fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38
  39 unsigned long page_mask;
  40 unsigned long page_size;
  41 #define ALIGN(buf)      \
  42         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  43
  44 int groupid = 0;
  45 int thread_number = 0;
  46 int nr_process = 0;
  47 int nr_thread = 0;
  48 int shm_id = 0;
  49 int temp_stall_ts;
  50
  51 static struct fio_sem *startup_sem;
  52 static volatile int fio_abort;
  53 static int exit_value;
  54
  55 struct io_log *agg_io_log[2];
  56
  57 #define TERMINATE_ALL           (-1)
  58 #define JOB_START_TIMEOUT       (5 * 1000)
  59
  60 static inline void td_set_runstate(struct thread_data *td, int runstate)
  61 {
  62         td->runstate = runstate;
  63 }
  64
  65 static void terminate_threads(int group_id)
  66 {
  67         struct thread_data *td;
  68         int i;
  69
  70         for_each_td(td, i) {
  71                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  72                         /*
  73                          * if the thread is running, just let it exit
  74                          */
  75                         if (td->runstate < TD_RUNNING)
  76                                 kill(td->pid, SIGQUIT);
  77                         td->terminate = 1;
  78                         td->o.start_delay = 0;
  79                 }
  80         }
  81 }
  82
  83 static void sig_handler(int sig)
  84 {
  85         switch (sig) {
  86                 case SIGALRM:
  87                         update_io_ticks();
  88                         disk_util_timer_arm();
  89                         print_thread_status();
  90                         break;
  91                 default:
  92                         printf("\nfio: terminating on signal %d\n", sig);
  93                         fflush(stdout);
  94                         terminate_threads(TERMINATE_ALL);
  95                         break;
  96         }
  97 }
  98
  99 /*
 100  * Check if we are above the minimum rate given.
 101  */
 102 static int check_min_rate(struct thread_data *td, struct timeval *now)
 103 {
 104         unsigned long long bytes = 0;
 105         unsigned long iops = 0;
 106         unsigned long spent;
 107         unsigned long rate;
 108
 109         /*
 110          * No minimum rate set, always ok
 111          */
 112         if (!td->o.ratemin && !td->o.rate_iops_min)
 113                 return 0;
 114
 115         /*
 116          * allow a 2 second settle period in the beginning
 117          */
 118         if (mtime_since(&td->start, now) < 2000)
 119                 return 0;
 120
 121         if (td_read(td)) {
 122                 iops += td->io_blocks[DDIR_READ];
 123                 bytes += td->this_io_bytes[DDIR_READ];
 124         }
 125         if (td_write(td)) {
 126                 iops += td->io_blocks[DDIR_WRITE];
 127                 bytes += td->this_io_bytes[DDIR_WRITE];
 128         }
 129
 130         /*
 131          * if rate blocks is set, sample is running
 132          */
 133         if (td->rate_bytes || td->rate_blocks) {
 134                 spent = mtime_since(&td->lastrate, now);
 135                 if (spent < td->o.ratecycle)
 136                         return 0;
 137
 138                 if (td->o.rate) {
 139                         /*
 140                          * check bandwidth specified rate
 141                          */
 142                         if (bytes < td->rate_bytes) {
 143                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
 144                                 return 1;
 145                         } else {
 146                                 rate = (bytes - td->rate_bytes) / spent;
 147                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
 148                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
 149                                         return 1;
 150                                 }
 151                         }
 152                 } else {
 153                         /*
 154                          * checks iops specified rate
 155                          */
 156                         if (iops < td->o.rate_iops) {
 157                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
 158                                 return 1;
 159                         } else {
 160                                 rate = (iops - td->rate_blocks) / spent;
 161                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
 162                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
 163                                 }
 164                         }
 165                 }
 166         }
 167
 168         td->rate_bytes = bytes;
 169         td->rate_blocks = iops;
 170         memcpy(&td->lastrate, now, sizeof(*now));
 171         return 0;
 172 }
 173
 174 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 175 {
 176         if (!td->o.timeout)
 177                 return 0;
 178         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 179                 return 1;
 180
 181         return 0;
 182 }
 183
 184 /*
 185  * When job exits, we can cancel the in-flight IO if we are using async
 186  * io. Attempt to do so.
