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                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 186                                 if (rate < ratemin ||
 187                                     bytes < td->rate_bytes[ddir]) {
 188                                         log_err("%s: min rate %u not met, got"
 189                                                 " %luKB/sec\n", td->o.name,
 190                                                         ratemin, rate);
 191                                         return 1;
 192                                 }
 193                         }
 194                 } else {
 195                         /*
 196                          * checks iops specified rate
 197                          */
 198                         if (iops < rate_iops) {
 199                                 log_err("%s: min iops rate %u not met\n",
 200                                                 td->o.name, rate_iops);
 201                                 return 1;
 202                         } else {
 203                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 204                                 if (rate < rate_iops_min ||
 205                                     iops < td->rate_blocks[ddir]) {
 206                                         log_err("%s: min iops rate %u not met,"
 207                                                 " got %lu\n", td->o.name,
 208                                                         rate_iops_min, rate);
 209                                 }
 210                         }
 211                 }
 212         }
 213
 214         td->rate_bytes[ddir] = bytes;
 215         td->rate_blocks[ddir] = iops;
 216         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 217         return 0;
 218 }
 219
 220 static int check_min_rate(struct thread_data *td, struct timeval *now,
 221                           uint64_t *bytes_done)
 222 {
 223         int ret = 0;
 224
 225         if (bytes_done[DDIR_READ])
 226                 ret |= __check_min_rate(td, now, DDIR_READ);
 227         if (bytes_done[DDIR_WRITE])
 228                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 229         if (bytes_done[DDIR_TRIM])
 230                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 231
 232         return ret;
 233 }
 234
 235 /*
 236  * When job exits, we can cancel the in-flight IO if we are using async
 237  * io. Attempt to do so.
 238  */
 239 static void cleanup_pending_aio(struct thread_data *td)
 240 {
 241         int r;
 242
 243         /*
 244          * get immediately available events, if any
 245          */
 246         r = io_u_queued_complete(td, 0, NULL);
 247         if (r < 0)
 248                 return;
 249
 250         /*
 251          * now cancel remaining active events
 252          */
 253         if (td->io_ops->cancel) {
 254                 struct io_u *io_u;
 255                 int i;
 256
 257                 io_u_qiter(&td->io_u_all, io_u, i) {
 258                         if (io_u->flags & IO_U_F_FLIGHT) {
 259                                 r = td->io_ops->cancel(td, io_u);
 260                                 if (!r)
 261                                         put_io_u(td, io_u);
 262                         }
 263                 }
 264         }
 265
 266         if (td->cur_depth)
 267                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 268 }
 269
 270 /*
 271  * Helper to handle the final sync of a file. Works just like the normal
 272  * io path, just does everything sync.
 273  */
 274 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 275 {
 276         struct io_u *io_u = __get_io_u(td);
 277         int ret;
 278
 279         if (!io_u)
 280                 return 1;
 281
 282         io_u->ddir = DDIR_SYNC;
 283         io_u->file = f;
 284
 285         if (td_io_prep(td, io_u)) {
 286                 put_io_u(td, io_u);
 287                 return 1;
 288         }
 289
 290 requeue:
 291         ret = td_io_queue(td, io_u);
 292         if (ret < 0) {
 293                 td_verror(td, io_u->error, "td_io_queue");
 294                 put_io_u(td, io_u);
 295                 return 1;
 296         } else if (ret == FIO_Q_QUEUED) {
 297                 if (io_u_queued_complete(td, 1, NULL) < 0)
 298                         return 1;
 299         } else if (ret == FIO_Q_COMPLETED) {
 300                 if (io_u->error) {
 301                         td_verror(td, io_u->error, "td_io_queue");
 302                         return 1;
 303                 }
 304
 305                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 306                         return 1;
 307         } else if (ret == FIO_Q_BUSY) {
 308                 if (td_io_commit(td))
 309                         return 1;
 310                 goto requeue;
 311         }
 312
 313         return 0;
 314 }
 315
 316 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 317 {
 318         int ret;
 319
 320         if (fio_file_open(f))
 321                 return fio_io_sync(td, f);
 322
 323         if (td_io_open_file(td, f))
 324                 return 1;
 325
 326         ret = fio_io_sync(td, f);
 327         td_io_close_file(td, f);
 328         return ret;
 329 }
 330
 331 static inline void __update_tv_cache(struct thread_data *td)
 332 {
 333         fio_gettime(&td->tv_cache, NULL);
 334 }
 335
 336 static inline void update_tv_cache(struct thread_data *td)
 337 {
 338         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 339                 __update_tv_cache(td);
 340 }
 341
 342 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 343 {
 344         if (in_ramp_time(td))
 345                 return 0;
 346         if (!td->o.timeout)
 347                 return 0;
 348         if (utime_since(&td->epoch, t) >= td->o.timeout)
 349                 return 1;
 350
 351         return 0;
 352 }
 353
 354 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 355                                int *retptr)
 356 {
 357         int ret = *retptr;
 358
 359         if (ret < 0 || td->error) {
 360                 int err = td->error;
 361                 enum error_type_bit eb;
 362
 363                 if (ret < 0)
 364                         err = -ret;
 365
 366                 eb = td_error_type(ddir, err);
 367                 if (!(td->o.continue_on_error & (1 << eb)))
 368                         return 1;
 369
 370                 if (td_non_fatal_error(td, eb, err)) {
 371                         /*
 372                          * Continue with the I/Os in case of
 373                          * a non fatal error.
 374                          */
 375                         update_error_count(td, err);
 376                         td_clear_error(td);
 377                         *retptr = 0;
 378                         return 0;
 379                 } else if (td->o.fill_device && err == ENOSPC) {
 380                         /*
 381                          * We expect to hit this error if
 382                          * fill_device option is set.
 383                          */
 384                         td_clear_error(td);
 385                         td->terminate = 1;
 386                         return 1;
 387                 } else {
 388                         /*
 389                          * Stop the I/O in case of a fatal
 390                          * error.
 391                          */
 392                         update_error_count(td, err);
 393                         return 1;
 394                 }
 395         }
 396
 397         return 0;
 398 }
 399
 400 static void check_update_rusage(struct thread_data *td)
 401 {
 402         if (td->update_rusage) {
 403                 td->update_rusage = 0;
 404                 update_rusage_stat(td);
 405                 fio_mutex_up(td->rusage_sem);
 406         }
 407 }
 408
 409 /*
 410  * The main verify engine. Runs over the writes we previously submitted,
 411  * reads the blocks back in, and checks the crc/md5 of the data.
 412  */
 413 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 414 {
 415         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 416         struct fio_file *f;
 417         struct io_u *io_u;
 418         int ret, min_events;
 419         unsigned int i;
 420
 421         dprint(FD_VERIFY, "starting loop\n");
 422
 423         /*
 424          * sync io first and invalidate cache, to make sure we really
 425          * read from disk.
