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
  39 unsigned long page_mask;
  40 unsigned long page_size;
  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 nr_process = 0;
  47 int nr_thread = 0;
  48 int shm_id = 0;
  49 int temp_stall_ts;
  50
  51 static struct fio_sem *startup_sem;
  52 static volatile int fio_abort;
  53 static int exit_value;
  54
  55 struct io_log *agg_io_log[2];
  56
  57 #define TERMINATE_ALL           (-1)
  58 #define JOB_START_TIMEOUT       (5 * 1000)
  59
  60 static inline void td_set_runstate(struct thread_data *td, int runstate)
  61 {
  62         td->runstate = runstate;
  63 }
  64
  65 static void terminate_threads(int group_id)
  66 {
  67         struct thread_data *td;
  68         int i;
  69
  70         for_each_td(td, i) {
  71                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  72                         /*
  73                          * if the thread is running, just let it exit
  74                          */
  75                         if (td->runstate < TD_RUNNING)
  76                                 kill(td->pid, SIGQUIT);
  77                         td->terminate = 1;
  78                         td->o.start_delay = 0;
  79                 }
  80         }
  81 }
  82
  83 static void sig_handler(int sig)
  84 {
  85         switch (sig) {
  86                 case SIGALRM:
  87                         update_io_ticks();
  88                         disk_util_timer_arm();
  89                         print_thread_status();
  90                         break;
  91                 default:
  92                         printf("\nfio: terminating on signal %d\n", sig);
  93                         fflush(stdout);
  94                         terminate_threads(TERMINATE_ALL);
  95                         break;
  96         }
  97 }
  98
  99 /*
 100  * Check if we are above the minimum rate given.
 101  */
 102 static int check_min_rate(struct thread_data *td, struct timeval *now)
 103 {
 104         unsigned long long bytes = 0;
 105         unsigned long iops = 0;
 106         unsigned long spent;
 107         unsigned long rate;
 108
 109         /*
 110          * No minimum rate set, always ok
 111          */
 112         if (!td->o.ratemin && !td->o.rate_iops_min)
 113                 return 0;
 114
 115         /*
 116          * allow a 2 second settle period in the beginning
 117          */
 118         if (mtime_since(&td->start, now) < 2000)
 119                 return 0;
 120
 121         if (td_read(td)) {
 122                 iops += td->io_blocks[DDIR_READ];
 123                 bytes += td->this_io_bytes[DDIR_READ];
 124         }
 125         if (td_write(td)) {
 126                 iops += td->io_blocks[DDIR_WRITE];
 127                 bytes += td->this_io_bytes[DDIR_WRITE];
 128         }
 129
 130         /*
 131          * if rate blocks is set, sample is running
 132          */
 133         if (td->rate_bytes || td->rate_blocks) {
 134                 spent = mtime_since(&td->lastrate, now);
 135                 if (spent < td->o.ratecycle)
 136                         return 0;
 137
 138                 if (td->o.rate) {
 139                         /*
 140                          * check bandwidth specified rate
 141                          */
 142                         if (bytes < td->rate_bytes) {
 143                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
 144                                 return 1;
 145                         } else {
 146                                 rate = (bytes - td->rate_bytes) / spent;
 147                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
 148                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
 149                                         return 1;
 150                                 }
 151                         }
 152                 } else {
 153                         /*
 154                          * checks iops specified rate
 155                          */
 156                         if (iops < td->o.rate_iops) {
 157                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
 158                                 return 1;
 159                         } else {
 160                                 rate = (iops - td->rate_blocks) / spent;
 161                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
 162                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
 163                                 }
 164                         }
 165                 }
 166         }
 167
 168         td->rate_bytes = bytes;
 169         td->rate_blocks = iops;
 170         memcpy(&td->lastrate, now, sizeof(*now));
 171         return 0;
 172 }
 173
 174 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 175 {
 176         if (!td->o.timeout)
 177                 return 0;
 178         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 179                 return 1;
 180
 181         return 0;
 182 }
 183
 184 /*
 185  * When job exits, we can cancel the in-flight IO if we are using async
 186  * io. Attempt to do so.
 187  */
 188 static void cleanup_pending_aio(struct thread_data *td)
 189 {
 190         struct list_head *entry, *n;
 191         struct io_u *io_u;
 192         int r;
 193
 194         /*
 195          * get immediately available events, if any
 196          */
 197         r = io_u_queued_complete(td, 0);
 198         if (r < 0)
 199                 return;
 200
 201         /*
 202          * now cancel remaining active events
 203          */
 204         if (td->io_ops->cancel) {
 205                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 206                         io_u = list_entry(entry, struct io_u, list);
 207
 208                         /*
 209                          * if the io_u isn't in flight, then that generally
 210                          * means someone leaked an io_u. complain but fix
 211                          * it up, so we don't stall here.
