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                 icd.nr = 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                         icd.nr = 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                 icd.nr = 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);
 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;
 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                 io_u = __get_io_u(td);
 280                 if (!io_u)
 281                         break;
 282
 283                 if (runtime_exceeded(td, &io_u->start_time))
 284                         break;
 285
 286                 if (get_next_verify(td, io_u))
 287                         break;
 288
 289                 if (td_io_prep(td, io_u))
 290                         break;
 291
 292 requeue:
 293                 ret = td_io_queue(td, io_u);
 294
 295                 switch (ret) {
 296                 case FIO_Q_COMPLETED:
 297                         if (io_u->error)
 298                                 ret = io_u->error;
 299                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 300                                 int bytes = io_u->xfer_buflen - io_u->resid;
 301
 302                                 io_u->xfer_buflen = io_u->resid;
 303                                 io_u->xfer_buf += bytes;
 304                                 goto requeue;
 305                         }
 306                         if (do_io_u_verify(td, &io_u)) {
 307                                 ret = -EIO;
 308                                 break;
 309                         }
 310                         continue;
 311                 case FIO_Q_QUEUED:
 312                         break;
 313                 default:
 314                         assert(ret < 0);
 315                         td_verror(td, ret);
 316                         break;
 317                 }
 318
 319                 /*
 320                  * We get here for a queued request, in the future we
 321                  * want to later make this take full advantage of
 322                  * keeping IO in flight while verifying others.
 323                  */
 324                 ret = td_io_getevents(td, 1, 1, NULL);
 325                 if (ret < 0)
 326                         break;
 327
 328                 assert(ret == 1);
 329                 io_u = td->io_ops->event(td, 0);
 330
 331                 if (do_io_u_verify(td, &io_u))
 332                         break;
 333         }
 334
 335         if (io_u)
 336                 put_io_u(td, io_u);
 337
 338         if (td->cur_depth)
 339                 cleanup_pending_aio(td);
 340
 341         td_set_runstate(td, TD_RUNNING);
 342 }
 343
 344 /*
 345  * Not really an io thread, all it does is burn CPU cycles in the specified
 346  * manner.
 347  */
 348 static void do_cpuio(struct thread_data *td)
 349 {
 350         struct timeval e;
 351         int split = 100 / td->cpuload;
 352         int i = 0;
 353
 354         while (!td->terminate) {
 355                 fio_gettime(&e, NULL);
 356
 357                 if (runtime_exceeded(td, &e))
 358                         break;
 359
 360                 if (!(i % split))
 361                         __usec_sleep(10000);
 362                 else
 363                         usec_sleep(td, 10000);
 364
 365                 i++;
 366         }
 367 }
 368
 369 /*
 370  * Main IO worker function. It retrieves io_u's to process and queues
 371  * and reaps them, checking for rate and errors along the way.
