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