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
  65 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  66
  67 int groupid = 0;
  68 unsigned int thread_number = 0;
  69 unsigned int stat_number = 0;
  70 unsigned int nr_process = 0;
  71 unsigned int nr_thread = 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         struct flist_head *entry, *n;
 242         struct io_u *io_u;
 243         int r;
 244
 245         /*
 246          * get immediately available events, if any
 247          */
 248         r = io_u_queued_complete(td, 0, NULL);
 249         if (r < 0)
 250                 return;
 251
 252         /*
 253          * now cancel remaining active events
 254          */
 255         if (td->io_ops->cancel) {
 256                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 257                         io_u = flist_entry(entry, struct io_u, list);
 258
 259                         /*
 260                          * if the io_u isn't in flight, then that generally
 261                          * means someone leaked an io_u. complain but fix
 262                          * it up, so we don't stall here.
 263                          */
 264                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 265                                 log_err("fio: non-busy IO on busy list\n");
 266                                 put_io_u(td, io_u);
 267                         } else {
 268                                 r = td->io_ops->cancel(td, io_u);
 269                                 if (!r)
 270                                         put_io_u(td, io_u);
 271                         }
 272                 }
 273         }
 274
 275         if (td->cur_depth)
 276                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 277 }
 278
 279 /*
 280  * Helper to handle the final sync of a file. Works just like the normal
 281  * io path, just does everything sync.
 282  */
 283 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 284 {
 285         struct io_u *io_u = __get_io_u(td);
 286         int ret;
 287
 288         if (!io_u)
 289                 return 1;
 290
 291         io_u->ddir = DDIR_SYNC;
 292         io_u->file = f;
 293
 294         if (td_io_prep(td, io_u)) {
 295                 put_io_u(td, io_u);
 296                 return 1;
 297         }
 298
 299 requeue:
 300         ret = td_io_queue(td, io_u);
 301         if (ret < 0) {
 302                 td_verror(td, io_u->error, "td_io_queue");
 303                 put_io_u(td, io_u);
 304                 return 1;
 305         } else if (ret == FIO_Q_QUEUED) {
 306                 if (io_u_queued_complete(td, 1, NULL) < 0)
 307                         return 1;
 308         } else if (ret == FIO_Q_COMPLETED) {
 309                 if (io_u->error) {
 310                         td_verror(td, io_u->error, "td_io_queue");
 311                         return 1;
 312                 }
 313
 314                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 315                         return 1;
 316         } else if (ret == FIO_Q_BUSY) {
 317                 if (td_io_commit(td))
 318                         return 1;
 319                 goto requeue;
 320         }
 321
 322         return 0;
 323 }
 324
 325 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 326 {
 327         int ret;
 328
 329         if (fio_file_open(f))
 330                 return fio_io_sync(td, f);
 331
 332         if (td_io_open_file(td, f))
 333                 return 1;
 334
 335         ret = fio_io_sync(td, f);
 336         td_io_close_file(td, f);
 337         return ret;
 338 }
 339
 340 static inline void __update_tv_cache(struct thread_data *td)
 341 {
 342         fio_gettime(&td->tv_cache, NULL);
 343 }
 344
 345 static inline void update_tv_cache(struct thread_data *td)
 346 {
 347         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 348                 __update_tv_cache(td);
 349 }
 350
 351 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 352 {
 353         if (in_ramp_time(td))
 354                 return 0;
 355         if (!td->o.timeout)
 356                 return 0;
 357         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 358                 return 1;
 359
 360         return 0;
 361 }
 362
 363 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 364                                int *retptr)
 365 {
 366         int ret = *retptr;
 367
 368         if (ret < 0 || td->error) {
 369                 int err = td->error;
 370                 enum error_type_bit eb;
 371
 372                 if (ret < 0)
 373                         err = -ret;
 374
 375                 eb = td_error_type(ddir, err);
 376                 if (!(td->o.continue_on_error & (1 << eb)))
 377                         return 1;
 378
 379                 if (td_non_fatal_error(td, eb, err)) {
 380                         /*
 381                          * Continue with the I/Os in case of
 382                          * a non fatal error.
 383                          */
 384                         update_error_count(td, err);
 385                         td_clear_error(td);
 386                         *retptr = 0;
 387                         return 0;
 388                 } else if (td->o.fill_device && err == ENOSPC) {
 389                         /*
 390                          * We expect to hit this error if
 391                          * fill_device option is set.
 392                          */
 393                         td_clear_error(td);
 394                         td->terminate = 1;
 395                         return 1;
 396                 } else {
 397                         /*
 398                          * Stop the I/O in case of a fatal
 399                          * error.
 400                          */
 401                         update_error_count(td, err);
 402                         return 1;
 403                 }
 404         }
 405
 406         return 0;
 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         if (td->error)
 437                 return;
 438
 439         td_set_runstate(td, TD_VERIFYING);
 440
 441         io_u = NULL;
 442         while (!td->terminate) {
 443                 enum fio_ddir ddir;
 444                 int ret2, full;
 445
 446                 update_tv_cache(td);
 447
 448                 if (runtime_exceeded(td, &td->tv_cache)) {
 449                         __update_tv_cache(td);
 450                         if (runtime_exceeded(td, &td->tv_cache)) {
 451                                 td->terminate = 1;
 452                                 break;
 453                         }
 454                 }
 455
 456                 if (flow_threshold_exceeded(td))
 457                         continue;
 458
 459                 if (!td->o.experimental_verify) {
 460                         io_u = __get_io_u(td);
 461                         if (!io_u)
 462                                 break;
 463
 464                         if (get_next_verify(td, io_u)) {
 465                                 put_io_u(td, io_u);
 466                                 break;
 467                         }
 468
 469                         if (td_io_prep(td, io_u)) {
 470                                 put_io_u(td, io_u);
 471                                 break;
 472                         }
 473                 } else {
 474                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 475                                 break;
 476
 477                         while ((io_u = get_io_u(td)) != NULL) {
 478                                 /*
 479                                  * We are only interested in the places where
 480                                  * we wrote or trimmed IOs. Turn those into
 481                                  * reads for verification purposes.
