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 JOB_START_TIMEOUT       (5 * 1000)
  79
  80 void td_set_runstate(struct thread_data *td, int runstate)
  81 {
  82         if (td->runstate == runstate)
  83                 return;
  84
  85         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  86                                                 td->runstate, runstate);
  87         td->runstate = runstate;
  88 }
  89
  90 void fio_terminate_threads(int group_id)
  91 {
  92         struct thread_data *td;
  93         int i;
  94
  95         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  96
  97         for_each_td(td, i) {
  98                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  99                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
 100                                                 td->o.name, (int) td->pid);
 101                         td->terminate = 1;
 102                         td->o.start_delay = 0;
 103
 104                         /*
 105                          * if the thread is running, just let it exit
 106                          */
 107                         if (!td->pid)
 108                                 continue;
 109                         else if (td->runstate < TD_RAMP)
 110                                 kill(td->pid, SIGTERM);
 111                         else {
 112                                 struct ioengine_ops *ops = td->io_ops;
 113
 114                                 if (ops && (ops->flags & FIO_SIGTERM))
 115                                         kill(td->pid, SIGTERM);
 116                         }
 117                 }
 118         }
 119 }
 120
 121 static void sig_int(int sig)
 122 {
 123         if (threads) {
 124                 log_info("\nfio: terminating on signal %d\n", sig);
 125                 exit_backend = 1;
 126                 fflush(stdout);
 127                 exit_value = 128;
 128                 fio_terminate_threads(TERMINATE_ALL);
 129         }
 130 }
 131
 132 static void *disk_thread_main(void *data)
 133 {
 134         fio_mutex_up(startup_mutex);
 135
 136         while (threads) {
 137                 usleep(DISK_UTIL_MSEC * 1000);
 138                 if (!threads)
 139                         break;
 140                 update_io_ticks();
 141
 142                 if (is_backend)
 143                         fio_server_send_status();
 144                 else
 145                         print_thread_status();
 146         }
 147
 148         return NULL;
 149 }
 150
 151 static int create_disk_util_thread(void)
 152 {
 153         int ret;
 154
 155         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
 156         if (ret) {
 157                 log_err("Can't create disk util thread: %s\n", strerror(ret));
 158                 return 1;
 159         }
 160
 161         ret = pthread_detach(disk_util_thread);
 162         if (ret) {
 163                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
 164                 return 1;
 165         }
 166
 167         dprint(FD_MUTEX, "wait on startup_mutex\n");
 168         fio_mutex_down(startup_mutex);
 169         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 170         return 0;
 171 }
 172
 173 static void set_sig_handlers(void)
 174 {
 175         struct sigaction act;
 176
 177         memset(&act, 0, sizeof(act));
 178         act.sa_handler = sig_int;
 179         act.sa_flags = SA_RESTART;
 180         sigaction(SIGINT, &act, NULL);
 181
 182         memset(&act, 0, sizeof(act));
 183         act.sa_handler = sig_int;
 184         act.sa_flags = SA_RESTART;
 185         sigaction(SIGTERM, &act, NULL);
 186
 187         if (is_backend) {
 188                 memset(&act, 0, sizeof(act));
 189                 act.sa_handler = sig_int;
 190                 act.sa_flags = SA_RESTART;
 191                 sigaction(SIGPIPE, &act, NULL);
 192         }
 193 }
 194
 195 /*
 196  * Check if we are above the minimum rate given.
 197  */
 198 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 199                             enum fio_ddir ddir)
 200 {
 201         unsigned long long bytes = 0;
 202         unsigned long iops = 0;
 203         unsigned long spent;
 204         unsigned long rate;
 205         unsigned int ratemin = 0;
 206         unsigned int rate_iops = 0;
 207         unsigned int rate_iops_min = 0;
 208
 209         assert(ddir_rw(ddir));
 210
 211         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 212                 return 0;
 213
 214         /*
 215          * allow a 2 second settle period in the beginning
 216          */
 217         if (mtime_since(&td->start, now) < 2000)
 218                 return 0;
 219
 220         iops += td->io_blocks[ddir];
 221         bytes += td->this_io_bytes[ddir];
 222         ratemin += td->o.ratemin[ddir];
 223         rate_iops += td->o.rate_iops[ddir];
 224         rate_iops_min += td->o.rate_iops_min[ddir];
 225
 226         /*
 227          * if rate blocks is set, sample is running
 228          */
 229         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 230                 spent = mtime_since(&td->lastrate[ddir], now);
 231                 if (spent < td->o.ratecycle)
 232                         return 0;
 233
 234                 if (td->o.rate[ddir]) {
 235                         /*
 236                          * check bandwidth specified rate
 237                          */
 238                         if (bytes < td->rate_bytes[ddir]) {
 239                                 log_err("%s: min rate %u not met\n", td->o.name,
 240                                                                 ratemin);
 241                                 return 1;
 242                         } else {
 243                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 244                                 if (rate < ratemin ||
 245                                     bytes < td->rate_bytes[ddir]) {
 246                                         log_err("%s: min rate %u not met, got"
 247                                                 " %luKB/sec\n", td->o.name,
 248                                                         ratemin, rate);
 249                                         return 1;
 250                                 }
 251                         }
 252                 } else {
 253                         /*
 254                          * checks iops specified rate
 255                          */
 256                         if (iops < rate_iops) {
 257                                 log_err("%s: min iops rate %u not met\n",
 258                                                 td->o.name, rate_iops);
 259                                 return 1;
 260                         } else {
 261                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 262                                 if (rate < rate_iops_min ||
 263                                     iops < td->rate_blocks[ddir]) {
 264                                         log_err("%s: min iops rate %u not met,"
 265                                                 " got %lu\n", td->o.name,
 266                                                         rate_iops_min, rate);
 267                                 }
 268                         }
 269                 }
 270         }
 271
 272         td->rate_bytes[ddir] = bytes;
 273         td->rate_blocks[ddir] = iops;
 274         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 275         return 0;
 276 }
 277
 278 static int check_min_rate(struct thread_data *td, struct timeval *now,
 279                           unsigned long *bytes_done)
 280 {
 281         int ret = 0;
 282
 283         if (bytes_done[0])
 284                 ret |= __check_min_rate(td, now, 0);
 285         if (bytes_done[1])
 286                 ret |= __check_min_rate(td, now, 1);
 287
 288         return ret;
 289 }
 290
 291 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 292 {
 293         if (!td->o.timeout)
 294                 return 0;
 295         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 296                 return 1;
 297
 298         return 0;
 299 }
 300
 301 /*
 302  * When job exits, we can cancel the in-flight IO if we are using async
 303  * io. Attempt to do so.