 426          */
 427         for_each_file(td, f, i) {
 428                 if (!fio_file_open(f))
 429                         continue;
 430                 if (fio_io_sync(td, f))
 431                         break;
 432                 if (file_invalidate_cache(td, f))
 433                         break;
 434         }
 435
 436         check_update_rusage(td);
 437
 438         if (td->error)
 439                 return;
 440
 441         td_set_runstate(td, TD_VERIFYING);
 442
 443         io_u = NULL;
 444         while (!td->terminate) {
 445                 enum fio_ddir ddir;
 446                 int ret2, full;
 447
 448                 update_tv_cache(td);
 449                 check_update_rusage(td);
 450
 451                 if (runtime_exceeded(td, &td->tv_cache)) {
 452                         __update_tv_cache(td);
 453                         if (runtime_exceeded(td, &td->tv_cache)) {
 454                                 td->terminate = 1;
 455                                 break;
 456                         }
 457                 }
 458
 459                 if (flow_threshold_exceeded(td))
 460                         continue;
 461
 462                 if (!td->o.experimental_verify) {
 463                         io_u = __get_io_u(td);
 464                         if (!io_u)
 465                                 break;
 466
 467                         if (get_next_verify(td, io_u)) {
 468                                 put_io_u(td, io_u);
 469                                 break;
 470                         }
 471
 472                         if (td_io_prep(td, io_u)) {
 473                                 put_io_u(td, io_u);
 474                                 break;
 475                         }
 476                 } else {
 477                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 478                                 break;
 479
 480                         while ((io_u = get_io_u(td)) != NULL) {
 481                                 if (IS_ERR(io_u)) {
 482                                         io_u = NULL;
 483                                         ret = FIO_Q_BUSY;
 484                                         goto reap;
 485                                 }
 486
 487                                 /*
 488                                  * We are only interested in the places where
 489                                  * we wrote or trimmed IOs. Turn those into
 490                                  * reads for verification purposes.
 491                                  */
 492                                 if (io_u->ddir == DDIR_READ) {
 493                                         /*
 494                                          * Pretend we issued it for rwmix
 495                                          * accounting
 496                                          */
 497                                         td->io_issues[DDIR_READ]++;
 498                                         put_io_u(td, io_u);
 499                                         continue;
 500                                 } else if (io_u->ddir == DDIR_TRIM) {
 501                                         io_u->ddir = DDIR_READ;
 502                                         io_u->flags |= IO_U_F_TRIMMED;
 503                                         break;
 504                                 } else if (io_u->ddir == DDIR_WRITE) {
 505                                         io_u->ddir = DDIR_READ;
 506                                         break;
 507                                 } else {
 508                                         put_io_u(td, io_u);
 509                                         continue;
 510                                 }
 511                         }
 512
 513                         if (!io_u)
 514                                 break;
 515                 }
 516
 517                 if (td->o.verify_async)
 518                         io_u->end_io = verify_io_u_async;
 519                 else
 520                         io_u->end_io = verify_io_u;
 521
 522                 ddir = io_u->ddir;
 523
 524                 ret = td_io_queue(td, io_u);
 525                 switch (ret) {
 526                 case FIO_Q_COMPLETED:
 527                         if (io_u->error) {
 528                                 ret = -io_u->error;
 529                                 clear_io_u(td, io_u);
 530                         } else if (io_u->resid) {
 531                                 int bytes = io_u->xfer_buflen - io_u->resid;
 532
 533                                 /*
 534                                  * zero read, fail
 535                                  */
 536                                 if (!bytes) {
 537                                         td_verror(td, EIO, "full resid");
 538                                         put_io_u(td, io_u);
 539                                         break;
 540                                 }
 541
 542                                 io_u->xfer_buflen = io_u->resid;
 543                                 io_u->xfer_buf += bytes;
 544                                 io_u->offset += bytes;
 545
 546                                 if (ddir_rw(io_u->ddir))
 547                                         td->ts.short_io_u[io_u->ddir]++;
 548
 549                                 f = io_u->file;
 550                                 if (io_u->offset == f->real_file_size)
 551                                         goto sync_done;
 552
 553                                 requeue_io_u(td, &io_u);
 554                         } else {
 555 sync_done:
 556                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 557                                 if (ret < 0)
 558                                         break;
 559                         }
 560                         continue;
 561                 case FIO_Q_QUEUED:
 562                         break;
 563                 case FIO_Q_BUSY:
 564                         requeue_io_u(td, &io_u);
 565                         ret2 = td_io_commit(td);
 566                         if (ret2 < 0)
 567                                 ret = ret2;
 568                         break;
 569                 default:
 570                         assert(ret < 0);
 571                         td_verror(td, -ret, "td_io_queue");
 572                         break;
 573                 }
 574
 575                 if (break_on_this_error(td, ddir, &ret))
 576                         break;
 577
 578                 /*
 579                  * if we can queue more, do so. but check if there are
 580                  * completed io_u's first. Note that we can get BUSY even
 581                  * without IO queued, if the system is resource starved.
 582                  */
 583 reap:
 584                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 585                 if (full || !td->o.iodepth_batch_complete) {
 586                         min_events = min(td->o.iodepth_batch_complete,
 587                                          td->cur_depth);
 588                         /*
 589                          * if the queue is full, we MUST reap at least 1 event
 590                          */
 591                         if (full && !min_events)
 592                                 min_events = 1;
 593
 594                         do {
 595                                 /*
 596                                  * Reap required number of io units, if any,
 597                                  * and do the verification on them through
 598                                  * the callback handler
 599                                  */
 600                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 601                                         ret = -1;
 602                                         break;
 603                                 }
 604                         } while (full && (td->cur_depth > td->o.iodepth_low));
 605                 }
 606                 if (ret < 0)
 607                         break;
 608         }
 609
 610         check_update_rusage(td);
 611
 612         if (!td->error) {
 613                 min_events = td->cur_depth;
 614
 615                 if (min_events)
 616                         ret = io_u_queued_complete(td, min_events, NULL);
 617         } else
 618                 cleanup_pending_aio(td);
 619
 620         td_set_runstate(td, TD_RUNNING);
 621
 622         dprint(FD_VERIFY, "exiting loop\n");
 623 }
 624
 625 static unsigned int exceeds_number_ios(struct thread_data *td)
 626 {
 627         unsigned long long number_ios;
 628
 629         if (!td->o.number_ios)
 630                 return 0;
 631
 632         number_ios = ddir_rw_sum(td->this_io_blocks);
 633         number_ios += td->io_u_queued + td->io_u_in_flight;
 634
 635         return number_ios >= td->o.number_ios;
 636 }
 637
 638 static int io_bytes_exceeded(struct thread_data *td)
 639 {
 640         unsigned long long bytes;
 641
 642         if (td_rw(td))
 643                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 644         else if (td_write(td))
 645                 bytes = td->this_io_bytes[DDIR_WRITE];
 646         else if (td_read(td))
 647                 bytes = td->this_io_bytes[DDIR_READ];
 648         else
 649                 bytes = td->this_io_bytes[DDIR_TRIM];
 650
 651         return bytes >= td->o.size || exceeds_number_ios(td);
 652 }
 653
 654 /*
 655  * Main IO worker function. It retrieves io_u's to process and queues
 656  * and reaps them, checking for rate and errors along the way.
 657  *
 658  * Returns number of bytes written and trimmed.
 659  */
 660 static uint64_t do_io(struct thread_data *td)
 661 {
 662         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 663         unsigned int i;
 664         int ret = 0;
 665         uint64_t total_bytes, bytes_issued = 0;
 666
 667         if (in_ramp_time(td))
 668                 td_set_runstate(td, TD_RAMP);
 669         else
 670                 td_set_runstate(td, TD_RUNNING);
 671
 672         lat_target_init(td);
 673
 674         /*
 675          * If verify_backlog is enabled, we'll run the verify in this
 676          * handler as well. For that case, we may need up to twice the
 677          * amount of bytes.
