fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "hash.h"
  39 #include "smalloc.h"
  40
  41 unsigned long page_mask;
  42 unsigned long page_size;
  43 #define ALIGN(buf)      \
  44         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  45
  46 int groupid = 0;
  47 int thread_number = 0;
  48 int nr_process = 0;
  49 int nr_thread = 0;
  50 int shm_id = 0;
  51 int temp_stall_ts;
  52 unsigned long done_secs = 0;
  53
  54 static struct fio_mutex *startup_mutex;
  55 static volatile int fio_abort;
  56 static int exit_value;
  57 static struct itimerval itimer;
  58
  59 struct io_log *agg_io_log[2];
  60
  61 #define TERMINATE_ALL           (-1)
  62 #define JOB_START_TIMEOUT       (5 * 1000)
  63
  64 static inline void td_set_runstate(struct thread_data *td, int runstate)
  65 {
  66         if (td->runstate == runstate)
  67                 return;
  68
  69         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  70                                                 td->runstate, runstate);
  71         td->runstate = runstate;
  72 }
  73
  74 static void terminate_threads(int group_id)
  75 {
  76         struct thread_data *td;
  77         int i;
  78
  79         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  80
  81         for_each_td(td, i) {
  82                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  83                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  84                                                 td->o.name, (int) td->pid);
  85                         td->terminate = 1;
  86                         td->o.start_delay = 0;
  87
  88                         /*
  89                          * if the thread is running, just let it exit
  90                          */
  91                         if (td->runstate < TD_RUNNING)
  92                                 kill(td->pid, SIGQUIT);
  93                         else {
  94                                 struct ioengine_ops *ops = td->io_ops;
  95
  96                                 if (ops && (ops->flags & FIO_SIGQUIT))
  97                                         kill(td->pid, SIGQUIT);
  98                         }
  99                 }
 100         }
 101 }
 102
 103 static void status_timer_arm(void)
 104 {
 105         itimer.it_value.tv_sec = 0;
 106         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 107         setitimer(ITIMER_REAL, &itimer, NULL);
 108 }
 109
 110 static void sig_alrm(int sig)
 111 {
 112         if (threads) {
 113                 update_io_ticks();
 114                 print_thread_status();
 115                 status_timer_arm();
 116         }
 117 }
 118
 119 static void sig_int(int sig)
 120 {
 121         if (threads) {
 122                 printf("\nfio: terminating on signal %d\n", sig);
 123                 fflush(stdout);
 124                 terminate_threads(TERMINATE_ALL);
 125         }
 126 }
 127
 128 static void set_sig_handlers(void)
 129 {
 130         struct sigaction act;
 131
 132         memset(&act, 0, sizeof(act));
 133         act.sa_handler = sig_alrm;
 134         act.sa_flags = SA_RESTART;
 135         sigaction(SIGALRM, &act, NULL);
 136
 137         memset(&act, 0, sizeof(act));
 138         act.sa_handler = sig_int;
 139         act.sa_flags = SA_RESTART;
 140         sigaction(SIGINT, &act, NULL);
 141 }
 142
 143 /*
 144  * Check if we are above the minimum rate given.
 145  */
 146 static int check_min_rate(struct thread_data *td, struct timeval *now)
 147 {
 148         unsigned long long bytes = 0;
 149         unsigned long iops = 0;
 150         unsigned long spent;
 151         unsigned long rate;
 152
 153         /*
 154          * No minimum rate set, always ok
 155          */
 156         if (!td->o.ratemin && !td->o.rate_iops_min)
 157                 return 0;
 158
 159         /*
 160          * allow a 2 second settle period in the beginning
 161          */
 162         if (mtime_since(&td->start, now) < 2000)
 163                 return 0;
 164
 165         if (td_read(td)) {
 166                 iops += td->io_blocks[DDIR_READ];
 167                 bytes += td->this_io_bytes[DDIR_READ];
 168         }
 169         if (td_write(td)) {
 170                 iops += td->io_blocks[DDIR_WRITE];
 171                 bytes += td->this_io_bytes[DDIR_WRITE];
 172         }
 173
 174         /*
 175          * if rate blocks is set, sample is running
 176          */
 177         if (td->rate_bytes || td->rate_blocks) {
 178                 spent = mtime_since(&td->lastrate, now);
 179                 if (spent < td->o.ratecycle)
 180                         return 0;
 181
 182                 if (td->o.rate) {
 183                         /*
 184                          * check bandwidth specified rate
 185                          */
 186                         if (bytes < td->rate_bytes) {
 187                                 log_err("%s: min rate %u not met\n", td->o.name,
 188                                                                 td->o.ratemin);
 189                                 return 1;
 190                         } else {
 191                                 rate = (bytes - td->rate_bytes) / spent;
 192                                 if (rate < td->o.ratemin ||
 193                                     bytes < td->rate_bytes) {
 194                                         log_err("%s: min rate %u not met, got"
 195                                                 " %luKiB/sec\n", td->o.name,
 196                                                         td->o.ratemin, rate);
 197                                         return 1;
 198                                 }
 199                         }
 200                 } else {
 201                         /*
 202                          * checks iops specified rate
 203                          */
 204                         if (iops < td->o.rate_iops) {
 205                                 log_err("%s: min iops rate %u not met\n",
 206                                                 td->o.name, td->o.rate_iops);
 207                                 return 1;
 208                         } else {
 209                                 rate = (iops - td->rate_blocks) / spent;
 210                                 if (rate < td->o.rate_iops_min ||
 211                                     iops < td->rate_blocks) {
 212                                         log_err("%s: min iops rate %u not met,"
 213                                                 " got %lu\n", td->o.name,
 214                                                         td->o.rate_iops_min,
 215                                                         rate);
 216                                 }
 217                         }
 218                 }
 219         }
 220
 221         td->rate_bytes = bytes;
 222         td->rate_blocks = iops;
 223         memcpy(&td->lastrate, now, sizeof(*now));
 224         return 0;
 225 }
 226
 227 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 228 {
 229         if (!td->o.timeout)
 230                 return 0;
 231         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 232                 return 1;
 233
 234         return 0;
 235 }
 236
 237 /*
 238  * When job exits, we can cancel the in-flight IO if we are using async
 239  * io. Attempt to do so.