 187  */
 188 static void cleanup_pending_aio(struct thread_data *td)
 189 {
 190         struct list_head *entry, *n;
 191         struct io_u *io_u;
 192         int r;
 193
 194         /*
 195          * get immediately available events, if any
 196          */
 197         r = io_u_queued_complete(td, 0);
 198         if (r < 0)
 199                 return;
 200
 201         /*
 202          * now cancel remaining active events
 203          */
 204         if (td->io_ops->cancel) {
 205                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 206                         io_u = list_entry(entry, struct io_u, list);
 207
 208                         /*
 209                          * if the io_u isn't in flight, then that generally
 210                          * means someone leaked an io_u. complain but fix
 211                          * it up, so we don't stall here.
 212                          */
 213                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 214                                 log_err("fio: non-busy IO on busy list\n");
 215                                 put_io_u(td, io_u);
 216                         } else {
 217                                 r = td->io_ops->cancel(td, io_u);
 218                                 if (!r)
 219                                         put_io_u(td, io_u);
 220                         }
 221                 }
 222         }
 223
 224         if (td->cur_depth)
 225                 r = io_u_queued_complete(td, td->cur_depth);
 226 }
 227
 228 /*
 229  * Helper to handle the final sync of a file. Works just like the normal
 230  * io path, just does everything sync.
 231  */
 232 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 233 {
 234         struct io_u *io_u = __get_io_u(td);
 235         int ret;
 236
 237         if (!io_u)
 238                 return 1;
 239
 240         io_u->ddir = DDIR_SYNC;
 241         io_u->file = f;
 242
 243         if (td_io_prep(td, io_u)) {
 244                 put_io_u(td, io_u);
 245                 return 1;
 246         }
 247
 248 requeue:
 249         ret = td_io_queue(td, io_u);
 250         if (ret < 0) {
 251                 td_verror(td, io_u->error, "td_io_queue");
 252                 put_io_u(td, io_u);
 253                 return 1;
 254         } else if (ret == FIO_Q_QUEUED) {
 255                 if (io_u_queued_complete(td, 1) < 0)
 256                         return 1;
 257         } else if (ret == FIO_Q_COMPLETED) {
 258                 if (io_u->error) {
 259                         td_verror(td, io_u->error, "td_io_queue");
 260                         return 1;
 261                 }
 262
 263                 if (io_u_sync_complete(td, io_u) < 0)
 264                         return 1;
 265         } else if (ret == FIO_Q_BUSY) {
 266                 if (td_io_commit(td))
 267                         return 1;
 268                 goto requeue;
 269         }
 270
 271         return 0;
 272 }
 273
 274 /*
 275  * The main verify engine. Runs over the writes we previously submitted,
 276  * reads the blocks back in, and checks the crc/md5 of the data.
 277  */
 278 static void do_verify(struct thread_data *td)
 279 {
 280         struct fio_file *f;
 281         struct io_u *io_u;
 282         int ret, min_events;
 283         unsigned int i;
 284
 285         /*
 286          * sync io first and invalidate cache, to make sure we really
 287          * read from disk.
 288          */
 289         for_each_file(td, f, i) {
 290                 if (!(f->flags & FIO_FILE_OPEN))
 291                         continue;
 292                 if (fio_io_sync(td, f))
 293                         break;
 294                 if (file_invalidate_cache(td, f))
 295                         break;
 296         }
 297
 298         if (td->error)
 299                 return;
 300
 301         td_set_runstate(td, TD_VERIFYING);
 302
 303         io_u = NULL;
 304         while (!td->terminate) {
 305                 int ret2;
 306
 307                 io_u = __get_io_u(td);
 308                 if (!io_u)
 309                         break;
 310
 311                 if (runtime_exceeded(td, &io_u->start_time)) {
 312                         put_io_u(td, io_u);
 313                         break;
 314                 }
 315
 316                 if (get_next_verify(td, io_u)) {
 317                         put_io_u(td, io_u);
 318                         break;
 319                 }
 320
 321                 if (td_io_prep(td, io_u)) {
 322                         put_io_u(td, io_u);
 323                         break;
 324                 }
 325
 326                 io_u->end_io = verify_io_u;
 327
 328                 ret = td_io_queue(td, io_u);
 329                 switch (ret) {
 330                 case FIO_Q_COMPLETED:
 331                         if (io_u->error)
 332                                 ret = -io_u->error;
 333                         else if (io_u->resid) {
 334                                 int bytes = io_u->xfer_buflen - io_u->resid;
 335                                 struct fio_file *f = io_u->file;
 336
 337                                 /*
 338                                  * zero read, fail
 339                                  */
 340                                 if (!bytes) {
 341                                         td_verror(td, ENODATA, "full resid");
 342                                         put_io_u(td, io_u);
 343                                         break;
 344                                 }
 345
 346                                 io_u->xfer_buflen = io_u->resid;
 347                                 io_u->xfer_buf += bytes;
 348                                 io_u->offset += bytes;
 349                                 f->last_completed_pos = io_u->offset;
 350
 351                                 td->ts.short_io_u[io_u->ddir]++;
 352
 353                                 if (io_u->offset == f->real_file_size)
 354                                         goto sync_done;
 355
 356                                 requeue_io_u(td, &io_u);
 357                         } else {
 358 sync_done:
 359                                 ret = io_u_sync_complete(td, io_u);
 360                                 if (ret < 0)
 361                                         break;
 362                         }
 363                         continue;
 364                 case FIO_Q_QUEUED:
 365                         break;
 366                 case FIO_Q_BUSY:
 367                         requeue_io_u(td, &io_u);
 368                         ret2 = td_io_commit(td);
 369                         if (ret2 < 0)
 370                                 ret = ret2;
 371                         break;
 372                 default:
 373                         assert(ret < 0);
 374                         td_verror(td, -ret, "td_io_queue");
 375                         break;
 376                 }
 377
 378                 if (ret < 0 || td->error)
 379                         break;
 380
 381                 /*
 382                  * if we can queue more, do so. but check if there are
 383                  * completed io_u's first.
