backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/shm.h>
  38 #include <sys/mman.h>
  39
  40 #include "fio.h"
  41 #include "hash.h"
  42 #include "smalloc.h"
  43 #include "verify.h"
  44 #include "trim.h"
  45 #include "diskutil.h"
  46 #include "cgroup.h"
  47 #include "profile.h"
  48 #include "lib/rand.h"
  49 #include "memalign.h"
  50 #include "server.h"
  51
  52 static pthread_t disk_util_thread;
  53 static struct fio_mutex *disk_thread_mutex;
  54 static struct fio_mutex *startup_mutex;
  55 static struct fio_mutex *writeout_mutex;
  56 static struct flist_head *cgroup_list;
  57 static char *cgroup_mnt;
  58 static int exit_value;
  59 static volatile int fio_abort;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int nr_process = 0;
  66 unsigned int nr_thread = 0;
  67 int shm_id = 0;
  68 int temp_stall_ts;
  69 unsigned long done_secs = 0;
  70 volatile int disk_util_exit = 0;
  71
  72 #define PAGE_ALIGN(buf) \
  73         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  74
  75 #define JOB_START_TIMEOUT       (5 * 1000)
  76
  77 static void sig_int(int sig)
  78 {
  79         if (threads) {
  80                 if (is_backend)
  81                         fio_server_got_signal(sig);
  82                 else {
  83                         log_info("\nfio: terminating on signal %d\n", sig);
  84                         fflush(stdout);
  85                         exit_value = 128;
  86                 }
  87
  88                 fio_terminate_threads(TERMINATE_ALL);
  89         }
  90 }
  91
  92 static void sig_show_status(int sig)
  93 {
  94         show_running_run_stats();
  95 }
  96
  97 static void set_sig_handlers(void)
  98 {
  99         struct sigaction act;
 100
 101         memset(&act, 0, sizeof(act));
 102         act.sa_handler = sig_int;
 103         act.sa_flags = SA_RESTART;
 104         sigaction(SIGINT, &act, NULL);
 105
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGTERM, &act, NULL);
 110
 111 /* Windows uses SIGBREAK as a quit signal from other applications */
 112 #ifdef WIN32
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_int;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGBREAK, &act, NULL);
 117 #endif
 118
 119         memset(&act, 0, sizeof(act));
 120         act.sa_handler = sig_show_status;
 121         act.sa_flags = SA_RESTART;
 122         sigaction(SIGUSR1, &act, NULL);
 123
 124         if (is_backend) {
 125                 memset(&act, 0, sizeof(act));
 126                 act.sa_handler = sig_int;
 127                 act.sa_flags = SA_RESTART;
 128                 sigaction(SIGPIPE, &act, NULL);
 129         }
 130 }
 131
 132 /*
 133  * Check if we are above the minimum rate given.
 134  */
 135 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 136                             enum fio_ddir ddir)
 137 {
 138         unsigned long long bytes = 0;
 139         unsigned long iops = 0;
 140         unsigned long spent;
 141         unsigned long rate;
 142         unsigned int ratemin = 0;
 143         unsigned int rate_iops = 0;
 144         unsigned int rate_iops_min = 0;
 145
 146         assert(ddir_rw(ddir));
 147
 148         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 149                 return 0;
 150
 151         /*
 152          * allow a 2 second settle period in the beginning
 153          */
 154         if (mtime_since(&td->start, now) < 2000)
 155                 return 0;
 156
 157         iops += td->this_io_blocks[ddir];
 158         bytes += td->this_io_bytes[ddir];
 159         ratemin += td->o.ratemin[ddir];
 160         rate_iops += td->o.rate_iops[ddir];
 161         rate_iops_min += td->o.rate_iops_min[ddir];
 162
 163         /*
 164          * if rate blocks is set, sample is running
 165          */
 166         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 167                 spent = mtime_since(&td->lastrate[ddir], now);
 168                 if (spent < td->o.ratecycle)
 169                         return 0;
 170
 171                 if (td->o.rate[ddir]) {
 172                         /*
 173                          * check bandwidth specified rate
 174                          */
 175                         if (bytes < td->rate_bytes[ddir]) {
 176                                 log_err("%s: min rate %u not met\n", td->o.name,
 177                                                                 ratemin);
 178                                 return 1;
 179                         } else {
 180                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 181                                 if (rate < ratemin ||
 182                                     bytes < td->rate_bytes[ddir]) {
 183                                         log_err("%s: min rate %u not met, got"
 184                                                 " %luKB/sec\n", td->o.name,
 185                                                         ratemin, rate);
 186                                         return 1;
 187                                 }
 188                         }
 189                 } else {
 190                         /*
 191                          * checks iops specified rate
 192                          */
 193                         if (iops < rate_iops) {
 194                                 log_err("%s: min iops rate %u not met\n",
 195                                                 td->o.name, rate_iops);
 196                                 return 1;
 197                         } else {
 198                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 199                                 if (rate < rate_iops_min ||
 200                                     iops < td->rate_blocks[ddir]) {
 201                                         log_err("%s: min iops rate %u not met,"
 202                                                 " got %lu\n", td->o.name,
 203                                                         rate_iops_min, rate);
 204                                 }
 205                         }
 206                 }
 207         }
 208
 209         td->rate_bytes[ddir] = bytes;
 210         td->rate_blocks[ddir] = iops;
 211         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 212         return 0;
 213 }
 214
 215 static int check_min_rate(struct thread_data *td, struct timeval *now,
 216                           unsigned long *bytes_done)
 217 {
 218         int ret = 0;
 219
 220         if (bytes_done[DDIR_READ])
 221                 ret |= __check_min_rate(td, now, DDIR_READ);
 222         if (bytes_done[DDIR_WRITE])
 223                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 224         if (bytes_done[DDIR_TRIM])
 225                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 226
 227         return ret;
 228 }
 229
 230 /*
 231  * When job exits, we can cancel the in-flight IO if we are using async
 232  * io. Attempt to do so.
 233  */
 234 static void cleanup_pending_aio(struct thread_data *td)
 235 {
 236         struct flist_head *entry, *n;
 237         struct io_u *io_u;
 238         int r;
 239
 240         /*
 241          * get immediately available events, if any
 242          */
 243         r = io_u_queued_complete(td, 0, NULL);
 244         if (r < 0)
 245                 return;
 246
 247         /*
 248          * now cancel remaining active events
 249          */
 250         if (td->io_ops->cancel) {
 251                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 252                         io_u = flist_entry(entry, struct io_u, list);
 253
 254                         /*
 255                          * if the io_u isn't in flight, then that generally
 256                          * means someone leaked an io_u. complain but fix
 257                          * it up, so we don't stall here.