 372  */
 373 static void do_io(struct thread_data *td)
 374 {
 375         struct io_completion_data icd;
 376         struct timeval s;
 377         unsigned long usec;
 378         struct fio_file *f;
 379         int i, ret = 0;
 380
 381         td_set_runstate(td, TD_RUNNING);
 382
 383         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 384                 struct timespec *timeout;
 385                 int min_evts = 0;
 386                 struct io_u *io_u;
 387
 388                 if (td->terminate)
 389                         break;
 390
 391                 f = get_next_file(td);
 392                 if (!f)
 393                         break;
 394
 395                 io_u = get_io_u(td, f);
 396                 if (!io_u)
 397                         break;
 398
 399                 memcpy(&s, &io_u->start_time, sizeof(s));
 400 requeue:
 401                 ret = td_io_queue(td, io_u);
 402
 403                 switch (ret) {
 404                 case FIO_Q_COMPLETED:
 405                         if (io_u->error) {
 406                                 ret = io_u->error;
 407                                 break;
 408                         }
 409                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 410                                 int bytes = io_u->xfer_buflen - io_u->resid;
 411
 412                                 io_u->xfer_buflen = io_u->resid;
 413                                 io_u->xfer_buf += bytes;
 414                                 goto requeue;
 415                         }
 416                         init_icd(&icd);
 417                         io_completed(td, io_u, &icd);
 418                         put_io_u(td, io_u);
 419                         break;
 420                 case FIO_Q_QUEUED:
 421                         break;
 422                 default:
 423                         assert(ret < 0);
 424                         put_io_u(td, io_u);
 425                         break;
 426                 }
 427
 428                 if (ret < 0)
 429                         break;
 430
 431                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 432
 433                 if (ret == FIO_Q_QUEUED) {
 434                         if (td->cur_depth < td->iodepth) {
 435                                 struct timespec ts;
 436
 437                                 ts.tv_sec = 0;
 438                                 ts.tv_nsec = 0;
 439                                 timeout = &ts;
 440                                 min_evts = 0;
 441                         } else {
 442                                 timeout = NULL;
 443                                 min_evts = 1;
 444                         }
 445
 446                         ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
 447                         if (ret < 0) {
 448                                 td_verror(td, ret);
 449                                 break;
 450                         } else if (!ret)
 451                                 continue;
 452
 453                         icd.nr = ret;
 454                         ios_completed(td, &icd);
 455                         if (icd.error) {
 456                                 td_verror(td, icd.error);
 457                                 break;
 458                         }
 459                 }
 460
 461                 /*
 462                  * the rate is batched for now, it should work for batches
 463                  * of completions except the very first one which may look
 464                  * a little bursty
 465                  */
 466                 usec = utime_since(&s, &icd.time);
 467
 468                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 469
 470                 if (check_min_rate(td, &icd.time)) {
 471                         if (exitall_on_terminate)
 472                                 terminate_threads(td->groupid, 0);
 473                         td_verror(td, ENODATA);
 474                         break;
 475                 }
 476
 477                 if (runtime_exceeded(td, &icd.time))
 478                         break;
 479
 480                 if (td->thinktime) {
 481                         unsigned long long b;
 482
 483                         b = td->io_blocks[0] + td->io_blocks[1];
 484                         if (!(b % td->thinktime_blocks)) {
 485                                 int left;
 486
 487                                 if (td->thinktime_spin)
 488                                         __usec_sleep(td->thinktime_spin);
 489
 490                                 left = td->thinktime - td->thinktime_spin;
 491                                 if (left)
 492                                         usec_sleep(td, left);
 493                         }
 494                 }
 495         }
 496
 497         if (!td->error) {
 498                 if (td->cur_depth)
 499                         cleanup_pending_aio(td);
 500
 501                 if (should_fsync(td) && td->end_fsync) {
 502                         td_set_runstate(td, TD_FSYNCING);
 503                         for_each_file(td, f, i)
 504                                 fio_io_sync(td, f);
 505                 }
 506         }
 507 }
 508
 509 static void cleanup_io_u(struct thread_data *td)
 510 {
 511         struct list_head *entry, *n;
 512         struct io_u *io_u;
 513
 514         list_for_each_safe(entry, n, &td->io_u_freelist) {
 515                 io_u = list_entry(entry, struct io_u, list);
 516
 517                 list_del(&io_u->list);
 518                 free(io_u);
 519         }
 520
 521         free_io_mem(td);
 522 }
 523
