backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/shm.h>
  38 #include <sys/mman.h>
  39
  40 #include "fio.h"
  41 #include "hash.h"
  42 #include "smalloc.h"
  43 #include "verify.h"
  44 #include "trim.h"
  45 #include "diskutil.h"
  46 #include "cgroup.h"
  47 #include "profile.h"
  48 #include "lib/rand.h"
  49 #include "memalign.h"
  50 #include "server.h"
  51
  52 static pthread_t disk_util_thread;
  53 static struct fio_mutex *disk_thread_mutex;
  54 static struct fio_mutex *startup_mutex;
  55 static struct fio_mutex *writeout_mutex;
  56 static struct flist_head *cgroup_list;
  57 static char *cgroup_mnt;
  58 static int exit_value;
  59 static volatile int fio_abort;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int nr_process = 0;
  66 unsigned int nr_thread = 0;
  67 int shm_id = 0;
  68 int temp_stall_ts;
  69 unsigned long done_secs = 0;
  70 volatile int disk_util_exit = 0;
  71
  72 #define PAGE_ALIGN(buf) \
  73         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  74
  75 #define JOB_START_TIMEOUT       (5 * 1000)
  76
  77 static void sig_int(int sig)
  78 {
  79         if (threads) {
  80                 if (is_backend)
  81                         fio_server_got_signal(sig);
  82                 else {
  83                         log_info("\nfio: terminating on signal %d\n", sig);
  84                         fflush(stdout);
  85                         exit_value = 128;
  86                 }
  87
  88                 fio_terminate_threads(TERMINATE_ALL);
  89         }
  90 }
  91
  92 static void sig_show_status(int sig)
  93 {
  94         show_running_run_stats();
  95 }
  96
  97 static void set_sig_handlers(void)
  98 {
  99         struct sigaction act;
 100
 101         memset(&act, 0, sizeof(act));
 102         act.sa_handler = sig_int;
 103         act.sa_flags = SA_RESTART;
 104         sigaction(SIGINT, &act, NULL);
 105
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGTERM, &act, NULL);
 110
 111         memset(&act, 0, sizeof(act));
 112         act.sa_handler = sig_show_status;
 113         act.sa_flags = SA_RESTART;
 114         sigaction(SIGUSR1, &act, NULL);
 115
 116         if (is_backend) {
 117                 memset(&act, 0, sizeof(act));
 118                 act.sa_handler = sig_int;
 119                 act.sa_flags = SA_RESTART;
 120                 sigaction(SIGPIPE, &act, NULL);
 121         }
 122 }
 123
 124 /*
 125  * Check if we are above the minimum rate given.
 126  */
 127 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 128                             enum fio_ddir ddir)
 129 {
 130         unsigned long long bytes = 0;
 131         unsigned long iops = 0;
 132         unsigned long spent;
 133         unsigned long rate;
 134         unsigned int ratemin = 0;
 135         unsigned int rate_iops = 0;
 136         unsigned int rate_iops_min = 0;
 137
 138         assert(ddir_rw(ddir));
 139
 140         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 141                 return 0;
 142
 143         /*
 144          * allow a 2 second settle period in the beginning
 145          */
 146         if (mtime_since(&td->start, now) < 2000)
 147                 return 0;
 148
 149         iops += td->this_io_blocks[ddir];
 150         bytes += td->this_io_bytes[ddir];
 151         ratemin += td->o.ratemin[ddir];
 152         rate_iops += td->o.rate_iops[ddir];
 153         rate_iops_min += td->o.rate_iops_min[ddir];
 154
 155         /*
 156          * if rate blocks is set, sample is running
 157          */
 158         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 159                 spent = mtime_since(&td->lastrate[ddir], now);
 160                 if (spent < td->o.ratecycle)
 161                         return 0;
 162
 163                 if (td->o.rate[ddir]) {
 164                         /*
 165                          * check bandwidth specified rate
 166                          */
 167                         if (bytes < td->rate_bytes[ddir]) {
 168                                 log_err("%s: min rate %u not met\n", td->o.name,
 169                                                                 ratemin);
 170                                 return 1;
 171                         } else {
 172                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 173                                 if (rate < ratemin ||
 174                                     bytes < td->rate_bytes[ddir]) {
 175                                         log_err("%s: min rate %u not met, got"
 176                                                 " %luKB/sec\n", td->o.name,
 177                                                         ratemin, rate);
 178                                         return 1;
 179                                 }
 180                         }
 181                 } else {
 182                         /*
 183                          * checks iops specified rate
 184                          */
 185                         if (iops < rate_iops) {
 186                                 log_err("%s: min iops rate %u not met\n",
 187                                                 td->o.name, rate_iops);
 188                                 return 1;
 189                         } else {
 190                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 191                                 if (rate < rate_iops_min ||
 192                                     iops < td->rate_blocks[ddir]) {
 193                                         log_err("%s: min iops rate %u not met,"
 194                                                 " got %lu\n", td->o.name,
 195                                                         rate_iops_min, rate);
 196                                 }
 197                         }
 198                 }
 199         }
 200
 201         td->rate_bytes[ddir] = bytes;
 202         td->rate_blocks[ddir] = iops;
 203         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 204         return 0;
 205 }
 206
 207 static int check_min_rate(struct thread_data *td, struct timeval *now,
 208                           unsigned long *bytes_done)
 209 {
 210         int ret = 0;
 211
 212         if (bytes_done[DDIR_READ])
 213                 ret |= __check_min_rate(td, now, DDIR_READ);
 214         if (bytes_done[DDIR_WRITE])
 215                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 216         if (bytes_done[DDIR_TRIM])
 217                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 218
 219         return ret;
 220 }
 221
 222 /*
 223  * When job exits, we can cancel the in-flight IO if we are using async
 224  * io. Attempt to do so.
 225  */
 226 static void cleanup_pending_aio(struct thread_data *td)
 227 {
 228         struct flist_head *entry, *n;
 229         struct io_u *io_u;
 230         int r;
 231
 232         /*
 233          * get immediately available events, if any
 234          */
 235         r = io_u_queued_complete(td, 0, NULL);
 236         if (r < 0)
 237                 return;
 238
 239         /*
 240          * now cancel remaining active events
 241          */
 242         if (td->io_ops->cancel) {
 243                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 244                         io_u = flist_entry(entry, struct io_u, list);
 245
 246                         /*
 247                          * if the io_u isn't in flight, then that generally
 248                          * means someone leaked an io_u. complain but fix
 249                          * it up, so we don't stall here.
