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                                 if (spent)
 208                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 209                                 else
 210                                         rate = 0;
 211
 212                                 if (rate < rate_iops_min ||
 213                                     iops < td->rate_blocks[ddir]) {
 214                                         log_err("%s: min iops rate %u not met,"
 215                                                 " got %lu\n", td->o.name,
 216                                                         rate_iops_min, rate);
 217                                 }
 218                         }
 219                 }
 220         }
 221
 222         td->rate_bytes[ddir] = bytes;
 223         td->rate_blocks[ddir] = iops;
 224         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 225         return 0;
 226 }
 227
 228 static int check_min_rate(struct thread_data *td, struct timeval *now,
 229                           uint64_t *bytes_done)
 230 {
 231         int ret = 0;
 232
 233         if (bytes_done[DDIR_READ])
 234                 ret |= __check_min_rate(td, now, DDIR_READ);
 235         if (bytes_done[DDIR_WRITE])
 236                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 237         if (bytes_done[DDIR_TRIM])
 238                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 239
 240         return ret;
 241 }
 242
 243 /*
 244  * When job exits, we can cancel the in-flight IO if we are using async
 245  * io. Attempt to do so.
 246  */
 247 static void cleanup_pending_aio(struct thread_data *td)
 248 {
 249         int r;
 250
 251         /*
 252          * get immediately available events, if any
 253          */
 254         r = io_u_queued_complete(td, 0, NULL);
 255         if (r < 0)
 256                 return;
 257
 258         /*
 259          * now cancel remaining active events
 260          */
 261         if (td->io_ops->cancel) {
 262                 struct io_u *io_u;
 263                 int i;
 264
 265                 io_u_qiter(&td->io_u_all, io_u, i) {
 266                         if (io_u->flags & IO_U_F_FLIGHT) {
 267                                 r = td->io_ops->cancel(td, io_u);
 268                                 if (!r)
 269                                         put_io_u(td, io_u);
 270                         }
 271                 }
 272         }
 273
 274         if (td->cur_depth)
 275                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 276 }
 277
 278 /*
 279  * Helper to handle the final sync of a file. Works just like the normal
 280  * io path, just does everything sync.
 281  */
 282 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 283 {
 284         struct io_u *io_u = __get_io_u(td);
 285         int ret;
 286
 287         if (!io_u)
 288                 return 1;
 289
 290         io_u->ddir = DDIR_SYNC;
 291         io_u->file = f;
 292
 293         if (td_io_prep(td, io_u)) {
 294                 put_io_u(td, io_u);
 295                 return 1;
 296         }
 297
 298 requeue:
 299         ret = td_io_queue(td, io_u);
 300         if (ret < 0) {
 301                 td_verror(td, io_u->error, "td_io_queue");
 302                 put_io_u(td, io_u);
 303                 return 1;
 304         } else if (ret == FIO_Q_QUEUED) {
 305                 if (io_u_queued_complete(td, 1, NULL) < 0)
 306                         return 1;
 307         } else if (ret == FIO_Q_COMPLETED) {
 308                 if (io_u->error) {
 309                         td_verror(td, io_u->error, "td_io_queue");
 310                         return 1;
 311                 }
 312
 313                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 314                         return 1;
 315         } else if (ret == FIO_Q_BUSY) {
 316                 if (td_io_commit(td))
 317                         return 1;
 318                 goto requeue;
 319         }
 320
 321         return 0;
 322 }
 323
 324 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 325 {
 326         int ret;
 327
 328         if (fio_file_open(f))
 329                 return fio_io_sync(td, f);
 330
 331         if (td_io_open_file(td, f))
 332                 return 1;
 333
 334         ret = fio_io_sync(td, f);
 335         td_io_close_file(td, f);
 336         return ret;
 337 }
 338
 339 static inline void __update_tv_cache(struct thread_data *td)
 340 {
 341         fio_gettime(&td->tv_cache, NULL);
 342 }
 343
 344 static inline void update_tv_cache(struct thread_data *td)
 345 {
 346         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 347                 __update_tv_cache(td);
 348 }
 349
 350 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 351 {
 352         if (in_ramp_time(td))
 353                 return 0;
 354         if (!td->o.timeout)
 355                 return 0;
 356         if (utime_since(&td->epoch, t) >= td->o.timeout)
 357                 return 1;
 358
 359         return 0;
 360 }
 361
 362 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 363                                int *retptr)
 364 {
 365         int ret = *retptr;
 366
 367         if (ret < 0 || td->error) {
 368                 int err = td->error;
 369                 enum error_type_bit eb;
 370
 371                 if (ret < 0)
 372                         err = -ret;
 373
 374                 eb = td_error_type(ddir, err);
 375                 if (!(td->o.continue_on_error & (1 << eb)))
 376                         return 1;
 377
 378                 if (td_non_fatal_error(td, eb, err)) {
 379                         /*
 380                          * Continue with the I/Os in case of
 381                          * a non fatal error.
 382                          */
 383                         update_error_count(td, err);
 384                         td_clear_error(td);
 385                         *retptr = 0;
 386                         return 0;
 387                 } else if (td->o.fill_device && err == ENOSPC) {
 388                         /*
 389                          * We expect to hit this error if
 390                          * fill_device option is set.
 391                          */
 392                         td_clear_error(td);
 393                         td->terminate = 1;
 394                         return 1;
 395                 } else {
 396                         /*
 397                          * Stop the I/O in case of a fatal
 398                          * error.
 399                          */
 400                         update_error_count(td, err);
 401                         return 1;
 402                 }
 403         }
 404
 405         return 0;
 406 }
 407
 408 static void check_update_rusage(struct thread_data *td)
 409 {
 410         if (td->update_rusage) {
 411                 td->update_rusage = 0;
 412                 update_rusage_stat(td);
 413                 fio_mutex_up(td->rusage_sem);
 414         }
 415 }
 416
 417 /*
 418  * The main verify engine. Runs over the writes we previously submitted,
 419  * reads the blocks back in, and checks the crc/md5 of the data.
 420  */
 421 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 422 {
 423         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 424         struct fio_file *f;
 425         struct io_u *io_u;
 426         int ret, min_events;
 427         unsigned int i;
 428
 429         dprint(FD_VERIFY, "starting loop\n");
 430
 431         /*
 432          * sync io first and invalidate cache, to make sure we really
 433          * read from disk.