 482                                  */
 483                                 if (io_u->ddir == DDIR_READ) {
 484                                         /*
 485                                          * Pretend we issued it for rwmix
 486                                          * accounting
 487                                          */
 488                                         td->io_issues[DDIR_READ]++;
 489                                         put_io_u(td, io_u);
 490                                         continue;
 491                                 } else if (io_u->ddir == DDIR_TRIM) {
 492                                         io_u->ddir = DDIR_READ;
 493                                         io_u->flags |= IO_U_F_TRIMMED;
 494                                         break;
 495                                 } else if (io_u->ddir == DDIR_WRITE) {
 496                                         io_u->ddir = DDIR_READ;
 497                                         break;
 498                                 } else {
 499                                         put_io_u(td, io_u);
 500                                         continue;
 501                                 }
 502                         }
 503
 504                         if (!io_u)
 505                                 break;
 506                 }
 507
 508                 if (td->o.verify_async)
 509                         io_u->end_io = verify_io_u_async;
 510                 else
 511                         io_u->end_io = verify_io_u;
 512
 513                 ddir = io_u->ddir;
 514
 515                 ret = td_io_queue(td, io_u);
 516                 switch (ret) {
 517                 case FIO_Q_COMPLETED:
 518                         if (io_u->error) {
 519                                 ret = -io_u->error;
 520                                 clear_io_u(td, io_u);
 521                         } else if (io_u->resid) {
 522                                 int bytes = io_u->xfer_buflen - io_u->resid;
 523
 524                                 /*
 525                                  * zero read, fail
 526                                  */
 527                                 if (!bytes) {
 528                                         td_verror(td, EIO, "full resid");
 529                                         put_io_u(td, io_u);
 530                                         break;
 531                                 }
 532
 533                                 io_u->xfer_buflen = io_u->resid;
 534                                 io_u->xfer_buf += bytes;
 535                                 io_u->offset += bytes;
 536
 537                                 if (ddir_rw(io_u->ddir))
 538                                         td->ts.short_io_u[io_u->ddir]++;
 539
 540                                 f = io_u->file;
 541                                 if (io_u->offset == f->real_file_size)
 542                                         goto sync_done;
 543
 544                                 requeue_io_u(td, &io_u);
 545                         } else {
 546 sync_done:
 547                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 548                                 if (ret < 0)
 549                                         break;
 550                         }
 551                         continue;
 552                 case FIO_Q_QUEUED:
 553                         break;
 554                 case FIO_Q_BUSY:
 555                         requeue_io_u(td, &io_u);
 556                         ret2 = td_io_commit(td);
 557                         if (ret2 < 0)
 558                                 ret = ret2;
 559                         break;
 560                 default:
 561                         assert(ret < 0);
 562                         td_verror(td, -ret, "td_io_queue");
 563                         break;
 564                 }
 565
 566                 if (break_on_this_error(td, ddir, &ret))
 567                         break;
 568
 569                 /*
 570                  * if we can queue more, do so. but check if there are
 571                  * completed io_u's first. Note that we can get BUSY even
 572                  * without IO queued, if the system is resource starved.
 573                  */
 574                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 575                 if (full || !td->o.iodepth_batch_complete) {
 576                         min_events = min(td->o.iodepth_batch_complete,
 577                                          td->cur_depth);
 578                         /*
 579                          * if the queue is full, we MUST reap at least 1 event
 580                          */
 581                         if (full && !min_events)
 582                                 min_events = 1;
 583
 584                         do {
 585                                 /*
 586                                  * Reap required number of io units, if any,
 587                                  * and do the verification on them through
 588                                  * the callback handler
 589                                  */
 590                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 591                                         ret = -1;
 592                                         break;
 593                                 }
 594                         } while (full && (td->cur_depth > td->o.iodepth_low));
 595                 }
 596                 if (ret < 0)
 597                         break;
 598         }
 599
 600         if (!td->error) {
 601                 min_events = td->cur_depth;
 602
 603                 if (min_events)
 604                         ret = io_u_queued_complete(td, min_events, NULL);
 605         } else
 606                 cleanup_pending_aio(td);
 607
 608         td_set_runstate(td, TD_RUNNING);
 609
 610         dprint(FD_VERIFY, "exiting loop\n");
 611 }
 612
 613 static int io_bytes_exceeded(struct thread_data *td)
 614 {
 615         unsigned long long bytes;
 616
 617         if (td_rw(td))
 618                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 619         else if (td_write(td))
 620                 bytes = td->this_io_bytes[DDIR_WRITE];
 621         else if (td_read(td))
 622                 bytes = td->this_io_bytes[DDIR_READ];
 623         else
 624                 bytes = td->this_io_bytes[DDIR_TRIM];
 625
 626         return bytes >= td->o.size;
 627 }
 628
 629 /*
 630  * Main IO worker function. It retrieves io_u's to process and queues
 631  * and reaps them, checking for rate and errors along the way.
 632  *
 633  * Returns number of bytes written and trimmed.
