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