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 stat_number = 0;
  66 unsigned int nr_process = 0;
  67 unsigned int nr_thread = 0;
  68 int shm_id = 0;
  69 int temp_stall_ts;
  70 unsigned long done_secs = 0;
  71 volatile int disk_util_exit = 0;
  72
  73 #define PAGE_ALIGN(buf) \
  74         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  75
  76 #define JOB_START_TIMEOUT       (5 * 1000)
  77
  78 static void sig_int(int sig)
  79 {
  80         if (threads) {
  81                 if (is_backend)
  82                         fio_server_got_signal(sig);
  83                 else {
  84                         log_info("\nfio: terminating on signal %d\n", sig);
  85                         fflush(stdout);
  86                         exit_value = 128;
  87                 }
  88
  89                 fio_terminate_threads(TERMINATE_ALL);
  90         }
  91 }
  92
  93 static void sig_show_status(int sig)
  94 {
  95         show_running_run_stats();
  96 }
  97
  98 static void set_sig_handlers(void)
  99 {
 100         struct sigaction act;
 101
 102         memset(&act, 0, sizeof(act));
 103         act.sa_handler = sig_int;
 104         act.sa_flags = SA_RESTART;
 105         sigaction(SIGINT, &act, NULL);
 106
 107         memset(&act, 0, sizeof(act));
 108         act.sa_handler = sig_int;
 109         act.sa_flags = SA_RESTART;
 110         sigaction(SIGTERM, &act, NULL);
 111
 112 /* Windows uses SIGBREAK as a quit signal from other applications */
 113 #ifdef WIN32
 114         memset(&act, 0, sizeof(act));
 115         act.sa_handler = sig_int;
 116         act.sa_flags = SA_RESTART;
 117         sigaction(SIGBREAK, &act, NULL);
 118 #endif
 119
 120         memset(&act, 0, sizeof(act));
 121         act.sa_handler = sig_show_status;
 122         act.sa_flags = SA_RESTART;
 123         sigaction(SIGUSR1, &act, NULL);
 124
 125         if (is_backend) {
 126                 memset(&act, 0, sizeof(act));
 127                 act.sa_handler = sig_int;
 128                 act.sa_flags = SA_RESTART;
 129                 sigaction(SIGPIPE, &act, NULL);
 130         }
 131 }
 132
 133 /*
 134  * Check if we are above the minimum rate given.
 135  */
 136 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 137                             enum fio_ddir ddir)
 138 {
 139         unsigned long long bytes = 0;
 140         unsigned long iops = 0;
 141         unsigned long spent;
 142         unsigned long rate;
 143         unsigned int ratemin = 0;
 144         unsigned int rate_iops = 0;
 145         unsigned int rate_iops_min = 0;
 146
 147         assert(ddir_rw(ddir));
 148
 149         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 150                 return 0;
 151
 152         /*
 153          * allow a 2 second settle period in the beginning
 154          */
 155         if (mtime_since(&td->start, now) < 2000)
 156                 return 0;
 157
 158         iops += td->this_io_blocks[ddir];
 159         bytes += td->this_io_bytes[ddir];
 160         ratemin += td->o.ratemin[ddir];
 161         rate_iops += td->o.rate_iops[ddir];
 162         rate_iops_min += td->o.rate_iops_min[ddir];
 163
 164         /*
 165          * if rate blocks is set, sample is running
 166          */
 167         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 168                 spent = mtime_since(&td->lastrate[ddir], now);
 169                 if (spent < td->o.ratecycle)
 170                         return 0;
 171
 172                 if (td->o.rate[ddir]) {
 173                         /*
 174                          * check bandwidth specified rate
 175                          */
 176                         if (bytes < td->rate_bytes[ddir]) {
 177                                 log_err("%s: min rate %u not met\n", td->o.name,
 178                                                                 ratemin);
 179                                 return 1;
 180                         } else {
 181                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 182                                 if (rate < ratemin ||
 183                                     bytes < td->rate_bytes[ddir]) {
 184                                         log_err("%s: min rate %u not met, got"
 185                                                 " %luKB/sec\n", td->o.name,
 186                                                         ratemin, rate);
 187                                         return 1;
 188                                 }
 189                         }
 190                 } else {
 191                         /*
 192                          * checks iops specified rate
 193                          */
 194                         if (iops < rate_iops) {
 195                                 log_err("%s: min iops rate %u not met\n",
 196                                                 td->o.name, rate_iops);
 197                                 return 1;
 198                         } else {
 199                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 200                                 if (rate < rate_iops_min ||
 201                                     iops < td->rate_blocks[ddir]) {
 202                                         log_err("%s: min iops rate %u not met,"
 203                                                 " got %lu\n", td->o.name,
 204                                                         rate_iops_min, rate);
 205                                 }
 206                         }
 207                 }
 208         }
 209
 210         td->rate_bytes[ddir] = bytes;
 211         td->rate_blocks[ddir] = iops;
 212         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 213         return 0;
 214 }
 215
 216 static int check_min_rate(struct thread_data *td, struct timeval *now,
 217                           unsigned long *bytes_done)
 218 {
 219         int ret = 0;
 220
 221         if (bytes_done[DDIR_READ])
 222                 ret |= __check_min_rate(td, now, DDIR_READ);
 223         if (bytes_done[DDIR_WRITE])
 224                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 225         if (bytes_done[DDIR_TRIM])
 226                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 227
 228         return ret;
 229 }
 230
 231 /*
 232  * When job exits, we can cancel the in-flight IO if we are using async
 233  * io. Attempt to do so.
 234  */
 235 static void cleanup_pending_aio(struct thread_data *td)
 236 {
 237         struct flist_head *entry, *n;
 238         struct io_u *io_u;
 239         int r;
 240
 241         /*
 242          * get immediately available events, if any
 243          */
 244         r = io_u_queued_complete(td, 0, NULL);
 245         if (r < 0)
 246                 return;
 247
 248         /*
 249          * now cancel remaining active events
 250          */
 251         if (td->io_ops->cancel) {
 252                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 253                         io_u = flist_entry(entry, struct io_u, list);
 254
 255                         /*
 256                          * if the io_u isn't in flight, then that generally
 257                          * means someone leaked an io_u. complain but fix
 258                          * it up, so we don't stall here.