 250                          */
 251                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 252                                 log_err("fio: non-busy IO on busy list\n");
 253                                 put_io_u(td, io_u);
 254                         } else {
 255                                 r = td->io_ops->cancel(td, io_u);
 256                                 if (!r)
 257                                         put_io_u(td, io_u);
 258                         }
 259                 }
 260         }
 261
 262         if (td->cur_depth)
 263                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 264 }
 265
 266 /*
 267  * Helper to handle the final sync of a file. Works just like the normal
 268  * io path, just does everything sync.
 269  */
 270 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 271 {
 272         struct io_u *io_u = __get_io_u(td);
 273         int ret;
 274
 275         if (!io_u)
 276                 return 1;
 277
 278         io_u->ddir = DDIR_SYNC;
 279         io_u->file = f;
 280
 281         if (td_io_prep(td, io_u)) {
 282                 put_io_u(td, io_u);
 283                 return 1;
 284         }
 285
 286 requeue:
 287         ret = td_io_queue(td, io_u);
 288         if (ret < 0) {
 289                 td_verror(td, io_u->error, "td_io_queue");
 290                 put_io_u(td, io_u);
 291                 return 1;
 292         } else if (ret == FIO_Q_QUEUED) {
 293                 if (io_u_queued_complete(td, 1, NULL) < 0)
 294                         return 1;
 295         } else if (ret == FIO_Q_COMPLETED) {
 296                 if (io_u->error) {
 297                         td_verror(td, io_u->error, "td_io_queue");
 298                         return 1;
 299                 }
 300
 301                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 302                         return 1;
 303         } else if (ret == FIO_Q_BUSY) {
 304                 if (td_io_commit(td))
 305                         return 1;
 306                 goto requeue;
 307         }
 308
 309         return 0;
 310 }
 311
 312 static inline void __update_tv_cache(struct thread_data *td)
 313 {
 314         fio_gettime(&td->tv_cache, NULL);
 315 }
 316
 317 static inline void update_tv_cache(struct thread_data *td)
 318 {
 319         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 320                 __update_tv_cache(td);
 321 }
 322
 323 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 324 {
 325         if (in_ramp_time(td))
 326                 return 0;
 327         if (!td->o.timeout)
 328                 return 0;
 329         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 330                 return 1;
 331
 332         return 0;
 333 }
 334
 335 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 336                                int *retptr)
 337 {
 338         int ret = *retptr;
 339
 340         if (ret < 0 || td->error) {
 341                 int err = td->error;
 342                 enum error_type_bit eb;
 343
 344                 if (ret < 0)
 345                         err = -ret;
 346
 347                 eb = td_error_type(ddir, err);
 348                 if (!(td->o.continue_on_error & (1 << eb)))
 349                         return 1;
 350
 351                 if (td_non_fatal_error(td, eb, err)) {
 352                         /*
 353                          * Continue with the I/Os in case of
 354                          * a non fatal error.
 355                          */
 356                         update_error_count(td, err);
 357                         td_clear_error(td);
 358                         *retptr = 0;
 359                         return 0;
 360                 } else if (td->o.fill_device && err == ENOSPC) {
 361                         /*
 362                          * We expect to hit this error if
 363                          * fill_device option is set.
 364                          */
 365                         td_clear_error(td);
 366                         td->terminate = 1;
 367                         return 1;
 368                 } else {
 369                         /*
 370                          * Stop the I/O in case of a fatal
 371                          * error.
 372                          */
 373                         update_error_count(td, err);
 374                         return 1;
 375                 }
 376         }
 377
 378         return 0;
 379 }
 380
 381 /*
 382  * The main verify engine. Runs over the writes we previously submitted,
 383  * reads the blocks back in, and checks the crc/md5 of the data.
 384  */
 385 static void do_verify(struct thread_data *td)
 386 {
 387         struct fio_file *f;
 388         struct io_u *io_u;
 389         int ret, min_events;
 390         unsigned int i;
 391
 392         dprint(FD_VERIFY, "starting loop\n");
 393
 394         /*
 395          * sync io first and invalidate cache, to make sure we really
 396          * read from disk.
 397          */
 398         for_each_file(td, f, i) {
 399                 if (!fio_file_open(f))
 400                         continue;
 401                 if (fio_io_sync(td, f))
 402                         break;
 403                 if (file_invalidate_cache(td, f))
 404                         break;
 405         }
 406
 407         if (td->error)
 408                 return;
 409
 410         td_set_runstate(td, TD_VERIFYING);
 411
 412         io_u = NULL;
 413         while (!td->terminate) {
 414                 int ret2, full;
 415
 416                 update_tv_cache(td);
 417
 418                 if (runtime_exceeded(td, &td->tv_cache)) {
 419                         __update_tv_cache(td);
 420                         if (runtime_exceeded(td, &td->tv_cache)) {
 421                                 td->terminate = 1;
 422                                 break;
 423                         }
 424                 }
 425
 426                 if (flow_threshold_exceeded(td))
 427                         continue;
 428
 429                 io_u = __get_io_u(td);
 430                 if (!io_u)
 431                         break;
 432
 433                 if (get_next_verify(td, io_u)) {
 434                         put_io_u(td, io_u);
 435                         break;
 436                 }
 437
 438                 if (td_io_prep(td, io_u)) {
 439                         put_io_u(td, io_u);
 440                         break;
 441                 }
 442
 443                 if (td->o.verify_async)
 444                         io_u->end_io = verify_io_u_async;
 445                 else
 446                         io_u->end_io = verify_io_u;
 447
 448                 ret = td_io_queue(td, io_u);
 449                 switch (ret) {
 450                 case FIO_Q_COMPLETED:
 451                         if (io_u->error) {
 452                                 ret = -io_u->error;
 453                                 clear_io_u(td, io_u);
 454                         } else if (io_u->resid) {
 455                                 int bytes = io_u->xfer_buflen - io_u->resid;
 456
 457                                 /*
 458                                  * zero read, fail
 459                                  */
 460                                 if (!bytes) {
 461                                         td_verror(td, EIO, "full resid");
 462                                         put_io_u(td, io_u);
 463                                         break;
 464                                 }
 465
 466                                 io_u->xfer_buflen = io_u->resid;
 467                                 io_u->xfer_buf += bytes;
 468                                 io_u->offset += bytes;
 469
 470                                 if (ddir_rw(io_u->ddir))
 471                                         td->ts.short_io_u[io_u->ddir]++;
 472
 473                                 f = io_u->file;
 474                                 if (io_u->offset == f->real_file_size)
 475                                         goto sync_done;
 476
 477                                 requeue_io_u(td, &io_u);
 478                         } else {
 479 sync_done:
 480                                 ret = io_u_sync_complete(td, io_u, NULL);
 481                                 if (ret < 0)
 482                                         break;
 483                         }
 484                         continue;
 485                 case FIO_Q_QUEUED:
 486                         break;
 487                 case FIO_Q_BUSY:
 488                         requeue_io_u(td, &io_u);
 489                         ret2 = td_io_commit(td);
 490                         if (ret2 < 0)
 491                                 ret = ret2;
 492                         break;
 493                 default:
 494                         assert(ret < 0);
 495                         td_verror(td, -ret, "td_io_queue");
 496                         break;
 497                 }
 498
 499                 if (break_on_this_error(td, io_u->ddir, &ret))
 500                         break;
 501
 502                 /*
 503                  * if we can queue more, do so. but check if there are
 504                  * completed io_u's first. Note that we can get BUSY even
 505                  * without IO queued, if the system is resource starved.
