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 /*
 140  * When job exits, we can cancel the in-flight IO if we are using async
 141  * io. Attempt to do so.
 142  */
 143 static void cleanup_pending_aio(struct thread_data *td)
 144 {
 145         struct list_head *entry, *n;
 146         struct io_u *io_u;
 147         int r;
 148
 149         /*
 150          * get immediately available events, if any
 151          */
 152         io_u_queued_complete(td, 0, NULL);
 153
 154         /*
 155          * now cancel remaining active events
 156          */
 157         if (td->io_ops->cancel) {
 158                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 159                         io_u = list_entry(entry, struct io_u, list);
 160
 161                         r = td->io_ops->cancel(td, io_u);
 162                         if (!r)
 163                                 put_io_u(td, io_u);
 164                 }
 165         }
 166
 167         if (td->cur_depth)
 168                 io_u_queued_complete(td, td->cur_depth, NULL);
 169 }
 170
 171 /*
 172  * Helper to handle the final sync of a file. Works just like the normal
 173  * io path, just does everything sync.
 174  */
 175 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 176 {
 177         struct io_u *io_u = __get_io_u(td);
 178         int ret;
 179
 180         if (!io_u)
 181                 return 1;
 182
 183         io_u->ddir = DDIR_SYNC;
 184         io_u->file = f;
 185
 186         if (td_io_prep(td, io_u)) {
 187                 put_io_u(td, io_u);
 188                 return 1;
 189         }
 190
 191 requeue:
 192         ret = td_io_queue(td, io_u);
 193         if (ret < 0) {
 194                 td_verror(td, io_u->error);
 195                 put_io_u(td, io_u);
 196                 return 1;
 197         } else if (ret == FIO_Q_QUEUED) {
 198                 if (io_u_queued_complete(td, 1, NULL))
 199                         return 1;
 200         } else if (ret == FIO_Q_COMPLETED) {
 201                 if (io_u->error) {
 202                         td_verror(td, io_u->error);
 203                         return 1;
 204                 }
 205
 206                 io_u_sync_complete(td, io_u, NULL);
 207         } else if (ret == FIO_Q_BUSY) {
 208                 if (td_io_commit(td))
 209                         return 1;
 210                 goto requeue;
 211         }
 212
 213         return 0;
 214 }
 215
 216 /*
 217  * The main verify engine. Runs over the writes we previusly submitted,
 218  * reads the blocks back in, and checks the crc/md5 of the data.
 219  */
 220 static void do_verify(struct thread_data *td)
 221 {
 222         struct fio_file *f;
 223         struct io_u *io_u;
 224         int ret, i, min_events;
 225
 226         /*
 227          * sync io first and invalidate cache, to make sure we really
 228          * read from disk.
 229          */
 230         for_each_file(td, f, i) {
 231                 fio_io_sync(td, f);
 232                 file_invalidate_cache(td, f);
 233         }
 234
 235         td_set_runstate(td, TD_VERIFYING);
 236
 237         io_u = NULL;
 238         while (!td->terminate) {
 239                 io_u = __get_io_u(td);
 240                 if (!io_u)
 241                         break;
 242
 243                 if (runtime_exceeded(td, &io_u->start_time))
 244                         break;
 245
 246                 if (get_next_verify(td, io_u))
 247                         break;
 248
 249                 if (td_io_prep(td, io_u))
 250                         break;
 251
 252 requeue:
 253                 ret = td_io_queue(td, io_u);
 254
 255                 switch (ret) {
 256                 case FIO_Q_COMPLETED:
 257                         if (io_u->error)
 258                                 ret = -io_u->error;
 259                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 260                                 int bytes = io_u->xfer_buflen - io_u->resid;
 261
 262                                 io_u->xfer_buflen = io_u->resid;
 263                                 io_u->xfer_buf += bytes;
 264                                 goto requeue;
 265                         }
 266                         ret = io_u_sync_complete(td, io_u, verify_io_u);
 267                         if (ret)
 268                                 break;
 269                         continue;
 270                 case FIO_Q_QUEUED:
 271                         break;
 272                 case FIO_Q_BUSY:
 273                         requeue_io_u(td, &io_u);
 274                         ret = td_io_commit(td);
 275                         break;
 276                 default:
 277                         assert(ret < 0);
 278                         td_verror(td, -ret);
 279                         break;
 280                 }
 281
 282                 if (ret < 0 || td->error)
 283                         break;
 284
 285                 /*
 286                  * if we can queue more, do so. but check if there are
 287                  * completed io_u's first.
