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 /*
 120  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
 121  */
 122 static void sig_quit(int sig)
 123 {
 124 }
 125
 126 static void sig_int(int sig)
 127 {
 128         if (threads) {
 129                 printf("\nfio: terminating on signal %d\n", sig);
 130                 fflush(stdout);
 131                 terminate_threads(TERMINATE_ALL);
 132         }
 133 }
 134
 135 static void sig_ill(int fio_unused sig)
 136 {
 137         if (!threads)
 138                 return;
 139
 140         log_err("fio: illegal instruction. your cpu does not support "
 141                 "the sse4.2 instruction for crc32c\n");
 142         terminate_threads(TERMINATE_ALL);
 143         exit(4);
 144 }
 145
 146 static void set_sig_handlers(void)
 147 {
 148         struct sigaction act;
 149
 150         memset(&act, 0, sizeof(act));
 151         act.sa_handler = sig_alrm;
 152         act.sa_flags = SA_RESTART;
 153         sigaction(SIGALRM, &act, NULL);
 154
 155         memset(&act, 0, sizeof(act));
 156         act.sa_handler = sig_int;
 157         act.sa_flags = SA_RESTART;
 158         sigaction(SIGINT, &act, NULL);
 159
 160         memset(&act, 0, sizeof(act));
 161         act.sa_handler = sig_ill;
 162         act.sa_flags = SA_RESTART;
 163         sigaction(SIGILL, &act, NULL);
 164
 165         memset(&act, 0, sizeof(act));
 166         act.sa_handler = sig_quit;
 167         act.sa_flags = SA_RESTART;
 168         sigaction(SIGQUIT, &act, NULL);
 169 }
 170
 171 static inline int should_check_rate(struct thread_data *td)
 172 {
 173         /*
 174          * No minimum rate set, always ok
 175          */
 176         if (!td->o.ratemin && !td->o.rate_iops_min)
 177                 return 0;
 178
 179         return 1;
 180 }
 181
 182 /*
 183  * Check if we are above the minimum rate given.
 184  */
 185 static int check_min_rate(struct thread_data *td, struct timeval *now)
 186 {
 187         unsigned long long bytes = 0;
 188         unsigned long iops = 0;
 189         unsigned long spent;
 190         unsigned long rate;
 191
 192         /*
 193          * allow a 2 second settle period in the beginning
 194          */
 195         if (mtime_since(&td->start, now) < 2000)
 196                 return 0;
 197
 198         if (td_read(td)) {
 199                 iops += td->io_blocks[DDIR_READ];
 200                 bytes += td->this_io_bytes[DDIR_READ];
 201         }
 202         if (td_write(td)) {
 203                 iops += td->io_blocks[DDIR_WRITE];
 204                 bytes += td->this_io_bytes[DDIR_WRITE];
 205         }
 206
 207         /*
 208          * if rate blocks is set, sample is running
 209          */
 210         if (td->rate_bytes || td->rate_blocks) {
 211                 spent = mtime_since(&td->lastrate, now);
 212                 if (spent < td->o.ratecycle)
 213                         return 0;
 214
 215                 if (td->o.rate) {
 216                         /*
 217                          * check bandwidth specified rate
 218                          */
 219                         if (bytes < td->rate_bytes) {
 220                                 log_err("%s: min rate %u not met\n", td->o.name,
 221                                                                 td->o.ratemin);
 222                                 return 1;
 223                         } else {
 224                                 rate = (bytes - td->rate_bytes) / spent;
 225                                 if (rate < td->o.ratemin ||
 226                                     bytes < td->rate_bytes) {
 227                                         log_err("%s: min rate %u not met, got"
 228                                                 " %luKiB/sec\n", td->o.name,
 229                                                         td->o.ratemin, rate);
 230                                         return 1;
 231                                 }
 232                         }
 233                 } else {
 234                         /*
 235                          * checks iops specified rate
 236                          */
 237                         if (iops < td->o.rate_iops) {
 238                                 log_err("%s: min iops rate %u not met\n",
 239                                                 td->o.name, td->o.rate_iops);
 240                                 return 1;
 241                         } else {
 242                                 rate = (iops - td->rate_blocks) / spent;
 243                                 if (rate < td->o.rate_iops_min ||
 244                                     iops < td->rate_blocks) {
 245                                         log_err("%s: min iops rate %u not met,"
 246                                                 " got %lu\n", td->o.name,
 247                                                         td->o.rate_iops_min,
 248                                                         rate);
 249                                 }
 250                         }
 251                 }
 252         }
 253
 254         td->rate_bytes = bytes;
 255         td->rate_blocks = iops;
 256         memcpy(&td->lastrate, now, sizeof(*now));
 257         return 0;
 258 }
 259
 260 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 261 {
 262         if (!td->o.timeout)
 263                 return 0;
 264         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 265                 return 1;
 266
 267         return 0;
 268 }
 269
 270 /*
 271  * When job exits, we can cancel the in-flight IO if we are using async
 272  * io. Attempt to do so.