 304  */
 305 static void cleanup_pending_aio(struct thread_data *td)
 306 {
 307         struct flist_head *entry, *n;
 308         struct io_u *io_u;
 309         int r;
 310
 311         /*
 312          * get immediately available events, if any
 313          */
 314         r = io_u_queued_complete(td, 0, NULL);
 315         if (r < 0)
 316                 return;
 317
 318         /*
 319          * now cancel remaining active events
 320          */
 321         if (td->io_ops->cancel) {
 322                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 323                         io_u = flist_entry(entry, struct io_u, list);
 324
 325                         /*
 326                          * if the io_u isn't in flight, then that generally
 327                          * means someone leaked an io_u. complain but fix
 328                          * it up, so we don't stall here.
 329                          */
 330                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 331                                 log_err("fio: non-busy IO on busy list\n");
 332                                 put_io_u(td, io_u);
 333                         } else {
 334                                 r = td->io_ops->cancel(td, io_u);
 335                                 if (!r)
 336                                         put_io_u(td, io_u);
 337                         }
 338                 }
 339         }
 340
 341         if (td->cur_depth)
 342                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 343 }
 344
 345 /*
 346  * Helper to handle the final sync of a file. Works just like the normal
 347  * io path, just does everything sync.
 348  */
 349 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 350 {
 351         struct io_u *io_u = __get_io_u(td);
 352         int ret;
 353
 354         if (!io_u)
 355                 return 1;
 356
 357         io_u->ddir = DDIR_SYNC;
 358         io_u->file = f;
 359
 360         if (td_io_prep(td, io_u)) {
 361                 put_io_u(td, io_u);
 362                 return 1;
 363         }
 364
 365 requeue:
 366         ret = td_io_queue(td, io_u);
 367         if (ret < 0) {
 368                 td_verror(td, io_u->error, "td_io_queue");
 369                 put_io_u(td, io_u);
 370                 return 1;
 371         } else if (ret == FIO_Q_QUEUED) {
 372                 if (io_u_queued_complete(td, 1, NULL) < 0)
 373                         return 1;
 374         } else if (ret == FIO_Q_COMPLETED) {
 375                 if (io_u->error) {
 376                         td_verror(td, io_u->error, "td_io_queue");
 377                         return 1;
 378                 }
 379
 380                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 381                         return 1;
 382         } else if (ret == FIO_Q_BUSY) {
 383                 if (td_io_commit(td))
 384                         return 1;
 385                 goto requeue;
 386         }
 387
 388         return 0;
 389 }
 390
 391 static inline void __update_tv_cache(struct thread_data *td)
 392 {
 393         fio_gettime(&td->tv_cache, NULL);
 394 }
 395
 396 static inline void update_tv_cache(struct thread_data *td)
 397 {
 398         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 399                 __update_tv_cache(td);
 400 }
 401
 402 static int break_on_this_error(struct thread_data *td, int *retptr)
 403 {
 404         int ret = *retptr;
 405
 406         if (ret < 0 || td->error) {
 407                 int err;
 408
 409                 if (!td->o.continue_on_error)
 410                         return 1;
 411
 412                 if (ret < 0)
 413                         err = -ret;
 414                 else
 415                         err = td->error;
 416
 417                 if (td_non_fatal_error(err)) {
 418                         /*
 419                          * Continue with the I/Os in case of
 420                          * a non fatal error.
 421                          */
 422                         update_error_count(td, err);
 423                         td_clear_error(td);
 424                         *retptr = 0;
 425                         return 0;
 426                 } else if (td->o.fill_device && err == ENOSPC) {
 427                         /*
 428                          * We expect to hit this error if
 429                          * fill_device option is set.
 430                          */
 431                         td_clear_error(td);
 432                         td->terminate = 1;
 433                         return 1;
 434                 } else {
 435                         /*
 436                          * Stop the I/O in case of a fatal
 437                          * error.
 438                          */
 439                         update_error_count(td, err);
 440                         return 1;
 441                 }
 442         }
 443
 444         return 0;
 445 }
 446
 447 /*
 448  * The main verify engine. Runs over the writes we previously submitted,
 449  * reads the blocks back in, and checks the crc/md5 of the data.
 450  */
 451 static void do_verify(struct thread_data *td)
 452 {
 453         struct fio_file *f;
 454         struct io_u *io_u;
 455         int ret, min_events;
 456         unsigned int i;
 457
 458         dprint(FD_VERIFY, "starting loop\n");
 459
 460         /*
 461          * sync io first and invalidate cache, to make sure we really
 462          * read from disk.
 463          */
 464         for_each_file(td, f, i) {
 465                 if (!fio_file_open(f))
 466                         continue;
 467                 if (fio_io_sync(td, f))
 468                         break;
 469                 if (file_invalidate_cache(td, f))
 470                         break;
 471         }
 472
 473         if (td->error)
 474                 return;
 475
 476         td_set_runstate(td, TD_VERIFYING);
 477
 478         io_u = NULL;
 479         while (!td->terminate) {
 480                 int ret2, full;
 481
 482                 update_tv_cache(td);
 483
 484                 if (runtime_exceeded(td, &td->tv_cache)) {
 485                         __update_tv_cache(td);
 486                         if (runtime_exceeded(td, &td->tv_cache)) {
 487                                 td->terminate = 1;
 488                                 break;
 489                         }
 490                 }
 491
 492                 io_u = __get_io_u(td);
 493                 if (!io_u)
 494                         break;
 495
 496                 if (get_next_verify(td, io_u)) {
 497                         put_io_u(td, io_u);
 498                         break;
 499                 }
 500
 501                 if (td_io_prep(td, io_u)) {
 502                         put_io_u(td, io_u);
 503                         break;
 504                 }
 505
 506                 if (td->o.verify_async)
 507                         io_u->end_io = verify_io_u_async;
 508                 else
 509                         io_u->end_io = verify_io_u;
 510
 511                 ret = td_io_queue(td, io_u);
 512                 switch (ret) {
 513                 case FIO_Q_COMPLETED:
 514                         if (io_u->error) {
 515                                 ret = -io_u->error;
 516                                 clear_io_u(td, io_u);
 517                         } else if (io_u->resid) {
 518                                 int bytes = io_u->xfer_buflen - io_u->resid;
 519
 520                                 /*
 521                                  * zero read, fail
 522                                  */
 523                                 if (!bytes) {
 524                                         td_verror(td, EIO, "full resid");
 525                                         put_io_u(td, io_u);
 526                                         break;
 527                                 }
 528
 529                                 io_u->xfer_buflen = io_u->resid;
 530                                 io_u->xfer_buf += bytes;
 531                                 io_u->offset += bytes;
 532
 533                                 if (ddir_rw(io_u->ddir))
 534                                         td->ts.short_io_u[io_u->ddir]++;
 535
 536                                 f = io_u->file;
 537                                 if (io_u->offset == f->real_file_size)
 538                                         goto sync_done;
 539
 540                                 requeue_io_u(td, &io_u);
 541                         } else {
 542 sync_done:
 543                                 ret = io_u_sync_complete(td, io_u, NULL);
 544                                 if (ret < 0)
 545                                         break;
 546                         }
 547                         continue;
 548                 case FIO_Q_QUEUED:
 549                         break;
 550                 case FIO_Q_BUSY:
 551                         requeue_io_u(td, &io_u);
 552                         ret2 = td_io_commit(td);
 553                         if (ret2 < 0)
 554                                 ret = ret2;
 555                         break;
 556                 default:
 557                         assert(ret < 0);
 558                         td_verror(td, -ret, "td_io_queue");
 559                         break;
 560                 }
 561
 562                 if (break_on_this_error(td, &ret))
 563                         break;
 564
 565                 /*
 566                  * if we can queue more, do so. but check if there are
 567                  * completed io_u's first. Note that we can get BUSY even
 568                  * without IO queued, if the system is resource starved.
