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 #ifndef FIO_NO_HAVE_SHM_H
  38 #include <sys/shm.h>
  39 #endif
  40 #include <sys/mman.h>
  41
  42 #include "fio.h"
  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
  56 static pthread_t disk_util_thread;
  57 static struct fio_mutex *disk_thread_mutex;
  58 static struct fio_mutex *startup_mutex;
  59 static struct fio_mutex *writeout_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
  65 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  66
  67 int groupid = 0;
  68 unsigned int thread_number = 0;
  69 unsigned int stat_number = 0;
  70 unsigned int nr_process = 0;
  71 unsigned int nr_thread = 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         struct flist_head *entry, *n;
 242         struct io_u *io_u;
 243         int r;
 244
 245         /*
 246          * get immediately available events, if any
 247          */
 248         r = io_u_queued_complete(td, 0, NULL);
 249         if (r < 0)
 250                 return;
 251
 252         /*
 253          * now cancel remaining active events
 254          */
 255         if (td->io_ops->cancel) {
 256                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 257                         io_u = flist_entry(entry, struct io_u, list);
 258
 259                         /*
 260                          * if the io_u isn't in flight, then that generally
 261                          * means someone leaked an io_u. complain but fix
 262                          * it up, so we don't stall here.
 263                          */
 264                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 265                                 log_err("fio: non-busy IO on busy list\n");
 266                                 put_io_u(td, io_u);
 267                         } else {
 268                                 r = td->io_ops->cancel(td, io_u);
 269                                 if (!r)
 270                                         put_io_u(td, io_u);
 271                         }
 272                 }
 273         }
 274
 275         if (td->cur_depth)
 276                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 277 }
 278
 279 /*
 280  * Helper to handle the final sync of a file. Works just like the normal
 281  * io path, just does everything sync.
 282  */
 283 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 284 {
 285         struct io_u *io_u = __get_io_u(td);
 286         int ret;
 287
 288         if (!io_u)
 289                 return 1;
 290
 291         io_u->ddir = DDIR_SYNC;
 292         io_u->file = f;
 293
 294         if (td_io_prep(td, io_u)) {
 295                 put_io_u(td, io_u);
 296                 return 1;
 297         }
 298
 299 requeue:
 300         ret = td_io_queue(td, io_u);
 301         if (ret < 0) {
 302                 td_verror(td, io_u->error, "td_io_queue");
 303                 put_io_u(td, io_u);
 304                 return 1;
 305         } else if (ret == FIO_Q_QUEUED) {
 306                 if (io_u_queued_complete(td, 1, NULL) < 0)
 307                         return 1;
 308         } else if (ret == FIO_Q_COMPLETED) {
 309                 if (io_u->error) {
 310                         td_verror(td, io_u->error, "td_io_queue");
 311                         return 1;
 312                 }
 313
 314                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 315                         return 1;
 316         } else if (ret == FIO_Q_BUSY) {
 317                 if (td_io_commit(td))
 318                         return 1;
 319                 goto requeue;
 320         }
 321
 322         return 0;
 323 }
 324
 325 static inline void __update_tv_cache(struct thread_data *td)
 326 {
 327         fio_gettime(&td->tv_cache, NULL);
 328 }
 329
 330 static inline void update_tv_cache(struct thread_data *td)
 331 {
 332         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 333                 __update_tv_cache(td);
 334 }
 335
 336 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 337 {
 338         if (in_ramp_time(td))
 339                 return 0;
 340         if (!td->o.timeout)
 341                 return 0;
 342         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 343                 return 1;
 344
 345         return 0;
 346 }
 347
 348 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 349                                int *retptr)
 350 {
 351         int ret = *retptr;
 352
 353         if (ret < 0 || td->error) {
 354                 int err = td->error;
 355                 enum error_type_bit eb;
 356
 357                 if (ret < 0)
 358                         err = -ret;
 359
 360                 eb = td_error_type(ddir, err);
 361                 if (!(td->o.continue_on_error & (1 << eb)))
 362                         return 1;
 363
 364                 if (td_non_fatal_error(td, eb, err)) {
 365                         /*
 366                          * Continue with the I/Os in case of
 367                          * a non fatal error.
 368                          */
 369                         update_error_count(td, err);
 370                         td_clear_error(td);
 371                         *retptr = 0;
 372                         return 0;
 373                 } else if (td->o.fill_device && err == ENOSPC) {
 374                         /*
 375                          * We expect to hit this error if
 376                          * fill_device option is set.
 377                          */
 378                         td_clear_error(td);
 379                         td->terminate = 1;
 380                         return 1;
 381                 } else {
 382                         /*
 383                          * Stop the I/O in case of a fatal
 384                          * error.
 385                          */
 386                         update_error_count(td, err);
 387                         return 1;
 388                 }
 389         }
 390
 391         return 0;
 392 }
 393
 394 /*
 395  * The main verify engine. Runs over the writes we previously submitted,
 396  * reads the blocks back in, and checks the crc/md5 of the data.
 397  */
 398 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 399 {
 400         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 401         struct fio_file *f;
 402         struct io_u *io_u;
 403         int ret, min_events;
 404         unsigned int i;
 405
 406         dprint(FD_VERIFY, "starting loop\n");
 407
 408         /*
 409          * sync io first and invalidate cache, to make sure we really
 410          * read from disk.
 411          */
 412         for_each_file(td, f, i) {
 413                 if (!fio_file_open(f))
 414                         continue;
 415                 if (fio_io_sync(td, f))
 416                         break;
 417                 if (file_invalidate_cache(td, f))
 418                         break;
 419         }
 420
 421         if (td->error)
 422                 return;
 423
 424         td_set_runstate(td, TD_VERIFYING);
 425
 426         io_u = NULL;
 427         while (!td->terminate) {
 428                 enum fio_ddir ddir;
 429                 int ret2, full;
 430
 431                 update_tv_cache(td);
 432
 433                 if (runtime_exceeded(td, &td->tv_cache)) {
 434                         __update_tv_cache(td);
 435                         if (runtime_exceeded(td, &td->tv_cache)) {
 436                                 td->terminate = 1;
 437                                 break;
 438                         }
 439                 }
 440
 441                 if (flow_threshold_exceeded(td))
 442                         continue;
 443
 444                 if (!td->o.experimental_verify) {
 445                         io_u = __get_io_u(td);
 446                         if (!io_u)
 447                                 break;
 448
 449                         if (get_next_verify(td, io_u)) {
 450                                 put_io_u(td, io_u);
 451                                 break;
 452                         }
 453
 454                         if (td_io_prep(td, io_u)) {
 455                                 put_io_u(td, io_u);
 456                                 break;
 457                         }
 458                 } else {
 459                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 460                                 break;
 461
 462                         while ((io_u = get_io_u(td)) != NULL) {
 463                                 /*
 464                                  * We are only interested in the places where
 465                                  * we wrote or trimmed IOs. Turn those into
 466                                  * reads for verification purposes.