 288                  */
 289                 min_events = 0;
 290                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 291                         min_events = 1;
 292
 293                         if (td->cur_depth > td->iodepth_low)
 294                                 min_events = td->cur_depth - td->iodepth_low;
 295                 }
 296
 297                 /*
 298                  * Reap required number of io units, if any, and do the
 299                  * verification on them through the callback handler
 300                  */
 301                 if (io_u_queued_complete(td, min_events, verify_io_u))
 302                         break;
 303         }
 304
 305         if (io_u)
 306                 put_io_u(td, io_u);
 307
 308         if (td->cur_depth)
 309                 cleanup_pending_aio(td);
 310
 311         td_set_runstate(td, TD_RUNNING);
 312 }
 313
 314 /*
 315  * Not really an io thread, all it does is burn CPU cycles in the specified
 316  * manner.
 317  */
 318 static void do_cpuio(struct thread_data *td)
 319 {
 320         struct timeval e;
 321         int split = 100 / td->cpuload;
 322         int i = 0;
 323
 324         while (!td->terminate) {
 325                 fio_gettime(&e, NULL);
 326
 327                 if (runtime_exceeded(td, &e))
 328                         break;
 329
 330                 if (!(i % split))
 331                         __usec_sleep(10000);
 332                 else
 333                         usec_sleep(td, 10000);
 334
 335                 i++;
 336         }
 337 }
 338
 339 /*
 340  * Main IO worker function. It retrieves io_u's to process and queues
 341  * and reaps them, checking for rate and errors along the way.
 342  */
 343 static void do_io(struct thread_data *td)
 344 {
 345         struct timeval s;
 346         unsigned long usec;
 347         int i, ret = 0;
 348
 349         td_set_runstate(td, TD_RUNNING);
 350
 351         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 352                 struct timeval comp_time;
 353                 long bytes_done = 0;
 354                 int min_evts = 0;
 355                 struct io_u *io_u;
 356
 357                 if (td->terminate)
 358                         break;
 359
 360                 io_u = get_io_u(td);
 361                 if (!io_u)
 362                         break;
 363
 364                 memcpy(&s, &io_u->start_time, sizeof(s));
 365
 366                 if (runtime_exceeded(td, &s)) {
 367                         put_io_u(td, io_u);
 368                         break;
 369                 }
 370 requeue:
 371                 ret = td_io_queue(td, io_u);
 372
 373                 switch (ret) {
 374                 case FIO_Q_COMPLETED:
 375                         if (io_u->error) {
 376                                 ret = io_u->error;
 377                                 break;
 378                         }
 379                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 380                                 int bytes = io_u->xfer_buflen - io_u->resid;
 381
 382                                 io_u->xfer_buflen = io_u->resid;
 383                                 io_u->xfer_buf += bytes;
 384                                 goto requeue;
 385                         }
 386                         fio_gettime(&comp_time, NULL);
 387                         bytes_done = io_u_sync_complete(td, io_u, NULL);
 388                         if (bytes_done < 0)
 389                                 ret = bytes_done;
 390                         break;
 391                 case FIO_Q_QUEUED:
 392                         /*
 393                          * if the engine doesn't have a commit hook,
 394                          * the io_u is really queued. if it does have such
 395                          * a hook, it has to call io_u_queued() itself.
