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 "hash.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                         td->terminate = 1;
 315                         break;
 316                 }
 317
 318                 if (get_next_verify(td, io_u)) {
 319                         put_io_u(td, io_u);
 320                         break;
 321                 }
 322
 323                 if (td_io_prep(td, io_u)) {
 324                         put_io_u(td, io_u);
 325                         break;
 326                 }
 327
 328                 io_u->end_io = verify_io_u;
 329
 330                 ret = td_io_queue(td, io_u);
 331                 switch (ret) {
 332                 case FIO_Q_COMPLETED:
 333                         if (io_u->error)
 334                                 ret = -io_u->error;
 335                         else if (io_u->resid) {
 336                                 int bytes = io_u->xfer_buflen - io_u->resid;
 337                                 struct fio_file *f = io_u->file;
 338
 339                                 /*
 340                                  * zero read, fail
 341                                  */
 342                                 if (!bytes) {
 343                                         td_verror(td, ENODATA, "full resid");
 344                                         put_io_u(td, io_u);
 345                                         break;
 346                                 }
 347
 348                                 io_u->xfer_buflen = io_u->resid;
 349                                 io_u->xfer_buf += bytes;
 350                                 io_u->offset += bytes;
 351                                 f->last_completed_pos = io_u->offset;
 352
 353                                 td->ts.short_io_u[io_u->ddir]++;
 354
 355                                 if (io_u->offset == f->real_file_size)
 356                                         goto sync_done;
 357
 358                                 requeue_io_u(td, &io_u);
 359                         } else {
 360 sync_done:
 361                                 ret = io_u_sync_complete(td, io_u);
 362                                 if (ret < 0)
 363                                         break;
 364                         }
 365                         continue;
 366                 case FIO_Q_QUEUED:
 367                         break;
 368                 case FIO_Q_BUSY:
 369                         requeue_io_u(td, &io_u);
 370                         ret2 = td_io_commit(td);
 371                         if (ret2 < 0)
 372                                 ret = ret2;
 373                         break;
 374                 default:
 375                         assert(ret < 0);
 376                         td_verror(td, -ret, "td_io_queue");
 377                         break;
 378                 }
 379
 380                 if (ret < 0 || td->error)
 381                         break;
 382
 383                 /*
 384                  * if we can queue more, do so. but check if there are
 385                  * completed io_u's first.
 386                  */
 387                 min_events = 0;
 388                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 389                         min_events = 1;
 390
 391                         if (td->cur_depth > td->o.iodepth_low)
 392                                 min_events = td->cur_depth - td->o.iodepth_low;
 393                 }
 394
 395                 /*
 396                  * Reap required number of io units, if any, and do the
 397                  * verification on them through the callback handler
 398                  */
 399                 if (io_u_queued_complete(td, min_events) < 0)
 400                         break;
 401         }
 402
 403         if (!td->error) {
 404                 min_events = td->cur_depth;
 405
 406                 if (min_events)
 407                         ret = io_u_queued_complete(td, min_events);
 408         } else
 409                 cleanup_pending_aio(td);
 410
 411         td_set_runstate(td, TD_RUNNING);
 412 }
 413
 414 /*
 415  * Main IO worker function. It retrieves io_u's to process and queues
 416  * and reaps them, checking for rate and errors along the way.
 417  */
 418 static void do_io(struct thread_data *td)
 419 {
 420         struct timeval s;
 421         unsigned long usec;
 422         unsigned int i;
 423         int ret = 0;
 424
 425         td_set_runstate(td, TD_RUNNING);
 426
 427         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 428                 struct timeval comp_time;
 429                 long bytes_done = 0;
 430                 int min_evts = 0;
 431                 struct io_u *io_u;
 432                 int ret2;
 433
 434                 if (td->terminate)
 435                         break;
 436
 437                 io_u = get_io_u(td);
 438                 if (!io_u)
 439                         break;
 440
 441                 memcpy(&s, &io_u->start_time, sizeof(s));
 442
 443                 if (runtime_exceeded(td, &s)) {
 444                         put_io_u(td, io_u);
 445                         td->terminate = 1;
 446                         break;
 447                 }
 448
 449                 /*
 450                  * Add verification end_io handler, if asked to verify
 451                  * a previously written file.