 467                                  */
 468                                 if (io_u->ddir == DDIR_READ) {
 469                                         /*
 470                                          * Pretend we issued it for rwmix
 471                                          * accounting
 472                                          */
 473                                         td->io_issues[DDIR_READ]++;
 474                                         put_io_u(td, io_u);
 475                                         continue;
 476                                 } else if (io_u->ddir == DDIR_TRIM) {
 477                                         io_u->ddir = DDIR_READ;
 478                                         io_u->flags |= IO_U_F_TRIMMED;
 479                                         break;
 480                                 } else if (io_u->ddir == DDIR_WRITE) {
 481                                         io_u->ddir = DDIR_READ;
 482                                         break;
 483                                 } else {
 484                                         put_io_u(td, io_u);
 485                                         continue;
 486                                 }
 487                         }
 488
 489                         if (!io_u)
 490                                 break;
 491                 }
 492
 493                 if (td->o.verify_async)
 494                         io_u->end_io = verify_io_u_async;
 495                 else
 496                         io_u->end_io = verify_io_u;
 497
 498                 ddir = io_u->ddir;
 499
 500                 ret = td_io_queue(td, io_u);
 501                 switch (ret) {
 502                 case FIO_Q_COMPLETED:
 503                         if (io_u->error) {
 504                                 ret = -io_u->error;
 505                                 clear_io_u(td, io_u);
 506                         } else if (io_u->resid) {
 507                                 int bytes = io_u->xfer_buflen - io_u->resid;
 508
 509                                 /*
 510                                  * zero read, fail
 511                                  */
 512                                 if (!bytes) {
 513                                         td_verror(td, EIO, "full resid");
 514                                         put_io_u(td, io_u);
 515                                         break;
 516                                 }
 517
 518                                 io_u->xfer_buflen = io_u->resid;
 519                                 io_u->xfer_buf += bytes;
 520                                 io_u->offset += bytes;
 521
 522                                 if (ddir_rw(io_u->ddir))
 523                                         td->ts.short_io_u[io_u->ddir]++;
 524
 525                                 f = io_u->file;
 526                                 if (io_u->offset == f->real_file_size)
 527                                         goto sync_done;
 528
 529                                 requeue_io_u(td, &io_u);
 530                         } else {
 531 sync_done:
 532                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 533                                 if (ret < 0)
 534                                         break;
 535                         }
 536                         continue;
 537                 case FIO_Q_QUEUED:
 538                         break;
 539                 case FIO_Q_BUSY:
 540                         requeue_io_u(td, &io_u);
 541                         ret2 = td_io_commit(td);
 542                         if (ret2 < 0)
 543                                 ret = ret2;
 544                         break;
 545                 default:
 546                         assert(ret < 0);
 547                         td_verror(td, -ret, "td_io_queue");
 548                         break;
 549                 }
 550
 551                 if (break_on_this_error(td, ddir, &ret))
 552                         break;
 553
 554                 /*
 555                  * if we can queue more, do so. but check if there are
 556                  * completed io_u's first. Note that we can get BUSY even
 557                  * without IO queued, if the system is resource starved.
 558                  */
 559                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 560                 if (full || !td->o.iodepth_batch_complete) {
 561                         min_events = min(td->o.iodepth_batch_complete,
 562                                          td->cur_depth);
 563                         /*
 564                          * if the queue is full, we MUST reap at least 1 event
 565                          */
 566                         if (full && !min_events)
 567                                 min_events = 1;
 568
 569                         do {
 570                                 /*
 571                                  * Reap required number of io units, if any,
 572                                  * and do the verification on them through
 573                                  * the callback handler
 574                                  */
 575                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 576                                         ret = -1;
 577                                         break;
 578                                 }
 579                         } while (full && (td->cur_depth > td->o.iodepth_low));
 580                 }
 581                 if (ret < 0)
 582                         break;
 583         }
 584
 585         if (!td->error) {
 586                 min_events = td->cur_depth;
 587
 588                 if (min_events)
 589                         ret = io_u_queued_complete(td, min_events, NULL);
 590         } else
 591                 cleanup_pending_aio(td);
 592
 593         td_set_runstate(td, TD_RUNNING);
 594
 595         dprint(FD_VERIFY, "exiting loop\n");
 596 }
 597
 598 static int io_bytes_exceeded(struct thread_data *td)
 599 {
 600         unsigned long long bytes;
 601
 602         if (td_rw(td))
 603                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 604         else if (td_write(td))
 605                 bytes = td->this_io_bytes[DDIR_WRITE];
 606         else if (td_read(td))
 607                 bytes = td->this_io_bytes[DDIR_READ];
 608         else
 609                 bytes = td->this_io_bytes[DDIR_TRIM];
 610
 611         return bytes >= td->o.size;
 612 }
 613
 614 /*
 615  * Main IO worker function. It retrieves io_u's to process and queues
 616  * and reaps them, checking for rate and errors along the way.
 617  *
 618  * Returns number of bytes written and trimmed.
