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
  39 unsigned long page_mask;
  40 unsigned long page_size;
  41 #define ALIGN(buf)      \
  42         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  43
  44 int groupid = 0;
  45 int thread_number = 0;
  46 int nr_process = 0;
  47 int nr_thread = 0;
  48 int shm_id = 0;
  49 int temp_stall_ts;
  50
  51 static struct fio_sem *startup_sem;
  52 static volatile int fio_abort;
  53 static int exit_value;
  54
  55 struct io_log *agg_io_log[2];
  56
  57 #define TERMINATE_ALL           (-1)
  58 #define JOB_START_TIMEOUT       (5 * 1000)
  59
  60 static inline void td_set_runstate(struct thread_data *td, int runstate)
  61 {
  62         td->runstate = runstate;
  63 }
  64
  65 static void terminate_threads(int group_id)
  66 {
  67         struct thread_data *td;
  68         int i;
  69
  70         for_each_td(td, i) {
  71                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  72                         if (td->runstate <= TD_RUNNING)
  73                                 kill(td->pid, SIGQUIT);
  74                         td->terminate = 1;
  75                         td->o.start_delay = 0;
  76                 }
  77         }
  78 }
  79
  80 static void sig_handler(int sig)
  81 {
  82         switch (sig) {
  83                 case SIGALRM:
  84                         update_io_ticks();
  85                         disk_util_timer_arm();
  86                         print_thread_status();
  87                         break;
  88                 default:
  89                         printf("\nfio: terminating on signal %d\n", sig);
  90                         fflush(stdout);
  91                         terminate_threads(TERMINATE_ALL);
  92                         break;
  93         }
  94 }
  95
  96 /*
  97  * Check if we are above the minimum rate given.
  98  */
  99 static int check_min_rate(struct thread_data *td, struct timeval *now)
 100 {
 101         unsigned long long bytes = 0;
 102         unsigned long iops = 0;
 103         unsigned long spent;
 104         unsigned long rate;
 105
 106         /*
 107          * No minimum rate set, always ok
 108          */
 109         if (!td->o.ratemin && !td->o.rate_iops_min)
 110                 return 0;
 111
 112         /*
 113          * allow a 2 second settle period in the beginning
 114          */
 115         if (mtime_since(&td->start, now) < 2000)
 116                 return 0;
 117
 118         if (td_read(td)) {
 119                 iops += td->io_blocks[DDIR_READ];
 120                 bytes += td->this_io_bytes[DDIR_READ];
 121         }
 122         if (td_write(td)) {
 123                 iops += td->io_blocks[DDIR_WRITE];
 124                 bytes += td->this_io_bytes[DDIR_WRITE];
 125         }
 126
 127         /*
 128          * if rate blocks is set, sample is running
 129          */
 130         if (td->rate_bytes || td->rate_blocks) {
 131                 spent = mtime_since(&td->lastrate, now);
 132                 if (spent < td->o.ratecycle)
 133                         return 0;
 134
 135                 if (td->o.rate) {
 136                         /*
 137                          * check bandwidth specified rate
 138                          */
 139                         if (bytes < td->rate_bytes) {
 140                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
 141                                 return 1;
 142                         } else {
 143                                 rate = (bytes - td->rate_bytes) / spent;
 144                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
 145                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
 146                                         return 1;
 147                                 }
 148                         }
 149                 } else {
 150                         /*
 151                          * checks iops specified rate
 152                          */
 153                         if (iops < td->o.rate_iops) {
 154                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
 155                                 return 1;
 156                         } else {
 157                                 rate = (iops - td->rate_blocks) / spent;
 158                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
 159                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
 160                                 }
 161                         }
 162                 }
 163         }
 164
 165         td->rate_bytes = bytes;
 166         td->rate_blocks = iops;
 167         memcpy(&td->lastrate, now, sizeof(*now));
 168         return 0;
 169 }
 170
 171 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 172 {
 173         if (!td->o.timeout)
 174                 return 0;
 175         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 176                 return 1;
 177
 178         return 0;
 179 }
 180
 181 /*
 182  * When job exits, we can cancel the in-flight IO if we are using async
 183  * io. Attempt to do so.
 184  */
 185 static void cleanup_pending_aio(struct thread_data *td)
 186 {
 187         struct list_head *entry, *n;
 188         struct io_u *io_u;
 189         int r;
 190
 191         /*
 192          * get immediately available events, if any
 193          */
 194         r = io_u_queued_complete(td, 0);
 195         if (r < 0)
 196                 return;
 197
 198         /*
 199          * now cancel remaining active events
 200          */
 201         if (td->io_ops->cancel) {
 202                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 203                         io_u = list_entry(entry, struct io_u, list);
 204
 205                         /*
 206                          * if the io_u isn't in flight, then that generally
 207                          * means someone leaked an io_u. complain but fix
 208                          * it up, so we don't stall here.
