backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 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 static pthread_t disk_util_thread;
  52 static struct fio_mutex *startup_mutex;
  53 static struct fio_mutex *writeout_mutex;
  54 static struct flist_head *cgroup_list;
  55 static char *cgroup_mnt;
  56 static int exit_value;
  57 static volatile int fio_abort;
  58
  59 struct io_log *agg_io_log[2];
  60
  61 #define PAGE_ALIGN(buf) \
  62         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  63
  64 #define JOB_START_TIMEOUT       (5 * 1000)
  65
  66 static void sig_int(int sig)
  67 {
  68         if (threads) {
  69                 if (is_backend)
  70                         fio_server_got_signal(sig);
  71                 else {
  72                         log_info("\nfio: terminating on signal %d\n", sig);
  73                         fflush(stdout);
  74                         exit_value = 128;
  75                 }
  76
  77                 fio_terminate_threads(TERMINATE_ALL);
  78         }
  79 }
  80
  81 static void set_sig_handlers(void)
  82 {
  83         struct sigaction act;
  84
  85         memset(&act, 0, sizeof(act));
  86         act.sa_handler = sig_int;
  87         act.sa_flags = SA_RESTART;
  88         sigaction(SIGINT, &act, NULL);
  89
  90         memset(&act, 0, sizeof(act));
  91         act.sa_handler = sig_int;
  92         act.sa_flags = SA_RESTART;
  93         sigaction(SIGTERM, &act, NULL);
  94
  95         if (is_backend) {
  96                 memset(&act, 0, sizeof(act));
  97                 act.sa_handler = sig_int;
  98                 act.sa_flags = SA_RESTART;
  99                 sigaction(SIGPIPE, &act, NULL);
 100         }
 101 }
 102
 103 /*
 104  * Check if we are above the minimum rate given.
 105  */
 106 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 107                             enum fio_ddir ddir)
 108 {
 109         unsigned long long bytes = 0;
 110         unsigned long iops = 0;
 111         unsigned long spent;
 112         unsigned long rate;
 113         unsigned int ratemin = 0;
 114         unsigned int rate_iops = 0;
 115         unsigned int rate_iops_min = 0;
 116
 117         assert(ddir_rw(ddir));
 118
 119         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 120                 return 0;
 121
 122         /*
 123          * allow a 2 second settle period in the beginning
 124          */
 125         if (mtime_since(&td->start, now) < 2000)
 126                 return 0;
 127
 128         iops += td->this_io_blocks[ddir];
 129         bytes += td->this_io_bytes[ddir];
 130         ratemin += td->o.ratemin[ddir];
 131         rate_iops += td->o.rate_iops[ddir];
 132         rate_iops_min += td->o.rate_iops_min[ddir];
 133
 134         /*
 135          * if rate blocks is set, sample is running
 136          */
 137         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 138                 spent = mtime_since(&td->lastrate[ddir], now);
 139                 if (spent < td->o.ratecycle)
 140                         return 0;
 141
 142                 if (td->o.rate[ddir]) {
 143                         /*
 144                          * check bandwidth specified rate
 145                          */
 146                         if (bytes < td->rate_bytes[ddir]) {
 147                                 log_err("%s: min rate %u not met\n", td->o.name,
 148                                                                 ratemin);
 149                                 return 1;
 150                         } else {
 151                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 152                                 if (rate < ratemin ||
 153                                     bytes < td->rate_bytes[ddir]) {
 154                                         log_err("%s: min rate %u not met, got"
 155                                                 " %luKB/sec\n", td->o.name,
 156                                                         ratemin, rate);
 157                                         return 1;
 158                                 }
 159                         }
 160                 } else {
 161                         /*
 162                          * checks iops specified rate
 163                          */
 164                         if (iops < rate_iops) {
 165                                 log_err("%s: min iops rate %u not met\n",
 166                                                 td->o.name, rate_iops);
 167                                 return 1;
 168                         } else {
 169                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 170                                 if (rate < rate_iops_min ||
 171                                     iops < td->rate_blocks[ddir]) {
 172                                         log_err("%s: min iops rate %u not met,"
 173                                                 " got %lu\n", td->o.name,
 174                                                         rate_iops_min, rate);
 175                                 }
 176                         }
 177                 }
 178         }
 179
 180         td->rate_bytes[ddir] = bytes;
 181         td->rate_blocks[ddir] = iops;
 182         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 183         return 0;
 184 }
 185
 186 static int check_min_rate(struct thread_data *td, struct timeval *now,
 187                           unsigned long *bytes_done)
 188 {
 189         int ret = 0;
 190
 191         if (bytes_done[0])
 192                 ret |= __check_min_rate(td, now, 0);
 193         if (bytes_done[1])
 194                 ret |= __check_min_rate(td, now, 1);
 195
 196         return ret;
 197 }
 198
 199 /*
 200  * When job exits, we can cancel the in-flight IO if we are using async
 201  * io. Attempt to do so.
 202  */
 203 static void cleanup_pending_aio(struct thread_data *td)
 204 {
 205         struct flist_head *entry, *n;
 206         struct io_u *io_u;
 207         int r;
 208
 209         /*
 210          * get immediately available events, if any
 211          */
 212         r = io_u_queued_complete(td, 0, NULL);
 213         if (r < 0)
 214                 return;
 215
 216         /*
 217          * now cancel remaining active events
 218          */
 219         if (td->io_ops->cancel) {
 220                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 221                         io_u = flist_entry(entry, struct io_u, list);
 222
 223                         /*
 224                          * if the io_u isn't in flight, then that generally
 225                          * means someone leaked an io_u. complain but fix
 226                          * it up, so we don't stall here.
 227                          */
 228                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 229                                 log_err("fio: non-busy IO on busy list\n");
 230                                 put_io_u(td, io_u);
 231                         } else {
 232                                 r = td->io_ops->cancel(td, io_u);
 233                                 if (!r)
 234                                         put_io_u(td, io_u);
 235                         }
 236                 }
 237         }
 238
 239         if (td->cur_depth)
 240                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 241 }
 242
 243 /*
 244  * Helper to handle the final sync of a file. Works just like the normal
 245  * io path, just does everything sync.
