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 int thread_number = 0;
  59 int nr_process = 0;
  60 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, int *retptr)
 456 {
 457         int ret = *retptr;
 458
 459         if (ret < 0 || td->error) {
 460                 int err;
 461
 462                 if (!td->o.continue_on_error)
 463                         return 1;
 464
 465                 if (ret < 0)
 466                         err = -ret;
 467                 else
 468                         err = td->error;
 469
 470                 if (td_non_fatal_error(err)) {
 471                         /*
 472                          * Continue with the I/Os in case of
 473                          * a non fatal error.
 474                          */
 475                         update_error_count(td, err);
 476                         td_clear_error(td);
 477                         *retptr = 0;
 478                         return 0;
 479                 } else if (td->o.fill_device && err == ENOSPC) {
 480                         /*
 481                          * We expect to hit this error if
 482                          * fill_device option is set.
 483                          */
 484                         td_clear_error(td);
 485                         td->terminate = 1;
 486                         return 1;
 487                 } else {
 488                         /*
 489                          * Stop the I/O in case of a fatal
 490                          * error.
 491                          */
 492                         update_error_count(td, err);
 493                         return 1;
 494                 }
 495         }
 496
 497         return 0;
 498 }
 499
 500 /*
 501  * The main verify engine. Runs over the writes we previously submitted,
 502  * reads the blocks back in, and checks the crc/md5 of the data.
 503  */
 504 static void do_verify(struct thread_data *td)
 505 {
 506         struct fio_file *f;
 507         struct io_u *io_u;
 508         int ret, min_events;
 509         unsigned int i;
 510
 511         dprint(FD_VERIFY, "starting loop\n");
 512
 513         /*
 514          * sync io first and invalidate cache, to make sure we really
 515          * read from disk.
 516          */
 517         for_each_file(td, f, i) {
 518                 if (!fio_file_open(f))
 519                         continue;
 520                 if (fio_io_sync(td, f))
 521                         break;
 522                 if (file_invalidate_cache(td, f))
 523                         break;
 524         }
 525
 526         if (td->error)
 527                 return;
 528
 529         td_set_runstate(td, TD_VERIFYING);
 530
 531         io_u = NULL;
 532         while (!td->terminate) {
 533                 int ret2, full;
 534
 535                 update_tv_cache(td);
 536
 537                 if (runtime_exceeded(td, &td->tv_cache)) {
 538                         __update_tv_cache(td);
 539                         if (runtime_exceeded(td, &td->tv_cache)) {
 540                                 td->terminate = 1;
 541                                 break;
 542                         }
 543                 }
 544
 545                 io_u = __get_io_u(td);
 546                 if (!io_u)
 547                         break;
 548
 549                 if (get_next_verify(td, io_u)) {
 550                         put_io_u(td, io_u);
 551                         break;
 552                 }
 553
 554                 if (td_io_prep(td, io_u)) {
 555                         put_io_u(td, io_u);
 556                         break;
 557                 }
 558
 559                 if (td->o.verify_async)
 560                         io_u->end_io = verify_io_u_async;
 561                 else
 562                         io_u->end_io = verify_io_u;
 563
 564                 ret = td_io_queue(td, io_u);
 565                 switch (ret) {
 566                 case FIO_Q_COMPLETED:
 567                         if (io_u->error) {
 568                                 ret = -io_u->error;
 569                                 clear_io_u(td, io_u);
 570                         } else if (io_u->resid) {
 571                                 int bytes = io_u->xfer_buflen - io_u->resid;
 572
 573                                 /*
 574                                  * zero read, fail
 575                                  */
 576                                 if (!bytes) {
 577                                         td_verror(td, EIO, "full resid");
 578                                         put_io_u(td, io_u);
 579                                         break;
 580                                 }
 581
 582                                 io_u->xfer_buflen = io_u->resid;
 583                                 io_u->xfer_buf += bytes;
 584                                 io_u->offset += bytes;
 585
 586                                 if (ddir_rw(io_u->ddir))
 587                                         td->ts.short_io_u[io_u->ddir]++;
 588
 589                                 f = io_u->file;
 590                                 if (io_u->offset == f->real_file_size)
 591                                         goto sync_done;
 592
 593                                 requeue_io_u(td, &io_u);
 594                         } else {
 595 sync_done:
 596                                 ret = io_u_sync_complete(td, io_u, NULL);
 597                                 if (ret < 0)
 598                                         break;
 599                         }
 600                         continue;
 601                 case FIO_Q_QUEUED:
 602                         break;
 603                 case FIO_Q_BUSY:
 604                         requeue_io_u(td, &io_u);
 605                         ret2 = td_io_commit(td);
 606                         if (ret2 < 0)
 607                                 ret = ret2;
 608                         break;
 609                 default:
 610                         assert(ret < 0);
 611                         td_verror(td, -ret, "td_io_queue");
 612                         break;
 613                 }
 614
 615                 if (break_on_this_error(td, &ret))
 616                         break;
 617
 618                 /*
 619                  * if we can queue more, do so. but check if there are
 620                  * completed io_u's first. Note that we can get BUSY even
 621                  * without IO queued, if the system is resource starved.
