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 <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <sys/stat.h>
  34 #include <sys/wait.h>
  35 #include <sys/ipc.h>
  36 #include <sys/shm.h>
  37 #include <sys/mman.h>
  38
  39 #include "fio.h"
  40 #include "hash.h"
  41 #include "smalloc.h"
  42 #include "verify.h"
  43 #include "trim.h"
  44 #include "diskutil.h"
  45 #include "cgroup.h"
  46 #include "profile.h"
  47 #include "lib/rand.h"
  48 #include "memalign.h"
  49
  50 unsigned long page_mask;
  51 unsigned long page_size;
  52
  53 #define PAGE_ALIGN(buf) \
  54         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  55
  56 int groupid = 0;
  57 int thread_number = 0;
  58 int nr_process = 0;
  59 int nr_thread = 0;
  60 int shm_id = 0;
  61 int temp_stall_ts;
  62 unsigned long done_secs = 0;
  63
  64 static struct fio_mutex *startup_mutex;
  65 static struct fio_mutex *writeout_mutex;
  66 static volatile int fio_abort;
  67 static int exit_value;
  68 static pthread_t gtod_thread;
  69 static pthread_t disk_util_thread;
  70 static struct flist_head *cgroup_list;
  71 static char *cgroup_mnt;
  72
  73 struct io_log *agg_io_log[2];
  74
  75 #define TERMINATE_ALL           (-1)
  76 #define JOB_START_TIMEOUT       (5 * 1000)
  77
  78 void td_set_runstate(struct thread_data *td, int runstate)
  79 {
  80         if (td->runstate == runstate)
  81                 return;
  82
  83         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  84                                                 td->runstate, runstate);
  85         td->runstate = runstate;
  86 }
  87
  88 static void terminate_threads(int group_id)
  89 {
  90         struct thread_data *td;
  91         int i;
  92
  93         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  94
  95         for_each_td(td, i) {
  96                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  97                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  98                                                 td->o.name, (int) td->pid);
  99                         td->terminate = 1;
 100                         td->o.start_delay = 0;
 101
 102                         /*
 103                          * if the thread is running, just let it exit
 104                          */
 105                         if (!td->pid)
 106                                 continue;
 107                         else if (td->runstate < TD_RAMP)
 108                                 kill(td->pid, SIGTERM);
 109                         else {
 110                                 struct ioengine_ops *ops = td->io_ops;
 111
 112                                 if (ops && (ops->flags & FIO_SIGTERM))
 113                                         kill(td->pid, SIGTERM);
 114                         }
 115                 }
 116         }
 117 }
 118
 119 static void sig_int(int sig)
 120 {
 121         if (threads) {
 122                 log_info("\nfio: terminating on signal %d\n", sig);
 123                 fflush(stdout);
 124                 exit_value = 128;
 125                 terminate_threads(TERMINATE_ALL);
 126         }
 127 }
 128
 129 static void *disk_thread_main(void *data)
 130 {
 131         fio_mutex_up(startup_mutex);
 132
 133         while (threads) {
 134                 usleep(DISK_UTIL_MSEC * 1000);
 135                 if (!threads)
 136                         break;
 137                 update_io_ticks();
 138                 print_thread_status();
 139         }
 140
 141         return NULL;
 142 }
 143
 144 static int create_disk_util_thread(void)
 145 {
 146         int ret;
 147
 148         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 149         if (ret) {
 150                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 151                 return 1;
 152         }
 153
 154         ret = pthread_detach(disk_util_thread);
 155         if (ret) {
 156                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 157                 return 1;
 158         }
 159
 160         dprint(FD_MUTEX, "wait on startup_mutex\n");
 161         fio_mutex_down(startup_mutex);
 162         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 163         return 0;
 164 }
 165
 166 static void set_sig_handlers(void)
 167 {
 168         struct sigaction act;
 169
 170         memset(&act, 0, sizeof(act));
 171         act.sa_handler = sig_int;
 172         act.sa_flags = SA_RESTART;
 173         sigaction(SIGINT, &act, NULL);
 174
 175         memset(&act, 0, sizeof(act));
 176         act.sa_handler = sig_int;
 177         act.sa_flags = SA_RESTART;
 178         sigaction(SIGTERM, &act, NULL);
 179 }
 180
 181 /*
 182  * Check if we are above the minimum rate given.
 183  */
 184 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 185                             enum fio_ddir ddir)
 186 {
 187         unsigned long long bytes = 0;
 188         unsigned long iops = 0;
 189         unsigned long spent;
 190         unsigned long rate;
 191         unsigned int ratemin = 0;
 192         unsigned int rate_iops = 0;
 193         unsigned int rate_iops_min = 0;
 194
 195         assert(ddir_rw(ddir));
 196
 197         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 198                 return 0;
 199
 200         /*
 201          * allow a 2 second settle period in the beginning
 202          */
 203         if (mtime_since(&td->start, now) < 2000)
 204                 return 0;
 205
 206         iops += td->io_blocks[ddir];
 207         bytes += td->this_io_bytes[ddir];
 208         ratemin += td->o.ratemin[ddir];
 209         rate_iops += td->o.rate_iops[ddir];
 210         rate_iops_min += td->o.rate_iops_min[ddir];
 211
 212         /*
 213          * if rate blocks is set, sample is running
 214          */
 215         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 216                 spent = mtime_since(&td->lastrate[ddir], now);
 217                 if (spent < td->o.ratecycle)
 218                         return 0;
 219
 220                 if (td->o.rate[ddir]) {
 221                         /*
 222                          * check bandwidth specified rate
 223                          */
 224                         if (bytes < td->rate_bytes[ddir]) {
 225                                 log_err("%s: min rate %u not met\n", td->o.name,
 226                                                                 ratemin);
 227                                 return 1;
 228                         } else {
 229                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 230                                 if (rate < ratemin ||
 231                                     bytes < td->rate_bytes[ddir]) {
 232                                         log_err("%s: min rate %u not met, got"
 233                                                 " %luKB/sec\n", td->o.name,
 234                                                         ratemin, rate);
 235                                         return 1;
 236                                 }
 237                         }
 238                 } else {
 239                         /*
 240                          * checks iops specified rate
 241                          */
 242                         if (iops < rate_iops) {
 243                                 log_err("%s: min iops rate %u not met\n",
 244                                                 td->o.name, rate_iops);
 245                                 return 1;
 246                         } else {
 247                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 248                                 if (rate < rate_iops_min ||
 249                                     iops < td->rate_blocks[ddir]) {
 250                                         log_err("%s: min iops rate %u not met,"
 251                                                 " got %lu\n", td->o.name,
 252                                                         rate_iops_min, rate);
 253                                 }
 254                         }
 255                 }
 256         }
 257
 258         td->rate_bytes[ddir] = bytes;
 259         td->rate_blocks[ddir] = iops;
 260         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 261         return 0;
 262 }
 263
 264 static int check_min_rate(struct thread_data *td, struct timeval *now,
 265                           unsigned long *bytes_done)
 266 {
 267         int ret = 0;
 268
 269         if (bytes_done[0])
 270                 ret |= __check_min_rate(td, now, 0);
 271         if (bytes_done[1])
 272                 ret |= __check_min_rate(td, now, 1);
 273
 274         return ret;
 275 }
 276
 277 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 278 {
 279         if (!td->o.timeout)
 280                 return 0;
 281         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 282                 return 1;
 283
 284         return 0;
 285 }
 286
 287 /*
 288  * When job exits, we can cancel the in-flight IO if we are using async
 289  * io. Attempt to do so.