 273  */
 274 static void cleanup_pending_aio(struct thread_data *td)
 275 {
 276         struct flist_head *entry, *n;
 277         struct io_u *io_u;
 278         int r;
 279
 280         /*
 281          * get immediately available events, if any
 282          */
 283         r = io_u_queued_complete(td, 0);
 284         if (r < 0)
 285                 return;
 286
 287         /*
 288          * now cancel remaining active events
 289          */
 290         if (td->io_ops->cancel) {
 291                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 292                         io_u = flist_entry(entry, struct io_u, list);
 293
 294                         /*
 295                          * if the io_u isn't in flight, then that generally
 296                          * means someone leaked an io_u. complain but fix
 297                          * it up, so we don't stall here.
 298                          */
 299                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 300                                 log_err("fio: non-busy IO on busy list\n");
 301                                 put_io_u(td, io_u);
 302                         } else {
 303                                 r = td->io_ops->cancel(td, io_u);
 304                                 if (!r)
 305                                         put_io_u(td, io_u);
 306                         }
 307                 }
 308         }
 309
 310         if (td->cur_depth)
 311                 r = io_u_queued_complete(td, td->cur_depth);
 312 }
 313
 314 /*
 315  * Helper to handle the final sync of a file. Works just like the normal
 316  * io path, just does everything sync.
 317  */
 318 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 319 {
 320         struct io_u *io_u = __get_io_u(td);
 321         int ret;
 322
 323         if (!io_u)
 324                 return 1;
 325
 326         io_u->ddir = DDIR_SYNC;
 327         io_u->file = f;
 328
 329         if (td_io_prep(td, io_u)) {
 330                 put_io_u(td, io_u);
 331                 return 1;
 332         }
 333
 334 requeue:
 335         ret = td_io_queue(td, io_u);
 336         if (ret < 0) {
 337                 td_verror(td, io_u->error, "td_io_queue");
 338                 put_io_u(td, io_u);
 339                 return 1;
 340         } else if (ret == FIO_Q_QUEUED) {
 341                 if (io_u_queued_complete(td, 1) < 0)
 342                         return 1;
 343         } else if (ret == FIO_Q_COMPLETED) {
 344                 if (io_u->error) {
 345                         td_verror(td, io_u->error, "td_io_queue");
 346                         return 1;
 347                 }
 348
 349                 if (io_u_sync_complete(td, io_u) < 0)
 350                         return 1;
 351         } else if (ret == FIO_Q_BUSY) {
 352                 if (td_io_commit(td))
 353                         return 1;
 354                 goto requeue;
 355         }
 356
 357         return 0;
 358 }
 359
 360 static inline void update_tv_cache(struct thread_data *td)
 361 {
 362         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 363                 fio_gettime(&td->tv_cache, NULL);
 364 }
 365
 366 /*
 367  * The main verify engine. Runs over the writes we previously submitted,
 368  * reads the blocks back in, and checks the crc/md5 of the data.
 369  */
 370 static void do_verify(struct thread_data *td)
 371 {
 372         struct fio_file *f;
 373         struct io_u *io_u;
 374         int ret, min_events;
 375         unsigned int i;
 376
 377         /*
 378          * sync io first and invalidate cache, to make sure we really
 379          * read from disk.
 380          */
 381         for_each_file(td, f, i) {
 382                 if (!(f->flags & FIO_FILE_OPEN))
 383                         continue;
 384                 if (fio_io_sync(td, f))
 385                         break;
 386                 if (file_invalidate_cache(td, f))
 387                         break;
 388         }
 389
 390         if (td->error)
 391                 return;
 392
 393         td_set_runstate(td, TD_VERIFYING);
 394
 395         io_u = NULL;
 396         while (!td->terminate) {
 397                 int ret2, full;
 398
 399                 io_u = __get_io_u(td);
 400                 if (!io_u)
 401                         break;
 402
 403                 update_tv_cache(td);
 404
 405                 if (runtime_exceeded(td, &td->tv_cache)) {
 406                         put_io_u(td, io_u);
 407                         td->terminate = 1;
 408                         break;
 409                 }
 410
 411                 if (get_next_verify(td, io_u)) {
 412                         put_io_u(td, io_u);
 413                         break;
 414                 }
 415
 416                 if (td_io_prep(td, io_u)) {
 417                         put_io_u(td, io_u);
 418                         break;
 419                 }
 420
 421                 io_u->end_io = verify_io_u;
 422
 423                 ret = td_io_queue(td, io_u);
 424                 switch (ret) {
 425                 case FIO_Q_COMPLETED:
 426                         if (io_u->error)
 427                                 ret = -io_u->error;
 428                         else if (io_u->resid) {
 429                                 int bytes = io_u->xfer_buflen - io_u->resid;
 430                                 struct fio_file *f = io_u->file;
 431
 432                                 /*
 433                                  * zero read, fail
 434                                  */
 435                                 if (!bytes) {
 436                                         td_verror(td, EIO, "full resid");
 437                                         put_io_u(td, io_u);
 438                                         break;
 439                                 }
 440
 441                                 io_u->xfer_buflen = io_u->resid;
 442                                 io_u->xfer_buf += bytes;
 443                                 io_u->offset += bytes;
 444
 445                                 td->ts.short_io_u[io_u->ddir]++;
 446
 447                                 if (io_u->offset == f->real_file_size)
 448                                         goto sync_done;
 449
 450                                 requeue_io_u(td, &io_u);
 451                         } else {
 452 sync_done:
 453                                 ret = io_u_sync_complete(td, io_u);
 454                                 if (ret < 0)
 455                                         break;
 456                         }
 457                         continue;
 458                 case FIO_Q_QUEUED:
 459                         break;
 460                 case FIO_Q_BUSY:
 461                         requeue_io_u(td, &io_u);
 462                         ret2 = td_io_commit(td);
 463                         if (ret2 < 0)
 464                                 ret = ret2;
 465                         break;
 466                 default:
 467                         assert(ret < 0);
 468                         td_verror(td, -ret, "td_io_queue");
 469                         break;
 470                 }
 471
 472                 if (ret < 0 || td->error)
 473                         break;
 474
 475                 /*
 476                  * if we can queue more, do so. but check if there are
 477                  * completed io_u's first.
