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. Note that we can get BUSY even
 554                  * without IO queued, if the system is resource starved.
 555                  */
 556                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 557                 if (full || !td->o.iodepth_batch_complete) {
 558                         min_events = min(td->o.iodepth_batch_complete,
 559                                          td->cur_depth);
 560                         if (full && !min_events)
 561                                 min_events = 1;
 562
 563                         do {
 564                                 /*
 565                                  * Reap required number of io units, if any,
 566                                  * and do the verification on them through
 567                                  * the callback handler
 568                                  */
 569                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 570                                         ret = -1;
 571                                         break;
 572                                 }
 573                         } while (full && (td->cur_depth > td->o.iodepth_low));
 574                 }
 575                 if (ret < 0)
 576                         break;
 577         }
 578
 579         if (!td->error) {
 580                 min_events = td->cur_depth;
 581
 582                 if (min_events)
 583                         ret = io_u_queued_complete(td, min_events, NULL);
 584         } else
 585                 cleanup_pending_aio(td);
 586
 587         td_set_runstate(td, TD_RUNNING);
 588
 589         dprint(FD_VERIFY, "exiting loop\n");
 590 }
 591
 592 /*
 593  * Main IO worker function. It retrieves io_u's to process and queues
 594  * and reaps them, checking for rate and errors along the way.
 595  */
 596 static void do_io(struct thread_data *td)
 597 {
 598         unsigned int i;
 599         int ret = 0;
 600
 601         if (in_ramp_time(td))
 602                 td_set_runstate(td, TD_RAMP);
 603         else
 604                 td_set_runstate(td, TD_RUNNING);
 605
 606         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 607                 (!flist_empty(&td->trim_list)) ||
 608                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 609                 struct timeval comp_time;
 610                 unsigned long bytes_done[2] = { 0, 0 };
 611                 int min_evts = 0;
 612                 struct io_u *io_u;
 613                 int ret2, full;
 614
 615                 if (td->terminate)
 616                         break;
 617
 618                 update_tv_cache(td);
 619
 620                 if (runtime_exceeded(td, &td->tv_cache)) {
 621                         __update_tv_cache(td);
 622                         if (runtime_exceeded(td, &td->tv_cache)) {
 623                                 td->terminate = 1;
 624                                 break;
 625                         }
 626                 }
 627
 628                 io_u = get_io_u(td);
 629                 if (!io_u)
 630                         break;
 631
 632                 /*
 633                  * Add verification end_io handler, if asked to verify
 634                  * a previously written file.
 635                  */
 636                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 637                     !td_rw(td)) {
 638                         if (td->o.verify_async)
 639                                 io_u->end_io = verify_io_u_async;
 640                         else
 641                                 io_u->end_io = verify_io_u;
 642                         td_set_runstate(td, TD_VERIFYING);
 643                 } else if (in_ramp_time(td))
 644                         td_set_runstate(td, TD_RAMP);
 645                 else
 646                         td_set_runstate(td, TD_RUNNING);
 647
 648                 ret = td_io_queue(td, io_u);
 649                 switch (ret) {
 650                 case FIO_Q_COMPLETED:
 651                         if (io_u->error) {
 652                                 ret = -io_u->error;
 653                                 clear_io_u(td, io_u);
 654                         } else if (io_u->resid) {
 655                                 int bytes = io_u->xfer_buflen - io_u->resid;
 656                                 struct fio_file *f = io_u->file;
 657
 658                                 /*
 659                                  * zero read, fail
 660                                  */
 661                                 if (!bytes) {
 662                                         td_verror(td, EIO, "full resid");
 663                                         put_io_u(td, io_u);
 664                                         break;
 665                                 }
 666
 667                                 io_u->xfer_buflen = io_u->resid;
 668                                 io_u->xfer_buf += bytes;
 669                                 io_u->offset += bytes;
 670
 671                                 if (ddir_rw(io_u->ddir))
 672                                         td->ts.short_io_u[io_u->ddir]++;
 673
 674                                 if (io_u->offset == f->real_file_size)
 675                                         goto sync_done;
 676
 677                                 requeue_io_u(td, &io_u);
 678                         } else {
 679 sync_done:
 680                                 if (__should_check_rate(td, 0) ||
 681                                     __should_check_rate(td, 1))
 682                                         fio_gettime(&comp_time, NULL);
 683
 684                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 685                                 if (ret < 0)
 686                                         break;
 687                         }
 688                         break;
 689                 case FIO_Q_QUEUED:
 690                         /*
 691                          * if the engine doesn't have a commit hook,
 692                          * the io_u is really queued. if it does have such
 693                          * a hook, it has to call io_u_queued() itself.
