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/mman.h>
  38
  39 #include "fio.h"
  40 #ifndef FIO_NO_HAVE_SHM_H
  41 #include <sys/shm.h>
  42 #endif
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53 #include "lib/getrusage.h"
  54 #include "idletime.h"
  55
  56 static pthread_t disk_util_thread;
  57 static struct fio_mutex *disk_thread_mutex;
  58 static struct fio_mutex *startup_mutex;
  59 static struct fio_mutex *writeout_mutex;
  60 static struct flist_head *cgroup_list;
  61 static char *cgroup_mnt;
  62 static int exit_value;
  63 static volatile int fio_abort;
  64 static unsigned int nr_process = 0;
  65 static unsigned int nr_thread = 0;
  66
  67 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  68
  69 int groupid = 0;
  70 unsigned int thread_number = 0;
  71 unsigned int stat_number = 0;
  72 int shm_id = 0;
  73 int temp_stall_ts;
  74 unsigned long done_secs = 0;
  75 volatile int disk_util_exit = 0;
  76
  77 #define PAGE_ALIGN(buf) \
  78         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  79
  80 #define JOB_START_TIMEOUT       (5 * 1000)
  81
  82 static void sig_int(int sig)
  83 {
  84         if (threads) {
  85                 if (is_backend)
  86                         fio_server_got_signal(sig);
  87                 else {
  88                         log_info("\nfio: terminating on signal %d\n", sig);
  89                         fflush(stdout);
  90                         exit_value = 128;
  91                 }
  92
  93                 fio_terminate_threads(TERMINATE_ALL);
  94         }
  95 }
  96
  97 static void sig_show_status(int sig)
  98 {
  99         show_running_run_stats();
 100 }
 101
 102 static void set_sig_handlers(void)
 103 {
 104         struct sigaction act;
 105
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGINT, &act, NULL);
 110
 111         memset(&act, 0, sizeof(act));
 112         act.sa_handler = sig_int;
 113         act.sa_flags = SA_RESTART;
 114         sigaction(SIGTERM, &act, NULL);
 115
 116 /* Windows uses SIGBREAK as a quit signal from other applications */
 117 #ifdef WIN32
 118         memset(&act, 0, sizeof(act));
 119         act.sa_handler = sig_int;
 120         act.sa_flags = SA_RESTART;
 121         sigaction(SIGBREAK, &act, NULL);
 122 #endif
 123
 124         memset(&act, 0, sizeof(act));
 125         act.sa_handler = sig_show_status;
 126         act.sa_flags = SA_RESTART;
 127         sigaction(SIGUSR1, &act, NULL);
 128
 129         if (is_backend) {
 130                 memset(&act, 0, sizeof(act));
 131                 act.sa_handler = sig_int;
 132                 act.sa_flags = SA_RESTART;
 133                 sigaction(SIGPIPE, &act, NULL);
 134         }
 135 }
 136
 137 /*
 138  * Check if we are above the minimum rate given.
 139  */
 140 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 141                             enum fio_ddir ddir)
 142 {
 143         unsigned long long bytes = 0;
 144         unsigned long iops = 0;
 145         unsigned long spent;
 146         unsigned long rate;
 147         unsigned int ratemin = 0;
 148         unsigned int rate_iops = 0;
 149         unsigned int rate_iops_min = 0;
 150
 151         assert(ddir_rw(ddir));
 152
 153         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 154                 return 0;
 155
 156         /*
 157          * allow a 2 second settle period in the beginning
 158          */
 159         if (mtime_since(&td->start, now) < 2000)
 160                 return 0;
 161
 162         iops += td->this_io_blocks[ddir];
 163         bytes += td->this_io_bytes[ddir];
 164         ratemin += td->o.ratemin[ddir];
 165         rate_iops += td->o.rate_iops[ddir];
 166         rate_iops_min += td->o.rate_iops_min[ddir];
 167
 168         /*
 169          * if rate blocks is set, sample is running
 170          */
 171         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 172                 spent = mtime_since(&td->lastrate[ddir], now);
 173                 if (spent < td->o.ratecycle)
 174                         return 0;
 175
 176                 if (td->o.rate[ddir]) {
 177                         /*
 178                          * check bandwidth specified rate
 179                          */
 180                         if (bytes < td->rate_bytes[ddir]) {
 181                                 log_err("%s: min rate %u not met\n", td->o.name,
 182                                                                 ratemin);
 183                                 return 1;
 184                         } else {
 185                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 186                                 if (rate < ratemin ||
 187                                     bytes < td->rate_bytes[ddir]) {
 188                                         log_err("%s: min rate %u not met, got"
 189                                                 " %luKB/sec\n", td->o.name,
 190                                                         ratemin, rate);
 191                                         return 1;
 192                                 }
 193                         }
 194                 } else {
 195                         /*
 196                          * checks iops specified rate
 197                          */
 198                         if (iops < rate_iops) {
 199                                 log_err("%s: min iops rate %u not met\n",
 200                                                 td->o.name, rate_iops);
 201                                 return 1;
 202                         } else {
 203                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 204                                 if (rate < rate_iops_min ||
 205                                     iops < td->rate_blocks[ddir]) {
 206                                         log_err("%s: min iops rate %u not met,"
 207                                                 " got %lu\n", td->o.name,
 208                                                         rate_iops_min, rate);
 209                                 }
 210                         }
 211                 }
 212         }
 213
 214         td->rate_bytes[ddir] = bytes;
 215         td->rate_blocks[ddir] = iops;
 216         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 217         return 0;
 218 }
 219
 220 static int check_min_rate(struct thread_data *td, struct timeval *now,
 221                           uint64_t *bytes_done)
 222 {
 223         int ret = 0;
 224
 225         if (bytes_done[DDIR_READ])
 226                 ret |= __check_min_rate(td, now, DDIR_READ);
 227         if (bytes_done[DDIR_WRITE])
 228                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 229         if (bytes_done[DDIR_TRIM])
 230                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 231
 232         return ret;
 233 }
 234
 235 /*
 236  * When job exits, we can cancel the in-flight IO if we are using async
 237  * io. Attempt to do so.
 238  */
 239 static void cleanup_pending_aio(struct thread_data *td)
 240 {
 241         int r;
 242
 243         /*
 244          * get immediately available events, if any
 245          */
 246         r = io_u_queued_complete(td, 0, NULL);
 247         if (r < 0)
 248                 return;
 249
 250         /*
 251          * now cancel remaining active events
 252          */
 253         if (td->io_ops->cancel) {
 254                 struct io_u *io_u;
 255                 int i;
 256
 257                 io_u_qiter(&td->io_u_all, io_u, i) {
 258                         if (io_u->flags & IO_U_F_FLIGHT) {
 259                                 r = td->io_ops->cancel(td, io_u);
 260                                 if (!r)
 261                                         put_io_u(td, io_u);
 262                         }
 263                 }
 264         }
 265
 266         if (td->cur_depth)
 267                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 268 }
 269
 270 /*
 271  * Helper to handle the final sync of a file. Works just like the normal
 272  * io path, just does everything sync.