 258                          */
 259                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 260                                 log_err("fio: non-busy IO on busy list\n");
 261                                 put_io_u(td, io_u);
 262                         } else {
 263                                 r = td->io_ops->cancel(td, io_u);
 264                                 if (!r)
 265                                         put_io_u(td, io_u);
 266                         }
 267                 }
 268         }
 269
 270         if (td->cur_depth)
 271                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 272 }
 273
 274 /*
 275  * Helper to handle the final sync of a file. Works just like the normal
 276  * io path, just does everything sync.
 277  */
 278 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 279 {
 280         struct io_u *io_u = __get_io_u(td);
 281         int ret;
 282
 283         if (!io_u)
 284                 return 1;
 285
 286         io_u->ddir = DDIR_SYNC;
 287         io_u->file = f;
 288
 289         if (td_io_prep(td, io_u)) {
 290                 put_io_u(td, io_u);
 291                 return 1;
 292         }
 293
 294 requeue:
 295         ret = td_io_queue(td, io_u);
 296         if (ret < 0) {
 297                 td_verror(td, io_u->error, "td_io_queue");
 298                 put_io_u(td, io_u);
 299                 return 1;
 300         } else if (ret == FIO_Q_QUEUED) {
 301                 if (io_u_queued_complete(td, 1, NULL) < 0)
 302                         return 1;
 303         } else if (ret == FIO_Q_COMPLETED) {
 304                 if (io_u->error) {
 305                         td_verror(td, io_u->error, "td_io_queue");
 306                         return 1;
 307                 }
 308
 309                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 310                         return 1;
 311         } else if (ret == FIO_Q_BUSY) {
 312                 if (td_io_commit(td))
 313                         return 1;
 314                 goto requeue;
 315         }
 316
 317         return 0;
 318 }
 319
 320 static inline void __update_tv_cache(struct thread_data *td)
 321 {
 322         fio_gettime(&td->tv_cache, NULL);
 323 }
 324
 325 static inline void update_tv_cache(struct thread_data *td)
 326 {
 327         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 328                 __update_tv_cache(td);
 329 }
 330
 331 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 332 {
 333         if (in_ramp_time(td))
 334                 return 0;
 335         if (!td->o.timeout)
 336                 return 0;
 337         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 338                 return 1;
 339
 340         return 0;
 341 }
 342
 343 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 344                                int *retptr)
 345 {
 346         int ret = *retptr;
 347
 348         if (ret < 0 || td->error) {
 349                 int err = td->error;
 350                 enum error_type_bit eb;
 351
 352                 if (ret < 0)
 353                         err = -ret;
 354
 355                 eb = td_error_type(ddir, err);
 356                 if (!(td->o.continue_on_error & (1 << eb)))
 357                         return 1;
 358
 359                 if (td_non_fatal_error(td, eb, err)) {
 360                         /*
 361                          * Continue with the I/Os in case of
 362                          * a non fatal error.
 363                          */
 364                         update_error_count(td, err);
 365                         td_clear_error(td);
 366                         *retptr = 0;
 367                         return 0;
 368                 } else if (td->o.fill_device && err == ENOSPC) {
 369                         /*
 370                          * We expect to hit this error if
 371                          * fill_device option is set.
 372                          */
 373                         td_clear_error(td);
 374                         td->terminate = 1;
 375                         return 1;
 376                 } else {
 377                         /*
 378                          * Stop the I/O in case of a fatal
 379                          * error.
 380                          */
 381                         update_error_count(td, err);
 382                         return 1;
 383                 }
 384         }
 385
 386         return 0;
 387 }
 388
 389 /*
 390  * The main verify engine. Runs over the writes we previously submitted,
 391  * reads the blocks back in, and checks the crc/md5 of the data.
 392  */
 393 static void do_verify(struct thread_data *td)
 394 {
 395         struct fio_file *f;
 396         struct io_u *io_u;
 397         int ret, min_events;
 398         unsigned int i;
 399
 400         dprint(FD_VERIFY, "starting loop\n");
 401
 402         /*
 403          * sync io first and invalidate cache, to make sure we really
 404          * read from disk.
 405          */
 406         for_each_file(td, f, i) {
 407                 if (!fio_file_open(f))
 408                         continue;
 409                 if (fio_io_sync(td, f))
 410                         break;
 411                 if (file_invalidate_cache(td, f))
 412                         break;
 413         }
 414
 415         if (td->error)
 416                 return;
 417
 418         td_set_runstate(td, TD_VERIFYING);
 419
 420         io_u = NULL;
 421         while (!td->terminate) {
 422                 int ret2, full;
 423
 424                 update_tv_cache(td);
 425
 426                 if (runtime_exceeded(td, &td->tv_cache)) {
 427                         __update_tv_cache(td);
 428                         if (runtime_exceeded(td, &td->tv_cache)) {
 429                                 td->terminate = 1;
 430                                 break;
 431                         }
 432                 }
 433
 434                 if (flow_threshold_exceeded(td))
 435                         continue;
 436
 437                 io_u = __get_io_u(td);
 438                 if (!io_u)
 439                         break;
 440
 441                 if (get_next_verify(td, io_u)) {
 442                         put_io_u(td, io_u);
 443                         break;
 444                 }
 445
 446                 if (td_io_prep(td, io_u)) {
 447                         put_io_u(td, io_u);
 448                         break;
 449                 }
 450
 451                 if (td->o.verify_async)
 452                         io_u->end_io = verify_io_u_async;
 453                 else
 454                         io_u->end_io = verify_io_u;
 455
 456                 ret = td_io_queue(td, io_u);
 457                 switch (ret) {
 458                 case FIO_Q_COMPLETED:
 459                         if (io_u->error) {
 460                                 ret = -io_u->error;
 461                                 clear_io_u(td, io_u);
 462                         } else if (io_u->resid) {
 463                                 int bytes = io_u->xfer_buflen - io_u->resid;
 464
 465                                 /*
 466                                  * zero read, fail
 467                                  */
 468                                 if (!bytes) {
 469                                         td_verror(td, EIO, "full resid");
 470                                         put_io_u(td, io_u);
 471                                         break;
 472                                 }
 473
 474                                 io_u->xfer_buflen = io_u->resid;
 475                                 io_u->xfer_buf += bytes;
 476                                 io_u->offset += bytes;
 477
 478                                 if (ddir_rw(io_u->ddir))
 479                                         td->ts.short_io_u[io_u->ddir]++;
 480
 481                                 f = io_u->file;
 482                                 if (io_u->offset == f->real_file_size)
 483                                         goto sync_done;
 484
 485                                 requeue_io_u(td, &io_u);
 486                         } else {
 487 sync_done:
 488                                 ret = io_u_sync_complete(td, io_u, NULL);
 489                                 if (ret < 0)
 490                                         break;
 491                         }
 492                         continue;
 493                 case FIO_Q_QUEUED:
 494                         break;
 495                 case FIO_Q_BUSY:
 496                         requeue_io_u(td, &io_u);
 497                         ret2 = td_io_commit(td);
 498                         if (ret2 < 0)
 499                                 ret = ret2;
 500                         break;
 501                 default:
 502                         assert(ret < 0);
 503                         td_verror(td, -ret, "td_io_queue");
 504                         break;
 505                 }
 506
 507                 if (break_on_this_error(td, io_u->ddir, &ret))
 508                         break;
 509
 510                 /*
 511                  * if we can queue more, do so. but check if there are
 512                  * completed io_u's first. Note that we can get BUSY even
 513                  * without IO queued, if the system is resource starved.
