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_mutex *startup_mutex;
  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                         td->terminate = 1;
  78                         td->o.start_delay = 0;
  79
  80                         /*
  81                          * if the thread is running, just let it exit
  82                          */
  83                         if (td->runstate < TD_RUNNING)
  84                                 kill(td->pid, SIGQUIT);
  85                         else {
  86                                 struct ioengine_ops *ops = td->io_ops;
  87
  88                                 if (ops && (ops->flags & FIO_SIGQUIT))
  89                                         kill(td->pid, SIGQUIT);
  90                         }
  91                 }
  92         }
  93 }
  94
  95 static void sig_handler(int sig)
  96 {
  97         switch (sig) {
  98                 case SIGALRM:
  99                         update_io_ticks();
 100                         disk_util_timer_arm();
 101                         print_thread_status();
 102                         break;
 103                 default:
 104                         printf("\nfio: terminating on signal %d\n", sig);
 105                         fflush(stdout);
 106                         terminate_threads(TERMINATE_ALL);
 107                         break;
 108         }
 109 }
 110
 111 /*
 112  * Check if we are above the minimum rate given.
 113  */
 114 static int check_min_rate(struct thread_data *td, struct timeval *now)
 115 {
 116         unsigned long long bytes = 0;
 117         unsigned long iops = 0;
 118         unsigned long spent;
 119         unsigned long rate;
 120
 121         /*
 122          * No minimum rate set, always ok
 123          */
 124         if (!td->o.ratemin && !td->o.rate_iops_min)
 125                 return 0;
 126
 127         /*
 128          * allow a 2 second settle period in the beginning
 129          */
 130         if (mtime_since(&td->start, now) < 2000)
 131                 return 0;
 132
 133         if (td_read(td)) {
 134                 iops += td->io_blocks[DDIR_READ];
 135                 bytes += td->this_io_bytes[DDIR_READ];
 136         }
 137         if (td_write(td)) {
 138                 iops += td->io_blocks[DDIR_WRITE];
 139                 bytes += td->this_io_bytes[DDIR_WRITE];
 140         }
 141
 142         /*
 143          * if rate blocks is set, sample is running
 144          */
 145         if (td->rate_bytes || td->rate_blocks) {
 146                 spent = mtime_since(&td->lastrate, now);
 147                 if (spent < td->o.ratecycle)
 148                         return 0;
 149
 150                 if (td->o.rate) {
 151                         /*
 152                          * check bandwidth specified rate
 153                          */
 154                         if (bytes < td->rate_bytes) {
 155                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
 156                                 return 1;
 157                         } else {
 158                                 rate = (bytes - td->rate_bytes) / spent;
 159                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
 160                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
 161                                         return 1;
 162                                 }
 163                         }
 164                 } else {
 165                         /*
 166                          * checks iops specified rate
 167                          */
 168                         if (iops < td->o.rate_iops) {
 169                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
 170                                 return 1;
 171                         } else {
 172                                 rate = (iops - td->rate_blocks) / spent;
 173                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
 174                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
 175                                 }
 176                         }
 177                 }
 178         }
 179
 180         td->rate_bytes = bytes;
 181         td->rate_blocks = iops;
 182         memcpy(&td->lastrate, now, sizeof(*now));
 183         return 0;
 184 }
 185
 186 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 187 {
 188         if (!td->o.timeout)
 189                 return 0;
 190         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 191                 return 1;
 192
 193         return 0;
 194 }
 195
 196 /*
 197  * When job exits, we can cancel the in-flight IO if we are using async
 198  * io. Attempt to do so.
 199  */
 200 static void cleanup_pending_aio(struct thread_data *td)
 201 {
 202         struct list_head *entry, *n;
 203         struct io_u *io_u;
 204         int r;
 205
 206         /*
 207          * get immediately available events, if any
 208          */
 209         r = io_u_queued_complete(td, 0);
 210         if (r < 0)
 211                 return;
 212
 213         /*
 214          * now cancel remaining active events
 215          */
 216         if (td->io_ops->cancel) {
 217                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 218                         io_u = list_entry(entry, struct io_u, list);
 219
 220                         /*
 221                          * if the io_u isn't in flight, then that generally
 222                          * means someone leaked an io_u. complain but fix
 223                          * it up, so we don't stall here.
 224                          */
 225                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 226                                 log_err("fio: non-busy IO on busy list\n");
 227                                 put_io_u(td, io_u);
 228                         } else {
 229                                 r = td->io_ops->cancel(td, io_u);
 230                                 if (!r)
 231                                         put_io_u(td, io_u);
 232                         }
 233                 }
 234         }
 235
 236         if (td->cur_depth)
 237                 r = io_u_queued_complete(td, td->cur_depth);
 238 }
 239
 240 /*
 241  * Helper to handle the final sync of a file. Works just like the normal
 242  * io path, just does everything sync.
