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 #include "lib/tp.h"
  57
  58 static pthread_t helper_thread;
  59 static pthread_mutex_t helper_lock;
  60 pthread_cond_t helper_cond;
  61 int helper_do_stat = 0;
  62
  63 static struct fio_mutex *startup_mutex;
  64 static struct flist_head *cgroup_list;
  65 static char *cgroup_mnt;
  66 static int exit_value;
  67 static volatile int fio_abort;
  68 static unsigned int nr_process = 0;
  69 static unsigned int nr_thread = 0;
  70
  71 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  72
  73 int groupid = 0;
  74 unsigned int thread_number = 0;
  75 unsigned int stat_number = 0;
  76 int shm_id = 0;
  77 int temp_stall_ts;
  78 unsigned long done_secs = 0;
  79 volatile int helper_exit = 0;
  80
  81 #define PAGE_ALIGN(buf) \
  82         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  83
  84 #define JOB_START_TIMEOUT       (5 * 1000)
  85
  86 static void sig_int(int sig)
  87 {
  88         if (threads) {
  89                 if (is_backend)
  90                         fio_server_got_signal(sig);
  91                 else {
  92                         log_info("\nfio: terminating on signal %d\n", sig);
  93                         log_info_flush();
  94                         exit_value = 128;
  95                 }
  96
  97                 fio_terminate_threads(TERMINATE_ALL);
  98         }
  99 }
 100
 101 static void sig_show_status(int sig)
 102 {
 103         show_running_run_stats();
 104 }
 105
 106 static void set_sig_handlers(void)
 107 {
 108         struct sigaction act;
 109
 110         memset(&act, 0, sizeof(act));
 111         act.sa_handler = sig_int;
 112         act.sa_flags = SA_RESTART;
 113         sigaction(SIGINT, &act, NULL);
 114
 115         memset(&act, 0, sizeof(act));
 116         act.sa_handler = sig_int;
 117         act.sa_flags = SA_RESTART;
 118         sigaction(SIGTERM, &act, NULL);
 119
 120 /* Windows uses SIGBREAK as a quit signal from other applications */
 121 #ifdef WIN32
 122         memset(&act, 0, sizeof(act));
 123         act.sa_handler = sig_int;
 124         act.sa_flags = SA_RESTART;
 125         sigaction(SIGBREAK, &act, NULL);
 126 #endif
 127
 128         memset(&act, 0, sizeof(act));
 129         act.sa_handler = sig_show_status;
 130         act.sa_flags = SA_RESTART;
 131         sigaction(SIGUSR1, &act, NULL);
 132
 133         if (is_backend) {
 134                 memset(&act, 0, sizeof(act));
 135                 act.sa_handler = sig_int;
 136                 act.sa_flags = SA_RESTART;
 137                 sigaction(SIGPIPE, &act, NULL);
 138         }
 139 }
 140
 141 /*
 142  * Check if we are above the minimum rate given.
 143  */
 144 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 145                             enum fio_ddir ddir)
 146 {
 147         unsigned long long bytes = 0;
 148         unsigned long iops = 0;
 149         unsigned long spent;
 150         unsigned long rate;
 151         unsigned int ratemin = 0;
 152         unsigned int rate_iops = 0;
 153         unsigned int rate_iops_min = 0;
 154
 155         assert(ddir_rw(ddir));
 156
 157         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 158                 return 0;
 159
 160         /*
 161          * allow a 2 second settle period in the beginning
 162          */
 163         if (mtime_since(&td->start, now) < 2000)
 164                 return 0;
 165
 166         iops += td->this_io_blocks[ddir];
 167         bytes += td->this_io_bytes[ddir];
 168         ratemin += td->o.ratemin[ddir];
 169         rate_iops += td->o.rate_iops[ddir];
 170         rate_iops_min += td->o.rate_iops_min[ddir];
 171
 172         /*
 173          * if rate blocks is set, sample is running
 174          */
 175         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 176                 spent = mtime_since(&td->lastrate[ddir], now);
 177                 if (spent < td->o.ratecycle)
 178                         return 0;
 179
 180                 if (td->o.rate[ddir]) {
 181                         /*
 182                          * check bandwidth specified rate
 183                          */
 184                         if (bytes < td->rate_bytes[ddir]) {
 185                                 log_err("%s: min rate %u not met\n", td->o.name,
 186                                                                 ratemin);
 187                                 return 1;
 188                         } else {
 189                                 if (spent)
 190                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 191                                 else
 192                                         rate = 0;
 193
 194                                 if (rate < ratemin ||
 195                                     bytes < td->rate_bytes[ddir]) {
 196                                         log_err("%s: min rate %u not met, got"
 197                                                 " %luKB/sec\n", td->o.name,
 198                                                         ratemin, rate);
 199                                         return 1;
 200                                 }
 201                         }
 202                 } else {
 203                         /*
 204                          * checks iops specified rate
 205                          */
 206                         if (iops < rate_iops) {
 207                                 log_err("%s: min iops rate %u not met\n",
 208                                                 td->o.name, rate_iops);
 209                                 return 1;
 210                         } else {
 211                                 if (spent)
 212                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 213                                 else
 214                                         rate = 0;
 215
 216                                 if (rate < rate_iops_min ||
 217                                     iops < td->rate_blocks[ddir]) {
 218                                         log_err("%s: min iops rate %u not met,"
 219                                                 " got %lu\n", td->o.name,
 220                                                         rate_iops_min, rate);
 221                                 }
 222                         }
 223                 }
 224         }
 225
 226         td->rate_bytes[ddir] = bytes;
 227         td->rate_blocks[ddir] = iops;
 228         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 229         return 0;
 230 }
 231
 232 static int check_min_rate(struct thread_data *td, struct timeval *now,
 233                           uint64_t *bytes_done)
 234 {
 235         int ret = 0;
 236
 237         if (bytes_done[DDIR_READ])
 238                 ret |= __check_min_rate(td, now, DDIR_READ);
 239         if (bytes_done[DDIR_WRITE])
 240                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 241         if (bytes_done[DDIR_TRIM])
 242                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 243
 244         return ret;
 245 }
 246
 247 /*
 248  * When job exits, we can cancel the in-flight IO if we are using async
 249  * io. Attempt to do so.
 250  */
 251 static void cleanup_pending_aio(struct thread_data *td)
 252 {
 253         int r;
 254
 255         /*
 256          * get immediately available events, if any
 257          */
 258         r = io_u_queued_complete(td, 0, NULL);
 259         if (r < 0)
 260                 return;
 261
 262         /*
 263          * now cancel remaining active events
 264          */
 265         if (td->io_ops->cancel) {
 266                 struct io_u *io_u;
 267                 int i;
 268
 269                 io_u_qiter(&td->io_u_all, io_u, i) {
 270                         if (io_u->flags & IO_U_F_FLIGHT) {
 271                                 r = td->io_ops->cancel(td, io_u);
 272                                 if (!r)
 273                                         put_io_u(td, io_u);
 274                         }
 275                 }
 276         }
 277
 278         if (td->cur_depth)
 279                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 280 }
 281
 282 /*
 283  * Helper to handle the final sync of a file. Works just like the normal
 284  * io path, just does everything sync.
