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 #include "server.h"
  50
  51 unsigned long page_mask;
  52 unsigned long page_size;
  53
  54 #define PAGE_ALIGN(buf) \
  55         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  56
  57 int groupid = 0;
  58 int thread_number = 0;
  59 int nr_process = 0;
  60 int nr_thread = 0;
  61 int shm_id = 0;
  62 int temp_stall_ts;
  63 unsigned long done_secs = 0;
  64
  65 static struct fio_mutex *startup_mutex;
  66 static struct fio_mutex *writeout_mutex;
  67 static volatile int fio_abort;
  68 static int exit_value;
  69 static pthread_t gtod_thread;
  70 static pthread_t disk_util_thread;
  71 static struct flist_head *cgroup_list;
  72 static char *cgroup_mnt;
  73
  74 unsigned long arch_flags = 0;
  75
  76 struct io_log *agg_io_log[2];
  77
  78 #define TERMINATE_ALL           (-1)
  79 #define JOB_START_TIMEOUT       (5 * 1000)
  80
  81 void td_set_runstate(struct thread_data *td, int runstate)
  82 {
  83         if (td->runstate == runstate)
  84                 return;
  85
  86         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  87                                                 td->runstate, runstate);
  88         td->runstate = runstate;
  89 }
  90
  91 static void terminate_threads(int group_id)
  92 {
  93         struct thread_data *td;
  94         int i;
  95
  96         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  97
  98         for_each_td(td, i) {
  99                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
 100                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
 101                                                 td->o.name, (int) td->pid);
 102                         td->terminate = 1;
 103                         td->o.start_delay = 0;
 104
 105                         /*
 106                          * if the thread is running, just let it exit
 107                          */
 108                         if (!td->pid)
 109                                 continue;
 110                         else if (td->runstate < TD_RAMP)
 111                                 kill(td->pid, SIGTERM);
 112                         else {
 113                                 struct ioengine_ops *ops = td->io_ops;
 114
 115                                 if (ops && (ops->flags & FIO_SIGTERM))
 116                                         kill(td->pid, SIGTERM);
 117                         }
 118                 }
 119         }
 120 }
 121
 122 static void sig_int(int sig)
 123 {
 124         if (threads) {
 125                 log_info("\nfio: terminating on signal %d\n", sig);
 126                 exit_backend = 1;
 127                 fflush(stdout);
 128                 exit_value = 128;
 129                 terminate_threads(TERMINATE_ALL);
 130         }
 131 }
 132
 133 static void *disk_thread_main(void *data)
 134 {
 135         fio_mutex_up(startup_mutex);
 136
 137         while (threads) {
 138                 usleep(DISK_UTIL_MSEC * 1000);
 139                 if (!threads)
 140                         break;
 141                 update_io_ticks();
 142
 143                 if (is_backend)
 144                         fio_server_send_status();
 145                 else
 146                         print_thread_status();
 147         }
 148
 149         return NULL;
 150 }
 151
 152 static int create_disk_util_thread(void)
 153 {
 154         int ret;
 155
 156         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 157         if (ret) {
 158                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 159                 return 1;
 160         }
 161
 162         ret = pthread_detach(disk_util_thread);
 163         if (ret) {
 164                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 165                 return 1;
 166         }
 167
 168         dprint(FD_MUTEX, "wait on startup_mutex\n");
 169         fio_mutex_down(startup_mutex);
 170         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 171         return 0;
 172 }
 173
 174 static void set_sig_handlers(void)
 175 {
 176         struct sigaction act;
 177
 178         memset(&act, 0, sizeof(act));
 179         act.sa_handler = sig_int;
 180         act.sa_flags = SA_RESTART;
 181         sigaction(SIGINT, &act, NULL);
 182
 183         memset(&act, 0, sizeof(act));
 184         act.sa_handler = sig_int;
 185         act.sa_flags = SA_RESTART;
 186         sigaction(SIGTERM, &act, NULL);
 187 }
 188
 189 /*
 190  * Check if we are above the minimum rate given.
 191  */
 192 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 193                             enum fio_ddir ddir)
 194 {
 195         unsigned long long bytes = 0;
 196         unsigned long iops = 0;
 197         unsigned long spent;
 198         unsigned long rate;
 199         unsigned int ratemin = 0;
 200         unsigned int rate_iops = 0;
 201         unsigned int rate_iops_min = 0;
 202
 203         assert(ddir_rw(ddir));
 204
 205         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 206                 return 0;
 207
 208         /*
 209          * allow a 2 second settle period in the beginning
 210          */
 211         if (mtime_since(&td->start, now) < 2000)
 212                 return 0;
 213
 214         iops += td->io_blocks[ddir];
 215         bytes += td->this_io_bytes[ddir];
 216         ratemin += td->o.ratemin[ddir];
 217         rate_iops += td->o.rate_iops[ddir];
 218         rate_iops_min += td->o.rate_iops_min[ddir];
 219
 220         /*
 221          * if rate blocks is set, sample is running
 222          */
 223         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 224                 spent = mtime_since(&td->lastrate[ddir], now);
 225                 if (spent < td->o.ratecycle)
 226                         return 0;
 227
 228                 if (td->o.rate[ddir]) {
 229                         /*
 230                          * check bandwidth specified rate
 231                          */
 232                         if (bytes < td->rate_bytes[ddir]) {
 233                                 log_err("%s: min rate %u not met\n", td->o.name,
 234                                                                 ratemin);
 235                                 return 1;
 236                         } else {
 237                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 238                                 if (rate < ratemin ||
 239                                     bytes < td->rate_bytes[ddir]) {
 240                                         log_err("%s: min rate %u not met, got"
 241                                                 " %luKB/sec\n", td->o.name,
 242                                                         ratemin, rate);
 243                                         return 1;
 244                                 }
 245                         }
 246                 } else {
 247                         /*
 248                          * checks iops specified rate
 249                          */
 250                         if (iops < rate_iops) {
 251                                 log_err("%s: min iops rate %u not met\n",
 252                                                 td->o.name, rate_iops);
 253                                 return 1;
 254                         } else {
 255                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 256                                 if (rate < rate_iops_min ||
 257                                     iops < td->rate_blocks[ddir]) {
 258                                         log_err("%s: min iops rate %u not met,"
 259                                                 " got %lu\n", td->o.name,
 260                                                         rate_iops_min, rate);
 261                                 }
 262                         }
 263                 }
 264         }
 265
 266         td->rate_bytes[ddir] = bytes;
 267         td->rate_blocks[ddir] = iops;
 268         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 269         return 0;
 270 }
 271
 272 static int check_min_rate(struct thread_data *td, struct timeval *now,
 273                           unsigned long *bytes_done)
 274 {
 275         int ret = 0;
 276
 277         if (bytes_done[0])
 278                 ret |= __check_min_rate(td, now, 0);
 279         if (bytes_done[1])
 280                 ret |= __check_min_rate(td, now, 1);
 281
 282         return ret;
 283 }
 284
 285 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 286 {
 287         if (!td->o.timeout)
 288                 return 0;
 289         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 290                 return 1;
 291
 292         return 0;
 293 }
 294
 295 /*
 296  * When job exits, we can cancel the in-flight IO if we are using async
 297  * io. Attempt to do so.