 273  */
 274 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 275 {
 276         struct io_u *io_u = __get_io_u(td);
 277         int ret;
 278
 279         if (!io_u)
 280                 return 1;
 281
 282         io_u->ddir = DDIR_SYNC;
 283         io_u->file = f;
 284
 285         if (td_io_prep(td, io_u)) {
 286                 put_io_u(td, io_u);
 287                 return 1;
 288         }
 289
 290 requeue:
 291         ret = td_io_queue(td, io_u);
 292         if (ret < 0) {
 293                 td_verror(td, io_u->error, "td_io_queue");
 294                 put_io_u(td, io_u);
 295                 return 1;
 296         } else if (ret == FIO_Q_QUEUED) {
 297                 if (io_u_queued_complete(td, 1, NULL) < 0)
 298                         return 1;
 299         } else if (ret == FIO_Q_COMPLETED) {
 300                 if (io_u->error) {
 301                         td_verror(td, io_u->error, "td_io_queue");
 302                         return 1;
 303                 }
 304
 305                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 306                         return 1;
 307         } else if (ret == FIO_Q_BUSY) {
 308                 if (td_io_commit(td))
 309                         return 1;
 310                 goto requeue;
 311         }
 312
 313         return 0;
 314 }
 315
 316 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 317 {
 318         int ret;
 319
 320         if (fio_file_open(f))
 321                 return fio_io_sync(td, f);
 322
 323         if (td_io_open_file(td, f))
 324                 return 1;
 325
 326         ret = fio_io_sync(td, f);
 327         td_io_close_file(td, f);
 328         return ret;
 329 }
 330
 331 static inline void __update_tv_cache(struct thread_data *td)
 332 {
 333         fio_gettime(&td->tv_cache, NULL);
 334 }
 335
 336 static inline void update_tv_cache(struct thread_data *td)
 337 {
 338         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 339                 __update_tv_cache(td);
 340 }
 341
 342 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 343 {
 344         if (in_ramp_time(td))
 345                 return 0;
 346         if (!td->o.timeout)
 347                 return 0;
 348         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 349                 return 1;
 350
 351         return 0;
 352 }
 353
 354 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 355                                int *retptr)
 356 {
 357         int ret = *retptr;
 358
 359         if (ret < 0 || td->error) {
 360                 int err = td->error;
 361                 enum error_type_bit eb;
 362
 363                 if (ret < 0)
 364                         err = -ret;
 365
 366                 eb = td_error_type(ddir, err);
 367                 if (!(td->o.continue_on_error & (1 << eb)))
 368                         return 1;
 369
 370                 if (td_non_fatal_error(td, eb, err)) {
 371                         /*
 372                          * Continue with the I/Os in case of
 373                          * a non fatal error.
 374                          */
 375                         update_error_count(td, err);
 376                         td_clear_error(td);
 377                         *retptr = 0;
 378                         return 0;
 379                 } else if (td->o.fill_device && err == ENOSPC) {
 380                         /*
 381                          * We expect to hit this error if
 382                          * fill_device option is set.
 383                          */
 384                         td_clear_error(td);
 385                         td->terminate = 1;
 386                         return 1;
 387                 } else {
 388                         /*
 389                          * Stop the I/O in case of a fatal
 390                          * error.
 391                          */
 392                         update_error_count(td, err);
 393                         return 1;
 394                 }
 395         }
 396
 397         return 0;
 398 }
 399
 400 static void check_update_rusage(struct thread_data *td)
 401 {
 402         if (td->update_rusage) {
 403                 td->update_rusage = 0;
 404                 update_rusage_stat(td);
 405                 fio_mutex_up(td->rusage_sem);
 406         }
 407 }
 408
 409 /*
 410  * The main verify engine. Runs over the writes we previously submitted,
 411  * reads the blocks back in, and checks the crc/md5 of the data.
 412  */
 413 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 414 {
 415         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 416         struct fio_file *f;
 417         struct io_u *io_u;
 418         int ret, min_events;
 419         unsigned int i;
 420
 421         dprint(FD_VERIFY, "starting loop\n");
 422
 423         /*
 424          * sync io first and invalidate cache, to make sure we really
 425          * read from disk.
 426          */
 427         for_each_file(td, f, i) {
 428                 if (!fio_file_open(f))
 429                         continue;
 430                 if (fio_io_sync(td, f))
 431                         break;
 432                 if (file_invalidate_cache(td, f))
 433                         break;
 434         }
 435
 436         check_update_rusage(td);
 437
 438         if (td->error)
 439                 return;
 440
 441         td_set_runstate(td, TD_VERIFYING);
 442
 443         io_u = NULL;
 444         while (!td->terminate) {
 445                 enum fio_ddir ddir;
 446                 int ret2, full;
 447
 448                 update_tv_cache(td);
 449                 check_update_rusage(td);
 450
 451                 if (runtime_exceeded(td, &td->tv_cache)) {
 452                         __update_tv_cache(td);
 453                         if (runtime_exceeded(td, &td->tv_cache)) {
 454                                 td->terminate = 1;
 455                                 break;
 456                         }
 457                 }
 458
 459                 if (flow_threshold_exceeded(td))
 460                         continue;
 461
 462                 if (!td->o.experimental_verify) {
 463                         io_u = __get_io_u(td);
 464                         if (!io_u)
 465                                 break;
 466
 467                         if (get_next_verify(td, io_u)) {
 468                                 put_io_u(td, io_u);
 469                                 break;
 470                         }
 471
 472                         if (td_io_prep(td, io_u)) {
 473                                 put_io_u(td, io_u);
 474                                 break;
 475                         }
 476                 } else {
 477                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 478                                 break;
 479
 480                         while ((io_u = get_io_u(td)) != NULL) {
 481                                 /*
 482                                  * We are only interested in the places where
 483                                  * we wrote or trimmed IOs. Turn those into
 484                                  * reads for verification purposes.
