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 unsigned long page_mask;
  41 unsigned long page_size;
  42 #define ALIGN(buf)      \
  43         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  44
  45 int groupid = 0;
  46 int thread_number = 0;
  47 int nr_process = 0;
  48 int nr_thread = 0;
  49 int shm_id = 0;
  50 int temp_stall_ts;
  51
  52 static struct fio_sem *startup_sem;
  53 static volatile int fio_abort;
  54 static int exit_value;
  55
  56 struct io_log *agg_io_log[2];
  57
  58 #define TERMINATE_ALL           (-1)
  59 #define JOB_START_TIMEOUT       (5 * 1000)
  60
  61 static inline void td_set_runstate(struct thread_data *td, int runstate)
  62 {
  63         td->runstate = runstate;
  64 }
  65
  66 static void terminate_threads(int group_id)
  67 {
  68         struct thread_data *td;
  69         int i;
  70
  71         for_each_td(td, i) {
  72                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  73                         /*
  74                          * if the thread is running, just let it exit
  75                          */
  76                         if (td->runstate < TD_RUNNING)
  77                                 kill(td->pid, SIGQUIT);
  78                         td->terminate = 1;
  79                         td->o.start_delay = 0;
  80                 }
  81         }
  82 }
  83
  84 static void sig_handler(int sig)
  85 {
  86         switch (sig) {
  87                 case SIGALRM:
  88                         update_io_ticks();
  89                         disk_util_timer_arm();
  90                         print_thread_status();
  91                         break;
  92                 default:
  93                         printf("\nfio: terminating on signal %d\n", sig);
  94                         fflush(stdout);
  95                         terminate_threads(TERMINATE_ALL);
  96                         break;
  97         }
  98 }
  99
 100 /*
 101  * Check if we are above the minimum rate given.
 102  */
 103 static int check_min_rate(struct thread_data *td, struct timeval *now)
 104 {
 105         unsigned long long bytes = 0;
 106         unsigned long iops = 0;
 107         unsigned long spent;
 108         unsigned long rate;
 109
 110         /*
 111          * No minimum rate set, always ok
 112          */
 113         if (!td->o.ratemin && !td->o.rate_iops_min)
 114                 return 0;
 115
 116         /*
 117          * allow a 2 second settle period in the beginning
 118          */
 119         if (mtime_since(&td->start, now) < 2000)
 120                 return 0;
 121
 122         if (td_read(td)) {
 123                 iops += td->io_blocks[DDIR_READ];
 124                 bytes += td->this_io_bytes[DDIR_READ];
 125         }
 126         if (td_write(td)) {
 127                 iops += td->io_blocks[DDIR_WRITE];
 128                 bytes += td->this_io_bytes[DDIR_WRITE];
 129         }
 130
 131         /*
 132          * if rate blocks is set, sample is running
 133          */
 134         if (td->rate_bytes || td->rate_blocks) {
 135                 spent = mtime_since(&td->lastrate, now);
 136                 if (spent < td->o.ratecycle)
 137                         return 0;
 138
 139                 if (td->o.rate) {
 140                         /*
 141                          * check bandwidth specified rate
 142                          */
 143                         if (bytes < td->rate_bytes) {
 144                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
 145                                 return 1;
 146                         } else {
 147                                 rate = (bytes - td->rate_bytes) / spent;
 148                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
 149                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
 150                                         return 1;
 151                                 }
 152                         }
 153                 } else {
 154                         /*
 155                          * checks iops specified rate
 156                          */
 157                         if (iops < td->o.rate_iops) {
 158                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
 159                                 return 1;
 160                         } else {
 161                                 rate = (iops - td->rate_blocks) / spent;
 162                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
 163                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
 164                                 }
 165                         }
 166                 }
 167         }
 168
 169         td->rate_bytes = bytes;
 170         td->rate_blocks = iops;
 171         memcpy(&td->lastrate, now, sizeof(*now));
 172         return 0;
 173 }
 174
 175 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 176 {
 177         if (!td->o.timeout)
 178                 return 0;
 179         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 180                 return 1;
 181
 182         return 0;
 183 }
 184
 185 /*
 186  * When job exits, we can cancel the in-flight IO if we are using async
 187  * io. Attempt to do so.