 506                  */
 507                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 508                 if (full || !td->o.iodepth_batch_complete) {
 509                         min_events = min(td->o.iodepth_batch_complete,
 510                                          td->cur_depth);
 511                         /*
 512                          * if the queue is full, we MUST reap at least 1 event
 513                          */
 514                         if (full && !min_events)
 515                                 min_events = 1;
 516
 517                         do {
 518                                 /*
 519                                  * Reap required number of io units, if any,
 520                                  * and do the verification on them through
 521                                  * the callback handler
 522                                  */
 523                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 524                                         ret = -1;
 525                                         break;
 526                                 }
 527                         } while (full && (td->cur_depth > td->o.iodepth_low));
 528                 }
 529                 if (ret < 0)
 530                         break;
 531         }
 532
 533         if (!td->error) {
 534                 min_events = td->cur_depth;
 535
 536                 if (min_events)
 537                         ret = io_u_queued_complete(td, min_events, NULL);
 538         } else
 539                 cleanup_pending_aio(td);
 540
 541         td_set_runstate(td, TD_RUNNING);
 542
 543         dprint(FD_VERIFY, "exiting loop\n");
 544 }
 545
 546 static int io_bytes_exceeded(struct thread_data *td)
 547 {
 548         unsigned long long bytes;
 549
 550         if (td_rw(td))
 551                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 552         else if (td_write(td))
 553                 bytes = td->this_io_bytes[DDIR_WRITE];
 554         else if (td_read(td))
 555                 bytes = td->this_io_bytes[DDIR_READ];
 556         else
 557                 bytes = td->this_io_bytes[DDIR_TRIM];
 558
 559         return bytes >= td->o.size;
 560 }
 561
 562 /*
 563  * Main IO worker function. It retrieves io_u's to process and queues
 564  * and reaps them, checking for rate and errors along the way.
 565  */
 566 static void do_io(struct thread_data *td)
 567 {
 568         unsigned int i;
 569         int ret = 0;
 570
 571         if (in_ramp_time(td))
 572                 td_set_runstate(td, TD_RAMP);
 573         else
 574                 td_set_runstate(td, TD_RUNNING);
 575
 576         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 577                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 578                 td->o.time_based) {
 579                 struct timeval comp_time;
 580                 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 581                 int min_evts = 0;
 582                 struct io_u *io_u;
 583                 int ret2, full;
 584                 enum fio_ddir ddir;
 585
 586                 if (td->terminate)
 587                         break;
 588
 589                 update_tv_cache(td);
 590
 591                 if (runtime_exceeded(td, &td->tv_cache)) {
 592                         __update_tv_cache(td);
 593                         if (runtime_exceeded(td, &td->tv_cache)) {
 594                                 td->terminate = 1;
 595                                 break;
 596                         }
 597                 }
 598
 599                 if (flow_threshold_exceeded(td))
 600                         continue;
 601
 602                 io_u = get_io_u(td);
 603                 if (!io_u)
 604                         break;
 605
 606                 ddir = io_u->ddir;
 607
 608                 /*
 609                  * Add verification end_io handler if:
 610                  *      - Asked to verify (!td_rw(td))
 611                  *      - Or the io_u is from our verify list (mixed write/ver)
 612                  */
 613                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 614                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 615                         if (td->o.verify_async)
 616                                 io_u->end_io = verify_io_u_async;
 617                         else
 618                                 io_u->end_io = verify_io_u;
 619                         td_set_runstate(td, TD_VERIFYING);
 620                 } else if (in_ramp_time(td))
 621                         td_set_runstate(td, TD_RAMP);
 622                 else
 623                         td_set_runstate(td, TD_RUNNING);
 624
 625                 ret = td_io_queue(td, io_u);
 626                 switch (ret) {
 627                 case FIO_Q_COMPLETED:
 628                         if (io_u->error) {
 629                                 ret = -io_u->error;
 630                                 clear_io_u(td, io_u);
 631                         } else if (io_u->resid) {
 632                                 int bytes = io_u->xfer_buflen - io_u->resid;
 633                                 struct fio_file *f = io_u->file;
 634
 635                                 /*
 636                                  * zero read, fail
 637                                  */
 638                                 if (!bytes) {
 639                                         td_verror(td, EIO, "full resid");
 640                                         put_io_u(td, io_u);
 641                                         break;
 642                                 }
 643
 644                                 io_u->xfer_buflen = io_u->resid;
 645                                 io_u->xfer_buf += bytes;
 646                                 io_u->offset += bytes;
 647
 648                                 if (ddir_rw(io_u->ddir))
 649                                         td->ts.short_io_u[io_u->ddir]++;
 650
 651                                 if (io_u->offset == f->real_file_size)
 652                                         goto sync_done;
 653
 654                                 requeue_io_u(td, &io_u);
 655                         } else {
 656 sync_done:
 657                                 if (__should_check_rate(td, DDIR_READ) ||
 658                                     __should_check_rate(td, DDIR_WRITE) ||
 659                                     __should_check_rate(td, DDIR_TRIM))
 660                                         fio_gettime(&comp_time, NULL);
 661
 662                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 663                                 if (ret < 0)
 664                                         break;
 665                         }
 666                         break;
 667                 case FIO_Q_QUEUED:
 668                         /*
 669                          * if the engine doesn't have a commit hook,
 670                          * the io_u is really queued. if it does have such
 671                          * a hook, it has to call io_u_queued() itself.