 290  */
 291 static void cleanup_pending_aio(struct thread_data *td)
 292 {
 293         struct flist_head *entry, *n;
 294         struct io_u *io_u;
 295         int r;
 296
 297         /*
 298          * get immediately available events, if any
 299          */
 300         r = io_u_queued_complete(td, 0, NULL);
 301         if (r < 0)
 302                 return;
 303
 304         /*
 305          * now cancel remaining active events
 306          */
 307         if (td->io_ops->cancel) {
 308                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 309                         io_u = flist_entry(entry, struct io_u, list);
 310
 311                         /*
 312                          * if the io_u isn't in flight, then that generally
 313                          * means someone leaked an io_u. complain but fix
 314                          * it up, so we don't stall here.
 315                          */
 316                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 317                                 log_err("fio: non-busy IO on busy list\n");
 318                                 put_io_u(td, io_u);
 319                         } else {
 320                                 r = td->io_ops->cancel(td, io_u);
 321                                 if (!r)
 322                                         put_io_u(td, io_u);
 323                         }
 324                 }
 325         }
 326
 327         if (td->cur_depth)
 328                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 329 }
 330
 331 /*
 332  * Helper to handle the final sync of a file. Works just like the normal
 333  * io path, just does everything sync.
 334  */
 335 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 336 {
 337         struct io_u *io_u = __get_io_u(td);
 338         int ret;
 339
 340         if (!io_u)
 341                 return 1;
 342
 343         io_u->ddir = DDIR_SYNC;
 344         io_u->file = f;
 345
 346         if (td_io_prep(td, io_u)) {
 347                 put_io_u(td, io_u);
 348                 return 1;
 349         }
 350
 351 requeue:
 352         ret = td_io_queue(td, io_u);
 353         if (ret < 0) {
 354                 td_verror(td, io_u->error, "td_io_queue");
 355                 put_io_u(td, io_u);
 356                 return 1;
 357         } else if (ret == FIO_Q_QUEUED) {
 358                 if (io_u_queued_complete(td, 1, NULL) < 0)
 359                         return 1;
 360         } else if (ret == FIO_Q_COMPLETED) {
 361                 if (io_u->error) {
 362                         td_verror(td, io_u->error, "td_io_queue");
 363                         return 1;
 364                 }
 365
 366                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 367                         return 1;
 368         } else if (ret == FIO_Q_BUSY) {
 369                 if (td_io_commit(td))
 370                         return 1;
 371                 goto requeue;
 372         }
 373
 374         return 0;
 375 }
 376
 377 static inline void __update_tv_cache(struct thread_data *td)
 378 {
 379         fio_gettime(&td->tv_cache, NULL);
 380 }
 381
 382 static inline void update_tv_cache(struct thread_data *td)
 383 {
 384         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 385                 __update_tv_cache(td);
 386 }
 387
 388 static int break_on_this_error(struct thread_data *td, int *retptr)
 389 {
 390         int ret = *retptr;
 391
 392         if (ret < 0 || td->error) {
 393                 int err;
 394
 395                 if (!td->o.continue_on_error)
 396                         return 1;
 397
 398                 if (ret < 0)
 399                         err = -ret;
 400                 else
 401                         err = td->error;
 402
 403                 if (td_non_fatal_error(err)) {
 404                         /*
 405                          * Continue with the I/Os in case of
 406                          * a non fatal error.
 407                          */
 408                         update_error_count(td, err);
 409                         td_clear_error(td);
 410                         *retptr = 0;
 411                         return 0;
 412                 } else if (td->o.fill_device && err == ENOSPC) {
 413                         /*
 414                          * We expect to hit this error if
 415                          * fill_device option is set.
 416                          */
 417                         td_clear_error(td);
 418                         td->terminate = 1;
 419                         return 1;
 420                 } else {
 421                         /*
 422                          * Stop the I/O in case of a fatal
 423                          * error.
 424                          */
 425                         update_error_count(td, err);
 426                         return 1;
 427                 }
 428         }
 429
 430         return 0;
 431 }
 432
 433 /*
 434  * The main verify engine. Runs over the writes we previously submitted,
 435  * reads the blocks back in, and checks the crc/md5 of the data.
 436  */
 437 static void do_verify(struct thread_data *td)
 438 {
 439         struct fio_file *f;
 440         struct io_u *io_u;
 441         int ret, min_events;
 442         unsigned int i;
 443
 444         dprint(FD_VERIFY, "starting loop\n");
 445
 446         /*
 447          * sync io first and invalidate cache, to make sure we really
 448          * read from disk.
