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 int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 326 {
 327         int ret;
 328
 329         if (fio_file_open(f))
 330                 return fio_io_sync(td, f);
 331
 332         if (td_io_open_file(td, f))
 333                 return 1;
 334
 335         ret = fio_io_sync(td, f);
 336         td_io_close_file(td, f);
 337         return ret;
 338 }
 339
 340 static inline void __update_tv_cache(struct thread_data *td)
 341 {
 342         fio_gettime(&td->tv_cache, NULL);
 343 }
 344
 345 static inline void update_tv_cache(struct thread_data *td)
 346 {
 347         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 348                 __update_tv_cache(td);
 349 }
 350
 351 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 352 {
 353         if (in_ramp_time(td))
 354                 return 0;
 355         if (!td->o.timeout)
 356                 return 0;
 357         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 358                 return 1;
 359
 360         return 0;
 361 }
 362
 363 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 364                                int *retptr)
 365 {
 366         int ret = *retptr;
 367
 368         if (ret < 0 || td->error) {
 369                 int err = td->error;
 370                 enum error_type_bit eb;
 371
 372                 if (ret < 0)
 373                         err = -ret;
 374
 375                 eb = td_error_type(ddir, err);
 376                 if (!(td->o.continue_on_error & (1 << eb)))
 377                         return 1;
 378
 379                 if (td_non_fatal_error(td, eb, err)) {
 380                         /*
 381                          * Continue with the I/Os in case of
 382                          * a non fatal error.
 383                          */
 384                         update_error_count(td, err);
 385                         td_clear_error(td);
 386                         *retptr = 0;
 387                         return 0;
 388                 } else if (td->o.fill_device && err == ENOSPC) {
 389                         /*
 390                          * We expect to hit this error if
 391                          * fill_device option is set.
 392                          */
 393                         td_clear_error(td);
 394                         td->terminate = 1;
 395                         return 1;
 396                 } else {
 397                         /*
 398                          * Stop the I/O in case of a fatal
 399                          * error.
 400                          */
 401                         update_error_count(td, err);
 402                         return 1;
 403                 }
 404         }
 405
 406         return 0;
 407 }
 408
 409 /*
 410  * The main verify engine. Runs over the writes we previously submitted,
 411  * reads the blocks back in, and checks the crc/md5 of the data.
 412  */
 413 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 414 {
 415         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 416         struct fio_file *f;
 417         struct io_u *io_u;
 418         int ret, min_events;
 419         unsigned int i;
 420
 421         dprint(FD_VERIFY, "starting loop\n");
 422
 423         /*
 424          * sync io first and invalidate cache, to make sure we really
 425          * read from disk.
 426          */
 427         for_each_file(td, f, i) {
 428                 if (!fio_file_open(f))
 429                         continue;
 430                 if (fio_io_sync(td, f))
 431                         break;
 432                 if (file_invalidate_cache(td, f))
 433                         break;
 434         }
 435
 436         if (td->error)
 437                 return;
 438
 439         td_set_runstate(td, TD_VERIFYING);
 440
 441         io_u = NULL;
 442         while (!td->terminate) {
 443                 enum fio_ddir ddir;
 444                 int ret2, full;
 445
 446                 update_tv_cache(td);
 447
 448                 if (runtime_exceeded(td, &td->tv_cache)) {
 449                         __update_tv_cache(td);
 450                         if (runtime_exceeded(td, &td->tv_cache)) {
 451                                 td->terminate = 1;
 452                                 break;
 453                         }
 454                 }
 455
 456                 if (flow_threshold_exceeded(td))
 457                         continue;
 458
 459                 if (!td->o.experimental_verify) {
 460                         io_u = __get_io_u(td);
 461                         if (!io_u)
 462                                 break;
 463
 464                         if (get_next_verify(td, io_u)) {
 465                                 put_io_u(td, io_u);
 466                                 break;
 467                         }
 468
 469                         if (td_io_prep(td, io_u)) {
 470                                 put_io_u(td, io_u);
 471                                 break;
 472                         }
 473                 } else {
 474                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 475                                 break;
 476
 477                         while ((io_u = get_io_u(td)) != NULL) {
 478                                 /*
 479                                  * We are only interested in the places where
 480                                  * we wrote or trimmed IOs. Turn those into
 481                                  * reads for verification purposes.
 482                                  */
 483                                 if (io_u->ddir == DDIR_READ) {
 484                                         /*
 485                                          * Pretend we issued it for rwmix
 486                                          * accounting
 487                                          */
 488                                         td->io_issues[DDIR_READ]++;
 489                                         put_io_u(td, io_u);
 490                                         continue;
 491                                 } else if (io_u->ddir == DDIR_TRIM) {
 492                                         io_u->ddir = DDIR_READ;
 493                                         io_u->flags |= IO_U_F_TRIMMED;
 494                                         break;
 495                                 } else if (io_u->ddir == DDIR_WRITE) {
 496                                         io_u->ddir = DDIR_READ;
 497                                         break;
 498                                 } else {
 499                                         put_io_u(td, io_u);
 500                                         continue;
 501                                 }
 502                         }
 503
 504                         if (!io_u)
 505                                 break;
 506                 }
 507
 508                 if (td->o.verify_async)
 509                         io_u->end_io = verify_io_u_async;
 510                 else
 511                         io_u->end_io = verify_io_u;
 512
 513                 ddir = io_u->ddir;
 514
 515                 ret = td_io_queue(td, io_u);
 516                 switch (ret) {
 517                 case FIO_Q_COMPLETED:
 518                         if (io_u->error) {
 519                                 ret = -io_u->error;
 520                                 clear_io_u(td, io_u);
 521                         } else if (io_u->resid) {
 522                                 int bytes = io_u->xfer_buflen - io_u->resid;
 523
 524                                 /*
 525                                  * zero read, fail
 526                                  */
 527                                 if (!bytes) {
 528                                         td_verror(td, EIO, "full resid");
 529                                         put_io_u(td, io_u);
 530                                         break;
 531                                 }
 532
 533                                 io_u->xfer_buflen = io_u->resid;
 534                                 io_u->xfer_buf += bytes;
 535                                 io_u->offset += bytes;
 536
 537                                 if (ddir_rw(io_u->ddir))
 538                                         td->ts.short_io_u[io_u->ddir]++;
 539
 540                                 f = io_u->file;
 541                                 if (io_u->offset == f->real_file_size)
 542                                         goto sync_done;
 543
 544                                 requeue_io_u(td, &io_u);
 545                         } else {
 546 sync_done:
 547                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 548                                 if (ret < 0)
 549                                         break;
 550                         }
 551                         continue;
 552                 case FIO_Q_QUEUED:
 553                         break;
 554                 case FIO_Q_BUSY:
 555                         requeue_io_u(td, &io_u);
 556                         ret2 = td_io_commit(td);
 557                         if (ret2 < 0)
 558                                 ret = ret2;
 559                         break;
 560                 default:
 561                         assert(ret < 0);
 562                         td_verror(td, -ret, "td_io_queue");
 563                         break;
 564                 }
 565
 566                 if (break_on_this_error(td, ddir, &ret))
 567                         break;
 568
 569                 /*
 570                  * if we can queue more, do so. but check if there are
 571                  * completed io_u's first. Note that we can get BUSY even
 572                  * without IO queued, if the system is resource starved.