 434          */
 435         for_each_file(td, f, i) {
 436                 if (!fio_file_open(f))
 437                         continue;
 438                 if (fio_io_sync(td, f))
 439                         break;
 440                 if (file_invalidate_cache(td, f))
 441                         break;
 442         }
 443
 444         check_update_rusage(td);
 445
 446         if (td->error)
 447                 return;
 448
 449         td_set_runstate(td, TD_VERIFYING);
 450
 451         io_u = NULL;
 452         while (!td->terminate) {
 453                 enum fio_ddir ddir;
 454                 int ret2, full;
 455
 456                 update_tv_cache(td);
 457                 check_update_rusage(td);
 458
 459                 if (runtime_exceeded(td, &td->tv_cache)) {
 460                         __update_tv_cache(td);
 461                         if (runtime_exceeded(td, &td->tv_cache)) {
 462                                 td->terminate = 1;
 463                                 break;
 464                         }
 465                 }
 466
 467                 if (flow_threshold_exceeded(td))
 468                         continue;
 469
 470                 if (!td->o.experimental_verify) {
 471                         io_u = __get_io_u(td);
 472                         if (!io_u)
 473                                 break;
 474
 475                         if (get_next_verify(td, io_u)) {
 476                                 put_io_u(td, io_u);
 477                                 break;
 478                         }
 479
 480                         if (td_io_prep(td, io_u)) {
 481                                 put_io_u(td, io_u);
 482                                 break;
 483                         }
 484                 } else {
 485                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 486                                 break;
 487
 488                         while ((io_u = get_io_u(td)) != NULL) {
 489                                 if (IS_ERR(io_u)) {
 490                                         io_u = NULL;
 491                                         ret = FIO_Q_BUSY;
 492                                         goto reap;
 493                                 }
 494
 495                                 /*
 496                                  * We are only interested in the places where
 497                                  * we wrote or trimmed IOs. Turn those into
 498                                  * reads for verification purposes.
 499                                  */
 500                                 if (io_u->ddir == DDIR_READ) {
 501                                         /*
 502                                          * Pretend we issued it for rwmix
 503                                          * accounting
 504                                          */
 505                                         td->io_issues[DDIR_READ]++;
 506                                         put_io_u(td, io_u);
 507                                         continue;
 508                                 } else if (io_u->ddir == DDIR_TRIM) {
 509                                         io_u->ddir = DDIR_READ;
 510                                         io_u->flags |= IO_U_F_TRIMMED;
 511                                         break;
 512                                 } else if (io_u->ddir == DDIR_WRITE) {
 513                                         io_u->ddir = DDIR_READ;
 514                                         break;
 515                                 } else {
 516                                         put_io_u(td, io_u);
 517                                         continue;
 518                                 }
 519                         }
 520
 521                         if (!io_u)
 522                                 break;
 523                 }
 524
 525                 if (td->o.verify_async)
 526                         io_u->end_io = verify_io_u_async;
 527                 else
 528                         io_u->end_io = verify_io_u;
 529
 530                 ddir = io_u->ddir;
 531
 532                 ret = td_io_queue(td, io_u);
 533                 switch (ret) {
 534                 case FIO_Q_COMPLETED:
 535                         if (io_u->error) {
 536                                 ret = -io_u->error;
 537                                 clear_io_u(td, io_u);
 538                         } else if (io_u->resid) {
 539                                 int bytes = io_u->xfer_buflen - io_u->resid;
 540
 541                                 /*
 542                                  * zero read, fail
 543                                  */
 544                                 if (!bytes) {
 545                                         td_verror(td, EIO, "full resid");
 546                                         put_io_u(td, io_u);
 547                                         break;
 548                                 }
 549
 550                                 io_u->xfer_buflen = io_u->resid;
 551                                 io_u->xfer_buf += bytes;
 552                                 io_u->offset += bytes;
 553
 554                                 if (ddir_rw(io_u->ddir))
 555                                         td->ts.short_io_u[io_u->ddir]++;
 556
 557                                 f = io_u->file;
 558                                 if (io_u->offset == f->real_file_size)
 559                                         goto sync_done;
 560
 561                                 requeue_io_u(td, &io_u);
 562                         } else {
 563 sync_done:
 564                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 565                                 if (ret < 0)
 566                                         break;
 567                         }
 568                         continue;
 569                 case FIO_Q_QUEUED:
 570                         break;
 571                 case FIO_Q_BUSY:
 572                         requeue_io_u(td, &io_u);
 573                         ret2 = td_io_commit(td);
 574                         if (ret2 < 0)
 575                                 ret = ret2;
 576                         break;
 577                 default:
 578                         assert(ret < 0);
 579                         td_verror(td, -ret, "td_io_queue");
 580                         break;
 581                 }
 582
 583                 if (break_on_this_error(td, ddir, &ret))
 584                         break;
 585
 586                 /*
 587                  * if we can queue more, do so. but check if there are
 588                  * completed io_u's first. Note that we can get BUSY even
 589                  * without IO queued, if the system is resource starved.
 590                  */
 591 reap:
 592                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 593                 if (full || !td->o.iodepth_batch_complete) {
 594                         min_events = min(td->o.iodepth_batch_complete,
 595                                          td->cur_depth);
 596                         /*
 597                          * if the queue is full, we MUST reap at least 1 event
 598                          */
 599                         if (full && !min_events)
 600                                 min_events = 1;
 601
 602                         do {
 603                                 /*
 604                                  * Reap required number of io units, if any,
 605                                  * and do the verification on them through
 606                                  * the callback handler
 607                                  */
 608                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 609                                         ret = -1;
 610                                         break;
 611                                 }
 612                         } while (full && (td->cur_depth > td->o.iodepth_low));
 613                 }
 614                 if (ret < 0)
 615                         break;
 616         }
 617
 618         check_update_rusage(td);
 619
 620         if (!td->error) {
 621                 min_events = td->cur_depth;
 622
 623                 if (min_events)
 624                         ret = io_u_queued_complete(td, min_events, NULL);
 625         } else
 626                 cleanup_pending_aio(td);
 627
 628         td_set_runstate(td, TD_RUNNING);
 629
 630         dprint(FD_VERIFY, "exiting loop\n");
 631 }
 632
 633 static unsigned int exceeds_number_ios(struct thread_data *td)
 634 {
 635         unsigned long long number_ios;
 636
 637         if (!td->o.number_ios)
 638                 return 0;
 639
 640         number_ios = ddir_rw_sum(td->this_io_blocks);
 641         number_ios += td->io_u_queued + td->io_u_in_flight;
 642
 643         return number_ios >= td->o.number_ios;
 644 }
 645
 646 static int io_bytes_exceeded(struct thread_data *td)
 647 {
 648         unsigned long long bytes, limit;
 649
 650         if (td_rw(td))
 651                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 652         else if (td_write(td))
 653                 bytes = td->this_io_bytes[DDIR_WRITE];
 654         else if (td_read(td))
 655                 bytes = td->this_io_bytes[DDIR_READ];
 656         else
 657                 bytes = td->this_io_bytes[DDIR_TRIM];
 658
 659         if (td->o.io_limit)
 660                 limit = td->o.io_limit;
 661         else
 662                 limit = td->o.size;
 663
 664         return bytes >= limit || exceeds_number_ios(td);
 665 }
 666
 667 /*
 668  * Main IO worker function. It retrieves io_u's to process and queues
 669  * and reaps them, checking for rate and errors along the way.