 622                  */
 623                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 624                 if (full || !td->o.iodepth_batch_complete) {
 625                         min_events = min(td->o.iodepth_batch_complete,
 626                                          td->cur_depth);
 627                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 628                                 min_events = 1;
 629
 630                         do {
 631                                 /*
 632                                  * Reap required number of io units, if any,
 633                                  * and do the verification on them through
 634                                  * the callback handler
 635                                  */
 636                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 637                                         ret = -1;
 638                                         break;
 639                                 }
 640                         } while (full && (td->cur_depth > td->o.iodepth_low));
 641                 }
 642                 if (ret < 0)
 643                         break;
 644         }
 645
 646         if (!td->error) {
 647                 min_events = td->cur_depth;
 648
 649                 if (min_events)
 650                         ret = io_u_queued_complete(td, min_events, NULL);
 651         } else
 652                 cleanup_pending_aio(td);
 653
 654         td_set_runstate(td, TD_RUNNING);
 655
 656         dprint(FD_VERIFY, "exiting loop\n");
 657 }
 658
 659 /*
 660  * Main IO worker function. It retrieves io_u's to process and queues
 661  * and reaps them, checking for rate and errors along the way.
 662  */
 663 static void do_io(struct thread_data *td)
 664 {
 665         unsigned int i;
 666         int ret = 0;
 667
 668         if (in_ramp_time(td))
 669                 td_set_runstate(td, TD_RAMP);
 670         else
 671                 td_set_runstate(td, TD_RUNNING);
 672
 673         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 674                 (!flist_empty(&td->trim_list)) ||
 675                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 676                 struct timeval comp_time;
 677                 unsigned long bytes_done[2] = { 0, 0 };
 678                 int min_evts = 0;
 679                 struct io_u *io_u;
 680                 int ret2, full;
 681
 682                 if (td->terminate)
 683                         break;
 684
 685                 update_tv_cache(td);
 686
 687                 if (runtime_exceeded(td, &td->tv_cache)) {
 688                         __update_tv_cache(td);
 689                         if (runtime_exceeded(td, &td->tv_cache)) {
 690                                 td->terminate = 1;
 691                                 break;
 692                         }
 693                 }
 694
 695                 io_u = get_io_u(td);
 696                 if (!io_u)
 697                         break;
 698
 699                 /*
 700                  * Add verification end_io handler, if asked to verify
 701                  * a previously written file.
 702                  */
 703                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 704                     !td_rw(td)) {
 705                         if (td->o.verify_async)
 706                                 io_u->end_io = verify_io_u_async;
 707                         else
 708                                 io_u->end_io = verify_io_u;
 709                         td_set_runstate(td, TD_VERIFYING);
 710                 } else if (in_ramp_time(td))
 711                         td_set_runstate(td, TD_RAMP);
 712                 else
 713                         td_set_runstate(td, TD_RUNNING);
 714
 715                 ret = td_io_queue(td, io_u);
 716                 switch (ret) {
 717                 case FIO_Q_COMPLETED:
 718                         if (io_u->error) {
 719                                 ret = -io_u->error;
 720                                 clear_io_u(td, io_u);
 721                         } else if (io_u->resid) {
 722                                 int bytes = io_u->xfer_buflen - io_u->resid;
 723                                 struct fio_file *f = io_u->file;
 724
 725                                 /*
 726                                  * zero read, fail
 727                                  */
 728                                 if (!bytes) {
 729                                         td_verror(td, EIO, "full resid");
 730                                         put_io_u(td, io_u);
 731                                         break;
 732                                 }
 733
 734                                 io_u->xfer_buflen = io_u->resid;
 735                                 io_u->xfer_buf += bytes;
 736                                 io_u->offset += bytes;
 737
 738                                 if (ddir_rw(io_u->ddir))
 739                                         td->ts.short_io_u[io_u->ddir]++;
 740
 741                                 if (io_u->offset == f->real_file_size)
 742                                         goto sync_done;
 743
 744                                 requeue_io_u(td, &io_u);
 745                         } else {
 746 sync_done:
 747                                 if (__should_check_rate(td, 0) ||
 748                                     __should_check_rate(td, 1))
 749                                         fio_gettime(&comp_time, NULL);
 750
 751                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 752                                 if (ret < 0)
 753                                         break;
 754                         }
 755                         break;
 756                 case FIO_Q_QUEUED:
 757                         /*
 758                          * if the engine doesn't have a commit hook,
 759                          * the io_u is really queued. if it does have such
 760                          * a hook, it has to call io_u_queued() itself.
 761                          */
 762                         if (td->io_ops->commit == NULL)
 763                                 io_u_queued(td, io_u);
 764                         break;
 765                 case FIO_Q_BUSY:
 766                         requeue_io_u(td, &io_u);
 767                         ret2 = td_io_commit(td);
 768                         if (ret2 < 0)
 769                                 ret = ret2;
 770                         break;
 771                 default:
 772                         assert(ret < 0);
 773                         put_io_u(td, io_u);
 774                         break;
 775                 }
 776
 777                 if (break_on_this_error(td, &ret))
 778                         break;
 779
 780                 /*
 781                  * See if we need to complete some commands. Note that we
 782                  * can get BUSY even without IO queued, if the system is
 783                  * resource starved.