 678          */
 679         total_bytes = td->o.size;
 680         if (td->o.verify != VERIFY_NONE &&
 681            (td_write(td) && td->o.verify_backlog))
 682                 total_bytes += td->o.size;
 683
 684         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 685                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 686                 td->o.time_based) {
 687                 struct timeval comp_time;
 688                 int min_evts = 0;
 689                 struct io_u *io_u;
 690                 int ret2, full;
 691                 enum fio_ddir ddir;
 692
 693                 check_update_rusage(td);
 694
 695                 if (td->terminate || td->done)
 696                         break;
 697
 698                 update_tv_cache(td);
 699
 700                 if (runtime_exceeded(td, &td->tv_cache)) {
 701                         __update_tv_cache(td);
 702                         if (runtime_exceeded(td, &td->tv_cache)) {
 703                                 td->terminate = 1;
 704                                 break;
 705                         }
 706                 }
 707
 708                 if (flow_threshold_exceeded(td))
 709                         continue;
 710
 711                 if (bytes_issued >= total_bytes)
 712                         break;
 713
 714                 io_u = get_io_u(td);
 715                 if (IS_ERR_OR_NULL(io_u)) {
 716                         int err = PTR_ERR(io_u);
 717
 718                         io_u = NULL;
 719                         if (err == -EBUSY) {
 720                                 ret = FIO_Q_BUSY;
 721                                 goto reap;
 722                         }
 723                         if (td->o.latency_target)
 724                                 goto reap;
 725                         break;
 726                 }
 727
 728                 ddir = io_u->ddir;
 729
 730                 /*
 731                  * Add verification end_io handler if:
 732                  *      - Asked to verify (!td_rw(td))
 733                  *      - Or the io_u is from our verify list (mixed write/ver)
 734                  */
 735                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 736                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 737
 738                         if (!td->o.verify_pattern_bytes) {
 739                                 io_u->rand_seed = __rand(&td->__verify_state);
 740                                 if (sizeof(int) != sizeof(long *))
 741                                         io_u->rand_seed *= __rand(&td->__verify_state);
 742                         }
 743
 744                         if (td->o.verify_async)
 745                                 io_u->end_io = verify_io_u_async;
 746                         else
 747                                 io_u->end_io = verify_io_u;
 748                         td_set_runstate(td, TD_VERIFYING);
 749                 } else if (in_ramp_time(td))
 750                         td_set_runstate(td, TD_RAMP);
 751                 else
 752                         td_set_runstate(td, TD_RUNNING);
 753
 754                 /*
 755                  * Always log IO before it's issued, so we know the specific
 756                  * order of it. The logged unit will track when the IO has
 757                  * completed.
 758                  */
 759                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 760                     td->o.do_verify &&
 761                     td->o.verify != VERIFY_NONE &&
 762                     !td->o.experimental_verify)
 763                         log_io_piece(td, io_u);
 764
 765                 ret = td_io_queue(td, io_u);
 766                 switch (ret) {
 767                 case FIO_Q_COMPLETED:
 768                         if (io_u->error) {
 769                                 ret = -io_u->error;
 770                                 clear_io_u(td, io_u);
 771                         } else if (io_u->resid) {
 772                                 int bytes = io_u->xfer_buflen - io_u->resid;
 773                                 struct fio_file *f = io_u->file;
 774
 775                                 bytes_issued += bytes;
 776                                 /*
 777                                  * zero read, fail
 778                                  */
 779                                 if (!bytes) {
 780                                         td_verror(td, EIO, "full resid");
 781                                         put_io_u(td, io_u);
 782                                         break;
 783                                 }
 784
 785                                 io_u->xfer_buflen = io_u->resid;
 786                                 io_u->xfer_buf += bytes;
 787                                 io_u->offset += bytes;
 788
 789                                 if (ddir_rw(io_u->ddir))
 790                                         td->ts.short_io_u[io_u->ddir]++;
 791
 792                                 if (io_u->offset == f->real_file_size)
 793                                         goto sync_done;
 794
 795                                 requeue_io_u(td, &io_u);
 796                         } else {
 797 sync_done:
 798                                 if (__should_check_rate(td, DDIR_READ) ||
 799                                     __should_check_rate(td, DDIR_WRITE) ||
 800                                     __should_check_rate(td, DDIR_TRIM))
 801                                         fio_gettime(&comp_time, NULL);
 802
 803                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 804                                 if (ret < 0)
 805                                         break;
 806                                 bytes_issued += io_u->xfer_buflen;
 807                         }
 808                         break;
 809                 case FIO_Q_QUEUED:
 810                         /*
 811                          * if the engine doesn't have a commit hook,
 812                          * the io_u is really queued. if it does have such
 813                          * a hook, it has to call io_u_queued() itself.
 814                          */
 815                         if (td->io_ops->commit == NULL)
 816                                 io_u_queued(td, io_u);
 817                         bytes_issued += io_u->xfer_buflen;
 818                         break;
 819                 case FIO_Q_BUSY:
 820                         requeue_io_u(td, &io_u);
 821                         ret2 = td_io_commit(td);
 822                         if (ret2 < 0)
 823                                 ret = ret2;
 824                         break;
 825                 default:
 826                         assert(ret < 0);
 827                         put_io_u(td, io_u);
 828                         break;
 829                 }
 830
 831                 if (break_on_this_error(td, ddir, &ret))
 832                         break;
 833
 834                 /*
 835                  * See if we need to complete some commands. Note that we
 836                  * can get BUSY even without IO queued, if the system is
 837                  * resource starved.