 384                  */
 385                 min_events = 0;
 386                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 387                         min_events = 1;
 388
 389                         if (td->cur_depth > td->o.iodepth_low)
 390                                 min_events = td->cur_depth - td->o.iodepth_low;
 391                 }
 392
 393                 /*
 394                  * Reap required number of io units, if any, and do the
 395                  * verification on them through the callback handler
 396                  */
 397                 if (io_u_queued_complete(td, min_events) < 0)
 398                         break;
 399         }
 400
 401         if (!td->error) {
 402                 min_events = td->cur_depth;
 403
 404                 if (min_events)
 405                         ret = io_u_queued_complete(td, min_events);
 406         } else
 407                 cleanup_pending_aio(td);
 408
 409         td_set_runstate(td, TD_RUNNING);
 410 }
 411
 412 /*
 413  * Main IO worker function. It retrieves io_u's to process and queues
 414  * and reaps them, checking for rate and errors along the way.
 415  */
 416 static void do_io(struct thread_data *td)
 417 {
 418         struct timeval s;
 419         unsigned long usec;
 420         unsigned int i;
 421         int ret = 0;
 422
 423         td_set_runstate(td, TD_RUNNING);
 424
 425         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 426                 struct timeval comp_time;
 427                 long bytes_done = 0;
 428                 int min_evts = 0;
 429                 struct io_u *io_u;
 430                 int ret2;
 431
 432                 if (td->terminate)
 433                         break;
 434
 435                 io_u = get_io_u(td);
 436                 if (!io_u)
 437                         break;
 438
 439                 memcpy(&s, &io_u->start_time, sizeof(s));
 440
 441                 if (runtime_exceeded(td, &s)) {
 442                         put_io_u(td, io_u);
 443                         break;
 444                 }
 445
 446                 ret = td_io_queue(td, io_u);
 447                 switch (ret) {
 448                 case FIO_Q_COMPLETED:
 449                         if (io_u->error)
 450                                 ret = -io_u->error;
 451                         else if (io_u->resid) {
 452                                 int bytes = io_u->xfer_buflen - io_u->resid;
 453                                 struct fio_file *f = io_u->file;
 454
 455                                 /*
 456                                  * zero read, fail
 457                                  */
 458                                 if (!bytes) {
 459                                         td_verror(td, ENODATA, "full resid");
 460                                         put_io_u(td, io_u);
 461                                         break;
 462                                 }
 463
 464                                 io_u->xfer_buflen = io_u->resid;
 465                                 io_u->xfer_buf += bytes;
 466                                 io_u->offset += bytes;
 467                                 f->last_completed_pos = io_u->offset;
 468
 469                                 td->ts.short_io_u[io_u->ddir]++;
 470
 471                                 if (io_u->offset == f->real_file_size)
 472                                         goto sync_done;
 473
 474                                 requeue_io_u(td, &io_u);
 475                         } else {
 476 sync_done:
 477                                 fio_gettime(&comp_time, NULL);
 478                                 bytes_done = io_u_sync_complete(td, io_u);
 479                                 if (bytes_done < 0)
 480                                         ret = bytes_done;
 481                         }
 482                         break;
 483                 case FIO_Q_QUEUED:
 484                         /*
 485                          * if the engine doesn't have a commit hook,
 486                          * the io_u is really queued. if it does have such
 487                          * a hook, it has to call io_u_queued() itself.