 569                  */
 570                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 571                 if (full || !td->o.iodepth_batch_complete) {
 572                         min_events = min(td->o.iodepth_batch_complete,
 573                                          td->cur_depth);
 574                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 575                                 min_events = 1;
 576
 577                         do {
 578                                 /*
 579                                  * Reap required number of io units, if any,
 580                                  * and do the verification on them through
 581                                  * the callback handler
 582                                  */
 583                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 584                                         ret = -1;
 585                                         break;
 586                                 }
 587                         } while (full && (td->cur_depth > td->o.iodepth_low));
 588                 }
 589                 if (ret < 0)
 590                         break;
 591         }
 592
 593         if (!td->error) {
 594                 min_events = td->cur_depth;
 595
 596                 if (min_events)
 597                         ret = io_u_queued_complete(td, min_events, NULL);
 598         } else
 599                 cleanup_pending_aio(td);
 600
 601         td_set_runstate(td, TD_RUNNING);
 602
 603         dprint(FD_VERIFY, "exiting loop\n");
 604 }
 605
 606 /*
 607  * Main IO worker function. It retrieves io_u's to process and queues
 608  * and reaps them, checking for rate and errors along the way.
 609  */
 610 static void do_io(struct thread_data *td)
 611 {
 612         unsigned int i;
 613         int ret = 0;
 614
 615         if (in_ramp_time(td))
 616                 td_set_runstate(td, TD_RAMP);
 617         else
 618                 td_set_runstate(td, TD_RUNNING);
 619
 620         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 621                 (!flist_empty(&td->trim_list)) ||
 622                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 623                 struct timeval comp_time;
 624                 unsigned long bytes_done[2] = { 0, 0 };
 625                 int min_evts = 0;
 626                 struct io_u *io_u;
 627                 int ret2, full;
 628
 629                 if (td->terminate)
 630                         break;
 631
 632                 update_tv_cache(td);
 633
 634                 if (runtime_exceeded(td, &td->tv_cache)) {
 635                         __update_tv_cache(td);
 636                         if (runtime_exceeded(td, &td->tv_cache)) {
 637                                 td->terminate = 1;
 638                                 break;
 639                         }
 640                 }
 641
 642                 io_u = get_io_u(td);
 643                 if (!io_u)
 644                         break;
 645
 646                 /*
 647                  * Add verification end_io handler, if asked to verify
 648                  * a previously written file.
 649                  */
 650                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 651                     !td_rw(td)) {
 652                         if (td->o.verify_async)
 653                                 io_u->end_io = verify_io_u_async;
 654                         else
 655                                 io_u->end_io = verify_io_u;
 656                         td_set_runstate(td, TD_VERIFYING);
 657                 } else if (in_ramp_time(td))
 658                         td_set_runstate(td, TD_RAMP);
 659                 else
 660                         td_set_runstate(td, TD_RUNNING);
 661
 662                 ret = td_io_queue(td, io_u);
 663                 switch (ret) {
 664                 case FIO_Q_COMPLETED:
 665                         if (io_u->error) {
 666                                 ret = -io_u->error;
 667                                 clear_io_u(td, io_u);
 668                         } else if (io_u->resid) {
 669                                 int bytes = io_u->xfer_buflen - io_u->resid;
 670                                 struct fio_file *f = io_u->file;
 671
 672                                 /*
 673                                  * zero read, fail
 674                                  */
 675                                 if (!bytes) {
 676                                         td_verror(td, EIO, "full resid");
 677                                         put_io_u(td, io_u);
 678                                         break;
 679                                 }
 680
 681                                 io_u->xfer_buflen = io_u->resid;
 682                                 io_u->xfer_buf += bytes;
 683                                 io_u->offset += bytes;
 684
 685                                 if (ddir_rw(io_u->ddir))
 686                                         td->ts.short_io_u[io_u->ddir]++;
 687
 688                                 if (io_u->offset == f->real_file_size)
 689                                         goto sync_done;
 690
 691                                 requeue_io_u(td, &io_u);
 692                         } else {
 693 sync_done:
 694                                 if (__should_check_rate(td, 0) ||
 695                                     __should_check_rate(td, 1))
 696                                         fio_gettime(&comp_time, NULL);
 697
 698                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 699                                 if (ret < 0)
 700                                         break;
 701                         }
 702                         break;
 703                 case FIO_Q_QUEUED:
 704                         /*
 705                          * if the engine doesn't have a commit hook,
 706                          * the io_u is really queued. if it does have such
 707                          * a hook, it has to call io_u_queued() itself.