 672                          */
 673                         if (td->io_ops->commit == NULL)
 674                                 io_u_queued(td, io_u);
 675                         break;
 676                 case FIO_Q_BUSY:
 677                         requeue_io_u(td, &io_u);
 678                         ret2 = td_io_commit(td);
 679                         if (ret2 < 0)
 680                                 ret = ret2;
 681                         break;
 682                 default:
 683                         assert(ret < 0);
 684                         put_io_u(td, io_u);
 685                         break;
 686                 }
 687
 688                 if (break_on_this_error(td, ddir, &ret))
 689                         break;
 690
 691                 /*
 692                  * See if we need to complete some commands. Note that we
 693                  * can get BUSY even without IO queued, if the system is
 694                  * resource starved.
 695                  */
 696                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 697                 if (full || !td->o.iodepth_batch_complete) {
 698                         min_evts = min(td->o.iodepth_batch_complete,
 699                                         td->cur_depth);
 700                         /*
 701                          * if the queue is full, we MUST reap at least 1 event
 702                          */
 703                         if (full && !min_evts)
 704                                 min_evts = 1;
 705
 706                         if (__should_check_rate(td, DDIR_READ) ||
 707                             __should_check_rate(td, DDIR_WRITE) ||
 708                             __should_check_rate(td, DDIR_TRIM))
 709                                 fio_gettime(&comp_time, NULL);
 710
 711                         do {
 712                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 713                                 if (ret < 0)
 714                                         break;
 715
 716                         } while (full && (td->cur_depth > td->o.iodepth_low));
 717                 }
 718
 719                 if (ret < 0)
 720                         break;
 721                 if (!ddir_rw_sum(bytes_done))
 722                         continue;
 723
 724                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 725                         if (check_min_rate(td, &comp_time, bytes_done)) {
 726                                 if (exitall_on_terminate)
 727                                         fio_terminate_threads(td->groupid);
 728                                 td_verror(td, EIO, "check_min_rate");
 729                                 break;
 730                         }
 731                 }
 732
 733                 if (td->o.thinktime) {
 734                         unsigned long long b;
 735
 736                         b = ddir_rw_sum(td->io_blocks);
 737                         if (!(b % td->o.thinktime_blocks)) {
 738                                 int left;
 739
 740                                 if (td->o.thinktime_spin)
 741                                         usec_spin(td->o.thinktime_spin);
 742
 743                                 left = td->o.thinktime - td->o.thinktime_spin;
 744                                 if (left)
 745                                         usec_sleep(td, left);
 746                         }
 747                 }
 748         }
 749
 750         if (td->trim_entries)
 751                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 752
 753         if (td->o.fill_device && td->error == ENOSPC) {
 754                 td->error = 0;
 755                 td->terminate = 1;
 756         }
 757         if (!td->error) {
 758                 struct fio_file *f;
 759
 760                 i = td->cur_depth;
 761                 if (i) {
 762                         ret = io_u_queued_complete(td, i, NULL);
 763                         if (td->o.fill_device && td->error == ENOSPC)
 764                                 td->error = 0;
 765                 }
 766
 767                 if (should_fsync(td) && td->o.end_fsync) {
 768                         td_set_runstate(td, TD_FSYNCING);
 769
 770                         for_each_file(td, f, i) {
 771                                 if (!fio_file_open(f))
 772                                         continue;
 773                                 fio_io_sync(td, f);
 774                         }
 775                 }
 776         } else
 777                 cleanup_pending_aio(td);
 778
 779         /*
 780          * stop job if we failed doing any IO
 781          */
 782         if (!ddir_rw_sum(td->this_io_bytes))
 783                 td->done = 1;
 784 }
 785
 786 static void cleanup_io_u(struct thread_data *td)
 787 {
 788         struct flist_head *entry, *n;
 789         struct io_u *io_u;
 790
 791         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 792                 io_u = flist_entry(entry, struct io_u, list);
 793
 794                 flist_del(&io_u->list);
 795                 fio_memfree(io_u, sizeof(*io_u));
 796         }
 797
 798         free_io_mem(td);
 799 }
 800
 801 static int init_io_u(struct thread_data *td)
 802 {
 803         struct io_u *io_u;
 804         unsigned int max_bs, min_write;
 805         int cl_align, i, max_units;
 806         int data_xfer = 1;
 807         char *p;
 808
 809         max_units = td->o.iodepth;
 810         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 811         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
 812         min_write = td->o.min_bs[DDIR_WRITE];
 813         td->orig_buffer_size = (unsigned long long) max_bs
 814                                         * (unsigned long long) max_units;
 815
 816         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 817                 data_xfer = 0;
 818
 819         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 820                 unsigned long bs;
 821
 822                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 823                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 824         }
 825
 826         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 827                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 828                 return 1;
 829         }
 830
 831         if (data_xfer && allocate_io_mem(td))
 832                 return 1;
 833
 834         if (td->o.odirect || td->o.mem_align ||
 835             (td->io_ops->flags & FIO_RAWIO))
 836                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 837         else
 838                 p = td->orig_buffer;
 839
 840         cl_align = os_cache_line_size();
 841
 842         for (i = 0; i < max_units; i++) {
 843                 void *ptr;
 844
 845                 if (td->terminate)
 846                         return 1;
 847
 848                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 849                 if (!ptr) {
 850                         log_err("fio: unable to allocate aligned memory\n");
 851                         break;
 852                 }
 853
 854                 io_u = ptr;
 855                 memset(io_u, 0, sizeof(*io_u));
 856                 INIT_FLIST_HEAD(&io_u->list);
 857                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 858
 859                 if (data_xfer) {
 860                         io_u->buf = p;
 861                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 862
 863                         if (td_write(td))
 864                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 865                         if (td_write(td) && td->o.verify_pattern_bytes) {
 866                                 /*
 867                                  * Fill the buffer with the pattern if we are
 868                                  * going to be doing writes.