 298  */
 299 static void cleanup_pending_aio(struct thread_data *td)
 300 {
 301         struct flist_head *entry, *n;
 302         struct io_u *io_u;
 303         int r;
 304
 305         /*
 306          * get immediately available events, if any
 307          */
 308         r = io_u_queued_complete(td, 0, NULL);
 309         if (r < 0)
 310                 return;
 311
 312         /*
 313          * now cancel remaining active events
 314          */
 315         if (td->io_ops->cancel) {
 316                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 317                         io_u = flist_entry(entry, struct io_u, list);
 318
 319                         /*
 320                          * if the io_u isn't in flight, then that generally
 321                          * means someone leaked an io_u. complain but fix
 322                          * it up, so we don't stall here.
 323                          */
 324                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 325                                 log_err("fio: non-busy IO on busy list\n");
 326                                 put_io_u(td, io_u);
 327                         } else {
 328                                 r = td->io_ops->cancel(td, io_u);
 329                                 if (!r)
 330                                         put_io_u(td, io_u);
 331                         }
 332                 }
 333         }
 334
 335         if (td->cur_depth)
 336                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 337 }
 338
 339 /*
 340  * Helper to handle the final sync of a file. Works just like the normal
 341  * io path, just does everything sync.
 342  */
 343 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 344 {
 345         struct io_u *io_u = __get_io_u(td);
 346         int ret;
 347
 348         if (!io_u)
 349                 return 1;
 350
 351         io_u->ddir = DDIR_SYNC;
 352         io_u->file = f;
 353
 354         if (td_io_prep(td, io_u)) {
 355                 put_io_u(td, io_u);
 356                 return 1;
 357         }
 358
 359 requeue:
 360         ret = td_io_queue(td, io_u);
 361         if (ret < 0) {
 362                 td_verror(td, io_u->error, "td_io_queue");
 363                 put_io_u(td, io_u);
 364                 return 1;
 365         } else if (ret == FIO_Q_QUEUED) {
 366                 if (io_u_queued_complete(td, 1, NULL) < 0)
 367                         return 1;
 368         } else if (ret == FIO_Q_COMPLETED) {
 369                 if (io_u->error) {
 370                         td_verror(td, io_u->error, "td_io_queue");
 371                         return 1;
 372                 }
 373
 374                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 375                         return 1;
 376         } else if (ret == FIO_Q_BUSY) {
 377                 if (td_io_commit(td))
 378                         return 1;
 379                 goto requeue;
 380         }
 381
 382         return 0;
 383 }
 384
 385 static inline void __update_tv_cache(struct thread_data *td)
 386 {
 387         fio_gettime(&td->tv_cache, NULL);
 388 }
 389
 390 static inline void update_tv_cache(struct thread_data *td)
 391 {
 392         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 393                 __update_tv_cache(td);
 394 }
 395
 396 static int break_on_this_error(struct thread_data *td, int *retptr)
 397 {
 398         int ret = *retptr;
 399
 400         if (ret < 0 || td->error) {
 401                 int err;
 402
 403                 if (!td->o.continue_on_error)
 404                         return 1;
 405
 406                 if (ret < 0)
 407                         err = -ret;
 408                 else
 409                         err = td->error;
 410
 411                 if (td_non_fatal_error(err)) {
 412                         /*
 413                          * Continue with the I/Os in case of
 414                          * a non fatal error.
 415                          */
 416                         update_error_count(td, err);
 417                         td_clear_error(td);
 418                         *retptr = 0;
 419                         return 0;
 420                 } else if (td->o.fill_device && err == ENOSPC) {
 421                         /*
 422                          * We expect to hit this error if
 423                          * fill_device option is set.
 424                          */
 425                         td_clear_error(td);
 426                         td->terminate = 1;
 427                         return 1;
 428                 } else {
 429                         /*
 430                          * Stop the I/O in case of a fatal
 431                          * error.
 432                          */
 433                         update_error_count(td, err);
 434                         return 1;
 435                 }
 436         }
 437
 438         return 0;
 439 }
 440
 441 /*
 442  * The main verify engine. Runs over the writes we previously submitted,
 443  * reads the blocks back in, and checks the crc/md5 of the data.
 444  */
 445 static void do_verify(struct thread_data *td)
 446 {
 447         struct fio_file *f;
 448         struct io_u *io_u;
 449         int ret, min_events;
 450         unsigned int i;
 451
 452         dprint(FD_VERIFY, "starting loop\n");
 453
 454         /*
 455          * sync io first and invalidate cache, to make sure we really
 456          * read from disk.
