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 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 fio_unused 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 fio_unused 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 static inline int should_check_rate(struct thread_data *td)
 160 {
 161         /*
 162          * No minimum rate set, always ok
 163          */
 164         if (!td->o.ratemin && !td->o.rate_iops_min)
 165                 return 0;
 166
 167         return 1;
 168 }
 169
 170 /*
 171  * Check if we are above the minimum rate given.
 172  */
 173 static int check_min_rate(struct thread_data *td, struct timeval *now)
 174 {
 175         unsigned long long bytes = 0;
 176         unsigned long iops = 0;
 177         unsigned long spent;
 178         unsigned long rate;
 179
 180         /*
 181          * allow a 2 second settle period in the beginning
 182          */
 183         if (mtime_since(&td->start, now) < 2000)
 184                 return 0;
 185
 186         if (td_read(td)) {
 187                 iops += td->io_blocks[DDIR_READ];
 188                 bytes += td->this_io_bytes[DDIR_READ];
 189         }
 190         if (td_write(td)) {
 191                 iops += td->io_blocks[DDIR_WRITE];
 192                 bytes += td->this_io_bytes[DDIR_WRITE];
 193         }
 194
 195         /*
 196          * if rate blocks is set, sample is running
 197          */
 198         if (td->rate_bytes || td->rate_blocks) {
 199                 spent = mtime_since(&td->lastrate, now);
 200                 if (spent < td->o.ratecycle)
 201                         return 0;
 202
 203                 if (td->o.rate) {
 204                         /*
 205                          * check bandwidth specified rate
 206                          */
 207                         if (bytes < td->rate_bytes) {
 208                                 log_err("%s: min rate %u not met\n", td->o.name,
 209                                                                 td->o.ratemin);
 210                                 return 1;
 211                         } else {
 212                                 rate = (bytes - td->rate_bytes) / spent;
 213                                 if (rate < td->o.ratemin ||
 214                                     bytes < td->rate_bytes) {
 215                                         log_err("%s: min rate %u not met, got"
 216                                                 " %luKiB/sec\n", td->o.name,
 217                                                         td->o.ratemin, rate);
 218                                         return 1;
 219                                 }
 220                         }
 221                 } else {
 222                         /*
 223                          * checks iops specified rate
 224                          */
 225                         if (iops < td->o.rate_iops) {
 226                                 log_err("%s: min iops rate %u not met\n",
 227                                                 td->o.name, td->o.rate_iops);
 228                                 return 1;
 229                         } else {
 230                                 rate = (iops - td->rate_blocks) / spent;
 231                                 if (rate < td->o.rate_iops_min ||
 232                                     iops < td->rate_blocks) {
 233                                         log_err("%s: min iops rate %u not met,"
 234                                                 " got %lu\n", td->o.name,
 235                                                         td->o.rate_iops_min,
 236                                                         rate);
 237                                 }
 238                         }
 239                 }
 240         }
 241
 242         td->rate_bytes = bytes;
 243         td->rate_blocks = iops;
 244         memcpy(&td->lastrate, now, sizeof(*now));
 245         return 0;
 246 }
 247
 248 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 249 {
 250         if (!td->o.timeout)
 251                 return 0;
 252         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 253                 return 1;
 254
 255         return 0;
 256 }
 257
 258 /*
 259  * When job exits, we can cancel the in-flight IO if we are using async
 260  * io. Attempt to do so.
 261  */
 262 static void cleanup_pending_aio(struct thread_data *td)
 263 {
 264         struct flist_head *entry, *n;
 265         struct io_u *io_u;
 266         int r;
 267
 268         /*
 269          * get immediately available events, if any
 270          */
 271         r = io_u_queued_complete(td, 0);
 272         if (r < 0)
 273                 return;
 274
 275         /*
 276          * now cancel remaining active events
 277          */
 278         if (td->io_ops->cancel) {
 279                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 280                         io_u = flist_entry(entry, struct io_u, list);
 281
 282                         /*
 283                          * if the io_u isn't in flight, then that generally
 284                          * means someone leaked an io_u. complain but fix
 285                          * it up, so we don't stall here.
 286                          */
 287                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 288                                 log_err("fio: non-busy IO on busy list\n");
 289                                 put_io_u(td, io_u);
 290                         } else {
 291                                 r = td->io_ops->cancel(td, io_u);
 292                                 if (!r)
 293                                         put_io_u(td, io_u);
 294                         }
 295                 }
 296         }
 297
 298         if (td->cur_depth)
 299                 r = io_u_queued_complete(td, td->cur_depth);
 300 }
 301
 302 /*
 303  * Helper to handle the final sync of a file. Works just like the normal
 304  * io path, just does everything sync.
