backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38
  39 #include "fio.h"
  40 #ifndef FIO_NO_HAVE_SHM_H
  41 #include <sys/shm.h>
  42 #endif
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53 #include "lib/getrusage.h"
  54 #include "idletime.h"
  55
  56 static pthread_t disk_util_thread;
  57 static struct fio_mutex *disk_thread_mutex;
  58 static struct fio_mutex *startup_mutex;
  59 static struct fio_mutex *writeout_mutex;
  60 static struct flist_head *cgroup_list;
  61 static char *cgroup_mnt;
  62 static int exit_value;
  63 static volatile int fio_abort;
  64 static unsigned int nr_process = 0;
  65 static unsigned int nr_thread = 0;
  66
  67 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  68
  69 int groupid = 0;
  70 unsigned int thread_number = 0;
  71 unsigned int stat_number = 0;
  72 int shm_id = 0;
  73 int temp_stall_ts;
  74 unsigned long done_secs = 0;
  75 volatile int disk_util_exit = 0;
  76
  77 #define PAGE_ALIGN(buf) \
  78         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  79
  80 #define JOB_START_TIMEOUT       (5 * 1000)
  81
  82 static void sig_int(int sig)
  83 {
  84         if (threads) {
  85                 if (is_backend)
  86                         fio_server_got_signal(sig);
  87                 else {
  88                         log_info("\nfio: terminating on signal %d\n", sig);
  89                         fflush(stdout);
  90                         exit_value = 128;
  91                 }
  92
  93                 fio_terminate_threads(TERMINATE_ALL);
  94         }
  95 }
  96
  97 static void sig_show_status(int sig)
  98 {
  99         show_running_run_stats();
 100 }
 101
 102 static void set_sig_handlers(void)
 103 {
 104         struct sigaction act;
 105
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGINT, &act, NULL);
 110
 111         memset(&act, 0, sizeof(act));
 112         act.sa_handler = sig_int;
 113         act.sa_flags = SA_RESTART;
 114         sigaction(SIGTERM, &act, NULL);
 115
 116 /* Windows uses SIGBREAK as a quit signal from other applications */
 117 #ifdef WIN32
 118         memset(&act, 0, sizeof(act));
 119         act.sa_handler = sig_int;
 120         act.sa_flags = SA_RESTART;
 121         sigaction(SIGBREAK, &act, NULL);
 122 #endif
 123
 124         memset(&act, 0, sizeof(act));
 125         act.sa_handler = sig_show_status;
 126         act.sa_flags = SA_RESTART;
 127         sigaction(SIGUSR1, &act, NULL);
 128
 129         if (is_backend) {
 130                 memset(&act, 0, sizeof(act));
 131                 act.sa_handler = sig_int;
 132                 act.sa_flags = SA_RESTART;
 133                 sigaction(SIGPIPE, &act, NULL);
 134         }
 135 }
 136
 137 /*
 138  * Check if we are above the minimum rate given.
 139  */
 140 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 141                             enum fio_ddir ddir)
 142 {
 143         unsigned long long bytes = 0;
 144         unsigned long iops = 0;
 145         unsigned long spent;
 146         unsigned long rate;
 147         unsigned int ratemin = 0;
 148         unsigned int rate_iops = 0;
 149         unsigned int rate_iops_min = 0;
 150
 151         assert(ddir_rw(ddir));
 152
 153         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 154                 return 0;
 155
 156         /*
 157          * allow a 2 second settle period in the beginning
 158          */
 159         if (mtime_since(&td->start, now) < 2000)
 160                 return 0;
 161
 162         iops += td->this_io_blocks[ddir];
 163         bytes += td->this_io_bytes[ddir];
 164         ratemin += td->o.ratemin[ddir];
 165         rate_iops += td->o.rate_iops[ddir];
 166         rate_iops_min += td->o.rate_iops_min[ddir];
 167
 168         /*
 169          * if rate blocks is set, sample is running
 170          */
 171         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 172                 spent = mtime_since(&td->lastrate[ddir], now);
 173                 if (spent < td->o.ratecycle)
 174                         return 0;
 175
 176                 if (td->o.rate[ddir]) {
 177                         /*
 178                          * check bandwidth specified rate
 179                          */
 180                         if (bytes < td->rate_bytes[ddir]) {
 181                                 log_err("%s: min rate %u not met\n", td->o.name,
 182                                                                 ratemin);
 183                                 return 1;
 184                         } else {
 185                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 186                                 if (rate < ratemin ||
 187                                     bytes < td->rate_bytes[ddir]) {
 188                                         log_err("%s: min rate %u not met, got"
 189                                                 " %luKB/sec\n", td->o.name,
 190                                                         ratemin, rate);
 191                                         return 1;
 192                                 }
 193                         }
 194                 } else {
 195                         /*
 196                          * checks iops specified rate
 197                          */
 198                         if (iops < rate_iops) {
 199                                 log_err("%s: min iops rate %u not met\n",
 200                                                 td->o.name, rate_iops);
 201                                 return 1;
 202                         } else {
 203                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 204                                 if (rate < rate_iops_min ||
 205                                     iops < td->rate_blocks[ddir]) {
 206                                         log_err("%s: min iops rate %u not met,"
 207                                                 " got %lu\n", td->o.name,
 208                                                         rate_iops_min, rate);
 209                                 }
 210                         }
 211                 }
 212         }
 213
 214         td->rate_bytes[ddir] = bytes;
 215         td->rate_blocks[ddir] = iops;
 216         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 217         return 0;
 218 }
 219
 220 static int check_min_rate(struct thread_data *td, struct timeval *now,
 221                           uint64_t *bytes_done)
 222 {
 223         int ret = 0;
 224
 225         if (bytes_done[DDIR_READ])
 226                 ret |= __check_min_rate(td, now, DDIR_READ);
 227         if (bytes_done[DDIR_WRITE])
 228                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 229         if (bytes_done[DDIR_TRIM])
 230                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 231
 232         return ret;
 233 }
 234
 235 /*
 236  * When job exits, we can cancel the in-flight IO if we are using async
 237  * io. Attempt to do so.
 238  */
 239 static void cleanup_pending_aio(struct thread_data *td)
 240 {
 241         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 static void check_update_rusage(struct thread_data *td)
 410 {
 411         if (td->update_rusage) {
 412                 td->update_rusage = 0;
 413                 update_rusage_stat(td);
 414                 fio_mutex_up(td->rusage_sem);
 415         }
 416 }
 417
 418 /*
 419  * The main verify engine. Runs over the writes we previously submitted,
 420  * reads the blocks back in, and checks the crc/md5 of the data.
