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 #ifndef FIO_NO_HAVE_SHM_H
  38 #include <sys/shm.h>
  39 #endif
  40 #include <sys/mman.h>
  41
  42 #include "fio.h"
  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
  54 static pthread_t disk_util_thread;
  55 static struct fio_mutex *disk_thread_mutex;
  56 static struct fio_mutex *startup_mutex;
  57 static struct fio_mutex *writeout_mutex;
  58 static struct flist_head *cgroup_list;
  59 static char *cgroup_mnt;
  60 static int exit_value;
  61 static volatile int fio_abort;
  62
  63 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  64
  65 int groupid = 0;
  66 unsigned int thread_number = 0;
  67 unsigned int stat_number = 0;
  68 unsigned int nr_process = 0;
  69 unsigned int nr_thread = 0;
  70 int shm_id = 0;
  71 int temp_stall_ts;
  72 unsigned long done_secs = 0;
  73 volatile int disk_util_exit = 0;
  74
  75 #define PAGE_ALIGN(buf) \
  76         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  77
  78 #define JOB_START_TIMEOUT       (5 * 1000)
  79
  80 static void sig_int(int sig)
  81 {
  82         if (threads) {
  83                 if (is_backend)
  84                         fio_server_got_signal(sig);
  85                 else {
  86                         log_info("\nfio: terminating on signal %d\n", sig);
  87                         fflush(stdout);
  88                         exit_value = 128;
  89                 }
  90
  91                 fio_terminate_threads(TERMINATE_ALL);
  92         }
  93 }
  94
  95 static void sig_show_status(int sig)
  96 {
  97         show_running_run_stats();
  98 }
  99
 100 static void set_sig_handlers(void)
 101 {
 102         struct sigaction act;
 103
 104         memset(&act, 0, sizeof(act));
 105         act.sa_handler = sig_int;
 106         act.sa_flags = SA_RESTART;
 107         sigaction(SIGINT, &act, NULL);
 108
 109         memset(&act, 0, sizeof(act));
 110         act.sa_handler = sig_int;
 111         act.sa_flags = SA_RESTART;
 112         sigaction(SIGTERM, &act, NULL);
 113
 114 /* Windows uses SIGBREAK as a quit signal from other applications */
 115 #ifdef WIN32
 116         memset(&act, 0, sizeof(act));
 117         act.sa_handler = sig_int;
 118         act.sa_flags = SA_RESTART;
 119         sigaction(SIGBREAK, &act, NULL);
 120 #endif
 121
 122         memset(&act, 0, sizeof(act));
 123         act.sa_handler = sig_show_status;
 124         act.sa_flags = SA_RESTART;
 125         sigaction(SIGUSR1, &act, NULL);
 126
 127         if (is_backend) {
 128                 memset(&act, 0, sizeof(act));
 129                 act.sa_handler = sig_int;
 130                 act.sa_flags = SA_RESTART;
 131                 sigaction(SIGPIPE, &act, NULL);
 132         }
 133 }
 134
 135 /*
 136  * Check if we are above the minimum rate given.
 137  */
 138 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 139                             enum fio_ddir ddir)
 140 {
 141         unsigned long long bytes = 0;
 142         unsigned long iops = 0;
 143         unsigned long spent;
 144         unsigned long rate;
 145         unsigned int ratemin = 0;
 146         unsigned int rate_iops = 0;
 147         unsigned int rate_iops_min = 0;
 148
 149         assert(ddir_rw(ddir));
 150
 151         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 152                 return 0;
 153
 154         /*
 155          * allow a 2 second settle period in the beginning
 156          */
 157         if (mtime_since(&td->start, now) < 2000)
 158                 return 0;
 159
 160         iops += td->this_io_blocks[ddir];
 161         bytes += td->this_io_bytes[ddir];
 162         ratemin += td->o.ratemin[ddir];
 163         rate_iops += td->o.rate_iops[ddir];
 164         rate_iops_min += td->o.rate_iops_min[ddir];
 165
 166         /*
 167          * if rate blocks is set, sample is running
 168          */
 169         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 170                 spent = mtime_since(&td->lastrate[ddir], now);
 171                 if (spent < td->o.ratecycle)
 172                         return 0;
 173
 174                 if (td->o.rate[ddir]) {
 175                         /*
 176                          * check bandwidth specified rate
 177                          */
 178                         if (bytes < td->rate_bytes[ddir]) {
 179                                 log_err("%s: min rate %u not met\n", td->o.name,
 180                                                                 ratemin);
 181                                 return 1;
 182                         } else {
 183                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 184                                 if (rate < ratemin ||
 185                                     bytes < td->rate_bytes[ddir]) {
 186                                         log_err("%s: min rate %u not met, got"
 187                                                 " %luKB/sec\n", td->o.name,
 188                                                         ratemin, rate);
 189                                         return 1;
 190                                 }
 191                         }
 192                 } else {
 193                         /*
 194                          * checks iops specified rate
 195                          */
 196                         if (iops < rate_iops) {
 197                                 log_err("%s: min iops rate %u not met\n",
 198                                                 td->o.name, rate_iops);
 199                                 return 1;
 200                         } else {
 201                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 202                                 if (rate < rate_iops_min ||
 203                                     iops < td->rate_blocks[ddir]) {
 204                                         log_err("%s: min iops rate %u not met,"
 205                                                 " got %lu\n", td->o.name,
 206                                                         rate_iops_min, rate);
 207                                 }
 208                         }
 209                 }
 210         }
 211
 212         td->rate_bytes[ddir] = bytes;
 213         td->rate_blocks[ddir] = iops;
 214         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 215         return 0;
 216 }
 217
 218 static int check_min_rate(struct thread_data *td, struct timeval *now,
 219                           unsigned long *bytes_done)
 220 {
 221         int ret = 0;
 222
 223         if (bytes_done[DDIR_READ])
 224                 ret |= __check_min_rate(td, now, DDIR_READ);
 225         if (bytes_done[DDIR_WRITE])
 226                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 227         if (bytes_done[DDIR_TRIM])
 228                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 229
 230         return ret;
 231 }
 232
 233 /*
 234  * When job exits, we can cancel the in-flight IO if we are using async
 235  * io. Attempt to do so.
 236  */
 237 static void cleanup_pending_aio(struct thread_data *td)
 238 {
 239         struct flist_head *entry, *n;
 240         struct io_u *io_u;
 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                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 255                         io_u = flist_entry(entry, struct io_u, list);
 256
 257                         /*
 258                          * if the io_u isn't in flight, then that generally
 259                          * means someone leaked an io_u. complain but fix
 260                          * it up, so we don't stall here.