 209                          */
 210                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 211                                 log_err("fio: non-busy IO on busy list\n");
 212                                 put_io_u(td, io_u);
 213                         } else {
 214                                 r = td->io_ops->cancel(td, io_u);
 215                                 if (!r)
 216                                         put_io_u(td, io_u);
 217                         }
 218                 }
 219         }
 220
 221         if (td->cur_depth)
 222                 r = io_u_queued_complete(td, td->cur_depth);
 223 }
 224
 225 /*
 226  * Helper to handle the final sync of a file. Works just like the normal
 227  * io path, just does everything sync.
 228  */
 229 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 230 {
 231         struct io_u *io_u = __get_io_u(td);
 232         int ret;
 233
 234         if (!io_u)
 235                 return 1;
 236
 237         io_u->ddir = DDIR_SYNC;
 238         io_u->file = f;
 239
 240         if (td_io_prep(td, io_u)) {
 241                 put_io_u(td, io_u);
 242                 return 1;
 243         }
 244
 245 requeue:
 246         ret = td_io_queue(td, io_u);
 247         if (ret < 0) {
 248                 td_verror(td, io_u->error, "td_io_queue");
 249                 put_io_u(td, io_u);
 250                 return 1;
 251         } else if (ret == FIO_Q_QUEUED) {
 252                 if (io_u_queued_complete(td, 1) < 0)
 253                         return 1;
 254         } else if (ret == FIO_Q_COMPLETED) {
 255                 if (io_u->error) {
 256                         td_verror(td, io_u->error, "td_io_queue");
 257                         return 1;
 258                 }
 259
 260                 if (io_u_sync_complete(td, io_u) < 0)
 261                         return 1;
 262         } else if (ret == FIO_Q_BUSY) {
 263                 if (td_io_commit(td))
 264                         return 1;
 265                 goto requeue;
 266         }
 267
 268         return 0;
 269 }
 270
 271 /*
 272  * The main verify engine. Runs over the writes we previously submitted,
 273  * reads the blocks back in, and checks the crc/md5 of the data.
 274  */
 275 static void do_verify(struct thread_data *td)
 276 {
 277         struct fio_file *f;
 278         struct io_u *io_u;
 279         int ret, min_events;
 280         unsigned int i;
 281
 282         /*
 283          * sync io first and invalidate cache, to make sure we really
 284          * read from disk.
 285          */
 286         for_each_file(td, f, i) {
 287                 if (!(f->flags & FIO_FILE_OPEN))
 288                         continue;
 289                 if (fio_io_sync(td, f))
 290                         break;
 291                 if (file_invalidate_cache(td, f))
 292                         break;
 293         }
 294
 295         if (td->error)
 296                 return;
 297
 298         td_set_runstate(td, TD_VERIFYING);
 299
 300         io_u = NULL;
 301         while (!td->terminate) {
 302                 int ret2;
 303
 304                 io_u = __get_io_u(td);
 305                 if (!io_u)
 306                         break;
 307
 308                 if (runtime_exceeded(td, &io_u->start_time)) {
 309                         put_io_u(td, io_u);
 310                         break;
 311                 }
 312
 313                 if (get_next_verify(td, io_u)) {
 314                         put_io_u(td, io_u);
 315                         break;
 316                 }
 317
 318                 if (td_io_prep(td, io_u)) {
 319                         put_io_u(td, io_u);
 320                         break;
 321                 }
 322
 323                 io_u->end_io = verify_io_u;
 324
 325                 ret = td_io_queue(td, io_u);
 326                 switch (ret) {
 327                 case FIO_Q_COMPLETED:
 328                         if (io_u->error)
 329                                 ret = -io_u->error;
 330                         else if (io_u->resid) {
 331                                 int bytes = io_u->xfer_buflen - io_u->resid;
 332                                 struct fio_file *f = io_u->file;
 333
 334                                 /*
 335                                  * zero read, fail
 336                                  */
 337                                 if (!bytes) {
 338                                         td_verror(td, ENODATA, "full resid");
 339                                         put_io_u(td, io_u);
 340                                         break;
 341                                 }
 342
 343                                 io_u->xfer_buflen = io_u->resid;
 344                                 io_u->xfer_buf += bytes;
 345                                 io_u->offset += bytes;
 346                                 f->last_completed_pos = io_u->offset;
 347
 348                                 td->ts.short_io_u[io_u->ddir]++;
 349
 350                                 if (io_u->offset == f->real_file_size)
 351                                         goto sync_done;
 352
 353                                 requeue_io_u(td, &io_u);
 354                         } else {
 355 sync_done:
 356                                 ret = io_u_sync_complete(td, io_u);
 357                                 if (ret < 0)
 358                                         break;
 359                         }
 360                         continue;
 361                 case FIO_Q_QUEUED:
 362                         break;
 363                 case FIO_Q_BUSY:
 364                         requeue_io_u(td, &io_u);
 365                         ret2 = td_io_commit(td);
 366                         if (ret2 < 0)
 367                                 ret = ret2;
 368                         break;
 369                 default:
 370                         assert(ret < 0);
 371                         td_verror(td, -ret, "td_io_queue");
 372                         break;
 373                 }
 374
 375                 if (ret < 0 || td->error)
 376                         break;
 377
 378                 /*
 379                  * if we can queue more, do so. but check if there are
 380                  * completed io_u's first.