 449          */
 450         for_each_file(td, f, i) {
 451                 if (!fio_file_open(f))
 452                         continue;
 453                 if (fio_io_sync(td, f))
 454                         break;
 455                 if (file_invalidate_cache(td, f))
 456                         break;
 457         }
 458
 459         if (td->error)
 460                 return;
 461
 462         td_set_runstate(td, TD_VERIFYING);
 463
 464         io_u = NULL;
 465         while (!td->terminate) {
 466                 int ret2, full;
 467
 468                 update_tv_cache(td);
 469
 470                 if (runtime_exceeded(td, &td->tv_cache)) {
 471                         __update_tv_cache(td);
 472                         if (runtime_exceeded(td, &td->tv_cache)) {
 473                                 td->terminate = 1;
 474                                 break;
 475                         }
 476                 }
 477
 478                 io_u = __get_io_u(td);
 479                 if (!io_u)
 480                         break;
 481
 482                 if (get_next_verify(td, io_u)) {
 483                         put_io_u(td, io_u);
 484                         break;
 485                 }
 486
 487                 if (td_io_prep(td, io_u)) {
 488                         put_io_u(td, io_u);
 489                         break;
 490                 }
 491
 492                 if (td->o.verify_async)
 493                         io_u->end_io = verify_io_u_async;
 494                 else
 495                         io_u->end_io = verify_io_u;
 496
 497                 ret = td_io_queue(td, io_u);
 498                 switch (ret) {
 499                 case FIO_Q_COMPLETED:
 500                         if (io_u->error) {
 501                                 ret = -io_u->error;
 502                                 clear_io_u(td, io_u);
 503                         } else if (io_u->resid) {
 504                                 int bytes = io_u->xfer_buflen - io_u->resid;
 505
 506                                 /*
 507                                  * zero read, fail
 508                                  */
 509                                 if (!bytes) {
 510                                         td_verror(td, EIO, "full resid");
 511                                         put_io_u(td, io_u);
 512                                         break;
 513                                 }
 514
 515                                 io_u->xfer_buflen = io_u->resid;
 516                                 io_u->xfer_buf += bytes;
 517                                 io_u->offset += bytes;
 518
 519                                 if (ddir_rw(io_u->ddir))
 520                                         td->ts.short_io_u[io_u->ddir]++;
 521
 522                                 f = io_u->file;
 523                                 if (io_u->offset == f->real_file_size)
 524                                         goto sync_done;
 525
 526                                 requeue_io_u(td, &io_u);
 527                         } else {
 528 sync_done:
 529                                 ret = io_u_sync_complete(td, io_u, NULL);
 530                                 if (ret < 0)
 531                                         break;
 532                         }
 533                         continue;
 534                 case FIO_Q_QUEUED:
 535                         break;
 536                 case FIO_Q_BUSY:
 537                         requeue_io_u(td, &io_u);
 538                         ret2 = td_io_commit(td);
 539                         if (ret2 < 0)
 540                                 ret = ret2;
 541                         break;
 542                 default:
 543                         assert(ret < 0);
 544                         td_verror(td, -ret, "td_io_queue");
 545                         break;
 546                 }
 547
 548                 if (break_on_this_error(td, &ret))
 549                         break;
 550
 551                 /*
 552                  * if we can queue more, do so. but check if there are
 553                  * completed io_u's first.
 554                  */
 555                 full = queue_full(td) || ret == FIO_Q_BUSY;
 556                 if (full || !td->o.iodepth_batch_complete) {
 557                         min_events = min(td->o.iodepth_batch_complete,
 558                                          td->cur_depth);
 559                         if (full && !min_events)
 560                                 min_events = 1;
 561
 562                         do {
 563                                 /*
 564                                  * Reap required number of io units, if any,
 565                                  * and do the verification on them through
 566                                  * the callback handler
 567                                  */
 568                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 569                                         ret = -1;
 570                                         break;
 571                                 }
 572                         } while (full && (td->cur_depth > td->o.iodepth_low));
 573                 }
 574                 if (ret < 0)
 575                         break;
 576         }
 577
 578         if (!td->error) {
 579                 min_events = td->cur_depth;
 580
 581                 if (min_events)
 582                         ret = io_u_queued_complete(td, min_events, NULL);
 583         } else
 584                 cleanup_pending_aio(td);
 585
 586         td_set_runstate(td, TD_RUNNING);
 587
 588         dprint(FD_VERIFY, "exiting loop\n");
 589 }
 590
 591 /*
 592  * Main IO worker function. It retrieves io_u's to process and queues
 593  * and reaps them, checking for rate and errors along the way.
 594  */
 595 static void do_io(struct thread_data *td)
 596 {
 597         unsigned int i;
 598         int ret = 0;
 599
 600         if (in_ramp_time(td))
 601                 td_set_runstate(td, TD_RAMP);
 602         else
 603                 td_set_runstate(td, TD_RUNNING);
 604
 605         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 606                 (!flist_empty(&td->trim_list)) ||
 607                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 608                 struct timeval comp_time;
 609                 unsigned long bytes_done[2] = { 0, 0 };
 610                 int min_evts = 0;
 611                 struct io_u *io_u;
 612                 int ret2, full;
 613
 614                 if (td->terminate)
 615                         break;
 616
 617                 update_tv_cache(td);
 618
 619                 if (runtime_exceeded(td, &td->tv_cache)) {
 620                         __update_tv_cache(td);
 621                         if (runtime_exceeded(td, &td->tv_cache)) {
 622                                 td->terminate = 1;
 623                                 break;
 624                         }
 625                 }
 626
 627                 io_u = get_io_u(td);
 628                 if (!io_u)
 629                         break;
 630
 631                 /*
 632                  * Add verification end_io handler, if asked to verify
 633                  * a previously written file.
 634                  */
 635                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 636                     !td_rw(td)) {
 637                         if (td->o.verify_async)
 638                                 io_u->end_io = verify_io_u_async;
 639                         else
 640                                 io_u->end_io = verify_io_u;
 641                         td_set_runstate(td, TD_VERIFYING);
 642                 } else if (in_ramp_time(td))
 643                         td_set_runstate(td, TD_RAMP);
 644                 else
 645                         td_set_runstate(td, TD_RUNNING);
 646
 647                 ret = td_io_queue(td, io_u);
 648                 switch (ret) {
 649                 case FIO_Q_COMPLETED:
 650                         if (io_u->error) {
 651                                 ret = -io_u->error;
 652                                 clear_io_u(td, io_u);
 653                         } else if (io_u->resid) {
 654                                 int bytes = io_u->xfer_buflen - io_u->resid;
 655                                 struct fio_file *f = io_u->file;
 656
 657                                 /*
 658                                  * zero read, fail
 659                                  */
 660                                 if (!bytes) {
 661                                         td_verror(td, EIO, "full resid");
 662                                         put_io_u(td, io_u);
 663                                         break;
 664                                 }
 665
 666                                 io_u->xfer_buflen = io_u->resid;
 667                                 io_u->xfer_buf += bytes;
 668                                 io_u->offset += bytes;
 669
 670                                 if (ddir_rw(io_u->ddir))
 671                                         td->ts.short_io_u[io_u->ddir]++;
 672
 673                                 if (io_u->offset == f->real_file_size)
 674                                         goto sync_done;
 675
 676                                 requeue_io_u(td, &io_u);
 677                         } else {
 678 sync_done:
 679                                 if (__should_check_rate(td, 0) ||
 680                                     __should_check_rate(td, 1))
 681                                         fio_gettime(&comp_time, NULL);
 682
 683                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 684                                 if (ret < 0)
 685                                         break;
 686                         }
 687                         break;
 688                 case FIO_Q_QUEUED:
 689                         /*
 690                          * if the engine doesn't have a commit hook,
 691                          * the io_u is really queued. if it does have such
 692                          * a hook, it has to call io_u_queued() itself.
