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 <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38
  39 #include "fio.h"
  40 #ifndef FIO_NO_HAVE_SHM_H
  41 #include <sys/shm.h>
  42 #endif
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53 #include "lib/getrusage.h"
  54 #include "idletime.h"
  55 #include "err.h"
  56
  57 static pthread_t disk_util_thread;
  58 static struct fio_mutex *disk_thread_mutex;
  59 static struct fio_mutex *startup_mutex;
  60 static struct fio_mutex *writeout_mutex;
  61 static struct flist_head *cgroup_list;
  62 static char *cgroup_mnt;
  63 static int exit_value;
  64 static volatile int fio_abort;
  65 static unsigned int nr_process = 0;
  66 static unsigned int nr_thread = 0;
  67
  68 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  69
  70 int groupid = 0;
  71 unsigned int thread_number = 0;
  72 unsigned int stat_number = 0;
  73 int shm_id = 0;
  74 int temp_stall_ts;
  75 unsigned long done_secs = 0;
  76 volatile int disk_util_exit = 0;
  77
  78 #define PAGE_ALIGN(buf) \
  79         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  80
  81 #define JOB_START_TIMEOUT       (5 * 1000)
  82
  83 static void sig_int(int sig)
  84 {
  85         if (threads) {
  86                 if (is_backend)
  87                         fio_server_got_signal(sig);
  88                 else {
  89                         log_info("\nfio: terminating on signal %d\n", sig);
  90                         fflush(stdout);
  91                         exit_value = 128;
  92                 }
  93
  94                 fio_terminate_threads(TERMINATE_ALL);
  95         }
  96 }
  97
  98 static void sig_show_status(int sig)
  99 {
 100         show_running_run_stats();
 101 }
 102
 103 static void set_sig_handlers(void)
 104 {
 105         struct sigaction act;
 106
 107         memset(&act, 0, sizeof(act));
 108         act.sa_handler = sig_int;
 109         act.sa_flags = SA_RESTART;
 110         sigaction(SIGINT, &act, NULL);
 111
 112         memset(&act, 0, sizeof(act));
 113         act.sa_handler = sig_int;
 114         act.sa_flags = SA_RESTART;
 115         sigaction(SIGTERM, &act, NULL);
 116
 117 /* Windows uses SIGBREAK as a quit signal from other applications */
 118 #ifdef WIN32
 119         memset(&act, 0, sizeof(act));
 120         act.sa_handler = sig_int;
 121         act.sa_flags = SA_RESTART;
 122         sigaction(SIGBREAK, &act, NULL);
 123 #endif
 124
 125         memset(&act, 0, sizeof(act));
 126         act.sa_handler = sig_show_status;
 127         act.sa_flags = SA_RESTART;
 128         sigaction(SIGUSR1, &act, NULL);
 129
 130         if (is_backend) {
 131                 memset(&act, 0, sizeof(act));
 132                 act.sa_handler = sig_int;
 133                 act.sa_flags = SA_RESTART;
 134                 sigaction(SIGPIPE, &act, NULL);
 135         }
 136 }
 137
 138 /*
 139  * Check if we are above the minimum rate given.
 140  */
 141 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 142                             enum fio_ddir ddir)
 143 {
 144         unsigned long long bytes = 0;
 145         unsigned long iops = 0;
 146         unsigned long spent;
 147         unsigned long rate;
 148         unsigned int ratemin = 0;
 149         unsigned int rate_iops = 0;
 150         unsigned int rate_iops_min = 0;
 151
 152         assert(ddir_rw(ddir));
 153
 154         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 155                 return 0;
 156
 157         /*
 158          * allow a 2 second settle period in the beginning
 159          */
 160         if (mtime_since(&td->start, now) < 2000)
 161                 return 0;
 162
 163         iops += td->this_io_blocks[ddir];
 164         bytes += td->this_io_bytes[ddir];
 165         ratemin += td->o.ratemin[ddir];
 166         rate_iops += td->o.rate_iops[ddir];
 167         rate_iops_min += td->o.rate_iops_min[ddir];
 168
 169         /*
 170          * if rate blocks is set, sample is running
 171          */
 172         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 173                 spent = mtime_since(&td->lastrate[ddir], now);
 174                 if (spent < td->o.ratecycle)
 175                         return 0;
 176
 177                 if (td->o.rate[ddir]) {
 178                         /*
 179                          * check bandwidth specified rate
 180                          */
 181                         if (bytes < td->rate_bytes[ddir]) {
 182                                 log_err("%s: min rate %u not met\n", td->o.name,
 183                                                                 ratemin);
 184                                 return 1;
 185                         } else {
 186                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 187                                 if (rate < ratemin ||
 188                                     bytes < td->rate_bytes[ddir]) {
 189                                         log_err("%s: min rate %u not met, got"
 190                                                 " %luKB/sec\n", td->o.name,
 191                                                         ratemin, rate);
 192                                         return 1;
 193                                 }
 194                         }
 195                 } else {
 196                         /*
 197                          * checks iops specified rate
 198                          */
 199                         if (iops < rate_iops) {
 200                                 log_err("%s: min iops rate %u not met\n",
 201                                                 td->o.name, rate_iops);
 202                                 return 1;
 203                         } else {
 204                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 205                                 if (rate < rate_iops_min ||
 206                                     iops < td->rate_blocks[ddir]) {
 207                                         log_err("%s: min iops rate %u not met,"
 208                                                 " got %lu\n", td->o.name,
 209                                                         rate_iops_min, rate);
 210                                 }
 211                         }
 212                 }
 213         }
 214
 215         td->rate_bytes[ddir] = bytes;
 216         td->rate_blocks[ddir] = iops;
 217         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 218         return 0;
 219 }
 220
 221 static int check_min_rate(struct thread_data *td, struct timeval *now,
 222                           uint64_t *bytes_done)
 223 {
 224         int ret = 0;
 225
 226         if (bytes_done[DDIR_READ])
 227                 ret |= __check_min_rate(td, now, DDIR_READ);
 228         if (bytes_done[DDIR_WRITE])
 229                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 230         if (bytes_done[DDIR_TRIM])
 231                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 232
 233         return ret;
 234 }
 235
 236 /*
 237  * When job exits, we can cancel the in-flight IO if we are using async
 238  * io. Attempt to do so.