 708                          */
 709                         if (td->io_ops->commit == NULL)
 710                                 io_u_queued(td, io_u);
 711                         break;
 712                 case FIO_Q_BUSY:
 713                         requeue_io_u(td, &io_u);
 714                         ret2 = td_io_commit(td);
 715                         if (ret2 < 0)
 716                                 ret = ret2;
 717                         break;
 718                 default:
 719                         assert(ret < 0);
 720                         put_io_u(td, io_u);
 721                         break;
 722                 }
 723
 724                 if (break_on_this_error(td, &ret))
 725                         break;
 726
 727                 /*
 728                  * See if we need to complete some commands. Note that we
 729                  * can get BUSY even without IO queued, if the system is
 730                  * resource starved.
 731                  */
 732                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 733                 if (full || !td->o.iodepth_batch_complete) {
 734                         min_evts = min(td->o.iodepth_batch_complete,
 735                                         td->cur_depth);
 736                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 737                                 min_evts = 1;
 738
 739                         if (__should_check_rate(td, 0) ||
 740                             __should_check_rate(td, 1))
 741                                 fio_gettime(&comp_time, NULL);
 742
 743                         do {
 744                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 745                                 if (ret < 0)
 746                                         break;
 747
 748                         } while (full && (td->cur_depth > td->o.iodepth_low));
 749                 }
 750
 751                 if (ret < 0)
 752                         break;
 753                 if (!(bytes_done[0] + bytes_done[1]))
 754                         continue;
 755
 756                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 757                         if (check_min_rate(td, &comp_time, bytes_done)) {
 758                                 if (exitall_on_terminate)
 759                                         fio_terminate_threads(td->groupid);
 760                                 td_verror(td, EIO, "check_min_rate");
 761                                 break;
 762                         }
 763                 }
 764
 765                 if (td->o.thinktime) {
 766                         unsigned long long b;
 767
 768                         b = td->io_blocks[0] + td->io_blocks[1];
 769                         if (!(b % td->o.thinktime_blocks)) {
 770                                 int left;
 771
 772                                 if (td->o.thinktime_spin)
 773                                         usec_spin(td->o.thinktime_spin);
 774
 775                                 left = td->o.thinktime - td->o.thinktime_spin;
 776                                 if (left)
 777                                         usec_sleep(td, left);
 778                         }
 779                 }
 780         }
 781
 782         if (td->trim_entries)
 783                 printf("trim entries %ld\n", td->trim_entries);
 784
 785         if (td->o.fill_device && td->error == ENOSPC) {
 786                 td->error = 0;
 787                 td->terminate = 1;
 788         }
 789         if (!td->error) {
 790                 struct fio_file *f;
 791
 792                 i = td->cur_depth;
 793                 if (i) {
 794                         ret = io_u_queued_complete(td, i, NULL);
 795                         if (td->o.fill_device && td->error == ENOSPC)
 796                                 td->error = 0;
 797                 }
 798
 799                 if (should_fsync(td) && td->o.end_fsync) {
 800                         td_set_runstate(td, TD_FSYNCING);
 801
 802                         for_each_file(td, f, i) {
 803                                 if (!fio_file_open(f))
 804                                         continue;
 805                                 fio_io_sync(td, f);
 806                         }
 807                 }
 808         } else
 809                 cleanup_pending_aio(td);
 810
 811         /*
 812          * stop job if we failed doing any IO
 813          */
 814         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 815                 td->done = 1;
 816 }
 817
 818 static void cleanup_io_u(struct thread_data *td)
 819 {
 820         struct flist_head *entry, *n;
 821         struct io_u *io_u;
 822
 823         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 824                 io_u = flist_entry(entry, struct io_u, list);
 825
 826                 flist_del(&io_u->list);
 827                 fio_memfree(io_u, sizeof(*io_u));
 828         }
 829
 830         free_io_mem(td);
 831 }
 832
 833 static int init_io_u(struct thread_data *td)
 834 {
 835         struct io_u *io_u;
 836         unsigned int max_bs;
 837         int cl_align, i, max_units;
 838         char *p;
 839
 840         max_units = td->o.iodepth;
 841         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 842         td->orig_buffer_size = (unsigned long long) max_bs
 843                                         * (unsigned long long) max_units;
 844
 845         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 846                 unsigned long bs;
 847
 848                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 849                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 850         }
 851
 852         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 853                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 854                 return 1;
 855         }
 856
 857         if (allocate_io_mem(td))
 858                 return 1;
 859
 860         if (td->o.odirect || td->o.mem_align ||
 861             (td->io_ops->flags & FIO_RAWIO))
 862                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 863         else
 864                 p = td->orig_buffer;
 865
 866         cl_align = os_cache_line_size();
 867
 868         for (i = 0; i < max_units; i++) {
 869                 void *ptr;
 870
 871                 if (td->terminate)
 872                         return 1;
 873
 874                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 875                 if (!ptr) {
 876                         log_err("fio: unable to allocate aligned memory\n");
 877                         break;
 878                 }
 879
 880                 io_u = ptr;
 881                 memset(io_u, 0, sizeof(*io_u));
 882                 INIT_FLIST_HEAD(&io_u->list);
 883                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 884
 885                 if (!(td->io_ops->flags & FIO_NOIO)) {
 886                         io_u->buf = p + max_bs * i;
 887                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 888
 889                         if (td_write(td))
 890                                 io_u_fill_buffer(td, io_u, max_bs);
 891                         if (td_write(td) && td->o.verify_pattern_bytes) {
 892                                 /*
 893                                  * Fill the buffer with the pattern if we are
 894                                  * going to be doing writes.
 895                                  */
 896                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 897                         }
 898                 }
 899
 900                 io_u->index = i;
 901                 io_u->flags = IO_U_F_FREE;
 902                 flist_add(&io_u->list, &td->io_u_freelist);
 903         }
 904
 905         return 0;
 906 }
 907
 908 static int switch_ioscheduler(struct thread_data *td)
 909 {
 910         char tmp[256], tmp2[128];
 911         FILE *f;
 912         int ret;
 913
 914         if (td->io_ops->flags & FIO_DISKLESSIO)
 915                 return 0;
 916
 917         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 918
 919         f = fopen(tmp, "r+");
 920         if (!f) {
 921                 if (errno == ENOENT) {
 922                         log_err("fio: os or kernel doesn't support IO scheduler"
 923                                 " switching\n");
 924                         return 0;
 925                 }
 926                 td_verror(td, errno, "fopen iosched");
 927                 return 1;
 928         }
 929
 930         /*
 931          * Set io scheduler.
 932          */
 933         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 934         if (ferror(f) || ret != 1) {
 935                 td_verror(td, errno, "fwrite");
 936                 fclose(f);
 937                 return 1;
 938         }
 939
 940         rewind(f);
 941
 942         /*
 943          * Read back and check that the selected scheduler is now the default.