 421  */
 422 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 423 {
 424         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 425         struct fio_file *f;
 426         struct io_u *io_u;
 427         int ret, min_events;
 428         unsigned int i;
 429
 430         dprint(FD_VERIFY, "starting loop\n");
 431
 432         /*
 433          * sync io first and invalidate cache, to make sure we really
 434          * read from disk.
 435          */
 436         for_each_file(td, f, i) {
 437                 if (!fio_file_open(f))
 438                         continue;
 439                 if (fio_io_sync(td, f))
 440                         break;
 441                 if (file_invalidate_cache(td, f))
 442                         break;
 443         }
 444
 445         check_update_rusage(td);
 446
 447         if (td->error)
 448                 return;
 449
 450         td_set_runstate(td, TD_VERIFYING);
 451
 452         io_u = NULL;
 453         while (!td->terminate) {
 454                 enum fio_ddir ddir;
 455                 int ret2, full;
 456
 457                 update_tv_cache(td);
 458                 check_update_rusage(td);
 459
 460                 if (runtime_exceeded(td, &td->tv_cache)) {
 461                         __update_tv_cache(td);
 462                         if (runtime_exceeded(td, &td->tv_cache)) {
 463                                 td->terminate = 1;
 464                                 break;
 465                         }
 466                 }
 467
 468                 if (flow_threshold_exceeded(td))
 469                         continue;
 470
 471                 if (!td->o.experimental_verify) {
 472                         io_u = __get_io_u(td);
 473                         if (!io_u)
 474                                 break;
 475
 476                         if (get_next_verify(td, io_u)) {
 477                                 put_io_u(td, io_u);
 478                                 break;
 479                         }
 480
 481                         if (td_io_prep(td, io_u)) {
 482                                 put_io_u(td, io_u);
 483                                 break;
 484                         }
 485                 } else {
 486                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 487                                 break;
 488
 489                         while ((io_u = get_io_u(td)) != NULL) {
 490                                 /*
 491                                  * We are only interested in the places where
 492                                  * we wrote or trimmed IOs. Turn those into
 493                                  * reads for verification purposes.
 494                                  */
 495                                 if (io_u->ddir == DDIR_READ) {
 496                                         /*
 497                                          * Pretend we issued it for rwmix
 498                                          * accounting
 499                                          */
 500                                         td->io_issues[DDIR_READ]++;
 501                                         put_io_u(td, io_u);
 502                                         continue;
 503                                 } else if (io_u->ddir == DDIR_TRIM) {
 504                                         io_u->ddir = DDIR_READ;
 505                                         io_u->flags |= IO_U_F_TRIMMED;
 506                                         break;
 507                                 } else if (io_u->ddir == DDIR_WRITE) {
 508                                         io_u->ddir = DDIR_READ;
 509                                         break;
 510                                 } else {
 511                                         put_io_u(td, io_u);
 512                                         continue;
 513                                 }
 514                         }
 515
 516                         if (!io_u)
 517                                 break;
 518                 }
 519
 520                 if (td->o.verify_async)
 521                         io_u->end_io = verify_io_u_async;
 522                 else
 523                         io_u->end_io = verify_io_u;
 524
 525                 ddir = io_u->ddir;
 526
 527                 ret = td_io_queue(td, io_u);
 528                 switch (ret) {
 529                 case FIO_Q_COMPLETED:
 530                         if (io_u->error) {
 531                                 ret = -io_u->error;
 532                                 clear_io_u(td, io_u);
 533                         } else if (io_u->resid) {
 534                                 int bytes = io_u->xfer_buflen - io_u->resid;
 535
 536                                 /*
 537                                  * zero read, fail
 538                                  */
 539                                 if (!bytes) {
 540                                         td_verror(td, EIO, "full resid");
 541                                         put_io_u(td, io_u);
 542                                         break;
 543                                 }
 544
 545                                 io_u->xfer_buflen = io_u->resid;
 546                                 io_u->xfer_buf += bytes;
 547                                 io_u->offset += bytes;
 548
 549                                 if (ddir_rw(io_u->ddir))
 550                                         td->ts.short_io_u[io_u->ddir]++;
 551
 552                                 f = io_u->file;
 553                                 if (io_u->offset == f->real_file_size)
 554                                         goto sync_done;
 555
 556                                 requeue_io_u(td, &io_u);
 557                         } else {
 558 sync_done:
 559                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 560                                 if (ret < 0)
 561                                         break;
 562                         }
 563                         continue;
 564                 case FIO_Q_QUEUED:
 565                         break;
 566                 case FIO_Q_BUSY:
 567                         requeue_io_u(td, &io_u);
 568                         ret2 = td_io_commit(td);
 569                         if (ret2 < 0)
 570                                 ret = ret2;
 571                         break;
 572                 default:
 573                         assert(ret < 0);
 574                         td_verror(td, -ret, "td_io_queue");
 575                         break;
 576                 }
 577
 578                 if (break_on_this_error(td, ddir, &ret))
 579                         break;
 580
 581                 /*
 582                  * if we can queue more, do so. but check if there are
 583                  * completed io_u's first. Note that we can get BUSY even
 584                  * without IO queued, if the system is resource starved.
 585                  */
 586                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 587                 if (full || !td->o.iodepth_batch_complete) {
 588                         min_events = min(td->o.iodepth_batch_complete,
 589                                          td->cur_depth);
 590                         /*
 591                          * if the queue is full, we MUST reap at least 1 event
 592                          */
 593                         if (full && !min_events)
 594                                 min_events = 1;
 595
 596                         do {
 597                                 /*
 598                                  * Reap required number of io units, if any,
 599                                  * and do the verification on them through
 600                                  * the callback handler
 601                                  */
 602                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 603                                         ret = -1;
 604                                         break;
 605                                 }
 606                         } while (full && (td->cur_depth > td->o.iodepth_low));
 607                 }
 608                 if (ret < 0)
 609                         break;
 610         }
 611
 612         check_update_rusage(td);
 613
 614         if (!td->error) {
 615                 min_events = td->cur_depth;
 616
 617                 if (min_events)
 618                         ret = io_u_queued_complete(td, min_events, NULL);
 619         } else
 620                 cleanup_pending_aio(td);
 621
 622         td_set_runstate(td, TD_RUNNING);
 623
 624         dprint(FD_VERIFY, "exiting loop\n");
 625 }
 626
 627 static int io_bytes_exceeded(struct thread_data *td)
 628 {
 629         unsigned long long bytes;
 630
 631         if (td_rw(td))
 632                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 633         else if (td_write(td))
 634                 bytes = td->this_io_bytes[DDIR_WRITE];
 635         else if (td_read(td))
 636                 bytes = td->this_io_bytes[DDIR_READ];
 637         else
 638                 bytes = td->this_io_bytes[DDIR_TRIM];
 639
 640         return bytes >= td->o.size;
 641 }
 642
 643 /*
 644  * Main IO worker function. It retrieves io_u's to process and queues
 645  * and reaps them, checking for rate and errors along the way.