 694                          */
 695                         if (td->io_ops->commit == NULL)
 696                                 io_u_queued(td, io_u);
 697                         break;
 698                 case FIO_Q_BUSY:
 699                         requeue_io_u(td, &io_u);
 700                         ret2 = td_io_commit(td);
 701                         if (ret2 < 0)
 702                                 ret = ret2;
 703                         break;
 704                 default:
 705                         assert(ret < 0);
 706                         put_io_u(td, io_u);
 707                         break;
 708                 }
 709
 710                 if (break_on_this_error(td, &ret))
 711                         break;
 712
 713                 /*
 714                  * See if we need to complete some commands. Note that we
 715                  * can get BUSY even without IO queued, if the system is
 716                  * resource starved.
 717                  */
 718                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 719                 if (full || !td->o.iodepth_batch_complete) {
 720                         min_evts = min(td->o.iodepth_batch_complete,
 721                                         td->cur_depth);
 722                         if (full && !min_evts)
 723                                 min_evts = 1;
 724
 725                         if (__should_check_rate(td, 0) ||
 726                             __should_check_rate(td, 1))
 727                                 fio_gettime(&comp_time, NULL);
 728
 729                         do {
 730                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 731                                 if (ret < 0)
 732                                         break;
 733
 734                         } while (full && (td->cur_depth > td->o.iodepth_low));
 735                 }
 736
 737                 if (ret < 0)
 738                         break;
 739                 if (!(bytes_done[0] + bytes_done[1]))
 740                         continue;
 741
 742                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 743                         if (check_min_rate(td, &comp_time, bytes_done)) {
 744                                 if (exitall_on_terminate)
 745                                         terminate_threads(td->groupid);
 746                                 td_verror(td, EIO, "check_min_rate");
 747                                 break;
 748                         }
 749                 }
 750
 751                 if (td->o.thinktime) {
 752                         unsigned long long b;
 753
 754                         b = td->io_blocks[0] + td->io_blocks[1];
 755                         if (!(b % td->o.thinktime_blocks)) {
 756                                 int left;
 757
 758                                 if (td->o.thinktime_spin)
 759                                         usec_spin(td->o.thinktime_spin);
 760
 761                                 left = td->o.thinktime - td->o.thinktime_spin;
 762                                 if (left)
 763                                         usec_sleep(td, left);
 764                         }
 765                 }
 766         }
 767
 768         if (td->trim_entries)
 769                 printf("trim entries %ld\n", td->trim_entries);
 770
 771         if (td->o.fill_device && td->error == ENOSPC) {
 772                 td->error = 0;
 773                 td->terminate = 1;
 774         }
 775         if (!td->error) {
 776                 struct fio_file *f;
 777
 778                 i = td->cur_depth;
 779                 if (i) {
 780                         ret = io_u_queued_complete(td, i, NULL);
 781                         if (td->o.fill_device && td->error == ENOSPC)
 782                                 td->error = 0;
 783                 }
 784
 785                 if (should_fsync(td) && td->o.end_fsync) {
 786                         td_set_runstate(td, TD_FSYNCING);
 787
 788                         for_each_file(td, f, i) {
 789                                 if (!fio_file_open(f))
 790                                         continue;
 791                                 fio_io_sync(td, f);
 792                         }
 793                 }
 794         } else
 795                 cleanup_pending_aio(td);
 796
 797         /*
 798          * stop job if we failed doing any IO
 799          */
 800         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 801                 td->done = 1;
 802 }
 803
 804 static void cleanup_io_u(struct thread_data *td)
 805 {
 806         struct flist_head *entry, *n;
 807         struct io_u *io_u;
 808
 809         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 810                 io_u = flist_entry(entry, struct io_u, list);
 811
 812                 flist_del(&io_u->list);
 813                 fio_memfree(io_u, sizeof(*io_u));
 814         }
 815
 816         free_io_mem(td);
 817 }
 818
 819 static int init_io_u(struct thread_data *td)
 820 {
 821         struct io_u *io_u;
 822         unsigned int max_bs;
 823         int cl_align, i, max_units;
 824         char *p;
 825
 826         max_units = td->o.iodepth;
 827         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 828         td->orig_buffer_size = (unsigned long long) max_bs
 829                                         * (unsigned long long) max_units;
 830
 831         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 832                 unsigned long bs;
 833
 834                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 835                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 836         }
 837
 838         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 839                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 840                 return 1;
 841         }
 842
 843         if (allocate_io_mem(td))
 844                 return 1;
 845
 846         if (td->o.odirect || td->o.mem_align ||
 847             (td->io_ops->flags & FIO_RAWIO))
 848                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 849         else
 850                 p = td->orig_buffer;
 851
 852         cl_align = os_cache_line_size();
 853
 854         for (i = 0; i < max_units; i++) {
 855                 void *ptr;
 856
 857                 if (td->terminate)
 858                         return 1;
 859
 860                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 861                 if (!ptr) {
 862                         log_err("fio: unable to allocate aligned memory\n");
 863                         break;
 864                 }
 865
 866                 io_u = ptr;
 867                 memset(io_u, 0, sizeof(*io_u));
 868                 INIT_FLIST_HEAD(&io_u->list);
 869                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 870
 871                 if (!(td->io_ops->flags & FIO_NOIO)) {
 872                         io_u->buf = p + max_bs * i;
 873                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 874
 875                         if (td_write(td) && !td->o.refill_buffers)
 876                                 io_u_fill_buffer(td, io_u, max_bs);
 877                         else if (td_write(td) && td->o.verify_pattern_bytes) {
 878                                 /*
 879                                  * Fill the buffer with the pattern if we are
 880                                  * going to be doing writes.