 188  */
 189 static void cleanup_pending_aio(struct thread_data *td)
 190 {
 191         struct list_head *entry, *n;
 192         struct io_u *io_u;
 193         int r;
 194
 195         /*
 196          * get immediately available events, if any
 197          */
 198         r = io_u_queued_complete(td, 0);
 199         if (r < 0)
 200                 return;
 201
 202         /*
 203          * now cancel remaining active events
 204          */
 205         if (td->io_ops->cancel) {
 206                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 207                         io_u = list_entry(entry, struct io_u, list);
 208
 209                         /*
 210                          * if the io_u isn't in flight, then that generally
 211                          * means someone leaked an io_u. complain but fix
 212                          * it up, so we don't stall here.
 213                          */
 214                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 215                                 log_err("fio: non-busy IO on busy list\n");
 216                                 put_io_u(td, io_u);
 217                         } else {
 218                                 r = td->io_ops->cancel(td, io_u);
 219                                 if (!r)
 220                                         put_io_u(td, io_u);
 221                         }
 222                 }
 223         }
 224
 225         if (td->cur_depth)
 226                 r = io_u_queued_complete(td, td->cur_depth);
 227 }
 228
 229 /*
 230  * Helper to handle the final sync of a file. Works just like the normal
 231  * io path, just does everything sync.
 232  */
 233 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 234 {
 235         struct io_u *io_u = __get_io_u(td);
 236         int ret;
 237
 238         if (!io_u)
 239                 return 1;
 240
 241         io_u->ddir = DDIR_SYNC;
 242         io_u->file = f;
 243
 244         if (td_io_prep(td, io_u)) {
 245                 put_io_u(td, io_u);
 246                 return 1;
 247         }
 248
 249 requeue:
 250         ret = td_io_queue(td, io_u);
 251         if (ret < 0) {
 252                 td_verror(td, io_u->error, "td_io_queue");
 253                 put_io_u(td, io_u);
 254                 return 1;
 255         } else if (ret == FIO_Q_QUEUED) {
 256                 if (io_u_queued_complete(td, 1) < 0)
 257                         return 1;
 258         } else if (ret == FIO_Q_COMPLETED) {
 259                 if (io_u->error) {
 260                         td_verror(td, io_u->error, "td_io_queue");
 261                         return 1;
 262                 }
 263
 264                 if (io_u_sync_complete(td, io_u) < 0)
 265                         return 1;
 266         } else if (ret == FIO_Q_BUSY) {
 267                 if (td_io_commit(td))
 268                         return 1;
 269                 goto requeue;
 270         }
 271
 272         return 0;
 273 }
 274
 275 /*
 276  * The main verify engine. Runs over the writes we previously submitted,
 277  * reads the blocks back in, and checks the crc/md5 of the data.
 278  */
 279 static void do_verify(struct thread_data *td)
 280 {
 281         struct fio_file *f;
 282         struct io_u *io_u;
 283         int ret, min_events;
 284         unsigned int i;
 285
 286         /*
 287          * sync io first and invalidate cache, to make sure we really
 288          * read from disk.
 289          */
 290         for_each_file(td, f, i) {
 291                 if (!(f->flags & FIO_FILE_OPEN))
 292                         continue;
 293                 if (fio_io_sync(td, f))
 294                         break;
 295                 if (file_invalidate_cache(td, f))
 296                         break;
 297         }
 298
 299         if (td->error)
 300                 return;
 301
 302         td_set_runstate(td, TD_VERIFYING);
 303
 304         io_u = NULL;
 305         while (!td->terminate) {
 306                 int ret2;
 307
 308                 io_u = __get_io_u(td);
 309                 if (!io_u)
 310                         break;
 311
 312                 if (runtime_exceeded(td, &io_u->start_time)) {
 313                         put_io_u(td, io_u);
 314                         break;
 315                 }
 316
 317                 if (get_next_verify(td, io_u)) {
 318                         put_io_u(td, io_u);
 319                         break;
 320                 }
 321
 322                 if (td_io_prep(td, io_u)) {
 323                         put_io_u(td, io_u);
 324                         break;
 325                 }
 326
 327                 io_u->end_io = verify_io_u;
 328
 329                 ret = td_io_queue(td, io_u);
 330                 switch (ret) {
 331                 case FIO_Q_COMPLETED:
 332                         if (io_u->error)
 333                                 ret = -io_u->error;
 334                         else if (io_u->resid) {
 335                                 int bytes = io_u->xfer_buflen - io_u->resid;
 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                                 io_u->xfer_buflen = io_u->resid;
 346                                 io_u->xfer_buf += bytes;
 347                                 requeue_io_u(td, &io_u);
 348                         } else {
 349                                 ret = io_u_sync_complete(td, io_u);
 350                                 if (ret < 0)
 351                                         break;
 352                         }
 353                         continue;
 354                 case FIO_Q_QUEUED:
 355                         break;
 356                 case FIO_Q_BUSY:
 357                         requeue_io_u(td, &io_u);
 358                         ret2 = td_io_commit(td);
 359                         if (ret2 < 0)
 360                                 ret = ret2;
 361                         break;
 362                 default:
 363                         assert(ret < 0);
 364                         td_verror(td, -ret, "td_io_queue");
 365                         break;
 366                 }
 367
 368                 if (ret < 0 || td->error)
 369                         break;
 370
 371                 /*
 372                  * if we can queue more, do so. but check if there are
 373                  * completed io_u's first.