 646  *
 647  * Returns number of bytes written and trimmed.
 648  */
 649 static uint64_t do_io(struct thread_data *td)
 650 {
 651         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 652         unsigned int i;
 653         int ret = 0;
 654
 655         if (in_ramp_time(td))
 656                 td_set_runstate(td, TD_RAMP);
 657         else
 658                 td_set_runstate(td, TD_RUNNING);
 659
 660         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 661                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 662                 td->o.time_based) {
 663                 struct timeval comp_time;
 664                 int min_evts = 0;
 665                 struct io_u *io_u;
 666                 int ret2, full;
 667                 enum fio_ddir ddir;
 668
 669                 check_update_rusage(td);
 670
 671                 if (td->terminate || td->done)
 672                         break;
 673
 674                 update_tv_cache(td);
 675
 676                 if (runtime_exceeded(td, &td->tv_cache)) {
 677                         __update_tv_cache(td);
 678                         if (runtime_exceeded(td, &td->tv_cache)) {
 679                                 td->terminate = 1;
 680                                 break;
 681                         }
 682                 }
 683
 684                 if (flow_threshold_exceeded(td))
 685                         continue;
 686
 687                 io_u = get_io_u(td);
 688                 if (!io_u)
 689                         break;
 690
 691                 ddir = io_u->ddir;
 692
 693                 /*
 694                  * Add verification end_io handler if:
 695                  *      - Asked to verify (!td_rw(td))
 696                  *      - Or the io_u is from our verify list (mixed write/ver)
 697                  */
 698                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 699                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 700                         if (td->o.verify_async)
 701                                 io_u->end_io = verify_io_u_async;
 702                         else
 703                                 io_u->end_io = verify_io_u;
 704                         td_set_runstate(td, TD_VERIFYING);
 705                 } else if (in_ramp_time(td))
 706                         td_set_runstate(td, TD_RAMP);
 707                 else
 708                         td_set_runstate(td, TD_RUNNING);
 709
 710                 ret = td_io_queue(td, io_u);
 711                 switch (ret) {
 712                 case FIO_Q_COMPLETED:
 713                         if (io_u->error) {
 714                                 ret = -io_u->error;
 715                                 clear_io_u(td, io_u);
 716                         } else if (io_u->resid) {
 717                                 int bytes = io_u->xfer_buflen - io_u->resid;
 718                                 struct fio_file *f = io_u->file;
 719
 720                                 /*
 721                                  * zero read, fail
 722                                  */
 723                                 if (!bytes) {
 724                                         td_verror(td, EIO, "full resid");
 725                                         put_io_u(td, io_u);
 726                                         break;
 727                                 }
 728
 729                                 io_u->xfer_buflen = io_u->resid;
 730                                 io_u->xfer_buf += bytes;
 731                                 io_u->offset += bytes;
 732
 733                                 if (ddir_rw(io_u->ddir))
 734                                         td->ts.short_io_u[io_u->ddir]++;
 735
 736                                 if (io_u->offset == f->real_file_size)
 737                                         goto sync_done;
 738
 739                                 requeue_io_u(td, &io_u);
 740                         } else {
 741 sync_done:
 742                                 if (__should_check_rate(td, DDIR_READ) ||
 743                                     __should_check_rate(td, DDIR_WRITE) ||
 744                                     __should_check_rate(td, DDIR_TRIM))
 745                                         fio_gettime(&comp_time, NULL);
 746
 747                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 748                                 if (ret < 0)
 749                                         break;
 750                         }
 751                         break;
 752                 case FIO_Q_QUEUED:
 753                         /*
 754                          * if the engine doesn't have a commit hook,
 755                          * the io_u is really queued. if it does have such
 756                          * a hook, it has to call io_u_queued() itself.
 757                          */
 758                         if (td->io_ops->commit == NULL)
 759                                 io_u_queued(td, io_u);
 760                         break;
 761                 case FIO_Q_BUSY:
 762                         requeue_io_u(td, &io_u);
 763                         ret2 = td_io_commit(td);
 764                         if (ret2 < 0)
 765                                 ret = ret2;
 766                         break;
 767                 default:
 768                         assert(ret < 0);
 769                         put_io_u(td, io_u);
 770                         break;
 771                 }
 772
 773                 if (break_on_this_error(td, ddir, &ret))
 774                         break;
 775
 776                 /*
 777                  * See if we need to complete some commands. Note that we
 778                  * can get BUSY even without IO queued, if the system is
 779                  * resource starved.