 944          */
 945         ret = fread(tmp, 1, sizeof(tmp), f);
 946         if (ferror(f) || ret < 0) {
 947                 td_verror(td, errno, "fread");
 948                 fclose(f);
 949                 return 1;
 950         }
 951
 952         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 953         if (!strstr(tmp, tmp2)) {
 954                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 955                 td_verror(td, EINVAL, "iosched_switch");
 956                 fclose(f);
 957                 return 1;
 958         }
 959
 960         fclose(f);
 961         return 0;
 962 }
 963
 964 static int keep_running(struct thread_data *td)
 965 {
 966         unsigned long long io_done;
 967
 968         if (td->done)
 969                 return 0;
 970         if (td->o.time_based)
 971                 return 1;
 972         if (td->o.loops) {
 973                 td->o.loops--;
 974                 return 1;
 975         }
 976
 977         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 978                         + td->io_skip_bytes;
 979         if (io_done < td->o.size)
 980                 return 1;
 981
 982         return 0;
 983 }
 984
 985 static void reset_io_counters(struct thread_data *td)
 986 {
 987         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 988         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 989         td->zone_bytes = 0;
 990         td->rate_bytes[0] = td->rate_bytes[1] = 0;
 991         td->rate_blocks[0] = td->rate_blocks[1] = 0;
 992
 993         td->last_was_sync = 0;
 994
 995         /*
 996          * reset file done count if we are to start over
 997          */
 998         if (td->o.time_based || td->o.loops)
 999                 td->nr_done_files = 0;
1000 }
1001
1002 void reset_all_stats(struct thread_data *td)
1003 {
1004         struct timeval tv;
1005         int i;
1006
1007         reset_io_counters(td);
1008
1009         for (i = 0; i < 2; i++) {
1010                 td->io_bytes[i] = 0;
1011                 td->io_blocks[i] = 0;
1012                 td->io_issues[i] = 0;
1013                 td->ts.total_io_u[i] = 0;
1014         }
1015
1016         fio_gettime(&tv, NULL);
1017         td->ts.runtime[0] = 0;
1018         td->ts.runtime[1] = 0;
1019         memcpy(&td->epoch, &tv, sizeof(tv));
1020         memcpy(&td->start, &tv, sizeof(tv));
1021 }
1022
1023 static void clear_io_state(struct thread_data *td)
1024 {
1025         struct fio_file *f;
1026         unsigned int i;
1027
1028         reset_io_counters(td);
1029
1030         close_files(td);
1031         for_each_file(td, f, i)
1032                 fio_file_clear_done(f);
1033
1034         /*
1035          * Set the same seed to get repeatable runs
1036          */
1037         td_fill_rand_seeds(td);
1038 }
1039
1040 static int exec_string(const char *string)
1041 {
1042         int ret, newlen = strlen(string) + 1 + 8;
1043         char *str;
1044
1045         str = malloc(newlen);
1046         sprintf(str, "sh -c %s", string);
1047
1048         ret = system(str);
1049         if (ret == -1)
1050                 log_err("fio: exec of cmd <%s> failed\n", str);
1051
1052         free(str);
1053         return ret;
1054 }
1055
1056 /*
1057  * Entry point for the thread based jobs. The process based jobs end up
1058  * here as well, after a little setup.
1059  */
1060 static void *thread_main(void *data)
1061 {
1062         unsigned long long elapsed;
1063         struct thread_data *td = data;
1064         pthread_condattr_t attr;
1065         int clear_state;
1066
1067         if (!td->o.use_thread) {
1068                 setsid();
1069                 td->pid = getpid();
1070         } else
1071                 td->pid = gettid();
1072
1073         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1074
1075         INIT_FLIST_HEAD(&td->io_u_freelist);
1076         INIT_FLIST_HEAD(&td->io_u_busylist);
1077         INIT_FLIST_HEAD(&td->io_u_requeues);
1078         INIT_FLIST_HEAD(&td->io_log_list);
1079         INIT_FLIST_HEAD(&td->io_hist_list);
1080         INIT_FLIST_HEAD(&td->verify_list);
1081         INIT_FLIST_HEAD(&td->trim_list);
1082         pthread_mutex_init(&td->io_u_lock, NULL);
1083         td->io_hist_tree = RB_ROOT;
1084
1085         pthread_condattr_init(&attr);
1086         pthread_cond_init(&td->verify_cond, &attr);
1087         pthread_cond_init(&td->free_cond, &attr);
1088
1089         td_set_runstate(td, TD_INITIALIZED);
1090         dprint(FD_MUTEX, "up startup_mutex\n");
1091         fio_mutex_up(startup_mutex);
1092         dprint(FD_MUTEX, "wait on td->mutex\n");
1093         fio_mutex_down(td->mutex);
1094         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1095
1096         /*
1097          * the ->mutex mutex is now no longer used, close it to avoid
1098          * eating a file descriptor
1099          */
1100         fio_mutex_remove(td->mutex);
1101
1102         /*
1103          * A new gid requires privilege, so we need to do this before setting
1104          * the uid.
1105          */
1106         if (td->o.gid != -1U && setgid(td->o.gid)) {
1107                 td_verror(td, errno, "setgid");
1108                 goto err;
1109         }
1110         if (td->o.uid != -1U && setuid(td->o.uid)) {
1111                 td_verror(td, errno, "setuid");
1112                 goto err;
1113         }
1114
1115         /*
1116          * If we have a gettimeofday() thread, make sure we exclude that
1117          * thread from this job
1118          */
1119         if (td->o.gtod_cpu)
1120                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1121
1122         /*
1123          * Set affinity first, in case it has an impact on the memory
1124          * allocations.
1125          */
1126         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1127                 td_verror(td, errno, "cpu_set_affinity");
1128                 goto err;
1129         }
1130
1131         /*
1132          * May alter parameters that init_io_u() will use, so we need to
1133          * do this first.
