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