 670  *
 671  * Returns number of bytes written and trimmed.
 672  */
 673 static uint64_t do_io(struct thread_data *td)
 674 {
 675         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 676         unsigned int i;
 677         int ret = 0;
 678         uint64_t total_bytes, bytes_issued = 0;
 679
 680         if (in_ramp_time(td))
 681                 td_set_runstate(td, TD_RAMP);
 682         else
 683                 td_set_runstate(td, TD_RUNNING);
 684
 685         lat_target_init(td);
 686
 687         /*
 688          * If verify_backlog is enabled, we'll run the verify in this
 689          * handler as well. For that case, we may need up to twice the
 690          * amount of bytes.
 691          */
 692         total_bytes = td->o.size;
 693         if (td->o.verify != VERIFY_NONE &&
 694            (td_write(td) && td->o.verify_backlog))
 695                 total_bytes += td->o.size;
 696
 697         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 698                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 699                 td->o.time_based) {
 700                 struct timeval comp_time;
 701                 int min_evts = 0;
 702                 struct io_u *io_u;
 703                 int ret2, full;
 704                 enum fio_ddir ddir;
 705
 706                 check_update_rusage(td);
 707
 708                 if (td->terminate || td->done)
 709                         break;
 710
 711                 update_tv_cache(td);
 712
 713                 if (runtime_exceeded(td, &td->tv_cache)) {
 714                         __update_tv_cache(td);
 715                         if (runtime_exceeded(td, &td->tv_cache)) {
 716                                 td->terminate = 1;
 717                                 break;
 718                         }
 719                 }
 720
 721                 if (flow_threshold_exceeded(td))
 722                         continue;
 723
 724                 if (bytes_issued >= total_bytes)
 725                         break;
 726
 727                 io_u = get_io_u(td);
 728                 if (IS_ERR_OR_NULL(io_u)) {
 729                         int err = PTR_ERR(io_u);
 730
 731                         io_u = NULL;
 732                         if (err == -EBUSY) {
 733                                 ret = FIO_Q_BUSY;
 734                                 goto reap;
 735                         }
 736                         if (td->o.latency_target)
 737                                 goto reap;
 738                         break;
 739                 }
 740
 741                 ddir = io_u->ddir;
 742
 743                 /*
 744                  * Add verification end_io handler if:
 745                  *      - Asked to verify (!td_rw(td))
 746                  *      - Or the io_u is from our verify list (mixed write/ver)
 747                  */
 748                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 749                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 750
 751                         if (!td->o.verify_pattern_bytes) {
 752                                 io_u->rand_seed = __rand(&td->__verify_state);
 753                                 if (sizeof(int) != sizeof(long *))
 754                                         io_u->rand_seed *= __rand(&td->__verify_state);
 755                         }
 756
 757                         if (td->o.verify_async)
 758                                 io_u->end_io = verify_io_u_async;
 759                         else
 760                                 io_u->end_io = verify_io_u;
 761                         td_set_runstate(td, TD_VERIFYING);
 762                 } else if (in_ramp_time(td))
 763                         td_set_runstate(td, TD_RAMP);
 764                 else
 765                         td_set_runstate(td, TD_RUNNING);
 766
 767                 /*
 768                  * Always log IO before it's issued, so we know the specific
 769                  * order of it. The logged unit will track when the IO has
 770                  * completed.
 771                  */
 772                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 773                     td->o.do_verify &&
 774                     td->o.verify != VERIFY_NONE &&
 775                     !td->o.experimental_verify)
 776                         log_io_piece(td, io_u);
 777
 778                 ret = td_io_queue(td, io_u);
 779                 switch (ret) {
 780                 case FIO_Q_COMPLETED:
 781                         if (io_u->error) {
 782                                 ret = -io_u->error;
 783                                 clear_io_u(td, io_u);
 784                         } else if (io_u->resid) {
 785                                 int bytes = io_u->xfer_buflen - io_u->resid;
 786                                 struct fio_file *f = io_u->file;
 787
 788                                 bytes_issued += bytes;
 789                                 /*
 790                                  * zero read, fail
 791                                  */
 792                                 if (!bytes) {
 793                                         td_verror(td, EIO, "full resid");
 794                                         put_io_u(td, io_u);
 795                                         break;
 796                                 }
 797
 798                                 io_u->xfer_buflen = io_u->resid;
 799                                 io_u->xfer_buf += bytes;
 800                                 io_u->offset += bytes;
 801
 802                                 if (ddir_rw(io_u->ddir))
 803                                         td->ts.short_io_u[io_u->ddir]++;
 804
 805                                 if (io_u->offset == f->real_file_size)
 806                                         goto sync_done;
 807
 808                                 requeue_io_u(td, &io_u);
 809                         } else {
 810 sync_done:
 811                                 if (__should_check_rate(td, DDIR_READ) ||
 812                                     __should_check_rate(td, DDIR_WRITE) ||
 813                                     __should_check_rate(td, DDIR_TRIM))
 814                                         fio_gettime(&comp_time, NULL);
 815
 816                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 817                                 if (ret < 0)
 818                                         break;
 819                                 bytes_issued += io_u->xfer_buflen;
 820                         }
 821                         break;
 822                 case FIO_Q_QUEUED:
 823                         /*
 824                          * if the engine doesn't have a commit hook,
 825                          * the io_u is really queued. if it does have such
 826                          * a hook, it has to call io_u_queued() itself.
 827                          */
 828                         if (td->io_ops->commit == NULL)
 829                                 io_u_queued(td, io_u);
 830                         bytes_issued += io_u->xfer_buflen;
 831                         break;
 832                 case FIO_Q_BUSY:
 833                         requeue_io_u(td, &io_u);
 834                         ret2 = td_io_commit(td);
 835                         if (ret2 < 0)
 836                                 ret = ret2;
 837                         break;
 838                 default:
 839                         assert(ret < 0);
 840                         put_io_u(td, io_u);
 841                         break;
 842                 }
 843
 844                 if (break_on_this_error(td, ddir, &ret))
 845                         break;
 846
 847                 /*
 848                  * See if we need to complete some commands. Note that we
 849                  * can get BUSY even without IO queued, if the system is
 850                  * resource starved.