 305  */
 306 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 307 {
 308         struct io_u *io_u = __get_io_u(td);
 309         int ret;
 310
 311         if (!io_u)
 312                 return 1;
 313
 314         io_u->ddir = DDIR_SYNC;
 315         io_u->file = f;
 316
 317         if (td_io_prep(td, io_u)) {
 318                 put_io_u(td, io_u);
 319                 return 1;
 320         }
 321
 322 requeue:
 323         ret = td_io_queue(td, io_u);
 324         if (ret < 0) {
 325                 td_verror(td, io_u->error, "td_io_queue");
 326                 put_io_u(td, io_u);
 327                 return 1;
 328         } else if (ret == FIO_Q_QUEUED) {
 329                 if (io_u_queued_complete(td, 1) < 0)
 330                         return 1;
 331         } else if (ret == FIO_Q_COMPLETED) {
 332                 if (io_u->error) {
 333                         td_verror(td, io_u->error, "td_io_queue");
 334                         return 1;
 335                 }
 336
 337                 if (io_u_sync_complete(td, io_u) < 0)
 338                         return 1;
 339         } else if (ret == FIO_Q_BUSY) {
 340                 if (td_io_commit(td))
 341                         return 1;
 342                 goto requeue;
 343         }
 344
 345         return 0;
 346 }
 347
 348 static inline void update_tv_cache(struct thread_data *td)
 349 {
 350         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 351                 fio_gettime(&td->tv_cache, NULL);
 352 }
 353
 354 /*
 355  * The main verify engine. Runs over the writes we previously submitted,
 356  * reads the blocks back in, and checks the crc/md5 of the data.
 357  */
 358 static void do_verify(struct thread_data *td)
 359 {
 360         struct fio_file *f;
 361         struct io_u *io_u;
 362         int ret, min_events;
 363         unsigned int i;
 364
 365         /*
 366          * sync io first and invalidate cache, to make sure we really
 367          * read from disk.
 368          */
 369         for_each_file(td, f, i) {
 370                 if (!(f->flags & FIO_FILE_OPEN))
 371                         continue;
 372                 if (fio_io_sync(td, f))
 373                         break;
 374                 if (file_invalidate_cache(td, f))
 375                         break;
 376         }
 377
 378         if (td->error)
 379                 return;
 380
 381         td_set_runstate(td, TD_VERIFYING);
 382
 383         io_u = NULL;
 384         while (!td->terminate) {
 385                 int ret2, full;
 386
 387                 io_u = __get_io_u(td);
 388                 if (!io_u)
 389                         break;
 390
 391                 update_tv_cache(td);
 392
 393                 if (runtime_exceeded(td, &td->tv_cache)) {
 394                         put_io_u(td, io_u);
 395                         td->terminate = 1;
 396                         break;
 397                 }
 398
 399                 if (get_next_verify(td, io_u)) {
 400                         put_io_u(td, io_u);
 401                         break;
 402                 }
 403
 404                 if (td_io_prep(td, io_u)) {
 405                         put_io_u(td, io_u);
 406                         break;
 407                 }
 408
 409                 io_u->end_io = verify_io_u;
 410
 411                 ret = td_io_queue(td, io_u);
 412                 switch (ret) {
 413                 case FIO_Q_COMPLETED:
 414                         if (io_u->error)
 415                                 ret = -io_u->error;
 416                         else if (io_u->resid) {
 417                                 int bytes = io_u->xfer_buflen - io_u->resid;
 418                                 struct fio_file *f = io_u->file;
 419
 420                                 /*
 421                                  * zero read, fail
 422                                  */
 423                                 if (!bytes) {
 424                                         td_verror(td, EIO, "full resid");
 425                                         put_io_u(td, io_u);
 426                                         break;
 427                                 }
 428
 429                                 io_u->xfer_buflen = io_u->resid;
 430                                 io_u->xfer_buf += bytes;
 431                                 io_u->offset += bytes;
 432
 433                                 td->ts.short_io_u[io_u->ddir]++;
 434
 435                                 if (io_u->offset == f->real_file_size)
 436                                         goto sync_done;
 437
 438                                 requeue_io_u(td, &io_u);
 439                         } else {
 440 sync_done:
 441                                 ret = io_u_sync_complete(td, io_u);
 442                                 if (ret < 0)
 443                                         break;
 444                         }
 445                         continue;
 446                 case FIO_Q_QUEUED:
 447                         break;
 448                 case FIO_Q_BUSY:
 449                         requeue_io_u(td, &io_u);
 450                         ret2 = td_io_commit(td);
 451                         if (ret2 < 0)
 452                                 ret = ret2;
 453                         break;
 454                 default:
 455                         assert(ret < 0);
 456                         td_verror(td, -ret, "td_io_queue");
 457                         break;
 458                 }
 459
 460                 if (ret < 0 || td->error)
 461                         break;
 462
 463                 /*
 464                  * if we can queue more, do so. but check if there are
 465                  * completed io_u's first.