 239  */
 240 static void cleanup_pending_aio(struct thread_data *td)
 241 {
 242         int r;
 243
 244         /*
 245          * get immediately available events, if any
 246          */
 247         r = io_u_queued_complete(td, 0, NULL);
 248         if (r < 0)
 249                 return;
 250
 251         /*
 252          * now cancel remaining active events
 253          */
 254         if (td->io_ops->cancel) {
 255                 struct io_u *io_u;
 256                 int i;
 257
 258                 io_u_qiter(&td->io_u_all, io_u, i) {
 259                         if (io_u->flags & IO_U_F_FLIGHT) {
 260                                 r = td->io_ops->cancel(td, io_u);
 261                                 if (!r)
 262                                         put_io_u(td, io_u);
 263                         }
 264                 }
 265         }
 266
 267         if (td->cur_depth)
 268                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 269 }
 270
 271 /*
 272  * Helper to handle the final sync of a file. Works just like the normal
 273  * io path, just does everything sync.
 274  */
 275 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 276 {
 277         struct io_u *io_u = __get_io_u(td);
 278         int ret;
 279
 280         if (!io_u)
 281                 return 1;
 282
 283         io_u->ddir = DDIR_SYNC;
 284         io_u->file = f;
 285
 286         if (td_io_prep(td, io_u)) {
 287                 put_io_u(td, io_u);
 288                 return 1;
 289         }
 290
 291 requeue:
 292         ret = td_io_queue(td, io_u);
 293         if (ret < 0) {
 294                 td_verror(td, io_u->error, "td_io_queue");
 295                 put_io_u(td, io_u);
 296                 return 1;
 297         } else if (ret == FIO_Q_QUEUED) {
 298                 if (io_u_queued_complete(td, 1, NULL) < 0)
 299                         return 1;
 300         } else if (ret == FIO_Q_COMPLETED) {
 301                 if (io_u->error) {
 302                         td_verror(td, io_u->error, "td_io_queue");
 303                         return 1;
 304                 }
 305
 306                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 307                         return 1;
 308         } else if (ret == FIO_Q_BUSY) {
 309                 if (td_io_commit(td))
 310                         return 1;
 311                 goto requeue;
 312         }
 313
 314         return 0;
 315 }
 316
 317 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 318 {
 319         int ret;
 320
 321         if (fio_file_open(f))
 322                 return fio_io_sync(td, f);
 323
 324         if (td_io_open_file(td, f))
 325                 return 1;
 326
 327         ret = fio_io_sync(td, f);
 328         td_io_close_file(td, f);
 329         return ret;
 330 }
 331
 332 static inline void __update_tv_cache(struct thread_data *td)
 333 {
 334         fio_gettime(&td->tv_cache, NULL);
 335 }
 336
 337 static inline void update_tv_cache(struct thread_data *td)
 338 {
 339         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 340                 __update_tv_cache(td);
 341 }
 342
 343 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 344 {
 345         if (in_ramp_time(td))
 346                 return 0;
 347         if (!td->o.timeout)
 348                 return 0;
 349         if (utime_since(&td->epoch, t) >= td->o.timeout)
 350                 return 1;
 351
 352         return 0;
 353 }
 354
 355 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 356                                int *retptr)
 357 {
 358         int ret = *retptr;
 359
 360         if (ret < 0 || td->error) {
 361                 int err = td->error;
 362                 enum error_type_bit eb;
 363
 364                 if (ret < 0)
 365                         err = -ret;
 366
 367                 eb = td_error_type(ddir, err);
 368                 if (!(td->o.continue_on_error & (1 << eb)))
 369                         return 1;
 370
 371                 if (td_non_fatal_error(td, eb, err)) {
 372                         /*
 373                          * Continue with the I/Os in case of
 374                          * a non fatal error.
 375                          */
 376                         update_error_count(td, err);
 377                         td_clear_error(td);
 378                         *retptr = 0;
 379                         return 0;
 380                 } else if (td->o.fill_device && err == ENOSPC) {
 381                         /*
 382                          * We expect to hit this error if
 383                          * fill_device option is set.
 384                          */
 385                         td_clear_error(td);
 386                         td->terminate = 1;
 387                         return 1;
 388                 } else {
 389                         /*
 390                          * Stop the I/O in case of a fatal
 391                          * error.
 392                          */
 393                         update_error_count(td, err);
 394                         return 1;
 395                 }
 396         }
 397
 398         return 0;
 399 }
 400
 401 static void check_update_rusage(struct thread_data *td)
 402 {
 403         if (td->update_rusage) {
 404                 td->update_rusage = 0;
 405                 update_rusage_stat(td);
 406                 fio_mutex_up(td->rusage_sem);
 407         }
 408 }
 409
 410 /*
 411  * The main verify engine. Runs over the writes we previously submitted,
 412  * reads the blocks back in, and checks the crc/md5 of the data.
 413  */
 414 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 415 {
 416         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 417         struct fio_file *f;
 418         struct io_u *io_u;
 419         int ret, min_events;
 420         unsigned int i;
 421
 422         dprint(FD_VERIFY, "starting loop\n");
 423
 424         /*
 425          * sync io first and invalidate cache, to make sure we really
 426          * read from disk.
 427          */
 428         for_each_file(td, f, i) {
 429                 if (!fio_file_open(f))
 430                         continue;
 431                 if (fio_io_sync(td, f))
 432                         break;
 433                 if (file_invalidate_cache(td, f))
 434                         break;
 435         }
 436
 437         check_update_rusage(td);
 438
 439         if (td->error)
 440                 return;
 441
 442         td_set_runstate(td, TD_VERIFYING);
 443
 444         io_u = NULL;
 445         while (!td->terminate) {
 446                 enum fio_ddir ddir;
 447                 int ret2, full;
 448
 449                 update_tv_cache(td);
 450                 check_update_rusage(td);
 451
 452                 if (runtime_exceeded(td, &td->tv_cache)) {
 453                         __update_tv_cache(td);
 454                         if (runtime_exceeded(td, &td->tv_cache)) {
 455                                 td->terminate = 1;
 456                                 break;
 457                         }
 458                 }
 459
 460                 if (flow_threshold_exceeded(td))
 461                         continue;
 462
 463                 if (!td->o.experimental_verify) {
 464                         io_u = __get_io_u(td);
 465                         if (!io_u)
 466                                 break;
 467
 468                         if (get_next_verify(td, io_u)) {
 469                                 put_io_u(td, io_u);
 470                                 break;
 471                         }
 472
 473                         if (td_io_prep(td, io_u)) {
 474                                 put_io_u(td, io_u);
 475                                 break;
 476                         }
 477                 } else {
 478                         if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
 479                                 break;
 480
 481                         while ((io_u = get_io_u(td)) != NULL) {
 482                                 if (IS_ERR(io_u)) {
 483                                         io_u = NULL;
 484                                         ret = FIO_Q_BUSY;
 485                                         goto reap;
 486                                 }
 487
 488                                 /*
 489                                  * We are only interested in the places where
 490                                  * we wrote or trimmed IOs. Turn those into
 491                                  * reads for verification purposes.