 452                  */
 453                 if (td->o.verify != VERIFY_NONE)
 454                         io_u->end_io = verify_io_u;
 455
 456                 ret = td_io_queue(td, io_u);
 457                 switch (ret) {
 458                 case FIO_Q_COMPLETED:
 459                         if (io_u->error)
 460                                 ret = -io_u->error;
 461                         else if (io_u->resid) {
 462                                 int bytes = io_u->xfer_buflen - io_u->resid;
 463                                 struct fio_file *f = io_u->file;
 464
 465                                 /*
 466                                  * zero read, fail
 467                                  */
 468                                 if (!bytes) {
 469                                         td_verror(td, ENODATA, "full resid");
 470                                         put_io_u(td, io_u);
 471                                         break;
 472                                 }
 473
 474                                 io_u->xfer_buflen = io_u->resid;
 475                                 io_u->xfer_buf += bytes;
 476                                 io_u->offset += bytes;
 477                                 f->last_completed_pos = io_u->offset;
 478
 479                                 td->ts.short_io_u[io_u->ddir]++;
 480
 481                                 if (io_u->offset == f->real_file_size)
 482                                         goto sync_done;
 483
 484                                 requeue_io_u(td, &io_u);
 485                         } else {
 486 sync_done:
 487                                 fio_gettime(&comp_time, NULL);
 488                                 bytes_done = io_u_sync_complete(td, io_u);
 489                                 if (bytes_done < 0)
 490                                         ret = bytes_done;
 491                         }
 492                         break;
 493                 case FIO_Q_QUEUED:
 494                         /*
 495                          * if the engine doesn't have a commit hook,
 496                          * the io_u is really queued. if it does have such
 497                          * a hook, it has to call io_u_queued() itself.
 498                          */
 499                         if (td->io_ops->commit == NULL)
 500                                 io_u_queued(td, io_u);
 501                         break;
 502                 case FIO_Q_BUSY:
 503                         requeue_io_u(td, &io_u);
 504                         ret2 = td_io_commit(td);
 505                         if (ret2 < 0)
 506                                 ret = ret2;
 507                         break;
 508                 default:
 509                         assert(ret < 0);
 510                         put_io_u(td, io_u);
 511                         break;
 512                 }
 513
 514                 if (ret < 0 || td->error)
 515                         break;
 516
 517                 /*
 518                  * See if we need to complete some commands
 519                  */
 520                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 521                         min_evts = 0;
 522                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 523                                 min_evts = 1;
 524
 525                                 if (td->cur_depth > td->o.iodepth_low)
 526                                         min_evts = td->cur_depth - td->o.iodepth_low;
 527                         }
 528
 529                         fio_gettime(&comp_time, NULL);
 530                         bytes_done = io_u_queued_complete(td, min_evts);
 531                         if (bytes_done < 0)
 532                                 break;
 533                 }
 534
 535                 if (!bytes_done)
 536                         continue;
 537
 538                 /*
 539                  * the rate is batched for now, it should work for batches
 540                  * of completions except the very first one which may look
 541                  * a little bursty
 542                  */
 543                 usec = utime_since(&s, &comp_time);
 544
 545                 rate_throttle(td, usec, bytes_done);
 546
 547                 if (check_min_rate(td, &comp_time)) {
 548                         if (exitall_on_terminate)
 549                                 terminate_threads(td->groupid);
 550                         td_verror(td, ENODATA, "check_min_rate");
 551                         break;
 552                 }
 553
 554                 if (td->o.thinktime) {
 555                         unsigned long long b;
 556
 557                         b = td->io_blocks[0] + td->io_blocks[1];
 558                         if (!(b % td->o.thinktime_blocks)) {
 559                                 int left;
 560
 561                                 if (td->o.thinktime_spin)
 562                                         __usec_sleep(td->o.thinktime_spin);
 563
 564                                 left = td->o.thinktime - td->o.thinktime_spin;