 374                  */
 375                 min_events = 0;
 376                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 377                         min_events = 1;
 378
 379                         if (td->cur_depth > td->o.iodepth_low)
 380                                 min_events = td->cur_depth - td->o.iodepth_low;
 381                 }
 382
 383                 /*
 384                  * Reap required number of io units, if any, and do the
 385                  * verification on them through the callback handler
 386                  */
 387                 if (io_u_queued_complete(td, min_events) < 0)
 388                         break;
 389         }
 390
 391         if (!td->error) {
 392                 min_events = td->cur_depth;
 393
 394                 if (min_events)
 395                         ret = io_u_queued_complete(td, min_events);
 396         } else
 397                 cleanup_pending_aio(td);
 398
 399         td_set_runstate(td, TD_RUNNING);
 400 }
 401
 402 /*
 403  * Main IO worker function. It retrieves io_u's to process and queues
 404  * and reaps them, checking for rate and errors along the way.
 405  */
 406 static void do_io(struct thread_data *td)
 407 {
 408         struct timeval s;
 409         unsigned long usec;
 410         unsigned int i;
 411         int ret = 0;
 412
 413         td_set_runstate(td, TD_RUNNING);
 414
 415         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 416                 struct timeval comp_time;
 417                 long bytes_done = 0;
 418                 int min_evts = 0;
 419                 struct io_u *io_u;
 420                 int ret2;
 421
 422                 if (td->terminate)
 423                         break;
 424
 425                 io_u = get_io_u(td);
 426                 if (!io_u)
 427                         break;
 428
 429                 memcpy(&s, &io_u->start_time, sizeof(s));
 430
 431                 if (runtime_exceeded(td, &s)) {
 432                         put_io_u(td, io_u);
 433                         break;
 434                 }
 435
 436                 ret = td_io_queue(td, io_u);
 437                 switch (ret) {
 438                 case FIO_Q_COMPLETED:
 439                         if (io_u->error)
 440                                 ret = -io_u->error;
 441                         else if (io_u->resid) {
 442                                 int bytes = io_u->xfer_buflen - io_u->resid;
 443
 444                                 /*
 445                                  * zero read, fail
 446                                  */
 447                                 if (!bytes) {
 448                                         td_verror(td, ENODATA, "full resid");
 449                                         put_io_u(td, io_u);
 450                                         break;
 451                                 }
 452
 453                                 io_u->xfer_buflen = io_u->resid;
 454                                 io_u->xfer_buf += bytes;
 455                                 requeue_io_u(td, &io_u);
 456                         } else {
 457                                 fio_gettime(&comp_time, NULL);
 458                                 bytes_done = io_u_sync_complete(td, io_u);
 459                                 if (bytes_done < 0)
 460                                         ret = bytes_done;
 461                         }
 462                         break;
 463                 case FIO_Q_QUEUED:
 464                         /*
 465                          * if the engine doesn't have a commit hook,
 466                          * the io_u is really queued. if it does have such
 467                          * a hook, it has to call io_u_queued() itself.