1134          */
1135         if (init_iolog(td))
1136                 goto err;
1137
1138         if (init_io_u(td))
1139                 goto err;
1140
1141         if (td->o.verify_async && verify_async_init(td))
1142                 goto err;
1143
1144         if (td->ioprio_set) {
1145                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1146                         td_verror(td, errno, "ioprio_set");
1147                         goto err;
1148                 }
1149         }
1150
1151         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1152                 goto err;
1153
1154         if (nice(td->o.nice) == -1) {
1155                 td_verror(td, errno, "nice");
1156                 goto err;
1157         }
1158
1159         if (td->o.ioscheduler && switch_ioscheduler(td))
1160                 goto err;
1161
1162         if (!td->o.create_serialize && setup_files(td))
1163                 goto err;
1164
1165         if (td_io_init(td))
1166                 goto err;
1167
1168         if (init_random_map(td))
1169                 goto err;
1170
1171         if (td->o.exec_prerun) {
1172                 if (exec_string(td->o.exec_prerun))
1173                         goto err;
1174         }
1175
1176         if (td->o.pre_read) {
1177                 if (pre_read_files(td) < 0)
1178                         goto err;
1179         }
1180
1181         fio_gettime(&td->epoch, NULL);
1182         getrusage(RUSAGE_SELF, &td->ru_start);
1183
1184         clear_state = 0;
1185         while (keep_running(td)) {
1186                 fio_gettime(&td->start, NULL);
1187                 memcpy(&td->stat_sample_time[0], &td->start, sizeof(td->start));
1188                 memcpy(&td->stat_sample_time[1], &td->start, sizeof(td->start));
1189                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1190
1191                 if (td->o.ratemin[0] || td->o.ratemin[1])
1192                         memcpy(&td->lastrate, &td->stat_sample_time,
1193                                                         sizeof(td->lastrate));
1194
1195                 if (clear_state)
1196                         clear_io_state(td);
1197
1198                 prune_io_piece_log(td);
1199
1200                 do_io(td);
1201
1202                 clear_state = 1;
1203
1204                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1205                         elapsed = utime_since_now(&td->start);
1206                         td->ts.runtime[DDIR_READ] += elapsed;
1207                 }
1208                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1209                         elapsed = utime_since_now(&td->start);
1210                         td->ts.runtime[DDIR_WRITE] += elapsed;
1211                 }
1212
1213                 if (td->error || td->terminate)
1214                         break;
1215
1216                 if (!td->o.do_verify ||
1217                     td->o.verify == VERIFY_NONE ||
1218                     (td->io_ops->flags & FIO_UNIDIR))
1219                         continue;
1220
1221                 clear_io_state(td);
1222
1223                 fio_gettime(&td->start, NULL);
1224
1225                 do_verify(td);
1226
1227                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1228
1229                 if (td->error || td->terminate)
1230                         break;
1231         }
1232
1233         update_rusage_stat(td);
1234         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1235         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1236         td->ts.total_run_time = mtime_since_now(&td->epoch);
1237         td->ts.io_bytes[0] = td->io_bytes[0];
1238         td->ts.io_bytes[1] = td->io_bytes[1];
1239
1240         fio_mutex_down(writeout_mutex);
1241         if (td->bw_log) {
1242                 if (td->o.bw_log_file) {
1243                         finish_log_named(td, td->bw_log,
1244                                                 td->o.bw_log_file, "bw");
1245                 } else
1246                         finish_log(td, td->bw_log, "bw");
1247         }
1248         if (td->lat_log) {
1249                 if (td->o.lat_log_file) {
1250                         finish_log_named(td, td->lat_log,
1251                                                 td->o.lat_log_file, "lat");
1252                 } else
1253                         finish_log(td, td->lat_log, "lat");
1254         }
1255         if (td->slat_log) {
1256                 if (td->o.lat_log_file) {
1257                         finish_log_named(td, td->slat_log,
1258                                                 td->o.lat_log_file, "slat");
1259                 } else
1260                         finish_log(td, td->slat_log, "slat");
1261         }
1262         if (td->clat_log) {
1263                 if (td->o.lat_log_file) {
1264                         finish_log_named(td, td->clat_log,
1265                                                 td->o.lat_log_file, "clat");
1266                 } else
1267                         finish_log(td, td->clat_log, "clat");
1268         }
1269         fio_mutex_up(writeout_mutex);
1270         if (td->o.exec_postrun)
1271                 exec_string(td->o.exec_postrun);
1272
1273         if (exitall_on_terminate)
1274                 fio_terminate_threads(td->groupid);
1275
1276 err:
1277         if (td->error)
1278                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1279                                                         td->verror);
1280
1281         if (td->o.verify_async)
1282                 verify_async_exit(td);
1283
1284         close_and_free_files(td);
1285         close_ioengine(td);
1286         cleanup_io_u(td);
1287         cgroup_shutdown(td, &cgroup_mnt);
1288
1289         if (td->o.cpumask_set) {
1290                 int ret = fio_cpuset_exit(&td->o.cpumask);
1291
1292                 td_verror(td, ret, "fio_cpuset_exit");
1293         }
1294
1295         /*
1296          * do this very late, it will log file closing as well
1297          */
1298         if (td->o.write_iolog_file)
1299                 write_iolog_close(td);
1300
1301         options_mem_free(td);
1302         td_set_runstate(td, TD_EXITED);
1303         return (void *) (unsigned long) td->error;
1304 }
1305
1306 /*
1307  * We cannot pass the td data into a forked process, so attach the td and
1308  * pass it to the thread worker.
1309  */
1310 static int fork_main(int shmid, int offset)
1311 {
1312         struct thread_data *td;
1313         void *data, *ret;
1314
1315 #ifndef __hpux
1316         data = shmat(shmid, NULL, 0);
1317         if (data == (void *) -1) {
1318                 int __err = errno;
1319
1320                 perror("shmat");
1321                 return __err;
1322         }
1323 #else
1324         /*
1325          * HP-UX inherits shm mappings?
1326          */
1327         data = threads;
1328 #endif
1329
1330         td = data + offset * sizeof(struct thread_data);
1331         ret = thread_main(td);
1332         shmdt(data);
1333         return (int) (unsigned long) ret;
1334 }
1335
1336 /*
1337  * Run over the job map and reap the threads that have exited, if any.