 261                          */
 262                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 263                                 log_err("fio: non-busy IO on busy list\n");
 264                                 put_io_u(td, io_u);
 265                         } else {
 266                                 r = td->io_ops->cancel(td, io_u);
 267                                 if (!r)
 268                                         put_io_u(td, io_u);
 269                         }
 270                 }
 271         }
 272
 273         if (td->cur_depth)
 274                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 275 }
 276
 277 /*
 278  * Helper to handle the final sync of a file. Works just like the normal
 279  * io path, just does everything sync.
 280  */
 281 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 282 {
 283         struct io_u *io_u = __get_io_u(td);
 284         int ret;
 285
 286         if (!io_u)
 287                 return 1;
 288
 289         io_u->ddir = DDIR_SYNC;
 290         io_u->file = f;
 291
 292         if (td_io_prep(td, io_u)) {
 293                 put_io_u(td, io_u);
 294                 return 1;
 295         }
 296
 297 requeue:
 298         ret = td_io_queue(td, io_u);
 299         if (ret < 0) {
 300                 td_verror(td, io_u->error, "td_io_queue");
 301                 put_io_u(td, io_u);
 302                 return 1;
 303         } else if (ret == FIO_Q_QUEUED) {
 304                 if (io_u_queued_complete(td, 1, NULL) < 0)
 305                         return 1;
 306         } else if (ret == FIO_Q_COMPLETED) {
 307                 if (io_u->error) {
 308                         td_verror(td, io_u->error, "td_io_queue");
 309                         return 1;
 310                 }
 311
 312                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 313                         return 1;
 314         } else if (ret == FIO_Q_BUSY) {
 315                 if (td_io_commit(td))
 316                         return 1;
 317                 goto requeue;
 318         }
 319
 320         return 0;
 321 }
 322
 323 static inline void __update_tv_cache(struct thread_data *td)
 324 {
 325         fio_gettime(&td->tv_cache, NULL);
 326 }
 327
 328 static inline void update_tv_cache(struct thread_data *td)
 329 {
 330         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 331                 __update_tv_cache(td);
 332 }
 333
 334 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 335 {
 336         if (in_ramp_time(td))
 337                 return 0;
 338         if (!td->o.timeout)
 339                 return 0;
 340         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 341                 return 1;
 342
 343         return 0;
 344 }
 345
 346 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 347                                int *retptr)
 348 {
 349         int ret = *retptr;
 350
 351         if (ret < 0 || td->error) {
 352                 int err = td->error;
 353                 enum error_type_bit eb;
 354
 355                 if (ret < 0)
 356                         err = -ret;
 357
 358                 eb = td_error_type(ddir, err);
 359                 if (!(td->o.continue_on_error & (1 << eb)))
 360                         return 1;
 361
 362                 if (td_non_fatal_error(td, eb, err)) {
 363                         /*
 364                          * Continue with the I/Os in case of
 365                          * a non fatal error.
 366                          */
 367                         update_error_count(td, err);
 368                         td_clear_error(td);
 369                         *retptr = 0;
 370                         return 0;
 371                 } else if (td->o.fill_device && err == ENOSPC) {
 372                         /*
 373                          * We expect to hit this error if
 374                          * fill_device option is set.
 375                          */
 376                         td_clear_error(td);
 377                         td->terminate = 1;
 378                         return 1;
 379                 } else {
 380                         /*
 381                          * Stop the I/O in case of a fatal
 382                          * error.
 383                          */
 384                         update_error_count(td, err);
 385                         return 1;
 386                 }
 387         }
 388
 389         return 0;
 390 }
 391
 392 /*
 393  * The main verify engine. Runs over the writes we previously submitted,
 394  * reads the blocks back in, and checks the crc/md5 of the data.
 395  */
 396 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 397 {
 398         struct fio_file *f;
 399         struct io_u *io_u;
 400         int ret, min_events;
 401         uint64_t io_bytes;
 402         unsigned int i;
 403
 404         dprint(FD_VERIFY, "starting loop\n");
 405
 406         /*
 407          * sync io first and invalidate cache, to make sure we really
 408          * read from disk.
 409          */
 410         for_each_file(td, f, i) {
 411                 if (!fio_file_open(f))
 412                         continue;
 413                 if (fio_io_sync(td, f))
 414                         break;
 415                 if (file_invalidate_cache(td, f))
 416                         break;
 417         }
 418
 419         if (td->error)
 420                 return;
 421
 422         td_set_runstate(td, TD_VERIFYING);
 423
 424         io_u = NULL;
 425         io_bytes = 0;
 426         while (!td->terminate) {
 427                 enum fio_ddir ddir;
 428                 int ret2, full;
 429
 430                 update_tv_cache(td);
 431
 432                 if (runtime_exceeded(td, &td->tv_cache)) {
 433                         __update_tv_cache(td);
 434                         if (runtime_exceeded(td, &td->tv_cache)) {
 435                                 td->terminate = 1;
 436                                 break;
 437                         }
 438                 }
 439
 440                 if (flow_threshold_exceeded(td))
 441                         continue;
 442
 443                 if (!td->o.experimental_verify) {
 444                         io_u = __get_io_u(td);
 445                         if (!io_u)
 446                                 break;
 447
 448                         if (get_next_verify(td, io_u)) {
 449                                 put_io_u(td, io_u);
 450                                 break;
 451                         }
 452
 453                         if (td_io_prep(td, io_u)) {
 454                                 put_io_u(td, io_u);
 455                                 break;
 456                         }
 457                 } else {
 458                         io_u = get_io_u(td);
 459                         if (!io_u)
 460                                 break;
 461
 462                         if (io_u->buflen + io_bytes > verify_bytes)
 463                                 break;
 464                 }
 465
 466                 if (td->o.verify_async)
 467                         io_u->end_io = verify_io_u_async;
 468                 else
 469                         io_u->end_io = verify_io_u;
 470
 471                 ddir = io_u->ddir;
 472
 473                 ret = td_io_queue(td, io_u);
 474                 switch (ret) {
 475                 case FIO_Q_COMPLETED:
 476                         if (io_u->error) {
 477                                 ret = -io_u->error;
 478                                 clear_io_u(td, io_u);
 479                         } else if (io_u->resid) {
 480                                 int bytes = io_u->xfer_buflen - io_u->resid;
 481
 482                                 /*
 483                                  * zero read, fail
 484                                  */
 485                                 if (!bytes) {
 486                                         td_verror(td, EIO, "full resid");
 487                                         put_io_u(td, io_u);
 488                                         break;
 489                                 }
 490
 491                                 io_u->xfer_buflen = io_u->resid;
 492                                 io_u->xfer_buf += bytes;
 493                                 io_u->offset += bytes;
 494                                 io_bytes += bytes;
 495
 496                                 if (ddir_rw(io_u->ddir))
 497                                         td->ts.short_io_u[io_u->ddir]++;
 498
 499                                 f = io_u->file;
 500                                 if (io_u->offset == f->real_file_size)
 501                                         goto sync_done;
 502
 503                                 requeue_io_u(td, &io_u);
 504                         } else {
 505 sync_done:
 506                                 ret = io_u_sync_complete(td, io_u, NULL);
 507                                 if (ret < 0)
 508                                         break;
 509                         }
 510                         io_bytes += io_u->xfer_buflen;
 511                         continue;
 512                 case FIO_Q_QUEUED:
 513                         break;
 514                 case FIO_Q_BUSY:
 515                         requeue_io_u(td, &io_u);
 516                         ret2 = td_io_commit(td);
 517                         if (ret2 < 0)
 518                                 ret = ret2;
 519                         break;
 520                 default:
 521                         assert(ret < 0);
 522                         td_verror(td, -ret, "td_io_queue");
 523                         break;
 524                 }
 525
 526                 if (break_on_this_error(td, ddir, &ret))
 527                         break;
 528
 529                 /*
 530                  * if we can queue more, do so. but check if there are
 531                  * completed io_u's first. Note that we can get BUSY even
 532                  * without IO queued, if the system is resource starved.