 246  */
 247 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 248 {
 249         struct io_u *io_u = __get_io_u(td);
 250         int ret;
 251
 252         if (!io_u)
 253                 return 1;
 254
 255         io_u->ddir = DDIR_SYNC;
 256         io_u->file = f;
 257
 258         if (td_io_prep(td, io_u)) {
 259                 put_io_u(td, io_u);
 260                 return 1;
 261         }
 262
 263 requeue:
 264         ret = td_io_queue(td, io_u);
 265         if (ret < 0) {
 266                 td_verror(td, io_u->error, "td_io_queue");
 267                 put_io_u(td, io_u);
 268                 return 1;
 269         } else if (ret == FIO_Q_QUEUED) {
 270                 if (io_u_queued_complete(td, 1, NULL) < 0)
 271                         return 1;
 272         } else if (ret == FIO_Q_COMPLETED) {
 273                 if (io_u->error) {
 274                         td_verror(td, io_u->error, "td_io_queue");
 275                         return 1;
 276                 }
 277
 278                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 279                         return 1;
 280         } else if (ret == FIO_Q_BUSY) {
 281                 if (td_io_commit(td))
 282                         return 1;
 283                 goto requeue;
 284         }
 285
 286         return 0;
 287 }
 288 static inline void __update_tv_cache(struct thread_data *td)
 289 {
 290         fio_gettime(&td->tv_cache, NULL);
 291 }
 292
 293 static inline void update_tv_cache(struct thread_data *td)
 294 {
 295         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 296                 __update_tv_cache(td);
 297 }
 298
 299 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 300 {
 301         if (in_ramp_time(td))
 302                 return 0;
 303         if (!td->o.timeout)
 304                 return 0;
 305         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 306                 return 1;
 307
 308         return 0;
 309 }
 310
 311 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 312                                int *retptr)
 313 {
 314         int ret = *retptr;
 315
 316         if (ret < 0 || td->error) {
 317                 int err;
 318
 319                 if (ret < 0)
 320                         err = -ret;
 321                 else
 322                         err = td->error;
 323
 324                 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
 325                         return 1;
 326
 327                 if (td_non_fatal_error(err)) {
 328                         /*
 329                          * Continue with the I/Os in case of
 330                          * a non fatal error.
 331                          */
 332                         update_error_count(td, err);
 333                         td_clear_error(td);
 334                         *retptr = 0;
 335                         return 0;
 336                 } else if (td->o.fill_device && err == ENOSPC) {
 337                         /*
 338                          * We expect to hit this error if
 339                          * fill_device option is set.
 340                          */
 341                         td_clear_error(td);
 342                         td->terminate = 1;
 343                         return 1;
 344                 } else {
 345                         /*
 346                          * Stop the I/O in case of a fatal
 347                          * error.
 348                          */
 349                         update_error_count(td, err);
 350                         return 1;
 351                 }
 352         }
 353
 354         return 0;
 355 }
 356
 357
 358
 359 /*
 360  * The main verify engine. Runs over the writes we previously submitted,
 361  * reads the blocks back in, and checks the crc/md5 of the data.
 362  */
 363 static void do_verify(struct thread_data *td)
 364 {
 365         struct fio_file *f;
 366         struct io_u *io_u;
 367         int ret, min_events;
 368         unsigned int i;
 369
 370         dprint(FD_VERIFY, "starting loop\n");
 371
 372         /*
 373          * sync io first and invalidate cache, to make sure we really
 374          * read from disk.
 375          */
 376         for_each_file(td, f, i) {
 377                 if (!fio_file_open(f))
 378                         continue;
 379                 if (fio_io_sync(td, f))
 380                         break;
 381                 if (file_invalidate_cache(td, f))
 382                         break;
 383         }
 384
 385         if (td->error)
 386                 return;
 387
 388         td_set_runstate(td, TD_VERIFYING);
 389
 390         io_u = NULL;
 391         while (!td->terminate) {
 392                 int ret2, full;
 393
 394                 update_tv_cache(td);
 395
 396                 if (runtime_exceeded(td, &td->tv_cache)) {
 397                         __update_tv_cache(td);
 398                         if (runtime_exceeded(td, &td->tv_cache)) {
 399                                 td->terminate = 1;
 400                                 break;
 401                         }
 402                 }
 403
 404                 io_u = __get_io_u(td);
 405                 if (!io_u)
 406                         break;
 407
 408                 if (get_next_verify(td, io_u)) {
 409                         put_io_u(td, io_u);
 410                         break;
 411                 }
 412
 413                 if (td_io_prep(td, io_u)) {
 414                         put_io_u(td, io_u);
 415                         break;
 416                 }
 417
 418                 if (td->o.verify_async)
 419                         io_u->end_io = verify_io_u_async;
 420                 else
 421                         io_u->end_io = verify_io_u;
 422
 423                 ret = td_io_queue(td, io_u);
 424                 switch (ret) {
 425                 case FIO_Q_COMPLETED:
 426                         if (io_u->error) {
 427                                 ret = -io_u->error;
 428                                 clear_io_u(td, io_u);
 429                         } else if (io_u->resid) {
 430                                 int bytes = io_u->xfer_buflen - io_u->resid;
 431
 432                                 /*
 433                                  * zero read, fail
 434                                  */
 435                                 if (!bytes) {
 436                                         td_verror(td, EIO, "full resid");
 437                                         put_io_u(td, io_u);
 438                                         break;
 439                                 }
 440
 441                                 io_u->xfer_buflen = io_u->resid;
 442                                 io_u->xfer_buf += bytes;
 443                                 io_u->offset += bytes;
 444
 445                                 if (ddir_rw(io_u->ddir))
 446                                         td->ts.short_io_u[io_u->ddir]++;
 447
 448                                 f = io_u->file;
 449                                 if (io_u->offset == f->real_file_size)
 450                                         goto sync_done;
 451
 452                                 requeue_io_u(td, &io_u);
 453                         } else {
 454 sync_done:
 455                                 ret = io_u_sync_complete(td, io_u, NULL);
 456                                 if (ret < 0)
 457                                         break;
 458                         }
 459                         continue;
 460                 case FIO_Q_QUEUED:
 461                         break;
 462                 case FIO_Q_BUSY:
 463                         requeue_io_u(td, &io_u);
 464                         ret2 = td_io_commit(td);
 465                         if (ret2 < 0)
 466                                 ret = ret2;
 467                         break;
 468                 default:
 469                         assert(ret < 0);
 470                         td_verror(td, -ret, "td_io_queue");
 471                         break;
 472                 }
 473
 474                 if (break_on_this_error(td, io_u->ddir, &ret))
 475                         break;
 476
 477                 /*
 478                  * if we can queue more, do so. but check if there are
 479                  * completed io_u's first. Note that we can get BUSY even
 480                  * without IO queued, if the system is resource starved.