 524 /*
 525  * "randomly" fill the buffer contents
 526  */
 527 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 528 {
 529         int *ptr = io_u->buf;
 530
 531         while ((void *) ptr - io_u->buf < max_bs) {
 532                 *ptr = rand() * 0x9e370001;
 533                 ptr++;
 534         }
 535 }
 536
 537 static int init_io_u(struct thread_data *td)
 538 {
 539         struct io_u *io_u;
 540         unsigned int max_bs;
 541         int i, max_units;
 542         char *p;
 543
 544         if (td->io_ops->flags & FIO_CPUIO)
 545                 return 0;
 546
 547         if (td->io_ops->flags & FIO_SYNCIO)
 548                 max_units = 1;
 549         else
 550                 max_units = td->iodepth;
 551
 552         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 553         td->orig_buffer_size = max_bs * max_units;
 554
 555         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 556                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 557         else
 558                 td->orig_buffer_size += page_mask;
 559
 560         if (allocate_io_mem(td))
 561                 return 1;
 562
 563         p = ALIGN(td->orig_buffer);
 564         for (i = 0; i < max_units; i++) {
 565                 io_u = malloc(sizeof(*io_u));
 566                 memset(io_u, 0, sizeof(*io_u));
 567                 INIT_LIST_HEAD(&io_u->list);
 568
 569                 io_u->buf = p + max_bs * i;
 570                 if (td_write(td) || td_rw(td))
 571                         fill_rand_buf(io_u, max_bs);
 572
 573                 io_u->index = i;
 574                 list_add(&io_u->list, &td->io_u_freelist);
 575         }
 576
 577         return 0;
 578 }
 579
 580 static int switch_ioscheduler(struct thread_data *td)
 581 {
 582         char tmp[256], tmp2[128];
 583         FILE *f;
 584         int ret;
 585
 586         if (td->io_ops->flags & FIO_CPUIO)
 587                 return 0;
 588
 589         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 590
 591         f = fopen(tmp, "r+");
 592         if (!f) {
 593                 td_verror(td, errno);
 594                 return 1;
 595         }
 596
 597         /*
 598          * Set io scheduler.
 599          */
 600         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 601         if (ferror(f) || ret != 1) {
 602                 td_verror(td, errno);
 603                 fclose(f);
 604                 return 1;
 605         }
 606
 607         rewind(f);
 608
 609         /*
 610          * Read back and check that the selected scheduler is now the default.
 611          */
 612         ret = fread(tmp, 1, sizeof(tmp), f);
 613         if (ferror(f) || ret < 0) {
 614                 td_verror(td, errno);
 615                 fclose(f);
 616                 return 1;
 617         }
 618
 619         sprintf(tmp2, "[%s]", td->ioscheduler);
 620         if (!strstr(tmp, tmp2)) {
 621                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 622                 td_verror(td, EINVAL);
 623                 fclose(f);
 624                 return 1;
 625         }
 626
 627         fclose(f);
 628         return 0;
 629 }
 630
 631 static void clear_io_state(struct thread_data *td)
 632 {
 633         struct fio_file *f;
 634         int i;
 635
 636         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 637         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 638         td->zone_bytes = 0;
 639
 640         for_each_file(td, f, i) {
 641                 f->last_pos = 0;
 642                 if (td->io_ops->flags & FIO_SYNCIO)
 643                         lseek(f->fd, SEEK_SET, 0);
 644
 645                 if (f->file_map)
 646                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 647         }
 648 }
 649
 650 /*
 651  * Entry point for the thread based jobs. The process based jobs end up
 652  * here as well, after a little setup.
 653  */
 654 static void *thread_main(void *data)
 655 {
 656         unsigned long long runtime[2];
 657         struct thread_data *td = data;
 658
 659         if (!td->use_thread)
 660                 setsid();
 661
 662         td->pid = getpid();
 663
 664         INIT_LIST_HEAD(&td->io_u_freelist);
 665         INIT_LIST_HEAD(&td->io_u_busylist);
 666         INIT_LIST_HEAD(&td->io_hist_list);
 667         INIT_LIST_HEAD(&td->io_log_list);
 668
 669         if (init_io_u(td))
 670                 goto err;
 671
 672         if (fio_setaffinity(td) == -1) {
 673                 td_verror(td, errno);
 674                 goto err;
 675         }
 676
 677         if (init_iolog(td))
 678                 goto err;
 679
 680         if (td->ioprio) {
 681                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 682                         td_verror(td, errno);
 683                         goto err;
 684                 }
 685         }
 686
 687         if (nice(td->nice) == -1) {
 688                 td_verror(td, errno);
 689                 goto err;
 690         }
 691
 692         if (init_random_state(td))
 693                 goto err;
 694
 695         if (td->ioscheduler && switch_ioscheduler(td))
 696                 goto err;
 697
 698         td_set_runstate(td, TD_INITIALIZED);
 699         fio_sem_up(&startup_sem);
 700         fio_sem_down(&td->mutex);
 701
 702         if (!td->create_serialize && setup_files(td))
 703                 goto err;
 704         if (open_files(td))
 705                 goto err;
 706
 707         /*
 708          * Do this late, as some IO engines would like to have the
 709          * files setup prior to initializing structures.