 573                  */
 574                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 575                 if (full || !td->o.iodepth_batch_complete) {
 576                         min_events = min(td->o.iodepth_batch_complete,
 577                                          td->cur_depth);
 578                         /*
 579                          * if the queue is full, we MUST reap at least 1 event
 580                          */
 581                         if (full && !min_events)
 582                                 min_events = 1;
 583
 584                         do {
 585                                 /*
 586                                  * Reap required number of io units, if any,
 587                                  * and do the verification on them through
 588                                  * the callback handler
 589                                  */
 590                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 591                                         ret = -1;
 592                                         break;
 593                                 }
 594                         } while (full && (td->cur_depth > td->o.iodepth_low));
 595                 }
 596                 if (ret < 0)
 597                         break;
 598         }
 599
 600         if (!td->error) {
 601                 min_events = td->cur_depth;
 602
 603                 if (min_events)
 604                         ret = io_u_queued_complete(td, min_events, NULL);
 605         } else
 606                 cleanup_pending_aio(td);
 607
 608         td_set_runstate(td, TD_RUNNING);
 609
 610         dprint(FD_VERIFY, "exiting loop\n");
 611 }
 612
 613 static int io_bytes_exceeded(struct thread_data *td)
 614 {
 615         unsigned long long bytes;
 616
 617         if (td_rw(td))
 618                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 619         else if (td_write(td))
 620                 bytes = td->this_io_bytes[DDIR_WRITE];
 621         else if (td_read(td))
 622                 bytes = td->this_io_bytes[DDIR_READ];
 623         else
 624                 bytes = td->this_io_bytes[DDIR_TRIM];
 625
 626         return bytes >= td->o.size;
 627 }
 628
 629 /*
 630  * Main IO worker function. It retrieves io_u's to process and queues
 631  * and reaps them, checking for rate and errors along the way.
 632  *
 633  * Returns number of bytes written and trimmed.
 634  */
 635 static uint64_t do_io(struct thread_data *td)
 636 {
 637         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 638         unsigned int i;
 639         int ret = 0;
 640
 641         if (in_ramp_time(td))
 642                 td_set_runstate(td, TD_RAMP);
 643         else
 644                 td_set_runstate(td, TD_RUNNING);
 645
 646         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 647                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 648                 td->o.time_based) {
 649                 struct timeval comp_time;
 650                 int min_evts = 0;
 651                 struct io_u *io_u;
 652                 int ret2, full;
 653                 enum fio_ddir ddir;
 654
 655                 if (td->terminate || td->done)
 656                         break;
 657
 658                 update_tv_cache(td);
 659
 660                 if (runtime_exceeded(td, &td->tv_cache)) {
 661                         __update_tv_cache(td);
 662                         if (runtime_exceeded(td, &td->tv_cache)) {
 663                                 td->terminate = 1;
 664                                 break;
 665                         }
 666                 }
 667
 668                 if (flow_threshold_exceeded(td))
 669                         continue;
 670
 671                 io_u = get_io_u(td);
 672                 if (!io_u)
 673                         break;
 674
 675                 ddir = io_u->ddir;
 676
 677                 /*
 678                  * Add verification end_io handler if:
 679                  *      - Asked to verify (!td_rw(td))
 680                  *      - Or the io_u is from our verify list (mixed write/ver)
 681                  */
 682                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 683                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 684                         if (td->o.verify_async)
 685                                 io_u->end_io = verify_io_u_async;
 686                         else
 687                                 io_u->end_io = verify_io_u;
 688                         td_set_runstate(td, TD_VERIFYING);
 689                 } else if (in_ramp_time(td))
 690                         td_set_runstate(td, TD_RAMP);
 691                 else
 692                         td_set_runstate(td, TD_RUNNING);
 693
 694                 ret = td_io_queue(td, io_u);
 695                 switch (ret) {
 696                 case FIO_Q_COMPLETED:
 697                         if (io_u->error) {
 698                                 ret = -io_u->error;
 699                                 clear_io_u(td, io_u);
 700                         } else if (io_u->resid) {
 701                                 int bytes = io_u->xfer_buflen - io_u->resid;
 702                                 struct fio_file *f = io_u->file;
 703
 704                                 /*
 705                                  * zero read, fail
 706                                  */
 707                                 if (!bytes) {
 708                                         td_verror(td, EIO, "full resid");
 709                                         put_io_u(td, io_u);
 710                                         break;
 711                                 }
 712
 713                                 io_u->xfer_buflen = io_u->resid;
 714                                 io_u->xfer_buf += bytes;
 715                                 io_u->offset += bytes;
 716
 717                                 if (ddir_rw(io_u->ddir))
 718                                         td->ts.short_io_u[io_u->ddir]++;
 719
 720                                 if (io_u->offset == f->real_file_size)
 721                                         goto sync_done;
 722
 723                                 requeue_io_u(td, &io_u);
 724                         } else {
 725 sync_done:
 726                                 if (__should_check_rate(td, DDIR_READ) ||
 727                                     __should_check_rate(td, DDIR_WRITE) ||
 728                                     __should_check_rate(td, DDIR_TRIM))
 729                                         fio_gettime(&comp_time, NULL);
 730
 731                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 732                                 if (ret < 0)
 733                                         break;
 734                         }
 735                         break;
 736                 case FIO_Q_QUEUED:
 737                         /*
 738                          * if the engine doesn't have a commit hook,
 739                          * the io_u is really queued. if it does have such
 740                          * a hook, it has to call io_u_queued() itself.
