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