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 fio_mutex *writeout_mutex;
  61 static struct flist_head *cgroup_list;
  62 static char *cgroup_mnt;
  63 static int exit_value;
  64 static volatile int fio_abort;
  65 static unsigned int nr_process = 0;
  66 static unsigned int nr_thread = 0;
  67
  68 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  69
  70 int groupid = 0;
  71 unsigned int thread_number = 0;
  72 unsigned int stat_number = 0;
  73 int shm_id = 0;
  74 int temp_stall_ts;
  75 unsigned long done_secs = 0;
  76 volatile int disk_util_exit = 0;
  77
  78 #define PAGE_ALIGN(buf) \
  79         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  80
  81 #define JOB_START_TIMEOUT       (5 * 1000)
  82
  83 static void sig_int(int sig)
  84 {
  85         if (threads) {
  86                 if (is_backend)
  87                         fio_server_got_signal(sig);
  88                 else {
  89                         log_info("\nfio: terminating on signal %d\n", sig);
  90                         fflush(stdout);
  91                         exit_value = 128;
  92                 }
  93
  94                 fio_terminate_threads(TERMINATE_ALL);
  95         }
  96 }
  97
  98 static void sig_show_status(int sig)
  99 {
 100         show_running_run_stats();
 101 }
 102
 103 static void set_sig_handlers(void)
 104 {
 105         struct sigaction act;
 106
 107         memset(&act, 0, sizeof(act));
 108         act.sa_handler = sig_int;
 109         act.sa_flags = SA_RESTART;
 110         sigaction(SIGINT, &act, NULL);
 111
 112         memset(&act, 0, sizeof(act));
 113         act.sa_handler = sig_int;
 114         act.sa_flags = SA_RESTART;
 115         sigaction(SIGTERM, &act, NULL);
 116
 117 /* Windows uses SIGBREAK as a quit signal from other applications */
 118 #ifdef WIN32
 119         memset(&act, 0, sizeof(act));
 120         act.sa_handler = sig_int;
 121         act.sa_flags = SA_RESTART;
 122         sigaction(SIGBREAK, &act, NULL);
 123 #endif
 124
 125         memset(&act, 0, sizeof(act));
 126         act.sa_handler = sig_show_status;
 127         act.sa_flags = SA_RESTART;
 128         sigaction(SIGUSR1, &act, NULL);
 129
 130         if (is_backend) {
 131                 memset(&act, 0, sizeof(act));
 132                 act.sa_handler = sig_int;
 133                 act.sa_flags = SA_RESTART;
 134                 sigaction(SIGPIPE, &act, NULL);
 135         }
 136 }
 137
 138 /*
 139  * Check if we are above the minimum rate given.
 140  */
 141 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 142                             enum fio_ddir ddir)
 143 {
 144         unsigned long long bytes = 0;
 145         unsigned long iops = 0;
 146         unsigned long spent;
 147         unsigned long rate;
 148         unsigned int ratemin = 0;
 149         unsigned int rate_iops = 0;
 150         unsigned int rate_iops_min = 0;
 151
 152         assert(ddir_rw(ddir));
 153
 154         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 155                 return 0;
 156
 157         /*
 158          * allow a 2 second settle period in the beginning
 159          */
 160         if (mtime_since(&td->start, now) < 2000)
 161                 return 0;
 162
 163         iops += td->this_io_blocks[ddir];
 164         bytes += td->this_io_bytes[ddir];
 165         ratemin += td->o.ratemin[ddir];
 166         rate_iops += td->o.rate_iops[ddir];
 167         rate_iops_min += td->o.rate_iops_min[ddir];
 168
 169         /*
 170          * if rate blocks is set, sample is running
 171          */
 172         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 173                 spent = mtime_since(&td->lastrate[ddir], now);
 174                 if (spent < td->o.ratecycle)
 175                         return 0;
 176
 177                 if (td->o.rate[ddir]) {
 178                         /*
 179                          * check bandwidth specified rate
 180                          */
 181                         if (bytes < td->rate_bytes[ddir]) {
 182                                 log_err("%s: min rate %u not met\n", td->o.name,
 183                                                                 ratemin);
 184                                 return 1;
 185                         } else {
 186                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 187                                 if (rate < ratemin ||
 188                                     bytes < td->rate_bytes[ddir]) {
 189                                         log_err("%s: min rate %u not met, got"
 190                                                 " %luKB/sec\n", td->o.name,
 191                                                         ratemin, rate);
 192                                         return 1;
 193                                 }
 194                         }
 195                 } else {
 196                         /*
 197                          * checks iops specified rate
 198                          */
 199                         if (iops < rate_iops) {
 200                                 log_err("%s: min iops rate %u not met\n",
 201                                                 td->o.name, rate_iops);
 202                                 return 1;
 203                         } else {
 204                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 205                                 if (rate < rate_iops_min ||
 206                                     iops < td->rate_blocks[ddir]) {
 207                                         log_err("%s: min iops rate %u not met,"
 208                                                 " got %lu\n", td->o.name,
 209                                                         rate_iops_min, rate);
 210                                 }
 211                         }
 212                 }
 213         }
 214
 215         td->rate_bytes[ddir] = bytes;
 216         td->rate_blocks[ddir] = iops;
 217         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 218         return 0;
 219 }
 220
 221 static int check_min_rate(struct thread_data *td, struct timeval *now,
 222                           uint64_t *bytes_done)
 223 {
 224         int ret = 0;
 225
 226         if (bytes_done[DDIR_READ])
 227                 ret |= __check_min_rate(td, now, DDIR_READ);
 228         if (bytes_done[DDIR_WRITE])
 229                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 230         if (bytes_done[DDIR_TRIM])
 231                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 232
 233         return ret;
 234 }
 235
 236 /*
 237  * When job exits, we can cancel the in-flight IO if we are using async
 238  * io. Attempt to do so.
 239  */
 240 static void cleanup_pending_aio(struct thread_data *td)
 241 {
 242         int r;
 243
 244         /*
 245          * get immediately available events, if any
 246          */
 247         r = io_u_queued_complete(td, 0, NULL);
 248         if (r < 0)
 249                 return;
 250
 251         /*
 252          * now cancel remaining active events
 253          */
 254         if (td->io_ops->cancel) {
 255                 struct io_u *io_u;
 256                 int i;
 257
 258                 io_u_qiter(&td->io_u_all, io_u, i) {
 259                         if (io_u->flags & IO_U_F_FLIGHT) {
 260                                 r = td->io_ops->cancel(td, io_u);
 261                                 if (!r)
 262                                         put_io_u(td, io_u);
 263                         }
 264                 }
 265         }
 266
 267         if (td->cur_depth)
 268                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 269 }
 270
 271 /*
 272  * Helper to handle the final sync of a file. Works just like the normal
 273  * io path, just does everything sync.
