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                         break;
 351                 else if (!ret)
 352                         continue;
 353
 354                 init_icd(&icd, verify_io_u, ret);
 355                 ios_completed(td, &icd);
 356
 357                 if (icd.error) {
 358                         td_verror(td, icd.error);
 359                         break;
 360                 }
 361         }
 362
 363         if (io_u)
 364                 put_io_u(td, io_u);
 365
 366         if (td->cur_depth)
 367                 cleanup_pending_aio(td);
 368
 369         td_set_runstate(td, TD_RUNNING);
 370 }
 371
 372 /*
 373  * Not really an io thread, all it does is burn CPU cycles in the specified
 374  * manner.
 375  */
 376 static void do_cpuio(struct thread_data *td)
 377 {
 378         struct timeval e;
 379         int split = 100 / td->cpuload;
 380         int i = 0;
 381
 382         while (!td->terminate) {
 383                 fio_gettime(&e, NULL);
 384
 385                 if (runtime_exceeded(td, &e))
 386                         break;
 387
 388                 if (!(i % split))
 389                         __usec_sleep(10000);
 390                 else
 391                         usec_sleep(td, 10000);
 392
 393                 i++;
 394         }
 395 }
 396
 397 /*
 398  * Main IO worker function. It retrieves io_u's to process and queues
 399  * and reaps them, checking for rate and errors along the way.
 400  */
 401 static void do_io(struct thread_data *td)
 402 {
 403         struct io_completion_data icd;
 404         struct timeval s;
 405         unsigned long usec;
 406         struct fio_file *f;
 407         int i, ret = 0;
 408
 409         td_set_runstate(td, TD_RUNNING);
 410
 411         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 412                 struct timespec *timeout;
 413                 int min_evts = 0;
 414                 struct io_u *io_u;
 415
 416                 if (td->terminate)
 417                         break;
 418
 419                 f = get_next_file(td);
 420                 if (!f)
 421                         break;
 422
 423                 io_u = get_io_u(td, f);
 424                 if (!io_u)
 425                         break;
 426
 427                 memcpy(&s, &io_u->start_time, sizeof(s));
 428 requeue:
 429                 ret = td_io_queue(td, io_u);
 430
 431                 switch (ret) {
 432                 case FIO_Q_COMPLETED:
 433                         if (io_u->error) {
 434                                 ret = io_u->error;
 435                                 break;
 436                         }
 437                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 438                                 int bytes = io_u->xfer_buflen - io_u->resid;
 439
 440                                 io_u->xfer_buflen = io_u->resid;
 441                                 io_u->xfer_buf += bytes;
 442                                 goto requeue;
 443                         }
 444                         init_icd(&icd, NULL, 1);
 445                         io_completed(td, io_u, &icd);
 446                         put_io_u(td, io_u);
 447                         break;
 448                 case FIO_Q_QUEUED:
 449                         break;
 450                 default:
 451                         assert(ret < 0);
 452                         put_io_u(td, io_u);
 453                         break;
 454                 }
 455
 456                 if (ret < 0)
 457                         break;
 458
 459                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 460
 461                 if (ret == FIO_Q_QUEUED) {
 462                         if (td->cur_depth < td->iodepth) {
 463                                 struct timespec ts;
 464
 465                                 ts.tv_sec = 0;
 466                                 ts.tv_nsec = 0;
 467                                 timeout = &ts;
 468                                 min_evts = 0;
 469                         } else {
 470                                 timeout = NULL;
 471                                 min_evts = 1;
 472                         }
 473
 474                         ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
 475                         if (ret < 0) {
 476                                 td_verror(td, ret);
 477                                 break;
 478                         } else if (!ret)
 479                                 continue;
 480
 481                         init_icd(&icd, NULL, ret);
 482                         ios_completed(td, &icd);
 483                         if (icd.error) {
 484                                 td_verror(td, icd.error);
 485                                 break;
 486                         }
 487                 }
 488
 489                 /*