 634  */
 635 static uint64_t do_io(struct thread_data *td)
 636 {
 637         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 638         unsigned int i;
 639         int ret = 0;
 640
 641         if (in_ramp_time(td))
 642                 td_set_runstate(td, TD_RAMP);
 643         else
 644                 td_set_runstate(td, TD_RUNNING);
 645
 646         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 647                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 648                 td->o.time_based) {
 649                 struct timeval comp_time;
 650                 int min_evts = 0;
 651                 struct io_u *io_u;
 652                 int ret2, full;
 653                 enum fio_ddir ddir;
 654
 655                 if (td->terminate || td->done)
 656                         break;
 657
 658                 update_tv_cache(td);
 659
 660                 if (runtime_exceeded(td, &td->tv_cache)) {
 661                         __update_tv_cache(td);
 662                         if (runtime_exceeded(td, &td->tv_cache)) {
 663                                 td->terminate = 1;
 664                                 break;
 665                         }
 666                 }
 667
 668                 if (flow_threshold_exceeded(td))
 669                         continue;
 670
 671                 io_u = get_io_u(td);
 672                 if (!io_u)
 673                         break;
 674
 675                 ddir = io_u->ddir;
 676
 677                 /*
 678                  * Add verification end_io handler if:
 679                  *      - Asked to verify (!td_rw(td))
 680                  *      - Or the io_u is from our verify list (mixed write/ver)
 681                  */
 682                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 683                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 684                         if (td->o.verify_async)
 685                                 io_u->end_io = verify_io_u_async;
 686                         else
 687                                 io_u->end_io = verify_io_u;
 688                         td_set_runstate(td, TD_VERIFYING);
 689                 } else if (in_ramp_time(td))
 690                         td_set_runstate(td, TD_RAMP);
 691                 else
 692                         td_set_runstate(td, TD_RUNNING);
 693
 694                 ret = td_io_queue(td, io_u);
 695                 switch (ret) {
 696                 case FIO_Q_COMPLETED:
 697                         if (io_u->error) {
 698                                 ret = -io_u->error;
 699                                 clear_io_u(td, io_u);
 700                         } else if (io_u->resid) {
 701                                 int bytes = io_u->xfer_buflen - io_u->resid;
 702                                 struct fio_file *f = io_u->file;
 703
 704                                 /*
 705                                  * zero read, fail
 706                                  */
 707                                 if (!bytes) {
 708                                         td_verror(td, EIO, "full resid");
 709                                         put_io_u(td, io_u);
 710                                         break;
 711                                 }
 712
 713                                 io_u->xfer_buflen = io_u->resid;
 714                                 io_u->xfer_buf += bytes;
 715                                 io_u->offset += bytes;
 716
 717                                 if (ddir_rw(io_u->ddir))
 718                                         td->ts.short_io_u[io_u->ddir]++;
 719
 720                                 if (io_u->offset == f->real_file_size)
 721                                         goto sync_done;
 722
 723                                 requeue_io_u(td, &io_u);
 724                         } else {
 725 sync_done:
 726                                 if (__should_check_rate(td, DDIR_READ) ||
 727                                     __should_check_rate(td, DDIR_WRITE) ||
 728                                     __should_check_rate(td, DDIR_TRIM))
 729                                         fio_gettime(&comp_time, NULL);
 730
 731                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 732                                 if (ret < 0)
 733                                         break;
 734                         }
 735                         break;
 736                 case FIO_Q_QUEUED:
 737                         /*
 738                          * if the engine doesn't have a commit hook,
 739                          * the io_u is really queued. if it does have such
 740                          * a hook, it has to call io_u_queued() itself.
 741                          */
 742                         if (td->io_ops->commit == NULL)
 743                                 io_u_queued(td, io_u);
 744                         break;
 745                 case FIO_Q_BUSY:
 746                         requeue_io_u(td, &io_u);
 747                         ret2 = td_io_commit(td);
 748                         if (ret2 < 0)
 749                                 ret = ret2;
 750                         break;
 751                 default:
 752                         assert(ret < 0);
 753                         put_io_u(td, io_u);
 754                         break;
 755                 }
 756
 757                 if (break_on_this_error(td, ddir, &ret))
 758                         break;
 759
 760                 /*
 761                  * See if we need to complete some commands. Note that we
 762                  * can get BUSY even without IO queued, if the system is
 763                  * resource starved.
 764                  */
 765                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 766                 if (full || !td->o.iodepth_batch_complete) {
 767                         min_evts = min(td->o.iodepth_batch_complete,
 768                                         td->cur_depth);
 769                         /*
 770                          * if the queue is full, we MUST reap at least 1 event
 771                          */
 772                         if (full && !min_evts)
 773                                 min_evts = 1;
 774
 775                         if (__should_check_rate(td, DDIR_READ) ||
 776                             __should_check_rate(td, DDIR_WRITE) ||
 777                             __should_check_rate(td, DDIR_TRIM))
 778                                 fio_gettime(&comp_time, NULL);
 779
 780                         do {
 781                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 782                                 if (ret < 0)
 783                                         break;
 784
 785                         } while (full && (td->cur_depth > td->o.iodepth_low));
 786                 }
 787
 788                 if (ret < 0)
 789                         break;
 790                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 791                         continue;
 792
 793                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 794                         if (check_min_rate(td, &comp_time, bytes_done)) {
 795                                 if (exitall_on_terminate)
 796                                         fio_terminate_threads(td->groupid);
 797                                 td_verror(td, EIO, "check_min_rate");
 798                                 break;
 799                         }
 800                 }
 801
 802                 if (td->o.thinktime) {
 803                         unsigned long long b;
 804
 805                         b = ddir_rw_sum(td->io_blocks);
 806                         if (!(b % td->o.thinktime_blocks)) {
 807                                 int left;
 808
 809                                 if (td->o.thinktime_spin)
 810                                         usec_spin(td->o.thinktime_spin);
 811
 812                                 left = td->o.thinktime - td->o.thinktime_spin;
 813                                 if (left)
 814                                         usec_sleep(td, left);
 815                         }
 816                 }
 817         }
 818
 819         if (td->trim_entries)
 820                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 821
 822         if (td->o.fill_device && td->error == ENOSPC) {