 466                  */
 467                 full = queue_full(td) || ret == FIO_Q_BUSY;
 468                 if (full || !td->o.iodepth_batch_complete) {
 469                         min_events = td->o.iodepth_batch_complete;
 470                         if (full && !min_events)
 471                                 min_events = 1;
 472
 473                         do {
 474                                 /*
 475                                  * Reap required number of io units, if any,
 476                                  * and do the verification on them through
 477                                  * the callback handler
 478                                  */
 479                                 if (io_u_queued_complete(td, min_events) < 0) {
 480                                         ret = -1;
 481                                         break;
 482                                 }
 483                         } while (full && (td->cur_depth > td->o.iodepth_low));
 484                 }
 485                 if (ret < 0)
 486                         break;
 487         }
 488
 489         if (!td->error) {
 490                 min_events = td->cur_depth;
 491
 492                 if (min_events)
 493                         ret = io_u_queued_complete(td, min_events);
 494         } else
 495                 cleanup_pending_aio(td);
 496
 497         td_set_runstate(td, TD_RUNNING);
 498 }
 499
 500 /*
 501  * Main IO worker function. It retrieves io_u's to process and queues
 502  * and reaps them, checking for rate and errors along the way.
 503  */
 504 static void do_io(struct thread_data *td)
 505 {
 506         unsigned long usec;
 507         unsigned int i;
 508         int ret = 0;
 509
 510         if (in_ramp_time(td))
 511                 td_set_runstate(td, TD_RAMP);
 512         else
 513                 td_set_runstate(td, TD_RUNNING);
 514
 515         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 516                 struct timeval comp_time;
 517                 long bytes_done = 0;
 518                 int min_evts = 0;
 519                 struct io_u *io_u;
 520                 int ret2, full;
 521
 522                 if (td->terminate)
 523                         break;
 524
 525                 io_u = get_io_u(td);
 526                 if (!io_u)
 527                         break;
 528
 529                 update_tv_cache(td);
 530
 531                 if (runtime_exceeded(td, &td->tv_cache)) {
 532                         put_io_u(td, io_u);
 533                         td->terminate = 1;
 534                         break;
 535                 }
 536
 537                 /*
 538                  * Add verification end_io handler, if asked to verify
 539                  * a previously written file.
 540                  */
 541                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 542                         io_u->end_io = verify_io_u;
 543                         td_set_runstate(td, TD_VERIFYING);
 544                 } else if (in_ramp_time(td))
 545                         td_set_runstate(td, TD_RAMP);
 546                 else
 547                         td_set_runstate(td, TD_RUNNING);
 548
 549                 ret = td_io_queue(td, io_u);
 550                 switch (ret) {
 551                 case FIO_Q_COMPLETED:
 552                         if (io_u->error)
 553                                 ret = -io_u->error;
 554                         else if (io_u->resid) {
 555                                 int bytes = io_u->xfer_buflen - io_u->resid;
 556                                 struct fio_file *f = io_u->file;
 557
 558                                 /*
 559                                  * zero read, fail
 560                                  */
 561                                 if (!bytes) {
 562                                         td_verror(td, EIO, "full resid");
 563                                         put_io_u(td, io_u);
 564                                         break;
 565                                 }
 566
 567                                 io_u->xfer_buflen = io_u->resid;
 568                                 io_u->xfer_buf += bytes;
 569                                 io_u->offset += bytes;
 570
 571                                 td->ts.short_io_u[io_u->ddir]++;
 572
 573                                 if (io_u->offset == f->real_file_size)
 574                                         goto sync_done;
 575
 576                                 requeue_io_u(td, &io_u);
 577                         } else {
 578 sync_done:
 579                                 if (should_check_rate(td))
 580                                         fio_gettime(&comp_time, NULL);
 581
 582                                 bytes_done = io_u_sync_complete(td, io_u);
 583                                 if (bytes_done < 0)
 584                                         ret = bytes_done;
 585                         }
 586                         break;
 587                 case FIO_Q_QUEUED:
 588                         /*
 589                          * if the engine doesn't have a commit hook,
 590                          * the io_u is really queued. if it does have such
 591                          * a hook, it has to call io_u_queued() itself.