 212                          */
 213                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 214                                 log_err("fio: non-busy IO on busy list\n");
 215                                 put_io_u(td, io_u);
 216                         } else {
 217                                 r = td->io_ops->cancel(td, io_u);
 218                                 if (!r)
 219                                         put_io_u(td, io_u);
 220                         }
 221                 }
 222         }
 223
 224         if (td->cur_depth)
 225                 r = io_u_queued_complete(td, td->cur_depth);
 226 }
 227
 228 /*
 229  * Helper to handle the final sync of a file. Works just like the normal
 230  * io path, just does everything sync.
 231  */
 232 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 233 {
 234         struct io_u *io_u = __get_io_u(td);
 235         int ret;
 236
 237         if (!io_u)
 238                 return 1;
 239
 240         io_u->ddir = DDIR_SYNC;
 241         io_u->file = f;
 242
 243         if (td_io_prep(td, io_u)) {
 244                 put_io_u(td, io_u);
 245                 return 1;
 246         }
 247
 248 requeue:
 249         ret = td_io_queue(td, io_u);
 250         if (ret < 0) {
 251                 td_verror(td, io_u->error, "td_io_queue");
 252                 put_io_u(td, io_u);
 253                 return 1;
 254         } else if (ret == FIO_Q_QUEUED) {
 255                 if (io_u_queued_complete(td, 1) < 0)
 256                         return 1;
 257         } else if (ret == FIO_Q_COMPLETED) {
 258                 if (io_u->error) {
 259                         td_verror(td, io_u->error, "td_io_queue");
 260                         return 1;
 261                 }
 262
 263                 if (io_u_sync_complete(td, io_u) < 0)
 264                         return 1;
 265         } else if (ret == FIO_Q_BUSY) {
 266                 if (td_io_commit(td))
 267                         return 1;
 268                 goto requeue;
 269         }
 270
 271         return 0;
 272 }
 273
 274 /*
 275  * The main verify engine. Runs over the writes we previously submitted,
 276  * reads the blocks back in, and checks the crc/md5 of the data.
 277  */
 278 static void do_verify(struct thread_data *td)
 279 {
 280         struct fio_file *f;
 281         struct io_u *io_u;
 282         int ret, min_events;
 283         unsigned int i;
 284
 285         /*
 286          * sync io first and invalidate cache, to make sure we really
 287          * read from disk.
 288          */
 289         for_each_file(td, f, i) {
 290                 if (!(f->flags & FIO_FILE_OPEN))
 291                         continue;
 292                 if (fio_io_sync(td, f))
 293                         break;
 294                 if (file_invalidate_cache(td, f))
 295                         break;
 296         }
 297
 298         if (td->error)
 299                 return;
 300
 301         td_set_runstate(td, TD_VERIFYING);
 302
 303         io_u = NULL;
 304         while (!td->terminate) {
 305                 int ret2;
 306
 307                 io_u = __get_io_u(td);
 308                 if (!io_u)
 309                         break;
 310
 311                 if (runtime_exceeded(td, &io_u->start_time)) {
 312                         put_io_u(td, io_u);
 313                         break;
 314                 }
 315
 316                 if (get_next_verify(td, io_u)) {
 317                         put_io_u(td, io_u);
 318                         break;
 319                 }
 320
 321                 if (td_io_prep(td, io_u)) {
 322                         put_io_u(td, io_u);
 323                         break;
 324                 }
 325
 326                 io_u->end_io = verify_io_u;
 327
 328                 ret = td_io_queue(td, io_u);
 329                 switch (ret) {
 330                 case FIO_Q_COMPLETED:
 331                         if (io_u->error)
 332                                 ret = -io_u->error;
 333                         else if (io_u->resid) {
 334                                 int bytes = io_u->xfer_buflen - io_u->resid;
 335                                 struct fio_file *f = io_u->file;
 336
 337                                 /*
 338                                  * zero read, fail
 339                                  */
 340                                 if (!bytes) {
 341                                         td_verror(td, ENODATA, "full resid");
 342                                         put_io_u(td, io_u);
 343                                         break;
 344                                 }
 345
 346                                 io_u->xfer_buflen = io_u->resid;
 347                                 io_u->xfer_buf += bytes;
 348                                 io_u->offset += bytes;
 349                                 f->last_completed_pos = io_u->offset;
 350
 351                                 if (io_u->offset == f->real_file_size)
 352                                         goto sync_done;
 353
 354                                 requeue_io_u(td, &io_u);
 355                         } else {
 356 sync_done:
 357                                 ret = io_u_sync_complete(td, io_u);
 358                                 if (ret < 0)
 359                                         break;
 360                         }
 361                         continue;
 362                 case FIO_Q_QUEUED:
 363                         break;
 364                 case FIO_Q_BUSY:
 365                         requeue_io_u(td, &io_u);
 366                         ret2 = td_io_commit(td);
 367                         if (ret2 < 0)
 368                                 ret = ret2;
 369                         break;
 370                 default:
 371                         assert(ret < 0);
 372                         td_verror(td, -ret, "td_io_queue");
 373                         break;
 374                 }
 375
 376                 if (ret < 0 || td->error)
 377                         break;
 378
 379                 /*
 380                  * if we can queue more, do so. but check if there are
 381                  * completed io_u's first.