 851                  */
 852 reap:
 853                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 854                 if (full || !td->o.iodepth_batch_complete) {
 855                         min_evts = min(td->o.iodepth_batch_complete,
 856                                         td->cur_depth);
 857                         /*
 858                          * if the queue is full, we MUST reap at least 1 event
 859                          */
 860                         if (full && !min_evts)
 861                                 min_evts = 1;
 862
 863                         if (__should_check_rate(td, DDIR_READ) ||
 864                             __should_check_rate(td, DDIR_WRITE) ||
 865                             __should_check_rate(td, DDIR_TRIM))
 866                                 fio_gettime(&comp_time, NULL);
 867
 868                         do {
 869                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 870                                 if (ret < 0)
 871                                         break;
 872
 873                         } while (full && (td->cur_depth > td->o.iodepth_low));
 874                 }
 875
 876                 if (ret < 0)
 877                         break;
 878                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 879                         continue;
 880
 881                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 882                         if (check_min_rate(td, &comp_time, bytes_done)) {
 883                                 if (exitall_on_terminate)
 884                                         fio_terminate_threads(td->groupid);
 885                                 td_verror(td, EIO, "check_min_rate");
 886                                 break;
 887                         }
 888                 }
 889                 if (!in_ramp_time(td) && td->o.latency_target)
 890                         lat_target_check(td);
 891
 892                 if (td->o.thinktime) {
 893                         unsigned long long b;
 894
 895                         b = ddir_rw_sum(td->io_blocks);
 896                         if (!(b % td->o.thinktime_blocks)) {
 897                                 int left;
 898
 899                                 io_u_quiesce(td);
 900
 901                                 if (td->o.thinktime_spin)
 902                                         usec_spin(td->o.thinktime_spin);
 903
 904                                 left = td->o.thinktime - td->o.thinktime_spin;
 905                                 if (left)
 906                                         usec_sleep(td, left);
 907                         }
 908                 }
 909         }
 910
 911         check_update_rusage(td);
 912
 913         if (td->trim_entries)
 914                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 915
 916         if (td->o.fill_device && td->error == ENOSPC) {
 917                 td->error = 0;
 918                 td->terminate = 1;
 919         }
 920         if (!td->error) {
 921                 struct fio_file *f;
 922
 923                 i = td->cur_depth;
 924                 if (i) {
 925                         ret = io_u_queued_complete(td, i, bytes_done);
 926                         if (td->o.fill_device && td->error == ENOSPC)
 927                                 td->error = 0;
 928                 }
 929
 930                 if (should_fsync(td) && td->o.end_fsync) {
 931                         td_set_runstate(td, TD_FSYNCING);
 932
 933                         for_each_file(td, f, i) {
 934                                 if (!fio_file_fsync(td, f))
 935                                         continue;
 936
 937                                 log_err("fio: end_fsync failed for file %s\n",
 938                                                                 f->file_name);
 939                         }
 940                 }
 941         } else
 942                 cleanup_pending_aio(td);
 943
 944         /*
 945          * stop job if we failed doing any IO
 946          */
 947         if (!ddir_rw_sum(td->this_io_bytes))
 948                 td->done = 1;
 949
 950         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 951 }
 952
 953 static void cleanup_io_u(struct thread_data *td)
 954 {
 955         struct io_u *io_u;
 956
 957         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 958
 959                 if (td->io_ops->io_u_free)
 960                         td->io_ops->io_u_free(td, io_u);
 961
 962                 fio_memfree(io_u, sizeof(*io_u));
 963         }
 964
 965         free_io_mem(td);
 966
 967         io_u_rexit(&td->io_u_requeues);
 968         io_u_qexit(&td->io_u_freelist);
 969         io_u_qexit(&td->io_u_all);
 970 }
 971
 972 static int init_io_u(struct thread_data *td)
 973 {
 974         struct io_u *io_u;
 975         unsigned int max_bs, min_write;
 976         int cl_align, i, max_units;
 977         int data_xfer = 1, err;
 978         char *p;
 979
 980         max_units = td->o.iodepth;
 981         max_bs = td_max_bs(td);
 982         min_write = td->o.min_bs[DDIR_WRITE];
 983         td->orig_buffer_size = (unsigned long long) max_bs
 984                                         * (unsigned long long) max_units;
 985
 986         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 987                 data_xfer = 0;
 988
 989         err = 0;
 990         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 991         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 992         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 993
 994         if (err) {
 995                 log_err("fio: failed setting up IO queues\n");
 996                 return 1;
 997         }
 998
 999         /*
1000          * if we may later need to do address alignment, then add any
1001          * possible adjustment here so that we don't cause a buffer
1002          * overflow later. this adjustment may be too much if we get
1003          * lucky and the allocator gives us an aligned address.
1004          */
1005         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1006             (td->io_ops->flags & FIO_RAWIO))
1007                 td->orig_buffer_size += page_mask + td->o.mem_align;
1008
1009         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1010                 unsigned long bs;
1011
1012                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1013                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1014         }
1015
1016         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1017                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1018                 return 1;
1019         }
1020
1021         if (data_xfer && allocate_io_mem(td))
1022                 return 1;
1023
1024         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1025             (td->io_ops->flags & FIO_RAWIO))
1026                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1027         else
1028                 p = td->orig_buffer;
1029
1030         cl_align = os_cache_line_size();
1031
1032         for (i = 0; i < max_units; i++) {
1033                 void *ptr;
1034
1035                 if (td->terminate)
1036                         return 1;
1037
1038                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1039                 if (!ptr) {
1040                         log_err("fio: unable to allocate aligned memory\n");
1041                         break;
1042                 }
1043
1044                 io_u = ptr;
1045                 memset(io_u, 0, sizeof(*io_u));
1046                 INIT_FLIST_HEAD(&io_u->verify_list);
1047                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1048
1049                 if (data_xfer) {
1050                         io_u->buf = p;
1051                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1052
1053                         if (td_write(td))
1054                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1055                         if (td_write(td) && td->o.verify_pattern_bytes) {
1056                                 /*
1057                                  * Fill the buffer with the pattern if we are
1058                                  * going to be doing writes.
1059                                  */
1060                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1061                         }
1062                 }
1063
1064                 io_u->index = i;
1065                 io_u->flags = IO_U_F_FREE;
1066                 io_u_qpush(&td->io_u_freelist, io_u);
1067
1068                 /*
1069                  * io_u never leaves this stack, used for iteration of all
1070                  * io_u buffers.
1071                  */
1072                 io_u_qpush(&td->io_u_all, io_u);
1073
1074                 if (td->io_ops->io_u_init) {
1075                         int ret = td->io_ops->io_u_init(td, io_u);
1076
1077                         if (ret) {
1078                                 log_err("fio: failed to init engine data: %d\n", ret);
1079                                 return 1;
1080                         }
1081                 }
1082
1083                 p += max_bs;
1084         }
1085
1086         return 0;
1087 }
1088
1089 static int switch_ioscheduler(struct thread_data *td)
1090 {
1091         char tmp[256], tmp2[128];
1092         FILE *f;
1093         int ret;
1094
1095         if (td->io_ops->flags & FIO_DISKLESSIO)
1096                 return 0;
1097
1098         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1099
1100         f = fopen(tmp, "r+");
1101         if (!f) {
1102                 if (errno == ENOENT) {
1103                         log_err("fio: os or kernel doesn't support IO scheduler"
1104                                 " switching\n");
1105                         return 0;
1106                 }
1107                 td_verror(td, errno, "fopen iosched");
1108                 return 1;
1109         }
1110
1111         /*
1112          * Set io scheduler.