 592                          */
 593                         if (td->io_ops->commit == NULL)
 594                                 io_u_queued(td, io_u);
 595                         break;
 596                 case FIO_Q_BUSY:
 597                         requeue_io_u(td, &io_u);
 598                         ret2 = td_io_commit(td);
 599                         if (ret2 < 0)
 600                                 ret = ret2;
 601                         break;
 602                 default:
 603                         assert(ret < 0);
 604                         put_io_u(td, io_u);
 605                         break;
 606                 }
 607
 608                 if (ret < 0 || td->error)
 609                         break;
 610
 611                 /*
 612                  * See if we need to complete some commands
 613                  */
 614                 full = queue_full(td) || ret == FIO_Q_BUSY;
 615                 if (full || !td->o.iodepth_batch_complete) {
 616                         min_evts = td->o.iodepth_batch_complete;
 617                         if (full && !min_evts)
 618                                 min_evts = 1;
 619
 620                         if (should_check_rate(td))
 621                                 fio_gettime(&comp_time, NULL);
 622
 623                         do {
 624                                 ret = io_u_queued_complete(td, min_evts);
 625                                 if (ret <= 0)
 626                                         break;
 627
 628                                 bytes_done += ret;
 629                         } while (full && (td->cur_depth > td->o.iodepth_low));
 630                 }
 631
 632                 if (ret < 0)
 633                         break;
 634                 if (!bytes_done)
 635                         continue;
 636
 637                 /*
 638                  * the rate is batched for now, it should work for batches
 639                  * of completions except the very first one which may look
 640                  * a little bursty
 641                  */
 642                 if (!in_ramp_time(td) && should_check_rate(td)) {
 643                         usec = utime_since(&td->tv_cache, &comp_time);
 644
 645                         rate_throttle(td, usec, bytes_done);
 646
 647                         if (check_min_rate(td, &comp_time)) {
 648                                 if (exitall_on_terminate)
 649                                         terminate_threads(td->groupid);
 650                                 td_verror(td, EIO, "check_min_rate");
 651                                 break;
 652                         }
 653                 }
 654
 655                 if (td->o.thinktime) {
 656                         unsigned long long b;
 657
 658                         b = td->io_blocks[0] + td->io_blocks[1];
 659                         if (!(b % td->o.thinktime_blocks)) {
 660                                 int left;
 661
 662                                 if (td->o.thinktime_spin)
 663                                         __usec_sleep(td->o.thinktime_spin);
 664
 665                                 left = td->o.thinktime - td->o.thinktime_spin;
 666                                 if (left)
 667                                         usec_sleep(td, left);
 668                         }
 669                 }
 670         }
 671
 672         if (td->o.fill_device && td->error == ENOSPC) {
 673                 td->error = 0;
 674                 td->terminate = 1;
 675         }
 676         if (!td->error) {
 677                 struct fio_file *f;
 678
 679                 i = td->cur_depth;
 680                 if (i)
 681                         ret = io_u_queued_complete(td, i);
 682
 683                 if (should_fsync(td) && td->o.end_fsync) {
 684                         td_set_runstate(td, TD_FSYNCING);
 685
 686                         for_each_file(td, f, i) {
 687                                 if (!(f->flags & FIO_FILE_OPEN))
 688                                         continue;
 689                                 fio_io_sync(td, f);
 690                         }
 691                 }
 692         } else
 693                 cleanup_pending_aio(td);
 694
 695         /*
 696          * stop job if we failed doing any IO
 697          */
 698         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 699                 td->done = 1;
 700 }
 701
 702 static void cleanup_io_u(struct thread_data *td)
 703 {
 704         struct flist_head *entry, *n;
 705         struct io_u *io_u;
 706
 707         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 708                 io_u = flist_entry(entry, struct io_u, list);
 709
 710                 flist_del(&io_u->list);
 711                 free(io_u);
 712         }
 713
 714         free_io_mem(td);
 715 }
 716
 717 static int init_io_u(struct thread_data *td)
 718 {
 719         struct io_u *io_u;
 720         unsigned int max_bs;
 721         int i, max_units;
 722         char *p;
 723
 724         max_units = td->o.iodepth;
 725         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 726         td->orig_buffer_size = (unsigned long long) max_bs
 727                                         * (unsigned long long) max_units;
 728
 729         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 730                 unsigned long bs;
 731
 732                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 733                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 734         }
 735
 736         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 737                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 738                 return 1;
 739         }
 740
 741         if (allocate_io_mem(td))
 742                 return 1;
 743
 744         if (td->o.odirect)
 745                 p = ALIGN(td->orig_buffer);
 746         else
 747                 p = td->orig_buffer;
 748
 749         for (i = 0; i < max_units; i++) {
 750                 if (td->terminate)
 751                         return 1;
 752                 io_u = malloc(sizeof(*io_u));
 753                 memset(io_u, 0, sizeof(*io_u));
 754                 INIT_FLIST_HEAD(&io_u->list);
 755
 756                 if (!(td->io_ops->flags & FIO_NOIO)) {
 757                         io_u->buf = p + max_bs * i;
 758
 759                         if (td_write(td) && !td->o.refill_buffers)
 760                                 io_u_fill_buffer(td, io_u, max_bs);
 761                 }
 762
 763                 io_u->index = i;
 764                 io_u->flags = IO_U_F_FREE;
 765                 flist_add(&io_u->list, &td->io_u_freelist);
 766         }
 767
 768         return 0;
 769 }
 770
 771 static int switch_ioscheduler(struct thread_data *td)
 772 {
 773         char tmp[256], tmp2[128];
 774         FILE *f;
 775         int ret;
 776
 777         if (td->io_ops->flags & FIO_DISKLESSIO)
 778                 return 0;
 779
 780         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 781
 782         f = fopen(tmp, "r+");
 783         if (!f) {
 784                 if (errno == ENOENT) {
 785                         log_err("fio: os or kernel doesn't support IO scheduler"
 786                                 " switching\n");
 787                         return 0;
 788                 }
 789                 td_verror(td, errno, "fopen iosched");
 790                 return 1;
 791         }
 792
 793         /*
 794          * Set io scheduler.