 881                                  */
 882                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 883                         }
 884                 }
 885
 886                 io_u->index = i;
 887                 io_u->flags = IO_U_F_FREE;
 888                 flist_add(&io_u->list, &td->io_u_freelist);
 889         }
 890
 891         return 0;
 892 }
 893
 894 static int switch_ioscheduler(struct thread_data *td)
 895 {
 896         char tmp[256], tmp2[128];
 897         FILE *f;
 898         int ret;
 899
 900         if (td->io_ops->flags & FIO_DISKLESSIO)
 901                 return 0;
 902
 903         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 904
 905         f = fopen(tmp, "r+");
 906         if (!f) {
 907                 if (errno == ENOENT) {
 908                         log_err("fio: os or kernel doesn't support IO scheduler"
 909                                 " switching\n");
 910                         return 0;
 911                 }
 912                 td_verror(td, errno, "fopen iosched");
 913                 return 1;
 914         }
 915
 916         /*
 917          * Set io scheduler.
 918          */
 919         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 920         if (ferror(f) || ret != 1) {
 921                 td_verror(td, errno, "fwrite");
 922                 fclose(f);
 923                 return 1;
 924         }
 925
 926         rewind(f);
 927
 928         /*
 929          * Read back and check that the selected scheduler is now the default.
 930          */
 931         ret = fread(tmp, 1, sizeof(tmp), f);
 932         if (ferror(f) || ret < 0) {
 933                 td_verror(td, errno, "fread");
 934                 fclose(f);
 935                 return 1;
 936         }
 937
 938         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 939         if (!strstr(tmp, tmp2)) {
 940                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 941                 td_verror(td, EINVAL, "iosched_switch");
 942                 fclose(f);
 943                 return 1;
 944         }
 945
 946         fclose(f);
 947         return 0;
 948 }
 949
 950 static int keep_running(struct thread_data *td)
 951 {
 952         unsigned long long io_done;
 953
 954         if (td->done)
 955                 return 0;
 956         if (td->o.time_based)
 957                 return 1;
 958         if (td->o.loops) {
 959                 td->o.loops--;
 960                 return 1;
 961         }
 962
 963         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 964                         + td->io_skip_bytes;
 965         if (io_done < td->o.size)
 966                 return 1;
 967
 968         return 0;
 969 }
 970
 971 static void reset_io_counters(struct thread_data *td)
 972 {
 973         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 974         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 975         td->zone_bytes = 0;
 976         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 977         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 978
 979         td->last_was_sync = 0;
 980
 981         /*
 982          * reset file done count if we are to start over
 983          */
 984         if (td->o.time_based || td->o.loops)
 985                 td->nr_done_files = 0;
 986 }
 987
 988 void reset_all_stats(struct thread_data *td)
 989 {
 990         struct timeval tv;
 991         int i;
 992
 993         reset_io_counters(td);
 994
 995         for (i = 0; i < 2; i++) {
 996                 td->io_bytes[i] = 0;
 997                 td->io_blocks[i] = 0;
 998                 td->io_issues[i] = 0;
 999                 td->ts.total_io_u[i] = 0;
1000         }
1001
1002         fio_gettime(&tv, NULL);
1003         td->ts.runtime[0] = 0;
1004         td->ts.runtime[1] = 0;
1005         memcpy(&td->epoch, &tv, sizeof(tv));
1006         memcpy(&td->start, &tv, sizeof(tv));
1007 }
1008
1009 static void clear_io_state(struct thread_data *td)
1010 {
1011         struct fio_file *f;
1012         unsigned int i;
1013
1014         reset_io_counters(td);
1015
1016         close_files(td);
1017         for_each_file(td, f, i)
1018                 fio_file_clear_done(f);
1019
1020         /*
1021          * Set the same seed to get repeatable runs
1022          */
1023         td_fill_rand_seeds(td);
1024 }
1025
1026 static int exec_string(const char *string)
1027 {
1028         int ret, newlen = strlen(string) + 1 + 8;
1029         char *str;
1030
1031         str = malloc(newlen);
1032         sprintf(str, "sh -c %s", string);
1033
1034         ret = system(str);
1035         if (ret == -1)
1036                 log_err("fio: exec of cmd <%s> failed\n", str);
1037
1038         free(str);
1039         return ret;
1040 }
1041
1042 /*
1043  * Entry point for the thread based jobs. The process based jobs end up
1044  * here as well, after a little setup.
