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 <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "hash.h"
  39 #include "smalloc.h"
  40
  41 unsigned long page_mask;
  42 unsigned long page_size;
  43 #define ALIGN(buf)      \
  44         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  45
  46 int groupid = 0;
  47 int thread_number = 0;
  48 int nr_process = 0;
  49 int nr_thread = 0;
  50 int shm_id = 0;
  51 int temp_stall_ts;
  52 unsigned long done_secs = 0;
  53
  54 static struct fio_mutex *startup_mutex;
  55 static volatile int fio_abort;
  56 static int exit_value;
  57 static struct itimerval itimer;
  58
  59 struct io_log *agg_io_log[2];
  60
  61 #define TERMINATE_ALL           (-1)
  62 #define JOB_START_TIMEOUT       (5 * 1000)
  63
  64 static inline void td_set_runstate(struct thread_data *td, int runstate)
  65 {
  66         if (td->runstate == runstate)
  67                 return;
  68
  69         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  70                                                 td->runstate, runstate);
  71         td->runstate = runstate;
  72 }
  73
  74 static void terminate_threads(int group_id)
  75 {
  76         struct thread_data *td;
  77         int i;
  78
  79         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  80
  81         for_each_td(td, i) {
  82                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  83                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  84                                                 td->o.name, (int) td->pid);
  85                         td->terminate = 1;
  86                         td->o.start_delay = 0;
  87
  88                         /*
  89                          * if the thread is running, just let it exit
  90                          */
  91                         if (td->runstate < TD_RUNNING)
  92                                 kill(td->pid, SIGQUIT);
  93                         else {
  94                                 struct ioengine_ops *ops = td->io_ops;
  95
  96                                 if (ops && (ops->flags & FIO_SIGQUIT))
  97                                         kill(td->pid, SIGQUIT);
  98                         }
  99                 }
 100         }
 101 }
 102
 103 static void status_timer_arm(void)
 104 {
 105         itimer.it_value.tv_sec = 0;
 106         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 107         setitimer(ITIMER_REAL, &itimer, NULL);
 108 }
 109
 110 static void sig_alrm(int sig)
 111 {
 112         if (threads) {
 113                 update_io_ticks();
 114                 print_thread_status();
 115                 status_timer_arm();
 116         }
 117 }
 118
 119 static void sig_int(int sig)
 120 {
 121         if (threads) {
 122                 printf("\nfio: terminating on signal %d\n", sig);
 123                 fflush(stdout);
 124                 terminate_threads(TERMINATE_ALL);
 125         }
 126 }
 127
 128 static void sig_ill(int sig)
 129 {
 130         if (!threads)
 131                 return;
 132
 133         log_err("fio: illegal instruction. your cpu does not support "
 134                 "the sse4.2 instruction for crc32c\n");
 135         terminate_threads(TERMINATE_ALL);
 136         exit(4);
 137 }
 138
 139 static void set_sig_handlers(void)
 140 {
 141         struct sigaction act;
 142
 143         memset(&act, 0, sizeof(act));
 144         act.sa_handler = sig_alrm;
 145         act.sa_flags = SA_RESTART;
 146         sigaction(SIGALRM, &act, NULL);
 147
 148         memset(&act, 0, sizeof(act));
 149         act.sa_handler = sig_int;
 150         act.sa_flags = SA_RESTART;
 151         sigaction(SIGINT, &act, NULL);
 152
 153         memset(&act, 0, sizeof(act));
 154         act.sa_handler = sig_ill;
 155         act.sa_flags = SA_RESTART;
 156         sigaction(SIGILL, &act, NULL);
 157 }
 158
 159 /*
 160  * Check if we are above the minimum rate given.
 161  */
 162 static int check_min_rate(struct thread_data *td, struct timeval *now)
 163 {
 164         unsigned long long bytes = 0;
 165         unsigned long iops = 0;
 166         unsigned long spent;
 167         unsigned long rate;
 168
 169         /*
 170          * No minimum rate set, always ok
 171          */
 172         if (!td->o.ratemin && !td->o.rate_iops_min)
 173                 return 0;
 174
 175         /*
 176          * allow a 2 second settle period in the beginning
 177          */
 178         if (mtime_since(&td->start, now) < 2000)
 179                 return 0;
 180
 181         if (td_read(td)) {
 182                 iops += td->io_blocks[DDIR_READ];
 183                 bytes += td->this_io_bytes[DDIR_READ];
 184         }
 185         if (td_write(td)) {
 186                 iops += td->io_blocks[DDIR_WRITE];
 187                 bytes += td->this_io_bytes[DDIR_WRITE];
 188         }
 189
 190         /*
 191          * if rate blocks is set, sample is running
 192          */
 193         if (td->rate_bytes || td->rate_blocks) {
 194                 spent = mtime_since(&td->lastrate, now);
 195                 if (spent < td->o.ratecycle)
 196                         return 0;
 197
 198                 if (td->o.rate) {
 199                         /*
 200                          * check bandwidth specified rate
 201                          */
 202                         if (bytes < td->rate_bytes) {
 203                                 log_err("%s: min rate %u not met\n", td->o.name,
 204                                                                 td->o.ratemin);
 205                                 return 1;
 206                         } else {
 207                                 rate = (bytes - td->rate_bytes) / spent;
 208                                 if (rate < td->o.ratemin ||
 209                                     bytes < td->rate_bytes) {
 210                                         log_err("%s: min rate %u not met, got"
 211                                                 " %luKiB/sec\n", td->o.name,
 212                                                         td->o.ratemin, rate);
 213                                         return 1;
 214                                 }
 215                         }
 216                 } else {
 217                         /*
 218                          * checks iops specified rate
 219                          */
 220                         if (iops < td->o.rate_iops) {
 221                                 log_err("%s: min iops rate %u not met\n",
 222                                                 td->o.name, td->o.rate_iops);
 223                                 return 1;
 224                         } else {
 225                                 rate = (iops - td->rate_blocks) / spent;
 226                                 if (rate < td->o.rate_iops_min ||
 227                                     iops < td->rate_blocks) {
 228                                         log_err("%s: min iops rate %u not met,"
 229                                                 " got %lu\n", td->o.name,
 230                                                         td->o.rate_iops_min,
 231                                                         rate);
 232                                 }
 233                         }
 234                 }
 235         }
 236
 237         td->rate_bytes = bytes;
 238         td->rate_blocks = iops;
 239         memcpy(&td->lastrate, now, sizeof(*now));
 240         return 0;
 241 }
 242
 243 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 244 {
 245         if (!td->o.timeout)
 246                 return 0;
 247         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 248                 return 1;
 249
 250         return 0;
 251 }
 252
 253 /*
 254  * When job exits, we can cancel the in-flight IO if we are using async
 255  * io. Attempt to do so.