 795          */
 796         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 797         if (ferror(f) || ret != 1) {
 798                 td_verror(td, errno, "fwrite");
 799                 fclose(f);
 800                 return 1;
 801         }
 802
 803         rewind(f);
 804
 805         /*
 806          * Read back and check that the selected scheduler is now the default.
 807          */
 808         ret = fread(tmp, 1, sizeof(tmp), f);
 809         if (ferror(f) || ret < 0) {
 810                 td_verror(td, errno, "fread");
 811                 fclose(f);
 812                 return 1;
 813         }
 814
 815         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 816         if (!strstr(tmp, tmp2)) {
 817                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 818                 td_verror(td, EINVAL, "iosched_switch");
 819                 fclose(f);
 820                 return 1;
 821         }
 822
 823         fclose(f);
 824         return 0;
 825 }
 826
 827 static int keep_running(struct thread_data *td)
 828 {
 829         unsigned long long io_done;
 830
 831         if (td->done)
 832                 return 0;
 833         if (td->o.time_based)
 834                 return 1;
 835         if (td->o.loops) {
 836                 td->o.loops--;
 837                 return 1;
 838         }
 839
 840         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 841                         + td->io_skip_bytes;
 842         if (io_done < td->o.size)
 843                 return 1;
 844
 845         return 0;
 846 }
 847
 848 static void reset_io_counters(struct thread_data *td)
 849 {
 850         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 851         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 852         td->zone_bytes = 0;
 853         td->rate_bytes = 0;
 854         td->rate_blocks = 0;
 855         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 856
 857         td->last_was_sync = 0;
 858
 859         /*
 860          * reset file done count if we are to start over
 861          */
 862         if (td->o.time_based || td->o.loops)
 863                 td->nr_done_files = 0;
 864 }
 865
 866 void reset_all_stats(struct thread_data *td)
 867 {
 868         struct timeval tv;
 869         int i;
 870
 871         reset_io_counters(td);
 872
 873         for (i = 0; i < 2; i++) {
 874                 td->io_bytes[i] = 0;
 875                 td->io_blocks[i] = 0;
 876                 td->io_issues[i] = 0;
 877                 td->ts.total_io_u[i] = 0;
 878         }
 879
 880         fio_gettime(&tv, NULL);
 881         memcpy(&td->epoch, &tv, sizeof(tv));
 882         memcpy(&td->start, &tv, sizeof(tv));
 883 }
 884
 885 static int clear_io_state(struct thread_data *td)
 886 {
 887         struct fio_file *f;
 888         unsigned int i;
 889         int ret;
 890
 891         reset_io_counters(td);
 892
 893         close_files(td);
 894
 895         ret = 0;
 896         for_each_file(td, f, i) {
 897                 f->flags &= ~FIO_FILE_DONE;
 898                 ret = td_io_open_file(td, f);
 899                 if (ret)
 900                         break;
 901         }
 902
 903         return ret;
 904 }
 905
 906 /*
 907  * Entry point for the thread based jobs. The process based jobs end up
 908  * here as well, after a little setup.
 909  */
 910 static void *thread_main(void *data)
 911 {
 912         unsigned long long runtime[2], elapsed;
 913         struct thread_data *td = data;
 914         int clear_state;
 915
 916         if (!td->o.use_thread)
 917                 setsid();
 918
 919         td->pid = getpid();
 920
 921         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 922
 923         INIT_FLIST_HEAD(&td->io_u_freelist);
 924         INIT_FLIST_HEAD(&td->io_u_busylist);
 925         INIT_FLIST_HEAD(&td->io_u_requeues);
 926         INIT_FLIST_HEAD(&td->io_log_list);
 927         INIT_FLIST_HEAD(&td->io_hist_list);
 928         td->io_hist_tree = RB_ROOT;
 929
 930         td_set_runstate(td, TD_INITIALIZED);
 931         fio_mutex_up(startup_mutex);
 932         fio_mutex_down(td->mutex);
 933
 934         /*
 935          * the ->mutex mutex is now no longer used, close it to avoid
 936          * eating a file descriptor
 937          */
 938         fio_mutex_remove(td->mutex);
 939
 940         /*
 941          * May alter parameters that init_io_u() will use, so we need to
 942          * do this first.