1338  */
1339 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1340 {
1341         struct thread_data *td;
1342         int i, cputhreads, realthreads, pending, status, ret;
1343
1344         /*
1345          * reap exited threads (TD_EXITED -> TD_REAPED)
1346          */
1347         realthreads = pending = cputhreads = 0;
1348         for_each_td(td, i) {
1349                 int flags = 0;
1350
1351                 /*
1352                  * ->io_ops is NULL for a thread that has closed its
1353                  * io engine
1354                  */
1355                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1356                         cputhreads++;
1357                 else
1358                         realthreads++;
1359
1360                 if (!td->pid) {
1361                         pending++;
1362                         continue;
1363                 }
1364                 if (td->runstate == TD_REAPED)
1365                         continue;
1366                 if (td->o.use_thread) {
1367                         if (td->runstate == TD_EXITED) {
1368                                 td_set_runstate(td, TD_REAPED);
1369                                 goto reaped;
1370                         }
1371                         continue;
1372                 }
1373
1374                 flags = WNOHANG;
1375                 if (td->runstate == TD_EXITED)
1376                         flags = 0;
1377
1378                 /*
1379                  * check if someone quit or got killed in an unusual way
1380                  */
1381                 ret = waitpid(td->pid, &status, flags);
1382                 if (ret < 0) {
1383                         if (errno == ECHILD) {
1384                                 log_err("fio: pid=%d disappeared %d\n",
1385                                                 (int) td->pid, td->runstate);
1386                                 td_set_runstate(td, TD_REAPED);
1387                                 goto reaped;
1388                         }
1389                         perror("waitpid");
1390                 } else if (ret == td->pid) {
1391                         if (WIFSIGNALED(status)) {
1392                                 int sig = WTERMSIG(status);
1393
1394                                 if (sig != SIGTERM)
1395                                         log_err("fio: pid=%d, got signal=%d\n",
1396                                                         (int) td->pid, sig);
1397                                 td_set_runstate(td, TD_REAPED);
1398                                 goto reaped;
1399                         }
1400                         if (WIFEXITED(status)) {
1401                                 if (WEXITSTATUS(status) && !td->error)
1402                                         td->error = WEXITSTATUS(status);
1403
1404                                 td_set_runstate(td, TD_REAPED);
1405                                 goto reaped;
1406                         }
1407                 }
1408
1409                 /*
1410                  * thread is not dead, continue
1411                  */
1412                 pending++;
1413                 continue;
1414 reaped:
1415                 (*nr_running)--;
1416                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1417                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1418                 if (!td->pid)
1419                         pending--;
1420
1421                 if (td->error)
1422                         exit_value++;
1423
1424                 done_secs += mtime_since_now(&td->epoch) / 1000;
1425         }
1426
1427         if (*nr_running == cputhreads && !pending && realthreads)
1428                 fio_terminate_threads(TERMINATE_ALL);
1429 }
1430
1431 static void *gtod_thread_main(void *data)
1432 {
1433         fio_mutex_up(startup_mutex);
1434
1435         /*
1436          * As long as we have jobs around, update the clock. It would be nice
1437          * to have some way of NOT hammering that CPU with gettimeofday(),
1438          * but I'm not sure what to use outside of a simple CPU nop to relax
1439          * it - we don't want to lose precision.
1440          */
1441         while (threads) {
1442                 fio_gtod_update();
1443                 nop;
1444         }
1445
1446         return NULL;
1447 }
1448
1449 static int fio_start_gtod_thread(void)
1450 {
1451         pthread_attr_t attr;
1452         int ret;
1453
1454         pthread_attr_init(&attr);
1455         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1456         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1457         pthread_attr_destroy(&attr);
1458         if (ret) {
1459                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1460                 return 1;
1461         }
1462
1463         ret = pthread_detach(gtod_thread);
1464         if (ret) {
1465                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1466                 return 1;
1467         }
1468
1469         dprint(FD_MUTEX, "wait on startup_mutex\n");
1470         fio_mutex_down(startup_mutex);
1471         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1472         return 0;
1473 }
1474
1475 /*
1476  * Main function for kicking off and reaping jobs, as needed.
1477  */
1478 static void run_threads(void)
1479 {
1480         struct thread_data *td;
1481         unsigned long spent;
1482         int i, todo, nr_running, m_rate, t_rate, nr_started;
1483
1484         if (fio_pin_memory())
1485                 return;
1486
1487         if (fio_gtod_offload && fio_start_gtod_thread())
1488                 return;
1489
1490         set_sig_handlers();
1491
1492         if (!terse_output) {
1493                 log_info("Starting ");
1494                 if (nr_thread)
1495                         log_info("%d thread%s", nr_thread,
1496                                                 nr_thread > 1 ? "s" : "");
1497                 if (nr_process) {
1498                         if (nr_thread)
1499                                 printf(" and ");
1500                         log_info("%d process%s", nr_process,
1501                                                 nr_process > 1 ? "es" : "");
1502                 }
1503                 log_info("\n");
1504                 fflush(stdout);
1505         }
1506
1507         todo = thread_number;
1508         nr_running = 0;
1509         nr_started = 0;
1510         m_rate = t_rate = 0;
1511
1512         for_each_td(td, i) {
1513                 print_status_init(td->thread_number - 1);
1514
1515                 if (!td->o.create_serialize)
1516                         continue;
1517
1518                 /*
1519                  * do file setup here so it happens sequentially,
1520                  * we don't want X number of threads getting their
1521                  * client data interspersed on disk
1522                  */
1523                 if (setup_files(td)) {
1524                         exit_value++;
1525                         if (td->error)
1526                                 log_err("fio: pid=%d, err=%d/%s\n",
1527                                         (int) td->pid, td->error, td->verror);
1528                         td_set_runstate(td, TD_REAPED);
1529                         todo--;
1530                 } else {
1531                         struct fio_file *f;
1532                         unsigned int j;
1533
1534                         /*
1535                          * for sharing to work, each job must always open
1536                          * its own files. so close them, if we opened them
1537                          * for creation
1538                          */
1539                         for_each_file(td, f, j) {
1540                                 if (fio_file_open(f))
1541                                         td_io_close_file(td, f);
1542                         }
1543                 }
1544         }
1545
1546         set_genesis_time();
1547
1548         while (todo) {
1549                 struct thread_data *map[REAL_MAX_JOBS];
1550                 struct timeval this_start;
1551                 int this_jobs = 0, left;
1552
1553                 /*
1554                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1555                  */
1556                 for_each_td(td, i) {
1557                         if (td->runstate != TD_NOT_CREATED)
1558                                 continue;
1559
1560                         /*
1561                          * never got a chance to start, killed by other
1562                          * thread for some reason
1563                          */
1564                         if (td->terminate) {
1565                                 todo--;
1566                                 continue;
1567                         }
1568
1569                         if (td->o.start_delay) {
1570                                 spent = mtime_since_genesis();
1571
1572                                 if (td->o.start_delay * 1000 > spent)
1573                                         continue;
1574                         }
1575
1576                         if (td->o.stonewall && (nr_started || nr_running)) {
1577                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1578                                                         td->o.name);
1579                                 break;
1580                         }
1581
1582                         init_disk_util(td);
1583
1584                         /*
1585                          * Set state to created. Thread will transition
1586                          * to TD_INITIALIZED when it's done setting up.