 565                                 if (left)
 566                                         usec_sleep(td, left);
 567                         }
 568                 }
 569         }
 570
 571         /*
 572          * stop job if we failed doing any IO
 573          */
 574         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 575                 td->done = 1;
 576
 577         if (td->o.fill_device && td->error == ENOSPC) {
 578                 td->error = 0;
 579                 td->terminate = 1;
 580         }
 581         if (!td->error) {
 582                 struct fio_file *f;
 583
 584                 i = td->cur_depth;
 585                 if (i)
 586                         ret = io_u_queued_complete(td, i);
 587
 588                 if (should_fsync(td) && td->o.end_fsync) {
 589                         td_set_runstate(td, TD_FSYNCING);
 590
 591                         for_each_file(td, f, i) {
 592                                 if (!(f->flags & FIO_FILE_OPEN))
 593                                         continue;
 594                                 fio_io_sync(td, f);
 595                         }
 596                 }
 597         } else
 598                 cleanup_pending_aio(td);
 599 }
 600
 601 static void cleanup_io_u(struct thread_data *td)
 602 {
 603         struct list_head *entry, *n;
 604         struct io_u *io_u;
 605
 606         list_for_each_safe(entry, n, &td->io_u_freelist) {
 607                 io_u = list_entry(entry, struct io_u, list);
 608
 609                 list_del(&io_u->list);
 610                 free(io_u);
 611         }
 612
 613         free_io_mem(td);
 614 }
 615
 616 /*
 617  * "randomly" fill the buffer contents
 618  */
 619 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 620 {
 621         long *ptr = io_u->buf;
 622
 623         if (!td->o.zero_buffers) {
 624                 while ((void *) ptr - io_u->buf < max_bs) {
 625                         *ptr = rand() * GOLDEN_RATIO_PRIME;
 626                         ptr++;
 627                 }
 628         } else
 629                 memset(ptr, 0, max_bs);
 630 }
 631
 632 static int init_io_u(struct thread_data *td)
 633 {
 634         struct io_u *io_u;
 635         unsigned int max_bs;
 636         int i, max_units;
 637         char *p;
 638
 639         if (td->io_ops->flags & FIO_SYNCIO)
 640                 max_units = 1;
 641         else
 642                 max_units = td->o.iodepth;
 643
 644         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 645         td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 646
 647         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 648                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 649
 650         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 651                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 652                 return 1;
 653         }
 654
 655         if (allocate_io_mem(td))
 656                 return 1;
 657
 658         if (td->o.odirect)
 659                 p = ALIGN(td->orig_buffer);
 660         else
 661                 p = td->orig_buffer;
 662
 663         for (i = 0; i < max_units; i++) {
 664                 if (td->terminate)
 665                         return 1;
 666                 io_u = malloc(sizeof(*io_u));
 667                 memset(io_u, 0, sizeof(*io_u));
 668                 INIT_LIST_HEAD(&io_u->list);
 669
 670                 if (!(td->io_ops->flags & FIO_NOIO)) {
 671                         io_u->buf = p + max_bs * i;
 672
 673                         if (td_write(td))
 674                                 fill_io_buf(td, io_u, max_bs);
 675                 }
 676
 677                 io_u->index = i;
 678                 io_u->flags = IO_U_F_FREE;
 679                 list_add(&io_u->list, &td->io_u_freelist);
 680         }
 681
 682         io_u_init_timeout();
 683
 684         return 0;
 685 }
 686
 687 static int switch_ioscheduler(struct thread_data *td)
 688 {
 689         char tmp[256], tmp2[128];
 690         FILE *f;
 691         int ret;
 692
 693         if (td->io_ops->flags & FIO_DISKLESSIO)
 694                 return 0;
 695
 696         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 697
 698         f = fopen(tmp, "r+");
 699         if (!f) {
 700                 if (errno == ENOENT) {
 701                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 702                         return 0;
 703                 }
 704                 td_verror(td, errno, "fopen iosched");
 705                 return 1;
 706         }
 707
 708         /*
 709          * Set io scheduler.
 710          */
 711         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 712         if (ferror(f) || ret != 1) {
 713                 td_verror(td, errno, "fwrite");
 714                 fclose(f);
 715                 return 1;
 716         }
 717
 718         rewind(f);
 719
 720         /*
 721          * Read back and check that the selected scheduler is now the default.