 741                          */
 742                         if (td->io_ops->commit == NULL)
 743                                 io_u_queued(td, io_u);
 744                         break;
 745                 case FIO_Q_BUSY:
 746                         requeue_io_u(td, &io_u);
 747                         ret2 = td_io_commit(td);
 748                         if (ret2 < 0)
 749                                 ret = ret2;
 750                         break;
 751                 default:
 752                         assert(ret < 0);
 753                         put_io_u(td, io_u);
 754                         break;
 755                 }
 756
 757                 if (break_on_this_error(td, ddir, &ret))
 758                         break;
 759
 760                 /*
 761                  * See if we need to complete some commands. Note that we
 762                  * can get BUSY even without IO queued, if the system is
 763                  * resource starved.
 764                  */
 765                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 766                 if (full || !td->o.iodepth_batch_complete) {
 767                         min_evts = min(td->o.iodepth_batch_complete,
 768                                         td->cur_depth);
 769                         /*
 770                          * if the queue is full, we MUST reap at least 1 event
 771                          */
 772                         if (full && !min_evts)
 773                                 min_evts = 1;
 774
 775                         if (__should_check_rate(td, DDIR_READ) ||
 776                             __should_check_rate(td, DDIR_WRITE) ||
 777                             __should_check_rate(td, DDIR_TRIM))
 778                                 fio_gettime(&comp_time, NULL);
 779
 780                         do {
 781                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 782                                 if (ret < 0)
 783                                         break;
 784
 785                         } while (full && (td->cur_depth > td->o.iodepth_low));
 786                 }
 787
 788                 if (ret < 0)
 789                         break;
 790                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 791                         continue;
 792
 793                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 794                         if (check_min_rate(td, &comp_time, bytes_done)) {
 795                                 if (exitall_on_terminate)
 796                                         fio_terminate_threads(td->groupid);
 797                                 td_verror(td, EIO, "check_min_rate");
 798                                 break;
 799                         }
 800                 }
 801
 802                 if (td->o.thinktime) {
 803                         unsigned long long b;
 804
 805                         b = ddir_rw_sum(td->io_blocks);
 806                         if (!(b % td->o.thinktime_blocks)) {
 807                                 int left;
 808
 809                                 if (td->o.thinktime_spin)
 810                                         usec_spin(td->o.thinktime_spin);
 811
 812                                 left = td->o.thinktime - td->o.thinktime_spin;
 813                                 if (left)
 814                                         usec_sleep(td, left);
 815                         }
 816                 }
 817         }
 818
 819         if (td->trim_entries)
 820                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 821
 822         if (td->o.fill_device && td->error == ENOSPC) {
 823                 td->error = 0;
 824                 td->terminate = 1;
 825         }
 826         if (!td->error) {
 827                 struct fio_file *f;
 828
 829                 i = td->cur_depth;
 830                 if (i) {
 831                         ret = io_u_queued_complete(td, i, bytes_done);
 832                         if (td->o.fill_device && td->error == ENOSPC)
 833                                 td->error = 0;
 834                 }
 835
 836                 if (should_fsync(td) && td->o.end_fsync) {
 837                         td_set_runstate(td, TD_FSYNCING);
 838
 839                         for_each_file(td, f, i) {
 840                                 if (!fio_file_fsync(td, f))
 841                                         continue;
 842
 843                                 log_err("fio: end_fsync failed for file %s\n",
 844                                                                 f->file_name);
 845                         }
 846                 }
 847         } else
 848                 cleanup_pending_aio(td);
 849
 850         /*
 851          * stop job if we failed doing any IO
 852          */
 853         if (!ddir_rw_sum(td->this_io_bytes))
 854                 td->done = 1;
 855
 856         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 857 }
 858
 859 static void cleanup_io_u(struct thread_data *td)
 860 {
 861         struct flist_head *entry, *n;
 862         struct io_u *io_u;
 863
 864         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 865                 io_u = flist_entry(entry, struct io_u, list);
 866
 867                 flist_del(&io_u->list);
 868
 869                 if (td->io_ops->io_u_free)
 870                         td->io_ops->io_u_free(td, io_u);
 871
 872                 fio_memfree(io_u, sizeof(*io_u));
 873         }
 874
 875         free_io_mem(td);
 876 }
 877
 878 static int init_io_u(struct thread_data *td)
 879 {
 880         struct io_u *io_u;
 881         unsigned int max_bs, min_write;
 882         int cl_align, i, max_units;
 883         int data_xfer = 1;
 884         char *p;
 885
 886         max_units = td->o.iodepth;
 887         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 888         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
 889         min_write = td->o.min_bs[DDIR_WRITE];
 890         td->orig_buffer_size = (unsigned long long) max_bs
 891                                         * (unsigned long long) max_units;
 892
 893         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 894                 data_xfer = 0;
 895
 896         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 897                 unsigned long bs;
 898
 899                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 900                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 901         }
 902
 903         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 904                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 905                 return 1;
 906         }
 907
 908         if (data_xfer && allocate_io_mem(td))
 909                 return 1;
 910
 911         if (td->o.odirect || td->o.mem_align ||
 912             (td->io_ops->flags & FIO_RAWIO))
 913                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 914         else
 915                 p = td->orig_buffer;
 916
 917         cl_align = os_cache_line_size();
 918
 919         for (i = 0; i < max_units; i++) {
 920                 void *ptr;
 921
 922                 if (td->terminate)
 923                         return 1;
 924
 925                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 926                 if (!ptr) {
 927                         log_err("fio: unable to allocate aligned memory\n");
 928                         break;
 929                 }
 930
 931                 io_u = ptr;
 932                 memset(io_u, 0, sizeof(*io_u));
 933                 INIT_FLIST_HEAD(&io_u->list);
 934                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 935
 936                 if (data_xfer) {
 937                         io_u->buf = p;
 938                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 939
 940                         if (td_write(td))
 941                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 942                         if (td_write(td) && td->o.verify_pattern_bytes) {
 943                                 /*
 944                                  * Fill the buffer with the pattern if we are
 945                                  * going to be doing writes.
