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
 346                                 io_u->xfer_buflen = io_u->resid;
 347                                 io_u->xfer_buf += bytes;
 348                                 io_u->offset += bytes;
 349
 350                                 if (io_u->offset == io_u->file->real_file_size)
 351                                         goto sync_done;
 352
 353                                 requeue_io_u(td, &io_u);
 354                         } else {
 355 sync_done:
 356                                 ret = io_u_sync_complete(td, io_u);
 357                                 if (ret < 0)
 358                                         break;
 359                         }
 360                         continue;
 361                 case FIO_Q_QUEUED:
 362                         break;
 363                 case FIO_Q_BUSY:
 364                         requeue_io_u(td, &io_u);
 365                         ret2 = td_io_commit(td);
 366                         if (ret2 < 0)
 367                                 ret = ret2;
 368                         break;
 369                 default:
 370                         assert(ret < 0);
 371                         td_verror(td, -ret, "td_io_queue");
 372                         break;
 373                 }
 374
 375                 if (ret < 0 || td->error)
 376                         break;
 377
 378                 /*
 379                  * if we can queue more, do so. but check if there are
 380                  * completed io_u's first.
 381                  */
 382                 min_events = 0;
 383                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 384                         min_events = 1;
 385
 386                         if (td->cur_depth > td->o.iodepth_low)
 387                                 min_events = td->cur_depth - td->o.iodepth_low;
 388                 }
 389
 390                 /*
 391                  * Reap required number of io units, if any, and do the
 392                  * verification on them through the callback handler
 393                  */
 394                 if (io_u_queued_complete(td, min_events) < 0)
 395                         break;
 396         }
 397
 398         if (!td->error) {
 399                 min_events = td->cur_depth;
 400
 401                 if (min_events)
 402                         ret = io_u_queued_complete(td, min_events);
 403         } else
 404                 cleanup_pending_aio(td);
 405
 406         td_set_runstate(td, TD_RUNNING);
 407 }
 408
 409 /*
 410  * Main IO worker function. It retrieves io_u's to process and queues
 411  * and reaps them, checking for rate and errors along the way.
 412  */
 413 static void do_io(struct thread_data *td)
 414 {
 415         struct timeval s;
 416         unsigned long usec;
 417         unsigned int i;
 418         int ret = 0;
 419
 420         td_set_runstate(td, TD_RUNNING);
 421
 422         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 423                 struct timeval comp_time;
 424                 long bytes_done = 0;
 425                 int min_evts = 0;
 426                 struct io_u *io_u;
 427                 int ret2;
 428
 429                 if (td->terminate)
 430                         break;
 431
 432                 io_u = get_io_u(td);
 433                 if (!io_u)
 434                         break;
 435
 436                 memcpy(&s, &io_u->start_time, sizeof(s));
 437
 438                 if (runtime_exceeded(td, &s)) {
 439                         put_io_u(td, io_u);
 440                         break;
 441                 }
 442
 443                 ret = td_io_queue(td, io_u);
 444                 switch (ret) {
 445                 case FIO_Q_COMPLETED:
 446                         if (io_u->error)
 447                                 ret = -io_u->error;
 448                         else if (io_u->resid) {
 449                                 int bytes = io_u->xfer_buflen - io_u->resid;
 450
 451                                 /*
 452                                  * zero read, fail
 453                                  */
 454                                 if (!bytes) {
 455                                         td_verror(td, ENODATA, "full resid");
 456                                         put_io_u(td, io_u);
 457                                         break;
 458                                 }
 459
 460                                 io_u->xfer_buflen = io_u->resid;
 461                                 io_u->xfer_buf += bytes;
 462                                 io_u->offset += bytes;
 463
 464                                 if (io_u->offset == io_u->file->real_file_size)
 465                                         goto sync_done;
 466
 467                                 requeue_io_u(td, &io_u);
 468                         } else {
 469 sync_done:
 470                                 fio_gettime(&comp_time, NULL);
 471                                 bytes_done = io_u_sync_complete(td, io_u);
 472                                 if (bytes_done < 0)
 473                                         ret = bytes_done;
 474                         }
 475                         break;
 476                 case FIO_Q_QUEUED:
 477                         /*
 478                          * if the engine doesn't have a commit hook,
 479                          * the io_u is really queued. if it does have such
 480                          * a hook, it has to call io_u_queued() itself.