 492                                  */
 493                                 if (io_u->ddir == DDIR_READ) {
 494                                         /*
 495                                          * Pretend we issued it for rwmix
 496                                          * accounting
 497                                          */
 498                                         td->io_issues[DDIR_READ]++;
 499                                         put_io_u(td, io_u);
 500                                         continue;
 501                                 } else if (io_u->ddir == DDIR_TRIM) {
 502                                         io_u->ddir = DDIR_READ;
 503                                         io_u->flags |= IO_U_F_TRIMMED;
 504                                         break;
 505                                 } else if (io_u->ddir == DDIR_WRITE) {
 506                                         io_u->ddir = DDIR_READ;
 507                                         break;
 508                                 } else {
 509                                         put_io_u(td, io_u);
 510                                         continue;
 511                                 }
 512                         }
 513
 514                         if (!io_u)
 515                                 break;
 516                 }
 517
 518                 if (td->o.verify_async)
 519                         io_u->end_io = verify_io_u_async;
 520                 else
 521                         io_u->end_io = verify_io_u;
 522
 523                 ddir = io_u->ddir;
 524
 525                 ret = td_io_queue(td, io_u);
 526                 switch (ret) {
 527                 case FIO_Q_COMPLETED:
 528                         if (io_u->error) {
 529                                 ret = -io_u->error;
 530                                 clear_io_u(td, io_u);
 531                         } else if (io_u->resid) {
 532                                 int bytes = io_u->xfer_buflen - io_u->resid;
 533
 534                                 /*
 535                                  * zero read, fail
 536                                  */
 537                                 if (!bytes) {
 538                                         td_verror(td, EIO, "full resid");
 539                                         put_io_u(td, io_u);
 540                                         break;
 541                                 }
 542
 543                                 io_u->xfer_buflen = io_u->resid;
 544                                 io_u->xfer_buf += bytes;
 545                                 io_u->offset += bytes;
 546
 547                                 if (ddir_rw(io_u->ddir))
 548                                         td->ts.short_io_u[io_u->ddir]++;
 549
 550                                 f = io_u->file;
 551                                 if (io_u->offset == f->real_file_size)
 552                                         goto sync_done;
 553
 554                                 requeue_io_u(td, &io_u);
 555                         } else {
 556 sync_done:
 557                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 558                                 if (ret < 0)
 559                                         break;
 560                         }
 561                         continue;
 562                 case FIO_Q_QUEUED:
 563                         break;
 564                 case FIO_Q_BUSY:
 565                         requeue_io_u(td, &io_u);
 566                         ret2 = td_io_commit(td);
 567                         if (ret2 < 0)
 568                                 ret = ret2;
 569                         break;
 570                 default:
 571                         assert(ret < 0);
 572                         td_verror(td, -ret, "td_io_queue");
 573                         break;
 574                 }
 575
 576                 if (break_on_this_error(td, ddir, &ret))
 577                         break;
 578
 579                 /*
 580                  * if we can queue more, do so. but check if there are
 581                  * completed io_u's first. Note that we can get BUSY even
 582                  * without IO queued, if the system is resource starved.
 583                  */
 584 reap:
 585                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 586                 if (full || !td->o.iodepth_batch_complete) {
 587                         min_events = min(td->o.iodepth_batch_complete,
 588                                          td->cur_depth);
 589                         /*
 590                          * if the queue is full, we MUST reap at least 1 event
 591                          */
 592                         if (full && !min_events)
 593                                 min_events = 1;
 594
 595                         do {
 596                                 /*
 597                                  * Reap required number of io units, if any,
 598                                  * and do the verification on them through
 599                                  * the callback handler
 600                                  */
 601                                 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
 602                                         ret = -1;
 603                                         break;
 604                                 }
 605                         } while (full && (td->cur_depth > td->o.iodepth_low));
 606                 }
 607                 if (ret < 0)
 608                         break;
 609         }
 610
 611         check_update_rusage(td);
 612
 613         if (!td->error) {
 614                 min_events = td->cur_depth;
 615
 616                 if (min_events)
 617                         ret = io_u_queued_complete(td, min_events, NULL);
 618         } else
 619                 cleanup_pending_aio(td);
 620
 621         td_set_runstate(td, TD_RUNNING);
 622
 623         dprint(FD_VERIFY, "exiting loop\n");
 624 }
 625
 626 static int io_bytes_exceeded(struct thread_data *td)
 627 {
 628         unsigned long long number_ios = 0;
 629         unsigned long long bytes;
 630
 631         if (td_rw(td))
 632                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 633         else if (td_write(td))
 634                 bytes = td->this_io_bytes[DDIR_WRITE];
 635         else if (td_read(td))
 636                 bytes = td->this_io_bytes[DDIR_READ];
 637         else
 638                 bytes = td->this_io_bytes[DDIR_TRIM];
 639
 640         if (td->o.number_ios) {
 641                 number_ios = ddir_rw_sum(td->this_io_blocks);
 642                 number_ios += td->io_u_queued + td->io_u_in_flight;
 643         }
 644
 645         return bytes >= td->o.size ||
 646                 (number_ios && number_ios >= td->o.number_ios);
 647 }
 648
 649 /*
 650  * Main IO worker function. It retrieves io_u's to process and queues
 651  * and reaps them, checking for rate and errors along the way.
 652  *
 653  * Returns number of bytes written and trimmed.
 654  */
 655 static uint64_t do_io(struct thread_data *td)
 656 {
 657         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 658         unsigned int i;
 659         int ret = 0;
 660         uint64_t total_bytes, bytes_issued = 0;
 661
 662         if (in_ramp_time(td))
 663                 td_set_runstate(td, TD_RAMP);
 664         else
 665                 td_set_runstate(td, TD_RUNNING);
 666
 667         lat_target_init(td);
 668
 669         /*
 670          * If verify_backlog is enabled, we'll run the verify in this
 671          * handler as well. For that case, we may need up to twice the
 672          * amount of bytes.