 619  */
 620 static uint64_t do_io(struct thread_data *td)
 621 {
 622         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 623         unsigned int i;
 624         int ret = 0;
 625
 626         if (in_ramp_time(td))
 627                 td_set_runstate(td, TD_RAMP);
 628         else
 629                 td_set_runstate(td, TD_RUNNING);
 630
 631         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 632                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 633                 td->o.time_based) {
 634                 struct timeval comp_time;
 635                 int min_evts = 0;
 636                 struct io_u *io_u;
 637                 int ret2, full;
 638                 enum fio_ddir ddir;
 639
 640                 if (td->terminate || td->done)
 641                         break;
 642
 643                 update_tv_cache(td);
 644
 645                 if (runtime_exceeded(td, &td->tv_cache)) {
 646                         __update_tv_cache(td);
 647                         if (runtime_exceeded(td, &td->tv_cache)) {
 648                                 td->terminate = 1;
 649                                 break;
 650                         }
 651                 }
 652
 653                 if (flow_threshold_exceeded(td))
 654                         continue;
 655
 656                 io_u = get_io_u(td);
 657                 if (!io_u)
 658                         break;
 659
 660                 ddir = io_u->ddir;
 661
 662                 /*
 663                  * Add verification end_io handler if:
 664                  *      - Asked to verify (!td_rw(td))
 665                  *      - Or the io_u is from our verify list (mixed write/ver)
 666                  */
 667                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 668                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 669                         if (td->o.verify_async)
 670                                 io_u->end_io = verify_io_u_async;
 671                         else
 672                                 io_u->end_io = verify_io_u;
 673                         td_set_runstate(td, TD_VERIFYING);
 674                 } else if (in_ramp_time(td))
 675                         td_set_runstate(td, TD_RAMP);
 676                 else
 677                         td_set_runstate(td, TD_RUNNING);
 678
 679                 ret = td_io_queue(td, io_u);
 680                 switch (ret) {
 681                 case FIO_Q_COMPLETED:
 682                         if (io_u->error) {
 683                                 ret = -io_u->error;
 684                                 clear_io_u(td, io_u);
 685                         } else if (io_u->resid) {
 686                                 int bytes = io_u->xfer_buflen - io_u->resid;
 687                                 struct fio_file *f = io_u->file;
 688
 689                                 /*
 690                                  * zero read, fail
 691                                  */
 692                                 if (!bytes) {
 693                                         td_verror(td, EIO, "full resid");
 694                                         put_io_u(td, io_u);
 695                                         break;
 696                                 }
 697
 698                                 io_u->xfer_buflen = io_u->resid;
 699                                 io_u->xfer_buf += bytes;
 700                                 io_u->offset += bytes;
 701
 702                                 if (ddir_rw(io_u->ddir))
 703                                         td->ts.short_io_u[io_u->ddir]++;
 704
 705                                 if (io_u->offset == f->real_file_size)
 706                                         goto sync_done;
 707
 708                                 requeue_io_u(td, &io_u);
 709                         } else {
 710 sync_done:
 711                                 if (__should_check_rate(td, DDIR_READ) ||
 712                                     __should_check_rate(td, DDIR_WRITE) ||
 713                                     __should_check_rate(td, DDIR_TRIM))
 714                                         fio_gettime(&comp_time, NULL);
 715
 716                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 717                                 if (ret < 0)
 718                                         break;
 719                         }
 720                         break;
 721                 case FIO_Q_QUEUED:
 722                         /*
 723                          * if the engine doesn't have a commit hook,
 724                          * the io_u is really queued. if it does have such
 725                          * a hook, it has to call io_u_queued() itself.
 726                          */
 727                         if (td->io_ops->commit == NULL)
 728                                 io_u_queued(td, io_u);
 729                         break;
 730                 case FIO_Q_BUSY:
 731                         requeue_io_u(td, &io_u);
 732                         ret2 = td_io_commit(td);
 733                         if (ret2 < 0)
 734                                 ret = ret2;
 735                         break;
 736                 default:
 737                         assert(ret < 0);
 738                         put_io_u(td, io_u);
 739                         break;
 740                 }
 741
 742                 if (break_on_this_error(td, ddir, &ret))
 743                         break;
 744
 745                 /*
 746                  * See if we need to complete some commands. Note that we
 747                  * can get BUSY even without IO queued, if the system is
 748                  * resource starved.
 749                  */
 750                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 751                 if (full || !td->o.iodepth_batch_complete) {
 752                         min_evts = min(td->o.iodepth_batch_complete,
 753                                         td->cur_depth);
 754                         /*
 755                          * if the queue is full, we MUST reap at least 1 event
 756                          */
 757                         if (full && !min_evts)
 758                                 min_evts = 1;
 759
 760                         if (__should_check_rate(td, DDIR_READ) ||
 761                             __should_check_rate(td, DDIR_WRITE) ||
 762                             __should_check_rate(td, DDIR_TRIM))
 763                                 fio_gettime(&comp_time, NULL);
 764
 765                         do {
 766                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 767                                 if (ret < 0)
 768                                         break;
 769
 770                         } while (full && (td->cur_depth > td->o.iodepth_low));
 771                 }
 772
 773                 if (ret < 0)
 774                         break;
 775                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 776                         continue;
 777
 778                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 779                         if (check_min_rate(td, &comp_time, bytes_done)) {
 780                                 if (exitall_on_terminate)
 781                                         fio_terminate_threads(td->groupid);
 782                                 td_verror(td, EIO, "check_min_rate");
 783                                 break;
 784                         }
 785                 }
 786
 787                 if (td->o.thinktime) {
 788                         unsigned long long b;
 789
 790                         b = ddir_rw_sum(td->io_blocks);
 791                         if (!(b % td->o.thinktime_blocks)) {
 792                                 int left;
 793
 794                                 if (td->o.thinktime_spin)
 795                                         usec_spin(td->o.thinktime_spin);
 796
 797                                 left = td->o.thinktime - td->o.thinktime_spin;
 798                                 if (left)
 799                                         usec_sleep(td, left);
 800                         }
 801                 }
 802         }
 803
 804         if (td->trim_entries)
 805                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 806
 807         if (td->o.fill_device && td->error == ENOSPC) {