 784                  */
 785                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 786                 if (full || !td->o.iodepth_batch_complete) {
 787                         min_evts = min(td->o.iodepth_batch_complete,
 788                                         td->cur_depth);
 789                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 790                                 min_evts = 1;
 791
 792                         if (__should_check_rate(td, 0) ||
 793                             __should_check_rate(td, 1))
 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 (!(bytes_done[0] + bytes_done[1]))
 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 = td->io_blocks[0] + td->io_blocks[1];
 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         if (td->trim_entries)
 836                 printf("trim entries %ld\n", td->trim_entries);
 837
 838         if (td->o.fill_device && td->error == ENOSPC) {
 839                 td->error = 0;
 840                 td->terminate = 1;
 841         }
 842         if (!td->error) {
 843                 struct fio_file *f;
 844
 845                 i = td->cur_depth;
 846                 if (i) {
 847                         ret = io_u_queued_complete(td, i, NULL);
 848                         if (td->o.fill_device && td->error == ENOSPC)
 849                                 td->error = 0;
 850                 }
 851
 852                 if (should_fsync(td) && td->o.end_fsync) {
 853                         td_set_runstate(td, TD_FSYNCING);
 854
 855                         for_each_file(td, f, i) {
 856                                 if (!fio_file_open(f))
 857                                         continue;
 858                                 fio_io_sync(td, f);
 859                         }
 860                 }
 861         } else
 862                 cleanup_pending_aio(td);
 863
 864         /*
 865          * stop job if we failed doing any IO
 866          */
 867         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 868                 td->done = 1;
 869 }
 870
 871 static void cleanup_io_u(struct thread_data *td)
 872 {
 873         struct flist_head *entry, *n;
 874         struct io_u *io_u;
 875
 876         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 877                 io_u = flist_entry(entry, struct io_u, list);
 878
 879                 flist_del(&io_u->list);
 880                 fio_memfree(io_u, sizeof(*io_u));
 881         }
 882
 883         free_io_mem(td);
 884 }
 885
 886 static int init_io_u(struct thread_data *td)
 887 {
 888         struct io_u *io_u;
 889         unsigned int max_bs;
 890         int cl_align, i, max_units;
 891         char *p;
 892
 893         max_units = td->o.iodepth;
 894         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 895         td->orig_buffer_size = (unsigned long long) max_bs
 896                                         * (unsigned long long) max_units;
 897
 898         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 899                 unsigned long bs;
 900
 901                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 902                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 903         }
 904
 905         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 906                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 907                 return 1;
 908         }
 909
 910         if (allocate_io_mem(td))
 911                 return 1;
 912
 913         if (td->o.odirect || td->o.mem_align ||
 914             (td->io_ops->flags & FIO_RAWIO))
 915                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 916         else
 917                 p = td->orig_buffer;
 918
 919         cl_align = os_cache_line_size();
 920
 921         for (i = 0; i < max_units; i++) {
 922                 void *ptr;
 923
 924                 if (td->terminate)
 925                         return 1;
 926
 927                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 928                 if (!ptr) {
 929                         log_err("fio: unable to allocate aligned memory\n");
 930                         break;
 931                 }
 932
 933                 io_u = ptr;
 934                 memset(io_u, 0, sizeof(*io_u));
 935                 INIT_FLIST_HEAD(&io_u->list);
 936                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 937
 938                 if (!(td->io_ops->flags & FIO_NOIO)) {
 939                         io_u->buf = p + max_bs * i;
 940                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 941
 942                         if (td_write(td))
 943                                 io_u_fill_buffer(td, io_u, max_bs);
 944                         if (td_write(td) && td->o.verify_pattern_bytes) {
 945                                 /*
 946                                  * Fill the buffer with the pattern if we are
 947                                  * going to be doing writes.
 948                                  */
 949                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 950                         }
 951                 }
 952
 953                 io_u->index = i;
 954                 io_u->flags = IO_U_F_FREE;
 955                 flist_add(&io_u->list, &td->io_u_freelist);
 956         }
 957
 958         return 0;
 959 }
 960
 961 static int switch_ioscheduler(struct thread_data *td)
 962 {
 963         char tmp[256], tmp2[128];
 964         FILE *f;
 965         int ret;
 966
 967         if (td->io_ops->flags & FIO_DISKLESSIO)
 968                 return 0;
 969
 970         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 971
 972         f = fopen(tmp, "r+");
 973         if (!f) {
 974                 if (errno == ENOENT) {
 975                         log_err("fio: os or kernel doesn't support IO scheduler"
 976                                 " switching\n");
 977                         return 0;
 978                 }
 979                 td_verror(td, errno, "fopen iosched");
 980                 return 1;
 981         }
 982
 983         /*
 984          * Set io scheduler.
 985          */
 986         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 987         if (ferror(f) || ret != 1) {
 988                 td_verror(td, errno, "fwrite");
 989                 fclose(f);
 990                 return 1;
 991         }
 992
 993         rewind(f);
 994
 995         /*
 996          * Read back and check that the selected scheduler is now the default.