 946                                  */
 947                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 948                         }
 949                 }
 950
 951                 io_u->index = i;
 952                 io_u->flags = IO_U_F_FREE;
 953                 flist_add(&io_u->list, &td->io_u_freelist);
 954
 955                 if (td->io_ops->io_u_init) {
 956                         int ret = td->io_ops->io_u_init(td, io_u);
 957
 958                         if (ret) {
 959                                 log_err("fio: failed to init engine data: %d\n", ret);
 960                                 return 1;
 961                         }
 962                 }
 963
 964                 p += max_bs;
 965         }
 966
 967         return 0;
 968 }
 969
 970 static int switch_ioscheduler(struct thread_data *td)
 971 {
 972         char tmp[256], tmp2[128];
 973         FILE *f;
 974         int ret;
 975
 976         if (td->io_ops->flags & FIO_DISKLESSIO)
 977                 return 0;
 978
 979         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 980
 981         f = fopen(tmp, "r+");
 982         if (!f) {
 983                 if (errno == ENOENT) {
 984                         log_err("fio: os or kernel doesn't support IO scheduler"
 985                                 " switching\n");
 986                         return 0;
 987                 }
 988                 td_verror(td, errno, "fopen iosched");
 989                 return 1;
 990         }
 991
 992         /*
 993          * Set io scheduler.
 994          */
 995         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 996         if (ferror(f) || ret != 1) {
 997                 td_verror(td, errno, "fwrite");
 998                 fclose(f);
 999                 return 1;
1000         }
1001
1002         rewind(f);
1003
1004         /*
1005          * Read back and check that the selected scheduler is now the default.
1006          */
1007         ret = fread(tmp, 1, sizeof(tmp), f);
1008         if (ferror(f) || ret < 0) {
1009                 td_verror(td, errno, "fread");
1010                 fclose(f);
1011                 return 1;
1012         }
1013
1014         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1015         if (!strstr(tmp, tmp2)) {
1016                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1017                 td_verror(td, EINVAL, "iosched_switch");
1018                 fclose(f);
1019                 return 1;
1020         }
1021
1022         fclose(f);
1023         return 0;
1024 }
1025
1026 static int keep_running(struct thread_data *td)
1027 {
1028         if (td->done)
1029                 return 0;
1030         if (td->o.time_based)
1031                 return 1;
1032         if (td->o.loops) {
1033                 td->o.loops--;
1034                 return 1;
1035         }
1036
1037         if (ddir_rw_sum(td->io_bytes) < td->o.size) {
1038                 uint64_t diff;
1039
1040                 /*
1041                  * If the difference is less than the minimum IO size, we
1042                  * are done.
1043                  */
1044                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1045                 if (diff < td->o.rw_min_bs)
1046                         return 0;
1047
1048                 return 1;
1049         }
1050
1051         return 0;
1052 }
1053
1054 static int exec_string(const char *string)
1055 {
1056         int ret, newlen = strlen(string) + 1 + 8;
1057         char *str;
1058
1059         str = malloc(newlen);
1060         sprintf(str, "sh -c %s", string);
1061
1062         ret = system(str);
1063         if (ret == -1)
1064                 log_err("fio: exec of cmd <%s> failed\n", str);
1065
1066         free(str);
1067         return ret;
1068 }
1069
1070 /*
1071  * Entry point for the thread based jobs. The process based jobs end up
1072  * here as well, after a little setup.
1073  */
1074 static void *thread_main(void *data)
1075 {
1076         unsigned long long elapsed;
1077         struct thread_data *td = data;
1078         pthread_condattr_t attr;
1079         int clear_state;
1080
1081         if (!td->o.use_thread) {
1082                 setsid();
1083                 td->pid = getpid();
1084         } else
1085                 td->pid = gettid();
1086
1087         fio_local_clock_init(td->o.use_thread);
1088
1089         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1090
1091         INIT_FLIST_HEAD(&td->io_u_freelist);
1092         INIT_FLIST_HEAD(&td->io_u_busylist);
1093         INIT_FLIST_HEAD(&td->io_u_requeues);
1094         INIT_FLIST_HEAD(&td->io_log_list);
1095         INIT_FLIST_HEAD(&td->io_hist_list);
1096         INIT_FLIST_HEAD(&td->verify_list);
1097         INIT_FLIST_HEAD(&td->trim_list);
1098         INIT_FLIST_HEAD(&td->next_rand_list);
1099         pthread_mutex_init(&td->io_u_lock, NULL);
1100         td->io_hist_tree = RB_ROOT;
1101
1102         pthread_condattr_init(&attr);
1103         pthread_cond_init(&td->verify_cond, &attr);
1104         pthread_cond_init(&td->free_cond, &attr);
1105
1106         td_set_runstate(td, TD_INITIALIZED);
1107         dprint(FD_MUTEX, "up startup_mutex\n");
1108         fio_mutex_up(startup_mutex);
1109         dprint(FD_MUTEX, "wait on td->mutex\n");
1110         fio_mutex_down(td->mutex);
1111         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1112
1113         /*
1114          * the ->mutex mutex is now no longer used, close it to avoid
1115          * eating a file descriptor
1116          */
1117         fio_mutex_remove(td->mutex);
1118
1119         /*
1120          * A new gid requires privilege, so we need to do this before setting
1121          * the uid.