 382                  */
 383                 min_events = 0;
 384                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 385                         min_events = 1;
 386
 387                         if (td->cur_depth > td->o.iodepth_low)
 388                                 min_events = td->cur_depth - td->o.iodepth_low;
 389                 }
 390
 391                 /*
 392                  * Reap required number of io units, if any, and do the
 393                  * verification on them through the callback handler
 394                  */
 395                 if (io_u_queued_complete(td, min_events) < 0)
 396                         break;
 397         }
 398
 399         if (!td->error) {
 400                 min_events = td->cur_depth;
 401
 402                 if (min_events)
 403                         ret = io_u_queued_complete(td, min_events);
 404         } else
 405                 cleanup_pending_aio(td);
 406
 407         td_set_runstate(td, TD_RUNNING);
 408 }
 409
 410 /*
 411  * Main IO worker function. It retrieves io_u's to process and queues
 412  * and reaps them, checking for rate and errors along the way.
 413  */
 414 static void do_io(struct thread_data *td)
 415 {
 416         struct timeval s;
 417         unsigned long usec;
 418         unsigned int i;
 419         int ret = 0;
 420
 421         td_set_runstate(td, TD_RUNNING);
 422
 423         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 424                 struct timeval comp_time;
 425                 long bytes_done = 0;
 426                 int min_evts = 0;
 427                 struct io_u *io_u;
 428                 int ret2;
 429
 430                 if (td->terminate)
 431                         break;
 432
 433                 io_u = get_io_u(td);
 434                 if (!io_u)
 435                         break;
 436
 437                 memcpy(&s, &io_u->start_time, sizeof(s));
 438
 439                 if (runtime_exceeded(td, &s)) {
 440                         put_io_u(td, io_u);
 441                         break;
 442                 }
 443
 444                 ret = td_io_queue(td, io_u);
 445                 switch (ret) {
 446                 case FIO_Q_COMPLETED:
 447                         if (io_u->error)
 448                                 ret = -io_u->error;
 449                         else if (io_u->resid) {
 450                                 int bytes = io_u->xfer_buflen - io_u->resid;
 451                                 struct fio_file *f = io_u->file;
 452
 453                                 /*
 454                                  * zero read, fail
 455                                  */
 456                                 if (!bytes) {
 457                                         td_verror(td, ENODATA, "full resid");
 458                                         put_io_u(td, io_u);
 459                                         break;
 460                                 }
 461
 462                                 io_u->xfer_buflen = io_u->resid;
 463                                 io_u->xfer_buf += bytes;
 464                                 io_u->offset += bytes;
 465                                 f->last_completed_pos = io_u->offset;
 466
 467                                 if (io_u->offset == f->real_file_size)
 468                                         goto sync_done;
 469
 470                                 requeue_io_u(td, &io_u);
 471                         } else {
 472 sync_done:
 473                                 fio_gettime(&comp_time, NULL);
 474                                 bytes_done = io_u_sync_complete(td, io_u);
 475                                 if (bytes_done < 0)
 476                                         ret = bytes_done;
 477                         }
 478                         break;
 479                 case FIO_Q_QUEUED:
 480                         /*
 481                          * if the engine doesn't have a commit hook,
 482                          * the io_u is really queued. if it does have such
 483                          * a hook, it has to call io_u_queued() itself.