 808                 td->error = 0;
 809                 td->terminate = 1;
 810         }
 811         if (!td->error) {
 812                 struct fio_file *f;
 813
 814                 i = td->cur_depth;
 815                 if (i) {
 816                         ret = io_u_queued_complete(td, i, bytes_done);
 817                         if (td->o.fill_device && td->error == ENOSPC)
 818                                 td->error = 0;
 819                 }
 820
 821                 if (should_fsync(td) && td->o.end_fsync) {
 822                         td_set_runstate(td, TD_FSYNCING);
 823
 824                         for_each_file(td, f, i) {
 825                                 if (!fio_file_open(f))
 826                                         continue;
 827                                 fio_io_sync(td, f);
 828                         }
 829                 }
 830         } else
 831                 cleanup_pending_aio(td);
 832
 833         /*
 834          * stop job if we failed doing any IO
 835          */
 836         if (!ddir_rw_sum(td->this_io_bytes))
 837                 td->done = 1;
 838
 839         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 840 }
 841
 842 static void cleanup_io_u(struct thread_data *td)
 843 {
 844         struct flist_head *entry, *n;
 845         struct io_u *io_u;
 846
 847         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 848                 io_u = flist_entry(entry, struct io_u, list);
 849
 850                 flist_del(&io_u->list);
 851
 852                 if (td->io_ops->io_u_free)
 853                         td->io_ops->io_u_free(td, io_u);
 854
 855                 fio_memfree(io_u, sizeof(*io_u));
 856         }
 857
 858         free_io_mem(td);
 859 }
 860
 861 static int init_io_u(struct thread_data *td)
 862 {
 863         struct io_u *io_u;
 864         unsigned int max_bs, min_write;
 865         int cl_align, i, max_units;
 866         int data_xfer = 1;
 867         char *p;
 868
 869         max_units = td->o.iodepth;
 870         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 871         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
 872         min_write = td->o.min_bs[DDIR_WRITE];
 873         td->orig_buffer_size = (unsigned long long) max_bs
 874                                         * (unsigned long long) max_units;
 875
 876         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 877                 data_xfer = 0;
 878
 879         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 880                 unsigned long bs;
 881
 882                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 883                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 884         }
 885
 886         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 887                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 888                 return 1;
 889         }
 890
 891         if (data_xfer && allocate_io_mem(td))
 892                 return 1;
 893
 894         if (td->o.odirect || td->o.mem_align ||
 895             (td->io_ops->flags & FIO_RAWIO))
 896                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 897         else
 898                 p = td->orig_buffer;
 899
 900         cl_align = os_cache_line_size();
 901
 902         for (i = 0; i < max_units; i++) {
 903                 void *ptr;
 904
 905                 if (td->terminate)
 906                         return 1;
 907
 908                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 909                 if (!ptr) {
 910                         log_err("fio: unable to allocate aligned memory\n");
 911                         break;
 912                 }
 913
 914                 io_u = ptr;
 915                 memset(io_u, 0, sizeof(*io_u));
 916                 INIT_FLIST_HEAD(&io_u->list);
 917                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 918
 919                 if (data_xfer) {
 920                         io_u->buf = p;
 921                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 922
 923                         if (td_write(td))
 924                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 925                         if (td_write(td) && td->o.verify_pattern_bytes) {
 926                                 /*
 927                                  * Fill the buffer with the pattern if we are
 928                                  * going to be doing writes.
 929                                  */
 930                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 931                         }
 932                 }
 933
 934                 io_u->index = i;
 935                 io_u->flags = IO_U_F_FREE;
 936                 flist_add(&io_u->list, &td->io_u_freelist);
 937
 938                 if (td->io_ops->io_u_init) {
 939                         int ret = td->io_ops->io_u_init(td, io_u);
 940
 941                         if (ret) {
 942                                 log_err("fio: failed to init engine data: %d\n", ret);
 943                                 return 1;
 944                         }
 945                 }
 946
 947                 p += max_bs;
 948         }
 949
 950         return 0;
 951 }
 952
 953 static int switch_ioscheduler(struct thread_data *td)
 954 {
 955         char tmp[256], tmp2[128];
 956         FILE *f;
 957         int ret;
 958
 959         if (td->io_ops->flags & FIO_DISKLESSIO)
 960                 return 0;
 961
 962         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 963
 964         f = fopen(tmp, "r+");
 965         if (!f) {
 966                 if (errno == ENOENT) {
 967                         log_err("fio: os or kernel doesn't support IO scheduler"
 968                                 " switching\n");
 969                         return 0;
 970                 }
 971                 td_verror(td, errno, "fopen iosched");
 972                 return 1;
 973         }
 974
 975         /*
 976          * Set io scheduler.
 977          */
 978         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 979         if (ferror(f) || ret != 1) {
 980                 td_verror(td, errno, "fwrite");
 981                 fclose(f);
 982                 return 1;
 983         }
 984
 985         rewind(f);
 986
 987         /*
 988          * Read back and check that the selected scheduler is now the default.
 989          */
 990         ret = fread(tmp, 1, sizeof(tmp), f);
 991         if (ferror(f) || ret < 0) {
 992                 td_verror(td, errno, "fread");
 993                 fclose(f);
 994                 return 1;
 995         }
 996
 997         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 998         if (!strstr(tmp, tmp2)) {
 999                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1000                 td_verror(td, EINVAL, "iosched_switch");
1001                 fclose(f);
1002                 return 1;
1003         }
1004
1005         fclose(f);
1006         return 0;
1007 }
1008
1009 static int keep_running(struct thread_data *td)
1010 {
1011         if (td->done)
1012                 return 0;
1013         if (td->o.time_based)
1014                 return 1;
1015         if (td->o.loops) {
1016                 td->o.loops--;
1017                 return 1;
1018         }
1019
1020         if (ddir_rw_sum(td->io_bytes) < td->o.size) {
1021                 uint64_t diff;
1022
1023                 /*
1024                  * If the difference is less than the minimum IO size, we
1025                  * are done.
1026                  */
1027                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1028                 if (diff < td->o.rw_min_bs)
1029                         return 0;
1030
1031                 return 1;
1032         }
1033
1034         return 0;
1035 }
1036
1037 static int exec_string(const char *string)
1038 {
1039         int ret, newlen = strlen(string) + 1 + 8;
1040         char *str;
1041
1042         str = malloc(newlen);
1043         sprintf(str, "sh -c %s", string);
1044
1045         ret = system(str);
1046         if (ret == -1)
1047                 log_err("fio: exec of cmd <%s> failed\n", str);
1048
1049         free(str);
1050         return ret;
1051 }
1052
1053 /*
1054  * Entry point for the thread based jobs. The process based jobs end up
1055  * here as well, after a little setup.