 396                          */
 397                         if (td->io_ops->commit == NULL)
 398                                 io_u_queued(td, io_u);
 399                         break;
 400                 case FIO_Q_BUSY:
 401                         requeue_io_u(td, &io_u);
 402                         ret = td_io_commit(td);
 403                         break;
 404                 default:
 405                         assert(ret < 0);
 406                         put_io_u(td, io_u);
 407                         break;
 408                 }
 409
 410                 if (ret < 0 || td->error)
 411                         break;
 412
 413                 /*
 414                  * See if we need to complete some commands
 415                  */
 416                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 417                         min_evts = 0;
 418                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 419                                 min_evts = 1;
 420
 421                                 if (td->cur_depth > td->iodepth_low)
 422                                         min_evts = td->cur_depth - td->iodepth_low;
 423                         }
 424
 425                         fio_gettime(&comp_time, NULL);
 426                         bytes_done = io_u_queued_complete(td, min_evts, NULL);
 427                         if (bytes_done < 0)
 428                                 break;
 429                 }
 430
 431                 if (!bytes_done)
 432                         continue;
 433
 434                 /*
 435                  * the rate is batched for now, it should work for batches
 436                  * of completions except the very first one which may look
 437                  * a little bursty
 438                  */
 439                 usec = utime_since(&s, &comp_time);
 440
 441                 rate_throttle(td, usec, bytes_done, td->ddir);
 442
 443                 if (check_min_rate(td, &comp_time)) {
 444                         if (exitall_on_terminate)
 445                                 terminate_threads(td->groupid, 0);
 446                         td_verror(td, ENODATA);
 447                         break;
 448                 }
 449
 450                 if (td->thinktime) {
 451                         unsigned long long b;
 452
 453                         b = td->io_blocks[0] + td->io_blocks[1];
 454                         if (!(b % td->thinktime_blocks)) {
 455                                 int left;
 456
 457                                 if (td->thinktime_spin)
 458                                         __usec_sleep(td->thinktime_spin);
 459
 460                                 left = td->thinktime - td->thinktime_spin;
 461                                 if (left)
 462                                         usec_sleep(td, left);
 463                         }
 464                 }
 465         }
 466
 467         if (!td->error) {
 468                 struct fio_file *f;
 469
 470                 if (td->cur_depth)
 471                         cleanup_pending_aio(td);
 472
 473                 if (should_fsync(td) && td->end_fsync) {
 474                         td_set_runstate(td, TD_FSYNCING);
 475                         for_each_file(td, f, i)
 476                                 fio_io_sync(td, f);
 477                 }
 478         }
 479 }
 480
 481 static void cleanup_io_u(struct thread_data *td)
 482 {
 483         struct list_head *entry, *n;
 484         struct io_u *io_u;
 485
 486         list_for_each_safe(entry, n, &td->io_u_freelist) {
 487                 io_u = list_entry(entry, struct io_u, list);
 488
 489                 list_del(&io_u->list);
 490                 free(io_u);
 491         }
 492
 493         free_io_mem(td);
 494 }
 495
 496 /*
 497  * "randomly" fill the buffer contents
 498  */
 499 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 500 {
 501         int *ptr = io_u->buf;
 502
 503         while ((void *) ptr - io_u->buf < max_bs) {
 504                 *ptr = rand() * 0x9e370001;
 505                 ptr++;
 506         }
 507 }
 508
 509 static int init_io_u(struct thread_data *td)
 510 {
 511         struct io_u *io_u;
 512         unsigned int max_bs;
 513         int i, max_units;
 514         char *p;
 515
 516         if (td->io_ops->flags & FIO_CPUIO)
 517                 return 0;
 518
 519         if (td->io_ops->flags & FIO_SYNCIO)
 520                 max_units = 1;
 521         else
 522                 max_units = td->iodepth;
 523
 524         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 525         td->orig_buffer_size = max_bs * max_units;
 526
 527         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 528                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 529         else
 530                 td->orig_buffer_size += page_mask;
 531
 532         if (allocate_io_mem(td))
 533                 return 1;
 534
 535         p = ALIGN(td->orig_buffer);
 536         for (i = 0; i < max_units; i++) {
 537                 io_u = malloc(sizeof(*io_u));
 538                 memset(io_u, 0, sizeof(*io_u));
 539                 INIT_LIST_HEAD(&io_u->list);
 540
 541                 io_u->buf = p + max_bs * i;
 542                 if (td_write(td) || td_rw(td))
 543                         fill_rand_buf(io_u, max_bs);
 544
 545                 io_u->index = i;
 546                 list_add(&io_u->list, &td->io_u_freelist);
 547         }
 548
 549         return 0;
 550 }
 551
 552 static int switch_ioscheduler(struct thread_data *td)
 553 {
 554         char tmp[256], tmp2[128];
 555         FILE *f;
 556         int ret;
 557
 558         if (td->io_ops->flags & FIO_CPUIO)
 559                 return 0;
 560
 561         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 562
 563         f = fopen(tmp, "r+");
 564         if (!f) {
 565                 td_verror(td, errno);
 566                 return 1;
 567         }
 568
 569         /*
 570          * Set io scheduler.