 533                  */
 534                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 535                 if (full || !td->o.iodepth_batch_complete) {
 536                         min_events = min(td->o.iodepth_batch_complete,
 537                                          td->cur_depth);
 538                         /*
 539                          * if the queue is full, we MUST reap at least 1 event
 540                          */
 541                         if (full && !min_events)
 542                                 min_events = 1;
 543
 544                         do {
 545                                 unsigned long bytes = 0;
 546
 547                                 /*
 548                                  * Reap required number of io units, if any,
 549                                  * and do the verification on them through
 550                                  * the callback handler
 551                                  */
 552                                 if (io_u_queued_complete(td, min_events, &bytes) < 0) {
 553                                         ret = -1;
 554                                         break;
 555                                 }
 556                                 io_bytes += bytes;
 557                         } while (full && (td->cur_depth > td->o.iodepth_low));
 558                 }
 559                 if (ret < 0)
 560                         break;
 561         }
 562
 563         if (!td->error) {
 564                 min_events = td->cur_depth;
 565
 566                 if (min_events)
 567                         ret = io_u_queued_complete(td, min_events, NULL);
 568         } else
 569                 cleanup_pending_aio(td);
 570
 571         td_set_runstate(td, TD_RUNNING);
 572
 573         dprint(FD_VERIFY, "exiting loop\n");
 574 }
 575
 576 static int io_bytes_exceeded(struct thread_data *td)
 577 {
 578         unsigned long long bytes;
 579
 580         if (td_rw(td))
 581                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 582         else if (td_write(td))
 583                 bytes = td->this_io_bytes[DDIR_WRITE];
 584         else if (td_read(td))
 585                 bytes = td->this_io_bytes[DDIR_READ];
 586         else
 587                 bytes = td->this_io_bytes[DDIR_TRIM];
 588
 589         return bytes >= td->o.size;
 590 }
 591
 592 /*
 593  * Main IO worker function. It retrieves io_u's to process and queues
 594  * and reaps them, checking for rate and errors along the way.
 595  */
 596 static void do_io(struct thread_data *td)
 597 {
 598         unsigned int i;
 599         int ret = 0;
 600
 601         if (in_ramp_time(td))
 602                 td_set_runstate(td, TD_RAMP);
 603         else
 604                 td_set_runstate(td, TD_RUNNING);
 605
 606         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 607                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 608                 td->o.time_based) {
 609                 struct timeval comp_time;
 610                 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 611                 int min_evts = 0;
 612                 struct io_u *io_u;
 613                 int ret2, full;
 614                 enum fio_ddir ddir;
 615
 616                 if (td->terminate || td->done)
 617                         break;
 618
 619                 update_tv_cache(td);
 620
 621                 if (runtime_exceeded(td, &td->tv_cache)) {
 622                         __update_tv_cache(td);
 623                         if (runtime_exceeded(td, &td->tv_cache)) {
 624                                 td->terminate = 1;
 625                                 break;
 626                         }
 627                 }
 628
 629                 if (flow_threshold_exceeded(td))
 630                         continue;
 631
 632                 io_u = get_io_u(td);
 633                 if (!io_u)
 634                         break;
 635
 636                 ddir = io_u->ddir;
 637
 638                 /*
 639                  * Add verification end_io handler if:
 640                  *      - Asked to verify (!td_rw(td))
 641                  *      - Or the io_u is from our verify list (mixed write/ver)
 642                  */
 643                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 644                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 645                         if (td->o.verify_async)
 646                                 io_u->end_io = verify_io_u_async;
 647                         else
 648                                 io_u->end_io = verify_io_u;
 649                         td_set_runstate(td, TD_VERIFYING);
 650                 } else if (in_ramp_time(td))
 651                         td_set_runstate(td, TD_RAMP);
 652                 else
 653                         td_set_runstate(td, TD_RUNNING);
 654
 655                 ret = td_io_queue(td, io_u);
 656                 switch (ret) {
 657                 case FIO_Q_COMPLETED:
 658                         if (io_u->error) {
 659                                 ret = -io_u->error;
 660                                 clear_io_u(td, io_u);
 661                         } else if (io_u->resid) {
 662                                 int bytes = io_u->xfer_buflen - io_u->resid;
 663                                 struct fio_file *f = io_u->file;
 664
 665                                 /*
 666                                  * zero read, fail
 667                                  */
 668                                 if (!bytes) {
 669                                         td_verror(td, EIO, "full resid");
 670                                         put_io_u(td, io_u);
 671                                         break;
 672                                 }
 673
 674                                 io_u->xfer_buflen = io_u->resid;
 675                                 io_u->xfer_buf += bytes;
 676                                 io_u->offset += bytes;
 677
 678                                 if (ddir_rw(io_u->ddir))
 679                                         td->ts.short_io_u[io_u->ddir]++;
 680
 681                                 if (io_u->offset == f->real_file_size)
 682                                         goto sync_done;
 683
 684                                 requeue_io_u(td, &io_u);
 685                         } else {
 686 sync_done:
 687                                 if (__should_check_rate(td, DDIR_READ) ||
 688                                     __should_check_rate(td, DDIR_WRITE) ||
 689                                     __should_check_rate(td, DDIR_TRIM))
 690                                         fio_gettime(&comp_time, NULL);
 691
 692                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 693                                 if (ret < 0)
 694                                         break;
 695                         }
 696                         break;
 697                 case FIO_Q_QUEUED:
 698                         /*
 699                          * if the engine doesn't have a commit hook,
 700                          * the io_u is really queued. if it does have such
 701                          * a hook, it has to call io_u_queued() itself.