1113          */
1114         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1115         if (ferror(f) || ret != 1) {
1116                 td_verror(td, errno, "fwrite");
1117                 fclose(f);
1118                 return 1;
1119         }
1120
1121         rewind(f);
1122
1123         /*
1124          * Read back and check that the selected scheduler is now the default.
1125          */
1126         ret = fread(tmp, sizeof(tmp), 1, f);
1127         if (ferror(f) || ret < 0) {
1128                 td_verror(td, errno, "fread");
1129                 fclose(f);
1130                 return 1;
1131         }
1132         tmp[sizeof(tmp) - 1] = '\0';
1133
1134
1135         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1136         if (!strstr(tmp, tmp2)) {
1137                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1138                 td_verror(td, EINVAL, "iosched_switch");
1139                 fclose(f);
1140                 return 1;
1141         }
1142
1143         fclose(f);
1144         return 0;
1145 }
1146
1147 static int keep_running(struct thread_data *td)
1148 {
1149         unsigned long long limit;
1150
1151         if (td->done)
1152                 return 0;
1153         if (td->o.time_based)
1154                 return 1;
1155         if (td->o.loops) {
1156                 td->o.loops--;
1157                 return 1;
1158         }
1159         if (exceeds_number_ios(td))
1160                 return 0;
1161
1162         if (td->o.io_limit)
1163                 limit = td->o.io_limit;
1164         else
1165                 limit = td->o.size;
1166
1167         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1168                 uint64_t diff;
1169
1170                 /*
1171                  * If the difference is less than the minimum IO size, we
1172                  * are done.
1173                  */
1174                 diff = limit - ddir_rw_sum(td->io_bytes);
1175                 if (diff < td_max_bs(td))
1176                         return 0;
1177
1178                 if (fio_files_done(td))
1179                         return 0;
1180
1181                 return 1;
1182         }
1183
1184         return 0;
1185 }
1186
1187 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1188 {
1189         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1190         char *str;
1191
1192         str = malloc(newlen);
1193         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1194
1195         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1196         ret = system(str);
1197         if (ret == -1)
1198                 log_err("fio: exec of cmd <%s> failed\n", str);
1199
1200         free(str);
1201         return ret;
1202 }
1203
1204 /*
1205  * Dry run to compute correct state of numberio for verification.
1206  */
1207 static uint64_t do_dry_run(struct thread_data *td)
1208 {
1209         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1210
1211         td_set_runstate(td, TD_RUNNING);
1212
1213         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1214                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1215                 struct io_u *io_u;
1216                 int ret;
1217
1218                 if (td->terminate || td->done)
1219                         break;
1220
1221                 io_u = get_io_u(td);
1222                 if (!io_u)
1223                         break;
1224
1225                 io_u->flags |= IO_U_F_FLIGHT;
1226                 io_u->error = 0;
1227                 io_u->resid = 0;
1228                 if (ddir_rw(acct_ddir(io_u)))
1229                         td->io_issues[acct_ddir(io_u)]++;
1230                 if (ddir_rw(io_u->ddir)) {
1231                         io_u_mark_depth(td, 1);
1232                         td->ts.total_io_u[io_u->ddir]++;
1233                 }
1234
1235                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1236                     td->o.do_verify &&
1237                     td->o.verify != VERIFY_NONE &&
1238                     !td->o.experimental_verify)
1239                         log_io_piece(td, io_u);
1240
1241                 ret = io_u_sync_complete(td, io_u, bytes_done);
1242                 (void) ret;
1243         }
1244
1245         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1246 }
1247
1248 /*
1249  * Entry point for the thread based jobs. The process based jobs end up
1250  * here as well, after a little setup.
1251  */
1252 static void *thread_main(void *data)
1253 {
1254         unsigned long long elapsed;
1255         struct thread_data *td = data;
1256         struct thread_options *o = &td->o;
1257         pthread_condattr_t attr;
1258         int clear_state;
1259         int ret;
1260
1261         if (!o->use_thread) {
1262                 setsid();
1263                 td->pid = getpid();
1264         } else
1265                 td->pid = gettid();
1266
1267         /*
1268          * fio_time_init() may not have been called yet if running as a server
1269          */
1270         fio_time_init();
1271
1272         fio_local_clock_init(o->use_thread);
1273
1274         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1275
1276         if (is_backend)
1277                 fio_server_send_start(td);
1278
1279         INIT_FLIST_HEAD(&td->io_log_list);
1280         INIT_FLIST_HEAD(&td->io_hist_list);
1281         INIT_FLIST_HEAD(&td->verify_list);
1282         INIT_FLIST_HEAD(&td->trim_list);
1283         INIT_FLIST_HEAD(&td->next_rand_list);
1284         pthread_mutex_init(&td->io_u_lock, NULL);
1285         td->io_hist_tree = RB_ROOT;
1286
1287         pthread_condattr_init(&attr);
1288         pthread_cond_init(&td->verify_cond, &attr);
1289         pthread_cond_init(&td->free_cond, &attr);
1290
1291         td_set_runstate(td, TD_INITIALIZED);
1292         dprint(FD_MUTEX, "up startup_mutex\n");
1293         fio_mutex_up(startup_mutex);
1294         dprint(FD_MUTEX, "wait on td->mutex\n");
1295         fio_mutex_down(td->mutex);
1296         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1297
1298         /*
1299          * A new gid requires privilege, so we need to do this before setting
1300          * the uid.
1301          */
1302         if (o->gid != -1U && setgid(o->gid)) {
1303                 td_verror(td, errno, "setgid");
1304                 goto err;
1305         }
1306         if (o->uid != -1U && setuid(o->uid)) {
1307                 td_verror(td, errno, "setuid");
1308                 goto err;
1309         }
1310
1311         /*
1312          * If we have a gettimeofday() thread, make sure we exclude that
1313          * thread from this job
1314          */
1315         if (o->gtod_cpu)
1316                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1317
1318         /*
1319          * Set affinity first, in case it has an impact on the memory
1320          * allocations.