 571          */
 572         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 573         if (ferror(f) || ret != 1) {
 574                 td_verror(td, errno);
 575                 fclose(f);
 576                 return 1;
 577         }
 578
 579         rewind(f);
 580
 581         /*
 582          * Read back and check that the selected scheduler is now the default.
 583          */
 584         ret = fread(tmp, 1, sizeof(tmp), f);
 585         if (ferror(f) || ret < 0) {
 586                 td_verror(td, errno);
 587                 fclose(f);
 588                 return 1;
 589         }
 590
 591         sprintf(tmp2, "[%s]", td->ioscheduler);
 592         if (!strstr(tmp, tmp2)) {
 593                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 594                 td_verror(td, EINVAL);
 595                 fclose(f);
 596                 return 1;
 597         }
 598
 599         fclose(f);
 600         return 0;
 601 }
 602
 603 static void clear_io_state(struct thread_data *td)
 604 {
 605         struct fio_file *f;
 606         int i;
 607
 608         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 609         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 610         td->zone_bytes = 0;
 611
 612         td->last_was_sync = 0;
 613
 614         for_each_file(td, f, i) {
 615                 f->last_completed_pos = 0;
 616
 617                 f->last_pos = 0;
 618                 if (td->io_ops->flags & FIO_SYNCIO)
 619                         lseek(f->fd, SEEK_SET, 0);
 620
 621                 if (f->file_map)
 622                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 623         }
 624 }
 625
 626 /*
 627  * Entry point for the thread based jobs. The process based jobs end up
 628  * here as well, after a little setup.
 629  */
 630 static void *thread_main(void *data)
 631 {
 632         unsigned long long runtime[2];
 633         struct thread_data *td = data;
 634
 635         if (!td->use_thread)
 636                 setsid();
 637
 638         td->pid = getpid();
 639
 640         INIT_LIST_HEAD(&td->io_u_freelist);
 641         INIT_LIST_HEAD(&td->io_u_busylist);
 642         INIT_LIST_HEAD(&td->io_u_requeues);
 643         INIT_LIST_HEAD(&td->io_hist_list);
 644         INIT_LIST_HEAD(&td->io_log_list);
 645
 646         if (init_io_u(td))
 647                 goto err;
 648
 649         if (fio_setaffinity(td) == -1) {
 650                 td_verror(td, errno);
 651                 goto err;
 652         }
 653
 654         if (init_iolog(td))
 655                 goto err;
 656
 657         if (td->ioprio) {
 658                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 659                         td_verror(td, errno);
 660                         goto err;
 661                 }
 662         }
 663
 664         if (nice(td->nice) == -1) {
 665                 td_verror(td, errno);
 666                 goto err;
 667         }
 668
 669         if (init_random_state(td))
 670                 goto err;
 671
 672         if (td->ioscheduler && switch_ioscheduler(td))
 673                 goto err;
 674
 675         td_set_runstate(td, TD_INITIALIZED);
 676         fio_sem_up(&startup_sem);
 677         fio_sem_down(&td->mutex);
 678
 679         if (!td->create_serialize && setup_files(td))
 680                 goto err;
 681         if (open_files(td))
 682                 goto err;
 683
 684         /*
 685          * Do this late, as some IO engines would like to have the
 686          * files setup prior to initializing structures.
