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 #include "err.h"
  56
  57 static pthread_t disk_util_thread;
  58 static struct fio_mutex *disk_thread_mutex;
  59 static struct fio_mutex *startup_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                                 if (spent)
 186                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 187                                 else
 188                                         rate = 0;
 189
 190                                 if (rate < ratemin ||
 191                                     bytes < td->rate_bytes[ddir]) {
 192                                         log_err("%s: min rate %u not met, got"
 193                                                 " %luKB/sec\n", td->o.name,
 194                                                         ratemin, rate);
 195                                         return 1;
 196                                 }
 197                         }
 198                 } else {
 199                         /*
 200                          * checks iops specified rate
 201                          */
 202                         if (iops < rate_iops) {
 203                                 log_err("%s: min iops rate %u not met\n",
 204                                                 td->o.name, rate_iops);
 205                                 return 1;
 206                         } else {
 207                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 208                                 if (rate < rate_iops_min ||
 209                                     iops < td->rate_blocks[ddir]) {
 210                                         log_err("%s: min iops rate %u not met,"
 211                                                 " got %lu\n", td->o.name,
 212                                                         rate_iops_min, rate);
 213                                 }
 214                         }
 215                 }
 216         }
 217
 218         td->rate_bytes[ddir] = bytes;
 219         td->rate_blocks[ddir] = iops;
 220         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 221         return 0;
 222 }
 223
 224 static int check_min_rate(struct thread_data *td, struct timeval *now,
 225                           uint64_t *bytes_done)
 226 {
 227         int ret = 0;
 228
 229         if (bytes_done[DDIR_READ])
 230                 ret |= __check_min_rate(td, now, DDIR_READ);
 231         if (bytes_done[DDIR_WRITE])
 232                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 233         if (bytes_done[DDIR_TRIM])
 234                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 235
 236         return ret;
 237 }
 238
 239 /*
 240  * When job exits, we can cancel the in-flight IO if we are using async
 241  * io. Attempt to do so.
 242  */
 243 static void cleanup_pending_aio(struct thread_data *td)
 244 {
 245         int r;
 246
 247         /*
 248          * get immediately available events, if any
 249          */
 250         r = io_u_queued_complete(td, 0, NULL);
 251         if (r < 0)
 252                 return;
 253
 254         /*
 255          * now cancel remaining active events
 256          */
 257         if (td->io_ops->cancel) {
 258                 struct io_u *io_u;
 259                 int i;
 260
 261                 io_u_qiter(&td->io_u_all, io_u, i) {
 262                         if (io_u->flags & IO_U_F_FLIGHT) {
 263                                 r = td->io_ops->cancel(td, io_u);
 264                                 if (!r)
 265                                         put_io_u(td, io_u);
 266                         }
 267                 }
 268         }
 269
 270         if (td->cur_depth)
 271                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 272 }
 273
 274 /*
 275  * Helper to handle the final sync of a file. Works just like the normal
 276  * io path, just does everything sync.
 277  */
 278 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 279 {
 280         struct io_u *io_u = __get_io_u(td);
 281         int ret;
 282
 283         if (!io_u)
 284                 return 1;
 285
 286         io_u->ddir = DDIR_SYNC;
 287         io_u->file = f;
 288
 289         if (td_io_prep(td, io_u)) {
 290                 put_io_u(td, io_u);
 291                 return 1;
 292         }
 293
 294 requeue:
 295         ret = td_io_queue(td, io_u);
 296         if (ret < 0) {
 297                 td_verror(td, io_u->error, "td_io_queue");
 298                 put_io_u(td, io_u);
 299                 return 1;
 300         } else if (ret == FIO_Q_QUEUED) {
 301                 if (io_u_queued_complete(td, 1, NULL) < 0)
 302                         return 1;
 303         } else if (ret == FIO_Q_COMPLETED) {
 304                 if (io_u->error) {
 305                         td_verror(td, io_u->error, "td_io_queue");
 306                         return 1;
 307                 }
 308
 309                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 310                         return 1;
 311         } else if (ret == FIO_Q_BUSY) {
 312                 if (td_io_commit(td))
 313                         return 1;
 314                 goto requeue;
 315         }
 316
 317         return 0;
 318 }
 319
 320 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 321 {
 322         int ret;
 323
 324         if (fio_file_open(f))
 325                 return fio_io_sync(td, f);
 326
 327         if (td_io_open_file(td, f))
 328                 return 1;
 329
 330         ret = fio_io_sync(td, f);
 331         td_io_close_file(td, f);
 332         return ret;
 333 }
 334
 335 static inline void __update_tv_cache(struct thread_data *td)
 336 {
 337         fio_gettime(&td->tv_cache, NULL);
 338 }
 339
 340 static inline void update_tv_cache(struct thread_data *td)
 341 {
 342         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 343                 __update_tv_cache(td);
 344 }
 345
 346 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 347 {
 348         if (in_ramp_time(td))
 349                 return 0;
 350         if (!td->o.timeout)
 351                 return 0;
 352         if (utime_since(&td->epoch, t) >= td->o.timeout)
 353                 return 1;
 354
 355         return 0;
 356 }
 357
 358 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 359                                int *retptr)
 360 {
 361         int ret = *retptr;
 362
 363         if (ret < 0 || td->error) {
 364                 int err = td->error;
 365                 enum error_type_bit eb;
 366
 367                 if (ret < 0)
 368                         err = -ret;
 369
 370                 eb = td_error_type(ddir, err);
 371                 if (!(td->o.continue_on_error & (1 << eb)))
 372                         return 1;
 373
 374                 if (td_non_fatal_error(td, eb, err)) {
 375                         /*
 376                          * Continue with the I/Os in case of
 377                          * a non fatal error.
 378                          */
 379                         update_error_count(td, err);
 380                         td_clear_error(td);
 381                         *retptr = 0;
 382                         return 0;
 383                 } else if (td->o.fill_device && err == ENOSPC) {
 384                         /*
 385                          * We expect to hit this error if
 386                          * fill_device option is set.
 387                          */
 388                         td_clear_error(td);
 389                         td->terminate = 1;
 390                         return 1;
 391                 } else {
 392                         /*
 393                          * Stop the I/O in case of a fatal
 394                          * error.
 395                          */
 396                         update_error_count(td, err);
 397                         return 1;
 398                 }
 399         }
 400
 401         return 0;
 402 }
 403
 404 static void check_update_rusage(struct thread_data *td)
 405 {
 406         if (td->update_rusage) {
 407                 td->update_rusage = 0;
 408                 update_rusage_stat(td);
 409                 fio_mutex_up(td->rusage_sem);
 410         }
 411 }
 412
 413 /*
 414  * The main verify engine. Runs over the writes we previously submitted,
 415  * reads the blocks back in, and checks the crc/md5 of the data.