 780                  */
 781                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 782                 if (full || !td->o.iodepth_batch_complete) {
 783                         min_evts = min(td->o.iodepth_batch_complete,
 784                                         td->cur_depth);
 785                         /*
 786                          * if the queue is full, we MUST reap at least 1 event
 787                          */
 788                         if (full && !min_evts)
 789                                 min_evts = 1;
 790
 791                         if (__should_check_rate(td, DDIR_READ) ||
 792                             __should_check_rate(td, DDIR_WRITE) ||
 793                             __should_check_rate(td, DDIR_TRIM))
 794                                 fio_gettime(&comp_time, NULL);
 795
 796                         do {
 797                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 798                                 if (ret < 0)
 799                                         break;
 800
 801                         } while (full && (td->cur_depth > td->o.iodepth_low));
 802                 }
 803
 804                 if (ret < 0)
 805                         break;
 806                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 807                         continue;
 808
 809                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 810                         if (check_min_rate(td, &comp_time, bytes_done)) {
 811                                 if (exitall_on_terminate)
 812                                         fio_terminate_threads(td->groupid);
 813                                 td_verror(td, EIO, "check_min_rate");
 814                                 break;
 815                         }
 816                 }
 817
 818                 if (td->o.thinktime) {
 819                         unsigned long long b;
 820
 821                         b = ddir_rw_sum(td->io_blocks);
 822                         if (!(b % td->o.thinktime_blocks)) {
 823                                 int left;
 824
 825                                 if (td->o.thinktime_spin)
 826                                         usec_spin(td->o.thinktime_spin);
 827
 828                                 left = td->o.thinktime - td->o.thinktime_spin;
 829                                 if (left)
 830                                         usec_sleep(td, left);
 831                         }
 832                 }
 833         }
 834
 835         check_update_rusage(td);
 836
 837         if (td->trim_entries)
 838                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 839
 840         if (td->o.fill_device && td->error == ENOSPC) {
 841                 td->error = 0;
 842                 td->terminate = 1;
 843         }
 844         if (!td->error) {
 845                 struct fio_file *f;
 846
 847                 i = td->cur_depth;
 848                 if (i) {
 849                         ret = io_u_queued_complete(td, i, bytes_done);
 850                         if (td->o.fill_device && td->error == ENOSPC)
 851                                 td->error = 0;
 852                 }
 853
 854                 if (should_fsync(td) && td->o.end_fsync) {
 855                         td_set_runstate(td, TD_FSYNCING);
 856
 857                         for_each_file(td, f, i) {
 858                                 if (!fio_file_fsync(td, f))
 859                                         continue;
 860
 861                                 log_err("fio: end_fsync failed for file %s\n",
 862                                                                 f->file_name);
 863                         }
 864                 }
 865         } else
 866                 cleanup_pending_aio(td);
 867
 868         /*
 869          * stop job if we failed doing any IO
 870          */
 871         if (!ddir_rw_sum(td->this_io_bytes))
 872                 td->done = 1;
 873
 874         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 875 }
 876
 877 static void cleanup_io_u(struct thread_data *td)
 878 {
 879         struct flist_head *entry, *n;
 880         struct io_u *io_u;
 881
 882         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 883                 io_u = flist_entry(entry, struct io_u, list);
 884
 885                 flist_del(&io_u->list);
 886
 887                 if (td->io_ops->io_u_free)
 888                         td->io_ops->io_u_free(td, io_u);
 889
 890                 fio_memfree(io_u, sizeof(*io_u));
 891         }
 892
 893         free_io_mem(td);
 894 }
 895
 896 static int init_io_u(struct thread_data *td)
 897 {
 898         struct io_u *io_u;
 899         unsigned int max_bs, min_write;
 900         int cl_align, i, max_units;
 901         int data_xfer = 1;
 902         char *p;
 903
 904         max_units = td->o.iodepth;
 905         max_bs = td_max_bs(td);
 906         min_write = td->o.min_bs[DDIR_WRITE];
 907         td->orig_buffer_size = (unsigned long long) max_bs
 908                                         * (unsigned long long) max_units;
 909
 910         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 911                 data_xfer = 0;
 912
 913         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 914                 unsigned long bs;
 915
 916                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 917                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 918         }
 919
 920         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 921                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 922                 return 1;
 923         }
 924
 925         if (data_xfer && allocate_io_mem(td))
 926                 return 1;
 927
 928         if (td->o.odirect || td->o.mem_align ||
 929             (td->io_ops->flags & FIO_RAWIO))
 930                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 931         else
 932                 p = td->orig_buffer;
 933
 934         cl_align = os_cache_line_size();
 935
 936         for (i = 0; i < max_units; i++) {
 937                 void *ptr;
 938
 939                 if (td->terminate)
 940                         return 1;
 941
 942                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 943                 if (!ptr) {
 944                         log_err("fio: unable to allocate aligned memory\n");
 945                         break;
 946                 }
 947
 948                 io_u = ptr;
 949                 memset(io_u, 0, sizeof(*io_u));
 950                 INIT_FLIST_HEAD(&io_u->list);
 951                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 952
 953                 if (data_xfer) {
 954                         io_u->buf = p;
 955                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 956
 957                         if (td_write(td))
 958                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 959                         if (td_write(td) && td->o.verify_pattern_bytes) {
 960                                 /*
 961                                  * Fill the buffer with the pattern if we are
 962                                  * going to be doing writes.
 963                                  */
 964                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 965                         }
 966                 }
 967
 968                 io_u->index = i;
 969                 io_u->flags = IO_U_F_FREE;
 970                 flist_add(&io_u->list, &td->io_u_freelist);
 971
 972                 if (td->io_ops->io_u_init) {
 973                         int ret = td->io_ops->io_u_init(td, io_u);
 974
 975                         if (ret) {
 976                                 log_err("fio: failed to init engine data: %d\n", ret);
 977                                 return 1;
 978                         }
 979                 }
 980
 981                 p += max_bs;
 982         }
 983
 984         return 0;
 985 }
 986
 987 static int switch_ioscheduler(struct thread_data *td)
 988 {
 989         char tmp[256], tmp2[128];
 990         FILE *f;
 991         int ret;
 992
 993         if (td->io_ops->flags & FIO_DISKLESSIO)
 994                 return 0;
 995
 996         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 997
 998         f = fopen(tmp, "r+");
 999         if (!f) {
1000                 if (errno == ENOENT) {
1001                         log_err("fio: os or kernel doesn't support IO scheduler"
1002                                 " switching\n");
1003                         return 0;
1004                 }
1005                 td_verror(td, errno, "fopen iosched");
1006                 return 1;
1007         }
1008
1009         /*
1010          * Set io scheduler.
1011          */
1012         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1013         if (ferror(f) || ret != 1) {
1014                 td_verror(td, errno, "fwrite");
1015                 fclose(f);
1016                 return 1;
1017         }
1018
1019         rewind(f);
1020
1021         /*
1022          * Read back and check that the selected scheduler is now the default.
1023          */
1024         ret = fread(tmp, 1, sizeof(tmp), f);
1025         if (ferror(f) || ret < 0) {
1026                 td_verror(td, errno, "fread");
1027                 fclose(f);
1028                 return 1;
1029         }
1030
1031         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1032         if (!strstr(tmp, tmp2)) {
1033                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1034                 td_verror(td, EINVAL, "iosched_switch");
1035                 fclose(f);
1036                 return 1;
1037         }
1038
1039         fclose(f);
1040         return 0;
1041 }
1042
1043 static int keep_running(struct thread_data *td)
1044 {
1045         if (td->done)
1046                 return 0;
1047         if (td->o.time_based)
1048                 return 1;
1049         if (td->o.loops) {
1050                 td->o.loops--;
1051                 return 1;
1052         }
1053
1054         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1055                 uint64_t diff;
1056
1057                 /*
1058                  * If the difference is less than the minimum IO size, we
1059                  * are done.