 869                                  */
 870                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 871                         }
 872                 }
 873
 874                 io_u->index = i;
 875                 io_u->flags = IO_U_F_FREE;
 876                 flist_add(&io_u->list, &td->io_u_freelist);
 877                 p += max_bs;
 878         }
 879
 880         return 0;
 881 }
 882
 883 static int switch_ioscheduler(struct thread_data *td)
 884 {
 885         char tmp[256], tmp2[128];
 886         FILE *f;
 887         int ret;
 888
 889         if (td->io_ops->flags & FIO_DISKLESSIO)
 890                 return 0;
 891
 892         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 893
 894         f = fopen(tmp, "r+");
 895         if (!f) {
 896                 if (errno == ENOENT) {
 897                         log_err("fio: os or kernel doesn't support IO scheduler"
 898                                 " switching\n");
 899                         return 0;
 900                 }
 901                 td_verror(td, errno, "fopen iosched");
 902                 return 1;
 903         }
 904
 905         /*
 906          * Set io scheduler.
 907          */
 908         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 909         if (ferror(f) || ret != 1) {
 910                 td_verror(td, errno, "fwrite");
 911                 fclose(f);
 912                 return 1;
 913         }
 914
 915         rewind(f);
 916
 917         /*
 918          * Read back and check that the selected scheduler is now the default.
 919          */
 920         ret = fread(tmp, 1, sizeof(tmp), f);
 921         if (ferror(f) || ret < 0) {
 922                 td_verror(td, errno, "fread");
 923                 fclose(f);
 924                 return 1;
 925         }
 926
 927         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 928         if (!strstr(tmp, tmp2)) {
 929                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 930                 td_verror(td, EINVAL, "iosched_switch");
 931                 fclose(f);
 932                 return 1;
 933         }
 934
 935         fclose(f);
 936         return 0;
 937 }
 938
 939 static int keep_running(struct thread_data *td)
 940 {
 941         if (td->done)
 942                 return 0;
 943         if (td->o.time_based)
 944                 return 1;
 945         if (td->o.loops) {
 946                 td->o.loops--;
 947                 return 1;
 948         }
 949
 950         if (ddir_rw_sum(td->io_bytes) < td->o.size)
 951                 return 1;
 952
 953         return 0;
 954 }
 955
 956 static int exec_string(const char *string)
 957 {
 958         int ret, newlen = strlen(string) + 1 + 8;
 959         char *str;
 960
 961         str = malloc(newlen);
 962         sprintf(str, "sh -c %s", string);
 963
 964         ret = system(str);
 965         if (ret == -1)
 966                 log_err("fio: exec of cmd <%s> failed\n", str);
 967
 968         free(str);
 969         return ret;
 970 }
 971
 972 /*
 973  * Entry point for the thread based jobs. The process based jobs end up
 974  * here as well, after a little setup.
 975  */
 976 static void *thread_main(void *data)
 977 {
 978         unsigned long long elapsed;
 979         struct thread_data *td = data;
 980         pthread_condattr_t attr;
 981         int clear_state;
 982
 983         if (!td->o.use_thread) {
 984                 setsid();
 985                 td->pid = getpid();
 986         } else
 987                 td->pid = gettid();
 988
 989         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 990
 991         INIT_FLIST_HEAD(&td->io_u_freelist);
 992         INIT_FLIST_HEAD(&td->io_u_busylist);
 993         INIT_FLIST_HEAD(&td->io_u_requeues);
 994         INIT_FLIST_HEAD(&td->io_log_list);
 995         INIT_FLIST_HEAD(&td->io_hist_list);
 996         INIT_FLIST_HEAD(&td->verify_list);
 997         INIT_FLIST_HEAD(&td->trim_list);
 998         pthread_mutex_init(&td->io_u_lock, NULL);
 999         td->io_hist_tree = RB_ROOT;
1000
1001         pthread_condattr_init(&attr);
1002         pthread_cond_init(&td->verify_cond, &attr);
1003         pthread_cond_init(&td->free_cond, &attr);
1004
1005         td_set_runstate(td, TD_INITIALIZED);
1006         dprint(FD_MUTEX, "up startup_mutex\n");
1007         fio_mutex_up(startup_mutex);
1008         dprint(FD_MUTEX, "wait on td->mutex\n");
1009         fio_mutex_down(td->mutex);
1010         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1011
1012         /*
1013          * the ->mutex mutex is now no longer used, close it to avoid
1014          * eating a file descriptor
1015          */
1016         fio_mutex_remove(td->mutex);
1017
1018         /*
1019          * A new gid requires privilege, so we need to do this before setting
1020          * the uid.
1021          */
1022         if (td->o.gid != -1U && setgid(td->o.gid)) {
1023                 td_verror(td, errno, "setgid");
1024                 goto err;
1025         }
1026         if (td->o.uid != -1U && setuid(td->o.uid)) {
1027                 td_verror(td, errno, "setuid");
1028                 goto err;
1029         }
1030
1031         /*
1032          * If we have a gettimeofday() thread, make sure we exclude that
1033          * thread from this job
1034          */
1035         if (td->o.gtod_cpu)
1036                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1037
1038         /*
1039          * Set affinity first, in case it has an impact on the memory
1040          * allocations.
1041          */
1042         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1043                 td_verror(td, errno, "cpu_set_affinity");
1044                 goto err;
1045         }
1046
1047         /*
1048          * May alter parameters that init_io_u() will use, so we need to
1049          * do this first.