 381                  */
 382                 min_events = 0;
 383                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 384                         min_events = 1;
 385
 386                         if (td->cur_depth > td->o.iodepth_low)
 387                                 min_events = td->cur_depth - td->o.iodepth_low;
 388                 }
 389
 390                 /*
 391                  * Reap required number of io units, if any, and do the
 392                  * verification on them through the callback handler
 393                  */
 394                 if (io_u_queued_complete(td, min_events) < 0)
 395                         break;
 396         }
 397
 398         if (!td->error) {
 399                 min_events = td->cur_depth;
 400
 401                 if (min_events)
 402                         ret = io_u_queued_complete(td, min_events);
 403         } else
 404                 cleanup_pending_aio(td);
 405
 406         td_set_runstate(td, TD_RUNNING);
 407 }
 408
 409 /*
 410  * Main IO worker function. It retrieves io_u's to process and queues
 411  * and reaps them, checking for rate and errors along the way.
 412  */
 413 static void do_io(struct thread_data *td)
 414 {
 415         struct timeval s;
 416         unsigned long usec;
 417         unsigned int i;
 418         int ret = 0;
 419
 420         td_set_runstate(td, TD_RUNNING);
 421
 422         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 423                 struct timeval comp_time;
 424                 long bytes_done = 0;
 425                 int min_evts = 0;
 426                 struct io_u *io_u;
 427                 int ret2;
 428
 429                 if (td->terminate)
 430                         break;
 431
 432                 io_u = get_io_u(td);
 433                 if (!io_u)
 434                         break;
 435
 436                 memcpy(&s, &io_u->start_time, sizeof(s));
 437
 438                 if (runtime_exceeded(td, &s)) {
 439                         put_io_u(td, io_u);
 440                         break;
 441                 }
 442
 443                 ret = td_io_queue(td, io_u);
 444                 switch (ret) {
 445                 case FIO_Q_COMPLETED:
 446                         if (io_u->error)
 447                                 ret = -io_u->error;
 448                         else if (io_u->resid) {
 449                                 int bytes = io_u->xfer_buflen - io_u->resid;
 450                                 struct fio_file *f = io_u->file;
 451
 452                                 /*
 453                                  * zero read, fail
 454                                  */
 455                                 if (!bytes) {
 456                                         td_verror(td, ENODATA, "full resid");
 457                                         put_io_u(td, io_u);
 458                                         break;
 459                                 }
 460
 461                                 io_u->xfer_buflen = io_u->resid;
 462                                 io_u->xfer_buf += bytes;
 463                                 io_u->offset += bytes;
 464                                 f->last_completed_pos = io_u->offset;
 465
 466                                 td->ts.short_io_u[io_u->ddir]++;
 467
 468                                 if (io_u->offset == f->real_file_size)
 469                                         goto sync_done;
 470
 471                                 requeue_io_u(td, &io_u);
 472                         } else {
 473 sync_done:
 474                                 fio_gettime(&comp_time, NULL);
 475                                 bytes_done = io_u_sync_complete(td, io_u);
 476                                 if (bytes_done < 0)
 477                                         ret = bytes_done;
 478                         }
 479                         break;
 480                 case FIO_Q_QUEUED:
 481                         /*
 482                          * if the engine doesn't have a commit hook,
 483                          * the io_u is really queued. if it does have such
 484                          * a hook, it has to call io_u_queued() itself.
