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         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
  64                                                                 runstate);
  65         td->runstate = runstate;
  66 }
  67
  68 static void terminate_threads(int group_id)
  69 {
  70         struct thread_data *td;
  71         int i;
  72
  73         for_each_td(td, i) {
  74                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  75                         dprint(FD_PROCESS, "setting terminate on %d\n",td->pid);
  76                         /*
  77                          * if the thread is running, just let it exit
  78                          */
  79                         if (td->runstate < TD_RUNNING)
  80                                 kill(td->pid, SIGQUIT);
  81                         td->terminate = 1;
  82                         td->o.start_delay = 0;
  83                 }
  84         }
  85 }
  86
  87 static void sig_handler(int sig)
  88 {
  89         switch (sig) {
  90                 case SIGALRM:
  91                         update_io_ticks();
  92                         disk_util_timer_arm();
  93                         print_thread_status();
  94                         break;
  95                 default:
  96                         printf("\nfio: terminating on signal %d\n", sig);
  97                         fflush(stdout);
  98                         terminate_threads(TERMINATE_ALL);
  99                         break;
 100         }
 101 }
 102
 103 /*
 104  * Check if we are above the minimum rate given.
 105  */
 106 static int check_min_rate(struct thread_data *td, struct timeval *now)
 107 {
 108         unsigned long long bytes = 0;
 109         unsigned long iops = 0;
 110         unsigned long spent;
 111         unsigned long rate;
 112
 113         /*
 114          * No minimum rate set, always ok
 115          */
 116         if (!td->o.ratemin && !td->o.rate_iops_min)
 117                 return 0;
 118
 119         /*
 120          * allow a 2 second settle period in the beginning
 121          */
 122         if (mtime_since(&td->start, now) < 2000)
 123                 return 0;
 124
 125         if (td_read(td)) {
 126                 iops += td->io_blocks[DDIR_READ];
 127                 bytes += td->this_io_bytes[DDIR_READ];
 128         }
 129         if (td_write(td)) {
 130                 iops += td->io_blocks[DDIR_WRITE];
 131                 bytes += td->this_io_bytes[DDIR_WRITE];
 132         }
 133
 134         /*
 135          * if rate blocks is set, sample is running
 136          */
 137         if (td->rate_bytes || td->rate_blocks) {
 138                 spent = mtime_since(&td->lastrate, now);
 139                 if (spent < td->o.ratecycle)
 140                         return 0;
 141
 142                 if (td->o.rate) {
 143                         /*
 144                          * check bandwidth specified rate
 145                          */
 146                         if (bytes < td->rate_bytes) {
 147                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
 148                                 return 1;
 149                         } else {
 150                                 rate = (bytes - td->rate_bytes) / spent;
 151                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
 152                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
 153                                         return 1;
 154                                 }
 155                         }
 156                 } else {
 157                         /*
 158                          * checks iops specified rate
 159                          */
 160                         if (iops < td->o.rate_iops) {
 161                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
 162                                 return 1;
 163                         } else {
 164                                 rate = (iops - td->rate_blocks) / spent;
 165                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
 166                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
 167                                 }
 168                         }
 169                 }
 170         }
 171
 172         td->rate_bytes = bytes;
 173         td->rate_blocks = iops;
 174         memcpy(&td->lastrate, now, sizeof(*now));
 175         return 0;
 176 }
 177
 178 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 179 {
 180         if (!td->o.timeout)
 181                 return 0;
 182         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 183                 return 1;
 184
 185         return 0;
 186 }
 187
 188 /*
 189  * When job exits, we can cancel the in-flight IO if we are using async
 190  * io. Attempt to do so.
 191  */
 192 static void cleanup_pending_aio(struct thread_data *td)
 193 {
 194         struct list_head *entry, *n;
 195         struct io_u *io_u;
 196         int r;
 197
 198         /*
 199          * get immediately available events, if any
 200          */
 201         r = io_u_queued_complete(td, 0);
 202         if (r < 0)
 203                 return;
 204
 205         /*
 206          * now cancel remaining active events
 207          */
 208         if (td->io_ops->cancel) {
 209                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 210                         io_u = list_entry(entry, struct io_u, list);
 211
 212                         /*
 213                          * if the io_u isn't in flight, then that generally
 214                          * means someone leaked an io_u. complain but fix
 215                          * it up, so we don't stall here.
 216                          */
 217                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 218                                 log_err("fio: non-busy IO on busy list\n");
 219                                 put_io_u(td, io_u);
 220                         } else {
 221                                 r = td->io_ops->cancel(td, io_u);
 222                                 if (!r)
 223                                         put_io_u(td, io_u);
 224                         }
 225                 }
 226         }
 227
 228         if (td->cur_depth)
 229                 r = io_u_queued_complete(td, td->cur_depth);
 230 }
 231
 232 /*
 233  * Helper to handle the final sync of a file. Works just like the normal
 234  * io path, just does everything sync.
