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