1050          */
1051         if (init_iolog(td))
1052                 goto err;
1053
1054         if (init_io_u(td))
1055                 goto err;
1056
1057         if (td->o.verify_async && verify_async_init(td))
1058                 goto err;
1059
1060         if (td->ioprio_set) {
1061                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1062                         td_verror(td, errno, "ioprio_set");
1063                         goto err;
1064                 }
1065         }
1066
1067         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1068                 goto err;
1069
1070         errno = 0;
1071         if (nice(td->o.nice) == -1 && errno != 0) {
1072                 td_verror(td, errno, "nice");
1073                 goto err;
1074         }
1075
1076         if (td->o.ioscheduler && switch_ioscheduler(td))
1077                 goto err;
1078
1079         if (!td->o.create_serialize && setup_files(td))
1080                 goto err;
1081
1082         if (td_io_init(td))
1083                 goto err;
1084
1085         if (init_random_map(td))
1086                 goto err;
1087
1088         if (td->o.exec_prerun) {
1089                 if (exec_string(td->o.exec_prerun))
1090                         goto err;
1091         }
1092
1093         if (td->o.pre_read) {
1094                 if (pre_read_files(td) < 0)
1095                         goto err;
1096         }
1097
1098         fio_gettime(&td->epoch, NULL);
1099         getrusage(RUSAGE_SELF, &td->ru_start);
1100
1101         clear_state = 0;
1102         while (keep_running(td)) {
1103                 fio_gettime(&td->start, NULL);
1104                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1105                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1106                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1107
1108                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1109                                 td->o.ratemin[DDIR_TRIM]) {
1110                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1111                                                 sizeof(td->bw_sample_time));
1112                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1113                                                 sizeof(td->bw_sample_time));
1114                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1115                                                 sizeof(td->bw_sample_time));
1116                 }
1117
1118                 if (clear_state)
1119                         clear_io_state(td);
1120
1121                 prune_io_piece_log(td);
1122
1123                 do_io(td);
1124
1125                 clear_state = 1;
1126
1127                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1128                         elapsed = utime_since_now(&td->start);
1129                         td->ts.runtime[DDIR_READ] += elapsed;
1130                 }
1131                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1132                         elapsed = utime_since_now(&td->start);
1133                         td->ts.runtime[DDIR_WRITE] += elapsed;
1134                 }
1135                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1136                         elapsed = utime_since_now(&td->start);
1137                         td->ts.runtime[DDIR_TRIM] += elapsed;
1138                 }
1139
1140                 if (td->error || td->terminate)
1141                         break;
1142
1143                 if (!td->o.do_verify ||
1144                     td->o.verify == VERIFY_NONE ||
1145                     (td->io_ops->flags & FIO_UNIDIR))
1146                         continue;
1147
1148                 clear_io_state(td);
1149
1150                 fio_gettime(&td->start, NULL);
1151
1152                 do_verify(td);
1153
1154                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1155
1156                 if (td->error || td->terminate)
1157                         break;
1158         }
1159
1160         update_rusage_stat(td);
1161         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1162         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1163         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1164         td->ts.total_run_time = mtime_since_now(&td->epoch);
1165         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1166         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1167         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1168
1169         fio_mutex_down(writeout_mutex);
1170         if (td->bw_log) {
1171                 if (td->o.bw_log_file) {
1172                         finish_log_named(td, td->bw_log,
1173                                                 td->o.bw_log_file, "bw");
1174                 } else
1175                         finish_log(td, td->bw_log, "bw");
1176         }
1177         if (td->lat_log) {
1178                 if (td->o.lat_log_file) {
1179                         finish_log_named(td, td->lat_log,
1180                                                 td->o.lat_log_file, "lat");
1181                 } else
1182                         finish_log(td, td->lat_log, "lat");
1183         }
1184         if (td->slat_log) {
1185                 if (td->o.lat_log_file) {
1186                         finish_log_named(td, td->slat_log,
1187                                                 td->o.lat_log_file, "slat");
1188                 } else
1189                         finish_log(td, td->slat_log, "slat");
1190         }
1191         if (td->clat_log) {
1192                 if (td->o.lat_log_file) {
1193                         finish_log_named(td, td->clat_log,
1194                                                 td->o.lat_log_file, "clat");
1195                 } else
1196                         finish_log(td, td->clat_log, "clat");
1197         }
1198         if (td->iops_log) {
1199                 if (td->o.iops_log_file) {
1200                         finish_log_named(td, td->iops_log,
1201                                                 td->o.iops_log_file, "iops");
1202                 } else
1203                         finish_log(td, td->iops_log, "iops");
1204         }
1205
1206         fio_mutex_up(writeout_mutex);
1207         if (td->o.exec_postrun)
1208                 exec_string(td->o.exec_postrun);
1209
1210         if (exitall_on_terminate)
1211                 fio_terminate_threads(td->groupid);
1212
1213 err:
1214         if (td->error)
1215                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1216                                                         td->verror);
1217
1218         if (td->o.verify_async)
1219                 verify_async_exit(td);
1220
1221         close_and_free_files(td);
1222         close_ioengine(td);
1223         cleanup_io_u(td);
1224         cgroup_shutdown(td, &cgroup_mnt);
1225
1226         if (td->o.cpumask_set) {
1227                 int ret = fio_cpuset_exit(&td->o.cpumask);
1228
1229                 td_verror(td, ret, "fio_cpuset_exit");
1230         }
1231
1232         /*
1233          * do this very late, it will log file closing as well
1234          */
1235         if (td->o.write_iolog_file)
1236                 write_iolog_close(td);
1237
1238         td_set_runstate(td, TD_EXITED);
1239         return (void *) (uintptr_t) td->error;
1240 }
1241
1242
1243 /*
1244  * We cannot pass the td data into a forked process, so attach the td and
1245  * pass it to the thread worker.
1246  */
1247 static int fork_main(int shmid, int offset)
1248 {
1249         struct thread_data *td;
1250         void *data, *ret;
1251
1252 #ifndef __hpux
1253         data = shmat(shmid, NULL, 0);
1254         if (data == (void *) -1) {
1255                 int __err = errno;
1256
1257                 perror("shmat");
1258                 return __err;
1259         }
1260 #else
1261         /*
1262          * HP-UX inherits shm mappings?
1263          */
1264         data = threads;
1265 #endif
1266
1267         td = data + offset * sizeof(struct thread_data);
1268         ret = thread_main(td);
1269         shmdt(data);
1270         return (int) (uintptr_t) ret;
1271 }
1272
1273 /*
1274  * Run over the job map and reap the threads that have exited, if any.
