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