 673          */
 674         total_bytes = td->o.size;
 675         if (td->o.verify != VERIFY_NONE &&
 676            (td_write(td) && td->o.verify_backlog))
 677                 total_bytes += td->o.size;
 678
 679         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 680                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 681                 td->o.time_based) {
 682                 struct timeval comp_time;
 683                 int min_evts = 0;
 684                 struct io_u *io_u;
 685                 int ret2, full;
 686                 enum fio_ddir ddir;
 687
 688                 check_update_rusage(td);
 689
 690                 if (td->terminate || td->done)
 691                         break;
 692
 693                 update_tv_cache(td);
 694
 695                 if (runtime_exceeded(td, &td->tv_cache)) {
 696                         __update_tv_cache(td);
 697                         if (runtime_exceeded(td, &td->tv_cache)) {
 698                                 td->terminate = 1;
 699                                 break;
 700                         }
 701                 }
 702
 703                 if (flow_threshold_exceeded(td))
 704                         continue;
 705
 706                 if (bytes_issued >= total_bytes)
 707                         break;
 708
 709                 io_u = get_io_u(td);
 710                 if (IS_ERR_OR_NULL(io_u)) {
 711                         int err = PTR_ERR(io_u);
 712
 713                         io_u = NULL;
 714                         if (err == -EBUSY) {
 715                                 ret = FIO_Q_BUSY;
 716                                 goto reap;
 717                         }
 718                         if (td->o.latency_target)
 719                                 goto reap;
 720                         break;
 721                 }
 722
 723                 ddir = io_u->ddir;
 724
 725                 /*
 726                  * Add verification end_io handler if:
 727                  *      - Asked to verify (!td_rw(td))
 728                  *      - Or the io_u is from our verify list (mixed write/ver)
 729                  */
 730                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 731                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 732
 733                         if (!td->o.verify_pattern_bytes) {
 734                                 io_u->rand_seed = __rand(&td->__verify_state);
 735                                 if (sizeof(int) != sizeof(long *))
 736                                         io_u->rand_seed *= __rand(&td->__verify_state);
 737                         }
 738
 739                         if (td->o.verify_async)
 740                                 io_u->end_io = verify_io_u_async;
 741                         else
 742                                 io_u->end_io = verify_io_u;
 743                         td_set_runstate(td, TD_VERIFYING);
 744                 } else if (in_ramp_time(td))
 745                         td_set_runstate(td, TD_RAMP);
 746                 else
 747                         td_set_runstate(td, TD_RUNNING);
 748
 749                 /*
 750                  * Always log IO before it's issued, so we know the specific
 751                  * order of it. The logged unit will track when the IO has
 752                  * completed.
 753                  */
 754                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
 755                     td->o.do_verify &&
 756                     td->o.verify != VERIFY_NONE &&
 757                     !td->o.experimental_verify)
 758                         log_io_piece(td, io_u);
 759
 760                 ret = td_io_queue(td, io_u);
 761                 switch (ret) {
 762                 case FIO_Q_COMPLETED:
 763                         if (io_u->error) {
 764                                 ret = -io_u->error;
 765                                 clear_io_u(td, io_u);
 766                         } else if (io_u->resid) {
 767                                 int bytes = io_u->xfer_buflen - io_u->resid;
 768                                 struct fio_file *f = io_u->file;
 769
 770                                 bytes_issued += bytes;
 771                                 /*
 772                                  * zero read, fail
 773                                  */
 774                                 if (!bytes) {
 775                                         td_verror(td, EIO, "full resid");
 776                                         put_io_u(td, io_u);
 777                                         break;
 778                                 }
 779
 780                                 io_u->xfer_buflen = io_u->resid;
 781                                 io_u->xfer_buf += bytes;
 782                                 io_u->offset += bytes;
 783
 784                                 if (ddir_rw(io_u->ddir))
 785                                         td->ts.short_io_u[io_u->ddir]++;
 786
 787                                 if (io_u->offset == f->real_file_size)
 788                                         goto sync_done;
 789
 790                                 requeue_io_u(td, &io_u);
 791                         } else {
 792 sync_done:
 793                                 if (__should_check_rate(td, DDIR_READ) ||
 794                                     __should_check_rate(td, DDIR_WRITE) ||
 795                                     __should_check_rate(td, DDIR_TRIM))
 796                                         fio_gettime(&comp_time, NULL);
 797
 798                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 799                                 if (ret < 0)
 800                                         break;
 801                                 bytes_issued += io_u->xfer_buflen;
 802                         }
 803                         break;
 804                 case FIO_Q_QUEUED:
 805                         /*
 806                          * if the engine doesn't have a commit hook,
 807                          * the io_u is really queued. if it does have such
 808                          * a hook, it has to call io_u_queued() itself.
 809                          */
 810                         if (td->io_ops->commit == NULL)
 811                                 io_u_queued(td, io_u);
 812                         bytes_issued += io_u->xfer_buflen;
 813                         break;
 814                 case FIO_Q_BUSY:
 815                         requeue_io_u(td, &io_u);
 816                         ret2 = td_io_commit(td);
 817                         if (ret2 < 0)
 818                                 ret = ret2;
 819                         break;
 820                 default:
 821                         assert(ret < 0);
 822                         put_io_u(td, io_u);
 823                         break;
 824                 }
 825
 826                 if (break_on_this_error(td, ddir, &ret))
 827                         break;
 828
 829                 /*
 830                  * See if we need to complete some commands. Note that we
 831                  * can get BUSY even without IO queued, if the system is
 832                  * resource starved.