 838                  */
 839 reap:
 840                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 841                 if (full || !td->o.iodepth_batch_complete) {
 842                         min_evts = min(td->o.iodepth_batch_complete,
 843                                         td->cur_depth);
 844                         /*
 845                          * if the queue is full, we MUST reap at least 1 event
 846                          */
 847                         if (full && !min_evts)
 848                                 min_evts = 1;
 849
 850                         if (__should_check_rate(td, DDIR_READ) ||
 851                             __should_check_rate(td, DDIR_WRITE) ||
 852                             __should_check_rate(td, DDIR_TRIM))
 853                                 fio_gettime(&comp_time, NULL);
 854
 855                         do {
 856                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 857                                 if (ret < 0)
 858                                         break;
 859
 860                         } while (full && (td->cur_depth > td->o.iodepth_low));
 861                 }
 862
 863                 if (ret < 0)
 864                         break;
 865                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 866                         continue;
 867
 868                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 869                         if (check_min_rate(td, &comp_time, bytes_done)) {
 870                                 if (exitall_on_terminate)
 871                                         fio_terminate_threads(td->groupid);
 872                                 td_verror(td, EIO, "check_min_rate");
 873                                 break;
 874                         }
 875                 }
 876                 if (!in_ramp_time(td) && td->o.latency_target)
 877                         lat_target_check(td);
 878
 879                 if (td->o.thinktime) {
 880                         unsigned long long b;
 881
 882                         b = ddir_rw_sum(td->io_blocks);
 883                         if (!(b % td->o.thinktime_blocks)) {
 884                                 int left;
 885
 886                                 io_u_quiesce(td);
 887
 888                                 if (td->o.thinktime_spin)
 889                                         usec_spin(td->o.thinktime_spin);
 890
 891                                 left = td->o.thinktime - td->o.thinktime_spin;
 892                                 if (left)
 893                                         usec_sleep(td, left);
 894                         }
 895                 }
 896         }
 897
 898         check_update_rusage(td);
 899
 900         if (td->trim_entries)
 901                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 902
 903         if (td->o.fill_device && td->error == ENOSPC) {
 904                 td->error = 0;
 905                 td->terminate = 1;
 906         }
 907         if (!td->error) {
 908                 struct fio_file *f;
 909
 910                 i = td->cur_depth;
 911                 if (i) {
 912                         ret = io_u_queued_complete(td, i, bytes_done);
 913                         if (td->o.fill_device && td->error == ENOSPC)
 914                                 td->error = 0;
 915                 }
 916
 917                 if (should_fsync(td) && td->o.end_fsync) {
 918                         td_set_runstate(td, TD_FSYNCING);
 919
 920                         for_each_file(td, f, i) {
 921                                 if (!fio_file_fsync(td, f))
 922                                         continue;
 923
 924                                 log_err("fio: end_fsync failed for file %s\n",
 925                                                                 f->file_name);
 926                         }
 927                 }
 928         } else
 929                 cleanup_pending_aio(td);
 930
 931         /*
 932          * stop job if we failed doing any IO
 933          */
 934         if (!ddir_rw_sum(td->this_io_bytes))
 935                 td->done = 1;
 936
 937         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 938 }
 939
 940 static void cleanup_io_u(struct thread_data *td)
 941 {
 942         struct io_u *io_u;
 943
 944         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 945
 946                 if (td->io_ops->io_u_free)
 947                         td->io_ops->io_u_free(td, io_u);
 948
 949                 fio_memfree(io_u, sizeof(*io_u));
 950         }
 951
 952         free_io_mem(td);
 953
 954         io_u_rexit(&td->io_u_requeues);
 955         io_u_qexit(&td->io_u_freelist);
 956         io_u_qexit(&td->io_u_all);
 957 }
 958
 959 static int init_io_u(struct thread_data *td)
 960 {
 961         struct io_u *io_u;
 962         unsigned int max_bs, min_write;
 963         int cl_align, i, max_units;
 964         int data_xfer = 1, err;
 965         char *p;
 966
 967         max_units = td->o.iodepth;
 968         max_bs = td_max_bs(td);
 969         min_write = td->o.min_bs[DDIR_WRITE];
 970         td->orig_buffer_size = (unsigned long long) max_bs
 971                                         * (unsigned long long) max_units;
 972
 973         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 974                 data_xfer = 0;
 975
 976         err = 0;
 977         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 978         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 979         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 980
 981         if (err) {
 982                 log_err("fio: failed setting up IO queues\n");
 983                 return 1;
 984         }
 985
 986         /*
 987          * if we may later need to do address alignment, then add any
 988          * possible adjustment here so that we don't cause a buffer
 989          * overflow later. this adjustment may be too much if we get
 990          * lucky and the allocator gives us an aligned address.
 991          */
 992         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 993             (td->io_ops->flags & FIO_RAWIO))
 994                 td->orig_buffer_size += page_mask + td->o.mem_align;
 995
 996         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 997                 unsigned long bs;
 998
 999                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1000                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1001         }
1002
1003         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1004                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1005                 return 1;
1006         }
1007
1008         if (data_xfer && allocate_io_mem(td))
1009                 return 1;
1010
1011         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1012             (td->io_ops->flags & FIO_RAWIO))
1013                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1014         else
1015                 p = td->orig_buffer;
1016
1017         cl_align = os_cache_line_size();
1018
1019         for (i = 0; i < max_units; i++) {
1020                 void *ptr;
1021
1022                 if (td->terminate)
1023                         return 1;
1024
1025                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1026                 if (!ptr) {
1027                         log_err("fio: unable to allocate aligned memory\n");
1028                         break;
1029                 }
1030
1031                 io_u = ptr;
1032                 memset(io_u, 0, sizeof(*io_u));
1033                 INIT_FLIST_HEAD(&io_u->verify_list);
1034                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1035
1036                 if (data_xfer) {
1037                         io_u->buf = p;
1038                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1039
1040                         if (td_write(td))
1041                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1042                         if (td_write(td) && td->o.verify_pattern_bytes) {
1043                                 /*
1044                                  * Fill the buffer with the pattern if we are
1045                                  * going to be doing writes.
1046                                  */
1047                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1048                         }
1049                 }
1050
1051                 io_u->index = i;
1052                 io_u->flags = IO_U_F_FREE;
1053                 io_u_qpush(&td->io_u_freelist, io_u);
1054
1055                 /*
1056                  * io_u never leaves this stack, used for iteration of all
1057                  * io_u buffers.
1058                  */
1059                 io_u_qpush(&td->io_u_all, io_u);
1060
1061                 if (td->io_ops->io_u_init) {
1062                         int ret = td->io_ops->io_u_init(td, io_u);
1063
1064                         if (ret) {
1065                                 log_err("fio: failed to init engine data: %d\n", ret);
1066                                 return 1;
1067                         }
1068                 }
1069
1070                 p += max_bs;
1071         }
1072
1073         return 0;
1074 }
1075
1076 static int switch_ioscheduler(struct thread_data *td)
1077 {
1078         char tmp[256], tmp2[128];
1079         FILE *f;
1080         int ret;
1081
1082         if (td->io_ops->flags & FIO_DISKLESSIO)
1083                 return 0;
1084
1085         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1086
1087         f = fopen(tmp, "r+");
1088         if (!f) {
1089                 if (errno == ENOENT) {
1090                         log_err("fio: os or kernel doesn't support IO scheduler"
1091                                 " switching\n");
1092                         return 0;
1093                 }
1094                 td_verror(td, errno, "fopen iosched");
1095                 return 1;
1096         }
1097
1098         /*
1099          * Set io scheduler.
1100          */
1101         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1102         if (ferror(f) || ret != 1) {
1103                 td_verror(td, errno, "fwrite");
1104                 fclose(f);
1105                 return 1;
1106         }
1107
1108         rewind(f);
1109
1110         /*
1111          * Read back and check that the selected scheduler is now the default.
1112          */
1113         ret = fread(tmp, 1, sizeof(tmp), f);
1114         if (ferror(f) || ret < 0) {
1115                 td_verror(td, errno, "fread");
1116                 fclose(f);
1117                 return 1;
1118         }
1119
1120         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1121         if (!strstr(tmp, tmp2)) {
1122                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1123                 td_verror(td, EINVAL, "iosched_switch");
1124                 fclose(f);
1125                 return 1;
1126         }
1127
1128         fclose(f);
1129         return 0;
1130 }
1131
1132 static int keep_running(struct thread_data *td)
1133 {
1134         if (td->done)
1135                 return 0;
1136         if (td->o.time_based)
1137                 return 1;
1138         if (td->o.loops) {
1139                 td->o.loops--;
1140                 return 1;
1141         }
1142         if (exceeds_number_ios(td))
1143                 return 0;
1144
1145         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1146                 uint64_t diff;
1147
1148                 /*
1149                  * If the difference is less than the minimum IO size, we
1150                  * are done.