 997          */
 998         ret = fread(tmp, 1, sizeof(tmp), f);
 999         if (ferror(f) || ret < 0) {
1000                 td_verror(td, errno, "fread");
1001                 fclose(f);
1002                 return 1;
1003         }
1004
1005         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1006         if (!strstr(tmp, tmp2)) {
1007                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1008                 td_verror(td, EINVAL, "iosched_switch");
1009                 fclose(f);
1010                 return 1;
1011         }
1012
1013         fclose(f);
1014         return 0;
1015 }
1016
1017 static int keep_running(struct thread_data *td)
1018 {
1019         unsigned long long io_done;
1020
1021         if (td->done)
1022                 return 0;
1023         if (td->o.time_based)
1024                 return 1;
1025         if (td->o.loops) {
1026                 td->o.loops--;
1027                 return 1;
1028         }
1029
1030         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1031                         + td->io_skip_bytes;
1032         if (io_done < td->o.size)
1033                 return 1;
1034
1035         return 0;
1036 }
1037
1038 static void reset_io_counters(struct thread_data *td)
1039 {
1040         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1041         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1042         td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1043         td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1044         td->zone_bytes = 0;
1045         td->rate_bytes[0] = td->rate_bytes[1] = 0;
1046         td->rate_blocks[0] = td->rate_blocks[1] = 0;
1047
1048         td->last_was_sync = 0;
1049
1050         /*
1051          * reset file done count if we are to start over
1052          */
1053         if (td->o.time_based || td->o.loops)
1054                 td->nr_done_files = 0;
1055 }
1056
1057 void reset_all_stats(struct thread_data *td)
1058 {
1059         struct timeval tv;
1060         int i;
1061
1062         reset_io_counters(td);
1063
1064         for (i = 0; i < 2; i++) {
1065                 td->io_bytes[i] = 0;
1066                 td->io_blocks[i] = 0;
1067                 td->io_issues[i] = 0;
1068                 td->ts.total_io_u[i] = 0;
1069         }
1070
1071         fio_gettime(&tv, NULL);
1072         td->ts.runtime[0] = 0;
1073         td->ts.runtime[1] = 0;
1074         memcpy(&td->epoch, &tv, sizeof(tv));
1075         memcpy(&td->start, &tv, sizeof(tv));
1076 }
1077
1078 static void clear_io_state(struct thread_data *td)
1079 {
1080         struct fio_file *f;
1081         unsigned int i;
1082
1083         reset_io_counters(td);
1084
1085         close_files(td);
1086         for_each_file(td, f, i)
1087                 fio_file_clear_done(f);
1088
1089         /*
1090          * Set the same seed to get repeatable runs
1091          */
1092         td_fill_rand_seeds(td);
1093 }
1094
1095 static int exec_string(const char *string)
1096 {
1097         int ret, newlen = strlen(string) + 1 + 8;
1098         char *str;
1099
1100         str = malloc(newlen);
1101         sprintf(str, "sh -c %s", string);
1102
1103         ret = system(str);
1104         if (ret == -1)
1105                 log_err("fio: exec of cmd <%s> failed\n", str);
1106
1107         free(str);
1108         return ret;
1109 }
1110
1111 /*
1112  * Entry point for the thread based jobs. The process based jobs end up
1113  * here as well, after a little setup.
1114  */
1115 static void *thread_main(void *data)
1116 {
1117         unsigned long long elapsed;
1118         struct thread_data *td = data;
1119         pthread_condattr_t attr;
1120         int clear_state;
1121
1122         if (!td->o.use_thread) {
1123                 setsid();
1124                 td->pid = getpid();
1125         } else
1126                 td->pid = gettid();
1127
1128         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1129
1130         INIT_FLIST_HEAD(&td->io_u_freelist);
1131         INIT_FLIST_HEAD(&td->io_u_busylist);
1132         INIT_FLIST_HEAD(&td->io_u_requeues);
1133         INIT_FLIST_HEAD(&td->io_log_list);
1134         INIT_FLIST_HEAD(&td->io_hist_list);
1135         INIT_FLIST_HEAD(&td->verify_list);
1136         INIT_FLIST_HEAD(&td->trim_list);
1137         pthread_mutex_init(&td->io_u_lock, NULL);
1138         td->io_hist_tree = RB_ROOT;
1139
1140         pthread_condattr_init(&attr);
1141         pthread_cond_init(&td->verify_cond, &attr);
1142         pthread_cond_init(&td->free_cond, &attr);
1143
1144         td_set_runstate(td, TD_INITIALIZED);
1145         dprint(FD_MUTEX, "up startup_mutex\n");
1146         fio_mutex_up(startup_mutex);
1147         dprint(FD_MUTEX, "wait on td->mutex\n");
1148         fio_mutex_down(td->mutex);
1149         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1150
1151         /*
1152          * the ->mutex mutex is now no longer used, close it to avoid
1153          * eating a file descriptor
1154          */
1155         fio_mutex_remove(td->mutex);
1156
1157         /*
1158          * A new gid requires privilege, so we need to do this before setting
1159          * the uid.
1160          */
1161         if (td->o.gid != -1U && setgid(td->o.gid)) {
1162                 td_verror(td, errno, "setgid");
1163                 goto err;
1164         }
1165         if (td->o.uid != -1U && setuid(td->o.uid)) {
1166                 td_verror(td, errno, "setuid");
1167                 goto err;
1168         }
1169
1170         /*
1171          * If we have a gettimeofday() thread, make sure we exclude that
1172          * thread from this job
1173          */
1174         if (td->o.gtod_cpu)
1175                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1176
1177         /*
1178          * Set affinity first, in case it has an impact on the memory
1179          * allocations.
1180          */
1181         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1182                 td_verror(td, errno, "cpu_set_affinity");
1183                 goto err;
1184         }
1185
1186         /*
1187          * May alter parameters that init_io_u() will use, so we need to
1188          * do this first.