 243  */
 244 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 245 {
 246         struct io_u *io_u = __get_io_u(td);
 247         int ret;
 248
 249         if (!io_u)
 250                 return 1;
 251
 252         io_u->ddir = DDIR_SYNC;
 253         io_u->file = f;
 254
 255         if (td_io_prep(td, io_u)) {
 256                 put_io_u(td, io_u);
 257                 return 1;
 258         }
 259
 260 requeue:
 261         ret = td_io_queue(td, io_u);
 262         if (ret < 0) {
 263                 td_verror(td, io_u->error, "td_io_queue");
 264                 put_io_u(td, io_u);
 265                 return 1;
 266         } else if (ret == FIO_Q_QUEUED) {
 267                 if (io_u_queued_complete(td, 1) < 0)
 268                         return 1;
 269         } else if (ret == FIO_Q_COMPLETED) {
 270                 if (io_u->error) {
 271                         td_verror(td, io_u->error, "td_io_queue");
 272                         return 1;
 273                 }
 274
 275                 if (io_u_sync_complete(td, io_u) < 0)
 276                         return 1;
 277         } else if (ret == FIO_Q_BUSY) {
 278                 if (td_io_commit(td))
 279                         return 1;
 280                 goto requeue;
 281         }
 282
 283         return 0;
 284 }
 285
 286 /*
 287  * The main verify engine. Runs over the writes we previously submitted,
 288  * reads the blocks back in, and checks the crc/md5 of the data.
 289  */
 290 static void do_verify(struct thread_data *td)
 291 {
 292         struct fio_file *f;
 293         struct io_u *io_u;
 294         int ret, min_events;
 295         unsigned int i;
 296
 297         /*
 298          * sync io first and invalidate cache, to make sure we really
 299          * read from disk.
 300          */
 301         for_each_file(td, f, i) {
 302                 if (!(f->flags & FIO_FILE_OPEN))
 303                         continue;
 304                 if (fio_io_sync(td, f))
 305                         break;
 306                 if (file_invalidate_cache(td, f))
 307                         break;
 308         }
 309
 310         if (td->error)
 311                 return;
 312
 313         td_set_runstate(td, TD_VERIFYING);
 314
 315         io_u = NULL;
 316         while (!td->terminate) {
 317                 int ret2;
 318
 319                 io_u = __get_io_u(td);
 320                 if (!io_u)
 321                         break;
 322
 323                 if (runtime_exceeded(td, &io_u->start_time)) {
 324                         put_io_u(td, io_u);
 325                         td->terminate = 1;
 326                         break;
 327                 }
 328
 329                 if (get_next_verify(td, io_u)) {
 330                         put_io_u(td, io_u);
 331                         break;
 332                 }
 333
 334                 if (td_io_prep(td, io_u)) {
 335                         put_io_u(td, io_u);
 336                         break;
 337                 }
 338
 339                 io_u->end_io = verify_io_u;
 340
 341                 ret = td_io_queue(td, io_u);
 342                 switch (ret) {
 343                 case FIO_Q_COMPLETED:
 344                         if (io_u->error)
 345                                 ret = -io_u->error;
 346                         else if (io_u->resid) {
 347                                 int bytes = io_u->xfer_buflen - io_u->resid;
 348                                 struct fio_file *f = io_u->file;
 349
 350                                 /*
 351                                  * zero read, fail
 352                                  */
 353                                 if (!bytes) {
 354                                         td_verror(td, ENODATA, "full resid");
 355                                         put_io_u(td, io_u);
 356                                         break;
 357                                 }
 358
 359                                 io_u->xfer_buflen = io_u->resid;
 360                                 io_u->xfer_buf += bytes;
 361                                 io_u->offset += bytes;
 362
 363                                 td->ts.short_io_u[io_u->ddir]++;
 364
 365                                 if (io_u->offset == f->real_file_size)
 366                                         goto sync_done;
 367
 368                                 requeue_io_u(td, &io_u);
 369                         } else {
 370 sync_done:
 371                                 ret = io_u_sync_complete(td, io_u);
 372                                 if (ret < 0)
 373                                         break;
 374                         }
 375                         continue;
 376                 case FIO_Q_QUEUED:
 377                         break;
 378                 case FIO_Q_BUSY:
 379                         requeue_io_u(td, &io_u);
 380                         ret2 = td_io_commit(td);
 381                         if (ret2 < 0)
 382                                 ret = ret2;
 383                         break;
 384                 default:
 385                         assert(ret < 0);
 386                         td_verror(td, -ret, "td_io_queue");
 387                         break;
 388                 }
 389
 390                 if (ret < 0 || td->error)
 391                         break;
 392
 393                 /*
 394                  * if we can queue more, do so. but check if there are
 395                  * completed io_u's first.