 259                          */
 260                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 261                                 log_err("fio: non-busy IO on busy list\n");
 262                                 put_io_u(td, io_u);
 263                         } else {
 264                                 r = td->io_ops->cancel(td, io_u);
 265                                 if (!r)
 266                                         put_io_u(td, io_u);
 267                         }
 268                 }
 269         }
 270
 271         if (td->cur_depth)
 272                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 273 }
 274
 275 /*
 276  * Helper to handle the final sync of a file. Works just like the normal
 277  * io path, just does everything sync.
 278  */
 279 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 280 {
 281         struct io_u *io_u = __get_io_u(td);
 282         int ret;
 283
 284         if (!io_u)
 285                 return 1;
 286
 287         io_u->ddir = DDIR_SYNC;
 288         io_u->file = f;
 289
 290         if (td_io_prep(td, io_u)) {
 291                 put_io_u(td, io_u);
 292                 return 1;
 293         }
 294
 295 requeue:
 296         ret = td_io_queue(td, io_u);
 297         if (ret < 0) {
 298                 td_verror(td, io_u->error, "td_io_queue");
 299                 put_io_u(td, io_u);
 300                 return 1;
 301         } else if (ret == FIO_Q_QUEUED) {
 302                 if (io_u_queued_complete(td, 1, NULL) < 0)
 303                         return 1;
 304         } else if (ret == FIO_Q_COMPLETED) {
 305                 if (io_u->error) {
 306                         td_verror(td, io_u->error, "td_io_queue");
 307                         return 1;
 308                 }
 309
 310                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 311                         return 1;
 312         } else if (ret == FIO_Q_BUSY) {
 313                 if (td_io_commit(td))
 314                         return 1;
 315                 goto requeue;
 316         }
 317
 318         return 0;
 319 }
 320
 321 static inline void __update_tv_cache(struct thread_data *td)
 322 {
 323         fio_gettime(&td->tv_cache, NULL);
 324 }
 325
 326 static inline void update_tv_cache(struct thread_data *td)
 327 {
 328         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 329                 __update_tv_cache(td);
 330 }
 331
 332 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 333 {
 334         if (in_ramp_time(td))
 335                 return 0;
 336         if (!td->o.timeout)
 337                 return 0;
 338         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 339                 return 1;
 340
 341         return 0;
 342 }
 343
 344 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 345                                int *retptr)
 346 {
 347         int ret = *retptr;
 348
 349         if (ret < 0 || td->error) {
 350                 int err = td->error;
 351                 enum error_type_bit eb;
 352
 353                 if (ret < 0)
 354                         err = -ret;
 355
 356                 eb = td_error_type(ddir, err);
 357                 if (!(td->o.continue_on_error & (1 << eb)))
 358                         return 1;
 359
 360                 if (td_non_fatal_error(td, eb, err)) {
 361                         /*
 362                          * Continue with the I/Os in case of
 363                          * a non fatal error.
 364                          */
 365                         update_error_count(td, err);
 366                         td_clear_error(td);
 367                         *retptr = 0;
 368                         return 0;
 369                 } else if (td->o.fill_device && err == ENOSPC) {
 370                         /*
 371                          * We expect to hit this error if
 372                          * fill_device option is set.
 373                          */
 374                         td_clear_error(td);
 375                         td->terminate = 1;
 376                         return 1;
 377                 } else {
 378                         /*
 379                          * Stop the I/O in case of a fatal
 380                          * error.
 381                          */
 382                         update_error_count(td, err);
 383                         return 1;
 384                 }
 385         }
 386
 387         return 0;
 388 }
 389
 390 /*
 391  * The main verify engine. Runs over the writes we previously submitted,
 392  * reads the blocks back in, and checks the crc/md5 of the data.
 393  */
 394 static void do_verify(struct thread_data *td)
 395 {
 396         struct fio_file *f;
 397         struct io_u *io_u;
 398         int ret, min_events;
 399         unsigned int i;
 400
 401         dprint(FD_VERIFY, "starting loop\n");
 402
 403         /*
 404          * sync io first and invalidate cache, to make sure we really
 405          * read from disk.
 406          */
 407         for_each_file(td, f, i) {
 408                 if (!fio_file_open(f))
 409                         continue;
 410                 if (fio_io_sync(td, f))
 411                         break;
 412                 if (file_invalidate_cache(td, f))
 413                         break;
 414         }
 415
 416         if (td->error)
 417                 return;
 418
 419         td_set_runstate(td, TD_VERIFYING);
 420
 421         io_u = NULL;
 422         while (!td->terminate) {
 423                 int ret2, full;
 424
 425                 update_tv_cache(td);
 426
 427                 if (runtime_exceeded(td, &td->tv_cache)) {
 428                         __update_tv_cache(td);
 429                         if (runtime_exceeded(td, &td->tv_cache)) {
 430                                 td->terminate = 1;
 431                                 break;
 432                         }
 433                 }
 434
 435                 if (flow_threshold_exceeded(td))
 436                         continue;
 437
 438                 io_u = __get_io_u(td);
 439                 if (!io_u)
 440                         break;
 441
 442                 if (get_next_verify(td, io_u)) {
 443                         put_io_u(td, io_u);
 444                         break;
 445                 }
 446
 447                 if (td_io_prep(td, io_u)) {
 448                         put_io_u(td, io_u);
 449                         break;
 450                 }
 451
 452                 if (td->o.verify_async)
 453                         io_u->end_io = verify_io_u_async;
 454                 else
 455                         io_u->end_io = verify_io_u;
 456
 457                 ret = td_io_queue(td, io_u);
 458                 switch (ret) {
 459                 case FIO_Q_COMPLETED:
 460                         if (io_u->error) {
 461                                 ret = -io_u->error;
 462                                 clear_io_u(td, io_u);
 463                         } else if (io_u->resid) {
 464                                 int bytes = io_u->xfer_buflen - io_u->resid;
 465
 466                                 /*
 467                                  * zero read, fail
 468                                  */
 469                                 if (!bytes) {
 470                                         td_verror(td, EIO, "full resid");
 471                                         put_io_u(td, io_u);
 472                                         break;
 473                                 }
 474
 475                                 io_u->xfer_buflen = io_u->resid;
 476                                 io_u->xfer_buf += bytes;
 477                                 io_u->offset += bytes;
 478
 479                                 if (ddir_rw(io_u->ddir))
 480                                         td->ts.short_io_u[io_u->ddir]++;
 481
 482                                 f = io_u->file;
 483                                 if (io_u->offset == f->real_file_size)
 484                                         goto sync_done;
 485
 486                                 requeue_io_u(td, &io_u);
 487                         } else {
 488 sync_done:
 489                                 ret = io_u_sync_complete(td, io_u, NULL);
 490                                 if (ret < 0)
 491                                         break;
 492                         }
 493                         continue;
 494                 case FIO_Q_QUEUED:
 495                         break;
 496                 case FIO_Q_BUSY:
 497                         requeue_io_u(td, &io_u);
 498                         ret2 = td_io_commit(td);
 499                         if (ret2 < 0)
 500                                 ret = ret2;
 501                         break;
 502                 default:
 503                         assert(ret < 0);
 504                         td_verror(td, -ret, "td_io_queue");
 505                         break;
 506                 }
 507
 508                 if (break_on_this_error(td, io_u->ddir, &ret))
 509                         break;
 510
 511                 /*
 512                  * if we can queue more, do so. but check if there are
 513                  * completed io_u's first. Note that we can get BUSY even
 514                  * without IO queued, if the system is resource starved.