1587                          */
1588                         td_set_runstate(td, TD_CREATED);
1589                         map[this_jobs++] = td;
1590                         nr_started++;
1591
1592                         if (td->o.use_thread) {
1593                                 int ret;
1594
1595                                 dprint(FD_PROCESS, "will pthread_create\n");
1596                                 ret = pthread_create(&td->thread, NULL,
1597                                                         thread_main, td);
1598                                 if (ret) {
1599                                         log_err("pthread_create: %s\n",
1600                                                         strerror(ret));
1601                                         nr_started--;
1602                                         break;
1603                                 }
1604                                 ret = pthread_detach(td->thread);
1605                                 if (ret)
1606                                         log_err("pthread_detach: %s",
1607                                                         strerror(ret));
1608                         } else {
1609                                 pid_t pid;
1610                                 dprint(FD_PROCESS, "will fork\n");
1611                                 pid = fork();
1612                                 if (!pid) {
1613                                         int ret = fork_main(shm_id, i);
1614
1615                                         _exit(ret);
1616                                 } else if (i == fio_debug_jobno)
1617                                         *fio_debug_jobp = pid;
1618                         }
1619                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1620                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1621                                 log_err("fio: job startup hung? exiting.\n");
1622                                 fio_terminate_threads(TERMINATE_ALL);
1623                                 fio_abort = 1;
1624                                 nr_started--;
1625                                 break;
1626                         }
1627                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1628                 }
1629
1630                 /*
1631                  * Wait for the started threads to transition to
1632                  * TD_INITIALIZED.
1633                  */
1634                 fio_gettime(&this_start, NULL);
1635                 left = this_jobs;
1636                 while (left && !fio_abort) {
1637                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1638                                 break;
1639
1640                         usleep(100000);
1641
1642                         for (i = 0; i < this_jobs; i++) {
1643                                 td = map[i];
1644                                 if (!td)
1645                                         continue;
1646                                 if (td->runstate == TD_INITIALIZED) {
1647                                         map[i] = NULL;
1648                                         left--;
1649                                 } else if (td->runstate >= TD_EXITED) {
1650                                         map[i] = NULL;
1651                                         left--;
1652                                         todo--;
1653                                         nr_running++; /* work-around... */
1654                                 }
1655                         }
1656                 }
1657
1658                 if (left) {
1659                         log_err("fio: %d jobs failed to start\n", left);
1660                         for (i = 0; i < this_jobs; i++) {
1661                                 td = map[i];
1662                                 if (!td)
1663                                         continue;
1664                                 kill(td->pid, SIGTERM);
1665                         }
1666                         break;
1667                 }
1668
1669                 /*
1670                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1671                  */
1672                 for_each_td(td, i) {
1673                         if (td->runstate != TD_INITIALIZED)
1674                                 continue;
1675
1676                         if (in_ramp_time(td))
1677                                 td_set_runstate(td, TD_RAMP);
1678                         else
1679                                 td_set_runstate(td, TD_RUNNING);
1680                         nr_running++;
1681                         nr_started--;
1682                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1683                         t_rate += td->o.rate[0] + td->o.rate[1];
1684                         todo--;
1685                         fio_mutex_up(td->mutex);
1686                 }
1687
1688                 reap_threads(&nr_running, &t_rate, &m_rate);
1689
1690                 if (todo)
1691                         usleep(100000);
1692         }
1693
1694         while (nr_running) {
1695                 reap_threads(&nr_running, &t_rate, &m_rate);
1696
1697                 if (is_backend)
1698                         fio_server_idle_loop();
1699                 else
1700                         usleep(10000);
1701         }
1702
1703         update_io_ticks();
1704         fio_unpin_memory();
1705 }
1706
1707 int exec_run(void)
1708 {
1709         if (nr_clients)
1710                 return fio_handle_clients();
1711         if (exec_profile && load_profile(exec_profile))
1712                 return 1;
1713
1714         if (!thread_number)
1715                 return 0;
1716
1717         if (write_bw_log) {
1718                 setup_log(&agg_io_log[DDIR_READ]);
1719                 setup_log(&agg_io_log[DDIR_WRITE]);
1720         }
1721
1722         startup_mutex = fio_mutex_init(0);
1723         if (startup_mutex == NULL)
1724                 return 1;
1725         writeout_mutex = fio_mutex_init(1);
1726         if (writeout_mutex == NULL)
1727                 return 1;
1728
1729         set_genesis_time();
1730         create_disk_util_thread();
1731
1732         cgroup_list = smalloc(sizeof(*cgroup_list));
1733         INIT_FLIST_HEAD(cgroup_list);
1734
1735         run_threads();
1736
1737         if (!fio_abort) {
1738                 show_run_stats();
1739                 if (write_bw_log) {
1740                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1741                         __finish_log(agg_io_log[DDIR_WRITE],
1742                                         "agg-write_bw.log");
1743                 }
1744         }
1745
1746         cgroup_kill(cgroup_list);
1747         sfree(cgroup_list);
1748         sfree(cgroup_mnt);
1749
1750         fio_mutex_remove(startup_mutex);
1751         fio_mutex_remove(writeout_mutex);
1752         return exit_value;
1753 }
1754
1755 void reset_fio_state(void)
1756 {
1757         groupid = 0;
1758         thread_number = 0;
1759         nr_process = 0;
1760         nr_thread = 0;
1761         done_secs = 0;
1762 }
1763
1764 int main(int argc, char *argv[], char *envp[])
1765 {
1766         long ps;
1767
1768         arch_init(envp);
1769
1770         sinit();
1771
1772         /*
1773          * We need locale for number printing, if it isn't set then just
1774          * go with the US format.
1775          */
1776         if (!getenv("LC_NUMERIC"))
1777                 setlocale(LC_NUMERIC, "en_US");
1778
1779         ps = sysconf(_SC_PAGESIZE);
1780         if (ps < 0) {
1781                 log_err("Failed to get page size\n");
1782                 return 1;
1783         }
1784
1785         page_size = ps;
1786         page_mask = ps - 1;
1787
1788         fio_keywords_init();
1789
1790         if (parse_options(argc, argv))
1791                 return 1;
1792
1793         return exec_run();
1794 }