 702                          */
 703                         if (td->io_ops->commit == NULL)
 704                                 io_u_queued(td, io_u);
 705                         break;
 706                 case FIO_Q_BUSY:
 707                         requeue_io_u(td, &io_u);
 708                         ret2 = td_io_commit(td);
 709                         if (ret2 < 0)
 710                                 ret = ret2;
 711                         break;
 712                 default:
 713                         assert(ret < 0);
 714                         put_io_u(td, io_u);
 715                         break;
 716                 }
 717
 718                 if (break_on_this_error(td, ddir, &ret))
 719                         break;
 720
 721                 /*
 722                  * See if we need to complete some commands. Note that we
 723                  * can get BUSY even without IO queued, if the system is
 724                  * resource starved.
 725                  */
 726                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 727                 if (full || !td->o.iodepth_batch_complete) {
 728                         min_evts = min(td->o.iodepth_batch_complete,
 729                                         td->cur_depth);
 730                         /*
 731                          * if the queue is full, we MUST reap at least 1 event
 732                          */
 733                         if (full && !min_evts)
 734                                 min_evts = 1;
 735
 736                         if (__should_check_rate(td, DDIR_READ) ||
 737                             __should_check_rate(td, DDIR_WRITE) ||
 738                             __should_check_rate(td, DDIR_TRIM))
 739                                 fio_gettime(&comp_time, NULL);
 740
 741                         do {
 742                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 743                                 if (ret < 0)
 744                                         break;
 745
 746                         } while (full && (td->cur_depth > td->o.iodepth_low));
 747                 }
 748
 749                 if (ret < 0)
 750                         break;
 751                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 752                         continue;
 753
 754                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 755                         if (check_min_rate(td, &comp_time, bytes_done)) {
 756                                 if (exitall_on_terminate)
 757                                         fio_terminate_threads(td->groupid);
 758                                 td_verror(td, EIO, "check_min_rate");
 759                                 break;
 760                         }
 761                 }
 762
 763                 if (td->o.thinktime) {
 764                         unsigned long long b;
 765
 766                         b = ddir_rw_sum(td->io_blocks);
 767                         if (!(b % td->o.thinktime_blocks)) {
 768                                 int left;
 769
 770                                 if (td->o.thinktime_spin)
 771                                         usec_spin(td->o.thinktime_spin);
 772
 773                                 left = td->o.thinktime - td->o.thinktime_spin;
 774                                 if (left)
 775                                         usec_sleep(td, left);
 776                         }
 777                 }
 778         }
 779
 780         if (td->trim_entries)
 781                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 782
 783         if (td->o.fill_device && td->error == ENOSPC) {
 784                 td->error = 0;
 785                 td->terminate = 1;
 786         }
 787         if (!td->error) {
 788                 struct fio_file *f;
 789
 790                 i = td->cur_depth;
 791                 if (i) {
 792                         ret = io_u_queued_complete(td, i, NULL);
 793                         if (td->o.fill_device && td->error == ENOSPC)
 794                                 td->error = 0;
 795                 }
 796
 797                 if (should_fsync(td) && td->o.end_fsync) {
 798                         td_set_runstate(td, TD_FSYNCING);
 799
 800                         for_each_file(td, f, i) {
 801                                 if (!fio_file_open(f))
 802                                         continue;
 803                                 fio_io_sync(td, f);
 804                         }
 805                 }
 806         } else
 807                 cleanup_pending_aio(td);
 808
 809         /*
 810          * stop job if we failed doing any IO
 811          */
 812         if (!ddir_rw_sum(td->this_io_bytes))
 813                 td->done = 1;
 814 }
 815
 816 static void cleanup_io_u(struct thread_data *td)
 817 {
 818         struct flist_head *entry, *n;
 819         struct io_u *io_u;
 820
 821         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 822                 io_u = flist_entry(entry, struct io_u, list);
 823
 824                 flist_del(&io_u->list);
 825
 826                 if (td->io_ops->io_u_free)
 827                         td->io_ops->io_u_free(td, io_u);
 828
 829                 fio_memfree(io_u, sizeof(*io_u));
 830         }
 831
 832         free_io_mem(td);
 833 }
 834
 835 static int init_io_u(struct thread_data *td)
 836 {
 837         struct io_u *io_u;
 838         unsigned int max_bs, min_write;
 839         int cl_align, i, max_units;
 840         int data_xfer = 1;
 841         char *p;
 842
 843         max_units = td->o.iodepth;
 844         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 845         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
 846         min_write = td->o.min_bs[DDIR_WRITE];
 847         td->orig_buffer_size = (unsigned long long) max_bs
 848                                         * (unsigned long long) max_units;
 849
 850         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 851                 data_xfer = 0;
 852
 853         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 854                 unsigned long bs;
 855
 856                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 857                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 858         }
 859
 860         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 861                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 862                 return 1;
 863         }
 864
 865         if (data_xfer && allocate_io_mem(td))
 866                 return 1;
 867
 868         if (td->o.odirect || td->o.mem_align ||
 869             (td->io_ops->flags & FIO_RAWIO))
 870                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 871         else
 872                 p = td->orig_buffer;
 873
 874         cl_align = os_cache_line_size();
 875
 876         for (i = 0; i < max_units; i++) {
 877                 void *ptr;
 878
 879                 if (td->terminate)
 880                         return 1;
 881
 882                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 883                 if (!ptr) {
 884                         log_err("fio: unable to allocate aligned memory\n");
 885                         break;
 886                 }
 887
 888                 io_u = ptr;
 889                 memset(io_u, 0, sizeof(*io_u));
 890                 INIT_FLIST_HEAD(&io_u->list);
 891                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 892
 893                 if (data_xfer) {
 894                         io_u->buf = p;
 895                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 896
 897                         if (td_write(td))
 898                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 899                         if (td_write(td) && td->o.verify_pattern_bytes) {
 900                                 /*
 901                                  * Fill the buffer with the pattern if we are
 902                                  * going to be doing writes.