1045  */
1046 static void *thread_main(void *data)
1047 {
1048         unsigned long long elapsed;
1049         struct thread_data *td = data;
1050         pthread_condattr_t attr;
1051         int clear_state;
1052
1053         if (!td->o.use_thread) {
1054                 setsid();
1055                 td->pid = getpid();
1056         } else
1057                 td->pid = gettid();
1058
1059         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1060
1061         INIT_FLIST_HEAD(&td->io_u_freelist);
1062         INIT_FLIST_HEAD(&td->io_u_busylist);
1063         INIT_FLIST_HEAD(&td->io_u_requeues);
1064         INIT_FLIST_HEAD(&td->io_log_list);
1065         INIT_FLIST_HEAD(&td->io_hist_list);
1066         INIT_FLIST_HEAD(&td->verify_list);
1067         INIT_FLIST_HEAD(&td->trim_list);
1068         pthread_mutex_init(&td->io_u_lock, NULL);
1069         td->io_hist_tree = RB_ROOT;
1070
1071         pthread_condattr_init(&attr);
1072         pthread_cond_init(&td->verify_cond, &attr);
1073         pthread_cond_init(&td->free_cond, &attr);
1074
1075         td_set_runstate(td, TD_INITIALIZED);
1076         dprint(FD_MUTEX, "up startup_mutex\n");
1077         fio_mutex_up(startup_mutex);
1078         dprint(FD_MUTEX, "wait on td->mutex\n");
1079         fio_mutex_down(td->mutex);
1080         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1081
1082         /*
1083          * the ->mutex mutex is now no longer used, close it to avoid
1084          * eating a file descriptor
1085          */
1086         fio_mutex_remove(td->mutex);
1087
1088         /*
1089          * A new gid requires privilege, so we need to do this before setting
1090          * the uid.
1091          */
1092         if (td->o.gid != -1U && setgid(td->o.gid)) {
1093                 td_verror(td, errno, "setgid");
1094                 goto err;
1095         }
1096         if (td->o.uid != -1U && setuid(td->o.uid)) {
1097                 td_verror(td, errno, "setuid");
1098                 goto err;
1099         }
1100
1101         /*
1102          * If we have a gettimeofday() thread, make sure we exclude that
1103          * thread from this job
1104          */
1105         if (td->o.gtod_cpu)
1106                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1107
1108         /*
1109          * Set affinity first, in case it has an impact on the memory
1110          * allocations.
1111          */
1112         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1113                 td_verror(td, errno, "cpu_set_affinity");
1114                 goto err;
1115         }
1116
1117         /*
1118          * May alter parameters that init_io_u() will use, so we need to
1119          * do this first.