 256  */
 257 static void cleanup_pending_aio(struct thread_data *td)
 258 {
 259         struct flist_head *entry, *n;
 260         struct io_u *io_u;
 261         int r;
 262
 263         /*
 264          * get immediately available events, if any
 265          */
 266         r = io_u_queued_complete(td, 0);
 267         if (r < 0)
 268                 return;
 269
 270         /*
 271          * now cancel remaining active events
 272          */
 273         if (td->io_ops->cancel) {
 274                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 275                         io_u = flist_entry(entry, struct io_u, list);
 276
 277                         /*
 278                          * if the io_u isn't in flight, then that generally
 279                          * means someone leaked an io_u. complain but fix
 280                          * it up, so we don't stall here.
 281                          */
 282                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 283                                 log_err("fio: non-busy IO on busy list\n");
 284                                 put_io_u(td, io_u);
 285                         } else {
 286                                 r = td->io_ops->cancel(td, io_u);
 287                                 if (!r)
 288                                         put_io_u(td, io_u);
 289                         }
 290                 }
 291         }
 292
 293         if (td->cur_depth)
 294                 r = io_u_queued_complete(td, td->cur_depth);
 295 }
 296
 297 /*
 298  * Helper to handle the final sync of a file. Works just like the normal
 299  * io path, just does everything sync.
 300  */
 301 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 302 {
 303         struct io_u *io_u = __get_io_u(td);
 304         int ret;
 305
 306         if (!io_u)
 307                 return 1;
 308
 309         io_u->ddir = DDIR_SYNC;
 310         io_u->file = f;
 311
 312         if (td_io_prep(td, io_u)) {
 313                 put_io_u(td, io_u);
 314                 return 1;
 315         }
 316
 317 requeue:
 318         ret = td_io_queue(td, io_u);
 319         if (ret < 0) {
 320                 td_verror(td, io_u->error, "td_io_queue");
 321                 put_io_u(td, io_u);
 322                 return 1;
 323         } else if (ret == FIO_Q_QUEUED) {
 324                 if (io_u_queued_complete(td, 1) < 0)
 325                         return 1;
 326         } else if (ret == FIO_Q_COMPLETED) {
 327                 if (io_u->error) {
 328                         td_verror(td, io_u->error, "td_io_queue");
 329                         return 1;
 330                 }
 331
 332                 if (io_u_sync_complete(td, io_u) < 0)
 333                         return 1;
 334         } else if (ret == FIO_Q_BUSY) {
 335                 if (td_io_commit(td))
 336                         return 1;
 337                 goto requeue;
 338         }
 339
 340         return 0;
 341 }
 342
 343 /*
 344  * The main verify engine. Runs over the writes we previously submitted,
 345  * reads the blocks back in, and checks the crc/md5 of the data.
 346  */
 347 static void do_verify(struct thread_data *td)
 348 {
 349         struct fio_file *f;
 350         struct io_u *io_u;
 351         int ret, min_events;
 352         unsigned int i;
 353
 354         /*
 355          * sync io first and invalidate cache, to make sure we really
 356          * read from disk.