 396                  */
 397                 min_events = 0;
 398                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 399                         min_events = 1;
 400
 401                         if (td->cur_depth > td->o.iodepth_low)
 402                                 min_events = td->cur_depth - td->o.iodepth_low;
 403                 }
 404
 405                 /*
 406                  * Reap required number of io units, if any, and do the
 407                  * verification on them through the callback handler
 408                  */
 409                 if (io_u_queued_complete(td, min_events) < 0)
 410                         break;
 411         }
 412
 413         if (!td->error) {
 414                 min_events = td->cur_depth;
 415
 416                 if (min_events)
 417                         ret = io_u_queued_complete(td, min_events);
 418         } else
 419                 cleanup_pending_aio(td);
 420
 421         td_set_runstate(td, TD_RUNNING);
 422 }
 423
 424 /*
 425  * Main IO worker function. It retrieves io_u's to process and queues
 426  * and reaps them, checking for rate and errors along the way.
 427  */
 428 static void do_io(struct thread_data *td)
 429 {
 430         struct timeval s;
 431         unsigned long usec;
 432         unsigned int i;
 433         int ret = 0;
 434
 435         td_set_runstate(td, TD_RUNNING);
 436
 437         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 438                 struct timeval comp_time;
 439                 long bytes_done = 0;
 440                 int min_evts = 0;
 441                 struct io_u *io_u;
 442                 int ret2;
 443
 444                 if (td->terminate)
 445                         break;
 446
 447                 io_u = get_io_u(td);
 448                 if (!io_u)
 449                         break;
 450
 451                 memcpy(&s, &io_u->start_time, sizeof(s));
 452
 453                 if (runtime_exceeded(td, &s)) {
 454                         put_io_u(td, io_u);
 455                         td->terminate = 1;
 456                         break;
 457                 }
 458
 459                 /*
 460                  * Add verification end_io handler, if asked to verify
 461                  * a previously written file.
 462                  */
 463                 if (td->o.verify != VERIFY_NONE)
 464                         io_u->end_io = verify_io_u;
 465
 466                 ret = td_io_queue(td, io_u);
 467                 switch (ret) {
 468                 case FIO_Q_COMPLETED:
 469                         if (io_u->error)
 470                                 ret = -io_u->error;
 471                         else if (io_u->resid) {
 472                                 int bytes = io_u->xfer_buflen - io_u->resid;
 473                                 struct fio_file *f = io_u->file;
 474
 475                                 /*
 476                                  * zero read, fail
 477                                  */
 478                                 if (!bytes) {
 479                                         td_verror(td, ENODATA, "full resid");
 480                                         put_io_u(td, io_u);
 481                                         break;
 482                                 }
 483
 484                                 io_u->xfer_buflen = io_u->resid;
 485                                 io_u->xfer_buf += bytes;
 486                                 io_u->offset += bytes;
 487
 488                                 td->ts.short_io_u[io_u->ddir]++;
 489
 490                                 if (io_u->offset == f->real_file_size)
 491                                         goto sync_done;
 492
 493                                 requeue_io_u(td, &io_u);
 494                         } else {
 495 sync_done:
 496                                 fio_gettime(&comp_time, NULL);
 497                                 bytes_done = io_u_sync_complete(td, io_u);
 498                                 if (bytes_done < 0)
 499                                         ret = bytes_done;
 500                         }
 501                         break;
 502                 case FIO_Q_QUEUED:
 503                         /*
 504                          * if the engine doesn't have a commit hook,
 505                          * the io_u is really queued. if it does have such
 506                          * a hook, it has to call io_u_queued() itself.