1151                  */
1152                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1153                 if (diff < td_max_bs(td))
1154                         return 0;
1155
1156                 if (fio_files_done(td))
1157                         return 0;
1158
1159                 return 1;
1160         }
1161
1162         return 0;
1163 }
1164
1165 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1166 {
1167         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1168         char *str;
1169
1170         str = malloc(newlen);
1171         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1172
1173         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1174         ret = system(str);
1175         if (ret == -1)
1176                 log_err("fio: exec of cmd <%s> failed\n", str);
1177
1178         free(str);
1179         return ret;
1180 }
1181
1182 /*
1183  * Dry run to compute correct state of numberio for verification.
1184  */
1185 static uint64_t do_dry_run(struct thread_data *td)
1186 {
1187         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1188
1189         td_set_runstate(td, TD_RUNNING);
1190
1191         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1192                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1193                 struct io_u *io_u;
1194                 int ret;
1195
1196                 if (td->terminate || td->done)
1197                         break;
1198
1199                 io_u = get_io_u(td);
1200                 if (!io_u)
1201                         break;
1202
1203                 io_u->flags |= IO_U_F_FLIGHT;
1204                 io_u->error = 0;
1205                 io_u->resid = 0;
1206                 if (ddir_rw(acct_ddir(io_u)))
1207                         td->io_issues[acct_ddir(io_u)]++;
1208                 if (ddir_rw(io_u->ddir)) {
1209                         io_u_mark_depth(td, 1);
1210                         td->ts.total_io_u[io_u->ddir]++;
1211                 }
1212
1213                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1214                     td->o.do_verify &&
1215                     td->o.verify != VERIFY_NONE &&
1216                     !td->o.experimental_verify)
1217                         log_io_piece(td, io_u);
1218
1219                 ret = io_u_sync_complete(td, io_u, bytes_done);
1220                 (void) ret;
1221         }
1222
1223         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1224 }
1225
1226 static int write_this_log(struct thread_data *td, struct io_log *log,
1227                           const char *log_file, const char *name, int try)
1228 {
1229         int ret;
1230
1231         if (!log)
1232                 return 0;
1233
1234         if (log_file)
1235                 ret = finish_log_named(td, log, log_file, name, try);
1236         else
1237                 ret = finish_log(td, log, name, try);
1238
1239         return ret;
1240 }
1241
1242 static int write_iops_log(struct thread_data *td, struct thread_options *o,
1243                           int try)
1244 {
1245         return write_this_log(td, td->iops_log, o->iops_log_file, "iops", try);
1246 }
1247
1248 static int write_slat_log(struct thread_data *td, struct thread_options *o,
1249                           int try)
1250 {
1251         return write_this_log(td, td->slat_log, o->lat_log_file, "slat", try);
1252 }
1253
1254 static int write_clat_log(struct thread_data *td, struct thread_options *o,
1255                           int try)
1256 {
1257         return write_this_log(td, td->clat_log, o->lat_log_file, "clat" , try);
1258 }
1259
1260 static int write_lat_log(struct thread_data *td, struct thread_options *o,
1261                          int try)
1262 {
1263         return write_this_log(td, td->lat_log, o->lat_log_file, "lat", try);
1264 }
1265
1266 static int write_bandw_log(struct thread_data *td, struct thread_options *o,
1267                         int try)
1268 {
1269         return write_this_log(td, td->bw_log, o->bw_log_file, "bw", try);
1270 }
1271
1272 enum {
1273         BW_LOG_MASK     = 1,
1274         LAT_LOG_MASK    = 2,
1275         SLAT_LOG_MASK   = 4,
1276         CLAT_LOG_MASK   = 8,
1277         IOPS_LOG_MASK   = 16,
1278
1279         ALL_LOG_MASK    = 31,
1280         ALL_LOG_NR      = 5,
1281 };
1282
1283 static void writeout_logs(struct thread_data *td)
1284 {
1285         struct thread_options *o = &td->o;
1286         unsigned int log_mask = ALL_LOG_MASK;
1287         unsigned int log_left = ALL_LOG_NR;
1288         int old_state;
1289
1290         old_state = td_bump_runstate(td, TD_FINISHING);
1291
1292         finalize_logs(td);
1293
1294         while (log_left) {
1295                 int ret, prev_log_left = log_left;
1296
1297                 if (log_mask & BW_LOG_MASK) {
1298                         ret = write_bandw_log(td, o, log_left != 1);
1299                         if (!ret) {
1300                                 log_left--;
1301                                 log_mask &= ~BW_LOG_MASK;
1302                         }
1303                 }
1304                 if (log_mask & LAT_LOG_MASK) {
1305                         ret = write_lat_log(td, o, log_left != 1);
1306                         if (!ret) {
1307                                 log_left--;
1308                                 log_mask &= ~LAT_LOG_MASK;
1309                         }
1310                 }
1311                 if (log_mask & SLAT_LOG_MASK) {
1312                         ret = write_slat_log(td, o, log_left != 1);
1313                         if (!ret) {
1314                                 log_left--;
1315                                 log_mask &= ~SLAT_LOG_MASK;
1316                         }
1317                 }
1318                 if (log_mask & CLAT_LOG_MASK) {
1319                         ret = write_clat_log(td, o, log_left != 1);
1320                         if (!ret) {
1321                                 log_left--;
1322                                 log_mask &= ~CLAT_LOG_MASK;
1323                         }
1324                 }
1325                 if (log_mask & IOPS_LOG_MASK) {
1326                         ret = write_iops_log(td, o, log_left != 1);
1327                         if (!ret) {
1328                                 log_left--;
1329                                 log_mask &= ~IOPS_LOG_MASK;
1330                         }
1331                 }
1332
1333                 if (prev_log_left == log_left)
1334                         usleep(5000);
1335         }
1336
1337         td_restore_runstate(td, old_state);
1338 }
1339
1340 /*
1341  * Entry point for the thread based jobs. The process based jobs end up
1342  * here as well, after a little setup.
1343  */
1344 static void *thread_main(void *data)
1345 {
1346         unsigned long long elapsed;
1347         struct thread_data *td = data;
1348         struct thread_options *o = &td->o;
1349         pthread_condattr_t attr;
1350         int clear_state;
1351         int ret;
1352
1353         if (!o->use_thread) {
1354                 setsid();
1355                 td->pid = getpid();
1356         } else
1357                 td->pid = gettid();
1358
1359         /*
1360          * fio_time_init() may not have been called yet if running as a server
1361          */
1362         fio_time_init();
1363
1364         fio_local_clock_init(o->use_thread);
1365
1366         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1367
1368         if (is_backend)
1369                 fio_server_send_start(td);
1370
1371         INIT_FLIST_HEAD(&td->io_log_list);
1372         INIT_FLIST_HEAD(&td->io_hist_list);
1373         INIT_FLIST_HEAD(&td->verify_list);
1374         INIT_FLIST_HEAD(&td->trim_list);
1375         INIT_FLIST_HEAD(&td->next_rand_list);
1376         pthread_mutex_init(&td->io_u_lock, NULL);
1377         td->io_hist_tree = RB_ROOT;
1378
1379         pthread_condattr_init(&attr);
1380         pthread_cond_init(&td->verify_cond, &attr);
1381         pthread_cond_init(&td->free_cond, &attr);
1382
1383         td_set_runstate(td, TD_INITIALIZED);
1384         dprint(FD_MUTEX, "up startup_mutex\n");
1385         fio_mutex_up(startup_mutex);
1386         dprint(FD_MUTEX, "wait on td->mutex\n");
1387         fio_mutex_down(td->mutex);
1388         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1389
1390         /*
1391          * A new gid requires privilege, so we need to do this before setting
1392          * the uid.