1120          */
1121         if (init_iolog(td))
1122                 goto err;
1123
1124         if (init_io_u(td))
1125                 goto err;
1126
1127         if (td->o.verify_async && verify_async_init(td))
1128                 goto err;
1129
1130         if (td->ioprio_set) {
1131                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1132                         td_verror(td, errno, "ioprio_set");
1133                         goto err;
1134                 }
1135         }
1136
1137         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1138                 goto err;
1139
1140         if (nice(td->o.nice) == -1) {
1141                 td_verror(td, errno, "nice");
1142                 goto err;
1143         }
1144
1145         if (td->o.ioscheduler && switch_ioscheduler(td))
1146                 goto err;
1147
1148         if (!td->o.create_serialize && setup_files(td))
1149                 goto err;
1150
1151         if (td_io_init(td))
1152                 goto err;
1153
1154         if (init_random_map(td))
1155                 goto err;
1156
1157         if (td->o.exec_prerun) {
1158                 if (exec_string(td->o.exec_prerun))
1159                         goto err;
1160         }
1161
1162         if (td->o.pre_read) {
1163                 if (pre_read_files(td) < 0)
1164                         goto err;
1165         }
1166
1167         fio_gettime(&td->epoch, NULL);
1168         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1169
1170         clear_state = 0;
1171         while (keep_running(td)) {
1172                 fio_gettime(&td->start, NULL);
1173                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1174                                 sizeof(td->start));
1175                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1176                                 sizeof(td->start));
1177                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1178
1179                 if (td->o.ratemin[0] || td->o.ratemin[1])
1180                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1181                                                         sizeof(td->lastrate));
1182
1183                 if (clear_state)
1184                         clear_io_state(td);
1185
1186                 prune_io_piece_log(td);
1187
1188                 do_io(td);
1189
1190                 clear_state = 1;
1191
1192                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1193                         elapsed = utime_since_now(&td->start);
1194                         td->ts.runtime[DDIR_READ] += elapsed;
1195                 }
1196                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1197                         elapsed = utime_since_now(&td->start);
1198                         td->ts.runtime[DDIR_WRITE] += elapsed;
1199                 }
1200
1201                 if (td->error || td->terminate)
1202                         break;
1203
1204                 if (!td->o.do_verify ||
1205                     td->o.verify == VERIFY_NONE ||
1206                     (td->io_ops->flags & FIO_UNIDIR))
1207                         continue;
1208
1209                 clear_io_state(td);
1210
1211                 fio_gettime(&td->start, NULL);
1212
1213                 do_verify(td);
1214
1215                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1216
1217                 if (td->error || td->terminate)
1218                         break;
1219         }
1220
1221         update_rusage_stat(td);
1222         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1223         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1224         td->ts.total_run_time = mtime_since_now(&td->epoch);
1225         td->ts.io_bytes[0] = td->io_bytes[0];
1226         td->ts.io_bytes[1] = td->io_bytes[1];
1227
1228         fio_mutex_down(writeout_mutex);
1229         if (td->ts.bw_log) {
1230                 if (td->o.bw_log_file) {
1231                         finish_log_named(td, td->ts.bw_log,
1232                                                 td->o.bw_log_file, "bw");
1233                 } else
1234                         finish_log(td, td->ts.bw_log, "bw");
1235         }
1236         if (td->ts.lat_log) {
1237                 if (td->o.lat_log_file) {
1238                         finish_log_named(td, td->ts.lat_log,
1239                                                 td->o.lat_log_file, "lat");
1240                 } else
1241                         finish_log(td, td->ts.lat_log, "lat");
1242         }
1243         if (td->ts.slat_log) {
1244                 if (td->o.lat_log_file) {
1245                         finish_log_named(td, td->ts.slat_log,
1246                                                 td->o.lat_log_file, "slat");
1247                 } else
1248                         finish_log(td, td->ts.slat_log, "slat");
1249         }
1250         if (td->ts.clat_log) {
1251                 if (td->o.lat_log_file) {
1252                         finish_log_named(td, td->ts.clat_log,
1253                                                 td->o.lat_log_file, "clat");
1254                 } else
1255                         finish_log(td, td->ts.clat_log, "clat");
1256         }
1257         fio_mutex_up(writeout_mutex);
1258         if (td->o.exec_postrun)
1259                 exec_string(td->o.exec_postrun);
1260
1261         if (exitall_on_terminate)
1262                 terminate_threads(td->groupid);
1263
1264 err:
1265         if (td->error)
1266                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1267                                                         td->verror);
1268
1269         if (td->o.verify_async)
1270                 verify_async_exit(td);
1271
1272         close_and_free_files(td);
1273         close_ioengine(td);
1274         cleanup_io_u(td);
1275         cgroup_shutdown(td, &cgroup_mnt);
1276
1277         if (td->o.cpumask_set) {
1278                 int ret = fio_cpuset_exit(&td->o.cpumask);
1279
1280                 td_verror(td, ret, "fio_cpuset_exit");
1281         }
1282
1283         /*
1284          * do this very late, it will log file closing as well
1285          */
1286         if (td->o.write_iolog_file)
1287                 write_iolog_close(td);
1288
1289         options_mem_free(td);
1290         td_set_runstate(td, TD_EXITED);
1291         return (void *) (unsigned long) td->error;
1292 }
1293
1294 /*
1295  * We cannot pass the td data into a forked process, so attach the td and
1296  * pass it to the thread worker.