 481                  */
 482                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 483                 if (full || !td->o.iodepth_batch_complete) {
 484                         min_events = min(td->o.iodepth_batch_complete,
 485                                          td->cur_depth);
 486                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
 487                                 min_events = 1;
 488
 489                         do {
 490                                 /*
 491                                  * Reap required number of io units, if any,
 492                                  * and do the verification on them through
 493                                  * the callback handler
 494                                  */
 495                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 496                                         ret = -1;
 497                                         break;
 498                                 }
 499                         } while (full && (td->cur_depth > td->o.iodepth_low));
 500                 }
 501                 if (ret < 0)
 502                         break;
 503         }
 504
 505         if (!td->error) {
 506                 min_events = td->cur_depth;
 507
 508                 if (min_events)
 509                         ret = io_u_queued_complete(td, min_events, NULL);
 510         } else
 511                 cleanup_pending_aio(td);
 512
 513         td_set_runstate(td, TD_RUNNING);
 514
 515         dprint(FD_VERIFY, "exiting loop\n");
 516 }
 517
 518 /*
 519  * Main IO worker function. It retrieves io_u's to process and queues
 520  * and reaps them, checking for rate and errors along the way.
 521  */
 522 static void do_io(struct thread_data *td)
 523 {
 524         unsigned int i;
 525         int ret = 0;
 526
 527         if (in_ramp_time(td))
 528                 td_set_runstate(td, TD_RAMP);
 529         else
 530                 td_set_runstate(td, TD_RUNNING);
 531
 532         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 533                 (!flist_empty(&td->trim_list)) ||
 534                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
 535                 struct timeval comp_time;
 536                 unsigned long bytes_done[2] = { 0, 0 };
 537                 int min_evts = 0;
 538                 struct io_u *io_u;
 539                 int ret2, full;
 540                 enum fio_ddir ddir;
 541
 542                 if (td->terminate)
 543                         break;
 544
 545                 update_tv_cache(td);
 546
 547                 if (runtime_exceeded(td, &td->tv_cache)) {
 548                         __update_tv_cache(td);
 549                         if (runtime_exceeded(td, &td->tv_cache)) {
 550                                 td->terminate = 1;
 551                                 break;
 552                         }
 553                 }
 554
 555                 io_u = get_io_u(td);
 556                 if (!io_u)
 557                         break;
 558
 559                 ddir = io_u->ddir;
 560
 561                 /*
 562                  * Add verification end_io handler, if asked to verify
 563                  * a previously written file.
 564                  */
 565                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 566                     !td_rw(td)) {
 567                         if (td->o.verify_async)
 568                                 io_u->end_io = verify_io_u_async;
 569                         else
 570                                 io_u->end_io = verify_io_u;
 571                         td_set_runstate(td, TD_VERIFYING);
 572                 } else if (in_ramp_time(td))
 573                         td_set_runstate(td, TD_RAMP);
 574                 else
 575                         td_set_runstate(td, TD_RUNNING);
 576
 577                 ret = td_io_queue(td, io_u);
 578                 switch (ret) {
 579                 case FIO_Q_COMPLETED:
 580                         if (io_u->error) {
 581                                 ret = -io_u->error;
 582                                 clear_io_u(td, io_u);
 583                         } else if (io_u->resid) {
 584                                 int bytes = io_u->xfer_buflen - io_u->resid;
 585                                 struct fio_file *f = io_u->file;
 586
 587                                 /*
 588                                  * zero read, fail
 589                                  */
 590                                 if (!bytes) {
 591                                         td_verror(td, EIO, "full resid");
 592                                         put_io_u(td, io_u);
 593                                         break;
 594                                 }
 595
 596                                 io_u->xfer_buflen = io_u->resid;
 597                                 io_u->xfer_buf += bytes;
 598                                 io_u->offset += bytes;
 599
 600                                 if (ddir_rw(io_u->ddir))
 601                                         td->ts.short_io_u[io_u->ddir]++;
 602
 603                                 if (io_u->offset == f->real_file_size)
 604                                         goto sync_done;
 605
 606                                 requeue_io_u(td, &io_u);
 607                         } else {
 608 sync_done:
 609                                 if (__should_check_rate(td, 0) ||
 610                                     __should_check_rate(td, 1))
 611                                         fio_gettime(&comp_time, NULL);
 612
 613                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 614                                 if (ret < 0)
 615                                         break;
 616                         }
 617                         break;
 618                 case FIO_Q_QUEUED:
 619                         /*
 620                          * if the engine doesn't have a commit hook,
 621                          * the io_u is really queued. if it does have such
 622                          * a hook, it has to call io_u_queued() itself.