1056  */
1057 static void *thread_main(void *data)
1058 {
1059         unsigned long long elapsed;
1060         struct thread_data *td = data;
1061         pthread_condattr_t attr;
1062         int clear_state;
1063
1064         if (!td->o.use_thread) {
1065                 setsid();
1066                 td->pid = getpid();
1067         } else
1068                 td->pid = gettid();
1069
1070         fio_local_clock_init(td->o.use_thread);
1071
1072         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1073
1074         INIT_FLIST_HEAD(&td->io_u_freelist);
1075         INIT_FLIST_HEAD(&td->io_u_busylist);
1076         INIT_FLIST_HEAD(&td->io_u_requeues);
1077         INIT_FLIST_HEAD(&td->io_log_list);
1078         INIT_FLIST_HEAD(&td->io_hist_list);
1079         INIT_FLIST_HEAD(&td->verify_list);
1080         INIT_FLIST_HEAD(&td->trim_list);
1081         INIT_FLIST_HEAD(&td->next_rand_list);
1082         pthread_mutex_init(&td->io_u_lock, NULL);
1083         td->io_hist_tree = RB_ROOT;
1084
1085         pthread_condattr_init(&attr);
1086         pthread_cond_init(&td->verify_cond, &attr);
1087         pthread_cond_init(&td->free_cond, &attr);
1088
1089         td_set_runstate(td, TD_INITIALIZED);
1090         dprint(FD_MUTEX, "up startup_mutex\n");
1091         fio_mutex_up(startup_mutex);
1092         dprint(FD_MUTEX, "wait on td->mutex\n");
1093         fio_mutex_down(td->mutex);
1094         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1095
1096         /*
1097          * the ->mutex mutex is now no longer used, close it to avoid
1098          * eating a file descriptor
1099          */
1100         fio_mutex_remove(td->mutex);
1101
1102         /*
1103          * A new gid requires privilege, so we need to do this before setting
1104          * the uid.
1105          */
1106         if (td->o.gid != -1U && setgid(td->o.gid)) {
1107                 td_verror(td, errno, "setgid");
1108                 goto err;
1109         }
1110         if (td->o.uid != -1U && setuid(td->o.uid)) {
1111                 td_verror(td, errno, "setuid");
1112                 goto err;
1113         }
1114
1115         /*
1116          * If we have a gettimeofday() thread, make sure we exclude that
1117          * thread from this job
1118          */
1119         if (td->o.gtod_cpu)
1120                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1121
1122         /*
1123          * Set affinity first, in case it has an impact on the memory
1124          * allocations.
1125          */
1126         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1127                 td_verror(td, errno, "cpu_set_affinity");
1128                 goto err;
1129         }
1130
1131 #ifdef CONFIG_LIBNUMA
1132         /* numa node setup */
1133         if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1134                 int ret;
1135
1136                 if (numa_available() < 0) {
1137                         td_verror(td, errno, "Does not support NUMA API\n");
1138                         goto err;
1139                 }
1140
1141                 if (td->o.numa_cpumask_set) {
1142                         ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1143                         if (ret == -1) {
1144                                 td_verror(td, errno, \
1145                                         "numa_run_on_node_mask failed\n");
1146                                 goto err;
1147                         }
1148                 }
1149
1150                 if (td->o.numa_memmask_set) {
1151
1152                         switch (td->o.numa_mem_mode) {
1153                         case MPOL_INTERLEAVE:
1154                                 numa_set_interleave_mask(td->o.numa_memnodesmask);
1155                                 break;
1156                         case MPOL_BIND:
1157                                 numa_set_membind(td->o.numa_memnodesmask);
1158                                 break;
1159                         case MPOL_LOCAL:
1160                                 numa_set_localalloc();
1161                                 break;
1162                         case MPOL_PREFERRED:
1163                                 numa_set_preferred(td->o.numa_mem_prefer_node);
1164                                 break;
1165                         case MPOL_DEFAULT:
1166                         default:
1167                                 break;
1168                         }
1169
1170                 }
1171         }
1172 #endif
1173
1174         /*
1175          * May alter parameters that init_io_u() will use, so we need to
1176          * do this first.