 240  */
 241 static void cleanup_pending_aio(struct thread_data *td)
 242 {
 243         struct flist_head *entry, *n;
 244         struct io_u *io_u;
 245         int r;
 246
 247         /*
 248          * get immediately available events, if any
 249          */
 250         r = io_u_queued_complete(td, 0);
 251         if (r < 0)
 252                 return;
 253
 254         /*
 255          * now cancel remaining active events
 256          */
 257         if (td->io_ops->cancel) {
 258                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 259                         io_u = flist_entry(entry, struct io_u, list);
 260
 261                         /*
 262                          * if the io_u isn't in flight, then that generally
 263                          * means someone leaked an io_u. complain but fix
 264                          * it up, so we don't stall here.
 265                          */
 266                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 267                                 log_err("fio: non-busy IO on busy list\n");
 268                                 put_io_u(td, io_u);
 269                         } else {
 270                                 r = td->io_ops->cancel(td, io_u);
 271                                 if (!r)
 272                                         put_io_u(td, io_u);
 273                         }
 274                 }
 275         }
 276
 277         if (td->cur_depth)
 278                 r = io_u_queued_complete(td, td->cur_depth);
 279 }
 280
 281 /*
 282  * Helper to handle the final sync of a file. Works just like the normal
 283  * io path, just does everything sync.
 284  */
 285 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 286 {
 287         struct io_u *io_u = __get_io_u(td);
 288         int ret;
 289
 290         if (!io_u)
 291                 return 1;
 292
 293         io_u->ddir = DDIR_SYNC;
 294         io_u->file = f;
 295
 296         if (td_io_prep(td, io_u)) {
 297                 put_io_u(td, io_u);
 298                 return 1;
 299         }
 300
 301 requeue:
 302         ret = td_io_queue(td, io_u);
 303         if (ret < 0) {
 304                 td_verror(td, io_u->error, "td_io_queue");
 305                 put_io_u(td, io_u);
 306                 return 1;
 307         } else if (ret == FIO_Q_QUEUED) {
 308                 if (io_u_queued_complete(td, 1) < 0)
 309                         return 1;
 310         } else if (ret == FIO_Q_COMPLETED) {
 311                 if (io_u->error) {
 312                         td_verror(td, io_u->error, "td_io_queue");
 313                         return 1;
 314                 }
 315
 316                 if (io_u_sync_complete(td, io_u) < 0)
 317                         return 1;
 318         } else if (ret == FIO_Q_BUSY) {
 319                 if (td_io_commit(td))
 320                         return 1;
 321                 goto requeue;
 322         }
 323
 324         return 0;
 325 }
 326
 327 /*
 328  * The main verify engine. Runs over the writes we previously submitted,
 329  * reads the blocks back in, and checks the crc/md5 of the data.
 330  */
 331 static void do_verify(struct thread_data *td)
 332 {
 333         struct fio_file *f;
 334         struct io_u *io_u;
 335         int ret, min_events;
 336         unsigned int i;
 337
 338         /*
 339          * sync io first and invalidate cache, to make sure we really
 340          * read from disk.
