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