 484                          */
 485                         if (td->io_ops->commit == NULL)
 486                                 io_u_queued(td, io_u);
 487                         break;
 488                 case FIO_Q_BUSY:
 489                         requeue_io_u(td, &io_u);
 490                         ret2 = td_io_commit(td);
 491                         if (ret2 < 0)
 492                                 ret = ret2;
 493                         break;
 494                 default:
 495                         assert(ret < 0);
 496                         put_io_u(td, io_u);
 497                         break;
 498                 }
 499
 500                 if (ret < 0 || td->error)
 501                         break;
 502
 503                 /*
 504                  * See if we need to complete some commands
 505                  */
 506                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 507                         min_evts = 0;
 508                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 509                                 min_evts = 1;
 510
 511                                 if (td->cur_depth > td->o.iodepth_low)
 512                                         min_evts = td->cur_depth - td->o.iodepth_low;
 513                         }
 514
 515                         fio_gettime(&comp_time, NULL);
 516                         bytes_done = io_u_queued_complete(td, min_evts);
 517                         if (bytes_done < 0)
 518                                 break;
 519                 }
 520
 521                 if (!bytes_done)
 522                         continue;
 523
 524                 /*
 525                  * the rate is batched for now, it should work for batches
 526                  * of completions except the very first one which may look
 527                  * a little bursty
 528                  */
 529                 usec = utime_since(&s, &comp_time);
 530
 531                 rate_throttle(td, usec, bytes_done);
 532
 533                 if (check_min_rate(td, &comp_time)) {
 534                         if (exitall_on_terminate)
 535                                 terminate_threads(td->groupid);
 536                         td_verror(td, ENODATA, "check_min_rate");
 537                         break;
 538                 }
 539
 540                 if (td->o.thinktime) {
 541                         unsigned long long b;
 542
 543                         b = td->io_blocks[0] + td->io_blocks[1];
 544                         if (!(b % td->o.thinktime_blocks)) {
 545                                 int left;
 546
 547                                 if (td->o.thinktime_spin)
 548                                         __usec_sleep(td->o.thinktime_spin);
 549
 550                                 left = td->o.thinktime - td->o.thinktime_spin;
 551                                 if (left)
 552                                         usec_sleep(td, left);
 553                         }
 554                 }
 555         }
 556
 557         if (!td->error) {
 558                 struct fio_file *f;
 559
 560                 i = td->cur_depth;
 561                 if (i)
 562                         ret = io_u_queued_complete(td, i);
 563
 564                 if (should_fsync(td) && td->o.end_fsync) {
 565                         td_set_runstate(td, TD_FSYNCING);
 566
 567                         for_each_file(td, f, i) {
 568                                 if (!(f->flags & FIO_FILE_OPEN))
 569                                         continue;
 570                                 fio_io_sync(td, f);
 571                         }
 572                 }
 573         } else
 574                 cleanup_pending_aio(td);
 575 }
 576
 577 static void cleanup_io_u(struct thread_data *td)
 578 {
 579         struct list_head *entry, *n;
 580         struct io_u *io_u;
 581
 582         list_for_each_safe(entry, n, &td->io_u_freelist) {
 583                 io_u = list_entry(entry, struct io_u, list);
 584
 585                 list_del(&io_u->list);
 586                 free(io_u);
 587         }
 588
 589         free_io_mem(td);
 590 }
 591
 592 /*
 593  * "randomly" fill the buffer contents
 594  */
 595 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 596 {
 597         int *ptr = io_u->buf;
 598
 599         while ((void *) ptr - io_u->buf < max_bs) {
 600                 *ptr = rand() * 0x9e370001;
 601                 ptr++;
 602         }
 603 }
 604
 605 static int init_io_u(struct thread_data *td)
 606 {
 607         unsigned long long buf_size;
 608         struct io_u *io_u;
 609         unsigned int max_bs;
 610         int i, max_units;
 611         char *p;
 612
 613         if (td->io_ops->flags & FIO_SYNCIO)
 614                 max_units = 1;
 615         else
 616                 max_units = td->o.iodepth;
 617
 618         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 619         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 620         buf_size += page_mask;
 621         if (buf_size != (size_t) buf_size) {
 622                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 623                 return 1;
 624         }
 625
 626         td->orig_buffer_size = buf_size;
 627
 628         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 629                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 630         else if (td->orig_buffer_size & page_mask)
 631                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
 632
 633         if (allocate_io_mem(td))
 634                 return 1;
 635
 636         p = ALIGN(td->orig_buffer);
 637         for (i = 0; i < max_units; i++) {
 638                 io_u = malloc(sizeof(*io_u));
 639                 memset(io_u, 0, sizeof(*io_u));
 640                 INIT_LIST_HEAD(&io_u->list);
 641
 642                 io_u->buf = p + max_bs * i;
 643                 if (td_write(td) || td_rw(td))
 644                         fill_rand_buf(io_u, max_bs);
 645
 646                 io_u->index = i;
 647                 io_u->flags = IO_U_F_FREE;
 648                 list_add(&io_u->list, &td->io_u_freelist);
 649         }
 650
 651         io_u_init_timeout();
 652
 653         return 0;
 654 }
 655
 656 static int switch_ioscheduler(struct thread_data *td)
 657 {
 658         char tmp[256], tmp2[128];
 659         FILE *f;
 660         int ret;
 661
 662         if (td->io_ops->flags & FIO_DISKLESSIO)
 663                 return 0;
 664
 665         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 666
 667         f = fopen(tmp, "r+");
 668         if (!f) {
 669                 if (errno == ENOENT) {
 670                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 671                         return 0;
 672                 }
 673                 td_verror(td, errno, "fopen iosched");
 674                 return 1;
 675         }
 676
 677         /*
 678          * Set io scheduler.
