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