 457          */
 458         for_each_file(td, f, i) {
 459                 if (!fio_file_open(f))
 460                         continue;
 461                 if (fio_io_sync(td, f))
 462                         break;
 463                 if (file_invalidate_cache(td, f))
 464                         break;
 465         }
 466
 467         if (td->error)
 468                 return;
 469
 470         td_set_runstate(td, TD_VERIFYING);
 471
 472         io_u = NULL;
 473         while (!td->terminate) {
 474                 int ret2, full;
 475
 476                 update_tv_cache(td);
 477
 478                 if (runtime_exceeded(td, &td->tv_cache)) {
 479                         __update_tv_cache(td);
 480                         if (runtime_exceeded(td, &td->tv_cache)) {
 481                                 td->terminate = 1;
 482                                 break;
 483                         }
 484                 }
 485
 486                 io_u = __get_io_u(td);
 487                 if (!io_u)
 488                         break;
 489
 490                 if (get_next_verify(td, io_u)) {
 491                         put_io_u(td, io_u);
 492                         break;
 493                 }
 494
 495                 if (td_io_prep(td, io_u)) {
 496                         put_io_u(td, io_u);
 497                         break;
 498                 }
 499
 500                 if (td->o.verify_async)
 501                         io_u->end_io = verify_io_u_async;
 502                 else
 503                         io_u->end_io = verify_io_u;
 504
 505                 ret = td_io_queue(td, io_u);
 506                 switch (ret) {
 507                 case FIO_Q_COMPLETED:
 508                         if (io_u->error) {
 509                                 ret = -io_u->error;
 510                                 clear_io_u(td, io_u);
 511                         } else if (io_u->resid) {
 512                                 int bytes = io_u->xfer_buflen - io_u->resid;
 513
 514                                 /*
 515                                  * zero read, fail
 516                                  */
 517                                 if (!bytes) {
 518                                         td_verror(td, EIO, "full resid");
 519                                         put_io_u(td, io_u);
 520                                         break;
 521                                 }
 522
 523                                 io_u->xfer_buflen = io_u->resid;
 524                                 io_u->xfer_buf += bytes;
 525                                 io_u->offset += bytes;
 526
 527                                 if (ddir_rw(io_u->ddir))
 528                                         td->ts.short_io_u[io_u->ddir]++;
 529
 530                                 f = io_u->file;
 531                                 if (io_u->offset == f->real_file_size)
 532                                         goto sync_done;
 533
 534                                 requeue_io_u(td, &io_u);
 535                         } else {
 536 sync_done:
 537                                 ret = io_u_sync_complete(td, io_u, NULL);
 538                                 if (ret < 0)
 539                                         break;
 540                         }
 541                         continue;
 542                 case FIO_Q_QUEUED:
 543                         break;
 544                 case FIO_Q_BUSY:
 545                         requeue_io_u(td, &io_u);
 546                         ret2 = td_io_commit(td);
 547                         if (ret2 < 0)
 548                                 ret = ret2;
 549                         break;
 550                 default:
 551                         assert(ret < 0);
 552                         td_verror(td, -ret, "td_io_queue");
 553                         break;
 554                 }
 555
 556                 if (break_on_this_error(td, &ret))
 557                         break;
 558
 559                 /*
 560                  * if we can queue more, do so. but check if there are
 561                  * completed io_u's first. Note that we can get BUSY even
 562                  * without IO queued, if the system is resource starved.
 563                  */
 564                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 565                 if (full || !td->o.iodepth_batch_complete) {
 566                         min_events = min(td->o.iodepth_batch_complete,
 567                                          td->cur_depth);
 568                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 569                                 min_events = 1;
 570
 571                         do {
 572                                 /*
 573                                  * Reap required number of io units, if any,
 574                                  * and do the verification on them through
 575                                  * the callback handler
 576                                  */
 577                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 578                                         ret = -1;
 579                                         break;
 580                                 }
 581                         } while (full && (td->cur_depth > td->o.iodepth_low));
 582                 }
 583                 if (ret < 0)
 584                         break;
 585         }
 586
 587         if (!td->error) {
 588                 min_events = td->cur_depth;
 589
 590                 if (min_events)
 591                         ret = io_u_queued_complete(td, min_events, NULL);
 592         } else
 593                 cleanup_pending_aio(td);
 594
 595         td_set_runstate(td, TD_RUNNING);
 596
 597         dprint(FD_VERIFY, "exiting loop\n");
 598 }
 599
 600 /*
 601  * Main IO worker function. It retrieves io_u's to process and queues
 602  * and reaps them, checking for rate and errors along the way.
 603  */
 604 static void do_io(struct thread_data *td)
 605 {
 606         unsigned int i;
 607         int ret = 0;
 608
 609         if (in_ramp_time(td))
 610                 td_set_runstate(td, TD_RAMP);
 611         else
 612                 td_set_runstate(td, TD_RUNNING);
 613
 614         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 615                 (!flist_empty(&td->trim_list)) ||
 616                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 617                 struct timeval comp_time;
 618                 unsigned long bytes_done[2] = { 0, 0 };
 619                 int min_evts = 0;
 620                 struct io_u *io_u;
 621                 int ret2, full;
 622
 623                 if (td->terminate)
 624                         break;
 625
 626                 update_tv_cache(td);
 627
 628                 if (runtime_exceeded(td, &td->tv_cache)) {
 629                         __update_tv_cache(td);
 630                         if (runtime_exceeded(td, &td->tv_cache)) {
 631                                 td->terminate = 1;
 632                                 break;
 633                         }
 634                 }
 635
 636                 io_u = get_io_u(td);
 637                 if (!io_u)
 638                         break;
 639
 640                 /*
 641                  * Add verification end_io handler, if asked to verify
 642                  * a previously written file.
 643                  */
 644                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 645                     !td_rw(td)) {
 646                         if (td->o.verify_async)
 647                                 io_u->end_io = verify_io_u_async;
 648                         else
 649                                 io_u->end_io = verify_io_u;
 650                         td_set_runstate(td, TD_VERIFYING);
 651                 } else if (in_ramp_time(td))
 652                         td_set_runstate(td, TD_RAMP);
 653                 else
 654                         td_set_runstate(td, TD_RUNNING);
 655
 656                 ret = td_io_queue(td, io_u);
 657                 switch (ret) {
 658                 case FIO_Q_COMPLETED:
 659                         if (io_u->error) {
 660                                 ret = -io_u->error;
 661                                 clear_io_u(td, io_u);
 662                         } else if (io_u->resid) {
 663                                 int bytes = io_u->xfer_buflen - io_u->resid;
 664                                 struct fio_file *f = io_u->file;
 665
 666                                 /*
 667                                  * zero read, fail
 668                                  */
 669                                 if (!bytes) {
 670                                         td_verror(td, EIO, "full resid");
 671                                         put_io_u(td, io_u);
 672                                         break;
 673                                 }
 674
 675                                 io_u->xfer_buflen = io_u->resid;
 676                                 io_u->xfer_buf += bytes;
 677                                 io_u->offset += bytes;
 678
 679                                 if (ddir_rw(io_u->ddir))
 680                                         td->ts.short_io_u[io_u->ddir]++;
 681
 682                                 if (io_u->offset == f->real_file_size)
 683                                         goto sync_done;
 684
 685                                 requeue_io_u(td, &io_u);
 686                         } else {
 687 sync_done:
 688                                 if (__should_check_rate(td, 0) ||
 689                                     __should_check_rate(td, 1))
 690                                         fio_gettime(&comp_time, NULL);
 691
 692                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 693                                 if (ret < 0)
 694                                         break;
 695                         }
 696                         break;
 697                 case FIO_Q_QUEUED:
 698                         /*
 699                          * if the engine doesn't have a commit hook,
 700                          * the io_u is really queued. if it does have such
 701                          * a hook, it has to call io_u_queued() itself.