 485                          */
 486                         if (td->io_ops->commit == NULL)
 487                                 io_u_queued(td, io_u);
 488                         break;
 489                 case FIO_Q_BUSY:
 490                         requeue_io_u(td, &io_u);
 491                         ret2 = td_io_commit(td);
 492                         if (ret2 < 0)
 493                                 ret = ret2;
 494                         break;
 495                 default:
 496                         assert(ret < 0);
 497                         put_io_u(td, io_u);
 498                         break;
 499                 }
 500
 501                 if (ret < 0 || td->error)
 502                         break;
 503
 504                 /*
 505                  * See if we need to complete some commands
 506                  */
 507                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 508                         min_evts = 0;
 509                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 510                                 min_evts = 1;
 511
 512                                 if (td->cur_depth > td->o.iodepth_low)
 513                                         min_evts = td->cur_depth - td->o.iodepth_low;
 514                         }
 515
 516                         fio_gettime(&comp_time, NULL);
 517                         bytes_done = io_u_queued_complete(td, min_evts);
 518                         if (bytes_done < 0)
 519                                 break;
 520                 }
 521
 522                 if (!bytes_done)
 523                         continue;
 524
 525                 /*
 526                  * the rate is batched for now, it should work for batches
 527                  * of completions except the very first one which may look
 528                  * a little bursty
 529                  */
 530                 usec = utime_since(&s, &comp_time);
 531
 532                 rate_throttle(td, usec, bytes_done);
 533
 534                 if (check_min_rate(td, &comp_time)) {
 535                         if (exitall_on_terminate)
 536                                 terminate_threads(td->groupid);
 537                         td_verror(td, ENODATA, "check_min_rate");
 538                         break;
 539                 }
 540
 541                 if (td->o.thinktime) {
 542                         unsigned long long b;
 543
 544                         b = td->io_blocks[0] + td->io_blocks[1];
 545                         if (!(b % td->o.thinktime_blocks)) {
 546                                 int left;
 547
 548                                 if (td->o.thinktime_spin)
 549                                         __usec_sleep(td->o.thinktime_spin);
 550
 551                                 left = td->o.thinktime - td->o.thinktime_spin;
 552                                 if (left)
 553                                         usec_sleep(td, left);
 554                         }
 555                 }
 556         }
 557
 558         if (!td->error) {
 559                 struct fio_file *f;
 560
 561                 i = td->cur_depth;
 562                 if (i)
 563                         ret = io_u_queued_complete(td, i);
 564
 565                 if (should_fsync(td) && td->o.end_fsync) {
 566                         td_set_runstate(td, TD_FSYNCING);
 567
 568                         for_each_file(td, f, i) {
 569                                 if (!(f->flags & FIO_FILE_OPEN))
 570                                         continue;
 571                                 fio_io_sync(td, f);
 572                         }
 573                 }
 574         } else
 575                 cleanup_pending_aio(td);
 576 }
 577
 578 static void cleanup_io_u(struct thread_data *td)
 579 {
 580         struct list_head *entry, *n;
 581         struct io_u *io_u;
 582
 583         list_for_each_safe(entry, n, &td->io_u_freelist) {
 584                 io_u = list_entry(entry, struct io_u, list);
 585
 586                 list_del(&io_u->list);
 587                 free(io_u);
 588         }
 589
 590         free_io_mem(td);
 591 }
 592
 593 /*
 594  * "randomly" fill the buffer contents
 595  */
 596 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 597 {
 598         int *ptr = io_u->buf;
 599
 600         if (!td->o.zero_buffers) {
 601                 while ((void *) ptr - io_u->buf < max_bs) {
 602                         *ptr = rand() * 0x9e370001;
 603                         ptr++;
 604                 }
 605         } else
 606                 memset(ptr, 0, max_bs);
 607 }
 608
 609 static int init_io_u(struct thread_data *td)
 610 {
 611         unsigned long long buf_size;
 612         struct io_u *io_u;
 613         unsigned int max_bs;
 614         int i, max_units;
 615         char *p;
 616
 617         if (td->io_ops->flags & FIO_SYNCIO)
 618                 max_units = 1;
 619         else
 620                 max_units = td->o.iodepth;
 621
 622         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 623         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 624         buf_size += page_mask;