 490                  * the rate is batched for now, it should work for batches
 491                  * of completions except the very first one which may look
 492                  * a little bursty
 493                  */
 494                 usec = utime_since(&s, &icd.time);
 495
 496                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 497
 498                 if (check_min_rate(td, &icd.time)) {
 499                         if (exitall_on_terminate)
 500                                 terminate_threads(td->groupid, 0);
 501                         td_verror(td, ENODATA);
 502                         break;
 503                 }
 504
 505                 if (runtime_exceeded(td, &icd.time))
 506                         break;
 507
 508                 if (td->thinktime) {
 509                         unsigned long long b;
 510
 511                         b = td->io_blocks[0] + td->io_blocks[1];
 512                         if (!(b % td->thinktime_blocks)) {
 513                                 int left;
 514
 515                                 if (td->thinktime_spin)
 516                                         __usec_sleep(td->thinktime_spin);
 517
 518                                 left = td->thinktime - td->thinktime_spin;
 519                                 if (left)
 520                                         usec_sleep(td, left);
 521                         }
 522                 }
 523         }
 524
 525         if (!td->error) {
 526                 if (td->cur_depth)
 527                         cleanup_pending_aio(td);
 528
 529                 if (should_fsync(td) && td->end_fsync) {
 530                         td_set_runstate(td, TD_FSYNCING);
 531                         for_each_file(td, f, i)
 532                                 fio_io_sync(td, f);
 533                 }
 534         }
 535 }
 536
 537 static void cleanup_io_u(struct thread_data *td)
 538 {
 539         struct list_head *entry, *n;
 540         struct io_u *io_u;
 541
 542         list_for_each_safe(entry, n, &td->io_u_freelist) {
 543                 io_u = list_entry(entry, struct io_u, list);
 544
 545                 list_del(&io_u->list);
 546                 free(io_u);
 547         }
 548
 549         free_io_mem(td);
 550 }
 551
 552 /*
 553  * "randomly" fill the buffer contents
 554  */
 555 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 556 {
 557         int *ptr = io_u->buf;
 558
 559         while ((void *) ptr - io_u->buf < max_bs) {
 560                 *ptr = rand() * 0x9e370001;
 561                 ptr++;
 562         }
 563 }
 564
 565 static int init_io_u(struct thread_data *td)
 566 {
 567         struct io_u *io_u;
 568         unsigned int max_bs;
 569         int i, max_units;
 570         char *p;
 571
 572         if (td->io_ops->flags & FIO_CPUIO)
 573                 return 0;
 574
 575         if (td->io_ops->flags & FIO_SYNCIO)
 576                 max_units = 1;
 577         else
 578                 max_units = td->iodepth;
 579
 580         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 581         td->orig_buffer_size = max_bs * max_units;
 582
 583         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 584                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 585         else
 586                 td->orig_buffer_size += page_mask;
 587
 588         if (allocate_io_mem(td))
 589                 return 1;
 590
 591         p = ALIGN(td->orig_buffer);
 592         for (i = 0; i < max_units; i++) {
 593                 io_u = malloc(sizeof(*io_u));
 594                 memset(io_u, 0, sizeof(*io_u));
 595                 INIT_LIST_HEAD(&io_u->list);
 596
 597                 io_u->buf = p + max_bs * i;
 598                 if (td_write(td) || td_rw(td))
 599                         fill_rand_buf(io_u, max_bs);
 600
 601                 io_u->index = i;
 602                 list_add(&io_u->list, &td->io_u_freelist);
 603         }
 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);
 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);
 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);
 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);
 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->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 665         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 666         td->zone_bytes = 0;
 667
 668         for_each_file(td, f, i) {
 669                 f->last_pos = 0;
 670                 if (td->io_ops->flags & FIO_SYNCIO)
 671                         lseek(f->fd, SEEK_SET, 0);
 672
 673                 if (f->file_map)
 674                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 675         }
 676 }
 677
 678 /*
 679  * Entry point for the thread based jobs. The process based jobs end up
 680  * here as well, after a little setup.