 702                          */
 703                         if (td->io_ops->commit == NULL)
 704                                 io_u_queued(td, io_u);
 705                         break;
 706                 case FIO_Q_BUSY:
 707                         requeue_io_u(td, &io_u);
 708                         ret2 = td_io_commit(td);
 709                         if (ret2 < 0)
 710                                 ret = ret2;
 711                         break;
 712                 default:
 713                         assert(ret < 0);
 714                         put_io_u(td, io_u);
 715                         break;
 716                 }
 717
 718                 if (break_on_this_error(td, &ret))
 719                         break;
 720
 721                 /*
 722                  * See if we need to complete some commands. Note that we
 723                  * can get BUSY even without IO queued, if the system is
 724                  * resource starved.
 725                  */
 726                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 727                 if (full || !td->o.iodepth_batch_complete) {
 728                         min_evts = min(td->o.iodepth_batch_complete,
 729                                         td->cur_depth);
 730                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 731                                 min_evts = 1;
 732
 733                         if (__should_check_rate(td, 0) ||
 734                             __should_check_rate(td, 1))
 735                                 fio_gettime(&comp_time, NULL);
 736
 737                         do {
 738                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 739                                 if (ret < 0)
 740                                         break;
 741
 742                         } while (full && (td->cur_depth > td->o.iodepth_low));
 743                 }
 744
 745                 if (ret < 0)
 746                         break;
 747                 if (!(bytes_done[0] + bytes_done[1]))
 748                         continue;
 749
 750                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 751                         if (check_min_rate(td, &comp_time, bytes_done)) {
 752                                 if (exitall_on_terminate)
 753                                         terminate_threads(td->groupid);
 754                                 td_verror(td, EIO, "check_min_rate");
 755                                 break;
 756                         }
 757                 }
 758
 759                 if (td->o.thinktime) {
 760                         unsigned long long b;
 761
 762                         b = td->io_blocks[0] + td->io_blocks[1];
 763                         if (!(b % td->o.thinktime_blocks)) {
 764                                 int left;
 765
 766                                 if (td->o.thinktime_spin)
 767                                         usec_spin(td->o.thinktime_spin);
 768
 769                                 left = td->o.thinktime - td->o.thinktime_spin;
 770                                 if (left)
 771                                         usec_sleep(td, left);
 772                         }
 773                 }
 774         }
 775
 776         if (td->trim_entries)
 777                 printf("trim entries %ld\n", td->trim_entries);
 778
 779         if (td->o.fill_device && td->error == ENOSPC) {
 780                 td->error = 0;
 781                 td->terminate = 1;
 782         }
 783         if (!td->error) {
 784                 struct fio_file *f;
 785
 786                 i = td->cur_depth;
 787                 if (i) {
 788                         ret = io_u_queued_complete(td, i, NULL);
 789                         if (td->o.fill_device && td->error == ENOSPC)
 790                                 td->error = 0;
 791                 }
 792
 793                 if (should_fsync(td) && td->o.end_fsync) {
 794                         td_set_runstate(td, TD_FSYNCING);
 795
 796                         for_each_file(td, f, i) {
 797                                 if (!fio_file_open(f))
 798                                         continue;
 799                                 fio_io_sync(td, f);
 800                         }
 801                 }
 802         } else
 803                 cleanup_pending_aio(td);
 804
 805         /*
 806          * stop job if we failed doing any IO
 807          */
 808         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 809                 td->done = 1;
 810 }
 811
 812 static void cleanup_io_u(struct thread_data *td)
 813 {
 814         struct flist_head *entry, *n;
 815         struct io_u *io_u;
 816
 817         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 818                 io_u = flist_entry(entry, struct io_u, list);
 819
 820                 flist_del(&io_u->list);
 821                 fio_memfree(io_u, sizeof(*io_u));
 822         }
 823
 824         free_io_mem(td);
 825 }
 826
 827 static int init_io_u(struct thread_data *td)
 828 {
 829         struct io_u *io_u;
 830         unsigned int max_bs;
 831         int cl_align, i, max_units;
 832         char *p;
 833
 834         max_units = td->o.iodepth;
 835         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 836         td->orig_buffer_size = (unsigned long long) max_bs
 837                                         * (unsigned long long) max_units;
 838
 839         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 840                 unsigned long bs;
 841
 842                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 843                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 844         }
 845
 846         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 847                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 848                 return 1;
 849         }
 850
 851         if (allocate_io_mem(td))
 852                 return 1;
 853
 854         if (td->o.odirect || td->o.mem_align ||
 855             (td->io_ops->flags & FIO_RAWIO))
 856                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 857         else
 858                 p = td->orig_buffer;
 859
 860         cl_align = os_cache_line_size();
 861
 862         for (i = 0; i < max_units; i++) {
 863                 void *ptr;
 864
 865                 if (td->terminate)
 866                         return 1;
 867
 868                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 869                 if (!ptr) {
 870                         log_err("fio: unable to allocate aligned memory\n");
 871                         break;
 872                 }
 873
 874                 io_u = ptr;
 875                 memset(io_u, 0, sizeof(*io_u));
 876                 INIT_FLIST_HEAD(&io_u->list);
 877                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 878
 879                 if (!(td->io_ops->flags & FIO_NOIO)) {
 880                         io_u->buf = p + max_bs * i;
 881                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 882
 883                         if (td_write(td))
 884                                 io_u_fill_buffer(td, io_u, max_bs);
 885                         if (td_write(td) && td->o.verify_pattern_bytes) {
 886                                 /*
 887                                  * Fill the buffer with the pattern if we are
 888                                  * going to be doing writes.
 889                                  */
 890                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 891                         }
 892                 }
 893
 894                 io_u->index = i;
 895                 io_u->flags = IO_U_F_FREE;
 896                 flist_add(&io_u->list, &td->io_u_freelist);
 897         }
 898
 899         return 0;
 900 }
 901
 902 static int switch_ioscheduler(struct thread_data *td)
 903 {
 904         char tmp[256], tmp2[128];
 905         FILE *f;
 906         int ret;
 907
 908         if (td->io_ops->flags & FIO_DISKLESSIO)
 909                 return 0;
 910
 911         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 912
 913         f = fopen(tmp, "r+");
 914         if (!f) {
 915                 if (errno == ENOENT) {
 916                         log_err("fio: os or kernel doesn't support IO scheduler"
 917                                 " switching\n");
 918                         return 0;
 919                 }
 920                 td_verror(td, errno, "fopen iosched");
 921                 return 1;
 922         }
 923
 924         /*
 925          * Set io scheduler.
