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