 623                          */
 624                         if (td->io_ops->commit == NULL)
 625                                 io_u_queued(td, io_u);
 626                         break;
 627                 case FIO_Q_BUSY:
 628                         requeue_io_u(td, &io_u);
 629                         ret2 = td_io_commit(td);
 630                         if (ret2 < 0)
 631                                 ret = ret2;
 632                         break;
 633                 default:
 634                         assert(ret < 0);
 635                         put_io_u(td, io_u);
 636                         break;
 637                 }
 638
 639                 if (break_on_this_error(td, ddir, &ret))
 640                         break;
 641
 642                 /*
 643                  * See if we need to complete some commands. Note that we
 644                  * can get BUSY even without IO queued, if the system is
 645                  * resource starved.
 646                  */
 647                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 648                 if (full || !td->o.iodepth_batch_complete) {
 649                         min_evts = min(td->o.iodepth_batch_complete,
 650                                         td->cur_depth);
 651                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
 652                                 min_evts = 1;
 653
 654                         if (__should_check_rate(td, 0) ||
 655                             __should_check_rate(td, 1))
 656                                 fio_gettime(&comp_time, NULL);
 657
 658                         do {
 659                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 660                                 if (ret < 0)
 661                                         break;
 662
 663                         } while (full && (td->cur_depth > td->o.iodepth_low));
 664                 }
 665
 666                 if (ret < 0)
 667                         break;
 668                 if (!(bytes_done[0] + bytes_done[1]))
 669                         continue;
 670
 671                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 672                         if (check_min_rate(td, &comp_time, bytes_done)) {
 673                                 if (exitall_on_terminate)
 674                                         fio_terminate_threads(td->groupid);
 675                                 td_verror(td, EIO, "check_min_rate");
 676                                 break;
 677                         }
 678                 }
 679
 680                 if (td->o.thinktime) {
 681                         unsigned long long b;
 682
 683                         b = td->io_blocks[0] + td->io_blocks[1];
 684                         if (!(b % td->o.thinktime_blocks)) {
 685                                 int left;
 686
 687                                 if (td->o.thinktime_spin)
 688                                         usec_spin(td->o.thinktime_spin);
 689
 690                                 left = td->o.thinktime - td->o.thinktime_spin;
 691                                 if (left)
 692                                         usec_sleep(td, left);
 693                         }
 694                 }
 695         }
 696
 697         if (td->trim_entries)
 698                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 699
 700         if (td->o.fill_device && td->error == ENOSPC) {
 701                 td->error = 0;
 702                 td->terminate = 1;
 703         }
 704         if (!td->error) {
 705                 struct fio_file *f;
 706
 707                 i = td->cur_depth;
 708                 if (i) {
 709                         ret = io_u_queued_complete(td, i, NULL);
 710                         if (td->o.fill_device && td->error == ENOSPC)
 711                                 td->error = 0;
 712                 }
 713
 714                 if (should_fsync(td) && td->o.end_fsync) {
 715                         td_set_runstate(td, TD_FSYNCING);
 716
 717                         for_each_file(td, f, i) {
 718                                 if (!fio_file_open(f))
 719                                         continue;
 720                                 fio_io_sync(td, f);
 721                         }
 722                 }
 723         } else
 724                 cleanup_pending_aio(td);
 725
 726         /*
 727          * stop job if we failed doing any IO
 728          */
 729         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 730                 td->done = 1;
 731 }
 732
 733 static void cleanup_io_u(struct thread_data *td)
 734 {
 735         struct flist_head *entry, *n;
 736         struct io_u *io_u;
 737
 738         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 739                 io_u = flist_entry(entry, struct io_u, list);
 740
 741                 flist_del(&io_u->list);
 742                 fio_memfree(io_u, sizeof(*io_u));
 743         }
 744
 745         free_io_mem(td);
 746 }
 747
 748 static int init_io_u(struct thread_data *td)
 749 {
 750         struct io_u *io_u;
 751         unsigned int max_bs;
 752         int cl_align, i, max_units;
 753         char *p;
 754
 755         max_units = td->o.iodepth;
 756         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 757         td->orig_buffer_size = (unsigned long long) max_bs
 758                                         * (unsigned long long) max_units;
 759
 760         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 761                 unsigned long bs;
 762
 763                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 764                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 765         }
 766
 767         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 768                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 769                 return 1;
 770         }
 771
 772         if (allocate_io_mem(td))
 773                 return 1;
 774
 775         if (td->o.odirect || td->o.mem_align ||
 776             (td->io_ops->flags & FIO_RAWIO))
 777                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 778         else
 779                 p = td->orig_buffer;
 780
 781         cl_align = os_cache_line_size();
 782
 783         for (i = 0; i < max_units; i++) {
 784                 void *ptr;
 785
 786                 if (td->terminate)
 787                         return 1;
 788
 789                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 790                 if (!ptr) {
 791                         log_err("fio: unable to allocate aligned memory\n");
 792                         break;
 793                 }
 794
 795                 io_u = ptr;
 796                 memset(io_u, 0, sizeof(*io_u));
 797                 INIT_FLIST_HEAD(&io_u->list);
 798                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 799
 800                 if (!(td->io_ops->flags & FIO_NOIO)) {
 801                         io_u->buf = p;
 802                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 803
 804                         if (td_write(td))
 805                                 io_u_fill_buffer(td, io_u, max_bs);
 806                         if (td_write(td) && td->o.verify_pattern_bytes) {
 807                                 /*
 808                                  * Fill the buffer with the pattern if we are
 809                                  * going to be doing writes.
 810                                  */
 811                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 812                         }
 813                 }
 814
 815                 io_u->index = i;
 816                 io_u->flags = IO_U_F_FREE;
 817                 flist_add(&io_u->list, &td->io_u_freelist);
 818                 p += max_bs;
 819         }
 820
 821         return 0;
 822 }
 823
 824 static int switch_ioscheduler(struct thread_data *td)
 825 {
 826         char tmp[256], tmp2[128];
 827         FILE *f;
 828         int ret;
 829
 830         if (td->io_ops->flags & FIO_DISKLESSIO)
 831                 return 0;
 832
 833         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 834
 835         f = fopen(tmp, "r+");
 836         if (!f) {
 837                 if (errno == ENOENT) {
 838                         log_err("fio: os or kernel doesn't support IO scheduler"
 839                                 " switching\n");
 840                         return 0;
 841                 }
 842                 td_verror(td, errno, "fopen iosched");
 843                 return 1;
 844         }
 845
 846         /*
 847          * Set io scheduler.