 722          */
 723         ret = fread(tmp, 1, sizeof(tmp), f);
 724         if (ferror(f) || ret < 0) {
 725                 td_verror(td, errno, "fread");
 726                 fclose(f);
 727                 return 1;
 728         }
 729
 730         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 731         if (!strstr(tmp, tmp2)) {
 732                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 733                 td_verror(td, EINVAL, "iosched_switch");
 734                 fclose(f);
 735                 return 1;
 736         }
 737
 738         fclose(f);
 739         return 0;
 740 }
 741
 742 static int keep_running(struct thread_data *td)
 743 {
 744         unsigned long long io_done;
 745
 746         if (td->done)
 747                 return 0;
 748         if (td->o.time_based)
 749                 return 1;
 750         if (td->o.loops) {
 751                 td->o.loops--;
 752                 return 1;
 753         }
 754
 755         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
 756         if (io_done < td->o.size)
 757                 return 1;
 758
 759         return 0;
 760 }
 761
 762 static int clear_io_state(struct thread_data *td)
 763 {
 764         struct fio_file *f;
 765         unsigned int i;
 766         int ret;
 767
 768         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 769         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 770         td->zone_bytes = 0;
 771         td->rate_bytes = 0;
 772         td->rate_blocks = 0;
 773         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 774
 775         td->last_was_sync = 0;
 776
 777         if (td->o.time_based)
 778                 td->nr_done_files = 0;
 779
 780         for_each_file(td, f, i)
 781                 td_io_close_file(td, f);
 782
 783         ret = 0;
 784         for_each_file(td, f, i) {
 785                 f->flags &= ~FIO_FILE_DONE;
 786                 ret = td_io_open_file(td, f);
 787                 if (ret)
 788                         break;
 789         }
 790
 791         return ret;
 792 }
 793
 794 /*
 795  * Entry point for the thread based jobs. The process based jobs end up
 796  * here as well, after a little setup.
 797  */
 798 static void *thread_main(void *data)
 799 {
 800         unsigned long long runtime[2], elapsed;
 801         struct thread_data *td = data;
 802         int clear_state;
 803
 804         if (!td->o.use_thread)
 805                 setsid();
 806
 807         td->pid = getpid();
 808
 809         INIT_LIST_HEAD(&td->io_u_freelist);
 810         INIT_LIST_HEAD(&td->io_u_busylist);
 811         INIT_LIST_HEAD(&td->io_u_requeues);
 812         INIT_LIST_HEAD(&td->io_log_list);
 813         INIT_LIST_HEAD(&td->io_hist_list);
 814         td->io_hist_tree = RB_ROOT;
 815
 816         td_set_runstate(td, TD_INITIALIZED);
 817         fio_sem_up(startup_sem);
 818         fio_sem_down(td->mutex);
 819
 820         /*
 821          * the ->mutex semaphore is now no longer used, close it to avoid
 822          * eating a file descriptor
 823          */
 824         fio_sem_remove(td->mutex);
 825
 826         /*
 827          * May alter parameters that init_io_u() will use, so we need to
 828          * do this first.
 829          */
 830         if (init_iolog(td))
 831                 goto err;
 832
 833         if (init_io_u(td))
 834                 goto err;
 835
 836         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 837                 td_verror(td, errno, "cpu_set_affinity");
 838                 goto err;
 839         }
 840
 841         if (td->ioprio_set) {
 842                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 843                         td_verror(td, errno, "ioprio_set");
 844                         goto err;
 845                 }
 846         }
 847
 848         if (nice(td->o.nice) == -1) {
 849                 td_verror(td, errno, "nice");
 850                 goto err;
 851         }
 852
 853         if (td->o.ioscheduler && switch_ioscheduler(td))
 854                 goto err;
 855
 856         if (!td->o.create_serialize && setup_files(td))
 857                 goto err;
 858
 859         if (td_io_init(td))
 860                 goto err;
 861
 862         if (open_files(td))
 863                 goto err;
 864
 865         if (init_random_map(td))
 866                 goto err;
 867
 868         if (td->o.exec_prerun) {
 869                 if (system(td->o.exec_prerun) < 0)