 481                          */
 482                         if (td->io_ops->commit == NULL)
 483                                 io_u_queued(td, io_u);
 484                         break;
 485                 case FIO_Q_BUSY:
 486                         requeue_io_u(td, &io_u);
 487                         ret2 = td_io_commit(td);
 488                         if (ret2 < 0)
 489                                 ret = ret2;
 490                         break;
 491                 default:
 492                         assert(ret < 0);
 493                         put_io_u(td, io_u);
 494                         break;
 495                 }
 496
 497                 if (ret < 0 || td->error)
 498                         break;
 499
 500                 /*
 501                  * See if we need to complete some commands
 502                  */
 503                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 504                         min_evts = 0;
 505                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 506                                 min_evts = 1;
 507
 508                                 if (td->cur_depth > td->o.iodepth_low)
 509                                         min_evts = td->cur_depth - td->o.iodepth_low;
 510                         }
 511
 512                         fio_gettime(&comp_time, NULL);
 513                         bytes_done = io_u_queued_complete(td, min_evts);
 514                         if (bytes_done < 0)
 515                                 break;
 516                 }
 517
 518                 if (!bytes_done)
 519                         continue;
 520
 521                 /*
 522                  * the rate is batched for now, it should work for batches
 523                  * of completions except the very first one which may look
 524                  * a little bursty
 525                  */
 526                 usec = utime_since(&s, &comp_time);
 527
 528                 rate_throttle(td, usec, bytes_done);
 529
 530                 if (check_min_rate(td, &comp_time)) {
 531                         if (exitall_on_terminate)
 532                                 terminate_threads(td->groupid);
 533                         td_verror(td, ENODATA, "check_min_rate");
 534                         break;
 535                 }
 536
 537                 if (td->o.thinktime) {
 538                         unsigned long long b;
 539
 540                         b = td->io_blocks[0] + td->io_blocks[1];
 541                         if (!(b % td->o.thinktime_blocks)) {
 542                                 int left;
 543
 544                                 if (td->o.thinktime_spin)
 545                                         __usec_sleep(td->o.thinktime_spin);
 546
 547                                 left = td->o.thinktime - td->o.thinktime_spin;
 548                                 if (left)
 549                                         usec_sleep(td, left);
 550                         }
 551                 }
 552         }
 553
 554         if (!td->error) {
 555                 struct fio_file *f;
 556
 557                 i = td->cur_depth;
 558                 if (i)
 559                         ret = io_u_queued_complete(td, i);
 560
 561                 if (should_fsync(td) && td->o.end_fsync) {
 562                         td_set_runstate(td, TD_FSYNCING);
 563
 564                         for_each_file(td, f, i) {
 565                                 if (!(f->flags & FIO_FILE_OPEN))
 566                                         continue;
 567                                 fio_io_sync(td, f);
 568                         }
 569                 }
 570         } else
 571                 cleanup_pending_aio(td);
 572 }
 573
 574 static void cleanup_io_u(struct thread_data *td)
 575 {
 576         struct list_head *entry, *n;
 577         struct io_u *io_u;
 578
 579         list_for_each_safe(entry, n, &td->io_u_freelist) {
 580                 io_u = list_entry(entry, struct io_u, list);
 581
 582                 list_del(&io_u->list);
 583                 free(io_u);
 584         }
 585
 586         free_io_mem(td);
 587 }
 588
 589 /*
 590  * "randomly" fill the buffer contents
 591  */
 592 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 593 {
 594         int *ptr = io_u->buf;
 595
 596         while ((void *) ptr - io_u->buf < max_bs) {
 597                 *ptr = rand() * 0x9e370001;
 598                 ptr++;
 599         }
 600 }
 601
 602 static int init_io_u(struct thread_data *td)
 603 {
 604         unsigned long long buf_size;
 605         struct io_u *io_u;
 606         unsigned int max_bs;
 607         int i, max_units;
 608         char *p;
 609
 610         if (td->io_ops->flags & FIO_SYNCIO)
 611                 max_units = 1;
 612         else
 613                 max_units = td->o.iodepth;
 614
 615         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 616         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 617         buf_size += page_mask;
 618         if (buf_size != (size_t) buf_size) {
 619                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 620                 return 1;
 621         }
 622
 623         td->orig_buffer_size = buf_size;
 624
 625         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 626                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 627         else if (td->orig_buffer_size & page_mask)
 628                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
 629
 630         if (allocate_io_mem(td))
 631                 return 1;
 632
 633         p = ALIGN(td->orig_buffer);
 634         for (i = 0; i < max_units; i++) {
 635                 io_u = malloc(sizeof(*io_u));
 636                 memset(io_u, 0, sizeof(*io_u));
 637                 INIT_LIST_HEAD(&io_u->list);
 638
 639                 io_u->buf = p + max_bs * i;
 640                 if (td_write(td) || td_rw(td))
 641                         fill_rand_buf(io_u, max_bs);
 642
 643                 io_u->index = i;
 644                 io_u->flags = IO_U_F_FREE;
 645                 list_add(&io_u->list, &td->io_u_freelist);
 646         }
 647
 648         io_u_init_timeout();
 649
 650         return 0;
 651 }
 652
 653 static int switch_ioscheduler(struct thread_data *td)
 654 {
 655         char tmp[256], tmp2[128];
 656         FILE *f;
 657         int ret;
 658
 659         if (td->io_ops->flags & FIO_DISKLESSIO)
 660                 return 0;
 661
 662         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 663
 664         f = fopen(tmp, "r+");
 665         if (!f) {
 666                 td_verror(td, errno, "fopen");
 667                 return 1;
 668         }
 669
 670         /*
 671          * Set io scheduler.