 833                  */
 834 reap:
 835                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 836                 if (full || !td->o.iodepth_batch_complete) {
 837                         min_evts = min(td->o.iodepth_batch_complete,
 838                                         td->cur_depth);
 839                         /*
 840                          * if the queue is full, we MUST reap at least 1 event
 841                          */
 842                         if (full && !min_evts)
 843                                 min_evts = 1;
 844
 845                         if (__should_check_rate(td, DDIR_READ) ||
 846                             __should_check_rate(td, DDIR_WRITE) ||
 847                             __should_check_rate(td, DDIR_TRIM))
 848                                 fio_gettime(&comp_time, NULL);
 849
 850                         do {
 851                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 852                                 if (ret < 0)
 853                                         break;
 854
 855                         } while (full && (td->cur_depth > td->o.iodepth_low));
 856                 }
 857
 858                 if (ret < 0)
 859                         break;
 860                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 861                         continue;
 862
 863                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 864                         if (check_min_rate(td, &comp_time, bytes_done)) {
 865                                 if (exitall_on_terminate)
 866                                         fio_terminate_threads(td->groupid);
 867                                 td_verror(td, EIO, "check_min_rate");
 868                                 break;
 869                         }
 870                 }
 871                 if (!in_ramp_time(td) && td->o.latency_target)
 872                         lat_target_check(td);
 873
 874                 if (td->o.thinktime) {
 875                         unsigned long long b;
 876
 877                         b = ddir_rw_sum(td->io_blocks);
 878                         if (!(b % td->o.thinktime_blocks)) {
 879                                 int left;
 880
 881                                 io_u_quiesce(td);
 882
 883                                 if (td->o.thinktime_spin)
 884                                         usec_spin(td->o.thinktime_spin);
 885
 886                                 left = td->o.thinktime - td->o.thinktime_spin;
 887                                 if (left)
 888                                         usec_sleep(td, left);
 889                         }
 890                 }
 891         }
 892
 893         check_update_rusage(td);
 894
 895         if (td->trim_entries)
 896                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
 897
 898         if (td->o.fill_device && td->error == ENOSPC) {
 899                 td->error = 0;
 900                 td->terminate = 1;
 901         }
 902         if (!td->error) {
 903                 struct fio_file *f;
 904
 905                 i = td->cur_depth;
 906                 if (i) {
 907                         ret = io_u_queued_complete(td, i, bytes_done);
 908                         if (td->o.fill_device && td->error == ENOSPC)
 909                                 td->error = 0;
 910                 }
 911
 912                 if (should_fsync(td) && td->o.end_fsync) {
 913                         td_set_runstate(td, TD_FSYNCING);
 914
 915                         for_each_file(td, f, i) {
 916                                 if (!fio_file_fsync(td, f))
 917                                         continue;
 918
 919                                 log_err("fio: end_fsync failed for file %s\n",
 920                                                                 f->file_name);
 921                         }
 922                 }
 923         } else
 924                 cleanup_pending_aio(td);
 925
 926         /*
 927          * stop job if we failed doing any IO
 928          */
 929         if (!ddir_rw_sum(td->this_io_bytes))
 930                 td->done = 1;
 931
 932         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 933 }
 934
 935 static void cleanup_io_u(struct thread_data *td)
 936 {
 937         struct io_u *io_u;
 938
 939         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 940
 941                 if (td->io_ops->io_u_free)
 942                         td->io_ops->io_u_free(td, io_u);
 943
 944                 fio_memfree(io_u, sizeof(*io_u));
 945         }
 946
 947         free_io_mem(td);
 948
 949         io_u_rexit(&td->io_u_requeues);
 950         io_u_qexit(&td->io_u_freelist);
 951         io_u_qexit(&td->io_u_all);
 952 }
 953
 954 static int init_io_u(struct thread_data *td)
 955 {
 956         struct io_u *io_u;
 957         unsigned int max_bs, min_write;
 958         int cl_align, i, max_units;
 959         int data_xfer = 1, err;
 960         char *p;
 961
 962         max_units = td->o.iodepth;
 963         max_bs = td_max_bs(td);
 964         min_write = td->o.min_bs[DDIR_WRITE];
 965         td->orig_buffer_size = (unsigned long long) max_bs
 966                                         * (unsigned long long) max_units;
 967
 968         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 969                 data_xfer = 0;
 970
 971         err = 0;
 972         err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
 973         err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
 974         err += io_u_qinit(&td->io_u_all, td->o.iodepth);
 975
 976         if (err) {
 977                 log_err("fio: failed setting up IO queues\n");
 978                 return 1;
 979         }
 980
 981         /*
 982          * if we may later need to do address alignment, then add any
 983          * possible adjustment here so that we don't cause a buffer
 984          * overflow later. this adjustment may be too much if we get
 985          * lucky and the allocator gives us an aligned address.
 986          */
 987         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
 988             (td->io_ops->flags & FIO_RAWIO))
 989                 td->orig_buffer_size += page_mask + td->o.mem_align;
 990
 991         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 992                 unsigned long bs;
 993
 994                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 995                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 996         }
 997
 998         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 999                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1000                 return 1;
1001         }
1002
1003         if (data_xfer && allocate_io_mem(td))
1004                 return 1;
1005
1006         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1007             (td->io_ops->flags & FIO_RAWIO))
1008                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1009         else
1010                 p = td->orig_buffer;
1011
1012         cl_align = os_cache_line_size();
1013
1014         for (i = 0; i < max_units; i++) {
1015                 void *ptr;
1016
1017                 if (td->terminate)
1018                         return 1;
1019
1020                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1021                 if (!ptr) {
1022                         log_err("fio: unable to allocate aligned memory\n");
1023                         break;
1024                 }
1025
1026                 io_u = ptr;
1027                 memset(io_u, 0, sizeof(*io_u));
1028                 INIT_FLIST_HEAD(&io_u->verify_list);
1029                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1030
1031                 if (data_xfer) {
1032                         io_u->buf = p;
1033                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1034
1035                         if (td_write(td))
1036                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1037                         if (td_write(td) && td->o.verify_pattern_bytes) {
1038                                 /*
1039                                  * Fill the buffer with the pattern if we are
1040                                  * going to be doing writes.
1041                                  */
1042                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1043                         }
1044                 }
1045
1046                 io_u->index = i;
1047                 io_u->flags = IO_U_F_FREE;
1048                 io_u_qpush(&td->io_u_freelist, io_u);
1049
1050                 /*
1051                  * io_u never leaves this stack, used for iteration of all
1052                  * io_u buffers.
1053                  */
1054                 io_u_qpush(&td->io_u_all, io_u);
1055
1056                 if (td->io_ops->io_u_init) {
1057                         int ret = td->io_ops->io_u_init(td, io_u);
1058
1059                         if (ret) {
1060                                 log_err("fio: failed to init engine data: %d\n", ret);
1061                                 return 1;
1062                         }
1063                 }
1064
1065                 p += max_bs;
1066         }
1067
1068         return 0;
1069 }
1070
1071 static int switch_ioscheduler(struct thread_data *td)
1072 {
1073         char tmp[256], tmp2[128];
1074         FILE *f;
1075         int ret;
1076
1077         if (td->io_ops->flags & FIO_DISKLESSIO)
1078                 return 0;
1079
1080         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1081
1082         f = fopen(tmp, "r+");
1083         if (!f) {
1084                 if (errno == ENOENT) {
1085                         log_err("fio: os or kernel doesn't support IO scheduler"
1086                                 " switching\n");
1087                         return 0;
1088                 }
1089                 td_verror(td, errno, "fopen iosched");
1090                 return 1;
1091         }
1092
1093         /*
1094          * Set io scheduler.
1095          */
1096         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1097         if (ferror(f) || ret != 1) {
1098                 td_verror(td, errno, "fwrite");
1099                 fclose(f);
1100                 return 1;
1101         }
1102
1103         rewind(f);
1104
1105         /*
1106          * Read back and check that the selected scheduler is now the default.