1060                  */
1061                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1062                 if (diff < td_max_bs(td))
1063                         return 0;
1064
1065                 return 1;
1066         }
1067
1068         return 0;
1069 }
1070
1071 static int exec_string(const char *string)
1072 {
1073         int ret, newlen = strlen(string) + 1 + 8;
1074         char *str;
1075
1076         str = malloc(newlen);
1077         sprintf(str, "sh -c %s", string);
1078
1079         ret = system(str);
1080         if (ret == -1)
1081                 log_err("fio: exec of cmd <%s> failed\n", str);
1082
1083         free(str);
1084         return ret;
1085 }
1086
1087 /*
1088  * Entry point for the thread based jobs. The process based jobs end up
1089  * here as well, after a little setup.
1090  */
1091 static void *thread_main(void *data)
1092 {
1093         unsigned long long elapsed;
1094         struct thread_data *td = data;
1095         struct thread_options *o = &td->o;
1096         pthread_condattr_t attr;
1097         int clear_state;
1098         int ret;
1099
1100         if (!o->use_thread) {
1101                 setsid();
1102                 td->pid = getpid();
1103         } else
1104                 td->pid = gettid();
1105
1106         fio_local_clock_init(td->o.use_thread);
1107
1108         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1109
1110         if (is_backend)
1111                 fio_server_send_start(td);
1112
1113         INIT_FLIST_HEAD(&td->io_u_freelist);
1114         INIT_FLIST_HEAD(&td->io_u_busylist);
1115         INIT_FLIST_HEAD(&td->io_u_requeues);
1116         INIT_FLIST_HEAD(&td->io_log_list);
1117         INIT_FLIST_HEAD(&td->io_hist_list);
1118         INIT_FLIST_HEAD(&td->verify_list);
1119         INIT_FLIST_HEAD(&td->trim_list);
1120         INIT_FLIST_HEAD(&td->next_rand_list);
1121         pthread_mutex_init(&td->io_u_lock, NULL);
1122         td->io_hist_tree = RB_ROOT;
1123
1124         pthread_condattr_init(&attr);
1125         pthread_cond_init(&td->verify_cond, &attr);
1126         pthread_cond_init(&td->free_cond, &attr);
1127
1128         td_set_runstate(td, TD_INITIALIZED);
1129         dprint(FD_MUTEX, "up startup_mutex\n");
1130         fio_mutex_up(startup_mutex);
1131         dprint(FD_MUTEX, "wait on td->mutex\n");
1132         fio_mutex_down(td->mutex);
1133         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1134
1135         /*
1136          * the ->mutex mutex is now no longer used, close it to avoid
1137          * eating a file descriptor
1138          */
1139         fio_mutex_remove(td->mutex);
1140         td->mutex = NULL;
1141
1142         /*
1143          * A new gid requires privilege, so we need to do this before setting
1144          * the uid.
1145          */
1146         if (o->gid != -1U && setgid(o->gid)) {
1147                 td_verror(td, errno, "setgid");
1148                 goto err;
1149         }
1150         if (o->uid != -1U && setuid(o->uid)) {
1151                 td_verror(td, errno, "setuid");
1152                 goto err;
1153         }
1154
1155         /*
1156          * If we have a gettimeofday() thread, make sure we exclude that
1157          * thread from this job
1158          */
1159         if (o->gtod_cpu)
1160                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1161
1162         /*
1163          * Set affinity first, in case it has an impact on the memory
1164          * allocations.
1165          */
1166         if (o->cpumask_set) {
1167                 ret = fio_setaffinity(td->pid, o->cpumask);
1168                 if (ret == -1) {
1169                         td_verror(td, errno, "cpu_set_affinity");
1170                         goto err;
1171                 }
1172         }
1173
1174 #ifdef CONFIG_LIBNUMA
1175         /* numa node setup */
1176         if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1177                 int ret;
1178
1179                 if (numa_available() < 0) {
1180                         td_verror(td, errno, "Does not support NUMA API\n");
1181                         goto err;
1182                 }
1183
1184                 if (td->o.numa_cpumask_set) {
1185                         ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1186                         if (ret == -1) {
1187                                 td_verror(td, errno, \
1188                                         "numa_run_on_node_mask failed\n");
1189                                 goto err;
1190                         }
1191                 }
1192
1193                 if (td->o.numa_memmask_set) {
1194
1195                         switch (td->o.numa_mem_mode) {
1196                         case MPOL_INTERLEAVE:
1197                                 numa_set_interleave_mask(td->o.numa_memnodesmask);
1198                                 break;
1199                         case MPOL_BIND:
1200                                 numa_set_membind(td->o.numa_memnodesmask);
1201                                 break;
1202                         case MPOL_LOCAL:
1203                                 numa_set_localalloc();
1204                                 break;
1205                         case MPOL_PREFERRED:
1206                                 numa_set_preferred(td->o.numa_mem_prefer_node);
1207                                 break;
1208                         case MPOL_DEFAULT:
1209                         default:
1210                                 break;
1211                         }
1212
1213                 }
1214         }
1215 #endif
1216
1217         /*
1218          * May alter parameters that init_io_u() will use, so we need to
1219          * do this first.