 341          */
 342         for_each_file(td, f, i) {
 343                 if (!(f->flags & FIO_FILE_OPEN))
 344                         continue;
 345                 if (fio_io_sync(td, f))
 346                         break;
 347                 if (file_invalidate_cache(td, f))
 348                         break;
 349         }
 350
 351         if (td->error)
 352                 return;
 353
 354         td_set_runstate(td, TD_VERIFYING);
 355
 356         io_u = NULL;
 357         while (!td->terminate) {
 358                 int ret2;
 359
 360                 io_u = __get_io_u(td);
 361                 if (!io_u)
 362                         break;
 363
 364                 if (runtime_exceeded(td, &io_u->start_time)) {
 365                         put_io_u(td, io_u);
 366                         td->terminate = 1;
 367                         break;
 368                 }
 369
 370                 if (get_next_verify(td, io_u)) {
 371                         put_io_u(td, io_u);
 372                         break;
 373                 }
 374
 375                 if (td_io_prep(td, io_u)) {
 376                         put_io_u(td, io_u);
 377                         break;
 378                 }
 379
 380                 io_u->end_io = verify_io_u;
 381
 382                 ret = td_io_queue(td, io_u);
 383                 switch (ret) {
 384                 case FIO_Q_COMPLETED:
 385                         if (io_u->error)
 386                                 ret = -io_u->error;
 387                         else if (io_u->resid) {
 388                                 int bytes = io_u->xfer_buflen - io_u->resid;
 389                                 struct fio_file *f = io_u->file;
 390
 391                                 /*
 392                                  * zero read, fail
 393                                  */
 394                                 if (!bytes) {
 395                                         td_verror(td, EIO, "full resid");
 396                                         put_io_u(td, io_u);
 397                                         break;
 398                                 }
 399
 400                                 io_u->xfer_buflen = io_u->resid;
 401                                 io_u->xfer_buf += bytes;
 402                                 io_u->offset += bytes;
 403
 404                                 td->ts.short_io_u[io_u->ddir]++;
 405
 406                                 if (io_u->offset == f->real_file_size)
 407                                         goto sync_done;
 408
 409                                 requeue_io_u(td, &io_u);
 410                         } else {
 411 sync_done:
 412                                 ret = io_u_sync_complete(td, io_u);
 413                                 if (ret < 0)
 414                                         break;
 415                         }
 416                         continue;
 417                 case FIO_Q_QUEUED:
 418                         break;
 419                 case FIO_Q_BUSY:
 420                         requeue_io_u(td, &io_u);
 421                         ret2 = td_io_commit(td);
 422                         if (ret2 < 0)
 423                                 ret = ret2;
 424                         break;
 425                 default:
 426                         assert(ret < 0);
 427                         td_verror(td, -ret, "td_io_queue");
 428                         break;
 429                 }
 430
 431                 if (ret < 0 || td->error)
 432                         break;
 433
 434                 /*
 435                  * if we can queue more, do so. but check if there are
 436                  * completed io_u's first.
 437                  */
 438                 min_events = 0;
 439                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 440                         if (td->cur_depth >= td->o.iodepth_low)
 441                                 min_events = td->cur_depth - td->o.iodepth_low;
 442                         if (!min_events)
 443                                 min_events = 1;
 444                 }
 445
 446                 /*
 447                  * Reap required number of io units, if any, and do the
 448                  * verification on them through the callback handler
 449                  */
 450                 if (io_u_queued_complete(td, min_events) < 0)
 451                         break;
 452         }
 453
 454         if (!td->error) {
 455                 min_events = td->cur_depth;
 456
 457                 if (min_events)
 458                         ret = io_u_queued_complete(td, min_events);
 459         } else
 460                 cleanup_pending_aio(td);
 461
 462         td_set_runstate(td, TD_RUNNING);
 463 }
 464
 465 /*
 466  * Main IO worker function. It retrieves io_u's to process and queues
 467  * and reaps them, checking for rate and errors along the way.
 468  */
 469 static void do_io(struct thread_data *td)
 470 {
 471         struct timeval s;
 472         unsigned long usec;
 473         unsigned int i;
 474         int ret = 0;
 475
 476         td_set_runstate(td, TD_RUNNING);
 477
 478         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 479                 struct timeval comp_time;
 480                 long bytes_done = 0;
 481                 int min_evts = 0;
 482                 struct io_u *io_u;
 483                 int ret2;
 484
 485                 if (td->terminate)
 486                         break;
 487
 488                 io_u = get_io_u(td);
 489                 if (!io_u)
 490                         break;
 491
 492                 memcpy(&s, &io_u->start_time, sizeof(s));
 493
 494                 if (runtime_exceeded(td, &s)) {
 495                         put_io_u(td, io_u);
 496                         td->terminate = 1;
 497                         break;
 498                 }
 499
 500                 /*
 501                  * Add verification end_io handler, if asked to verify
 502                  * a previously written file.
 503                  */
 504                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 505                         io_u->end_io = verify_io_u;
 506                         td_set_runstate(td, TD_VERIFYING);
 507                 } else
 508                         td_set_runstate(td, TD_RUNNING);
 509
 510                 ret = td_io_queue(td, io_u);
 511                 switch (ret) {
 512                 case FIO_Q_COMPLETED:
 513                         if (io_u->error)
 514                                 ret = -io_u->error;
 515                         else if (io_u->resid) {
 516                                 int bytes = io_u->xfer_buflen - io_u->resid;
 517                                 struct fio_file *f = io_u->file;
 518
 519                                 /*
 520                                  * zero read, fail
 521                                  */
 522                                 if (!bytes) {
 523                                         td_verror(td, EIO, "full resid");
 524                                         put_io_u(td, io_u);
 525                                         break;
 526                                 }
 527
 528                                 io_u->xfer_buflen = io_u->resid;
 529                                 io_u->xfer_buf += bytes;
 530                                 io_u->offset += bytes;
 531
 532                                 td->ts.short_io_u[io_u->ddir]++;
 533
 534                                 if (io_u->offset == f->real_file_size)
 535                                         goto sync_done;
 536
 537                                 requeue_io_u(td, &io_u);
 538                         } else {
 539 sync_done:
 540                                 fio_gettime(&comp_time, NULL);
 541                                 bytes_done = io_u_sync_complete(td, io_u);
 542                                 if (bytes_done < 0)
 543                                         ret = bytes_done;
 544                         }
 545                         break;
 546                 case FIO_Q_QUEUED:
 547                         /*
 548                          * if the engine doesn't have a commit hook,
 549                          * the io_u is really queued. if it does have such
 550                          * a hook, it has to call io_u_queued() itself.