1393          */
1394         if (o->gid != -1U && setgid(o->gid)) {
1395                 td_verror(td, errno, "setgid");
1396                 goto err;
1397         }
1398         if (o->uid != -1U && setuid(o->uid)) {
1399                 td_verror(td, errno, "setuid");
1400                 goto err;
1401         }
1402
1403         /*
1404          * If we have a gettimeofday() thread, make sure we exclude that
1405          * thread from this job
1406          */
1407         if (o->gtod_cpu)
1408                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1409
1410         /*
1411          * Set affinity first, in case it has an impact on the memory
1412          * allocations.
1413          */
1414         if (o->cpumask_set) {
1415                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1416                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1417                         if (!ret) {
1418                                 log_err("fio: no CPUs set\n");
1419                                 log_err("fio: Try increasing number of available CPUs\n");
1420                                 td_verror(td, EINVAL, "cpus_split");
1421                                 goto err;
1422                         }
1423                 }
1424                 ret = fio_setaffinity(td->pid, o->cpumask);
1425                 if (ret == -1) {
1426                         td_verror(td, errno, "cpu_set_affinity");
1427                         goto err;
1428                 }
1429         }
1430
1431 #ifdef CONFIG_LIBNUMA
1432         /* numa node setup */
1433         if (o->numa_cpumask_set || o->numa_memmask_set) {
1434                 int ret;
1435
1436                 if (numa_available() < 0) {
1437                         td_verror(td, errno, "Does not support NUMA API\n");
1438                         goto err;
1439                 }
1440
1441                 if (o->numa_cpumask_set) {
1442                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1443                         if (ret == -1) {
1444                                 td_verror(td, errno, \
1445                                         "numa_run_on_node_mask failed\n");
1446                                 goto err;
1447                         }
1448                 }
1449
1450                 if (o->numa_memmask_set) {
1451
1452                         switch (o->numa_mem_mode) {
1453                         case MPOL_INTERLEAVE:
1454                                 numa_set_interleave_mask(o->numa_memnodesmask);
1455                                 break;
1456                         case MPOL_BIND:
1457                                 numa_set_membind(o->numa_memnodesmask);
1458                                 break;
1459                         case MPOL_LOCAL:
1460                                 numa_set_localalloc();
1461                                 break;
1462                         case MPOL_PREFERRED:
1463                                 numa_set_preferred(o->numa_mem_prefer_node);
1464                                 break;
1465                         case MPOL_DEFAULT:
1466                         default:
1467                                 break;
1468                         }
1469
1470                 }
1471         }
1472 #endif
1473
1474         if (fio_pin_memory(td))
1475                 goto err;
1476
1477         /*
1478          * May alter parameters that init_io_u() will use, so we need to
1479          * do this first.
1480          */
1481         if (init_iolog(td))
1482                 goto err;
1483
1484         if (init_io_u(td))
1485                 goto err;
1486
1487         if (o->verify_async && verify_async_init(td))
1488                 goto err;
1489
1490         if (o->ioprio) {
1491                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1492                 if (ret == -1) {
1493                         td_verror(td, errno, "ioprio_set");
1494                         goto err;
1495                 }
1496         }
1497
1498         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1499                 goto err;
1500
1501         errno = 0;
1502         if (nice(o->nice) == -1 && errno != 0) {
1503                 td_verror(td, errno, "nice");
1504                 goto err;
1505         }
1506
1507         if (o->ioscheduler && switch_ioscheduler(td))
1508                 goto err;
1509
1510         if (!o->create_serialize && setup_files(td))
1511                 goto err;
1512
1513         if (td_io_init(td))
1514                 goto err;
1515
1516         if (init_random_map(td))
1517                 goto err;
1518
1519         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1520                 goto err;
1521
1522         if (o->pre_read) {
1523                 if (pre_read_files(td) < 0)
1524                         goto err;
1525         }
1526
1527         fio_verify_init(td);
1528
1529         fio_gettime(&td->epoch, NULL);
1530         fio_getrusage(&td->ru_start);
1531         clear_state = 0;
1532         while (keep_running(td)) {
1533                 uint64_t verify_bytes;
1534
1535                 fio_gettime(&td->start, NULL);
1536                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1537                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1538                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1539
1540                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1541                                 o->ratemin[DDIR_TRIM]) {
1542                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1543                                                 sizeof(td->bw_sample_time));
1544                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1545                                                 sizeof(td->bw_sample_time));
1546                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1547                                                 sizeof(td->bw_sample_time));
1548                 }
1549
1550                 if (clear_state)
1551                         clear_io_state(td);
1552
1553                 prune_io_piece_log(td);
1554
1555                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1556                         verify_bytes = do_dry_run(td);
1557                 else
1558                         verify_bytes = do_io(td);
1559
1560                 clear_state = 1;
1561
1562                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1563                         elapsed = utime_since_now(&td->start);
1564                         td->ts.runtime[DDIR_READ] += elapsed;
1565                 }
1566                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1567                         elapsed = utime_since_now(&td->start);
1568                         td->ts.runtime[DDIR_WRITE] += elapsed;
1569                 }
1570                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1571                         elapsed = utime_since_now(&td->start);
1572                         td->ts.runtime[DDIR_TRIM] += elapsed;
1573                 }
1574
1575                 if (td->error || td->terminate)
1576                         break;
1577
1578                 if (!o->do_verify ||
1579                     o->verify == VERIFY_NONE ||
1580                     (td->io_ops->flags & FIO_UNIDIR))
1581                         continue;
1582
1583                 clear_io_state(td);
1584
1585                 fio_gettime(&td->start, NULL);
1586
1587                 do_verify(td, verify_bytes);
1588
1589                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1590
1591                 if (td->error || td->terminate)
1592                         break;
1593         }
1594
1595         update_rusage_stat(td);
1596         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1597         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1598         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1599         td->ts.total_run_time = mtime_since_now(&td->epoch);
1600         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1601         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1602         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1603
1604         fio_unpin_memory(td);
1605
1606         writeout_logs(td);
1607
1608         if (o->exec_postrun)
1609                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1610
1611         if (exitall_on_terminate)
1612                 fio_terminate_threads(td->groupid);
1613
1614 err:
1615         if (td->error)
1616                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1617                                                         td->verror);
1618
1619         if (o->verify_async)
1620                 verify_async_exit(td);
1621
1622         close_and_free_files(td);
1623         cleanup_io_u(td);
1624         close_ioengine(td);
1625         cgroup_shutdown(td, &cgroup_mnt);
1626
1627         if (o->cpumask_set) {
1628                 int ret = fio_cpuset_exit(&o->cpumask);
1629
1630                 td_verror(td, ret, "fio_cpuset_exit");
1631         }
1632
1633         /*
1634          * do this very late, it will log file closing as well
1635          */
1636         if (o->write_iolog_file)
1637                 write_iolog_close(td);
1638
1639         fio_mutex_remove(td->rusage_sem);
1640         td->rusage_sem = NULL;
1641
1642         fio_mutex_remove(td->mutex);
1643         td->mutex = NULL;
1644
1645         td_set_runstate(td, TD_EXITED);
1646         return (void *) (uintptr_t) td->error;
1647 }
1648
1649
1650 /*
1651  * We cannot pass the td data into a forked process, so attach the td and
1652  * pass it to the thread worker.