 274  */
 275 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 276 {
 277         struct io_u *io_u = __get_io_u(td);
 278         int ret;
 279
 280         if (!io_u)
 281                 return 1;
 282
 283         io_u->ddir = DDIR_SYNC;
 284         io_u->file = f;
 285
 286         if (td_io_prep(td, io_u)) {
 287                 put_io_u(td, io_u);
 288                 return 1;
 289         }
 290
 291 requeue:
 292         ret = td_io_queue(td, io_u);
 293         if (ret < 0) {
 294                 td_verror(td, io_u->error, "td_io_queue");
 295                 put_io_u(td, io_u);
 296                 return 1;
 297         } else if (ret == FIO_Q_QUEUED) {
 298                 if (io_u_queued_complete(td, 1, NULL) < 0)
 299                         return 1;
 300         } else if (ret == FIO_Q_COMPLETED) {
 301                 if (io_u->error) {
 302                         td_verror(td, io_u->error, "td_io_queue");
 303                         return 1;
 304                 }
 305
 306                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 307                         return 1;
 308         } else if (ret == FIO_Q_BUSY) {
 309                 if (td_io_commit(td))
 310                         return 1;
 311                 goto requeue;
 312         }
 313
 314         return 0;
 315 }
 316
 317 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 318 {
 319         int ret;
 320
 321         if (fio_file_open(f))
 322                 return fio_io_sync(td, f);
 323
 324         if (td_io_open_file(td, f))
 325                 return 1;
 326
 327         ret = fio_io_sync(td, f);
 328         td_io_close_file(td, f);
 329         return ret;
 330 }
 331
 332 static inline void __update_tv_cache(struct thread_data *td)
 333 {
 334         fio_gettime(&td->tv_cache, NULL);
 335 }
 336
 337 static inline void update_tv_cache(struct thread_data *td)
 338 {
 339         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 340                 __update_tv_cache(td);
 341 }
 342
 343 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 344 {
 345         if (in_ramp_time(td))
 346                 return 0;
 347         if (!td->o.timeout)
 348                 return 0;
 349         if (utime_since(&td->epoch, t) >= td->o.timeout)
 350                 return 1;
 351
 352         return 0;
 353 }
 354
 355 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 356                                int *retptr)
 357 {
 358         int ret = *retptr;
 359
 360         if (ret < 0 || td->error) {
 361                 int err = td->error;
 362                 enum error_type_bit eb;
 363
 364                 if (ret < 0)
 365                         err = -ret;
 366
 367                 eb = td_error_type(ddir, err);
 368                 if (!(td->o.continue_on_error & (1 << eb)))
 369                         return 1;
 370
 371                 if (td_non_fatal_error(td, eb, err)) {
 372                         /*
 373                          * Continue with the I/Os in case of
 374                          * a non fatal error.
 375                          */
 376                         update_error_count(td, err);
 377                         td_clear_error(td);
 378                         *retptr = 0;
 379                         return 0;
 380                 } else if (td->o.fill_device && err == ENOSPC) {
 381                         /*
 382                          * We expect to hit this error if
 383                          * fill_device option is set.
 384                          */
 385                         td_clear_error(td);
 386                         td->terminate = 1;
 387                         return 1;
 388                 } else {
 389                         /*
 390                          * Stop the I/O in case of a fatal
 391                          * error.
 392                          */
 393                         update_error_count(td, err);
 394                         return 1;
 395                 }
 396         }
 397
 398         return 0;
 399 }
 400
 401 static void check_update_rusage(struct thread_data *td)
 402 {
 403         if (td->update_rusage) {
 404                 td->update_rusage = 0;
 405                 update_rusage_stat(td);
 406                 fio_mutex_up(td->rusage_sem);
 407         }
 408 }
 409
 410 /*
 411  * The main verify engine. Runs over the writes we previously submitted,
 412  * reads the blocks back in, and checks the crc/md5 of the data.
 413  */
 414 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 415 {
 416         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 417         struct fio_file *f;
 418         struct io_u *io_u;
 419         int ret, min_events;
 420         unsigned int i;
 421
 422         dprint(FD_VERIFY, "starting loop\n");
 423
 424         /*
 425          * sync io first and invalidate cache, to make sure we really
 426          * read from disk.
 427          */
 428         for_each_file(td, f, i) {
 429                 if (!fio_file_open(f))
 430                         continue;
 431                 if (fio_io_sync(td, f))
 432                         break;
 433                 if (file_invalidate_cache(td, f))
 434                         break;
 435         }
 436
 437         check_update_rusage(td);
 438
 439         if (td->error)
 440                 return;
 441
 442         td_set_runstate(td, TD_VERIFYING);
 443
 444         io_u = NULL;
 445         while (!td->terminate) {
 446                 enum fio_ddir ddir;
 447                 int ret2, full;
 448
 449                 update_tv_cache(td);
 450                 check_update_rusage(td);
 451
 452                 if (runtime_exceeded(td, &td->tv_cache)) {
 453                         __update_tv_cache(td);
 454                         if (runtime_exceeded(td, &td->tv_cache)) {
 455                                 td->terminate = 1;
 456                                 break;
 457                         }
 458                 }
 459
 460                 if (flow_threshold_exceeded(td))
 461                         continue;
 462
 463                 if (!td->o.experimental_verify) {
 464                         io_u = __get_io_u(td);
 465                         if (!io_u)
 466                                 break;
 467
 468                         if (get_next_verify(td, io_u)) {
 469                                 put_io_u(td, io_u);
 470                                 break;
 471                         }
 472
 473                         if (td_io_prep(td, io_u)) {
 474                                 put_io_u(td, io_u);
 475                                 break;
 476                         }
 477                 } else {
 478                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 479                                 break;
 480
 481                         while ((io_u = get_io_u(td)) != NULL) {
 482                                 if (IS_ERR(io_u)) {
 483                                         io_u = NULL;
 484                                         ret = FIO_Q_BUSY;
 485                                         goto reap;
 486                                 }
 487
 488                                 /*
 489                                  * We are only interested in the places where
 490                                  * we wrote or trimmed IOs. Turn those into
 491                                  * reads for verification purposes.
 492                                  */
 493                                 if (io_u->ddir == DDIR_READ) {
 494                                         /*
 495                                          * Pretend we issued it for rwmix
 496                                          * accounting
 497                                          */
 498                                         td->io_issues[DDIR_READ]++;
 499                                         put_io_u(td, io_u);
 500                                         continue;
 501                                 } else if (io_u->ddir == DDIR_TRIM) {
 502                                         io_u->ddir = DDIR_READ;
 503                                         io_u->flags |= IO_U_F_TRIMMED;
 504                                         break;
 505                                 } else if (io_u->ddir == DDIR_WRITE) {
 506                                         io_u->ddir = DDIR_READ;
 507                                         break;
 508                                 } else {
 509                                         put_io_u(td, io_u);
 510                                         continue;
 511                                 }
 512                         }
 513
 514                         if (!io_u)
 515                                 break;
 516                 }
 517
 518                 if (td->o.verify_async)
 519                         io_u->end_io = verify_io_u_async;
 520                 else
 521                         io_u->end_io = verify_io_u;
 522
 523                 ddir = io_u->ddir;
 524
 525                 ret = td_io_queue(td, io_u);
 526                 switch (ret) {
 527                 case FIO_Q_COMPLETED:
 528                         if (io_u->error) {
 529                                 ret = -io_u->error;
 530                                 clear_io_u(td, io_u);
 531                         } else if (io_u->resid) {
 532                                 int bytes = io_u->xfer_buflen - io_u->resid;
 533
 534                                 /*
 535                                  * zero read, fail
 536                                  */
 537                                 if (!bytes) {
 538                                         td_verror(td, EIO, "full resid");
 539                                         put_io_u(td, io_u);
 540                                         break;
 541                                 }
 542
 543                                 io_u->xfer_buflen = io_u->resid;
 544                                 io_u->xfer_buf += bytes;
 545                                 io_u->offset += bytes;
 546
 547                                 if (ddir_rw(io_u->ddir))
 548                                         td->ts.short_io_u[io_u->ddir]++;
 549
 550                                 f = io_u->file;
 551                                 if (io_u->offset == f->real_file_size)
 552                                         goto sync_done;
 553
 554                                 requeue_io_u(td, &io_u);
 555                         } else {
 556 sync_done:
 557                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 558                                 if (ret < 0)
 559                                         break;
 560                         }
 561                         continue;
 562                 case FIO_Q_QUEUED:
 563                         break;
 564                 case FIO_Q_BUSY:
 565                         requeue_io_u(td, &io_u);
 566                         ret2 = td_io_commit(td);
 567                         if (ret2 < 0)
 568                                 ret = ret2;
 569                         break;
 570                 default:
 571                         assert(ret < 0);
 572                         td_verror(td, -ret, "td_io_queue");
 573                         break;
 574                 }
 575
 576                 if (break_on_this_error(td, ddir, &ret))
 577                         break;
 578
 579                 /*
 580                  * if we can queue more, do so. but check if there are
 581                  * completed io_u's first. Note that we can get BUSY even
 582                  * without IO queued, if the system is resource starved.