 285  */
 286 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 287 {
 288         struct io_u *io_u = __get_io_u(td);
 289         int ret;
 290
 291         if (!io_u)
 292                 return 1;
 293
 294         io_u->ddir = DDIR_SYNC;
 295         io_u->file = f;
 296
 297         if (td_io_prep(td, io_u)) {
 298                 put_io_u(td, io_u);
 299                 return 1;
 300         }
 301
 302 requeue:
 303         ret = td_io_queue(td, io_u);
 304         if (ret < 0) {
 305                 td_verror(td, io_u->error, "td_io_queue");
 306                 put_io_u(td, io_u);
 307                 return 1;
 308         } else if (ret == FIO_Q_QUEUED) {
 309                 if (io_u_queued_complete(td, 1, NULL) < 0)
 310                         return 1;
 311         } else if (ret == FIO_Q_COMPLETED) {
 312                 if (io_u->error) {
 313                         td_verror(td, io_u->error, "td_io_queue");
 314                         return 1;
 315                 }
 316
 317                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 318                         return 1;
 319         } else if (ret == FIO_Q_BUSY) {
 320                 if (td_io_commit(td))
 321                         return 1;
 322                 goto requeue;
 323         }
 324
 325         return 0;
 326 }
 327
 328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 329 {
 330         int ret;
 331
 332         if (fio_file_open(f))
 333                 return fio_io_sync(td, f);
 334
 335         if (td_io_open_file(td, f))
 336                 return 1;
 337
 338         ret = fio_io_sync(td, f);
 339         td_io_close_file(td, f);
 340         return ret;
 341 }
 342
 343 static inline void __update_tv_cache(struct thread_data *td)
 344 {
 345         fio_gettime(&td->tv_cache, NULL);
 346 }
 347
 348 static inline void update_tv_cache(struct thread_data *td)
 349 {
 350         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 351                 __update_tv_cache(td);
 352 }
 353
 354 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 355 {
 356         if (in_ramp_time(td))
 357                 return 0;
 358         if (!td->o.timeout)
 359                 return 0;
 360         if (utime_since(&td->epoch, t) >= td->o.timeout)
 361                 return 1;
 362
 363         return 0;
 364 }
 365
 366 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 367                                int *retptr)
 368 {
 369         int ret = *retptr;
 370
 371         if (ret < 0 || td->error) {
 372                 int err = td->error;
 373                 enum error_type_bit eb;
 374
 375                 if (ret < 0)
 376                         err = -ret;
 377
 378                 eb = td_error_type(ddir, err);
 379                 if (!(td->o.continue_on_error & (1 << eb)))
 380                         return 1;
 381
 382                 if (td_non_fatal_error(td, eb, err)) {
 383                         /*
 384                          * Continue with the I/Os in case of
 385                          * a non fatal error.
 386                          */
 387                         update_error_count(td, err);
 388                         td_clear_error(td);
 389                         *retptr = 0;
 390                         return 0;
 391                 } else if (td->o.fill_device && err == ENOSPC) {
 392                         /*
 393                          * We expect to hit this error if
 394                          * fill_device option is set.
 395                          */
 396                         td_clear_error(td);
 397                         fio_mark_td_terminate(td);
 398                         return 1;
 399                 } else {
 400                         /*
 401                          * Stop the I/O in case of a fatal
 402                          * error.
 403                          */
 404                         update_error_count(td, err);
 405                         return 1;
 406                 }
 407         }
 408
 409         return 0;
 410 }
 411
 412 static void check_update_rusage(struct thread_data *td)
 413 {
 414         if (td->update_rusage) {
 415                 td->update_rusage = 0;
 416                 update_rusage_stat(td);
 417                 fio_mutex_up(td->rusage_sem);
 418         }
 419 }
 420
 421 /*
 422  * The main verify engine. Runs over the writes we previously submitted,
 423  * reads the blocks back in, and checks the crc/md5 of the data.
 424  */
 425 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 426 {
 427         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 428         struct fio_file *f;
 429         struct io_u *io_u;
 430         int ret, min_events;
 431         unsigned int i;
 432
 433         dprint(FD_VERIFY, "starting loop\n");
 434
 435         /*
 436          * sync io first and invalidate cache, to make sure we really
 437          * read from disk.
 438          */
 439         for_each_file(td, f, i) {
 440                 if (!fio_file_open(f))
 441                         continue;
 442                 if (fio_io_sync(td, f))
 443                         break;
 444                 if (file_invalidate_cache(td, f))
 445                         break;
 446         }
 447
 448         check_update_rusage(td);
 449
 450         if (td->error)
 451                 return;
 452
 453         td_set_runstate(td, TD_VERIFYING);
 454
 455         io_u = NULL;
 456         while (!td->terminate) {
 457                 enum fio_ddir ddir;
 458                 int ret2, full;
 459
 460                 update_tv_cache(td);
 461                 check_update_rusage(td);
 462
 463                 if (runtime_exceeded(td, &td->tv_cache)) {
 464                         __update_tv_cache(td);
 465                         if (runtime_exceeded(td, &td->tv_cache)) {
 466                                 fio_mark_td_terminate(td);
 467                                 break;
 468                         }
 469                 }
 470
 471                 if (flow_threshold_exceeded(td))
 472                         continue;
 473
 474                 if (!td->o.experimental_verify) {
 475                         io_u = __get_io_u(td);
 476                         if (!io_u)
 477                                 break;
 478
 479                         if (get_next_verify(td, io_u)) {
 480                                 put_io_u(td, io_u);
 481                                 break;
 482                         }
 483
 484                         if (td_io_prep(td, io_u)) {
 485                                 put_io_u(td, io_u);
 486                                 break;
 487                         }
 488                 } else {
 489                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 490                                 break;
 491
 492                         while ((io_u = get_io_u(td)) != NULL) {
 493                                 if (IS_ERR(io_u)) {
 494                                         io_u = NULL;
 495                                         ret = FIO_Q_BUSY;
 496                                         goto reap;
 497                                 }
 498
 499                                 /*
 500                                  * We are only interested in the places where
 501                                  * we wrote or trimmed IOs. Turn those into
 502                                  * reads for verification purposes.
 503                                  */
 504                                 if (io_u->ddir == DDIR_READ) {
 505                                         /*
 506                                          * Pretend we issued it for rwmix
 507                                          * accounting
 508                                          */
 509                                         td->io_issues[DDIR_READ]++;
 510                                         put_io_u(td, io_u);
 511                                         continue;
 512                                 } else if (io_u->ddir == DDIR_TRIM) {
 513                                         io_u->ddir = DDIR_READ;
 514                                         io_u->flags |= IO_U_F_TRIMMED;
 515                                         break;
 516                                 } else if (io_u->ddir == DDIR_WRITE) {
 517                                         io_u->ddir = DDIR_READ;
 518                                         break;
 519                                 } else {
 520                                         put_io_u(td, io_u);
 521                                         continue;
 522                                 }
 523                         }
 524
 525                         if (!io_u)
 526                                 break;
 527                 }
 528
 529                 if (td->o.verify_async)
 530                         io_u->end_io = verify_io_u_async;
 531                 else
 532                         io_u->end_io = verify_io_u;
 533
 534                 ddir = io_u->ddir;
 535
 536                 ret = td_io_queue(td, io_u);
 537                 switch (ret) {
 538                 case FIO_Q_COMPLETED:
 539                         if (io_u->error) {
 540                                 ret = -io_u->error;
 541                                 clear_io_u(td, io_u);
 542                         } else if (io_u->resid) {
 543                                 int bytes = io_u->xfer_buflen - io_u->resid;
 544
 545                                 /*
 546                                  * zero read, fail
 547                                  */
 548                                 if (!bytes) {
 549                                         td_verror(td, EIO, "full resid");
 550                                         put_io_u(td, io_u);
 551                                         break;
 552                                 }
 553
 554                                 io_u->xfer_buflen = io_u->resid;
 555                                 io_u->xfer_buf += bytes;
 556                                 io_u->offset += bytes;
 557
 558                                 if (ddir_rw(io_u->ddir))
 559                                         td->ts.short_io_u[io_u->ddir]++;
 560
 561                                 f = io_u->file;
 562                                 if (io_u->offset == f->real_file_size)
 563                                         goto sync_done;
 564
 565                                 requeue_io_u(td, &io_u);
 566                         } else {
 567 sync_done:
 568                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 569                                 if (ret < 0)
 570                                         break;
 571                         }
 572                         continue;
 573                 case FIO_Q_QUEUED:
 574                         break;
 575                 case FIO_Q_BUSY:
 576                         requeue_io_u(td, &io_u);
 577                         ret2 = td_io_commit(td);
 578                         if (ret2 < 0)
 579                                 ret = ret2;
 580                         break;
 581                 default:
 582                         assert(ret < 0);
 583                         td_verror(td, -ret, "td_io_queue");
 584                         break;
 585                 }
 586
 587                 if (break_on_this_error(td, ddir, &ret))
 588                         break;
 589
 590                 /*
 591                  * if we can queue more, do so. but check if there are
 592                  * completed io_u's first. Note that we can get BUSY even
 593                  * without IO queued, if the system is resource starved.