 679          */
 680         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 681         if (ferror(f) || ret != 1) {
 682                 td_verror(td, errno, "fwrite");
 683                 fclose(f);
 684                 return 1;
 685         }
 686
 687         rewind(f);
 688
 689         /*
 690          * Read back and check that the selected scheduler is now the default.
 691          */
 692         ret = fread(tmp, 1, sizeof(tmp), f);
 693         if (ferror(f) || ret < 0) {
 694                 td_verror(td, errno, "fread");
 695                 fclose(f);
 696                 return 1;
 697         }
 698
 699         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 700         if (!strstr(tmp, tmp2)) {
 701                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 702                 td_verror(td, EINVAL, "iosched_switch");
 703                 fclose(f);
 704                 return 1;
 705         }
 706
 707         fclose(f);
 708         return 0;
 709 }
 710
 711 static int clear_io_state(struct thread_data *td)
 712 {
 713         struct fio_file *f;
 714         unsigned int i;
 715         int ret;
 716
 717         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 718         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 719         td->zone_bytes = 0;
 720         td->rate_bytes = 0;
 721         td->rate_blocks = 0;
 722         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 723
 724         td->last_was_sync = 0;
 725
 726         for_each_file(td, f, i)
 727                 td_io_close_file(td, f);
 728
 729         ret = 0;
 730         for_each_file(td, f, i) {
 731                 ret = td_io_open_file(td, f);
 732                 if (ret)
 733                         break;
 734         }
 735
 736         return ret;
 737 }
 738
 739 /*
 740  * Entry point for the thread based jobs. The process based jobs end up
 741  * here as well, after a little setup.
 742  */
 743 static void *thread_main(void *data)
 744 {
 745         unsigned long long runtime[2];
 746         struct thread_data *td = data;
 747         unsigned long elapsed;
 748         int clear_state;
 749
 750         if (!td->o.use_thread)
 751                 setsid();
 752
 753         td->pid = getpid();
 754
 755         INIT_LIST_HEAD(&td->io_u_freelist);
 756         INIT_LIST_HEAD(&td->io_u_busylist);
 757         INIT_LIST_HEAD(&td->io_u_requeues);
 758         INIT_LIST_HEAD(&td->io_log_list);
 759         INIT_LIST_HEAD(&td->io_hist_list);
 760         td->io_hist_tree = RB_ROOT;
 761
 762         if (init_io_u(td))
 763                 goto err_sem;
 764
 765         if (fio_setaffinity(td) == -1) {
 766                 td_verror(td, errno, "cpu_set_affinity");
 767                 goto err_sem;
 768         }
 769
 770         if (init_iolog(td))
 771                 goto err_sem;
 772
 773         if (td->ioprio) {
 774                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 775                         td_verror(td, errno, "ioprio_set");
 776                         goto err_sem;
 777                 }
 778         }
 779
 780         if (nice(td->o.nice) == -1) {
 781                 td_verror(td, errno, "nice");
 782                 goto err_sem;
 783         }
 784
 785         if (td->o.ioscheduler && switch_ioscheduler(td))
 786                 goto err_sem;
 787
 788         td_set_runstate(td, TD_INITIALIZED);
 789         fio_sem_up(startup_sem);
 790         fio_sem_down(td->mutex);
 791
 792         /*
 793          * the ->mutex semaphore is now no longer used, close it to avoid
 794          * eating a file descriptor
 795          */
 796         fio_sem_remove(td->mutex);
 797
 798         if (!td->o.create_serialize && setup_files(td))
 799                 goto err;
 800
 801         if (td_io_init(td))
 802                 goto err;
 803
 804         if (open_files(td))
 805                 goto err;
 806
 807         if (init_random_map(td))
 808                 goto err;
 809
 810         if (td->o.exec_prerun) {
 811                 if (system(td->o.exec_prerun) < 0)
 812                         goto err;
 813         }
 814
 815         fio_gettime(&td->epoch, NULL);