 903                                  */
 904                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 905                         }
 906                 }
 907
 908                 io_u->index = i;
 909                 io_u->flags = IO_U_F_FREE;
 910                 flist_add(&io_u->list, &td->io_u_freelist);
 911
 912                 if (td->io_ops->io_u_init) {
 913                         int ret = td->io_ops->io_u_init(td, io_u);
 914
 915                         if (ret) {
 916                                 log_err("fio: failed to init engine data: %d\n", ret);
 917                                 return 1;
 918                         }
 919                 }
 920
 921                 p += max_bs;
 922         }
 923
 924         return 0;
 925 }
 926
 927 static int switch_ioscheduler(struct thread_data *td)
 928 {
 929         char tmp[256], tmp2[128];
 930         FILE *f;
 931         int ret;
 932
 933         if (td->io_ops->flags & FIO_DISKLESSIO)
 934                 return 0;
 935
 936         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 937
 938         f = fopen(tmp, "r+");
 939         if (!f) {
 940                 if (errno == ENOENT) {
 941                         log_err("fio: os or kernel doesn't support IO scheduler"
 942                                 " switching\n");
 943                         return 0;
 944                 }
 945                 td_verror(td, errno, "fopen iosched");
 946                 return 1;
 947         }
 948
 949         /*
 950          * Set io scheduler.
 951          */
 952         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 953         if (ferror(f) || ret != 1) {
 954                 td_verror(td, errno, "fwrite");
 955                 fclose(f);
 956                 return 1;
 957         }
 958
 959         rewind(f);
 960
 961         /*
 962          * Read back and check that the selected scheduler is now the default.
 963          */
 964         ret = fread(tmp, 1, sizeof(tmp), f);
 965         if (ferror(f) || ret < 0) {
 966                 td_verror(td, errno, "fread");
 967                 fclose(f);
 968                 return 1;
 969         }
 970
 971         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 972         if (!strstr(tmp, tmp2)) {
 973                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 974                 td_verror(td, EINVAL, "iosched_switch");
 975                 fclose(f);
 976                 return 1;
 977         }
 978
 979         fclose(f);
 980         return 0;
 981 }
 982
 983 static int keep_running(struct thread_data *td)
 984 {
 985         if (td->done)
 986                 return 0;
 987         if (td->o.time_based)
 988                 return 1;
 989         if (td->o.loops) {
 990                 td->o.loops--;
 991                 return 1;
 992         }
 993
 994         if (ddir_rw_sum(td->io_bytes) < td->o.size)
 995                 return 1;
 996
 997         return 0;
 998 }
 999
1000 static int exec_string(const char *string)
1001 {
1002         int ret, newlen = strlen(string) + 1 + 8;
1003         char *str;
1004
1005         str = malloc(newlen);
1006         sprintf(str, "sh -c %s", string);
1007
1008         ret = system(str);
1009         if (ret == -1)
1010                 log_err("fio: exec of cmd <%s> failed\n", str);
1011
1012         free(str);
1013         return ret;
1014 }
1015
1016 /*
1017  * Entry point for the thread based jobs. The process based jobs end up
1018  * here as well, after a little setup.
1019  */
1020 static void *thread_main(void *data)
1021 {
1022         unsigned long long elapsed;
1023         struct thread_data *td = data;
1024         pthread_condattr_t attr;
1025         int clear_state;
1026
1027         if (!td->o.use_thread) {
1028                 setsid();
1029                 td->pid = getpid();
1030         } else
1031                 td->pid = gettid();
1032
1033         fio_local_clock_init(td->o.use_thread);
1034
1035         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1036
1037         INIT_FLIST_HEAD(&td->io_u_freelist);
1038         INIT_FLIST_HEAD(&td->io_u_busylist);
1039         INIT_FLIST_HEAD(&td->io_u_requeues);
1040         INIT_FLIST_HEAD(&td->io_log_list);
1041         INIT_FLIST_HEAD(&td->io_hist_list);
1042         INIT_FLIST_HEAD(&td->verify_list);
1043         INIT_FLIST_HEAD(&td->trim_list);
1044         INIT_FLIST_HEAD(&td->next_rand_list);
1045         pthread_mutex_init(&td->io_u_lock, NULL);
1046         td->io_hist_tree = RB_ROOT;
1047
1048         pthread_condattr_init(&attr);
1049         pthread_cond_init(&td->verify_cond, &attr);
1050         pthread_cond_init(&td->free_cond, &attr);
1051
1052         td_set_runstate(td, TD_INITIALIZED);
1053         dprint(FD_MUTEX, "up startup_mutex\n");
1054         fio_mutex_up(startup_mutex);
1055         dprint(FD_MUTEX, "wait on td->mutex\n");
1056         fio_mutex_down(td->mutex);
1057         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1058
1059         /*
1060          * the ->mutex mutex is now no longer used, close it to avoid
1061          * eating a file descriptor
1062          */
1063         fio_mutex_remove(td->mutex);
1064
1065         /*
1066          * A new gid requires privilege, so we need to do this before setting
1067          * the uid.
1068          */
1069         if (td->o.gid != -1U && setgid(td->o.gid)) {
1070                 td_verror(td, errno, "setgid");
1071                 goto err;
1072         }
1073         if (td->o.uid != -1U && setuid(td->o.uid)) {
1074                 td_verror(td, errno, "setuid");
1075                 goto err;
1076         }
1077
1078         /*
1079          * If we have a gettimeofday() thread, make sure we exclude that
1080          * thread from this job
1081          */
1082         if (td->o.gtod_cpu)
1083                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1084
1085         /*
1086          * Set affinity first, in case it has an impact on the memory
1087          * allocations.
1088          */
1089         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1090                 td_verror(td, errno, "cpu_set_affinity");
1091                 goto err;
1092         }
1093
1094 #ifdef CONFIG_LIBNUMA
1095         /* numa node setup */
1096         if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1097                 int ret;
1098
1099                 if (numa_available() < 0) {
1100                         td_verror(td, errno, "Does not support NUMA API\n");
1101                         goto err;
1102                 }
1103
1104                 if (td->o.numa_cpumask_set) {
1105                         ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1106                         if (ret == -1) {
1107                                 td_verror(td, errno, \
1108                                         "numa_run_on_node_mask failed\n");
1109                                 goto err;
1110                         }
1111                 }
1112
1113                 if (td->o.numa_memmask_set) {
1114
1115                         switch (td->o.numa_mem_mode) {
1116                         case MPOL_INTERLEAVE:
1117                                 numa_set_interleave_mask(td->o.numa_memnodesmask);
1118                                 break;
1119                         case MPOL_BIND:
1120                                 numa_set_membind(td->o.numa_memnodesmask);
1121                                 break;
1122                         case MPOL_LOCAL:
1123                                 numa_set_localalloc();
1124                                 break;
1125                         case MPOL_PREFERRED:
1126                                 numa_set_preferred(td->o.numa_mem_prefer_node);
1127                                 break;
1128                         case MPOL_DEFAULT:
1129                         default:
1130                                 break;
1131                         }
1132
1133                 }
1134         }
1135 #endif
1136
1137         /*
1138          * May alter parameters that init_io_u() will use, so we need to
1139          * do this first.