 625         if (buf_size != (size_t) buf_size) {
 626                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 627                 return 1;
 628         }
 629
 630         td->orig_buffer_size = buf_size;
 631
 632         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 633                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 634         else if (td->orig_buffer_size & page_mask)
 635                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
 636
 637         if (allocate_io_mem(td))
 638                 return 1;
 639
 640         p = ALIGN(td->orig_buffer);
 641         for (i = 0; i < max_units; i++) {
 642                 io_u = malloc(sizeof(*io_u));
 643                 memset(io_u, 0, sizeof(*io_u));
 644                 INIT_LIST_HEAD(&io_u->list);
 645
 646                 io_u->buf = p + max_bs * i;
 647
 648                 if (td_write(td))
 649                         fill_io_buf(td, io_u, max_bs);
 650
 651                 io_u->index = i;
 652                 io_u->flags = IO_U_F_FREE;
 653                 list_add(&io_u->list, &td->io_u_freelist);
 654         }
 655
 656         io_u_init_timeout();
 657
 658         return 0;
 659 }
 660
 661 static int switch_ioscheduler(struct thread_data *td)
 662 {
 663         char tmp[256], tmp2[128];
 664         FILE *f;
 665         int ret;
 666
 667         if (td->io_ops->flags & FIO_DISKLESSIO)
 668                 return 0;
 669
 670         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 671
 672         f = fopen(tmp, "r+");
 673         if (!f) {
 674                 if (errno == ENOENT) {
 675                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 676                         return 0;
 677                 }
 678                 td_verror(td, errno, "fopen iosched");
 679                 return 1;
 680         }
 681
 682         /*
 683          * Set io scheduler.
 684          */
 685         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 686         if (ferror(f) || ret != 1) {
 687                 td_verror(td, errno, "fwrite");
 688                 fclose(f);
 689                 return 1;
 690         }
 691
 692         rewind(f);
 693
 694         /*
 695          * Read back and check that the selected scheduler is now the default.
 696          */
 697         ret = fread(tmp, 1, sizeof(tmp), f);
 698         if (ferror(f) || ret < 0) {
 699                 td_verror(td, errno, "fread");
 700                 fclose(f);
 701                 return 1;
 702         }
 703
 704         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 705         if (!strstr(tmp, tmp2)) {
 706                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 707                 td_verror(td, EINVAL, "iosched_switch");
 708                 fclose(f);
 709                 return 1;
 710         }
 711
 712         fclose(f);
 713         return 0;
 714 }
 715
 716 static int clear_io_state(struct thread_data *td)
 717 {
 718         struct fio_file *f;
 719         unsigned int i;
 720         int ret;
 721
 722         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 723         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 724         td->zone_bytes = 0;
 725         td->rate_bytes = 0;
 726         td->rate_blocks = 0;
 727         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 728
 729         td->last_was_sync = 0;
 730
 731         for_each_file(td, f, i)
 732                 td_io_close_file(td, f);
 733
 734         ret = 0;
 735         for_each_file(td, f, i) {
 736                 ret = td_io_open_file(td, f);
 737                 if (ret)
 738                         break;
 739         }
 740
 741         return ret;
 742 }
 743
 744 /*
 745  * Entry point for the thread based jobs. The process based jobs end up
 746  * here as well, after a little setup.
 747  */
 748 static void *thread_main(void *data)
 749 {
 750         unsigned long long runtime[2];
 751         struct thread_data *td = data;
 752         unsigned long elapsed;
 753         int clear_state;
 754
 755         if (!td->o.use_thread)
 756                 setsid();
 757
 758         td->pid = getpid();
 759
 760         INIT_LIST_HEAD(&td->io_u_freelist);
 761         INIT_LIST_HEAD(&td->io_u_busylist);
 762         INIT_LIST_HEAD(&td->io_u_requeues);
 763         INIT_LIST_HEAD(&td->io_log_list);
 764         INIT_LIST_HEAD(&td->io_hist_list);
 765         td->io_hist_tree = RB_ROOT;
 766
 767         if (init_io_u(td))
 768                 goto err_sem;
 769
 770         if (fio_setaffinity(td) == -1) {
 771                 td_verror(td, errno, "cpu_set_affinity");
 772                 goto err_sem;
 773         }
 774
 775         if (init_iolog(td))
 776                 goto err_sem;
 777
 778         if (td->ioprio) {
 779                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 780                         td_verror(td, errno, "ioprio_set");
 781                         goto err_sem;
 782                 }
 783         }
 784
 785         if (nice(td->o.nice) == -1) {
 786                 td_verror(td, errno, "nice");
 787                 goto err_sem;