 478                  */
 479                 full = queue_full(td) || ret == FIO_Q_BUSY;
 480                 if (full || !td->o.iodepth_batch_complete) {
 481                         min_events = td->o.iodepth_batch_complete;
 482                         if (full && !min_events)
 483                                 min_events = 1;
 484
 485                         do {
 486                                 /*
 487                                  * Reap required number of io units, if any,
 488                                  * and do the verification on them through
 489                                  * the callback handler
 490                                  */
 491                                 if (io_u_queued_complete(td, min_events) < 0) {
 492                                         ret = -1;
 493                                         break;
 494                                 }
 495                         } while (full && (td->cur_depth > td->o.iodepth_low));
 496                 }
 497                 if (ret < 0)
 498                         break;
 499         }
 500
 501         if (!td->error) {
 502                 min_events = td->cur_depth;
 503
 504                 if (min_events)
 505                         ret = io_u_queued_complete(td, min_events);
 506         } else
 507                 cleanup_pending_aio(td);
 508
 509         td_set_runstate(td, TD_RUNNING);
 510 }
 511
 512 /*
 513  * Main IO worker function. It retrieves io_u's to process and queues
 514  * and reaps them, checking for rate and errors along the way.
 515  */
 516 static void do_io(struct thread_data *td)
 517 {
 518         unsigned long usec;
 519         unsigned int i;
 520         int ret = 0;
 521
 522         if (in_ramp_time(td))
 523                 td_set_runstate(td, TD_RAMP);
 524         else
 525                 td_set_runstate(td, TD_RUNNING);
 526
 527         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 528                 struct timeval comp_time;
 529                 long bytes_done = 0;
 530                 int min_evts = 0;
 531                 struct io_u *io_u;
 532                 int ret2, full;
 533
 534                 if (td->terminate)
 535                         break;
 536
 537                 io_u = get_io_u(td);
 538                 if (!io_u)
 539                         break;
 540
 541                 update_tv_cache(td);
 542
 543                 if (runtime_exceeded(td, &td->tv_cache)) {
 544                         put_io_u(td, io_u);
 545                         td->terminate = 1;
 546                         break;
 547                 }
 548
 549                 /*
 550                  * Add verification end_io handler, if asked to verify
 551                  * a previously written file.
 552                  */
 553                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 554                         io_u->end_io = verify_io_u;
 555                         td_set_runstate(td, TD_VERIFYING);
 556                 } else if (in_ramp_time(td))
 557                         td_set_runstate(td, TD_RAMP);
 558                 else
 559                         td_set_runstate(td, TD_RUNNING);
 560
 561                 ret = td_io_queue(td, io_u);
 562                 switch (ret) {
 563                 case FIO_Q_COMPLETED:
 564                         if (io_u->error)
 565                                 ret = -io_u->error;
 566                         else if (io_u->resid) {
 567                                 int bytes = io_u->xfer_buflen - io_u->resid;
 568                                 struct fio_file *f = io_u->file;
 569
 570                                 /*
 571                                  * zero read, fail
 572                                  */
 573                                 if (!bytes) {
 574                                         td_verror(td, EIO, "full resid");
 575                                         put_io_u(td, io_u);
 576                                         break;
 577                                 }
 578
 579                                 io_u->xfer_buflen = io_u->resid;
 580                                 io_u->xfer_buf += bytes;
 581                                 io_u->offset += bytes;
 582
 583                                 td->ts.short_io_u[io_u->ddir]++;
 584
 585                                 if (io_u->offset == f->real_file_size)
 586                                         goto sync_done;
 587
 588                                 requeue_io_u(td, &io_u);
 589                         } else {
 590 sync_done:
 591                                 if (should_check_rate(td))
 592                                         fio_gettime(&comp_time, NULL);
 593
 594                                 bytes_done = io_u_sync_complete(td, io_u);
 595                                 if (bytes_done < 0)
 596                                         ret = bytes_done;
 597                         }
 598                         break;
 599                 case FIO_Q_QUEUED:
 600                         /*
 601                          * if the engine doesn't have a commit hook,
 602                          * the io_u is really queued. if it does have such
 603                          * a hook, it has to call io_u_queued() itself.