 594                  */
 595 reap:
 596                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 597                 if (full || !td->o.iodepth_batch_complete) {
 598                         min_events = min(td->o.iodepth_batch_complete,
 599                                          td->cur_depth);
 600                         /*
 601                          * if the queue is full, we MUST reap at least 1 event
 602                          */
 603                         if (full && !min_events)
 604                                 min_events = 1;
 605
 606                         do {
 607                                 /*
 608                                  * Reap required number of io units, if any,
 609                                  * and do the verification on them through
 610                                  * the callback handler
 611                                  */
 612                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 613                                         ret = -1;
 614                                         break;
 615                                 }
 616                         } while (full && (td->cur_depth > td->o.iodepth_low));
 617                 }
 618                 if (ret < 0)
 619                         break;
 620         }
 621
 622         check_update_rusage(td);
 623
 624         if (!td->error) {
 625                 min_events = td->cur_depth;
 626
 627                 if (min_events)
 628                         ret = io_u_queued_complete(td, min_events, NULL);
 629         } else
 630                 cleanup_pending_aio(td);
 631
 632         td_set_runstate(td, TD_RUNNING);
 633
 634         dprint(FD_VERIFY, "exiting loop\n");
 635 }
 636
 637 static unsigned int exceeds_number_ios(struct thread_data *td)
 638 {
 639         unsigned long long number_ios;
 640
 641         if (!td->o.number_ios)
 642                 return 0;
 643
 644         number_ios = ddir_rw_sum(td->this_io_blocks);
 645         number_ios += td->io_u_queued + td->io_u_in_flight;
 646
 647         return number_ios >= td->o.number_ios;
 648 }
 649
 650 static int io_bytes_exceeded(struct thread_data *td)
 651 {
 652         unsigned long long bytes, limit;
 653
 654         if (td_rw(td))
 655                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 656         else if (td_write(td))
 657                 bytes = td->this_io_bytes[DDIR_WRITE];
 658         else if (td_read(td))
 659                 bytes = td->this_io_bytes[DDIR_READ];
 660         else
 661                 bytes = td->this_io_bytes[DDIR_TRIM];
 662
 663         if (td->o.io_limit)
 664                 limit = td->o.io_limit;
 665         else
 666                 limit = td->o.size;
 667
 668         return bytes >= limit || exceeds_number_ios(td);
 669 }
 670
 671 /*
 672  * Main IO worker function. It retrieves io_u's to process and queues
 673  * and reaps them, checking for rate and errors along the way.
 674  *
 675  * Returns number of bytes written and trimmed.
 676  */
 677 static uint64_t do_io(struct thread_data *td)
 678 {
 679         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 680         unsigned int i;
 681         int ret = 0;
 682         uint64_t total_bytes, bytes_issued = 0;
 683
 684         if (in_ramp_time(td))
 685                 td_set_runstate(td, TD_RAMP);
 686         else
 687                 td_set_runstate(td, TD_RUNNING);
 688
 689         lat_target_init(td);
 690
 691         /*
 692          * If verify_backlog is enabled, we'll run the verify in this
 693          * handler as well. For that case, we may need up to twice the
 694          * amount of bytes.
 695          */
 696         total_bytes = td->o.size;
 697         if (td->o.verify != VERIFY_NONE &&
 698            (td_write(td) && td->o.verify_backlog))
 699                 total_bytes += td->o.size;
 700
 701         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 702                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 703                 td->o.time_based) {
 704                 struct timeval comp_time;
 705                 int min_evts = 0;
 706                 struct io_u *io_u;
 707                 int ret2, full;
 708                 enum fio_ddir ddir;
 709
 710                 check_update_rusage(td);
 711
 712                 if (td->terminate || td->done)
 713                         break;
 714
 715                 update_tv_cache(td);
 716
 717                 if (runtime_exceeded(td, &td->tv_cache)) {
 718                         __update_tv_cache(td);
 719                         if (runtime_exceeded(td, &td->tv_cache)) {
 720                                 fio_mark_td_terminate(td);
 721                                 break;
 722                         }
 723                 }
 724
 725                 if (flow_threshold_exceeded(td))
 726                         continue;
 727
 728                 if (bytes_issued >= total_bytes)
 729                         break;
 730
 731                 io_u = get_io_u(td);
 732                 if (IS_ERR_OR_NULL(io_u)) {
 733                         int err = PTR_ERR(io_u);
 734
 735                         io_u = NULL;
 736                         if (err == -EBUSY) {
 737                                 ret = FIO_Q_BUSY;
 738                                 goto reap;
 739                         }
 740                         if (td->o.latency_target)
 741                                 goto reap;
 742                         break;
 743                 }
 744
 745                 ddir = io_u->ddir;
 746
 747                 /*
 748                  * Add verification end_io handler if:
 749                  *      - Asked to verify (!td_rw(td))
 750                  *      - Or the io_u is from our verify list (mixed write/ver)
 751                  */
 752                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 753                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 754
 755                         if (!td->o.verify_pattern_bytes) {
 756                                 io_u->rand_seed = __rand(&td->verify_state);
 757                                 if (sizeof(int) != sizeof(long *))
 758                                         io_u->rand_seed *= __rand(&td->verify_state);
 759                         }
 760
 761                         if (td->o.verify_async)
 762                                 io_u->end_io = verify_io_u_async;
 763                         else
 764                                 io_u->end_io = verify_io_u;
 765                         td_set_runstate(td, TD_VERIFYING);
 766                 } else if (in_ramp_time(td))
 767                         td_set_runstate(td, TD_RAMP);
 768                 else
 769                         td_set_runstate(td, TD_RUNNING);
 770
 771                 /*
 772                  * Always log IO before it's issued, so we know the specific
 773                  * order of it. The logged unit will track when the IO has
 774                  * completed.
