fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 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 <sys/stat.h>
  34 #include <sys/wait.h>
  35 #include <sys/ipc.h>
  36 #include <sys/shm.h>
  37 #include <sys/mman.h>
  38
  39 #include "fio.h"
  40 #include "hash.h"
  41 #include "smalloc.h"
  42 #include "verify.h"
  43 #include "trim.h"
  44 #include "diskutil.h"
  45 #include "cgroup.h"
  46 #include "profile.h"
  47 #include "lib/rand.h"
  48 #include "memalign.h"
  49 #include "server.h"
  50
  51 unsigned long page_mask;
  52 unsigned long page_size;
  53
  54 #define PAGE_ALIGN(buf) \
  55         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  56
  57 int groupid = 0;
  58 unsigned int thread_number = 0;
  59 unsigned int nr_process = 0;
  60 unsigned int nr_thread = 0;
  61 int shm_id = 0;
  62 int temp_stall_ts;
  63 unsigned long done_secs = 0;
  64
  65 /*
  66  * Just expose an empty list, if the OS does not support disk util stats
  67  */
  68 #ifndef FIO_HAVE_DISK_UTIL
  69 FLIST_HEAD(disk_list);
  70 #endif
  71
  72 static struct fio_mutex *startup_mutex;
  73 static struct fio_mutex *writeout_mutex;
  74 static volatile int fio_abort;
  75 static int exit_value;
  76 static pthread_t gtod_thread;
  77 static pthread_t disk_util_thread;
  78 static struct flist_head *cgroup_list;
  79 static char *cgroup_mnt;
  80
  81 unsigned long arch_flags = 0;
  82
  83 struct io_log *agg_io_log[2];
  84
  85 #define JOB_START_TIMEOUT       (5 * 1000)
  86
  87 static const char *fio_os_strings[os_nr] = {
  88         "Invalid",
  89         "Linux",
  90         "AIX",
  91         "FreeBSD",
  92         "HP-UX",
  93         "OSX",
  94         "NetBSD",
  95         "Solaris",
  96         "Windows"
  97 };
  98
  99 static const char *fio_arch_strings[arch_nr] = {
 100         "Invalid",
 101         "x86-64",
 102         "x86",
 103         "ppc",
 104         "ia64",
 105         "s390",
 106         "alpha",
 107         "sparc",
 108         "sparc64",
 109         "arm",
 110         "sh",
 111         "hppa",
 112         "generic"
 113 };
 114
 115 const char *fio_get_os_string(int nr)
 116 {
 117         if (nr < os_nr)
 118                 return fio_os_strings[nr];
 119
 120         return NULL;
 121 }
 122
 123 const char *fio_get_arch_string(int nr)
 124 {
 125         if (nr < arch_nr)
 126                 return fio_arch_strings[nr];
 127
 128         return NULL;
 129 }
 130
 131 void td_set_runstate(struct thread_data *td, int runstate)
 132 {
 133         if (td->runstate == runstate)
 134                 return;
 135
 136         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
 137                                                 td->runstate, runstate);
 138         td->runstate = runstate;
 139 }
 140
 141 void fio_terminate_threads(int group_id)
 142 {
 143         struct thread_data *td;
 144         int i;
 145
 146         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
 147
 148         for_each_td(td, i) {
 149                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
 150                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
 151                                                 td->o.name, (int) td->pid);
 152                         td->terminate = 1;
 153                         td->o.start_delay = 0;
 154
 155                         /*
 156                          * if the thread is running, just let it exit
 157                          */
 158                         if (!td->pid)
 159                                 continue;
 160                         else if (td->runstate < TD_RAMP)
 161                                 kill(td->pid, SIGTERM);
 162                         else {
 163                                 struct ioengine_ops *ops = td->io_ops;
 164
 165                                 if (ops && (ops->flags & FIO_SIGTERM))
 166                                         kill(td->pid, SIGTERM);
 167                         }
 168                 }
 169         }
 170 }
 171
 172 static void sig_int(int sig)
 173 {
 174         if (threads) {
 175                 if (is_backend)
 176                         fio_server_got_signal(sig);
 177                 else {
 178                         log_info("\nfio: terminating on signal %d\n", sig);
 179                         fflush(stdout);
 180                         exit_value = 128;
 181                 }
 182
 183                 fio_terminate_threads(TERMINATE_ALL);
 184         }
 185 }
 186
 187 static void *disk_thread_main(void *data)
 188 {
 189         fio_mutex_up(startup_mutex);
 190
 191         while (threads) {
 192                 usleep(DISK_UTIL_MSEC * 1000);
 193                 if (!threads)
 194                         break;
 195                 update_io_ticks();
 196
 197                 if (!is_backend)
 198                         print_thread_status();
 199         }
 200
 201         return NULL;
 202 }
 203
 204 static int create_disk_util_thread(void)
 205 {
 206         int ret;
 207
 208         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 209         if (ret) {
 210                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 211                 return 1;
 212         }
 213
 214         ret = pthread_detach(disk_util_thread);
 215         if (ret) {
 216                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 217                 return 1;
 218         }
 219
 220         dprint(FD_MUTEX, "wait on startup_mutex\n");
 221         fio_mutex_down(startup_mutex);
 222         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 223         return 0;
 224 }
 225
 226 static void set_sig_handlers(void)
 227 {
 228         struct sigaction act;
 229
 230         memset(&act, 0, sizeof(act));
 231         act.sa_handler = sig_int;
 232         act.sa_flags = SA_RESTART;
 233         sigaction(SIGINT, &act, NULL);
 234
 235         memset(&act, 0, sizeof(act));
 236         act.sa_handler = sig_int;
 237         act.sa_flags = SA_RESTART;
 238         sigaction(SIGTERM, &act, NULL);
 239
 240         if (is_backend) {
 241                 memset(&act, 0, sizeof(act));
 242                 act.sa_handler = sig_int;
 243                 act.sa_flags = SA_RESTART;
 244                 sigaction(SIGPIPE, &act, NULL);
 245         }
 246 }
 247
 248 /*
 249  * Check if we are above the minimum rate given.