 681  */
 682 static void *thread_main(void *data)
 683 {
 684         unsigned long long runtime[2];
 685         struct thread_data *td = data;
 686
 687         if (!td->use_thread)
 688                 setsid();
 689
 690         td->pid = getpid();
 691
 692         INIT_LIST_HEAD(&td->io_u_freelist);
 693         INIT_LIST_HEAD(&td->io_u_busylist);
 694         INIT_LIST_HEAD(&td->io_hist_list);
 695         INIT_LIST_HEAD(&td->io_log_list);
 696
 697         if (init_io_u(td))
 698                 goto err;
 699
 700         if (fio_setaffinity(td) == -1) {
 701                 td_verror(td, errno);
 702                 goto err;
 703         }
 704
 705         if (init_iolog(td))
 706                 goto err;
 707
 708         if (td->ioprio) {
 709                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 710                         td_verror(td, errno);
 711                         goto err;
 712                 }
 713         }
 714
 715         if (nice(td->nice) == -1) {
 716                 td_verror(td, errno);
 717                 goto err;
 718         }
 719
 720         if (init_random_state(td))
 721                 goto err;
 722
 723         if (td->ioscheduler && switch_ioscheduler(td))
 724                 goto err;
 725
 726         td_set_runstate(td, TD_INITIALIZED);
 727         fio_sem_up(&startup_sem);
 728         fio_sem_down(&td->mutex);
 729
 730         if (!td->create_serialize && setup_files(td))
 731                 goto err;
 732         if (open_files(td))
 733                 goto err;
 734
 735         /*
 736          * Do this late, as some IO engines would like to have the
 737          * files setup prior to initializing structures.
 738          */
 739         if (td_io_init(td))
 740                 goto err;
 741
 742         if (td->exec_prerun) {
 743                 if (system(td->exec_prerun) < 0)
 744                         goto err;
 745         }
 746
 747         fio_gettime(&td->epoch, NULL);
 748         getrusage(RUSAGE_SELF, &td->ru_start);
 749
 750         runtime[0] = runtime[1] = 0;
 751         while (td->loops--) {
 752                 fio_gettime(&td->start, NULL);
 753                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 754
 755                 if (td->ratemin)
 756                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 757
 758                 clear_io_state(td);
 759                 prune_io_piece_log(td);
 760
 761                 if (td->io_ops->flags & FIO_CPUIO)
 762                         do_cpuio(td);
 763                 else
 764                         do_io(td);
 765
 766                 runtime[td->ddir] += utime_since_now(&td->start);
 767                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 768                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 769
 770                 if (td->error || td->terminate)
 771                         break;
 772
 773                 if (td->verify == VERIFY_NONE)
 774                         continue;
 775
 776                 clear_io_state(td);
 777                 fio_gettime(&td->start, NULL);
 778
 779                 do_verify(td);
 780
 781                 runtime[DDIR_READ] += utime_since_now(&td->start);
 782
 783                 if (td->error || td->terminate)
 784                         break;
 785         }
 786
 787         update_rusage_stat(td);
 788         fio_gettime(&td->end_time, NULL);
 789         td->runtime[0] = runtime[0] / 1000;
 790         td->runtime[1] = runtime[1] / 1000;
 791
 792         if (td->bw_log)
 793                 finish_log(td, td->bw_log, "bw");
 794         if (td->slat_log)
 795                 finish_log(td, td->slat_log, "slat");
 796         if (td->clat_log)
 797                 finish_log(td, td->clat_log, "clat");
 798         if (td->write_iolog_file)
 799                 write_iolog_close(td);
 800         if (td->exec_postrun) {
 801                 if (system(td->exec_postrun) < 0)
 802                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 803         }
 804
 805         if (exitall_on_terminate)
 806                 terminate_threads(td->groupid, 0);
 807
 808 err:
 809         if (td->error)
 810                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 811         close_files(td);
 812         close_ioengine(td);
 813         cleanup_io_u(td);
 814         td_set_runstate(td, TD_EXITED);
 815         return (void *) td->error;
 816 }
 817
 818 /*
 819  * We cannot pass the td data into a forked process, so attach the td and
 820  * pass it to the thread worker.
 821  */
 822 static int fork_main(int shmid, int offset)
 823 {
 824         struct thread_data *td;
 825         void *data, *ret;
 826
 827         data = shmat(shmid, NULL, 0);
 828         if (data == (void *) -1) {
 829                 int __err = errno;
 830
 831                 perror("shmat");
 832                 return __err;
 833         }
 834
 835         td = data + offset * sizeof(struct thread_data);
 836         ret = thread_main(td);
 837         shmdt(data);
 838         return (int) ret;
 839 }
 840
 841 /*
 842  * Run over the job map and reap the threads that have exited, if any.