 515                  */
 516                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 517                 if (full || !td->o.iodepth_batch_complete) {
 518                         min_events = min(td->o.iodepth_batch_complete,
 519                                          td->cur_depth);
 520                         /*
 521                          * if the queue is full, we MUST reap at least 1 event
 522                          */
 523                         if (full && !min_events)
 524                                 min_events = 1;
 525
 526                         do {
 527                                 /*
 528                                  * Reap required number of io units, if any,
 529                                  * and do the verification on them through
 530                                  * the callback handler
 531                                  */
 532                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 533                                         ret = -1;
 534                                         break;
 535                                 }
 536                         } while (full && (td->cur_depth > td->o.iodepth_low));
 537                 }
 538                 if (ret < 0)
 539                         break;
 540         }
 541
 542         if (!td->error) {
 543                 min_events = td->cur_depth;
 544
 545                 if (min_events)
 546                         ret = io_u_queued_complete(td, min_events, NULL);
 547         } else
 548                 cleanup_pending_aio(td);
 549
 550         td_set_runstate(td, TD_RUNNING);
 551
 552         dprint(FD_VERIFY, "exiting loop\n");
 553 }
 554
 555 static int io_bytes_exceeded(struct thread_data *td)
 556 {
 557         unsigned long long bytes;
 558
 559         if (td_rw(td))
 560                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 561         else if (td_write(td))
 562                 bytes = td->this_io_bytes[DDIR_WRITE];
 563         else if (td_read(td))
 564                 bytes = td->this_io_bytes[DDIR_READ];
 565         else
 566                 bytes = td->this_io_bytes[DDIR_TRIM];
 567
 568         return bytes >= td->o.size;
 569 }
 570
 571 /*
 572  * Main IO worker function. It retrieves io_u's to process and queues
 573  * and reaps them, checking for rate and errors along the way.
 574  */
 575 static void do_io(struct thread_data *td)
 576 {
 577         unsigned int i;
 578         int ret = 0;
 579
 580         if (in_ramp_time(td))
 581                 td_set_runstate(td, TD_RAMP);
 582         else
 583                 td_set_runstate(td, TD_RUNNING);
 584
 585         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 586                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 587                 td->o.time_based) {
 588                 struct timeval comp_time;
 589                 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 590                 int min_evts = 0;
 591                 struct io_u *io_u;
 592                 int ret2, full;
 593                 enum fio_ddir ddir;
 594
 595                 if (td->terminate || td->done)
 596                         break;
 597
 598                 update_tv_cache(td);
 599
 600                 if (runtime_exceeded(td, &td->tv_cache)) {
 601                         __update_tv_cache(td);
 602                         if (runtime_exceeded(td, &td->tv_cache)) {
 603                                 td->terminate = 1;
 604                                 break;
 605                         }
 606                 }
 607
 608                 if (flow_threshold_exceeded(td))
 609                         continue;
 610
 611                 io_u = get_io_u(td);
 612                 if (!io_u)
 613                         break;
 614
 615                 ddir = io_u->ddir;
 616
 617                 /*
 618                  * Add verification end_io handler if:
 619                  *      - Asked to verify (!td_rw(td))
 620                  *      - Or the io_u is from our verify list (mixed write/ver)
 621                  */
 622                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 623                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 624                         if (td->o.verify_async)
 625                                 io_u->end_io = verify_io_u_async;
 626                         else
 627                                 io_u->end_io = verify_io_u;
 628                         td_set_runstate(td, TD_VERIFYING);
 629                 } else if (in_ramp_time(td))
 630                         td_set_runstate(td, TD_RAMP);
 631                 else
 632                         td_set_runstate(td, TD_RUNNING);
 633
 634                 ret = td_io_queue(td, io_u);
 635                 switch (ret) {
 636                 case FIO_Q_COMPLETED:
 637                         if (io_u->error) {
 638                                 ret = -io_u->error;
 639                                 clear_io_u(td, io_u);
 640                         } else if (io_u->resid) {
 641                                 int bytes = io_u->xfer_buflen - io_u->resid;
 642                                 struct fio_file *f = io_u->file;
 643
 644                                 /*
 645                                  * zero read, fail
 646                                  */
 647                                 if (!bytes) {
 648                                         td_verror(td, EIO, "full resid");
 649                                         put_io_u(td, io_u);
 650                                         break;
 651                                 }
 652
 653                                 io_u->xfer_buflen = io_u->resid;
 654                                 io_u->xfer_buf += bytes;
 655                                 io_u->offset += bytes;
 656
 657                                 if (ddir_rw(io_u->ddir))
 658                                         td->ts.short_io_u[io_u->ddir]++;
 659
 660                                 if (io_u->offset == f->real_file_size)
 661                                         goto sync_done;
 662
 663                                 requeue_io_u(td, &io_u);
 664                         } else {
 665 sync_done:
 666                                 if (__should_check_rate(td, DDIR_READ) ||
 667                                     __should_check_rate(td, DDIR_WRITE) ||
 668                                     __should_check_rate(td, DDIR_TRIM))
 669                                         fio_gettime(&comp_time, NULL);
 670
 671                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 672                                 if (ret < 0)
 673                                         break;
 674                         }
 675                         break;
 676                 case FIO_Q_QUEUED:
 677                         /*
 678                          * if the engine doesn't have a commit hook,
 679                          * the io_u is really queued. if it does have such
 680                          * a hook, it has to call io_u_queued() itself.