1653  */
1654 static int fork_main(int shmid, int offset)
1655 {
1656         struct thread_data *td;
1657         void *data, *ret;
1658
1659 #ifndef __hpux
1660         data = shmat(shmid, NULL, 0);
1661         if (data == (void *) -1) {
1662                 int __err = errno;
1663
1664                 perror("shmat");
1665                 return __err;
1666         }
1667 #else
1668         /*
1669          * HP-UX inherits shm mappings?
1670          */
1671         data = threads;
1672 #endif
1673
1674         td = data + offset * sizeof(struct thread_data);
1675         ret = thread_main(td);
1676         shmdt(data);
1677         return (int) (uintptr_t) ret;
1678 }
1679
1680 /*
1681  * Run over the job map and reap the threads that have exited, if any.
1682  */
1683 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1684                          unsigned int *m_rate)
1685 {
1686         struct thread_data *td;
1687         unsigned int cputhreads, realthreads, pending;
1688         int i, status, ret;
1689
1690         /*
1691          * reap exited threads (TD_EXITED -> TD_REAPED)
1692          */
1693         realthreads = pending = cputhreads = 0;
1694         for_each_td(td, i) {
1695                 int flags = 0;
1696
1697                 /*
1698                  * ->io_ops is NULL for a thread that has closed its
1699                  * io engine
1700                  */
1701                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1702                         cputhreads++;
1703                 else
1704                         realthreads++;
1705
1706                 if (!td->pid) {
1707                         pending++;
1708                         continue;
1709                 }
1710                 if (td->runstate == TD_REAPED)
1711                         continue;
1712                 if (td->o.use_thread) {
1713                         if (td->runstate == TD_EXITED) {
1714                                 td_set_runstate(td, TD_REAPED);
1715                                 goto reaped;
1716                         }
1717                         continue;
1718                 }
1719
1720                 flags = WNOHANG;
1721                 if (td->runstate == TD_EXITED)
1722                         flags = 0;
1723
1724                 /*
1725                  * check if someone quit or got killed in an unusual way
1726                  */
1727                 ret = waitpid(td->pid, &status, flags);
1728                 if (ret < 0) {
1729                         if (errno == ECHILD) {
1730                                 log_err("fio: pid=%d disappeared %d\n",
1731                                                 (int) td->pid, td->runstate);
1732                                 td->sig = ECHILD;
1733                                 td_set_runstate(td, TD_REAPED);
1734                                 goto reaped;
1735                         }
1736                         perror("waitpid");
1737                 } else if (ret == td->pid) {
1738                         if (WIFSIGNALED(status)) {
1739                                 int sig = WTERMSIG(status);
1740
1741                                 if (sig != SIGTERM && sig != SIGUSR2)
1742                                         log_err("fio: pid=%d, got signal=%d\n",
1743                                                         (int) td->pid, sig);
1744                                 td->sig = sig;
1745                                 td_set_runstate(td, TD_REAPED);
1746                                 goto reaped;
1747                         }
1748                         if (WIFEXITED(status)) {
1749                                 if (WEXITSTATUS(status) && !td->error)
1750                                         td->error = WEXITSTATUS(status);
1751
1752                                 td_set_runstate(td, TD_REAPED);
1753                                 goto reaped;
1754                         }
1755                 }
1756
1757                 /*
1758                  * thread is not dead, continue
1759                  */
1760                 pending++;
1761                 continue;
1762 reaped:
1763                 (*nr_running)--;
1764                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1765                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1766                 if (!td->pid)
1767                         pending--;
1768
1769                 if (td->error)
1770                         exit_value++;
1771
1772                 done_secs += mtime_since_now(&td->epoch) / 1000;
1773                 profile_td_exit(td);
1774         }
1775
1776         if (*nr_running == cputhreads && !pending && realthreads)
1777                 fio_terminate_threads(TERMINATE_ALL);
1778 }
1779
1780 static void do_usleep(unsigned int usecs)
1781 {
1782         check_for_running_stats();
1783         usleep(usecs);
1784 }
1785
1786 /*
1787  * Main function for kicking off and reaping jobs, as needed.
1788  */
1789 static void run_threads(void)
1790 {
1791         struct thread_data *td;
1792         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1793         uint64_t spent;
1794
1795         if (fio_gtod_offload && fio_start_gtod_thread())
1796                 return;
1797
1798         fio_idle_prof_init();
1799
1800         set_sig_handlers();
1801
1802         nr_thread = nr_process = 0;
1803         for_each_td(td, i) {
1804                 if (td->o.use_thread)
1805                         nr_thread++;
1806                 else
1807                         nr_process++;
1808         }
1809
1810         if (output_format == FIO_OUTPUT_NORMAL) {
1811                 log_info("Starting ");
1812                 if (nr_thread)
1813                         log_info("%d thread%s", nr_thread,
1814                                                 nr_thread > 1 ? "s" : "");
1815                 if (nr_process) {
1816                         if (nr_thread)
1817                                 log_info(" and ");
1818                         log_info("%d process%s", nr_process,
1819                                                 nr_process > 1 ? "es" : "");
1820                 }
1821                 log_info("\n");
1822                 fflush(stdout);
1823         }
1824
1825         todo = thread_number;
1826         nr_running = 0;
1827         nr_started = 0;
1828         m_rate = t_rate = 0;
1829
1830         for_each_td(td, i) {
1831                 print_status_init(td->thread_number - 1);
1832
1833                 if (!td->o.create_serialize)
1834                         continue;
1835
1836                 /*
1837                  * do file setup here so it happens sequentially,
1838                  * we don't want X number of threads getting their
1839                  * client data interspersed on disk
1840                  */
1841                 if (setup_files(td)) {
1842                         exit_value++;
1843                         if (td->error)
1844                                 log_err("fio: pid=%d, err=%d/%s\n",
1845                                         (int) td->pid, td->error, td->verror);
1846                         td_set_runstate(td, TD_REAPED);
1847                         todo--;
1848                 } else {
1849                         struct fio_file *f;
1850                         unsigned int j;
1851
1852                         /*
1853                          * for sharing to work, each job must always open
1854                          * its own files. so close them, if we opened them
1855                          * for creation
1856                          */
1857                         for_each_file(td, f, j) {
1858                                 if (fio_file_open(f))
1859                                         td_io_close_file(td, f);
1860                         }
1861                 }
1862         }
1863
1864         /* start idle threads before io threads start to run */
1865         fio_idle_prof_start();
1866
1867         set_genesis_time();
1868
1869         while (todo) {
1870                 struct thread_data *map[REAL_MAX_JOBS];
1871                 struct timeval this_start;
1872                 int this_jobs = 0, left;
1873
1874                 /*
1875                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1876                  */
1877                 for_each_td(td, i) {
1878                         if (td->runstate != TD_NOT_CREATED)
1879                                 continue;
1880
1881                         /*
1882                          * never got a chance to start, killed by other
1883                          * thread for some reason
1884                          */
1885                         if (td->terminate) {
1886                                 todo--;
1887                                 continue;
1888                         }
1889
1890                         if (td->o.start_delay) {
1891                                 spent = utime_since_genesis();
1892
1893                                 if (td->o.start_delay > spent)
1894                                         continue;
1895                         }
1896
1897                         if (td->o.stonewall && (nr_started || nr_running)) {
1898                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1899                                                         td->o.name);
1900                                 break;
1901                         }
1902
1903                         init_disk_util(td);
1904
1905                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1906                         td->update_rusage = 0;
1907
1908                         /*
1909                          * Set state to created. Thread will transition
1910                          * to TD_INITIALIZED when it's done setting up.