 693                          */
 694                         if (td->io_ops->commit == NULL)
 695                                 io_u_queued(td, io_u);
 696                         break;
 697                 case FIO_Q_BUSY:
 698                         requeue_io_u(td, &io_u);
 699                         ret2 = td_io_commit(td);
 700                         if (ret2 < 0)
 701                                 ret = ret2;
 702                         break;
 703                 default:
 704                         assert(ret < 0);
 705                         put_io_u(td, io_u);
 706                         break;
 707                 }
 708
 709                 if (break_on_this_error(td, &ret))
 710                         break;
 711
 712                 /*
 713                  * See if we need to complete some commands
 714                  */
 715                 full = queue_full(td) || ret == FIO_Q_BUSY;
 716                 if (full || !td->o.iodepth_batch_complete) {
 717                         min_evts = min(td->o.iodepth_batch_complete,
 718                                         td->cur_depth);
 719                         if (full && !min_evts)
 720                                 min_evts = 1;
 721
 722                         if (__should_check_rate(td, 0) ||
 723                             __should_check_rate(td, 1))
 724                                 fio_gettime(&comp_time, NULL);
 725
 726                         do {
 727                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 728                                 if (ret < 0)
 729                                         break;
 730
 731                         } while (full && (td->cur_depth > td->o.iodepth_low));
 732                 }
 733
 734                 if (ret < 0)
 735                         break;
 736                 if (!(bytes_done[0] + bytes_done[1]))
 737                         continue;
 738
 739                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 740                         if (check_min_rate(td, &comp_time, bytes_done)) {
 741                                 if (exitall_on_terminate)
 742                                         terminate_threads(td->groupid);
 743                                 td_verror(td, EIO, "check_min_rate");
 744                                 break;
 745                         }
 746                 }
 747
 748                 if (td->o.thinktime) {
 749                         unsigned long long b;
 750
 751                         b = td->io_blocks[0] + td->io_blocks[1];
 752                         if (!(b % td->o.thinktime_blocks)) {
 753                                 int left;
 754
 755                                 if (td->o.thinktime_spin)
 756                                         usec_spin(td->o.thinktime_spin);
 757
 758                                 left = td->o.thinktime - td->o.thinktime_spin;
 759                                 if (left)
 760                                         usec_sleep(td, left);
 761                         }
 762                 }
 763         }
 764
 765         if (td->trim_entries)
 766                 printf("trim entries %ld\n", td->trim_entries);
 767
 768         if (td->o.fill_device && td->error == ENOSPC) {
 769                 td->error = 0;
 770                 td->terminate = 1;
 771         }
 772         if (!td->error) {
 773                 struct fio_file *f;
 774
 775                 i = td->cur_depth;
 776                 if (i) {
 777                         ret = io_u_queued_complete(td, i, NULL);
 778                         if (td->o.fill_device && td->error == ENOSPC)
 779                                 td->error = 0;
 780                 }
 781
 782                 if (should_fsync(td) && td->o.end_fsync) {
 783                         td_set_runstate(td, TD_FSYNCING);
 784
 785                         for_each_file(td, f, i) {
 786                                 if (!fio_file_open(f))
 787                                         continue;
 788                                 fio_io_sync(td, f);
 789                         }
 790                 }
 791         } else
 792                 cleanup_pending_aio(td);
 793
 794         /*
 795          * stop job if we failed doing any IO
 796          */
 797         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 798                 td->done = 1;
 799 }
 800
 801 static void cleanup_io_u(struct thread_data *td)
 802 {
 803         struct flist_head *entry, *n;
 804         struct io_u *io_u;
 805
 806         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 807                 io_u = flist_entry(entry, struct io_u, list);
 808
 809                 flist_del(&io_u->list);
 810                 fio_memfree(io_u, sizeof(*io_u));
 811         }
 812
 813         free_io_mem(td);
 814 }
 815
 816 static int init_io_u(struct thread_data *td)
 817 {
 818         struct io_u *io_u;
 819         unsigned int max_bs;
 820         int cl_align, i, max_units;
 821         char *p;
 822
 823         max_units = td->o.iodepth;
 824         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 825         td->orig_buffer_size = (unsigned long long) max_bs
 826                                         * (unsigned long long) max_units;
 827
 828         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 829                 unsigned long bs;
 830
 831                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 832                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 833         }
 834
 835         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 836                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 837                 return 1;
 838         }
 839
 840         if (allocate_io_mem(td))
 841                 return 1;
 842
 843         if (td->o.odirect || td->o.mem_align ||
 844             (td->io_ops->flags & FIO_RAWIO))
 845                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 846         else
 847                 p = td->orig_buffer;
 848
 849         cl_align = os_cache_line_size();
 850
 851         for (i = 0; i < max_units; i++) {
 852                 void *ptr;
 853
 854                 if (td->terminate)
 855                         return 1;
 856
 857                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 858                 if (!ptr) {
 859                         log_err("fio: unable to allocate aligned memory\n");
 860                         break;
 861                 }
 862
 863                 io_u = ptr;
 864                 memset(io_u, 0, sizeof(*io_u));
 865                 INIT_FLIST_HEAD(&io_u->list);
 866                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 867
 868                 if (!(td->io_ops->flags & FIO_NOIO)) {
 869                         io_u->buf = p + max_bs * i;
 870                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 871
 872                         if (td_write(td) && !td->o.refill_buffers)
 873                                 io_u_fill_buffer(td, io_u, max_bs);
 874                         else if (td_write(td) && td->o.verify_pattern_bytes) {
 875                                 /*
 876                                  * Fill the buffer with the pattern if we are
 877                                  * going to be doing writes.