1321          */
1322         if (o->cpumask_set) {
1323                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1324                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1325                         if (!ret) {
1326                                 log_err("fio: no CPUs set\n");
1327                                 log_err("fio: Try increasing number of available CPUs\n");
1328                                 td_verror(td, EINVAL, "cpus_split");
1329                                 goto err;
1330                         }
1331                 }
1332                 ret = fio_setaffinity(td->pid, o->cpumask);
1333                 if (ret == -1) {
1334                         td_verror(td, errno, "cpu_set_affinity");
1335                         goto err;
1336                 }
1337         }
1338
1339 #ifdef CONFIG_LIBNUMA
1340         /* numa node setup */
1341         if (o->numa_cpumask_set || o->numa_memmask_set) {
1342                 int ret;
1343
1344                 if (numa_available() < 0) {
1345                         td_verror(td, errno, "Does not support NUMA API\n");
1346                         goto err;
1347                 }
1348
1349                 if (o->numa_cpumask_set) {
1350                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1351                         if (ret == -1) {
1352                                 td_verror(td, errno, \
1353                                         "numa_run_on_node_mask failed\n");
1354                                 goto err;
1355                         }
1356                 }
1357
1358                 if (o->numa_memmask_set) {
1359
1360                         switch (o->numa_mem_mode) {
1361                         case MPOL_INTERLEAVE:
1362                                 numa_set_interleave_mask(o->numa_memnodesmask);
1363                                 break;
1364                         case MPOL_BIND:
1365                                 numa_set_membind(o->numa_memnodesmask);
1366                                 break;
1367                         case MPOL_LOCAL:
1368                                 numa_set_localalloc();
1369                                 break;
1370                         case MPOL_PREFERRED:
1371                                 numa_set_preferred(o->numa_mem_prefer_node);
1372                                 break;
1373                         case MPOL_DEFAULT:
1374                         default:
1375                                 break;
1376                         }
1377
1378                 }
1379         }
1380 #endif
1381
1382         if (fio_pin_memory(td))
1383                 goto err;
1384
1385         /*
1386          * May alter parameters that init_io_u() will use, so we need to
1387          * do this first.
1388          */
1389         if (init_iolog(td))
1390                 goto err;
1391
1392         if (init_io_u(td))
1393                 goto err;
1394
1395         if (o->verify_async && verify_async_init(td))
1396                 goto err;
1397
1398         if (o->ioprio) {
1399                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1400                 if (ret == -1) {
1401                         td_verror(td, errno, "ioprio_set");
1402                         goto err;
1403                 }
1404         }
1405
1406         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1407                 goto err;
1408
1409         errno = 0;
1410         if (nice(o->nice) == -1 && errno != 0) {
1411                 td_verror(td, errno, "nice");
1412                 goto err;
1413         }
1414
1415         if (o->ioscheduler && switch_ioscheduler(td))
1416                 goto err;
1417
1418         if (!o->create_serialize && setup_files(td))
1419                 goto err;
1420
1421         if (td_io_init(td))
1422                 goto err;
1423
1424         if (init_random_map(td))
1425                 goto err;
1426
1427         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1428                 goto err;
1429
1430         if (o->pre_read) {
1431                 if (pre_read_files(td) < 0)
1432                         goto err;
1433         }
1434
1435         fio_verify_init(td);
1436
1437         fio_gettime(&td->epoch, NULL);
1438         fio_getrusage(&td->ru_start);
1439         clear_state = 0;
1440         while (keep_running(td)) {
1441                 uint64_t verify_bytes;
1442
1443                 fio_gettime(&td->start, NULL);
1444                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1445                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1446                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1447
1448                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1449                                 o->ratemin[DDIR_TRIM]) {
1450                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1451                                                 sizeof(td->bw_sample_time));
1452                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1453                                                 sizeof(td->bw_sample_time));
1454                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1455                                                 sizeof(td->bw_sample_time));
1456                 }
1457
1458                 if (clear_state)
1459                         clear_io_state(td);
1460
1461                 prune_io_piece_log(td);
1462
1463                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1464                         verify_bytes = do_dry_run(td);
1465                 else
1466                         verify_bytes = do_io(td);
1467
1468                 clear_state = 1;
1469
1470                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1471                         elapsed = utime_since_now(&td->start);
1472                         td->ts.runtime[DDIR_READ] += elapsed;
1473                 }
1474                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1475                         elapsed = utime_since_now(&td->start);
1476                         td->ts.runtime[DDIR_WRITE] += elapsed;
1477                 }
1478                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1479                         elapsed = utime_since_now(&td->start);
1480                         td->ts.runtime[DDIR_TRIM] += elapsed;
1481                 }
1482
1483                 if (td->error || td->terminate)
1484                         break;
1485
1486                 if (!o->do_verify ||
1487                     o->verify == VERIFY_NONE ||
1488                     (td->io_ops->flags & FIO_UNIDIR))
1489                         continue;
1490
1491                 clear_io_state(td);
1492
1493                 fio_gettime(&td->start, NULL);
1494
1495                 do_verify(td, verify_bytes);
1496
1497                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1498
1499                 if (td->error || td->terminate)
1500                         break;
1501         }
1502
1503         update_rusage_stat(td);
1504         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1505         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1506         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1507         td->ts.total_run_time = mtime_since_now(&td->epoch);
1508         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1509         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1510         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1511
1512         fio_unpin_memory(td);
1513
1514         fio_writeout_logs(td);
1515
1516         if (o->exec_postrun)
1517                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1518
1519         if (exitall_on_terminate)
1520                 fio_terminate_threads(td->groupid);
1521
1522 err:
1523         if (td->error)
1524                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1525                                                         td->verror);
1526
1527         if (o->verify_async)
1528                 verify_async_exit(td);
1529
1530         close_and_free_files(td);
1531         cleanup_io_u(td);
1532         close_ioengine(td);
1533         cgroup_shutdown(td, &cgroup_mnt);
1534
1535         if (o->cpumask_set) {
1536                 int ret = fio_cpuset_exit(&o->cpumask);
1537
1538                 td_verror(td, ret, "fio_cpuset_exit");
1539         }
1540
1541         /*
1542          * do this very late, it will log file closing as well
1543          */
1544         if (o->write_iolog_file)
1545                 write_iolog_close(td);
1546
1547         fio_mutex_remove(td->rusage_sem);
1548         td->rusage_sem = NULL;
1549
1550         fio_mutex_remove(td->mutex);
1551         td->mutex = NULL;
1552
1553         td_set_runstate(td, TD_EXITED);
1554         return (void *) (uintptr_t) td->error;
1555 }
1556
1557
1558 /*
1559  * We cannot pass the td data into a forked process, so attach the td and
1560  * pass it to the thread worker.
1561  */
1562 static int fork_main(int shmid, int offset)
1563 {
1564         struct thread_data *td;
1565         void *data, *ret;
1566
1567 #ifndef __hpux
1568         data = shmat(shmid, NULL, 0);
1569         if (data == (void *) -1) {
1570                 int __err = errno;
1571
1572                 perror("shmat");
1573                 return __err;
1574         }
1575 #else
1576         /*
1577          * HP-UX inherits shm mappings?
1578          */
1579         data = threads;
1580 #endif
1581
1582         td = data + offset * sizeof(struct thread_data);
1583         ret = thread_main(td);
1584         shmdt(data);
1585         return (int) (uintptr_t) ret;
1586 }
1587
1588 /*
1589  * Run over the job map and reap the threads that have exited, if any.