 583                  */
 584 reap:
 585                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 586                 if (full || !td->o.iodepth_batch_complete) {
 587                         min_events = min(td->o.iodepth_batch_complete,
 588                                          td->cur_depth);
 589                         /*
 590                          * if the queue is full, we MUST reap at least 1 event
 591                          */
 592                         if (full && !min_events)
 593                                 min_events = 1;
 594
 595                         do {
 596                                 /*
 597                                  * Reap required number of io units, if any,
 598                                  * and do the verification on them through
 599                                  * the callback handler
 600                                  */
 601                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 602                                         ret = -1;
 603                                         break;
 604                                 }
 605                         } while (full && (td->cur_depth > td->o.iodepth_low));
 606                 }
 607                 if (ret < 0)
 608                         break;
 609         }
 610
 611         check_update_rusage(td);
 612
 613         if (!td->error) {
 614                 min_events = td->cur_depth;
 615
 616                 if (min_events)
 617                         ret = io_u_queued_complete(td, min_events, NULL);
 618         } else
 619                 cleanup_pending_aio(td);
 620
 621         td_set_runstate(td, TD_RUNNING);
 622
 623         dprint(FD_VERIFY, "exiting loop\n");
 624 }
 625
 626 static int io_bytes_exceeded(struct thread_data *td)
 627 {
 628         unsigned long long bytes;
 629
 630         if (td_rw(td))
 631                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 632         else if (td_write(td))
 633                 bytes = td->this_io_bytes[DDIR_WRITE];
 634         else if (td_read(td))
 635                 bytes = td->this_io_bytes[DDIR_READ];
 636         else
 637                 bytes = td->this_io_bytes[DDIR_TRIM];
 638
 639         return bytes >= td->o.size;
 640 }
 641
 642 /*
 643  * Main IO worker function. It retrieves io_u's to process and queues
 644  * and reaps them, checking for rate and errors along the way.
 645  *
 646  * Returns number of bytes written and trimmed.
 647  */
 648 static uint64_t do_io(struct thread_data *td)
 649 {
 650         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 651         unsigned int i;
 652         int ret = 0;
 653         uint64_t total_bytes, bytes_issued = 0;
 654
 655         if (in_ramp_time(td))
 656                 td_set_runstate(td, TD_RAMP);
 657         else
 658                 td_set_runstate(td, TD_RUNNING);
 659
 660         lat_target_init(td);
 661
 662         /*
 663          * If verify_backlog is enabled, we'll run the verify in this
 664          * handler as well. For that case, we may need up to twice the
 665          * amount of bytes.
 666          */
 667         total_bytes = td->o.size;
 668         if (td->o.verify != VERIFY_NONE &&
 669            (td_write(td) && td->o.verify_backlog))
 670                 total_bytes += td->o.size;
 671
 672         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 673                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 674                 td->o.time_based) {
 675                 struct timeval comp_time;
 676                 int min_evts = 0;
 677                 struct io_u *io_u;
 678                 int ret2, full;
 679                 enum fio_ddir ddir;
 680
 681                 check_update_rusage(td);
 682
 683                 if (td->terminate || td->done)
 684                         break;
 685
 686                 update_tv_cache(td);
 687
 688                 if (runtime_exceeded(td, &td->tv_cache)) {
 689                         __update_tv_cache(td);
 690                         if (runtime_exceeded(td, &td->tv_cache)) {
 691                                 td->terminate = 1;
 692                                 break;
 693                         }
 694                 }
 695
 696                 if (flow_threshold_exceeded(td))
 697                         continue;
 698
 699                 if (bytes_issued >= total_bytes)
 700                         break;
 701
 702                 io_u = get_io_u(td);
 703                 if (IS_ERR_OR_NULL(io_u)) {
 704                         int err = PTR_ERR(io_u);
 705
 706                         io_u = NULL;
 707                         if (err == -EBUSY) {
 708                                 ret = FIO_Q_BUSY;
 709                                 goto reap;
 710                         }
 711                         if (td->o.latency_target)
 712                                 goto reap;
 713                         break;
 714                 }
 715
 716                 ddir = io_u->ddir;
 717
 718                 /*
 719                  * Add verification end_io handler if:
 720                  *      - Asked to verify (!td_rw(td))
 721                  *      - Or the io_u is from our verify list (mixed write/ver)
 722                  */
 723                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 724                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 725
 726                         if (!td->o.verify_pattern_bytes) {
 727                                 io_u->rand_seed = __rand(&td->__verify_state);
 728                                 if (sizeof(int) != sizeof(long *))
 729                                         io_u->rand_seed *= __rand(&td->__verify_state);
 730                         }
 731
 732                         if (td->o.verify_async)
 733                                 io_u->end_io = verify_io_u_async;
 734                         else
 735                                 io_u->end_io = verify_io_u;
 736                         td_set_runstate(td, TD_VERIFYING);
 737                 } else if (in_ramp_time(td))
 738                         td_set_runstate(td, TD_RAMP);
 739                 else
 740                         td_set_runstate(td, TD_RUNNING);
 741
 742                 /*
 743                  * Always log IO before it's issued, so we know the specific
 744                  * order of it. The logged unit will track when the IO has
 745                  * completed.
 746                  */
 747                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 748                     td->o.do_verify &&
 749                     td->o.verify != VERIFY_NONE &&
 750                     !td->o.experimental_verify)
 751                         log_io_piece(td, io_u);
 752
 753                 ret = td_io_queue(td, io_u);
 754                 switch (ret) {
 755                 case FIO_Q_COMPLETED:
 756                         if (io_u->error) {
 757                                 ret = -io_u->error;
 758                                 clear_io_u(td, io_u);
 759                         } else if (io_u->resid) {
 760                                 int bytes = io_u->xfer_buflen - io_u->resid;
 761                                 struct fio_file *f = io_u->file;
 762
 763                                 bytes_issued += bytes;
 764                                 /*
 765                                  * zero read, fail
 766                                  */
 767                                 if (!bytes) {
 768                                         td_verror(td, EIO, "full resid");
 769                                         put_io_u(td, io_u);
 770                                         break;
 771                                 }
 772
 773                                 io_u->xfer_buflen = io_u->resid;
 774                                 io_u->xfer_buf += bytes;
 775                                 io_u->offset += bytes;
 776
 777                                 if (ddir_rw(io_u->ddir))
 778                                         td->ts.short_io_u[io_u->ddir]++;
 779
 780                                 if (io_u->offset == f->real_file_size)
 781                                         goto sync_done;
 782
 783                                 requeue_io_u(td, &io_u);
 784                         } else {
 785 sync_done:
 786                                 if (__should_check_rate(td, DDIR_READ) ||
 787                                     __should_check_rate(td, DDIR_WRITE) ||
 788                                     __should_check_rate(td, DDIR_TRIM))
 789                                         fio_gettime(&comp_time, NULL);
 790
 791                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 792                                 if (ret < 0)
 793                                         break;
 794                                 bytes_issued += io_u->xfer_buflen;
 795                         }
 796                         break;
 797                 case FIO_Q_QUEUED:
 798                         /*
 799                          * if the engine doesn't have a commit hook,
 800                          * the io_u is really queued. if it does have such
 801                          * a hook, it has to call io_u_queued() itself.