 823                 td->error = 0;
 824                 td->terminate = 1;
 825         }
 826         if (!td->error) {
 827                 struct fio_file *f;
 828
 829                 i = td->cur_depth;
 830                 if (i) {
 831                         ret = io_u_queued_complete(td, i, bytes_done);
 832                         if (td->o.fill_device && td->error == ENOSPC)
 833                                 td->error = 0;
 834                 }
 835
 836                 if (should_fsync(td) && td->o.end_fsync) {
 837                         td_set_runstate(td, TD_FSYNCING);
 838
 839                         for_each_file(td, f, i) {
 840                                 if (!fio_file_fsync(td, f))
 841                                         continue;
 842
 843                                 log_err("fio: end_fsync failed for file %s\n",
 844                                                                 f->file_name);
 845                         }
 846                 }
 847         } else
 848                 cleanup_pending_aio(td);
 849
 850         /*
 851          * stop job if we failed doing any IO
 852          */
 853         if (!ddir_rw_sum(td->this_io_bytes))
 854                 td->done = 1;
 855
 856         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 857 }
 858
 859 static void cleanup_io_u(struct thread_data *td)
 860 {
 861         struct flist_head *entry, *n;
 862         struct io_u *io_u;
 863
 864         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 865                 io_u = flist_entry(entry, struct io_u, list);
 866
 867                 flist_del(&io_u->list);
 868
 869                 if (td->io_ops->io_u_free)
 870                         td->io_ops->io_u_free(td, io_u);
 871
 872                 fio_memfree(io_u, sizeof(*io_u));
 873         }
 874
 875         free_io_mem(td);
 876 }
 877
 878 static int init_io_u(struct thread_data *td)
 879 {
 880         struct io_u *io_u;
 881         unsigned int max_bs, min_write;
 882         int cl_align, i, max_units;
 883         int data_xfer = 1;
 884         char *p;
 885
 886         max_units = td->o.iodepth;
 887         max_bs = td_max_bs(td);
 888         min_write = td->o.min_bs[DDIR_WRITE];
 889         td->orig_buffer_size = (unsigned long long) max_bs
 890                                         * (unsigned long long) max_units;
 891
 892         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 893                 data_xfer = 0;
 894
 895         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 896                 unsigned long bs;
 897
 898                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 899                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 900         }
 901
 902         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 903                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 904                 return 1;
 905         }
 906
 907         if (data_xfer && allocate_io_mem(td))
 908                 return 1;
 909
 910         if (td->o.odirect || td->o.mem_align ||
 911             (td->io_ops->flags & FIO_RAWIO))
 912                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 913         else
 914                 p = td->orig_buffer;
 915
 916         cl_align = os_cache_line_size();
 917
 918         for (i = 0; i < max_units; i++) {
 919                 void *ptr;
 920
 921                 if (td->terminate)
 922                         return 1;
 923
 924                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 925                 if (!ptr) {
 926                         log_err("fio: unable to allocate aligned memory\n");
 927                         break;
 928                 }
 929
 930                 io_u = ptr;
 931                 memset(io_u, 0, sizeof(*io_u));
 932                 INIT_FLIST_HEAD(&io_u->list);
 933                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 934
 935                 if (data_xfer) {
 936                         io_u->buf = p;
 937                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 938
 939                         if (td_write(td))
 940                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 941                         if (td_write(td) && td->o.verify_pattern_bytes) {
 942                                 /*
 943                                  * Fill the buffer with the pattern if we are
 944                                  * going to be doing writes.
 945                                  */
 946                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 947                         }
 948                 }
 949
 950                 io_u->index = i;
 951                 io_u->flags = IO_U_F_FREE;
 952                 flist_add(&io_u->list, &td->io_u_freelist);
 953
 954                 if (td->io_ops->io_u_init) {
 955                         int ret = td->io_ops->io_u_init(td, io_u);
 956
 957                         if (ret) {
 958                                 log_err("fio: failed to init engine data: %d\n", ret);
 959                                 return 1;
 960                         }
 961                 }
 962
 963                 p += max_bs;
 964         }
 965
 966         return 0;
 967 }
 968
 969 static int switch_ioscheduler(struct thread_data *td)
 970 {
 971         char tmp[256], tmp2[128];
 972         FILE *f;
 973         int ret;
 974
 975         if (td->io_ops->flags & FIO_DISKLESSIO)
 976                 return 0;
 977
 978         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 979
 980         f = fopen(tmp, "r+");
 981         if (!f) {
 982                 if (errno == ENOENT) {
 983                         log_err("fio: os or kernel doesn't support IO scheduler"
 984                                 " switching\n");
 985                         return 0;
 986                 }
 987                 td_verror(td, errno, "fopen iosched");
 988                 return 1;
 989         }
 990
 991         /*
 992          * Set io scheduler.
 993          */
 994         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 995         if (ferror(f) || ret != 1) {
 996                 td_verror(td, errno, "fwrite");
 997                 fclose(f);
 998                 return 1;
 999         }
1000
1001         rewind(f);
1002
1003         /*
1004          * Read back and check that the selected scheduler is now the default.
1005          */
1006         ret = fread(tmp, 1, sizeof(tmp), f);
1007         if (ferror(f) || ret < 0) {
1008                 td_verror(td, errno, "fread");
1009                 fclose(f);
1010                 return 1;
1011         }
1012
1013         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1014         if (!strstr(tmp, tmp2)) {
1015                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1016                 td_verror(td, EINVAL, "iosched_switch");
1017                 fclose(f);
1018                 return 1;
1019         }
1020
1021         fclose(f);
1022         return 0;
1023 }
1024
1025 static int keep_running(struct thread_data *td)
1026 {
1027         if (td->done)
1028                 return 0;
1029         if (td->o.time_based)
1030                 return 1;
1031         if (td->o.loops) {
1032                 td->o.loops--;
1033                 return 1;
1034         }
1035
1036         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1037                 uint64_t diff;
1038
1039                 /*
1040                  * If the difference is less than the minimum IO size, we
1041                  * are done.
1042                  */
1043                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1044                 if (diff < td_max_bs(td))
1045                         return 0;
1046
1047                 return 1;
1048         }
1049
1050         return 0;
1051 }
1052
1053 static int exec_string(const char *string)
1054 {
1055         int ret, newlen = strlen(string) + 1 + 8;
1056         char *str;
1057
1058         str = malloc(newlen);
1059         sprintf(str, "sh -c %s", string);
1060
1061         ret = system(str);
1062         if (ret == -1)
1063                 log_err("fio: exec of cmd <%s> failed\n", str);
1064
1065         free(str);
1066         return ret;
1067 }
1068
1069 /*
1070  * Entry point for the thread based jobs. The process based jobs end up
1071  * here as well, after a little setup.