 468                          */
 469                         if (td->io_ops->commit == NULL)
 470                                 io_u_queued(td, io_u);
 471                         break;
 472                 case FIO_Q_BUSY:
 473                         requeue_io_u(td, &io_u);
 474                         ret2 = td_io_commit(td);
 475                         if (ret2 < 0)
 476                                 ret = ret2;
 477                         break;
 478                 default:
 479                         assert(ret < 0);
 480                         put_io_u(td, io_u);
 481                         break;
 482                 }
 483
 484                 if (ret < 0 || td->error)
 485                         break;
 486
 487                 /*
 488                  * See if we need to complete some commands
 489                  */
 490                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 491                         min_evts = 0;
 492                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 493                                 min_evts = 1;
 494
 495                                 if (td->cur_depth > td->o.iodepth_low)
 496                                         min_evts = td->cur_depth - td->o.iodepth_low;
 497                         }
 498
 499                         fio_gettime(&comp_time, NULL);
 500                         bytes_done = io_u_queued_complete(td, min_evts);
 501                         if (bytes_done < 0)
 502                                 break;
 503                 }
 504
 505                 if (!bytes_done)
 506                         continue;
 507
 508                 /*
 509                  * the rate is batched for now, it should work for batches
 510                  * of completions except the very first one which may look
 511                  * a little bursty
 512                  */
 513                 usec = utime_since(&s, &comp_time);
 514
 515                 rate_throttle(td, usec, bytes_done);
 516
 517                 if (check_min_rate(td, &comp_time)) {
 518                         if (exitall_on_terminate)
 519                                 terminate_threads(td->groupid);
 520                         td_verror(td, ENODATA, "check_min_rate");
 521                         break;
 522                 }
 523
 524                 if (td->o.thinktime) {
 525                         unsigned long long b;
 526
 527                         b = td->io_blocks[0] + td->io_blocks[1];
 528                         if (!(b % td->o.thinktime_blocks)) {
 529                                 int left;
 530
 531                                 if (td->o.thinktime_spin)
 532                                         __usec_sleep(td->o.thinktime_spin);
 533
 534                                 left = td->o.thinktime - td->o.thinktime_spin;
 535                                 if (left)
 536                                         usec_sleep(td, left);
 537                         }
 538                 }
 539         }
 540
 541         if (!td->error) {
 542                 struct fio_file *f;
 543
 544                 i = td->cur_depth;
 545                 if (i)
 546                         ret = io_u_queued_complete(td, i);
 547
 548                 if (should_fsync(td) && td->o.end_fsync) {
 549                         td_set_runstate(td, TD_FSYNCING);
 550
 551                         for_each_file(td, f, i) {
 552                                 if (!(f->flags & FIO_FILE_OPEN))
 553                                         continue;
 554                                 fio_io_sync(td, f);
 555                         }
 556                 }
 557         } else
 558                 cleanup_pending_aio(td);
 559 }
 560
 561 static void cleanup_io_u(struct thread_data *td)
 562 {
 563         struct list_head *entry, *n;
 564         struct io_u *io_u;
 565
 566         list_for_each_safe(entry, n, &td->io_u_freelist) {
 567                 io_u = list_entry(entry, struct io_u, list);
 568
 569                 list_del(&io_u->list);
 570                 free(io_u);
 571         }
 572
 573         free_io_mem(td);
 574 }
 575
 576 /*
 577  * "randomly" fill the buffer contents
 578  */
 579 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 580 {
 581         int *ptr = io_u->buf;
 582
 583         while ((void *) ptr - io_u->buf < max_bs) {
 584                 *ptr = rand() * 0x9e370001;
 585                 ptr++;
 586         }
 587 }
 588
 589 static int init_io_u(struct thread_data *td)
 590 {
 591         struct io_u *io_u;
 592         unsigned int max_bs;
 593         int i, max_units;
 594         char *p;
 595
 596         if (td->io_ops->flags & FIO_SYNCIO)
 597                 max_units = 1;
 598         else
 599                 max_units = td->o.iodepth;
 600
 601         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 602         td->orig_buffer_size = max_bs * max_units;
 603
 604         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 605                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 606         else
 607                 td->orig_buffer_size += page_mask;
 608
 609         if (allocate_io_mem(td))
 610                 return 1;
 611
 612         p = ALIGN(td->orig_buffer);
 613         for (i = 0; i < max_units; i++) {
 614                 io_u = malloc(sizeof(*io_u));
 615                 memset(io_u, 0, sizeof(*io_u));
 616                 INIT_LIST_HEAD(&io_u->list);
 617
 618                 io_u->buf = p + max_bs * i;
 619                 if (td_write(td) || td_rw(td))
 620                         fill_rand_buf(io_u, max_bs);
 621
 622                 io_u->index = i;
 623                 io_u->flags = IO_U_F_FREE;
 624                 list_add(&io_u->list, &td->io_u_freelist);
 625         }
 626
 627         io_u_init_timeout();
 628
 629         return 0;
 630 }
 631
 632 static int switch_ioscheduler(struct thread_data *td)
 633 {
 634         char tmp[256], tmp2[128];
 635         FILE *f;
 636         int ret;
 637
 638         if (td->io_ops->flags & FIO_DISKLESSIO)
 639                 return 0;
 640
 641         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 642
 643         f = fopen(tmp, "r+");
 644         if (!f) {
 645                 td_verror(td, errno, "fopen");
 646                 return 1;
 647         }
 648
 649         /*
 650          * Set io scheduler.