 672          */
 673         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 674         if (ferror(f) || ret != 1) {
 675                 td_verror(td, errno, "fwrite");
 676                 fclose(f);
 677                 return 1;
 678         }
 679
 680         rewind(f);
 681
 682         /*
 683          * Read back and check that the selected scheduler is now the default.
 684          */
 685         ret = fread(tmp, 1, sizeof(tmp), f);
 686         if (ferror(f) || ret < 0) {
 687                 td_verror(td, errno, "fread");
 688                 fclose(f);
 689                 return 1;
 690         }
 691
 692         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 693         if (!strstr(tmp, tmp2)) {
 694                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 695                 td_verror(td, EINVAL, "iosched_switch");
 696                 fclose(f);
 697                 return 1;
 698         }
 699
 700         fclose(f);
 701         return 0;
 702 }
 703
 704 static int clear_io_state(struct thread_data *td)
 705 {
 706         struct fio_file *f;
 707         unsigned int i;
 708         int ret;
 709
 710         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 711         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 712         td->zone_bytes = 0;
 713         td->rate_bytes = 0;
 714         td->rate_blocks = 0;
 715         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 716
 717         td->last_was_sync = 0;
 718
 719         for_each_file(td, f, i)
 720                 td_io_close_file(td, f);
 721
 722         ret = 0;
 723         for_each_file(td, f, i) {
 724                 ret = td_io_open_file(td, f);
 725                 if (ret)
 726                         break;
 727         }
 728
 729         return ret;
 730 }
 731
 732 /*
 733  * Entry point for the thread based jobs. The process based jobs end up
 734  * here as well, after a little setup.
 735  */
 736 static void *thread_main(void *data)
 737 {
 738         unsigned long long runtime[2];
 739         struct thread_data *td = data;
 740         unsigned long elapsed;
 741         int clear_state;
 742
 743         if (!td->o.use_thread)
 744                 setsid();
 745
 746         td->pid = getpid();
 747
 748         INIT_LIST_HEAD(&td->io_u_freelist);
 749         INIT_LIST_HEAD(&td->io_u_busylist);
 750         INIT_LIST_HEAD(&td->io_u_requeues);
 751         INIT_LIST_HEAD(&td->io_log_list);
 752         INIT_LIST_HEAD(&td->io_hist_list);
 753         td->io_hist_tree = RB_ROOT;
 754
 755         if (init_io_u(td))
 756                 goto err_sem;
 757
 758         if (fio_setaffinity(td) == -1) {
 759                 td_verror(td, errno, "cpu_set_affinity");
 760                 goto err_sem;
 761         }
 762
 763         if (init_iolog(td))
 764                 goto err_sem;
 765
 766         if (td->ioprio) {
 767                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 768                         td_verror(td, errno, "ioprio_set");
 769                         goto err_sem;
 770                 }
 771         }
 772
 773         if (nice(td->o.nice) == -1) {
 774                 td_verror(td, errno, "nice");
 775                 goto err_sem;
 776         }
 777
 778         if (td->o.ioscheduler && switch_ioscheduler(td))
 779                 goto err_sem;
 780
 781         td_set_runstate(td, TD_INITIALIZED);
 782         fio_sem_up(startup_sem);
 783         fio_sem_down(td->mutex);
 784
 785         /*
 786          * the ->mutex semaphore is now no longer used, close it to avoid
 787          * eating a file descriptor
 788          */
 789         fio_sem_remove(td->mutex);
 790
 791         if (!td->o.create_serialize && setup_files(td))
 792                 goto err;
 793
 794         if (td_io_init(td))
 795                 goto err;
 796
 797         if (open_files(td))
 798                 goto err;
 799
 800         if (init_random_map(td))
 801                 goto err;
 802
 803         if (td->o.exec_prerun) {
 804                 if (system(td->o.exec_prerun) < 0)
 805                         goto err;
 806         }
 807
 808         fio_gettime(&td->epoch, NULL);