 788         }
 789
 790         if (td->o.ioscheduler && switch_ioscheduler(td))
 791                 goto err_sem;
 792
 793         td_set_runstate(td, TD_INITIALIZED);
 794         fio_sem_up(startup_sem);
 795         fio_sem_down(td->mutex);
 796
 797         /*
 798          * the ->mutex semaphore is now no longer used, close it to avoid
 799          * eating a file descriptor
 800          */
 801         fio_sem_remove(td->mutex);
 802
 803         if (!td->o.create_serialize && setup_files(td))
 804                 goto err;
 805
 806         if (td_io_init(td))
 807                 goto err;
 808
 809         if (open_files(td))
 810                 goto err;
 811
 812         if (init_random_map(td))
 813                 goto err;
 814
 815         if (td->o.exec_prerun) {
 816                 if (system(td->o.exec_prerun) < 0)
 817                         goto err;
 818         }
 819
 820         fio_gettime(&td->epoch, NULL);
 821         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 822         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 823
 824         runtime[0] = runtime[1] = 0;
 825         clear_state = 0;
 826         while (td->o.loops--) {
 827                 fio_gettime(&td->start, NULL);
 828                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 829
 830                 if (td->o.ratemin)
 831                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 832
 833                 if (clear_state && clear_io_state(td))
 834                         break;
 835
 836                 prune_io_piece_log(td);
 837
 838                 do_io(td);
 839
 840                 clear_state = 1;
 841
 842                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 843                         if (td->rw_end_set[DDIR_READ])
 844                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 845                         else
 846                                 elapsed = utime_since_now(&td->start);
 847
 848                         runtime[DDIR_READ] += elapsed;
 849                 }
 850                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 851                         if (td->rw_end_set[DDIR_WRITE])
 852                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 853                         else
 854                                 elapsed = utime_since_now(&td->start);
 855
 856                         runtime[DDIR_WRITE] += elapsed;
 857                 }
 858
 859                 if (td->error || td->terminate)
 860                         break;
 861
 862                 if (td->o.verify == VERIFY_NONE)
 863                         continue;
 864
 865                 if (clear_io_state(td))
 866                         break;
 867
 868                 fio_gettime(&td->start, NULL);
 869
 870                 do_verify(td);
 871
 872                 runtime[DDIR_READ] += utime_since_now(&td->start);
 873
 874                 if (td->error || td->terminate)
 875                         break;
 876         }
 877
 878         update_rusage_stat(td);
 879         td->ts.runtime[0] = runtime[0] / 1000;
 880         td->ts.runtime[1] = runtime[1] / 1000;
 881         td->ts.total_run_time = mtime_since_now(&td->epoch);
 882         td->ts.io_bytes[0] = td->io_bytes[0];
 883         td->ts.io_bytes[1] = td->io_bytes[1];
 884
 885         if (td->ts.bw_log)
 886                 finish_log(td, td->ts.bw_log, "bw");
 887         if (td->ts.slat_log)
 888                 finish_log(td, td->ts.slat_log, "slat");
 889         if (td->ts.clat_log)
 890                 finish_log(td, td->ts.clat_log, "clat");
 891         if (td->o.write_iolog_file)
 892                 write_iolog_close(td);
 893         if (td->o.exec_postrun) {
 894                 if (system(td->o.exec_postrun) < 0)
 895                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 896         }
 897
 898         if (exitall_on_terminate)
 899                 terminate_threads(td->groupid);
 900
 901 err:
 902         if (td->error)
 903                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 904         close_files(td);
 905         close_ioengine(td);
 906         cleanup_io_u(td);
 907         options_mem_free(td);
 908         td_set_runstate(td, TD_EXITED);
 909         return (void *) (unsigned long) td->error;
 910 err_sem:
 911         fio_sem_up(startup_sem);
 912         goto err;
 913 }
 914
 915 /*
 916  * We cannot pass the td data into a forked process, so attach the td and
 917  * pass it to the thread worker.
 918  */
 919 static int fork_main(int shmid, int offset)
 920 {
 921         struct thread_data *td;
 922         void *data, *ret;
 923
 924         data = shmat(shmid, NULL, 0);
 925         if (data == (void *) -1) {
 926                 int __err = errno;
 927
 928                 perror("shmat");
 929                 return __err;
 930         }
 931
 932         td = data + offset * sizeof(struct thread_data);
 933         ret = thread_main(td);
 934         shmdt(data);
 935         return (int) (unsigned long) ret;
 936 }
 937
 938 /*
 939  * Run over the job map and reap the threads that have exited, if any.