 250  */
 251 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 252                             enum fio_ddir ddir)
 253 {
 254         unsigned long long bytes = 0;
 255         unsigned long iops = 0;
 256         unsigned long spent;
 257         unsigned long rate;
 258         unsigned int ratemin = 0;
 259         unsigned int rate_iops = 0;
 260         unsigned int rate_iops_min = 0;
 261
 262         assert(ddir_rw(ddir));
 263
 264         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 265                 return 0;
 266
 267         /*
 268          * allow a 2 second settle period in the beginning
 269          */
 270         if (mtime_since(&td->start, now) < 2000)
 271                 return 0;
 272
 273         iops += td->this_io_blocks[ddir];
 274         bytes += td->this_io_bytes[ddir];
 275         ratemin += td->o.ratemin[ddir];
 276         rate_iops += td->o.rate_iops[ddir];
 277         rate_iops_min += td->o.rate_iops_min[ddir];
 278
 279         /*
 280          * if rate blocks is set, sample is running
 281          */
 282         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 283                 spent = mtime_since(&td->lastrate[ddir], now);
 284                 if (spent < td->o.ratecycle)
 285                         return 0;
 286
 287                 if (td->o.rate[ddir]) {
 288                         /*
 289                          * check bandwidth specified rate
 290                          */
 291                         if (bytes < td->rate_bytes[ddir]) {
 292                                 log_err("%s: min rate %u not met\n", td->o.name,
 293                                                                 ratemin);
 294                                 return 1;
 295                         } else {
 296                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 297                                 if (rate < ratemin ||
 298                                     bytes < td->rate_bytes[ddir]) {
 299                                         log_err("%s: min rate %u not met, got"
 300                                                 " %luKB/sec\n", td->o.name,
 301                                                         ratemin, rate);
 302                                         return 1;
 303                                 }
 304                         }
 305                 } else {
 306                         /*
 307                          * checks iops specified rate
 308                          */
 309                         if (iops < rate_iops) {
 310                                 log_err("%s: min iops rate %u not met\n",
 311                                                 td->o.name, rate_iops);
 312                                 return 1;
 313                         } else {
 314                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 315                                 if (rate < rate_iops_min ||
 316                                     iops < td->rate_blocks[ddir]) {
 317                                         log_err("%s: min iops rate %u not met,"
 318                                                 " got %lu\n", td->o.name,
 319                                                         rate_iops_min, rate);
 320                                 }
 321                         }
 322                 }
 323         }
 324
 325         td->rate_bytes[ddir] = bytes;
 326         td->rate_blocks[ddir] = iops;
 327         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 328         return 0;
 329 }
 330
 331 static int check_min_rate(struct thread_data *td, struct timeval *now,
 332                           unsigned long *bytes_done)
 333 {
 334         int ret = 0;
 335
 336         if (bytes_done[0])
 337                 ret |= __check_min_rate(td, now, 0);
 338         if (bytes_done[1])
 339                 ret |= __check_min_rate(td, now, 1);
 340
 341         return ret;
 342 }
 343
 344 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 345 {
 346         if (!td->o.timeout)
 347                 return 0;
 348         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 349                 return 1;
 350
 351         return 0;
 352 }
 353
 354 /*
 355  * When job exits, we can cancel the in-flight IO if we are using async
 356  * io. Attempt to do so.
 357  */
 358 static void cleanup_pending_aio(struct thread_data *td)
 359 {
 360         struct flist_head *entry, *n;
 361         struct io_u *io_u;
 362         int r;
 363
 364         /*
 365          * get immediately available events, if any
 366          */
 367         r = io_u_queued_complete(td, 0, NULL);
 368         if (r < 0)
 369                 return;
 370
 371         /*
 372          * now cancel remaining active events
 373          */
 374         if (td->io_ops->cancel) {
 375                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 376                         io_u = flist_entry(entry, struct io_u, list);
 377
 378                         /*
 379                          * if the io_u isn't in flight, then that generally
 380                          * means someone leaked an io_u. complain but fix
 381                          * it up, so we don't stall here.
 382                          */
 383                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 384                                 log_err("fio: non-busy IO on busy list\n");
 385                                 put_io_u(td, io_u);
 386                         } else {
 387                                 r = td->io_ops->cancel(td, io_u);
 388                                 if (!r)
 389                                         put_io_u(td, io_u);
 390                         }
 391                 }
 392         }
 393
 394         if (td->cur_depth)
 395                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 396 }
 397
 398 /*
 399  * Helper to handle the final sync of a file. Works just like the normal
 400  * io path, just does everything sync.
 401  */
 402 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 403 {
 404         struct io_u *io_u = __get_io_u(td);
 405         int ret;
 406
 407         if (!io_u)
 408                 return 1;
 409
 410         io_u->ddir = DDIR_SYNC;
 411         io_u->file = f;
 412
 413         if (td_io_prep(td, io_u)) {
 414                 put_io_u(td, io_u);
 415                 return 1;
 416         }
 417
 418 requeue:
 419         ret = td_io_queue(td, io_u);
 420         if (ret < 0) {
 421                 td_verror(td, io_u->error, "td_io_queue");
 422                 put_io_u(td, io_u);
 423                 return 1;
 424         } else if (ret == FIO_Q_QUEUED) {
 425                 if (io_u_queued_complete(td, 1, NULL) < 0)
 426                         return 1;
 427         } else if (ret == FIO_Q_COMPLETED) {
 428                 if (io_u->error) {
 429                         td_verror(td, io_u->error, "td_io_queue");
 430                         return 1;
 431                 }
 432
 433                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 434                         return 1;
 435         } else if (ret == FIO_Q_BUSY) {
 436                 if (td_io_commit(td))
 437                         return 1;
 438                 goto requeue;
 439         }
 440
 441         return 0;
 442 }
 443
 444 static inline void __update_tv_cache(struct thread_data *td)
 445 {
 446         fio_gettime(&td->tv_cache, NULL);
 447 }
 448
 449 static inline void update_tv_cache(struct thread_data *td)
 450 {
 451         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 452                 __update_tv_cache(td);
 453 }
 454
 455 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 456                                int *retptr)
 457 {
 458         int ret = *retptr;
 459
 460         if (ret < 0 || td->error) {
 461                 int err;
 462
 463                 if (ret < 0)
 464                         err = -ret;
 465                 else
 466                         err = td->error;
 467
 468                 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
 469                         return 1;
 470
 471                 if (td_non_fatal_error(err)) {
 472                         /*
 473                          * Continue with the I/Os in case of
 474                          * a non fatal error.
 475                          */
 476                         update_error_count(td, err);
 477                         td_clear_error(td);
 478                         *retptr = 0;
 479                         return 0;
 480                 } else if (td->o.fill_device && err == ENOSPC) {
 481                         /*
 482                          * We expect to hit this error if
 483                          * fill_device option is set.
 484                          */
 485                         td_clear_error(td);
 486                         td->terminate = 1;
 487                         return 1;
 488                 } else {
 489                         /*
 490                          * Stop the I/O in case of a fatal
 491                          * error.
 492                          */
 493                         update_error_count(td, err);
 494                         return 1;
 495                 }
 496         }
 497
 498         return 0;
 499 }
 500
 501 /*
 502  * The main verify engine. Runs over the writes we previously submitted,
 503  * reads the blocks back in, and checks the crc/md5 of the data.
 504  */
 505 static void do_verify(struct thread_data *td)
 506 {
 507         struct fio_file *f;
 508         struct io_u *io_u;
 509         int ret, min_events;
 510         unsigned int i;
 511
 512         dprint(FD_VERIFY, "starting loop\n");
 513
 514         /*
 515          * sync io first and invalidate cache, to make sure we really
 516          * read from disk.