 687          */
 688         if (td_io_init(td))
 689                 goto err;
 690
 691         if (td->exec_prerun) {
 692                 if (system(td->exec_prerun) < 0)
 693                         goto err;
 694         }
 695
 696         fio_gettime(&td->epoch, NULL);
 697         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 698
 699         runtime[0] = runtime[1] = 0;
 700         while (td->loops--) {
 701                 fio_gettime(&td->start, NULL);
 702                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 703
 704                 if (td->ratemin)
 705                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 706
 707                 clear_io_state(td);
 708                 prune_io_piece_log(td);
 709
 710                 if (td->io_ops->flags & FIO_CPUIO)
 711                         do_cpuio(td);
 712                 else
 713                         do_io(td);
 714
 715                 runtime[td->ddir] += utime_since_now(&td->start);
 716                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 717                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 718
 719                 if (td->error || td->terminate)
 720                         break;
 721
 722                 if (td->verify == VERIFY_NONE)
 723                         continue;
 724
 725                 clear_io_state(td);
 726                 fio_gettime(&td->start, NULL);
 727
 728                 do_verify(td);
 729
 730                 runtime[DDIR_READ] += utime_since_now(&td->start);
 731
 732                 if (td->error || td->terminate)
 733                         break;
 734         }
 735
 736         update_rusage_stat(td);
 737         fio_gettime(&td->end_time, NULL);
 738         td->runtime[0] = runtime[0] / 1000;
 739         td->runtime[1] = runtime[1] / 1000;
 740
 741         if (td->ts.bw_log)
 742                 finish_log(td, td->ts.bw_log, "bw");
 743         if (td->ts.slat_log)
 744                 finish_log(td, td->ts.slat_log, "slat");
 745         if (td->ts.clat_log)
 746                 finish_log(td, td->ts.clat_log, "clat");
 747         if (td->write_iolog_file)
 748                 write_iolog_close(td);
 749         if (td->exec_postrun) {
 750                 if (system(td->exec_postrun) < 0)
 751                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 752         }
 753
 754         if (exitall_on_terminate)
 755                 terminate_threads(td->groupid, 0);
 756
 757 err:
 758         if (td->error)
 759                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 760         close_files(td);
 761         close_ioengine(td);
 762         cleanup_io_u(td);
 763         td_set_runstate(td, TD_EXITED);
 764         return (void *) td->error;
 765 }
 766
 767 /*
 768  * We cannot pass the td data into a forked process, so attach the td and
 769  * pass it to the thread worker.
 770  */
 771 static int fork_main(int shmid, int offset)
 772 {
 773         struct thread_data *td;
 774         void *data, *ret;
 775
 776         data = shmat(shmid, NULL, 0);
 777         if (data == (void *) -1) {
 778                 int __err = errno;
 779
 780                 perror("shmat");
 781                 return __err;
 782         }
 783
 784         td = data + offset * sizeof(struct thread_data);
 785         ret = thread_main(td);
 786         shmdt(data);
 787         return (int) ret;
 788 }
 789
 790 /*
 791  * Run over the job map and reap the threads that have exited, if any.