1275  */
1276 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1277                          unsigned int *m_rate)
1278 {
1279         struct thread_data *td;
1280         unsigned int cputhreads, realthreads, pending;
1281         int i, status, ret;
1282
1283         /*
1284          * reap exited threads (TD_EXITED -> TD_REAPED)
1285          */
1286         realthreads = pending = cputhreads = 0;
1287         for_each_td(td, i) {
1288                 int flags = 0;
1289
1290                 /*
1291                  * ->io_ops is NULL for a thread that has closed its
1292                  * io engine
1293                  */
1294                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1295                         cputhreads++;
1296                 else
1297                         realthreads++;
1298
1299                 if (!td->pid) {
1300                         pending++;
1301                         continue;
1302                 }
1303                 if (td->runstate == TD_REAPED)
1304                         continue;
1305                 if (td->o.use_thread) {
1306                         if (td->runstate == TD_EXITED) {
1307                                 td_set_runstate(td, TD_REAPED);
1308                                 goto reaped;
1309                         }
1310                         continue;
1311                 }
1312
1313                 flags = WNOHANG;
1314                 if (td->runstate == TD_EXITED)
1315                         flags = 0;
1316
1317                 /*
1318                  * check if someone quit or got killed in an unusual way
1319                  */
1320                 ret = waitpid(td->pid, &status, flags);
1321                 if (ret < 0) {
1322                         if (errno == ECHILD) {
1323                                 log_err("fio: pid=%d disappeared %d\n",
1324                                                 (int) td->pid, td->runstate);
1325                                 td->sig = ECHILD;
1326                                 td_set_runstate(td, TD_REAPED);
1327                                 goto reaped;
1328                         }
1329                         perror("waitpid");
1330                 } else if (ret == td->pid) {
1331                         if (WIFSIGNALED(status)) {
1332                                 int sig = WTERMSIG(status);
1333
1334                                 if (sig != SIGTERM)
1335                                         log_err("fio: pid=%d, got signal=%d\n",
1336                                                         (int) td->pid, sig);
1337                                 td->sig = sig;
1338                                 td_set_runstate(td, TD_REAPED);
1339                                 goto reaped;
1340                         }
1341                         if (WIFEXITED(status)) {
1342                                 if (WEXITSTATUS(status) && !td->error)
1343                                         td->error = WEXITSTATUS(status);
1344
1345                                 td_set_runstate(td, TD_REAPED);
1346                                 goto reaped;
1347                         }
1348                 }
1349
1350                 /*
1351                  * thread is not dead, continue
1352                  */
1353                 pending++;
1354                 continue;
1355 reaped:
1356                 (*nr_running)--;
1357                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1358                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1359                 if (!td->pid)
1360                         pending--;
1361
1362                 if (td->error)
1363                         exit_value++;
1364
1365                 done_secs += mtime_since_now(&td->epoch) / 1000;
1366         }
1367
1368         if (*nr_running == cputhreads && !pending && realthreads)
1369                 fio_terminate_threads(TERMINATE_ALL);
1370 }
1371
1372 /*
1373  * Main function for kicking off and reaping jobs, as needed.
1374  */
1375 static void run_threads(void)
1376 {
1377         struct thread_data *td;
1378         unsigned long spent;
1379         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1380
1381         if (fio_pin_memory())
1382                 return;
1383
1384         if (fio_gtod_offload && fio_start_gtod_thread())
1385                 return;
1386
1387         set_sig_handlers();
1388
1389         if (output_format == FIO_OUTPUT_NORMAL) {
1390                 log_info("Starting ");
1391                 if (nr_thread)
1392                         log_info("%d thread%s", nr_thread,
1393                                                 nr_thread > 1 ? "s" : "");
1394                 if (nr_process) {
1395                         if (nr_thread)
1396                                 log_info(" and ");
1397                         log_info("%d process%s", nr_process,
1398                                                 nr_process > 1 ? "es" : "");
1399                 }
1400                 log_info("\n");
1401                 fflush(stdout);
1402         }
1403
1404         todo = thread_number;
1405         nr_running = 0;
1406         nr_started = 0;
1407         m_rate = t_rate = 0;
1408
1409         for_each_td(td, i) {
1410                 print_status_init(td->thread_number - 1);
1411
1412                 if (!td->o.create_serialize)
1413                         continue;
1414
1415                 /*
1416                  * do file setup here so it happens sequentially,
1417                  * we don't want X number of threads getting their
1418                  * client data interspersed on disk
1419                  */
1420                 if (setup_files(td)) {
1421                         exit_value++;
1422                         if (td->error)
1423                                 log_err("fio: pid=%d, err=%d/%s\n",
1424                                         (int) td->pid, td->error, td->verror);
1425                         td_set_runstate(td, TD_REAPED);
1426                         todo--;
1427                 } else {
1428                         struct fio_file *f;
1429                         unsigned int j;
1430
1431                         /*
1432                          * for sharing to work, each job must always open
1433                          * its own files. so close them, if we opened them
1434                          * for creation
1435                          */
1436                         for_each_file(td, f, j) {
1437                                 if (fio_file_open(f))
1438                                         td_io_close_file(td, f);
1439                         }
1440                 }
1441         }
1442
1443         set_genesis_time();
1444
1445         while (todo) {
1446                 struct thread_data *map[REAL_MAX_JOBS];
1447                 struct timeval this_start;
1448                 int this_jobs = 0, left;
1449
1450                 /*
1451                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1452                  */
1453                 for_each_td(td, i) {
1454                         if (td->runstate != TD_NOT_CREATED)
1455                                 continue;
1456
1457                         /*
1458                          * never got a chance to start, killed by other
1459                          * thread for some reason
1460                          */
1461                         if (td->terminate) {
1462                                 todo--;
1463                                 continue;
1464                         }
1465
1466                         if (td->o.start_delay) {
1467                                 spent = mtime_since_genesis();
1468
1469                                 if (td->o.start_delay * 1000 > spent)
1470                                         continue;
1471                         }
1472
1473                         if (td->o.stonewall && (nr_started || nr_running)) {
1474                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1475                                                         td->o.name);
1476                                 break;
1477                         }
1478
1479                         init_disk_util(td);
1480
1481                         /*
1482                          * Set state to created. Thread will transition
1483                          * to TD_INITIALIZED when it's done setting up.