1189          */
1190         if (init_iolog(td))
1191                 goto err;
1192
1193         if (init_io_u(td))
1194                 goto err;
1195
1196         if (td->o.verify_async && verify_async_init(td))
1197                 goto err;
1198
1199         if (td->ioprio_set) {
1200                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1201                         td_verror(td, errno, "ioprio_set");
1202                         goto err;
1203                 }
1204         }
1205
1206         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1207                 goto err;
1208
1209         if (nice(td->o.nice) == -1) {
1210                 td_verror(td, errno, "nice");
1211                 goto err;
1212         }
1213
1214         if (td->o.ioscheduler && switch_ioscheduler(td))
1215                 goto err;
1216
1217         if (!td->o.create_serialize && setup_files(td))
1218                 goto err;
1219
1220         if (td_io_init(td))
1221                 goto err;
1222
1223         if (init_random_map(td))
1224                 goto err;
1225
1226         if (td->o.exec_prerun) {
1227                 if (exec_string(td->o.exec_prerun))
1228                         goto err;
1229         }
1230
1231         if (td->o.pre_read) {
1232                 if (pre_read_files(td) < 0)
1233                         goto err;
1234         }
1235
1236         fio_gettime(&td->epoch, NULL);
1237         getrusage(RUSAGE_SELF, &td->ru_start);
1238
1239         clear_state = 0;
1240         while (keep_running(td)) {
1241                 fio_gettime(&td->start, NULL);
1242                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1243                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1244                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1245
1246                 if (td->o.ratemin[0] || td->o.ratemin[1])
1247                         memcpy(&td->lastrate, &td->bw_sample_time,
1248                                                         sizeof(td->lastrate));
1249
1250                 if (clear_state)
1251                         clear_io_state(td);
1252
1253                 prune_io_piece_log(td);
1254
1255                 do_io(td);
1256
1257                 clear_state = 1;
1258
1259                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1260                         elapsed = utime_since_now(&td->start);
1261                         td->ts.runtime[DDIR_READ] += elapsed;
1262                 }
1263                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1264                         elapsed = utime_since_now(&td->start);
1265                         td->ts.runtime[DDIR_WRITE] += elapsed;
1266                 }
1267
1268                 if (td->error || td->terminate)
1269                         break;
1270
1271                 if (!td->o.do_verify ||
1272                     td->o.verify == VERIFY_NONE ||
1273                     (td->io_ops->flags & FIO_UNIDIR))
1274                         continue;
1275
1276                 clear_io_state(td);
1277
1278                 fio_gettime(&td->start, NULL);
1279
1280                 do_verify(td);
1281
1282                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1283
1284                 if (td->error || td->terminate)
1285                         break;
1286         }
1287
1288         update_rusage_stat(td);
1289         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1290         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1291         td->ts.total_run_time = mtime_since_now(&td->epoch);
1292         td->ts.io_bytes[0] = td->io_bytes[0];
1293         td->ts.io_bytes[1] = td->io_bytes[1];
1294
1295         fio_mutex_down(writeout_mutex);
1296         if (td->bw_log) {
1297                 if (td->o.bw_log_file) {
1298                         finish_log_named(td, td->bw_log,
1299                                                 td->o.bw_log_file, "bw");
1300                 } else
1301                         finish_log(td, td->bw_log, "bw");
1302         }
1303         if (td->lat_log) {
1304                 if (td->o.lat_log_file) {
1305                         finish_log_named(td, td->lat_log,
1306                                                 td->o.lat_log_file, "lat");
1307                 } else
1308                         finish_log(td, td->lat_log, "lat");
1309         }
1310         if (td->slat_log) {
1311                 if (td->o.lat_log_file) {
1312                         finish_log_named(td, td->slat_log,
1313                                                 td->o.lat_log_file, "slat");
1314                 } else
1315                         finish_log(td, td->slat_log, "slat");
1316         }
1317         if (td->clat_log) {
1318                 if (td->o.lat_log_file) {
1319                         finish_log_named(td, td->clat_log,
1320                                                 td->o.lat_log_file, "clat");
1321                 } else
1322                         finish_log(td, td->clat_log, "clat");
1323         }
1324         if (td->iops_log) {
1325                 if (td->o.iops_log_file) {
1326                         finish_log_named(td, td->iops_log,
1327                                                 td->o.iops_log_file, "iops");
1328                 } else
1329                         finish_log(td, td->iops_log, "iops");
1330         }
1331
1332         fio_mutex_up(writeout_mutex);
1333         if (td->o.exec_postrun)
1334                 exec_string(td->o.exec_postrun);
1335
1336         if (exitall_on_terminate)
1337                 fio_terminate_threads(td->groupid);
1338
1339 err:
1340         if (td->error)
1341                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1342                                                         td->verror);
1343
1344         if (td->o.verify_async)
1345                 verify_async_exit(td);
1346
1347         close_and_free_files(td);
1348         close_ioengine(td);
1349         cleanup_io_u(td);
1350         cgroup_shutdown(td, &cgroup_mnt);
1351
1352         if (td->o.cpumask_set) {
1353                 int ret = fio_cpuset_exit(&td->o.cpumask);
1354
1355                 td_verror(td, ret, "fio_cpuset_exit");
1356         }
1357
1358         /*
1359          * do this very late, it will log file closing as well
1360          */
1361         if (td->o.write_iolog_file)
1362                 write_iolog_close(td);
1363
1364         td_set_runstate(td, TD_EXITED);
1365         return (void *) (unsigned long) td->error;
1366 }
1367
1368 /*
1369  * We cannot pass the td data into a forked process, so attach the td and
1370  * pass it to the thread worker.