1140          */
1141         if (init_iolog(td))
1142                 goto err;
1143
1144         if (init_io_u(td))
1145                 goto err;
1146
1147         if (td->o.verify_async && verify_async_init(td))
1148                 goto err;
1149
1150         if (td->ioprio_set) {
1151                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1152                         td_verror(td, errno, "ioprio_set");
1153                         goto err;
1154                 }
1155         }
1156
1157         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1158                 goto err;
1159
1160         errno = 0;
1161         if (nice(td->o.nice) == -1 && errno != 0) {
1162                 td_verror(td, errno, "nice");
1163                 goto err;
1164         }
1165
1166         if (td->o.ioscheduler && switch_ioscheduler(td))
1167                 goto err;
1168
1169         if (!td->o.create_serialize && setup_files(td))
1170                 goto err;
1171
1172         if (td_io_init(td))
1173                 goto err;
1174
1175         if (init_random_map(td))
1176                 goto err;
1177
1178         if (td->o.exec_prerun) {
1179                 if (exec_string(td->o.exec_prerun))
1180                         goto err;
1181         }
1182
1183         if (td->o.pre_read) {
1184                 if (pre_read_files(td) < 0)
1185                         goto err;
1186         }
1187
1188         fio_gettime(&td->epoch, NULL);
1189         getrusage(RUSAGE_SELF, &td->ru_start);
1190
1191         clear_state = 0;
1192         while (keep_running(td)) {
1193                 uint64_t write_bytes;
1194
1195                 fio_gettime(&td->start, NULL);
1196                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1197                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1198                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1199
1200                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1201                                 td->o.ratemin[DDIR_TRIM]) {
1202                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1203                                                 sizeof(td->bw_sample_time));
1204                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1205                                                 sizeof(td->bw_sample_time));
1206                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1207                                                 sizeof(td->bw_sample_time));
1208                 }
1209
1210                 if (clear_state)
1211                         clear_io_state(td);
1212
1213                 prune_io_piece_log(td);
1214
1215                 write_bytes = td->io_bytes[DDIR_WRITE];
1216                 do_io(td);
1217                 write_bytes = td->io_bytes[DDIR_WRITE] - write_bytes;
1218
1219                 clear_state = 1;
1220
1221                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1222                         elapsed = utime_since_now(&td->start);
1223                         td->ts.runtime[DDIR_READ] += elapsed;
1224                 }
1225                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1226                         elapsed = utime_since_now(&td->start);
1227                         td->ts.runtime[DDIR_WRITE] += elapsed;
1228                 }
1229                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1230                         elapsed = utime_since_now(&td->start);
1231                         td->ts.runtime[DDIR_TRIM] += elapsed;
1232                 }
1233
1234                 if (td->error || td->terminate)
1235                         break;
1236
1237                 if (!td->o.do_verify ||
1238                     td->o.verify == VERIFY_NONE ||
1239                     (td->io_ops->flags & FIO_UNIDIR))
1240                         continue;
1241
1242                 clear_io_state(td);
1243
1244                 fio_gettime(&td->start, NULL);
1245
1246                 do_verify(td, write_bytes);
1247
1248                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1249
1250                 if (td->error || td->terminate)
1251                         break;
1252         }
1253
1254         update_rusage_stat(td);
1255         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1256         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1257         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1258         td->ts.total_run_time = mtime_since_now(&td->epoch);
1259         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1260         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1261         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1262
1263         fio_mutex_down(writeout_mutex);
1264         if (td->bw_log) {
1265                 if (td->o.bw_log_file) {
1266                         finish_log_named(td, td->bw_log,
1267                                                 td->o.bw_log_file, "bw");
1268                 } else
1269                         finish_log(td, td->bw_log, "bw");
1270         }
1271         if (td->lat_log) {
1272                 if (td->o.lat_log_file) {
1273                         finish_log_named(td, td->lat_log,
1274                                                 td->o.lat_log_file, "lat");
1275                 } else
1276                         finish_log(td, td->lat_log, "lat");
1277         }
1278         if (td->slat_log) {
1279                 if (td->o.lat_log_file) {
1280                         finish_log_named(td, td->slat_log,
1281                                                 td->o.lat_log_file, "slat");
1282                 } else
1283                         finish_log(td, td->slat_log, "slat");
1284         }
1285         if (td->clat_log) {
1286                 if (td->o.lat_log_file) {
1287                         finish_log_named(td, td->clat_log,
1288                                                 td->o.lat_log_file, "clat");
1289                 } else
1290                         finish_log(td, td->clat_log, "clat");
1291         }
1292         if (td->iops_log) {
1293                 if (td->o.iops_log_file) {
1294                         finish_log_named(td, td->iops_log,
1295                                                 td->o.iops_log_file, "iops");
1296                 } else
1297                         finish_log(td, td->iops_log, "iops");
1298         }
1299
1300         fio_mutex_up(writeout_mutex);
1301         if (td->o.exec_postrun)
1302                 exec_string(td->o.exec_postrun);
1303
1304         if (exitall_on_terminate)
1305                 fio_terminate_threads(td->groupid);
1306
1307 err:
1308         if (td->error)
1309                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1310                                                         td->verror);
1311
1312         if (td->o.verify_async)
1313                 verify_async_exit(td);
1314
1315         close_and_free_files(td);
1316         cleanup_io_u(td);
1317         close_ioengine(td);
1318         cgroup_shutdown(td, &cgroup_mnt);
1319
1320         if (td->o.cpumask_set) {
1321                 int ret = fio_cpuset_exit(&td->o.cpumask);
1322
1323                 td_verror(td, ret, "fio_cpuset_exit");
1324         }
1325
1326         /*
1327          * do this very late, it will log file closing as well
1328          */
1329         if (td->o.write_iolog_file)
1330                 write_iolog_close(td);
1331
1332         td_set_runstate(td, TD_EXITED);
1333         return (void *) (uintptr_t) td->error;
1334 }
1335
1336
1337 /*
1338  * We cannot pass the td data into a forked process, so attach the td and
1339  * pass it to the thread worker.
1340  */
1341 static int fork_main(int shmid, int offset)
1342 {
1343         struct thread_data *td;
1344         void *data, *ret;
1345
1346 #ifndef __hpux
1347         data = shmat(shmid, NULL, 0);
1348         if (data == (void *) -1) {
1349                 int __err = errno;
1350
1351                 perror("shmat");
1352                 return __err;
1353         }
1354 #else
1355         /*
1356          * HP-UX inherits shm mappings?
1357          */
1358         data = threads;
1359 #endif
1360
1361         td = data + offset * sizeof(struct thread_data);
1362         ret = thread_main(td);
1363         shmdt(data);
1364         return (int) (uintptr_t) ret;
1365 }
1366
1367 /*
1368  * Run over the job map and reap the threads that have exited, if any.