1484                          */
1485                         td_set_runstate(td, TD_CREATED);
1486                         map[this_jobs++] = td;
1487                         nr_started++;
1488
1489                         if (td->o.use_thread) {
1490                                 int ret;
1491
1492                                 dprint(FD_PROCESS, "will pthread_create\n");
1493                                 ret = pthread_create(&td->thread, NULL,
1494                                                         thread_main, td);
1495                                 if (ret) {
1496                                         log_err("pthread_create: %s\n",
1497                                                         strerror(ret));
1498                                         nr_started--;
1499                                         break;
1500                                 }
1501                                 ret = pthread_detach(td->thread);
1502                                 if (ret)
1503                                         log_err("pthread_detach: %s",
1504                                                         strerror(ret));
1505                         } else {
1506                                 pid_t pid;
1507                                 dprint(FD_PROCESS, "will fork\n");
1508                                 pid = fork();
1509                                 if (!pid) {
1510                                         int ret = fork_main(shm_id, i);
1511
1512                                         _exit(ret);
1513                                 } else if (i == fio_debug_jobno)
1514                                         *fio_debug_jobp = pid;
1515                         }
1516                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1517                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1518                                 log_err("fio: job startup hung? exiting.\n");
1519                                 fio_terminate_threads(TERMINATE_ALL);
1520                                 fio_abort = 1;
1521                                 nr_started--;
1522                                 break;
1523                         }
1524                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1525                 }
1526
1527                 /*
1528                  * Wait for the started threads to transition to
1529                  * TD_INITIALIZED.
1530                  */
1531                 fio_gettime(&this_start, NULL);
1532                 left = this_jobs;
1533                 while (left && !fio_abort) {
1534                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1535                                 break;
1536
1537                         usleep(100000);
1538
1539                         for (i = 0; i < this_jobs; i++) {
1540                                 td = map[i];
1541                                 if (!td)
1542                                         continue;
1543                                 if (td->runstate == TD_INITIALIZED) {
1544                                         map[i] = NULL;
1545                                         left--;
1546                                 } else if (td->runstate >= TD_EXITED) {
1547                                         map[i] = NULL;
1548                                         left--;
1549                                         todo--;
1550                                         nr_running++; /* work-around... */
1551                                 }
1552                         }
1553                 }
1554
1555                 if (left) {
1556                         log_err("fio: %d job%s failed to start\n", left,
1557                                         left > 1 ? "s" : "");
1558                         for (i = 0; i < this_jobs; i++) {
1559                                 td = map[i];
1560                                 if (!td)
1561                                         continue;
1562                                 kill(td->pid, SIGTERM);
1563                         }
1564                         break;
1565                 }
1566
1567                 /*
1568                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1569                  */
1570                 for_each_td(td, i) {
1571                         if (td->runstate != TD_INITIALIZED)
1572                                 continue;
1573
1574                         if (in_ramp_time(td))
1575                                 td_set_runstate(td, TD_RAMP);
1576                         else
1577                                 td_set_runstate(td, TD_RUNNING);
1578                         nr_running++;
1579                         nr_started--;
1580                         m_rate += ddir_rw_sum(td->o.ratemin);
1581                         t_rate += ddir_rw_sum(td->o.rate);
1582                         todo--;
1583                         fio_mutex_up(td->mutex);
1584                 }
1585
1586                 reap_threads(&nr_running, &t_rate, &m_rate);
1587
1588                 if (todo) {
1589                         if (is_backend)
1590                                 fio_server_idle_loop();
1591                         else
1592                                 usleep(100000);
1593                 }
1594         }
1595
1596         while (nr_running) {
1597                 reap_threads(&nr_running, &t_rate, &m_rate);
1598
1599                 if (is_backend)
1600                         fio_server_idle_loop();
1601                 else
1602                         usleep(10000);
1603         }
1604
1605         update_io_ticks();
1606         fio_unpin_memory();
1607 }
1608
1609 void wait_for_disk_thread_exit(void)
1610 {
1611         fio_mutex_down(disk_thread_mutex);
1612 }
1613
1614 static void free_disk_util(void)
1615 {
1616         disk_util_start_exit();
1617         wait_for_disk_thread_exit();
1618         disk_util_prune_entries();
1619 }
1620
1621 static void *disk_thread_main(void *data)
1622 {
1623         int ret = 0;
1624
1625         fio_mutex_up(startup_mutex);
1626
1627         while (threads && !ret) {
1628                 usleep(DISK_UTIL_MSEC * 1000);
1629                 if (!threads)
1630                         break;
1631                 ret = update_io_ticks();
1632
1633                 if (!is_backend)
1634                         print_thread_status();
1635         }
1636
1637         fio_mutex_up(disk_thread_mutex);
1638         return NULL;
1639 }
1640
1641 static int create_disk_util_thread(void)
1642 {
1643         int ret;
1644
1645         setup_disk_util();
1646
1647         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1648
1649         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1650         if (ret) {
1651                 fio_mutex_remove(disk_thread_mutex);
1652                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1653                 return 1;
1654         }
1655
1656         ret = pthread_detach(disk_util_thread);
1657         if (ret) {
1658                 fio_mutex_remove(disk_thread_mutex);
1659                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1660                 return 1;
1661         }
1662
1663         dprint(FD_MUTEX, "wait on startup_mutex\n");
1664         fio_mutex_down(startup_mutex);
1665         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1666         return 0;
1667 }
1668
1669 int fio_backend(void)
1670 {
1671         struct thread_data *td;
1672         int i;
1673
1674         if (exec_profile) {
1675                 if (load_profile(exec_profile))
1676                         return 1;
1677                 free(exec_profile);
1678                 exec_profile = NULL;
1679         }
1680         if (!thread_number)
1681                 return 0;
1682
1683         if (write_bw_log) {
1684                 setup_log(&agg_io_log[DDIR_READ], 0);
1685                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1686                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1687         }
1688
1689         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1690         if (startup_mutex == NULL)
1691                 return 1;
1692         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1693         if (writeout_mutex == NULL)
1694                 return 1;
1695
1696         set_genesis_time();
1697         create_disk_util_thread();
1698
1699         cgroup_list = smalloc(sizeof(*cgroup_list));
1700         INIT_FLIST_HEAD(cgroup_list);
1701
1702         run_threads();
1703
1704         if (!fio_abort) {
1705                 show_run_stats();
1706                 if (write_bw_log) {
1707                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1708                         __finish_log(agg_io_log[DDIR_WRITE],
1709                                         "agg-write_bw.log");
1710                         __finish_log(agg_io_log[DDIR_TRIM],
1711                                         "agg-write_bw.log");
1712                 }
1713         }
1714
1715         for_each_td(td, i)
1716                 fio_options_free(td);
1717
1718         free_disk_util();
1719         cgroup_kill(cgroup_list);
1720         sfree(cgroup_list);
1721         sfree(cgroup_mnt);
1722
1723         fio_mutex_remove(startup_mutex);
1724         fio_mutex_remove(writeout_mutex);
1725         fio_mutex_remove(disk_thread_mutex);
1726         return exit_value;
1727 }