 416  */
 417 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 418 {
 419         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 420         struct fio_file *f;
 421         struct io_u *io_u;
 422         int ret, min_events;
 423         unsigned int i;
 424
 425         dprint(FD_VERIFY, "starting loop\n");
 426
 427         /*
 428          * sync io first and invalidate cache, to make sure we really
 429          * read from disk.
 430          */
 431         for_each_file(td, f, i) {
 432                 if (!fio_file_open(f))
 433                         continue;
 434                 if (fio_io_sync(td, f))
 435                         break;
 436                 if (file_invalidate_cache(td, f))
 437                         break;
 438         }
 439
 440         check_update_rusage(td);
 441
 442         if (td->error)
 443                 return;
 444
 445         td_set_runstate(td, TD_VERIFYING);
 446
 447         io_u = NULL;
 448         while (!td->terminate) {
 449                 enum fio_ddir ddir;
 450                 int ret2, full;
 451
 452                 update_tv_cache(td);
 453                 check_update_rusage(td);
 454
 455                 if (runtime_exceeded(td, &td->tv_cache)) {
 456                         __update_tv_cache(td);
 457                         if (runtime_exceeded(td, &td->tv_cache)) {
 458                                 td->terminate = 1;
 459                                 break;
 460                         }
 461                 }
 462
 463                 if (flow_threshold_exceeded(td))
 464                         continue;
 465
 466                 if (!td->o.experimental_verify) {
 467                         io_u = __get_io_u(td);
 468                         if (!io_u)
 469                                 break;
 470
 471                         if (get_next_verify(td, io_u)) {
 472                                 put_io_u(td, io_u);
 473                                 break;
 474                         }
 475
 476                         if (td_io_prep(td, io_u)) {
 477                                 put_io_u(td, io_u);
 478                                 break;
 479                         }
 480                 } else {
 481                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 482                                 break;
 483
 484                         while ((io_u = get_io_u(td)) != NULL) {
 485                                 if (IS_ERR(io_u)) {
 486                                         io_u = NULL;
 487                                         ret = FIO_Q_BUSY;
 488                                         goto reap;
 489                                 }
 490
 491                                 /*
 492                                  * We are only interested in the places where
 493                                  * we wrote or trimmed IOs. Turn those into
 494                                  * reads for verification purposes.
 495                                  */
 496                                 if (io_u->ddir == DDIR_READ) {
 497                                         /*
 498                                          * Pretend we issued it for rwmix
 499                                          * accounting
 500                                          */
 501                                         td->io_issues[DDIR_READ]++;
 502                                         put_io_u(td, io_u);
 503                                         continue;
 504                                 } else if (io_u->ddir == DDIR_TRIM) {
 505                                         io_u->ddir = DDIR_READ;
 506                                         io_u->flags |= IO_U_F_TRIMMED;
 507                                         break;
 508                                 } else if (io_u->ddir == DDIR_WRITE) {
 509                                         io_u->ddir = DDIR_READ;
 510                                         break;
 511                                 } else {
 512                                         put_io_u(td, io_u);
 513                                         continue;
 514                                 }
 515                         }
 516
 517                         if (!io_u)
 518                                 break;
 519                 }
 520
 521                 if (td->o.verify_async)
 522                         io_u->end_io = verify_io_u_async;
 523                 else
 524                         io_u->end_io = verify_io_u;
 525
 526                 ddir = io_u->ddir;
 527
 528                 ret = td_io_queue(td, io_u);
 529                 switch (ret) {
 530                 case FIO_Q_COMPLETED:
 531                         if (io_u->error) {
 532                                 ret = -io_u->error;
 533                                 clear_io_u(td, io_u);
 534                         } else if (io_u->resid) {
 535                                 int bytes = io_u->xfer_buflen - io_u->resid;
 536
 537                                 /*
 538                                  * zero read, fail
 539                                  */
 540                                 if (!bytes) {
 541                                         td_verror(td, EIO, "full resid");
 542                                         put_io_u(td, io_u);
 543                                         break;
 544                                 }
 545
 546                                 io_u->xfer_buflen = io_u->resid;
 547                                 io_u->xfer_buf += bytes;
 548                                 io_u->offset += bytes;
 549
 550                                 if (ddir_rw(io_u->ddir))
 551                                         td->ts.short_io_u[io_u->ddir]++;
 552
 553                                 f = io_u->file;
 554                                 if (io_u->offset == f->real_file_size)
 555                                         goto sync_done;
 556
 557                                 requeue_io_u(td, &io_u);
 558                         } else {
 559 sync_done:
 560                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 561                                 if (ret < 0)
 562                                         break;
 563                         }
 564                         continue;
 565                 case FIO_Q_QUEUED:
 566                         break;
 567                 case FIO_Q_BUSY:
 568                         requeue_io_u(td, &io_u);
 569                         ret2 = td_io_commit(td);
 570                         if (ret2 < 0)
 571                                 ret = ret2;
 572                         break;
 573                 default:
 574                         assert(ret < 0);
 575                         td_verror(td, -ret, "td_io_queue");
 576                         break;
 577                 }
 578
 579                 if (break_on_this_error(td, ddir, &ret))
 580                         break;
 581
 582                 /*
 583                  * if we can queue more, do so. but check if there are
 584                  * completed io_u's first. Note that we can get BUSY even
 585                  * without IO queued, if the system is resource starved.
 586                  */
 587 reap:
 588                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 589                 if (full || !td->o.iodepth_batch_complete) {
 590                         min_events = min(td->o.iodepth_batch_complete,
 591                                          td->cur_depth);
 592                         /*
 593                          * if the queue is full, we MUST reap at least 1 event
 594                          */
 595                         if (full && !min_events)
 596                                 min_events = 1;
 597
 598                         do {
 599                                 /*
 600                                  * Reap required number of io units, if any,
 601                                  * and do the verification on them through
 602                                  * the callback handler
 603                                  */
 604                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 605                                         ret = -1;
 606                                         break;
 607                                 }
 608                         } while (full && (td->cur_depth > td->o.iodepth_low));
 609                 }
 610                 if (ret < 0)
 611                         break;
 612         }
 613
 614         check_update_rusage(td);
 615
 616         if (!td->error) {
 617                 min_events = td->cur_depth;
 618
 619                 if (min_events)
 620                         ret = io_u_queued_complete(td, min_events, NULL);
 621         } else
 622                 cleanup_pending_aio(td);
 623
 624         td_set_runstate(td, TD_RUNNING);
 625
 626         dprint(FD_VERIFY, "exiting loop\n");
 627 }
 628
 629 static unsigned int exceeds_number_ios(struct thread_data *td)
 630 {
 631         unsigned long long number_ios;
 632
 633         if (!td->o.number_ios)
 634                 return 0;
 635
 636         number_ios = ddir_rw_sum(td->this_io_blocks);
 637         number_ios += td->io_u_queued + td->io_u_in_flight;
 638
 639         return number_ios >= td->o.number_ios;
 640 }
 641
 642 static int io_bytes_exceeded(struct thread_data *td)
 643 {
 644         unsigned long long bytes;
 645
 646         if (td_rw(td))
 647                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 648         else if (td_write(td))
 649                 bytes = td->this_io_bytes[DDIR_WRITE];
 650         else if (td_read(td))
 651                 bytes = td->this_io_bytes[DDIR_READ];
 652         else
 653                 bytes = td->this_io_bytes[DDIR_TRIM];
 654
 655         return bytes >= td->o.size || exceeds_number_ios(td);
 656 }
 657
 658 /*
 659  * Main IO worker function. It retrieves io_u's to process and queues
 660  * and reaps them, checking for rate and errors along the way.
