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