 926          */
 927         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 928         if (ferror(f) || ret != 1) {
 929                 td_verror(td, errno, "fwrite");
 930                 fclose(f);
 931                 return 1;
 932         }
 933
 934         rewind(f);
 935
 936         /*
 937          * Read back and check that the selected scheduler is now the default.
 938          */
 939         ret = fread(tmp, 1, sizeof(tmp), f);
 940         if (ferror(f) || ret < 0) {
 941                 td_verror(td, errno, "fread");
 942                 fclose(f);
 943                 return 1;
 944         }
 945
 946         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 947         if (!strstr(tmp, tmp2)) {
 948                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 949                 td_verror(td, EINVAL, "iosched_switch");
 950                 fclose(f);
 951                 return 1;
 952         }
 953
 954         fclose(f);
 955         return 0;
 956 }
 957
 958 static int keep_running(struct thread_data *td)
 959 {
 960         unsigned long long io_done;
 961
 962         if (td->done)
 963                 return 0;
 964         if (td->o.time_based)
 965                 return 1;
 966         if (td->o.loops) {
 967                 td->o.loops--;
 968                 return 1;
 969         }
 970
 971         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 972                         + td->io_skip_bytes;
 973         if (io_done < td->o.size)
 974                 return 1;
 975
 976         return 0;
 977 }
 978
 979 static void reset_io_counters(struct thread_data *td)
 980 {
 981         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 982         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 983         td->zone_bytes = 0;
 984         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 985         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 986
 987         td->last_was_sync = 0;
 988
 989         /*
 990          * reset file done count if we are to start over
 991          */
 992         if (td->o.time_based || td->o.loops)
 993                 td->nr_done_files = 0;
 994 }
 995
 996 void reset_all_stats(struct thread_data *td)
 997 {
 998         struct timeval tv;
 999         int i;
1000
1001         reset_io_counters(td);
1002
1003         for (i = 0; i < 2; i++) {
1004                 td->io_bytes[i] = 0;
1005                 td->io_blocks[i] = 0;
1006                 td->io_issues[i] = 0;
1007                 td->ts.total_io_u[i] = 0;
1008         }
1009
1010         fio_gettime(&tv, NULL);
1011         td->ts.runtime[0] = 0;
1012         td->ts.runtime[1] = 0;
1013         memcpy(&td->epoch, &tv, sizeof(tv));
1014         memcpy(&td->start, &tv, sizeof(tv));
1015 }
1016
1017 static void clear_io_state(struct thread_data *td)
1018 {
1019         struct fio_file *f;
1020         unsigned int i;
1021
1022         reset_io_counters(td);
1023
1024         close_files(td);
1025         for_each_file(td, f, i)
1026                 fio_file_clear_done(f);
1027
1028         /*
1029          * Set the same seed to get repeatable runs
1030          */
1031         td_fill_rand_seeds(td);
1032 }
1033
1034 static int exec_string(const char *string)
1035 {
1036         int ret, newlen = strlen(string) + 1 + 8;
1037         char *str;
1038
1039         str = malloc(newlen);
1040         sprintf(str, "sh -c %s", string);
1041
1042         ret = system(str);
1043         if (ret == -1)
1044                 log_err("fio: exec of cmd <%s> failed\n", str);
1045
1046         free(str);
1047         return ret;
1048 }
1049
1050 /*
1051  * Entry point for the thread based jobs. The process based jobs end up
1052  * here as well, after a little setup.
1053  */
1054 static void *thread_main(void *data)
1055 {
1056         unsigned long long elapsed;
1057         struct thread_data *td = data;
1058         pthread_condattr_t attr;
1059         int clear_state;
1060
1061         if (!td->o.use_thread) {
1062                 setsid();
1063                 td->pid = getpid();
1064         } else
1065                 td->pid = gettid();
1066
1067         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1068
1069         INIT_FLIST_HEAD(&td->io_u_freelist);
1070         INIT_FLIST_HEAD(&td->io_u_busylist);
1071         INIT_FLIST_HEAD(&td->io_u_requeues);
1072         INIT_FLIST_HEAD(&td->io_log_list);
1073         INIT_FLIST_HEAD(&td->io_hist_list);
1074         INIT_FLIST_HEAD(&td->verify_list);
1075         INIT_FLIST_HEAD(&td->trim_list);
1076         pthread_mutex_init(&td->io_u_lock, NULL);
1077         td->io_hist_tree = RB_ROOT;
1078
1079         pthread_condattr_init(&attr);
1080         pthread_cond_init(&td->verify_cond, &attr);
1081         pthread_cond_init(&td->free_cond, &attr);
1082
1083         td_set_runstate(td, TD_INITIALIZED);
1084         dprint(FD_MUTEX, "up startup_mutex\n");
1085         fio_mutex_up(startup_mutex);
1086         dprint(FD_MUTEX, "wait on td->mutex\n");
1087         fio_mutex_down(td->mutex);
1088         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1089
1090         /*
1091          * the ->mutex mutex is now no longer used, close it to avoid
1092          * eating a file descriptor
1093          */
1094         fio_mutex_remove(td->mutex);
1095
1096         /*
1097          * A new gid requires privilege, so we need to do this before setting
1098          * the uid.
1099          */
1100         if (td->o.gid != -1U && setgid(td->o.gid)) {
1101                 td_verror(td, errno, "setgid");
1102                 goto err;
1103         }
1104         if (td->o.uid != -1U && setuid(td->o.uid)) {
1105                 td_verror(td, errno, "setuid");
1106                 goto err;
1107         }
1108
1109         /*
1110          * If we have a gettimeofday() thread, make sure we exclude that
1111          * thread from this job
1112          */
1113         if (td->o.gtod_cpu)
1114                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1115
1116         /*
1117          * Set affinity first, in case it has an impact on the memory
1118          * allocations.
1119          */
1120         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1121                 td_verror(td, errno, "cpu_set_affinity");
1122                 goto err;
1123         }
1124
1125         /*
1126          * May alter parameters that init_io_u() will use, so we need to
1127          * do this first.