 816         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 817         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 818
 819         runtime[0] = runtime[1] = 0;
 820         clear_state = 0;
 821         while (td->o.loops--) {
 822                 fio_gettime(&td->start, NULL);
 823                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 824
 825                 if (td->o.ratemin)
 826                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 827
 828                 if (clear_state && clear_io_state(td))
 829                         break;
 830
 831                 prune_io_piece_log(td);
 832
 833                 do_io(td);
 834
 835                 clear_state = 1;
 836
 837                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 838                         if (td->rw_end_set[DDIR_READ])
 839                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 840                         else
 841                                 elapsed = utime_since_now(&td->start);
 842
 843                         runtime[DDIR_READ] += elapsed;
 844                 }
 845                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 846                         if (td->rw_end_set[DDIR_WRITE])
 847                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 848                         else
 849                                 elapsed = utime_since_now(&td->start);
 850
 851                         runtime[DDIR_WRITE] += elapsed;
 852                 }
 853
 854                 if (td->error || td->terminate)
 855                         break;
 856
 857                 if (td->o.verify == VERIFY_NONE)
 858                         continue;
 859
 860                 if (clear_io_state(td))
 861                         break;
 862
 863                 fio_gettime(&td->start, NULL);
 864
 865                 do_verify(td);
 866
 867                 runtime[DDIR_READ] += utime_since_now(&td->start);
 868
 869                 if (td->error || td->terminate)
 870                         break;
 871         }
 872
 873         update_rusage_stat(td);
 874         td->ts.runtime[0] = runtime[0] / 1000;
 875         td->ts.runtime[1] = runtime[1] / 1000;
 876         td->ts.total_run_time = mtime_since_now(&td->epoch);
 877         td->ts.io_bytes[0] = td->io_bytes[0];
 878         td->ts.io_bytes[1] = td->io_bytes[1];
 879
 880         if (td->ts.bw_log)
 881                 finish_log(td, td->ts.bw_log, "bw");
 882         if (td->ts.slat_log)
 883                 finish_log(td, td->ts.slat_log, "slat");
 884         if (td->ts.clat_log)
 885                 finish_log(td, td->ts.clat_log, "clat");
 886         if (td->o.write_iolog_file)
 887                 write_iolog_close(td);
 888         if (td->o.exec_postrun) {
 889                 if (system(td->o.exec_postrun) < 0)
 890                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 891         }
 892
 893         if (exitall_on_terminate)
 894                 terminate_threads(td->groupid);
 895
 896 err:
 897         if (td->error)
 898                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 899         close_files(td);
 900         close_ioengine(td);
 901         cleanup_io_u(td);
 902         options_mem_free(td);
 903         td_set_runstate(td, TD_EXITED);
 904         return (void *) (unsigned long) td->error;
 905 err_sem:
 906         fio_sem_up(startup_sem);
 907         goto err;
 908 }
 909
 910 /*
 911  * We cannot pass the td data into a forked process, so attach the td and
 912  * pass it to the thread worker.
 913  */
 914 static int fork_main(int shmid, int offset)
 915 {
 916         struct thread_data *td;
 917         void *data, *ret;
 918
 919         data = shmat(shmid, NULL, 0);
 920         if (data == (void *) -1) {
 921                 int __err = errno;
 922
 923                 perror("shmat");
 924                 return __err;
 925         }
 926
 927         td = data + offset * sizeof(struct thread_data);
 928         ret = thread_main(td);
 929         shmdt(data);
 930         return (int) (unsigned long) ret;
 931 }
 932
 933 /*
 934  * Run over the job map and reap the threads that have exited, if any.