 485                                  */
 486                                 if (io_u->ddir == DDIR_READ) {
 487                                         /*
 488                                          * Pretend we issued it for rwmix
 489                                          * accounting
 490                                          */
 491                                         td->io_issues[DDIR_READ]++;
 492                                         put_io_u(td, io_u);
 493                                         continue;
 494                                 } else if (io_u->ddir == DDIR_TRIM) {
 495                                         io_u->ddir = DDIR_READ;
 496                                         io_u->flags |= IO_U_F_TRIMMED;
 497                                         break;
 498                                 } else if (io_u->ddir == DDIR_WRITE) {
 499                                         io_u->ddir = DDIR_READ;
 500                                         break;
 501                                 } else {
 502                                         put_io_u(td, io_u);
 503                                         continue;
 504                                 }
 505                         }
 506
 507                         if (!io_u)
 508                                 break;
 509                 }
 510
 511                 if (td->o.verify_async)
 512                         io_u->end_io = verify_io_u_async;
 513                 else
 514                         io_u->end_io = verify_io_u;
 515
 516                 ddir = io_u->ddir;
 517
 518                 ret = td_io_queue(td, io_u);
 519                 switch (ret) {
 520                 case FIO_Q_COMPLETED:
 521                         if (io_u->error) {
 522                                 ret = -io_u->error;
 523                                 clear_io_u(td, io_u);
 524                         } else if (io_u->resid) {
 525                                 int bytes = io_u->xfer_buflen - io_u->resid;
 526
 527                                 /*
 528                                  * zero read, fail
 529                                  */
 530                                 if (!bytes) {
 531                                         td_verror(td, EIO, "full resid");
 532                                         put_io_u(td, io_u);
 533                                         break;
 534                                 }
 535
 536                                 io_u->xfer_buflen = io_u->resid;
 537                                 io_u->xfer_buf += bytes;
 538                                 io_u->offset += bytes;
 539
 540                                 if (ddir_rw(io_u->ddir))
 541                                         td->ts.short_io_u[io_u->ddir]++;
 542
 543                                 f = io_u->file;
 544                                 if (io_u->offset == f->real_file_size)
 545                                         goto sync_done;
 546
 547                                 requeue_io_u(td, &io_u);
 548                         } else {
 549 sync_done:
 550                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 551                                 if (ret < 0)
 552                                         break;
 553                         }
 554                         continue;
 555                 case FIO_Q_QUEUED:
 556                         break;
 557                 case FIO_Q_BUSY:
 558                         requeue_io_u(td, &io_u);
 559                         ret2 = td_io_commit(td);
 560                         if (ret2 < 0)
 561                                 ret = ret2;
 562                         break;
 563                 default:
 564                         assert(ret < 0);
 565                         td_verror(td, -ret, "td_io_queue");
 566                         break;
 567                 }
 568
 569                 if (break_on_this_error(td, ddir, &ret))
 570                         break;
 571
 572                 /*
 573                  * if we can queue more, do so. but check if there are
 574                  * completed io_u's first. Note that we can get BUSY even
 575                  * without IO queued, if the system is resource starved.
 576                  */
 577                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 578                 if (full || !td->o.iodepth_batch_complete) {
 579                         min_events = min(td->o.iodepth_batch_complete,
 580                                          td->cur_depth);
 581                         /*
 582                          * if the queue is full, we MUST reap at least 1 event
 583                          */
 584                         if (full && !min_events)
 585                                 min_events = 1;
 586
 587                         do {
 588                                 /*
 589                                  * Reap required number of io units, if any,
 590                                  * and do the verification on them through
 591                                  * the callback handler
 592                                  */
 593                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 594                                         ret = -1;
 595                                         break;
 596                                 }
 597                         } while (full && (td->cur_depth > td->o.iodepth_low));
 598                 }
 599                 if (ret < 0)
 600                         break;
 601         }
 602
 603         check_update_rusage(td);
 604
 605         if (!td->error) {
 606                 min_events = td->cur_depth;
 607
 608                 if (min_events)
 609                         ret = io_u_queued_complete(td, min_events, NULL);
 610         } else
 611                 cleanup_pending_aio(td);
 612
 613         td_set_runstate(td, TD_RUNNING);
 614
 615         dprint(FD_VERIFY, "exiting loop\n");
 616 }
 617
 618 static int io_bytes_exceeded(struct thread_data *td)
 619 {
 620         unsigned long long bytes;
 621
 622         if (td_rw(td))
 623                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 624         else if (td_write(td))
 625                 bytes = td->this_io_bytes[DDIR_WRITE];
 626         else if (td_read(td))
 627                 bytes = td->this_io_bytes[DDIR_READ];
 628         else
 629                 bytes = td->this_io_bytes[DDIR_TRIM];
 630
 631         return bytes >= td->o.size;
 632 }
 633
 634 /*
 635  * Main IO worker function. It retrieves io_u's to process and queues
 636  * and reaps them, checking for rate and errors along the way.
 637  *
 638  * Returns number of bytes written and trimmed.
 639  */
 640 static uint64_t do_io(struct thread_data *td)
 641 {
 642         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 643         unsigned int i;
 644         int ret = 0;
 645         uint64_t bytes_issued = 0;
 646
 647         if (in_ramp_time(td))
 648                 td_set_runstate(td, TD_RAMP);
 649         else
 650                 td_set_runstate(td, TD_RUNNING);
 651
 652         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 653                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 654                 td->o.time_based) {
 655                 struct timeval comp_time;
 656                 int min_evts = 0;
 657                 struct io_u *io_u;
 658                 int ret2, full;
 659                 enum fio_ddir ddir;
 660
 661                 check_update_rusage(td);
 662
 663                 if (td->terminate || td->done)
 664                         break;
 665
 666                 update_tv_cache(td);
 667
 668                 if (runtime_exceeded(td, &td->tv_cache)) {
 669                         __update_tv_cache(td);
 670                         if (runtime_exceeded(td, &td->tv_cache)) {
 671                                 td->terminate = 1;
 672                                 break;
 673                         }
 674                 }
 675
 676                 if (flow_threshold_exceeded(td))
 677                         continue;
 678
 679                 if (bytes_issued >= (uint64_t) td->o.size)
 680                         break;
 681
 682                 io_u = get_io_u(td);
 683                 if (!io_u)
 684                         break;
 685
 686                 ddir = io_u->ddir;
 687
 688                 /*
 689                  * Add verification end_io handler if:
 690                  *      - Asked to verify (!td_rw(td))
 691                  *      - Or the io_u is from our verify list (mixed write/ver)
 692                  */
 693                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 694                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 695                         if (td->o.verify_async)
 696                                 io_u->end_io = verify_io_u_async;
 697                         else
 698                                 io_u->end_io = verify_io_u;
 699                         td_set_runstate(td, TD_VERIFYING);
 700                 } else if (in_ramp_time(td))
 701                         td_set_runstate(td, TD_RAMP);
 702                 else
 703                         td_set_runstate(td, TD_RUNNING);
 704
 705                 ret = td_io_queue(td, io_u);
 706                 switch (ret) {
 707                 case FIO_Q_COMPLETED:
 708                         if (io_u->error) {
 709                                 ret = -io_u->error;
 710                                 clear_io_u(td, io_u);
 711                         } else if (io_u->resid) {
 712                                 int bytes = io_u->xfer_buflen - io_u->resid;
 713                                 struct fio_file *f = io_u->file;
 714
 715                                 bytes_issued += bytes;
 716                                 /*
 717                                  * zero read, fail
 718                                  */
 719                                 if (!bytes) {
 720                                         td_verror(td, EIO, "full resid");
 721                                         put_io_u(td, io_u);
 722                                         break;
 723                                 }
 724
 725                                 io_u->xfer_buflen = io_u->resid;
 726                                 io_u->xfer_buf += bytes;
 727                                 io_u->offset += bytes;
 728
 729                                 if (ddir_rw(io_u->ddir))
 730                                         td->ts.short_io_u[io_u->ddir]++;
 731
 732                                 if (io_u->offset == f->real_file_size)
 733                                         goto sync_done;
 734
 735                                 requeue_io_u(td, &io_u);
 736                         } else {
 737 sync_done:
 738                                 if (__should_check_rate(td, DDIR_READ) ||
 739                                     __should_check_rate(td, DDIR_WRITE) ||
 740                                     __should_check_rate(td, DDIR_TRIM))
 741                                         fio_gettime(&comp_time, NULL);
 742
 743                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 744                                 if (ret < 0)
 745                                         break;
 746                                 bytes_issued += io_u->xfer_buflen;
 747                         }
 748                         break;
 749                 case FIO_Q_QUEUED:
 750                         /*
 751                          * if the engine doesn't have a commit hook,
 752                          * the io_u is really queued. if it does have such
 753                          * a hook, it has to call io_u_queued() itself.