 514                  */
 515                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 516                 if (full || !td->o.iodepth_batch_complete) {
 517                         min_events = min(td->o.iodepth_batch_complete,
 518                                          td->cur_depth);
 519                         /*
 520                          * if the queue is full, we MUST reap at least 1 event
 521                          */
 522                         if (full && !min_events)
 523                                 min_events = 1;
 524
 525                         do {
 526                                 /*
 527                                  * Reap required number of io units, if any,
 528                                  * and do the verification on them through
 529                                  * the callback handler
 530                                  */
 531                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 532                                         ret = -1;
 533                                         break;
 534                                 }
 535                         } while (full && (td->cur_depth > td->o.iodepth_low));
 536                 }
 537                 if (ret < 0)
 538                         break;
 539         }
 540
 541         if (!td->error) {
 542                 min_events = td->cur_depth;
 543
 544                 if (min_events)
 545                         ret = io_u_queued_complete(td, min_events, NULL);
 546         } else
 547                 cleanup_pending_aio(td);
 548
 549         td_set_runstate(td, TD_RUNNING);
 550
 551         dprint(FD_VERIFY, "exiting loop\n");
 552 }
 553
 554 static int io_bytes_exceeded(struct thread_data *td)
 555 {
 556         unsigned long long bytes;
 557
 558         if (td_rw(td))
 559                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 560         else if (td_write(td))
 561                 bytes = td->this_io_bytes[DDIR_WRITE];
 562         else if (td_read(td))
 563                 bytes = td->this_io_bytes[DDIR_READ];
 564         else
 565                 bytes = td->this_io_bytes[DDIR_TRIM];
 566
 567         return bytes >= td->o.size;
 568 }
 569
 570 /*
 571  * Main IO worker function. It retrieves io_u's to process and queues
 572  * and reaps them, checking for rate and errors along the way.
 573  */
 574 static void do_io(struct thread_data *td)
 575 {
 576         unsigned int i;
 577         int ret = 0;
 578
 579         if (in_ramp_time(td))
 580                 td_set_runstate(td, TD_RAMP);
 581         else
 582                 td_set_runstate(td, TD_RUNNING);
 583
 584         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 585                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 586                 td->o.time_based) {
 587                 struct timeval comp_time;
 588                 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 589                 int min_evts = 0;
 590                 struct io_u *io_u;
 591                 int ret2, full;
 592                 enum fio_ddir ddir;
 593
 594                 if (td->terminate || td->done)
 595                         break;
 596
 597                 update_tv_cache(td);
 598
 599                 if (runtime_exceeded(td, &td->tv_cache)) {
 600                         __update_tv_cache(td);
 601                         if (runtime_exceeded(td, &td->tv_cache)) {
 602                                 td->terminate = 1;
 603                                 break;
 604                         }
 605                 }
 606
 607                 if (flow_threshold_exceeded(td))
 608                         continue;
 609
 610                 io_u = get_io_u(td);
 611                 if (!io_u)
 612                         break;
 613
 614                 ddir = io_u->ddir;
 615
 616                 /*
 617                  * Add verification end_io handler if:
 618                  *      - Asked to verify (!td_rw(td))
 619                  *      - Or the io_u is from our verify list (mixed write/ver)
 620                  */
 621                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 622                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 623                         if (td->o.verify_async)
 624                                 io_u->end_io = verify_io_u_async;
 625                         else
 626                                 io_u->end_io = verify_io_u;
 627                         td_set_runstate(td, TD_VERIFYING);
 628                 } else if (in_ramp_time(td))
 629                         td_set_runstate(td, TD_RAMP);
 630                 else
 631                         td_set_runstate(td, TD_RUNNING);
 632
 633                 ret = td_io_queue(td, io_u);
 634                 switch (ret) {
 635                 case FIO_Q_COMPLETED:
 636                         if (io_u->error) {
 637                                 ret = -io_u->error;
 638                                 clear_io_u(td, io_u);
 639                         } else if (io_u->resid) {
 640                                 int bytes = io_u->xfer_buflen - io_u->resid;
 641                                 struct fio_file *f = io_u->file;
 642
 643                                 /*
 644                                  * zero read, fail
 645                                  */
 646                                 if (!bytes) {
 647                                         td_verror(td, EIO, "full resid");
 648                                         put_io_u(td, io_u);
 649                                         break;
 650                                 }
 651
 652                                 io_u->xfer_buflen = io_u->resid;
 653                                 io_u->xfer_buf += bytes;
 654                                 io_u->offset += bytes;
 655
 656                                 if (ddir_rw(io_u->ddir))
 657                                         td->ts.short_io_u[io_u->ddir]++;
 658
 659                                 if (io_u->offset == f->real_file_size)
 660                                         goto sync_done;
 661
 662                                 requeue_io_u(td, &io_u);
 663                         } else {
 664 sync_done:
 665                                 if (__should_check_rate(td, DDIR_READ) ||
 666                                     __should_check_rate(td, DDIR_WRITE) ||
 667                                     __should_check_rate(td, DDIR_TRIM))
 668                                         fio_gettime(&comp_time, NULL);
 669
 670                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 671                                 if (ret < 0)
 672                                         break;
 673                         }
 674                         break;
 675                 case FIO_Q_QUEUED:
 676                         /*
 677                          * if the engine doesn't have a commit hook,
 678                          * the io_u is really queued. if it does have such
 679                          * a hook, it has to call io_u_queued() itself.