1371  */
1372 static int fork_main(int shmid, int offset)
1373 {
1374         struct thread_data *td;
1375         void *data, *ret;
1376
1377 #ifndef __hpux
1378         data = shmat(shmid, NULL, 0);
1379         if (data == (void *) -1) {
1380                 int __err = errno;
1381
1382                 perror("shmat");
1383                 return __err;
1384         }
1385 #else
1386         /*
1387          * HP-UX inherits shm mappings?
1388          */
1389         data = threads;
1390 #endif
1391
1392         td = data + offset * sizeof(struct thread_data);
1393         ret = thread_main(td);
1394         shmdt(data);
1395         return (int) (unsigned long) ret;
1396 }
1397
1398 /*
1399  * Run over the job map and reap the threads that have exited, if any.
1400  */
1401 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1402 {
1403         struct thread_data *td;
1404         int i, cputhreads, realthreads, pending, status, ret;
1405
1406         /*
1407          * reap exited threads (TD_EXITED -> TD_REAPED)
1408          */
1409         realthreads = pending = cputhreads = 0;
1410         for_each_td(td, i) {
1411                 int flags = 0;
1412
1413                 /*
1414                  * ->io_ops is NULL for a thread that has closed its
1415                  * io engine
1416                  */
1417                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1418                         cputhreads++;
1419                 else
1420                         realthreads++;
1421
1422                 if (!td->pid) {
1423                         pending++;
1424                         continue;
1425                 }
1426                 if (td->runstate == TD_REAPED)
1427                         continue;
1428                 if (td->o.use_thread) {
1429                         if (td->runstate == TD_EXITED) {
1430                                 td_set_runstate(td, TD_REAPED);
1431                                 goto reaped;
1432                         }
1433                         continue;
1434                 }
1435
1436                 flags = WNOHANG;
1437                 if (td->runstate == TD_EXITED)
1438                         flags = 0;
1439
1440                 /*
1441                  * check if someone quit or got killed in an unusual way
1442                  */
1443                 ret = waitpid(td->pid, &status, flags);
1444                 if (ret < 0) {
1445                         if (errno == ECHILD) {
1446                                 log_err("fio: pid=%d disappeared %d\n",
1447                                                 (int) td->pid, td->runstate);
1448                                 td_set_runstate(td, TD_REAPED);
1449                                 goto reaped;
1450                         }
1451                         perror("waitpid");
1452                 } else if (ret == td->pid) {
1453                         if (WIFSIGNALED(status)) {
1454                                 int sig = WTERMSIG(status);
1455
1456                                 if (sig != SIGTERM)
1457                                         log_err("fio: pid=%d, got signal=%d\n",
1458                                                         (int) td->pid, sig);
1459                                 td_set_runstate(td, TD_REAPED);
1460                                 goto reaped;
1461                         }
1462                         if (WIFEXITED(status)) {
1463                                 if (WEXITSTATUS(status) && !td->error)
1464                                         td->error = WEXITSTATUS(status);
1465
1466                                 td_set_runstate(td, TD_REAPED);
1467                                 goto reaped;
1468                         }
1469                 }
1470
1471                 /*
1472                  * thread is not dead, continue
1473                  */
1474                 pending++;
1475                 continue;
1476 reaped:
1477                 (*nr_running)--;
1478                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1479                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1480                 if (!td->pid)
1481                         pending--;
1482
1483                 if (td->error)
1484                         exit_value++;
1485
1486                 done_secs += mtime_since_now(&td->epoch) / 1000;
1487         }
1488
1489         if (*nr_running == cputhreads && !pending && realthreads)
1490                 fio_terminate_threads(TERMINATE_ALL);
1491 }
1492
1493 static void *gtod_thread_main(void *data)
1494 {
1495         fio_mutex_up(startup_mutex);
1496
1497         /*
1498          * As long as we have jobs around, update the clock. It would be nice
1499          * to have some way of NOT hammering that CPU with gettimeofday(),
1500          * but I'm not sure what to use outside of a simple CPU nop to relax
1501          * it - we don't want to lose precision.
1502          */
1503         while (threads) {
1504                 fio_gtod_update();
1505                 nop;
1506         }
1507
1508         return NULL;
1509 }
1510
1511 static int fio_start_gtod_thread(void)
1512 {
1513         pthread_attr_t attr;
1514         int ret;
1515
1516         pthread_attr_init(&attr);
1517         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1518         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1519         pthread_attr_destroy(&attr);
1520         if (ret) {
1521                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1522                 return 1;
1523         }
1524
1525         ret = pthread_detach(gtod_thread);
1526         if (ret) {
1527                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1528                 return 1;
1529         }
1530
1531         dprint(FD_MUTEX, "wait on startup_mutex\n");
1532         fio_mutex_down(startup_mutex);
1533         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1534         return 0;
1535 }
1536
1537 /*
1538  * Main function for kicking off and reaping jobs, as needed.