 517          */
 518         for_each_file(td, f, i) {
 519                 if (!fio_file_open(f))
 520                         continue;
 521                 if (fio_io_sync(td, f))
 522                         break;
 523                 if (file_invalidate_cache(td, f))
 524                         break;
 525         }
 526
 527         if (td->error)
 528                 return;
 529
 530         td_set_runstate(td, TD_VERIFYING);
 531
 532         io_u = NULL;
 533         while (!td->terminate) {
 534                 int ret2, full;
 535
 536                 update_tv_cache(td);
 537
 538                 if (runtime_exceeded(td, &td->tv_cache)) {
 539                         __update_tv_cache(td);
 540                         if (runtime_exceeded(td, &td->tv_cache)) {
 541                                 td->terminate = 1;
 542                                 break;
 543                         }
 544                 }
 545
 546                 io_u = __get_io_u(td);
 547                 if (!io_u)
 548                         break;
 549
 550                 if (get_next_verify(td, io_u)) {
 551                         put_io_u(td, io_u);
 552                         break;
 553                 }
 554
 555                 if (td_io_prep(td, io_u)) {
 556                         put_io_u(td, io_u);
 557                         break;
 558                 }
 559
 560                 if (td->o.verify_async)
 561                         io_u->end_io = verify_io_u_async;
 562                 else
 563                         io_u->end_io = verify_io_u;
 564
 565                 ret = td_io_queue(td, io_u);
 566                 switch (ret) {
 567                 case FIO_Q_COMPLETED:
 568                         if (io_u->error) {
 569                                 ret = -io_u->error;
 570                                 clear_io_u(td, io_u);
 571                         } else if (io_u->resid) {
 572                                 int bytes = io_u->xfer_buflen - io_u->resid;
 573
 574                                 /*
 575                                  * zero read, fail
 576                                  */
 577                                 if (!bytes) {
 578                                         td_verror(td, EIO, "full resid");
 579                                         put_io_u(td, io_u);
 580                                         break;
 581                                 }
 582
 583                                 io_u->xfer_buflen = io_u->resid;
 584                                 io_u->xfer_buf += bytes;
 585                                 io_u->offset += bytes;
 586
 587                                 if (ddir_rw(io_u->ddir))
 588                                         td->ts.short_io_u[io_u->ddir]++;
 589
 590                                 f = io_u->file;
 591                                 if (io_u->offset == f->real_file_size)
 592                                         goto sync_done;
 593
 594                                 requeue_io_u(td, &io_u);
 595                         } else {
 596 sync_done:
 597                                 ret = io_u_sync_complete(td, io_u, NULL);
 598                                 if (ret < 0)
 599                                         break;
 600                         }
 601                         continue;
 602                 case FIO_Q_QUEUED:
 603                         break;
 604                 case FIO_Q_BUSY:
 605                         requeue_io_u(td, &io_u);
 606                         ret2 = td_io_commit(td);
 607                         if (ret2 < 0)
 608                                 ret = ret2;
 609                         break;
 610                 default:
 611                         assert(ret < 0);
 612                         td_verror(td, -ret, "td_io_queue");
 613                         break;
 614                 }
 615
 616                 if (break_on_this_error(td, io_u->ddir, &ret))
 617                         break;
 618
 619                 /*
 620                  * if we can queue more, do so. but check if there are
 621                  * completed io_u's first. Note that we can get BUSY even
 622                  * without IO queued, if the system is resource starved.
 623                  */
 624                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 625                 if (full || !td->o.iodepth_batch_complete) {
 626                         min_events = min(td->o.iodepth_batch_complete,
 627                                          td->cur_depth);
 628                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 629                                 min_events = 1;
 630
 631                         do {
 632                                 /*
 633                                  * Reap required number of io units, if any,
 634                                  * and do the verification on them through
 635                                  * the callback handler
 636                                  */
 637                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 638                                         ret = -1;
 639                                         break;
 640                                 }
 641                         } while (full && (td->cur_depth > td->o.iodepth_low));
 642                 }
 643                 if (ret < 0)
 644                         break;
 645         }
 646
 647         if (!td->error) {
 648                 min_events = td->cur_depth;
 649
 650                 if (min_events)
 651                         ret = io_u_queued_complete(td, min_events, NULL);
 652         } else
 653                 cleanup_pending_aio(td);
 654
 655         td_set_runstate(td, TD_RUNNING);
 656
 657         dprint(FD_VERIFY, "exiting loop\n");
 658 }
 659
 660 /*
 661  * Main IO worker function. It retrieves io_u's to process and queues
 662  * and reaps them, checking for rate and errors along the way.
 663  */
 664 static void do_io(struct thread_data *td)
 665 {
 666         unsigned int i;
 667         int ret = 0;
 668
 669         if (in_ramp_time(td))
 670                 td_set_runstate(td, TD_RAMP);
 671         else
 672                 td_set_runstate(td, TD_RUNNING);
 673
 674         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 675                 (!flist_empty(&td->trim_list)) ||
 676                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 677                 struct timeval comp_time;
 678                 unsigned long bytes_done[2] = { 0, 0 };
 679                 int min_evts = 0;
 680                 struct io_u *io_u;
 681                 int ret2, full;
 682                 enum fio_ddir ddir;
 683
 684                 if (td->terminate)
 685                         break;
 686
 687                 update_tv_cache(td);
 688
 689                 if (runtime_exceeded(td, &td->tv_cache)) {
 690                         __update_tv_cache(td);
 691                         if (runtime_exceeded(td, &td->tv_cache)) {
 692                                 td->terminate = 1;
 693                                 break;
 694                         }
 695                 }
 696
 697                 io_u = get_io_u(td);
 698                 if (!io_u)
 699                         break;
 700
 701                 ddir = io_u->ddir;
 702
 703                 /*
 704                  * Add verification end_io handler, if asked to verify
 705                  * a previously written file.