 754                          */
 755                         if (td->io_ops->commit == NULL)
 756                                 io_u_queued(td, io_u);
 757                         bytes_issued += io_u->xfer_buflen;
 758                         break;
 759                 case FIO_Q_BUSY:
 760                         requeue_io_u(td, &io_u);
 761                         ret2 = td_io_commit(td);
 762                         if (ret2 < 0)
 763                                 ret = ret2;
 764                         break;
 765                 default:
 766                         assert(ret < 0);
 767                         put_io_u(td, io_u);
 768                         break;
 769                 }
 770
 771                 if (break_on_this_error(td, ddir, &ret))
 772                         break;
 773
 774                 /*
 775                  * See if we need to complete some commands. Note that we
 776                  * can get BUSY even without IO queued, if the system is
 777                  * resource starved.
 778                  */
 779                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 780                 if (full || !td->o.iodepth_batch_complete) {
 781                         min_evts = min(td->o.iodepth_batch_complete,
 782                                         td->cur_depth);
 783                         /*
 784                          * if the queue is full, we MUST reap at least 1 event
 785                          */
 786                         if (full && !min_evts)
 787                                 min_evts = 1;
 788
 789                         if (__should_check_rate(td, DDIR_READ) ||
 790                             __should_check_rate(td, DDIR_WRITE) ||
 791                             __should_check_rate(td, DDIR_TRIM))
 792                                 fio_gettime(&comp_time, NULL);
 793
 794                         do {
 795                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 796                                 if (ret < 0)
 797                                         break;
 798
 799                         } while (full && (td->cur_depth > td->o.iodepth_low));
 800                 }
 801
 802                 if (ret < 0)
 803                         break;
 804                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 805                         continue;
 806
 807                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 808                         if (check_min_rate(td, &comp_time, bytes_done)) {
 809                                 if (exitall_on_terminate)
 810                                         fio_terminate_threads(td->groupid);
 811                                 td_verror(td, EIO, "check_min_rate");
 812                                 break;
 813                         }
 814                 }
 815
 816                 if (td->o.thinktime) {
 817                         unsigned long long b;
 818
 819                         b = ddir_rw_sum(td->io_blocks);
 820                         if (!(b % td->o.thinktime_blocks)) {
 821                                 int left;
 822
 823                                 io_u_quiesce(td);
 824
 825                                 if (td->o.thinktime_spin)
 826                                         usec_spin(td->o.thinktime_spin);
 827
 828                                 left = td->o.thinktime - td->o.thinktime_spin;
 829                                 if (left)
 830                                         usec_sleep(td, left);
 831                         }
 832                 }
 833         }
 834
 835         check_update_rusage(td);
 836
 837         if (td->trim_entries)
 838                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 839
 840         if (td->o.fill_device && td->error == ENOSPC) {
 841                 td->error = 0;
 842                 td->terminate = 1;
 843         }
 844         if (!td->error) {
 845                 struct fio_file *f;
 846
 847                 i = td->cur_depth;
 848                 if (i) {
 849                         ret = io_u_queued_complete(td, i, bytes_done);
 850                         if (td->o.fill_device && td->error == ENOSPC)
 851                                 td->error = 0;
 852                 }
 853
 854                 if (should_fsync(td) && td->o.end_fsync) {
 855                         td_set_runstate(td, TD_FSYNCING);
 856
 857                         for_each_file(td, f, i) {
 858                                 if (!fio_file_fsync(td, f))
 859                                         continue;
 860
 861                                 log_err("fio: end_fsync failed for file %s\n",
 862                                                                 f->file_name);
 863                         }
 864                 }
 865         } else
 866                 cleanup_pending_aio(td);
 867
 868         /*
 869          * stop job if we failed doing any IO
 870          */
 871         if (!ddir_rw_sum(td->this_io_bytes))
 872                 td->done = 1;
 873
 874         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 875 }
 876
 877 static void cleanup_io_u(struct thread_data *td)
 878 {
 879         struct io_u *io_u;
 880
 881         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 882
 883                 if (td->io_ops->io_u_free)
 884                         td->io_ops->io_u_free(td, io_u);
 885
 886                 fio_memfree(io_u, sizeof(*io_u));
 887         }
 888
 889         free_io_mem(td);
 890
 891         io_u_rexit(&td->io_u_requeues);
 892         io_u_qexit(&td->io_u_freelist);
 893         io_u_qexit(&td->io_u_all);
 894 }
 895
 896 static int init_io_u(struct thread_data *td)
 897 {
 898         struct io_u *io_u;
 899         unsigned int max_bs, min_write;
 900         int cl_align, i, max_units;
 901         int data_xfer = 1, err;
 902         char *p;
 903
 904         max_units = td->o.iodepth;
 905         max_bs = td_max_bs(td);
 906         min_write = td->o.min_bs[DDIR_WRITE];
 907         td->orig_buffer_size = (unsigned long long) max_bs
 908                                         * (unsigned long long) max_units;
 909
 910         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 911                 data_xfer = 0;
 912
 913         err = 0;
 914         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 915         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 916         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 917
 918         if (err) {
 919                 log_err("fio: failed setting up IO queues\n");
 920                 return 1;
 921         }
 922
 923         /*
 924          * if we may later need to do address alignment, then add any
 925          * possible adjustment here so that we don't cause a buffer
 926          * overflow later. this adjustment may be too much if we get
 927          * lucky and the allocator gives us an aligned address.
 928          */
 929         if (td->o.odirect || td->o.mem_align || (td->io_ops->flags & FIO_RAWIO))
 930                 td->orig_buffer_size += page_mask + td->o.mem_align;
 931
 932         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 933                 unsigned long bs;
 934
 935                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 936                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 937         }
 938
 939         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 940                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 941                 return 1;
 942         }
 943
 944         if (data_xfer && allocate_io_mem(td))
 945                 return 1;
 946
 947         if (td->o.odirect || td->o.mem_align ||
 948             (td->io_ops->flags & FIO_RAWIO))
 949                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 950         else
 951                 p = td->orig_buffer;
 952
 953         cl_align = os_cache_line_size();
 954
 955         for (i = 0; i < max_units; i++) {
 956                 void *ptr;
 957
 958                 if (td->terminate)
 959                         return 1;
 960
 961                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 962                 if (!ptr) {
 963                         log_err("fio: unable to allocate aligned memory\n");
 964                         break;
 965                 }
 966
 967                 io_u = ptr;
 968                 memset(io_u, 0, sizeof(*io_u));
 969                 INIT_FLIST_HEAD(&io_u->verify_list);
 970                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 971
 972                 if (data_xfer) {
 973                         io_u->buf = p;
 974                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 975
 976                         if (td_write(td))
 977                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 978                         if (td_write(td) && td->o.verify_pattern_bytes) {
 979                                 /*
 980                                  * Fill the buffer with the pattern if we are
 981                                  * going to be doing writes.