 802                          */
 803                         if (td->io_ops->commit == NULL)
 804                                 io_u_queued(td, io_u);
 805                         bytes_issued += io_u->xfer_buflen;
 806                         break;
 807                 case FIO_Q_BUSY:
 808                         requeue_io_u(td, &io_u);
 809                         ret2 = td_io_commit(td);
 810                         if (ret2 < 0)
 811                                 ret = ret2;
 812                         break;
 813                 default:
 814                         assert(ret < 0);
 815                         put_io_u(td, io_u);
 816                         break;
 817                 }
 818
 819                 if (break_on_this_error(td, ddir, &ret))
 820                         break;
 821
 822                 /*
 823                  * See if we need to complete some commands. Note that we
 824                  * can get BUSY even without IO queued, if the system is
 825                  * resource starved.
 826                  */
 827 reap:
 828                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 829                 if (full || !td->o.iodepth_batch_complete) {
 830                         min_evts = min(td->o.iodepth_batch_complete,
 831                                         td->cur_depth);
 832                         /*
 833                          * if the queue is full, we MUST reap at least 1 event
 834                          */
 835                         if (full && !min_evts)
 836                                 min_evts = 1;
 837
 838                         if (__should_check_rate(td, DDIR_READ) ||
 839                             __should_check_rate(td, DDIR_WRITE) ||
 840                             __should_check_rate(td, DDIR_TRIM))
 841                                 fio_gettime(&comp_time, NULL);
 842
 843                         do {
 844                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 845                                 if (ret < 0)
 846                                         break;
 847
 848                         } while (full && (td->cur_depth > td->o.iodepth_low));
 849                 }
 850
 851                 if (ret < 0)
 852                         break;
 853                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 854                         continue;
 855
 856                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 857                         if (check_min_rate(td, &comp_time, bytes_done)) {
 858                                 if (exitall_on_terminate)
 859                                         fio_terminate_threads(td->groupid);
 860                                 td_verror(td, EIO, "check_min_rate");
 861                                 break;
 862                         }
 863                 }
 864                 if (!in_ramp_time(td) && td->o.latency_target)
 865                         lat_target_check(td);
 866
 867                 if (td->o.thinktime) {
 868                         unsigned long long b;
 869
 870                         b = ddir_rw_sum(td->io_blocks);
 871                         if (!(b % td->o.thinktime_blocks)) {
 872                                 int left;
 873
 874                                 io_u_quiesce(td);
 875
 876                                 if (td->o.thinktime_spin)
 877                                         usec_spin(td->o.thinktime_spin);
 878
 879                                 left = td->o.thinktime - td->o.thinktime_spin;
 880                                 if (left)
 881                                         usec_sleep(td, left);
 882                         }
 883                 }
 884         }
 885
 886         check_update_rusage(td);
 887
 888         if (td->trim_entries)
 889                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 890
 891         if (td->o.fill_device && td->error == ENOSPC) {
 892                 td->error = 0;
 893                 td->terminate = 1;
 894         }
 895         if (!td->error) {
 896                 struct fio_file *f;
 897
 898                 i = td->cur_depth;
 899                 if (i) {
 900                         ret = io_u_queued_complete(td, i, bytes_done);
 901                         if (td->o.fill_device && td->error == ENOSPC)
 902                                 td->error = 0;
 903                 }
 904
 905                 if (should_fsync(td) && td->o.end_fsync) {
 906                         td_set_runstate(td, TD_FSYNCING);
 907
 908                         for_each_file(td, f, i) {
 909                                 if (!fio_file_fsync(td, f))
 910                                         continue;
 911
 912                                 log_err("fio: end_fsync failed for file %s\n",
 913                                                                 f->file_name);
 914                         }
 915                 }
 916         } else
 917                 cleanup_pending_aio(td);
 918
 919         /*
 920          * stop job if we failed doing any IO
 921          */
 922         if (!ddir_rw_sum(td->this_io_bytes))
 923                 td->done = 1;
 924
 925         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 926 }
 927
 928 static void cleanup_io_u(struct thread_data *td)
 929 {
 930         struct io_u *io_u;
 931
 932         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 933
 934                 if (td->io_ops->io_u_free)
 935                         td->io_ops->io_u_free(td, io_u);
 936
 937                 fio_memfree(io_u, sizeof(*io_u));
 938         }
 939
 940         free_io_mem(td);
 941
 942         io_u_rexit(&td->io_u_requeues);
 943         io_u_qexit(&td->io_u_freelist);
 944         io_u_qexit(&td->io_u_all);
 945 }
 946
 947 static int init_io_u(struct thread_data *td)
 948 {
 949         struct io_u *io_u;
 950         unsigned int max_bs, min_write;
 951         int cl_align, i, max_units;
 952         int data_xfer = 1, err;
 953         char *p;
 954
 955         max_units = td->o.iodepth;
 956         max_bs = td_max_bs(td);
 957         min_write = td->o.min_bs[DDIR_WRITE];
 958         td->orig_buffer_size = (unsigned long long) max_bs
 959                                         * (unsigned long long) max_units;
 960
 961         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 962                 data_xfer = 0;
 963
 964         err = 0;
 965         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 966         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 967         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 968
 969         if (err) {
 970                 log_err("fio: failed setting up IO queues\n");
 971                 return 1;
 972         }
 973
 974         /*
 975          * if we may later need to do address alignment, then add any
 976          * possible adjustment here so that we don't cause a buffer
 977          * overflow later. this adjustment may be too much if we get
 978          * lucky and the allocator gives us an aligned address.
 979          */
 980         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 981             (td->io_ops->flags & FIO_RAWIO))
 982                 td->orig_buffer_size += page_mask + td->o.mem_align;
 983
 984         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 985                 unsigned long bs;
 986
 987                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 988                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 989         }
 990
 991         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 992                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 993                 return 1;
 994         }
 995
 996         if (data_xfer && allocate_io_mem(td))
 997                 return 1;
 998
 999         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1000             (td->io_ops->flags & FIO_RAWIO))
1001                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1002         else
1003                 p = td->orig_buffer;
1004
1005         cl_align = os_cache_line_size();
1006
1007         for (i = 0; i < max_units; i++) {
1008                 void *ptr;
1009
1010                 if (td->terminate)
1011                         return 1;
1012
1013                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1014                 if (!ptr) {
1015                         log_err("fio: unable to allocate aligned memory\n");
1016                         break;
1017                 }
1018
1019                 io_u = ptr;
1020                 memset(io_u, 0, sizeof(*io_u));
1021                 INIT_FLIST_HEAD(&io_u->verify_list);
1022                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1023
1024                 if (data_xfer) {
1025                         io_u->buf = p;
1026                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1027
1028                         if (td_write(td))
1029                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1030                         if (td_write(td) && td->o.verify_pattern_bytes) {
1031                                 /*
1032                                  * Fill the buffer with the pattern if we are
1033                                  * going to be doing writes.