 706                  */
 707                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 708                     !td_rw(td)) {
 709                         if (td->o.verify_async)
 710                                 io_u->end_io = verify_io_u_async;
 711                         else
 712                                 io_u->end_io = verify_io_u;
 713                         td_set_runstate(td, TD_VERIFYING);
 714                 } else if (in_ramp_time(td))
 715                         td_set_runstate(td, TD_RAMP);
 716                 else
 717                         td_set_runstate(td, TD_RUNNING);
 718
 719                 ret = td_io_queue(td, io_u);
 720                 switch (ret) {
 721                 case FIO_Q_COMPLETED:
 722                         if (io_u->error) {
 723                                 ret = -io_u->error;
 724                                 clear_io_u(td, io_u);
 725                         } else if (io_u->resid) {
 726                                 int bytes = io_u->xfer_buflen - io_u->resid;
 727                                 struct fio_file *f = io_u->file;
 728
 729                                 /*
 730                                  * zero read, fail
 731                                  */
 732                                 if (!bytes) {
 733                                         td_verror(td, EIO, "full resid");
 734                                         put_io_u(td, io_u);
 735                                         break;
 736                                 }
 737
 738                                 io_u->xfer_buflen = io_u->resid;
 739                                 io_u->xfer_buf += bytes;
 740                                 io_u->offset += bytes;
 741
 742                                 if (ddir_rw(io_u->ddir))
 743                                         td->ts.short_io_u[io_u->ddir]++;
 744
 745                                 if (io_u->offset == f->real_file_size)
 746                                         goto sync_done;
 747
 748                                 requeue_io_u(td, &io_u);
 749                         } else {
 750 sync_done:
 751                                 if (__should_check_rate(td, 0) ||
 752                                     __should_check_rate(td, 1))
 753                                         fio_gettime(&comp_time, NULL);
 754
 755                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 756                                 if (ret < 0)
 757                                         break;
 758                         }
 759                         break;
 760                 case FIO_Q_QUEUED:
 761                         /*
 762                          * if the engine doesn't have a commit hook,
 763                          * the io_u is really queued. if it does have such
 764                          * a hook, it has to call io_u_queued() itself.
 765                          */
 766                         if (td->io_ops->commit == NULL)
 767                                 io_u_queued(td, io_u);
 768                         break;
 769                 case FIO_Q_BUSY:
 770                         requeue_io_u(td, &io_u);
 771                         ret2 = td_io_commit(td);
 772                         if (ret2 < 0)
 773                                 ret = ret2;
 774                         break;
 775                 default:
 776                         assert(ret < 0);
 777                         put_io_u(td, io_u);
 778                         break;
 779                 }
 780
 781                 if (break_on_this_error(td, ddir, &ret))
 782                         break;
 783
 784                 /*
 785                  * See if we need to complete some commands. Note that we
 786                  * can get BUSY even without IO queued, if the system is
 787                  * resource starved.
 788                  */
 789                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 790                 if (full || !td->o.iodepth_batch_complete) {
 791                         min_evts = min(td->o.iodepth_batch_complete,
 792                                         td->cur_depth);
 793                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 794                                 min_evts = 1;
 795
 796                         if (__should_check_rate(td, 0) ||
 797                             __should_check_rate(td, 1))
 798                                 fio_gettime(&comp_time, NULL);
 799
 800                         do {
 801                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 802                                 if (ret < 0)
 803                                         break;
 804
 805                         } while (full && (td->cur_depth > td->o.iodepth_low));
 806                 }
 807
 808                 if (ret < 0)
 809                         break;
 810                 if (!(bytes_done[0] + bytes_done[1]))
 811                         continue;
 812
 813                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 814                         if (check_min_rate(td, &comp_time, bytes_done)) {
 815                                 if (exitall_on_terminate)
 816                                         fio_terminate_threads(td->groupid);
 817                                 td_verror(td, EIO, "check_min_rate");
 818                                 break;
 819                         }
 820                 }
 821
 822                 if (td->o.thinktime) {
 823                         unsigned long long b;
 824
 825                         b = td->io_blocks[0] + td->io_blocks[1];
 826                         if (!(b % td->o.thinktime_blocks)) {
 827                                 int left;
 828
 829                                 if (td->o.thinktime_spin)
 830                                         usec_spin(td->o.thinktime_spin);
 831
 832                                 left = td->o.thinktime - td->o.thinktime_spin;
 833                                 if (left)
 834                                         usec_sleep(td, left);
 835                         }
 836                 }
 837         }
 838
 839         if (td->trim_entries)
 840                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 841
 842         if (td->o.fill_device && td->error == ENOSPC) {
 843                 td->error = 0;
 844                 td->terminate = 1;
 845         }
 846         if (!td->error) {
 847                 struct fio_file *f;
 848
 849                 i = td->cur_depth;
 850                 if (i) {
 851                         ret = io_u_queued_complete(td, i, NULL);
 852                         if (td->o.fill_device && td->error == ENOSPC)
 853                                 td->error = 0;
 854                 }
 855
 856                 if (should_fsync(td) && td->o.end_fsync) {
 857                         td_set_runstate(td, TD_FSYNCING);
 858
 859                         for_each_file(td, f, i) {
 860                                 if (!fio_file_open(f))
 861                                         continue;
 862                                 fio_io_sync(td, f);
 863                         }
 864                 }
 865         } else
 866                 cleanup_pending_aio(td);
 867
 868         /*
 869          * stop job if we failed doing any IO
 870          */
 871         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 872                 td->done = 1;
 873 }
 874
 875 static void cleanup_io_u(struct thread_data *td)
 876 {
 877         struct flist_head *entry, *n;
 878         struct io_u *io_u;
 879
 880         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 881                 io_u = flist_entry(entry, struct io_u, list);
 882
 883                 flist_del(&io_u->list);
 884                 fio_memfree(io_u, sizeof(*io_u));
 885         }
 886
 887         free_io_mem(td);
 888 }
 889
 890 static int init_io_u(struct thread_data *td)
 891 {
 892         struct io_u *io_u;
 893         unsigned int max_bs;
 894         int cl_align, i, max_units;
 895         char *p;
 896
 897         max_units = td->o.iodepth;
 898         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 899         td->orig_buffer_size = (unsigned long long) max_bs
 900                                         * (unsigned long long) max_units;
 901
 902         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 903                 unsigned long bs;
 904
 905                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 906                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 907         }
 908
 909         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 910                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 911                 return 1;
 912         }
 913
 914         if (allocate_io_mem(td))
 915                 return 1;
 916
 917         if (td->o.odirect || td->o.mem_align ||
 918             (td->io_ops->flags & FIO_RAWIO))
 919                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 920         else
 921                 p = td->orig_buffer;
 922
 923         cl_align = os_cache_line_size();
 924
 925         for (i = 0; i < max_units; i++) {
 926                 void *ptr;
 927
 928                 if (td->terminate)
 929                         return 1;
 930
 931                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 932                 if (!ptr) {
 933                         log_err("fio: unable to allocate aligned memory\n");
 934                         break;
 935                 }
 936
 937                 io_u = ptr;
 938                 memset(io_u, 0, sizeof(*io_u));
 939                 INIT_FLIST_HEAD(&io_u->list);
 940                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 941
 942                 if (!(td->io_ops->flags & FIO_NOIO)) {
 943                         io_u->buf = p;
 944                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 945
 946                         if (td_write(td))
 947                                 io_u_fill_buffer(td, io_u, max_bs);
 948                         if (td_write(td) && td->o.verify_pattern_bytes) {
 949                                 /*
 950                                  * Fill the buffer with the pattern if we are
 951                                  * going to be doing writes.