 661  *
 662  * Returns number of bytes written and trimmed.
 663  */
 664 static uint64_t do_io(struct thread_data *td)
 665 {
 666         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 667         unsigned int i;
 668         int ret = 0;
 669         uint64_t total_bytes, bytes_issued = 0;
 670
 671         if (in_ramp_time(td))
 672                 td_set_runstate(td, TD_RAMP);
 673         else
 674                 td_set_runstate(td, TD_RUNNING);
 675
 676         lat_target_init(td);
 677
 678         /*
 679          * If verify_backlog is enabled, we'll run the verify in this
 680          * handler as well. For that case, we may need up to twice the
 681          * amount of bytes.
 682          */
 683         total_bytes = td->o.size;
 684         if (td->o.verify != VERIFY_NONE &&
 685            (td_write(td) && td->o.verify_backlog))
 686                 total_bytes += td->o.size;
 687
 688         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 689                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 690                 td->o.time_based) {
 691                 struct timeval comp_time;
 692                 int min_evts = 0;
 693                 struct io_u *io_u;
 694                 int ret2, full;
 695                 enum fio_ddir ddir;
 696
 697                 check_update_rusage(td);
 698
 699                 if (td->terminate || td->done)
 700                         break;
 701
 702                 update_tv_cache(td);
 703
 704                 if (runtime_exceeded(td, &td->tv_cache)) {
 705                         __update_tv_cache(td);
 706                         if (runtime_exceeded(td, &td->tv_cache)) {
 707                                 td->terminate = 1;
 708                                 break;
 709                         }
 710                 }
 711
 712                 if (flow_threshold_exceeded(td))
 713                         continue;
 714
 715                 if (bytes_issued >= total_bytes)
 716                         break;
 717
 718                 io_u = get_io_u(td);
 719                 if (IS_ERR_OR_NULL(io_u)) {
 720                         int err = PTR_ERR(io_u);
 721
 722                         io_u = NULL;
 723                         if (err == -EBUSY) {
 724                                 ret = FIO_Q_BUSY;
 725                                 goto reap;
 726                         }
 727                         if (td->o.latency_target)
 728                                 goto reap;
 729                         break;
 730                 }
 731
 732                 ddir = io_u->ddir;
 733
 734                 /*
 735                  * Add verification end_io handler if:
 736                  *      - Asked to verify (!td_rw(td))
 737                  *      - Or the io_u is from our verify list (mixed write/ver)
 738                  */
 739                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 740                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 741
 742                         if (!td->o.verify_pattern_bytes) {
 743                                 io_u->rand_seed = __rand(&td->__verify_state);
 744                                 if (sizeof(int) != sizeof(long *))
 745                                         io_u->rand_seed *= __rand(&td->__verify_state);
 746                         }
 747
 748                         if (td->o.verify_async)
 749                                 io_u->end_io = verify_io_u_async;
 750                         else
 751                                 io_u->end_io = verify_io_u;
 752                         td_set_runstate(td, TD_VERIFYING);
 753                 } else if (in_ramp_time(td))
 754                         td_set_runstate(td, TD_RAMP);
 755                 else
 756                         td_set_runstate(td, TD_RUNNING);
 757
 758                 /*
 759                  * Always log IO before it's issued, so we know the specific
 760                  * order of it. The logged unit will track when the IO has
 761                  * completed.
 762                  */
 763                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 764                     td->o.do_verify &&
 765                     td->o.verify != VERIFY_NONE &&
 766                     !td->o.experimental_verify)
 767                         log_io_piece(td, io_u);
 768
 769                 ret = td_io_queue(td, io_u);
 770                 switch (ret) {
 771                 case FIO_Q_COMPLETED:
 772                         if (io_u->error) {
 773                                 ret = -io_u->error;
 774                                 clear_io_u(td, io_u);
 775                         } else if (io_u->resid) {
 776                                 int bytes = io_u->xfer_buflen - io_u->resid;
 777                                 struct fio_file *f = io_u->file;
 778
 779                                 bytes_issued += bytes;
 780                                 /*
 781                                  * zero read, fail
 782                                  */
 783                                 if (!bytes) {
 784                                         td_verror(td, EIO, "full resid");
 785                                         put_io_u(td, io_u);
 786                                         break;
 787                                 }
 788
 789                                 io_u->xfer_buflen = io_u->resid;
 790                                 io_u->xfer_buf += bytes;
 791                                 io_u->offset += bytes;
 792
 793                                 if (ddir_rw(io_u->ddir))
 794                                         td->ts.short_io_u[io_u->ddir]++;
 795
 796                                 if (io_u->offset == f->real_file_size)
 797                                         goto sync_done;
 798
 799                                 requeue_io_u(td, &io_u);
 800                         } else {
 801 sync_done:
 802                                 if (__should_check_rate(td, DDIR_READ) ||
 803                                     __should_check_rate(td, DDIR_WRITE) ||
 804                                     __should_check_rate(td, DDIR_TRIM))
 805                                         fio_gettime(&comp_time, NULL);
 806
 807                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 808                                 if (ret < 0)
 809                                         break;
 810                                 bytes_issued += io_u->xfer_buflen;
 811                         }
 812                         break;
 813                 case FIO_Q_QUEUED:
 814                         /*
 815                          * if the engine doesn't have a commit hook,
 816                          * the io_u is really queued. if it does have such
 817                          * a hook, it has to call io_u_queued() itself.
 818                          */
 819                         if (td->io_ops->commit == NULL)
 820                                 io_u_queued(td, io_u);
 821                         bytes_issued += io_u->xfer_buflen;
 822                         break;
 823                 case FIO_Q_BUSY:
 824                         requeue_io_u(td, &io_u);
 825                         ret2 = td_io_commit(td);
 826                         if (ret2 < 0)
 827                                 ret = ret2;
 828                         break;
 829                 default:
 830                         assert(ret < 0);
 831                         put_io_u(td, io_u);
 832                         break;
 833                 }
 834
 835                 if (break_on_this_error(td, ddir, &ret))
 836                         break;
 837
 838                 /*
 839                  * See if we need to complete some commands. Note that we
 840                  * can get BUSY even without IO queued, if the system is
 841                  * resource starved.