1034                                  */
1035                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1036                         }
1037                 }
1038
1039                 io_u->index = i;
1040                 io_u->flags = IO_U_F_FREE;
1041                 io_u_qpush(&td->io_u_freelist, io_u);
1042
1043                 /*
1044                  * io_u never leaves this stack, used for iteration of all
1045                  * io_u buffers.
1046                  */
1047                 io_u_qpush(&td->io_u_all, io_u);
1048
1049                 if (td->io_ops->io_u_init) {
1050                         int ret = td->io_ops->io_u_init(td, io_u);
1051
1052                         if (ret) {
1053                                 log_err("fio: failed to init engine data: %d\n", ret);
1054                                 return 1;
1055                         }
1056                 }
1057
1058                 p += max_bs;
1059         }
1060
1061         return 0;
1062 }
1063
1064 static int switch_ioscheduler(struct thread_data *td)
1065 {
1066         char tmp[256], tmp2[128];
1067         FILE *f;
1068         int ret;
1069
1070         if (td->io_ops->flags & FIO_DISKLESSIO)
1071                 return 0;
1072
1073         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1074
1075         f = fopen(tmp, "r+");
1076         if (!f) {
1077                 if (errno == ENOENT) {
1078                         log_err("fio: os or kernel doesn't support IO scheduler"
1079                                 " switching\n");
1080                         return 0;
1081                 }
1082                 td_verror(td, errno, "fopen iosched");
1083                 return 1;
1084         }
1085
1086         /*
1087          * Set io scheduler.
1088          */
1089         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1090         if (ferror(f) || ret != 1) {
1091                 td_verror(td, errno, "fwrite");
1092                 fclose(f);
1093                 return 1;
1094         }
1095
1096         rewind(f);
1097
1098         /*
1099          * Read back and check that the selected scheduler is now the default.
1100          */
1101         ret = fread(tmp, 1, sizeof(tmp), f);
1102         if (ferror(f) || ret < 0) {
1103                 td_verror(td, errno, "fread");
1104                 fclose(f);
1105                 return 1;
1106         }
1107
1108         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1109         if (!strstr(tmp, tmp2)) {
1110                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1111                 td_verror(td, EINVAL, "iosched_switch");
1112                 fclose(f);
1113                 return 1;
1114         }
1115
1116         fclose(f);
1117         return 0;
1118 }
1119
1120 static int keep_running(struct thread_data *td)
1121 {
1122         if (td->done)
1123                 return 0;
1124         if (td->o.time_based)
1125                 return 1;
1126         if (td->o.loops) {
1127                 td->o.loops--;
1128                 return 1;
1129         }
1130
1131         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1132                 uint64_t diff;
1133
1134                 /*
1135                  * If the difference is less than the minimum IO size, we
1136                  * are done.
1137                  */
1138                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1139                 if (diff < td_max_bs(td))
1140                         return 0;
1141
1142                 if (fio_files_done(td))
1143                         return 0;
1144
1145                 return 1;
1146         }
1147
1148         return 0;
1149 }
1150
1151 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1152 {
1153         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1154         char *str;
1155
1156         str = malloc(newlen);
1157         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1158
1159         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1160         ret = system(str);
1161         if (ret == -1)
1162                 log_err("fio: exec of cmd <%s> failed\n", str);
1163
1164         free(str);
1165         return ret;
1166 }
1167
1168 /*
1169  * Dry run to compute correct state of numberio for verification.
1170  */
1171 static uint64_t do_dry_run(struct thread_data *td)
1172 {
1173         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1174
1175         td_set_runstate(td, TD_RUNNING);
1176
1177         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1178                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1179                 struct io_u *io_u;
1180                 int ret;
1181
1182                 if (td->terminate || td->done)
1183                         break;
1184
1185                 io_u = get_io_u(td);
1186                 if (!io_u)
1187                         break;
1188
1189                 io_u->flags |= IO_U_F_FLIGHT;
1190                 io_u->error = 0;
1191                 io_u->resid = 0;
1192                 if (ddir_rw(acct_ddir(io_u)))
1193                         td->io_issues[acct_ddir(io_u)]++;
1194                 if (ddir_rw(io_u->ddir)) {
1195                         io_u_mark_depth(td, 1);
1196                         td->ts.total_io_u[io_u->ddir]++;
1197                 }
1198
1199                 ret = io_u_sync_complete(td, io_u, bytes_done);
1200                 (void) ret;
1201         }
1202
1203         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1204 }
1205
1206 /*
1207  * Entry point for the thread based jobs. The process based jobs end up
1208  * here as well, after a little setup.
1209  */
1210 static void *thread_main(void *data)
1211 {
1212         unsigned long long elapsed;
1213         struct thread_data *td = data;
1214         struct thread_options *o = &td->o;
1215         pthread_condattr_t attr;
1216         int clear_state;
1217         int ret;
1218
1219         if (!o->use_thread) {
1220                 setsid();
1221                 td->pid = getpid();
1222         } else
1223                 td->pid = gettid();
1224
1225         /*
1226          * fio_time_init() may not have been called yet if running as a server
1227          */
1228         fio_time_init();
1229
1230         fio_local_clock_init(o->use_thread);
1231
1232         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1233
1234         if (is_backend)
1235                 fio_server_send_start(td);
1236
1237         INIT_FLIST_HEAD(&td->io_log_list);
1238         INIT_FLIST_HEAD(&td->io_hist_list);
1239         INIT_FLIST_HEAD(&td->verify_list);
1240         INIT_FLIST_HEAD(&td->trim_list);
1241         INIT_FLIST_HEAD(&td->next_rand_list);
1242         pthread_mutex_init(&td->io_u_lock, NULL);
1243         td->io_hist_tree = RB_ROOT;
1244
1245         pthread_condattr_init(&attr);
1246         pthread_cond_init(&td->verify_cond, &attr);
1247         pthread_cond_init(&td->free_cond, &attr);
1248
1249         td_set_runstate(td, TD_INITIALIZED);
1250         dprint(FD_MUTEX, "up startup_mutex\n");
1251         fio_mutex_up(startup_mutex);
1252         dprint(FD_MUTEX, "wait on td->mutex\n");
1253         fio_mutex_down(td->mutex);
1254         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1255
1256         /*
1257          * A new gid requires privilege, so we need to do this before setting
1258          * the uid.
1259          */
1260         if (o->gid != -1U && setgid(o->gid)) {
1261                 td_verror(td, errno, "setgid");
1262                 goto err;
1263         }
1264         if (o->uid != -1U && setuid(o->uid)) {
1265                 td_verror(td, errno, "setuid");
1266                 goto err;
1267         }
1268
1269         /*
1270          * If we have a gettimeofday() thread, make sure we exclude that
1271          * thread from this job
1272          */
1273         if (o->gtod_cpu)
1274                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1275
1276         /*
1277          * Set affinity first, in case it has an impact on the memory
1278          * allocations.