 551                          */
 552                         if (td->io_ops->commit == NULL)
 553                                 io_u_queued(td, io_u);
 554                         break;
 555                 case FIO_Q_BUSY:
 556                         requeue_io_u(td, &io_u);
 557                         ret2 = td_io_commit(td);
 558                         if (ret2 < 0)
 559                                 ret = ret2;
 560                         break;
 561                 default:
 562                         assert(ret < 0);
 563                         put_io_u(td, io_u);
 564                         break;
 565                 }
 566
 567                 if (ret < 0 || td->error)
 568                         break;
 569
 570                 /*
 571                  * See if we need to complete some commands
 572                  */
 573                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 574                         min_evts = 0;
 575                         if (td->cur_depth >= td->o.iodepth_low)
 576                                 min_evts = td->cur_depth - td->o.iodepth_low;
 577                         if (!min_evts)
 578                                 min_evts = 1;
 579                         fio_gettime(&comp_time, NULL);
 580                         bytes_done = io_u_queued_complete(td, min_evts);
 581                         if (bytes_done < 0)
 582                                 break;
 583                 }
 584
 585                 if (!bytes_done)
 586                         continue;
 587
 588                 /*
 589                  * the rate is batched for now, it should work for batches
 590                  * of completions except the very first one which may look
 591                  * a little bursty
 592                  */
 593                 usec = utime_since(&s, &comp_time);
 594
 595                 rate_throttle(td, usec, bytes_done);
 596
 597                 if (check_min_rate(td, &comp_time)) {
 598                         if (exitall_on_terminate)
 599                                 terminate_threads(td->groupid);
 600                         td_verror(td, EIO, "check_min_rate");
 601                         break;
 602                 }
 603
 604                 if (td->o.thinktime) {
 605                         unsigned long long b;
 606
 607                         b = td->io_blocks[0] + td->io_blocks[1];
 608                         if (!(b % td->o.thinktime_blocks)) {
 609                                 int left;
 610
 611                                 if (td->o.thinktime_spin)
 612                                         __usec_sleep(td->o.thinktime_spin);
 613
 614                                 left = td->o.thinktime - td->o.thinktime_spin;
 615                                 if (left)
 616                                         usec_sleep(td, left);
 617                         }
 618                 }
 619         }
 620
 621         if (td->o.fill_device && td->error == ENOSPC) {
 622                 td->error = 0;
 623                 td->terminate = 1;
 624         }
 625         if (!td->error) {
 626                 struct fio_file *f;
 627
 628                 i = td->cur_depth;
 629                 if (i)
 630                         ret = io_u_queued_complete(td, i);
 631
 632                 if (should_fsync(td) && td->o.end_fsync) {
 633                         td_set_runstate(td, TD_FSYNCING);
 634
 635                         for_each_file(td, f, i) {
 636                                 if (!(f->flags & FIO_FILE_OPEN))
 637                                         continue;
 638                                 fio_io_sync(td, f);
 639                         }
 640                 }
 641         } else
 642                 cleanup_pending_aio(td);
 643
 644         /*
 645          * stop job if we failed doing any IO
 646          */
 647         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 648                 td->done = 1;
 649 }
 650
 651 static void cleanup_io_u(struct thread_data *td)
 652 {
 653         struct flist_head *entry, *n;
 654         struct io_u *io_u;
 655
 656         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 657                 io_u = flist_entry(entry, struct io_u, list);
 658
 659                 flist_del(&io_u->list);
 660                 free(io_u);
 661         }
 662
 663         free_io_mem(td);
 664 }
 665
 666 static int init_io_u(struct thread_data *td)
 667 {
 668         struct io_u *io_u;
 669         unsigned int max_bs;
 670         int i, max_units;
 671         char *p;
 672
 673         max_units = td->o.iodepth;
 674         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 675         td->orig_buffer_size = (unsigned long long) max_bs
 676                                         * (unsigned long long) max_units;
 677
 678         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 679                 unsigned long bs;
 680
 681                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 682                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 683         }
 684
 685         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 686                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 687                 return 1;
 688         }
 689
 690         if (allocate_io_mem(td))
 691                 return 1;
 692
 693         if (td->o.odirect)
 694                 p = ALIGN(td->orig_buffer);
 695         else
 696                 p = td->orig_buffer;
 697
 698         for (i = 0; i < max_units; i++) {
 699                 if (td->terminate)
 700                         return 1;
 701                 io_u = malloc(sizeof(*io_u));
 702                 memset(io_u, 0, sizeof(*io_u));
 703                 INIT_FLIST_HEAD(&io_u->list);
 704
 705                 if (!(td->io_ops->flags & FIO_NOIO)) {
 706                         io_u->buf = p + max_bs * i;
 707
 708                         if (td_write(td) && !td->o.refill_buffers)
 709                                 io_u_fill_buffer(td, io_u, max_bs);
 710                 }
 711
 712                 io_u->index = i;
 713                 io_u->flags = IO_U_F_FREE;
 714                 flist_add(&io_u->list, &td->io_u_freelist);
 715         }
 716
 717         io_u_init_timeout();
 718
 719         return 0;
 720 }
 721
 722 static int switch_ioscheduler(struct thread_data *td)
 723 {
 724         char tmp[256], tmp2[128];
 725         FILE *f;
 726         int ret;
 727
 728         if (td->io_ops->flags & FIO_DISKLESSIO)
 729                 return 0;
 730
 731         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 732
 733         f = fopen(tmp, "r+");
 734         if (!f) {
 735                 if (errno == ENOENT) {
 736                         log_err("fio: os or kernel doesn't support IO scheduler"
 737                                 " switching\n");
 738                         return 0;
 739                 }
 740                 td_verror(td, errno, "fopen iosched");
 741                 return 1;
 742         }
 743
 744         /*
 745          * Set io scheduler.