 357          */
 358         for_each_file(td, f, i) {
 359                 if (!(f->flags & FIO_FILE_OPEN))
 360                         continue;
 361                 if (fio_io_sync(td, f))
 362                         break;
 363                 if (file_invalidate_cache(td, f))
 364                         break;
 365         }
 366
 367         if (td->error)
 368                 return;
 369
 370         td_set_runstate(td, TD_VERIFYING);
 371
 372         io_u = NULL;
 373         while (!td->terminate) {
 374                 int ret2, full;
 375
 376                 io_u = __get_io_u(td);
 377                 if (!io_u)
 378                         break;
 379
 380                 if (runtime_exceeded(td, &io_u->start_time)) {
 381                         put_io_u(td, io_u);
 382                         td->terminate = 1;
 383                         break;
 384                 }
 385
 386                 if (get_next_verify(td, io_u)) {
 387                         put_io_u(td, io_u);
 388                         break;
 389                 }
 390
 391                 if (td_io_prep(td, io_u)) {
 392                         put_io_u(td, io_u);
 393                         break;
 394                 }
 395
 396                 io_u->end_io = verify_io_u;
 397
 398                 ret = td_io_queue(td, io_u);
 399                 switch (ret) {
 400                 case FIO_Q_COMPLETED:
 401                         if (io_u->error)
 402                                 ret = -io_u->error;
 403                         else if (io_u->resid) {
 404                                 int bytes = io_u->xfer_buflen - io_u->resid;
 405                                 struct fio_file *f = io_u->file;
 406
 407                                 /*
 408                                  * zero read, fail
 409                                  */
 410                                 if (!bytes) {
 411                                         td_verror(td, EIO, "full resid");
 412                                         put_io_u(td, io_u);
 413                                         break;
 414                                 }
 415
 416                                 io_u->xfer_buflen = io_u->resid;
 417                                 io_u->xfer_buf += bytes;
 418                                 io_u->offset += bytes;
 419
 420                                 td->ts.short_io_u[io_u->ddir]++;
 421
 422                                 if (io_u->offset == f->real_file_size)
 423                                         goto sync_done;
 424
 425                                 requeue_io_u(td, &io_u);
 426                         } else {
 427 sync_done:
 428                                 ret = io_u_sync_complete(td, io_u);
 429                                 if (ret < 0)
 430                                         break;
 431                         }
 432                         continue;
 433                 case FIO_Q_QUEUED:
 434                         break;
 435                 case FIO_Q_BUSY:
 436                         requeue_io_u(td, &io_u);
 437                         ret2 = td_io_commit(td);
 438                         if (ret2 < 0)
 439                                 ret = ret2;
 440                         break;
 441                 default:
 442                         assert(ret < 0);
 443                         td_verror(td, -ret, "td_io_queue");
 444                         break;
 445                 }
 446
 447                 if (ret < 0 || td->error)
 448                         break;
 449
 450                 /*
 451                  * if we can queue more, do so. but check if there are
 452                  * completed io_u's first.
 453                  */
 454                 full = queue_full(td) || ret == FIO_Q_BUSY;
 455                 if (full || !td->o.iodepth_batch_complete) {
 456                         min_events = td->o.iodepth_batch_complete;
 457                         if (full && !min_events)
 458                                 min_events = 1;
 459
 460                         do {
 461                                 /*
 462                                  * Reap required number of io units, if any,
 463                                  * and do the verification on them through
 464                                  * the callback handler
 465                                  */
 466                                 if (io_u_queued_complete(td, min_events) < 0) {
 467                                         ret = -1;
 468                                         break;
 469                                 }
 470                         } while (full && (td->cur_depth > td->o.iodepth_low));
 471                 }
 472                 if (ret < 0)
 473                         break;
 474         }
 475
 476         if (!td->error) {
 477                 min_events = td->cur_depth;
 478
 479                 if (min_events)
 480                         ret = io_u_queued_complete(td, min_events);
 481         } else
 482                 cleanup_pending_aio(td);
 483
 484         td_set_runstate(td, TD_RUNNING);
 485 }
 486
 487 /*
 488  * Main IO worker function. It retrieves io_u's to process and queues
 489  * and reaps them, checking for rate and errors along the way.
 490  */
 491 static void do_io(struct thread_data *td)
 492 {
 493         struct timeval s;
 494         unsigned long usec;
 495         unsigned int i;
 496         int ret = 0;
 497
 498         td_set_runstate(td, TD_RUNNING);
 499
 500         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 501                 struct timeval comp_time;
 502                 long bytes_done = 0;
 503                 int min_evts = 0;
 504                 struct io_u *io_u;
 505                 int ret2, full;
 506
 507                 if (td->terminate)
 508                         break;
 509
 510                 io_u = get_io_u(td);
 511                 if (!io_u)
 512                         break;
 513
 514                 memcpy(&s, &io_u->start_time, sizeof(s));
 515
 516                 if (runtime_exceeded(td, &s)) {
 517                         put_io_u(td, io_u);
 518                         td->terminate = 1;
 519                         break;
 520                 }
 521
 522                 /*
 523                  * Add verification end_io handler, if asked to verify
 524                  * a previously written file.