 878                                  */
 879                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 880                         }
 881                 }
 882
 883                 io_u->index = i;
 884                 io_u->flags = IO_U_F_FREE;
 885                 flist_add(&io_u->list, &td->io_u_freelist);
 886         }
 887
 888         return 0;
 889 }
 890
 891 static int switch_ioscheduler(struct thread_data *td)
 892 {
 893         char tmp[256], tmp2[128];
 894         FILE *f;
 895         int ret;
 896
 897         if (td->io_ops->flags & FIO_DISKLESSIO)
 898                 return 0;
 899
 900         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 901
 902         f = fopen(tmp, "r+");
 903         if (!f) {
 904                 if (errno == ENOENT) {
 905                         log_err("fio: os or kernel doesn't support IO scheduler"
 906                                 " switching\n");
 907                         return 0;
 908                 }
 909                 td_verror(td, errno, "fopen iosched");
 910                 return 1;
 911         }
 912
 913         /*
 914          * Set io scheduler.
 915          */
 916         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 917         if (ferror(f) || ret != 1) {
 918                 td_verror(td, errno, "fwrite");
 919                 fclose(f);
 920                 return 1;
 921         }
 922
 923         rewind(f);
 924
 925         /*
 926          * Read back and check that the selected scheduler is now the default.
 927          */
 928         ret = fread(tmp, 1, sizeof(tmp), f);
 929         if (ferror(f) || ret < 0) {
 930                 td_verror(td, errno, "fread");
 931                 fclose(f);
 932                 return 1;
 933         }
 934
 935         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 936         if (!strstr(tmp, tmp2)) {
 937                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 938                 td_verror(td, EINVAL, "iosched_switch");
 939                 fclose(f);
 940                 return 1;
 941         }
 942
 943         fclose(f);
 944         return 0;
 945 }
 946
 947 static int keep_running(struct thread_data *td)
 948 {
 949         unsigned long long io_done;
 950
 951         if (td->done)
 952                 return 0;
 953         if (td->o.time_based)
 954                 return 1;
 955         if (td->o.loops) {
 956                 td->o.loops--;
 957                 return 1;
 958         }
 959
 960         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 961                         + td->io_skip_bytes;
 962         if (io_done < td->o.size)
 963                 return 1;
 964
 965         return 0;
 966 }
 967
 968 static void reset_io_counters(struct thread_data *td)
 969 {
 970         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 971         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 972         td->zone_bytes = 0;
 973         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 974         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 975
 976         td->last_was_sync = 0;
 977
 978         /*
 979          * reset file done count if we are to start over
 980          */
 981         if (td->o.time_based || td->o.loops)
 982                 td->nr_done_files = 0;
 983 }
 984
 985 void reset_all_stats(struct thread_data *td)
 986 {
 987         struct timeval tv;
 988         int i;
 989
 990         reset_io_counters(td);
 991
 992         for (i = 0; i < 2; i++) {
 993                 td->io_bytes[i] = 0;
 994                 td->io_blocks[i] = 0;
 995                 td->io_issues[i] = 0;
 996                 td->ts.total_io_u[i] = 0;
 997         }
 998
 999         fio_gettime(&tv, NULL);
1000         td->ts.runtime[0] = 0;
1001         td->ts.runtime[1] = 0;
1002         memcpy(&td->epoch, &tv, sizeof(tv));
1003         memcpy(&td->start, &tv, sizeof(tv));
1004 }
1005
1006 static void clear_io_state(struct thread_data *td)
1007 {
1008         struct fio_file *f;
1009         unsigned int i;
1010
1011         reset_io_counters(td);
1012
1013         close_files(td);
1014         for_each_file(td, f, i)
1015                 fio_file_clear_done(f);
1016
1017         /*
1018          * Set the same seed to get repeatable runs
1019          */
1020         td_fill_rand_seeds(td);
1021 }
1022
1023 static int exec_string(const char *string)
1024 {
1025         int ret, newlen = strlen(string) + 1 + 8;
1026         char *str;
1027
1028         str = malloc(newlen);
1029         sprintf(str, "sh -c %s", string);
1030
1031         ret = system(str);
1032         if (ret == -1)
1033                 log_err("fio: exec of cmd <%s> failed\n", str);
1034
1035         free(str);
1036         return ret;
1037 }
1038
1039 /*
1040  * Entry point for the thread based jobs. The process based jobs end up
1041  * here as well, after a little setup.
1042  */
1043 static void *thread_main(void *data)
1044 {
1045         unsigned long long elapsed;
1046         struct thread_data *td = data;
1047         pthread_condattr_t attr;
1048         int clear_state;
1049
1050         if (!td->o.use_thread) {
1051                 setsid();
1052                 td->pid = getpid();
1053         } else
1054                 td->pid = gettid();
1055
1056         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1057
1058         INIT_FLIST_HEAD(&td->io_u_freelist);
1059         INIT_FLIST_HEAD(&td->io_u_busylist);
1060         INIT_FLIST_HEAD(&td->io_u_requeues);
1061         INIT_FLIST_HEAD(&td->io_log_list);
1062         INIT_FLIST_HEAD(&td->io_hist_list);
1063         INIT_FLIST_HEAD(&td->verify_list);
1064         INIT_FLIST_HEAD(&td->trim_list);
1065         pthread_mutex_init(&td->io_u_lock, NULL);
1066         td->io_hist_tree = RB_ROOT;
1067
1068         pthread_condattr_init(&attr);
1069         pthread_cond_init(&td->verify_cond, &attr);
1070         pthread_cond_init(&td->free_cond, &attr);
1071
1072         td_set_runstate(td, TD_INITIALIZED);
1073         dprint(FD_MUTEX, "up startup_mutex\n");
1074         fio_mutex_up(startup_mutex);
1075         dprint(FD_MUTEX, "wait on td->mutex\n");
1076         fio_mutex_down(td->mutex);
1077         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1078
1079         /*
1080          * the ->mutex mutex is now no longer used, close it to avoid
1081          * eating a file descriptor
1082          */
1083         fio_mutex_remove(td->mutex);
1084
1085         /*
1086          * A new gid requires privilege, so we need to do this before setting
1087          * the uid.
1088          */
1089         if (td->o.gid != -1U && setgid(td->o.gid)) {
1090                 td_verror(td, errno, "setgid");
1091                 goto err;
1092         }
1093         if (td->o.uid != -1U && setuid(td->o.uid)) {
1094                 td_verror(td, errno, "setuid");
1095                 goto err;
1096         }
1097
1098         /*
1099          * If we have a gettimeofday() thread, make sure we exclude that
1100          * thread from this job
1101          */
1102         if (td->o.gtod_cpu)
1103                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1104
1105         /*
1106          * Set affinity first, in case it has an impact on the memory
1107          * allocations.