 746          */
 747         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 748         if (ferror(f) || ret != 1) {
 749                 td_verror(td, errno, "fwrite");
 750                 fclose(f);
 751                 return 1;
 752         }
 753
 754         rewind(f);
 755
 756         /*
 757          * Read back and check that the selected scheduler is now the default.
 758          */
 759         ret = fread(tmp, 1, sizeof(tmp), f);
 760         if (ferror(f) || ret < 0) {
 761                 td_verror(td, errno, "fread");
 762                 fclose(f);
 763                 return 1;
 764         }
 765
 766         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 767         if (!strstr(tmp, tmp2)) {
 768                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 769                 td_verror(td, EINVAL, "iosched_switch");
 770                 fclose(f);
 771                 return 1;
 772         }
 773
 774         fclose(f);
 775         return 0;
 776 }
 777
 778 static int keep_running(struct thread_data *td)
 779 {
 780         unsigned long long io_done;
 781
 782         if (td->done)
 783                 return 0;
 784         if (td->o.time_based)
 785                 return 1;
 786         if (td->o.loops) {
 787                 td->o.loops--;
 788                 return 1;
 789         }
 790
 791         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 792                         + td->io_skip_bytes;
 793         if (io_done < td->o.size)
 794                 return 1;
 795
 796         return 0;
 797 }
 798
 799 static int clear_io_state(struct thread_data *td)
 800 {
 801         struct fio_file *f;
 802         unsigned int i;
 803         int ret;
 804
 805         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 806         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 807         td->zone_bytes = 0;
 808         td->rate_bytes = 0;
 809         td->rate_blocks = 0;
 810         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 811
 812         td->last_was_sync = 0;
 813
 814         /*
 815          * reset file done count if we are to start over
 816          */
 817         if (td->o.time_based || td->o.loops)
 818                 td->nr_done_files = 0;
 819
 820         close_files(td);
 821
 822         ret = 0;
 823         for_each_file(td, f, i) {
 824                 f->flags &= ~FIO_FILE_DONE;
 825                 ret = td_io_open_file(td, f);
 826                 if (ret)
 827                         break;
 828         }
 829
 830         return ret;
 831 }
 832
 833 /*
 834  * Entry point for the thread based jobs. The process based jobs end up
 835  * here as well, after a little setup.
 836  */
 837 static void *thread_main(void *data)
 838 {
 839         unsigned long long runtime[2], elapsed;
 840         struct thread_data *td = data;
 841         int clear_state;
 842
 843         if (!td->o.use_thread)
 844                 setsid();
 845
 846         td->pid = getpid();
 847
 848         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 849
 850         INIT_FLIST_HEAD(&td->io_u_freelist);
 851         INIT_FLIST_HEAD(&td->io_u_busylist);
 852         INIT_FLIST_HEAD(&td->io_u_requeues);
 853         INIT_FLIST_HEAD(&td->io_log_list);
 854         INIT_FLIST_HEAD(&td->io_hist_list);
 855         td->io_hist_tree = RB_ROOT;
 856
 857         td_set_runstate(td, TD_INITIALIZED);
 858         fio_mutex_up(startup_mutex);
 859         fio_mutex_down(td->mutex);
 860
 861         /*
 862          * the ->mutex mutex is now no longer used, close it to avoid
 863          * eating a file descriptor
 864          */
 865         fio_mutex_remove(td->mutex);
 866
 867         /*
 868          * May alter parameters that init_io_u() will use, so we need to
 869          * do this first.
