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