 488                          */
 489                         if (td->io_ops->commit == NULL)
 490                                 io_u_queued(td, io_u);
 491                         break;
 492                 case FIO_Q_BUSY:
 493                         requeue_io_u(td, &io_u);
 494                         ret2 = td_io_commit(td);
 495                         if (ret2 < 0)
 496                                 ret = ret2;
 497                         break;
 498                 default:
 499                         assert(ret < 0);
 500                         put_io_u(td, io_u);
 501                         break;
 502                 }
 503
 504                 if (ret < 0 || td->error)
 505                         break;
 506
 507                 /*
 508                  * See if we need to complete some commands
 509                  */
 510                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 511                         min_evts = 0;
 512                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 513                                 min_evts = 1;
 514
 515                                 if (td->cur_depth > td->o.iodepth_low)
 516                                         min_evts = td->cur_depth - td->o.iodepth_low;
 517                         }
 518
 519                         fio_gettime(&comp_time, NULL);
 520                         bytes_done = io_u_queued_complete(td, min_evts);
 521                         if (bytes_done < 0)
 522                                 break;
 523                 }
 524
 525                 if (!bytes_done)
 526                         continue;
 527
 528                 /*
 529                  * the rate is batched for now, it should work for batches
 530                  * of completions except the very first one which may look
 531                  * a little bursty
 532                  */
 533                 usec = utime_since(&s, &comp_time);
 534
 535                 rate_throttle(td, usec, bytes_done);
 536
 537                 if (check_min_rate(td, &comp_time)) {
 538                         if (exitall_on_terminate)
 539                                 terminate_threads(td->groupid);
 540                         td_verror(td, ENODATA, "check_min_rate");
 541                         break;
 542                 }
 543
 544                 if (td->o.thinktime) {
 545                         unsigned long long b;
 546
 547                         b = td->io_blocks[0] + td->io_blocks[1];
 548                         if (!(b % td->o.thinktime_blocks)) {
 549                                 int left;
 550
 551                                 if (td->o.thinktime_spin)
 552                                         __usec_sleep(td->o.thinktime_spin);
 553
 554                                 left = td->o.thinktime - td->o.thinktime_spin;
 555                                 if (left)
 556                                         usec_sleep(td, left);
 557                         }
 558                 }
 559         }
 560
 561         if (!td->error) {
 562                 struct fio_file *f;
 563
 564                 i = td->cur_depth;
 565                 if (i)
 566                         ret = io_u_queued_complete(td, i);
 567
 568                 if (should_fsync(td) && td->o.end_fsync) {
 569                         td_set_runstate(td, TD_FSYNCING);
 570
 571                         for_each_file(td, f, i) {
 572                                 if (!(f->flags & FIO_FILE_OPEN))
 573                                         continue;
 574                                 fio_io_sync(td, f);
 575                         }
 576                 }
 577         } else
 578                 cleanup_pending_aio(td);
 579 }
 580
 581 static void cleanup_io_u(struct thread_data *td)
 582 {
 583         struct list_head *entry, *n;
 584         struct io_u *io_u;
 585
 586         list_for_each_safe(entry, n, &td->io_u_freelist) {
 587                 io_u = list_entry(entry, struct io_u, list);
 588
 589                 list_del(&io_u->list);
 590                 free(io_u);
 591         }
 592
 593         free_io_mem(td);
 594 }
 595
 596 /*
 597  * "randomly" fill the buffer contents
 598  */
 599 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 600 {
 601         int *ptr = io_u->buf;
 602
 603         while ((void *) ptr - io_u->buf < max_bs) {
 604                 *ptr = rand() * 0x9e370001;
 605                 ptr++;
 606         }
 607 }
 608
 609 static int init_io_u(struct thread_data *td)
 610 {
 611         unsigned long long buf_size;
 612         struct io_u *io_u;
 613         unsigned int max_bs;
 614         int i, max_units;
 615         char *p;
 616
 617         if (td->io_ops->flags & FIO_SYNCIO)
 618                 max_units = 1;
 619         else
 620                 max_units = td->o.iodepth;
 621
 622         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 623         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 624         buf_size += page_mask;
 625         if (buf_size != (size_t) buf_size) {
 626                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 627                 return 1;