 235  */
 236 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 237 {
 238         struct io_u *io_u = __get_io_u(td);
 239         int ret;
 240
 241         if (!io_u)
 242                 return 1;
 243
 244         io_u->ddir = DDIR_SYNC;
 245         io_u->file = f;
 246
 247         if (td_io_prep(td, io_u)) {
 248                 put_io_u(td, io_u);
 249                 return 1;
 250         }
 251
 252 requeue:
 253         ret = td_io_queue(td, io_u);
 254         if (ret < 0) {
 255                 td_verror(td, io_u->error, "td_io_queue");
 256                 put_io_u(td, io_u);
 257                 return 1;
 258         } else if (ret == FIO_Q_QUEUED) {
 259                 if (io_u_queued_complete(td, 1) < 0)
 260                         return 1;
 261         } else if (ret == FIO_Q_COMPLETED) {
 262                 if (io_u->error) {
 263                         td_verror(td, io_u->error, "td_io_queue");
 264                         return 1;
 265                 }
 266
 267                 if (io_u_sync_complete(td, io_u) < 0)
 268                         return 1;
 269         } else if (ret == FIO_Q_BUSY) {
 270                 if (td_io_commit(td))
 271                         return 1;
 272                 goto requeue;
 273         }
 274
 275         return 0;
 276 }
 277
 278 /*
 279  * The main verify engine. Runs over the writes we previously submitted,
 280  * reads the blocks back in, and checks the crc/md5 of the data.
 281  */
 282 static void do_verify(struct thread_data *td)
 283 {
 284         struct fio_file *f;
 285         struct io_u *io_u;
 286         int ret, min_events;
 287         unsigned int i;
 288
 289         /*
 290          * sync io first and invalidate cache, to make sure we really
 291          * read from disk.
 292          */
 293         for_each_file(td, f, i) {
 294                 if (!(f->flags & FIO_FILE_OPEN))
 295                         continue;
 296                 if (fio_io_sync(td, f))
 297                         break;
 298                 if (file_invalidate_cache(td, f))
 299                         break;
 300         }
 301
 302         if (td->error)
 303                 return;
 304
 305         td_set_runstate(td, TD_VERIFYING);
 306
 307         io_u = NULL;
 308         while (!td->terminate) {
 309                 int ret2;
 310
 311                 io_u = __get_io_u(td);
 312                 if (!io_u)
 313                         break;
 314
 315                 if (runtime_exceeded(td, &io_u->start_time)) {
 316                         put_io_u(td, io_u);
 317                         td->terminate = 1;
 318                         break;
 319                 }
 320
 321                 if (get_next_verify(td, io_u)) {
 322                         put_io_u(td, io_u);
 323                         break;
 324                 }
 325
 326                 if (td_io_prep(td, io_u)) {
 327                         put_io_u(td, io_u);
 328                         break;
 329                 }
 330
 331                 io_u->end_io = verify_io_u;
 332
 333                 ret = td_io_queue(td, io_u);
 334                 switch (ret) {
 335                 case FIO_Q_COMPLETED:
 336                         if (io_u->error)
 337                                 ret = -io_u->error;
 338                         else if (io_u->resid) {
 339                                 int bytes = io_u->xfer_buflen - io_u->resid;
 340                                 struct fio_file *f = io_u->file;
 341
 342                                 /*
 343                                  * zero read, fail
 344                                  */
 345                                 if (!bytes) {
 346                                         td_verror(td, ENODATA, "full resid");
 347                                         put_io_u(td, io_u);
 348                                         break;
 349                                 }
 350
 351                                 io_u->xfer_buflen = io_u->resid;
 352                                 io_u->xfer_buf += bytes;
 353                                 io_u->offset += bytes;
 354                                 f->last_completed_pos = io_u->offset;
 355
 356                                 td->ts.short_io_u[io_u->ddir]++;
 357
 358                                 if (io_u->offset == f->real_file_size)
 359                                         goto sync_done;
 360
 361                                 requeue_io_u(td, &io_u);
 362                         } else {
 363 sync_done:
 364                                 ret = io_u_sync_complete(td, io_u);
 365                                 if (ret < 0)
 366                                         break;
 367                         }
 368                         continue;
 369                 case FIO_Q_QUEUED:
 370                         break;
 371                 case FIO_Q_BUSY:
 372                         requeue_io_u(td, &io_u);
 373                         ret2 = td_io_commit(td);
 374                         if (ret2 < 0)
 375                                 ret = ret2;
 376                         break;
 377                 default:
 378                         assert(ret < 0);
 379                         td_verror(td, -ret, "td_io_queue");
 380                         break;
 381                 }
 382
 383                 if (ret < 0 || td->error)
 384                         break;
 385
 386                 /*
 387                  * if we can queue more, do so. but check if there are
 388                  * completed io_u's first.