 982                                  */
 983                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 984                         }
 985                 }
 986
 987                 io_u->index = i;
 988                 io_u->flags = IO_U_F_FREE;
 989                 io_u_qpush(&td->io_u_freelist, io_u);
 990
 991                 /*
 992                  * io_u never leaves this stack, used for iteration of all
 993                  * io_u buffers.
 994                  */
 995                 io_u_qpush(&td->io_u_all, io_u);
 996
 997                 if (td->io_ops->io_u_init) {
 998                         int ret = td->io_ops->io_u_init(td, io_u);
 999
1000                         if (ret) {
1001                                 log_err("fio: failed to init engine data: %d\n", ret);
1002                                 return 1;
1003                         }
1004                 }
1005
1006                 p += max_bs;
1007         }
1008
1009         return 0;
1010 }
1011
1012 static int switch_ioscheduler(struct thread_data *td)
1013 {
1014         char tmp[256], tmp2[128];
1015         FILE *f;
1016         int ret;
1017
1018         if (td->io_ops->flags & FIO_DISKLESSIO)
1019                 return 0;
1020
1021         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1022
1023         f = fopen(tmp, "r+");
1024         if (!f) {
1025                 if (errno == ENOENT) {
1026                         log_err("fio: os or kernel doesn't support IO scheduler"
1027                                 " switching\n");
1028                         return 0;
1029                 }
1030                 td_verror(td, errno, "fopen iosched");
1031                 return 1;
1032         }
1033
1034         /*
1035          * Set io scheduler.
1036          */
1037         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1038         if (ferror(f) || ret != 1) {
1039                 td_verror(td, errno, "fwrite");
1040                 fclose(f);
1041                 return 1;
1042         }
1043
1044         rewind(f);
1045
1046         /*
1047          * Read back and check that the selected scheduler is now the default.
1048          */
1049         ret = fread(tmp, 1, sizeof(tmp), f);
1050         if (ferror(f) || ret < 0) {
1051                 td_verror(td, errno, "fread");
1052                 fclose(f);
1053                 return 1;
1054         }
1055
1056         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1057         if (!strstr(tmp, tmp2)) {
1058                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1059                 td_verror(td, EINVAL, "iosched_switch");
1060                 fclose(f);
1061                 return 1;
1062         }
1063
1064         fclose(f);
1065         return 0;
1066 }
1067
1068 static int keep_running(struct thread_data *td)
1069 {
1070         if (td->done)
1071                 return 0;
1072         if (td->o.time_based)
1073                 return 1;
1074         if (td->o.loops) {
1075                 td->o.loops--;
1076                 return 1;
1077         }
1078
1079         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1080                 uint64_t diff;
1081
1082                 /*
1083                  * If the difference is less than the minimum IO size, we
1084                  * are done.
1085                  */
1086                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1087                 if (diff < td_max_bs(td))
1088                         return 0;
1089
1090                 return 1;
1091         }
1092
1093         return 0;
1094 }
1095
1096 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1097 {
1098         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1099         char *str;
1100
1101         str = malloc(newlen);
1102         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1103
1104         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1105         ret = system(str);
1106         if (ret == -1)
1107                 log_err("fio: exec of cmd <%s> failed\n", str);
1108
1109         free(str);
1110         return ret;
1111 }
1112
1113 /*
1114  * Entry point for the thread based jobs. The process based jobs end up
1115  * here as well, after a little setup.
1116  */
1117 static void *thread_main(void *data)
1118 {
1119         unsigned long long elapsed;
1120         struct thread_data *td = data;
1121         struct thread_options *o = &td->o;
1122         pthread_condattr_t attr;
1123         int clear_state;
1124         int ret;
1125
1126         if (!o->use_thread) {
1127                 setsid();
1128                 td->pid = getpid();
1129         } else
1130                 td->pid = gettid();
1131
1132         /*
1133          * fio_time_init() may not have been called yet if running as a server
1134          */
1135         fio_time_init();
1136
1137         fio_local_clock_init(o->use_thread);
1138
1139         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1140
1141         if (is_backend)
1142                 fio_server_send_start(td);
1143
1144         INIT_FLIST_HEAD(&td->io_log_list);
1145         INIT_FLIST_HEAD(&td->io_hist_list);
1146         INIT_FLIST_HEAD(&td->verify_list);
1147         INIT_FLIST_HEAD(&td->trim_list);
1148         INIT_FLIST_HEAD(&td->next_rand_list);
1149         pthread_mutex_init(&td->io_u_lock, NULL);
1150         td->io_hist_tree = RB_ROOT;
1151
1152         pthread_condattr_init(&attr);
1153         pthread_cond_init(&td->verify_cond, &attr);
1154         pthread_cond_init(&td->free_cond, &attr);
1155
1156         td_set_runstate(td, TD_INITIALIZED);
1157         dprint(FD_MUTEX, "up startup_mutex\n");
1158         fio_mutex_up(startup_mutex);
1159         dprint(FD_MUTEX, "wait on td->mutex\n");
1160         fio_mutex_down(td->mutex);
1161         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1162
1163         /*
1164          * the ->mutex mutex is now no longer used, close it to avoid
1165          * eating a file descriptor
1166          */
1167         fio_mutex_remove(td->mutex);
1168         td->mutex = NULL;
1169
1170         /*
1171          * A new gid requires privilege, so we need to do this before setting
1172          * the uid.
1173          */
1174         if (o->gid != -1U && setgid(o->gid)) {
1175                 td_verror(td, errno, "setgid");
1176                 goto err;
1177         }
1178         if (o->uid != -1U && setuid(o->uid)) {
1179                 td_verror(td, errno, "setuid");
1180                 goto err;
1181         }
1182
1183         /*
1184          * If we have a gettimeofday() thread, make sure we exclude that
1185          * thread from this job
1186          */
1187         if (o->gtod_cpu)
1188                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1189
1190         /*
1191          * Set affinity first, in case it has an impact on the memory
1192          * allocations.