 604                          */
 605                         if (td->io_ops->commit == NULL)
 606                                 io_u_queued(td, io_u);
 607                         break;
 608                 case FIO_Q_BUSY:
 609                         requeue_io_u(td, &io_u);
 610                         ret2 = td_io_commit(td);
 611                         if (ret2 < 0)
 612                                 ret = ret2;
 613                         break;
 614                 default:
 615                         assert(ret < 0);
 616                         put_io_u(td, io_u);
 617                         break;
 618                 }
 619
 620                 if (ret < 0 || td->error)
 621                         break;
 622
 623                 /*
 624                  * See if we need to complete some commands
 625                  */
 626                 full = queue_full(td) || ret == FIO_Q_BUSY;
 627                 if (full || !td->o.iodepth_batch_complete) {
 628                         min_evts = td->o.iodepth_batch_complete;
 629                         if (full && !min_evts)
 630                                 min_evts = 1;
 631
 632                         if (should_check_rate(td))
 633                                 fio_gettime(&comp_time, NULL);
 634
 635                         do {
 636                                 ret = io_u_queued_complete(td, min_evts);
 637                                 if (ret <= 0)
 638                                         break;
 639
 640                                 bytes_done += ret;
 641                         } while (full && (td->cur_depth > td->o.iodepth_low));
 642                 }
 643
 644                 if (ret < 0)
 645                         break;
 646                 if (!bytes_done)
 647                         continue;
 648
 649                 /*
 650                  * the rate is batched for now, it should work for batches
 651                  * of completions except the very first one which may look
 652                  * a little bursty
 653                  */
 654                 if (!in_ramp_time(td) && should_check_rate(td)) {
 655                         usec = utime_since(&td->tv_cache, &comp_time);
 656
 657                         rate_throttle(td, usec, bytes_done);
 658
 659                         if (check_min_rate(td, &comp_time)) {
 660                                 if (exitall_on_terminate)
 661                                         terminate_threads(td->groupid);
 662                                 td_verror(td, EIO, "check_min_rate");
 663                                 break;
 664                         }
 665                 }
 666
 667                 if (td->o.thinktime) {
 668                         unsigned long long b;
 669
 670                         b = td->io_blocks[0] + td->io_blocks[1];
 671                         if (!(b % td->o.thinktime_blocks)) {
 672                                 int left;
 673
 674                                 if (td->o.thinktime_spin)
 675                                         __usec_sleep(td->o.thinktime_spin);
 676
 677                                 left = td->o.thinktime - td->o.thinktime_spin;
 678                                 if (left)
 679                                         usec_sleep(td, left);
 680                         }
 681                 }
 682         }
 683
 684         if (td->o.fill_device && td->error == ENOSPC) {
 685                 td->error = 0;
 686                 td->terminate = 1;
 687         }
 688         if (!td->error) {
 689                 struct fio_file *f;
 690
 691                 i = td->cur_depth;
 692                 if (i)
 693                         ret = io_u_queued_complete(td, i);
 694
 695                 if (should_fsync(td) && td->o.end_fsync) {
 696                         td_set_runstate(td, TD_FSYNCING);
 697
 698                         for_each_file(td, f, i) {
 699                                 if (!(f->flags & FIO_FILE_OPEN))
 700                                         continue;
 701                                 fio_io_sync(td, f);
 702                         }
 703                 }
 704         } else
 705                 cleanup_pending_aio(td);
 706
 707         /*
 708          * stop job if we failed doing any IO
 709          */
 710         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 711                 td->done = 1;
 712 }
 713
 714 static void cleanup_io_u(struct thread_data *td)
 715 {
 716         struct flist_head *entry, *n;
 717         struct io_u *io_u;
 718
 719         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 720                 io_u = flist_entry(entry, struct io_u, list);
 721
 722                 flist_del(&io_u->list);
 723                 free(io_u);
 724         }
 725
 726         free_io_mem(td);
 727 }
 728
 729 static int init_io_u(struct thread_data *td)
 730 {
 731         struct io_u *io_u;
 732         unsigned int max_bs;
 733         int i, max_units;
 734         char *p;
 735
 736         max_units = td->o.iodepth;
 737         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 738         td->orig_buffer_size = (unsigned long long) max_bs
 739                                         * (unsigned long long) max_units;
 740
 741         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 742                 unsigned long bs;
 743
 744                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 745                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 746         }
 747
 748         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 749                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 750                 return 1;
 751         }
 752
 753         if (allocate_io_mem(td))
 754                 return 1;
 755
 756         if (td->o.odirect)
 757                 p = ALIGN(td->orig_buffer);
 758         else
 759                 p = td->orig_buffer;
 760
 761         for (i = 0; i < max_units; i++) {
 762                 if (td->terminate)
 763                         return 1;
 764                 io_u = malloc(sizeof(*io_u));
 765                 memset(io_u, 0, sizeof(*io_u));
 766                 INIT_FLIST_HEAD(&io_u->list);
 767
 768                 if (!(td->io_ops->flags & FIO_NOIO)) {
 769                         io_u->buf = p + max_bs * i;
 770
 771                         if (td_write(td) && !td->o.refill_buffers)
 772                                 io_u_fill_buffer(td, io_u, max_bs);
 773                 }
 774
 775                 io_u->index = i;
 776                 io_u->flags = IO_U_F_FREE;
 777                 flist_add(&io_u->list, &td->io_u_freelist);
 778         }
 779
 780         return 0;
 781 }
 782
 783 static int switch_ioscheduler(struct thread_data *td)
 784 {
 785         char tmp[256], tmp2[128];
 786         FILE *f;
 787         int ret;
 788
 789         if (td->io_ops->flags & FIO_DISKLESSIO)
 790                 return 0;
 791
 792         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 793
 794         f = fopen(tmp, "r+");
 795         if (!f) {
 796                 if (errno == ENOENT) {
 797                         log_err("fio: os or kernel doesn't support IO scheduler"
 798                                 " switching\n");
 799                         return 0;
 800                 }
 801                 td_verror(td, errno, "fopen iosched");
 802                 return 1;
 803         }
 804
 805         /*
 806          * Set io scheduler.