1369  */
1370 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1371                          unsigned int *m_rate)
1372 {
1373         struct thread_data *td;
1374         unsigned int cputhreads, realthreads, pending;
1375         int i, status, ret;
1376
1377         /*
1378          * reap exited threads (TD_EXITED -> TD_REAPED)
1379          */
1380         realthreads = pending = cputhreads = 0;
1381         for_each_td(td, i) {
1382                 int flags = 0;
1383
1384                 /*
1385                  * ->io_ops is NULL for a thread that has closed its
1386                  * io engine
1387                  */
1388                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1389                         cputhreads++;
1390                 else
1391                         realthreads++;
1392
1393                 if (!td->pid) {
1394                         pending++;
1395                         continue;
1396                 }
1397                 if (td->runstate == TD_REAPED)
1398                         continue;
1399                 if (td->o.use_thread) {
1400                         if (td->runstate == TD_EXITED) {
1401                                 td_set_runstate(td, TD_REAPED);
1402                                 goto reaped;
1403                         }
1404                         continue;
1405                 }
1406
1407                 flags = WNOHANG;
1408                 if (td->runstate == TD_EXITED)
1409                         flags = 0;
1410
1411                 /*
1412                  * check if someone quit or got killed in an unusual way
1413                  */
1414                 ret = waitpid(td->pid, &status, flags);
1415                 if (ret < 0) {
1416                         if (errno == ECHILD) {
1417                                 log_err("fio: pid=%d disappeared %d\n",
1418                                                 (int) td->pid, td->runstate);
1419                                 td->sig = ECHILD;
1420                                 td_set_runstate(td, TD_REAPED);
1421                                 goto reaped;
1422                         }
1423                         perror("waitpid");
1424                 } else if (ret == td->pid) {
1425                         if (WIFSIGNALED(status)) {
1426                                 int sig = WTERMSIG(status);
1427
1428                                 if (sig != SIGTERM && sig != SIGUSR2)
1429                                         log_err("fio: pid=%d, got signal=%d\n",
1430                                                         (int) td->pid, sig);
1431                                 td->sig = sig;
1432                                 td_set_runstate(td, TD_REAPED);
1433                                 goto reaped;
1434                         }
1435                         if (WIFEXITED(status)) {
1436                                 if (WEXITSTATUS(status) && !td->error)
1437                                         td->error = WEXITSTATUS(status);
1438
1439                                 td_set_runstate(td, TD_REAPED);
1440                                 goto reaped;
1441                         }
1442                 }
1443
1444                 /*
1445                  * thread is not dead, continue
1446                  */
1447                 pending++;
1448                 continue;
1449 reaped:
1450                 (*nr_running)--;
1451                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1452                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1453                 if (!td->pid)
1454                         pending--;
1455
1456                 if (td->error)
1457                         exit_value++;
1458
1459                 done_secs += mtime_since_now(&td->epoch) / 1000;
1460         }
1461
1462         if (*nr_running == cputhreads && !pending && realthreads)
1463                 fio_terminate_threads(TERMINATE_ALL);
1464 }
1465
1466 /*
1467  * Main function for kicking off and reaping jobs, as needed.
1468  */
1469 static void run_threads(void)
1470 {
1471         struct thread_data *td;
1472         unsigned long spent;
1473         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1474
1475         if (fio_pin_memory())
1476                 return;
1477
1478         if (fio_gtod_offload && fio_start_gtod_thread())
1479                 return;
1480
1481         set_sig_handlers();
1482
1483         if (output_format == FIO_OUTPUT_NORMAL) {
1484                 log_info("Starting ");
1485                 if (nr_thread)
1486                         log_info("%d thread%s", nr_thread,
1487                                                 nr_thread > 1 ? "s" : "");
1488                 if (nr_process) {
1489                         if (nr_thread)
1490                                 log_info(" and ");
1491                         log_info("%d process%s", nr_process,
1492                                                 nr_process > 1 ? "es" : "");
1493                 }
1494                 log_info("\n");
1495                 fflush(stdout);
1496         }
1497
1498         todo = thread_number;
1499         nr_running = 0;
1500         nr_started = 0;
1501         m_rate = t_rate = 0;
1502
1503         for_each_td(td, i) {
1504                 print_status_init(td->thread_number - 1);
1505
1506                 if (!td->o.create_serialize)
1507                         continue;
1508
1509                 /*
1510                  * do file setup here so it happens sequentially,
1511                  * we don't want X number of threads getting their
1512                  * client data interspersed on disk
1513                  */
1514                 if (setup_files(td)) {
1515                         exit_value++;
1516                         if (td->error)
1517                                 log_err("fio: pid=%d, err=%d/%s\n",
1518                                         (int) td->pid, td->error, td->verror);
1519                         td_set_runstate(td, TD_REAPED);
1520                         todo--;
1521                 } else {
1522                         struct fio_file *f;
1523                         unsigned int j;
1524
1525                         /*
1526                          * for sharing to work, each job must always open
1527                          * its own files. so close them, if we opened them
1528                          * for creation
1529                          */
1530                         for_each_file(td, f, j) {
1531                                 if (fio_file_open(f))
1532                                         td_io_close_file(td, f);
1533                         }
1534                 }
1535         }
1536
1537         set_genesis_time();
1538
1539         while (todo) {
1540                 struct thread_data *map[REAL_MAX_JOBS];
1541                 struct timeval this_start;
1542                 int this_jobs = 0, left;
1543
1544                 /*
1545                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1546                  */
1547                 for_each_td(td, i) {
1548                         if (td->runstate != TD_NOT_CREATED)
1549                                 continue;
1550
1551                         /*
1552                          * never got a chance to start, killed by other
1553                          * thread for some reason
1554                          */
1555                         if (td->terminate) {
1556                                 todo--;
1557                                 continue;
1558                         }
1559
1560                         if (td->o.start_delay) {
1561                                 spent = mtime_since_genesis();
1562
1563                                 if (td->o.start_delay * 1000 > spent)
1564                                         continue;
1565                         }
1566
1567                         if (td->o.stonewall && (nr_started || nr_running)) {
1568                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1569                                                         td->o.name);
1570                                 break;
1571                         }
1572
1573                         init_disk_util(td);
1574
1575                         /*
1576                          * Set state to created. Thread will transition
1577                          * to TD_INITIALIZED when it's done setting up.