 710          */
 711         if (td_io_init(td))
 712                 goto err;
 713
 714         if (td->exec_prerun) {
 715                 if (system(td->exec_prerun) < 0)
 716                         goto err;
 717         }
 718
 719         fio_gettime(&td->epoch, NULL);
 720         getrusage(RUSAGE_SELF, &td->ru_start);
 721
 722         runtime[0] = runtime[1] = 0;
 723         while (td->loops--) {
 724                 fio_gettime(&td->start, NULL);
 725                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 726
 727                 if (td->ratemin)
 728                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
 729
 730                 clear_io_state(td);
 731                 prune_io_piece_log(td);
 732
 733                 if (td->io_ops->flags & FIO_CPUIO)
 734                         do_cpuio(td);
 735                 else
 736                         do_io(td);
 737
 738                 runtime[td->ddir] += utime_since_now(&td->start);
 739                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
 740                         runtime[td->ddir ^ 1] = runtime[td->ddir];
 741
 742                 if (td->error || td->terminate)
 743                         break;
 744
 745                 if (td->verify == VERIFY_NONE)
 746                         continue;
 747
 748                 clear_io_state(td);
 749                 fio_gettime(&td->start, NULL);
 750
 751                 do_verify(td);
 752
 753                 runtime[DDIR_READ] += utime_since_now(&td->start);
 754
 755                 if (td->error || td->terminate)
 756                         break;
 757         }
 758
 759         update_rusage_stat(td);
 760         fio_gettime(&td->end_time, NULL);
 761         td->runtime[0] = runtime[0] / 1000;
 762         td->runtime[1] = runtime[1] / 1000;
 763
 764         if (td->bw_log)
 765                 finish_log(td, td->bw_log, "bw");
 766         if (td->slat_log)
 767                 finish_log(td, td->slat_log, "slat");
 768         if (td->clat_log)
 769                 finish_log(td, td->clat_log, "clat");
 770         if (td->write_iolog_file)
 771                 write_iolog_close(td);
 772         if (td->exec_postrun) {
 773                 if (system(td->exec_postrun) < 0)
 774                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 775         }
 776
 777         if (exitall_on_terminate)
 778                 terminate_threads(td->groupid, 0);
 779
 780 err:
 781         if (td->error)
 782                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 783         close_files(td);
 784         close_ioengine(td);
 785         cleanup_io_u(td);
 786         td_set_runstate(td, TD_EXITED);
 787         return (void *) td->error;
 788 }
 789
 790 /*
 791  * We cannot pass the td data into a forked process, so attach the td and
 792  * pass it to the thread worker.
 793  */
 794 static int fork_main(int shmid, int offset)
 795 {
 796         struct thread_data *td;
 797         void *data, *ret;
 798
 799         data = shmat(shmid, NULL, 0);
 800         if (data == (void *) -1) {
 801                 int __err = errno;
 802
 803                 perror("shmat");
 804                 return __err;
 805         }
 806
 807         td = data + offset * sizeof(struct thread_data);
 808         ret = thread_main(td);
 809         shmdt(data);
 810         return (int) ret;
 811 }
 812
 813 /*
 814  * Run over the job map and reap the threads that have exited, if any.