1279          */
1280         if (o->cpumask_set) {
1281                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1282                         ret = fio_cpus_split(&o->cpumask, td->thread_number);
1283                         if (!ret) {
1284                                 log_err("fio: no CPUs set\n");
1285                                 log_err("fio: Try increasing number of available CPUs\n");
1286                                 td_verror(td, EINVAL, "cpus_split");
1287                                 goto err;
1288                         }
1289                 }
1290                 ret = fio_setaffinity(td->pid, o->cpumask);
1291                 if (ret == -1) {
1292                         td_verror(td, errno, "cpu_set_affinity");
1293                         goto err;
1294                 }
1295         }
1296
1297 #ifdef CONFIG_LIBNUMA
1298         /* numa node setup */
1299         if (o->numa_cpumask_set || o->numa_memmask_set) {
1300                 int ret;
1301
1302                 if (numa_available() < 0) {
1303                         td_verror(td, errno, "Does not support NUMA API\n");
1304                         goto err;
1305                 }
1306
1307                 if (o->numa_cpumask_set) {
1308                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1309                         if (ret == -1) {
1310                                 td_verror(td, errno, \
1311                                         "numa_run_on_node_mask failed\n");
1312                                 goto err;
1313                         }
1314                 }
1315
1316                 if (o->numa_memmask_set) {
1317
1318                         switch (o->numa_mem_mode) {
1319                         case MPOL_INTERLEAVE:
1320                                 numa_set_interleave_mask(o->numa_memnodesmask);
1321                                 break;
1322                         case MPOL_BIND:
1323                                 numa_set_membind(o->numa_memnodesmask);
1324                                 break;
1325                         case MPOL_LOCAL:
1326                                 numa_set_localalloc();
1327                                 break;
1328                         case MPOL_PREFERRED:
1329                                 numa_set_preferred(o->numa_mem_prefer_node);
1330                                 break;
1331                         case MPOL_DEFAULT:
1332                         default:
1333                                 break;
1334                         }
1335
1336                 }
1337         }
1338 #endif
1339
1340         if (fio_pin_memory(td))
1341                 goto err;
1342
1343         /*
1344          * May alter parameters that init_io_u() will use, so we need to
1345          * do this first.
1346          */
1347         if (init_iolog(td))
1348                 goto err;
1349
1350         if (init_io_u(td))
1351                 goto err;
1352
1353         if (o->verify_async && verify_async_init(td))
1354                 goto err;
1355
1356         if (o->ioprio) {
1357                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1358                 if (ret == -1) {
1359                         td_verror(td, errno, "ioprio_set");
1360                         goto err;
1361                 }
1362         }
1363
1364         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1365                 goto err;
1366
1367         errno = 0;
1368         if (nice(o->nice) == -1 && errno != 0) {
1369                 td_verror(td, errno, "nice");
1370                 goto err;
1371         }
1372
1373         if (o->ioscheduler && switch_ioscheduler(td))
1374                 goto err;
1375
1376         if (!o->create_serialize && setup_files(td))
1377                 goto err;
1378
1379         if (td_io_init(td))
1380                 goto err;
1381
1382         if (init_random_map(td))
1383                 goto err;
1384
1385         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1386                 goto err;
1387
1388         if (o->pre_read) {
1389                 if (pre_read_files(td) < 0)
1390                         goto err;
1391         }
1392
1393         fio_verify_init(td);
1394
1395         fio_gettime(&td->epoch, NULL);
1396         fio_getrusage(&td->ru_start);
1397         clear_state = 0;
1398         while (keep_running(td)) {
1399                 uint64_t verify_bytes;
1400
1401                 fio_gettime(&td->start, NULL);
1402                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1403                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1404                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1405
1406                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1407                                 o->ratemin[DDIR_TRIM]) {
1408                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1409                                                 sizeof(td->bw_sample_time));
1410                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1411                                                 sizeof(td->bw_sample_time));
1412                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1413                                                 sizeof(td->bw_sample_time));
1414                 }
1415
1416                 if (clear_state)
1417                         clear_io_state(td);
1418
1419                 prune_io_piece_log(td);
1420
1421                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1422                         verify_bytes = do_dry_run(td);
1423                 else
1424                         verify_bytes = do_io(td);
1425
1426                 clear_state = 1;
1427
1428                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1429                         elapsed = utime_since_now(&td->start);
1430                         td->ts.runtime[DDIR_READ] += elapsed;
1431                 }
1432                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1433                         elapsed = utime_since_now(&td->start);
1434                         td->ts.runtime[DDIR_WRITE] += elapsed;
1435                 }
1436                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1437                         elapsed = utime_since_now(&td->start);
1438                         td->ts.runtime[DDIR_TRIM] += elapsed;
1439                 }
1440
1441                 if (td->error || td->terminate)
1442                         break;
1443
1444                 if (!o->do_verify ||
1445                     o->verify == VERIFY_NONE ||
1446                     (td->io_ops->flags & FIO_UNIDIR))
1447                         continue;
1448
1449                 clear_io_state(td);
1450
1451                 fio_gettime(&td->start, NULL);
1452
1453                 do_verify(td, verify_bytes);
1454
1455                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1456
1457                 if (td->error || td->terminate)
1458                         break;
1459         }
1460
1461         update_rusage_stat(td);
1462         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1463         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1464         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1465         td->ts.total_run_time = mtime_since_now(&td->epoch);
1466         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1467         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1468         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1469
1470         fio_unpin_memory(td);
1471
1472         fio_mutex_down(writeout_mutex);
1473         finalize_logs(td);
1474         if (td->bw_log) {
1475                 if (o->bw_log_file) {
1476                         finish_log_named(td, td->bw_log,
1477                                                 o->bw_log_file, "bw");
1478                 } else
1479                         finish_log(td, td->bw_log, "bw");
1480         }
1481         if (td->lat_log) {
1482                 if (o->lat_log_file) {
1483                         finish_log_named(td, td->lat_log,
1484                                                 o->lat_log_file, "lat");
1485                 } else
1486                         finish_log(td, td->lat_log, "lat");
1487         }
1488         if (td->slat_log) {
1489                 if (o->lat_log_file) {
1490                         finish_log_named(td, td->slat_log,
1491                                                 o->lat_log_file, "slat");
1492                 } else
1493                         finish_log(td, td->slat_log, "slat");
1494         }
1495         if (td->clat_log) {
1496                 if (o->lat_log_file) {
1497                         finish_log_named(td, td->clat_log,
1498                                                 o->lat_log_file, "clat");
1499                 } else
1500                         finish_log(td, td->clat_log, "clat");
1501         }
1502         if (td->iops_log) {
1503                 if (o->iops_log_file) {
1504                         finish_log_named(td, td->iops_log,
1505                                                 o->iops_log_file, "iops");
1506                 } else
1507                         finish_log(td, td->iops_log, "iops");
1508         }
1509
1510         fio_mutex_up(writeout_mutex);
1511         if (o->exec_postrun)
1512                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1513
1514         if (exitall_on_terminate)
1515                 fio_terminate_threads(td->groupid);
1516
1517 err:
1518         if (td->error)
1519                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1520                                                         td->verror);
1521
1522         if (o->verify_async)
1523                 verify_async_exit(td);
1524
1525         close_and_free_files(td);
1526         cleanup_io_u(td);
1527         close_ioengine(td);
1528         cgroup_shutdown(td, &cgroup_mnt);
1529
1530         if (o->cpumask_set) {
1531                 int ret = fio_cpuset_exit(&o->cpumask);
1532
1533                 td_verror(td, ret, "fio_cpuset_exit");
1534         }
1535
1536         /*
1537          * do this very late, it will log file closing as well
1538          */
1539         if (o->write_iolog_file)
1540                 write_iolog_close(td);
1541
1542         fio_mutex_remove(td->rusage_sem);
1543         td->rusage_sem = NULL;
1544
1545         fio_mutex_remove(td->mutex);
1546         td->mutex = NULL;
1547
1548         td_set_runstate(td, TD_EXITED);
1549         return (void *) (uintptr_t) td->error;
1550 }
1551
1552
1553 /*
1554  * We cannot pass the td data into a forked process, so attach the td and
1555  * pass it to the thread worker.