 809         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 810         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 811
 812         runtime[0] = runtime[1] = 0;
 813         clear_state = 0;
 814         while (td->o.loops--) {
 815                 fio_gettime(&td->start, NULL);
 816                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 817
 818                 if (td->o.ratemin)
 819                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 820
 821                 if (clear_state && clear_io_state(td))
 822                         break;
 823
 824                 prune_io_piece_log(td);
 825
 826                 do_io(td);
 827
 828                 clear_state = 1;
 829
 830                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 831                         if (td->rw_end_set[DDIR_READ])
 832                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 833                         else
 834                                 elapsed = utime_since_now(&td->start);
 835
 836                         runtime[DDIR_READ] += elapsed;
 837                 }
 838                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 839                         if (td->rw_end_set[DDIR_WRITE])
 840                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 841                         else
 842                                 elapsed = utime_since_now(&td->start);
 843
 844                         runtime[DDIR_WRITE] += elapsed;
 845                 }
 846
 847                 if (td->error || td->terminate)
 848                         break;
 849
 850                 if (td->o.verify == VERIFY_NONE)
 851                         continue;
 852
 853                 if (clear_io_state(td))
 854                         break;
 855
 856                 fio_gettime(&td->start, NULL);
 857
 858                 do_verify(td);
 859
 860                 runtime[DDIR_READ] += utime_since_now(&td->start);
 861
 862                 if (td->error || td->terminate)
 863                         break;
 864         }
 865
 866         update_rusage_stat(td);
 867         td->ts.runtime[0] = runtime[0] / 1000;
 868         td->ts.runtime[1] = runtime[1] / 1000;
 869         td->ts.total_run_time = mtime_since_now(&td->epoch);
 870         td->ts.io_bytes[0] = td->io_bytes[0];
 871         td->ts.io_bytes[1] = td->io_bytes[1];
 872
 873         if (td->ts.bw_log)
 874                 finish_log(td, td->ts.bw_log, "bw");
 875         if (td->ts.slat_log)
 876                 finish_log(td, td->ts.slat_log, "slat");
 877         if (td->ts.clat_log)
 878                 finish_log(td, td->ts.clat_log, "clat");
 879         if (td->o.write_iolog_file)
 880                 write_iolog_close(td);
 881         if (td->o.exec_postrun) {
 882                 if (system(td->o.exec_postrun) < 0)
 883                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 884         }
 885
 886         if (exitall_on_terminate)
 887                 terminate_threads(td->groupid);
 888
 889 err:
 890         if (td->error)
 891                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 892         close_files(td);
 893         close_ioengine(td);
 894         cleanup_io_u(td);
 895         options_mem_free(td);
 896         td_set_runstate(td, TD_EXITED);
 897         return (void *) (unsigned long) td->error;
 898 err_sem:
 899         fio_sem_up(startup_sem);
 900         goto err;
 901 }
 902
 903 /*
 904  * We cannot pass the td data into a forked process, so attach the td and
 905  * pass it to the thread worker.
 906  */
 907 static int fork_main(int shmid, int offset)
 908 {
 909         struct thread_data *td;
 910         void *data, *ret;
 911
 912         data = shmat(shmid, NULL, 0);
 913         if (data == (void *) -1) {
 914                 int __err = errno;
 915
 916                 perror("shmat");
 917                 return __err;
 918         }
 919
 920         td = data + offset * sizeof(struct thread_data);
 921         ret = thread_main(td);
 922         shmdt(data);
 923         return (int) (unsigned long) ret;
 924 }
 925
 926 /*
 927  * Run over the job map and reap the threads that have exited, if any.