1297  */
1298 static int fork_main(int shmid, int offset)
1299 {
1300         struct thread_data *td;
1301         void *data, *ret;
1302
1303 #ifndef __hpux
1304         data = shmat(shmid, NULL, 0);
1305         if (data == (void *) -1) {
1306                 int __err = errno;
1307
1308                 perror("shmat");
1309                 return __err;
1310         }
1311 #else
1312         /*
1313          * HP-UX inherits shm mappings?
1314          */
1315         data = threads;
1316 #endif
1317
1318         td = data + offset * sizeof(struct thread_data);
1319         ret = thread_main(td);
1320         shmdt(data);
1321         return (int) (unsigned long) ret;
1322 }
1323
1324 /*
1325  * Run over the job map and reap the threads that have exited, if any.
1326  */
1327 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1328 {
1329         struct thread_data *td;
1330         int i, cputhreads, realthreads, pending, status, ret;
1331
1332         /*
1333          * reap exited threads (TD_EXITED -> TD_REAPED)
1334          */
1335         realthreads = pending = cputhreads = 0;
1336         for_each_td(td, i) {
1337                 int flags = 0;
1338
1339                 /*
1340                  * ->io_ops is NULL for a thread that has closed its
1341                  * io engine
1342                  */
1343                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1344                         cputhreads++;
1345                 else
1346                         realthreads++;
1347
1348                 if (!td->pid) {
1349                         pending++;
1350                         continue;
1351                 }
1352                 if (td->runstate == TD_REAPED)
1353                         continue;
1354                 if (td->o.use_thread) {
1355                         if (td->runstate == TD_EXITED) {
1356                                 td_set_runstate(td, TD_REAPED);
1357                                 goto reaped;
1358                         }
1359                         continue;
1360                 }
1361
1362                 flags = WNOHANG;
1363                 if (td->runstate == TD_EXITED)
1364                         flags = 0;
1365
1366                 /*
1367                  * check if someone quit or got killed in an unusual way
1368                  */
1369                 ret = waitpid(td->pid, &status, flags);
1370                 if (ret < 0) {
1371                         if (errno == ECHILD) {
1372                                 log_err("fio: pid=%d disappeared %d\n",
1373                                                 (int) td->pid, td->runstate);
1374                                 td_set_runstate(td, TD_REAPED);
1375                                 goto reaped;
1376                         }
1377                         perror("waitpid");
1378                 } else if (ret == td->pid) {
1379                         if (WIFSIGNALED(status)) {
1380                                 int sig = WTERMSIG(status);
1381
1382                                 if (sig != SIGTERM)
1383                                         log_err("fio: pid=%d, got signal=%d\n",
1384                                                         (int) td->pid, sig);
1385                                 td_set_runstate(td, TD_REAPED);
1386                                 goto reaped;
1387                         }
1388                         if (WIFEXITED(status)) {
1389                                 if (WEXITSTATUS(status) && !td->error)
1390                                         td->error = WEXITSTATUS(status);
1391
1392                                 td_set_runstate(td, TD_REAPED);
1393                                 goto reaped;
1394                         }
1395                 }
1396
1397                 /*
1398                  * thread is not dead, continue
1399                  */
1400                 pending++;
1401                 continue;
1402 reaped:
1403                 (*nr_running)--;
1404                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1405                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1406                 if (!td->pid)
1407                         pending--;
1408
1409                 if (td->error)
1410                         exit_value++;
1411
1412                 done_secs += mtime_since_now(&td->epoch) / 1000;
1413         }
1414
1415         if (*nr_running == cputhreads && !pending && realthreads)
1416                 terminate_threads(TERMINATE_ALL);
1417 }
1418
1419 static void *gtod_thread_main(void *data)
1420 {
1421         fio_mutex_up(startup_mutex);
1422
1423         /*
1424          * As long as we have jobs around, update the clock. It would be nice
1425          * to have some way of NOT hammering that CPU with gettimeofday(),
1426          * but I'm not sure what to use outside of a simple CPU nop to relax
1427          * it - we don't want to lose precision.
1428          */
1429         while (threads) {
1430                 fio_gtod_update();
1431                 nop;
1432         }
1433
1434         return NULL;
1435 }
1436
1437 static int fio_start_gtod_thread(void)
1438 {
1439         pthread_attr_t attr;
1440         int ret;
1441
1442         pthread_attr_init(&attr);
1443         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1444         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1445         pthread_attr_destroy(&attr);
1446         if (ret) {
1447                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1448                 return 1;
1449         }
1450
1451         ret = pthread_detach(gtod_thread);
1452         if (ret) {
1453                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1454                 return 1;
1455         }
1456
1457         dprint(FD_MUTEX, "wait on startup_mutex\n");
1458         fio_mutex_down(startup_mutex);
1459         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1460         return 0;
1461 }
1462
1463 /*
1464  * Main function for kicking off and reaping jobs, as needed.