 651          */
 652         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 653         if (ferror(f) || ret != 1) {
 654                 td_verror(td, errno, "fwrite");
 655                 fclose(f);
 656                 return 1;
 657         }
 658
 659         rewind(f);
 660
 661         /*
 662          * Read back and check that the selected scheduler is now the default.
 663          */
 664         ret = fread(tmp, 1, sizeof(tmp), f);
 665         if (ferror(f) || ret < 0) {
 666                 td_verror(td, errno, "fread");
 667                 fclose(f);
 668                 return 1;
 669         }
 670
 671         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 672         if (!strstr(tmp, tmp2)) {
 673                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 674                 td_verror(td, EINVAL, "iosched_switch");
 675                 fclose(f);
 676                 return 1;
 677         }
 678
 679         fclose(f);
 680         return 0;
 681 }
 682
 683 static int clear_io_state(struct thread_data *td)
 684 {
 685         struct fio_file *f;
 686         unsigned int i;
 687         int ret;
 688
 689         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 690         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 691         td->zone_bytes = 0;
 692         td->rate_bytes = 0;
 693         td->rate_blocks = 0;
 694
 695         td->last_was_sync = 0;
 696
 697         for_each_file(td, f, i)
 698                 td_io_close_file(td, f);
 699
 700         ret = 0;
 701         for_each_file(td, f, i) {
 702                 ret = td_io_open_file(td, f);
 703                 if (ret)
 704                         break;
 705         }
 706
 707         return ret;
 708 }
 709
 710 /*
 711  * Entry point for the thread based jobs. The process based jobs end up
 712  * here as well, after a little setup.
 713  */
 714 static void *thread_main(void *data)
 715 {
 716         unsigned long long runtime[2];
 717         struct thread_data *td = data;
 718         int clear_state;
 719
 720         if (!td->o.use_thread)
 721                 setsid();
 722
 723         td->pid = getpid();
 724
 725         INIT_LIST_HEAD(&td->io_u_freelist);
 726         INIT_LIST_HEAD(&td->io_u_busylist);
 727         INIT_LIST_HEAD(&td->io_u_requeues);
 728         INIT_LIST_HEAD(&td->io_hist_list);
 729         INIT_LIST_HEAD(&td->io_log_list);
 730
 731         if (init_io_u(td))
 732                 goto err_sem;
 733
 734         if (fio_setaffinity(td) == -1) {
 735                 td_verror(td, errno, "cpu_set_affinity");
 736                 goto err_sem;
 737         }
 738
 739         if (init_iolog(td))
 740                 goto err_sem;
 741
 742         if (td->ioprio) {
 743                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 744                         td_verror(td, errno, "ioprio_set");
 745                         goto err_sem;
 746                 }
 747         }
 748
 749         if (nice(td->o.nice) == -1) {
 750                 td_verror(td, errno, "nice");
 751                 goto err_sem;
 752         }
 753
 754         if (td->o.ioscheduler && switch_ioscheduler(td))
 755                 goto err_sem;
 756
 757         td_set_runstate(td, TD_INITIALIZED);
 758         fio_sem_up(startup_sem);
 759         fio_sem_down(td->mutex);
 760
 761         /*
 762          * the ->mutex semaphore is now no longer used, close it to avoid
 763          * eating a file descriptor
 764          */
 765         fio_sem_remove(td->mutex);
 766
 767         if (!td->o.create_serialize && setup_files(td))
 768                 goto err;
 769
 770         if (td_io_init(td))
 771                 goto err;
 772
 773         if (open_files(td))
 774                 goto err;
 775
 776         if (init_random_map(td))
 777                 goto err;
 778