 525                  */
 526                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 527                         io_u->end_io = verify_io_u;
 528                         td_set_runstate(td, TD_VERIFYING);
 529                 } else
 530                         td_set_runstate(td, TD_RUNNING);
 531
 532                 ret = td_io_queue(td, io_u);
 533                 switch (ret) {
 534                 case FIO_Q_COMPLETED:
 535                         if (io_u->error)
 536                                 ret = -io_u->error;
 537                         else if (io_u->resid) {
 538                                 int bytes = io_u->xfer_buflen - io_u->resid;
 539                                 struct fio_file *f = io_u->file;
 540
 541                                 /*
 542                                  * zero read, fail
 543                                  */
 544                                 if (!bytes) {
 545                                         td_verror(td, EIO, "full resid");
 546                                         put_io_u(td, io_u);
 547                                         break;
 548                                 }
 549
 550                                 io_u->xfer_buflen = io_u->resid;
 551                                 io_u->xfer_buf += bytes;
 552                                 io_u->offset += bytes;
 553
 554                                 td->ts.short_io_u[io_u->ddir]++;
 555
 556                                 if (io_u->offset == f->real_file_size)
 557                                         goto sync_done;
 558
 559                                 requeue_io_u(td, &io_u);
 560                         } else {
 561 sync_done:
 562                                 fio_gettime(&comp_time, NULL);
 563                                 bytes_done = io_u_sync_complete(td, io_u);
 564                                 if (bytes_done < 0)
 565                                         ret = bytes_done;
 566                         }
 567                         break;
 568                 case FIO_Q_QUEUED:
 569                         /*
 570                          * if the engine doesn't have a commit hook,
 571                          * the io_u is really queued. if it does have such
 572                          * a hook, it has to call io_u_queued() itself.
 573                          */
 574                         if (td->io_ops->commit == NULL)
 575                                 io_u_queued(td, io_u);
 576                         break;
 577                 case FIO_Q_BUSY:
 578                         requeue_io_u(td, &io_u);
 579                         ret2 = td_io_commit(td);
 580                         if (ret2 < 0)
 581                                 ret = ret2;
 582                         break;
 583                 default:
 584                         assert(ret < 0);
 585                         put_io_u(td, io_u);
 586                         break;
 587                 }
 588
 589                 if (ret < 0 || td->error)
 590                         break;
 591
 592                 /*
 593                  * See if we need to complete some commands
 594                  */
 595                 full = queue_full(td) || ret == FIO_Q_BUSY;
 596                 if (full || !td->o.iodepth_batch_complete) {
 597                         min_evts = td->o.iodepth_batch_complete;
 598                         if (full && !min_evts)
 599                                 min_evts = 1;
 600
 601                         fio_gettime(&comp_time, NULL);
 602
 603                         do {
 604                                 ret = io_u_queued_complete(td, min_evts);
 605                                 if (ret <= 0)
 606                                         break;
 607
 608                                 bytes_done += ret;
 609                         } while (full && (td->cur_depth > td->o.iodepth_low));
 610                 }
 611
 612                 if (ret < 0)
 613                         break;
 614                 if (!bytes_done)
 615                         continue;
 616
 617                 /*
 618                  * the rate is batched for now, it should work for batches
 619                  * of completions except the very first one which may look
 620                  * a little bursty
 621                  */
 622                 if (ramp_time_over(td)) {
 623                         usec = utime_since(&s, &comp_time);
 624
 625                         rate_throttle(td, usec, bytes_done);
 626
 627                         if (check_min_rate(td, &comp_time)) {
 628                                 if (exitall_on_terminate)
 629                                         terminate_threads(td->groupid);
 630                                 td_verror(td, EIO, "check_min_rate");
 631                                 break;
 632                         }
 633                 }
 634
 635                 if (td->o.thinktime) {
 636                         unsigned long long b;
 637
 638                         b = td->io_blocks[0] + td->io_blocks[1];
 639                         if (!(b % td->o.thinktime_blocks)) {
 640                                 int left;
 641
 642                                 if (td->o.thinktime_spin)
 643                                         __usec_sleep(td->o.thinktime_spin);
 644
 645                                 left = td->o.thinktime - td->o.thinktime_spin;
 646                                 if (left)
 647                                         usec_sleep(td, left);
 648                         }
 649                 }
 650         }
 651
 652         if (td->o.fill_device && td->error == ENOSPC) {
 653                 td->error = 0;
 654                 td->terminate = 1;
 655         }
 656         if (!td->error) {
 657                 struct fio_file *f;
 658
 659                 i = td->cur_depth;
 660                 if (i)
 661                         ret = io_u_queued_complete(td, i);
 662
 663                 if (should_fsync(td) && td->o.end_fsync) {
 664                         td_set_runstate(td, TD_FSYNCING);
 665
 666                         for_each_file(td, f, i) {
 667                                 if (!(f->flags & FIO_FILE_OPEN))
 668                                         continue;
 669                                 fio_io_sync(td, f);
 670                         }
 671                 }
 672         } else
 673                 cleanup_pending_aio(td);
 674
 675         /*
 676          * stop job if we failed doing any IO
 677          */
 678         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 679                 td->done = 1;
 680 }
 681
 682 static void cleanup_io_u(struct thread_data *td)
 683 {
 684         struct flist_head *entry, *n;
 685         struct io_u *io_u;
 686
