by other
1747                          * thread for some reason
1748                          */
1749                         if (td->terminate) {
1750                                 todo--;
1751                                 continue;
1752                         }
1753
1754                         if (td->o.start_delay) {
1755                                 spent = mtime_since_genesis();
1756
1757                                 if (td->o.start_delay * 1000 > spent)
1758                                         continue;
1759                         }
1760
1761                         if (td->o.stonewall && (nr_started || nr_running)) {
1762                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1763                                                         td->o.name);
1764                                 break;
1765                         }
1766
1767                         init_disk_util(td);
1768
1769                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1770                         td->update_rusage = 0;
1771
1772                         /*
1773                          * Set state to created. Thread will transition
1774                          * to TD_INITIALIZED when it's done setting up.
1775                          */
1776                         td_set_runstate(td, TD_CREATED);
1777                         map[this_jobs++] = td;
1778                         nr_started++;
1779
1780                         if (td->o.use_thread) {
1781                                 int ret;
1782
1783                                 dprint(FD_PROCESS, "will pthread_create\n");
1784                                 ret = pthread_create(&td->thread, NULL,
1785                                                         thread_main, td);
1786                                 if (ret) {
1787                                         log_err("pthread_create: %s\n",
1788                                                         strerror(ret));
1789                                         nr_started--;
1790                                         break;
1791                                 }
1792                                 ret = pthread_detach(td->thread);
1793                                 if (ret)
1794                                         log_err("pthread_detach: %s",
1795                                                         strerror(ret));
1796                         } else {
1797                                 pid_t pid;
1798                                 dprint(FD_PROCESS, "will fork\n");
1799                                 pid = fork();
1800                                 if (!pid) {
1801                                         int ret = fork_main(shm_id, i);
1802
1803                                         _exit(ret);
1804                                 } else if (i == fio_debug_jobno)
1805                                         *fio_debug_jobp = pid;
1806                         }
1807                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1808                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1809                                 log_err("fio: job startup hung? exiting.\n");
1810                                 fio_terminate_threads(TERMINATE_ALL);
1811                                 fio_abort = 1;
1812                                 nr_started--;
1813                                 break;
1814                         }
1815                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1816                 }
1817
1818                 /*
1819                  * Wait for the started threads to transition to
1820                  * TD_INITIALIZED.
1821                  */
1822                 fio_gettime(&this_start, NULL);
1823                 left = this_jobs;
1824                 while (left && !fio_abort) {
1825                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1826                                 break;
1827
1828                         do_usleep(100000);
1829
1830                         for (i = 0; i < this_jobs; i++) {
1831                                 td = map[i];
1832                                 if (!td)
1833                                         continue;
1834                                 if (td->runstate == TD_INITIALIZED) {
1835                                         map[i] = NULL;
1836                                         left--;
1837                                 } else if (td->runstate >= TD_EXITED) {
1838                                         map[i] = NULL;
1839                                         left--;
1840                                         todo--;
1841                                         nr_running++; /* work-around... */
1842                                 }
1843                         }
1844                 }
1845
1846                 if (left) {
1847                         log_err("fio: %d job%s failed to start\n", left,
1848                                         left > 1 ? "s" : "");
1849                         for (i = 0; i < this_jobs; i++) {
1850                                 td = map[i];
1851                                 if (!td)
1852                                         continue;
1853                                 kill(td->pid, SIGTERM);
1854                         }
1855                         break;
1856                 }
1857
1858                 /*
1859                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1860                  */
1861                 for_each_td(td, i) {
1862                         if (td->runstate != TD_INITIALIZED)
1863                                 continue;
1864
1865                         if (in_ramp_time(td))
1866                                 td_set_runstate(td, TD_RAMP);
1867                         else
1868                                 td_set_runstate(td, TD_RUNNING);
1869                         nr_running++;
1870                         nr_started--;
1871                         m_rate += ddir_rw_sum(td->o.ratemin);
1872                         t_rate += ddir_rw_sum(td->o.rate);
1873                         todo--;
1874                         fio_mutex_up(td->mutex);
1875                 }
1876
1877                 reap_threads(&nr_running, &t_rate, &m_rate);
1878
1879                 if (todo)
1880                         do_usleep(100000);
1881         }
1882
1883         while (nr_running) {
1884                 reap_threads(&nr_running, &t_rate, &m_rate);
1885                 do_usleep(10000);
1886         }
1887
1888         fio_idle_prof_stop();
1889
1890         update_io_ticks();
1891 }
1892
1893 void wait_for_disk_thread_exit(void)
1894 {
1895         fio_mutex_down(disk_thread_mutex);
1896 }
1897
1898 static void free_disk_util(void)
1899 {
1900         disk_util_start_exit();
1901         wait_for_disk_thread_exit();
1902         disk_util_prune_entries();
1903 }
1904
1905 static void *disk_thread_main(void *data)
1906 {
1907         int ret = 0;
1908
1909         fio_mutex_up(startup_mutex);
1910
1911         while (threads && !ret) {
1912                 usleep(DISK_UTIL_MSEC * 1000);
1913                 if (!threads)
1914                         break;
1915                 ret = update_io_ticks();
1916
1917                 if (!is_backend)
1918                         print_thread_status();
1919         }
1920
1921         fio_mutex_up(disk_thread_mutex);
1922         return NULL;
1923 }
1924
1925 static int create_disk_util_thread(void)
1926 {
1927         int ret;
1928
1929         setup_disk_util();
1930
1931         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1932
1933         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1934         if (ret) {
1935                 fio_mutex_remove(disk_thread_mutex);
1936                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1937                 return 1;
1938         }
1939
1940         ret = pthread_detach(disk_util_thread);
1941         if (ret) {
1942                 fio_mutex_remove(disk_thread_mutex);
1943                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1944                 return 1;
1945         }
1946
1947         dprint(FD_MUTEX, "wait on startup_mutex\n");
1948         fio_mutex_down(startup_mutex);
1949         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1950         return 0;
1951 }
1952
1953 int fio_backend(void)
1954 {
1955         struct thread_data *td;
1956         int i;
1957
1958         if (exec_profile) {
1959                 if (load_profile(exec_profile))
1960                         return 1;
1961                 free(exec_profile);
1962                 exec_profile = NULL;
1963         }
1964         if (!thread_number)
1965                 return 0;
1966
1967         if (write_bw_log) {
1968                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1969                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1970                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1971         }
1972
1973         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1974         if (startup_mutex == NULL)
1975                 return 1;
1976         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1977         if (writeout_mutex == NULL)
1978                 return 1;
1979
1980         set_genesis_time();
1981         stat_init();
1982         create_disk_util_thread();
1983
1984         cgroup_list = smalloc(sizeof(*cgroup_list));
1985         INIT_FLIST_HEAD(cgroup_list);
1986
1987         run_threads();
1988
1989         if (!fio_abort) {
1990                 show_run_stats();
1991                 if (write_bw_log) {
1992                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1993                         __finish_log(agg_io_log[DDIR_WRITE],
1994                                         "agg-write_bw.log");
1995                         __finish_log(agg_io_log[DDIR_TRIM],
1996                                         "agg-write_bw.log");
1997                 }
1998         }
1999
2000         for_each_td(td, i)
2001                 fio_options_free(td);
2002
2003         free_disk_util();
2004         cgroup_kill(cgroup_list);
2005         sfree(cgroup_list);
2006         sfree(cgroup_mnt);
2007
2008         fio_mutex_remove(startup_mutex);
2009         fio_mutex_remove(writeout_mutex);
2010         fio_mutex_remove(disk_thread_mutex);
2011         stat_exit();
2012         return exit_value;
2013 }