1578                          */
1579                         td_set_runstate(td, TD_CREATED);
1580                         map[this_jobs++] = td;
1581                         nr_started++;
1582
1583                         if (td->o.use_thread) {
1584                                 int ret;
1585
1586                                 dprint(FD_PROCESS, "will pthread_create\n");
1587                                 ret = pthread_create(&td->thread, NULL,
1588                                                         thread_main, td);
1589                                 if (ret) {
1590                                         log_err("pthread_create: %s\n",
1591                                                         strerror(ret));
1592                                         nr_started--;
1593                                         break;
1594                                 }
1595                                 ret = pthread_detach(td->thread);
1596                                 if (ret)
1597                                         log_err("pthread_detach: %s",
1598                                                         strerror(ret));
1599                         } else {
1600                                 pid_t pid;
1601                                 dprint(FD_PROCESS, "will fork\n");
1602                                 pid = fork();
1603                                 if (!pid) {
1604                                         int ret = fork_main(shm_id, i);
1605
1606                                         _exit(ret);
1607                                 } else if (i == fio_debug_jobno)
1608                                         *fio_debug_jobp = pid;
1609                         }
1610                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1611                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1612                                 log_err("fio: job startup hung? exiting.\n");
1613                                 fio_terminate_threads(TERMINATE_ALL);
1614                                 fio_abort = 1;
1615                                 nr_started--;
1616                                 break;
1617                         }
1618                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1619                 }
1620
1621                 /*
1622                  * Wait for the started threads to transition to
1623                  * TD_INITIALIZED.
1624                  */
1625                 fio_gettime(&this_start, NULL);
1626                 left = this_jobs;
1627                 while (left && !fio_abort) {
1628                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1629                                 break;
1630
1631                         usleep(100000);
1632
1633                         for (i = 0; i < this_jobs; i++) {
1634                                 td = map[i];
1635                                 if (!td)
1636                                         continue;
1637                                 if (td->runstate == TD_INITIALIZED) {
1638                                         map[i] = NULL;
1639                                         left--;
1640                                 } else if (td->runstate >= TD_EXITED) {
1641                                         map[i] = NULL;
1642                                         left--;
1643                                         todo--;
1644                                         nr_running++; /* work-around... */
1645                                 }
1646                         }
1647                 }
1648
1649                 if (left) {
1650                         log_err("fio: %d job%s failed to start\n", left,
1651                                         left > 1 ? "s" : "");
1652                         for (i = 0; i < this_jobs; i++) {
1653                                 td = map[i];
1654                                 if (!td)
1655                                         continue;
1656                                 kill(td->pid, SIGTERM);
1657                         }
1658                         break;
1659                 }
1660
1661                 /*
1662                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1663                  */
1664                 for_each_td(td, i) {
1665                         if (td->runstate != TD_INITIALIZED)
1666                                 continue;
1667
1668                         if (in_ramp_time(td))
1669                                 td_set_runstate(td, TD_RAMP);
1670                         else
1671                                 td_set_runstate(td, TD_RUNNING);
1672                         nr_running++;
1673                         nr_started--;
1674                         m_rate += ddir_rw_sum(td->o.ratemin);
1675                         t_rate += ddir_rw_sum(td->o.rate);
1676                         todo--;
1677                         fio_mutex_up(td->mutex);
1678                 }
1679
1680                 reap_threads(&nr_running, &t_rate, &m_rate);
1681
1682                 if (todo) {
1683                         if (is_backend)
1684                                 fio_server_idle_loop();
1685                         else
1686                                 usleep(100000);
1687                 }
1688         }
1689
1690         while (nr_running) {
1691                 reap_threads(&nr_running, &t_rate, &m_rate);
1692
1693                 if (is_backend)
1694                         fio_server_idle_loop();
1695                 else
1696                         usleep(10000);
1697         }
1698
1699         update_io_ticks();
1700         fio_unpin_memory();
1701 }
1702
1703 void wait_for_disk_thread_exit(void)
1704 {
1705         fio_mutex_down(disk_thread_mutex);
1706 }
1707
1708 static void free_disk_util(void)
1709 {
1710         disk_util_start_exit();
1711         wait_for_disk_thread_exit();
1712         disk_util_prune_entries();
1713 }
1714
1715 static void *disk_thread_main(void *data)
1716 {
1717         int ret = 0;
1718
1719         fio_mutex_up(startup_mutex);
1720
1721         while (threads && !ret) {
1722                 usleep(DISK_UTIL_MSEC * 1000);
1723                 if (!threads)
1724                         break;
1725                 ret = update_io_ticks();
1726
1727                 if (!is_backend)
1728                         print_thread_status();
1729         }
1730
1731         fio_mutex_up(disk_thread_mutex);
1732         return NULL;
1733 }
1734
1735 static int create_disk_util_thread(void)
1736 {
1737         int ret;
1738
1739         setup_disk_util();
1740
1741         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1742
1743         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1744         if (ret) {
1745                 fio_mutex_remove(disk_thread_mutex);
1746                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1747                 return 1;
1748         }
1749
1750         ret = pthread_detach(disk_util_thread);
1751         if (ret) {
1752                 fio_mutex_remove(disk_thread_mutex);
1753                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1754                 return 1;
1755         }
1756
1757         dprint(FD_MUTEX, "wait on startup_mutex\n");
1758         fio_mutex_down(startup_mutex);
1759         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1760         return 0;
1761 }
1762
1763 int fio_backend(void)
1764 {
1765         struct thread_data *td;
1766         int i;
1767
1768         if (exec_profile) {
1769                 if (load_profile(exec_profile))
1770                         return 1;
1771                 free(exec_profile);
1772                 exec_profile = NULL;
1773         }
1774         if (!thread_number)
1775                 return 0;
1776
1777         if (write_bw_log) {
1778                 setup_log(&agg_io_log[DDIR_READ], 0);
1779                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1780                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1781         }
1782
1783         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1784         if (startup_mutex == NULL)
1785                 return 1;
1786         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1787         if (writeout_mutex == NULL)
1788                 return 1;
1789
1790         set_genesis_time();
1791         create_disk_util_thread();
1792
1793         cgroup_list = smalloc(sizeof(*cgroup_list));
1794         INIT_FLIST_HEAD(cgroup_list);
1795
1796         run_threads();
1797
1798         if (!fio_abort) {
1799                 show_run_stats();
1800                 if (write_bw_log) {
1801                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1802                         __finish_log(agg_io_log[DDIR_WRITE],
1803                                         "agg-write_bw.log");
1804                         __finish_log(agg_io_log[DDIR_TRIM],
1805                                         "agg-write_bw.log");
1806                 }
1807         }
1808
1809         for_each_td(td, i)
1810                 fio_options_free(td);
1811
1812         free_disk_util();
1813         cgroup_kill(cgroup_list);
1814         sfree(cgroup_list);
1815         sfree(cgroup_mnt);
1816
1817         fio_mutex_remove(startup_mutex);
1818         fio_mutex_remove(writeout_mutex);
1819         fio_mutex_remove(disk_thread_mutex);
1820         return exit_value;
1821 }