 943          */
 944         if (init_iolog(td))
 945                 goto err;
 946
 947         if (init_io_u(td))
 948                 goto err;
 949
 950         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 951                 td_verror(td, errno, "cpu_set_affinity");
 952                 goto err;
 953         }
 954
 955         if (td->ioprio_set) {
 956                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 957                         td_verror(td, errno, "ioprio_set");
 958                         goto err;
 959                 }
 960         }
 961
 962         if (nice(td->o.nice) == -1) {
 963                 td_verror(td, errno, "nice");
 964                 goto err;
 965         }
 966
 967         if (td->o.ioscheduler && switch_ioscheduler(td))
 968                 goto err;
 969
 970         if (!td->o.create_serialize && setup_files(td))
 971                 goto err;
 972
 973         if (td_io_init(td))
 974                 goto err;
 975
 976         if (open_files(td))
 977                 goto err;
 978
 979         if (init_random_map(td))
 980                 goto err;
 981
 982         if (td->o.exec_prerun) {
 983                 if (system(td->o.exec_prerun) < 0)
 984                         goto err;
 985         }
 986
 987         fio_gettime(&td->epoch, NULL);
 988         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 989
 990         runtime[0] = runtime[1] = 0;
 991         clear_state = 0;
 992         while (keep_running(td)) {
 993                 fio_gettime(&td->start, NULL);
 994                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 995                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
 996
 997                 if (td->o.ratemin)
 998                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 999                                                         sizeof(td->lastrate));
1000
1001                 if (clear_state && clear_io_state(td))
1002                         break;
1003
1004                 prune_io_piece_log(td);
1005
1006                 do_io(td);
1007
1008                 clear_state = 1;
1009
1010                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1011                         if (td->rw_end_set[DDIR_READ])
1012                                 elapsed = utime_since(&td->start,
1013                                                       &td->rw_end[DDIR_READ]);
1014                         else
1015                                 elapsed = utime_since_now(&td->start);
1016
1017                         runtime[DDIR_READ] += elapsed;
1018                 }
1019                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1020                         if (td->rw_end_set[DDIR_WRITE])
1021                                 elapsed = utime_since(&td->start,
1022                                                       &td->rw_end[DDIR_WRITE]);
1023                         else
1024                                 elapsed = utime_since_now(&td->start);
1025
1026                         runtime[DDIR_WRITE] += elapsed;
1027                 }
1028
1029                 if (td->error || td->terminate)
1030                         break;
1031
1032                 if (!td->o.do_verify ||
1033                     td->o.verify == VERIFY_NONE ||
1034                     (td->io_ops->flags & FIO_UNIDIR))
1035                         continue;
1036
1037                 if (clear_io_state(td))
1038                         break;
1039
1040                 fio_gettime(&td->start, NULL);
1041
1042                 do_verify(td);
1043
1044                 runtime[DDIR_READ] += utime_since_now(&td->start);
1045
1046                 if (td->error || td->terminate)
1047                         break;
1048         }
1049
1050         update_rusage_stat(td);
1051         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1052         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1053         td->ts.total_run_time = mtime_since_now(&td->epoch);
1054         td->ts.io_bytes[0] = td->io_bytes[0];
1055         td->ts.io_bytes[1] = td->io_bytes[1];
1056
1057         if (td->ts.bw_log)
1058                 finish_log(td, td->ts.bw_log, "bw");
1059         if (td->ts.slat_log)
1060                 finish_log(td, td->ts.slat_log, "slat");
1061         if (td->ts.clat_log)
1062                 finish_log(td, td->ts.clat_log, "clat");
1063         if (td->o.exec_postrun) {
1064                 if (system(td->o.exec_postrun) < 0)
1065                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
1066         }
1067
1068         if (exitall_on_terminate)
1069                 terminate_threads(td->groupid);
1070
1071 err:
1072         if (td->error)
1073                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1074                                                         td->verror);
1075         close_and_free_files(td);
1076         close_ioengine(td);
1077         cleanup_io_u(td);
1078
1079         /*
1080          * do this very late, it will log file closing as well
1081          */
1082         if (td->o.write_iolog_file)
1083                 write_iolog_close(td);
1084
1085         options_mem_free(td);
1086         td_set_runstate(td, TD_EXITED);
1087         return (void *) (unsigned long) td->error;
1088 }
1089
1090 /*
1091  * We cannot pass the td data into a forked process, so attach the td and
1092  * pass it to the thread worker.
1093  */
1094 static int fork_main(int shmid, int offset)
1095 {
1096         struct thread_data *td;
1097         void *data, *ret;
1098
1099         data = shmat(shmid, NULL, 0);
1100         if (data == (void *) -1) {
1101                 int __err = errno;
1102
1103                 perror("shmat");
1104                 return __err;
1105         }
1106
1107         td = data + offset * sizeof(struct thread_data);
1108         ret = thread_main(td);
1109         shmdt(data);
1110         return (int) (unsigned long) ret;
1111 }
1112
1113 /*
1114  * Run over the job map and reap the threads that have exited, if any.