 775                  */
 776                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 777                     td->o.do_verify &&
 778                     td->o.verify != VERIFY_NONE &&
 779                     !td->o.experimental_verify)
 780                         log_io_piece(td, io_u);
 781
 782                 ret = td_io_queue(td, io_u);
 783                 switch (ret) {
 784                 case FIO_Q_COMPLETED:
 785                         if (io_u->error) {
 786                                 ret = -io_u->error;
 787                                 unlog_io_piece(td, io_u);
 788                                 clear_io_u(td, io_u);
 789                         } else if (io_u->resid) {
 790                                 int bytes = io_u->xfer_buflen - io_u->resid;
 791                                 struct fio_file *f = io_u->file;
 792
 793                                 bytes_issued += bytes;
 794
 795                                 trim_io_piece(td, io_u);
 796
 797                                 /*
 798                                  * zero read, fail
 799                                  */
 800                                 if (!bytes) {
 801                                         unlog_io_piece(td, io_u);
 802                                         td_verror(td, EIO, "full resid");
 803                                         put_io_u(td, io_u);
 804                                         break;
 805                                 }
 806
 807                                 io_u->xfer_buflen = io_u->resid;
 808                                 io_u->xfer_buf += bytes;
 809                                 io_u->offset += bytes;
 810
 811                                 if (ddir_rw(io_u->ddir))
 812                                         td->ts.short_io_u[io_u->ddir]++;
 813
 814                                 if (io_u->offset == f->real_file_size)
 815                                         goto sync_done;
 816
 817                                 requeue_io_u(td, &io_u);
 818                         } else {
 819 sync_done:
 820                                 if (__should_check_rate(td, DDIR_READ) ||
 821                                     __should_check_rate(td, DDIR_WRITE) ||
 822                                     __should_check_rate(td, DDIR_TRIM))
 823                                         fio_gettime(&comp_time, NULL);
 824
 825                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 826                                 if (ret < 0)
 827                                         break;
 828                                 bytes_issued += io_u->xfer_buflen;
 829                         }
 830                         break;
 831                 case FIO_Q_QUEUED:
 832                         /*
 833                          * if the engine doesn't have a commit hook,
 834                          * the io_u is really queued. if it does have such
 835                          * a hook, it has to call io_u_queued() itself.
 836                          */
 837                         if (td->io_ops->commit == NULL)
 838                                 io_u_queued(td, io_u);
 839                         bytes_issued += io_u->xfer_buflen;
 840                         break;
 841                 case FIO_Q_BUSY:
 842                         unlog_io_piece(td, io_u);
 843                         requeue_io_u(td, &io_u);
 844                         ret2 = td_io_commit(td);
 845                         if (ret2 < 0)
 846                                 ret = ret2;
 847                         break;
 848                 default:
 849                         assert(ret < 0);
 850                         put_io_u(td, io_u);
 851                         break;
 852                 }
 853
 854                 if (break_on_this_error(td, ddir, &ret))
 855                         break;
 856
 857                 /*
 858                  * See if we need to complete some commands. Note that we
 859                  * can get BUSY even without IO queued, if the system is
 860                  * resource starved.
 861                  */
 862 reap:
 863                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 864                 if (full || !td->o.iodepth_batch_complete) {
 865                         min_evts = min(td->o.iodepth_batch_complete,
 866                                         td->cur_depth);
 867                         /*
 868                          * if the queue is full, we MUST reap at least 1 event
 869                          */
 870                         if (full && !min_evts)
 871                                 min_evts = 1;
 872
 873                         if (__should_check_rate(td, DDIR_READ) ||
 874                             __should_check_rate(td, DDIR_WRITE) ||
 875                             __should_check_rate(td, DDIR_TRIM))
 876                                 fio_gettime(&comp_time, NULL);
 877
 878                         do {
 879                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 880                                 if (ret < 0)
 881                                         break;
 882
 883                         } while (full && (td->cur_depth > td->o.iodepth_low));
 884                 }
 885
 886                 if (ret < 0)
 887                         break;
 888                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 889                         continue;
 890
 891                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 892                         if (check_min_rate(td, &comp_time, bytes_done)) {
 893                                 if (exitall_on_terminate)
 894                                         fio_terminate_threads(td->groupid);
 895                                 td_verror(td, EIO, "check_min_rate");
 896                                 break;
 897                         }
 898                 }
 899                 if (!in_ramp_time(td) && td->o.latency_target)
 900                         lat_target_check(td);
 901
 902                 if (td->o.thinktime) {
 903                         unsigned long long b;
 904
 905                         b = ddir_rw_sum(td->io_blocks);
 906                         if (!(b % td->o.thinktime_blocks)) {
 907                                 int left;
 908
 909                                 io_u_quiesce(td);
 910
 911                                 if (td->o.thinktime_spin)
 912                                         usec_spin(td->o.thinktime_spin);
 913
 914                                 left = td->o.thinktime - td->o.thinktime_spin;
 915                                 if (left)
 916                                         usec_sleep(td, left);
 917                         }
 918                 }
 919         }
 920
 921         check_update_rusage(td);
 922
 923         if (td->trim_entries)
 924                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 925
 926         if (td->o.fill_device && td->error == ENOSPC) {
 927                 td->error = 0;
 928                 fio_mark_td_terminate(td);
 929         }
 930         if (!td->error) {
 931                 struct fio_file *f;
 932
 933                 i = td->cur_depth;
 934                 if (i) {
 935                         ret = io_u_queued_complete(td, i, bytes_done);
 936                         if (td->o.fill_device && td->error == ENOSPC)
 937                                 td->error = 0;
 938                 }
 939
 940                 if (should_fsync(td) && td->o.end_fsync) {
 941                         td_set_runstate(td, TD_FSYNCING);
 942
 943                         for_each_file(td, f, i) {
 944                                 if (!fio_file_fsync(td, f))
 945                                         continue;
 946
 947                                 log_err("fio: end_fsync failed for file %s\n",
 948                                                                 f->file_name);
 949                         }
 950                 }
 951         } else
 952                 cleanup_pending_aio(td);
 953
 954         /*
 955          * stop job if we failed doing any IO
 956          */
 957         if (!ddir_rw_sum(td->this_io_bytes))
 958                 td->done = 1;
 959
 960         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 961 }
 962
 963 static void cleanup_io_u(struct thread_data *td)
 964 {
 965         struct io_u *io_u;
 966
 967         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 968
 969                 if (td->io_ops->io_u_free)
 970                         td->io_ops->io_u_free(td, io_u);
 971
 972                 fio_memfree(io_u, sizeof(*io_u));
 973         }
 974
 975         free_io_mem(td);
 976
 977         io_u_rexit(&td->io_u_requeues);
 978         io_u_qexit(&td->io_u_freelist);
 979         io_u_qexit(&td->io_u_all);
 980 }
 981
 982 static int init_io_u(struct thread_data *td)
 983 {
 984         struct io_u *io_u;
 985         unsigned int max_bs, min_write;
 986         int cl_align, i, max_units;
 987         int data_xfer = 1, err;
 988         char *p;
 989
 990         max_units = td->o.iodepth;
 991         max_bs = td_max_bs(td);
 992         min_write = td->o.min_bs[DDIR_WRITE];
 993         td->orig_buffer_size = (unsigned long long) max_bs
 994                                         * (unsigned long long) max_units;
 995
 996         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 997                 data_xfer = 0;
 998
 999         err = 0;
1000         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1001         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1002         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1003
1004         if (err) {
1005                 log_err("fio: failed setting up IO queues\n");
1006                 return 1;
1007         }
1008
1009         /*
1010          * if we may later need to do address alignment, then add any
1011          * possible adjustment here so that we don't cause a buffer
1012          * overflow later. this adjustment may be too much if we get
1013          * lucky and the allocator gives us an aligned address.
1014          */
1015         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1016             (td->io_ops->flags & FIO_RAWIO))
1017                 td->orig_buffer_size += page_mask + td->o.mem_align;
1018
1019         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1020                 unsigned long bs;
1021
1022                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1023                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1024         }
1025
1026         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1027                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1028                 return 1;
1029         }
1030
1031         if (data_xfer && allocate_io_mem(td))
1032                 return 1;
1033
1034         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1035             (td->io_ops->flags & FIO_RAWIO))
1036                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1037         else
1038                 p = td->orig_buffer;
1039
1040         cl_align = os_cache_line_size();
1041
1042         for (i = 0; i < max_units; i++) {
1043                 void *ptr;
1044
1045                 if (td->terminate)
1046                         return 1;
1047
1048                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1049                 if (!ptr) {
1050                         log_err("fio: unable to allocate aligned memory\n");
1051                         break;
1052                 }
1053
1054                 io_u = ptr;
1055                 memset(io_u, 0, sizeof(*io_u));
1056                 INIT_FLIST_HEAD(&io_u->verify_list);
1057                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1058
1059                 if (data_xfer) {
1060                         io_u->buf = p;
1061                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1062
1063                         if (td_write(td))
1064                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1065                         if (td_write(td) && td->o.verify_pattern_bytes) {
1066                                 /*
1067                                  * Fill the buffer with the pattern if we are
1068                                  * going to be doing writes.