 680                          */
 681                         if (td->io_ops->commit == NULL)
 682                                 io_u_queued(td, io_u);
 683                         break;
 684                 case FIO_Q_BUSY:
 685                         requeue_io_u(td, &io_u);
 686                         ret2 = td_io_commit(td);
 687                         if (ret2 < 0)
 688                                 ret = ret2;
 689                         break;
 690                 default:
 691                         assert(ret < 0);
 692                         put_io_u(td, io_u);
 693                         break;
 694                 }
 695
 696                 if (break_on_this_error(td, ddir, &ret))
 697                         break;
 698
 699                 /*
 700                  * See if we need to complete some commands. Note that we
 701                  * can get BUSY even without IO queued, if the system is
 702                  * resource starved.
 703                  */
 704                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 705                 if (full || !td->o.iodepth_batch_complete) {
 706                         min_evts = min(td->o.iodepth_batch_complete,
 707                                         td->cur_depth);
 708                         /*
 709                          * if the queue is full, we MUST reap at least 1 event
 710                          */
 711                         if (full && !min_evts)
 712                                 min_evts = 1;
 713
 714                         if (__should_check_rate(td, DDIR_READ) ||
 715                             __should_check_rate(td, DDIR_WRITE) ||
 716                             __should_check_rate(td, DDIR_TRIM))
 717                                 fio_gettime(&comp_time, NULL);
 718
 719                         do {
 720                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 721                                 if (ret < 0)
 722                                         break;
 723
 724                         } while (full && (td->cur_depth > td->o.iodepth_low));
 725                 }
 726
 727                 if (ret < 0)
 728                         break;
 729                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 730                         continue;
 731
 732                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 733                         if (check_min_rate(td, &comp_time, bytes_done)) {
 734                                 if (exitall_on_terminate)
 735                                         fio_terminate_threads(td->groupid);
 736                                 td_verror(td, EIO, "check_min_rate");
 737                                 break;
 738                         }
 739                 }
 740
 741                 if (td->o.thinktime) {
 742                         unsigned long long b;
 743
 744                         b = ddir_rw_sum(td->io_blocks);
 745                         if (!(b % td->o.thinktime_blocks)) {
 746                                 int left;
 747
 748                                 if (td->o.thinktime_spin)
 749                                         usec_spin(td->o.thinktime_spin);
 750
 751                                 left = td->o.thinktime - td->o.thinktime_spin;
 752                                 if (left)
 753                                         usec_sleep(td, left);
 754                         }
 755                 }
 756         }
 757
 758         if (td->trim_entries)
 759                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 760
 761         if (td->o.fill_device && td->error == ENOSPC) {
 762                 td->error = 0;
 763                 td->terminate = 1;
 764         }
 765         if (!td->error) {
 766                 struct fio_file *f;
 767
 768                 i = td->cur_depth;
 769                 if (i) {
 770                         ret = io_u_queued_complete(td, i, NULL);
 771                         if (td->o.fill_device && td->error == ENOSPC)
 772                                 td->error = 0;
 773                 }
 774
 775                 if (should_fsync(td) && td->o.end_fsync) {
 776                         td_set_runstate(td, TD_FSYNCING);
 777
 778                         for_each_file(td, f, i) {
 779                                 if (!fio_file_open(f))
 780                                         continue;
 781                                 fio_io_sync(td, f);
 782                         }
 783                 }
 784         } else
 785                 cleanup_pending_aio(td);
 786
 787         /*
 788          * stop job if we failed doing any IO
 789          */
 790         if (!ddir_rw_sum(td->this_io_bytes))
 791                 td->done = 1;
 792 }
 793
 794 static void cleanup_io_u(struct thread_data *td)
 795 {
 796         struct flist_head *entry, *n;
 797         struct io_u *io_u;
 798
 799         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 800                 io_u = flist_entry(entry, struct io_u, list);
 801
 802                 flist_del(&io_u->list);
 803                 fio_memfree(io_u, sizeof(*io_u));
 804         }
 805
 806         free_io_mem(td);
 807 }
 808
 809 static int init_io_u(struct thread_data *td)
 810 {
 811         struct io_u *io_u;
 812         unsigned int max_bs, min_write;
 813         int cl_align, i, max_units;
 814         int data_xfer = 1;
 815         char *p;
 816
 817         max_units = td->o.iodepth;
 818         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 819         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
 820         min_write = td->o.min_bs[DDIR_WRITE];
 821         td->orig_buffer_size = (unsigned long long) max_bs
 822                                         * (unsigned long long) max_units;
 823
 824         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 825                 data_xfer = 0;
 826
 827         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 828                 unsigned long bs;
 829
 830                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 831                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 832         }
 833
 834         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 835                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 836                 return 1;
 837         }
 838
 839         if (data_xfer && allocate_io_mem(td))
 840                 return 1;
 841
 842         if (td->o.odirect || td->o.mem_align ||
 843             (td->io_ops->flags & FIO_RAWIO))
 844                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 845         else
 846                 p = td->orig_buffer;
 847
 848         cl_align = os_cache_line_size();
 849
 850         for (i = 0; i < max_units; i++) {
 851                 void *ptr;
 852
 853                 if (td->terminate)
 854                         return 1;
 855
 856                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 857                 if (!ptr) {
 858                         log_err("fio: unable to allocate aligned memory\n");
 859                         break;
 860                 }
 861
 862                 io_u = ptr;
 863                 memset(io_u, 0, sizeof(*io_u));
 864                 INIT_FLIST_HEAD(&io_u->list);
 865                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 866
 867                 if (data_xfer) {
 868                         io_u->buf = p;
 869                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 870
 871                         if (td_write(td))
 872                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 873                         if (td_write(td) && td->o.verify_pattern_bytes) {
 874                                 /*
 875                                  * Fill the buffer with the pattern if we are
 876                                  * going to be doing writes.
 877                                  */
 878                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 879                         }
 880                 }
 881
 882                 io_u->index = i;
 883                 io_u->flags = IO_U_F_FREE;
 884                 flist_add(&io_u->list, &td->io_u_freelist);
 885                 p += max_bs;
 886         }
 887
 888         return 0;
 889 }
 890
 891 static int switch_ioscheduler(struct thread_data *td)
 892 {
 893         char tmp[256], tmp2[128];
 894         FILE *f;
 895         int ret;
 896
 897         if (td->io_ops->flags & FIO_DISKLESSIO)
 898                 return 0;
 899
 900         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 901
 902         f = fopen(tmp, "r+");
 903         if (!f) {
 904                 if (errno == ENOENT) {
 905                         log_err("fio: os or kernel doesn't support IO scheduler"
 906                                 " switching\n");
 907                         return 0;
 908                 }
 909                 td_verror(td, errno, "fopen iosched");
 910                 return 1;
 911         }
 912
 913         /*
 914          * Set io scheduler.