1590  */
1591 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1592                          unsigned int *m_rate)
1593 {
1594         struct thread_data *td;
1595         unsigned int cputhreads, realthreads, pending;
1596         int i, status, ret;
1597
1598         /*
1599          * reap exited threads (TD_EXITED -> TD_REAPED)
1600          */
1601         realthreads = pending = cputhreads = 0;
1602         for_each_td(td, i) {
1603                 int flags = 0;
1604
1605                 /*
1606                  * ->io_ops is NULL for a thread that has closed its
1607                  * io engine
1608                  */
1609                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1610                         cputhreads++;
1611                 else
1612                         realthreads++;
1613
1614                 if (!td->pid) {
1615                         pending++;
1616                         continue;
1617                 }
1618                 if (td->runstate == TD_REAPED)
1619                         continue;
1620                 if (td->o.use_thread) {
1621                         if (td->runstate == TD_EXITED) {
1622                                 td_set_runstate(td, TD_REAPED);
1623                                 goto reaped;
1624                         }
1625                         continue;
1626                 }
1627
1628                 flags = WNOHANG;
1629                 if (td->runstate == TD_EXITED)
1630                         flags = 0;
1631
1632                 /*
1633                  * check if someone quit or got killed in an unusual way
1634                  */
1635                 ret = waitpid(td->pid, &status, flags);
1636                 if (ret < 0) {
1637                         if (errno == ECHILD) {
1638                                 log_err("fio: pid=%d disappeared %d\n",
1639                                                 (int) td->pid, td->runstate);
1640                                 td->sig = ECHILD;
1641                                 td_set_runstate(td, TD_REAPED);
1642                                 goto reaped;
1643                         }
1644                         perror("waitpid");
1645                 } else if (ret == td->pid) {
1646                         if (WIFSIGNALED(status)) {
1647                                 int sig = WTERMSIG(status);
1648
1649                                 if (sig != SIGTERM && sig != SIGUSR2)
1650                                         log_err("fio: pid=%d, got signal=%d\n",
1651                                                         (int) td->pid, sig);
1652                                 td->sig = sig;
1653                                 td_set_runstate(td, TD_REAPED);
1654                                 goto reaped;
1655                         }
1656                         if (WIFEXITED(status)) {
1657                                 if (WEXITSTATUS(status) && !td->error)
1658                                         td->error = WEXITSTATUS(status);
1659
1660                                 td_set_runstate(td, TD_REAPED);
1661                                 goto reaped;
1662                         }
1663                 }
1664
1665                 /*
1666                  * thread is not dead, continue
1667                  */
1668                 pending++;
1669                 continue;
1670 reaped:
1671                 (*nr_running)--;
1672                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1673                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1674                 if (!td->pid)
1675                         pending--;
1676
1677                 if (td->error)
1678                         exit_value++;
1679
1680                 done_secs += mtime_since_now(&td->epoch) / 1000;
1681                 profile_td_exit(td);
1682         }
1683
1684         if (*nr_running == cputhreads && !pending && realthreads)
1685                 fio_terminate_threads(TERMINATE_ALL);
1686 }
1687
1688 static void do_usleep(unsigned int usecs)
1689 {
1690         check_for_running_stats();
1691         usleep(usecs);
1692 }
1693
1694 /*
1695  * Main function for kicking off and reaping jobs, as needed.
1696  */
1697 static void run_threads(void)
1698 {
1699         struct thread_data *td;
1700         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1701         uint64_t spent;
1702
1703         if (fio_gtod_offload && fio_start_gtod_thread())
1704                 return;
1705
1706         fio_idle_prof_init();
1707
1708         set_sig_handlers();
1709
1710         nr_thread = nr_process = 0;
1711         for_each_td(td, i) {
1712                 if (td->o.use_thread)
1713                         nr_thread++;
1714                 else
1715                         nr_process++;
1716         }
1717
1718         if (output_format == FIO_OUTPUT_NORMAL) {
1719                 log_info("Starting ");
1720                 if (nr_thread)
1721                         log_info("%d thread%s", nr_thread,
1722                                                 nr_thread > 1 ? "s" : "");
1723                 if (nr_process) {
1724                         if (nr_thread)
1725                                 log_info(" and ");
1726                         log_info("%d process%s", nr_process,
1727                                                 nr_process > 1 ? "es" : "");
1728                 }
1729                 log_info("\n");
1730                 fflush(stdout);
1731         }
1732
1733         todo = thread_number;
1734         nr_running = 0;
1735         nr_started = 0;
1736         m_rate = t_rate = 0;
1737
1738         for_each_td(td, i) {
1739                 print_status_init(td->thread_number - 1);
1740
1741                 if (!td->o.create_serialize)
1742                         continue;
1743
1744                 /*
1745                  * do file setup here so it happens sequentially,
1746                  * we don't want X number of threads getting their
1747                  * client data interspersed on disk
1748                  */
1749                 if (setup_files(td)) {
1750                         exit_value++;
1751                         if (td->error)
1752                                 log_err("fio: pid=%d, err=%d/%s\n",
1753                                         (int) td->pid, td->error, td->verror);
1754                         td_set_runstate(td, TD_REAPED);
1755                         todo--;
1756                 } else {
1757                         struct fio_file *f;
1758                         unsigned int j;
1759
1760                         /*
1761                          * for sharing to work, each job must always open
1762                          * its own files. so close them, if we opened them
1763                          * for creation
1764                          */
1765                         for_each_file(td, f, j) {
1766                                 if (fio_file_open(f))
1767                                         td_io_close_file(td, f);
1768                         }
1769                 }
1770         }
1771
1772         /* start idle threads before io threads start to run */
1773         fio_idle_prof_start();
1774
1775         set_genesis_time();
1776
1777         while (todo) {
1778                 struct thread_data *map[REAL_MAX_JOBS];
1779                 struct timeval this_start;
1780                 int this_jobs = 0, left;
1781
1782                 /*
1783                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1784                  */
1785                 for_each_td(td, i) {
1786                         if (td->runstate != TD_NOT_CREATED)
1787                                 continue;
1788
1789                         /*
1790                          * never got a chance to start, killed by other
1791                          * thread for some reason
1792                          */
1793                         if (td->terminate) {
1794                                 todo--;
1795                                 continue;
1796                         }
1797
1798                         if (td->o.start_delay) {
1799                                 spent = utime_since_genesis();
1800
1801                                 if (td->o.start_delay > spent)
1802                                         continue;
1803                         }
1804
1805                         if (td->o.stonewall && (nr_started || nr_running)) {
1806                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1807                                                         td->o.name);
1808                                 break;
1809                         }
1810
1811                         init_disk_util(td);
1812
1813                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1814                         td->update_rusage = 0;
1815
1816                         /*
1817                          * Set state to created. Thread will transition
1818                          * to TD_INITIALIZED when it's done setting up.