 628         }
 629
 630         td->orig_buffer_size = buf_size;
 631
 632         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 633                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 634         else if (td->orig_buffer_size & page_mask)
 635                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
 636
 637         if (allocate_io_mem(td))
 638                 return 1;
 639
 640         p = ALIGN(td->orig_buffer);
 641         for (i = 0; i < max_units; i++) {
 642                 io_u = malloc(sizeof(*io_u));
 643                 memset(io_u, 0, sizeof(*io_u));
 644                 INIT_LIST_HEAD(&io_u->list);
 645
 646                 io_u->buf = p + max_bs * i;
 647                 if (td_write(td) || td_rw(td))
 648                         fill_rand_buf(io_u, max_bs);
 649
 650                 io_u->index = i;
 651                 io_u->flags = IO_U_F_FREE;
 652                 list_add(&io_u->list, &td->io_u_freelist);
 653         }
 654
 655         io_u_init_timeout();
 656
 657         return 0;
 658 }
 659
 660 static int switch_ioscheduler(struct thread_data *td)
 661 {
 662         char tmp[256], tmp2[128];
 663         FILE *f;
 664         int ret;
 665
 666         if (td->io_ops->flags & FIO_DISKLESSIO)
 667                 return 0;
 668
 669         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 670
 671         f = fopen(tmp, "r+");
 672         if (!f) {
 673                 if (errno == ENOENT) {
 674                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 675                         return 0;
 676                 }
 677                 td_verror(td, errno, "fopen iosched");
 678                 return 1;
 679         }
 680
 681         /*
 682          * Set io scheduler.
 683          */
 684         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 685         if (ferror(f) || ret != 1) {
 686                 td_verror(td, errno, "fwrite");
 687                 fclose(f);
 688                 return 1;
 689         }
 690
 691         rewind(f);
 692
 693         /*
 694          * Read back and check that the selected scheduler is now the default.
 695          */
 696         ret = fread(tmp, 1, sizeof(tmp), f);
 697         if (ferror(f) || ret < 0) {
 698                 td_verror(td, errno, "fread");
 699                 fclose(f);
 700                 return 1;
 701         }
 702
 703         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 704         if (!strstr(tmp, tmp2)) {
 705                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 706                 td_verror(td, EINVAL, "iosched_switch");
 707                 fclose(f);
 708                 return 1;
 709         }
 710
 711         fclose(f);
 712         return 0;
 713 }
 714
 715 static int clear_io_state(struct thread_data *td)
 716 {
 717         struct fio_file *f;
 718         unsigned int i;
 719         int ret;
 720
 721         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 722         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 723         td->zone_bytes = 0;
 724         td->rate_bytes = 0;
 725         td->rate_blocks = 0;
 726         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 727
 728         td->last_was_sync = 0;
 729
 730         for_each_file(td, f, i)
 731                 td_io_close_file(td, f);
 732
 733         ret = 0;
 734         for_each_file(td, f, i) {
 735                 ret = td_io_open_file(td, f);
 736                 if (ret)
 737                         break;
 738         }
 739
 740         return ret;
 741 }
 742
 743 /*
 744  * Entry point for the thread based jobs. The process based jobs end up
 745  * here as well, after a little setup.
 746  */
 747 static void *thread_main(void *data)
 748 {
 749         unsigned long long runtime[2];
 750         struct thread_data *td = data;
 751         unsigned long elapsed;
 752         int clear_state;
 753
 754         if (!td->o.use_thread)
 755                 setsid();
 756
 757         td->pid = getpid();
 758
 759         INIT_LIST_HEAD(&td->io_u_freelist);
 760         INIT_LIST_HEAD(&td->io_u_busylist);
 761         INIT_LIST_HEAD(&td->io_u_requeues);
 762         INIT_LIST_HEAD(&td->io_log_list);
 763         INIT_LIST_HEAD(&td->io_hist_list);
 764         td->io_hist_tree = RB_ROOT;
 765
 766         if (init_io_u(td))
 767                 goto err_sem;
 768
 769         if (fio_setaffinity(td) == -1) {
 770                 td_verror(td, errno, "cpu_set_affinity");
 771                 goto err_sem;
 772         }
 773
 774         if (init_iolog(td))
 775                 goto err_sem;
 776
 777         if (td->ioprio) {
 778                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 779                         td_verror(td, errno, "ioprio_set");
 780                         goto err_sem;
 781                 }
 782         }
 783
 784         if (nice(td->o.nice) == -1) {
 785                 td_verror(td, errno, "nice");
 786                 goto err_sem;
 787         }
 788