 687         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 688                 io_u = flist_entry(entry, struct io_u, list);
 689
 690                 flist_del(&io_u->list);
 691                 free(io_u);
 692         }
 693
 694         free_io_mem(td);
 695 }
 696
 697 static int init_io_u(struct thread_data *td)
 698 {
 699         struct io_u *io_u;
 700         unsigned int max_bs;
 701         int i, max_units;
 702         char *p;
 703
 704         max_units = td->o.iodepth;
 705         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 706         td->orig_buffer_size = (unsigned long long) max_bs
 707                                         * (unsigned long long) max_units;
 708
 709         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 710                 unsigned long bs;
 711
 712                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 713                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 714         }
 715
 716         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 717                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 718                 return 1;
 719         }
 720
 721         if (allocate_io_mem(td))
 722                 return 1;
 723
 724         if (td->o.odirect)
 725                 p = ALIGN(td->orig_buffer);
 726         else
 727                 p = td->orig_buffer;
 728
 729         for (i = 0; i < max_units; i++) {
 730                 if (td->terminate)
 731                         return 1;
 732                 io_u = malloc(sizeof(*io_u));
 733                 memset(io_u, 0, sizeof(*io_u));
 734                 INIT_FLIST_HEAD(&io_u->list);
 735
 736                 if (!(td->io_ops->flags & FIO_NOIO)) {
 737                         io_u->buf = p + max_bs * i;
 738
 739                         if (td_write(td) && !td->o.refill_buffers)
 740                                 io_u_fill_buffer(td, io_u, max_bs);
 741                 }
 742
 743                 io_u->index = i;
 744                 io_u->flags = IO_U_F_FREE;
 745                 flist_add(&io_u->list, &td->io_u_freelist);
 746         }
 747
 748         io_u_init_timeout();
 749
 750         return 0;
 751 }
 752
 753 static int switch_ioscheduler(struct thread_data *td)
 754 {
 755         char tmp[256], tmp2[128];
 756         FILE *f;
 757         int ret;
 758
 759         if (td->io_ops->flags & FIO_DISKLESSIO)
 760                 return 0;
 761
 762         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 763
 764         f = fopen(tmp, "r+");
 765         if (!f) {
 766                 if (errno == ENOENT) {
 767                         log_err("fio: os or kernel doesn't support IO scheduler"
 768                                 " switching\n");
 769                         return 0;
 770                 }
 771                 td_verror(td, errno, "fopen iosched");
 772                 return 1;
 773         }
 774
 775         /*
 776          * Set io scheduler.
 777          */
 778         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 779         if (ferror(f) || ret != 1) {
 780                 td_verror(td, errno, "fwrite");
 781                 fclose(f);
 782                 return 1;
 783         }
 784
 785         rewind(f);
 786
 787         /*
 788          * Read back and check that the selected scheduler is now the default.
 789          */
 790         ret = fread(tmp, 1, sizeof(tmp), f);
 791         if (ferror(f) || ret < 0) {
 792                 td_verror(td, errno, "fread");
 793                 fclose(f);
 794                 return 1;
 795         }
 796
 797         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 798         if (!strstr(tmp, tmp2)) {
 799                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 800                 td_verror(td, EINVAL, "iosched_switch");
 801                 fclose(f);
 802                 return 1;
 803         }
 804
 805         fclose(f);
 806         return 0;
 807 }
 808
 809 static int keep_running(struct thread_data *td)
 810 {
 811         unsigned long long io_done;
 812
 813         if (td->done)
 814                 return 0;
 815         if (td->o.time_based)
 816                 return 1;
 817         if (td->o.loops) {
 818                 td->o.loops--;
 819                 return 1;
 820         }
 821
 822         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 823                         + td->io_skip_bytes;
 824         if (io_done < td->o.size)
 825                 return 1;
 826
 827         return 0;
 828 }
 829
 830 static int clear_io_state(struct thread_data *td)
 831 {
 832         struct fio_file *f;
 833         unsigned int i;
 834         int ret;
 835
 836         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 837         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 838         td->zone_bytes = 0;
 839         td->rate_bytes = 0;
 840         td->rate_blocks = 0;
 841         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 842
 843         td->last_was_sync = 0;
 844
 845         /*
 846          * reset file done count if we are to start over
 847          */
 848         if (td->o.time_based || td->o.loops)
 849                 td->nr_done_files = 0;
 850
 851         close_files(td);
 852
 853         ret = 0;
 854         for_each_file(td, f, i) {
 855                 f->flags &= ~FIO_FILE_DONE;
 856                 ret = td_io_open_file(td, f);
 857                 if (ret)
 858                         break;
 859         }
 860
 861         return ret;
 862 }
 863
 864 /*
 865  * Entry point for the thread based jobs. The process based jobs end up
 866  * here as well, after a little setup.
 867  */
 868 static void *thread_main(void *data)
 869 {
 870         unsigned long long runtime[2], elapsed;
 871         struct thread_data *td = data;
 872         int clear_state;
 873
 874         if (!td->o.use_thread)
 875                 setsid();
 876
 877         td->pid = getpid();
 878
 879         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 880
 881         INIT_FLIST_HEAD(&td->io_u_freelist);
 882         INIT_FLIST_HEAD(&td->io_u_busylist);
 883         INIT_FLIST_HEAD(&td->io_u_requeues);
 884         INIT_FLIST_HEAD(&td->io_log_list);
 885         INIT_FLIST_HEAD(&td->io_hist_list);
 886         td->io_hist_tree = RB_ROOT;
 887
 888         td_set_runstate(td, TD_INITIALIZED);
 889         fio_mutex_up(startup_mutex);
 890         fio_mutex_down(td->mutex);
 891
 892         /*
 893          * the ->mutex mutex is now no longer used, close it to avoid
 894          * eating a file descriptor
 895          */
 896         fio_mutex_remove(td->mutex);
 897
 898         /*
 899          * May alter parameters that init_io_u() will use, so we need to
 900          * do this first.