 940  */
 941 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 942 {
 943         struct thread_data *td;
 944         int i, cputhreads, pending, status, ret;
 945
 946         /*
 947          * reap exited threads (TD_EXITED -> TD_REAPED)
 948          */
 949         pending = cputhreads = 0;
 950         for_each_td(td, i) {
 951                 int flags = 0;
 952
 953                 /*
 954                  * ->io_ops is NULL for a thread that has closed its
 955                  * io engine
 956                  */
 957                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 958                         cputhreads++;
 959
 960                 if (!td->pid || td->runstate == TD_REAPED)
 961                         continue;
 962                 if (td->o.use_thread) {
 963                         if (td->runstate == TD_EXITED) {
 964                                 td_set_runstate(td, TD_REAPED);
 965                                 goto reaped;
 966                         }
 967                         continue;
 968                 }
 969
 970                 flags = WNOHANG;
 971                 if (td->runstate == TD_EXITED)
 972                         flags = 0;
 973
 974                 /*
 975                  * check if someone quit or got killed in an unusual way
 976                  */
 977                 ret = waitpid(td->pid, &status, flags);
 978                 if (ret < 0) {
 979                         if (errno == ECHILD) {
 980                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 981                                 td_set_runstate(td, TD_REAPED);
 982                                 goto reaped;
 983                         }
 984                         perror("waitpid");
 985                 } else if (ret == td->pid) {
 986                         if (WIFSIGNALED(status)) {
 987                                 int sig = WTERMSIG(status);
 988
 989                                 if (sig != SIGQUIT)
 990                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 991                                 td_set_runstate(td, TD_REAPED);
 992                                 goto reaped;
 993                         }
 994                         if (WIFEXITED(status)) {
 995                                 if (WEXITSTATUS(status) && !td->error)
 996                                         td->error = WEXITSTATUS(status);
 997
 998                                 td_set_runstate(td, TD_REAPED);
 999                                 goto reaped;
1000                         }
1001                 }
1002
1003                 /*
1004                  * thread is not dead, continue
1005                  */
1006                 pending++;
1007                 continue;
1008 reaped:
1009                 if (td->o.use_thread) {
1010                         long ret;
1011
1012                         if (pthread_join(td->thread, (void *) &ret))
1013                                 perror("pthread_join");
1014                 }
1015
1016                 (*nr_running)--;
1017                 (*m_rate) -= td->o.ratemin;
1018                 (*t_rate) -= td->o.rate;
1019                 pending--;
1020
1021                 if (td->error)
1022                         exit_value++;
1023         }
1024
1025         if (*nr_running == cputhreads && !pending)
1026                 terminate_threads(TERMINATE_ALL);
1027 }
1028
1029 /*
1030  * Main function for kicking off and reaping jobs, as needed.
1031  */
1032 static void run_threads(void)
1033 {
1034         struct thread_data *td;
1035         unsigned long spent;
1036         int i, todo, nr_running, m_rate, t_rate, nr_started;
1037
1038         if (fio_pin_memory())
1039                 return;
1040
1041         if (!terse_output) {
1042                 printf("Starting ");
1043                 if (nr_thread)
1044                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1045                 if (nr_process) {
1046                         if (nr_thread)
1047                                 printf(" and ");
1048                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1049                 }
1050                 printf("\n");
1051                 fflush(stdout);
1052         }
1053
1054         signal(SIGINT, sig_handler);
1055         signal(SIGALRM, sig_handler);
1056
1057         todo = thread_number;
1058         nr_running = 0;
1059         nr_started = 0;
1060         m_rate = t_rate = 0;
1061
1062         for_each_td(td, i) {
1063                 print_status_init(td->thread_number - 1);
1064
1065                 if (!td->o.create_serialize) {
1066                         init_disk_util(td);
1067                         continue;
1068                 }
1069
1070                 /*
1071                  * do file setup here so it happens sequentially,
1072                  * we don't want X number of threads getting their
1073                  * client data interspersed on disk
1074                  */
1075                 if (setup_files(td)) {
1076                         exit_value++;
1077                         if (td->error)
1078                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1079                         td_set_runstate(td, TD_REAPED);
1080                         todo--;
1081                 }
1082
1083                 init_disk_util(td);
1084         }
1085
1086         set_genesis_time();
1087
1088         while (todo) {
1089                 struct thread_data *map[MAX_JOBS];
1090                 struct timeval this_start;
1091                 int this_jobs = 0, left;
1092
1093                 /*
1094                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1095                  */
1096                 for_each_td(td, i) {
1097                         if (td->runstate != TD_NOT_CREATED)
1098                                 continue;
1099
1100                         /*
1101                          * never got a chance to start, killed by other
1102                          * thread for some reason
1103                          */
1104                         if (td->terminate) {
1105                                 todo--;
1106                                 continue;
1107                         }
1108
1109                         if (td->o.start_delay) {
1110                                 spent = mtime_since_genesis();
1111
1112                                 if (td->o.start_delay * 1000 > spent)
1113                                         continue;
1114                         }
1115
1116                         if (td->o.stonewall && (nr_started || nr_running))
1117                                 break;
1118
1119                         /*
1120                          * Set state to created. Thread will transition
1121                          * to TD_INITIALIZED when it's done setting up.