 681                          */
 682                         if (td->io_ops->commit == NULL)
 683                                 io_u_queued(td, io_u);
 684                         break;
 685                 case FIO_Q_BUSY:
 686                         requeue_io_u(td, &io_u);
 687                         ret2 = td_io_commit(td);
 688                         if (ret2 < 0)
 689                                 ret = ret2;
 690                         break;
 691                 default:
 692                         assert(ret < 0);
 693                         put_io_u(td, io_u);
 694                         break;
 695                 }
 696
 697                 if (break_on_this_error(td, ddir, &ret))
 698                         break;
 699
 700                 /*
 701                  * See if we need to complete some commands. Note that we
 702                  * can get BUSY even without IO queued, if the system is
 703                  * resource starved.
 704                  */
 705                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 706                 if (full || !td->o.iodepth_batch_complete) {
 707                         min_evts = min(td->o.iodepth_batch_complete,
 708                                         td->cur_depth);
 709                         /*
 710                          * if the queue is full, we MUST reap at least 1 event
 711                          */
 712                         if (full && !min_evts)
 713                                 min_evts = 1;
 714
 715                         if (__should_check_rate(td, DDIR_READ) ||
 716                             __should_check_rate(td, DDIR_WRITE) ||
 717                             __should_check_rate(td, DDIR_TRIM))
 718                                 fio_gettime(&comp_time, NULL);
 719
 720                         do {
 721                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 722                                 if (ret < 0)
 723                                         break;
 724
 725                         } while (full && (td->cur_depth > td->o.iodepth_low));
 726                 }
 727
 728                 if (ret < 0)
 729                         break;
 730                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 731                         continue;
 732
 733                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 734                         if (check_min_rate(td, &comp_time, bytes_done)) {
 735                                 if (exitall_on_terminate)
 736                                         fio_terminate_threads(td->groupid);
 737                                 td_verror(td, EIO, "check_min_rate");
 738                                 break;
 739                         }
 740                 }
 741
 742                 if (td->o.thinktime) {
 743                         unsigned long long b;
 744
 745                         b = ddir_rw_sum(td->io_blocks);
 746                         if (!(b % td->o.thinktime_blocks)) {
 747                                 int left;
 748
 749                                 if (td->o.thinktime_spin)
 750                                         usec_spin(td->o.thinktime_spin);
 751
 752                                 left = td->o.thinktime - td->o.thinktime_spin;
 753                                 if (left)
 754                                         usec_sleep(td, left);
 755                         }
 756                 }
 757         }
 758
 759         if (td->trim_entries)
 760                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 761
 762         if (td->o.fill_device && td->error == ENOSPC) {
 763                 td->error = 0;
 764                 td->terminate = 1;
 765         }
 766         if (!td->error) {
 767                 struct fio_file *f;
 768
 769                 i = td->cur_depth;
 770                 if (i) {
 771                         ret = io_u_queued_complete(td, i, NULL);
 772                         if (td->o.fill_device && td->error == ENOSPC)
 773                                 td->error = 0;
 774                 }
 775
 776                 if (should_fsync(td) && td->o.end_fsync) {
 777                         td_set_runstate(td, TD_FSYNCING);
 778
 779                         for_each_file(td, f, i) {
 780                                 if (!fio_file_open(f))
 781                                         continue;
 782                                 fio_io_sync(td, f);
 783                         }
 784                 }
 785         } else
 786                 cleanup_pending_aio(td);
 787
 788         /*
 789          * stop job if we failed doing any IO
 790          */
 791         if (!ddir_rw_sum(td->this_io_bytes))
 792                 td->done = 1;
 793 }
 794
 795 static void cleanup_io_u(struct thread_data *td)
 796 {
 797         struct flist_head *entry, *n;
 798         struct io_u *io_u;
 799
 800         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 801                 io_u = flist_entry(entry, struct io_u, list);
 802
 803                 flist_del(&io_u->list);
 804                 fio_memfree(io_u, sizeof(*io_u));
 805         }
 806
 807         free_io_mem(td);
 808 }
 809
 810 static int init_io_u(struct thread_data *td)
 811 {
 812         struct io_u *io_u;
 813         unsigned int max_bs, min_write;
 814         int cl_align, i, max_units;
 815         int data_xfer = 1;
 816         char *p;
 817
 818         max_units = td->o.iodepth;
 819         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 820         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
 821         min_write = td->o.min_bs[DDIR_WRITE];
 822         td->orig_buffer_size = (unsigned long long) max_bs
 823                                         * (unsigned long long) max_units;
 824
 825         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 826                 data_xfer = 0;
 827
 828         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 829                 unsigned long bs;
 830
 831                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 832                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 833         }
 834
 835         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 836                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 837                 return 1;
 838         }
 839
 840         if (data_xfer && allocate_io_mem(td))
 841                 return 1;
 842
 843         if (td->o.odirect || td->o.mem_align ||
 844             (td->io_ops->flags & FIO_RAWIO))
 845                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 846         else
 847                 p = td->orig_buffer;
 848
 849         cl_align = os_cache_line_size();
 850
 851         for (i = 0; i < max_units; i++) {
 852                 void *ptr;
 853
 854                 if (td->terminate)
 855                         return 1;
 856
 857                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 858                 if (!ptr) {
 859                         log_err("fio: unable to allocate aligned memory\n");
 860                         break;
 861                 }
 862
 863                 io_u = ptr;
 864                 memset(io_u, 0, sizeof(*io_u));
 865                 INIT_FLIST_HEAD(&io_u->list);
 866                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 867
 868                 if (data_xfer) {
 869                         io_u->buf = p;
 870                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 871
 872                         if (td_write(td))
 873                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 874                         if (td_write(td) && td->o.verify_pattern_bytes) {
 875                                 /*
 876                                  * Fill the buffer with the pattern if we are
 877                                  * going to be doing writes.
 878                                  */
 879                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 880                         }
 881                 }
 882
 883                 io_u->index = i;
 884                 io_u->flags = IO_U_F_FREE;
 885                 flist_add(&io_u->list, &td->io_u_freelist);
 886                 p += max_bs;
 887         }
 888
 889         return 0;
 890 }
 891
 892 static int switch_ioscheduler(struct thread_data *td)
 893 {
 894         char tmp[256], tmp2[128];
 895         FILE *f;
 896         int ret;
 897
 898         if (td->io_ops->flags & FIO_DISKLESSIO)
 899                 return 0;
 900
 901         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 902
 903         f = fopen(tmp, "r+");
 904         if (!f) {
 905                 if (errno == ENOENT) {
 906                         log_err("fio: os or kernel doesn't support IO scheduler"
 907                                 " switching\n");
 908                         return 0;
 909                 }
 910                 td_verror(td, errno, "fopen iosched");
 911                 return 1;
 912         }
 913
 914         /*
 915          * Set io scheduler.