1122          */
1123         if (td->o.gid != -1U && setgid(td->o.gid)) {
1124                 td_verror(td, errno, "setgid");
1125                 goto err;
1126         }
1127         if (td->o.uid != -1U && setuid(td->o.uid)) {
1128                 td_verror(td, errno, "setuid");
1129                 goto err;
1130         }
1131
1132         /*
1133          * If we have a gettimeofday() thread, make sure we exclude that
1134          * thread from this job
1135          */
1136         if (td->o.gtod_cpu)
1137                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1138
1139         /*
1140          * Set affinity first, in case it has an impact on the memory
1141          * allocations.
1142          */
1143         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1144                 td_verror(td, errno, "cpu_set_affinity");
1145                 goto err;
1146         }
1147
1148 #ifdef CONFIG_LIBNUMA
1149         /* numa node setup */
1150         if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1151                 int ret;
1152
1153                 if (numa_available() < 0) {
1154                         td_verror(td, errno, "Does not support NUMA API\n");
1155                         goto err;
1156                 }
1157
1158                 if (td->o.numa_cpumask_set) {
1159                         ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1160                         if (ret == -1) {
1161                                 td_verror(td, errno, \
1162                                         "numa_run_on_node_mask failed\n");
1163                                 goto err;
1164                         }
1165                 }
1166
1167                 if (td->o.numa_memmask_set) {
1168
1169                         switch (td->o.numa_mem_mode) {
1170                         case MPOL_INTERLEAVE:
1171                                 numa_set_interleave_mask(td->o.numa_memnodesmask);
1172                                 break;
1173                         case MPOL_BIND:
1174                                 numa_set_membind(td->o.numa_memnodesmask);
1175                                 break;
1176                         case MPOL_LOCAL:
1177                                 numa_set_localalloc();
1178                                 break;
1179                         case MPOL_PREFERRED:
1180                                 numa_set_preferred(td->o.numa_mem_prefer_node);
1181                                 break;
1182                         case MPOL_DEFAULT:
1183                         default:
1184                                 break;
1185                         }
1186
1187                 }
1188         }
1189 #endif
1190
1191         /*
1192          * May alter parameters that init_io_u() will use, so we need to
1193          * do this first.
1194          */
1195         if (init_iolog(td))
1196                 goto err;
1197
1198         if (init_io_u(td))
1199                 goto err;
1200
1201         if (td->o.verify_async && verify_async_init(td))
1202                 goto err;
1203
1204         if (td->ioprio_set) {
1205                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1206                         td_verror(td, errno, "ioprio_set");
1207                         goto err;
1208                 }
1209         }
1210
1211         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1212                 goto err;
1213
1214         errno = 0;
1215         if (nice(td->o.nice) == -1 && errno != 0) {
1216                 td_verror(td, errno, "nice");
1217                 goto err;
1218         }
1219
1220         if (td->o.ioscheduler && switch_ioscheduler(td))
1221                 goto err;
1222
1223         if (!td->o.create_serialize && setup_files(td))
1224                 goto err;
1225
1226         if (td_io_init(td))
1227                 goto err;
1228
1229         if (init_random_map(td))
1230                 goto err;
1231
1232         if (td->o.exec_prerun) {
1233                 if (exec_string(td->o.exec_prerun))
1234                         goto err;
1235         }
1236
1237         if (td->o.pre_read) {
1238                 if (pre_read_files(td) < 0)
1239                         goto err;
1240         }
1241
1242         fio_gettime(&td->epoch, NULL);
1243         fio_getrusage(&td->ru_start);
1244         clear_state = 0;
1245         while (keep_running(td)) {
1246                 uint64_t verify_bytes;
1247
1248                 fio_gettime(&td->start, NULL);
1249                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1250                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1251                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1252
1253                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1254                                 td->o.ratemin[DDIR_TRIM]) {
1255                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1256                                                 sizeof(td->bw_sample_time));
1257                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1258                                                 sizeof(td->bw_sample_time));
1259                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1260                                                 sizeof(td->bw_sample_time));
1261                 }
1262
1263                 if (clear_state)
1264                         clear_io_state(td);
1265
1266                 prune_io_piece_log(td);
1267
1268                 verify_bytes = do_io(td);
1269
1270                 clear_state = 1;
1271
1272                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1273                         elapsed = utime_since_now(&td->start);
1274                         td->ts.runtime[DDIR_READ] += elapsed;
1275                 }
1276                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1277                         elapsed = utime_since_now(&td->start);
1278                         td->ts.runtime[DDIR_WRITE] += elapsed;
1279                 }
1280                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1281                         elapsed = utime_since_now(&td->start);
1282                         td->ts.runtime[DDIR_TRIM] += elapsed;
1283                 }
1284
1285                 if (td->error || td->terminate)
1286                         break;
1287
1288                 if (!td->o.do_verify ||
1289                     td->o.verify == VERIFY_NONE ||
1290                     (td->io_ops->flags & FIO_UNIDIR))
1291                         continue;
1292
1293                 clear_io_state(td);
1294
1295                 fio_gettime(&td->start, NULL);
1296
1297                 do_verify(td, verify_bytes);
1298
1299                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1300
1301                 if (td->error || td->terminate)
1302                         break;
1303         }
1304
1305         update_rusage_stat(td);
1306         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1307         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1308         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1309         td->ts.total_run_time = mtime_since_now(&td->epoch);
1310         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1311         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1312         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1313
1314         fio_mutex_down(writeout_mutex);
1315         if (td->bw_log) {
1316                 if (td->o.bw_log_file) {
1317                         finish_log_named(td, td->bw_log,