1556  */
1557 static int fork_main(int shmid, int offset)
1558 {
1559         struct thread_data *td;
1560         void *data, *ret;
1561
1562 #ifndef __hpux
1563         data = shmat(shmid, NULL, 0);
1564         if (data == (void *) -1) {
1565                 int __err = errno;
1566
1567                 perror("shmat");
1568                 return __err;
1569         }
1570 #else
1571         /*
1572          * HP-UX inherits shm mappings?
1573          */
1574         data = threads;
1575 #endif
1576
1577         td = data + offset * sizeof(struct thread_data);
1578         ret = thread_main(td);
1579         shmdt(data);
1580         return (int) (uintptr_t) ret;
1581 }
1582
1583 /*
1584  * Run over the job map and reap the threads that have exited, if any.
1585  */
1586 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1587                          unsigned int *m_rate)
1588 {
1589         struct thread_data *td;
1590         unsigned int cputhreads, realthreads, pending;
1591         int i, status, ret;
1592
1593         /*
1594          * reap exited threads (TD_EXITED -> TD_REAPED)
1595          */
1596         realthreads = pending = cputhreads = 0;
1597         for_each_td(td, i) {
1598                 int flags = 0;
1599
1600                 /*
1601                  * ->io_ops is NULL for a thread that has closed its
1602                  * io engine
1603                  */
1604                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1605                         cputhreads++;
1606                 else
1607                         realthreads++;
1608
1609                 if (!td->pid) {
1610                         pending++;
1611                         continue;
1612                 }
1613                 if (td->runstate == TD_REAPED)
1614                         continue;
1615                 if (td->o.use_thread) {
1616                         if (td->runstate == TD_EXITED) {
1617                                 td_set_runstate(td, TD_REAPED);
1618                                 goto reaped;
1619                         }
1620                         continue;
1621                 }
1622
1623                 flags = WNOHANG;
1624                 if (td->runstate == TD_EXITED)
1625                         flags = 0;
1626
1627                 /*
1628                  * check if someone quit or got killed in an unusual way
1629                  */
1630                 ret = waitpid(td->pid, &status, flags);
1631                 if (ret < 0) {
1632                         if (errno == ECHILD) {
1633                                 log_err("fio: pid=%d disappeared %d\n",
1634                                                 (int) td->pid, td->runstate);
1635                                 td->sig = ECHILD;
1636                                 td_set_runstate(td, TD_REAPED);
1637                                 goto reaped;
1638                         }
1639                         perror("waitpid");
1640                 } else if (ret == td->pid) {
1641                         if (WIFSIGNALED(status)) {
1642                                 int sig = WTERMSIG(status);
1643
1644                                 if (sig != SIGTERM && sig != SIGUSR2)
1645                                         log_err("fio: pid=%d, got signal=%d\n",
1646                                                         (int) td->pid, sig);
1647                                 td->sig = sig;
1648                                 td_set_runstate(td, TD_REAPED);
1649                                 goto reaped;
1650                         }
1651                         if (WIFEXITED(status)) {
1652                                 if (WEXITSTATUS(status) && !td->error)
1653                                         td->error = WEXITSTATUS(status);
1654
1655                                 td_set_runstate(td, TD_REAPED);
1656                                 goto reaped;
1657                         }
1658                 }
1659
1660                 /*
1661                  * thread is not dead, continue
1662                  */
1663                 pending++;
1664                 continue;
1665 reaped:
1666                 (*nr_running)--;
1667                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1668                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1669                 if (!td->pid)
1670                         pending--;
1671
1672                 if (td->error)
1673                         exit_value++;
1674
1675                 done_secs += mtime_since_now(&td->epoch) / 1000;
1676                 profile_td_exit(td);
1677         }
1678
1679         if (*nr_running == cputhreads && !pending && realthreads)
1680                 fio_terminate_threads(TERMINATE_ALL);
1681 }
1682
1683 static void do_usleep(unsigned int usecs)
1684 {
1685         check_for_running_stats();
1686         usleep(usecs);
1687 }
1688
1689 /*
1690  * Main function for kicking off and reaping jobs, as needed.
1691  */
1692 static void run_threads(void)
1693 {
1694         struct thread_data *td;
1695         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1696         uint64_t spent;
1697
1698         if (fio_gtod_offload && fio_start_gtod_thread())
1699                 return;
1700
1701         fio_idle_prof_init();
1702
1703         set_sig_handlers();
1704
1705         nr_thread = nr_process = 0;
1706         for_each_td(td, i) {
1707                 if (td->o.use_thread)
1708                         nr_thread++;
1709                 else
1710                         nr_process++;
1711         }
1712
1713         if (output_format == FIO_OUTPUT_NORMAL) {
1714                 log_info("Starting ");
1715                 if (nr_thread)
1716                         log_info("%d thread%s", nr_thread,
1717                                                 nr_thread > 1 ? "s" : "");
1718                 if (nr_process) {
1719                         if (nr_thread)
1720                                 log_info(" and ");
1721                         log_info("%d process%s", nr_process,
1722                                                 nr_process > 1 ? "es" : "");
1723                 }
1724                 log_info("\n");
1725                 fflush(stdout);
1726         }
1727
1728         todo = thread_number;
1729         nr_running = 0;
1730         nr_started = 0;
1731         m_rate = t_rate = 0;
1732
1733         for_each_td(td, i) {
1734                 print_status_init(td->thread_number - 1);
1735
1736                 if (!td->o.create_serialize)
1737                         continue;
1738
1739                 /*
1740                  * do file setup here so it happens sequentially,
1741                  * we don't want X number of threads getting their
1742                  * client data interspersed on disk
1743                  */
1744                 if (setup_files(td)) {
1745                         exit_value++;
1746                         if (td->error)
1747                                 log_err("fio: pid=%d, err=%d/%s\n",
1748                                         (int) td->pid, td->error, td->verror);
1749                         td_set_runstate(td, TD_REAPED);
1750                         todo--;
1751                 } else {
1752                         struct fio_file *f;
1753                         unsigned int j;
1754
1755                         /*
1756                          * for sharing to work, each job must always open
1757                          * its own files. so close them, if we opened them
1758                          * for creation
1759                          */
1760                         for_each_file(td, f, j) {
1761                                 if (fio_file_open(f))
1762                                         td_io_close_file(td, f);
1763                         }
1764                 }
1765         }
1766
1767         /* start idle threads before io threads start to run */
1768         fio_idle_prof_start();
1769
1770         set_genesis_time();
1771
1772         while (todo) {
1773                 struct thread_data *map[REAL_MAX_JOBS];
1774                 struct timeval this_start;
1775                 int this_jobs = 0, left;
1776
1777                 /*
1778                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1779                  */
1780                 for_each_td(td, i) {
1781                         if (td->runstate != TD_NOT_CREATED)
1782                                 continue;
1783
1784                         /*
1785                          * never got a chance to start, killed by other
1786                          * thread for some reason
1787                          */
1788                         if (td->terminate) {
1789                                 todo--;
1790                                 continue;
1791                         }
1792
1793                         if (td->o.start_delay) {
1794                                 spent = utime_since_genesis();
1795
1796                                 if (td->o.start_delay > spent)
1797                                         continue;
1798                         }
1799
1800                         if (td->o.stonewall && (nr_started || nr_running)) {
1801                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1802                                                         td->o.name);
1803                                 break;
1804                         }
1805
1806                         init_disk_util(td);
1807
1808                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1809                         td->update_rusage = 0;
1810
1811                         /*
1812                          * Set state to created. Thread will transition
1813                          * to TD_INITIALIZED when it's done setting up.