1819                          */
1820                         td_set_runstate(td, TD_CREATED);
1821                         map[this_jobs++] = td;
1822                         nr_started++;
1823
1824                         if (td->o.use_thread) {
1825                                 int ret;
1826
1827                                 dprint(FD_PROCESS, "will pthread_create\n");
1828                                 ret = pthread_create(&td->thread, NULL,
1829                                                         thread_main, td);
1830                                 if (ret) {
1831                                         log_err("pthread_create: %s\n",
1832                                                         strerror(ret));
1833                                         nr_started--;
1834                                         break;
1835                                 }
1836                                 ret = pthread_detach(td->thread);
1837                                 if (ret)
1838                                         log_err("pthread_detach: %s",
1839                                                         strerror(ret));
1840                         } else {
1841                                 pid_t pid;
1842                                 dprint(FD_PROCESS, "will fork\n");
1843                                 pid = fork();
1844                                 if (!pid) {
1845                                         int ret = fork_main(shm_id, i);
1846
1847                                         _exit(ret);
1848                                 } else if (i == fio_debug_jobno)
1849                                         *fio_debug_jobp = pid;
1850                         }
1851                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1852                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1853                                 log_err("fio: job startup hung? exiting.\n");
1854                                 fio_terminate_threads(TERMINATE_ALL);
1855                                 fio_abort = 1;
1856                                 nr_started--;
1857                                 break;
1858                         }
1859                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1860                 }
1861
1862                 /*
1863                  * Wait for the started threads to transition to
1864                  * TD_INITIALIZED.
1865                  */
1866                 fio_gettime(&this_start, NULL);
1867                 left = this_jobs;
1868                 while (left && !fio_abort) {
1869                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1870                                 break;
1871
1872                         do_usleep(100000);
1873
1874                         for (i = 0; i < this_jobs; i++) {
1875                                 td = map[i];
1876                                 if (!td)
1877                                         continue;
1878                                 if (td->runstate == TD_INITIALIZED) {
1879                                         map[i] = NULL;
1880                                         left--;
1881                                 } else if (td->runstate >= TD_EXITED) {
1882                                         map[i] = NULL;
1883                                         left--;
1884                                         todo--;
1885                                         nr_running++; /* work-around... */
1886                                 }
1887                         }
1888                 }
1889
1890                 if (left) {
1891                         log_err("fio: %d job%s failed to start\n", left,
1892                                         left > 1 ? "s" : "");
1893                         for (i = 0; i < this_jobs; i++) {
1894                                 td = map[i];
1895                                 if (!td)
1896                                         continue;
1897                                 kill(td->pid, SIGTERM);
1898                         }
1899                         break;
1900                 }
1901
1902                 /*
1903                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1904                  */
1905                 for_each_td(td, i) {
1906                         if (td->runstate != TD_INITIALIZED)
1907                                 continue;
1908
1909                         if (in_ramp_time(td))
1910                                 td_set_runstate(td, TD_RAMP);
1911                         else
1912                                 td_set_runstate(td, TD_RUNNING);
1913                         nr_running++;
1914                         nr_started--;
1915                         m_rate += ddir_rw_sum(td->o.ratemin);
1916                         t_rate += ddir_rw_sum(td->o.rate);
1917                         todo--;
1918                         fio_mutex_up(td->mutex);
1919                 }
1920
1921                 reap_threads(&nr_running, &t_rate, &m_rate);
1922
1923                 if (todo)
1924                         do_usleep(100000);
1925         }
1926
1927         while (nr_running) {
1928                 reap_threads(&nr_running, &t_rate, &m_rate);
1929                 do_usleep(10000);
1930         }
1931
1932         fio_idle_prof_stop();
1933
1934         update_io_ticks();
1935 }
1936
1937 void wait_for_disk_thread_exit(void)
1938 {
1939         fio_mutex_down(disk_thread_mutex);
1940 }
1941
1942 static void free_disk_util(void)
1943 {
1944         disk_util_start_exit();
1945         wait_for_disk_thread_exit();
1946         disk_util_prune_entries();
1947 }
1948
1949 static void *disk_thread_main(void *data)
1950 {
1951         int ret = 0;
1952
1953         fio_mutex_up(startup_mutex);
1954
1955         while (threads && !ret) {
1956                 usleep(DISK_UTIL_MSEC * 1000);
1957                 if (!threads)
1958                         break;
1959                 ret = update_io_ticks();
1960
1961                 if (!is_backend)
1962                         print_thread_status();
1963         }
1964
1965         fio_mutex_up(disk_thread_mutex);
1966         return NULL;
1967 }
1968
1969 static int create_disk_util_thread(void)
1970 {
1971         int ret;
1972
1973         setup_disk_util();
1974
1975         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1976
1977         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1978         if (ret) {
1979                 fio_mutex_remove(disk_thread_mutex);
1980                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1981                 return 1;
1982         }
1983
1984         ret = pthread_detach(disk_util_thread);
1985         if (ret) {
1986                 fio_mutex_remove(disk_thread_mutex);
1987                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1988                 return 1;
1989         }
1990
1991         dprint(FD_MUTEX, "wait on startup_mutex\n");
1992         fio_mutex_down(startup_mutex);
1993         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1994         return 0;
1995 }
1996
1997 int fio_backend(void)
1998 {
1999         struct thread_data *td;
2000         int i;
2001
2002         if (exec_profile) {
2003                 if (load_profile(exec_profile))
2004                         return 1;
2005                 free(exec_profile);
2006                 exec_profile = NULL;
2007         }
2008         if (!thread_number)
2009                 return 0;
2010
2011         if (write_bw_log) {
2012                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
2013                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
2014                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2015         }
2016
2017         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2018         if (startup_mutex == NULL)
2019                 return 1;
2020
2021         set_genesis_time();
2022         stat_init();
2023         create_disk_util_thread();
2024
2025         cgroup_list = smalloc(sizeof(*cgroup_list));
2026         INIT_FLIST_HEAD(cgroup_list);
2027
2028         run_threads();
2029
2030         if (!fio_abort) {
2031                 show_run_stats();
2032                 if (write_bw_log) {
2033                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2034                         __finish_log(agg_io_log[DDIR_WRITE],
2035                                         "agg-write_bw.log");
2036                         __finish_log(agg_io_log[DDIR_TRIM],
2037                                         "agg-write_bw.log");
2038                 }
2039         }
2040
2041         for_each_td(td, i)
2042                 fio_options_free(td);
2043
2044         free_disk_util();
2045         cgroup_kill(cgroup_list);
2046         sfree(cgroup_list);
2047         sfree(cgroup_mnt);
2048
2049         fio_mutex_remove(startup_mutex);
2050         fio_mutex_remove(disk_thread_mutex);
2051         stat_exit();
2052         return exit_value;
2053 }