 870          */
 871         if (init_iolog(td))
 872                 goto err;
 873
 874         if (init_io_u(td))
 875                 goto err;
 876
 877         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 878                 td_verror(td, errno, "cpu_set_affinity");
 879                 goto err;
 880         }
 881
 882         if (td->ioprio_set) {
 883                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 884                         td_verror(td, errno, "ioprio_set");
 885                         goto err;
 886                 }
 887         }
 888
 889         if (nice(td->o.nice) == -1) {
 890                 td_verror(td, errno, "nice");
 891                 goto err;
 892         }
 893
 894         if (td->o.ioscheduler && switch_ioscheduler(td))
 895                 goto err;
 896
 897         if (!td->o.create_serialize && setup_files(td))
 898                 goto err;
 899
 900         if (td_io_init(td))
 901                 goto err;
 902
 903         if (open_files(td))
 904                 goto err;
 905
 906         if (init_random_map(td))
 907                 goto err;
 908
 909         if (td->o.exec_prerun) {
 910                 if (system(td->o.exec_prerun) < 0)
 911                         goto err;
 912         }
 913
 914         fio_gettime(&td->epoch, NULL);
 915         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 916         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 917
 918         runtime[0] = runtime[1] = 0;
 919         clear_state = 0;
 920         while (keep_running(td)) {
 921                 fio_gettime(&td->start, NULL);
 922                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 923
 924                 if (td->o.ratemin)
 925                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 926                                                         sizeof(td->lastrate));
 927
 928                 if (clear_state && clear_io_state(td))
 929                         break;
 930
 931                 prune_io_piece_log(td);
 932
 933                 do_io(td);
 934
 935                 clear_state = 1;
 936
 937                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 938                         if (td->rw_end_set[DDIR_READ])
 939                                 elapsed = utime_since(&td->start,
 940                                                       &td->rw_end[DDIR_READ]);
 941                         else
 942                                 elapsed = utime_since_now(&td->start);
 943
 944                         runtime[DDIR_READ] += elapsed;
 945                 }
 946                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 947                         if (td->rw_end_set[DDIR_WRITE])
 948                                 elapsed = utime_since(&td->start,
 949                                                       &td->rw_end[DDIR_WRITE]);
 950                         else
 951                                 elapsed = utime_since_now(&td->start);
 952
 953                         runtime[DDIR_WRITE] += elapsed;
 954                 }
 955
 956                 if (td->error || td->terminate)
 957                         break;
 958
 959                 if (!td->o.do_verify ||
 960                     td->o.verify == VERIFY_NONE ||
 961                     (td->io_ops->flags & FIO_UNIDIR))
 962                         continue;
 963
 964                 if (clear_io_state(td))
 965                         break;
 966
 967                 fio_gettime(&td->start, NULL);
 968
 969                 do_verify(td);
 970
 971                 runtime[DDIR_READ] += utime_since_now(&td->start);
 972
 973                 if (td->error || td->terminate)
 974                         break;
 975         }
 976
 977         update_rusage_stat(td);
 978         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 979         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 980         td->ts.total_run_time = mtime_since_now(&td->epoch);
 981         td->ts.io_bytes[0] = td->io_bytes[0];
 982         td->ts.io_bytes[1] = td->io_bytes[1];
 983
 984         if (td->ts.bw_log)
 985                 finish_log(td, td->ts.bw_log, "bw");
 986         if (td->ts.slat_log)
 987                 finish_log(td, td->ts.slat_log, "slat");
 988         if (td->ts.clat_log)
 989                 finish_log(td, td->ts.clat_log, "clat");
 990         if (td->o.exec_postrun) {
 991                 if (system(td->o.exec_postrun) < 0)
 992                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 993         }
 994
 995         if (exitall_on_terminate)
 996                 terminate_threads(td->groupid);
 997
 998 err:
 999         if (td->error)
1000                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1001                                                         td->verror);
1002         close_and_free_files(td);
1003         close_ioengine(td);
1004         cleanup_io_u(td);
1005
1006         /*
1007          * do this very late, it will log file closing as well
1008          */
1009         if (td->o.write_iolog_file)
1010                 write_iolog_close(td);
1011
1012         options_mem_free(td);
1013         td_set_runstate(td, TD_EXITED);
1014         return (void *) (unsigned long) td->error;
1015 }
1016
1017 /*
1018  * We cannot pass the td data into a forked process, so attach the td and
1019  * pass it to the thread worker.
1020  */
1021 static int fork_main(int shmid, int offset)
1022 {
1023         struct thread_data *td;
1024         void *data, *ret;
1025
1026         data = shmat(shmid, NULL, 0);
1027         if (data == (void *) -1) {
1028                 int __err = errno;
1029
1030                 perror("shmat");
1031                 return __err;
1032         }
1033
1034         td = data + offset * sizeof(struct thread_data);
1035         ret = thread_main(td);
1036         shmdt(data);
1037         return (int) (unsigned long) ret;
1038 }
1039
1040 /*
1041  * Run over the job map and reap the threads that have exited, if any.