 815  */
 816 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 817 {
 818         struct thread_data *td;
 819         int i, cputhreads, pending, status, ret;
 820
 821         /*
 822          * reap exited threads (TD_EXITED -> TD_REAPED)
 823          */
 824         pending = cputhreads = 0;
 825         for_each_td(td, i) {
 826                 /*
 827                  * ->io_ops is NULL for a thread that has closed its
 828                  * io engine
 829                  */
 830                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 831                         cputhreads++;
 832
 833                 if (td->runstate < TD_EXITED) {
 834                         /*
 835                          * check if someone quit or got killed in an unusual way
 836                          */
 837                         ret = waitpid(td->pid, &status, WNOHANG);
 838                         if (ret < 0)
 839                                 perror("waitpid");
 840                         else if ((ret == td->pid) && WIFSIGNALED(status)) {
 841                                 int sig = WTERMSIG(status);
 842
 843                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 844                                 td_set_runstate(td, TD_REAPED);
 845                                 goto reaped;
 846                         }
 847                 }
 848
 849                 if (td->runstate != TD_EXITED) {
 850                         if (td->runstate < TD_RUNNING)
 851                                 pending++;
 852
 853                         continue;
 854                 }
 855
 856                 if (td->error)
 857                         exit_value++;
 858
 859                 td_set_runstate(td, TD_REAPED);
 860
 861                 if (td->use_thread) {
 862                         long ret;
 863
 864                         if (pthread_join(td->thread, (void *) &ret))
 865                                 perror("thread_join");
 866                 } else {
 867                         int status;
 868
 869                         ret = waitpid(td->pid, &status, 0);
 870                         if (ret < 0)
 871                                 perror("waitpid");
 872                         else if (WIFEXITED(status) && WEXITSTATUS(status)) {
 873                                 if (!exit_value)
 874                                         exit_value++;
 875                         }
 876                 }
 877
 878 reaped:
 879                 (*nr_running)--;
 880                 (*m_rate) -= td->ratemin;
 881                 (*t_rate) -= td->rate;
 882         }
 883
 884         if (*nr_running == cputhreads && !pending)
 885                 terminate_threads(TERMINATE_ALL, 0);
 886 }
 887
 888 /*
 889  * Main function for kicking off and reaping jobs, as needed.
 890  */
 891 static void run_threads(void)
 892 {
 893         struct thread_data *td;
 894         unsigned long spent;
 895         int i, todo, nr_running, m_rate, t_rate, nr_started;
 896
 897         if (fio_pin_memory())
 898                 return;
 899
 900         if (!terse_output) {
 901                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 902                 fflush(stdout);
 903         }
 904
 905         signal(SIGINT, sig_handler);
 906         signal(SIGALRM, sig_handler);
 907
 908         todo = thread_number;
 909         nr_running = 0;
 910         nr_started = 0;
 911         m_rate = t_rate = 0;
 912
 913         for_each_td(td, i) {
 914                 print_status_init(td->thread_number - 1);
 915
 916                 if (!td->create_serialize) {
 917                         init_disk_util(td);
 918                         continue;
 919                 }
 920
 921                 /*
 922                  * do file setup here so it happens sequentially,
 923                  * we don't want X number of threads getting their
 924                  * client data interspersed on disk
 925                  */
 926                 if (setup_files(td)) {
 927                         exit_value++;
 928                         if (td->error)
 929                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 930                         td_set_runstate(td, TD_REAPED);
 931                         todo--;
 932                 }
 933
 934                 init_disk_util(td);
 935         }
 936
 937         while (todo) {
 938                 struct thread_data *map[MAX_JOBS];
 939                 struct timeval this_start;
 940                 int this_jobs = 0, left;
 941
 942                 /*
 943                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 944                  */
 945                 for_each_td(td, i) {
 946                         if (td->runstate != TD_NOT_CREATED)
 947                                 continue;
 948
 949                         /*
 950                          * never got a chance to start, killed by other
 951                          * thread for some reason
 952                          */
 953                         if (td->terminate) {
 954                                 todo--;
 955                                 continue;
 956                         }
 957
 958                         if (td->start_delay) {
 959                                 spent = mtime_since_genesis();
 960
 961                                 if (td->start_delay * 1000 > spent)
 962                                         continue;
 963                         }
 964
 965                         if (td->stonewall && (nr_started || nr_running))
 966                                 break;
 967
 968                         /*
 969                          * Set state to created. Thread will transition
 970                          * to TD_INITIALIZED when it's done setting up.