1107          */
1108         ret = fread(tmp, 1, sizeof(tmp), f);
1109         if (ferror(f) || ret < 0) {
1110                 td_verror(td, errno, "fread");
1111                 fclose(f);
1112                 return 1;
1113         }
1114
1115         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1116         if (!strstr(tmp, tmp2)) {
1117                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1118                 td_verror(td, EINVAL, "iosched_switch");
1119                 fclose(f);
1120                 return 1;
1121         }
1122
1123         fclose(f);
1124         return 0;
1125 }
1126
1127 static int keep_running(struct thread_data *td)
1128 {
1129         if (td->done)
1130                 return 0;
1131         if (td->o.time_based)
1132                 return 1;
1133         if (td->o.loops) {
1134                 td->o.loops--;
1135                 return 1;
1136         }
1137
1138         if (td->o.number_ios) {
1139                 unsigned long long number_ios = ddir_rw_sum(td->this_io_blocks);
1140
1141                 number_ios += td->io_u_queued + td->io_u_in_flight;
1142                 if (number_ios >= td->o.number_ios)
1143                         return 0;
1144         }
1145
1146         if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1147                 uint64_t diff;
1148
1149                 /*
1150                  * If the difference is less than the minimum IO size, we
1151                  * are done.
1152                  */
1153                 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1154                 if (diff < td_max_bs(td))
1155                         return 0;
1156
1157                 if (fio_files_done(td))
1158                         return 0;
1159
1160                 return 1;
1161         }
1162
1163         return 0;
1164 }
1165
1166 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1167 {
1168         int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1169         char *str;
1170
1171         str = malloc(newlen);
1172         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1173
1174         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1175         ret = system(str);
1176         if (ret == -1)
1177                 log_err("fio: exec of cmd <%s> failed\n", str);
1178
1179         free(str);
1180         return ret;
1181 }
1182
1183 /*
1184  * Dry run to compute correct state of numberio for verification.
1185  */
1186 static uint64_t do_dry_run(struct thread_data *td)
1187 {
1188         uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1189
1190         td_set_runstate(td, TD_RUNNING);
1191
1192         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1193                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1194                 struct io_u *io_u;
1195                 int ret;
1196
1197                 if (td->terminate || td->done)
1198                         break;
1199
1200                 io_u = get_io_u(td);
1201                 if (!io_u)
1202                         break;
1203
1204                 io_u->flags |= IO_U_F_FLIGHT;
1205                 io_u->error = 0;
1206                 io_u->resid = 0;
1207                 if (ddir_rw(acct_ddir(io_u)))
1208                         td->io_issues[acct_ddir(io_u)]++;
1209                 if (ddir_rw(io_u->ddir)) {
1210                         io_u_mark_depth(td, 1);
1211                         td->ts.total_io_u[io_u->ddir]++;
1212                 }
1213
1214                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1215                     td->o.do_verify &&
1216                     td->o.verify != VERIFY_NONE &&
1217                     !td->o.experimental_verify)
1218                         log_io_piece(td, io_u);
1219
1220                 ret = io_u_sync_complete(td, io_u, bytes_done);
1221                 (void) ret;
1222         }
1223
1224         return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1225 }
1226
1227 /*
1228  * Entry point for the thread based jobs. The process based jobs end up
1229  * here as well, after a little setup.
1230  */
1231 static void *thread_main(void *data)
1232 {
1233         unsigned long long elapsed;
1234         struct thread_data *td = data;
1235         struct thread_options *o = &td->o;
1236         pthread_condattr_t attr;
1237         int clear_state;
1238         int ret;
1239
1240         if (!o->use_thread) {
1241                 setsid();
1242                 td->pid = getpid();
1243         } else
1244                 td->pid = gettid();
1245
1246         /*
1247          * fio_time_init() may not have been called yet if running as a server
1248          */
1249         fio_time_init();
1250
1251         fio_local_clock_init(o->use_thread);
1252
1253         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1254
1255         if (is_backend)
1256                 fio_server_send_start(td);
1257
1258         INIT_FLIST_HEAD(&td->io_log_list);
1259         INIT_FLIST_HEAD(&td->io_hist_list);
1260         INIT_FLIST_HEAD(&td->verify_list);
1261         INIT_FLIST_HEAD(&td->trim_list);
1262         INIT_FLIST_HEAD(&td->next_rand_list);
1263         pthread_mutex_init(&td->io_u_lock, NULL);
1264         td->io_hist_tree = RB_ROOT;
1265
1266         pthread_condattr_init(&attr);
1267         pthread_cond_init(&td->verify_cond, &attr);
1268         pthread_cond_init(&td->free_cond, &attr);
1269
1270         td_set_runstate(td, TD_INITIALIZED);
1271         dprint(FD_MUTEX, "up startup_mutex\n");
1272         fio_mutex_up(startup_mutex);
1273         dprint(FD_MUTEX, "wait on td->mutex\n");
1274         fio_mutex_down(td->mutex);
1275         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1276
1277         /*
1278          * A new gid requires privilege, so we need to do this before setting
1279          * the uid.
1280          */
1281         if (o->gid != -1U && setgid(o->gid)) {
1282                 td_verror(td, errno, "setgid");
1283                 goto err;
1284         }
1285         if (o->uid != -1U && setuid(o->uid)) {
1286                 td_verror(td, errno, "setuid");
1287                 goto err;
1288         }
1289
1290         /*
1291          * If we have a gettimeofday() thread, make sure we exclude that
1292          * thread from this job
1293          */
1294         if (o->gtod_cpu)
1295                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1296
1297         /*
1298          * Set affinity first, in case it has an impact on the memory
1299          * allocations.
1300          */
1301         if (o->cpumask_set) {
1302                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1303                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1304                         if (!ret) {
1305                                 log_err("fio: no CPUs set\n");
1306                                 log_err("fio: Try increasing number of available CPUs\n");
1307                                 td_verror(td, EINVAL, "cpus_split");
1308                                 goto err;
1309                         }
1310                 }
1311                 ret = fio_setaffinity(td->pid, o->cpumask);
1312                 if (ret == -1) {
1313                         td_verror(td, errno, "cpu_set_affinity");
1314                         goto err;
1315                 }
1316         }
1317
1318 #ifdef CONFIG_LIBNUMA
1319         /* numa node setup */
1320         if (o->numa_cpumask_set || o->numa_memmask_set) {
1321                 int ret;
1322
1323                 if (numa_available() < 0) {
1324                         td_verror(td, errno, "Does not support NUMA API\n");
1325                         goto err;
1326                 }
1327
1328                 if (o->numa_cpumask_set) {
1329                         ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1330                         if (ret == -1) {
1331                                 td_verror(td, errno, \
1332                                         "numa_run_on_node_mask failed\n");
1333                                 goto err;
1334                         }
1335                 }
1336
1337                 if (o->numa_memmask_set) {
1338
1339                         switch (o->numa_mem_mode) {
1340                         case MPOL_INTERLEAVE:
1341                                 numa_set_interleave_mask(o->numa_memnodesmask);
1342                                 break;
1343                         case MPOL_BIND:
1344                                 numa_set_membind(o->numa_memnodesmask);
1345                                 break;
1346                         case MPOL_LOCAL:
1347                                 numa_set_localalloc();
1348                                 break;
1349                         case MPOL_PREFERRED:
1350                                 numa_set_preferred(o->numa_mem_prefer_node);
1351                                 break;
1352                         case MPOL_DEFAULT:
1353                         default:
1354                                 break;
1355                         }
1356
1357                 }
1358         }
1359 #endif
1360
1361         if (fio_pin_memory(td))
1362                 goto err;
1363
1364         /*
1365          * May alter parameters that init_io_u() will use, so we need to
1366          * do this first.