1072  */
1073 static void *thread_main(void *data)
1074 {
1075         unsigned long long elapsed;
1076         struct thread_data *td = data;
1077         pthread_condattr_t attr;
1078         int clear_state;
1079
1080         if (!td->o.use_thread) {
1081                 setsid();
1082                 td->pid = getpid();
1083         } else
1084                 td->pid = gettid();
1085
1086         fio_local_clock_init(td->o.use_thread);
1087
1088         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1089
1090         INIT_FLIST_HEAD(&td->io_u_freelist);
1091         INIT_FLIST_HEAD(&td->io_u_busylist);
1092         INIT_FLIST_HEAD(&td->io_u_requeues);
1093         INIT_FLIST_HEAD(&td->io_log_list);
1094         INIT_FLIST_HEAD(&td->io_hist_list);
1095         INIT_FLIST_HEAD(&td->verify_list);
1096         INIT_FLIST_HEAD(&td->trim_list);
1097         INIT_FLIST_HEAD(&td->next_rand_list);
1098         pthread_mutex_init(&td->io_u_lock, NULL);
1099         td->io_hist_tree = RB_ROOT;
1100
1101         pthread_condattr_init(&attr);
1102         pthread_cond_init(&td->verify_cond, &attr);
1103         pthread_cond_init(&td->free_cond, &attr);
1104
1105         td_set_runstate(td, TD_INITIALIZED);
1106         dprint(FD_MUTEX, "up startup_mutex\n");
1107         fio_mutex_up(startup_mutex);
1108         dprint(FD_MUTEX, "wait on td->mutex\n");
1109         fio_mutex_down(td->mutex);
1110         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1111
1112         /*
1113          * the ->mutex mutex is now no longer used, close it to avoid
1114          * eating a file descriptor
1115          */
1116         fio_mutex_remove(td->mutex);
1117
1118         /*
1119          * A new gid requires privilege, so we need to do this before setting
1120          * the uid.
1121          */
1122         if (td->o.gid != -1U && setgid(td->o.gid)) {
1123                 td_verror(td, errno, "setgid");
1124                 goto err;
1125         }
1126         if (td->o.uid != -1U && setuid(td->o.uid)) {
1127                 td_verror(td, errno, "setuid");
1128                 goto err;
1129         }
1130
1131         /*
1132          * If we have a gettimeofday() thread, make sure we exclude that
1133          * thread from this job
1134          */
1135         if (td->o.gtod_cpu)
1136                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1137
1138         /*
1139          * Set affinity first, in case it has an impact on the memory
1140          * allocations.
1141          */
1142         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1143                 td_verror(td, errno, "cpu_set_affinity");
1144                 goto err;
1145         }
1146
1147 #ifdef CONFIG_LIBNUMA
1148         /* numa node setup */
1149         if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1150                 int ret;
1151
1152                 if (numa_available() < 0) {
1153                         td_verror(td, errno, "Does not support NUMA API\n");
1154                         goto err;
1155                 }
1156
1157                 if (td->o.numa_cpumask_set) {
1158                         ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1159                         if (ret == -1) {
1160                                 td_verror(td, errno, \
1161                                         "numa_run_on_node_mask failed\n");
1162                                 goto err;
1163                         }
1164                 }
1165
1166                 if (td->o.numa_memmask_set) {
1167
1168                         switch (td->o.numa_mem_mode) {
1169                         case MPOL_INTERLEAVE:
1170                                 numa_set_interleave_mask(td->o.numa_memnodesmask);
1171                                 break;
1172                         case MPOL_BIND:
1173                                 numa_set_membind(td->o.numa_memnodesmask);
1174                                 break;
1175                         case MPOL_LOCAL:
1176                                 numa_set_localalloc();
1177                                 break;
1178                         case MPOL_PREFERRED:
1179                                 numa_set_preferred(td->o.numa_mem_prefer_node);
1180                                 break;
1181                         case MPOL_DEFAULT:
1182                         default:
1183                                 break;
1184                         }
1185
1186                 }
1187         }
1188 #endif
1189
1190         /*
1191          * May alter parameters that init_io_u() will use, so we need to
1192          * do this first.
1193          */
1194         if (init_iolog(td))
1195                 goto err;
1196
1197         if (init_io_u(td))
1198                 goto err;
1199
1200         if (td->o.verify_async && verify_async_init(td))
1201                 goto err;
1202
1203         if (td->ioprio_set) {
1204                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1205                         td_verror(td, errno, "ioprio_set");
1206                         goto err;
1207                 }
1208         }
1209
1210         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1211                 goto err;
1212
1213         errno = 0;
1214         if (nice(td->o.nice) == -1 && errno != 0) {
1215                 td_verror(td, errno, "nice");
1216                 goto err;
1217         }
1218
1219         if (td->o.ioscheduler && switch_ioscheduler(td))
1220                 goto err;
1221
1222         if (!td->o.create_serialize && setup_files(td))
1223                 goto err;
1224
1225         if (td_io_init(td))
1226                 goto err;
1227
1228         if (init_random_map(td))
1229                 goto err;
1230
1231         if (td->o.exec_prerun) {
1232                 if (exec_string(td->o.exec_prerun))
1233                         goto err;
1234         }
1235
1236         if (td->o.pre_read) {
1237                 if (pre_read_files(td) < 0)
1238                         goto err;
1239         }
1240
1241         fio_gettime(&td->epoch, NULL);
1242         fio_getrusage(&td->ru_start);
1243         clear_state = 0;
1244         while (keep_running(td)) {
1245                 uint64_t verify_bytes;
1246
1247                 fio_gettime(&td->start, NULL);
1248                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1249                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1250                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1251