 848          */
 849         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 850         if (ferror(f) || ret != 1) {
 851                 td_verror(td, errno, "fwrite");
 852                 fclose(f);
 853                 return 1;
 854         }
 855
 856         rewind(f);
 857
 858         /*
 859          * Read back and check that the selected scheduler is now the default.
 860          */
 861         ret = fread(tmp, 1, sizeof(tmp), f);
 862         if (ferror(f) || ret < 0) {
 863                 td_verror(td, errno, "fread");
 864                 fclose(f);
 865                 return 1;
 866         }
 867
 868         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 869         if (!strstr(tmp, tmp2)) {
 870                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 871                 td_verror(td, EINVAL, "iosched_switch");
 872                 fclose(f);
 873                 return 1;
 874         }
 875
 876         fclose(f);
 877         return 0;
 878 }
 879
 880 static int keep_running(struct thread_data *td)
 881 {
 882         unsigned long long io_done;
 883
 884         if (td->done)
 885                 return 0;
 886         if (td->o.time_based)
 887                 return 1;
 888         if (td->o.loops) {
 889                 td->o.loops--;
 890                 return 1;
 891         }
 892
 893         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 894                         + td->io_skip_bytes;
 895         if (io_done < td->o.size)
 896                 return 1;
 897
 898         return 0;
 899 }
 900
 901 static int exec_string(const char *string)
 902 {
 903         int ret, newlen = strlen(string) + 1 + 8;
 904         char *str;
 905
 906         str = malloc(newlen);
 907         sprintf(str, "sh -c %s", string);
 908
 909         ret = system(str);
 910         if (ret == -1)
 911                 log_err("fio: exec of cmd <%s> failed\n", str);
 912
 913         free(str);
 914         return ret;
 915 }
 916
 917 /*
 918  * Entry point for the thread based jobs. The process based jobs end up
 919  * here as well, after a little setup.
 920  */
 921 static void *thread_main(void *data)
 922 {
 923         unsigned long long elapsed;
 924         struct thread_data *td = data;
 925         pthread_condattr_t attr;
 926         int clear_state;
 927
 928         if (!td->o.use_thread) {
 929                 setsid();
 930                 td->pid = getpid();
 931         } else
 932                 td->pid = gettid();
 933
 934         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 935
 936         INIT_FLIST_HEAD(&td->io_u_freelist);
 937         INIT_FLIST_HEAD(&td->io_u_busylist);
 938         INIT_FLIST_HEAD(&td->io_u_requeues);
 939         INIT_FLIST_HEAD(&td->io_log_list);
 940         INIT_FLIST_HEAD(&td->io_hist_list);
 941         INIT_FLIST_HEAD(&td->verify_list);
 942         INIT_FLIST_HEAD(&td->trim_list);
 943         pthread_mutex_init(&td->io_u_lock, NULL);
 944         td->io_hist_tree = RB_ROOT;
 945
 946         pthread_condattr_init(&attr);
 947         pthread_cond_init(&td->verify_cond, &attr);
 948         pthread_cond_init(&td->free_cond, &attr);
 949
 950         td_set_runstate(td, TD_INITIALIZED);
 951         dprint(FD_MUTEX, "up startup_mutex\n");
 952         fio_mutex_up(startup_mutex);
 953         dprint(FD_MUTEX, "wait on td->mutex\n");
 954         fio_mutex_down(td->mutex);
 955         dprint(FD_MUTEX, "done waiting on td->mutex\n");
 956
 957         /*
 958          * the ->mutex mutex is now no longer used, close it to avoid
 959          * eating a file descriptor
 960          */
 961         fio_mutex_remove(td->mutex);
 962
 963         /*
 964          * A new gid requires privilege, so we need to do this before setting
 965          * the uid.
 966          */
 967         if (td->o.gid != -1U && setgid(td->o.gid)) {
 968                 td_verror(td, errno, "setgid");
 969                 goto err;
 970         }
 971         if (td->o.uid != -1U && setuid(td->o.uid)) {
 972                 td_verror(td, errno, "setuid");
 973                 goto err;
 974         }
 975
 976         /*
 977          * If we have a gettimeofday() thread, make sure we exclude that
 978          * thread from this job
 979          */
 980         if (td->o.gtod_cpu)
 981                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
 982
 983         /*
 984          * Set affinity first, in case it has an impact on the memory
 985          * allocations.
 986          */
 987         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
 988                 td_verror(td, errno, "cpu_set_affinity");
 989                 goto err;
 990         }
 991
 992         /*
 993          * May alter parameters that init_io_u() will use, so we need to
 994          * do this first.