 916          */
 917         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 918         if (ferror(f) || ret != 1) {
 919                 td_verror(td, errno, "fwrite");
 920                 fclose(f);
 921                 return 1;
 922         }
 923
 924         rewind(f);
 925
 926         /*
 927          * Read back and check that the selected scheduler is now the default.
 928          */
 929         ret = fread(tmp, 1, sizeof(tmp), f);
 930         if (ferror(f) || ret < 0) {
 931                 td_verror(td, errno, "fread");
 932                 fclose(f);
 933                 return 1;
 934         }
 935
 936         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 937         if (!strstr(tmp, tmp2)) {
 938                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 939                 td_verror(td, EINVAL, "iosched_switch");
 940                 fclose(f);
 941                 return 1;
 942         }
 943
 944         fclose(f);
 945         return 0;
 946 }
 947
 948 static int keep_running(struct thread_data *td)
 949 {
 950         if (td->done)
 951                 return 0;
 952         if (td->o.time_based)
 953                 return 1;
 954         if (td->o.loops) {
 955                 td->o.loops--;
 956                 return 1;
 957         }
 958
 959         if (ddir_rw_sum(td->io_bytes) < td->o.size)
 960                 return 1;
 961
 962         return 0;
 963 }
 964
 965 static int exec_string(const char *string)
 966 {
 967         int ret, newlen = strlen(string) + 1 + 8;
 968         char *str;
 969
 970         str = malloc(newlen);
 971         sprintf(str, "sh -c %s", string);
 972
 973         ret = system(str);
 974         if (ret == -1)
 975                 log_err("fio: exec of cmd <%s> failed\n", str);
 976
 977         free(str);
 978         return ret;
 979 }
 980
 981 /*
 982  * Entry point for the thread based jobs. The process based jobs end up
 983  * here as well, after a little setup.
 984  */
 985 static void *thread_main(void *data)
 986 {
 987         unsigned long long elapsed;
 988         struct thread_data *td = data;
 989         pthread_condattr_t attr;
 990         int clear_state;
 991
 992         if (!td->o.use_thread) {
 993                 setsid();
 994                 td->pid = getpid();
 995         } else
 996                 td->pid = gettid();
 997
 998         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 999
1000         INIT_FLIST_HEAD(&td->io_u_freelist);
1001         INIT_FLIST_HEAD(&td->io_u_busylist);
1002         INIT_FLIST_HEAD(&td->io_u_requeues);
1003         INIT_FLIST_HEAD(&td->io_log_list);
1004         INIT_FLIST_HEAD(&td->io_hist_list);
1005         INIT_FLIST_HEAD(&td->verify_list);
1006         INIT_FLIST_HEAD(&td->trim_list);
1007         pthread_mutex_init(&td->io_u_lock, NULL);
1008         td->io_hist_tree = RB_ROOT;
1009
1010         pthread_condattr_init(&attr);
1011         pthread_cond_init(&td->verify_cond, &attr);
1012         pthread_cond_init(&td->free_cond, &attr);
1013
1014         td_set_runstate(td, TD_INITIALIZED);
1015         dprint(FD_MUTEX, "up startup_mutex\n");
1016         fio_mutex_up(startup_mutex);
1017         dprint(FD_MUTEX, "wait on td->mutex\n");
1018         fio_mutex_down(td->mutex);
1019         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1020
1021         /*
1022          * the ->mutex mutex is now no longer used, close it to avoid
1023          * eating a file descriptor
1024          */
1025         fio_mutex_remove(td->mutex);
1026
1027         /*
1028          * A new gid requires privilege, so we need to do this before setting
1029          * the uid.
1030          */
1031         if (td->o.gid != -1U && setgid(td->o.gid)) {
1032                 td_verror(td, errno, "setgid");
1033                 goto err;
1034         }
1035         if (td->o.uid != -1U && setuid(td->o.uid)) {
1036                 td_verror(td, errno, "setuid");
1037                 goto err;
1038         }
1039
1040         /*
1041          * If we have a gettimeofday() thread, make sure we exclude that
1042          * thread from this job
1043          */
1044         if (td->o.gtod_cpu)
1045                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1046
1047         /*
1048          * Set affinity first, in case it has an impact on the memory
1049          * allocations.
1050          */
1051         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1052                 td_verror(td, errno, "cpu_set_affinity");
1053                 goto err;
1054         }
1055
1056 #ifdef FIO_HAVE_LIBNUMA
1057         /* numa node setup */
1058         if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1059                 int ret;
1060
1061                 if (numa_available() < 0) {
1062                         td_verror(td, errno, "Does not support NUMA API\n");
1063                         goto err;
1064                 }
1065
1066                 if (td->o.numa_cpumask_set) {
1067                         ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1068                         if (ret == -1) {
1069                                 td_verror(td, errno, \
1070                                         "numa_run_on_node_mask failed\n");
1071                                 goto err;
1072                         }
1073                 }
1074
1075                 if (td->o.numa_memmask_set) {
1076
1077                         switch (td->o.numa_mem_mode) {
1078                         case MPOL_INTERLEAVE:
1079                                 numa_set_interleave_mask(td->o.numa_memnodesmask);
1080                                 break;
1081                         case MPOL_BIND:
1082                                 numa_set_membind(td->o.numa_memnodesmask);
1083                                 break;
1084                         case MPOL_LOCAL:
1085                                 numa_set_localalloc();
1086                                 break;
1087                         case MPOL_PREFERRED:
1088                                 numa_set_preferred(td->o.numa_mem_prefer_node);
1089                                 break;
1090                         case MPOL_DEFAULT:
1091                         default:
1092                                 break;
1093                         }
1094
1095                 }
1096         }
1097 #endif
1098
1099         /*
1100          * May alter parameters that init_io_u() will use, so we need to
1101          * do this first.