1465  */
1466 static void run_threads(void)
1467 {
1468         struct thread_data *td;
1469         unsigned long spent;
1470         int i, todo, nr_running, m_rate, t_rate, nr_started;
1471
1472         if (fio_pin_memory())
1473                 return;
1474
1475         if (fio_gtod_offload && fio_start_gtod_thread())
1476                 return;
1477
1478         if (!terse_output) {
1479                 log_info("Starting ");
1480                 if (nr_thread)
1481                         log_info("%d thread%s", nr_thread,
1482                                                 nr_thread > 1 ? "s" : "");
1483                 if (nr_process) {
1484                         if (nr_thread)
1485                                 printf(" and ");
1486                         log_info("%d process%s", nr_process,
1487                                                 nr_process > 1 ? "es" : "");
1488                 }
1489                 log_info("\n");
1490                 fflush(stdout);
1491         }
1492
1493         set_sig_handlers();
1494
1495         todo = thread_number;
1496         nr_running = 0;
1497         nr_started = 0;
1498         m_rate = t_rate = 0;
1499
1500         for_each_td(td, i) {
1501                 print_status_init(td->thread_number - 1);
1502
1503                 if (!td->o.create_serialize) {
1504                         init_disk_util(td);
1505                         continue;
1506                 }
1507
1508                 /*
1509                  * do file setup here so it happens sequentially,
1510                  * we don't want X number of threads getting their
1511                  * client data interspersed on disk
1512                  */
1513                 if (setup_files(td)) {
1514                         exit_value++;
1515                         if (td->error)
1516                                 log_err("fio: pid=%d, err=%d/%s\n",
1517                                         (int) td->pid, td->error, td->verror);
1518                         td_set_runstate(td, TD_REAPED);
1519                         todo--;
1520                 } else {
1521                         struct fio_file *f;
1522                         unsigned int j;
1523
1524                         /*
1525                          * for sharing to work, each job must always open
1526                          * its own files. so close them, if we opened them
1527                          * for creation
1528                          */
1529                         for_each_file(td, f, j) {
1530                                 if (fio_file_open(f))
1531                                         td_io_close_file(td, f);
1532                         }
1533                 }
1534
1535                 init_disk_util(td);
1536         }
1537
1538         set_genesis_time();
1539
1540         while (todo) {
1541                 struct thread_data *map[REAL_MAX_JOBS];
1542                 struct timeval this_start;
1543                 int this_jobs = 0, left;
1544
1545                 /*
1546                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1547                  */
1548                 for_each_td(td, i) {
1549                         if (td->runstate != TD_NOT_CREATED)
1550                                 continue;
1551
1552                         /*
1553                          * never got a chance to start, killed by other
1554                          * thread for some reason
1555                          */
1556                         if (td->terminate) {
1557                                 todo--;
1558                                 continue;
1559                         }
1560
1561                         if (td->o.start_delay) {
1562                                 spent = mtime_since_genesis();
1563
1564                                 if (td->o.start_delay * 1000 > spent)
1565                                         continue;
1566                         }
1567
1568                         if (td->o.stonewall && (nr_started || nr_running)) {
1569                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1570                                                         td->o.name);
1571                                 break;
1572                         }
1573
1574                         /*
1575                          * Set state to created. Thread will transition
1576                          * to TD_INITIALIZED when it's done setting up.
1577                          */
1578                         td_set_runstate(td, TD_CREATED);
1579                         map[this_jobs++] = td;
1580                         nr_started++;
1581
1582                         if (td->o.use_thread) {
1583                                 int ret;
1584
1585                                 dprint(FD_PROCESS, "will pthread_create\n");
1586                                 ret = pthread_create(&td->thread, NULL,
1587                                                         thread_main, td);
1588                                 if (ret) {
1589                                         log_err("pthread_create: %s\n",
1590                                                         strerror(ret));
1591                                         nr_started--;
1592                                         break;
1593                                 }
1594                                 ret = pthread_detach(td->thread);
1595                                 if (ret)
1596                                         log_err("pthread_detach: %s",
1597                                                         strerror(ret));
1598                         } else {
1599                                 pid_t pid;
1600                                 dprint(FD_PROCESS, "will fork\n");
1601                                 pid = fork();
1602                                 if (!pid) {
1603                                         int ret = fork_main(shm_id, i);
1604
1605                                         _exit(ret);
1606                                 } else if (i == fio_debug_jobno)
1607                                         *fio_debug_jobp = pid;
1608                         }
1609                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1610                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1611                                 log_err("fio: job startup hung? exiting.\n");
1612                                 terminate_threads(TERMINATE_ALL);
1613                                 fio_abort = 1;
1614                                 nr_started--;
1615                                 break;
1616                         }
1617                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1618                 }
1619
1620                 /*
1621                  * Wait for the started threads to transition to
1622                  * TD_INITIALIZED.