 843  */
 844 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 845 {
 846         struct thread_data *td;
 847         int i, cputhreads, pending, status, ret;
 848
 849         /*
 850          * reap exited threads (TD_EXITED -> TD_REAPED)
 851          */
 852         pending = cputhreads = 0;
 853         for_each_td(td, i) {
 854                 /*
 855                  * ->io_ops is NULL for a thread that has closed its
 856                  * io engine
 857                  */
 858                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 859                         cputhreads++;
 860
 861                 if (td->runstate < TD_EXITED) {
 862                         /*
 863                          * check if someone quit or got killed in an unusual way
 864                          */
 865                         ret = waitpid(td->pid, &status, WNOHANG);
 866                         if (ret < 0)
 867                                 perror("waitpid");
 868                         else if ((ret == td->pid) && WIFSIGNALED(status)) {
 869                                 int sig = WTERMSIG(status);
 870
 871                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 872                                 td_set_runstate(td, TD_REAPED);
 873                                 goto reaped;
 874                         }
 875                 }
 876
 877                 if (td->runstate != TD_EXITED) {
 878                         if (td->runstate < TD_RUNNING)
 879                                 pending++;
 880
 881                         continue;
 882                 }
 883
 884                 if (td->error)
 885                         exit_value++;
 886
 887                 td_set_runstate(td, TD_REAPED);
 888
 889                 if (td->use_thread) {
 890                         long ret;
 891
 892                         if (pthread_join(td->thread, (void *) &ret))
 893                                 perror("thread_join");
 894                 } else {
 895                         int status;
 896
 897                         ret = waitpid(td->pid, &status, 0);
 898                         if (ret < 0)
 899                                 perror("waitpid");
 900                         else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 901                                 if (!exit_value)
 902                                         exit_value++;
 903                         }
 904                 }
 905
 906 reaped:
 907                 (*nr_running)--;
 908                 (*m_rate) -= td->ratemin;
 909                 (*t_rate) -= td->rate;
 910         }
 911
 912         if (*nr_running == cputhreads && !pending)
 913                 terminate_threads(TERMINATE_ALL, 0);
 914 }
 915
 916 /*
 917  * Main function for kicking off and reaping jobs, as needed.
 918  */
 919 static void run_threads(void)
 920 {
 921         struct thread_data *td;
 922         unsigned long spent;
 923         int i, todo, nr_running, m_rate, t_rate, nr_started;
 924
 925         if (fio_pin_memory())
 926                 return;
 927
 928         if (!terse_output) {
 929                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 930                 fflush(stdout);
 931         }
 932
 933         signal(SIGINT, sig_handler);
 934         signal(SIGALRM, sig_handler);
 935
 936         todo = thread_number;
 937         nr_running = 0;
 938         nr_started = 0;
 939         m_rate = t_rate = 0;
 940
 941         for_each_td(td, i) {
 942                 print_status_init(td->thread_number - 1);
 943
 944                 if (!td->create_serialize) {
 945                         init_disk_util(td);
 946                         continue;
 947                 }
 948
 949                 /*
 950                  * do file setup here so it happens sequentially,
 951                  * we don't want X number of threads getting their
 952                  * client data interspersed on disk
 953                  */
 954                 if (setup_files(td)) {
 955                         exit_value++;
 956                         if (td->error)
 957                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 958                         td_set_runstate(td, TD_REAPED);
 959                         todo--;
 960                 }
 961
 962                 init_disk_util(td);
 963         }
 964
 965         while (todo) {
 966                 struct thread_data *map[MAX_JOBS];
 967                 struct timeval this_start;
 968                 int this_jobs = 0, left;
 969
 970                 /*
 971                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 972                  */
 973                 for_each_td(td, i) {
 974                         if (td->runstate != TD_NOT_CREATED)
 975                                 continue;
 976
 977                         /*
 978                          * never got a chance to start, killed by other
 979                          * thread for some reason
 980                          */
 981                         if (td->terminate) {
 982                                 todo--;
 983                                 continue;
 984                         }
 985
 986                         if (td->start_delay) {
 987                                 spent = mtime_since_genesis();
 988
 989                                 if (td->start_delay * 1000 > spent)
 990                                         continue;
 991                         }
 992
 993                         if (td->stonewall && (nr_started || nr_running))
 994                                 break;
 995
 996                         /*
 997                          * Set state to created. Thread will transition
 998                          * to TD_INITIALIZED when it's done setting up.