 915          */
 916         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 917         if (ferror(f) || ret != 1) {
 918                 td_verror(td, errno, "fwrite");
 919                 fclose(f);
 920                 return 1;
 921         }
 922
 923         rewind(f);
 924
 925         /*
 926          * Read back and check that the selected scheduler is now the default.
 927          */
 928         ret = fread(tmp, 1, sizeof(tmp), f);
 929         if (ferror(f) || ret < 0) {
 930                 td_verror(td, errno, "fread");
 931                 fclose(f);
 932                 return 1;
 933         }
 934
 935         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 936         if (!strstr(tmp, tmp2)) {
 937                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 938                 td_verror(td, EINVAL, "iosched_switch");
 939                 fclose(f);
 940                 return 1;
 941         }
 942
 943         fclose(f);
 944         return 0;
 945 }
 946
 947 static int keep_running(struct thread_data *td)
 948 {
 949         if (td->done)
 950                 return 0;
 951         if (td->o.time_based)
 952                 return 1;
 953         if (td->o.loops) {
 954                 td->o.loops--;
 955                 return 1;
 956         }
 957
 958         if (ddir_rw_sum(td->io_bytes) < td->o.size)
 959                 return 1;
 960
 961         return 0;
 962 }
 963
 964 static int exec_string(const char *string)
 965 {
 966         int ret, newlen = strlen(string) + 1 + 8;
 967         char *str;
 968
 969         str = malloc(newlen);
 970         sprintf(str, "sh -c %s", string);
 971
 972         ret = system(str);
 973         if (ret == -1)
 974                 log_err("fio: exec of cmd <%s> failed\n", str);
 975
 976         free(str);
 977         return ret;
 978 }
 979
 980 /*
 981  * Entry point for the thread based jobs. The process based jobs end up
 982  * here as well, after a little setup.
 983  */
 984 static void *thread_main(void *data)
 985 {
 986         unsigned long long elapsed;
 987         struct thread_data *td = data;
 988         pthread_condattr_t attr;
 989         int clear_state;
 990
 991         if (!td->o.use_thread) {
 992                 setsid();
 993                 td->pid = getpid();
 994         } else
 995                 td->pid = gettid();
 996
 997         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 998
 999         INIT_FLIST_HEAD(&td->io_u_freelist);
1000         INIT_FLIST_HEAD(&td->io_u_busylist);
1001         INIT_FLIST_HEAD(&td->io_u_requeues);
1002         INIT_FLIST_HEAD(&td->io_log_list);
1003         INIT_FLIST_HEAD(&td->io_hist_list);
1004         INIT_FLIST_HEAD(&td->verify_list);
1005         INIT_FLIST_HEAD(&td->trim_list);
1006         pthread_mutex_init(&td->io_u_lock, NULL);
1007         td->io_hist_tree = RB_ROOT;
1008
1009         pthread_condattr_init(&attr);
1010         pthread_cond_init(&td->verify_cond, &attr);
1011         pthread_cond_init(&td->free_cond, &attr);
1012
1013         td_set_runstate(td, TD_INITIALIZED);
1014         dprint(FD_MUTEX, "up startup_mutex\n");
1015         fio_mutex_up(startup_mutex);
1016         dprint(FD_MUTEX, "wait on td->mutex\n");
1017         fio_mutex_down(td->mutex);
1018         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1019
1020         /*
1021          * the ->mutex mutex is now no longer used, close it to avoid
1022          * eating a file descriptor
1023          */
1024         fio_mutex_remove(td->mutex);
1025
1026         /*
1027          * A new gid requires privilege, so we need to do this before setting
1028          * the uid.
1029          */
1030         if (td->o.gid != -1U && setgid(td->o.gid)) {
1031                 td_verror(td, errno, "setgid");
1032                 goto err;
1033         }
1034         if (td->o.uid != -1U && setuid(td->o.uid)) {
1035                 td_verror(td, errno, "setuid");
1036                 goto err;
1037         }
1038
1039         /*
1040          * If we have a gettimeofday() thread, make sure we exclude that
1041          * thread from this job
1042          */
1043         if (td->o.gtod_cpu)
1044                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1045
1046         /*
1047          * Set affinity first, in case it has an impact on the memory
1048          * allocations.
1049          */
1050         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1051                 td_verror(td, errno, "cpu_set_affinity");
1052                 goto err;
1053         }
1054
1055 #ifdef FIO_HAVE_LIBNUMA
1056         /* numa node setup */
1057         if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1058                 int ret;
1059
1060                 if (numa_available() < 0) {
1061                         td_verror(td, errno, "Does not support NUMA API\n");
1062                         goto err;
1063                 }
1064
1065                 if (td->o.numa_cpumask_set) {
1066                         ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1067                         if (ret == -1) {
1068                                 td_verror(td, errno, \
1069                                         "numa_run_on_node_mask failed\n");
1070                                 goto err;
1071                         }
1072                 }
1073
1074                 if (td->o.numa_memmask_set) {
1075
1076                         switch (td->o.numa_mem_mode) {
1077                         case MPOL_INTERLEAVE:
1078                                 numa_set_interleave_mask(td->o.numa_memnodesmask);
1079                                 break;
1080                         case MPOL_BIND:
1081                                 numa_set_membind(td->o.numa_memnodesmask);
1082                                 break;
1083                         case MPOL_LOCAL:
1084                                 numa_set_localalloc();
1085                                 break;
1086                         case MPOL_PREFERRED:
1087                                 numa_set_preferred(td->o.numa_mem_prefer_node);
1088                                 break;
1089                         case MPOL_DEFAULT:
1090                         default:
1091                                 break;
1092                         }
1093
1094                 }
1095         }
1096 #endif
1097
1098         /*
1099          * May alter parameters that init_io_u() will use, so we need to
1100          * do this first.