1193          */
1194         if (o->cpumask_set) {
1195                 ret = fio_setaffinity(td->pid, o->cpumask);
1196                 if (ret == -1) {
1197                         td_verror(td, errno, "cpu_set_affinity");
1198                         goto err;
1199                 }
1200         }
1201
1202 #ifdef CONFIG_LIBNUMA
1203         /* numa node setup */
1204         if (o->numa_cpumask_set || o->numa_memmask_set) {
1205                 int ret;
1206
1207                 if (numa_available() < 0) {
1208                         td_verror(td, errno, "Does not support NUMA API\n");
1209                         goto err;
1210                 }
1211
1212                 if (o->numa_cpumask_set) {
1213                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1214                         if (ret == -1) {
1215                                 td_verror(td, errno, \
1216                                         "numa_run_on_node_mask failed\n");
1217                                 goto err;
1218                         }
1219                 }
1220
1221                 if (o->numa_memmask_set) {
1222
1223                         switch (o->numa_mem_mode) {
1224                         case MPOL_INTERLEAVE:
1225                                 numa_set_interleave_mask(o->numa_memnodesmask);
1226                                 break;
1227                         case MPOL_BIND:
1228                                 numa_set_membind(o->numa_memnodesmask);
1229                                 break;
1230                         case MPOL_LOCAL:
1231                                 numa_set_localalloc();
1232                                 break;
1233                         case MPOL_PREFERRED:
1234                                 numa_set_preferred(o->numa_mem_prefer_node);
1235                                 break;
1236                         case MPOL_DEFAULT:
1237                         default:
1238                                 break;
1239                         }
1240
1241                 }
1242         }
1243 #endif
1244
1245         if (fio_pin_memory(td))
1246                 goto err;
1247
1248         /*
1249          * May alter parameters that init_io_u() will use, so we need to
1250          * do this first.
1251          */
1252         if (init_iolog(td))
1253                 goto err;
1254
1255         if (init_io_u(td))
1256                 goto err;
1257
1258         if (o->verify_async && verify_async_init(td))
1259                 goto err;
1260
1261         if (o->ioprio) {
1262                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1263                 if (ret == -1) {
1264                         td_verror(td, errno, "ioprio_set");
1265                         goto err;
1266                 }
1267         }
1268
1269         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1270                 goto err;
1271
1272         errno = 0;
1273         if (nice(o->nice) == -1 && errno != 0) {
1274                 td_verror(td, errno, "nice");
1275                 goto err;
1276         }
1277
1278         if (o->ioscheduler && switch_ioscheduler(td))
1279                 goto err;
1280
1281         if (!o->create_serialize && setup_files(td))
1282                 goto err;
1283
1284         if (td_io_init(td))
1285                 goto err;
1286
1287         if (init_random_map(td))
1288                 goto err;
1289
1290         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1291                 goto err;
1292
1293         if (o->pre_read) {
1294                 if (pre_read_files(td) < 0)
1295                         goto err;
1296         }
1297
1298         fio_verify_init(td);
1299
1300         fio_gettime(&td->epoch, NULL);
1301         fio_getrusage(&td->ru_start);
1302         clear_state = 0;
1303         while (keep_running(td)) {
1304                 uint64_t verify_bytes;
1305
1306                 fio_gettime(&td->start, NULL);
1307                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1308                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1309                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1310
1311                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1312                                 o->ratemin[DDIR_TRIM]) {
1313                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1314                                                 sizeof(td->bw_sample_time));
1315                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1316                                                 sizeof(td->bw_sample_time));
1317                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1318                                                 sizeof(td->bw_sample_time));
1319                 }
1320
1321                 if (clear_state)
1322                         clear_io_state(td);
1323
1324                 prune_io_piece_log(td);
1325
1326                 verify_bytes = do_io(td);
1327
1328                 clear_state = 1;
1329
1330                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1331                         elapsed = utime_since_now(&td->start);
1332                         td->ts.runtime[DDIR_READ] += elapsed;
1333                 }
1334                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1335                         elapsed = utime_since_now(&td->start);
1336                         td->ts.runtime[DDIR_WRITE] += elapsed;
1337                 }
1338                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1339                         elapsed = utime_since_now(&td->start);
1340                         td->ts.runtime[DDIR_TRIM] += elapsed;
1341                 }
1342
1343                 if (td->error || td->terminate)
1344                         break;
1345
1346                 if (!o->do_verify ||
1347                     o->verify == VERIFY_NONE ||
1348                     (td->io_ops->flags & FIO_UNIDIR))
1349                         continue;
1350
1351                 clear_io_state(td);
1352
1353                 fio_gettime(&td->start, NULL);
1354
1355                 do_verify(td, verify_bytes);
1356
1357                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1358
1359                 if (td->error || td->terminate)
1360                         break;
1361         }
1362
1363         update_rusage_stat(td);
1364         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1365         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1366         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1367         td->ts.total_run_time = mtime_since_now(&td->epoch);
1368         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1369         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1370         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1371
1372         fio_unpin_memory(td);
1373
1374         fio_mutex_down(writeout_mutex);
1375         if (td->bw_log) {
1376                 if (o->bw_log_file) {
1377                         finish_log_named(td, td->bw_log,
1378                                                 o->bw_log_file, "bw");
1379                 } else
1380                         finish_log(td, td->bw_log, "bw");
1381         }
1382         if (td->lat_log) {
1383                 if (o->lat_log_file) {
1384                         finish_log_named(td, td->lat_log,
1385                                                 o->lat_log_file, "lat");
1386                 } else
1387                         finish_log(td, td->lat_log, "lat");
1388         }
1389         if (td->slat_log) {
1390                 if (o->lat_log_file) {
1391                         finish_log_named(td, td->slat_log,
1392                                                 o->lat_log_file, "slat");
1393                 } else
1394                         finish_log(td, td->slat_log, "slat");
1395         }
1396         if (td->clat_log) {
1397                 if (o->lat_log_file) {
1398                         finish_log_named(td, td->clat_log,
1399                                                 o->lat_log_file, "clat");
1400                 } else
1401                         finish_log(td, td->clat_log, "clat");
1402         }
1403         if (td->iops_log) {
1404                 if (o->iops_log_file) {
1405                         finish_log_named(td, td->iops_log,
1406                                                 o->iops_log_file, "iops");
1407                 } else
1408                         finish_log(td, td->iops_log, "iops");
1409         }
1410
1411         fio_mutex_up(writeout_mutex);
1412         if (o->exec_postrun)
1413                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1414
1415         if (exitall_on_terminate)
1416                 fio_terminate_threads(td->groupid);
1417
1418 err:
1419         if (td->error)
1420                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1421                                                         td->verror);
1422
1423         if (o->verify_async)
1424                 verify_async_exit(td);
1425
1426         close_and_free_files(td);
1427         cleanup_io_u(td);
1428         close_ioengine(td);
1429         cgroup_shutdown(td, &cgroup_mnt);
1430
1431         if (o->cpumask_set) {
1432                 int ret = fio_cpuset_exit(&o->cpumask);
1433
1434                 td_verror(td, ret, "fio_cpuset_exit");
1435         }
1436
1437         /*
1438          * do this very late, it will log file closing as well
1439          */
1440         if (o->write_iolog_file)
1441                 write_iolog_close(td);
1442
1443         fio_mutex_remove(td->rusage_sem);
1444         td->rusage_sem = NULL;
1445
1446         td_set_runstate(td, TD_EXITED);
1447         return (void *) (uintptr_t) td->error;
1448 }
1449
1450
1451 /*
1452  * We cannot pass the td data into a forked process, so attach the td and
1453  * pass it to the thread worker.
1454  */
1455 static int fork_main(int shmid, int offset)
1456 {
1457         struct thread_data *td;
1458         void *data, *ret;
1459
1460 #ifndef __hpux
1461         data = shmat(shmid, NULL, 0);
1462         if (data == (void *) -1) {
1463                 int __err = errno;
1464
1465                 perror("shmat");
1466                 return __err;
1467         }
1468 #else
1469         /*
1470          * HP-UX inherits shm mappings?