1042  */
1043 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1044 {
1045         struct thread_data *td;
1046         int i, cputhreads, realthreads, pending, status, ret;
1047
1048         /*
1049          * reap exited threads (TD_EXITED -> TD_REAPED)
1050          */
1051         realthreads = pending = cputhreads = 0;
1052         for_each_td(td, i) {
1053                 int flags = 0;
1054
1055                 /*
1056                  * ->io_ops is NULL for a thread that has closed its
1057                  * io engine
1058                  */
1059                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1060                         cputhreads++;
1061                 else
1062                         realthreads++;
1063
1064                 if (!td->pid) {
1065                         pending++;
1066                         continue;
1067                 }
1068                 if (td->runstate == TD_REAPED)
1069                         continue;
1070                 if (td->o.use_thread) {
1071                         if (td->runstate == TD_EXITED) {
1072                                 td_set_runstate(td, TD_REAPED);
1073                                 goto reaped;
1074                         }
1075                         continue;
1076                 }
1077
1078                 flags = WNOHANG;
1079                 if (td->runstate == TD_EXITED)
1080                         flags = 0;
1081
1082                 /*
1083                  * check if someone quit or got killed in an unusual way
1084                  */
1085                 ret = waitpid(td->pid, &status, flags);
1086                 if (ret < 0) {
1087                         if (errno == ECHILD) {
1088                                 log_err("fio: pid=%d disappeared %d\n",
1089                                                 (int) td->pid, td->runstate);
1090                                 td_set_runstate(td, TD_REAPED);
1091                                 goto reaped;
1092                         }
1093                         perror("waitpid");
1094                 } else if (ret == td->pid) {
1095                         if (WIFSIGNALED(status)) {
1096                                 int sig = WTERMSIG(status);
1097
1098                                 if (sig != SIGQUIT)
1099                                         log_err("fio: pid=%d, got signal=%d\n",
1100                                                         (int) td->pid, sig);
1101                                 td_set_runstate(td, TD_REAPED);
1102                                 goto reaped;
1103                         }
1104                         if (WIFEXITED(status)) {
1105                                 if (WEXITSTATUS(status) && !td->error)
1106                                         td->error = WEXITSTATUS(status);
1107
1108                                 td_set_runstate(td, TD_REAPED);
1109                                 goto reaped;
1110                         }
1111                 }
1112
1113                 /*
1114                  * thread is not dead, continue
1115                  */
1116                 pending++;
1117                 continue;
1118 reaped:
1119                 (*nr_running)--;
1120                 (*m_rate) -= td->o.ratemin;
1121                 (*t_rate) -= td->o.rate;
1122                 if (!td->pid)
1123                         pending--;
1124
1125                 if (td->error)
1126                         exit_value++;
1127
1128                 done_secs += mtime_since_now(&td->epoch) / 1000;
1129         }
1130
1131         if (*nr_running == cputhreads && !pending && realthreads)
1132                 terminate_threads(TERMINATE_ALL);
1133 }
1134
1135 /*
1136  * Main function for kicking off and reaping jobs, as needed.
1137  */
1138 static void run_threads(void)
1139 {
1140         struct thread_data *td;
1141         unsigned long spent;
1142         int i, todo, nr_running, m_rate, t_rate, nr_started;
1143
1144         if (fio_pin_memory())
1145                 return;
1146
1147         if (!terse_output) {
1148                 printf("Starting ");
1149                 if (nr_thread)
1150                         printf("%d thread%s", nr_thread,
1151                                                 nr_thread > 1 ? "s" : "");
1152                 if (nr_process) {
1153                         if (nr_thread)
1154                                 printf(" and ");
1155                         printf("%d process%s", nr_process,
1156                                                 nr_process > 1 ? "es" : "");
1157                 }
1158                 printf("\n");
1159                 fflush(stdout);
1160         }
1161
1162         set_sig_handlers();
1163
1164         todo = thread_number;
1165         nr_running = 0;
1166         nr_started = 0;
1167         m_rate = t_rate = 0;
1168
1169         for_each_td(td, i) {
1170                 print_status_init(td->thread_number - 1);
1171
1172                 if (!td->o.create_serialize) {
1173                         init_disk_util(td);
1174                         continue;
1175                 }
1176
1177                 /*
1178                  * do file setup here so it happens sequentially,
1179                  * we don't want X number of threads getting their
1180                  * client data interspersed on disk
1181                  */
1182                 if (setup_files(td)) {
1183                         exit_value++;
1184                         if (td->error)
1185                                 log_err("fio: pid=%d, err=%d/%s\n",
1186                                         (int) td->pid, td->error, td->verror);
1187                         td_set_runstate(td, TD_REAPED);
1188                         todo--;
1189                 } else {
1190                         struct fio_file *f;
1191                         unsigned int i;
1192
1193                         /*
1194                          * for sharing to work, each job must always open
1195                          * its own files. so close them, if we opened them
1196                          * for creation
1197                          */
1198                         for_each_file(td, f, i)
1199                                 td_io_close_file(td, f);
1200                 }
1201
1202                 init_disk_util(td);
1203         }
1204
1205         set_genesis_time();
1206
1207         while (todo) {
1208                 struct thread_data *map[MAX_JOBS];
1209                 struct timeval this_start;
1210                 int this_jobs = 0, left;
1211
1212                 /*
1213                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1214                  */
1215                 for_each_td(td, i) {
1216                         if (td->runstate != TD_NOT_CREATED)
1217                                 continue;
1218
1219                         /*
1220                          * never got a chance to start, killed by other
1221                          * thread for some reason
1222                          */
1223                         if (td->terminate) {
1224                                 todo--;
1225                                 continue;
1226                         }
1227
1228                         if (td->o.start_delay) {
1229                                 spent = mtime_since_genesis();
1230
1231                                 if (td->o.start_delay * 1000 > spent)
1232                                         continue;
1233                         }
1234
1235                         if (td->o.stonewall && (nr_started || nr_running)) {
1236                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1237                                                         td->o.name);
1238                                 break;
1239                         }
1240
1241                         /*
1242                          * Set state to created. Thread will transition
1243                          * to TD_INITIALIZED when it's done setting up.