1177          */
1178         if (init_iolog(td))
1179                 goto err;
1180
1181         if (init_io_u(td))
1182                 goto err;
1183
1184         if (td->o.verify_async && verify_async_init(td))
1185                 goto err;
1186
1187         if (td->ioprio_set) {
1188                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1189                         td_verror(td, errno, "ioprio_set");
1190                         goto err;
1191                 }
1192         }
1193
1194         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1195                 goto err;
1196
1197         errno = 0;
1198         if (nice(td->o.nice) == -1 && errno != 0) {
1199                 td_verror(td, errno, "nice");
1200                 goto err;
1201         }
1202
1203         if (td->o.ioscheduler && switch_ioscheduler(td))
1204                 goto err;
1205
1206         if (!td->o.create_serialize && setup_files(td))
1207                 goto err;
1208
1209         if (td_io_init(td))
1210                 goto err;
1211
1212         if (init_random_map(td))
1213                 goto err;
1214
1215         if (td->o.exec_prerun) {
1216                 if (exec_string(td->o.exec_prerun))
1217                         goto err;
1218         }
1219
1220         if (td->o.pre_read) {
1221                 if (pre_read_files(td) < 0)
1222                         goto err;
1223         }
1224
1225         fio_gettime(&td->epoch, NULL);
1226         fio_getrusage(&td->ru_start);
1227         clear_state = 0;
1228         while (keep_running(td)) {
1229                 uint64_t verify_bytes;
1230
1231                 fio_gettime(&td->start, NULL);
1232                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1233                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1234                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1235
1236                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1237                                 td->o.ratemin[DDIR_TRIM]) {
1238                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1239                                                 sizeof(td->bw_sample_time));
1240                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1241                                                 sizeof(td->bw_sample_time));
1242                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1243                                                 sizeof(td->bw_sample_time));
1244                 }
1245
1246                 if (clear_state)
1247                         clear_io_state(td);
1248
1249                 prune_io_piece_log(td);
1250
1251                 verify_bytes = do_io(td);
1252
1253                 clear_state = 1;
1254
1255                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1256                         elapsed = utime_since_now(&td->start);
1257                         td->ts.runtime[DDIR_READ] += elapsed;
1258                 }
1259                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1260                         elapsed = utime_since_now(&td->start);
1261                         td->ts.runtime[DDIR_WRITE] += elapsed;
1262                 }
1263                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1264                         elapsed = utime_since_now(&td->start);
1265                         td->ts.runtime[DDIR_TRIM] += elapsed;
1266                 }
1267
1268                 if (td->error || td->terminate)
1269                         break;
1270
1271                 if (!td->o.do_verify ||
1272                     td->o.verify == VERIFY_NONE ||
1273                     (td->io_ops->flags & FIO_UNIDIR))
1274                         continue;
1275
1276                 clear_io_state(td);
1277
1278                 fio_gettime(&td->start, NULL);
1279
1280                 do_verify(td, verify_bytes);
1281
1282                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1283
1284                 if (td->error || td->terminate)
1285                         break;
1286         }
1287
1288         update_rusage_stat(td);
1289         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1290         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1291         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1292         td->ts.total_run_time = mtime_since_now(&td->epoch);
1293         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1294         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1295         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1296
1297         fio_mutex_down(writeout_mutex);
1298         if (td->bw_log) {
1299                 if (td->o.bw_log_file) {
1300                         finish_log_named(td, td->bw_log,
1301                                                 td->o.bw_log_file, "bw");
1302                 } else
1303                         finish_log(td, td->bw_log, "bw");
1304         }
1305         if (td->lat_log) {
1306                 if (td->o.lat_log_file) {
1307                         finish_log_named(td, td->lat_log,
1308                                                 td->o.lat_log_file, "lat");
1309                 } else
1310                         finish_log(td, td->lat_log, "lat");
1311         }
1312         if (td->slat_log) {
1313                 if (td->o.lat_log_file) {
1314                         finish_log_named(td, td->slat_log,
1315                                                 td->o.lat_log_file, "slat");
1316                 } else
1317                         finish_log(td, td->slat_log, "slat");
1318         }
1319         if (td->clat_log) {
1320                 if (td->o.lat_log_file) {
1321                         finish_log_named(td, td->clat_log,
1322                                                 td->o.lat_log_file, "clat");
1323                 } else
1324                         finish_log(td, td->clat_log, "clat");
1325         }
1326         if (td->iops_log) {
1327                 if (td->o.iops_log_file) {
1328                         finish_log_named(td, td->iops_log,
1329                                                 td->o.iops_log_file, "iops");
1330                 } else
1331                         finish_log(td, td->iops_log, "iops");
1332         }
1333
1334         fio_mutex_up(writeout_mutex);
1335         if (td->o.exec_postrun)
1336                 exec_string(td->o.exec_postrun);
1337
1338         if (exitall_on_terminate)
1339                 fio_terminate_threads(td->groupid);
1340
1341 err:
1342         if (td->error)
1343                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1344                                                         td->verror);
1345
1346         if (td->o.verify_async)
1347                 verify_async_exit(td);
1348
1349         close_and_free_files(td);
1350         cleanup_io_u(td);
1351         close_ioengine(td);
1352         cgroup_shutdown(td, &cgroup_mnt);
1353
1354         if (td->o.cpumask_set) {
1355                 int ret = fio_cpuset_exit(&td->o.cpumask);
1356
1357                 td_verror(td, ret, "fio_cpuset_exit");
1358         }
1359
1360         /*
1361          * do this very late, it will log file closing as well
1362          */
1363         if (td->o.write_iolog_file)
1364                 write_iolog_close(td);
1365
1366         td_set_runstate(td, TD_EXITED);
1367         return (void *) (uintptr_t) td->error;
1368 }
1369
1370
1371 /*
1372  * We cannot pass the td data into a forked process, so attach the td and
1373  * pass it to the thread worker.
1374  */
1375 static int fork_main(int shmid, int offset)
1376 {
1377         struct thread_data *td;
1378         void *data, *ret;
1379
1380 #ifndef __hpux
1381         data = shmat(shmid, NULL, 0);
1382         if (data == (void *) -1) {
1383                 int __err = errno;
1384
1385                 perror("shmat");
1386                 return __err;
1387         }
1388 #else
1389         /*
1390          * HP-UX inherits shm mappings?
1391          */
1392         data = threads;
1393 #endif
1394
1395         td = data + offset * sizeof(struct thread_data);
1396         ret = thread_main(td);
1397         shmdt(data);
1398         return (int) (uintptr_t) ret;
1399 }
1400
1401 /*
1402  * Run over the job map and reap the threads that have exited, if any.