 995          */
 996         if (init_iolog(td))
 997                 goto err;
 998
 999         if (init_io_u(td))
1000                 goto err;
1001
1002         if (td->o.verify_async && verify_async_init(td))
1003                 goto err;
1004
1005         if (td->ioprio_set) {
1006                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1007                         td_verror(td, errno, "ioprio_set");
1008                         goto err;
1009                 }
1010         }
1011
1012         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1013                 goto err;
1014
1015         if (nice(td->o.nice) == -1) {
1016                 td_verror(td, errno, "nice");
1017                 goto err;
1018         }
1019
1020         if (td->o.ioscheduler && switch_ioscheduler(td))
1021                 goto err;
1022
1023         if (!td->o.create_serialize && setup_files(td))
1024                 goto err;
1025
1026         if (td_io_init(td))
1027                 goto err;
1028
1029         if (init_random_map(td))
1030                 goto err;
1031
1032         if (td->o.exec_prerun) {
1033                 if (exec_string(td->o.exec_prerun))
1034                         goto err;
1035         }
1036
1037         if (td->o.pre_read) {
1038                 if (pre_read_files(td) < 0)
1039                         goto err;
1040         }
1041
1042         fio_gettime(&td->epoch, NULL);
1043         getrusage(RUSAGE_SELF, &td->ru_start);
1044
1045         clear_state = 0;
1046         while (keep_running(td)) {
1047                 fio_gettime(&td->start, NULL);
1048                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1049                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1050                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1051
1052                 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1053                         memcpy(&td->lastrate[0], &td->bw_sample_time,
1054                                                 sizeof(td->bw_sample_time));
1055                         memcpy(&td->lastrate[1], &td->bw_sample_time,
1056                                                 sizeof(td->bw_sample_time));
1057                 }
1058
1059                 if (clear_state)
1060                         clear_io_state(td);
1061
1062                 prune_io_piece_log(td);
1063
1064                 do_io(td);
1065
1066                 clear_state = 1;
1067
1068                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1069                         elapsed = utime_since_now(&td->start);
1070                         td->ts.runtime[DDIR_READ] += elapsed;
1071                 }
1072                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1073                         elapsed = utime_since_now(&td->start);
1074                         td->ts.runtime[DDIR_WRITE] += elapsed;
1075                 }
1076
1077                 if (td->error || td->terminate)
1078                         break;
1079
1080                 if (!td->o.do_verify ||
1081                     td->o.verify == VERIFY_NONE ||
1082                     (td->io_ops->flags & FIO_UNIDIR))
1083                         continue;
1084
1085                 clear_io_state(td);
1086
1087                 fio_gettime(&td->start, NULL);
1088
1089                 do_verify(td);
1090
1091                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1092
1093                 if (td->error || td->terminate)
1094                         break;
1095         }
1096
1097         update_rusage_stat(td);
1098         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1099         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1100         td->ts.total_run_time = mtime_since_now(&td->epoch);
1101         td->ts.io_bytes[0] = td->io_bytes[0];
1102         td->ts.io_bytes[1] = td->io_bytes[1];
1103
1104         fio_mutex_down(writeout_mutex);
1105         if (td->bw_log) {
1106                 if (td->o.bw_log_file) {
1107                         finish_log_named(td, td->bw_log,
1108                                                 td->o.bw_log_file, "bw");
1109                 } else
1110                         finish_log(td, td->bw_log, "bw");
1111         }
1112         if (td->lat_log) {
1113                 if (td->o.lat_log_file) {
1114                         finish_log_named(td, td->lat_log,
1115                                                 td->o.lat_log_file, "lat");
1116                 } else
1117                         finish_log(td, td->lat_log, "lat");
1118         }
1119         if (td->slat_log) {
1120                 if (td->o.lat_log_file) {
1121                         finish_log_named(td, td->slat_log,
1122                                                 td->o.lat_log_file, "slat");
1123                 } else
1124                         finish_log(td, td->slat_log, "slat");
1125         }
1126         if (td->clat_log) {
1127                 if (td->o.lat_log_file) {
1128                         finish_log_named(td, td->clat_log,
1129                                                 td->o.lat_log_file, "clat");
1130                 } else
1131                         finish_log(td, td->clat_log, "clat");
1132         }
1133         if (td->iops_log) {
1134                 if (td->o.iops_log_file) {
1135                         finish_log_named(td, td->iops_log,
1136                                                 td->o.iops_log_file, "iops");
1137                 } else
1138                         finish_log(td, td->iops_log, "iops");
1139         }
1140
1141         fio_mutex_up(writeout_mutex);
1142         if (td->o.exec_postrun)
1143                 exec_string(td->o.exec_postrun);
1144
1145         if (exitall_on_terminate)
1146                 fio_terminate_threads(td->groupid);
1147
1148 err:
1149         if (td->error)
1150                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1151                                                         td->verror);
1152
1153         if (td->o.verify_async)
1154                 verify_async_exit(td);
1155
1156         close_and_free_files(td);
1157         close_ioengine(td);
1158         cleanup_io_u(td);
1159         cgroup_shutdown(td, &cgroup_mnt);
1160
1161         if (td->o.cpumask_set) {
1162                 int ret = fio_cpuset_exit(&td->o.cpumask);
1163
1164                 td_verror(td, ret, "fio_cpuset_exit");
1165         }
1166
1167         /*
1168          * do this very late, it will log file closing as well
1169          */
1170         if (td->o.write_iolog_file)
1171                 write_iolog_close(td);
1172
1173         td_set_runstate(td, TD_EXITED);
1174         return (void *) (unsigned long) td->error;
1175 }
1176
1177
1178 /*
1179  * We cannot pass the td data into a forked process, so attach the td and
1180  * pass it to the thread worker.
1181  */
1182 static int fork_main(int shmid, int offset)
1183 {
1184         struct thread_data *td;
1185         void *data, *ret;
1186
1187 #ifndef __hpux
1188         data = shmat(shmid, NULL, 0);
1189         if (data == (void *) -1) {
1190                 int __err = errno;
1191
1192                 perror("shmat");
1193                 return __err;
1194         }
1195 #else
1196         /*
1197          * HP-UX inherits shm mappings?
1198          */
1199         data = threads;
1200 #endif
1201
1202         td = data + offset * sizeof(struct thread_data);
1203         ret = thread_main(td);
1204         shmdt(data);
1205         return (int) (unsigned long) ret;
1206 }
1207
1208 /*
1209  * Run over the job map and reap the threads that have exited, if any.