 901          */
 902         if (init_iolog(td))
 903                 goto err;
 904
 905         if (init_io_u(td))
 906                 goto err;
 907
 908         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 909                 td_verror(td, errno, "cpu_set_affinity");
 910                 goto err;
 911         }
 912
 913         if (td->ioprio_set) {
 914                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 915                         td_verror(td, errno, "ioprio_set");
 916                         goto err;
 917                 }
 918         }
 919
 920         if (nice(td->o.nice) == -1) {
 921                 td_verror(td, errno, "nice");
 922                 goto err;
 923         }
 924
 925         if (td->o.ioscheduler && switch_ioscheduler(td))
 926                 goto err;
 927
 928         if (!td->o.create_serialize && setup_files(td))
 929                 goto err;
 930
 931         if (td_io_init(td))
 932                 goto err;
 933
 934         if (open_files(td))
 935                 goto err;
 936
 937         if (init_random_map(td))
 938                 goto err;
 939
 940         if (td->o.exec_prerun) {
 941                 if (system(td->o.exec_prerun) < 0)
 942                         goto err;
 943         }
 944
 945         fio_gettime(&td->epoch, NULL);
 946         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 947         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 948
 949         runtime[0] = runtime[1] = 0;
 950         clear_state = 0;
 951         while (keep_running(td)) {
 952                 fio_gettime(&td->start, NULL);
 953                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 954
 955                 if (td->o.ratemin)
 956                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 957                                                         sizeof(td->lastrate));
 958
 959                 if (clear_state && clear_io_state(td))
 960                         break;
 961
 962                 prune_io_piece_log(td);
 963
 964                 do_io(td);
 965
 966                 clear_state = 1;
 967
 968                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 969                         if (td->rw_end_set[DDIR_READ])
 970                                 elapsed = utime_since(&td->start,
 971                                                       &td->rw_end[DDIR_READ]);
 972                         else
 973                                 elapsed = utime_since_now(&td->start);
 974
 975                         runtime[DDIR_READ] += elapsed;
 976                 }
 977                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 978                         if (td->rw_end_set[DDIR_WRITE])
 979                                 elapsed = utime_since(&td->start,
 980                                                       &td->rw_end[DDIR_WRITE]);
 981                         else
 982                                 elapsed = utime_since_now(&td->start);
 983
 984                         runtime[DDIR_WRITE] += elapsed;
 985                 }
 986
 987                 if (td->error || td->terminate)
 988                         break;
 989
 990                 if (!td->o.do_verify ||
 991                     td->o.verify == VERIFY_NONE ||
 992                     (td->io_ops->flags & FIO_UNIDIR))
 993                         continue;
 994
 995                 if (clear_io_state(td))
 996                         break;
 997
 998                 fio_gettime(&td->start, NULL);
 999
1000                 do_verify(td);
1001
1002                 runtime[DDIR_READ] += utime_since_now(&td->start);
1003
1004                 if (td->error || td->terminate)
1005                         break;
1006         }
1007
1008         update_rusage_stat(td);
1009         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1010         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1011         td->ts.total_run_time = mtime_since_now(&td->epoch);
1012         td->ts.io_bytes[0] = td->io_bytes[0];
1013         td->ts.io_bytes[1] = td->io_bytes[1];
1014
1015         if (td->ts.bw_log)
1016                 finish_log(td, td->ts.bw_log, "bw");
1017         if (td->ts.slat_log)
1018                 finish_log(td, td->ts.slat_log, "slat");
1019         if (td->ts.clat_log)
1020                 finish_log(td, td->ts.clat_log, "clat");
1021         if (td->o.exec_postrun) {
1022                 if (system(td->o.exec_postrun) < 0)
1023                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
1024         }
1025
1026         if (exitall_on_terminate)
1027                 terminate_threads(td->groupid);
1028
1029 err:
1030         if (td->error)
1031                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1032                                                         td->verror);
1033         close_and_free_files(td);
1034         close_ioengine(td);
1035         cleanup_io_u(td);
1036
1037         /*
1038          * do this very late, it will log file closing as well
1039          */
1040         if (td->o.write_iolog_file)
1041                 write_iolog_close(td);
1042
1043         options_mem_free(td);
1044         td_set_runstate(td, TD_EXITED);
1045         return (void *) (unsigned long) td->error;
1046 }
1047
1048 /*
1049  * We cannot pass the td data into a forked process, so attach the td and
1050  * pass it to the thread worker.
1051  */
1052 static int fork_main(int shmid, int offset)
1053 {
1054         struct thread_data *td;
1055         void *data, *ret;
1056
1057         data = shmat(shmid, NULL, 0);
1058         if (data == (void *) -1) {
1059                 int __err = errno;
1060
1061                 perror("shmat");
1062                 return __err;
1063         }
1064
1065         td = data + offset * sizeof(struct thread_data);
1066         ret = thread_main(td);
1067         shmdt(data);
1068         return (int) (unsigned long) ret;
1069 }
1070
1071 /*
1072  * Run over the job map and reap the threads that have exited, if any.