1403  */
1404 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1405                          unsigned int *m_rate)
1406 {
1407         struct thread_data *td;
1408         unsigned int cputhreads, realthreads, pending;
1409         int i, status, ret;
1410
1411         /*
1412          * reap exited threads (TD_EXITED -> TD_REAPED)
1413          */
1414         realthreads = pending = cputhreads = 0;
1415         for_each_td(td, i) {
1416                 int flags = 0;
1417
1418                 /*
1419                  * ->io_ops is NULL for a thread that has closed its
1420                  * io engine
1421                  */
1422                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1423                         cputhreads++;
1424                 else
1425                         realthreads++;
1426
1427                 if (!td->pid) {
1428                         pending++;
1429                         continue;
1430                 }
1431                 if (td->runstate == TD_REAPED)
1432                         continue;
1433                 if (td->o.use_thread) {
1434                         if (td->runstate == TD_EXITED) {
1435                                 td_set_runstate(td, TD_REAPED);
1436                                 goto reaped;
1437                         }
1438                         continue;
1439                 }
1440
1441                 flags = WNOHANG;
1442                 if (td->runstate == TD_EXITED)
1443                         flags = 0;
1444
1445                 /*
1446                  * check if someone quit or got killed in an unusual way
1447                  */
1448                 ret = waitpid(td->pid, &status, flags);
1449                 if (ret < 0) {
1450                         if (errno == ECHILD) {
1451                                 log_err("fio: pid=%d disappeared %d\n",
1452                                                 (int) td->pid, td->runstate);
1453                                 td->sig = ECHILD;
1454                                 td_set_runstate(td, TD_REAPED);
1455                                 goto reaped;
1456                         }
1457                         perror("waitpid");
1458                 } else if (ret == td->pid) {
1459                         if (WIFSIGNALED(status)) {
1460                                 int sig = WTERMSIG(status);
1461
1462                                 if (sig != SIGTERM && sig != SIGUSR2)
1463                                         log_err("fio: pid=%d, got signal=%d\n",
1464                                                         (int) td->pid, sig);
1465                                 td->sig = sig;
1466                                 td_set_runstate(td, TD_REAPED);
1467                                 goto reaped;
1468                         }
1469                         if (WIFEXITED(status)) {
1470                                 if (WEXITSTATUS(status) && !td->error)
1471                                         td->error = WEXITSTATUS(status);
1472
1473                                 td_set_runstate(td, TD_REAPED);
1474                                 goto reaped;
1475                         }
1476                 }
1477
1478                 /*
1479                  * thread is not dead, continue
1480                  */
1481                 pending++;
1482                 continue;
1483 reaped:
1484                 (*nr_running)--;
1485                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1486                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1487                 if (!td->pid)
1488                         pending--;
1489
1490                 if (td->error)
1491                         exit_value++;
1492
1493                 done_secs += mtime_since_now(&td->epoch) / 1000;
1494         }
1495
1496         if (*nr_running == cputhreads && !pending && realthreads)
1497                 fio_terminate_threads(TERMINATE_ALL);
1498 }
1499
1500 /*
1501  * Main function for kicking off and reaping jobs, as needed.
1502  */
1503 static void run_threads(void)
1504 {
1505         struct thread_data *td;
1506         unsigned long spent;
1507         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1508
1509         if (fio_pin_memory())
1510                 return;
1511
1512         if (fio_gtod_offload && fio_start_gtod_thread())
1513                 return;
1514
1515         fio_idle_prof_init();
1516
1517         set_sig_handlers();
1518
1519         if (output_format == FIO_OUTPUT_NORMAL) {
1520                 log_info("Starting ");
1521                 if (nr_thread)
1522                         log_info("%d thread%s", nr_thread,
1523                                                 nr_thread > 1 ? "s" : "");
1524                 if (nr_process) {
1525                         if (nr_thread)
1526                                 log_info(" and ");
1527                         log_info("%d process%s", nr_process,
1528                                                 nr_process > 1 ? "es" : "");
1529                 }
1530                 log_info("\n");
1531                 fflush(stdout);
1532         }
1533
1534         todo = thread_number;
1535         nr_running = 0;
1536         nr_started = 0;
1537         m_rate = t_rate = 0;
1538
1539         for_each_td(td, i) {
1540                 print_status_init(td->thread_number - 1);
1541
1542                 if (!td->o.create_serialize)
1543                         continue;
1544
1545                 /*
1546                  * do file setup here so it happens sequentially,
1547                  * we don't want X number of threads getting their
1548                  * client data interspersed on disk
1549                  */
1550                 if (setup_files(td)) {
1551                         exit_value++;
1552                         if (td->error)
1553                                 log_err("fio: pid=%d, err=%d/%s\n",
1554                                         (int) td->pid, td->error, td->verror);
1555                         td_set_runstate(td, TD_REAPED);
1556                         todo--;
1557                 } else {
1558                         struct fio_file *f;
1559                         unsigned int j;
1560
1561                         /*
1562                          * for sharing to work, each job must always open
1563                          * its own files. so close them, if we opened them
1564                          * for creation
1565                          */
1566                         for_each_file(td, f, j) {
1567                                 if (fio_file_open(f))
1568                                         td_io_close_file(td, f);
1569                         }
1570                 }
1571         }
1572
1573         /* start idle threads before io threads start to run */
1574         fio_idle_prof_start();
1575
1576         set_genesis_time();
1577
1578         while (todo) {
1579                 struct thread_data *map[REAL_MAX_JOBS];
1580                 struct timeval this_start;
1581                 int this_jobs = 0, left;
1582
1583                 /*
1584                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1585                  */
1586                 for_each_td(td, i) {
1587                         if (td->runstate != TD_NOT_CREATED)
1588                                 continue;
1589
1590                         /*
1591                          * never got a chance to start, killed by other
1592                          * thread for some reason
1593                          */
1594                         if (td->terminate) {
1595                                 todo--;
1596                                 continue;
1597                         }
1598
1599                         if (td->o.start_delay) {
1600                                 spent = mtime_since_genesis();
1601
1602                                 if (td->o.start_delay * 1000 > spent)
1603                                         continue;
1604                         }
1605
1606                         if (td->o.stonewall && (nr_started || nr_running)) {
1607                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1608                                                         td->o.name);
1609                                 break;
1610                         }
1611
1612                         init_disk_util(td);
1613
1614                         /*
1615                          * Set state to created. Thread will transition
1616                          * to TD_INITIALIZED when it's done setting up.
1617                          */
1618                         td_set_runstate(td, TD_CREATED);
1619                         map[this_jobs++] = td;
1620                         nr_started++;
1621
1622                         if (td->o.use_thread) {
1623                                 int ret;
1624
1625                                 dprint(FD_PROCESS, "will pthread_create\n");
1626                                 ret = pthread_create(&td->thread, NULL,
1627                                                         thread_main, td);
1628                                 if (ret) {
1629                                         log_err("pthread_create: %s\n",
1630                                                         strerror(ret));
1631                                         nr_started--;
1632                                         break;
1633                                 }
1634                                 ret = pthread_detach(td->thread);
1635                                 if (ret)
1636                                         log_err("pthread_detach: %s",
1637                                                         strerror(ret));
1638                         } else {
1639                                 pid_t pid;
1640                                 dprint(FD_PROCESS, "will fork\n");
1641                                 pid = fork();
1642                                 if (!pid) {
1643                                         int ret = fork_main(shm_id, i);
1644
1645                                         _exit(ret);
1646                                 } else if (i == fio_debug_jobno)
1647                                         *fio_debug_jobp = pid;
1648                         }
1649                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1650                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1651                                 log_err("fio: job startup hung? exiting.\n");
1652                                 fio_terminate_threads(TERMINATE_ALL);
1653                                 fio_abort = 1;
1654                                 nr_started--;
1655                                 break;
1656                         }
1657                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1658                 }
1659
1660                 /*
1661                  * Wait for the started threads to transition to
1662                  * TD_INITIALIZED.