 792  */
 793 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 794 {
 795         struct thread_data *td;
 796         int i, cputhreads, pending, status, ret;
 797
 798         /*
 799          * reap exited threads (TD_EXITED -> TD_REAPED)
 800          */
 801         pending = cputhreads = 0;
 802         for_each_td(td, i) {
 803                 /*
 804                  * ->io_ops is NULL for a thread that has closed its
 805                  * io engine
 806                  */
 807                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 808                         cputhreads++;
 809
 810                 if (td->runstate < TD_EXITED) {
 811                         /*
 812                          * check if someone quit or got killed in an unusual way
 813                          */
 814                         ret = waitpid(td->pid, &status, WNOHANG);
 815                         if (ret < 0)
 816                                 perror("waitpid");
 817                         else if ((ret == td->pid) && WIFSIGNALED(status)) {
 818                                 int sig = WTERMSIG(status);
 819
 820                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 821                                 td_set_runstate(td, TD_REAPED);
 822                                 goto reaped;
 823                         }
 824                 }
 825
 826                 if (td->runstate != TD_EXITED) {
 827                         if (td->runstate < TD_RUNNING)
 828                                 pending++;
 829
 830                         continue;
 831                 }
 832
 833                 if (td->error)
 834                         exit_value++;
 835
 836                 td_set_runstate(td, TD_REAPED);
 837
 838                 if (td->use_thread) {
 839                         long ret;
 840
 841                         if (pthread_join(td->thread, (void *) &ret))
 842                                 perror("thread_join");
 843                 } else {
 844                         int status;
 845
 846                         ret = waitpid(td->pid, &status, 0);
 847                         if (ret < 0)
 848                                 perror("waitpid");
 849                         else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 850                                 if (!exit_value)
 851                                         exit_value++;
 852                         }
 853                 }
 854
 855 reaped:
 856                 (*nr_running)--;
 857                 (*m_rate) -= td->ratemin;
 858                 (*t_rate) -= td->rate;
 859         }
 860
 861         if (*nr_running == cputhreads && !pending)
 862                 terminate_threads(TERMINATE_ALL, 0);
 863 }
 864
 865 /*
 866  * Main function for kicking off and reaping jobs, as needed.
 867  */
 868 static void run_threads(void)
 869 {
 870         struct thread_data *td;
 871         unsigned long spent;
 872         int i, todo, nr_running, m_rate, t_rate, nr_started;
 873
 874         if (fio_pin_memory())
 875                 return;
 876
 877         if (!terse_output) {
 878                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 879                 fflush(stdout);
 880         }
 881
 882         signal(SIGINT, sig_handler);
 883         signal(SIGALRM, sig_handler);
 884
 885         todo = thread_number;
 886         nr_running = 0;
 887         nr_started = 0;
 888         m_rate = t_rate = 0;
 889
 890         for_each_td(td, i) {
 891                 print_status_init(td->thread_number - 1);
 892
 893                 if (!td->create_serialize) {
 894                         init_disk_util(td);
 895                         continue;
 896                 }
 897
 898                 /*
 899                  * do file setup here so it happens sequentially,
 900                  * we don't want X number of threads getting their
 901                  * client data interspersed on disk
 902                  */
 903                 if (setup_files(td)) {
 904                         exit_value++;
 905                         if (td->error)
 906                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 907                         td_set_runstate(td, TD_REAPED);
 908                         todo--;
 909                 }
 910
 911                 init_disk_util(td);
 912         }
 913
 914         while (todo) {
 915                 struct thread_data *map[MAX_JOBS];
 916                 struct timeval this_start;
 917                 int this_jobs = 0, left;
 918
 919                 /*
 920                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 921                  */
 922                 for_each_td(td, i) {
 923                         if (td->runstate != TD_NOT_CREATED)
 924                                 continue;
 925
 926                         /*
 927                          * never got a chance to start, killed by other
 928                          * thread for some reason
 929                          */
 930                         if (td->terminate) {
 931                                 todo--;
 932                                 continue;
 933                         }
 934
 935                         if (td->start_delay) {
 936                                 spent = mtime_since_genesis();
 937
 938                                 if (td->start_delay * 1000 > spent)
 939                                         continue;
 940                         }
 941
 942                         if (td->stonewall && (nr_started || nr_running))
 943                                 break;
 944
 945                         /*
 946                          * Set state to created. Thread will transition
 947                          * to TD_INITIALIZED when it's done setting up.