1318                                                 td->o.bw_log_file, "bw");
1319                 } else
1320                         finish_log(td, td->bw_log, "bw");
1321         }
1322         if (td->lat_log) {
1323                 if (td->o.lat_log_file) {
1324                         finish_log_named(td, td->lat_log,
1325                                                 td->o.lat_log_file, "lat");
1326                 } else
1327                         finish_log(td, td->lat_log, "lat");
1328         }
1329         if (td->slat_log) {
1330                 if (td->o.lat_log_file) {
1331                         finish_log_named(td, td->slat_log,
1332                                                 td->o.lat_log_file, "slat");
1333                 } else
1334                         finish_log(td, td->slat_log, "slat");
1335         }
1336         if (td->clat_log) {
1337                 if (td->o.lat_log_file) {
1338                         finish_log_named(td, td->clat_log,
1339                                                 td->o.lat_log_file, "clat");
1340                 } else
1341                         finish_log(td, td->clat_log, "clat");
1342         }
1343         if (td->iops_log) {
1344                 if (td->o.iops_log_file) {
1345                         finish_log_named(td, td->iops_log,
1346                                                 td->o.iops_log_file, "iops");
1347                 } else
1348                         finish_log(td, td->iops_log, "iops");
1349         }
1350
1351         fio_mutex_up(writeout_mutex);
1352         if (td->o.exec_postrun)
1353                 exec_string(td->o.exec_postrun);
1354
1355         if (exitall_on_terminate)
1356                 fio_terminate_threads(td->groupid);
1357
1358 err:
1359         if (td->error)
1360                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1361                                                         td->verror);
1362
1363         if (td->o.verify_async)
1364                 verify_async_exit(td);
1365
1366         close_and_free_files(td);
1367         cleanup_io_u(td);
1368         close_ioengine(td);
1369         cgroup_shutdown(td, &cgroup_mnt);
1370
1371         if (td->o.cpumask_set) {
1372                 int ret = fio_cpuset_exit(&td->o.cpumask);
1373
1374                 td_verror(td, ret, "fio_cpuset_exit");
1375         }
1376
1377         /*
1378          * do this very late, it will log file closing as well
1379          */
1380         if (td->o.write_iolog_file)
1381                 write_iolog_close(td);
1382
1383         td_set_runstate(td, TD_EXITED);
1384         return (void *) (uintptr_t) td->error;
1385 }
1386
1387
1388 /*
1389  * We cannot pass the td data into a forked process, so attach the td and
1390  * pass it to the thread worker.
1391  */
1392 static int fork_main(int shmid, int offset)
1393 {
1394         struct thread_data *td;
1395         void *data, *ret;
1396
1397 #ifndef __hpux
1398         data = shmat(shmid, NULL, 0);
1399         if (data == (void *) -1) {
1400                 int __err = errno;
1401
1402                 perror("shmat");
1403                 return __err;
1404         }
1405 #else
1406         /*
1407          * HP-UX inherits shm mappings?
1408          */
1409         data = threads;
1410 #endif
1411
1412         td = data + offset * sizeof(struct thread_data);
1413         ret = thread_main(td);
1414         shmdt(data);
1415         return (int) (uintptr_t) ret;
1416 }
1417
1418 /*
1419  * Run over the job map and reap the threads that have exited, if any.
1420  */
1421 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1422                          unsigned int *m_rate)
1423 {
1424         struct thread_data *td;
1425         unsigned int cputhreads, realthreads, pending;
1426         int i, status, ret;
1427
1428         /*
1429          * reap exited threads (TD_EXITED -> TD_REAPED)
1430          */
1431         realthreads = pending = cputhreads = 0;
1432         for_each_td(td, i) {
1433                 int flags = 0;
1434
1435                 /*
1436                  * ->io_ops is NULL for a thread that has closed its
1437                  * io engine
1438                  */
1439                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1440                         cputhreads++;
1441                 else
1442                         realthreads++;
1443
1444                 if (!td->pid) {
1445                         pending++;
1446                         continue;
1447                 }
1448                 if (td->runstate == TD_REAPED)
1449                         continue;
1450                 if (td->o.use_thread) {
1451                         if (td->runstate == TD_EXITED) {
1452                                 td_set_runstate(td, TD_REAPED);
1453                                 goto reaped;
1454                         }
1455                         continue;
1456                 }
1457
1458                 flags = WNOHANG;
1459                 if (td->runstate == TD_EXITED)
1460                         flags = 0;
1461
1462                 /*
1463                  * check if someone quit or got killed in an unusual way
1464                  */
1465                 ret = waitpid(td->pid, &status, flags);
1466                 if (ret < 0) {
1467                         if (errno == ECHILD) {
1468                                 log_err("fio: pid=%d disappeared %d\n",
1469                                                 (int) td->pid, td->runstate);
1470                                 td->sig = ECHILD;
1471                                 td_set_runstate(td, TD_REAPED);
1472                                 goto reaped;
1473                         }
1474                         perror("waitpid");
1475                 } else if (ret == td->pid) {
1476                         if (WIFSIGNALED(status)) {
1477                                 int sig = WTERMSIG(status);
1478
1479                                 if (sig != SIGTERM && sig != SIGUSR2)
1480                                         log_err("fio: pid=%d, got signal=%d\n",
1481                                                         (int) td->pid, sig);
1482                                 td->sig = sig;
1483                                 td_set_runstate(td, TD_REAPED);
1484                                 goto reaped;
1485                         }
1486                         if (WIFEXITED(status)) {
1487                                 if (WEXITSTATUS(status) && !td->error)
1488                                         td->error = WEXITSTATUS(status);
1489
1490                                 td_set_runstate(td, TD_REAPED);
1491                                 goto reaped;
1492                         }
1493                 }
1494
1495                 /*
1496                  * thread is not dead, continue
1497                  */
1498                 pending++;
1499                 continue;
1500 reaped:
1501                 (*nr_running)--;
1502                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1503                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1504                 if (!td->pid)
1505                         pending--;
1506
1507                 if (td->error)
1508                         exit_value++;
1509
1510                 done_secs += mtime_since_now(&td->epoch) / 1000;
1511         }
1512
1513         if (*nr_running == cputhreads && !pending && realthreads)
1514                 fio_terminate_threads(TERMINATE_ALL);
1515 }
1516
1517 /*
1518  * Main function for kicking off and reaping jobs, as needed.
