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