 952                                  */
 953                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 954                         }
 955                 }
 956
 957                 io_u->index = i;
 958                 io_u->flags = IO_U_F_FREE;
 959                 flist_add(&io_u->list, &td->io_u_freelist);
 960                 p += max_bs;
 961         }
 962
 963         return 0;
 964 }
 965
 966 static int switch_ioscheduler(struct thread_data *td)
 967 {
 968         char tmp[256], tmp2[128];
 969         FILE *f;
 970         int ret;
 971
 972         if (td->io_ops->flags & FIO_DISKLESSIO)
 973                 return 0;
 974
 975         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 976
 977         f = fopen(tmp, "r+");
 978         if (!f) {
 979                 if (errno == ENOENT) {
 980                         log_err("fio: os or kernel doesn't support IO scheduler"
 981                                 " switching\n");
 982                         return 0;
 983                 }
 984                 td_verror(td, errno, "fopen iosched");
 985                 return 1;
 986         }
 987
 988         /*
 989          * Set io scheduler.
 990          */
 991         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 992         if (ferror(f) || ret != 1) {
 993                 td_verror(td, errno, "fwrite");
 994                 fclose(f);
 995                 return 1;
 996         }
 997
 998         rewind(f);
 999
1000         /*
1001          * Read back and check that the selected scheduler is now the default.
1002          */
1003         ret = fread(tmp, 1, sizeof(tmp), f);
1004         if (ferror(f) || ret < 0) {
1005                 td_verror(td, errno, "fread");
1006                 fclose(f);
1007                 return 1;
1008         }
1009
1010         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1011         if (!strstr(tmp, tmp2)) {
1012                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1013                 td_verror(td, EINVAL, "iosched_switch");
1014                 fclose(f);
1015                 return 1;
1016         }
1017
1018         fclose(f);
1019         return 0;
1020 }
1021
1022 static int keep_running(struct thread_data *td)
1023 {
1024         unsigned long long io_done;
1025
1026         if (td->done)
1027                 return 0;
1028         if (td->o.time_based)
1029                 return 1;
1030         if (td->o.loops) {
1031                 td->o.loops--;
1032                 return 1;
1033         }
1034
1035         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1036                         + td->io_skip_bytes;
1037         if (io_done < td->o.size)
1038                 return 1;
1039
1040         return 0;
1041 }
1042
1043 static void reset_io_counters(struct thread_data *td)
1044 {
1045         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1046         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1047         td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1048         td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1049         td->zone_bytes = 0;
1050         td->rate_bytes[0] = td->rate_bytes[1] = 0;
1051         td->rate_blocks[0] = td->rate_blocks[1] = 0;
1052
1053         td->last_was_sync = 0;
1054
1055         /*
1056          * reset file done count if we are to start over
1057          */
1058         if (td->o.time_based || td->o.loops)
1059                 td->nr_done_files = 0;
1060 }
1061
1062 void reset_all_stats(struct thread_data *td)
1063 {
1064         struct timeval tv;
1065         int i;
1066
1067         reset_io_counters(td);
1068
1069         for (i = 0; i < 2; i++) {
1070                 td->io_bytes[i] = 0;
1071                 td->io_blocks[i] = 0;
1072                 td->io_issues[i] = 0;
1073                 td->ts.total_io_u[i] = 0;
1074         }
1075
1076         fio_gettime(&tv, NULL);
1077         td->ts.runtime[0] = 0;
1078         td->ts.runtime[1] = 0;
1079         memcpy(&td->epoch, &tv, sizeof(tv));
1080         memcpy(&td->start, &tv, sizeof(tv));
1081 }
1082
1083 static void clear_io_state(struct thread_data *td)
1084 {
1085         struct fio_file *f;
1086         unsigned int i;
1087
1088         reset_io_counters(td);
1089
1090         close_files(td);
1091         for_each_file(td, f, i)
1092                 fio_file_clear_done(f);
1093
1094         /*
1095          * Set the same seed to get repeatable runs
1096          */
1097         td_fill_rand_seeds(td);
1098 }
1099
1100 static int exec_string(const char *string)
1101 {
1102         int ret, newlen = strlen(string) + 1 + 8;
1103         char *str;
1104
1105         str = malloc(newlen);
1106         sprintf(str, "sh -c %s", string);
1107
1108         ret = system(str);
1109         if (ret == -1)
1110                 log_err("fio: exec of cmd <%s> failed\n", str);
1111
1112         free(str);
1113         return ret;
1114 }
1115
1116 /*
1117  * Entry point for the thread based jobs. The process based jobs end up
1118  * here as well, after a little setup.
1119  */
1120 static void *thread_main(void *data)
1121 {
1122         unsigned long long elapsed;
1123         struct thread_data *td = data;
1124         pthread_condattr_t attr;
1125         int clear_state;
1126
1127         if (!td->o.use_thread) {
1128                 setsid();
1129                 td->pid = getpid();
1130         } else
1131                 td->pid = gettid();
1132
1133         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1134
1135         INIT_FLIST_HEAD(&td->io_u_freelist);
1136         INIT_FLIST_HEAD(&td->io_u_busylist);
1137         INIT_FLIST_HEAD(&td->io_u_requeues);
1138         INIT_FLIST_HEAD(&td->io_log_list);
1139         INIT_FLIST_HEAD(&td->io_hist_list);
1140         INIT_FLIST_HEAD(&td->verify_list);
1141         INIT_FLIST_HEAD(&td->trim_list);
1142         pthread_mutex_init(&td->io_u_lock, NULL);
1143         td->io_hist_tree = RB_ROOT;
1144
1145         pthread_condattr_init(&attr);
1146         pthread_cond_init(&td->verify_cond, &attr);
1147         pthread_cond_init(&td->free_cond, &attr);
1148
1149         td_set_runstate(td, TD_INITIALIZED);
1150         dprint(FD_MUTEX, "up startup_mutex\n");
1151         fio_mutex_up(startup_mutex);
1152         dprint(FD_MUTEX, "wait on td->mutex\n");
1153         fio_mutex_down(td->mutex);
1154         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1155
1156         /*
1157          * the ->mutex mutex is now no longer used, close it to avoid
1158          * eating a file descriptor
1159          */
1160         fio_mutex_remove(td->mutex);
1161
1162         /*
1163          * A new gid requires privilege, so we need to do this before setting
1164          * the uid.
1165          */
1166         if (td->o.gid != -1U && setgid(td->o.gid)) {
1167                 td_verror(td, errno, "setgid");
1168                 goto err;
1169         }
1170         if (td->o.uid != -1U && setuid(td->o.uid)) {
1171                 td_verror(td, errno, "setuid");
1172                 goto err;
1173         }
1174
1175         /*
1176          * If we have a gettimeofday() thread, make sure we exclude that
1177          * thread from this job
1178          */
1179         if (td->o.gtod_cpu)
1180                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1181
1182         /*
1183          * Set affinity first, in case it has an impact on the memory
1184          * allocations.