1108          */
1109         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1110                 td_verror(td, errno, "cpu_set_affinity");
1111                 goto err;
1112         }
1113
1114         /*
1115          * May alter parameters that init_io_u() will use, so we need to
1116          * do this first.
1117          */
1118         if (init_iolog(td))
1119                 goto err;
1120
1121         if (init_io_u(td))
1122                 goto err;
1123
1124         if (td->o.verify_async && verify_async_init(td))
1125                 goto err;
1126
1127         if (td->ioprio_set) {
1128                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1129                         td_verror(td, errno, "ioprio_set");
1130                         goto err;
1131                 }
1132         }
1133
1134         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1135                 goto err;
1136
1137         if (nice(td->o.nice) == -1) {
1138                 td_verror(td, errno, "nice");
1139                 goto err;
1140         }
1141
1142         if (td->o.ioscheduler && switch_ioscheduler(td))
1143                 goto err;
1144
1145         if (!td->o.create_serialize && setup_files(td))
1146                 goto err;
1147
1148         if (td_io_init(td))
1149                 goto err;
1150
1151         if (init_random_map(td))
1152                 goto err;
1153
1154         if (td->o.exec_prerun) {
1155                 if (exec_string(td->o.exec_prerun))
1156                         goto err;
1157         }
1158
1159         if (td->o.pre_read) {
1160                 if (pre_read_files(td) < 0)
1161                         goto err;
1162         }
1163
1164         fio_gettime(&td->epoch, NULL);
1165         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1166
1167         clear_state = 0;
1168         while (keep_running(td)) {
1169                 fio_gettime(&td->start, NULL);
1170                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1171                                 sizeof(td->start));
1172                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1173                                 sizeof(td->start));
1174                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1175
1176                 if (td->o.ratemin[0] || td->o.ratemin[1])
1177                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1178                                                         sizeof(td->lastrate));
1179
1180                 if (clear_state)
1181                         clear_io_state(td);
1182
1183                 prune_io_piece_log(td);
1184
1185                 do_io(td);
1186
1187                 clear_state = 1;
1188
1189                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1190                         elapsed = utime_since_now(&td->start);
1191                         td->ts.runtime[DDIR_READ] += elapsed;
1192                 }
1193                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1194                         elapsed = utime_since_now(&td->start);
1195                         td->ts.runtime[DDIR_WRITE] += elapsed;
1196                 }
1197
1198                 if (td->error || td->terminate)
1199                         break;
1200
1201                 if (!td->o.do_verify ||
1202                     td->o.verify == VERIFY_NONE ||
1203                     (td->io_ops->flags & FIO_UNIDIR))
1204                         continue;
1205
1206                 clear_io_state(td);
1207
1208                 fio_gettime(&td->start, NULL);
1209
1210                 do_verify(td);
1211
1212                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1213
1214                 if (td->error || td->terminate)
1215                         break;
1216         }
1217
1218         update_rusage_stat(td);
1219         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1220         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1221         td->ts.total_run_time = mtime_since_now(&td->epoch);
1222         td->ts.io_bytes[0] = td->io_bytes[0];
1223         td->ts.io_bytes[1] = td->io_bytes[1];
1224
1225         fio_mutex_down(writeout_mutex);
1226         if (td->ts.bw_log) {
1227                 if (td->o.bw_log_file) {
1228                         finish_log_named(td, td->ts.bw_log,
1229                                                 td->o.bw_log_file, "bw");
1230                 } else
1231                         finish_log(td, td->ts.bw_log, "bw");
1232         }
1233         if (td->ts.lat_log) {
1234                 if (td->o.lat_log_file) {
1235                         finish_log_named(td, td->ts.lat_log,
1236                                                 td->o.lat_log_file, "lat");
1237                 } else
1238                         finish_log(td, td->ts.lat_log, "lat");
1239         }
1240         if (td->ts.slat_log) {
1241                 if (td->o.lat_log_file) {
1242                         finish_log_named(td, td->ts.slat_log,
1243                                                 td->o.lat_log_file, "slat");
1244                 } else
1245                         finish_log(td, td->ts.slat_log, "slat");
1246         }
1247         if (td->ts.clat_log) {
1248                 if (td->o.lat_log_file) {
1249                         finish_log_named(td, td->ts.clat_log,
1250                                                 td->o.lat_log_file, "clat");
1251                 } else
1252                         finish_log(td, td->ts.clat_log, "clat");
1253         }
1254         fio_mutex_up(writeout_mutex);
1255         if (td->o.exec_postrun)
1256                 exec_string(td->o.exec_postrun);
1257
1258         if (exitall_on_terminate)
1259                 terminate_threads(td->groupid);
1260
1261 err:
1262         if (td->error)
1263                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1264                                                         td->verror);
1265
1266         if (td->o.verify_async)
1267                 verify_async_exit(td);
1268
1269         close_and_free_files(td);
1270         close_ioengine(td);
1271         cleanup_io_u(td);
1272         cgroup_shutdown(td, &cgroup_mnt);
1273
1274         if (td->o.cpumask_set) {
1275                 int ret = fio_cpuset_exit(&td->o.cpumask);
1276
1277                 td_verror(td, ret, "fio_cpuset_exit");
1278         }
1279
1280         /*
1281          * do this very late, it will log file closing as well
1282          */
1283         if (td->o.write_iolog_file)
1284                 write_iolog_close(td);
1285
1286         options_mem_free(td);
1287         td_set_runstate(td, TD_EXITED);
1288         return (void *) (unsigned long) td->error;
1289 }
1290
1291 /*
1292  * We cannot pass the td data into a forked process, so attach the td and
1293  * pass it to the thread worker.
1294  */
1295 static int fork_main(int shmid, int offset)
1296 {
1297         struct thread_data *td;
1298         void *data, *ret;
1299
1300 #ifndef __hpux
1301         data = shmat(shmid, NULL, 0);
1302         if (data == (void *) -1) {
1303                 int __err = errno;
1304
1305                 perror("shmat");
1306                 return __err;
1307         }
1308 #else
1309         /*
1310          * HP-UX inherits shm mappings?
1311          */
1312         data = threads;
1313 #endif
1314
1315         td = data + offset * sizeof(struct thread_data);
1316         ret = thread_main(td);
1317         shmdt(data);
1318         return (int) (unsigned long) ret;
1319 }
1320
1321 /*
1322  * Run over the job map and reap the threads that have exited, if any.