1252                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1253                                 td->o.ratemin[DDIR_TRIM]) {
1254                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1255                                                 sizeof(td->bw_sample_time));
1256                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1257                                                 sizeof(td->bw_sample_time));
1258                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1259                                                 sizeof(td->bw_sample_time));
1260                 }
1261
1262                 if (clear_state)
1263                         clear_io_state(td);
1264
1265                 prune_io_piece_log(td);
1266
1267                 verify_bytes = do_io(td);
1268
1269                 clear_state = 1;
1270
1271                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1272                         elapsed = utime_since_now(&td->start);
1273                         td->ts.runtime[DDIR_READ] += elapsed;
1274                 }
1275                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1276                         elapsed = utime_since_now(&td->start);
1277                         td->ts.runtime[DDIR_WRITE] += elapsed;
1278                 }
1279                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1280                         elapsed = utime_since_now(&td->start);
1281                         td->ts.runtime[DDIR_TRIM] += elapsed;
1282                 }
1283
1284                 if (td->error || td->terminate)
1285                         break;
1286
1287                 if (!td->o.do_verify ||
1288                     td->o.verify == VERIFY_NONE ||
1289                     (td->io_ops->flags & FIO_UNIDIR))
1290                         continue;
1291
1292                 clear_io_state(td);
1293
1294                 fio_gettime(&td->start, NULL);
1295
1296                 do_verify(td, verify_bytes);
1297
1298                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1299
1300                 if (td->error || td->terminate)
1301                         break;
1302         }
1303
1304         update_rusage_stat(td);
1305         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1306         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1307         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1308         td->ts.total_run_time = mtime_since_now(&td->epoch);
1309         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1310         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1311         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1312
1313         fio_mutex_down(writeout_mutex);
1314         if (td->bw_log) {
1315                 if (td->o.bw_log_file) {
1316                         finish_log_named(td, td->bw_log,
1317                                                 td->o.bw_log_file, "bw");
1318                 } else
1319                         finish_log(td, td->bw_log, "bw");
1320         }
1321         if (td->lat_log) {
1322                 if (td->o.lat_log_file) {
1323                         finish_log_named(td, td->lat_log,
1324                                                 td->o.lat_log_file, "lat");
1325                 } else
1326                         finish_log(td, td->lat_log, "lat");
1327         }
1328         if (td->slat_log) {
1329                 if (td->o.lat_log_file) {
1330                         finish_log_named(td, td->slat_log,
1331                                                 td->o.lat_log_file, "slat");
1332                 } else
1333                         finish_log(td, td->slat_log, "slat");
1334         }
1335         if (td->clat_log) {
1336                 if (td->o.lat_log_file) {
1337                         finish_log_named(td, td->clat_log,
1338                                                 td->o.lat_log_file, "clat");
1339                 } else
1340                         finish_log(td, td->clat_log, "clat");
1341         }
1342         if (td->iops_log) {
1343                 if (td->o.iops_log_file) {
1344                         finish_log_named(td, td->iops_log,
1345                                                 td->o.iops_log_file, "iops");
1346                 } else
1347                         finish_log(td, td->iops_log, "iops");
1348         }
1349
1350         fio_mutex_up(writeout_mutex);
1351         if (td->o.exec_postrun)
1352                 exec_string(td->o.exec_postrun);
1353
1354         if (exitall_on_terminate)
1355                 fio_terminate_threads(td->groupid);
1356
1357 err:
1358         if (td->error)
1359                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1360                                                         td->verror);
1361
1362         if (td->o.verify_async)
1363                 verify_async_exit(td);
1364
1365         close_and_free_files(td);
1366         cleanup_io_u(td);
1367         close_ioengine(td);
1368         cgroup_shutdown(td, &cgroup_mnt);
1369
1370         if (td->o.cpumask_set) {
1371                 int ret = fio_cpuset_exit(&td->o.cpumask);
1372
1373                 td_verror(td, ret, "fio_cpuset_exit");
1374         }
1375
1376         /*
1377          * do this very late, it will log file closing as well
1378          */
1379         if (td->o.write_iolog_file)
1380                 write_iolog_close(td);
1381
1382         td_set_runstate(td, TD_EXITED);
1383         return (void *) (uintptr_t) td->error;
1384 }
1385
1386
1387 /*
1388  * We cannot pass the td data into a forked process, so attach the td and
1389  * pass it to the thread worker.
1390  */
1391 static int fork_main(int shmid, int offset)
1392 {
1393         struct thread_data *td;
1394         void *data, *ret;
1395
1396 #ifndef __hpux
1397         data = shmat(shmid, NULL, 0);
1398         if (data == (void *) -1) {
1399                 int __err = errno;
1400
1401                 perror("shmat");
1402                 return __err;
1403         }
1404 #else
1405         /*
1406          * HP-UX inherits shm mappings?
1407          */
1408         data = threads;
1409 #endif
1410
1411         td = data + offset * sizeof(struct thread_data);
1412         ret = thread_main(td);
1413         shmdt(data);
1414         return (int) (uintptr_t) ret;
1415 }
1416
1417 /*
1418  * Run over the job map and reap the threads that have exited, if any.