1073  */
1074 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1075 {
1076         struct thread_data *td;
1077         int i, cputhreads, realthreads, pending, status, ret;
1078
1079         /*
1080          * reap exited threads (TD_EXITED -> TD_REAPED)
1081          */
1082         realthreads = pending = cputhreads = 0;
1083         for_each_td(td, i) {
1084                 int flags = 0;
1085
1086                 /*
1087                  * ->io_ops is NULL for a thread that has closed its
1088                  * io engine
1089                  */
1090                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1091                         cputhreads++;
1092                 else
1093                         realthreads++;
1094
1095                 if (!td->pid) {
1096                         pending++;
1097                         continue;
1098                 }
1099                 if (td->runstate == TD_REAPED)
1100                         continue;
1101                 if (td->o.use_thread) {
1102                         if (td->runstate == TD_EXITED) {
1103                                 td_set_runstate(td, TD_REAPED);
1104                                 goto reaped;
1105                         }
1106                         continue;
1107                 }
1108
1109                 flags = WNOHANG;
1110                 if (td->runstate == TD_EXITED)
1111                         flags = 0;
1112
1113                 /*
1114                  * check if someone quit or got killed in an unusual way
1115                  */
1116                 ret = waitpid(td->pid, &status, flags);
1117                 if (ret < 0) {
1118                         if (errno == ECHILD) {
1119                                 log_err("fio: pid=%d disappeared %d\n",
1120                                                 (int) td->pid, td->runstate);
1121                                 td_set_runstate(td, TD_REAPED);
1122                                 goto reaped;
1123                         }
1124                         perror("waitpid");
1125                 } else if (ret == td->pid) {
1126                         if (WIFSIGNALED(status)) {
1127                                 int sig = WTERMSIG(status);
1128
1129                                 if (sig != SIGQUIT)
1130                                         log_err("fio: pid=%d, got signal=%d\n",
1131                                                         (int) td->pid, sig);
1132                                 td_set_runstate(td, TD_REAPED);
1133                                 goto reaped;
1134                         }
1135                         if (WIFEXITED(status)) {
1136                                 if (WEXITSTATUS(status) && !td->error)
1137                                         td->error = WEXITSTATUS(status);
1138
1139                                 td_set_runstate(td, TD_REAPED);
1140                                 goto reaped;
1141                         }
1142                 }
1143
1144                 /*
1145                  * thread is not dead, continue
1146                  */
1147                 pending++;
1148                 continue;
1149 reaped:
1150                 (*nr_running)--;
1151                 (*m_rate) -= td->o.ratemin;
1152                 (*t_rate) -= td->o.rate;
1153                 if (!td->pid)
1154                         pending--;
1155
1156                 if (td->error)
1157                         exit_value++;
1158
1159                 done_secs += mtime_since_now(&td->epoch) / 1000;
1160         }
1161
1162         if (*nr_running == cputhreads && !pending && realthreads)
1163                 terminate_threads(TERMINATE_ALL);
1164 }
1165
1166 /*
1167  * Main function for kicking off and reaping jobs, as needed.
1168  */
1169 static void run_threads(void)
1170 {
1171         struct thread_data *td;
1172         unsigned long spent;
1173         int i, todo, nr_running, m_rate, t_rate, nr_started;
1174
1175         if (fio_pin_memory())
1176                 return;
1177
1178         if (!terse_output) {
1179                 printf("Starting ");
1180                 if (nr_thread)
1181                         printf("%d thread%s", nr_thread,
1182                                                 nr_thread > 1 ? "s" : "");
1183                 if (nr_process) {
1184                         if (nr_thread)
1185                                 printf(" and ");
1186                         printf("%d process%s", nr_process,
1187                                                 nr_process > 1 ? "es" : "");
1188                 }
1189                 printf("\n");
1190                 fflush(stdout);
1191         }
1192
1193         set_sig_handlers();
1194
1195         todo = thread_number;
1196         nr_running = 0;
1197         nr_started = 0;
1198         m_rate = t_rate = 0;
1199
1200         for_each_td(td, i) {
1201                 print_status_init(td->thread_number - 1);
1202
1203                 if (!td->o.create_serialize) {
1204                         init_disk_util(td);
1205                         continue;
1206                 }
1207
1208                 /*
1209                  * do file setup here so it happens sequentially,
1210                  * we don't want X number of threads getting their
1211                  * client data interspersed on disk
1212                  */
1213                 if (setup_files(td)) {
1214                         exit_value++;
1215                         if (td->error)
1216                                 log_err("fio: pid=%d, err=%d/%s\n",
1217                                         (int) td->pid, td->error, td->verror);
1218                         td_set_runstate(td, TD_REAPED);
1219                         todo--;
1220                 } else {
1221                         struct fio_file *f;
1222                         unsigned int i;
1223
1224                         /*
1225                          * for sharing to work, each job must always open
1226                          * its own files. so close them, if we opened them
1227                          * for creation
1228                          */
1229                         for_each_file(td, f, i)
1230                                 td_io_close_file(td, f);
1231                 }
1232
1233                 init_disk_util(td);
1234         }
1235
1236         set_genesis_time();
1237
1238         while (todo) {
1239                 struct thread_data *map[MAX_JOBS];
1240                 struct timeval this_start;
1241                 int this_jobs = 0, left;
1242
1243                 /*
1244                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1245                  */
1246                 for_each_td(td, i) {
1247                         if (td->runstate != TD_NOT_CREATED)
1248                                 continue;
1249
1250                         /*
1251                          * never got a chance to start, killed by other
1252                          * thread for some reason
1253                          */
1254                         if (td->terminate) {
1255                                 todo--;
1256                                 continue;
1257                         }
1258
1259                         if (td->o.start_delay) {
1260                                 spent = mtime_since_genesis();
1261
1262                                 if (td->o.start_delay * 1000 > spent)
1263                                         continue;
1264                         }
1265
1266                         if (td->o.stonewall && (nr_started || nr_running)) {
1267                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1268                                                         td->o.name);
1269                                 break;
1270                         }
1271
1272                         /*
1273                          * Set state to created. Thread will transition
1274                          * to TD_INITIALIZED when it's done setting up.