 928  */
 929 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 930 {
 931         struct thread_data *td;
 932         int i, cputhreads, pending, status, ret;
 933
 934         /*
 935          * reap exited threads (TD_EXITED -> TD_REAPED)
 936          */
 937         pending = cputhreads = 0;
 938         for_each_td(td, i) {
 939                 int flags = 0;
 940
 941                 /*
 942                  * ->io_ops is NULL for a thread that has closed its
 943                  * io engine
 944                  */
 945                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 946                         cputhreads++;
 947
 948                 if (!td->pid || td->runstate == TD_REAPED)
 949                         continue;
 950                 if (td->o.use_thread) {
 951                         if (td->runstate == TD_EXITED) {
 952                                 td_set_runstate(td, TD_REAPED);
 953                                 goto reaped;
 954                         }
 955                         continue;
 956                 }
 957
 958                 flags = WNOHANG;
 959                 if (td->runstate == TD_EXITED)
 960                         flags = 0;
 961
 962                 /*
 963                  * check if someone quit or got killed in an unusual way
 964                  */
 965                 ret = waitpid(td->pid, &status, flags);
 966                 if (ret < 0) {
 967                         if (errno == ECHILD) {
 968                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 969                                 td_set_runstate(td, TD_REAPED);
 970                                 goto reaped;
 971                         }
 972                         perror("waitpid");
 973                 } else if (ret == td->pid) {
 974                         if (WIFSIGNALED(status)) {
 975                                 int sig = WTERMSIG(status);
 976
 977                                 if (sig != SIGQUIT)
 978                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 979                                 td_set_runstate(td, TD_REAPED);
 980                                 goto reaped;
 981                         }
 982                         if (WIFEXITED(status)) {
 983                                 if (WEXITSTATUS(status) && !td->error)
 984                                         td->error = WEXITSTATUS(status);
 985
 986                                 td_set_runstate(td, TD_REAPED);
 987                                 goto reaped;
 988                         }
 989                 }
 990
 991                 /*
 992                  * thread is not dead, continue
 993                  */
 994                 continue;
 995 reaped:
 996                 if (td->o.use_thread) {
 997                         long ret;
 998
 999                         if (pthread_join(td->thread, (void *) &ret))
1000                                 perror("pthread_join");
1001                 }
1002
1003                 (*nr_running)--;
1004                 (*m_rate) -= td->o.ratemin;
1005                 (*t_rate) -= td->o.rate;
1006
1007                 if (td->error)
1008                         exit_value++;
1009         }
1010
1011         if (*nr_running == cputhreads && !pending)
1012                 terminate_threads(TERMINATE_ALL);
1013 }
1014
1015 /*
1016  * Main function for kicking off and reaping jobs, as needed.
1017  */
1018 static void run_threads(void)
1019 {
1020         struct thread_data *td;
1021         unsigned long spent;
1022         int i, todo, nr_running, m_rate, t_rate, nr_started;
1023
1024         if (fio_pin_memory())
1025                 return;
1026
1027         if (!terse_output) {
1028                 printf("Starting ");
1029                 if (nr_thread)
1030                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1031                 if (nr_process) {
1032                         if (nr_thread)
1033                                 printf(" and ");
1034                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1035                 }
1036                 printf("\n");
1037                 fflush(stdout);
1038         }
1039
1040         signal(SIGINT, sig_handler);
1041         signal(SIGALRM, sig_handler);
1042
1043         todo = thread_number;
1044         nr_running = 0;
1045         nr_started = 0;
1046         m_rate = t_rate = 0;
1047
1048         for_each_td(td, i) {
1049                 print_status_init(td->thread_number - 1);
1050
1051                 if (!td->o.create_serialize) {
1052                         init_disk_util(td);
1053                         continue;
1054                 }
1055
1056                 /*
1057                  * do file setup here so it happens sequentially,
1058                  * we don't want X number of threads getting their
1059                  * client data interspersed on disk
1060                  */
1061                 if (setup_files(td)) {
1062                         exit_value++;
1063                         if (td->error)
1064                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1065                         td_set_runstate(td, TD_REAPED);
1066                         todo--;
1067                 }
1068
1069                 init_disk_util(td);
1070         }
1071
1072         set_genesis_time();
1073
1074         while (todo) {
1075                 struct thread_data *map[MAX_JOBS];
1076                 struct timeval this_start;
1077                 int this_jobs = 0, left;
1078
1079                 /*
1080                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1081                  */
1082                 for_each_td(td, i) {
1083                         if (td->runstate != TD_NOT_CREATED)
1084                                 continue;
1085
1086                         /*
1087                          * never got a chance to start, killed by other
1088                          * thread for some reason
1089                          */
1090                         if (td->terminate) {
1091                                 todo--;
1092                                 continue;
1093                         }
1094
1095                         if (td->o.start_delay) {
1096                                 spent = mtime_since_genesis();
1097
1098                                 if (td->o.start_delay * 1000 > spent)
1099                                         continue;
1100                         }
1101
1102                         if (td->o.stonewall && (nr_started || nr_running))
1103                                 break;
1104
1105                         /*
1106                          * Set state to created. Thread will transition
1107                          * to TD_INITIALIZED when it's done setting up.