1101          */
1102         if (init_iolog(td))
1103                 goto err;
1104
1105         if (init_io_u(td))
1106                 goto err;
1107
1108         if (td->o.verify_async && verify_async_init(td))
1109                 goto err;
1110
1111         if (td->ioprio_set) {
1112                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1113                         td_verror(td, errno, "ioprio_set");
1114                         goto err;
1115                 }
1116         }
1117
1118         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1119                 goto err;
1120
1121         errno = 0;
1122         if (nice(td->o.nice) == -1 && errno != 0) {
1123                 td_verror(td, errno, "nice");
1124                 goto err;
1125         }
1126
1127         if (td->o.ioscheduler && switch_ioscheduler(td))
1128                 goto err;
1129
1130         if (!td->o.create_serialize && setup_files(td))
1131                 goto err;
1132
1133         if (td_io_init(td))
1134                 goto err;
1135
1136         if (init_random_map(td))
1137                 goto err;
1138
1139         if (td->o.exec_prerun) {
1140                 if (exec_string(td->o.exec_prerun))
1141                         goto err;
1142         }
1143
1144         if (td->o.pre_read) {
1145                 if (pre_read_files(td) < 0)
1146                         goto err;
1147         }
1148
1149         fio_gettime(&td->epoch, NULL);
1150         getrusage(RUSAGE_SELF, &td->ru_start);
1151
1152         clear_state = 0;
1153         while (keep_running(td)) {
1154                 fio_gettime(&td->start, NULL);
1155                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1156                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1157                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1158
1159                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1160                                 td->o.ratemin[DDIR_TRIM]) {
1161                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1162                                                 sizeof(td->bw_sample_time));
1163                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1164                                                 sizeof(td->bw_sample_time));
1165                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1166                                                 sizeof(td->bw_sample_time));
1167                 }
1168
1169                 if (clear_state)
1170                         clear_io_state(td);
1171
1172                 prune_io_piece_log(td);
1173
1174                 do_io(td);
1175
1176                 clear_state = 1;
1177
1178                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1179                         elapsed = utime_since_now(&td->start);
1180                         td->ts.runtime[DDIR_READ] += elapsed;
1181                 }
1182                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1183                         elapsed = utime_since_now(&td->start);
1184                         td->ts.runtime[DDIR_WRITE] += elapsed;
1185                 }
1186                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1187                         elapsed = utime_since_now(&td->start);
1188                         td->ts.runtime[DDIR_TRIM] += elapsed;
1189                 }
1190
1191                 if (td->error || td->terminate)
1192                         break;
1193
1194                 if (!td->o.do_verify ||
1195                     td->o.verify == VERIFY_NONE ||
1196                     (td->io_ops->flags & FIO_UNIDIR))
1197                         continue;
1198
1199                 clear_io_state(td);
1200
1201                 fio_gettime(&td->start, NULL);
1202
1203                 do_verify(td);
1204
1205                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1206
1207                 if (td->error || td->terminate)
1208                         break;
1209         }
1210
1211         update_rusage_stat(td);
1212         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1213         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1214         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1215         td->ts.total_run_time = mtime_since_now(&td->epoch);
1216         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1217         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1218         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1219
1220         fio_mutex_down(writeout_mutex);
1221         if (td->bw_log) {
1222                 if (td->o.bw_log_file) {
1223                         finish_log_named(td, td->bw_log,
1224                                                 td->o.bw_log_file, "bw");
1225                 } else
1226                         finish_log(td, td->bw_log, "bw");
1227         }
1228         if (td->lat_log) {
1229                 if (td->o.lat_log_file) {
1230                         finish_log_named(td, td->lat_log,
1231                                                 td->o.lat_log_file, "lat");
1232                 } else
1233                         finish_log(td, td->lat_log, "lat");
1234         }
1235         if (td->slat_log) {
1236                 if (td->o.lat_log_file) {
1237                         finish_log_named(td, td->slat_log,
1238                                                 td->o.lat_log_file, "slat");
1239                 } else
1240                         finish_log(td, td->slat_log, "slat");
1241         }
1242         if (td->clat_log) {
1243                 if (td->o.lat_log_file) {
1244                         finish_log_named(td, td->clat_log,
1245                                                 td->o.lat_log_file, "clat");
1246                 } else
1247                         finish_log(td, td->clat_log, "clat");
1248         }
1249         if (td->iops_log) {
1250                 if (td->o.iops_log_file) {
1251                         finish_log_named(td, td->iops_log,
1252                                                 td->o.iops_log_file, "iops");
1253                 } else
1254                         finish_log(td, td->iops_log, "iops");
1255         }
1256
1257         fio_mutex_up(writeout_mutex);
1258         if (td->o.exec_postrun)
1259                 exec_string(td->o.exec_postrun);
1260
1261         if (exitall_on_terminate)
1262                 fio_terminate_threads(td->groupid);
1263
1264 err:
1265         if (td->error)
1266                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1267                                                         td->verror);
1268
1269         if (td->o.verify_async)
1270                 verify_async_exit(td);
1271
1272         close_and_free_files(td);
1273         close_ioengine(td);
1274         cleanup_io_u(td);
1275         cgroup_shutdown(td, &cgroup_mnt);
1276
1277         if (td->o.cpumask_set) {
1278                 int ret = fio_cpuset_exit(&td->o.cpumask);
1279
1280                 td_verror(td, ret, "fio_cpuset_exit");
1281         }
1282
1283         /*
1284          * do this very late, it will log file closing as well
1285          */
1286         if (td->o.write_iolog_file)
1287                 write_iolog_close(td);
1288
1289         td_set_runstate(td, TD_EXITED);
1290         return (void *) (uintptr_t) td->error;
1291 }
1292
1293
1294 /*
1295  * We cannot pass the td data into a forked process, so attach the td and
1296  * pass it to the thread worker.
1297  */
1298 static int fork_main(int shmid, int offset)
1299 {
1300         struct thread_data *td;
1301         void *data, *ret;
1302
1303 #ifndef __hpux
1304         data = shmat(shmid, NULL, 0);
1305         if (data == (void *) -1) {
1306                 int __err = errno;
1307
1308                 perror("shmat");
1309                 return __err;
1310         }
1311 #else
1312         /*
1313          * HP-UX inherits shm mappings?
1314          */
1315         data = threads;
1316 #endif
1317
1318         td = data + offset * sizeof(struct thread_data);
1319         ret = thread_main(td);
1320         shmdt(data);
1321         return (int) (uintptr_t) ret;
1322 }
1323
1324 /*
1325  * Run over the job map and reap the threads that have exited, if any.