 389                  */
 390                 min_events = 0;
 391                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 392                         min_events = 1;
 393
 394                         if (td->cur_depth > td->o.iodepth_low)
 395                                 min_events = td->cur_depth - td->o.iodepth_low;
 396                 }
 397
 398                 /*
 399                  * Reap required number of io units, if any, and do the
 400                  * verification on them through the callback handler
 401                  */
 402                 if (io_u_queued_complete(td, min_events) < 0)
 403                         break;
 404         }
 405
 406         if (!td->error) {
 407                 min_events = td->cur_depth;
 408
 409                 if (min_events)
 410                         ret = io_u_queued_complete(td, min_events);
 411         } else
 412                 cleanup_pending_aio(td);
 413
 414         td_set_runstate(td, TD_RUNNING);
 415 }
 416
 417 /*
 418  * Main IO worker function. It retrieves io_u's to process and queues
 419  * and reaps them, checking for rate and errors along the way.
 420  */
 421 static void do_io(struct thread_data *td)
 422 {
 423         struct timeval s;
 424         unsigned long usec;
 425         unsigned int i;
 426         int ret = 0;
 427
 428         td_set_runstate(td, TD_RUNNING);
 429
 430         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 431                 struct timeval comp_time;
 432                 long bytes_done = 0;
 433                 int min_evts = 0;
 434                 struct io_u *io_u;
 435                 int ret2;
 436
 437                 if (td->terminate)
 438                         break;
 439
 440                 io_u = get_io_u(td);
 441                 if (!io_u)
 442                         break;
 443
 444                 memcpy(&s, &io_u->start_time, sizeof(s));
 445
 446                 if (runtime_exceeded(td, &s)) {
 447                         put_io_u(td, io_u);
 448                         td->terminate = 1;
 449                         break;
 450                 }
 451
 452                 /*
 453                  * Add verification end_io handler, if asked to verify
 454                  * a previously written file.
 455                  */
 456                 if (td->o.verify != VERIFY_NONE)
 457                         io_u->end_io = verify_io_u;
 458
 459                 ret = td_io_queue(td, io_u);
 460                 switch (ret) {
 461                 case FIO_Q_COMPLETED:
 462                         if (io_u->error)
 463                                 ret = -io_u->error;
 464                         else if (io_u->resid) {
 465                                 int bytes = io_u->xfer_buflen - io_u->resid;
 466                                 struct fio_file *f = io_u->file;
 467
 468                                 /*
 469                                  * zero read, fail
 470                                  */
 471                                 if (!bytes) {
 472                                         td_verror(td, ENODATA, "full resid");
 473                                         put_io_u(td, io_u);
 474                                         break;
 475                                 }
 476
 477                                 io_u->xfer_buflen = io_u->resid;
 478                                 io_u->xfer_buf += bytes;
 479                                 io_u->offset += bytes;
 480                                 f->last_completed_pos = io_u->offset;
 481
 482                                 td->ts.short_io_u[io_u->ddir]++;
 483
 484                                 if (io_u->offset == f->real_file_size)
 485                                         goto sync_done;
 486
 487                                 requeue_io_u(td, &io_u);
 488                         } else {
 489 sync_done:
 490                                 fio_gettime(&comp_time, NULL);
 491                                 bytes_done = io_u_sync_complete(td, io_u);
 492                                 if (bytes_done < 0)
 493                                         ret = bytes_done;
 494                         }
 495                         break;
 496                 case FIO_Q_QUEUED:
 497                         /*
 498                          * if the engine doesn't have a commit hook,
 499                          * the io_u is really queued. if it does have such
 500                          * a hook, it has to call io_u_queued() itself.