 779         if (td->o.exec_prerun) {
 780                 if (system(td->o.exec_prerun) < 0)
 781                         goto err;
 782         }
 783
 784         fio_gettime(&td->epoch, NULL);
 785         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 786         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 787
 788         runtime[0] = runtime[1] = 0;
 789         clear_state = 0;
 790         while (td->o.loops--) {
 791                 fio_gettime(&td->start, NULL);
 792                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 793
 794                 if (td->o.ratemin)
 795                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 796
 797                 if (clear_state && clear_io_state(td))
 798                         break;
 799
 800                 prune_io_piece_log(td);
 801
 802                 do_io(td);
 803
 804                 clear_state = 1;
 805
 806                 if (td_read(td) && td->io_bytes[DDIR_READ])
 807                         runtime[DDIR_READ] += utime_since_now(&td->start);
 808                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
 809                         runtime[DDIR_WRITE] += utime_since_now(&td->start);
 810
 811                 if (td->error || td->terminate)
 812                         break;
 813
 814                 if (td->o.verify == VERIFY_NONE)
 815                         continue;
 816
 817                 if (clear_io_state(td))
 818                         break;
 819
 820                 fio_gettime(&td->start, NULL);
 821
 822                 do_verify(td);
 823
 824                 runtime[DDIR_READ] += utime_since_now(&td->start);
 825
 826                 if (td->error || td->terminate)
 827                         break;
 828         }
 829
 830         update_rusage_stat(td);
 831         td->ts.runtime[0] = runtime[0] / 1000;
 832         td->ts.runtime[1] = runtime[1] / 1000;
 833         td->ts.total_run_time = mtime_since_now(&td->epoch);
 834         td->ts.io_bytes[0] = td->io_bytes[0];
 835         td->ts.io_bytes[1] = td->io_bytes[1];
 836
 837         if (td->ts.bw_log)
 838                 finish_log(td, td->ts.bw_log, "bw");
 839         if (td->ts.slat_log)
 840                 finish_log(td, td->ts.slat_log, "slat");
 841         if (td->ts.clat_log)
 842                 finish_log(td, td->ts.clat_log, "clat");
 843         if (td->o.write_iolog_file)
 844                 write_iolog_close(td);
 845         if (td->o.exec_postrun) {
 846                 if (system(td->o.exec_postrun) < 0)
 847                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 848         }
 849
 850         if (exitall_on_terminate)
 851                 terminate_threads(td->groupid);
 852
 853 err:
 854         if (td->error)
 855                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 856         close_files(td);
 857         close_ioengine(td);
 858         cleanup_io_u(td);
 859         td_set_runstate(td, TD_EXITED);
 860         return (void *) (unsigned long) td->error;
 861 err_sem:
 862         fio_sem_up(startup_sem);
 863         goto err;
 864 }
 865
 866 /*
 867  * We cannot pass the td data into a forked process, so attach the td and
 868  * pass it to the thread worker.
 869  */
 870 static int fork_main(int shmid, int offset)
 871 {
 872         struct thread_data *td;
 873         void *data, *ret;
 874
 875         data = shmat(shmid, NULL, 0);
 876         if (data == (void *) -1) {
 877                 int __err = errno;
 878
 879                 perror("shmat");
 880                 return __err;
 881         }
 882
 883         td = data + offset * sizeof(struct thread_data);
 884         ret = thread_main(td);
 885         shmdt(data);
 886         return (int) (unsigned long) ret;
 887 }
 888
 889 /*
 890  * Run over the job map and reap the threads that have exited, if any.