1069                                  */
1070                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1071                         }
1072                 }
1073
1074                 io_u->index = i;
1075                 io_u->flags = IO_U_F_FREE;
1076                 io_u_qpush(&td->io_u_freelist, io_u);
1077
1078                 /*
1079                  * io_u never leaves this stack, used for iteration of all
1080                  * io_u buffers.
1081                  */
1082                 io_u_qpush(&td->io_u_all, io_u);
1083
1084                 if (td->io_ops->io_u_init) {
1085                         int ret = td->io_ops->io_u_init(td, io_u);
1086
1087                         if (ret) {
1088                                 log_err("fio: failed to init engine data: %d\n", ret);
1089                                 return 1;
1090                         }
1091                 }
1092
1093                 p += max_bs;
1094         }
1095
1096         return 0;
1097 }
1098
1099 static int switch_ioscheduler(struct thread_data *td)
1100 {
1101         char tmp[256], tmp2[128];
1102         FILE *f;
1103         int ret;
1104
1105         if (td->io_ops->flags & FIO_DISKLESSIO)
1106                 return 0;
1107
1108         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1109
1110         f = fopen(tmp, "r+");
1111         if (!f) {
1112                 if (errno == ENOENT) {
1113                         log_err("fio: os or kernel doesn't support IO scheduler"
1114                                 " switching\n");
1115                         return 0;
1116                 }
1117                 td_verror(td, errno, "fopen iosched");
1118                 return 1;
1119         }
1120
1121         /*
1122          * Set io scheduler.
1123          */
1124         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1125         if (ferror(f) || ret != 1) {
1126                 td_verror(td, errno, "fwrite");
1127                 fclose(f);
1128                 return 1;
1129         }
1130
1131         rewind(f);
1132
1133         /*
1134          * Read back and check that the selected scheduler is now the default.
1135          */
1136         ret = fread(tmp, sizeof(tmp), 1, f);
1137         if (ferror(f) || ret < 0) {
1138                 td_verror(td, errno, "fread");
1139                 fclose(f);
1140                 return 1;
1141         }
1142         tmp[sizeof(tmp) - 1] = '\0';
1143
1144
1145         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1146         if (!strstr(tmp, tmp2)) {
1147                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1148                 td_verror(td, EINVAL, "iosched_switch");
1149                 fclose(f);
1150                 return 1;
1151         }
1152
1153         fclose(f);
1154         return 0;
1155 }
1156
1157 static int keep_running(struct thread_data *td)
1158 {
1159         unsigned long long limit;
1160
1161         if (td->done)
1162                 return 0;
1163         if (td->o.time_based)
1164                 return 1;
1165         if (td->o.loops) {
1166                 td->o.loops--;
1167                 return 1;
1168         }
1169         if (exceeds_number_ios(td))
1170                 return 0;
1171
1172         if (td->o.io_limit)
1173                 limit = td->o.io_limit;
1174         else
1175                 limit = td->o.size;
1176
1177         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1178                 uint64_t diff;
1179
1180                 /*
1181                  * If the difference is less than the minimum IO size, we
1182                  * are done.
1183                  */
1184                 diff = limit - ddir_rw_sum(td->io_bytes);
1185                 if (diff < td_max_bs(td))
1186                         return 0;
1187
1188                 if (fio_files_done(td))
1189                         return 0;
1190
1191                 return 1;
1192         }
1193
1194         return 0;
1195 }
1196
1197 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1198 {
1199         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1200         char *str;
1201
1202         str = malloc(newlen);
1203         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1204
1205         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1206         ret = system(str);
1207         if (ret == -1)
1208                 log_err("fio: exec of cmd <%s> failed\n", str);
1209
1210         free(str);
1211         return ret;
1212 }
1213
1214 /*
1215  * Dry run to compute correct state of numberio for verification.
1216  */
1217 static uint64_t do_dry_run(struct thread_data *td)
1218 {
1219         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1220
1221         td_set_runstate(td, TD_RUNNING);
1222
1223         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1224                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1225                 struct io_u *io_u;
1226                 int ret;
1227
1228                 if (td->terminate || td->done)
1229                         break;
1230
1231                 io_u = get_io_u(td);
1232                 if (!io_u)
1233                         break;
1234
1235                 io_u->flags |= IO_U_F_FLIGHT;
1236                 io_u->error = 0;
1237                 io_u->resid = 0;
1238                 if (ddir_rw(acct_ddir(io_u)))
1239                         td->io_issues[acct_ddir(io_u)]++;
1240                 if (ddir_rw(io_u->ddir)) {
1241                         io_u_mark_depth(td, 1);
1242                         td->ts.total_io_u[io_u->ddir]++;
1243                 }
1244
1245                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1246                     td->o.do_verify &&
1247                     td->o.verify != VERIFY_NONE &&
1248                     !td->o.experimental_verify)
1249                         log_io_piece(td, io_u);
1250
1251                 ret = io_u_sync_complete(td, io_u, bytes_done);
1252                 (void) ret;
1253         }
1254
1255         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1256 }
1257
1258 /*
1259  * Entry point for the thread based jobs. The process based jobs end up
1260  * here as well, after a little setup.
1261  */
1262 static void *thread_main(void *data)
1263 {
1264         unsigned long long elapsed;
1265         struct thread_data *td = data;
1266         struct thread_options *o = &td->o;
1267         pthread_condattr_t attr;
1268         int clear_state;
1269         int ret;
1270
1271         if (!o->use_thread) {
1272                 setsid();
1273                 td->pid = getpid();
1274         } else
1275                 td->pid = gettid();
1276
1277         fio_local_clock_init(o->use_thread);
1278
1279         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1280
1281         if (is_backend)
1282                 fio_server_send_start(td);
1283
1284         INIT_FLIST_HEAD(&td->io_log_list);
1285         INIT_FLIST_HEAD(&td->io_hist_list);
1286         INIT_FLIST_HEAD(&td->verify_list);
1287         INIT_FLIST_HEAD(&td->trim_list);
1288         INIT_FLIST_HEAD(&td->next_rand_list);
1289         pthread_mutex_init(&td->io_u_lock, NULL);
1290         td->io_hist_tree = RB_ROOT;
1291
1292         pthread_condattr_init(&attr);
1293         pthread_cond_init(&td->verify_cond, &attr);
1294         pthread_cond_init(&td->free_cond, &attr);
1295
1296         td_set_runstate(td, TD_INITIALIZED);
1297         dprint(FD_MUTEX, "up startup_mutex\n");
1298         fio_mutex_up(startup_mutex);
1299         dprint(FD_MUTEX, "wait on td->mutex\n");
1300         fio_mutex_down(td->mutex);
1301         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1302
1303         /*
1304          * A new gid requires privilege, so we need to do this before setting
1305          * the uid.
1306          */
1307         if (o->gid != -1U && setgid(o->gid)) {
1308                 td_verror(td, errno, "setgid");
1309                 goto err;
1310         }
1311         if (o->uid != -1U && setuid(o->uid)) {
1312                 td_verror(td, errno, "setuid");
1313                 goto err;
1314         }
1315
1316         /*
1317          * If we have a gettimeofday() thread, make sure we exclude that
1318          * thread from this job
1319          */
1320         if (o->gtod_cpu)
1321                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1322
1323         /*
1324          * Set affinity first, in case it has an impact on the memory
1325          * allocations.