1519  */
1520 static void run_threads(void)
1521 {
1522         struct thread_data *td;
1523         unsigned long spent;
1524         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1525
1526         if (fio_pin_memory())
1527                 return;
1528
1529         if (fio_gtod_offload && fio_start_gtod_thread())
1530                 return;
1531
1532         fio_idle_prof_init();
1533
1534         set_sig_handlers();
1535
1536         if (output_format == FIO_OUTPUT_NORMAL) {
1537                 log_info("Starting ");
1538                 if (nr_thread)
1539                         log_info("%d thread%s", nr_thread,
1540                                                 nr_thread > 1 ? "s" : "");
1541                 if (nr_process) {
1542                         if (nr_thread)
1543                                 log_info(" and ");
1544                         log_info("%d process%s", nr_process,
1545                                                 nr_process > 1 ? "es" : "");
1546                 }
1547                 log_info("\n");
1548                 fflush(stdout);
1549         }
1550
1551         todo = thread_number;
1552         nr_running = 0;
1553         nr_started = 0;
1554         m_rate = t_rate = 0;
1555
1556         for_each_td(td, i) {
1557                 print_status_init(td->thread_number - 1);
1558
1559                 if (!td->o.create_serialize)
1560                         continue;
1561
1562                 /*
1563                  * do file setup here so it happens sequentially,
1564                  * we don't want X number of threads getting their
1565                  * client data interspersed on disk
1566                  */
1567                 if (setup_files(td)) {
1568                         exit_value++;
1569                         if (td->error)
1570                                 log_err("fio: pid=%d, err=%d/%s\n",
1571                                         (int) td->pid, td->error, td->verror);
1572                         td_set_runstate(td, TD_REAPED);
1573                         todo--;
1574                 } else {
1575                         struct fio_file *f;
1576                         unsigned int j;
1577
1578                         /*
1579                          * for sharing to work, each job must always open
1580                          * its own files. so close them, if we opened them
1581                          * for creation
1582                          */
1583                         for_each_file(td, f, j) {
1584                                 if (fio_file_open(f))
1585                                         td_io_close_file(td, f);
1586                         }
1587                 }
1588         }
1589
1590         /* start idle threads before io threads start to run */
1591         fio_idle_prof_start();
1592
1593         set_genesis_time();
1594
1595         while (todo) {
1596                 struct thread_data *map[REAL_MAX_JOBS];
1597                 struct timeval this_start;
1598                 int this_jobs = 0, left;
1599
1600                 /*
1601                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1602                  */
1603                 for_each_td(td, i) {
1604                         if (td->runstate != TD_NOT_CREATED)
1605                                 continue;
1606
1607                         /*
1608                          * never got a chance to start, killed by other
1609                          * thread for some reason
1610                          */
1611                         if (td->terminate) {
1612                                 todo--;
1613                                 continue;
1614                         }
1615
1616                         if (td->o.start_delay) {
1617                                 spent = mtime_since_genesis();
1618
1619                                 if (td->o.start_delay * 1000 > spent)
1620                                         continue;
1621                         }
1622
1623                         if (td->o.stonewall && (nr_started || nr_running)) {
1624                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1625                                                         td->o.name);
1626                                 break;
1627                         }
1628
1629                         init_disk_util(td);
1630
1631                         /*
1632                          * Set state to created. Thread will transition
1633                          * to TD_INITIALIZED when it's done setting up.
1634                          */
1635                         td_set_runstate(td, TD_CREATED);
1636                         map[this_jobs++] = td;
1637                         nr_started++;
1638
1639                         if (td->o.use_thread) {
1640                                 int ret;
1641
1642                                 dprint(FD_PROCESS, "will pthread_create\n");
1643                                 ret = pthread_create(&td->thread, NULL,
1644                                                         thread_main, td);
1645                                 if (ret) {
1646                                         log_err("pthread_create: %s\n",
1647                                                         strerror(ret));
1648                                         nr_started--;
1649                                         break;
1650                                 }
1651                                 ret = pthread_detach(td->thread);
1652                                 if (ret)
1653                                         log_err("pthread_detach: %s",
1654                                                         strerror(ret));
1655                         } else {
1656                                 pid_t pid;
1657                                 dprint(FD_PROCESS, "will fork\n");
1658                                 pid = fork();
1659                                 if (!pid) {
1660                                         int ret = fork_main(shm_id, i);
1661
1662                                         _exit(ret);
1663                                 } else if (i == fio_debug_jobno)
1664                                         *fio_debug_jobp = pid;
1665                         }
1666                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1667                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1668                                 log_err("fio: job startup hung? exiting.\n");
1669                                 fio_terminate_threads(TERMINATE_ALL);
1670                                 fio_abort = 1;
1671                                 nr_started--;
1672                                 break;
1673                         }
1674                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1675                 }
1676
1677                 /*
1678                  * Wait for the started threads to transition to
1679                  * TD_INITIALIZED.