1275                          */
1276                         td_set_runstate(td, TD_CREATED);
1277                         map[this_jobs++] = td;
1278                         nr_started++;
1279
1280                         if (td->o.use_thread) {
1281                                 dprint(FD_PROCESS, "will pthread_create\n");
1282                                 if (pthread_create(&td->thread, NULL,
1283                                                    thread_main, td)) {
1284                                         perror("pthread_create");
1285                                         nr_started--;
1286                                         break;
1287                                 }
1288                                 if (pthread_detach(td->thread) < 0)
1289                                         perror("pthread_detach");
1290                         } else {
1291                                 pid_t pid;
1292                                 dprint(FD_PROCESS, "will fork\n");
1293                                 pid = fork();
1294                                 if (!pid) {
1295                                         int ret = fork_main(shm_id, i);
1296
1297                                         _exit(ret);
1298                                 } else if (i == fio_debug_jobno)
1299                                         *fio_debug_jobp = pid;
1300                         }
1301                         fio_mutex_down(startup_mutex);
1302                 }
1303
1304                 /*
1305                  * Wait for the started threads to transition to
1306                  * TD_INITIALIZED.
1307                  */
1308                 fio_gettime(&this_start, NULL);
1309                 left = this_jobs;
1310                 while (left && !fio_abort) {
1311                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1312                                 break;
1313
1314                         usleep(100000);
1315
1316                         for (i = 0; i < this_jobs; i++) {
1317                                 td = map[i];
1318                                 if (!td)
1319                                         continue;
1320                                 if (td->runstate == TD_INITIALIZED) {
1321                                         map[i] = NULL;
1322                                         left--;
1323                                 } else if (td->runstate >= TD_EXITED) {
1324                                         map[i] = NULL;
1325                                         left--;
1326                                         todo--;
1327                                         nr_running++; /* work-around... */
1328                                 }
1329                         }
1330                 }
1331
1332                 if (left) {
1333                         log_err("fio: %d jobs failed to start\n", left);
1334                         for (i = 0; i < this_jobs; i++) {
1335                                 td = map[i];
1336                                 if (!td)
1337                                         continue;
1338                                 kill(td->pid, SIGTERM);
1339                         }
1340                         break;
1341                 }
1342
1343                 /*
1344                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1345                  */
1346                 for_each_td(td, i) {
1347                         if (td->runstate != TD_INITIALIZED)
1348                                 continue;
1349
1350                         td_set_runstate(td, TD_RUNNING);
1351                         nr_running++;
1352                         nr_started--;
1353                         m_rate += td->o.ratemin;
1354                         t_rate += td->o.rate;
1355                         todo--;
1356                         fio_mutex_up(td->mutex);
1357                 }
1358
1359                 reap_threads(&nr_running, &t_rate, &m_rate);
1360
1361                 if (todo)
1362                         usleep(100000);
1363         }
1364
1365         while (nr_running) {
1366                 reap_threads(&nr_running, &t_rate, &m_rate);
1367                 usleep(10000);
1368         }
1369
1370         update_io_ticks();
1371         fio_unpin_memory();
1372 }
1373
1374 int main(int argc, char *argv[])
1375 {
1376         long ps;
1377
1378         sinit();
1379
1380         /*
1381          * We need locale for number printing, if it isn't set then just
1382          * go with the US format.
1383          */
1384         if (!getenv("LC_NUMERIC"))
1385                 setlocale(LC_NUMERIC, "en_US");
1386
1387         if (parse_options(argc, argv))
1388                 return 1;
1389
1390         if (!thread_number)
1391                 return 0;
1392
1393         ps = sysconf(_SC_PAGESIZE);
1394         if (ps < 0) {
1395                 log_err("Failed to get page size\n");
1396                 return 1;
1397         }
1398
1399         page_size = ps;
1400         page_mask = ps - 1;
1401
1402         if (write_bw_log) {
1403                 setup_log(&agg_io_log[DDIR_READ]);
1404                 setup_log(&agg_io_log[DDIR_WRITE]);
1405         }
1406
1407         startup_mutex = fio_mutex_init(0);
1408
1409         set_genesis_time();
1410
1411         status_timer_arm();
1412
1413         run_threads();
1414
1415         if (!fio_abort) {
1416                 show_run_stats();
1417                 if (write_bw_log) {
1418                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1419                         __finish_log(agg_io_log[DDIR_WRITE],
1420                                         "agg-write_bw.log");
1421                 }
1422         }
1423
1424         fio_mutex_remove(startup_mutex);
1425         return exit_value;
1426 }