 789         if (td->o.ioscheduler && switch_ioscheduler(td))
 790                 goto err_sem;
 791
 792         td_set_runstate(td, TD_INITIALIZED);
 793         fio_sem_up(startup_sem);
 794         fio_sem_down(td->mutex);
 795
 796         /*
 797          * the ->mutex semaphore is now no longer used, close it to avoid
 798          * eating a file descriptor
 799          */
 800         fio_sem_remove(td->mutex);
 801
 802         if (!td->o.create_serialize && setup_files(td))
 803                 goto err;
 804
 805         if (td_io_init(td))
 806                 goto err;
 807
 808         if (open_files(td))
 809                 goto err;
 810
 811         if (init_random_map(td))
 812                 goto err;
 813
 814         if (td->o.exec_prerun) {
 815                 if (system(td->o.exec_prerun) < 0)
 816                         goto err;
 817         }
 818
 819         fio_gettime(&td->epoch, NULL);
 820         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 821         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 822
 823         runtime[0] = runtime[1] = 0;
 824         clear_state = 0;
 825         while (td->o.loops--) {
 826                 fio_gettime(&td->start, NULL);
 827                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 828
 829                 if (td->o.ratemin)
 830                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 831
 832                 if (clear_state && clear_io_state(td))
 833                         break;
 834
 835                 prune_io_piece_log(td);
 836
 837                 do_io(td);
 838
 839                 clear_state = 1;
 840
 841                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 842                         if (td->rw_end_set[DDIR_READ])
 843                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 844                         else
 845                                 elapsed = utime_since_now(&td->start);
 846
 847                         runtime[DDIR_READ] += elapsed;
 848                 }
 849                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 850                         if (td->rw_end_set[DDIR_WRITE])
 851                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 852                         else
 853                                 elapsed = utime_since_now(&td->start);
 854
 855                         runtime[DDIR_WRITE] += elapsed;
 856                 }
 857
 858                 if (td->error || td->terminate)
 859                         break;
 860
 861                 if (td->o.verify == VERIFY_NONE)
 862                         continue;
 863
 864                 if (clear_io_state(td))
 865                         break;
 866
 867                 fio_gettime(&td->start, NULL);
 868
 869                 do_verify(td);
 870
 871                 runtime[DDIR_READ] += utime_since_now(&td->start);
 872
 873                 if (td->error || td->terminate)
 874                         break;
 875         }
 876
 877         update_rusage_stat(td);
 878         td->ts.runtime[0] = runtime[0] / 1000;
 879         td->ts.runtime[1] = runtime[1] / 1000;
 880         td->ts.total_run_time = mtime_since_now(&td->epoch);
 881         td->ts.io_bytes[0] = td->io_bytes[0];
 882         td->ts.io_bytes[1] = td->io_bytes[1];
 883
 884         if (td->ts.bw_log)
 885                 finish_log(td, td->ts.bw_log, "bw");
 886         if (td->ts.slat_log)
 887                 finish_log(td, td->ts.slat_log, "slat");
 888         if (td->ts.clat_log)
 889                 finish_log(td, td->ts.clat_log, "clat");
 890         if (td->o.write_iolog_file)
 891                 write_iolog_close(td);
 892         if (td->o.exec_postrun) {
 893                 if (system(td->o.exec_postrun) < 0)
 894                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 895         }
 896
 897         if (exitall_on_terminate)
 898                 terminate_threads(td->groupid);
 899
 900 err:
 901         if (td->error)
 902                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 903         close_files(td);
 904         close_ioengine(td);
 905         cleanup_io_u(td);
 906         options_mem_free(td);
 907         td_set_runstate(td, TD_EXITED);
 908         return (void *) (unsigned long) td->error;
 909 err_sem:
 910         fio_sem_up(startup_sem);
 911         goto err;
 912 }
 913
 914 /*
 915  * We cannot pass the td data into a forked process, so attach the td and
 916  * pass it to the thread worker.
 917  */
 918 static int fork_main(int shmid, int offset)
 919 {
 920         struct thread_data *td;
 921         void *data, *ret;
 922
 923         data = shmat(shmid, NULL, 0);
 924         if (data == (void *) -1) {
 925                 int __err = errno;
 926
 927                 perror("shmat");
 928                 return __err;
 929         }
 930
 931         td = data + offset * sizeof(struct thread_data);
 932         ret = thread_main(td);
 933         shmdt(data);
 934         return (int) (unsigned long) ret;
 935 }
 936
 937 /*
 938  * Run over the job map and reap the threads that have exited, if any.