1326          */
1327         if (o->cpumask_set) {
1328                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1329                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1330                         if (!ret) {
1331                                 log_err("fio: no CPUs set\n");
1332                                 log_err("fio: Try increasing number of available CPUs\n");
1333                                 td_verror(td, EINVAL, "cpus_split");
1334                                 goto err;
1335                         }
1336                 }
1337                 ret = fio_setaffinity(td->pid, o->cpumask);
1338                 if (ret == -1) {
1339                         td_verror(td, errno, "cpu_set_affinity");
1340                         goto err;
1341                 }
1342         }
1343
1344 #ifdef CONFIG_LIBNUMA
1345         /* numa node setup */
1346         if (o->numa_cpumask_set || o->numa_memmask_set) {
1347                 struct bitmask *mask;
1348
1349                 if (numa_available() < 0) {
1350                         td_verror(td, errno, "Does not support NUMA API\n");
1351                         goto err;
1352                 }
1353
1354                 if (o->numa_cpumask_set) {
1355                         mask = numa_parse_nodestring(o->numa_cpunodes);
1356                         ret = numa_run_on_node_mask(mask);
1357                         numa_free_nodemask(mask);
1358                         if (ret == -1) {
1359                                 td_verror(td, errno, \
1360                                         "numa_run_on_node_mask failed\n");
1361                                 goto err;
1362                         }
1363                 }
1364
1365                 if (o->numa_memmask_set) {
1366
1367                         mask = NULL;
1368                         if (o->numa_memnodes)
1369                                 mask = numa_parse_nodestring(o->numa_memnodes);
1370
1371                         switch (o->numa_mem_mode) {
1372                         case MPOL_INTERLEAVE:
1373                                 numa_set_interleave_mask(mask);
1374                                 break;
1375                         case MPOL_BIND:
1376                                 numa_set_membind(mask);
1377                                 break;
1378                         case MPOL_LOCAL:
1379                                 numa_set_localalloc();
1380                                 break;
1381                         case MPOL_PREFERRED:
1382                                 numa_set_preferred(o->numa_mem_prefer_node);
1383                                 break;
1384                         case MPOL_DEFAULT:
1385                         default:
1386                                 break;
1387                         }
1388
1389                         if (mask)
1390                                 numa_free_nodemask(mask);
1391
1392                 }
1393         }
1394 #endif
1395
1396         if (fio_pin_memory(td))
1397                 goto err;
1398
1399         /*
1400          * May alter parameters that init_io_u() will use, so we need to
1401          * do this first.
1402          */
1403         if (init_iolog(td))
1404                 goto err;
1405
1406         if (init_io_u(td))
1407                 goto err;
1408
1409         if (o->verify_async && verify_async_init(td))
1410                 goto err;
1411
1412         if (o->ioprio) {
1413                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1414                 if (ret == -1) {
1415                         td_verror(td, errno, "ioprio_set");
1416                         goto err;
1417                 }
1418         }
1419
1420         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1421                 goto err;
1422
1423         errno = 0;
1424         if (nice(o->nice) == -1 && errno != 0) {
1425                 td_verror(td, errno, "nice");
1426                 goto err;
1427         }
1428
1429         if (o->ioscheduler && switch_ioscheduler(td))
1430                 goto err;
1431
1432         if (!o->create_serialize && setup_files(td))
1433                 goto err;
1434
1435         if (td_io_init(td))
1436                 goto err;
1437
1438         if (init_random_map(td))
1439                 goto err;
1440
1441         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1442                 goto err;
1443
1444         if (o->pre_read) {
1445                 if (pre_read_files(td) < 0)
1446                         goto err;
1447         }
1448
1449         if (td->flags & TD_F_COMPRESS_LOG)
1450                 tp_init(&td->tp_data);
1451
1452         fio_verify_init(td);
1453
1454         fio_gettime(&td->epoch, NULL);
1455         fio_getrusage(&td->ru_start);
1456         clear_state = 0;
1457         while (keep_running(td)) {
1458                 uint64_t verify_bytes;
1459
1460                 fio_gettime(&td->start, NULL);
1461                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1462                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1463                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1464
1465                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1466                                 o->ratemin[DDIR_TRIM]) {
1467                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1468                                                 sizeof(td->bw_sample_time));
1469                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1470                                                 sizeof(td->bw_sample_time));
1471                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1472                                                 sizeof(td->bw_sample_time));
1473                 }
1474
1475                 if (clear_state)
1476                         clear_io_state(td);
1477
1478                 prune_io_piece_log(td);
1479
1480                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1481                         verify_bytes = do_dry_run(td);
1482                 else
1483                         verify_bytes = do_io(td);
1484
1485                 clear_state = 1;
1486
1487                 fio_mutex_down(stat_mutex);
1488                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1489                         elapsed = mtime_since_now(&td->start);
1490                         td->ts.runtime[DDIR_READ] += elapsed;
1491                 }
1492                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1493                         elapsed = mtime_since_now(&td->start);
1494                         td->ts.runtime[DDIR_WRITE] += elapsed;
1495                 }
1496                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1497                         elapsed = mtime_since_now(&td->start);
1498                         td->ts.runtime[DDIR_TRIM] += elapsed;
1499                 }
1500                 fio_gettime(&td->start, NULL);
1501                 fio_mutex_up(stat_mutex);
1502
1503                 if (td->error || td->terminate)
1504                         break;
1505
1506                 if (!o->do_verify ||
1507                     o->verify == VERIFY_NONE ||
1508                     (td->io_ops->flags & FIO_UNIDIR))
1509                         continue;
1510
1511                 clear_io_state(td);
1512
1513                 fio_gettime(&td->start, NULL);
1514
1515                 do_verify(td, verify_bytes);
1516
1517                 fio_mutex_down(stat_mutex);
1518                 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1519                 fio_gettime(&td->start, NULL);
1520                 fio_mutex_up(stat_mutex);
1521
1522                 if (td->error || td->terminate)
1523                         break;
1524         }
1525
1526         update_rusage_stat(td);
1527         td->ts.total_run_time = mtime_since_now(&td->epoch);
1528         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1529         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1530         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1531
1532         fio_unpin_memory(td);
1533
1534         fio_writeout_logs(td);
1535
1536         if (td->flags & TD_F_COMPRESS_LOG)
1537                 tp_exit(&td->tp_data);
1538
1539         if (o->exec_postrun)
1540                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1541
1542         if (exitall_on_terminate)
1543                 fio_terminate_threads(td->groupid);
1544
1545 err:
1546         if (td->error)
1547                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1548                                                         td->verror);
1549
1550         if (o->verify_async)
1551                 verify_async_exit(td);
1552
1553         close_and_free_files(td);
1554         cleanup_io_u(td);
1555         close_ioengine(td);
1556         cgroup_shutdown(td, &cgroup_mnt);
1557
1558         if (o->cpumask_set) {
1559                 ret = fio_cpuset_exit(&o->cpumask);
1560                 if (ret)
1561                         td_verror(td, ret, "fio_cpuset_exit");
1562         }
1563
1564         /*
1565          * do this very late, it will log file closing as well
1566          */
1567         if (o->write_iolog_file)
1568                 write_iolog_close(td);
1569
1570         fio_mutex_remove(td->mutex);
1571         td->mutex = NULL;
1572
1573         td_set_runstate(td, TD_EXITED);
1574
1575         /*
1576          * Do this last after setting our runstate to exited, so we
1577          * know that the stat thread is signaled.
1578          */
1579         check_update_rusage(td);
1580
1581         return (void *) (uintptr_t) td->error;
1582 }
1583
1584
1585 /*
1586  * We cannot pass the td data into a forked process, so attach the td and
1587  * pass it to the thread worker.