1367          */
1368         if (init_iolog(td))
1369                 goto err;
1370
1371         if (init_io_u(td))
1372                 goto err;
1373
1374         if (o->verify_async && verify_async_init(td))
1375                 goto err;
1376
1377         if (o->ioprio) {
1378                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1379                 if (ret == -1) {
1380                         td_verror(td, errno, "ioprio_set");
1381                         goto err;
1382                 }
1383         }
1384
1385         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1386                 goto err;
1387
1388         errno = 0;
1389         if (nice(o->nice) == -1 && errno != 0) {
1390                 td_verror(td, errno, "nice");
1391                 goto err;
1392         }
1393
1394         if (o->ioscheduler && switch_ioscheduler(td))
1395                 goto err;
1396
1397         if (!o->create_serialize && setup_files(td))
1398                 goto err;
1399
1400         if (td_io_init(td))
1401                 goto err;
1402
1403         if (init_random_map(td))
1404                 goto err;
1405
1406         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1407                 goto err;
1408
1409         if (o->pre_read) {
1410                 if (pre_read_files(td) < 0)
1411                         goto err;
1412         }
1413
1414         fio_verify_init(td);
1415
1416         fio_gettime(&td->epoch, NULL);
1417         fio_getrusage(&td->ru_start);
1418         clear_state = 0;
1419         while (keep_running(td)) {
1420                 uint64_t verify_bytes;
1421
1422                 fio_gettime(&td->start, NULL);
1423                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1424                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1425                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1426
1427                 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1428                                 o->ratemin[DDIR_TRIM]) {
1429                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1430                                                 sizeof(td->bw_sample_time));
1431                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1432                                                 sizeof(td->bw_sample_time));
1433                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1434                                                 sizeof(td->bw_sample_time));
1435                 }
1436
1437                 if (clear_state)
1438                         clear_io_state(td);
1439
1440                 prune_io_piece_log(td);
1441
1442                 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1443                         verify_bytes = do_dry_run(td);
1444                 else
1445                         verify_bytes = do_io(td);
1446
1447                 clear_state = 1;
1448
1449                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1450                         elapsed = utime_since_now(&td->start);
1451                         td->ts.runtime[DDIR_READ] += elapsed;
1452                 }
1453                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1454                         elapsed = utime_since_now(&td->start);
1455                         td->ts.runtime[DDIR_WRITE] += elapsed;
1456                 }
1457                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1458                         elapsed = utime_since_now(&td->start);
1459                         td->ts.runtime[DDIR_TRIM] += elapsed;
1460                 }
1461
1462                 if (td->error || td->terminate)
1463                         break;
1464
1465                 if (!o->do_verify ||
1466                     o->verify == VERIFY_NONE ||
1467                     (td->io_ops->flags & FIO_UNIDIR))
1468                         continue;
1469
1470                 clear_io_state(td);
1471
1472                 fio_gettime(&td->start, NULL);
1473
1474                 do_verify(td, verify_bytes);
1475
1476                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1477
1478                 if (td->error || td->terminate)
1479                         break;
1480         }
1481
1482         update_rusage_stat(td);
1483         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1484         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1485         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1486         td->ts.total_run_time = mtime_since_now(&td->epoch);
1487         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1488         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1489         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1490
1491         fio_unpin_memory(td);
1492
1493         fio_mutex_down(writeout_mutex);
1494         finalize_logs(td);
1495         if (td->bw_log) {
1496                 if (o->bw_log_file) {
1497                         finish_log_named(td, td->bw_log,
1498                                                 o->bw_log_file, "bw");
1499                 } else
1500                         finish_log(td, td->bw_log, "bw");
1501         }
1502         if (td->lat_log) {
1503                 if (o->lat_log_file) {
1504                         finish_log_named(td, td->lat_log,
1505                                                 o->lat_log_file, "lat");
1506                 } else
1507                         finish_log(td, td->lat_log, "lat");
1508         }
1509         if (td->slat_log) {
1510                 if (o->lat_log_file) {
1511                         finish_log_named(td, td->slat_log,
1512                                                 o->lat_log_file, "slat");
1513                 } else
1514                         finish_log(td, td->slat_log, "slat");
1515         }
1516         if (td->clat_log) {
1517                 if (o->lat_log_file) {
1518                         finish_log_named(td, td->clat_log,
1519                                                 o->lat_log_file, "clat");
1520                 } else
1521                         finish_log(td, td->clat_log, "clat");
1522         }
1523         if (td->iops_log) {
1524                 if (o->iops_log_file) {
1525                         finish_log_named(td, td->iops_log,
1526                                                 o->iops_log_file, "iops");
1527                 } else
1528                         finish_log(td, td->iops_log, "iops");
1529         }
1530
1531         fio_mutex_up(writeout_mutex);
1532         if (o->exec_postrun)
1533                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1534
1535         if (exitall_on_terminate)
1536                 fio_terminate_threads(td->groupid);
1537
1538 err:
1539         if (td->error)
1540                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1541                                                         td->verror);
1542
1543         if (o->verify_async)
1544                 verify_async_exit(td);
1545
1546         close_and_free_files(td);
1547         cleanup_io_u(td);
1548         close_ioengine(td);
1549         cgroup_shutdown(td, &cgroup_mnt);
1550
1551         if (o->cpumask_set) {
1552                 int ret = fio_cpuset_exit(&o->cpumask);
1553
1554                 td_verror(td, ret, "fio_cpuset_exit");
1555         }
1556
1557         /*
1558          * do this very late, it will log file closing as well
1559          */
1560         if (o->write_iolog_file)
1561                 write_iolog_close(td);
1562
1563         fio_mutex_remove(td->rusage_sem);
1564         td->rusage_sem = NULL;
1565
1566         fio_mutex_remove(td->mutex);
1567         td->mutex = NULL;
1568
1569         td_set_runstate(td, TD_EXITED);
1570         return (void *) (uintptr_t) td->error;
1571 }
1572
1573
1574 /*
1575  * We cannot pass the td data into a forked process, so attach the td and
1576  * pass it to the thread worker.