1210  */
1211 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1212                          unsigned int *m_rate)
1213 {
1214         struct thread_data *td;
1215         unsigned int cputhreads, realthreads, pending;
1216         int i, status, ret;
1217
1218         /*
1219          * reap exited threads (TD_EXITED -> TD_REAPED)
1220          */
1221         realthreads = pending = cputhreads = 0;
1222         for_each_td(td, i) {
1223                 int flags = 0;
1224
1225                 /*
1226                  * ->io_ops is NULL for a thread that has closed its
1227                  * io engine
1228                  */
1229                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1230                         cputhreads++;
1231                 else
1232                         realthreads++;
1233
1234                 if (!td->pid) {
1235                         pending++;
1236                         continue;
1237                 }
1238                 if (td->runstate == TD_REAPED)
1239                         continue;
1240                 if (td->o.use_thread) {
1241                         if (td->runstate == TD_EXITED) {
1242                                 td_set_runstate(td, TD_REAPED);
1243                                 goto reaped;
1244                         }
1245                         continue;
1246                 }
1247
1248                 flags = WNOHANG;
1249                 if (td->runstate == TD_EXITED)
1250                         flags = 0;
1251
1252                 /*
1253                  * check if someone quit or got killed in an unusual way
1254                  */
1255                 ret = waitpid(td->pid, &status, flags);
1256                 if (ret < 0) {
1257                         if (errno == ECHILD) {
1258                                 log_err("fio: pid=%d disappeared %d\n",
1259                                                 (int) td->pid, td->runstate);
1260                                 td_set_runstate(td, TD_REAPED);
1261                                 goto reaped;
1262                         }
1263                         perror("waitpid");
1264                 } else if (ret == td->pid) {
1265                         if (WIFSIGNALED(status)) {
1266                                 int sig = WTERMSIG(status);
1267
1268                                 if (sig != SIGTERM)
1269                                         log_err("fio: pid=%d, got signal=%d\n",
1270                                                         (int) td->pid, sig);
1271                                 td_set_runstate(td, TD_REAPED);
1272                                 goto reaped;
1273                         }
1274                         if (WIFEXITED(status)) {
1275                                 if (WEXITSTATUS(status) && !td->error)
1276                                         td->error = WEXITSTATUS(status);
1277
1278                                 td_set_runstate(td, TD_REAPED);
1279                                 goto reaped;
1280                         }
1281                 }
1282
1283                 /*
1284                  * thread is not dead, continue
1285                  */
1286                 pending++;
1287                 continue;
1288 reaped:
1289                 (*nr_running)--;
1290                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1291                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1292                 if (!td->pid)
1293                         pending--;
1294
1295                 if (td->error)
1296                         exit_value++;
1297
1298                 done_secs += mtime_since_now(&td->epoch) / 1000;
1299         }
1300
1301         if (*nr_running == cputhreads && !pending && realthreads)
1302                 fio_terminate_threads(TERMINATE_ALL);
1303 }
1304
1305
1306
1307 /*
1308  * Main function for kicking off and reaping jobs, as needed.
1309  */
1310 static void run_threads(void)
1311 {
1312         struct thread_data *td;
1313         unsigned long spent;
1314         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1315
1316         if (fio_pin_memory())
1317                 return;
1318
1319         if (fio_gtod_offload && fio_start_gtod_thread())
1320                 return;
1321
1322         set_sig_handlers();
1323
1324         if (!terse_output) {
1325                 log_info("Starting ");
1326                 if (nr_thread)
1327                         log_info("%d thread%s", nr_thread,
1328                                                 nr_thread > 1 ? "s" : "");
1329                 if (nr_process) {
1330                         if (nr_thread)
1331                                 log_info(" and ");
1332                         log_info("%d process%s", nr_process,
1333                                                 nr_process > 1 ? "es" : "");
1334                 }
1335                 log_info("\n");
1336                 fflush(stdout);
1337         }
1338
1339         todo = thread_number;
1340         nr_running = 0;
1341         nr_started = 0;
1342         m_rate = t_rate = 0;
1343
1344         for_each_td(td, i) {
1345                 print_status_init(td->thread_number - 1);
1346
1347                 if (!td->o.create_serialize)
1348                         continue;
1349
1350                 /*
1351                  * do file setup here so it happens sequentially,
1352                  * we don't want X number of threads getting their
1353                  * client data interspersed on disk
1354                  */
1355                 if (setup_files(td)) {
1356                         exit_value++;
1357                         if (td->error)
1358                                 log_err("fio: pid=%d, err=%d/%s\n",
1359                                         (int) td->pid, td->error, td->verror);
1360                         td_set_runstate(td, TD_REAPED);
1361                         todo--;
1362                 } else {
1363                         struct fio_file *f;
1364                         unsigned int j;
1365
1366                         /*
1367                          * for sharing to work, each job must always open
1368                          * its own files. so close them, if we opened them
1369                          * for creation
1370                          */
1371                         for_each_file(td, f, j) {
1372                                 if (fio_file_open(f))
1373                                         td_io_close_file(td, f);
1374                         }
1375                 }
1376         }
1377
1378         set_genesis_time();
1379
1380         while (todo) {
1381                 struct thread_data *map[REAL_MAX_JOBS];
1382                 struct timeval this_start;
1383                 int this_jobs = 0, left;
1384
1385                 /*
1386                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1387                  */
1388                 for_each_td(td, i) {
1389                         if (td->runstate != TD_NOT_CREATED)
1390                                 continue;
1391
1392                         /*
1393                          * never got a chance to start, killed by other
1394                          * thread for some reason
1395                          */
1396                         if (td->terminate) {
1397                                 todo--;
1398                                 continue;
1399                         }
1400
1401                         if (td->o.start_delay) {
1402                                 spent = mtime_since_genesis();
1403
1404                                 if (td->o.start_delay * 1000 > spent)
1405                                         continue;
1406                         }
1407
1408                         if (td->o.stonewall && (nr_started || nr_running)) {
1409                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1410                                                         td->o.name);
1411                                 break;
1412                         }
1413
1414                         init_disk_util(td);
1415
1416                         /*
1417                          * Set state to created. Thread will transition
1418                          * to TD_INITIALIZED when it's done setting up.