 870                         goto err;
 871         }
 872
 873         fio_gettime(&td->epoch, NULL);
 874         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 875         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 876
 877         runtime[0] = runtime[1] = 0;
 878         clear_state = 0;
 879         while (keep_running(td)) {
 880                 fio_gettime(&td->start, NULL);
 881                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 882
 883                 if (td->o.ratemin)
 884                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 885
 886                 if (clear_state && clear_io_state(td))
 887                         break;
 888
 889                 prune_io_piece_log(td);
 890
 891                 do_io(td);
 892
 893                 clear_state = 1;
 894
 895                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 896                         if (td->rw_end_set[DDIR_READ])
 897                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 898                         else
 899                                 elapsed = utime_since_now(&td->start);
 900
 901                         runtime[DDIR_READ] += elapsed;
 902                 }
 903                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 904                         if (td->rw_end_set[DDIR_WRITE])
 905                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 906                         else
 907                                 elapsed = utime_since_now(&td->start);
 908
 909                         runtime[DDIR_WRITE] += elapsed;
 910                 }
 911
 912                 if (td->error || td->terminate)
 913                         break;
 914
 915                 if (!td->o.do_verify ||
 916                     td->o.verify == VERIFY_NONE ||
 917                     (td->io_ops->flags & FIO_UNIDIR))
 918                         continue;
 919
 920                 if (clear_io_state(td))
 921                         break;
 922
 923                 fio_gettime(&td->start, NULL);
 924
 925                 do_verify(td);
 926
 927                 runtime[DDIR_READ] += utime_since_now(&td->start);
 928
 929                 if (td->error || td->terminate)
 930                         break;
 931         }
 932
 933         update_rusage_stat(td);
 934         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 935         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 936         td->ts.total_run_time = mtime_since_now(&td->epoch);
 937         td->ts.io_bytes[0] = td->io_bytes[0];
 938         td->ts.io_bytes[1] = td->io_bytes[1];
 939
 940         if (td->ts.bw_log)
 941                 finish_log(td, td->ts.bw_log, "bw");
 942         if (td->ts.slat_log)
 943                 finish_log(td, td->ts.slat_log, "slat");
 944         if (td->ts.clat_log)
 945                 finish_log(td, td->ts.clat_log, "clat");
 946         if (td->o.exec_postrun) {
 947                 if (system(td->o.exec_postrun) < 0)
 948                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 949         }
 950
 951         if (exitall_on_terminate)
 952                 terminate_threads(td->groupid);
 953
 954 err:
 955         if (td->error)
 956                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 957         close_files(td);
 958         close_ioengine(td);
 959         cleanup_io_u(td);
 960
 961         /*
 962          * do this very late, it will log file closing as well
 963          */
 964         if (td->o.write_iolog_file)
 965                 write_iolog_close(td);
 966
 967         options_mem_free(td);
 968         td_set_runstate(td, TD_EXITED);
 969         return (void *) (unsigned long) td->error;
 970 }
 971
 972 /*
 973  * We cannot pass the td data into a forked process, so attach the td and
 974  * pass it to the thread worker.
 975  */
 976 static int fork_main(int shmid, int offset)
 977 {
 978         struct thread_data *td;
 979         void *data, *ret;
 980
 981         data = shmat(shmid, NULL, 0);
 982         if (data == (void *) -1) {
 983                 int __err = errno;
 984
 985                 perror("shmat");
 986                 return __err;
 987         }
 988
 989         td = data + offset * sizeof(struct thread_data);
 990         ret = thread_main(td);
 991         shmdt(data);
 992         return (int) (unsigned long) ret;
 993 }
 994
 995 /*
 996  * Run over the job map and reap the threads that have exited, if any.