1326  */
1327 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1328                          unsigned int *m_rate)
1329 {
1330         struct thread_data *td;
1331         unsigned int cputhreads, realthreads, pending;
1332         int i, status, ret;
1333
1334         /*
1335          * reap exited threads (TD_EXITED -> TD_REAPED)
1336          */
1337         realthreads = pending = cputhreads = 0;
1338         for_each_td(td, i) {
1339                 int flags = 0;
1340
1341                 /*
1342                  * ->io_ops is NULL for a thread that has closed its
1343                  * io engine
1344                  */
1345                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1346                         cputhreads++;
1347                 else
1348                         realthreads++;
1349
1350                 if (!td->pid) {
1351                         pending++;
1352                         continue;
1353                 }
1354                 if (td->runstate == TD_REAPED)
1355                         continue;
1356                 if (td->o.use_thread) {
1357                         if (td->runstate == TD_EXITED) {
1358                                 td_set_runstate(td, TD_REAPED);
1359                                 goto reaped;
1360                         }
1361                         continue;
1362                 }
1363
1364                 flags = WNOHANG;
1365                 if (td->runstate == TD_EXITED)
1366                         flags = 0;
1367
1368                 /*
1369                  * check if someone quit or got killed in an unusual way
1370                  */
1371                 ret = waitpid(td->pid, &status, flags);
1372                 if (ret < 0) {
1373                         if (errno == ECHILD) {
1374                                 log_err("fio: pid=%d disappeared %d\n",
1375                                                 (int) td->pid, td->runstate);
1376                                 td->sig = ECHILD;
1377                                 td_set_runstate(td, TD_REAPED);
1378                                 goto reaped;
1379                         }
1380                         perror("waitpid");
1381                 } else if (ret == td->pid) {
1382                         if (WIFSIGNALED(status)) {
1383                                 int sig = WTERMSIG(status);
1384
1385                                 if (sig != SIGTERM)
1386                                         log_err("fio: pid=%d, got signal=%d\n",
1387                                                         (int) td->pid, sig);
1388                                 td->sig = sig;
1389                                 td_set_runstate(td, TD_REAPED);
1390                                 goto reaped;
1391                         }
1392                         if (WIFEXITED(status)) {
1393                                 if (WEXITSTATUS(status) && !td->error)
1394                                         td->error = WEXITSTATUS(status);
1395
1396                                 td_set_runstate(td, TD_REAPED);
1397                                 goto reaped;
1398                         }
1399                 }
1400
1401                 /*
1402                  * thread is not dead, continue
1403                  */
1404                 pending++;
1405                 continue;
1406 reaped:
1407                 (*nr_running)--;
1408                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1409                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1410                 if (!td->pid)
1411                         pending--;
1412
1413                 if (td->error)
1414                         exit_value++;
1415
1416                 done_secs += mtime_since_now(&td->epoch) / 1000;
1417         }
1418
1419         if (*nr_running == cputhreads && !pending && realthreads)
1420                 fio_terminate_threads(TERMINATE_ALL);
1421 }
1422
1423 /*
1424  * Main function for kicking off and reaping jobs, as needed.
1425  */
1426 static void run_threads(void)
1427 {
1428         struct thread_data *td;
1429         unsigned long spent;
1430         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1431
1432         if (fio_pin_memory())
1433                 return;
1434
1435         if (fio_gtod_offload && fio_start_gtod_thread())
1436                 return;
1437
1438         set_sig_handlers();
1439
1440         if (output_format == FIO_OUTPUT_NORMAL) {
1441                 log_info("Starting ");
1442                 if (nr_thread)
1443                         log_info("%d thread%s", nr_thread,
1444                                                 nr_thread > 1 ? "s" : "");
1445                 if (nr_process) {
1446                         if (nr_thread)
1447                                 log_info(" and ");
1448                         log_info("%d process%s", nr_process,
1449                                                 nr_process > 1 ? "es" : "");
1450                 }
1451                 log_info("\n");
1452                 fflush(stdout);
1453         }
1454
1455         todo = thread_number;
1456         nr_running = 0;
1457         nr_started = 0;
1458         m_rate = t_rate = 0;
1459
1460         for_each_td(td, i) {
1461                 print_status_init(td->thread_number - 1);
1462
1463                 if (!td->o.create_serialize)
1464                         continue;
1465
1466                 /*
1467                  * do file setup here so it happens sequentially,
1468                  * we don't want X number of threads getting their
1469                  * client data interspersed on disk
1470                  */
1471                 if (setup_files(td)) {
1472                         exit_value++;
1473                         if (td->error)
1474                                 log_err("fio: pid=%d, err=%d/%s\n",
1475                                         (int) td->pid, td->error, td->verror);
1476                         td_set_runstate(td, TD_REAPED);
1477                         todo--;
1478                 } else {
1479                         struct fio_file *f;
1480                         unsigned int j;
1481
1482                         /*
1483                          * for sharing to work, each job must always open
1484                          * its own files. so close them, if we opened them
1485                          * for creation
1486                          */
1487                         for_each_file(td, f, j) {
1488                                 if (fio_file_open(f))
1489                                         td_io_close_file(td, f);
1490                         }
1491                 }
1492         }
1493
1494         set_genesis_time();
1495
1496         while (todo) {
1497                 struct thread_data *map[REAL_MAX_JOBS];
1498                 struct timeval this_start;
1499                 int this_jobs = 0, left;
1500
1501                 /*
1502                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1503                  */
1504                 for_each_td(td, i) {
1505                         if (td->runstate != TD_NOT_CREATED)
1506                                 continue;
1507
1508                         /*
1509                          * never got a chance to start, killed by other
1510                          * thread for some reason
1511                          */
1512                         if (td->terminate) {
1513                                 todo--;
1514                                 continue;
1515                         }
1516
1517                         if (td->o.start_delay) {
1518                                 spent = mtime_since_genesis();
1519
1520                                 if (td->o.start_delay * 1000 > spent)
1521                                         continue;
1522                         }
1523
1524                         if (td->o.stonewall && (nr_started || nr_running)) {
1525                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1526                                                         td->o.name);
1527                                 break;
1528                         }
1529
1530                         init_disk_util(td);
1531
1532                         /*
1533                          * Set state to created. Thread will transition
1534                          * to TD_INITIALIZED when it's done setting up.