1419  */
1420 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1421                          unsigned int *m_rate)
1422 {
1423         struct thread_data *td;
1424         unsigned int cputhreads, realthreads, pending;
1425         int i, status, ret;
1426
1427         /*
1428          * reap exited threads (TD_EXITED -> TD_REAPED)
1429          */
1430         realthreads = pending = cputhreads = 0;
1431         for_each_td(td, i) {
1432                 int flags = 0;
1433
1434                 /*
1435                  * ->io_ops is NULL for a thread that has closed its
1436                  * io engine
1437                  */
1438                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1439                         cputhreads++;
1440                 else
1441                         realthreads++;
1442
1443                 if (!td->pid) {
1444                         pending++;
1445                         continue;
1446                 }
1447                 if (td->runstate == TD_REAPED)
1448                         continue;
1449                 if (td->o.use_thread) {
1450                         if (td->runstate == TD_EXITED) {
1451                                 td_set_runstate(td, TD_REAPED);
1452                                 goto reaped;
1453                         }
1454                         continue;
1455                 }
1456
1457                 flags = WNOHANG;
1458                 if (td->runstate == TD_EXITED)
1459                         flags = 0;
1460
1461                 /*
1462                  * check if someone quit or got killed in an unusual way
1463                  */
1464                 ret = waitpid(td->pid, &status, flags);
1465                 if (ret < 0) {
1466                         if (errno == ECHILD) {
1467                                 log_err("fio: pid=%d disappeared %d\n",
1468                                                 (int) td->pid, td->runstate);
1469                                 td->sig = ECHILD;
1470                                 td_set_runstate(td, TD_REAPED);
1471                                 goto reaped;
1472                         }
1473                         perror("waitpid");
1474                 } else if (ret == td->pid) {
1475                         if (WIFSIGNALED(status)) {
1476                                 int sig = WTERMSIG(status);
1477
1478                                 if (sig != SIGTERM && sig != SIGUSR2)
1479                                         log_err("fio: pid=%d, got signal=%d\n",
1480                                                         (int) td->pid, sig);
1481                                 td->sig = sig;
1482                                 td_set_runstate(td, TD_REAPED);
1483                                 goto reaped;
1484                         }
1485                         if (WIFEXITED(status)) {
1486                                 if (WEXITSTATUS(status) && !td->error)
1487                                         td->error = WEXITSTATUS(status);
1488
1489                                 td_set_runstate(td, TD_REAPED);
1490                                 goto reaped;
1491                         }
1492                 }
1493
1494                 /*
1495                  * thread is not dead, continue
1496                  */
1497                 pending++;
1498                 continue;
1499 reaped:
1500                 (*nr_running)--;
1501                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1502                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1503                 if (!td->pid)
1504                         pending--;
1505
1506                 if (td->error)
1507                         exit_value++;
1508
1509                 done_secs += mtime_since_now(&td->epoch) / 1000;
1510         }
1511
1512         if (*nr_running == cputhreads && !pending && realthreads)
1513                 fio_terminate_threads(TERMINATE_ALL);
1514 }
1515
1516 /*
1517  * Main function for kicking off and reaping jobs, as needed.
1518  */
1519 static void run_threads(void)
1520 {
1521         struct thread_data *td;
1522         unsigned long spent;
1523         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1524
1525         if (fio_pin_memory())
1526                 return;
1527
1528         if (fio_gtod_offload && fio_start_gtod_thread())
1529                 return;
1530
1531         fio_idle_prof_init();
1532
1533         set_sig_handlers();
1534
1535         if (output_format == FIO_OUTPUT_NORMAL) {
1536                 log_info("Starting ");
1537                 if (nr_thread)
1538                         log_info("%d thread%s", nr_thread,
1539                                                 nr_thread > 1 ? "s" : "");
1540                 if (nr_process) {
1541                         if (nr_thread)
1542                                 log_info(" and ");
1543                         log_info("%d process%s", nr_process,
1544                                                 nr_process > 1 ? "es" : "");
1545                 }
1546                 log_info("\n");
1547                 fflush(stdout);
1548         }
1549
1550         todo = thread_number;
1551         nr_running = 0;
1552         nr_started = 0;
1553         m_rate = t_rate = 0;
1554
1555         for_each_td(td, i) {
1556                 print_status_init(td->thread_number - 1);
1557
1558                 if (!td->o.create_serialize)
1559                         continue;
1560
1561                 /*
1562                  * do file setup here so it happens sequentially,
1563                  * we don't want X number of threads getting their
1564                  * client data interspersed on disk
1565                  */
1566                 if (setup_files(td)) {
1567                         exit_value++;
1568                         if (td->error)
1569                                 log_err("fio: pid=%d, err=%d/%s\n",
1570                                         (int) td->pid, td->error, td->verror);
1571                         td_set_runstate(td, TD_REAPED);
1572                         todo--;
1573                 } else {
1574                         struct fio_file *f;
1575                         unsigned int j;
1576
1577                         /*
1578                          * for sharing to work, each job must always open
1579                          * its own files. so close them, if we opened them
1580                          * for creation
1581                          */
1582                         for_each_file(td, f, j) {
1583                                 if (fio_file_open(f))
1584                                         td_io_close_file(td, f);
1585                         }
1586                 }
1587         }
1588
1589         /* start idle threads before io threads start to run */
1590         fio_idle_prof_start();
1591
1592         set_genesis_time();
1593
1594         while (todo) {
1595                 struct thread_data *map[REAL_MAX_JOBS];
1596                 struct timeval this_start;
1597                 int this_jobs = 0, left;
1598
1599                 /*
1600                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1601                  */
1602                 for_each_td(td, i) {
1603                         if (td->runstate != TD_NOT_CREATED)
1604                                 continue;
1605
1606                         /*
1607                          * never got a chance to start, killed by other
1608                          * thread for some reason
1609                          */
1610                         if (td->terminate) {
1611                                 todo--;
1612                                 continue;
1613                         }
1614
1615                         if (td->o.start_delay) {
1616                                 spent = mtime_since_genesis();
1617
1618                                 if (td->o.start_delay * 1000 > spent)
1619                                         continue;
1620                         }
1621
1622                         if (td->o.stonewall && (nr_started || nr_running)) {
1623                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1624                                                         td->o.name);
1625                                 break;
1626                         }
1627
1628                         init_disk_util(td);
1629
1630                         /*
1631                          * Set state to created. Thread will transition
1632                          * to TD_INITIALIZED when it's done setting up.
1633                          */
1634                         td_set_runstate(td, TD_CREATED);
1635                         map[this_jobs++] = td;
1636                         nr_started++;
1637
1638                         if (td->o.use_thread) {
1639                                 int ret;
1640
1641                                 dprint(FD_PROCESS, "will pthread_create\n");
1642                                 ret = pthread_create(&td->thread, NULL,
1643                                                         thread_main, td);
1644                                 if (ret) {
1645                                         log_err("pthread_create: %s\n",
1646                                                         strerror(ret));
1647                                         nr_started--;
1648                                         break;
1649                                 }
1650                                 ret = pthread_detach(td->thread);
1651                                 if (ret)
1652                                         log_err("pthread_detach: %s",
1653                                                         strerror(ret));
1654                         } else {
1655                                 pid_t pid;
1656                                 dprint(FD_PROCESS, "will fork\n");
1657                                 pid = fork();
1658                                 if (!pid) {
1659                                         int ret = fork_main(shm_id, i);
1660
1661                                         _exit(ret);
1662                                 } else if (i == fio_debug_jobno)
1663                                         *fio_debug_jobp = pid;
1664                         }
1665                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1666                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1667                                 log_err("fio: job startup hung? exiting.\n");
1668                                 fio_terminate_threads(TERMINATE_ALL);
1669                                 fio_abort = 1;
1670                                 nr_started--;
1671                                 break;
1672                         }
1673                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1674                 }
1675
1676                 /*
1677                  * Wait for the started threads to transition to
1678                  * TD_INITIALIZED.