1680                  */
1681                 fio_gettime(&this_start, NULL);
1682                 left = this_jobs;
1683                 while (left && !fio_abort) {
1684                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1685                                 break;
1686
1687                         usleep(100000);
1688
1689                         for (i = 0; i < this_jobs; i++) {
1690                                 td = map[i];
1691                                 if (!td)
1692                                         continue;
1693                                 if (td->runstate == TD_INITIALIZED) {
1694                                         map[i] = NULL;
1695                                         left--;
1696                                 } else if (td->runstate >= TD_EXITED) {
1697                                         map[i] = NULL;
1698                                         left--;
1699                                         todo--;
1700                                         nr_running++; /* work-around... */
1701                                 }
1702                         }
1703                 }
1704
1705                 if (left) {
1706                         log_err("fio: %d job%s failed to start\n", left,
1707                                         left > 1 ? "s" : "");
1708                         for (i = 0; i < this_jobs; i++) {
1709                                 td = map[i];
1710                                 if (!td)
1711                                         continue;
1712                                 kill(td->pid, SIGTERM);
1713                         }
1714                         break;
1715                 }
1716
1717                 /*
1718                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1719                  */
1720                 for_each_td(td, i) {
1721                         if (td->runstate != TD_INITIALIZED)
1722                                 continue;
1723
1724                         if (in_ramp_time(td))
1725                                 td_set_runstate(td, TD_RAMP);
1726                         else
1727                                 td_set_runstate(td, TD_RUNNING);
1728                         nr_running++;
1729                         nr_started--;
1730                         m_rate += ddir_rw_sum(td->o.ratemin);
1731                         t_rate += ddir_rw_sum(td->o.rate);
1732                         todo--;
1733                         fio_mutex_up(td->mutex);
1734                 }
1735
1736                 reap_threads(&nr_running, &t_rate, &m_rate);
1737
1738                 if (todo) {
1739                         if (is_backend)
1740                                 fio_server_idle_loop();
1741                         else
1742                                 usleep(100000);
1743                 }
1744         }
1745
1746         while (nr_running) {
1747                 reap_threads(&nr_running, &t_rate, &m_rate);
1748
1749                 if (is_backend)
1750                         fio_server_idle_loop();
1751                 else
1752                         usleep(10000);
1753         }
1754
1755         fio_idle_prof_stop();
1756
1757         update_io_ticks();
1758         fio_unpin_memory();
1759 }
1760
1761 void wait_for_disk_thread_exit(void)
1762 {
1763         fio_mutex_down(disk_thread_mutex);
1764 }
1765
1766 static void free_disk_util(void)
1767 {
1768         disk_util_start_exit();
1769         wait_for_disk_thread_exit();
1770         disk_util_prune_entries();
1771 }
1772
1773 static void *disk_thread_main(void *data)
1774 {
1775         int ret = 0;
1776
1777         fio_mutex_up(startup_mutex);
1778
1779         while (threads && !ret) {
1780                 usleep(DISK_UTIL_MSEC * 1000);
1781                 if (!threads)
1782                         break;
1783                 ret = update_io_ticks();
1784
1785                 if (!is_backend)
1786                         print_thread_status();
1787         }
1788
1789         fio_mutex_up(disk_thread_mutex);
1790         return NULL;
1791 }
1792
1793 static int create_disk_util_thread(void)
1794 {
1795         int ret;
1796
1797         setup_disk_util();
1798
1799         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1800
1801         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1802         if (ret) {
1803                 fio_mutex_remove(disk_thread_mutex);
1804                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1805                 return 1;
1806         }
1807
1808         ret = pthread_detach(disk_util_thread);
1809         if (ret) {
1810                 fio_mutex_remove(disk_thread_mutex);
1811                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1812                 return 1;
1813         }
1814
1815         dprint(FD_MUTEX, "wait on startup_mutex\n");
1816         fio_mutex_down(startup_mutex);
1817         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1818         return 0;
1819 }
1820
1821 int fio_backend(void)
1822 {
1823         struct thread_data *td;
1824         int i;
1825
1826         if (exec_profile) {
1827                 if (load_profile(exec_profile))
1828                         return 1;
1829                 free(exec_profile);
1830                 exec_profile = NULL;
1831         }
1832         if (!thread_number)
1833                 return 0;
1834
1835         if (write_bw_log) {
1836                 setup_log(&agg_io_log[DDIR_READ], 0);
1837                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1838                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1839         }
1840
1841         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1842         if (startup_mutex == NULL)
1843                 return 1;
1844         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1845         if (writeout_mutex == NULL)
1846                 return 1;
1847
1848         set_genesis_time();
1849         create_disk_util_thread();
1850
1851         cgroup_list = smalloc(sizeof(*cgroup_list));
1852         INIT_FLIST_HEAD(cgroup_list);
1853
1854         run_threads();
1855
1856         if (!fio_abort) {
1857                 show_run_stats();
1858                 if (write_bw_log) {
1859                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1860                         __finish_log(agg_io_log[DDIR_WRITE],
1861                                         "agg-write_bw.log");
1862                         __finish_log(agg_io_log[DDIR_TRIM],
1863                                         "agg-write_bw.log");
1864                 }
1865         }
1866
1867         for_each_td(td, i)
1868                 fio_options_free(td);
1869
1870         free_disk_util();
1871         cgroup_kill(cgroup_list);
1872         sfree(cgroup_list);
1873         sfree(cgroup_mnt);
1874
1875         fio_mutex_remove(startup_mutex);
1876         fio_mutex_remove(writeout_mutex);
1877         fio_mutex_remove(disk_thread_mutex);
1878         return exit_value;
1879 }