1471          */
1472         data = threads;
1473 #endif
1474
1475         td = data + offset * sizeof(struct thread_data);
1476         ret = thread_main(td);
1477         shmdt(data);
1478         return (int) (uintptr_t) ret;
1479 }
1480
1481 /*
1482  * Run over the job map and reap the threads that have exited, if any.
1483  */
1484 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1485                          unsigned int *m_rate)
1486 {
1487         struct thread_data *td;
1488         unsigned int cputhreads, realthreads, pending;
1489         int i, status, ret;
1490
1491         /*
1492          * reap exited threads (TD_EXITED -> TD_REAPED)
1493          */
1494         realthreads = pending = cputhreads = 0;
1495         for_each_td(td, i) {
1496                 int flags = 0;
1497
1498                 /*
1499                  * ->io_ops is NULL for a thread that has closed its
1500                  * io engine
1501                  */
1502                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1503                         cputhreads++;
1504                 else
1505                         realthreads++;
1506
1507                 if (!td->pid) {
1508                         pending++;
1509                         continue;
1510                 }
1511                 if (td->runstate == TD_REAPED)
1512                         continue;
1513                 if (td->o.use_thread) {
1514                         if (td->runstate == TD_EXITED) {
1515                                 td_set_runstate(td, TD_REAPED);
1516                                 goto reaped;
1517                         }
1518                         continue;
1519                 }
1520
1521                 flags = WNOHANG;
1522                 if (td->runstate == TD_EXITED)
1523                         flags = 0;
1524
1525                 /*
1526                  * check if someone quit or got killed in an unusual way
1527                  */
1528                 ret = waitpid(td->pid, &status, flags);
1529                 if (ret < 0) {
1530                         if (errno == ECHILD) {
1531                                 log_err("fio: pid=%d disappeared %d\n",
1532                                                 (int) td->pid, td->runstate);
1533                                 td->sig = ECHILD;
1534                                 td_set_runstate(td, TD_REAPED);
1535                                 goto reaped;
1536                         }
1537                         perror("waitpid");
1538                 } else if (ret == td->pid) {
1539                         if (WIFSIGNALED(status)) {
1540                                 int sig = WTERMSIG(status);
1541
1542                                 if (sig != SIGTERM && sig != SIGUSR2)
1543                                         log_err("fio: pid=%d, got signal=%d\n",
1544                                                         (int) td->pid, sig);
1545                                 td->sig = sig;
1546                                 td_set_runstate(td, TD_REAPED);
1547                                 goto reaped;
1548                         }
1549                         if (WIFEXITED(status)) {
1550                                 if (WEXITSTATUS(status) && !td->error)
1551                                         td->error = WEXITSTATUS(status);
1552
1553                                 td_set_runstate(td, TD_REAPED);
1554                                 goto reaped;
1555                         }
1556                 }
1557
1558                 /*
1559                  * thread is not dead, continue
1560                  */
1561                 pending++;
1562                 continue;
1563 reaped:
1564                 (*nr_running)--;
1565                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1566                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1567                 if (!td->pid)
1568                         pending--;
1569
1570                 if (td->error)
1571                         exit_value++;
1572
1573                 done_secs += mtime_since_now(&td->epoch) / 1000;
1574                 profile_td_exit(td);
1575         }
1576
1577         if (*nr_running == cputhreads && !pending && realthreads)
1578                 fio_terminate_threads(TERMINATE_ALL);
1579 }
1580
1581 static void do_usleep(unsigned int usecs)
1582 {
1583         check_for_running_stats();
1584         usleep(usecs);
1585 }
1586
1587 /*
1588  * Main function for kicking off and reaping jobs, as needed.
1589  */
1590 static void run_threads(void)
1591 {
1592         struct thread_data *td;
1593         unsigned long spent;
1594         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1595
1596         if (fio_gtod_offload && fio_start_gtod_thread())
1597                 return;
1598
1599         fio_idle_prof_init();
1600
1601         set_sig_handlers();
1602
1603         nr_thread = nr_process = 0;
1604         for_each_td(td, i) {
1605                 if (td->o.use_thread)
1606                         nr_thread++;
1607                 else
1608                         nr_process++;
1609         }
1610
1611         if (output_format == FIO_OUTPUT_NORMAL) {
1612                 log_info("Starting ");
1613                 if (nr_thread)
1614                         log_info("%d thread%s", nr_thread,
1615                                                 nr_thread > 1 ? "s" : "");
1616                 if (nr_process) {
1617                         if (nr_thread)
1618                                 log_info(" and ");
1619                         log_info("%d process%s", nr_process,
1620                                                 nr_process > 1 ? "es" : "");
1621                 }
1622                 log_info("\n");
1623                 fflush(stdout);
1624         }
1625
1626         todo = thread_number;
1627         nr_running = 0;
1628         nr_started = 0;
1629         m_rate = t_rate = 0;
1630
1631         for_each_td(td, i) {
1632                 print_status_init(td->thread_number - 1);
1633
1634                 if (!td->o.create_serialize)
1635                         continue;
1636
1637                 /*
1638                  * do file setup here so it happens sequentially,
1639                  * we don't want X number of threads getting their
1640                  * client data interspersed on disk
1641                  */
1642                 if (setup_files(td)) {
1643                         exit_value++;
1644                         if (td->error)
1645                                 log_err("fio: pid=%d, err=%d/%s\n",
1646                                         (int) td->pid, td->error, td->verror);
1647                         td_set_runstate(td, TD_REAPED);
1648                         todo--;
1649                 } else {
1650                         struct fio_file *f;
1651                         unsigned int j;
1652
1653                         /*
1654                          * for sharing to work, each job must always open
1655                          * its own files. so close them, if we opened them
1656                          * for creation
1657                          */
1658                         for_each_file(td, f, j) {
1659                                 if (fio_file_open(f))
1660                                         td_io_close_file(td, f);
1661                         }
1662                 }
1663         }
1664
1665         /* start idle threads before io threads start to run */
1666         fio_idle_prof_start();
1667
1668         set_genesis_time();
1669
1670         while (todo) {
1671                 struct thread_data *map[REAL_MAX_JOBS];
1672                 struct timeval this_start;
1673                 int this_jobs = 0, left;
1674
1675                 /*
1676                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1677                  */
1678                 for_each_td(td, i) {
1679                         if (td->runstate != TD_NOT_CREATED)
1680                                 continue;
1681
1682                         /*
1683                          * never got a chance to start, killed by other
1684                          * thread for some reason
1685                          */
1686                         if (td->terminate) {
1687                                 todo--;
1688                                 continue;
1689                         }
1690
1691                         if (td->o.start_delay) {
1692                                 spent = mtime_since_genesis();
1693
1694                                 if (td->o.start_delay * 1000 > spent)
1695                                         continue;
1696                         }
1697
1698                         if (td->o.stonewall && (nr_started || nr_running)) {
1699                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1700                                                         td->o.name);
1701                                 break;
1702                         }
1703
1704                         init_disk_util(td);
1705
1706                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1707                         td->update_rusage = 0;
1708
1709                         /*
1710                          * Set state to created. Thread will transition
1711                          * to TD_INITIALIZED when it's done setting up.