1102          */
1103         if (init_iolog(td))
1104                 goto err;
1105
1106         if (init_io_u(td))
1107                 goto err;
1108
1109         if (td->o.verify_async && verify_async_init(td))
1110                 goto err;
1111
1112         if (td->ioprio_set) {
1113                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1114                         td_verror(td, errno, "ioprio_set");
1115                         goto err;
1116                 }
1117         }
1118
1119         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1120                 goto err;
1121
1122         errno = 0;
1123         if (nice(td->o.nice) == -1 && errno != 0) {
1124                 td_verror(td, errno, "nice");
1125                 goto err;
1126         }
1127
1128         if (td->o.ioscheduler && switch_ioscheduler(td))
1129                 goto err;
1130
1131         if (!td->o.create_serialize && setup_files(td))
1132                 goto err;
1133
1134         if (td_io_init(td))
1135                 goto err;
1136
1137         if (init_random_map(td))
1138                 goto err;
1139
1140         if (td->o.exec_prerun) {
1141                 if (exec_string(td->o.exec_prerun))
1142                         goto err;
1143         }
1144
1145         if (td->o.pre_read) {
1146                 if (pre_read_files(td) < 0)
1147                         goto err;
1148         }
1149
1150         fio_gettime(&td->epoch, NULL);
1151         getrusage(RUSAGE_SELF, &td->ru_start);
1152
1153         clear_state = 0;
1154         while (keep_running(td)) {
1155                 fio_gettime(&td->start, NULL);
1156                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1157                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1158                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1159
1160                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1161                                 td->o.ratemin[DDIR_TRIM]) {
1162                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1163                                                 sizeof(td->bw_sample_time));
1164                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1165                                                 sizeof(td->bw_sample_time));
1166                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1167                                                 sizeof(td->bw_sample_time));
1168                 }
1169
1170                 if (clear_state)
1171                         clear_io_state(td);
1172
1173                 prune_io_piece_log(td);
1174
1175                 do_io(td);
1176
1177                 clear_state = 1;
1178
1179                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1180                         elapsed = utime_since_now(&td->start);
1181                         td->ts.runtime[DDIR_READ] += elapsed;
1182                 }
1183                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1184                         elapsed = utime_since_now(&td->start);
1185                         td->ts.runtime[DDIR_WRITE] += elapsed;
1186                 }
1187                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1188                         elapsed = utime_since_now(&td->start);
1189                         td->ts.runtime[DDIR_TRIM] += elapsed;
1190                 }
1191
1192                 if (td->error || td->terminate)
1193                         break;
1194
1195                 if (!td->o.do_verify ||
1196                     td->o.verify == VERIFY_NONE ||
1197                     (td->io_ops->flags & FIO_UNIDIR))
1198                         continue;
1199
1200                 clear_io_state(td);
1201
1202                 fio_gettime(&td->start, NULL);
1203
1204                 do_verify(td);
1205
1206                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1207
1208                 if (td->error || td->terminate)
1209                         break;
1210         }
1211
1212         update_rusage_stat(td);
1213         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1214         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1215         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1216         td->ts.total_run_time = mtime_since_now(&td->epoch);
1217         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1218         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1219         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1220
1221         fio_mutex_down(writeout_mutex);
1222         if (td->bw_log) {
1223                 if (td->o.bw_log_file) {
1224                         finish_log_named(td, td->bw_log,
1225                                                 td->o.bw_log_file, "bw");
1226                 } else
1227                         finish_log(td, td->bw_log, "bw");
1228         }
1229         if (td->lat_log) {
1230                 if (td->o.lat_log_file) {
1231                         finish_log_named(td, td->lat_log,
1232                                                 td->o.lat_log_file, "lat");
1233                 } else
1234                         finish_log(td, td->lat_log, "lat");
1235         }
1236         if (td->slat_log) {
1237                 if (td->o.lat_log_file) {
1238                         finish_log_named(td, td->slat_log,
1239                                                 td->o.lat_log_file, "slat");
1240                 } else
1241                         finish_log(td, td->slat_log, "slat");
1242         }
1243         if (td->clat_log) {
1244                 if (td->o.lat_log_file) {
1245                         finish_log_named(td, td->clat_log,
1246                                                 td->o.lat_log_file, "clat");
1247                 } else
1248                         finish_log(td, td->clat_log, "clat");
1249         }
1250         if (td->iops_log) {
1251                 if (td->o.iops_log_file) {
1252                         finish_log_named(td, td->iops_log,
1253                                                 td->o.iops_log_file, "iops");
1254                 } else
1255                         finish_log(td, td->iops_log, "iops");
1256         }
1257
1258         fio_mutex_up(writeout_mutex);
1259         if (td->o.exec_postrun)
1260                 exec_string(td->o.exec_postrun);
1261
1262         if (exitall_on_terminate)
1263                 fio_terminate_threads(td->groupid);
1264
1265 err:
1266         if (td->error)
1267                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1268                                                         td->verror);
1269
1270         if (td->o.verify_async)
1271                 verify_async_exit(td);
1272
1273         close_and_free_files(td);
1274         close_ioengine(td);
1275         cleanup_io_u(td);
1276         cgroup_shutdown(td, &cgroup_mnt);
1277
1278         if (td->o.cpumask_set) {
1279                 int ret = fio_cpuset_exit(&td->o.cpumask);
1280
1281                 td_verror(td, ret, "fio_cpuset_exit");
1282         }
1283
1284         /*
1285          * do this very late, it will log file closing as well
1286          */
1287         if (td->o.write_iolog_file)
1288                 write_iolog_close(td);
1289
1290         td_set_runstate(td, TD_EXITED);
1291         return (void *) (uintptr_t) td->error;
1292 }
1293
1294
1295 /*
1296  * We cannot pass the td data into a forked process, so attach the td and
1297  * pass it to the thread worker.
1298  */
1299 static int fork_main(int shmid, int offset)
1300 {
1301         struct thread_data *td;
1302         void *data, *ret;
1303
1304 #ifndef __hpux
1305         data = shmat(shmid, NULL, 0);
1306         if (data == (void *) -1) {
1307                 int __err = errno;
1308
1309                 perror("shmat");
1310                 return __err;
1311         }
1312 #else
1313         /*
1314          * HP-UX inherits shm mappings?
1315          */
1316         data = threads;
1317 #endif
1318
1319         td = data + offset * sizeof(struct thread_data);
1320         ret = thread_main(td);
1321         shmdt(data);
1322         return (int) (uintptr_t) ret;
1323 }
1324
1325 /*
1326  * Run over the job map and reap the threads that have exited, if any.
