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