1115  */
1116 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1117 {
1118         struct thread_data *td;
1119         int i, cputhreads, realthreads, pending, status, ret;
1120
1121         /*
1122          * reap exited threads (TD_EXITED -> TD_REAPED)
1123          */
1124         realthreads = pending = cputhreads = 0;
1125         for_each_td(td, i) {
1126                 int flags = 0;
1127
1128                 /*
1129                  * ->io_ops is NULL for a thread that has closed its
1130                  * io engine
1131                  */
1132                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1133                         cputhreads++;
1134                 else
1135                         realthreads++;
1136
1137                 if (!td->pid) {
1138                         pending++;
1139                         continue;
1140                 }
1141                 if (td->runstate == TD_REAPED)
1142                         continue;
1143                 if (td->o.use_thread) {
1144                         if (td->runstate == TD_EXITED) {
1145                                 td_set_runstate(td, TD_REAPED);
1146                                 goto reaped;
1147                         }
1148                         continue;
1149                 }
1150
1151                 flags = WNOHANG;
1152                 if (td->runstate == TD_EXITED)
1153                         flags = 0;
1154
1155                 /*
1156                  * check if someone quit or got killed in an unusual way
1157                  */
1158                 ret = waitpid(td->pid, &status, flags);
1159                 if (ret < 0) {
1160                         if (errno == ECHILD) {
1161                                 log_err("fio: pid=%d disappeared %d\n",
1162                                                 (int) td->pid, td->runstate);
1163                                 td_set_runstate(td, TD_REAPED);
1164                                 goto reaped;
1165                         }
1166                         perror("waitpid");
1167                 } else if (ret == td->pid) {
1168                         if (WIFSIGNALED(status)) {
1169                                 int sig = WTERMSIG(status);
1170
1171                                 if (sig != SIGQUIT)
1172                                         log_err("fio: pid=%d, got signal=%d\n",
1173                                                         (int) td->pid, sig);
1174                                 td_set_runstate(td, TD_REAPED);
1175                                 goto reaped;
1176                         }
1177                         if (WIFEXITED(status)) {
1178                                 if (WEXITSTATUS(status) && !td->error)
1179                                         td->error = WEXITSTATUS(status);
1180
1181                                 td_set_runstate(td, TD_REAPED);
1182                                 goto reaped;
1183                         }
1184                 }
1185
1186                 /*
1187                  * thread is not dead, continue
1188                  */
1189                 pending++;
1190                 continue;
1191 reaped:
1192                 (*nr_running)--;
1193                 (*m_rate) -= td->o.ratemin;
1194                 (*t_rate) -= td->o.rate;
1195                 if (!td->pid)
1196                         pending--;
1197
1198                 if (td->error)
1199                         exit_value++;
1200
1201                 done_secs += mtime_since_now(&td->epoch) / 1000;
1202         }
1203
1204         if (*nr_running == cputhreads && !pending && realthreads)
1205                 terminate_threads(TERMINATE_ALL);
1206 }
1207
1208 /*
1209  * Main function for kicking off and reaping jobs, as needed.
1210  */
1211 static void run_threads(void)
1212 {
1213         struct thread_data *td;
1214         unsigned long spent;
1215         int i, todo, nr_running, m_rate, t_rate, nr_started;
1216
1217         if (fio_pin_memory())
1218                 return;
1219
1220         if (!terse_output) {
1221                 printf("Starting ");
1222                 if (nr_thread)
1223                         printf("%d thread%s", nr_thread,
1224                                                 nr_thread > 1 ? "s" : "");
1225                 if (nr_process) {
1226                         if (nr_thread)
1227                                 printf(" and ");
1228                         printf("%d process%s", nr_process,
1229                                                 nr_process > 1 ? "es" : "");
1230                 }
1231                 printf("\n");
1232                 fflush(stdout);
1233         }
1234
1235         set_sig_handlers();
1236
1237         todo = thread_number;
1238         nr_running = 0;
1239         nr_started = 0;
1240         m_rate = t_rate = 0;
1241
1242         for_each_td(td, i) {
1243                 print_status_init(td->thread_number - 1);
1244
1245                 if (!td->o.create_serialize) {
1246                         init_disk_util(td);
1247                         continue;
1248                 }
1249
1250                 /*
1251                  * do file setup here so it happens sequentially,
1252                  * we don't want X number of threads getting their
1253                  * client data interspersed on disk
1254                  */
1255                 if (setup_files(td)) {
1256                         exit_value++;
1257                         if (td->error)
1258                                 log_err("fio: pid=%d, err=%d/%s\n",
1259                                         (int) td->pid, td->error, td->verror);
1260                         td_set_runstate(td, TD_REAPED);
1261                         todo--;
1262                 } else {
1263                         struct fio_file *f;
1264                         unsigned int i;
1265
1266                         /*
1267                          * for sharing to work, each job must always open
1268                          * its own files. so close them, if we opened them
1269                          * for creation
1270                          */
1271                         for_each_file(td, f, i)
1272                                 td_io_close_file(td, f);
1273                 }
1274
1275                 init_disk_util(td);
1276         }
1277
1278         set_genesis_time();
1279
1280         while (todo) {
1281                 struct thread_data *map[MAX_JOBS];
1282                 struct timeval this_start;
1283                 int this_jobs = 0, left;
1284
1285                 /*
1286                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1287                  */
1288                 for_each_td(td, i) {
1289                         if (td->runstate != TD_NOT_CREATED)
1290                                 continue;
1291
1292                         /*
1293                          * never got a chance to start, killed by other
1294                          * thread for some reason
1295                          */
1296                         if (td->terminate) {
1297                                 todo--;
1298                                 continue;
1299                         }
1300
1301                         if (td->o.start_delay) {
1302                                 spent = mtime_since_genesis();
1303
1304                                 if (td->o.start_delay * 1000 > spent)
1305                                         continue;
1306                         }
1307
1308                         if (td->o.stonewall && (nr_started || nr_running)) {
1309                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1310                                                         td->o.name);
1311                                 break;
1312                         }
1313
1314                         /*
1315                          * Set state to created. Thread will transition
1316                          * to TD_INITIALIZED when it's done setting up.