 807          */
 808         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 809         if (ferror(f) || ret != 1) {
 810                 td_verror(td, errno, "fwrite");
 811                 fclose(f);
 812                 return 1;
 813         }
 814
 815         rewind(f);
 816
 817         /*
 818          * Read back and check that the selected scheduler is now the default.
 819          */
 820         ret = fread(tmp, 1, sizeof(tmp), f);
 821         if (ferror(f) || ret < 0) {
 822                 td_verror(td, errno, "fread");
 823                 fclose(f);
 824                 return 1;
 825         }
 826
 827         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 828         if (!strstr(tmp, tmp2)) {
 829                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 830                 td_verror(td, EINVAL, "iosched_switch");
 831                 fclose(f);
 832                 return 1;
 833         }
 834
 835         fclose(f);
 836         return 0;
 837 }
 838
 839 static int keep_running(struct thread_data *td)
 840 {
 841         unsigned long long io_done;
 842
 843         if (td->done)
 844                 return 0;
 845         if (td->o.time_based)
 846                 return 1;
 847         if (td->o.loops) {
 848                 td->o.loops--;
 849                 return 1;
 850         }
 851
 852         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 853                         + td->io_skip_bytes;
 854         if (io_done < td->o.size)
 855                 return 1;
 856
 857         return 0;
 858 }
 859
 860 static void reset_io_counters(struct thread_data *td)
 861 {
 862         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 863         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 864         td->zone_bytes = 0;
 865         td->rate_bytes = 0;
 866         td->rate_blocks = 0;
 867         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 868
 869         td->last_was_sync = 0;
 870
 871         /*
 872          * reset file done count if we are to start over
 873          */
 874         if (td->o.time_based || td->o.loops)
 875                 td->nr_done_files = 0;
 876 }
 877
 878 void reset_all_stats(struct thread_data *td)
 879 {
 880         struct timeval tv;
 881         int i;
 882
 883         reset_io_counters(td);
 884
 885         for (i = 0; i < 2; i++) {
 886                 td->io_bytes[i] = 0;
 887                 td->io_blocks[i] = 0;
 888                 td->io_issues[i] = 0;
 889                 td->ts.total_io_u[i] = 0;
 890         }
 891
 892         fio_gettime(&tv, NULL);
 893         memcpy(&td->epoch, &tv, sizeof(tv));
 894         memcpy(&td->start, &tv, sizeof(tv));
 895 }
 896
 897 static int clear_io_state(struct thread_data *td)
 898 {
 899         struct fio_file *f;
 900         unsigned int i;
 901         int ret;
 902
 903         reset_io_counters(td);
 904
 905         close_files(td);
 906
 907         ret = 0;
 908         for_each_file(td, f, i) {
 909                 f->flags &= ~FIO_FILE_DONE;
 910                 ret = td_io_open_file(td, f);
 911                 if (ret)
 912                         break;
 913         }
 914
 915         return ret;
 916 }
 917
 918 /*
 919  * Entry point for the thread based jobs. The process based jobs end up
 920  * here as well, after a little setup.
 921  */
 922 static void *thread_main(void *data)
 923 {
 924         unsigned long long runtime[2], elapsed;
 925         struct thread_data *td = data;
 926         int clear_state;
 927
 928         if (!td->o.use_thread)
 929                 setsid();
 930
 931         td->pid = getpid();
 932
 933         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 934
 935         INIT_FLIST_HEAD(&td->io_u_freelist);
 936         INIT_FLIST_HEAD(&td->io_u_busylist);
 937         INIT_FLIST_HEAD(&td->io_u_requeues);
 938         INIT_FLIST_HEAD(&td->io_log_list);
 939         INIT_FLIST_HEAD(&td->io_hist_list);
 940         td->io_hist_tree = RB_ROOT;
 941
 942         td_set_runstate(td, TD_INITIALIZED);
 943         fio_mutex_up(startup_mutex);
 944         fio_mutex_down(td->mutex);
 945
 946         /*
 947          * the ->mutex mutex is now no longer used, close it to avoid
 948          * eating a file descriptor
 949          */
 950         fio_mutex_remove(td->mutex);
 951
 952         /*
 953          * May alter parameters that init_io_u() will use, so we need to
 954          * do this first.