 501                          */
 502                         if (td->io_ops->commit == NULL)
 503                                 io_u_queued(td, io_u);
 504                         break;
 505                 case FIO_Q_BUSY:
 506                         requeue_io_u(td, &io_u);
 507                         ret2 = td_io_commit(td);
 508                         if (ret2 < 0)
 509                                 ret = ret2;
 510                         break;
 511                 default:
 512                         assert(ret < 0);
 513                         put_io_u(td, io_u);
 514                         break;
 515                 }
 516
 517                 if (ret < 0 || td->error)
 518                         break;
 519
 520                 /*
 521                  * See if we need to complete some commands
 522                  */
 523                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 524                         min_evts = 0;
 525                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 526                                 min_evts = 1;
 527
 528                                 if (td->cur_depth > td->o.iodepth_low)
 529                                         min_evts = td->cur_depth - td->o.iodepth_low;
 530                         }
 531
 532                         fio_gettime(&comp_time, NULL);
 533                         bytes_done = io_u_queued_complete(td, min_evts);
 534                         if (bytes_done < 0)
 535                                 break;
 536                 }
 537
 538                 if (!bytes_done)
 539                         continue;
 540
 541                 /*
 542                  * the rate is batched for now, it should work for batches
 543                  * of completions except the very first one which may look
 544                  * a little bursty
 545                  */
 546                 usec = utime_since(&s, &comp_time);
 547
 548                 rate_throttle(td, usec, bytes_done);
 549
 550                 if (check_min_rate(td, &comp_time)) {
 551                         if (exitall_on_terminate)
 552                                 terminate_threads(td->groupid);
 553                         td_verror(td, ENODATA, "check_min_rate");
 554                         break;
 555                 }
 556
 557                 if (td->o.thinktime) {
 558                         unsigned long long b;
 559
 560                         b = td->io_blocks[0] + td->io_blocks[1];
 561                         if (!(b % td->o.thinktime_blocks)) {
 562                                 int left;
 563
 564                                 if (td->o.thinktime_spin)
 565                                         __usec_sleep(td->o.thinktime_spin);
 566
 567                                 left = td->o.thinktime - td->o.thinktime_spin;
 568                                 if (left)
 569                                         usec_sleep(td, left);
 570                         }
 571                 }
 572         }
 573
 574         if (td->o.fill_device && td->error == ENOSPC) {
 575                 td->error = 0;
 576                 td->terminate = 1;
 577         }
 578         if (!td->error) {
 579                 struct fio_file *f;
 580
 581                 i = td->cur_depth;
 582                 if (i)
 583                         ret = io_u_queued_complete(td, i);
 584
 585                 if (should_fsync(td) && td->o.end_fsync) {
 586                         td_set_runstate(td, TD_FSYNCING);
 587
 588                         for_each_file(td, f, i) {
 589                                 if (!(f->flags & FIO_FILE_OPEN))
 590                                         continue;
 591                                 fio_io_sync(td, f);
 592                         }
 593                 }
 594         } else
 595                 cleanup_pending_aio(td);
 596
 597         /*
 598          * stop job if we failed doing any IO
 599          */
 600         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 601                 td->done = 1;
 602 }
 603
 604 static void cleanup_io_u(struct thread_data *td)
 605 {
 606         struct list_head *entry, *n;
 607         struct io_u *io_u;
 608
 609         list_for_each_safe(entry, n, &td->io_u_freelist) {
 610                 io_u = list_entry(entry, struct io_u, list);
 611
 612                 list_del(&io_u->list);
 613                 free(io_u);
 614         }
 615
 616         free_io_mem(td);
 617 }
 618
 619 /*
 620  * "randomly" fill the buffer contents
 621  */
 622 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 623 {
 624         long *ptr = io_u->buf;
 625
 626         if (!td->o.zero_buffers) {
 627                 while ((void *) ptr - io_u->buf < max_bs) {
 628                         *ptr = rand() * GOLDEN_RATIO_PRIME;
 629                         ptr++;
 630                 }
 631         } else
 632                 memset(ptr, 0, max_bs);
 633 }
 634
 635 static int init_io_u(struct thread_data *td)
 636 {
 637         struct io_u *io_u;
 638         unsigned int max_bs;
 639         int i, max_units;
 640         char *p;
 641
 642         max_units = td->o.iodepth;
 643         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 644         td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 645
 646         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 647                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 648
 649         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 650                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 651                 return 1;
 652         }
 653
 654         if (allocate_io_mem(td))
 655                 return 1;
 656
 657         if (td->o.odirect)
 658                 p = ALIGN(td->orig_buffer);
 659         else
 660                 p = td->orig_buffer;
 661
 662         for (i = 0; i < max_units; i++) {
 663                 if (td->terminate)
 664                         return 1;
 665                 io_u = malloc(sizeof(*io_u));
 666                 memset(io_u, 0, sizeof(*io_u));
 667                 INIT_LIST_HEAD(&io_u->list);
 668
 669                 if (!(td->io_ops->flags & FIO_NOIO)) {
 670                         io_u->buf = p + max_bs * i;
 671
 672                         if (td_write(td))
 673                                 fill_io_buf(td, io_u, max_bs);
 674                 }
 675
 676                 io_u->index = i;
 677                 io_u->flags = IO_U_F_FREE;
 678                 list_add(&io_u->list, &td->io_u_freelist);
 679         }
 680
 681         io_u_init_timeout();
 682
 683         return 0;
 684 }
 685
 686 static int switch_ioscheduler(struct thread_data *td)
 687 {
 688         char tmp[256], tmp2[128];
 689         FILE *f;
 690         int ret;
 691
 692         if (td->io_ops->flags & FIO_DISKLESSIO)
 693                 return 0;
 694
 695         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 696
 697         f = fopen(tmp, "r+");
 698         if (!f) {
 699                 if (errno == ENOENT) {
 700                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 701                         return 0;
 702                 }
 703                 td_verror(td, errno, "fopen iosched");
 704                 return 1;
 705         }
 706
 707         /*
 708          * Set io scheduler.