1317                          */
1318                         td_set_runstate(td, TD_CREATED);
1319                         map[this_jobs++] = td;
1320                         nr_started++;
1321
1322                         if (td->o.use_thread) {
1323                                 dprint(FD_PROCESS, "will pthread_create\n");
1324                                 if (pthread_create(&td->thread, NULL,
1325                                                    thread_main, td)) {
1326                                         perror("pthread_create");
1327                                         nr_started--;
1328                                         break;
1329                                 }
1330                                 if (pthread_detach(td->thread) < 0)
1331                                         perror("pthread_detach");
1332                         } else {
1333                                 pid_t pid;
1334                                 dprint(FD_PROCESS, "will fork\n");
1335                                 pid = fork();
1336                                 if (!pid) {
1337                                         int ret = fork_main(shm_id, i);
1338
1339                                         _exit(ret);
1340                                 } else if (i == fio_debug_jobno)
1341                                         *fio_debug_jobp = pid;
1342                         }
1343                         fio_mutex_down(startup_mutex);
1344                 }
1345
1346                 /*
1347                  * Wait for the started threads to transition to
1348                  * TD_INITIALIZED.
1349                  */
1350                 fio_gettime(&this_start, NULL);
1351                 left = this_jobs;
1352                 while (left && !fio_abort) {
1353                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1354                                 break;
1355
1356                         usleep(100000);
1357
1358                         for (i = 0; i < this_jobs; i++) {
1359                                 td = map[i];
1360                                 if (!td)
1361                                         continue;
1362                                 if (td->runstate == TD_INITIALIZED) {
1363                                         map[i] = NULL;
1364                                         left--;
1365                                 } else if (td->runstate >= TD_EXITED) {
1366                                         map[i] = NULL;
1367                                         left--;
1368                                         todo--;
1369                                         nr_running++; /* work-around... */
1370                                 }
1371                         }
1372                 }
1373
1374                 if (left) {
1375                         log_err("fio: %d jobs failed to start\n", left);
1376                         for (i = 0; i < this_jobs; i++) {
1377                                 td = map[i];
1378                                 if (!td)
1379                                         continue;
1380                                 kill(td->pid, SIGTERM);
1381                         }
1382                         break;
1383                 }
1384
1385                 /*
1386                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1387                  */
1388                 for_each_td(td, i) {
1389                         if (td->runstate != TD_INITIALIZED)
1390                                 continue;
1391
1392                         if (in_ramp_time(td))
1393                                 td_set_runstate(td, TD_RAMP);
1394                         else
1395                                 td_set_runstate(td, TD_RUNNING);
1396                         nr_running++;
1397                         nr_started--;
1398                         m_rate += td->o.ratemin;
1399                         t_rate += td->o.rate;
1400                         todo--;
1401                         fio_mutex_up(td->mutex);
1402                 }
1403
1404                 reap_threads(&nr_running, &t_rate, &m_rate);
1405
1406                 if (todo)
1407                         usleep(100000);
1408         }
1409
1410         while (nr_running) {
1411                 reap_threads(&nr_running, &t_rate, &m_rate);
1412                 usleep(10000);
1413         }
1414
1415         update_io_ticks();
1416         fio_unpin_memory();
1417 }
1418
1419 int main(int argc, char *argv[])
1420 {
1421         long ps;
1422
1423         sinit();
1424
1425         /*
1426          * We need locale for number printing, if it isn't set then just
1427          * go with the US format.
1428          */
1429         if (!getenv("LC_NUMERIC"))
1430                 setlocale(LC_NUMERIC, "en_US");
1431
1432         if (parse_options(argc, argv))
1433                 return 1;
1434
1435         if (!thread_number)
1436                 return 0;
1437
1438         ps = sysconf(_SC_PAGESIZE);
1439         if (ps < 0) {
1440                 log_err("Failed to get page size\n");
1441                 return 1;
1442         }
1443
1444         page_size = ps;
1445         page_mask = ps - 1;
1446
1447         if (write_bw_log) {
1448                 setup_log(&agg_io_log[DDIR_READ]);
1449                 setup_log(&agg_io_log[DDIR_WRITE]);
1450         }
1451
1452         startup_mutex = fio_mutex_init(0);
1453
1454         set_genesis_time();
1455
1456         status_timer_arm();
1457
1458         run_threads();
1459
1460         if (!fio_abort) {
1461                 show_run_stats();
1462                 if (write_bw_log) {
1463                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1464                         __finish_log(agg_io_log[DDIR_WRITE],
1465                                         "agg-write_bw.log");
1466                 }
1467         }
1468
1469         fio_mutex_remove(startup_mutex);
1470         return exit_value;
1471 }