1911                          */
1912                         td_set_runstate(td, TD_CREATED);
1913                         map[this_jobs++] = td;
1914                         nr_started++;
1915
1916                         if (td->o.use_thread) {
1917                                 int ret;
1918
1919                                 dprint(FD_PROCESS, "will pthread_create\n");
1920                                 ret = pthread_create(&td->thread, NULL,
1921                                                         thread_main, td);
1922                                 if (ret) {
1923                                         log_err("pthread_create: %s\n",
1924                                                         strerror(ret));
1925                                         nr_started--;
1926                                         break;
1927                                 }
1928                                 ret = pthread_detach(td->thread);
1929                                 if (ret)
1930                                         log_err("pthread_detach: %s",
1931                                                         strerror(ret));
1932                         } else {
1933                                 pid_t pid;
1934                                 dprint(FD_PROCESS, "will fork\n");
1935                                 pid = fork();
1936                                 if (!pid) {
1937                                         int ret = fork_main(shm_id, i);
1938
1939                                         _exit(ret);
1940                                 } else if (i == fio_debug_jobno)
1941                                         *fio_debug_jobp = pid;
1942                         }
1943                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1944                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1945                                 log_err("fio: job startup hung? exiting.\n");
1946                                 fio_terminate_threads(TERMINATE_ALL);
1947                                 fio_abort = 1;
1948                                 nr_started--;
1949                                 break;
1950                         }
1951                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1952                 }
1953
1954                 /*
1955                  * Wait for the started threads to transition to
1956                  * TD_INITIALIZED.
1957                  */
1958                 fio_gettime(&this_start, NULL);
1959                 left = this_jobs;
1960                 while (left && !fio_abort) {
1961                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1962                                 break;
1963
1964                         do_usleep(100000);
1965
1966                         for (i = 0; i < this_jobs; i++) {
1967                                 td = map[i];
1968                                 if (!td)
1969                                         continue;
1970                                 if (td->runstate == TD_INITIALIZED) {
1971                                         map[i] = NULL;
1972                                         left--;
1973                                 } else if (td->runstate >= TD_EXITED) {
1974                                         map[i] = NULL;
1975                                         left--;
1976                                         todo--;
1977                                         nr_running++; /* work-around... */
1978                                 }
1979                         }
1980                 }
1981
1982                 if (left) {
1983                         log_err("fio: %d job%s failed to start\n", left,
1984                                         left > 1 ? "s" : "");
1985                         for (i = 0; i < this_jobs; i++) {
1986                                 td = map[i];
1987                                 if (!td)
1988                                         continue;
1989                                 kill(td->pid, SIGTERM);
1990                         }
1991                         break;
1992                 }
1993
1994                 /*
1995                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1996                  */
1997                 for_each_td(td, i) {
1998                         if (td->runstate != TD_INITIALIZED)
1999                                 continue;
2000
2001                         if (in_ramp_time(td))
2002                                 td_set_runstate(td, TD_RAMP);
2003                         else
2004                                 td_set_runstate(td, TD_RUNNING);
2005                         nr_running++;
2006                         nr_started--;
2007                         m_rate += ddir_rw_sum(td->o.ratemin);
2008                         t_rate += ddir_rw_sum(td->o.rate);
2009                         todo--;
2010                         fio_mutex_up(td->mutex);
2011                 }
2012
2013                 reap_threads(&nr_running, &t_rate, &m_rate);
2014
2015                 if (todo)
2016                         do_usleep(100000);
2017         }
2018
2019         while (nr_running) {
2020                 reap_threads(&nr_running, &t_rate, &m_rate);
2021                 do_usleep(10000);
2022         }
2023
2024         fio_idle_prof_stop();
2025
2026         update_io_ticks();
2027 }
2028
2029 void wait_for_disk_thread_exit(void)
2030 {
2031         fio_mutex_down(disk_thread_mutex);
2032 }
2033
2034 static void free_disk_util(void)
2035 {
2036         disk_util_start_exit();
2037         wait_for_disk_thread_exit();
2038         disk_util_prune_entries();
2039 }
2040
2041 static void *disk_thread_main(void *data)
2042 {
2043         int ret = 0;
2044
2045         fio_mutex_up(startup_mutex);
2046
2047         while (threads && !ret) {
2048                 usleep(DISK_UTIL_MSEC * 1000);
2049                 if (!threads)
2050                         break;
2051                 ret = update_io_ticks();
2052
2053                 if (!is_backend)
2054                         print_thread_status();
2055         }
2056
2057         fio_mutex_up(disk_thread_mutex);
2058         return NULL;
2059 }
2060
2061 static int create_disk_util_thread(void)
2062 {
2063         int ret;
2064
2065         setup_disk_util();
2066
2067         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2068
2069         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
2070         if (ret) {
2071                 fio_mutex_remove(disk_thread_mutex);
2072                 log_err("Can't create disk util thread: %s\n", strerror(ret));
2073                 return 1;
2074         }
2075
2076         ret = pthread_detach(disk_util_thread);
2077         if (ret) {
2078                 fio_mutex_remove(disk_thread_mutex);
2079                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
2080                 return 1;
2081         }
2082
2083         dprint(FD_MUTEX, "wait on startup_mutex\n");
2084         fio_mutex_down(startup_mutex);
2085         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2086         return 0;
2087 }
2088
2089 int fio_backend(void)
2090 {
2091         struct thread_data *td;
2092         int i;
2093
2094         if (exec_profile) {
2095                 if (load_profile(exec_profile))
2096                         return 1;
2097                 free(exec_profile);
2098                 exec_profile = NULL;
2099         }
2100         if (!thread_number)
2101                 return 0;
2102
2103         if (write_bw_log) {
2104                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
2105                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
2106                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2107         }
2108
2109         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2110         if (startup_mutex == NULL)
2111                 return 1;
2112
2113         set_genesis_time();
2114         stat_init();
2115         create_disk_util_thread();
2116
2117         cgroup_list = smalloc(sizeof(*cgroup_list));
2118         INIT_FLIST_HEAD(cgroup_list);
2119
2120         run_threads();
2121
2122         if (!fio_abort) {
2123                 show_run_stats();
2124                 if (write_bw_log) {
2125                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2126                         __finish_log(agg_io_log[DDIR_WRITE],
2127                                         "agg-write_bw.log");
2128                         __finish_log(agg_io_log[DDIR_TRIM],
2129                                         "agg-write_bw.log");
2130                 }
2131         }
2132
2133         for_each_td(td, i)
2134                 fio_options_free(td);
2135
2136         free_disk_util();
2137         cgroup_kill(cgroup_list);
2138         sfree(cgroup_list);
2139         sfree(cgroup_mnt);
2140
2141         fio_mutex_remove(startup_mutex);
2142         fio_mutex_remove(disk_thread_mutex);
2143         stat_exit();
2144         return exit_value;
2145 }