1220          */
1221         if (init_iolog(td))
1222                 goto err;
1223
1224         if (init_io_u(td))
1225                 goto err;
1226
1227         if (o->verify_async && verify_async_init(td))
1228                 goto err;
1229
1230         if (o->ioprio) {
1231                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1232                 if (ret == -1) {
1233                         td_verror(td, errno, "ioprio_set");
1234                         goto err;
1235                 }
1236         }
1237
1238         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1239                 goto err;
1240
1241         errno = 0;
1242         if (nice(o->nice) == -1 && errno != 0) {
1243                 td_verror(td, errno, "nice");
1244                 goto err;
1245         }
1246
1247         if (o->ioscheduler && switch_ioscheduler(td))
1248                 goto err;
1249
1250         if (!o->create_serialize && setup_files(td))
1251                 goto err;
1252
1253         if (td_io_init(td))
1254                 goto err;
1255
1256         if (init_random_map(td))
1257                 goto err;
1258
1259         if (o->exec_prerun && exec_string(o->exec_prerun))
1260                 goto err;
1261
1262         if (o->pre_read) {
1263                 if (pre_read_files(td) < 0)
1264                         goto err;
1265         }
1266
1267         fio_verify_init(td);
1268
1269         fio_gettime(&td->epoch, NULL);
1270         fio_getrusage(&td->ru_start);
1271         clear_state = 0;
1272         while (keep_running(td)) {
1273                 uint64_t verify_bytes;
1274
1275                 fio_gettime(&td->start, NULL);
1276                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1277                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1278                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1279
1280                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1281                                 td->o.ratemin[DDIR_TRIM]) {
1282                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1283                                                 sizeof(td->bw_sample_time));
1284                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1285                                                 sizeof(td->bw_sample_time));
1286                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1287                                                 sizeof(td->bw_sample_time));
1288                 }
1289
1290                 if (clear_state)
1291                         clear_io_state(td);
1292
1293                 prune_io_piece_log(td);
1294
1295                 verify_bytes = do_io(td);
1296
1297                 clear_state = 1;
1298
1299                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1300                         elapsed = utime_since_now(&td->start);
1301                         td->ts.runtime[DDIR_READ] += elapsed;
1302                 }
1303                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1304                         elapsed = utime_since_now(&td->start);
1305                         td->ts.runtime[DDIR_WRITE] += elapsed;
1306                 }
1307                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1308                         elapsed = utime_since_now(&td->start);
1309                         td->ts.runtime[DDIR_TRIM] += elapsed;
1310                 }
1311
1312                 if (td->error || td->terminate)
1313                         break;
1314
1315                 if (!td->o.do_verify ||
1316                     td->o.verify == VERIFY_NONE ||
1317                     (td->io_ops->flags & FIO_UNIDIR))
1318                         continue;
1319
1320                 clear_io_state(td);
1321
1322                 fio_gettime(&td->start, NULL);
1323
1324                 do_verify(td, verify_bytes);
1325
1326                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1327
1328                 if (td->error || td->terminate)
1329                         break;
1330         }
1331
1332         update_rusage_stat(td);
1333         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1334         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1335         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1336         td->ts.total_run_time = mtime_since_now(&td->epoch);
1337         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1338         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1339         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1340
1341         fio_unpin_memory(td);
1342
1343         fio_mutex_down(writeout_mutex);
1344         if (td->bw_log) {
1345                 if (td->o.bw_log_file) {
1346                         finish_log_named(td, td->bw_log,
1347                                                 td->o.bw_log_file, "bw");
1348                 } else
1349                         finish_log(td, td->bw_log, "bw");
1350         }
1351         if (td->lat_log) {
1352                 if (td->o.lat_log_file) {
1353                         finish_log_named(td, td->lat_log,
1354                                                 td->o.lat_log_file, "lat");
1355                 } else
1356                         finish_log(td, td->lat_log, "lat");
1357         }
1358         if (td->slat_log) {
1359                 if (td->o.lat_log_file) {
1360                         finish_log_named(td, td->slat_log,
1361                                                 td->o.lat_log_file, "slat");
1362                 } else
1363                         finish_log(td, td->slat_log, "slat");
1364         }
1365         if (td->clat_log) {
1366                 if (td->o.lat_log_file) {
1367                         finish_log_named(td, td->clat_log,
1368                                                 td->o.lat_log_file, "clat");
1369                 } else
1370                         finish_log(td, td->clat_log, "clat");
1371         }
1372         if (td->iops_log) {
1373                 if (td->o.iops_log_file) {
1374                         finish_log_named(td, td->iops_log,
1375                                                 td->o.iops_log_file, "iops");
1376                 } else
1377                         finish_log(td, td->iops_log, "iops");
1378         }
1379
1380         fio_mutex_up(writeout_mutex);
1381         if (td->o.exec_postrun)
1382                 exec_string(td->o.exec_postrun);
1383
1384         if (exitall_on_terminate)
1385                 fio_terminate_threads(td->groupid);
1386
1387 err:
1388         if (td->error)
1389                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1390                                                         td->verror);
1391
1392         if (td->o.verify_async)
1393                 verify_async_exit(td);
1394
1395         close_and_free_files(td);
1396         cleanup_io_u(td);
1397         close_ioengine(td);
1398         cgroup_shutdown(td, &cgroup_mnt);
1399
1400         if (o->cpumask_set) {
1401                 int ret = fio_cpuset_exit(&o->cpumask);
1402
1403                 td_verror(td, ret, "fio_cpuset_exit");
1404         }
1405
1406         /*
1407          * do this very late, it will log file closing as well
1408          */
1409         if (td->o.write_iolog_file)
1410                 write_iolog_close(td);
1411
1412         fio_mutex_remove(td->rusage_sem);
1413         td->rusage_sem = NULL;
1414
1415         td_set_runstate(td, TD_EXITED);
1416         return (void *) (uintptr_t) td->error;
1417 }
1418
1419
1420 /*
1421  * We cannot pass the td data into a forked process, so attach the td and
1422  * pass it to the thread worker.
1423  */
1424 static int fork_main(int shmid, int offset)
1425 {
1426         struct thread_data *td;
1427         void *data, *ret;
1428
1429 #ifndef __hpux
1430         data = shmat(shmid, NULL, 0);
1431         if (data == (void *) -1) {
1432                 int __err = errno;
1433
1434                 perror("shmat");
1435                 return __err;
1436         }
1437 #else
1438         /*
1439          * HP-UX inherits shm mappings?
1440          */
1441         data = threads;
1442 #endif
1443
1444         td = data + offset * sizeof(struct thread_data);
1445         ret = thread_main(td);
1446         shmdt(data);
1447         return (int) (uintptr_t) ret;
1448 }
1449
1450 /*
1451  * Run over the job map and reap the threads that have exited, if any.