1539  */
1540 static void run_threads(void)
1541 {
1542         struct thread_data *td;
1543         unsigned long spent;
1544         int i, todo, nr_running, m_rate, t_rate, nr_started;
1545
1546         if (fio_pin_memory())
1547                 return;
1548
1549         if (fio_gtod_offload && fio_start_gtod_thread())
1550                 return;
1551
1552         set_sig_handlers();
1553
1554         if (!terse_output) {
1555                 log_info("Starting ");
1556                 if (nr_thread)
1557                         log_info("%d thread%s", nr_thread,
1558                                                 nr_thread > 1 ? "s" : "");
1559                 if (nr_process) {
1560                         if (nr_thread)
1561                                 printf(" and ");
1562                         log_info("%d process%s", nr_process,
1563                                                 nr_process > 1 ? "es" : "");
1564                 }
1565                 log_info("\n");
1566                 fflush(stdout);
1567         }
1568
1569         todo = thread_number;
1570         nr_running = 0;
1571         nr_started = 0;
1572         m_rate = t_rate = 0;
1573
1574         for_each_td(td, i) {
1575                 print_status_init(td->thread_number - 1);
1576
1577                 if (!td->o.create_serialize)
1578                         continue;
1579
1580                 /*
1581                  * do file setup here so it happens sequentially,
1582                  * we don't want X number of threads getting their
1583                  * client data interspersed on disk
1584                  */
1585                 if (setup_files(td)) {
1586                         exit_value++;
1587                         if (td->error)
1588                                 log_err("fio: pid=%d, err=%d/%s\n",
1589                                         (int) td->pid, td->error, td->verror);
1590                         td_set_runstate(td, TD_REAPED);
1591                         todo--;
1592                 } else {
1593                         struct fio_file *f;
1594                         unsigned int j;
1595
1596                         /*
1597                          * for sharing to work, each job must always open
1598                          * its own files. so close them, if we opened them
1599                          * for creation
1600                          */
1601                         for_each_file(td, f, j) {
1602                                 if (fio_file_open(f))
1603                                         td_io_close_file(td, f);
1604                         }
1605                 }
1606         }
1607
1608         set_genesis_time();
1609
1610         while (todo) {
1611                 struct thread_data *map[REAL_MAX_JOBS];
1612                 struct timeval this_start;
1613                 int this_jobs = 0, left;
1614
1615                 /*
1616                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1617                  */
1618                 for_each_td(td, i) {
1619                         if (td->runstate != TD_NOT_CREATED)
1620                                 continue;
1621
1622                         /*
1623                          * never got a chance to start, killed by other
1624                          * thread for some reason
1625                          */
1626                         if (td->terminate) {
1627                                 todo--;
1628                                 continue;
1629                         }
1630
1631                         if (td->o.start_delay) {
1632                                 spent = mtime_since_genesis();
1633
1634                                 if (td->o.start_delay * 1000 > spent)
1635                                         continue;
1636                         }
1637
1638                         if (td->o.stonewall && (nr_started || nr_running)) {
1639                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1640                                                         td->o.name);
1641                                 break;
1642                         }
1643
1644                         init_disk_util(td);
1645
1646                         /*
1647                          * Set state to created. Thread will transition
1648                          * to TD_INITIALIZED when it's done setting up.
1649                          */
1650                         td_set_runstate(td, TD_CREATED);
1651                         map[this_jobs++] = td;
1652                         nr_started++;
1653
1654                         if (td->o.use_thread) {
1655                                 int ret;
1656
1657                                 dprint(FD_PROCESS, "will pthread_create\n");
1658                                 ret = pthread_create(&td->thread, NULL,
1659                                                         thread_main, td);
1660                                 if (ret) {
1661                                         log_err("pthread_create: %s\n",
1662                                                         strerror(ret));
1663                                         nr_started--;
1664                                         break;
1665                                 }
1666                                 ret = pthread_detach(td->thread);
1667                                 if (ret)
1668                                         log_err("pthread_detach: %s",
1669                                                         strerror(ret));
1670                         } else {
1671                                 pid_t pid;
1672                                 dprint(FD_PROCESS, "will fork\n");
1673                                 pid = fork();
1674                                 if (!pid) {
1675                                         int ret = fork_main(shm_id, i);
1676
1677                                         _exit(ret);
1678                                 } else if (i == fio_debug_jobno)
1679                                         *fio_debug_jobp = pid;
1680                         }
1681                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1682                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1683                                 log_err("fio: job startup hung? exiting.\n");
1684                                 fio_terminate_threads(TERMINATE_ALL);
1685                                 fio_abort = 1;
1686                                 nr_started--;
1687                                 break;
1688                         }
1689                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1690                 }
1691
1692                 /*
1693                  * Wait for the started threads to transition to
1694                  * TD_INITIALIZED.