1623                  */
1624                 fio_gettime(&this_start, NULL);
1625                 left = this_jobs;
1626                 while (left && !fio_abort) {
1627                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1628                                 break;
1629
1630                         usleep(100000);
1631
1632                         for (i = 0; i < this_jobs; i++) {
1633                                 td = map[i];
1634                                 if (!td)
1635                                         continue;
1636                                 if (td->runstate == TD_INITIALIZED) {
1637                                         map[i] = NULL;
1638                                         left--;
1639                                 } else if (td->runstate >= TD_EXITED) {
1640                                         map[i] = NULL;
1641                                         left--;
1642                                         todo--;
1643                                         nr_running++; /* work-around... */
1644                                 }
1645                         }
1646                 }
1647
1648                 if (left) {
1649                         log_err("fio: %d jobs failed to start\n", left);
1650                         for (i = 0; i < this_jobs; i++) {
1651                                 td = map[i];
1652                                 if (!td)
1653                                         continue;
1654                                 kill(td->pid, SIGTERM);
1655                         }
1656                         break;
1657                 }
1658
1659                 /*
1660                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1661                  */
1662                 for_each_td(td, i) {
1663                         if (td->runstate != TD_INITIALIZED)
1664                                 continue;
1665
1666                         if (in_ramp_time(td))
1667                                 td_set_runstate(td, TD_RAMP);
1668                         else
1669                                 td_set_runstate(td, TD_RUNNING);
1670                         nr_running++;
1671                         nr_started--;
1672                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1673                         t_rate += td->o.rate[0] + td->o.rate[1];
1674                         todo--;
1675                         fio_mutex_up(td->mutex);
1676                 }
1677
1678                 reap_threads(&nr_running, &t_rate, &m_rate);
1679
1680                 if (todo)
1681                         usleep(100000);
1682         }
1683
1684         while (nr_running) {
1685                 reap_threads(&nr_running, &t_rate, &m_rate);
1686                 usleep(10000);
1687         }
1688
1689         update_io_ticks();
1690         fio_unpin_memory();
1691 }
1692
1693 int main(int argc, char *argv[])
1694 {
1695         long ps;
1696
1697         sinit();
1698         init_rand(&__fio_rand_state);
1699
1700         /*
1701          * We need locale for number printing, if it isn't set then just
1702          * go with the US format.
1703          */
1704         if (!getenv("LC_NUMERIC"))
1705                 setlocale(LC_NUMERIC, "en_US");
1706
1707         ps = sysconf(_SC_PAGESIZE);
1708         if (ps < 0) {
1709                 log_err("Failed to get page size\n");
1710                 return 1;
1711         }
1712
1713         page_size = ps;
1714         page_mask = ps - 1;
1715
1716         fio_keywords_init();
1717
1718         if (parse_options(argc, argv))
1719                 return 1;
1720
1721         if (exec_profile && load_profile(exec_profile))
1722                 return 1;
1723
1724         if (!thread_number)
1725                 return 0;
1726
1727         if (write_bw_log) {
1728                 setup_log(&agg_io_log[DDIR_READ]);
1729                 setup_log(&agg_io_log[DDIR_WRITE]);
1730         }
1731
1732         startup_mutex = fio_mutex_init(0);
1733         if (startup_mutex == NULL)
1734                 return 1;
1735         writeout_mutex = fio_mutex_init(1);
1736         if (writeout_mutex == NULL)
1737                 return 1;
1738
1739         set_genesis_time();
1740         create_disk_util_thread();
1741
1742         cgroup_list = smalloc(sizeof(*cgroup_list));
1743         INIT_FLIST_HEAD(cgroup_list);
1744
1745         run_threads();
1746
1747         if (!fio_abort) {
1748                 show_run_stats();
1749                 if (write_bw_log) {
1750                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1751                         __finish_log(agg_io_log[DDIR_WRITE],
1752                                         "agg-write_bw.log");
1753                 }
1754         }
1755
1756         cgroup_kill(cgroup_list);
1757         sfree(cgroup_list);
1758         sfree(cgroup_mnt);
1759
1760         fio_mutex_remove(startup_mutex);
1761         fio_mutex_remove(writeout_mutex);
1762         return exit_value;
1763 }