1452  */
1453 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1454                          unsigned int *m_rate)
1455 {
1456         struct thread_data *td;
1457         unsigned int cputhreads, realthreads, pending;
1458         int i, status, ret;
1459
1460         /*
1461          * reap exited threads (TD_EXITED -> TD_REAPED)
1462          */
1463         realthreads = pending = cputhreads = 0;
1464         for_each_td(td, i) {
1465                 int flags = 0;
1466
1467                 /*
1468                  * ->io_ops is NULL for a thread that has closed its
1469                  * io engine
1470                  */
1471                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1472                         cputhreads++;
1473                 else
1474                         realthreads++;
1475
1476                 if (!td->pid) {
1477                         pending++;
1478                         continue;
1479                 }
1480                 if (td->runstate == TD_REAPED)
1481                         continue;
1482                 if (td->o.use_thread) {
1483                         if (td->runstate == TD_EXITED) {
1484                                 td_set_runstate(td, TD_REAPED);
1485                                 goto reaped;
1486                         }
1487                         continue;
1488                 }
1489
1490                 flags = WNOHANG;
1491                 if (td->runstate == TD_EXITED)
1492                         flags = 0;
1493
1494                 /*
1495                  * check if someone quit or got killed in an unusual way
1496                  */
1497                 ret = waitpid(td->pid, &status, flags);
1498                 if (ret < 0) {
1499                         if (errno == ECHILD) {
1500                                 log_err("fio: pid=%d disappeared %d\n",
1501                                                 (int) td->pid, td->runstate);
1502                                 td->sig = ECHILD;
1503                                 td_set_runstate(td, TD_REAPED);
1504                                 goto reaped;
1505                         }
1506                         perror("waitpid");
1507                 } else if (ret == td->pid) {
1508                         if (WIFSIGNALED(status)) {
1509                                 int sig = WTERMSIG(status);
1510
1511                                 if (sig != SIGTERM && sig != SIGUSR2)
1512                                         log_err("fio: pid=%d, got signal=%d\n",
1513                                                         (int) td->pid, sig);
1514                                 td->sig = sig;
1515                                 td_set_runstate(td, TD_REAPED);
1516                                 goto reaped;
1517                         }
1518                         if (WIFEXITED(status)) {
1519                                 if (WEXITSTATUS(status) && !td->error)
1520                                         td->error = WEXITSTATUS(status);
1521
1522                                 td_set_runstate(td, TD_REAPED);
1523                                 goto reaped;
1524                         }
1525                 }
1526
1527                 /*
1528                  * thread is not dead, continue
1529                  */
1530                 pending++;
1531                 continue;
1532 reaped:
1533                 (*nr_running)--;
1534                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1535                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1536                 if (!td->pid)
1537                         pending--;
1538
1539                 if (td->error)
1540                         exit_value++;
1541
1542                 done_secs += mtime_since_now(&td->epoch) / 1000;
1543         }
1544
1545         if (*nr_running == cputhreads && !pending && realthreads)
1546                 fio_terminate_threads(TERMINATE_ALL);
1547 }
1548
1549 /*
1550  * Main function for kicking off and reaping jobs, as needed.
1551  */
1552 static void run_threads(void)
1553 {
1554         struct thread_data *td;
1555         unsigned long spent;
1556         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1557
1558         if (fio_gtod_offload && fio_start_gtod_thread())
1559                 return;
1560
1561         fio_idle_prof_init();
1562
1563         set_sig_handlers();
1564
1565         nr_thread = nr_process = 0;
1566         for_each_td(td, i) {
1567                 if (td->o.use_thread)
1568                         nr_thread++;
1569                 else
1570                         nr_process++;
1571         }
1572
1573         if (output_format == FIO_OUTPUT_NORMAL) {
1574                 log_info("Starting ");
1575                 if (nr_thread)
1576                         log_info("%d thread%s", nr_thread,
1577                                                 nr_thread > 1 ? "s" : "");
1578                 if (nr_process) {
1579                         if (nr_thread)
1580                                 log_info(" and ");
1581                         log_info("%d process%s", nr_process,
1582                                                 nr_process > 1 ? "es" : "");
1583                 }
1584                 log_info("\n");
1585                 fflush(stdout);
1586         }
1587
1588         todo = thread_number;
1589         nr_running = 0;
1590         nr_started = 0;
1591         m_rate = t_rate = 0;
1592
1593         for_each_td(td, i) {
1594                 print_status_init(td->thread_number - 1);
1595
1596                 if (!td->o.create_serialize)
1597                         continue;
1598
1599                 /*
1600                  * do file setup here so it happens sequentially,
1601                  * we don't want X number of threads getting their
1602                  * client data interspersed on disk
1603                  */
1604                 if (setup_files(td)) {
1605                         exit_value++;
1606                         if (td->error)
1607                                 log_err("fio: pid=%d, err=%d/%s\n",
1608                                         (int) td->pid, td->error, td->verror);
1609                         td_set_runstate(td, TD_REAPED);
1610                         todo--;
1611                 } else {
1612                         struct fio_file *f;
1613                         unsigned int j;
1614
1615                         /*
1616                          * for sharing to work, each job must always open
1617                          * its own files. so close them, if we opened them
1618                          * for creation
1619                          */
1620                         for_each_file(td, f, j) {
1621                                 if (fio_file_open(f))
1622                                         td_io_close_file(td, f);
1623                         }
1624                 }
1625         }
1626
1627         /* start idle threads before io threads start to run */
1628         fio_idle_prof_start();
1629
1630         set_genesis_time();
1631
1632         while (todo) {
1633                 struct thread_data *map[REAL_MAX_JOBS];
1634                 struct timeval this_start;
1635                 int this_jobs = 0, left;
1636
1637                 /*
1638                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1639                  */
1640                 for_each_td(td, i) {
1641                         if (td->runstate != TD_NOT_CREATED)
1642                                 continue;
1643
1644                         /*
1645                          * never got a chance to start, killed by other
1646                          * thread for some reason
1647                          */
1648                         if (td->terminate) {
1649                                 todo--;
1650                                 continue;
1651                         }
1652
1653                         if (td->o.start_delay) {
1654                                 spent = mtime_since_genesis();
1655
1656                                 if (td->o.start_delay * 1000 > spent)
1657                                         continue;
1658                         }
1659
1660                         if (td->o.stonewall && (nr_started || nr_running)) {
1661                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1662                                                         td->o.name);
1663                                 break;
1664                         }
1665
1666                         init_disk_util(td);
1667
1668                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1669                         td->update_rusage = 0;
1670
1671                         /*
1672                          * Set state to created. Thread will transition
1673                          * to TD_INITIALIZED when it's done setting up.