1185          */
1186         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1187                 td_verror(td, errno, "cpu_set_affinity");
1188                 goto err;
1189         }
1190
1191         /*
1192          * May alter parameters that init_io_u() will use, so we need to
1193          * do this first.
1194          */
1195         if (init_iolog(td))
1196                 goto err;
1197
1198         if (init_io_u(td))
1199                 goto err;
1200
1201         if (td->o.verify_async && verify_async_init(td))
1202                 goto err;
1203
1204         if (td->ioprio_set) {
1205                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1206                         td_verror(td, errno, "ioprio_set");
1207                         goto err;
1208                 }
1209         }
1210
1211         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1212                 goto err;
1213
1214         if (nice(td->o.nice) == -1) {
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         getrusage(RUSAGE_SELF, &td->ru_start);
1243
1244         clear_state = 0;
1245         while (keep_running(td)) {
1246                 fio_gettime(&td->start, NULL);
1247                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1248                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1249                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1250
1251                 if (td->o.ratemin[0] || td->o.ratemin[1])
1252                         memcpy(&td->lastrate, &td->bw_sample_time,
1253                                                         sizeof(td->lastrate));
1254
1255                 if (clear_state)
1256                         clear_io_state(td);
1257
1258                 prune_io_piece_log(td);
1259
1260                 do_io(td);
1261
1262                 clear_state = 1;
1263
1264                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1265                         elapsed = utime_since_now(&td->start);
1266                         td->ts.runtime[DDIR_READ] += elapsed;
1267                 }
1268                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1269                         elapsed = utime_since_now(&td->start);
1270                         td->ts.runtime[DDIR_WRITE] += elapsed;
1271                 }
1272
1273                 if (td->error || td->terminate)
1274                         break;
1275
1276                 if (!td->o.do_verify ||
1277                     td->o.verify == VERIFY_NONE ||
1278                     (td->io_ops->flags & FIO_UNIDIR))
1279                         continue;
1280
1281                 clear_io_state(td);
1282
1283                 fio_gettime(&td->start, NULL);
1284
1285                 do_verify(td);
1286
1287                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1288
1289                 if (td->error || td->terminate)
1290                         break;
1291         }
1292
1293         update_rusage_stat(td);
1294         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1295         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1296         td->ts.total_run_time = mtime_since_now(&td->epoch);
1297         td->ts.io_bytes[0] = td->io_bytes[0];
1298         td->ts.io_bytes[1] = td->io_bytes[1];
1299
1300         fio_mutex_down(writeout_mutex);
1301         if (td->bw_log) {
1302                 if (td->o.bw_log_file) {
1303                         finish_log_named(td, td->bw_log,
1304                                                 td->o.bw_log_file, "bw");
1305                 } else
1306                         finish_log(td, td->bw_log, "bw");
1307         }
1308         if (td->lat_log) {
1309                 if (td->o.lat_log_file) {
1310                         finish_log_named(td, td->lat_log,
1311                                                 td->o.lat_log_file, "lat");
1312                 } else
1313                         finish_log(td, td->lat_log, "lat");
1314         }
1315         if (td->slat_log) {
1316                 if (td->o.lat_log_file) {
1317                         finish_log_named(td, td->slat_log,
1318                                                 td->o.lat_log_file, "slat");
1319                 } else
1320                         finish_log(td, td->slat_log, "slat");
1321         }
1322         if (td->clat_log) {
1323                 if (td->o.lat_log_file) {
1324                         finish_log_named(td, td->clat_log,
1325                                                 td->o.lat_log_file, "clat");
1326                 } else
1327                         finish_log(td, td->clat_log, "clat");
1328         }
1329         if (td->iops_log) {
1330                 if (td->o.iops_log_file) {
1331                         finish_log_named(td, td->iops_log,
1332                                                 td->o.iops_log_file, "iops");
1333                 } else
1334                         finish_log(td, td->iops_log, "iops");
1335         }
1336
1337         fio_mutex_up(writeout_mutex);
1338         if (td->o.exec_postrun)
1339                 exec_string(td->o.exec_postrun);
1340
1341         if (exitall_on_terminate)
1342                 fio_terminate_threads(td->groupid);
1343
1344 err:
1345         if (td->error)
1346                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1347                                                         td->verror);
1348
1349         if (td->o.verify_async)
1350                 verify_async_exit(td);
1351
1352         close_and_free_files(td);
1353         close_ioengine(td);
1354         cleanup_io_u(td);
1355         cgroup_shutdown(td, &cgroup_mnt);
1356
1357         if (td->o.cpumask_set) {
1358                 int ret = fio_cpuset_exit(&td->o.cpumask);
1359
1360                 td_verror(td, ret, "fio_cpuset_exit");
1361         }
1362
1363         /*
1364          * do this very late, it will log file closing as well
1365          */
1366         if (td->o.write_iolog_file)
1367                 write_iolog_close(td);
1368
1369         td_set_runstate(td, TD_EXITED);
1370         return (void *) (unsigned long) td->error;
1371 }
1372
1373 /*
1374  * We cannot pass the td data into a forked process, so attach the td and
1375  * pass it to the thread worker.
1376  */
1377 static int fork_main(int shmid, int offset)
1378 {
1379         struct thread_data *td;
1380         void *data, *ret;
1381
1382 #ifndef __hpux
1383         data = shmat(shmid, NULL, 0);
1384         if (data == (void *) -1) {
1385                 int __err = errno;
1386
1387                 perror("shmat");
1388                 return __err;
1389         }
1390 #else
1391         /*
1392          * HP-UX inherits shm mappings?
1393          */
1394         data = threads;
1395 #endif
1396
1397         td = data + offset * sizeof(struct thread_data);
1398         ret = thread_main(td);
1399         shmdt(data);
1400         return (int) (unsigned long) ret;
1401 }
1402
1403 /*
1404  * Run over the job map and reap the threads that have exited, if any.