 955          */
 956         if (init_iolog(td))
 957                 goto err;
 958
 959         if (init_io_u(td))
 960                 goto err;
 961
 962         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 963                 td_verror(td, errno, "cpu_set_affinity");
 964                 goto err;
 965         }
 966
 967         if (td->ioprio_set) {
 968                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 969                         td_verror(td, errno, "ioprio_set");
 970                         goto err;
 971                 }
 972         }
 973
 974         if (nice(td->o.nice) == -1) {
 975                 td_verror(td, errno, "nice");
 976                 goto err;
 977         }
 978
 979         if (td->o.ioscheduler && switch_ioscheduler(td))
 980                 goto err;
 981
 982         if (!td->o.create_serialize && setup_files(td))
 983                 goto err;
 984
 985         if (td_io_init(td))
 986                 goto err;
 987
 988         if (open_files(td))
 989                 goto err;
 990
 991         if (init_random_map(td))
 992                 goto err;
 993
 994         if (td->o.exec_prerun) {
 995                 if (system(td->o.exec_prerun) < 0)
 996                         goto err;
 997         }
 998
 999         fio_gettime(&td->epoch, NULL);
1000         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1001
1002         runtime[0] = runtime[1] = 0;
1003         clear_state = 0;
1004         while (keep_running(td)) {
1005                 fio_gettime(&td->start, NULL);
1006                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1007                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1008
1009                 if (td->o.ratemin)
1010                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1011                                                         sizeof(td->lastrate));
1012
1013                 if (clear_state && clear_io_state(td))
1014                         break;
1015
1016                 prune_io_piece_log(td);
1017
1018                 do_io(td);
1019
1020                 clear_state = 1;
1021
1022                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1023                         if (td->rw_end_set[DDIR_READ])
1024                                 elapsed = utime_since(&td->start,
1025                                                       &td->rw_end[DDIR_READ]);
1026                         else
1027                                 elapsed = utime_since_now(&td->start);
1028
1029                         runtime[DDIR_READ] += elapsed;
1030                 }
1031                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1032                         if (td->rw_end_set[DDIR_WRITE])
1033                                 elapsed = utime_since(&td->start,
1034                                                       &td->rw_end[DDIR_WRITE]);
1035                         else
1036                                 elapsed = utime_since_now(&td->start);
1037
1038                         runtime[DDIR_WRITE] += elapsed;
1039                 }
1040
1041                 if (td->error || td->terminate)
1042                         break;
1043
1044                 if (!td->o.do_verify ||
1045                     td->o.verify == VERIFY_NONE ||
1046                     (td->io_ops->flags & FIO_UNIDIR))
1047                         continue;
1048
1049                 if (clear_io_state(td))
1050                         break;
1051
1052                 fio_gettime(&td->start, NULL);
1053
1054                 do_verify(td);
1055
1056                 runtime[DDIR_READ] += utime_since_now(&td->start);
1057
1058                 if (td->error || td->terminate)
1059                         break;
1060         }
1061
1062         update_rusage_stat(td);
1063         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1064         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1065         td->ts.total_run_time = mtime_since_now(&td->epoch);
1066         td->ts.io_bytes[0] = td->io_bytes[0];
1067         td->ts.io_bytes[1] = td->io_bytes[1];
1068
1069         if (td->ts.bw_log)
1070                 finish_log(td, td->ts.bw_log, "bw");
1071         if (td->ts.slat_log)
1072                 finish_log(td, td->ts.slat_log, "slat");
1073         if (td->ts.clat_log)
1074                 finish_log(td, td->ts.clat_log, "clat");
1075         if (td->o.exec_postrun) {
1076                 if (system(td->o.exec_postrun) < 0)
1077                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
1078         }
1079
1080         if (exitall_on_terminate)
1081                 terminate_threads(td->groupid);
1082
1083 err:
1084         if (td->error)
1085                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1086                                                         td->verror);
1087         close_and_free_files(td);
1088         close_ioengine(td);
1089         cleanup_io_u(td);
1090
1091         /*
1092          * do this very late, it will log file closing as well
1093          */
1094         if (td->o.write_iolog_file)
1095                 write_iolog_close(td);
1096
1097         options_mem_free(td);
1098         td_set_runstate(td, TD_EXITED);
1099         return (void *) (unsigned long) td->error;
1100 }
1101
1102 /*
1103  * We cannot pass the td data into a forked process, so attach the td and
1104  * pass it to the thread worker.
1105  */
1106 static int fork_main(int shmid, int offset)
1107 {
1108         struct thread_data *td;
1109         void *data, *ret;
1110
1111         data = shmat(shmid, NULL, 0);
1112         if (data == (void *) -1) {
1113                 int __err = errno;
1114
1115                 perror("shmat");
1116                 return __err;
1117         }
1118
1119         td = data + offset * sizeof(struct thread_data);
1120         ret = thread_main(td);
1121         shmdt(data);
1122         return (int) (unsigned long) ret;
1123 }
1124
1125 /*
1126  * Run over the job map and reap the threads that have exited, if any.