1588  */
1589 static int fork_main(int shmid, int offset)
1590 {
1591         struct thread_data *td;
1592         void *data, *ret;
1593
1594 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1595         data = shmat(shmid, NULL, 0);
1596         if (data == (void *) -1) {
1597                 int __err = errno;
1598
1599                 perror("shmat");
1600                 return __err;
1601         }
1602 #else
1603         /*
1604          * HP-UX inherits shm mappings?
1605          */
1606         data = threads;
1607 #endif
1608
1609         td = data + offset * sizeof(struct thread_data);
1610         ret = thread_main(td);
1611         shmdt(data);
1612         return (int) (uintptr_t) ret;
1613 }
1614
1615 static void dump_td_info(struct thread_data *td)
1616 {
1617         log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1618                 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1619                         (unsigned long) time_since_now(&td->terminate_time));
1620 }
1621
1622 /*
1623  * Run over the job map and reap the threads that have exited, if any.
1624  */
1625 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1626                          unsigned int *m_rate)
1627 {
1628         struct thread_data *td;
1629         unsigned int cputhreads, realthreads, pending;
1630         int i, status, ret;
1631
1632         /*
1633          * reap exited threads (TD_EXITED -> TD_REAPED)
1634          */
1635         realthreads = pending = cputhreads = 0;
1636         for_each_td(td, i) {
1637                 int flags = 0;
1638
1639                 /*
1640                  * ->io_ops is NULL for a thread that has closed its
1641                  * io engine
1642                  */
1643                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1644                         cputhreads++;
1645                 else
1646                         realthreads++;
1647
1648                 if (!td->pid) {
1649                         pending++;
1650                         continue;
1651                 }
1652                 if (td->runstate == TD_REAPED)
1653                         continue;
1654                 if (td->o.use_thread) {
1655                         if (td->runstate == TD_EXITED) {
1656                                 td_set_runstate(td, TD_REAPED);
1657                                 goto reaped;
1658                         }
1659                         continue;
1660                 }
1661
1662                 flags = WNOHANG;
1663                 if (td->runstate == TD_EXITED)
1664                         flags = 0;
1665
1666                 /*
1667                  * check if someone quit or got killed in an unusual way
1668                  */
1669                 ret = waitpid(td->pid, &status, flags);
1670                 if (ret < 0) {
1671                         if (errno == ECHILD) {
1672                                 log_err("fio: pid=%d disappeared %d\n",
1673                                                 (int) td->pid, td->runstate);
1674                                 td->sig = ECHILD;
1675                                 td_set_runstate(td, TD_REAPED);
1676                                 goto reaped;
1677                         }
1678                         perror("waitpid");
1679                 } else if (ret == td->pid) {
1680                         if (WIFSIGNALED(status)) {
1681                                 int sig = WTERMSIG(status);
1682
1683                                 if (sig != SIGTERM && sig != SIGUSR2)
1684                                         log_err("fio: pid=%d, got signal=%d\n",
1685                                                         (int) td->pid, sig);
1686                                 td->sig = sig;
1687                                 td_set_runstate(td, TD_REAPED);
1688                                 goto reaped;
1689                         }
1690                         if (WIFEXITED(status)) {
1691                                 if (WEXITSTATUS(status) && !td->error)
1692                                         td->error = WEXITSTATUS(status);
1693
1694                                 td_set_runstate(td, TD_REAPED);
1695                                 goto reaped;
1696                         }
1697                 }
1698
1699                 /*
1700                  * If the job is stuck, do a forceful timeout of it and
1701                  * move on.
1702                  */
1703                 if (td->terminate &&
1704                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1705                         dump_td_info(td);
1706                         td_set_runstate(td, TD_REAPED);
1707                         goto reaped;
1708                 }
1709
1710                 /*
1711                  * thread is not dead, continue
1712                  */
1713                 pending++;
1714                 continue;
1715 reaped:
1716                 (*nr_running)--;
1717                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1718                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1719                 if (!td->pid)
1720                         pending--;
1721
1722                 if (td->error)
1723                         exit_value++;
1724
1725                 done_secs += mtime_since_now(&td->epoch) / 1000;
1726                 profile_td_exit(td);
1727         }
1728
1729         if (*nr_running == cputhreads && !pending && realthreads)
1730                 fio_terminate_threads(TERMINATE_ALL);
1731 }
1732
1733 static void do_usleep(unsigned int usecs)
1734 {
1735         check_for_running_stats();
1736         usleep(usecs);
1737 }
1738
1739 /*
1740  * Main function for kicking off and reaping jobs, as needed.
1741  */
1742 static void run_threads(void)
1743 {
1744         struct thread_data *td;
1745         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1746         uint64_t spent;
1747
1748         if (fio_gtod_offload && fio_start_gtod_thread())
1749                 return;
1750
1751         fio_idle_prof_init();
1752
1753         set_sig_handlers();
1754
1755         nr_thread = nr_process = 0;
1756         for_each_td(td, i) {
1757                 if (td->o.use_thread)
1758                         nr_thread++;
1759                 else
1760                         nr_process++;
1761         }
1762
1763         if (output_format == FIO_OUTPUT_NORMAL) {
1764                 log_info("Starting ");
1765                 if (nr_thread)
1766                         log_info("%d thread%s", nr_thread,
1767                                                 nr_thread > 1 ? "s" : "");
1768                 if (nr_process) {
1769                         if (nr_thread)
1770                                 log_info(" and ");
1771                         log_info("%d process%s", nr_process,
1772                                                 nr_process > 1 ? "es" : "");
1773                 }
1774                 log_info("\n");
1775                 log_info_flush();
1776         }
1777
1778         todo = thread_number;
1779         nr_running = 0;
1780         nr_started = 0;
1781         m_rate = t_rate = 0;
1782
1783         for_each_td(td, i) {
1784                 print_status_init(td->thread_number - 1);
1785
1786                 if (!td->o.create_serialize)
1787                         continue;
1788
1789                 /*
1790                  * do file setup here so it happens sequentially,
1791                  * we don't want X number of threads getting their
1792                  * client data interspersed on disk
1793                  */
1794                 if (setup_files(td)) {
1795                         exit_value++;
1796                         if (td->error)
1797                                 log_err("fio: pid=%d, err=%d/%s\n",
1798                                         (int) td->pid, td->error, td->verror);
1799                         td_set_runstate(td, TD_REAPED);
1800                         todo--;
1801                 } else {
1802                         struct fio_file *f;
1803                         unsigned int j;
1804
1805                         /*
1806                          * for sharing to work, each job must always open
1807                          * its own files. so close them, if we opened them
1808                          * for creation
1809                          */
1810                         for_each_file(td, f, j) {
1811                                 if (fio_file_open(f))
1812                                         td_io_close_file(td, f);
1813                         }
1814                 }
1815         }
1816
1817         /* start idle threads before io threads start to run */
1818         fio_idle_prof_start();
1819
1820         set_genesis_time();
1821
1822         while (todo) {
1823                 struct thread_data *map[REAL_MAX_JOBS];
1824                 struct timeval this_start;
1825                 int this_jobs = 0, left;
1826
1827                 /*
1828                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1829                  */
1830                 for_each_td(td, i) {
1831                         if (td->runstate != TD_NOT_CREATED)
1832                                 continue;
1833
1834                         /*
1835                          * never got a chance to start, killed by other
1836                          * thread for some reason
1837                          */
1838                         if (td->terminate) {
1839                                 todo--;
1840                                 continue;
1841                         }
1842
1843                         if (td->o.start_delay) {
1844                                 spent = utime_since_genesis();
1845
1846                                 if (td->o.start_delay > spent)
1847                                         continue;
1848                         }
1849
1850                         if (td->o.stonewall && (nr_started || nr_running)) {
1851                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1852                                                         td->o.name);
1853                                 break;
1854                         }
1855
1856                         init_disk_util(td);
1857
1858                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1859                         td->update_rusage = 0;
1860
1861                         /*
1862                          * Set state to created. Thread will transition
1863                          * to TD_INITIALIZED when it's done setting up.