1405  */
1406 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1407                          unsigned int *m_rate)
1408 {
1409         struct thread_data *td;
1410         unsigned int cputhreads, realthreads, pending;
1411         int i, status, ret;
1412
1413         /*
1414          * reap exited threads (TD_EXITED -> TD_REAPED)
1415          */
1416         realthreads = pending = cputhreads = 0;
1417         for_each_td(td, i) {
1418                 int flags = 0;
1419
1420                 /*
1421                  * ->io_ops is NULL for a thread that has closed its
1422                  * io engine
1423                  */
1424                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1425                         cputhreads++;
1426                 else
1427                         realthreads++;
1428
1429                 if (!td->pid) {
1430                         pending++;
1431                         continue;
1432                 }
1433                 if (td->runstate == TD_REAPED)
1434                         continue;
1435                 if (td->o.use_thread) {
1436                         if (td->runstate == TD_EXITED) {
1437                                 td_set_runstate(td, TD_REAPED);
1438                                 goto reaped;
1439                         }
1440                         continue;
1441                 }
1442
1443                 flags = WNOHANG;
1444                 if (td->runstate == TD_EXITED)
1445                         flags = 0;
1446
1447                 /*
1448                  * check if someone quit or got killed in an unusual way
1449                  */
1450                 ret = waitpid(td->pid, &status, flags);
1451                 if (ret < 0) {
1452                         if (errno == ECHILD) {
1453                                 log_err("fio: pid=%d disappeared %d\n",
1454                                                 (int) td->pid, td->runstate);
1455                                 td_set_runstate(td, TD_REAPED);
1456                                 goto reaped;
1457                         }
1458                         perror("waitpid");
1459                 } else if (ret == td->pid) {
1460                         if (WIFSIGNALED(status)) {
1461                                 int sig = WTERMSIG(status);
1462
1463                                 if (sig != SIGTERM)
1464                                         log_err("fio: pid=%d, got signal=%d\n",
1465                                                         (int) td->pid, sig);
1466                                 td_set_runstate(td, TD_REAPED);
1467                                 goto reaped;
1468                         }
1469                         if (WIFEXITED(status)) {
1470                                 if (WEXITSTATUS(status) && !td->error)
1471                                         td->error = WEXITSTATUS(status);
1472
1473                                 td_set_runstate(td, TD_REAPED);
1474                                 goto reaped;
1475                         }
1476                 }
1477
1478                 /*
1479                  * thread is not dead, continue
1480                  */
1481                 pending++;
1482                 continue;
1483 reaped:
1484                 (*nr_running)--;
1485                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1486                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1487                 if (!td->pid)
1488                         pending--;
1489
1490                 if (td->error)
1491                         exit_value++;
1492
1493                 done_secs += mtime_since_now(&td->epoch) / 1000;
1494         }
1495
1496         if (*nr_running == cputhreads && !pending && realthreads)
1497                 fio_terminate_threads(TERMINATE_ALL);
1498 }
1499
1500 static void *gtod_thread_main(void *data)
1501 {
1502         fio_mutex_up(startup_mutex);
1503
1504         /*
1505          * As long as we have jobs around, update the clock. It would be nice
1506          * to have some way of NOT hammering that CPU with gettimeofday(),
1507          * but I'm not sure what to use outside of a simple CPU nop to relax
1508          * it - we don't want to lose precision.
1509          */
1510         while (threads) {
1511                 fio_gtod_update();
1512                 nop;
1513         }
1514
1515         return NULL;
1516 }
1517
1518 static int fio_start_gtod_thread(void)
1519 {
1520         pthread_attr_t attr;
1521         int ret;
1522
1523         pthread_attr_init(&attr);
1524         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1525         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1526         pthread_attr_destroy(&attr);
1527         if (ret) {
1528                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1529                 return 1;
1530         }
1531
1532         ret = pthread_detach(gtod_thread);
1533         if (ret) {
1534                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1535                 return 1;
1536         }
1537
1538         dprint(FD_MUTEX, "wait on startup_mutex\n");
1539         fio_mutex_down(startup_mutex);
1540         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1541         return 0;
1542 }
1543
1544 /*
1545  * Main function for kicking off and reaping jobs, as needed.
1546  */
1547 static void run_threads(void)
1548 {
1549         struct thread_data *td;
1550         unsigned long spent;
1551         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1552
1553         if (fio_pin_memory())
1554                 return;
1555
1556         if (fio_gtod_offload && fio_start_gtod_thread())
1557                 return;
1558
1559         set_sig_handlers();
1560
1561         if (!terse_output) {
1562                 log_info("Starting ");
1563                 if (nr_thread)
1564                         log_info("%d thread%s", nr_thread,
1565                                                 nr_thread > 1 ? "s" : "");
1566                 if (nr_process) {
1567                         if (nr_thread)
1568                                 log_info(" and ");
1569                         log_info("%d process%s", nr_process,
1570                                                 nr_process > 1 ? "es" : "");
1571                 }
1572                 log_info("\n");
1573                 fflush(stdout);
1574         }
1575
1576         todo = thread_number;
1577         nr_running = 0;
1578         nr_started = 0;
1579         m_rate = t_rate = 0;
1580
1581         for_each_td(td, i) {
1582                 print_status_init(td->thread_number - 1);
1583
1584                 if (!td->o.create_serialize)
1585                         continue;
1586
1587                 /*
1588                  * do file setup here so it happens sequentially,
1589                  * we don't want X number of threads getting their
1590                  * client data interspersed on disk
1591                  */
1592                 if (setup_files(td)) {
1593                         exit_value++;
1594                         if (td->error)
1595                                 log_err("fio: pid=%d, err=%d/%s\n",
1596                                         (int) td->pid, td->error, td->verror);
1597                         td_set_runstate(td, TD_REAPED);
1598                         todo--;
1599                 } else {
1600                         struct fio_file *f;
1601                         unsigned int j;
1602
1603                         /*
1604                          * for sharing to work, each job must always open
1605                          * its own files. so close them, if we opened them
1606                          * for creation
1607                          */
1608                         for_each_file(td, f, j) {
1609                                 if (fio_file_open(f))
1610                                         td_io_close_file(td, f);
1611                         }
1612                 }
1613         }
1614
1615         set_genesis_time();
1616
1617         while (todo) {
1618                 struct thread_data *map[REAL_MAX_JOBS];
1619                 struct timeval this_start;
1620                 int this_jobs = 0, left;
1621
1622                 /*
1623                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1624                  */
1625                 for_each_td(td, i) {
1626                         if (td->runstate != TD_NOT_CREATED)
1627                                 continue;
1628
1629                         /*
1630                          * never got a chance to start, killed by other
1631                          * thread for some reason
1632                          */
1633                         if (td->terminate) {
1634                                 todo--;
1635                                 continue;
1636                         }
1637
1638                         if (td->o.start_delay) {
1639                                 spent = mtime_since_genesis();
1640
1641                                 if (td->o.start_delay * 1000 > spent)
1642                                         continue;
1643                         }
1644
1645                         if (td->o.stonewall && (nr_started || nr_running)) {
1646                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1647                                                         td->o.name);
1648                                 break;
1649                         }
1650
1651                         init_disk_util(td);
1652
1653                         /*
1654                          * Set state to created. Thread will transition
1655                          * to TD_INITIALIZED when it's done setting up.