 997  */
 998 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 999 {
1000         struct thread_data *td;
1001         int i, cputhreads, realthreads, pending, status, ret;
1002
1003         /*
1004          * reap exited threads (TD_EXITED -> TD_REAPED)
1005          */
1006         realthreads = pending = cputhreads = 0;
1007         for_each_td(td, i) {
1008                 int flags = 0;
1009
1010                 /*
1011                  * ->io_ops is NULL for a thread that has closed its
1012                  * io engine
1013                  */
1014                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1015                         cputhreads++;
1016                 else
1017                         realthreads++;
1018
1019                 if (!td->pid || td->runstate == TD_REAPED)
1020                         continue;
1021                 if (td->o.use_thread) {
1022                         if (td->runstate == TD_EXITED) {
1023                                 td_set_runstate(td, TD_REAPED);
1024                                 goto reaped;
1025                         }
1026                         continue;
1027                 }
1028
1029                 flags = WNOHANG;
1030                 if (td->runstate == TD_EXITED)
1031                         flags = 0;
1032
1033                 /*
1034                  * check if someone quit or got killed in an unusual way
1035                  */
1036                 ret = waitpid(td->pid, &status, flags);
1037                 if (ret < 0) {
1038                         if (errno == ECHILD) {
1039                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1040                                 td_set_runstate(td, TD_REAPED);
1041                                 goto reaped;
1042                         }
1043                         perror("waitpid");
1044                 } else if (ret == td->pid) {
1045                         if (WIFSIGNALED(status)) {
1046                                 int sig = WTERMSIG(status);
1047
1048                                 if (sig != SIGQUIT)
1049                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1050                                 td_set_runstate(td, TD_REAPED);
1051                                 goto reaped;
1052                         }
1053                         if (WIFEXITED(status)) {
1054                                 if (WEXITSTATUS(status) && !td->error)
1055                                         td->error = WEXITSTATUS(status);
1056
1057                                 td_set_runstate(td, TD_REAPED);
1058                                 goto reaped;
1059                         }
1060                 }
1061
1062                 /*
1063                  * thread is not dead, continue
1064                  */
1065                 pending++;
1066                 continue;
1067 reaped:
1068                 if (td->o.use_thread) {
1069                         long ret;
1070
1071                         if (pthread_join(td->thread, (void *) &ret))
1072                                 perror("pthread_join");
1073                 }
1074
1075                 (*nr_running)--;
1076                 (*m_rate) -= td->o.ratemin;
1077                 (*t_rate) -= td->o.rate;
1078                 pending--;
1079
1080                 if (td->error)
1081                         exit_value++;
1082         }
1083
1084         if (*nr_running == cputhreads && !pending && realthreads)
1085                 terminate_threads(TERMINATE_ALL);
1086 }
1087
1088 /*
1089  * Main function for kicking off and reaping jobs, as needed.
1090  */
1091 static void run_threads(void)
1092 {
1093         struct thread_data *td;
1094         unsigned long spent;
1095         int i, todo, nr_running, m_rate, t_rate, nr_started;
1096
1097         if (fio_pin_memory())
1098                 return;
1099
1100         if (!terse_output) {
1101                 printf("Starting ");
1102                 if (nr_thread)
1103                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1104                 if (nr_process) {
1105                         if (nr_thread)
1106                                 printf(" and ");
1107                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1108                 }
1109                 printf("\n");
1110                 fflush(stdout);
1111         }
1112
1113         signal(SIGINT, sig_handler);
1114         signal(SIGALRM, sig_handler);
1115
1116         todo = thread_number;
1117         nr_running = 0;
1118         nr_started = 0;
1119         m_rate = t_rate = 0;
1120
1121         for_each_td(td, i) {
1122                 print_status_init(td->thread_number - 1);
1123
1124                 if (!td->o.create_serialize) {
1125                         init_disk_util(td);
1126                         continue;
1127                 }
1128
1129                 /*
1130                  * do file setup here so it happens sequentially,
1131                  * we don't want X number of threads getting their
1132                  * client data interspersed on disk
1133                  */
1134                 if (setup_files(td)) {
1135                         exit_value++;
1136                         if (td->error)
1137                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1138                         td_set_runstate(td, TD_REAPED);
1139                         todo--;
1140                 }
1141
1142                 init_disk_util(td);
1143         }
1144
1145         set_genesis_time();
1146
1147         while (todo) {
1148                 struct thread_data *map[MAX_JOBS];
1149                 struct timeval this_start;
1150                 int this_jobs = 0, left;
1151
1152                 /*
1153                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1154                  */
1155                 for_each_td(td, i) {
1156                         if (td->runstate != TD_NOT_CREATED)
1157                                 continue;
1158
1159                         /*
1160                          * never got a chance to start, killed by other
1161                          * thread for some reason
1162                          */
1163                         if (td->terminate) {
1164                                 todo--;
1165                                 continue;
1166                         }
1167
1168                         if (td->o.start_delay) {
1169                                 spent = mtime_since_genesis();
1170
1171                                 if (td->o.start_delay * 1000 > spent)
1172                                         continue;
1173                         }
1174
1175                         if (td->o.stonewall && (nr_started || nr_running))
1176                                 break;
1177
1178                         /*
1179                          * Set state to created. Thread will transition
1180                          * to TD_INITIALIZED when it's done setting up.