 999                          */
1000                         td_set_runstate(td, TD_CREATED);
1001                         map[this_jobs++] = td;
1002                         fio_sem_init(&startup_sem, 1);
1003                         nr_started++;
1004
1005                         if (td->use_thread) {
1006                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1007                                         perror("thread_create");
1008                                         nr_started--;
1009                                 }
1010                         } else {
1011                                 if (fork())
1012                                         fio_sem_down(&startup_sem);
1013                                 else {
1014                                         int ret = fork_main(shm_id, i);
1015
1016                                         exit(ret);
1017                                 }
1018                         }
1019                 }
1020
1021                 /*
1022                  * Wait for the started threads to transition to
1023                  * TD_INITIALIZED.
1024                  */
1025                 fio_gettime(&this_start, NULL);
1026                 left = this_jobs;
1027                 while (left && !fio_abort) {
1028                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1029                                 break;
1030
1031                         usleep(100000);
1032
1033                         for (i = 0; i < this_jobs; i++) {
1034                                 td = map[i];
1035                                 if (!td)
1036                                         continue;
1037                                 if (td->runstate == TD_INITIALIZED) {
1038                                         map[i] = NULL;
1039                                         left--;
1040                                 } else if (td->runstate >= TD_EXITED) {
1041                                         map[i] = NULL;
1042                                         left--;
1043                                         todo--;
1044                                         nr_running++; /* work-around... */
1045                                 }
1046                         }
1047                 }
1048
1049                 if (left) {
1050                         log_err("fio: %d jobs failed to start\n", left);
1051                         for (i = 0; i < this_jobs; i++) {
1052                                 td = map[i];
1053                                 if (!td)
1054                                         continue;
1055                                 kill(td->pid, SIGTERM);
1056                         }
1057                         break;
1058                 }
1059
1060                 /*
1061                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1062                  */
1063                 for_each_td(td, i) {
1064                         if (td->runstate != TD_INITIALIZED)
1065                                 continue;
1066
1067                         td_set_runstate(td, TD_RUNNING);
1068                         nr_running++;
1069                         nr_started--;
1070                         m_rate += td->ratemin;
1071                         t_rate += td->rate;
1072                         todo--;
1073                         fio_sem_up(&td->mutex);
1074                 }
1075
1076                 reap_threads(&nr_running, &t_rate, &m_rate);
1077
1078                 if (todo)
1079                         usleep(100000);
1080         }
1081
1082         while (nr_running) {
1083                 reap_threads(&nr_running, &t_rate, &m_rate);
1084                 usleep(10000);
1085         }
1086
1087         update_io_ticks();
1088         fio_unpin_memory();
1089 }
1090
1091 int main(int argc, char *argv[])
1092 {
1093         long ps;
1094
1095         /*
1096          * We need locale for number printing, if it isn't set then just
1097          * go with the US format.
1098          */
1099         if (!getenv("LC_NUMERIC"))
1100                 setlocale(LC_NUMERIC, "en_US");
1101
1102         if (parse_options(argc, argv))
1103                 return 1;
1104
1105         if (!thread_number) {
1106                 log_err("Nothing to do\n");
1107                 return 1;
1108         }
1109
1110         ps = sysconf(_SC_PAGESIZE);
1111         if (ps < 0) {
1112                 log_err("Failed to get page size\n");
1113                 return 1;
1114         }
1115
1116         page_mask = ps - 1;
1117
1118         if (write_bw_log) {
1119                 setup_log(&agg_io_log[DDIR_READ]);
1120                 setup_log(&agg_io_log[DDIR_WRITE]);
1121         }
1122
1123         disk_util_timer_arm();
1124
1125         run_threads();
1126
1127         if (!fio_abort) {
1128                 show_run_stats();
1129                 if (write_bw_log) {
1130                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1131                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1132                 }
1133         }
1134
1135         return exit_value;
1136 }