1128          */
1129         if (init_iolog(td))
1130                 goto err;
1131
1132         if (init_io_u(td))
1133                 goto err;
1134
1135         if (td->o.verify_async && verify_async_init(td))
1136                 goto err;
1137
1138         if (td->ioprio_set) {
1139                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1140                         td_verror(td, errno, "ioprio_set");
1141                         goto err;
1142                 }
1143         }
1144
1145         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1146                 goto err;
1147
1148         if (nice(td->o.nice) == -1) {
1149                 td_verror(td, errno, "nice");
1150                 goto err;
1151         }
1152
1153         if (td->o.ioscheduler && switch_ioscheduler(td))
1154                 goto err;
1155
1156         if (!td->o.create_serialize && setup_files(td))
1157                 goto err;
1158
1159         if (td_io_init(td))
1160                 goto err;
1161
1162         if (init_random_map(td))
1163                 goto err;
1164
1165         if (td->o.exec_prerun) {
1166                 if (exec_string(td->o.exec_prerun))
1167                         goto err;
1168         }
1169
1170         if (td->o.pre_read) {
1171                 if (pre_read_files(td) < 0)
1172                         goto err;
1173         }
1174
1175         fio_gettime(&td->epoch, NULL);
1176         getrusage(RUSAGE_SELF, &td->ru_start);
1177
1178         clear_state = 0;
1179         while (keep_running(td)) {
1180                 fio_gettime(&td->start, NULL);
1181                 memcpy(&td->stat_sample_time[0], &td->start, sizeof(td->start));
1182                 memcpy(&td->stat_sample_time[1], &td->start, sizeof(td->start));
1183                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1184
1185                 if (td->o.ratemin[0] || td->o.ratemin[1])
1186                         memcpy(&td->lastrate, &td->stat_sample_time,
1187                                                         sizeof(td->lastrate));
1188
1189                 if (clear_state)
1190                         clear_io_state(td);
1191
1192                 prune_io_piece_log(td);
1193
1194                 do_io(td);
1195
1196                 clear_state = 1;
1197
1198                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1199                         elapsed = utime_since_now(&td->start);
1200                         td->ts.runtime[DDIR_READ] += elapsed;
1201                 }
1202                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1203                         elapsed = utime_since_now(&td->start);
1204                         td->ts.runtime[DDIR_WRITE] += elapsed;
1205                 }
1206
1207                 if (td->error || td->terminate)
1208                         break;
1209
1210                 if (!td->o.do_verify ||
1211                     td->o.verify == VERIFY_NONE ||
1212                     (td->io_ops->flags & FIO_UNIDIR))
1213                         continue;
1214
1215                 clear_io_state(td);
1216
1217                 fio_gettime(&td->start, NULL);
1218
1219                 do_verify(td);
1220
1221                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1222
1223                 if (td->error || td->terminate)
1224                         break;
1225         }
1226
1227         update_rusage_stat(td);
1228         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1229         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1230         td->ts.total_run_time = mtime_since_now(&td->epoch);
1231         td->ts.io_bytes[0] = td->io_bytes[0];
1232         td->ts.io_bytes[1] = td->io_bytes[1];
1233
1234         fio_mutex_down(writeout_mutex);
1235         if (td->bw_log) {
1236                 if (td->o.bw_log_file) {
1237                         finish_log_named(td, td->bw_log,
1238                                                 td->o.bw_log_file, "bw");
1239                 } else
1240                         finish_log(td, td->bw_log, "bw");
1241         }
1242         if (td->lat_log) {
1243                 if (td->o.lat_log_file) {
1244                         finish_log_named(td, td->lat_log,
1245                                                 td->o.lat_log_file, "lat");
1246                 } else
1247                         finish_log(td, td->lat_log, "lat");
1248         }
1249         if (td->slat_log) {
1250                 if (td->o.lat_log_file) {
1251                         finish_log_named(td, td->slat_log,
1252                                                 td->o.lat_log_file, "slat");
1253                 } else
1254                         finish_log(td, td->slat_log, "slat");
1255         }
1256         if (td->clat_log) {
1257                 if (td->o.lat_log_file) {
1258                         finish_log_named(td, td->clat_log,
1259                                                 td->o.lat_log_file, "clat");
1260                 } else
1261                         finish_log(td, td->clat_log, "clat");
1262         }
1263         fio_mutex_up(writeout_mutex);
1264         if (td->o.exec_postrun)
1265                 exec_string(td->o.exec_postrun);
1266
1267         if (exitall_on_terminate)
1268                 terminate_threads(td->groupid);
1269
1270 err:
1271         if (td->error)
1272                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1273                                                         td->verror);
1274
1275         if (td->o.verify_async)
1276                 verify_async_exit(td);
1277
1278         close_and_free_files(td);
1279         close_ioengine(td);
1280         cleanup_io_u(td);
1281         cgroup_shutdown(td, &cgroup_mnt);
1282
1283         if (td->o.cpumask_set) {
1284                 int ret = fio_cpuset_exit(&td->o.cpumask);
1285
1286                 td_verror(td, ret, "fio_cpuset_exit");
1287         }
1288
1289         /*
1290          * do this very late, it will log file closing as well
1291          */
1292         if (td->o.write_iolog_file)
1293                 write_iolog_close(td);
1294
1295         options_mem_free(td);
1296         td_set_runstate(td, TD_EXITED);
1297         return (void *) (unsigned long) td->error;
1298 }
1299
1300 /*
1301  * We cannot pass the td data into a forked process, so attach the td and
1302  * pass it to the thread worker.
1303  */
1304 static int fork_main(int shmid, int offset)
1305 {
1306         struct thread_data *td;
1307         void *data, *ret;
1308
1309 #ifndef __hpux
1310         data = shmat(shmid, NULL, 0);
1311         if (data == (void *) -1) {
1312                 int __err = errno;
1313
1314                 perror("shmat");
1315                 return __err;
1316         }
1317 #else
1318         /*
1319          * HP-UX inherits shm mappings?
1320          */
1321         data = threads;
1322 #endif
1323
1324         td = data + offset * sizeof(struct thread_data);
1325         ret = thread_main(td);
1326         shmdt(data);
1327         return (int) (unsigned long) ret;
1328 }
1329
1330 /*
1331  * Run over the job map and reap the threads that have exited, if any.
