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