1327  */
1328 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1329                          unsigned int *m_rate)
1330 {
1331         struct thread_data *td;
1332         unsigned int cputhreads, realthreads, pending;
1333         int i, status, ret;
1334
1335         /*
1336          * reap exited threads (TD_EXITED -> TD_REAPED)
1337          */
1338         realthreads = pending = cputhreads = 0;
1339         for_each_td(td, i) {
1340                 int flags = 0;
1341
1342                 /*
1343                  * ->io_ops is NULL for a thread that has closed its
1344                  * io engine
1345                  */
1346                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1347                         cputhreads++;
1348                 else
1349                         realthreads++;
1350
1351                 if (!td->pid) {
1352                         pending++;
1353                         continue;
1354                 }
1355                 if (td->runstate == TD_REAPED)
1356                         continue;
1357                 if (td->o.use_thread) {
1358                         if (td->runstate == TD_EXITED) {
1359                                 td_set_runstate(td, TD_REAPED);
1360                                 goto reaped;
1361                         }
1362                         continue;
1363                 }
1364
1365                 flags = WNOHANG;
1366                 if (td->runstate == TD_EXITED)
1367                         flags = 0;
1368
1369                 /*
1370                  * check if someone quit or got killed in an unusual way
1371                  */
1372                 ret = waitpid(td->pid, &status, flags);
1373                 if (ret < 0) {
1374                         if (errno == ECHILD) {
1375                                 log_err("fio: pid=%d disappeared %d\n",
1376                                                 (int) td->pid, td->runstate);
1377                                 td->sig = ECHILD;
1378                                 td_set_runstate(td, TD_REAPED);
1379                                 goto reaped;
1380                         }
1381                         perror("waitpid");
1382                 } else if (ret == td->pid) {
1383                         if (WIFSIGNALED(status)) {
1384                                 int sig = WTERMSIG(status);
1385
1386                                 if (sig != SIGTERM)
1387                                         log_err("fio: pid=%d, got signal=%d\n",
1388                                                         (int) td->pid, sig);
1389                                 td->sig = sig;
1390                                 td_set_runstate(td, TD_REAPED);
1391                                 goto reaped;
1392                         }
1393                         if (WIFEXITED(status)) {
1394                                 if (WEXITSTATUS(status) && !td->error)
1395                                         td->error = WEXITSTATUS(status);
1396
1397                                 td_set_runstate(td, TD_REAPED);
1398                                 goto reaped;
1399                         }
1400                 }
1401
1402                 /*
1403                  * thread is not dead, continue
1404                  */
1405                 pending++;
1406                 continue;
1407 reaped:
1408                 (*nr_running)--;
1409                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1410                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1411                 if (!td->pid)
1412                         pending--;
1413
1414                 if (td->error)
1415                         exit_value++;
1416
1417                 done_secs += mtime_since_now(&td->epoch) / 1000;
1418         }
1419
1420         if (*nr_running == cputhreads && !pending && realthreads)
1421                 fio_terminate_threads(TERMINATE_ALL);
1422 }
1423
1424 /*
1425  * Main function for kicking off and reaping jobs, as needed.
1426  */
1427 static void run_threads(void)
1428 {
1429         struct thread_data *td;
1430         unsigned long spent;
1431         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1432
1433         if (fio_pin_memory())
1434                 return;
1435
1436         if (fio_gtod_offload && fio_start_gtod_thread())
1437                 return;
1438
1439         set_sig_handlers();
1440
1441         if (output_format == FIO_OUTPUT_NORMAL) {
1442                 log_info("Starting ");
1443                 if (nr_thread)
1444                         log_info("%d thread%s", nr_thread,
1445                                                 nr_thread > 1 ? "s" : "");
1446                 if (nr_process) {
1447                         if (nr_thread)
1448                                 log_info(" and ");
1449                         log_info("%d process%s", nr_process,
1450                                                 nr_process > 1 ? "es" : "");
1451                 }
1452                 log_info("\n");
1453                 fflush(stdout);
1454         }
1455
1456         todo = thread_number;
1457         nr_running = 0;
1458         nr_started = 0;
1459         m_rate = t_rate = 0;
1460
1461         for_each_td(td, i) {
1462                 print_status_init(td->thread_number - 1);
1463
1464                 if (!td->o.create_serialize)
1465                         continue;
1466
1467                 /*
1468                  * do file setup here so it happens sequentially,
1469                  * we don't want X number of threads getting their
1470                  * client data interspersed on disk
1471                  */
1472                 if (setup_files(td)) {
1473                         exit_value++;
1474                         if (td->error)
1475                                 log_err("fio: pid=%d, err=%d/%s\n",
1476                                         (int) td->pid, td->error, td->verror);
1477                         td_set_runstate(td, TD_REAPED);
1478                         todo--;
1479                 } else {
1480                         struct fio_file *f;
1481                         unsigned int j;
1482
1483                         /*
1484                          * for sharing to work, each job must always open
1485                          * its own files. so close them, if we opened them
1486                          * for creation
1487                          */
1488                         for_each_file(td, f, j) {
1489                                 if (fio_file_open(f))
1490                                         td_io_close_file(td, f);
1491                         }
1492                 }
1493         }
1494
1495         set_genesis_time();
1496
1497         while (todo) {
1498                 struct thread_data *map[REAL_MAX_JOBS];
1499                 struct timeval this_start;
1500                 int this_jobs = 0, left;
1501
1502                 /*
1503                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1504                  */
1505                 for_each_td(td, i) {
1506                         if (td->runstate != TD_NOT_CREATED)
1507                                 continue;
1508
1509                         /*
1510                          * never got a chance to start, killed by other
1511                          * thread for some reason
1512                          */
1513                         if (td->terminate) {
1514                                 todo--;
1515                                 continue;
1516                         }
1517
1518                         if (td->o.start_delay) {
1519                                 spent = mtime_since_genesis();
1520
1521                                 if (td->o.start_delay * 1000 > spent)
1522                                         continue;
1523                         }
1524
1525                         if (td->o.stonewall && (nr_started || nr_running)) {
1526                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1527                                                         td->o.name);
1528                                 break;
1529                         }
1530
1531                         init_disk_util(td);
1532
1533                         /*
1534                          * Set state to created. Thread will transition
1535                          * to TD_INITIALIZED when it's done setting up.