1323  */
1324 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1325 {
1326         struct thread_data *td;
1327         int i, cputhreads, realthreads, pending, status, ret;
1328
1329         /*
1330          * reap exited threads (TD_EXITED -> TD_REAPED)
1331          */
1332         realthreads = pending = cputhreads = 0;
1333         for_each_td(td, i) {
1334                 int flags = 0;
1335
1336                 /*
1337                  * ->io_ops is NULL for a thread that has closed its
1338                  * io engine
1339                  */
1340                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1341                         cputhreads++;
1342                 else
1343                         realthreads++;
1344
1345                 if (!td->pid) {
1346                         pending++;
1347                         continue;
1348                 }
1349                 if (td->runstate == TD_REAPED)
1350                         continue;
1351                 if (td->o.use_thread) {
1352                         if (td->runstate == TD_EXITED) {
1353                                 td_set_runstate(td, TD_REAPED);
1354                                 goto reaped;
1355                         }
1356                         continue;
1357                 }
1358
1359                 flags = WNOHANG;
1360                 if (td->runstate == TD_EXITED)
1361                         flags = 0;
1362
1363                 /*
1364                  * check if someone quit or got killed in an unusual way
1365                  */
1366                 ret = waitpid(td->pid, &status, flags);
1367                 if (ret < 0) {
1368                         if (errno == ECHILD) {
1369                                 log_err("fio: pid=%d disappeared %d\n",
1370                                                 (int) td->pid, td->runstate);
1371                                 td_set_runstate(td, TD_REAPED);
1372                                 goto reaped;
1373                         }
1374                         perror("waitpid");
1375                 } else if (ret == td->pid) {
1376                         if (WIFSIGNALED(status)) {
1377                                 int sig = WTERMSIG(status);
1378
1379                                 if (sig != SIGTERM)
1380                                         log_err("fio: pid=%d, got signal=%d\n",
1381                                                         (int) td->pid, sig);
1382                                 td_set_runstate(td, TD_REAPED);
1383                                 goto reaped;
1384                         }
1385                         if (WIFEXITED(status)) {
1386                                 if (WEXITSTATUS(status) && !td->error)
1387                                         td->error = WEXITSTATUS(status);
1388
1389                                 td_set_runstate(td, TD_REAPED);
1390                                 goto reaped;
1391                         }
1392                 }
1393
1394                 /*
1395                  * thread is not dead, continue
1396                  */
1397                 pending++;
1398                 continue;
1399 reaped:
1400                 (*nr_running)--;
1401                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1402                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1403                 if (!td->pid)
1404                         pending--;
1405
1406                 if (td->error)
1407                         exit_value++;
1408
1409                 done_secs += mtime_since_now(&td->epoch) / 1000;
1410         }
1411
1412         if (*nr_running == cputhreads && !pending && realthreads)
1413                 terminate_threads(TERMINATE_ALL);
1414 }
1415
1416 static void *gtod_thread_main(void *data)
1417 {
1418         fio_mutex_up(startup_mutex);
1419
1420         /*
1421          * As long as we have jobs around, update the clock. It would be nice
1422          * to have some way of NOT hammering that CPU with gettimeofday(),
1423          * but I'm not sure what to use outside of a simple CPU nop to relax
1424          * it - we don't want to lose precision.
1425          */
1426         while (threads) {
1427                 fio_gtod_update();
1428                 nop;
1429         }
1430
1431         return NULL;
1432 }
1433
1434 static int fio_start_gtod_thread(void)
1435 {
1436         pthread_attr_t attr;
1437         int ret;
1438
1439         pthread_attr_init(&attr);
1440         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1441         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1442         pthread_attr_destroy(&attr);
1443         if (ret) {
1444                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1445                 return 1;
1446         }
1447
1448         ret = pthread_detach(gtod_thread);
1449         if (ret) {
1450                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1451                 return 1;
1452         }
1453
1454         dprint(FD_MUTEX, "wait on startup_mutex\n");
1455         fio_mutex_down(startup_mutex);
1456         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1457         return 0;
1458 }
1459
1460 /*
1461  * Main function for kicking off and reaping jobs, as needed.
1462  */
1463 static void run_threads(void)
1464 {
1465         struct thread_data *td;
1466         unsigned long spent;
1467         int i, todo, nr_running, m_rate, t_rate, nr_started;
1468
1469         if (fio_pin_memory())
1470                 return;
1471
1472         if (fio_gtod_offload && fio_start_gtod_thread())
1473                 return;
1474
1475         if (!terse_output) {
1476                 log_info("Starting ");
1477                 if (nr_thread)
1478                         log_info("%d thread%s", nr_thread,
1479                                                 nr_thread > 1 ? "s" : "");
1480                 if (nr_process) {
1481                         if (nr_thread)
1482                                 printf(" and ");
1483                         log_info("%d process%s", nr_process,
1484                                                 nr_process > 1 ? "es" : "");
1485                 }
1486                 log_info("\n");
1487                 fflush(stdout);
1488         }
1489
1490         set_sig_handlers();
1491
1492         todo = thread_number;
1493         nr_running = 0;
1494         nr_started = 0;
1495         m_rate = t_rate = 0;
1496
1497         for_each_td(td, i) {
1498                 print_status_init(td->thread_number - 1);
1499
1500                 if (!td->o.create_serialize) {
1501                         init_disk_util(td);
1502                         continue;
1503                 }
1504
1505                 /*
1506                  * do file setup here so it happens sequentially,
1507                  * we don't want X number of threads getting their
1508                  * client data interspersed on disk
1509                  */
1510                 if (setup_files(td)) {
1511                         exit_value++;
1512                         if (td->error)
1513                                 log_err("fio: pid=%d, err=%d/%s\n",
1514                                         (int) td->pid, td->error, td->verror);
1515                         td_set_runstate(td, TD_REAPED);
1516                         todo--;
1517                 } else {
1518                         struct fio_file *f;
1519                         unsigned int j;
1520
1521                         /*
1522                          * for sharing to work, each job must always open
1523                          * its own files. so close them, if we opened them
1524                          * for creation
1525                          */
1526                         for_each_file(td, f, j) {
1527                                 if (fio_file_open(f))
1528                                         td_io_close_file(td, f);
1529                         }
1530                 }
1531
1532                 init_disk_util(td);
1533         }
1534
1535         set_genesis_time();
1536
1537         while (todo) {
1538                 struct thread_data *map[REAL_MAX_JOBS];
1539                 struct timeval this_start;
1540                 int this_jobs = 0, left;
1541
1542                 /*
1543                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1544                  */
1545                 for_each_td(td, i) {
1546                         if (td->runstate != TD_NOT_CREATED)
1547                                 continue;
1548
1549                         /*
1550                          * never got a chance to start, killed by other
1551                          * thread for some reason
1552                          */
1553                         if (td->terminate) {
1554                                 todo--;
1555                                 continue;
1556                         }
1557
1558                         if (td->o.start_delay) {
1559                                 spent = mtime_since_genesis();
1560
1561                                 if (td->o.start_delay * 1000 > spent)
1562                                         continue;
1563                         }
1564
1565                         if (td->o.stonewall && (nr_started || nr_running)) {
1566                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1567                                                         td->o.name);
1568                                 break;
1569                         }
1570
1571                         /*
1572                          * Set state to created. Thread will transition
1573                          * to TD_INITIALIZED when it's done setting up.