 709          */
 710         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 711         if (ferror(f) || ret != 1) {
 712                 td_verror(td, errno, "fwrite");
 713                 fclose(f);
 714                 return 1;
 715         }
 716
 717         rewind(f);
 718
 719         /*
 720          * Read back and check that the selected scheduler is now the default.
 721          */
 722         ret = fread(tmp, 1, sizeof(tmp), f);
 723         if (ferror(f) || ret < 0) {
 724                 td_verror(td, errno, "fread");
 725                 fclose(f);
 726                 return 1;
 727         }
 728
 729         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 730         if (!strstr(tmp, tmp2)) {
 731                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 732                 td_verror(td, EINVAL, "iosched_switch");
 733                 fclose(f);
 734                 return 1;
 735         }
 736
 737         fclose(f);
 738         return 0;
 739 }
 740
 741 static int keep_running(struct thread_data *td)
 742 {
 743         unsigned long long io_done;
 744
 745         if (td->done)
 746                 return 0;
 747         if (td->o.time_based)
 748                 return 1;
 749         if (td->o.loops) {
 750                 td->o.loops--;
 751                 return 1;
 752         }
 753
 754         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
 755         if (io_done < td->o.size)
 756                 return 1;
 757
 758         return 0;
 759 }
 760
 761 static int clear_io_state(struct thread_data *td)
 762 {
 763         struct fio_file *f;
 764         unsigned int i;
 765         int ret;
 766
 767         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 768         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 769         td->zone_bytes = 0;
 770         td->rate_bytes = 0;
 771         td->rate_blocks = 0;
 772         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 773
 774         td->last_was_sync = 0;
 775
 776         /*
 777          * reset file done count if we are to start over
 778          */
 779         if (td->o.time_based || td->o.loops)
 780                 td->nr_done_files = 0;
 781
 782         for_each_file(td, f, i)
 783                 td_io_close_file(td, f);
 784
 785         ret = 0;
 786         for_each_file(td, f, i) {
 787                 f->flags &= ~FIO_FILE_DONE;
 788                 ret = td_io_open_file(td, f);
 789                 if (ret)
 790                         break;
 791         }
 792
 793         return ret;
 794 }
 795
 796 /*
 797  * Entry point for the thread based jobs. The process based jobs end up
 798  * here as well, after a little setup.
 799  */
 800 static void *thread_main(void *data)
 801 {
 802         unsigned long long runtime[2], elapsed;
 803         struct thread_data *td = data;
 804         int clear_state;
 805
 806         if (!td->o.use_thread)
 807                 setsid();
 808
 809         td->pid = getpid();
 810
 811         dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
 812
 813         INIT_LIST_HEAD(&td->io_u_freelist);
 814         INIT_LIST_HEAD(&td->io_u_busylist);
 815         INIT_LIST_HEAD(&td->io_u_requeues);
 816         INIT_LIST_HEAD(&td->io_log_list);
 817         INIT_LIST_HEAD(&td->io_hist_list);
 818         td->io_hist_tree = RB_ROOT;
 819
 820         td_set_runstate(td, TD_INITIALIZED);
 821         fio_sem_up(startup_sem);
 822         fio_sem_down(td->mutex);
 823
 824         /*
 825          * the ->mutex semaphore is now no longer used, close it to avoid
 826          * eating a file descriptor
 827          */
 828         fio_sem_remove(td->mutex);
 829
 830         /*
 831          * May alter parameters that init_io_u() will use, so we need to
 832          * do this first.