 939  */
 940 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 941 {
 942         struct thread_data *td;
 943         int i, cputhreads, pending, status, ret;
 944
 945         /*
 946          * reap exited threads (TD_EXITED -> TD_REAPED)
 947          */
 948         pending = cputhreads = 0;
 949         for_each_td(td, i) {
 950                 int flags = 0;
 951
 952                 /*
 953                  * ->io_ops is NULL for a thread that has closed its
 954                  * io engine
 955                  */
 956                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 957                         cputhreads++;
 958
 959                 if (!td->pid || td->runstate == TD_REAPED)
 960                         continue;
 961                 if (td->o.use_thread) {
 962                         if (td->runstate == TD_EXITED) {
 963                                 td_set_runstate(td, TD_REAPED);
 964                                 goto reaped;
 965                         }
 966                         continue;
 967                 }
 968
 969                 flags = WNOHANG;
 970                 if (td->runstate == TD_EXITED)
 971                         flags = 0;
 972
 973                 /*
 974                  * check if someone quit or got killed in an unusual way
 975                  */
 976                 ret = waitpid(td->pid, &status, flags);
 977                 if (ret < 0) {
 978                         if (errno == ECHILD) {
 979                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 980                                 td_set_runstate(td, TD_REAPED);
 981                                 goto reaped;
 982                         }
 983                         perror("waitpid");
 984                 } else if (ret == td->pid) {
 985                         if (WIFSIGNALED(status)) {
 986                                 int sig = WTERMSIG(status);
 987
 988                                 if (sig != SIGQUIT)
 989                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 990                                 td_set_runstate(td, TD_REAPED);
 991                                 goto reaped;
 992                         }
 993                         if (WIFEXITED(status)) {
 994                                 if (WEXITSTATUS(status) && !td->error)
 995                                         td->error = WEXITSTATUS(status);
 996
 997                                 td_set_runstate(td, TD_REAPED);
 998                                 goto reaped;
 999                         }
1000                 }
1001
1002                 /*
1003                  * thread is not dead, continue
1004                  */
1005                 pending++;
1006                 continue;
1007 reaped:
1008                 if (td->o.use_thread) {
1009                         long ret;
1010
1011                         if (pthread_join(td->thread, (void *) &ret))
1012                                 perror("pthread_join");
1013                 }
1014
1015                 (*nr_running)--;
1016                 (*m_rate) -= td->o.ratemin;
1017                 (*t_rate) -= td->o.rate;
1018                 pending--;
1019
1020                 if (td->error)
1021                         exit_value++;
1022         }
1023
1024         if (*nr_running == cputhreads && !pending)
1025                 terminate_threads(TERMINATE_ALL);
1026 }
1027
1028 /*
1029  * Main function for kicking off and reaping jobs, as needed.
1030  */
1031 static void run_threads(void)
1032 {
1033         struct thread_data *td;
1034         unsigned long spent;
1035         int i, todo, nr_running, m_rate, t_rate, nr_started;
1036
1037         if (fio_pin_memory())
1038                 return;
1039
1040         if (!terse_output) {
1041                 printf("Starting ");
1042                 if (nr_thread)
1043                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1044                 if (nr_process) {
1045                         if (nr_thread)
1046                                 printf(" and ");
1047                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1048                 }
1049                 printf("\n");
1050                 fflush(stdout);
1051         }
1052
1053         signal(SIGINT, sig_handler);
1054         signal(SIGALRM, sig_handler);
1055
1056         todo = thread_number;
1057         nr_running = 0;
1058         nr_started = 0;
1059         m_rate = t_rate = 0;
1060
1061         for_each_td(td, i) {
1062                 print_status_init(td->thread_number - 1);
1063
1064                 if (!td->o.create_serialize) {
1065                         init_disk_util(td);
1066                         continue;
1067                 }
1068
1069                 /*
1070                  * do file setup here so it happens sequentially,
1071                  * we don't want X number of threads getting their
1072                  * client data interspersed on disk
1073                  */
1074                 if (setup_files(td)) {
1075                         exit_value++;
1076                         if (td->error)
1077                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1078                         td_set_runstate(td, TD_REAPED);
1079                         todo--;
1080                 }
1081
1082                 init_disk_util(td);
1083         }
1084
1085         set_genesis_time();
1086
1087         while (todo) {
1088                 struct thread_data *map[MAX_JOBS];
1089                 struct timeval this_start;
1090                 int this_jobs = 0, left;
1091
1092                 /*
1093                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1094                  */
1095                 for_each_td(td, i) {
1096                         if (td->runstate != TD_NOT_CREATED)
1097                                 continue;
1098
1099                         /*
1100                          * never got a chance to start, killed by other
1101                          * thread for some reason
1102                          */
1103                         if (td->terminate) {
1104                                 todo--;
1105                                 continue;
1106                         }
1107
1108                         if (td->o.start_delay) {
1109                                 spent = mtime_since_genesis();
1110
1111                                 if (td->o.start_delay * 1000 > spent)
1112                                         continue;
1113                         }
1114
1115                         if (td->o.stonewall && (nr_started || nr_running))
1116                                 break;
1117
1118                         /*
1119                          * Set state to created. Thread will transition
1120                          * to TD_INITIALIZED when it's done setting up.