1864                          */
1865                         td_set_runstate(td, TD_CREATED);
1866                         map[this_jobs++] = td;
1867                         nr_started++;
1868
1869                         if (td->o.use_thread) {
1870                                 int ret;
1871
1872                                 dprint(FD_PROCESS, "will pthread_create\n");
1873                                 ret = pthread_create(&td->thread, NULL,
1874                                                         thread_main, td);
1875                                 if (ret) {
1876                                         log_err("pthread_create: %s\n",
1877                                                         strerror(ret));
1878                                         nr_started--;
1879                                         break;
1880                                 }
1881                                 ret = pthread_detach(td->thread);
1882                                 if (ret)
1883                                         log_err("pthread_detach: %s",
1884                                                         strerror(ret));
1885                         } else {
1886                                 pid_t pid;
1887                                 dprint(FD_PROCESS, "will fork\n");
1888                                 pid = fork();
1889                                 if (!pid) {
1890                                         int ret = fork_main(shm_id, i);
1891
1892                                         _exit(ret);
1893                                 } else if (i == fio_debug_jobno)
1894                                         *fio_debug_jobp = pid;
1895                         }
1896                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1897                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1898                                 log_err("fio: job startup hung? exiting.\n");
1899                                 fio_terminate_threads(TERMINATE_ALL);
1900                                 fio_abort = 1;
1901                                 nr_started--;
1902                                 break;
1903                         }
1904                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1905                 }
1906
1907                 /*
1908                  * Wait for the started threads to transition to
1909                  * TD_INITIALIZED.
1910                  */
1911                 fio_gettime(&this_start, NULL);
1912                 left = this_jobs;
1913                 while (left && !fio_abort) {
1914                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1915                                 break;
1916
1917                         do_usleep(100000);
1918
1919                         for (i = 0; i < this_jobs; i++) {
1920                                 td = map[i];
1921                                 if (!td)
1922                                         continue;
1923                                 if (td->runstate == TD_INITIALIZED) {
1924                                         map[i] = NULL;
1925                                         left--;
1926                                 } else if (td->runstate >= TD_EXITED) {
1927                                         map[i] = NULL;
1928                                         left--;
1929                                         todo--;
1930                                         nr_running++; /* work-around... */
1931                                 }
1932                         }
1933                 }
1934
1935                 if (left) {
1936                         log_err("fio: %d job%s failed to start\n", left,
1937                                         left > 1 ? "s" : "");
1938                         for (i = 0; i < this_jobs; i++) {
1939                                 td = map[i];
1940                                 if (!td)
1941                                         continue;
1942                                 kill(td->pid, SIGTERM);
1943                         }
1944                         break;
1945                 }
1946
1947                 /*
1948                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1949                  */
1950                 for_each_td(td, i) {
1951                         if (td->runstate != TD_INITIALIZED)
1952                                 continue;
1953
1954                         if (in_ramp_time(td))
1955                                 td_set_runstate(td, TD_RAMP);
1956                         else
1957                                 td_set_runstate(td, TD_RUNNING);
1958                         nr_running++;
1959                         nr_started--;
1960                         m_rate += ddir_rw_sum(td->o.ratemin);
1961                         t_rate += ddir_rw_sum(td->o.rate);
1962                         todo--;
1963                         fio_mutex_up(td->mutex);
1964                 }
1965
1966                 reap_threads(&nr_running, &t_rate, &m_rate);
1967
1968                 if (todo)
1969                         do_usleep(100000);
1970         }
1971
1972         while (nr_running) {
1973                 reap_threads(&nr_running, &t_rate, &m_rate);
1974                 do_usleep(10000);
1975         }
1976
1977         fio_idle_prof_stop();
1978
1979         update_io_ticks();
1980 }
1981
1982 static void wait_for_helper_thread_exit(void)
1983 {
1984         void *ret;
1985
1986         helper_exit = 1;
1987         pthread_cond_signal(&helper_cond);
1988         pthread_join(helper_thread, &ret);
1989 }
1990
1991 static void free_disk_util(void)
1992 {
1993         disk_util_prune_entries();
1994
1995         pthread_cond_destroy(&helper_cond);
1996 }
1997
1998 static void *helper_thread_main(void *data)
1999 {
2000         int ret = 0;
2001
2002         fio_mutex_up(startup_mutex);
2003
2004         while (!ret) {
2005                 uint64_t sec = DISK_UTIL_MSEC / 1000;
2006                 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2007                 struct timespec ts;
2008                 struct timeval tv;
2009
2010                 gettimeofday(&tv, NULL);
2011                 ts.tv_sec = tv.tv_sec + sec;
2012                 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2013                 if (ts.tv_nsec > 1000000000ULL) {
2014                         ts.tv_nsec -= 1000000000ULL;
2015                         ts.tv_sec++;
2016                 }
2017
2018                 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2019
2020                 ret = update_io_ticks();
2021
2022                 if (helper_do_stat) {
2023                         helper_do_stat = 0;
2024                         __show_running_run_stats();
2025                 }
2026
2027                 if (!is_backend)
2028                         print_thread_status();
2029         }
2030
2031         return NULL;
2032 }
2033
2034 static int create_helper_thread(void)
2035 {
2036         int ret;
2037
2038         setup_disk_util();
2039
2040         pthread_cond_init(&helper_cond, NULL);
2041         pthread_mutex_init(&helper_lock, NULL);
2042
2043         ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2044         if (ret) {
2045                 log_err("Can't create helper thread: %s\n", strerror(ret));
2046                 return 1;
2047         }
2048
2049         dprint(FD_MUTEX, "wait on startup_mutex\n");
2050         fio_mutex_down(startup_mutex);
2051         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2052         return 0;
2053 }
2054
2055 int fio_backend(void)
2056 {
2057         struct thread_data *td;
2058         int i;
2059
2060         if (exec_profile) {
2061                 if (load_profile(exec_profile))
2062                         return 1;
2063                 free(exec_profile);
2064                 exec_profile = NULL;
2065         }
2066         if (!thread_number)
2067                 return 0;
2068
2069         if (write_bw_log) {
2070                 struct log_params p = {
2071                         .log_type = IO_LOG_TYPE_BW,
2072                 };
2073
2074                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2075                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2076                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2077         }
2078
2079         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2080         if (startup_mutex == NULL)
2081                 return 1;
2082
2083         set_genesis_time();
2084         stat_init();
2085         create_helper_thread();
2086
2087         cgroup_list = smalloc(sizeof(*cgroup_list));
2088         INIT_FLIST_HEAD(cgroup_list);
2089
2090         run_threads();
2091
2092         wait_for_helper_thread_exit();
2093
2094         if (!fio_abort) {
2095                 __show_run_stats();
2096                 if (write_bw_log) {
2097                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2098                                 struct io_log *log = agg_io_log[i];
2099
2100                                 flush_log(log);
2101                                 free_log(log);
2102                         }
2103                 }
2104         }
2105
2106         for_each_td(td, i) {
2107                 fio_options_free(td);
2108                 fio_mutex_remove(td->rusage_sem);
2109                 td->rusage_sem = NULL;
2110         }
2111
2112         free_disk_util();
2113         cgroup_kill(cgroup_list);
2114         sfree(cgroup_list);
2115         sfree(cgroup_mnt);
2116
2117         fio_mutex_remove(startup_mutex);
2118         stat_exit();
2119         return exit_value;
2120 }