1674                          */
1675                         td_set_runstate(td, TD_CREATED);
1676                         map[this_jobs++] = td;
1677                         nr_started++;
1678
1679                         if (td->o.use_thread) {
1680                                 int ret;
1681
1682                                 dprint(FD_PROCESS, "will pthread_create\n");
1683                                 ret = pthread_create(&td->thread, NULL,
1684                                                         thread_main, td);
1685                                 if (ret) {
1686                                         log_err("pthread_create: %s\n",
1687                                                         strerror(ret));
1688                                         nr_started--;
1689                                         break;
1690                                 }
1691                                 ret = pthread_detach(td->thread);
1692                                 if (ret)
1693                                         log_err("pthread_detach: %s",
1694                                                         strerror(ret));
1695                         } else {
1696                                 pid_t pid;
1697                                 dprint(FD_PROCESS, "will fork\n");
1698                                 pid = fork();
1699                                 if (!pid) {
1700                                         int ret = fork_main(shm_id, i);
1701
1702                                         _exit(ret);
1703                                 } else if (i == fio_debug_jobno)
1704                                         *fio_debug_jobp = pid;
1705                         }
1706                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1707                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1708                                 log_err("fio: job startup hung? exiting.\n");
1709                                 fio_terminate_threads(TERMINATE_ALL);
1710                                 fio_abort = 1;
1711                                 nr_started--;
1712                                 break;
1713                         }
1714                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1715                 }
1716
1717                 /*
1718                  * Wait for the started threads to transition to
1719                  * TD_INITIALIZED.
1720                  */
1721                 fio_gettime(&this_start, NULL);
1722                 left = this_jobs;
1723                 while (left && !fio_abort) {
1724                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1725                                 break;
1726
1727                         usleep(100000);
1728
1729                         for (i = 0; i < this_jobs; i++) {
1730                                 td = map[i];
1731                                 if (!td)
1732                                         continue;
1733                                 if (td->runstate == TD_INITIALIZED) {
1734                                         map[i] = NULL;
1735                                         left--;
1736                                 } else if (td->runstate >= TD_EXITED) {
1737                                         map[i] = NULL;
1738                                         left--;
1739                                         todo--;
1740                                         nr_running++; /* work-around... */
1741                                 }
1742                         }
1743                 }
1744
1745                 if (left) {
1746                         log_err("fio: %d job%s failed to start\n", left,
1747                                         left > 1 ? "s" : "");
1748                         for (i = 0; i < this_jobs; i++) {
1749                                 td = map[i];
1750                                 if (!td)
1751                                         continue;
1752                                 kill(td->pid, SIGTERM);
1753                         }
1754                         break;
1755                 }
1756
1757                 /*
1758                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1759                  */
1760                 for_each_td(td, i) {
1761                         if (td->runstate != TD_INITIALIZED)
1762                                 continue;
1763
1764                         if (in_ramp_time(td))
1765                                 td_set_runstate(td, TD_RAMP);
1766                         else
1767                                 td_set_runstate(td, TD_RUNNING);
1768                         nr_running++;
1769                         nr_started--;
1770                         m_rate += ddir_rw_sum(td->o.ratemin);
1771                         t_rate += ddir_rw_sum(td->o.rate);
1772                         todo--;
1773                         fio_mutex_up(td->mutex);
1774                 }
1775
1776                 reap_threads(&nr_running, &t_rate, &m_rate);
1777
1778                 if (todo)
1779                         usleep(100000);
1780         }
1781
1782         while (nr_running) {
1783                 reap_threads(&nr_running, &t_rate, &m_rate);
1784                 usleep(10000);
1785         }
1786
1787         fio_idle_prof_stop();
1788
1789         update_io_ticks();
1790 }
1791
1792 void wait_for_disk_thread_exit(void)
1793 {
1794         fio_mutex_down(disk_thread_mutex);
1795 }
1796
1797 static void free_disk_util(void)
1798 {
1799         disk_util_start_exit();
1800         wait_for_disk_thread_exit();
1801         disk_util_prune_entries();
1802 }
1803
1804 static void *disk_thread_main(void *data)
1805 {
1806         int ret = 0;
1807
1808         fio_mutex_up(startup_mutex);
1809
1810         while (threads && !ret) {
1811                 usleep(DISK_UTIL_MSEC * 1000);
1812                 if (!threads)
1813                         break;
1814                 ret = update_io_ticks();
1815
1816                 if (!is_backend)
1817                         print_thread_status();
1818         }
1819
1820         fio_mutex_up(disk_thread_mutex);
1821         return NULL;
1822 }
1823
1824 static int create_disk_util_thread(void)
1825 {
1826         int ret;
1827
1828         setup_disk_util();
1829
1830         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1831
1832         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1833         if (ret) {
1834                 fio_mutex_remove(disk_thread_mutex);
1835                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1836                 return 1;
1837         }
1838
1839         ret = pthread_detach(disk_util_thread);
1840         if (ret) {
1841                 fio_mutex_remove(disk_thread_mutex);
1842                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1843                 return 1;
1844         }
1845
1846         dprint(FD_MUTEX, "wait on startup_mutex\n");
1847         fio_mutex_down(startup_mutex);
1848         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1849         return 0;
1850 }
1851
1852 int fio_backend(void)
1853 {
1854         struct thread_data *td;
1855         int i;
1856
1857         if (exec_profile) {
1858                 if (load_profile(exec_profile))
1859                         return 1;
1860                 free(exec_profile);
1861                 exec_profile = NULL;
1862         }
1863         if (!thread_number)
1864                 return 0;
1865
1866         if (write_bw_log) {
1867                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1868                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1869                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1870         }
1871
1872         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1873         if (startup_mutex == NULL)
1874                 return 1;
1875         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1876         if (writeout_mutex == NULL)
1877                 return 1;
1878
1879         set_genesis_time();
1880         create_disk_util_thread();
1881
1882         cgroup_list = smalloc(sizeof(*cgroup_list));
1883         INIT_FLIST_HEAD(cgroup_list);
1884
1885         run_threads();
1886
1887         if (!fio_abort) {
1888                 show_run_stats();
1889                 if (write_bw_log) {
1890                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1891                         __finish_log(agg_io_log[DDIR_WRITE],
1892                                         "agg-write_bw.log");
1893                         __finish_log(agg_io_log[DDIR_TRIM],
1894                                         "agg-write_bw.log");
1895                 }
1896         }
1897
1898         for_each_td(td, i)
1899                 fio_options_free(td);
1900
1901         free_disk_util();
1902         cgroup_kill(cgroup_list);
1903         sfree(cgroup_list);
1904         sfree(cgroup_mnt);
1905
1906         fio_mutex_remove(startup_mutex);
1907         fio_mutex_remove(writeout_mutex);
1908         fio_mutex_remove(disk_thread_mutex);
1909         return exit_value;
1910 }