1712                          */
1713                         td_set_runstate(td, TD_CREATED);
1714                         map[this_jobs++] = td;
1715                         nr_started++;
1716
1717                         if (td->o.use_thread) {
1718                                 int ret;
1719
1720                                 dprint(FD_PROCESS, "will pthread_create\n");
1721                                 ret = pthread_create(&td->thread, NULL,
1722                                                         thread_main, td);
1723                                 if (ret) {
1724                                         log_err("pthread_create: %s\n",
1725                                                         strerror(ret));
1726                                         nr_started--;
1727                                         break;
1728                                 }
1729                                 ret = pthread_detach(td->thread);
1730                                 if (ret)
1731                                         log_err("pthread_detach: %s",
1732                                                         strerror(ret));
1733                         } else {
1734                                 pid_t pid;
1735                                 dprint(FD_PROCESS, "will fork\n");
1736                                 pid = fork();
1737                                 if (!pid) {
1738                                         int ret = fork_main(shm_id, i);
1739
1740                                         _exit(ret);
1741                                 } else if (i == fio_debug_jobno)
1742                                         *fio_debug_jobp = pid;
1743                         }
1744                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1745                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1746                                 log_err("fio: job startup hung? exiting.\n");
1747                                 fio_terminate_threads(TERMINATE_ALL);
1748                                 fio_abort = 1;
1749                                 nr_started--;
1750                                 break;
1751                         }
1752                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1753                 }
1754
1755                 /*
1756                  * Wait for the started threads to transition to
1757                  * TD_INITIALIZED.
1758                  */
1759                 fio_gettime(&this_start, NULL);
1760                 left = this_jobs;
1761                 while (left && !fio_abort) {
1762                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1763                                 break;
1764
1765                         do_usleep(100000);
1766
1767                         for (i = 0; i < this_jobs; i++) {
1768                                 td = map[i];
1769                                 if (!td)
1770                                         continue;
1771                                 if (td->runstate == TD_INITIALIZED) {
1772                                         map[i] = NULL;
1773                                         left--;
1774                                 } else if (td->runstate >= TD_EXITED) {
1775                                         map[i] = NULL;
1776                                         left--;
1777                                         todo--;
1778                                         nr_running++; /* work-around... */
1779                                 }
1780                         }
1781                 }
1782
1783                 if (left) {
1784                         log_err("fio: %d job%s failed to start\n", left,
1785                                         left > 1 ? "s" : "");
1786                         for (i = 0; i < this_jobs; i++) {
1787                                 td = map[i];
1788                                 if (!td)
1789                                         continue;
1790                                 kill(td->pid, SIGTERM);
1791                         }
1792                         break;
1793                 }
1794
1795                 /*
1796                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1797                  */
1798                 for_each_td(td, i) {
1799                         if (td->runstate != TD_INITIALIZED)
1800                                 continue;
1801
1802                         if (in_ramp_time(td))
1803                                 td_set_runstate(td, TD_RAMP);
1804                         else
1805                                 td_set_runstate(td, TD_RUNNING);
1806                         nr_running++;
1807                         nr_started--;
1808                         m_rate += ddir_rw_sum(td->o.ratemin);
1809                         t_rate += ddir_rw_sum(td->o.rate);
1810                         todo--;
1811                         fio_mutex_up(td->mutex);
1812                 }
1813
1814                 reap_threads(&nr_running, &t_rate, &m_rate);
1815
1816                 if (todo)
1817                         do_usleep(100000);
1818         }
1819
1820         while (nr_running) {
1821                 reap_threads(&nr_running, &t_rate, &m_rate);
1822                 do_usleep(10000);
1823         }
1824
1825         fio_idle_prof_stop();
1826
1827         update_io_ticks();
1828 }
1829
1830 void wait_for_disk_thread_exit(void)
1831 {
1832         fio_mutex_down(disk_thread_mutex);
1833 }
1834
1835 static void free_disk_util(void)
1836 {
1837         disk_util_start_exit();
1838         wait_for_disk_thread_exit();
1839         disk_util_prune_entries();
1840 }
1841
1842 static void *disk_thread_main(void *data)
1843 {
1844         int ret = 0;
1845
1846         fio_mutex_up(startup_mutex);
1847
1848         while (threads && !ret) {
1849                 usleep(DISK_UTIL_MSEC * 1000);
1850                 if (!threads)
1851                         break;
1852                 ret = update_io_ticks();
1853
1854                 if (!is_backend)
1855                         print_thread_status();
1856         }
1857
1858         fio_mutex_up(disk_thread_mutex);
1859         return NULL;
1860 }
1861
1862 static int create_disk_util_thread(void)
1863 {
1864         int ret;
1865
1866         setup_disk_util();
1867
1868         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1869
1870         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1871         if (ret) {
1872                 fio_mutex_remove(disk_thread_mutex);
1873                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1874                 return 1;
1875         }
1876
1877         ret = pthread_detach(disk_util_thread);
1878         if (ret) {
1879                 fio_mutex_remove(disk_thread_mutex);
1880                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1881                 return 1;
1882         }
1883
1884         dprint(FD_MUTEX, "wait on startup_mutex\n");
1885         fio_mutex_down(startup_mutex);
1886         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1887         return 0;
1888 }
1889
1890 int fio_backend(void)
1891 {
1892         struct thread_data *td;
1893         int i;
1894
1895         if (exec_profile) {
1896                 if (load_profile(exec_profile))
1897                         return 1;
1898                 free(exec_profile);
1899                 exec_profile = NULL;
1900         }
1901         if (!thread_number)
1902                 return 0;
1903
1904         if (write_bw_log) {
1905                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1906                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1907                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1908         }
1909
1910         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1911         if (startup_mutex == NULL)
1912                 return 1;
1913         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1914         if (writeout_mutex == NULL)
1915                 return 1;
1916
1917         set_genesis_time();
1918         stat_init();
1919         create_disk_util_thread();
1920
1921         cgroup_list = smalloc(sizeof(*cgroup_list));
1922         INIT_FLIST_HEAD(cgroup_list);
1923
1924         run_threads();
1925
1926         if (!fio_abort) {
1927                 show_run_stats();
1928                 if (write_bw_log) {
1929                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1930                         __finish_log(agg_io_log[DDIR_WRITE],
1931                                         "agg-write_bw.log");
1932                         __finish_log(agg_io_log[DDIR_TRIM],
1933                                         "agg-write_bw.log");
1934                 }
1935         }
1936
1937         for_each_td(td, i)
1938                 fio_options_free(td);
1939
1940         free_disk_util();
1941         cgroup_kill(cgroup_list);
1942         sfree(cgroup_list);
1943         sfree(cgroup_mnt);
1944
1945         fio_mutex_remove(startup_mutex);
1946         fio_mutex_remove(writeout_mutex);
1947         fio_mutex_remove(disk_thread_mutex);
1948         stat_exit();
1949         return exit_value;
1950 }