 507                          */
 508                         if (td->io_ops->commit == NULL)
 509                                 io_u_queued(td, io_u);
 510                         break;
 511                 case FIO_Q_BUSY:
 512                         requeue_io_u(td, &io_u);
 513                         ret2 = td_io_commit(td);
 514                         if (ret2 < 0)
 515                                 ret = ret2;
 516                         break;
 517                 default:
 518                         assert(ret < 0);
 519                         put_io_u(td, io_u);
 520                         break;
 521                 }
 522
 523                 if (ret < 0 || td->error)
 524                         break;
 525
 526                 /*
 527                  * See if we need to complete some commands
 528                  */
 529                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 530                         min_evts = 0;
 531                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 532                                 min_evts = 1;
 533
 534                                 if (td->cur_depth > td->o.iodepth_low)
 535                                         min_evts = td->cur_depth - td->o.iodepth_low;
 536                         }
 537
 538                         fio_gettime(&comp_time, NULL);
 539                         bytes_done = io_u_queued_complete(td, min_evts);
 540                         if (bytes_done < 0)
 541                                 break;
 542                 }
 543
 544                 if (!bytes_done)
 545                         continue;
 546
 547                 /*
 548                  * the rate is batched for now, it should work for batches
 549                  * of completions except the very first one which may look
 550                  * a little bursty
 551                  */
 552                 usec = utime_since(&s, &comp_time);
 553
 554                 rate_throttle(td, usec, bytes_done);
 555
 556                 if (check_min_rate(td, &comp_time)) {
 557                         if (exitall_on_terminate)
 558                                 terminate_threads(td->groupid);
 559                         td_verror(td, ENODATA, "check_min_rate");
 560                         break;
 561                 }
 562
 563                 if (td->o.thinktime) {
 564                         unsigned long long b;
 565
 566                         b = td->io_blocks[0] + td->io_blocks[1];
 567                         if (!(b % td->o.thinktime_blocks)) {
 568                                 int left;
 569
 570                                 if (td->o.thinktime_spin)
 571                                         __usec_sleep(td->o.thinktime_spin);
 572
 573                                 left = td->o.thinktime - td->o.thinktime_spin;
 574                                 if (left)
 575                                         usec_sleep(td, left);
 576                         }
 577                 }
 578         }
 579
 580         if (td->o.fill_device && td->error == ENOSPC) {
 581                 td->error = 0;
 582                 td->terminate = 1;
 583         }
 584         if (!td->error) {
 585                 struct fio_file *f;
 586
 587                 i = td->cur_depth;
 588                 if (i)
 589                         ret = io_u_queued_complete(td, i);
 590
 591                 if (should_fsync(td) && td->o.end_fsync) {
 592                         td_set_runstate(td, TD_FSYNCING);
 593
 594                         for_each_file(td, f, i) {
 595                                 if (!(f->flags & FIO_FILE_OPEN))
 596                                         continue;
 597                                 fio_io_sync(td, f);
 598                         }
 599                 }
 600         } else
 601                 cleanup_pending_aio(td);
 602
 603         /*
 604          * stop job if we failed doing any IO
 605          */
 606         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 607                 td->done = 1;
 608 }
 609
 610 static void cleanup_io_u(struct thread_data *td)
 611 {
 612         struct list_head *entry, *n;
 613         struct io_u *io_u;
 614
 615         list_for_each_safe(entry, n, &td->io_u_freelist) {
 616                 io_u = list_entry(entry, struct io_u, list);
 617
 618                 list_del(&io_u->list);
 619                 free(io_u);
 620         }
 621
 622         free_io_mem(td);
 623 }
 624
 625 /*
 626  * "randomly" fill the buffer contents
 627  */
 628 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 629 {
 630         long *ptr = io_u->buf;
 631
 632         if (!td->o.zero_buffers) {
 633                 while ((void *) ptr - io_u->buf < max_bs) {
 634                         *ptr = rand() * GOLDEN_RATIO_PRIME;
 635                         ptr++;
 636                 }
 637         } else
 638                 memset(ptr, 0, max_bs);
 639 }
 640
 641 static int init_io_u(struct thread_data *td)
 642 {
 643         struct io_u *io_u;
 644         unsigned int max_bs;
 645         int i, max_units;
 646         char *p;
 647
 648         max_units = td->o.iodepth;
 649         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 650         td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 651
 652         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 653                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 654
 655         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 656                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 657                 return 1;
 658         }
 659
 660         if (allocate_io_mem(td))
 661                 return 1;
 662
 663         if (td->o.odirect)
 664                 p = ALIGN(td->orig_buffer);
 665         else
 666                 p = td->orig_buffer;
 667
 668         for (i = 0; i < max_units; i++) {
 669                 if (td->terminate)
 670                         return 1;
 671                 io_u = malloc(sizeof(*io_u));
 672                 memset(io_u, 0, sizeof(*io_u));
 673                 INIT_LIST_HEAD(&io_u->list);
 674
 675                 if (!(td->io_ops->flags & FIO_NOIO)) {
 676                         io_u->buf = p + max_bs * i;
 677
 678                         if (td_write(td))
 679                                 fill_io_buf(td, io_u, max_bs);
 680                 }
 681
 682                 io_u->index = i;
 683                 io_u->flags = IO_U_F_FREE;
 684                 list_add(&io_u->list, &td->io_u_freelist);
 685         }
 686
 687         io_u_init_timeout();
 688
 689         return 0;
 690 }
 691
 692 static int switch_ioscheduler(struct thread_data *td)
 693 {
 694         char tmp[256], tmp2[128];
 695         FILE *f;
 696         int ret;
 697
 698         if (td->io_ops->flags & FIO_DISKLESSIO)
 699                 return 0;
 700
 701         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 702
 703         f = fopen(tmp, "r+");
 704         if (!f) {
 705                 if (errno == ENOENT) {
 706                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 707                         return 0;
 708                 }
 709                 td_verror(td, errno, "fopen iosched");
 710                 return 1;
 711         }
 712
 713         /*
 714          * Set io scheduler.