 842                  */
 843 reap:
 844                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 845                 if (full || !td->o.iodepth_batch_complete) {
 846                         min_evts = min(td->o.iodepth_batch_complete,
 847                                         td->cur_depth);
 848                         /*
 849                          * if the queue is full, we MUST reap at least 1 event
 850                          */
 851                         if (full && !min_evts)
 852                                 min_evts = 1;
 853
 854                         if (__should_check_rate(td, DDIR_READ) ||
 855                             __should_check_rate(td, DDIR_WRITE) ||
 856                             __should_check_rate(td, DDIR_TRIM))
 857                                 fio_gettime(&comp_time, NULL);
 858
 859                         do {
 860                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 861                                 if (ret < 0)
 862                                         break;
 863
 864                         } while (full && (td->cur_depth > td->o.iodepth_low));
 865                 }
 866
 867                 if (ret < 0)
 868                         break;
 869                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 870                         continue;
 871
 872                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 873                         if (check_min_rate(td, &comp_time, bytes_done)) {
 874                                 if (exitall_on_terminate)
 875                                         fio_terminate_threads(td->groupid);
 876                                 td_verror(td, EIO, "check_min_rate");
 877                                 break;
 878                         }
 879                 }
 880                 if (!in_ramp_time(td) && td->o.latency_target)
 881                         lat_target_check(td);
 882
 883                 if (td->o.thinktime) {
 884                         unsigned long long b;
 885
 886                         b = ddir_rw_sum(td->io_blocks);
 887                         if (!(b % td->o.thinktime_blocks)) {
 888                                 int left;
 889
 890                                 io_u_quiesce(td);
 891
 892                                 if (td->o.thinktime_spin)
 893                                         usec_spin(td->o.thinktime_spin);
 894
 895                                 left = td->o.thinktime - td->o.thinktime_spin;
 896                                 if (left)
 897                                         usec_sleep(td, left);
 898                         }
 899                 }
 900         }
 901
 902         check_update_rusage(td);
 903
 904         if (td->trim_entries)
 905                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 906
 907         if (td->o.fill_device && td->error == ENOSPC) {
 908                 td->error = 0;
 909                 td->terminate = 1;
 910         }
 911         if (!td->error) {
 912                 struct fio_file *f;
 913
 914                 i = td->cur_depth;
 915                 if (i) {
 916                         ret = io_u_queued_complete(td, i, bytes_done);
 917                         if (td->o.fill_device && td->error == ENOSPC)
 918                                 td->error = 0;
 919                 }
 920
 921                 if (should_fsync(td) && td->o.end_fsync) {
 922                         td_set_runstate(td, TD_FSYNCING);
 923
 924                         for_each_file(td, f, i) {
 925                                 if (!fio_file_fsync(td, f))
 926                                         continue;
 927
 928                                 log_err("fio: end_fsync failed for file %s\n",
 929                                                                 f->file_name);
 930                         }
 931                 }
 932         } else
 933                 cleanup_pending_aio(td);
 934
 935         /*
 936          * stop job if we failed doing any IO
 937          */
 938         if (!ddir_rw_sum(td->this_io_bytes))
 939                 td->done = 1;
 940
 941         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 942 }
 943
 944 static void cleanup_io_u(struct thread_data *td)
 945 {
 946         struct io_u *io_u;
 947
 948         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 949
 950                 if (td->io_ops->io_u_free)
 951                         td->io_ops->io_u_free(td, io_u);
 952
 953                 fio_memfree(io_u, sizeof(*io_u));
 954         }
 955
 956         free_io_mem(td);
 957
 958         io_u_rexit(&td->io_u_requeues);
 959         io_u_qexit(&td->io_u_freelist);
 960         io_u_qexit(&td->io_u_all);
 961 }
 962
 963 static int init_io_u(struct thread_data *td)
 964 {
 965         struct io_u *io_u;
 966         unsigned int max_bs, min_write;
 967         int cl_align, i, max_units;
 968         int data_xfer = 1, err;
 969         char *p;
 970
 971         max_units = td->o.iodepth;
 972         max_bs = td_max_bs(td);
 973         min_write = td->o.min_bs[DDIR_WRITE];
 974         td->orig_buffer_size = (unsigned long long) max_bs
 975                                         * (unsigned long long) max_units;
 976
 977         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 978                 data_xfer = 0;
 979
 980         err = 0;
 981         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 982         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 983         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 984
 985         if (err) {
 986                 log_err("fio: failed setting up IO queues\n");
 987                 return 1;
 988         }
 989
 990         /*
 991          * if we may later need to do address alignment, then add any
 992          * possible adjustment here so that we don't cause a buffer
 993          * overflow later. this adjustment may be too much if we get
 994          * lucky and the allocator gives us an aligned address.
 995          */
 996         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 997             (td->io_ops->flags & FIO_RAWIO))
 998                 td->orig_buffer_size += page_mask + td->o.mem_align;
 999
1000         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1001                 unsigned long bs;
1002
1003                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1004                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1005         }
1006
1007         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1008                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1009                 return 1;
1010         }
1011
1012         if (data_xfer && allocate_io_mem(td))
1013                 return 1;
1014
1015         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1016             (td->io_ops->flags & FIO_RAWIO))
1017                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1018         else
1019                 p = td->orig_buffer;
1020
1021         cl_align = os_cache_line_size();
1022
1023         for (i = 0; i < max_units; i++) {
1024                 void *ptr;
1025
1026                 if (td->terminate)
1027                         return 1;
1028
1029                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1030                 if (!ptr) {
1031                         log_err("fio: unable to allocate aligned memory\n");
1032                         break;
1033                 }
1034
1035                 io_u = ptr;
1036                 memset(io_u, 0, sizeof(*io_u));
1037                 INIT_FLIST_HEAD(&io_u->verify_list);
1038                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1039
1040                 if (data_xfer) {
1041                         io_u->buf = p;
1042                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1043
1044                         if (td_write(td))
1045                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1046                         if (td_write(td) && td->o.verify_pattern_bytes) {
1047                                 /*
1048                                  * Fill the buffer with the pattern if we are
1049                                  * going to be doing writes.
1050                                  */
1051                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1052                         }
1053                 }
1054
1055                 io_u->index = i;
1056                 io_u->flags = IO_U_F_FREE;
1057                 io_u_qpush(&td->io_u_freelist, io_u);
1058
1059                 /*
1060                  * io_u never leaves this stack, used for iteration of all
1061                  * io_u buffers.