 833          */
 834         if (init_iolog(td))
 835                 goto err;
 836
 837         if (init_io_u(td))
 838                 goto err;
 839
 840         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 841                 td_verror(td, errno, "cpu_set_affinity");
 842                 goto err;
 843         }
 844
 845         if (td->ioprio_set) {
 846                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 847                         td_verror(td, errno, "ioprio_set");
 848                         goto err;
 849                 }
 850         }
 851
 852         if (nice(td->o.nice) == -1) {
 853                 td_verror(td, errno, "nice");
 854                 goto err;
 855         }
 856
 857         if (td->o.ioscheduler && switch_ioscheduler(td))
 858                 goto err;
 859
 860         if (!td->o.create_serialize && setup_files(td))
 861                 goto err;
 862
 863         if (td_io_init(td))
 864                 goto err;
 865
 866         if (open_files(td))
 867                 goto err;
 868
 869         if (init_random_map(td))
 870                 goto err;
 871
 872         if (td->o.exec_prerun) {
 873                 if (system(td->o.exec_prerun) < 0)
 874                         goto err;
 875         }
 876
 877         fio_gettime(&td->epoch, NULL);
 878         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 879         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 880
 881         runtime[0] = runtime[1] = 0;
 882         clear_state = 0;
 883         while (keep_running(td)) {
 884                 fio_gettime(&td->start, NULL);
 885                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 886
 887                 if (td->o.ratemin)
 888                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 889
 890                 if (clear_state && clear_io_state(td))
 891                         break;
 892
 893                 prune_io_piece_log(td);
 894
 895                 do_io(td);
 896
 897                 clear_state = 1;
 898
 899                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 900                         if (td->rw_end_set[DDIR_READ])
 901                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 902                         else
 903                                 elapsed = utime_since_now(&td->start);
 904
 905                         runtime[DDIR_READ] += elapsed;
 906                 }
 907                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 908                         if (td->rw_end_set[DDIR_WRITE])
 909                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 910                         else
 911                                 elapsed = utime_since_now(&td->start);
 912
 913                         runtime[DDIR_WRITE] += elapsed;
 914                 }
 915
 916                 if (td->error || td->terminate)
 917                         break;
 918
 919                 if (!td->o.do_verify ||
 920                     td->o.verify == VERIFY_NONE ||
 921                     (td->io_ops->flags & FIO_UNIDIR))
 922                         continue;
 923
 924                 if (clear_io_state(td))
 925                         break;
 926
 927                 fio_gettime(&td->start, NULL);
 928
 929                 do_verify(td);
 930
 931                 runtime[DDIR_READ] += utime_since_now(&td->start);
 932
 933                 if (td->error || td->terminate)
 934                         break;
 935         }
 936
 937         update_rusage_stat(td);
 938         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 939         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 940         td->ts.total_run_time = mtime_since_now(&td->epoch);
 941         td->ts.io_bytes[0] = td->io_bytes[0];
 942         td->ts.io_bytes[1] = td->io_bytes[1];
 943
 944         if (td->ts.bw_log)
 945                 finish_log(td, td->ts.bw_log, "bw");
 946         if (td->ts.slat_log)
 947                 finish_log(td, td->ts.slat_log, "slat");
 948         if (td->ts.clat_log)
 949                 finish_log(td, td->ts.clat_log, "clat");
 950         if (td->o.exec_postrun) {
 951                 if (system(td->o.exec_postrun) < 0)
 952                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 953         }
 954
 955         if (exitall_on_terminate)
 956                 terminate_threads(td->groupid);
 957
 958 err:
 959         if (td->error)
 960                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 961         close_files(td);
 962         close_ioengine(td);
 963         cleanup_io_u(td);
 964
 965         /*
 966          * do this very late, it will log file closing as well
 967          */
 968         if (td->o.write_iolog_file)
 969                 write_iolog_close(td);
 970
 971         options_mem_free(td);
 972         td_set_runstate(td, TD_EXITED);
 973         return (void *) (unsigned long) td->error;
 974 }
 975
 976 /*
 977  * We cannot pass the td data into a forked process, so attach the td and
 978  * pass it to the thread worker.
 979  */
 980 static int fork_main(int shmid, int offset)
 981 {
 982         struct thread_data *td;
 983         void *data, *ret;
 984
 985         data = shmat(shmid, NULL, 0);
 986         if (data == (void *) -1) {
 987                 int __err = errno;
 988
 989                 perror("shmat");
 990                 return __err;
 991         }
 992
 993         td = data + offset * sizeof(struct thread_data);
 994         ret = thread_main(td);
 995         shmdt(data);
 996         return (int) (unsigned long) ret;
 997 }
 998
 999 /*
1000  * Run over the job map and reap the threads that have exited, if any.