1814                          */
1815                         td_set_runstate(td, TD_CREATED);
1816                         map[this_jobs++] = td;
1817                         nr_started++;
1818
1819                         if (td->o.use_thread) {
1820                                 int ret;
1821
1822                                 dprint(FD_PROCESS, "will pthread_create\n");
1823                                 ret = pthread_create(&td->thread, NULL,
1824                                                         thread_main, td);
1825                                 if (ret) {
1826                                         log_err("pthread_create: %s\n",
1827                                                         strerror(ret));
1828                                         nr_started--;
1829                                         break;
1830                                 }
1831                                 ret = pthread_detach(td->thread);
1832                                 if (ret)
1833                                         log_err("pthread_detach: %s",
1834                                                         strerror(ret));
1835                         } else {
1836                                 pid_t pid;
1837                                 dprint(FD_PROCESS, "will fork\n");
1838                                 pid = fork();
1839                                 if (!pid) {
1840                                         int ret = fork_main(shm_id, i);
1841
1842                                         _exit(ret);
1843                                 } else if (i == fio_debug_jobno)
1844                                         *fio_debug_jobp = pid;
1845                         }
1846                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1847                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1848                                 log_err("fio: job startup hung? exiting.\n");
1849                                 fio_terminate_threads(TERMINATE_ALL);
1850                                 fio_abort = 1;
1851                                 nr_started--;
1852                                 break;
1853                         }
1854                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1855                 }
1856
1857                 /*
1858                  * Wait for the started threads to transition to
1859                  * TD_INITIALIZED.
1860                  */
1861                 fio_gettime(&this_start, NULL);
1862                 left = this_jobs;
1863                 while (left && !fio_abort) {
1864                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1865                                 break;
1866
1867                         do_usleep(100000);
1868
1869                         for (i = 0; i < this_jobs; i++) {
1870                                 td = map[i];
1871                                 if (!td)
1872                                         continue;
1873                                 if (td->runstate == TD_INITIALIZED) {
1874                                         map[i] = NULL;
1875                                         left--;
1876                                 } else if (td->runstate >= TD_EXITED) {
1877                                         map[i] = NULL;
1878                                         left--;
1879                                         todo--;
1880                                         nr_running++; /* work-around... */
1881                                 }
1882                         }
1883                 }
1884
1885                 if (left) {
1886                         log_err("fio: %d job%s failed to start\n", left,
1887                                         left > 1 ? "s" : "");
1888                         for (i = 0; i < this_jobs; i++) {
1889                                 td = map[i];
1890                                 if (!td)
1891                                         continue;
1892                                 kill(td->pid, SIGTERM);
1893                         }
1894                         break;
1895                 }
1896
1897                 /*
1898                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1899                  */
1900                 for_each_td(td, i) {
1901                         if (td->runstate != TD_INITIALIZED)
1902                                 continue;
1903
1904                         if (in_ramp_time(td))
1905                                 td_set_runstate(td, TD_RAMP);
1906                         else
1907                                 td_set_runstate(td, TD_RUNNING);
1908                         nr_running++;
1909                         nr_started--;
1910                         m_rate += ddir_rw_sum(td->o.ratemin);
1911                         t_rate += ddir_rw_sum(td->o.rate);
1912                         todo--;
1913                         fio_mutex_up(td->mutex);
1914                 }
1915
1916                 reap_threads(&nr_running, &t_rate, &m_rate);
1917
1918                 if (todo)
1919                         do_usleep(100000);
1920         }
1921
1922         while (nr_running) {
1923                 reap_threads(&nr_running, &t_rate, &m_rate);
1924                 do_usleep(10000);
1925         }
1926
1927         fio_idle_prof_stop();
1928
1929         update_io_ticks();
1930 }
1931
1932 void wait_for_disk_thread_exit(void)
1933 {
1934         fio_mutex_down(disk_thread_mutex);
1935 }
1936
1937 static void free_disk_util(void)
1938 {
1939         disk_util_start_exit();
1940         wait_for_disk_thread_exit();
1941         disk_util_prune_entries();
1942 }
1943
1944 static void *disk_thread_main(void *data)
1945 {
1946         int ret = 0;
1947
1948         fio_mutex_up(startup_mutex);
1949
1950         while (threads && !ret) {
1951                 usleep(DISK_UTIL_MSEC * 1000);
1952                 if (!threads)
1953                         break;
1954                 ret = update_io_ticks();
1955
1956                 if (!is_backend)
1957                         print_thread_status();
1958         }
1959
1960         fio_mutex_up(disk_thread_mutex);
1961         return NULL;
1962 }
1963
1964 static int create_disk_util_thread(void)
1965 {
1966         int ret;
1967
1968         setup_disk_util();
1969
1970         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1971
1972         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1973         if (ret) {
1974                 fio_mutex_remove(disk_thread_mutex);
1975                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1976                 return 1;
1977         }
1978
1979         ret = pthread_detach(disk_util_thread);
1980         if (ret) {
1981                 fio_mutex_remove(disk_thread_mutex);
1982                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1983                 return 1;
1984         }
1985
1986         dprint(FD_MUTEX, "wait on startup_mutex\n");
1987         fio_mutex_down(startup_mutex);
1988         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1989         return 0;
1990 }
1991
1992 int fio_backend(void)
1993 {
1994         struct thread_data *td;
1995         int i;
1996
1997         if (exec_profile) {
1998                 if (load_profile(exec_profile))
1999                         return 1;
2000                 free(exec_profile);
2001                 exec_profile = NULL;
2002         }
2003         if (!thread_number)
2004                 return 0;
2005
2006         if (write_bw_log) {
2007                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
2008                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
2009                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2010         }
2011
2012         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2013         if (startup_mutex == NULL)
2014                 return 1;
2015         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2016         if (writeout_mutex == NULL)
2017                 return 1;
2018
2019         set_genesis_time();
2020         stat_init();
2021         create_disk_util_thread();
2022
2023         cgroup_list = smalloc(sizeof(*cgroup_list));
2024         INIT_FLIST_HEAD(cgroup_list);
2025
2026         run_threads();
2027
2028         if (!fio_abort) {
2029                 show_run_stats();
2030                 if (write_bw_log) {
2031                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2032                         __finish_log(agg_io_log[DDIR_WRITE],
2033                                         "agg-write_bw.log");
2034                         __finish_log(agg_io_log[DDIR_TRIM],
2035                                         "agg-write_bw.log");
2036                 }
2037         }
2038
2039         for_each_td(td, i)
2040                 fio_options_free(td);
2041
2042         free_disk_util();
2043         cgroup_kill(cgroup_list);
2044         sfree(cgroup_list);
2045         sfree(cgroup_mnt);
2046
2047         fio_mutex_remove(startup_mutex);
2048         fio_mutex_remove(writeout_mutex);
2049         fio_mutex_remove(disk_thread_mutex);
2050         stat_exit();
2051         return exit_value;
2052 }