1332  */
1333 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1334 {
1335         struct thread_data *td;
1336         int i, cputhreads, realthreads, pending, status, ret;
1337
1338         /*
1339          * reap exited threads (TD_EXITED -> TD_REAPED)
1340          */
1341         realthreads = pending = cputhreads = 0;
1342         for_each_td(td, i) {
1343                 int flags = 0;
1344
1345                 /*
1346                  * ->io_ops is NULL for a thread that has closed its
1347                  * io engine
1348                  */
1349                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1350                         cputhreads++;
1351                 else
1352                         realthreads++;
1353
1354                 if (!td->pid) {
1355                         pending++;
1356                         continue;
1357                 }
1358                 if (td->runstate == TD_REAPED)
1359                         continue;
1360                 if (td->o.use_thread) {
1361                         if (td->runstate == TD_EXITED) {
1362                                 td_set_runstate(td, TD_REAPED);
1363                                 goto reaped;
1364                         }
1365                         continue;
1366                 }
1367
1368                 flags = WNOHANG;
1369                 if (td->runstate == TD_EXITED)
1370                         flags = 0;
1371
1372                 /*
1373                  * check if someone quit or got killed in an unusual way
1374                  */
1375                 ret = waitpid(td->pid, &status, flags);
1376                 if (ret < 0) {
1377                         if (errno == ECHILD) {
1378                                 log_err("fio: pid=%d disappeared %d\n",
1379                                                 (int) td->pid, td->runstate);
1380                                 td_set_runstate(td, TD_REAPED);
1381                                 goto reaped;
1382                         }
1383                         perror("waitpid");
1384                 } else if (ret == td->pid) {
1385                         if (WIFSIGNALED(status)) {
1386                                 int sig = WTERMSIG(status);
1387
1388                                 if (sig != SIGTERM)
1389                                         log_err("fio: pid=%d, got signal=%d\n",
1390                                                         (int) td->pid, sig);
1391                                 td_set_runstate(td, TD_REAPED);
1392                                 goto reaped;
1393                         }
1394                         if (WIFEXITED(status)) {
1395                                 if (WEXITSTATUS(status) && !td->error)
1396                                         td->error = WEXITSTATUS(status);
1397
1398                                 td_set_runstate(td, TD_REAPED);
1399                                 goto reaped;
1400                         }
1401                 }
1402
1403                 /*
1404                  * thread is not dead, continue
1405                  */
1406                 pending++;
1407                 continue;
1408 reaped:
1409                 (*nr_running)--;
1410                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1411                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1412                 if (!td->pid)
1413                         pending--;
1414
1415                 if (td->error)
1416                         exit_value++;
1417
1418                 done_secs += mtime_since_now(&td->epoch) / 1000;
1419         }
1420
1421         if (*nr_running == cputhreads && !pending && realthreads)
1422                 terminate_threads(TERMINATE_ALL);
1423 }
1424
1425 static void *gtod_thread_main(void *data)
1426 {
1427         fio_mutex_up(startup_mutex);
1428
1429         /*
1430          * As long as we have jobs around, update the clock. It would be nice
1431          * to have some way of NOT hammering that CPU with gettimeofday(),
1432          * but I'm not sure what to use outside of a simple CPU nop to relax
1433          * it - we don't want to lose precision.
1434          */
1435         while (threads) {
1436                 fio_gtod_update();
1437                 nop;
1438         }
1439
1440         return NULL;
1441 }
1442
1443 static int fio_start_gtod_thread(void)
1444 {
1445         pthread_attr_t attr;
1446         int ret;
1447
1448         pthread_attr_init(&attr);
1449         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1450         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1451         pthread_attr_destroy(&attr);
1452         if (ret) {
1453                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1454                 return 1;
1455         }
1456
1457         ret = pthread_detach(gtod_thread);
1458         if (ret) {
1459                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1460                 return 1;
1461         }
1462
1463         dprint(FD_MUTEX, "wait on startup_mutex\n");
1464         fio_mutex_down(startup_mutex);
1465         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1466         return 0;
1467 }
1468
1469 /*
1470  * Main function for kicking off and reaping jobs, as needed.
1471  */
1472 static void run_threads(void)
1473 {
1474         struct thread_data *td;
1475         unsigned long spent;
1476         int i, todo, nr_running, m_rate, t_rate, nr_started;
1477
1478         if (fio_pin_memory())
1479                 return;
1480
1481         if (fio_gtod_offload && fio_start_gtod_thread())
1482                 return;
1483
1484         if (!terse_output) {
1485                 log_info("Starting ");
1486                 if (nr_thread)
1487                         log_info("%d thread%s", nr_thread,
1488                                                 nr_thread > 1 ? "s" : "");
1489                 if (nr_process) {
1490                         if (nr_thread)
1491                                 printf(" and ");
1492                         log_info("%d process%s", nr_process,
1493                                                 nr_process > 1 ? "es" : "");
1494                 }
1495                 log_info("\n");
1496                 fflush(stdout);
1497         }
1498
1499         set_sig_handlers();
1500
1501         todo = thread_number;
1502         nr_running = 0;
1503         nr_started = 0;
1504         m_rate = t_rate = 0;
1505
1506         for_each_td(td, i) {
1507                 print_status_init(td->thread_number - 1);
1508
1509                 if (!td->o.create_serialize)
1510                         continue;
1511
1512                 /*
1513                  * do file setup here so it happens sequentially,
1514                  * we don't want X number of threads getting their
1515                  * client data interspersed on disk
1516                  */
1517                 if (setup_files(td)) {
1518                         exit_value++;
1519                         if (td->error)
1520                                 log_err("fio: pid=%d, err=%d/%s\n",
1521                                         (int) td->pid, td->error, td->verror);
1522                         td_set_runstate(td, TD_REAPED);
1523                         todo--;
1524                 } else {
1525                         struct fio_file *f;
1526                         unsigned int j;
1527
1528                         /*
1529                          * for sharing to work, each job must always open
1530                          * its own files. so close them, if we opened them
1531                          * for creation
1532                          */
1533                         for_each_file(td, f, j) {
1534                                 if (fio_file_open(f))
1535                                         td_io_close_file(td, f);
1536                         }
1537                 }
1538         }
1539
1540         set_genesis_time();
1541
1542         while (todo) {
1543                 struct thread_data *map[REAL_MAX_JOBS];
1544                 struct timeval this_start;
1545                 int this_jobs = 0, left;
1546
1547                 /*
1548                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1549                  */
1550                 for_each_td(td, i) {
1551                         if (td->runstate != TD_NOT_CREATED)
1552                                 continue;
1553
1554                         /*
1555                          * never got a chance to start, killed by other
1556                          * thread for some reason
1557                          */
1558                         if (td->terminate) {
1559                                 todo--;
1560                                 continue;
1561                         }
1562
1563                         if (td->o.start_delay) {
1564                                 spent = mtime_since_genesis();
1565
1566                                 if (td->o.start_delay * 1000 > spent)
1567                                         continue;
1568                         }
1569
1570                         if (td->o.stonewall && (nr_started || nr_running)) {
1571                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1572                                                         td->o.name);
1573                                 break;
1574                         }
1575
1576                         init_disk_util(td);
1577
1578                         /*
1579                          * Set state to created. Thread will transition
1580                          * to TD_INITIALIZED when it's done setting up.