1656                          */
1657                         td_set_runstate(td, TD_CREATED);
1658                         map[this_jobs++] = td;
1659                         nr_started++;
1660
1661                         if (td->o.use_thread) {
1662                                 int ret;
1663
1664                                 dprint(FD_PROCESS, "will pthread_create\n");
1665                                 ret = pthread_create(&td->thread, NULL,
1666                                                         thread_main, td);
1667                                 if (ret) {
1668                                         log_err("pthread_create: %s\n",
1669                                                         strerror(ret));
1670                                         nr_started--;
1671                                         break;
1672                                 }
1673                                 ret = pthread_detach(td->thread);
1674                                 if (ret)
1675                                         log_err("pthread_detach: %s",
1676                                                         strerror(ret));
1677                         } else {
1678                                 pid_t pid;
1679                                 dprint(FD_PROCESS, "will fork\n");
1680                                 pid = fork();
1681                                 if (!pid) {
1682                                         int ret = fork_main(shm_id, i);
1683
1684                                         _exit(ret);
1685                                 } else if (i == fio_debug_jobno)
1686                                         *fio_debug_jobp = pid;
1687                         }
1688                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1689                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1690                                 log_err("fio: job startup hung? exiting.\n");
1691                                 fio_terminate_threads(TERMINATE_ALL);
1692                                 fio_abort = 1;
1693                                 nr_started--;
1694                                 break;
1695                         }
1696                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1697                 }
1698
1699                 /*
1700                  * Wait for the started threads to transition to
1701                  * TD_INITIALIZED.
1702                  */
1703                 fio_gettime(&this_start, NULL);
1704                 left = this_jobs;
1705                 while (left && !fio_abort) {
1706                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1707                                 break;
1708
1709                         usleep(100000);
1710
1711                         for (i = 0; i < this_jobs; i++) {
1712                                 td = map[i];
1713                                 if (!td)
1714                                         continue;
1715                                 if (td->runstate == TD_INITIALIZED) {
1716                                         map[i] = NULL;
1717                                         left--;
1718                                 } else if (td->runstate >= TD_EXITED) {
1719                                         map[i] = NULL;
1720                                         left--;
1721                                         todo--;
1722                                         nr_running++; /* work-around... */
1723                                 }
1724                         }
1725                 }
1726
1727                 if (left) {
1728                         log_err("fio: %d jobs failed to start\n", left);
1729                         for (i = 0; i < this_jobs; i++) {
1730                                 td = map[i];
1731                                 if (!td)
1732                                         continue;
1733                                 kill(td->pid, SIGTERM);
1734                         }
1735                         break;
1736                 }
1737
1738                 /*
1739                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1740                  */
1741                 for_each_td(td, i) {
1742                         if (td->runstate != TD_INITIALIZED)
1743                                 continue;
1744
1745                         if (in_ramp_time(td))
1746                                 td_set_runstate(td, TD_RAMP);
1747                         else
1748                                 td_set_runstate(td, TD_RUNNING);
1749                         nr_running++;
1750                         nr_started--;
1751                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1752                         t_rate += td->o.rate[0] + td->o.rate[1];
1753                         todo--;
1754                         fio_mutex_up(td->mutex);
1755                 }
1756
1757                 reap_threads(&nr_running, &t_rate, &m_rate);
1758
1759                 if (todo) {
1760                         if (is_backend)
1761                                 fio_server_idle_loop();
1762                         else
1763                                 usleep(100000);
1764                 }
1765         }
1766
1767         while (nr_running) {
1768                 reap_threads(&nr_running, &t_rate, &m_rate);
1769
1770                 if (is_backend)
1771                         fio_server_idle_loop();
1772                 else
1773                         usleep(10000);
1774         }
1775
1776         update_io_ticks();
1777         fio_unpin_memory();
1778 }
1779
1780 int exec_run(void)
1781 {
1782         struct thread_data *td;
1783         int i;
1784
1785         if (nr_clients)
1786                 return fio_handle_clients();
1787         if (exec_profile) {
1788                 if (load_profile(exec_profile))
1789                         return 1;
1790                 free(exec_profile);
1791                 exec_profile = NULL;
1792         }
1793         if (!thread_number)
1794                 return 0;
1795
1796         if (write_bw_log) {
1797                 setup_log(&agg_io_log[DDIR_READ]);
1798                 setup_log(&agg_io_log[DDIR_WRITE]);
1799         }
1800
1801         startup_mutex = fio_mutex_init(0);
1802         if (startup_mutex == NULL)
1803                 return 1;
1804         writeout_mutex = fio_mutex_init(1);
1805         if (writeout_mutex == NULL)
1806                 return 1;
1807
1808         set_genesis_time();
1809         create_disk_util_thread();
1810
1811         cgroup_list = smalloc(sizeof(*cgroup_list));
1812         INIT_FLIST_HEAD(cgroup_list);
1813
1814         run_threads();
1815
1816         if (!fio_abort) {
1817                 show_run_stats();
1818                 if (write_bw_log) {
1819                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1820                         __finish_log(agg_io_log[DDIR_WRITE],
1821                                         "agg-write_bw.log");
1822                 }
1823         }
1824
1825         for_each_td(td, i)
1826                 fio_options_free(td);
1827
1828         cgroup_kill(cgroup_list);
1829         sfree(cgroup_list);
1830         sfree(cgroup_mnt);
1831
1832         fio_mutex_remove(startup_mutex);
1833         fio_mutex_remove(writeout_mutex);
1834         return exit_value;
1835 }
1836
1837 void reset_fio_state(void)
1838 {
1839         groupid = 0;
1840         thread_number = 0;
1841         nr_process = 0;
1842         nr_thread = 0;
1843         done_secs = 0;
1844 }
1845
1846 static int endian_check(void)
1847 {
1848         union {
1849                 uint8_t c[8];
1850                 uint64_t v;
1851         } u;
1852         int le = 0, be = 0;
1853
1854         u.v = 0x12;
1855         if (u.c[7] == 0x12)
1856                 be = 1;
1857         else if (u.c[0] == 0x12)
1858                 le = 1;
1859
1860 #if defined(FIO_LITTLE_ENDIAN)
1861         if (be)
1862                 return 1;
1863 #elif defined(FIO_BIG_ENDIAN)
1864         if (le)
1865                 return 1;
1866 #else
1867         return 1;
1868 #endif
1869
1870         if (!le && !be)
1871                 return 1;
1872
1873         return 0;
1874 }
1875
1876 int main(int argc, char *argv[], char *envp[])
1877 {
1878         long ps;
1879
1880         if (endian_check()) {
1881                 log_err("fio: endianness settings appear wrong.\n");
1882                 log_err("fio: please report this to fio@vger.kernel.org\n");
1883                 return 1;
1884         }
1885
1886         arch_init(envp);
1887
1888         sinit();
1889
1890         /*
1891          * We need locale for number printing, if it isn't set then just
1892          * go with the US format.
1893          */
1894         if (!getenv("LC_NUMERIC"))
1895                 setlocale(LC_NUMERIC, "en_US");
1896
1897         ps = sysconf(_SC_PAGESIZE);
1898         if (ps < 0) {
1899                 log_err("Failed to get page size\n");
1900                 return 1;
1901         }
1902
1903         page_size = ps;
1904         page_mask = ps - 1;
1905
1906         fio_keywords_init();
1907
1908         if (parse_options(argc, argv))
1909                 return 1;
1910
1911         return exec_run();
1912 }