 935  */
 936 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 937 {
 938         struct thread_data *td;
 939         int i, cputhreads, pending, status, ret;
 940
 941         /*
 942          * reap exited threads (TD_EXITED -> TD_REAPED)
 943          */
 944         pending = cputhreads = 0;
 945         for_each_td(td, i) {
 946                 int flags = 0;
 947
 948                 /*
 949                  * ->io_ops is NULL for a thread that has closed its
 950                  * io engine
 951                  */
 952                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 953                         cputhreads++;
 954
 955                 if (!td->pid || td->runstate == TD_REAPED)
 956                         continue;
 957                 if (td->o.use_thread) {
 958                         if (td->runstate == TD_EXITED) {
 959                                 td_set_runstate(td, TD_REAPED);
 960                                 goto reaped;
 961                         }
 962                         continue;
 963                 }
 964
 965                 flags = WNOHANG;
 966                 if (td->runstate == TD_EXITED)
 967                         flags = 0;
 968
 969                 /*
 970                  * check if someone quit or got killed in an unusual way
 971                  */
 972                 ret = waitpid(td->pid, &status, flags);
 973                 if (ret < 0) {
 974                         if (errno == ECHILD) {
 975                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 976                                 td_set_runstate(td, TD_REAPED);
 977                                 goto reaped;
 978                         }
 979                         perror("waitpid");
 980                 } else if (ret == td->pid) {
 981                         if (WIFSIGNALED(status)) {
 982                                 int sig = WTERMSIG(status);
 983
 984                                 if (sig != SIGQUIT)
 985                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 986                                 td_set_runstate(td, TD_REAPED);
 987                                 goto reaped;
 988                         }
 989                         if (WIFEXITED(status)) {
 990                                 if (WEXITSTATUS(status) && !td->error)
 991                                         td->error = WEXITSTATUS(status);
 992
 993                                 td_set_runstate(td, TD_REAPED);
 994                                 goto reaped;
 995                         }
 996                 }
 997
 998                 /*
 999                  * thread is not dead, continue
1000                  */
1001                 pending++;
1002                 continue;
1003 reaped:
1004                 if (td->o.use_thread) {
1005                         long ret;
1006
1007                         if (pthread_join(td->thread, (void *) &ret))
1008                                 perror("pthread_join");
1009                 }
1010
1011                 (*nr_running)--;
1012                 (*m_rate) -= td->o.ratemin;
1013                 (*t_rate) -= td->o.rate;
1014                 pending--;
1015
1016                 if (td->error)
1017                         exit_value++;
1018         }
1019
1020         if (*nr_running == cputhreads && !pending)
1021                 terminate_threads(TERMINATE_ALL);
1022 }
1023
1024 /*
1025  * Main function for kicking off and reaping jobs, as needed.
1026  */
1027 static void run_threads(void)
1028 {
1029         struct thread_data *td;
1030         unsigned long spent;
1031         int i, todo, nr_running, m_rate, t_rate, nr_started;
1032
1033         if (fio_pin_memory())
1034                 return;
1035
1036         if (!terse_output) {
1037                 printf("Starting ");
1038                 if (nr_thread)
1039                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1040                 if (nr_process) {
1041                         if (nr_thread)
1042                                 printf(" and ");
1043                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1044                 }
1045                 printf("\n");
1046                 fflush(stdout);
1047         }
1048
1049         signal(SIGINT, sig_handler);
1050         signal(SIGALRM, sig_handler);
1051
1052         todo = thread_number;
1053         nr_running = 0;
1054         nr_started = 0;
1055         m_rate = t_rate = 0;
1056
1057         for_each_td(td, i) {
1058                 print_status_init(td->thread_number - 1);
1059
1060                 if (!td->o.create_serialize) {
1061                         init_disk_util(td);
1062                         continue;
1063                 }
1064
1065                 /*
1066                  * do file setup here so it happens sequentially,
1067                  * we don't want X number of threads getting their
1068                  * client data interspersed on disk
1069                  */
1070                 if (setup_files(td)) {
1071                         exit_value++;
1072                         if (td->error)
1073                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1074                         td_set_runstate(td, TD_REAPED);
1075                         todo--;
1076                 }
1077
1078                 init_disk_util(td);
1079         }
1080
1081         set_genesis_time();
1082
1083         while (todo) {
1084                 struct thread_data *map[MAX_JOBS];
1085                 struct timeval this_start;
1086                 int this_jobs = 0, left;
1087
1088                 /*
1089                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1090                  */
1091                 for_each_td(td, i) {
1092                         if (td->runstate != TD_NOT_CREATED)
1093                                 continue;
1094
1095                         /*
1096                          * never got a chance to start, killed by other
1097                          * thread for some reason
1098                          */
1099                         if (td->terminate) {
1100                                 todo--;
1101                                 continue;
1102                         }
1103
1104                         if (td->o.start_delay) {
1105                                 spent = mtime_since_genesis();
1106
1107                                 if (td->o.start_delay * 1000 > spent)
1108                                         continue;
1109                         }
1110
1111                         if (td->o.stonewall && (nr_started || nr_running))
1112                                 break;
1113
1114                         /*
1115                          * Set state to created. Thread will transition
1116                          * to TD_INITIALIZED when it's done setting up.