1121                          */
1122                         td_set_runstate(td, TD_CREATED);
1123                         map[this_jobs++] = td;
1124                         nr_started++;
1125
1126                         if (td->o.use_thread) {
1127                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1128                                         perror("thread_create");
1129                                         nr_started--;
1130                                         break;
1131                                 }
1132                         } else {
1133                                 if (!fork()) {
1134                                         int ret = fork_main(shm_id, i);
1135
1136                                         exit(ret);
1137                                 }
1138                         }
1139                         fio_sem_down(startup_sem);
1140                 }
1141
1142                 /*
1143                  * Wait for the started threads to transition to
1144                  * TD_INITIALIZED.
1145                  */
1146                 fio_gettime(&this_start, NULL);
1147                 left = this_jobs;
1148                 while (left && !fio_abort) {
1149                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1150                                 break;
1151
1152                         usleep(100000);
1153
1154                         for (i = 0; i < this_jobs; i++) {
1155                                 td = map[i];
1156                                 if (!td)
1157                                         continue;
1158                                 if (td->runstate == TD_INITIALIZED) {
1159                                         map[i] = NULL;
1160                                         left--;
1161                                 } else if (td->runstate >= TD_EXITED) {
1162                                         map[i] = NULL;
1163                                         left--;
1164                                         todo--;
1165                                         nr_running++; /* work-around... */
1166                                 }
1167                         }
1168                 }
1169
1170                 if (left) {
1171                         log_err("fio: %d jobs failed to start\n", left);
1172                         for (i = 0; i < this_jobs; i++) {
1173                                 td = map[i];
1174                                 if (!td)
1175                                         continue;
1176                                 kill(td->pid, SIGTERM);
1177                         }
1178                         break;
1179                 }
1180
1181                 /*
1182                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1183                  */
1184                 for_each_td(td, i) {
1185                         if (td->runstate != TD_INITIALIZED)
1186                                 continue;
1187
1188                         td_set_runstate(td, TD_RUNNING);
1189                         nr_running++;
1190                         nr_started--;
1191                         m_rate += td->o.ratemin;
1192                         t_rate += td->o.rate;
1193                         todo--;
1194                         fio_sem_up(td->mutex);
1195                 }
1196
1197                 reap_threads(&nr_running, &t_rate, &m_rate);
1198
1199                 if (todo)
1200                         usleep(100000);
1201         }
1202
1203         while (nr_running) {
1204                 reap_threads(&nr_running, &t_rate, &m_rate);
1205                 usleep(10000);
1206         }
1207
1208         update_io_ticks();
1209         fio_unpin_memory();
1210 }
1211
1212 int main(int argc, char *argv[])
1213 {
1214         long ps;
1215
1216         /*
1217          * We need locale for number printing, if it isn't set then just
1218          * go with the US format.
1219          */
1220         if (!getenv("LC_NUMERIC"))
1221                 setlocale(LC_NUMERIC, "en_US");
1222
1223         if (parse_options(argc, argv))
1224                 return 1;
1225
1226         if (!thread_number)
1227                 return 0;
1228
1229         ps = sysconf(_SC_PAGESIZE);
1230         if (ps < 0) {
1231                 log_err("Failed to get page size\n");
1232                 return 1;
1233         }
1234
1235         page_size = ps;
1236         page_mask = ps - 1;
1237
1238         if (write_bw_log) {
1239                 setup_log(&agg_io_log[DDIR_READ]);
1240                 setup_log(&agg_io_log[DDIR_WRITE]);
1241         }
1242
1243         startup_sem = fio_sem_init(0);
1244
1245         set_genesis_time();
1246
1247         disk_util_timer_arm();
1248
1249         run_threads();
1250
1251         if (!fio_abort) {
1252                 show_run_stats();
1253                 if (write_bw_log) {
1254                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1255                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1256                 }
1257         }
1258
1259         fio_sem_remove(startup_sem);
1260         return exit_value;
1261 }