1663                  */
1664                 fio_gettime(&this_start, NULL);
1665                 left = this_jobs;
1666                 while (left && !fio_abort) {
1667                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1668                                 break;
1669
1670                         usleep(100000);
1671
1672                         for (i = 0; i < this_jobs; i++) {
1673                                 td = map[i];
1674                                 if (!td)
1675                                         continue;
1676                                 if (td->runstate == TD_INITIALIZED) {
1677                                         map[i] = NULL;
1678                                         left--;
1679                                 } else if (td->runstate >= TD_EXITED) {
1680                                         map[i] = NULL;
1681                                         left--;
1682                                         todo--;
1683                                         nr_running++; /* work-around... */
1684                                 }
1685                         }
1686                 }
1687
1688                 if (left) {
1689                         log_err("fio: %d job%s failed to start\n", left,
1690                                         left > 1 ? "s" : "");
1691                         for (i = 0; i < this_jobs; i++) {
1692                                 td = map[i];
1693                                 if (!td)
1694                                         continue;
1695                                 kill(td->pid, SIGTERM);
1696                         }
1697                         break;
1698                 }
1699
1700                 /*
1701                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1702                  */
1703                 for_each_td(td, i) {
1704                         if (td->runstate != TD_INITIALIZED)
1705                                 continue;
1706
1707                         if (in_ramp_time(td))
1708                                 td_set_runstate(td, TD_RAMP);
1709                         else
1710                                 td_set_runstate(td, TD_RUNNING);
1711                         nr_running++;
1712                         nr_started--;
1713                         m_rate += ddir_rw_sum(td->o.ratemin);
1714                         t_rate += ddir_rw_sum(td->o.rate);
1715                         todo--;
1716                         fio_mutex_up(td->mutex);
1717                 }
1718
1719                 reap_threads(&nr_running, &t_rate, &m_rate);
1720
1721                 if (todo) {
1722                         if (is_backend)
1723                                 fio_server_idle_loop();
1724                         else
1725                                 usleep(100000);
1726                 }
1727         }
1728
1729         while (nr_running) {
1730                 reap_threads(&nr_running, &t_rate, &m_rate);
1731
1732                 if (is_backend)
1733                         fio_server_idle_loop();
1734                 else
1735                         usleep(10000);
1736         }
1737
1738         fio_idle_prof_stop();
1739
1740         update_io_ticks();
1741         fio_unpin_memory();
1742 }
1743
1744 void wait_for_disk_thread_exit(void)
1745 {
1746         fio_mutex_down(disk_thread_mutex);
1747 }
1748
1749 static void free_disk_util(void)
1750 {
1751         disk_util_start_exit();
1752         wait_for_disk_thread_exit();
1753         disk_util_prune_entries();
1754 }
1755
1756 static void *disk_thread_main(void *data)
1757 {
1758         int ret = 0;
1759
1760         fio_mutex_up(startup_mutex);
1761
1762         while (threads && !ret) {
1763                 usleep(DISK_UTIL_MSEC * 1000);
1764                 if (!threads)
1765                         break;
1766                 ret = update_io_ticks();
1767
1768                 if (!is_backend)
1769                         print_thread_status();
1770         }
1771
1772         fio_mutex_up(disk_thread_mutex);
1773         return NULL;
1774 }
1775
1776 static int create_disk_util_thread(void)
1777 {
1778         int ret;
1779
1780         setup_disk_util();
1781
1782         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1783
1784         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1785         if (ret) {
1786                 fio_mutex_remove(disk_thread_mutex);
1787                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1788                 return 1;
1789         }
1790
1791         ret = pthread_detach(disk_util_thread);
1792         if (ret) {
1793                 fio_mutex_remove(disk_thread_mutex);
1794                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1795                 return 1;
1796         }
1797
1798         dprint(FD_MUTEX, "wait on startup_mutex\n");
1799         fio_mutex_down(startup_mutex);
1800         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1801         return 0;
1802 }
1803
1804 int fio_backend(void)
1805 {
1806         struct thread_data *td;
1807         int i;
1808
1809         if (exec_profile) {
1810                 if (load_profile(exec_profile))
1811                         return 1;
1812                 free(exec_profile);
1813                 exec_profile = NULL;
1814         }
1815         if (!thread_number)
1816                 return 0;
1817
1818         if (write_bw_log) {
1819                 setup_log(&agg_io_log[DDIR_READ], 0);
1820                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1821                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1822         }
1823
1824         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1825         if (startup_mutex == NULL)
1826                 return 1;
1827         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1828         if (writeout_mutex == NULL)
1829                 return 1;
1830
1831         set_genesis_time();
1832         create_disk_util_thread();
1833
1834         cgroup_list = smalloc(sizeof(*cgroup_list));
1835         INIT_FLIST_HEAD(cgroup_list);
1836
1837         run_threads();
1838
1839         if (!fio_abort) {
1840                 show_run_stats();
1841                 if (write_bw_log) {
1842                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1843                         __finish_log(agg_io_log[DDIR_WRITE],
1844                                         "agg-write_bw.log");
1845                         __finish_log(agg_io_log[DDIR_TRIM],
1846                                         "agg-write_bw.log");
1847                 }
1848         }
1849
1850         for_each_td(td, i)
1851                 fio_options_free(td);
1852
1853         free_disk_util();
1854         cgroup_kill(cgroup_list);
1855         sfree(cgroup_list);
1856         sfree(cgroup_mnt);
1857
1858         fio_mutex_remove(startup_mutex);
1859         fio_mutex_remove(writeout_mutex);
1860         fio_mutex_remove(disk_thread_mutex);
1861         return exit_value;
1862 }