1062                  */
1063                 io_u_qpush(&td->io_u_all, io_u);
1064
1065                 if (td->io_ops->io_u_init) {
1066                         int ret = td->io_ops->io_u_init(td, io_u);
1067
1068                         if (ret) {
1069                                 log_err("fio: failed to init engine data: %d\n", ret);
1070                                 return 1;
1071                         }
1072                 }
1073
1074                 p += max_bs;
1075         }
1076
1077         return 0;
1078 }
1079
1080 static int switch_ioscheduler(struct thread_data *td)
1081 {
1082         char tmp[256], tmp2[128];
1083         FILE *f;
1084         int ret;
1085
1086         if (td->io_ops->flags & FIO_DISKLESSIO)
1087                 return 0;
1088
1089         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1090
1091         f = fopen(tmp, "r+");
1092         if (!f) {
1093                 if (errno == ENOENT) {
1094                         log_err("fio: os or kernel doesn't support IO scheduler"
1095                                 " switching\n");
1096                         return 0;
1097                 }
1098                 td_verror(td, errno, "fopen iosched");
1099                 return 1;
1100         }
1101
1102         /*
1103          * Set io scheduler.
1104          */
1105         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1106         if (ferror(f) || ret != 1) {
1107                 td_verror(td, errno, "fwrite");
1108                 fclose(f);
1109                 return 1;
1110         }
1111
1112         rewind(f);
1113
1114         /*
1115          * Read back and check that the selected scheduler is now the default.
1116          */
1117         memset(tmp, 0, sizeof(tmp));
1118         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1119         if (ferror(f) || ret < 0) {
1120                 td_verror(td, errno, "fread");
1121                 fclose(f);
1122                 return 1;
1123         }
1124
1125         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1126         if (!strstr(tmp, tmp2)) {
1127                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1128                 td_verror(td, EINVAL, "iosched_switch");
1129                 fclose(f);
1130                 return 1;
1131         }
1132
1133         fclose(f);
1134         return 0;
1135 }
1136
1137 static int keep_running(struct thread_data *td)
1138 {
1139         if (td->done)
1140                 return 0;
1141         if (td->o.time_based)
1142                 return 1;
1143         if (td->o.loops) {
1144                 td->o.loops--;
1145                 return 1;
1146         }
1147         if (exceeds_number_ios(td))
1148                 return 0;
1149
1150         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1151                 uint64_t diff;
1152
1153                 /*
1154                  * If the difference is less than the minimum IO size, we
1155                  * are done.
1156                  */
1157                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1158                 if (diff < td_max_bs(td))
1159                         return 0;
1160
1161                 if (fio_files_done(td))
1162                         return 0;
1163
1164                 return 1;
1165         }
1166
1167         return 0;
1168 }
1169
1170 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1171 {
1172         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1173         char *str;
1174
1175         str = malloc(newlen);
1176         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1177
1178         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1179         ret = system(str);
1180         if (ret == -1)
1181                 log_err("fio: exec of cmd <%s> failed\n", str);
1182
1183         free(str);
1184         return ret;
1185 }
1186
1187 /*
1188  * Dry run to compute correct state of numberio for verification.
1189  */
1190 static uint64_t do_dry_run(struct thread_data *td)
1191 {
1192         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1193
1194         td_set_runstate(td, TD_RUNNING);
1195
1196         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1197                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1198                 struct io_u *io_u;
1199                 int ret;
1200
1201                 if (td->terminate || td->done)
1202                         break;
1203
1204                 io_u = get_io_u(td);
1205                 if (!io_u)
1206                         break;
1207
1208                 io_u->flags |= IO_U_F_FLIGHT;
1209                 io_u->error = 0;
1210                 io_u->resid = 0;
1211                 if (ddir_rw(acct_ddir(io_u)))
1212                         td->io_issues[acct_ddir(io_u)]++;
1213                 if (ddir_rw(io_u->ddir)) {
1214                         io_u_mark_depth(td, 1);
1215                         td->ts.total_io_u[io_u->ddir]++;
1216                 }
1217
1218                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1219                     td->o.do_verify &&
1220                     td->o.verify != VERIFY_NONE &&
1221                     !td->o.experimental_verify)
1222                         log_io_piece(td, io_u);
1223
1224                 ret = io_u_sync_complete(td, io_u, bytes_done);
1225                 (void) ret;
1226         }
1227
1228         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1229 }
1230
1231 /*
1232  * Entry point for the thread based jobs. The process based jobs end up
1233  * here as well, after a little setup.
1234  */
1235 static void *thread_main(void *data)
1236 {
1237         unsigned long long elapsed;
1238         struct thread_data *td = data;
1239         struct thread_options *o = &td->o;
1240         pthread_condattr_t attr;
1241         int clear_state;
1242         int ret;
1243
1244         if (!o->use_thread) {
1245                 setsid();
1246                 td->pid = getpid();
1247         } else
1248                 td->pid = gettid();
1249
1250         /*
1251          * fio_time_init() may not have been called yet if running as a server
1252          */
1253         fio_time_init();
1254
1255         fio_local_clock_init(o->use_thread);
1256
1257         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1258
1259         if (is_backend)
1260                 fio_server_send_start(td);
1261
1262         INIT_FLIST_HEAD(&td->io_log_list);
1263         INIT_FLIST_HEAD(&td->io_hist_list);
1264         INIT_FLIST_HEAD(&td->verify_list);
1265         INIT_FLIST_HEAD(&td->trim_list);
1266         INIT_FLIST_HEAD(&td->next_rand_list);
1267         pthread_mutex_init(&td->io_u_lock, NULL);
1268         td->io_hist_tree = RB_ROOT;
1269
1270         pthread_condattr_init(&attr);
1271         pthread_cond_init(&td->verify_cond, &attr);
1272         pthread_cond_init(&td->free_cond, &attr);
1273
1274         td_set_runstate(td, TD_INITIALIZED);
1275         dprint(FD_MUTEX, "up startup_mutex\n");
1276         fio_mutex_up(startup_mutex);
1277         dprint(FD_MUTEX, "wait on td->mutex\n");
1278         fio_mutex_down(td->mutex);
1279         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1280
1281         /*
1282          * A new gid requires privilege, so we need to do this before setting
1283          * the uid.
1284          */
1285         if (o->gid != -1U && setgid(o->gid)) {
1286                 td_verror(td, errno, "setgid");
1287                 goto err;
1288         }
1289         if (o->uid != -1U && setuid(o->uid)) {
1290                 td_verror(td, errno, "setuid");
1291                 goto err;
1292         }
1293
1294         /*
1295          * If we have a gettimeofday() thread, make sure we exclude that
1296          * thread from this job
1297          */
1298         if (o->gtod_cpu)
1299                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1300
1301         /*
1302          * Set affinity first, in case it has an impact on the memory
1303          * allocations.