1695                  */
1696                 fio_gettime(&this_start, NULL);
1697                 left = this_jobs;
1698                 while (left && !fio_abort) {
1699                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1700                                 break;
1701
1702                         usleep(100000);
1703
1704                         for (i = 0; i < this_jobs; i++) {
1705                                 td = map[i];
1706                                 if (!td)
1707                                         continue;
1708                                 if (td->runstate == TD_INITIALIZED) {
1709                                         map[i] = NULL;
1710                                         left--;
1711                                 } else if (td->runstate >= TD_EXITED) {
1712                                         map[i] = NULL;
1713                                         left--;
1714                                         todo--;
1715                                         nr_running++; /* work-around... */
1716                                 }
1717                         }
1718                 }
1719
1720                 if (left) {
1721                         log_err("fio: %d jobs failed to start\n", left);
1722                         for (i = 0; i < this_jobs; i++) {
1723                                 td = map[i];
1724                                 if (!td)
1725                                         continue;
1726                                 kill(td->pid, SIGTERM);
1727                         }
1728                         break;
1729                 }
1730
1731                 /*
1732                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1733                  */
1734                 for_each_td(td, i) {
1735                         if (td->runstate != TD_INITIALIZED)
1736                                 continue;
1737
1738                         if (in_ramp_time(td))
1739                                 td_set_runstate(td, TD_RAMP);
1740                         else
1741                                 td_set_runstate(td, TD_RUNNING);
1742                         nr_running++;
1743                         nr_started--;
1744                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1745                         t_rate += td->o.rate[0] + td->o.rate[1];
1746                         todo--;
1747                         fio_mutex_up(td->mutex);
1748                 }
1749
1750                 reap_threads(&nr_running, &t_rate, &m_rate);
1751
1752                 if (todo) {
1753                         if (is_backend)
1754                                 fio_server_idle_loop();
1755                         else
1756                                 usleep(100000);
1757                 }
1758         }
1759
1760         while (nr_running) {
1761                 reap_threads(&nr_running, &t_rate, &m_rate);
1762
1763                 if (is_backend)
1764                         fio_server_idle_loop();
1765                 else
1766                         usleep(10000);
1767         }
1768
1769         update_io_ticks();
1770         fio_unpin_memory();
1771 }
1772
1773 int exec_run(void)
1774 {
1775         struct thread_data *td;
1776         int i;
1777
1778         if (nr_clients)
1779                 return fio_handle_clients();
1780         if (exec_profile) {
1781                 if (load_profile(exec_profile))
1782                         return 1;
1783                 free(exec_profile);
1784                 exec_profile = NULL;
1785         }
1786         if (!thread_number)
1787                 return 0;
1788
1789         if (write_bw_log) {
1790                 setup_log(&agg_io_log[DDIR_READ]);
1791                 setup_log(&agg_io_log[DDIR_WRITE]);
1792         }
1793
1794         startup_mutex = fio_mutex_init(0);
1795         if (startup_mutex == NULL)
1796                 return 1;
1797         writeout_mutex = fio_mutex_init(1);
1798         if (writeout_mutex == NULL)
1799                 return 1;
1800
1801         set_genesis_time();
1802         create_disk_util_thread();
1803
1804         cgroup_list = smalloc(sizeof(*cgroup_list));
1805         INIT_FLIST_HEAD(cgroup_list);
1806
1807         run_threads();
1808
1809         if (!fio_abort) {
1810                 show_run_stats();
1811                 if (write_bw_log) {
1812                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1813                         __finish_log(agg_io_log[DDIR_WRITE],
1814                                         "agg-write_bw.log");
1815                 }
1816         }
1817
1818         for_each_td(td, i)
1819                 fio_options_free(td);
1820
1821         cgroup_kill(cgroup_list);
1822         sfree(cgroup_list);
1823         sfree(cgroup_mnt);
1824
1825         fio_mutex_remove(startup_mutex);
1826         fio_mutex_remove(writeout_mutex);
1827         return exit_value;
1828 }
1829
1830 void reset_fio_state(void)
1831 {
1832         groupid = 0;
1833         thread_number = 0;
1834         nr_process = 0;
1835         nr_thread = 0;
1836         done_secs = 0;
1837 }
1838
1839 static int endian_check(void)
1840 {
1841         union {
1842                 uint8_t c[8];
1843                 uint64_t v;
1844         } u;
1845         int le = 0, be = 0;
1846
1847         u.v = 0x12;
1848         if (u.c[7] == 0x12)
1849                 be = 1;
1850         else if (u.c[0] == 0x12)
1851                 le = 1;
1852
1853 #if defined(FIO_LITTLE_ENDIAN)
1854         if (be)
1855                 return 1;
1856 #elif defined(FIO_BIG_ENDIAN)
1857         if (le)
1858                 return 1;
1859 #else
1860         return 1;
1861 #endif
1862
1863         if (!le && !be)
1864                 return 1;
1865
1866         return 0;
1867 }
1868
1869 int main(int argc, char *argv[], char *envp[])
1870 {
1871         long ps;
1872
1873         if (endian_check()) {
1874                 log_err("fio: endianness settings appear wrong.\n");
1875                 log_err("fio: please report this to fio@vger.kernel.org\n");
1876                 return 1;
1877         }
1878
1879         arch_init(envp);
1880
1881         sinit();
1882
1883         /*
1884          * We need locale for number printing, if it isn't set then just
1885          * go with the US format.
1886          */
1887         if (!getenv("LC_NUMERIC"))
1888                 setlocale(LC_NUMERIC, "en_US");
1889
1890         ps = sysconf(_SC_PAGESIZE);
1891         if (ps < 0) {
1892                 log_err("Failed to get page size\n");
1893                 return 1;
1894         }
1895
1896         page_size = ps;
1897         page_mask = ps - 1;
1898
1899         fio_keywords_init();
1900
1901         if (parse_options(argc, argv))
1902                 return 1;
1903
1904         return exec_run();
1905 }