1419                          */
1420                         td_set_runstate(td, TD_CREATED);
1421                         map[this_jobs++] = td;
1422                         nr_started++;
1423
1424                         if (td->o.use_thread) {
1425                                 int ret;
1426
1427                                 dprint(FD_PROCESS, "will pthread_create\n");
1428                                 ret = pthread_create(&td->thread, NULL,
1429                                                         thread_main, td);
1430                                 if (ret) {
1431                                         log_err("pthread_create: %s\n",
1432                                                         strerror(ret));
1433                                         nr_started--;
1434                                         break;
1435                                 }
1436                                 ret = pthread_detach(td->thread);
1437                                 if (ret)
1438                                         log_err("pthread_detach: %s",
1439                                                         strerror(ret));
1440                         } else {
1441                                 pid_t pid;
1442                                 dprint(FD_PROCESS, "will fork\n");
1443                                 pid = fork();
1444                                 if (!pid) {
1445                                         int ret = fork_main(shm_id, i);
1446
1447                                         _exit(ret);
1448                                 } else if (i == fio_debug_jobno)
1449                                         *fio_debug_jobp = pid;
1450                         }
1451                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1452                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1453                                 log_err("fio: job startup hung? exiting.\n");
1454                                 fio_terminate_threads(TERMINATE_ALL);
1455                                 fio_abort = 1;
1456                                 nr_started--;
1457                                 break;
1458                         }
1459                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1460                 }
1461
1462                 /*
1463                  * Wait for the started threads to transition to
1464                  * TD_INITIALIZED.
1465                  */
1466                 fio_gettime(&this_start, NULL);
1467                 left = this_jobs;
1468                 while (left && !fio_abort) {
1469                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1470                                 break;
1471
1472                         usleep(100000);
1473
1474                         for (i = 0; i < this_jobs; i++) {
1475                                 td = map[i];
1476                                 if (!td)
1477                                         continue;
1478                                 if (td->runstate == TD_INITIALIZED) {
1479                                         map[i] = NULL;
1480                                         left--;
1481                                 } else if (td->runstate >= TD_EXITED) {
1482                                         map[i] = NULL;
1483                                         left--;
1484                                         todo--;
1485                                         nr_running++; /* work-around... */
1486                                 }
1487                         }
1488                 }
1489
1490                 if (left) {
1491                         log_err("fio: %d jobs failed to start\n", left);
1492                         for (i = 0; i < this_jobs; i++) {
1493                                 td = map[i];
1494                                 if (!td)
1495                                         continue;
1496                                 kill(td->pid, SIGTERM);
1497                         }
1498                         break;
1499                 }
1500
1501                 /*
1502                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1503                  */
1504                 for_each_td(td, i) {
1505                         if (td->runstate != TD_INITIALIZED)
1506                                 continue;
1507
1508                         if (in_ramp_time(td))
1509                                 td_set_runstate(td, TD_RAMP);
1510                         else
1511                                 td_set_runstate(td, TD_RUNNING);
1512                         nr_running++;
1513                         nr_started--;
1514                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1515                         t_rate += td->o.rate[0] + td->o.rate[1];
1516                         todo--;
1517                         fio_mutex_up(td->mutex);
1518                 }
1519
1520                 reap_threads(&nr_running, &t_rate, &m_rate);
1521
1522                 if (todo) {
1523                         if (is_backend)
1524                                 fio_server_idle_loop();
1525                         else
1526                                 usleep(100000);
1527                 }
1528         }
1529
1530         while (nr_running) {
1531                 reap_threads(&nr_running, &t_rate, &m_rate);
1532
1533                 if (is_backend)
1534                         fio_server_idle_loop();
1535                 else
1536                         usleep(10000);
1537         }
1538
1539         update_io_ticks();
1540         fio_unpin_memory();
1541 }
1542
1543 static void *disk_thread_main(void *data)
1544 {
1545         fio_mutex_up(startup_mutex);
1546
1547         while (threads) {
1548                 usleep(DISK_UTIL_MSEC * 1000);
1549                 if (!threads)
1550                         break;
1551                 update_io_ticks();
1552
1553                 if (!is_backend)
1554                         print_thread_status();
1555         }
1556
1557         return NULL;
1558 }
1559
1560 static int create_disk_util_thread(void)
1561 {
1562         int ret;
1563
1564         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1565         if (ret) {
1566                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1567                 return 1;
1568         }
1569
1570         ret = pthread_detach(disk_util_thread);
1571         if (ret) {
1572                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1573                 return 1;
1574         }
1575
1576         dprint(FD_MUTEX, "wait on startup_mutex\n");
1577         fio_mutex_down(startup_mutex);
1578         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1579         return 0;
1580 }
1581
1582
1583 int fio_backend(void)
1584 {
1585         struct thread_data *td;
1586         int i;
1587
1588         if (exec_profile) {
1589                 if (load_profile(exec_profile))
1590                         return 1;
1591                 free(exec_profile);
1592                 exec_profile = NULL;
1593         }
1594         if (!thread_number)
1595                 return 0;
1596
1597         if (write_bw_log) {
1598                 setup_log(&agg_io_log[DDIR_READ], 0);
1599                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1600         }
1601
1602         startup_mutex = fio_mutex_init(0);
1603         if (startup_mutex == NULL)
1604                 return 1;
1605         writeout_mutex = fio_mutex_init(1);
1606         if (writeout_mutex == NULL)
1607                 return 1;
1608
1609         set_genesis_time();
1610         create_disk_util_thread();
1611
1612         cgroup_list = smalloc(sizeof(*cgroup_list));
1613         INIT_FLIST_HEAD(cgroup_list);
1614
1615         run_threads();
1616
1617         if (!fio_abort) {
1618                 show_run_stats();
1619                 if (write_bw_log) {
1620                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1621                         __finish_log(agg_io_log[DDIR_WRITE],
1622                                         "agg-write_bw.log");
1623                 }
1624         }
1625
1626         for_each_td(td, i)
1627                 fio_options_free(td);
1628
1629         cgroup_kill(cgroup_list);
1630         sfree(cgroup_list);
1631         sfree(cgroup_mnt);
1632
1633         fio_mutex_remove(startup_mutex);
1634         fio_mutex_remove(writeout_mutex);
1635         return exit_value;
1636 }
1637
1638