1304          */
1305         if (o->cpumask_set) {
1306                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1307                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1308                         if (!ret) {
1309                                 log_err("fio: no CPUs set\n");
1310                                 log_err("fio: Try increasing number of available CPUs\n");
1311                                 td_verror(td, EINVAL, "cpus_split");
1312                                 goto err;
1313                         }
1314                 }
1315                 ret = fio_setaffinity(td->pid, o->cpumask);
1316                 if (ret == -1) {
1317                         td_verror(td, errno, "cpu_set_affinity");
1318                         goto err;
1319                 }
1320         }
1321
1322 #ifdef CONFIG_LIBNUMA
1323         /* numa node setup */
1324         if (o->numa_cpumask_set || o->numa_memmask_set) {
1325                 int ret;
1326
1327                 if (numa_available() < 0) {
1328                         td_verror(td, errno, "Does not support NUMA API\n");
1329                         goto err;
1330                 }
1331
1332                 if (o->numa_cpumask_set) {
1333                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1334                         if (ret == -1) {
1335                                 td_verror(td, errno, \
1336                                         "numa_run_on_node_mask failed\n");
1337                                 goto err;
1338                         }
1339                 }
1340
1341                 if (o->numa_memmask_set) {
1342
1343                         switch (o->numa_mem_mode) {
1344                         case MPOL_INTERLEAVE:
1345                                 numa_set_interleave_mask(o->numa_memnodesmask);
1346                                 break;
1347                         case MPOL_BIND:
1348                                 numa_set_membind(o->numa_memnodesmask);
1349                                 break;
1350                         case MPOL_LOCAL:
1351                                 numa_set_localalloc();
1352                                 break;
1353                         case MPOL_PREFERRED:
1354                                 numa_set_preferred(o->numa_mem_prefer_node);
1355                                 break;
1356                         case MPOL_DEFAULT:
1357                         default:
1358                                 break;
1359                         }
1360
1361                 }
1362         }
1363 #endif
1364
1365         if (fio_pin_memory(td))
1366                 goto err;
1367
1368         /*
1369          * May alter parameters that init_io_u() will use, so we need to
1370          * do this first.
1371          */
1372         if (init_iolog(td))
1373                 goto err;
1374
1375         if (init_io_u(td))
1376                 goto err;
1377
1378         if (o->verify_async && verify_async_init(td))
1379                 goto err;
1380
1381         if (o->ioprio) {
1382                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1383                 if (ret == -1) {
1384                         td_verror(td, errno, "ioprio_set");
1385                         goto err;
1386                 }
1387         }
1388
1389         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1390                 goto err;
1391
1392         errno = 0;
1393         if (nice(o->nice) == -1 && errno != 0) {
1394                 td_verror(td, errno, "nice");
1395                 goto err;
1396         }
1397
1398         if (o->ioscheduler && switch_ioscheduler(td))
1399                 goto err;
1400
1401         if (!o->create_serialize && setup_files(td))
1402                 goto err;
1403
1404         if (td_io_init(td))
1405                 goto err;
1406
1407         if (init_random_map(td))
1408                 goto err;
1409
1410         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1411                 goto err;
1412
1413         if (o->pre_read) {
1414                 if (pre_read_files(td) < 0)
1415                         goto err;
1416         }
1417
1418         fio_verify_init(td);
1419
1420         fio_gettime(&td->epoch, NULL);
1421         fio_getrusage(&td->ru_start);
1422         clear_state = 0;
1423         while (keep_running(td)) {
1424                 uint64_t verify_bytes;
1425
1426                 fio_gettime(&td->start, NULL);
1427                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1428                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1429                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1430
1431                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1432                                 o->ratemin[DDIR_TRIM]) {
1433                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1434                                                 sizeof(td->bw_sample_time));
1435                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1436                                                 sizeof(td->bw_sample_time));
1437                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1438                                                 sizeof(td->bw_sample_time));
1439                 }
1440
1441                 if (clear_state)
1442                         clear_io_state(td);
1443
1444                 prune_io_piece_log(td);
1445
1446                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1447                         verify_bytes = do_dry_run(td);
1448                 else
1449                         verify_bytes = do_io(td);
1450
1451                 clear_state = 1;
1452
1453                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1454                         elapsed = utime_since_now(&td->start);
1455                         td->ts.runtime[DDIR_READ] += elapsed;
1456                 }
1457                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1458                         elapsed = utime_since_now(&td->start);
1459                         td->ts.runtime[DDIR_WRITE] += elapsed;
1460                 }
1461                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1462                         elapsed = utime_since_now(&td->start);
1463                         td->ts.runtime[DDIR_TRIM] += elapsed;
1464                 }
1465
1466                 if (td->error || td->terminate)
1467                         break;
1468
1469                 if (!o->do_verify ||
1470                     o->verify == VERIFY_NONE ||
1471                     (td->io_ops->flags & FIO_UNIDIR))
1472                         continue;
1473
1474                 clear_io_state(td);
1475
1476                 fio_gettime(&td->start, NULL);
1477
1478                 do_verify(td, verify_bytes);
1479
1480                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1481
1482                 if (td->error || td->terminate)
1483                         break;
1484         }
1485
1486         update_rusage_stat(td);
1487         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1488         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1489         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1490         td->ts.total_run_time = mtime_since_now(&td->epoch);
1491         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1492         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1493         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1494
1495         fio_unpin_memory(td);
1496
1497         fio_writeout_logs(td);
1498
1499         if (o->exec_postrun)
1500                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1501
1502         if (exitall_on_terminate)
1503                 fio_terminate_threads(td->groupid);
1504
1505 err:
1506         if (td->error)
1507                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1508                                                         td->verror);
1509
1510         if (o->verify_async)
1511                 verify_async_exit(td);
1512
1513         close_and_free_files(td);
1514         cleanup_io_u(td);
1515         close_ioengine(td);
1516         cgroup_shutdown(td, &cgroup_mnt);
1517
1518         if (o->cpumask_set) {
1519                 int ret = fio_cpuset_exit(&o->cpumask);
1520
1521                 td_verror(td, ret, "fio_cpuset_exit");
1522         }
1523
1524         /*
1525          * do this very late, it will log file closing as well
1526          */
1527         if (o->write_iolog_file)
1528                 write_iolog_close(td);
1529
1530         fio_mutex_remove(td->rusage_sem);
1531         td->rusage_sem = NULL;
1532
1533         fio_mutex_remove(td->mutex);
1534         td->mutex = NULL;
1535
1536         td_set_runstate(td, TD_EXITED);
1537         return (void *) (uintptr_t) td->error;
1538 }
1539
1540
1541 /*
1542  * We cannot pass the td data into a forked process, so attach the td and
1543  * pass it to the thread worker.
1544  */
1545 static int fork_main(int shmid, int offset)
1546 {
1547         struct thread_data *td;
1548         void *data, *ret;
1549
1550 #ifndef __hpux
1551         data = shmat(shmid, NULL, 0);
1552         if (data == (void *) -1) {
1553                 int __err = errno;
1554
1555                 perror("shmat");
1556                 return __err;
1557         }
1558 #else
1559         /*
1560          * HP-UX inherits shm mappings?
1561          */
1562         data = threads;
1563 #endif
1564
1565         td = data + offset * sizeof(struct thread_data);
1566         ret = thread_main(td);
1567         shmdt(data);
1568         return (int) (uintptr_t) ret;
1569 }
1570
1571 /*
1572  * Run over the job map and reap the threads that have exited, if any.