1117                          */
1118                         td_set_runstate(td, TD_CREATED);
1119                         map[this_jobs++] = td;
1120                         nr_started++;
1121
1122                         if (td->o.use_thread) {
1123                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1124                                         perror("thread_create");
1125                                         nr_started--;
1126                                         break;
1127                                 }
1128                         } else {
1129                                 if (!fork()) {
1130                                         int ret = fork_main(shm_id, i);
1131
1132                                         exit(ret);
1133                                 }
1134                         }
1135                         fio_sem_down(startup_sem);
1136                 }
1137
1138                 /*
1139                  * Wait for the started threads to transition to
1140                  * TD_INITIALIZED.
1141                  */
1142                 fio_gettime(&this_start, NULL);
1143                 left = this_jobs;
1144                 while (left && !fio_abort) {
1145                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1146                                 break;
1147
1148                         usleep(100000);
1149
1150                         for (i = 0; i < this_jobs; i++) {
1151                                 td = map[i];
1152                                 if (!td)
1153                                         continue;
1154                                 if (td->runstate == TD_INITIALIZED) {
1155                                         map[i] = NULL;
1156                                         left--;
1157                                 } else if (td->runstate >= TD_EXITED) {
1158                                         map[i] = NULL;
1159                                         left--;
1160                                         todo--;
1161                                         nr_running++; /* work-around... */
1162                                 }
1163                         }
1164                 }
1165
1166                 if (left) {
1167                         log_err("fio: %d jobs failed to start\n", left);
1168                         for (i = 0; i < this_jobs; i++) {
1169                                 td = map[i];
1170                                 if (!td)
1171                                         continue;
1172                                 kill(td->pid, SIGTERM);
1173                         }
1174                         break;
1175                 }
1176
1177                 /*
1178                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1179                  */
1180                 for_each_td(td, i) {
1181                         if (td->runstate != TD_INITIALIZED)
1182                                 continue;
1183
1184                         td_set_runstate(td, TD_RUNNING);
1185                         nr_running++;
1186                         nr_started--;
1187                         m_rate += td->o.ratemin;
1188                         t_rate += td->o.rate;
1189                         todo--;
1190                         fio_sem_up(td->mutex);
1191                 }
1192
1193                 reap_threads(&nr_running, &t_rate, &m_rate);
1194
1195                 if (todo)
1196                         usleep(100000);
1197         }
1198
1199         while (nr_running) {
1200                 reap_threads(&nr_running, &t_rate, &m_rate);
1201                 usleep(10000);
1202         }
1203
1204         update_io_ticks();
1205         fio_unpin_memory();
1206 }
1207
1208 int main(int argc, char *argv[])
1209 {
1210         long ps;
1211
1212         /*
1213          * We need locale for number printing, if it isn't set then just
1214          * go with the US format.
1215          */
1216         if (!getenv("LC_NUMERIC"))
1217                 setlocale(LC_NUMERIC, "en_US");
1218
1219         if (parse_options(argc, argv))
1220                 return 1;
1221
1222         if (!thread_number)
1223                 return 0;
1224
1225         ps = sysconf(_SC_PAGESIZE);
1226         if (ps < 0) {
1227                 log_err("Failed to get page size\n");
1228                 return 1;
1229         }
1230
1231         page_size = ps;
1232         page_mask = ps - 1;
1233
1234         if (write_bw_log) {
1235                 setup_log(&agg_io_log[DDIR_READ]);
1236                 setup_log(&agg_io_log[DDIR_WRITE]);
1237         }
1238
1239         startup_sem = fio_sem_init(0);
1240
1241         set_genesis_time();
1242
1243         disk_util_timer_arm();
1244
1245         run_threads();
1246
1247         if (!fio_abort) {
1248                 show_run_stats();
1249                 if (write_bw_log) {
1250                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1251                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1252                 }
1253         }
1254
1255         fio_sem_remove(startup_sem);
1256         return exit_value;
1257 }