 715          */
 716         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 717         if (ferror(f) || ret != 1) {
 718                 td_verror(td, errno, "fwrite");
 719                 fclose(f);
 720                 return 1;
 721         }
 722
 723         rewind(f);
 724
 725         /*
 726          * Read back and check that the selected scheduler is now the default.
 727          */
 728         ret = fread(tmp, 1, sizeof(tmp), f);
 729         if (ferror(f) || ret < 0) {
 730                 td_verror(td, errno, "fread");
 731                 fclose(f);
 732                 return 1;
 733         }
 734
 735         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 736         if (!strstr(tmp, tmp2)) {
 737                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 738                 td_verror(td, EINVAL, "iosched_switch");
 739                 fclose(f);
 740                 return 1;
 741         }
 742
 743         fclose(f);
 744         return 0;
 745 }
 746
 747 static int keep_running(struct thread_data *td)
 748 {
 749         unsigned long long io_done;
 750
 751         if (td->done)
 752                 return 0;
 753         if (td->o.time_based)
 754                 return 1;
 755         if (td->o.loops) {
 756                 td->o.loops--;
 757                 return 1;
 758         }
 759
 760         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
 761         if (io_done < td->o.size)
 762                 return 1;
 763
 764         return 0;
 765 }
 766
 767 static int clear_io_state(struct thread_data *td)
 768 {
 769         struct fio_file *f;
 770         unsigned int i;
 771         int ret;
 772
 773         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 774         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 775         td->zone_bytes = 0;
 776         td->rate_bytes = 0;
 777         td->rate_blocks = 0;
 778         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 779
 780         td->last_was_sync = 0;
 781
 782         /*
 783          * reset file done count if we are to start over
 784          */
 785         if (td->o.time_based || td->o.loops)
 786                 td->nr_done_files = 0;
 787
 788         close_files(td);
 789
 790         ret = 0;
 791         for_each_file(td, f, i) {
 792                 f->flags &= ~FIO_FILE_DONE;
 793                 ret = td_io_open_file(td, f);
 794                 if (ret)
 795                         break;
 796         }
 797
 798         return ret;
 799 }
 800
 801 /*
 802  * Entry point for the thread based jobs. The process based jobs end up
 803  * here as well, after a little setup.
 804  */
 805 static void *thread_main(void *data)
 806 {
 807         unsigned long long runtime[2], elapsed;
 808         struct thread_data *td = data;
 809         int clear_state;
 810
 811         if (!td->o.use_thread)
 812                 setsid();
 813
 814         td->pid = getpid();
 815
 816         dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
 817
 818         INIT_LIST_HEAD(&td->io_u_freelist);
 819         INIT_LIST_HEAD(&td->io_u_busylist);
 820         INIT_LIST_HEAD(&td->io_u_requeues);
 821         INIT_LIST_HEAD(&td->io_log_list);
 822         INIT_LIST_HEAD(&td->io_hist_list);
 823         td->io_hist_tree = RB_ROOT;
 824
 825         td_set_runstate(td, TD_INITIALIZED);
 826         fio_mutex_up(startup_mutex);
 827         fio_mutex_down(td->mutex);
 828
 829         /*
 830          * the ->mutex mutex is now no longer used, close it to avoid
 831          * eating a file descriptor
 832          */
 833         fio_mutex_remove(td->mutex);
 834
 835         /*
 836          * May alter parameters that init_io_u() will use, so we need to
 837          * do this first.