1574                          */
1575                         td_set_runstate(td, TD_CREATED);
1576                         map[this_jobs++] = td;
1577                         nr_started++;
1578
1579                         if (td->o.use_thread) {
1580                                 int ret;
1581
1582                                 dprint(FD_PROCESS, "will pthread_create\n");
1583                                 ret = pthread_create(&td->thread, NULL,
1584                                                         thread_main, td);
1585                                 if (ret) {
1586                                         log_err("pthread_create: %s\n",
1587                                                         strerror(ret));
1588                                         nr_started--;
1589                                         break;
1590                                 }
1591                                 ret = pthread_detach(td->thread);
1592                                 if (ret)
1593                                         log_err("pthread_detach: %s",
1594                                                         strerror(ret));
1595                         } else {
1596                                 pid_t pid;
1597                                 dprint(FD_PROCESS, "will fork\n");
1598                                 pid = fork();
1599                                 if (!pid) {
1600                                         int ret = fork_main(shm_id, i);
1601
1602                                         _exit(ret);
1603                                 } else if (i == fio_debug_jobno)
1604                                         *fio_debug_jobp = pid;
1605                         }
1606                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1607                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1608                                 log_err("fio: job startup hung? exiting.\n");
1609                                 terminate_threads(TERMINATE_ALL);
1610                                 fio_abort = 1;
1611                                 nr_started--;
1612                                 break;
1613                         }
1614                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1615                 }
1616
1617                 /*
1618                  * Wait for the started threads to transition to
1619                  * TD_INITIALIZED.
1620                  */
1621                 fio_gettime(&this_start, NULL);
1622                 left = this_jobs;
1623                 while (left && !fio_abort) {
1624                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1625                                 break;
1626
1627                         usleep(100000);
1628
1629                         for (i = 0; i < this_jobs; i++) {
1630                                 td = map[i];
1631                                 if (!td)
1632                                         continue;
1633                                 if (td->runstate == TD_INITIALIZED) {
1634                                         map[i] = NULL;
1635                                         left--;
1636                                 } else if (td->runstate >= TD_EXITED) {
1637                                         map[i] = NULL;
1638                                         left--;
1639                                         todo--;
1640                                         nr_running++; /* work-around... */
1641                                 }
1642                         }
1643                 }
1644
1645                 if (left) {
1646                         log_err("fio: %d jobs failed to start\n", left);
1647                         for (i = 0; i < this_jobs; i++) {
1648                                 td = map[i];
1649                                 if (!td)
1650                                         continue;
1651                                 kill(td->pid, SIGTERM);
1652                         }
1653                         break;
1654                 }
1655
1656                 /*
1657                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1658                  */
1659                 for_each_td(td, i) {
1660                         if (td->runstate != TD_INITIALIZED)
1661                                 continue;
1662
1663                         if (in_ramp_time(td))
1664                                 td_set_runstate(td, TD_RAMP);
1665                         else
1666                                 td_set_runstate(td, TD_RUNNING);
1667                         nr_running++;
1668                         nr_started--;
1669                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1670                         t_rate += td->o.rate[0] + td->o.rate[1];
1671                         todo--;
1672                         fio_mutex_up(td->mutex);
1673                 }
1674
1675                 reap_threads(&nr_running, &t_rate, &m_rate);
1676
1677                 if (todo)
1678                         usleep(100000);
1679         }
1680
1681         while (nr_running) {
1682                 reap_threads(&nr_running, &t_rate, &m_rate);
1683                 usleep(10000);
1684         }
1685
1686         update_io_ticks();
1687         fio_unpin_memory();
1688 }
1689
1690 int main(int argc, char *argv[])
1691 {
1692         long ps;
1693
1694         sinit();
1695         init_rand(&__fio_rand_state);
1696
1697         /*
1698          * We need locale for number printing, if it isn't set then just
1699          * go with the US format.
1700          */
1701         if (!getenv("LC_NUMERIC"))
1702                 setlocale(LC_NUMERIC, "en_US");
1703
1704         ps = sysconf(_SC_PAGESIZE);
1705         if (ps < 0) {
1706                 log_err("Failed to get page size\n");
1707                 return 1;
1708         }
1709
1710         page_size = ps;
1711         page_mask = ps - 1;
1712
1713         fio_keywords_init();
1714
1715         if (parse_options(argc, argv))
1716                 return 1;
1717
1718         if (exec_profile && load_profile(exec_profile))
1719                 return 1;
1720
1721         if (!thread_number)
1722                 return 0;
1723
1724         if (write_bw_log) {
1725                 setup_log(&agg_io_log[DDIR_READ]);
1726                 setup_log(&agg_io_log[DDIR_WRITE]);
1727         }
1728
1729         startup_mutex = fio_mutex_init(0);
1730         if (startup_mutex == NULL)
1731                 return 1;
1732         writeout_mutex = fio_mutex_init(1);
1733         if (writeout_mutex == NULL)
1734                 return 1;
1735
1736         set_genesis_time();
1737         create_disk_util_thread();
1738
1739         cgroup_list = smalloc(sizeof(*cgroup_list));
1740         INIT_FLIST_HEAD(cgroup_list);
1741
1742         run_threads();
1743
1744         if (!fio_abort) {
1745                 show_run_stats();
1746                 if (write_bw_log) {
1747                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1748                         __finish_log(agg_io_log[DDIR_WRITE],
1749                                         "agg-write_bw.log");
1750                 }
1751         }
1752
1753         cgroup_kill(cgroup_list);
1754         sfree(cgroup_list);
1755         sfree(cgroup_mnt);
1756
1757         fio_mutex_remove(startup_mutex);
1758         fio_mutex_remove(writeout_mutex);
1759         return exit_value;
1760 }