1581                          */
1582                         td_set_runstate(td, TD_CREATED);
1583                         map[this_jobs++] = td;
1584                         nr_started++;
1585
1586                         if (td->o.use_thread) {
1587                                 int ret;
1588
1589                                 dprint(FD_PROCESS, "will pthread_create\n");
1590                                 ret = pthread_create(&td->thread, NULL,
1591                                                         thread_main, td);
1592                                 if (ret) {
1593                                         log_err("pthread_create: %s\n",
1594                                                         strerror(ret));
1595                                         nr_started--;
1596                                         break;
1597                                 }
1598                                 ret = pthread_detach(td->thread);
1599                                 if (ret)
1600                                         log_err("pthread_detach: %s",
1601                                                         strerror(ret));
1602                         } else {
1603                                 pid_t pid;
1604                                 dprint(FD_PROCESS, "will fork\n");
1605                                 pid = fork();
1606                                 if (!pid) {
1607                                         int ret = fork_main(shm_id, i);
1608
1609                                         _exit(ret);
1610                                 } else if (i == fio_debug_jobno)
1611                                         *fio_debug_jobp = pid;
1612                         }
1613                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1614                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1615                                 log_err("fio: job startup hung? exiting.\n");
1616                                 terminate_threads(TERMINATE_ALL);
1617                                 fio_abort = 1;
1618                                 nr_started--;
1619                                 break;
1620                         }
1621                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1622                 }
1623
1624                 /*
1625                  * Wait for the started threads to transition to
1626                  * TD_INITIALIZED.
1627                  */
1628                 fio_gettime(&this_start, NULL);
1629                 left = this_jobs;
1630                 while (left && !fio_abort) {
1631                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1632                                 break;
1633
1634                         usleep(100000);
1635
1636                         for (i = 0; i < this_jobs; i++) {
1637                                 td = map[i];
1638                                 if (!td)
1639                                         continue;
1640                                 if (td->runstate == TD_INITIALIZED) {
1641                                         map[i] = NULL;
1642                                         left--;
1643                                 } else if (td->runstate >= TD_EXITED) {
1644                                         map[i] = NULL;
1645                                         left--;
1646                                         todo--;
1647                                         nr_running++; /* work-around... */
1648                                 }
1649                         }
1650                 }
1651
1652                 if (left) {
1653                         log_err("fio: %d jobs failed to start\n", left);
1654                         for (i = 0; i < this_jobs; i++) {
1655                                 td = map[i];
1656                                 if (!td)
1657                                         continue;
1658                                 kill(td->pid, SIGTERM);
1659                         }
1660                         break;
1661                 }
1662
1663                 /*
1664                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1665                  */
1666                 for_each_td(td, i) {
1667                         if (td->runstate != TD_INITIALIZED)
1668                                 continue;
1669
1670                         if (in_ramp_time(td))
1671                                 td_set_runstate(td, TD_RAMP);
1672                         else
1673                                 td_set_runstate(td, TD_RUNNING);
1674                         nr_running++;
1675                         nr_started--;
1676                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1677                         t_rate += td->o.rate[0] + td->o.rate[1];
1678                         todo--;
1679                         fio_mutex_up(td->mutex);
1680                 }
1681
1682                 reap_threads(&nr_running, &t_rate, &m_rate);
1683
1684                 if (todo)
1685                         usleep(100000);
1686         }
1687
1688         while (nr_running) {
1689                 reap_threads(&nr_running, &t_rate, &m_rate);
1690                 usleep(10000);
1691         }
1692
1693         update_io_ticks();
1694         fio_unpin_memory();
1695 }
1696
1697 int exec_run(void)
1698 {
1699         if (nr_clients)
1700                 return fio_handle_clients();
1701         if (exec_profile && load_profile(exec_profile))
1702                 return 1;
1703
1704         if (!thread_number)
1705                 return 0;
1706
1707         if (write_bw_log) {
1708                 setup_log(&agg_io_log[DDIR_READ]);
1709                 setup_log(&agg_io_log[DDIR_WRITE]);
1710         }
1711
1712         startup_mutex = fio_mutex_init(0);
1713         if (startup_mutex == NULL)
1714                 return 1;
1715         writeout_mutex = fio_mutex_init(1);
1716         if (writeout_mutex == NULL)
1717                 return 1;
1718
1719         set_genesis_time();
1720         create_disk_util_thread();
1721
1722         cgroup_list = smalloc(sizeof(*cgroup_list));
1723         INIT_FLIST_HEAD(cgroup_list);
1724
1725         run_threads();
1726
1727         if (!fio_abort) {
1728                 show_run_stats();
1729                 if (write_bw_log) {
1730                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1731                         __finish_log(agg_io_log[DDIR_WRITE],
1732                                         "agg-write_bw.log");
1733                 }
1734         }
1735
1736         cgroup_kill(cgroup_list);
1737         sfree(cgroup_list);
1738         sfree(cgroup_mnt);
1739
1740         fio_mutex_remove(startup_mutex);
1741         fio_mutex_remove(writeout_mutex);
1742         return exit_value;
1743 }
1744
1745 void reset_fio_state(void)
1746 {
1747         groupid = 0;
1748         thread_number = 0;
1749         nr_process = 0;
1750         nr_thread = 0;
1751         done_secs = 0;
1752 }
1753
1754 int main(int argc, char *argv[], char *envp[])
1755 {
1756         long ps;
1757
1758         arch_init(envp);
1759
1760         sinit();
1761
1762         /*
1763          * We need locale for number printing, if it isn't set then just
1764          * go with the US format.
1765          */
1766         if (!getenv("LC_NUMERIC"))
1767                 setlocale(LC_NUMERIC, "en_US");
1768
1769         ps = sysconf(_SC_PAGESIZE);
1770         if (ps < 0) {
1771                 log_err("Failed to get page size\n");
1772                 return 1;
1773         }
1774
1775         page_size = ps;
1776         page_mask = ps - 1;
1777
1778         fio_keywords_init();
1779
1780         if (parse_options(argc, argv))
1781                 return 1;
1782
1783         return exec_run();
1784 }