1536                          */
1537                         td_set_runstate(td, TD_CREATED);
1538                         map[this_jobs++] = td;
1539                         nr_started++;
1540
1541                         if (td->o.use_thread) {
1542                                 int ret;
1543
1544                                 dprint(FD_PROCESS, "will pthread_create\n");
1545                                 ret = pthread_create(&td->thread, NULL,
1546                                                         thread_main, td);
1547                                 if (ret) {
1548                                         log_err("pthread_create: %s\n",
1549                                                         strerror(ret));
1550                                         nr_started--;
1551                                         break;
1552                                 }
1553                                 ret = pthread_detach(td->thread);
1554                                 if (ret)
1555                                         log_err("pthread_detach: %s",
1556                                                         strerror(ret));
1557                         } else {
1558                                 pid_t pid;
1559                                 dprint(FD_PROCESS, "will fork\n");
1560                                 pid = fork();
1561                                 if (!pid) {
1562                                         int ret = fork_main(shm_id, i);
1563
1564                                         _exit(ret);
1565                                 } else if (i == fio_debug_jobno)
1566                                         *fio_debug_jobp = pid;
1567                         }
1568                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1569                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1570                                 log_err("fio: job startup hung? exiting.\n");
1571                                 fio_terminate_threads(TERMINATE_ALL);
1572                                 fio_abort = 1;
1573                                 nr_started--;
1574                                 break;
1575                         }
1576                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1577                 }
1578
1579                 /*
1580                  * Wait for the started threads to transition to
1581                  * TD_INITIALIZED.
1582                  */
1583                 fio_gettime(&this_start, NULL);
1584                 left = this_jobs;
1585                 while (left && !fio_abort) {
1586                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1587                                 break;
1588
1589                         usleep(100000);
1590
1591                         for (i = 0; i < this_jobs; i++) {
1592                                 td = map[i];
1593                                 if (!td)
1594                                         continue;
1595                                 if (td->runstate == TD_INITIALIZED) {
1596                                         map[i] = NULL;
1597                                         left--;
1598                                 } else if (td->runstate >= TD_EXITED) {
1599                                         map[i] = NULL;
1600                                         left--;
1601                                         todo--;
1602                                         nr_running++; /* work-around... */
1603                                 }
1604                         }
1605                 }
1606
1607                 if (left) {
1608                         log_err("fio: %d job%s failed to start\n", left,
1609                                         left > 1 ? "s" : "");
1610                         for (i = 0; i < this_jobs; i++) {
1611                                 td = map[i];
1612                                 if (!td)
1613                                         continue;
1614                                 kill(td->pid, SIGTERM);
1615                         }
1616                         break;
1617                 }
1618
1619                 /*
1620                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1621                  */
1622                 for_each_td(td, i) {
1623                         if (td->runstate != TD_INITIALIZED)
1624                                 continue;
1625
1626                         if (in_ramp_time(td))
1627                                 td_set_runstate(td, TD_RAMP);
1628                         else
1629                                 td_set_runstate(td, TD_RUNNING);
1630                         nr_running++;
1631                         nr_started--;
1632                         m_rate += ddir_rw_sum(td->o.ratemin);
1633                         t_rate += ddir_rw_sum(td->o.rate);
1634                         todo--;
1635                         fio_mutex_up(td->mutex);
1636                 }
1637
1638                 reap_threads(&nr_running, &t_rate, &m_rate);
1639
1640                 if (todo) {
1641                         if (is_backend)
1642                                 fio_server_idle_loop();
1643                         else
1644                                 usleep(100000);
1645                 }
1646         }
1647
1648         while (nr_running) {
1649                 reap_threads(&nr_running, &t_rate, &m_rate);
1650
1651                 if (is_backend)
1652                         fio_server_idle_loop();
1653                 else
1654                         usleep(10000);
1655         }
1656
1657         update_io_ticks();
1658         fio_unpin_memory();
1659 }
1660
1661 void wait_for_disk_thread_exit(void)
1662 {
1663         fio_mutex_down(disk_thread_mutex);
1664 }
1665
1666 static void free_disk_util(void)
1667 {
1668         disk_util_start_exit();
1669         wait_for_disk_thread_exit();
1670         disk_util_prune_entries();
1671 }
1672
1673 static void *disk_thread_main(void *data)
1674 {
1675         int ret = 0;
1676
1677         fio_mutex_up(startup_mutex);
1678
1679         while (threads && !ret) {
1680                 usleep(DISK_UTIL_MSEC * 1000);
1681                 if (!threads)
1682                         break;
1683                 ret = update_io_ticks();
1684
1685                 if (!is_backend)
1686                         print_thread_status();
1687         }
1688
1689         fio_mutex_up(disk_thread_mutex);
1690         return NULL;
1691 }
1692
1693 static int create_disk_util_thread(void)
1694 {
1695         int ret;
1696
1697         setup_disk_util();
1698
1699         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1700
1701         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1702         if (ret) {
1703                 fio_mutex_remove(disk_thread_mutex);
1704                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1705                 return 1;
1706         }
1707
1708         ret = pthread_detach(disk_util_thread);
1709         if (ret) {
1710                 fio_mutex_remove(disk_thread_mutex);
1711                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1712                 return 1;
1713         }
1714
1715         dprint(FD_MUTEX, "wait on startup_mutex\n");
1716         fio_mutex_down(startup_mutex);
1717         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1718         return 0;
1719 }
1720
1721 int fio_backend(void)
1722 {
1723         struct thread_data *td;
1724         int i;
1725
1726         if (exec_profile) {
1727                 if (load_profile(exec_profile))
1728                         return 1;
1729                 free(exec_profile);
1730                 exec_profile = NULL;
1731         }
1732         if (!thread_number)
1733                 return 0;
1734
1735         if (write_bw_log) {
1736                 setup_log(&agg_io_log[DDIR_READ], 0);
1737                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1738                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1739         }
1740
1741         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1742         if (startup_mutex == NULL)
1743                 return 1;
1744         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1745         if (writeout_mutex == NULL)
1746                 return 1;
1747
1748         set_genesis_time();
1749         create_disk_util_thread();
1750
1751         cgroup_list = smalloc(sizeof(*cgroup_list));
1752         INIT_FLIST_HEAD(cgroup_list);
1753
1754         run_threads();
1755
1756         if (!fio_abort) {
1757                 show_run_stats();
1758                 if (write_bw_log) {
1759                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1760                         __finish_log(agg_io_log[DDIR_WRITE],
1761                                         "agg-write_bw.log");
1762                         __finish_log(agg_io_log[DDIR_TRIM],
1763                                         "agg-write_bw.log");
1764                 }
1765         }
1766
1767         for_each_td(td, i)
1768                 fio_options_free(td);
1769
1770         free_disk_util();
1771         cgroup_kill(cgroup_list);
1772         sfree(cgroup_list);
1773         sfree(cgroup_mnt);
1774
1775         fio_mutex_remove(startup_mutex);
1776         fio_mutex_remove(writeout_mutex);
1777         fio_mutex_remove(disk_thread_mutex);
1778         return exit_value;
1779 }