1244                          */
1245                         td_set_runstate(td, TD_CREATED);
1246                         map[this_jobs++] = td;
1247                         nr_started++;
1248
1249                         if (td->o.use_thread) {
1250                                 dprint(FD_PROCESS, "will pthread_create\n");
1251                                 if (pthread_create(&td->thread, NULL,
1252                                                    thread_main, td)) {
1253                                         perror("pthread_create");
1254                                         nr_started--;
1255                                         break;
1256                                 }
1257                                 if (pthread_detach(td->thread) < 0)
1258                                         perror("pthread_detach");
1259                         } else {
1260                                 pid_t pid;
1261                                 dprint(FD_PROCESS, "will fork\n");
1262                                 pid = fork();
1263                                 if (!pid) {
1264                                         int ret = fork_main(shm_id, i);
1265
1266                                         _exit(ret);
1267                                 } else if (i == fio_debug_jobno)
1268                                         *fio_debug_jobp = pid;
1269                         }
1270                         fio_mutex_down(startup_mutex);
1271                 }
1272
1273                 /*
1274                  * Wait for the started threads to transition to
1275                  * TD_INITIALIZED.
1276                  */
1277                 fio_gettime(&this_start, NULL);
1278                 left = this_jobs;
1279                 while (left && !fio_abort) {
1280                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1281                                 break;
1282
1283                         usleep(100000);
1284
1285                         for (i = 0; i < this_jobs; i++) {
1286                                 td = map[i];
1287                                 if (!td)
1288                                         continue;
1289                                 if (td->runstate == TD_INITIALIZED) {
1290                                         map[i] = NULL;
1291                                         left--;
1292                                 } else if (td->runstate >= TD_EXITED) {
1293                                         map[i] = NULL;
1294                                         left--;
1295                                         todo--;
1296                                         nr_running++; /* work-around... */
1297                                 }
1298                         }
1299                 }
1300
1301                 if (left) {
1302                         log_err("fio: %d jobs failed to start\n", left);
1303                         for (i = 0; i < this_jobs; i++) {
1304                                 td = map[i];
1305                                 if (!td)
1306                                         continue;
1307                                 kill(td->pid, SIGTERM);
1308                         }
1309                         break;
1310                 }
1311
1312                 /*
1313                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1314                  */
1315                 for_each_td(td, i) {
1316                         if (td->runstate != TD_INITIALIZED)
1317                                 continue;
1318
1319                         td_set_runstate(td, TD_RUNNING);
1320                         nr_running++;
1321                         nr_started--;
1322                         m_rate += td->o.ratemin;
1323                         t_rate += td->o.rate;
1324                         todo--;
1325                         fio_mutex_up(td->mutex);
1326                 }
1327
1328                 reap_threads(&nr_running, &t_rate, &m_rate);
1329
1330                 if (todo)
1331                         usleep(100000);
1332         }
1333
1334         while (nr_running) {
1335                 reap_threads(&nr_running, &t_rate, &m_rate);
1336                 usleep(10000);
1337         }
1338
1339         update_io_ticks();
1340         fio_unpin_memory();
1341 }
1342
1343 int main(int argc, char *argv[])
1344 {
1345         long ps;
1346
1347         sinit();
1348
1349         /*
1350          * We need locale for number printing, if it isn't set then just
1351          * go with the US format.
1352          */
1353         if (!getenv("LC_NUMERIC"))
1354                 setlocale(LC_NUMERIC, "en_US");
1355
1356         if (parse_options(argc, argv))
1357                 return 1;
1358
1359         if (!thread_number)
1360                 return 0;
1361
1362         ps = sysconf(_SC_PAGESIZE);
1363         if (ps < 0) {
1364                 log_err("Failed to get page size\n");
1365                 return 1;
1366         }
1367
1368         page_size = ps;
1369         page_mask = ps - 1;
1370
1371         if (write_bw_log) {
1372                 setup_log(&agg_io_log[DDIR_READ]);
1373                 setup_log(&agg_io_log[DDIR_WRITE]);
1374         }
1375
1376         startup_mutex = fio_mutex_init(0);
1377
1378         set_genesis_time();
1379
1380         status_timer_arm();
1381
1382         run_threads();
1383
1384         if (!fio_abort) {
1385                 show_run_stats();
1386                 if (write_bw_log) {
1387                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1388                         __finish_log(agg_io_log[DDIR_WRITE],
1389                                         "agg-write_bw.log");
1390                 }
1391         }
1392
1393         fio_mutex_remove(startup_mutex);
1394         return exit_value;
1395 }