1122                          */
1123                         td_set_runstate(td, TD_CREATED);
1124                         map[this_jobs++] = td;
1125                         nr_started++;
1126
1127                         if (td->o.use_thread) {
1128                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1129                                         perror("thread_create");
1130                                         nr_started--;
1131                                         break;
1132                                 }
1133                         } else {
1134                                 if (!fork()) {
1135                                         int ret = fork_main(shm_id, i);
1136
1137                                         exit(ret);
1138                                 }
1139                         }
1140                         fio_sem_down(startup_sem);
1141                 }
1142
1143                 /*
1144                  * Wait for the started threads to transition to
1145                  * TD_INITIALIZED.
1146                  */
1147                 fio_gettime(&this_start, NULL);
1148                 left = this_jobs;
1149                 while (left && !fio_abort) {
1150                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1151                                 break;
1152
1153                         usleep(100000);
1154
1155                         for (i = 0; i < this_jobs; i++) {
1156                                 td = map[i];
1157                                 if (!td)
1158                                         continue;
1159                                 if (td->runstate == TD_INITIALIZED) {
1160                                         map[i] = NULL;
1161                                         left--;
1162                                 } else if (td->runstate >= TD_EXITED) {
1163                                         map[i] = NULL;
1164                                         left--;
1165                                         todo--;
1166                                         nr_running++; /* work-around... */
1167                                 }
1168                         }
1169                 }
1170
1171                 if (left) {
1172                         log_err("fio: %d jobs failed to start\n", left);
1173                         for (i = 0; i < this_jobs; i++) {
1174                                 td = map[i];
1175                                 if (!td)
1176                                         continue;
1177                                 kill(td->pid, SIGTERM);
1178                         }
1179                         break;
1180                 }
1181
1182                 /*
1183                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1184                  */
1185                 for_each_td(td, i) {
1186                         if (td->runstate != TD_INITIALIZED)
1187                                 continue;
1188
1189                         td_set_runstate(td, TD_RUNNING);
1190                         nr_running++;
1191                         nr_started--;
1192                         m_rate += td->o.ratemin;
1193                         t_rate += td->o.rate;
1194                         todo--;
1195                         fio_sem_up(td->mutex);
1196                 }
1197
1198                 reap_threads(&nr_running, &t_rate, &m_rate);
1199
1200                 if (todo)
1201                         usleep(100000);
1202         }
1203
1204         while (nr_running) {
1205                 reap_threads(&nr_running, &t_rate, &m_rate);
1206                 usleep(10000);
1207         }
1208
1209         update_io_ticks();
1210         fio_unpin_memory();
1211 }
1212
1213 int main(int argc, char *argv[])
1214 {
1215         long ps;
1216
1217         /*
1218          * We need locale for number printing, if it isn't set then just
1219          * go with the US format.
1220          */
1221         if (!getenv("LC_NUMERIC"))
1222                 setlocale(LC_NUMERIC, "en_US");
1223
1224         if (parse_options(argc, argv))
1225                 return 1;
1226
1227         if (!thread_number)
1228                 return 0;
1229
1230         ps = sysconf(_SC_PAGESIZE);
1231         if (ps < 0) {
1232                 log_err("Failed to get page size\n");
1233                 return 1;
1234         }
1235
1236         page_size = ps;
1237         page_mask = ps - 1;
1238
1239         if (write_bw_log) {
1240                 setup_log(&agg_io_log[DDIR_READ]);
1241                 setup_log(&agg_io_log[DDIR_WRITE]);
1242         }
1243
1244         startup_sem = fio_sem_init(0);
1245
1246         set_genesis_time();
1247
1248         disk_util_timer_arm();
1249
1250         run_threads();
1251
1252         if (!fio_abort) {
1253                 show_run_stats();
1254                 if (write_bw_log) {
1255                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1256                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1257                 }
1258         }
1259
1260         fio_sem_remove(startup_sem);
1261         return exit_value;
1262 }