 838          */
 839         if (init_iolog(td))
 840                 goto err;
 841
 842         if (init_io_u(td))
 843                 goto err;
 844
 845         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 846                 td_verror(td, errno, "cpu_set_affinity");
 847                 goto err;
 848         }
 849
 850         if (td->ioprio_set) {
 851                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 852                         td_verror(td, errno, "ioprio_set");
 853                         goto err;
 854                 }
 855         }
 856
 857         if (nice(td->o.nice) == -1) {
 858                 td_verror(td, errno, "nice");
 859                 goto err;
 860         }
 861
 862         if (td->o.ioscheduler && switch_ioscheduler(td))
 863                 goto err;
 864
 865         if (!td->o.create_serialize && setup_files(td))
 866                 goto err;
 867
 868         if (td_io_init(td))
 869                 goto err;
 870
 871         if (open_files(td))
 872                 goto err;
 873
 874         if (init_random_map(td))
 875                 goto err;
 876
 877         if (td->o.exec_prerun) {
 878                 if (system(td->o.exec_prerun) < 0)
 879                         goto err;
 880         }
 881
 882         fio_gettime(&td->epoch, NULL);
 883         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 884         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 885
 886         runtime[0] = runtime[1] = 0;
 887         clear_state = 0;
 888         while (keep_running(td)) {
 889                 fio_gettime(&td->start, NULL);
 890                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 891
 892                 if (td->o.ratemin)
 893                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 894
 895                 if (clear_state && clear_io_state(td))
 896                         break;
 897
 898                 prune_io_piece_log(td);
 899
 900                 do_io(td);
 901
 902                 clear_state = 1;
 903
 904                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 905                         if (td->rw_end_set[DDIR_READ])
 906                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 907                         else
 908                                 elapsed = utime_since_now(&td->start);
 909
 910                         runtime[DDIR_READ] += elapsed;
 911                 }
 912                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 913                         if (td->rw_end_set[DDIR_WRITE])
 914                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 915                         else
 916                                 elapsed = utime_since_now(&td->start);
 917
 918                         runtime[DDIR_WRITE] += elapsed;
 919                 }
 920
 921                 if (td->error || td->terminate)
 922                         break;
 923
 924                 if (!td->o.do_verify ||
 925                     td->o.verify == VERIFY_NONE ||
 926                     (td->io_ops->flags & FIO_UNIDIR))
 927                         continue;
 928
 929                 if (clear_io_state(td))
 930                         break;
 931
 932                 fio_gettime(&td->start, NULL);
 933
 934                 do_verify(td);
 935
 936                 runtime[DDIR_READ] += utime_since_now(&td->start);
 937
 938                 if (td->error || td->terminate)
 939                         break;
 940         }
 941
 942         update_rusage_stat(td);
 943         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 944         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 945         td->ts.total_run_time = mtime_since_now(&td->epoch);
 946         td->ts.io_bytes[0] = td->io_bytes[0];
 947         td->ts.io_bytes[1] = td->io_bytes[1];
 948
 949         if (td->ts.bw_log)
 950                 finish_log(td, td->ts.bw_log, "bw");
 951         if (td->ts.slat_log)
 952                 finish_log(td, td->ts.slat_log, "slat");
 953         if (td->ts.clat_log)
 954                 finish_log(td, td->ts.clat_log, "clat");
 955         if (td->o.exec_postrun) {
 956                 if (system(td->o.exec_postrun) < 0)
 957                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 958         }
 959
 960         if (exitall_on_terminate)
 961                 terminate_threads(td->groupid);
 962
 963 err:
 964         if (td->error)
 965                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 966         close_and_free_files(td);
 967         close_ioengine(td);
 968         cleanup_io_u(td);
 969
 970         /*
 971          * do this very late, it will log file closing as well
 972          */
 973         if (td->o.write_iolog_file)
 974                 write_iolog_close(td);
 975
 976         options_mem_free(td);
 977         td_set_runstate(td, TD_EXITED);
 978         return (void *) (unsigned long) td->error;
 979 }
 980
 981 /*
 982  * We cannot pass the td data into a forked process, so attach the td and
 983  * pass it to the thread worker.
 984  */
 985 static int fork_main(int shmid, int offset)
 986 {
 987         struct thread_data *td;
 988         void *data, *ret;
 989
 990         data = shmat(shmid, NULL, 0);
 991         if (data == (void *) -1) {
 992                 int __err = errno;
 993
 994                 perror("shmat");
 995                 return __err;
 996         }
 997
 998         td = data + offset * sizeof(struct thread_data);
 999         ret = thread_main(td);
1000         shmdt(data);
1001         return (int) (unsigned long) ret;
1002 }
1003
1004 /*
1005  * Run over the job map and reap the threads that have exited, if any.