1001  */
1002 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1003 {
1004         struct thread_data *td;
1005         int i, cputhreads, realthreads, pending, status, ret;
1006
1007         /*
1008          * reap exited threads (TD_EXITED -> TD_REAPED)
1009          */
1010         realthreads = pending = cputhreads = 0;
1011         for_each_td(td, i) {
1012                 int flags = 0;
1013
1014                 /*
1015                  * ->io_ops is NULL for a thread that has closed its
1016                  * io engine
1017                  */
1018                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1019                         cputhreads++;
1020                 else
1021                         realthreads++;
1022
1023                 if (!td->pid || td->runstate == TD_REAPED)
1024                         continue;
1025                 if (td->o.use_thread) {
1026                         if (td->runstate == TD_EXITED) {
1027                                 td_set_runstate(td, TD_REAPED);
1028                                 goto reaped;
1029                         }
1030                         continue;
1031                 }
1032
1033                 flags = WNOHANG;
1034                 if (td->runstate == TD_EXITED)
1035                         flags = 0;
1036
1037                 /*
1038                  * check if someone quit or got killed in an unusual way
1039                  */
1040                 ret = waitpid(td->pid, &status, flags);
1041                 if (ret < 0) {
1042                         if (errno == ECHILD) {
1043                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1044                                 td_set_runstate(td, TD_REAPED);
1045                                 goto reaped;
1046                         }
1047                         perror("waitpid");
1048                 } else if (ret == td->pid) {
1049                         if (WIFSIGNALED(status)) {
1050                                 int sig = WTERMSIG(status);
1051
1052                                 if (sig != SIGQUIT)
1053                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1054                                 td_set_runstate(td, TD_REAPED);
1055                                 goto reaped;
1056                         }
1057                         if (WIFEXITED(status)) {
1058                                 if (WEXITSTATUS(status) && !td->error)
1059                                         td->error = WEXITSTATUS(status);
1060
1061                                 td_set_runstate(td, TD_REAPED);
1062                                 goto reaped;
1063                         }
1064                 }
1065
1066                 /*
1067                  * thread is not dead, continue
1068                  */
1069                 pending++;
1070                 continue;
1071 reaped:
1072                 if (td->o.use_thread) {
1073                         long ret;
1074
1075                         dprint(FD_PROCESS, "joining tread %d\n", td->pid);
1076                         if (pthread_join(td->thread, (void *) &ret)) {
1077                                 dprint(FD_PROCESS, "join failed %ld\n", ret);
1078                                 perror("pthread_join");
1079                         }
1080                 }
1081
1082                 (*nr_running)--;
1083                 (*m_rate) -= td->o.ratemin;
1084                 (*t_rate) -= td->o.rate;
1085                 pending--;
1086
1087                 if (td->error)
1088                         exit_value++;
1089         }
1090
1091         if (*nr_running == cputhreads && !pending && realthreads)
1092                 terminate_threads(TERMINATE_ALL);
1093 }
1094
1095 /*
1096  * Main function for kicking off and reaping jobs, as needed.
1097  */
1098 static void run_threads(void)
1099 {
1100         struct thread_data *td;
1101         unsigned long spent;
1102         int i, todo, nr_running, m_rate, t_rate, nr_started;
1103
1104         if (fio_pin_memory())
1105                 return;
1106
1107         if (!terse_output) {
1108                 printf("Starting ");
1109                 if (nr_thread)
1110                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1111                 if (nr_process) {
1112                         if (nr_thread)
1113                                 printf(" and ");
1114                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1115                 }
1116                 printf("\n");
1117                 fflush(stdout);
1118         }
1119
1120         signal(SIGINT, sig_handler);
1121         signal(SIGALRM, sig_handler);
1122
1123         todo = thread_number;
1124         nr_running = 0;
1125         nr_started = 0;
1126         m_rate = t_rate = 0;
1127
1128         for_each_td(td, i) {
1129                 print_status_init(td->thread_number - 1);
1130
1131                 if (!td->o.create_serialize) {
1132                         init_disk_util(td);
1133                         continue;
1134                 }
1135
1136                 /*
1137                  * do file setup here so it happens sequentially,
1138                  * we don't want X number of threads getting their
1139                  * client data interspersed on disk
1140                  */
1141                 if (setup_files(td)) {
1142                         exit_value++;
1143                         if (td->error)
1144                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1145                         td_set_runstate(td, TD_REAPED);
1146                         todo--;
1147                 }
1148
1149                 init_disk_util(td);
1150         }
1151
1152         set_genesis_time();
1153
1154         while (todo) {
1155                 struct thread_data *map[MAX_JOBS];
1156                 struct timeval this_start;
1157                 int this_jobs = 0, left;
1158
1159                 /*
1160                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1161                  */
1162                 for_each_td(td, i) {
1163                         if (td->runstate != TD_NOT_CREATED)
1164                                 continue;
1165
1166                         /*
1167                          * never got a chance to start, killed by other
1168                          * thread for some reason
1169                          */
1170                         if (td->terminate) {
1171                                 todo--;
1172                                 continue;
1173                         }
1174
1175                         if (td->o.start_delay) {
1176                                 spent = mtime_since_genesis();
1177
1178                                 if (td->o.start_delay * 1000 > spent)
1179                                         continue;
1180                         }
1181
1182                         if (td->o.stonewall && (nr_started || nr_running))
1183                                 break;
1184
1185                         /*
1186                          * Set state to created. Thread will transition
1187                          * to TD_INITIALIZED when it's done setting up.