1679                  */
1680                 fio_gettime(&this_start, NULL);
1681                 left = this_jobs;
1682                 while (left && !fio_abort) {
1683                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1684                                 break;
1685
1686                         usleep(100000);
1687
1688                         for (i = 0; i < this_jobs; i++) {
1689                                 td = map[i];
1690                                 if (!td)
1691                                         continue;
1692                                 if (td->runstate == TD_INITIALIZED) {
1693                                         map[i] = NULL;
1694                                         left--;
1695                                 } else if (td->runstate >= TD_EXITED) {
1696                                         map[i] = NULL;
1697                                         left--;
1698                                         todo--;
1699                                         nr_running++; /* work-around... */
1700                                 }
1701                         }
1702                 }
1703
1704                 if (left) {
1705                         log_err("fio: %d job%s failed to start\n", left,
1706                                         left > 1 ? "s" : "");
1707                         for (i = 0; i < this_jobs; i++) {
1708                                 td = map[i];
1709                                 if (!td)
1710                                         continue;
1711                                 kill(td->pid, SIGTERM);
1712                         }
1713                         break;
1714                 }
1715
1716                 /*
1717                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1718                  */
1719                 for_each_td(td, i) {
1720                         if (td->runstate != TD_INITIALIZED)
1721                                 continue;
1722
1723                         if (in_ramp_time(td))
1724                                 td_set_runstate(td, TD_RAMP);
1725                         else
1726                                 td_set_runstate(td, TD_RUNNING);
1727                         nr_running++;
1728                         nr_started--;
1729                         m_rate += ddir_rw_sum(td->o.ratemin);
1730                         t_rate += ddir_rw_sum(td->o.rate);
1731                         todo--;
1732                         fio_mutex_up(td->mutex);
1733                 }
1734
1735                 reap_threads(&nr_running, &t_rate, &m_rate);
1736
1737                 if (todo) {
1738                         if (is_backend)
1739                                 fio_server_idle_loop();
1740                         else
1741                                 usleep(100000);
1742                 }
1743         }
1744
1745         while (nr_running) {
1746                 reap_threads(&nr_running, &t_rate, &m_rate);
1747
1748                 if (is_backend)
1749                         fio_server_idle_loop();
1750                 else
1751                         usleep(10000);
1752         }
1753
1754         fio_idle_prof_stop();
1755
1756         update_io_ticks();
1757         fio_unpin_memory();
1758 }
1759
1760 void wait_for_disk_thread_exit(void)
1761 {
1762         fio_mutex_down(disk_thread_mutex);
1763 }
1764
1765 static void free_disk_util(void)
1766 {
1767         disk_util_start_exit();
1768         wait_for_disk_thread_exit();
1769         disk_util_prune_entries();
1770 }
1771
1772 static void *disk_thread_main(void *data)
1773 {
1774         int ret = 0;
1775
1776         fio_mutex_up(startup_mutex);
1777
1778         while (threads && !ret) {
1779                 usleep(DISK_UTIL_MSEC * 1000);
1780                 if (!threads)
1781                         break;
1782                 ret = update_io_ticks();
1783
1784                 if (!is_backend)
1785                         print_thread_status();
1786         }
1787
1788         fio_mutex_up(disk_thread_mutex);
1789         return NULL;
1790 }
1791
1792 static int create_disk_util_thread(void)
1793 {
1794         int ret;
1795
1796         setup_disk_util();
1797
1798         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1799
1800         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1801         if (ret) {
1802                 fio_mutex_remove(disk_thread_mutex);
1803                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1804                 return 1;
1805         }
1806
1807         ret = pthread_detach(disk_util_thread);
1808         if (ret) {
1809                 fio_mutex_remove(disk_thread_mutex);
1810                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1811                 return 1;
1812         }
1813
1814         dprint(FD_MUTEX, "wait on startup_mutex\n");
1815         fio_mutex_down(startup_mutex);
1816         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1817         return 0;
1818 }
1819
1820 int fio_backend(void)
1821 {
1822         struct thread_data *td;
1823         int i;
1824
1825         if (exec_profile) {
1826                 if (load_profile(exec_profile))
1827                         return 1;
1828                 free(exec_profile);
1829                 exec_profile = NULL;
1830         }
1831         if (!thread_number)
1832                 return 0;
1833
1834         if (write_bw_log) {
1835                 setup_log(&agg_io_log[DDIR_READ], 0);
1836                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1837                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1838         }
1839
1840         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1841         if (startup_mutex == NULL)
1842                 return 1;
1843         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1844         if (writeout_mutex == NULL)
1845                 return 1;
1846
1847         set_genesis_time();
1848         create_disk_util_thread();
1849
1850         cgroup_list = smalloc(sizeof(*cgroup_list));
1851         INIT_FLIST_HEAD(cgroup_list);
1852
1853         run_threads();
1854
1855         if (!fio_abort) {
1856                 show_run_stats();
1857                 if (write_bw_log) {
1858                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1859                         __finish_log(agg_io_log[DDIR_WRITE],
1860                                         "agg-write_bw.log");
1861                         __finish_log(agg_io_log[DDIR_TRIM],
1862                                         "agg-write_bw.log");
1863                 }
1864         }
1865
1866         for_each_td(td, i)
1867                 fio_options_free(td);
1868
1869         free_disk_util();
1870         cgroup_kill(cgroup_list);
1871         sfree(cgroup_list);
1872         sfree(cgroup_mnt);
1873
1874         fio_mutex_remove(startup_mutex);
1875         fio_mutex_remove(writeout_mutex);
1876         fio_mutex_remove(disk_thread_mutex);
1877         return exit_value;
1878 }