1006  */
1007 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1008 {
1009         struct thread_data *td;
1010         int i, cputhreads, realthreads, pending, status, ret;
1011
1012         /*
1013          * reap exited threads (TD_EXITED -> TD_REAPED)
1014          */
1015         realthreads = pending = cputhreads = 0;
1016         for_each_td(td, i) {
1017                 int flags = 0;
1018
1019                 /*
1020                  * ->io_ops is NULL for a thread that has closed its
1021                  * io engine
1022                  */
1023                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1024                         cputhreads++;
1025                 else
1026                         realthreads++;
1027
1028                 if (!td->pid) {
1029                         pending++;
1030                         continue;
1031                 }
1032                 if (td->runstate == TD_REAPED)
1033                         continue;
1034                 if (td->o.use_thread) {
1035                         if (td->runstate == TD_EXITED) {
1036                                 td_set_runstate(td, TD_REAPED);
1037                                 goto reaped;
1038                         }
1039                         continue;
1040                 }
1041
1042                 flags = WNOHANG;
1043                 if (td->runstate == TD_EXITED)
1044                         flags = 0;
1045
1046                 /*
1047                  * check if someone quit or got killed in an unusual way
1048                  */
1049                 ret = waitpid(td->pid, &status, flags);
1050                 if (ret < 0) {
1051                         if (errno == ECHILD) {
1052                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1053                                 td_set_runstate(td, TD_REAPED);
1054                                 goto reaped;
1055                         }
1056                         perror("waitpid");
1057                 } else if (ret == td->pid) {
1058                         if (WIFSIGNALED(status)) {
1059                                 int sig = WTERMSIG(status);
1060
1061                                 if (sig != SIGQUIT)
1062                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1063                                 td_set_runstate(td, TD_REAPED);
1064                                 goto reaped;
1065                         }
1066                         if (WIFEXITED(status)) {
1067                                 if (WEXITSTATUS(status) && !td->error)
1068                                         td->error = WEXITSTATUS(status);
1069
1070                                 td_set_runstate(td, TD_REAPED);
1071                                 goto reaped;
1072                         }
1073                 }
1074
1075                 /*
1076                  * thread is not dead, continue
1077                  */
1078                 pending++;
1079                 continue;
1080 reaped:
1081                 if (td->o.use_thread) {
1082                         long ret;
1083
1084                         dprint(FD_PROCESS, "joining tread %d\n", td->pid);
1085                         if (pthread_join(td->thread, (void *) &ret)) {
1086                                 dprint(FD_PROCESS, "join failed %ld\n", ret);
1087                                 perror("pthread_join");
1088                         }
1089                 }
1090
1091                 (*nr_running)--;
1092                 (*m_rate) -= td->o.ratemin;
1093                 (*t_rate) -= td->o.rate;
1094                 pending--;
1095
1096                 if (td->error)
1097                         exit_value++;
1098         }
1099
1100         if (*nr_running == cputhreads && !pending && realthreads)
1101                 terminate_threads(TERMINATE_ALL);
1102 }
1103
1104 /*
1105  * Main function for kicking off and reaping jobs, as needed.
1106  */
1107 static void run_threads(void)
1108 {
1109         struct thread_data *td;
1110         unsigned long spent;
1111         int i, todo, nr_running, m_rate, t_rate, nr_started;
1112
1113         if (fio_pin_memory())
1114                 return;
1115
1116         if (!terse_output) {
1117                 printf("Starting ");
1118                 if (nr_thread)
1119                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1120                 if (nr_process) {
1121                         if (nr_thread)
1122                                 printf(" and ");
1123                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1124                 }
1125                 printf("\n");
1126                 fflush(stdout);
1127         }
1128
1129         signal(SIGINT, sig_handler);
1130         signal(SIGALRM, sig_handler);
1131
1132         todo = thread_number;
1133         nr_running = 0;
1134         nr_started = 0;
1135         m_rate = t_rate = 0;
1136
1137         for_each_td(td, i) {
1138                 print_status_init(td->thread_number - 1);
1139
1140                 if (!td->o.create_serialize) {
1141                         init_disk_util(td);
1142                         continue;
1143                 }
1144
1145                 /*
1146                  * do file setup here so it happens sequentially,
1147                  * we don't want X number of threads getting their
1148                  * client data interspersed on disk
1149                  */
1150                 if (setup_files(td)) {
1151                         exit_value++;
1152                         if (td->error)
1153                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1154                         td_set_runstate(td, TD_REAPED);
1155                         todo--;
1156                 }
1157
1158                 init_disk_util(td);
1159         }
1160
1161         set_genesis_time();
1162
1163         while (todo) {
1164                 struct thread_data *map[MAX_JOBS];
1165                 struct timeval this_start;
1166                 int this_jobs = 0, left;
1167
1168                 /*
1169                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1170                  */
1171                 for_each_td(td, i) {
1172                         if (td->runstate != TD_NOT_CREATED)
1173                                 continue;
1174
1175                         /*
1176                          * never got a chance to start, killed by other
1177                          * thread for some reason
1178                          */
1179                         if (td->terminate) {
1180                                 todo--;
1181                                 continue;
1182                         }
1183
1184                         if (td->o.start_delay) {
1185                                 spent = mtime_since_genesis();
1186
1187                                 if (td->o.start_delay * 1000 > spent)
1188                                         continue;
1189                         }
1190
1191                         if (td->o.stonewall && (nr_started || nr_running))
1192                                 break;
1193
1194                         /*
1195                          * Set state to created. Thread will transition
1196                          * to TD_INITIALIZED when it's done setting up.