 971                          */
 972                         td_set_runstate(td, TD_CREATED);
 973                         map[this_jobs++] = td;
 974                         fio_sem_init(&startup_sem, 1);
 975                         nr_started++;
 976
 977                         if (td->use_thread) {
 978                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
 979                                         perror("thread_create");
 980                                         nr_started--;
 981                                 }
 982                         } else {
 983                                 if (fork())
 984                                         fio_sem_down(&startup_sem);
 985                                 else {
 986                                         int ret = fork_main(shm_id, i);
 987
 988                                         exit(ret);
 989                                 }
 990                         }
 991                 }
 992
 993                 /*
 994                  * Wait for the started threads to transition to
 995                  * TD_INITIALIZED.
 996                  */
 997                 fio_gettime(&this_start, NULL);
 998                 left = this_jobs;
 999                 while (left && !fio_abort) {
1000                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1001                                 break;
1002
1003                         usleep(100000);
1004
1005                         for (i = 0; i < this_jobs; i++) {
1006                                 td = map[i];
1007                                 if (!td)
1008                                         continue;
1009                                 if (td->runstate == TD_INITIALIZED) {
1010                                         map[i] = NULL;
1011                                         left--;
1012                                 } else if (td->runstate >= TD_EXITED) {
1013                                         map[i] = NULL;
1014                                         left--;
1015                                         todo--;
1016                                         nr_running++; /* work-around... */
1017                                 }
1018                         }
1019                 }
1020
1021                 if (left) {
1022                         log_err("fio: %d jobs failed to start\n", left);
1023                         for (i = 0; i < this_jobs; i++) {
1024                                 td = map[i];
1025                                 if (!td)
1026                                         continue;
1027                                 kill(td->pid, SIGTERM);
1028                         }
1029                         break;
1030                 }
1031
1032                 /*
1033                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1034                  */
1035                 for_each_td(td, i) {
1036                         if (td->runstate != TD_INITIALIZED)
1037                                 continue;
1038
1039                         td_set_runstate(td, TD_RUNNING);
1040                         nr_running++;
1041                         nr_started--;
1042                         m_rate += td->ratemin;
1043                         t_rate += td->rate;
1044                         todo--;
1045                         fio_sem_up(&td->mutex);
1046                 }
1047
1048                 reap_threads(&nr_running, &t_rate, &m_rate);
1049
1050                 if (todo)
1051                         usleep(100000);
1052         }
1053
1054         while (nr_running) {
1055                 reap_threads(&nr_running, &t_rate, &m_rate);
1056                 usleep(10000);
1057         }
1058
1059         update_io_ticks();
1060         fio_unpin_memory();
1061 }
1062
1063 int main(int argc, char *argv[])
1064 {
1065         long ps;
1066
1067         /*
1068          * We need locale for number printing, if it isn't set then just
1069          * go with the US format.
1070          */
1071         if (!getenv("LC_NUMERIC"))
1072                 setlocale(LC_NUMERIC, "en_US");
1073
1074         if (parse_options(argc, argv))
1075                 return 1;
1076
1077         if (!thread_number) {
1078                 log_err("Nothing to do\n");
1079                 return 1;
1080         }
1081
1082         ps = sysconf(_SC_PAGESIZE);
1083         if (ps < 0) {
1084                 log_err("Failed to get page size\n");
1085                 return 1;
1086         }
1087
1088         page_mask = ps - 1;
1089
1090         if (write_bw_log) {
1091                 setup_log(&agg_io_log[DDIR_READ]);
1092                 setup_log(&agg_io_log[DDIR_WRITE]);
1093         }
1094
1095         disk_util_timer_arm();
1096
1097         run_threads();
1098
1099         if (!fio_abort) {
1100                 show_run_stats();
1101                 if (write_bw_log) {
1102                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1103                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1104                 }
1105         }
1106
1107         return exit_value;
1108 }