1181                          */
1182                         td_set_runstate(td, TD_CREATED);
1183                         map[this_jobs++] = td;
1184                         nr_started++;
1185
1186                         if (td->o.use_thread) {
1187                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1188                                         perror("thread_create");
1189                                         nr_started--;
1190                                         break;
1191                                 }
1192                         } else {
1193                                 if (!fork()) {
1194                                         int ret = fork_main(shm_id, i);
1195
1196                                         exit(ret);
1197                                 }
1198                         }
1199                         fio_sem_down(startup_sem);
1200                 }
1201
1202                 /*
1203                  * Wait for the started threads to transition to
1204                  * TD_INITIALIZED.
1205                  */
1206                 fio_gettime(&this_start, NULL);
1207                 left = this_jobs;
1208                 while (left && !fio_abort) {
1209                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1210                                 break;
1211
1212                         usleep(100000);
1213
1214                         for (i = 0; i < this_jobs; i++) {
1215                                 td = map[i];
1216                                 if (!td)
1217                                         continue;
1218                                 if (td->runstate == TD_INITIALIZED) {
1219                                         map[i] = NULL;
1220                                         left--;
1221                                 } else if (td->runstate >= TD_EXITED) {
1222                                         map[i] = NULL;
1223                                         left--;
1224                                         todo--;
1225                                         nr_running++; /* work-around... */
1226                                 }
1227                         }
1228                 }
1229
1230                 if (left) {
1231                         log_err("fio: %d jobs failed to start\n", left);
1232                         for (i = 0; i < this_jobs; i++) {
1233                                 td = map[i];
1234                                 if (!td)
1235                                         continue;
1236                                 kill(td->pid, SIGTERM);
1237                         }
1238                         break;
1239                 }
1240
1241                 /*
1242                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1243                  */
1244                 for_each_td(td, i) {
1245                         if (td->runstate != TD_INITIALIZED)
1246                                 continue;
1247
1248                         td_set_runstate(td, TD_RUNNING);
1249                         nr_running++;
1250                         nr_started--;
1251                         m_rate += td->o.ratemin;
1252                         t_rate += td->o.rate;
1253                         todo--;
1254                         fio_sem_up(td->mutex);
1255                 }
1256
1257                 reap_threads(&nr_running, &t_rate, &m_rate);
1258
1259                 if (todo)
1260                         usleep(100000);
1261         }
1262
1263         while (nr_running) {
1264                 reap_threads(&nr_running, &t_rate, &m_rate);
1265                 usleep(10000);
1266         }
1267
1268         update_io_ticks();
1269         fio_unpin_memory();
1270 }
1271
1272 int main(int argc, char *argv[])
1273 {
1274         long ps;
1275
1276         /*
1277          * We need locale for number printing, if it isn't set then just
1278          * go with the US format.
1279          */
1280         if (!getenv("LC_NUMERIC"))
1281                 setlocale(LC_NUMERIC, "en_US");
1282
1283         if (parse_options(argc, argv))
1284                 return 1;
1285
1286         if (!thread_number)
1287                 return 0;
1288
1289         ps = sysconf(_SC_PAGESIZE);
1290         if (ps < 0) {
1291                 log_err("Failed to get page size\n");
1292                 return 1;
1293         }
1294
1295         page_size = ps;
1296         page_mask = ps - 1;
1297
1298         if (write_bw_log) {
1299                 setup_log(&agg_io_log[DDIR_READ]);
1300                 setup_log(&agg_io_log[DDIR_WRITE]);
1301         }
1302
1303         startup_sem = fio_sem_init(0);
1304
1305         set_genesis_time();
1306
1307         disk_util_timer_arm();
1308
1309         run_threads();
1310
1311         if (!fio_abort) {
1312                 show_run_stats();
1313                 if (write_bw_log) {
1314                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1315                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1316                 }
1317         }
1318
1319         fio_sem_remove(startup_sem);
1320         return exit_value;
1321 }