1197                          */
1198                         td_set_runstate(td, TD_CREATED);
1199                         map[this_jobs++] = td;
1200                         nr_started++;
1201
1202                         if (td->o.use_thread) {
1203                                 dprint(FD_PROCESS, "will pthread_create\n");
1204                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1205                                         perror("thread_create");
1206                                         nr_started--;
1207                                         break;
1208                                 }
1209                         } else {
1210                                 dprint(FD_PROCESS, "will fork\n");
1211                                 if (!fork()) {
1212                                         int ret = fork_main(shm_id, i);
1213
1214                                         exit(ret);
1215                                 }
1216                         }
1217                         fio_mutex_down(startup_mutex);
1218                 }
1219
1220                 /*
1221                  * Wait for the started threads to transition to
1222                  * TD_INITIALIZED.
1223                  */
1224                 fio_gettime(&this_start, NULL);
1225                 left = this_jobs;
1226                 while (left && !fio_abort) {
1227                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1228                                 break;
1229
1230                         usleep(100000);
1231
1232                         for (i = 0; i < this_jobs; i++) {
1233                                 td = map[i];
1234                                 if (!td)
1235                                         continue;
1236                                 if (td->runstate == TD_INITIALIZED) {
1237                                         map[i] = NULL;
1238                                         left--;
1239                                 } else if (td->runstate >= TD_EXITED) {
1240                                         map[i] = NULL;
1241                                         left--;
1242                                         todo--;
1243                                         nr_running++; /* work-around... */
1244                                 }
1245                         }
1246                 }
1247
1248                 if (left) {
1249                         log_err("fio: %d jobs failed to start\n", left);
1250                         for (i = 0; i < this_jobs; i++) {
1251                                 td = map[i];
1252                                 if (!td)
1253                                         continue;
1254                                 kill(td->pid, SIGTERM);
1255                         }
1256                         break;
1257                 }
1258
1259                 /*
1260                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1261                  */
1262                 for_each_td(td, i) {
1263                         if (td->runstate != TD_INITIALIZED)
1264                                 continue;
1265
1266                         td_set_runstate(td, TD_RUNNING);
1267                         nr_running++;
1268                         nr_started--;
1269                         m_rate += td->o.ratemin;
1270                         t_rate += td->o.rate;
1271                         todo--;
1272                         fio_mutex_up(td->mutex);
1273                 }
1274
1275                 reap_threads(&nr_running, &t_rate, &m_rate);
1276
1277                 if (todo)
1278                         usleep(100000);
1279         }
1280
1281         while (nr_running) {
1282                 reap_threads(&nr_running, &t_rate, &m_rate);
1283                 usleep(10000);
1284         }
1285
1286         update_io_ticks();
1287         fio_unpin_memory();
1288 }
1289
1290 int main(int argc, char *argv[])
1291 {
1292         long ps;
1293
1294         /*
1295          * We need locale for number printing, if it isn't set then just
1296          * go with the US format.
1297          */
1298         if (!getenv("LC_NUMERIC"))
1299                 setlocale(LC_NUMERIC, "en_US");
1300
1301         if (parse_options(argc, argv))
1302                 return 1;
1303
1304         if (!thread_number)
1305                 return 0;
1306
1307         ps = sysconf(_SC_PAGESIZE);
1308         if (ps < 0) {
1309                 log_err("Failed to get page size\n");
1310                 return 1;
1311         }
1312
1313         page_size = ps;
1314         page_mask = ps - 1;
1315
1316         if (write_bw_log) {
1317                 setup_log(&agg_io_log[DDIR_READ]);
1318                 setup_log(&agg_io_log[DDIR_WRITE]);
1319         }
1320
1321         startup_mutex = fio_mutex_init(0);
1322
1323         set_genesis_time();
1324
1325         disk_util_timer_arm();
1326
1327         run_threads();
1328
1329         if (!fio_abort) {
1330                 show_run_stats();
1331                 if (write_bw_log) {
1332                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1333                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1334                 }
1335         }
1336
1337         fio_mutex_remove(startup_mutex);
1338         return exit_value;
1339 }