1188                          */
1189                         td_set_runstate(td, TD_CREATED);
1190                         map[this_jobs++] = td;
1191                         nr_started++;
1192
1193                         if (td->o.use_thread) {
1194                                 dprint(FD_PROCESS, "will pthread_create\n");
1195                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1196                                         perror("thread_create");
1197                                         nr_started--;
1198                                         break;
1199                                 }
1200                         } else {
1201                                 dprint(FD_PROCESS, "will fork\n");
1202                                 if (!fork()) {
1203                                         int ret = fork_main(shm_id, i);
1204
1205                                         exit(ret);
1206                                 }
1207                         }
1208                         fio_sem_down(startup_sem);
1209                 }
1210
1211                 /*
1212                  * Wait for the started threads to transition to
1213                  * TD_INITIALIZED.
1214                  */
1215                 fio_gettime(&this_start, NULL);
1216                 left = this_jobs;
1217                 while (left && !fio_abort) {
1218                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1219                                 break;
1220
1221                         usleep(100000);
1222
1223                         for (i = 0; i < this_jobs; i++) {
1224                                 td = map[i];
1225                                 if (!td)
1226                                         continue;
1227                                 if (td->runstate == TD_INITIALIZED) {
1228                                         map[i] = NULL;
1229                                         left--;
1230                                 } else if (td->runstate >= TD_EXITED) {
1231                                         map[i] = NULL;
1232                                         left--;
1233                                         todo--;
1234                                         nr_running++; /* work-around... */
1235                                 }
1236                         }
1237                 }
1238
1239                 if (left) {
1240                         log_err("fio: %d jobs failed to start\n", left);
1241                         for (i = 0; i < this_jobs; i++) {
1242                                 td = map[i];
1243                                 if (!td)
1244                                         continue;
1245                                 kill(td->pid, SIGTERM);
1246                         }
1247                         break;
1248                 }
1249
1250                 /*
1251                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1252                  */
1253                 for_each_td(td, i) {
1254                         if (td->runstate != TD_INITIALIZED)
1255                                 continue;
1256
1257                         td_set_runstate(td, TD_RUNNING);
1258                         nr_running++;
1259                         nr_started--;
1260                         m_rate += td->o.ratemin;
1261                         t_rate += td->o.rate;
1262                         todo--;
1263                         fio_sem_up(td->mutex);
1264                 }
1265
1266                 reap_threads(&nr_running, &t_rate, &m_rate);
1267
1268                 if (todo)
1269                         usleep(100000);
1270         }
1271
1272         while (nr_running) {
1273                 reap_threads(&nr_running, &t_rate, &m_rate);
1274                 usleep(10000);
1275         }
1276
1277         update_io_ticks();
1278         fio_unpin_memory();
1279 }
1280
1281 int main(int argc, char *argv[])
1282 {
1283         long ps;
1284
1285         /*
1286          * We need locale for number printing, if it isn't set then just
1287          * go with the US format.
1288          */
1289         if (!getenv("LC_NUMERIC"))
1290                 setlocale(LC_NUMERIC, "en_US");
1291
1292         if (parse_options(argc, argv))
1293                 return 1;
1294
1295         if (!thread_number)
1296                 return 0;
1297
1298         ps = sysconf(_SC_PAGESIZE);
1299         if (ps < 0) {
1300                 log_err("Failed to get page size\n");
1301                 return 1;
1302         }
1303
1304         page_size = ps;
1305         page_mask = ps - 1;
1306
1307         if (write_bw_log) {
1308                 setup_log(&agg_io_log[DDIR_READ]);
1309                 setup_log(&agg_io_log[DDIR_WRITE]);
1310         }
1311
1312         startup_sem = fio_sem_init(0);
1313
1314         set_genesis_time();
1315
1316         disk_util_timer_arm();
1317
1318         run_threads();
1319
1320         if (!fio_abort) {
1321                 show_run_stats();
1322                 if (write_bw_log) {
1323                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1324                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1325                 }
1326         }
1327
1328         fio_sem_remove(startup_sem);
1329         return exit_value;
1330 }