 891  */
 892 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 893 {
 894         struct thread_data *td;
 895         int i, cputhreads, pending, status, ret;
 896
 897         /*
 898          * reap exited threads (TD_EXITED -> TD_REAPED)
 899          */
 900         pending = cputhreads = 0;
 901         for_each_td(td, i) {
 902                 int flags = 0;
 903
 904                 /*
 905                  * ->io_ops is NULL for a thread that has closed its
 906                  * io engine
 907                  */
 908                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 909                         cputhreads++;
 910
 911                 if (!td->pid || td->runstate == TD_REAPED)
 912                         continue;
 913                 if (td->o.use_thread) {
 914                         if (td->runstate == TD_EXITED) {
 915                                 td_set_runstate(td, TD_REAPED);
 916                                 goto reaped;
 917                         }
 918                         continue;
 919                 }
 920
 921                 flags = WNOHANG;
 922                 if (td->runstate == TD_EXITED)
 923                         flags = 0;
 924
 925                 /*
 926                  * check if someone quit or got killed in an unusual way
 927                  */
 928                 ret = waitpid(td->pid, &status, flags);
 929                 if (ret < 0) {
 930                         if (errno == ECHILD) {
 931                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 932                                 td_set_runstate(td, TD_REAPED);
 933                                 goto reaped;
 934                         }
 935                         perror("waitpid");
 936                 } else if (ret == td->pid) {
 937                         if (WIFSIGNALED(status)) {
 938                                 int sig = WTERMSIG(status);
 939
 940                                 if (sig != SIGQUIT)
 941                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 942                                 td_set_runstate(td, TD_REAPED);
 943                                 goto reaped;
 944                         }
 945                         if (WIFEXITED(status)) {
 946                                 if (WEXITSTATUS(status) && !td->error)
 947                                         td->error = WEXITSTATUS(status);
 948
 949                                 td_set_runstate(td, TD_REAPED);
 950                                 goto reaped;
 951                         }
 952                 }
 953
 954                 /*
 955                  * thread is not dead, continue
 956                  */
 957                 continue;
 958 reaped:
 959                 if (td->o.use_thread) {
 960                         long ret;
 961
 962                         if (pthread_join(td->thread, (void *) &ret))
 963                                 perror("pthread_join");
 964                 }
 965
 966                 (*nr_running)--;
 967                 (*m_rate) -= td->o.ratemin;
 968                 (*t_rate) -= td->o.rate;
 969
 970                 if (td->error)
 971                         exit_value++;
 972         }
 973
 974         if (*nr_running == cputhreads && !pending)
 975                 terminate_threads(TERMINATE_ALL);
 976 }
 977
 978 /*
 979  * Main function for kicking off and reaping jobs, as needed.
 980  */
 981 static void run_threads(void)
 982 {
 983         struct thread_data *td;
 984         unsigned long spent;
 985         int i, todo, nr_running, m_rate, t_rate, nr_started;
 986
 987         if (fio_pin_memory())
 988                 return;
 989
 990         if (!terse_output) {
 991                 printf("Starting ");
 992                 if (nr_thread)
 993                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
 994                 if (nr_process) {
 995                         if (nr_thread)
 996                                 printf(" and ");
 997                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
 998                 }
 999                 printf("\n");
1000                 fflush(stdout);
1001         }
1002
1003         signal(SIGINT, sig_handler);
1004         signal(SIGALRM, sig_handler);
1005
1006         todo = thread_number;
1007         nr_running = 0;
1008         nr_started = 0;
1009         m_rate = t_rate = 0;
1010
1011         for_each_td(td, i) {
1012                 print_status_init(td->thread_number - 1);
1013
1014                 if (!td->o.create_serialize) {
1015                         init_disk_util(td);
1016                         continue;
1017                 }
1018
1019                 /*
1020                  * do file setup here so it happens sequentially,
1021                  * we don't want X number of threads getting their
1022                  * client data interspersed on disk
1023                  */
1024                 if (setup_files(td)) {
1025                         exit_value++;
1026                         if (td->error)
1027                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1028                         td_set_runstate(td, TD_REAPED);
1029                         todo--;
1030                 }
1031
1032                 init_disk_util(td);
1033         }
1034
1035         set_genesis_time();
1036
1037         while (todo) {
1038                 struct thread_data *map[MAX_JOBS];
1039                 struct timeval this_start;
1040                 int this_jobs = 0, left;
1041
1042                 /*
1043                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1044                  */
1045                 for_each_td(td, i) {
1046                         if (td->runstate != TD_NOT_CREATED)
1047                                 continue;
1048
1049                         /*
1050                          * never got a chance to start, killed by other
1051                          * thread for some reason
1052                          */
1053                         if (td->terminate) {
1054                                 todo--;
1055                                 continue;
1056                         }
1057
1058                         if (td->o.start_delay) {
1059                                 spent = mtime_since_genesis();
1060
1061                                 if (td->o.start_delay * 1000 > spent)
1062                                         continue;
1063                         }
1064
1065                         if (td->o.stonewall && (nr_started || nr_running))
1066                                 break;
1067
1068                         /*
1069                          * Set state to created. Thread will transition
1070                          * to TD_INITIALIZED when it's done setting up.