1535                          */
1536                         td_set_runstate(td, TD_CREATED);
1537                         map[this_jobs++] = td;
1538                         nr_started++;
1539
1540                         if (td->o.use_thread) {
1541                                 int ret;
1542
1543                                 dprint(FD_PROCESS, "will pthread_create\n");
1544                                 ret = pthread_create(&td->thread, NULL,
1545                                                         thread_main, td);
1546                                 if (ret) {
1547                                         log_err("pthread_create: %s\n",
1548                                                         strerror(ret));
1549                                         nr_started--;
1550                                         break;
1551                                 }
1552                                 ret = pthread_detach(td->thread);
1553                                 if (ret)
1554                                         log_err("pthread_detach: %s",
1555                                                         strerror(ret));
1556                         } else {
1557                                 pid_t pid;
1558                                 dprint(FD_PROCESS, "will fork\n");
1559                                 pid = fork();
1560                                 if (!pid) {
1561                                         int ret = fork_main(shm_id, i);
1562
1563                                         _exit(ret);
1564                                 } else if (i == fio_debug_jobno)
1565                                         *fio_debug_jobp = pid;
1566                         }
1567                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1568                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1569                                 log_err("fio: job startup hung? exiting.\n");
1570                                 fio_terminate_threads(TERMINATE_ALL);
1571                                 fio_abort = 1;
1572                                 nr_started--;
1573                                 break;
1574                         }
1575                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1576                 }
1577
1578                 /*
1579                  * Wait for the started threads to transition to
1580                  * TD_INITIALIZED.
1581                  */
1582                 fio_gettime(&this_start, NULL);
1583                 left = this_jobs;
1584                 while (left && !fio_abort) {
1585                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1586                                 break;
1587
1588                         usleep(100000);
1589
1590                         for (i = 0; i < this_jobs; i++) {
1591                                 td = map[i];
1592                                 if (!td)
1593                                         continue;
1594                                 if (td->runstate == TD_INITIALIZED) {
1595                                         map[i] = NULL;
1596                                         left--;
1597                                 } else if (td->runstate >= TD_EXITED) {
1598                                         map[i] = NULL;
1599                                         left--;
1600                                         todo--;
1601                                         nr_running++; /* work-around... */
1602                                 }
1603                         }
1604                 }
1605
1606                 if (left) {
1607                         log_err("fio: %d job%s failed to start\n", left,
1608                                         left > 1 ? "s" : "");
1609                         for (i = 0; i < this_jobs; i++) {
1610                                 td = map[i];
1611                                 if (!td)
1612                                         continue;
1613                                 kill(td->pid, SIGTERM);
1614                         }
1615                         break;
1616                 }
1617
1618                 /*
1619                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1620                  */
1621                 for_each_td(td, i) {
1622                         if (td->runstate != TD_INITIALIZED)
1623                                 continue;
1624
1625                         if (in_ramp_time(td))
1626                                 td_set_runstate(td, TD_RAMP);
1627                         else
1628                                 td_set_runstate(td, TD_RUNNING);
1629                         nr_running++;
1630                         nr_started--;
1631                         m_rate += ddir_rw_sum(td->o.ratemin);
1632                         t_rate += ddir_rw_sum(td->o.rate);
1633                         todo--;
1634                         fio_mutex_up(td->mutex);
1635                 }
1636
1637                 reap_threads(&nr_running, &t_rate, &m_rate);
1638
1639                 if (todo) {
1640                         if (is_backend)
1641                                 fio_server_idle_loop();
1642                         else
1643                                 usleep(100000);
1644                 }
1645         }
1646
1647         while (nr_running) {
1648                 reap_threads(&nr_running, &t_rate, &m_rate);
1649
1650                 if (is_backend)
1651                         fio_server_idle_loop();
1652                 else
1653                         usleep(10000);
1654         }
1655
1656         update_io_ticks();
1657         fio_unpin_memory();
1658 }
1659
1660 void wait_for_disk_thread_exit(void)
1661 {
1662         fio_mutex_down(disk_thread_mutex);
1663 }
1664
1665 static void free_disk_util(void)
1666 {
1667         disk_util_start_exit();
1668         wait_for_disk_thread_exit();
1669         disk_util_prune_entries();
1670 }
1671
1672 static void *disk_thread_main(void *data)
1673 {
1674         int ret = 0;
1675
1676         fio_mutex_up(startup_mutex);
1677
1678         while (threads && !ret) {
1679                 usleep(DISK_UTIL_MSEC * 1000);
1680                 if (!threads)
1681                         break;
1682                 ret = update_io_ticks();
1683
1684                 if (!is_backend)
1685                         print_thread_status();
1686         }
1687
1688         fio_mutex_up(disk_thread_mutex);
1689         return NULL;
1690 }
1691
1692 static int create_disk_util_thread(void)
1693 {
1694         int ret;
1695
1696         setup_disk_util();
1697
1698         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1699
1700         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1701         if (ret) {
1702                 fio_mutex_remove(disk_thread_mutex);
1703                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1704                 return 1;
1705         }
1706
1707         ret = pthread_detach(disk_util_thread);
1708         if (ret) {
1709                 fio_mutex_remove(disk_thread_mutex);
1710                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1711                 return 1;
1712         }
1713
1714         dprint(FD_MUTEX, "wait on startup_mutex\n");
1715         fio_mutex_down(startup_mutex);
1716         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1717         return 0;
1718 }
1719
1720 int fio_backend(void)
1721 {
1722         struct thread_data *td;
1723         int i;
1724
1725         if (exec_profile) {
1726                 if (load_profile(exec_profile))
1727                         return 1;
1728                 free(exec_profile);
1729                 exec_profile = NULL;
1730         }
1731         if (!thread_number)
1732                 return 0;
1733
1734         if (write_bw_log) {
1735                 setup_log(&agg_io_log[DDIR_READ], 0);
1736                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1737                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1738         }
1739
1740         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1741         if (startup_mutex == NULL)
1742                 return 1;
1743         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1744         if (writeout_mutex == NULL)
1745                 return 1;
1746
1747         set_genesis_time();
1748         create_disk_util_thread();
1749
1750         cgroup_list = smalloc(sizeof(*cgroup_list));
1751         INIT_FLIST_HEAD(cgroup_list);
1752
1753         run_threads();
1754
1755         if (!fio_abort) {
1756                 show_run_stats();
1757                 if (write_bw_log) {
1758                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1759                         __finish_log(agg_io_log[DDIR_WRITE],
1760                                         "agg-write_bw.log");
1761                         __finish_log(agg_io_log[DDIR_TRIM],
1762                                         "agg-write_bw.log");
1763                 }
1764         }
1765
1766         for_each_td(td, i)
1767                 fio_options_free(td);
1768
1769         free_disk_util();
1770         cgroup_kill(cgroup_list);
1771         sfree(cgroup_list);
1772         sfree(cgroup_mnt);
1773
1774         fio_mutex_remove(startup_mutex);
1775         fio_mutex_remove(writeout_mutex);
1776         fio_mutex_remove(disk_thread_mutex);
1777         return exit_value;
1778 }