1108                          */
1109                         td_set_runstate(td, TD_CREATED);
1110                         map[this_jobs++] = td;
1111                         nr_started++;
1112
1113                         if (td->o.use_thread) {
1114                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1115                                         perror("thread_create");
1116                                         nr_started--;
1117                                         break;
1118                                 }
1119                         } else {
1120                                 if (!fork()) {
1121                                         int ret = fork_main(shm_id, i);
1122
1123                                         exit(ret);
1124                                 }
1125                         }
1126                         fio_sem_down(startup_sem);
1127                 }
1128
1129                 /*
1130                  * Wait for the started threads to transition to
1131                  * TD_INITIALIZED.
1132                  */
1133                 fio_gettime(&this_start, NULL);
1134                 left = this_jobs;
1135                 while (left && !fio_abort) {
1136                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1137                                 break;
1138
1139                         usleep(100000);
1140
1141                         for (i = 0; i < this_jobs; i++) {
1142                                 td = map[i];
1143                                 if (!td)
1144                                         continue;
1145                                 if (td->runstate == TD_INITIALIZED) {
1146                                         map[i] = NULL;
1147                                         left--;
1148                                 } else if (td->runstate >= TD_EXITED) {
1149                                         map[i] = NULL;
1150                                         left--;
1151                                         todo--;
1152                                         nr_running++; /* work-around... */
1153                                 }
1154                         }
1155                 }
1156
1157                 if (left) {
1158                         log_err("fio: %d jobs failed to start\n", left);
1159                         for (i = 0; i < this_jobs; i++) {
1160                                 td = map[i];
1161                                 if (!td)
1162                                         continue;
1163                                 kill(td->pid, SIGTERM);
1164                         }
1165                         break;
1166                 }
1167
1168                 /*
1169                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1170                  */
1171                 for_each_td(td, i) {
1172                         if (td->runstate != TD_INITIALIZED)
1173                                 continue;
1174
1175                         td_set_runstate(td, TD_RUNNING);
1176                         nr_running++;
1177                         nr_started--;
1178                         m_rate += td->o.ratemin;
1179                         t_rate += td->o.rate;
1180                         todo--;
1181                         fio_sem_up(td->mutex);
1182                 }
1183
1184                 reap_threads(&nr_running, &t_rate, &m_rate);
1185
1186                 if (todo)
1187                         usleep(100000);
1188         }
1189
1190         while (nr_running) {
1191                 reap_threads(&nr_running, &t_rate, &m_rate);
1192                 usleep(10000);
1193         }
1194
1195         update_io_ticks();
1196         fio_unpin_memory();
1197 }
1198
1199 int main(int argc, char *argv[])
1200 {
1201         long ps;
1202
1203         /*
1204          * We need locale for number printing, if it isn't set then just
1205          * go with the US format.
1206          */
1207         if (!getenv("LC_NUMERIC"))
1208                 setlocale(LC_NUMERIC, "en_US");
1209
1210         if (parse_options(argc, argv))
1211                 return 1;
1212
1213         if (!thread_number) {
1214                 log_err("Nothing to do\n");
1215                 return 1;
1216         }
1217
1218         ps = sysconf(_SC_PAGESIZE);
1219         if (ps < 0) {
1220                 log_err("Failed to get page size\n");
1221                 return 1;
1222         }
1223
1224         page_size = ps;
1225         page_mask = ps - 1;
1226
1227         if (write_bw_log) {
1228                 setup_log(&agg_io_log[DDIR_READ]);
1229                 setup_log(&agg_io_log[DDIR_WRITE]);
1230         }
1231
1232         startup_sem = fio_sem_init(0);
1233
1234         set_genesis_time();
1235
1236         disk_util_timer_arm();
1237
1238         run_threads();
1239
1240         if (!fio_abort) {
1241                 show_run_stats();
1242                 if (write_bw_log) {
1243                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1244                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1245                 }
1246         }
1247
1248         fio_sem_remove(startup_sem);
1249         return exit_value;
1250 }