 948                          */
 949                         td_set_runstate(td, TD_CREATED);
 950                         map[this_jobs++] = td;
 951                         fio_sem_init(&startup_sem, 1);
 952                         nr_started++;
 953
 954                         if (td->use_thread) {
 955                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 956                                         perror("thread_create");
 957                                         nr_started--;
 958                                 }
 959                         } else {
 960                                 if (fork())
 961                                         fio_sem_down(&startup_sem);
 962                                 else {
 963                                         int ret = fork_main(shm_id, i);
 964
 965                                         exit(ret);
 966                                 }
 967                         }
 968                 }
 969
 970                 /*
 971                  * Wait for the started threads to transition to
 972                  * TD_INITIALIZED.
 973                  */
 974                 fio_gettime(&this_start, NULL);
 975                 left = this_jobs;
 976                 while (left && !fio_abort) {
 977                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 978                                 break;
 979
 980                         usleep(100000);
 981
 982                         for (i = 0; i < this_jobs; i++) {
 983                                 td = map[i];
 984                                 if (!td)
 985                                         continue;
 986                                 if (td->runstate == TD_INITIALIZED) {
 987                                         map[i] = NULL;
 988                                         left--;
 989                                 } else if (td->runstate >= TD_EXITED) {
 990                                         map[i] = NULL;
 991                                         left--;
 992                                         todo--;
 993                                         nr_running++; /* work-around... */
 994                                 }
 995                         }
 996                 }
 997
 998                 if (left) {
 999                         log_err("fio: %d jobs failed to start\n", left);
1000                         for (i = 0; i < this_jobs; i++) {
1001                                 td = map[i];
1002                                 if (!td)
1003                                         continue;
1004                                 kill(td->pid, SIGTERM);
1005                         }
1006                         break;
1007                 }
1008
1009                 /*
1010                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1011                  */
1012                 for_each_td(td, i) {
1013                         if (td->runstate != TD_INITIALIZED)
1014                                 continue;
1015
1016                         td_set_runstate(td, TD_RUNNING);
1017                         nr_running++;
1018                         nr_started--;
1019                         m_rate += td->ratemin;
1020                         t_rate += td->rate;
1021                         todo--;
1022                         fio_sem_up(&td->mutex);
1023                 }
1024
1025                 reap_threads(&nr_running, &t_rate, &m_rate);
1026
1027                 if (todo)
1028                         usleep(100000);
1029         }
1030
1031         while (nr_running) {
1032                 reap_threads(&nr_running, &t_rate, &m_rate);
1033                 usleep(10000);
1034         }
1035
1036         update_io_ticks();
1037         fio_unpin_memory();
1038 }
1039
1040 int main(int argc, char *argv[])
1041 {
1042         long ps;
1043
1044         /*
1045          * We need locale for number printing, if it isn't set then just
1046          * go with the US format.
1047          */
1048         if (!getenv("LC_NUMERIC"))
1049                 setlocale(LC_NUMERIC, "en_US");
1050
1051         if (parse_options(argc, argv))
1052                 return 1;
1053
1054         if (!thread_number) {
1055                 log_err("Nothing to do\n");
1056                 return 1;
1057         }
1058
1059         ps = sysconf(_SC_PAGESIZE);
1060         if (ps < 0) {
1061                 log_err("Failed to get page size\n");
1062                 return 1;
1063         }
1064
1065         page_mask = ps - 1;
1066
1067         if (write_bw_log) {
1068                 setup_log(&agg_io_log[DDIR_READ]);
1069                 setup_log(&agg_io_log[DDIR_WRITE]);
1070         }
1071
1072         disk_util_timer_arm();
1073
1074         run_threads();
1075
1076         if (!fio_abort) {
1077                 show_run_stats();
1078                 if (write_bw_log) {
1079                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1080                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1081                 }
1082         }
1083
1084         return exit_value;
1085 }