1127  */
1128 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1129 {
1130         struct thread_data *td;
1131         int i, cputhreads, realthreads, pending, status, ret;
1132
1133         /*
1134          * reap exited threads (TD_EXITED -> TD_REAPED)
1135          */
1136         realthreads = pending = cputhreads = 0;
1137         for_each_td(td, i) {
1138                 int flags = 0;
1139
1140                 /*
1141                  * ->io_ops is NULL for a thread that has closed its
1142                  * io engine
1143                  */
1144                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1145                         cputhreads++;
1146                 else
1147                         realthreads++;
1148
1149                 if (!td->pid) {
1150                         pending++;
1151                         continue;
1152                 }
1153                 if (td->runstate == TD_REAPED)
1154                         continue;
1155                 if (td->o.use_thread) {
1156                         if (td->runstate == TD_EXITED) {
1157                                 td_set_runstate(td, TD_REAPED);
1158                                 goto reaped;
1159                         }
1160                         continue;
1161                 }
1162
1163                 flags = WNOHANG;
1164                 if (td->runstate == TD_EXITED)
1165                         flags = 0;
1166
1167                 /*
1168                  * check if someone quit or got killed in an unusual way
1169                  */
1170                 ret = waitpid(td->pid, &status, flags);
1171                 if (ret < 0) {
1172                         if (errno == ECHILD) {
1173                                 log_err("fio: pid=%d disappeared %d\n",
1174                                                 (int) td->pid, td->runstate);
1175                                 td_set_runstate(td, TD_REAPED);
1176                                 goto reaped;
1177                         }
1178                         perror("waitpid");
1179                 } else if (ret == td->pid) {
1180                         if (WIFSIGNALED(status)) {
1181                                 int sig = WTERMSIG(status);
1182
1183                                 if (sig != SIGQUIT)
1184                                         log_err("fio: pid=%d, got signal=%d\n",
1185                                                         (int) td->pid, sig);
1186                                 td_set_runstate(td, TD_REAPED);
1187                                 goto reaped;
1188                         }
1189                         if (WIFEXITED(status)) {
1190                                 if (WEXITSTATUS(status) && !td->error)
1191                                         td->error = WEXITSTATUS(status);
1192
1193                                 td_set_runstate(td, TD_REAPED);
1194                                 goto reaped;
1195                         }
1196                 }
1197
1198                 /*
1199                  * thread is not dead, continue
1200                  */
1201                 pending++;
1202                 continue;
1203 reaped:
1204                 (*nr_running)--;
1205                 (*m_rate) -= td->o.ratemin;
1206                 (*t_rate) -= td->o.rate;
1207                 if (!td->pid)
1208                         pending--;
1209
1210                 if (td->error)
1211                         exit_value++;
1212
1213                 done_secs += mtime_since_now(&td->epoch) / 1000;
1214         }
1215
1216         if (*nr_running == cputhreads && !pending && realthreads)
1217                 terminate_threads(TERMINATE_ALL);
1218 }
1219
1220 /*
1221  * Main function for kicking off and reaping jobs, as needed.
1222  */
1223 static void run_threads(void)
1224 {
1225         struct thread_data *td;
1226         unsigned long spent;
1227         int i, todo, nr_running, m_rate, t_rate, nr_started;
1228
1229         if (fio_pin_memory())
1230                 return;
1231
1232         if (!terse_output) {
1233                 printf("Starting ");
1234                 if (nr_thread)
1235                         printf("%d thread%s", nr_thread,
1236                                                 nr_thread > 1 ? "s" : "");
1237                 if (nr_process) {
1238                         if (nr_thread)
1239                                 printf(" and ");
1240                         printf("%d process%s", nr_process,
1241                                                 nr_process > 1 ? "es" : "");
1242                 }
1243                 printf("\n");
1244                 fflush(stdout);
1245         }
1246
1247         set_sig_handlers();
1248
1249         todo = thread_number;
1250         nr_running = 0;
1251         nr_started = 0;
1252         m_rate = t_rate = 0;
1253
1254         for_each_td(td, i) {
1255                 print_status_init(td->thread_number - 1);
1256
1257                 if (!td->o.create_serialize) {
1258                         init_disk_util(td);
1259                         continue;
1260                 }
1261
1262                 /*
1263                  * do file setup here so it happens sequentially,
1264                  * we don't want X number of threads getting their
1265                  * client data interspersed on disk
1266                  */
1267                 if (setup_files(td)) {
1268                         exit_value++;
1269                         if (td->error)
1270                                 log_err("fio: pid=%d, err=%d/%s\n",
1271                                         (int) td->pid, td->error, td->verror);
1272                         td_set_runstate(td, TD_REAPED);
1273                         todo--;
1274                 } else {
1275                         struct fio_file *f;
1276                         unsigned int i;
1277
1278                         /*
1279                          * for sharing to work, each job must always open
1280                          * its own files. so close them, if we opened them
1281                          * for creation
1282                          */
1283                         for_each_file(td, f, i)
1284                                 td_io_close_file(td, f);
1285                 }
1286
1287                 init_disk_util(td);
1288         }
1289
1290         set_genesis_time();
1291
1292         while (todo) {
1293                 struct thread_data *map[MAX_JOBS];
1294                 struct timeval this_start;
1295                 int this_jobs = 0, left;
1296
1297                 /*
1298                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1299                  */
1300                 for_each_td(td, i) {
1301                         if (td->runstate != TD_NOT_CREATED)
1302                                 continue;
1303
1304                         /*
1305                          * never got a chance to start, killed by other
1306                          * thread for some reason
1307                          */
1308                         if (td->terminate) {
1309                                 todo--;
1310                                 continue;
1311                         }
1312
1313                         if (td->o.start_delay) {
1314                                 spent = mtime_since_genesis();
1315
1316                                 if (td->o.start_delay * 1000 > spent)
1317                                         continue;
1318                         }
1319
1320                         if (td->o.stonewall && (nr_started || nr_running)) {
1321                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1322                                                         td->o.name);
1323                                 break;
1324                         }
1325
1326                         /*
1327                          * Set state to created. Thread will transition
1328                          * to TD_INITIALIZED when it's done setting up.