1071                          */
1072                         td_set_runstate(td, TD_CREATED);
1073                         map[this_jobs++] = td;
1074                         nr_started++;
1075
1076                         if (td->o.use_thread) {
1077                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1078                                         perror("thread_create");
1079                                         nr_started--;
1080                                         break;
1081                                 }
1082                         } else {
1083                                 if (!fork()) {
1084                                         int ret = fork_main(shm_id, i);
1085
1086                                         exit(ret);
1087                                 }
1088                         }
1089                         fio_sem_down(startup_sem);
1090                 }
1091
1092                 /*
1093                  * Wait for the started threads to transition to
1094                  * TD_INITIALIZED.
1095                  */
1096                 fio_gettime(&this_start, NULL);
1097                 left = this_jobs;
1098                 while (left && !fio_abort) {
1099                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1100                                 break;
1101
1102                         usleep(100000);
1103
1104                         for (i = 0; i < this_jobs; i++) {
1105                                 td = map[i];
1106                                 if (!td)
1107                                         continue;
1108                                 if (td->runstate == TD_INITIALIZED) {
1109                                         map[i] = NULL;
1110                                         left--;
1111                                 } else if (td->runstate >= TD_EXITED) {
1112                                         map[i] = NULL;
1113                                         left--;
1114                                         todo--;
1115                                         nr_running++; /* work-around... */
1116                                 }
1117                         }
1118                 }
1119
1120                 if (left) {
1121                         log_err("fio: %d jobs failed to start\n", left);
1122                         for (i = 0; i < this_jobs; i++) {
1123                                 td = map[i];
1124                                 if (!td)
1125                                         continue;
1126                                 kill(td->pid, SIGTERM);
1127                         }
1128                         break;
1129                 }
1130
1131                 /*
1132                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1133                  */
1134                 for_each_td(td, i) {
1135                         if (td->runstate != TD_INITIALIZED)
1136                                 continue;
1137
1138                         td_set_runstate(td, TD_RUNNING);
1139                         nr_running++;
1140                         nr_started--;
1141                         m_rate += td->o.ratemin;
1142                         t_rate += td->o.rate;
1143                         todo--;
1144                         fio_sem_up(td->mutex);
1145                 }
1146
1147                 reap_threads(&nr_running, &t_rate, &m_rate);
1148
1149                 if (todo)
1150                         usleep(100000);
1151         }
1152
1153         while (nr_running) {
1154                 reap_threads(&nr_running, &t_rate, &m_rate);
1155                 usleep(10000);
1156         }
1157
1158         update_io_ticks();
1159         fio_unpin_memory();
1160 }
1161
1162 int main(int argc, char *argv[])
1163 {
1164         long ps;
1165
1166         /*
1167          * We need locale for number printing, if it isn't set then just
1168          * go with the US format.
1169          */
1170         if (!getenv("LC_NUMERIC"))
1171                 setlocale(LC_NUMERIC, "en_US");
1172
1173         if (parse_options(argc, argv))
1174                 return 1;
1175
1176         if (!thread_number) {
1177                 log_err("Nothing to do\n");
1178                 return 1;
1179         }
1180
1181         ps = sysconf(_SC_PAGESIZE);
1182         if (ps < 0) {
1183                 log_err("Failed to get page size\n");
1184                 return 1;
1185         }
1186
1187         page_size = ps;
1188         page_mask = ps - 1;
1189
1190         if (write_bw_log) {
1191                 setup_log(&agg_io_log[DDIR_READ]);
1192                 setup_log(&agg_io_log[DDIR_WRITE]);
1193         }
1194
1195         startup_sem = fio_sem_init(0);
1196
1197         set_genesis_time();
1198
1199         disk_util_timer_arm();
1200
1201         run_threads();
1202
1203         if (!fio_abort) {
1204                 show_run_stats();
1205                 if (write_bw_log) {
1206                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1207                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1208                 }
1209         }
1210
1211         fio_sem_remove(startup_sem);
1212         return exit_value;
1213 }