1573  */
1574 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1575                          unsigned int *m_rate)
1576 {
1577         struct thread_data *td;
1578         unsigned int cputhreads, realthreads, pending;
1579         int i, status, ret;
1580
1581         /*
1582          * reap exited threads (TD_EXITED -> TD_REAPED)
1583          */
1584         realthreads = pending = cputhreads = 0;
1585         for_each_td(td, i) {
1586                 int flags = 0;
1587
1588                 /*
1589                  * ->io_ops is NULL for a thread that has closed its
1590                  * io engine
1591                  */
1592                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1593                         cputhreads++;
1594                 else
1595                         realthreads++;
1596
1597                 if (!td->pid) {
1598                         pending++;
1599                         continue;
1600                 }
1601                 if (td->runstate == TD_REAPED)
1602                         continue;
1603                 if (td->o.use_thread) {
1604                         if (td->runstate == TD_EXITED) {
1605                                 td_set_runstate(td, TD_REAPED);
1606                                 goto reaped;
1607                         }
1608                         continue;
1609                 }
1610
1611                 flags = WNOHANG;
1612                 if (td->runstate == TD_EXITED)
1613                         flags = 0;
1614
1615                 /*
1616                  * check if someone quit or got killed in an unusual way
1617                  */
1618                 ret = waitpid(td->pid, &status, flags);
1619                 if (ret < 0) {
1620                         if (errno == ECHILD) {
1621                                 log_err("fio: pid=%d disappeared %d\n",
1622                                                 (int) td->pid, td->runstate);
1623                                 td->sig = ECHILD;
1624                                 td_set_runstate(td, TD_REAPED);
1625                                 goto reaped;
1626                         }
1627                         perror("waitpid");
1628                 } else if (ret == td->pid) {
1629                         if (WIFSIGNALED(status)) {
1630                                 int sig = WTERMSIG(status);
1631
1632                                 if (sig != SIGTERM && sig != SIGUSR2)
1633                                         log_err("fio: pid=%d, got signal=%d\n",
1634                                                         (int) td->pid, sig);
1635                                 td->sig = sig;
1636                                 td_set_runstate(td, TD_REAPED);
1637                                 goto reaped;
1638                         }
1639                         if (WIFEXITED(status)) {
1640                                 if (WEXITSTATUS(status) && !td->error)
1641                                         td->error = WEXITSTATUS(status);
1642
1643                                 td_set_runstate(td, TD_REAPED);
1644                                 goto reaped;
1645                         }
1646                 }
1647
1648                 /*
1649                  * thread is not dead, continue
1650                  */
1651                 pending++;
1652                 continue;
1653 reaped:
1654                 (*nr_running)--;
1655                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1656                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1657                 if (!td->pid)
1658                         pending--;
1659
1660                 if (td->error)
1661                         exit_value++;
1662
1663                 done_secs += mtime_since_now(&td->epoch) / 1000;
1664                 profile_td_exit(td);
1665         }
1666
1667         if (*nr_running == cputhreads && !pending && realthreads)
1668                 fio_terminate_threads(TERMINATE_ALL);
1669 }
1670
1671 static void do_usleep(unsigned int usecs)
1672 {
1673         check_for_running_stats();
1674         usleep(usecs);
1675 }
1676
1677 /*
1678  * Main function for kicking off and reaping jobs, as needed.
1679  */
1680 static void run_threads(void)
1681 {
1682         struct thread_data *td;
1683         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1684         uint64_t spent;
1685
1686         if (fio_gtod_offload && fio_start_gtod_thread())
1687                 return;
1688
1689         fio_idle_prof_init();
1690
1691         set_sig_handlers();
1692
1693         nr_thread = nr_process = 0;
1694         for_each_td(td, i) {
1695                 if (td->o.use_thread)
1696                         nr_thread++;
1697                 else
1698                         nr_process++;
1699         }
1700
1701         if (output_format == FIO_OUTPUT_NORMAL) {
1702                 log_info("Starting ");
1703                 if (nr_thread)
1704                         log_info("%d thread%s", nr_thread,
1705                                                 nr_thread > 1 ? "s" : "");
1706                 if (nr_process) {
1707                         if (nr_thread)
1708                                 log_info(" and ");
1709                         log_info("%d process%s", nr_process,
1710                                                 nr_process > 1 ? "es" : "");
1711                 }
1712                 log_info("\n");
1713                 fflush(stdout);
1714         }
1715
1716         todo = thread_number;
1717         nr_running = 0;
1718         nr_started = 0;
1719         m_rate = t_rate = 0;
1720
1721         for_each_td(td, i) {
1722                 print_status_init(td->thread_number - 1);
1723
1724                 if (!td->o.create_serialize)
1725                         continue;
1726
1727                 /*
1728                  * do file setup here so it happens sequentially,
1729                  * we don't want X number of threads getting their
1730                  * client data interspersed on disk
1731                  */
1732                 if (setup_files(td)) {
1733                         exit_value++;
1734                         if (td->error)
1735                                 log_err("fio: pid=%d, err=%d/%s\n",
1736                                         (int) td->pid, td->error, td->verror);
1737                         td_set_runstate(td, TD_REAPED);
1738                         todo--;
1739                 } else {
1740                         struct fio_file *f;
1741                         unsigned int j;
1742
1743                         /*
1744                          * for sharing to work, each job must always open
1745                          * its own files. so close them, if we opened them
1746                          * for creation
1747                          */
1748                         for_each_file(td, f, j) {
1749                                 if (fio_file_open(f))
1750                                         td_io_close_file(td, f);
1751                         }
1752                 }
1753         }
1754
1755         /* start idle threads before io threads start to run */
1756         fio_idle_prof_start();
1757
1758         set_genesis_time();
1759
1760         while (todo) {
1761                 struct thread_data *map[REAL_MAX_JOBS];
1762                 struct timeval this_start;
1763                 int this_jobs = 0, left;
1764
1765                 /*
1766                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1767                  */
1768                 for_each_td(td, i) {
1769                         if (td->runstate != TD_NOT_CREATED)
1770                                 continue;
1771
1772                         /*
1773                          * never got a chance to start, killed by other
1774                          * thread for some reason
1775                          */
1776                         if (td->terminate) {
1777                                 todo--;
1778                                 continue;
1779                         }
1780
1781                         if (td->o.start_delay) {
1782                                 spent = utime_since_genesis();
1783
1784                                 if (td->o.start_delay > spent)
1785                                         continue;
1786                         }
1787
1788                         if (td->o.stonewall && (nr_started || nr_running)) {
1789                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1790                                                         td->o.name);
1791                                 break;
1792                         }
1793
1794                         init_disk_util(td);
1795
1796                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1797                         td->update_rusage = 0;
1798
1799                         /*
1800                          * Set state to created. Thread will transition
1801                          * to TD_INITIALIZED when it's done setting up.