1329                          */
1330                         td_set_runstate(td, TD_CREATED);
1331                         map[this_jobs++] = td;
1332                         nr_started++;
1333
1334                         if (td->o.use_thread) {
1335                                 dprint(FD_PROCESS, "will pthread_create\n");
1336                                 if (pthread_create(&td->thread, NULL,
1337                                                    thread_main, td)) {
1338                                         perror("pthread_create");
1339                                         nr_started--;
1340                                         break;
1341                                 }
1342                                 if (pthread_detach(td->thread) < 0)
1343                                         perror("pthread_detach");
1344                         } else {
1345                                 pid_t pid;
1346                                 dprint(FD_PROCESS, "will fork\n");
1347                                 pid = fork();
1348                                 if (!pid) {
1349                                         int ret = fork_main(shm_id, i);
1350
1351                                         _exit(ret);
1352                                 } else if (i == fio_debug_jobno)
1353                                         *fio_debug_jobp = pid;
1354                         }
1355                         fio_mutex_down(startup_mutex);
1356                 }
1357
1358                 /*
1359                  * Wait for the started threads to transition to
1360                  * TD_INITIALIZED.
1361                  */
1362                 fio_gettime(&this_start, NULL);
1363                 left = this_jobs;
1364                 while (left && !fio_abort) {
1365                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1366                                 break;
1367
1368                         usleep(100000);
1369
1370                         for (i = 0; i < this_jobs; i++) {
1371                                 td = map[i];
1372                                 if (!td)
1373                                         continue;
1374                                 if (td->runstate == TD_INITIALIZED) {
1375                                         map[i] = NULL;
1376                                         left--;
1377                                 } else if (td->runstate >= TD_EXITED) {
1378                                         map[i] = NULL;
1379                                         left--;
1380                                         todo--;
1381                                         nr_running++; /* work-around... */
1382                                 }
1383                         }
1384                 }
1385
1386                 if (left) {
1387                         log_err("fio: %d jobs failed to start\n", left);
1388                         for (i = 0; i < this_jobs; i++) {
1389                                 td = map[i];
1390                                 if (!td)
1391                                         continue;
1392                                 kill(td->pid, SIGTERM);
1393                         }
1394                         break;
1395                 }
1396
1397                 /*
1398                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1399                  */
1400                 for_each_td(td, i) {
1401                         if (td->runstate != TD_INITIALIZED)
1402                                 continue;
1403
1404                         if (in_ramp_time(td))
1405                                 td_set_runstate(td, TD_RAMP);
1406                         else
1407                                 td_set_runstate(td, TD_RUNNING);
1408                         nr_running++;
1409                         nr_started--;
1410                         m_rate += td->o.ratemin;
1411                         t_rate += td->o.rate;
1412                         todo--;
1413                         fio_mutex_up(td->mutex);
1414                 }
1415
1416                 reap_threads(&nr_running, &t_rate, &m_rate);
1417
1418                 if (todo)
1419                         usleep(100000);
1420         }
1421
1422         while (nr_running) {
1423                 reap_threads(&nr_running, &t_rate, &m_rate);
1424                 usleep(10000);
1425         }
1426
1427         update_io_ticks();
1428         fio_unpin_memory();
1429 }
1430
1431 int main(int argc, char *argv[])
1432 {
1433         long ps;
1434
1435         sinit();
1436
1437         /*
1438          * We need locale for number printing, if it isn't set then just
1439          * go with the US format.
1440          */
1441         if (!getenv("LC_NUMERIC"))
1442                 setlocale(LC_NUMERIC, "en_US");
1443
1444         if (parse_options(argc, argv))
1445                 return 1;
1446
1447         if (!thread_number)
1448                 return 0;
1449
1450         ps = sysconf(_SC_PAGESIZE);
1451         if (ps < 0) {
1452                 log_err("Failed to get page size\n");
1453                 return 1;
1454         }
1455
1456         page_size = ps;
1457         page_mask = ps - 1;
1458
1459         if (write_bw_log) {
1460                 setup_log(&agg_io_log[DDIR_READ]);
1461                 setup_log(&agg_io_log[DDIR_WRITE]);
1462         }
1463
1464         startup_mutex = fio_mutex_init(0);
1465
1466         set_genesis_time();
1467
1468         status_timer_arm();
1469
1470         run_threads();
1471
1472         if (!fio_abort) {
1473                 show_run_stats();
1474                 if (write_bw_log) {
1475                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1476                         __finish_log(agg_io_log[DDIR_WRITE],
1477                                         "agg-write_bw.log");
1478                 }
1479         }
1480
1481         fio_mutex_remove(startup_mutex);
1482         return exit_value;
1483 }