1802                          */
1803                         td_set_runstate(td, TD_CREATED);
1804                         map[this_jobs++] = td;
1805                         nr_started++;
1806
1807                         if (td->o.use_thread) {
1808                                 int ret;
1809
1810                                 dprint(FD_PROCESS, "will pthread_create\n");
1811                                 ret = pthread_create(&td->thread, NULL,
1812                                                         thread_main, td);
1813                                 if (ret) {
1814                                         log_err("pthread_create: %s\n",
1815                                                         strerror(ret));
1816                                         nr_started--;
1817                                         break;
1818                                 }
1819                                 ret = pthread_detach(td->thread);
1820                                 if (ret)
1821                                         log_err("pthread_detach: %s",
1822                                                         strerror(ret));
1823                         } else {
1824                                 pid_t pid;
1825                                 dprint(FD_PROCESS, "will fork\n");
1826                                 pid = fork();
1827                                 if (!pid) {
1828                                         int ret = fork_main(shm_id, i);
1829
1830                                         _exit(ret);
1831                                 } else if (i == fio_debug_jobno)
1832                                         *fio_debug_jobp = pid;
1833                         }
1834                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1835                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1836                                 log_err("fio: job startup hung? exiting.\n");
1837                                 fio_terminate_threads(TERMINATE_ALL);
1838                                 fio_abort = 1;
1839                                 nr_started--;
1840                                 break;
1841                         }
1842                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1843                 }
1844
1845                 /*
1846                  * Wait for the started threads to transition to
1847                  * TD_INITIALIZED.
1848                  */
1849                 fio_gettime(&this_start, NULL);
1850                 left = this_jobs;
1851                 while (left && !fio_abort) {
1852                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1853                                 break;
1854
1855                         do_usleep(100000);
1856
1857                         for (i = 0; i < this_jobs; i++) {
1858                                 td = map[i];
1859                                 if (!td)
1860                                         continue;
1861                                 if (td->runstate == TD_INITIALIZED) {
1862                                         map[i] = NULL;
1863                                         left--;
1864                                 } else if (td->runstate >= TD_EXITED) {
1865                                         map[i] = NULL;
1866                                         left--;
1867                                         todo--;
1868                                         nr_running++; /* work-around... */
1869                                 }
1870                         }
1871                 }
1872
1873                 if (left) {
1874                         log_err("fio: %d job%s failed to start\n", left,
1875                                         left > 1 ? "s" : "");
1876                         for (i = 0; i < this_jobs; i++) {
1877                                 td = map[i];
1878                                 if (!td)
1879                                         continue;
1880                                 kill(td->pid, SIGTERM);
1881                         }
1882                         break;
1883                 }
1884
1885                 /*
1886                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1887                  */
1888                 for_each_td(td, i) {
1889                         if (td->runstate != TD_INITIALIZED)
1890                                 continue;
1891
1892                         if (in_ramp_time(td))
1893                                 td_set_runstate(td, TD_RAMP);
1894                         else
1895                                 td_set_runstate(td, TD_RUNNING);
1896                         nr_running++;
1897                         nr_started--;
1898                         m_rate += ddir_rw_sum(td->o.ratemin);
1899                         t_rate += ddir_rw_sum(td->o.rate);
1900                         todo--;
1901                         fio_mutex_up(td->mutex);
1902                 }
1903
1904                 reap_threads(&nr_running, &t_rate, &m_rate);
1905
1906                 if (todo)
1907                         do_usleep(100000);
1908         }
1909
1910         while (nr_running) {
1911                 reap_threads(&nr_running, &t_rate, &m_rate);
1912                 do_usleep(10000);
1913         }
1914
1915         fio_idle_prof_stop();
1916
1917         update_io_ticks();
1918 }
1919
1920 void wait_for_disk_thread_exit(void)
1921 {
1922         fio_mutex_down(disk_thread_mutex);
1923 }
1924
1925 static void free_disk_util(void)
1926 {
1927         disk_util_start_exit();
1928         wait_for_disk_thread_exit();
1929         disk_util_prune_entries();
1930 }
1931
1932 static void *disk_thread_main(void *data)
1933 {
1934         int ret = 0;
1935
1936         fio_mutex_up(startup_mutex);
1937
1938         while (threads && !ret) {
1939                 usleep(DISK_UTIL_MSEC * 1000);
1940                 if (!threads)
1941                         break;
1942                 ret = update_io_ticks();
1943
1944                 if (!is_backend)
1945                         print_thread_status();
1946         }
1947
1948         fio_mutex_up(disk_thread_mutex);
1949         return NULL;
1950 }
1951
1952 static int create_disk_util_thread(void)
1953 {
1954         int ret;
1955
1956         setup_disk_util();
1957
1958         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1959
1960         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1961         if (ret) {
1962                 fio_mutex_remove(disk_thread_mutex);
1963                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1964                 return 1;
1965         }
1966
1967         ret = pthread_detach(disk_util_thread);
1968         if (ret) {
1969                 fio_mutex_remove(disk_thread_mutex);
1970                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1971                 return 1;
1972         }
1973
1974         dprint(FD_MUTEX, "wait on startup_mutex\n");
1975         fio_mutex_down(startup_mutex);
1976         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1977         return 0;
1978 }
1979
1980 int fio_backend(void)
1981 {
1982         struct thread_data *td;
1983         int i;
1984
1985         if (exec_profile) {
1986                 if (load_profile(exec_profile))
1987                         return 1;
1988                 free(exec_profile);
1989                 exec_profile = NULL;
1990         }
1991         if (!thread_number)
1992                 return 0;
1993
1994         if (write_bw_log) {
1995                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1996                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1997                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1998         }
1999
2000         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2001         if (startup_mutex == NULL)
2002                 return 1;
2003
2004         set_genesis_time();
2005         stat_init();
2006         create_disk_util_thread();
2007
2008         cgroup_list = smalloc(sizeof(*cgroup_list));
2009         INIT_FLIST_HEAD(cgroup_list);
2010
2011         run_threads();
2012
2013         if (!fio_abort) {
2014                 show_run_stats();
2015                 if (write_bw_log) {
2016                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2017                         __finish_log(agg_io_log[DDIR_WRITE],
2018                                         "agg-write_bw.log");
2019                         __finish_log(agg_io_log[DDIR_TRIM],
2020                                         "agg-write_bw.log");
2021                 }
2022         }
2023
2024         for_each_td(td, i)
2025                 fio_options_free(td);
2026
2027         free_disk_util();
2028         cgroup_kill(cgroup_list);
2029         sfree(cgroup_list);
2030         sfree(cgroup_mnt);
2031
2032         fio_mutex_remove(startup_mutex);
2033         fio_mutex_remove(disk_thread_mutex);
2034         stat_exit();
2035         return exit_value;
2036 }