1577  */
1578 static int fork_main(int shmid, int offset)
1579 {
1580         struct thread_data *td;
1581         void *data, *ret;
1582
1583 #ifndef __hpux
1584         data = shmat(shmid, NULL, 0);
1585         if (data == (void *) -1) {
1586                 int __err = errno;
1587
1588                 perror("shmat");
1589                 return __err;
1590         }
1591 #else
1592         /*
1593          * HP-UX inherits shm mappings?
1594          */
1595         data = threads;
1596 #endif
1597
1598         td = data + offset * sizeof(struct thread_data);
1599         ret = thread_main(td);
1600         shmdt(data);
1601         return (int) (uintptr_t) ret;
1602 }
1603
1604 /*
1605  * Run over the job map and reap the threads that have exited, if any.
1606  */
1607 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1608                          unsigned int *m_rate)
1609 {
1610         struct thread_data *td;
1611         unsigned int cputhreads, realthreads, pending;
1612         int i, status, ret;
1613
1614         /*
1615          * reap exited threads (TD_EXITED -> TD_REAPED)
1616          */
1617         realthreads = pending = cputhreads = 0;
1618         for_each_td(td, i) {
1619                 int flags = 0;
1620
1621                 /*
1622                  * ->io_ops is NULL for a thread that has closed its
1623                  * io engine
1624                  */
1625                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1626                         cputhreads++;
1627                 else
1628                         realthreads++;
1629
1630                 if (!td->pid) {
1631                         pending++;
1632                         continue;
1633                 }
1634                 if (td->runstate == TD_REAPED)
1635                         continue;
1636                 if (td->o.use_thread) {
1637                         if (td->runstate == TD_EXITED) {
1638                                 td_set_runstate(td, TD_REAPED);
1639                                 goto reaped;
1640                         }
1641                         continue;
1642                 }
1643
1644                 flags = WNOHANG;
1645                 if (td->runstate == TD_EXITED)
1646                         flags = 0;
1647
1648                 /*
1649                  * check if someone quit or got killed in an unusual way
1650                  */
1651                 ret = waitpid(td->pid, &status, flags);
1652                 if (ret < 0) {
1653                         if (errno == ECHILD) {
1654                                 log_err("fio: pid=%d disappeared %d\n",
1655                                                 (int) td->pid, td->runstate);
1656                                 td->sig = ECHILD;
1657                                 td_set_runstate(td, TD_REAPED);
1658                                 goto reaped;
1659                         }
1660                         perror("waitpid");
1661                 } else if (ret == td->pid) {
1662                         if (WIFSIGNALED(status)) {
1663                                 int sig = WTERMSIG(status);
1664
1665                                 if (sig != SIGTERM && sig != SIGUSR2)
1666                                         log_err("fio: pid=%d, got signal=%d\n",
1667                                                         (int) td->pid, sig);
1668                                 td->sig = sig;
1669                                 td_set_runstate(td, TD_REAPED);
1670                                 goto reaped;
1671                         }
1672                         if (WIFEXITED(status)) {
1673                                 if (WEXITSTATUS(status) && !td->error)
1674                                         td->error = WEXITSTATUS(status);
1675
1676                                 td_set_runstate(td, TD_REAPED);
1677                                 goto reaped;
1678                         }
1679                 }
1680
1681                 /*
1682                  * thread is not dead, continue
1683                  */
1684                 pending++;
1685                 continue;
1686 reaped:
1687                 (*nr_running)--;
1688                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1689                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1690                 if (!td->pid)
1691                         pending--;
1692
1693                 if (td->error)
1694                         exit_value++;
1695
1696                 done_secs += mtime_since_now(&td->epoch) / 1000;
1697                 profile_td_exit(td);
1698         }
1699
1700         if (*nr_running == cputhreads && !pending && realthreads)
1701                 fio_terminate_threads(TERMINATE_ALL);
1702 }
1703
1704 static void do_usleep(unsigned int usecs)
1705 {
1706         check_for_running_stats();
1707         usleep(usecs);
1708 }
1709
1710 /*
1711  * Main function for kicking off and reaping jobs, as needed.
1712  */
1713 static void run_threads(void)
1714 {
1715         struct thread_data *td;
1716         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1717         uint64_t spent;
1718
1719         if (fio_gtod_offload && fio_start_gtod_thread())
1720                 return;
1721
1722         fio_idle_prof_init();
1723
1724         set_sig_handlers();
1725
1726         nr_thread = nr_process = 0;
1727         for_each_td(td, i) {
1728                 if (td->o.use_thread)
1729                         nr_thread++;
1730                 else
1731                         nr_process++;
1732         }
1733
1734         if (output_format == FIO_OUTPUT_NORMAL) {
1735                 log_info("Starting ");
1736                 if (nr_thread)
1737                         log_info("%d thread%s", nr_thread,
1738                                                 nr_thread > 1 ? "s" : "");
1739                 if (nr_process) {
1740                         if (nr_thread)
1741                                 log_info(" and ");
1742                         log_info("%d process%s", nr_process,
1743                                                 nr_process > 1 ? "es" : "");
1744                 }
1745                 log_info("\n");
1746                 fflush(stdout);
1747         }
1748
1749         todo = thread_number;
1750         nr_running = 0;
1751         nr_started = 0;
1752         m_rate = t_rate = 0;
1753
1754         for_each_td(td, i) {
1755                 print_status_init(td->thread_number - 1);
1756
1757                 if (!td->o.create_serialize)
1758                         continue;
1759
1760                 /*
1761                  * do file setup here so it happens sequentially,
1762                  * we don't want X number of threads getting their
1763                  * client data interspersed on disk
1764                  */
1765                 if (setup_files(td)) {
1766                         exit_value++;
1767                         if (td->error)
1768                                 log_err("fio: pid=%d, err=%d/%s\n",
1769                                         (int) td->pid, td->error, td->verror);
1770                         td_set_runstate(td, TD_REAPED);
1771                         todo--;
1772                 } else {
1773                         struct fio_file *f;
1774                         unsigned int j;
1775
1776                         /*
1777                          * for sharing to work, each job must always open
1778                          * its own files. so close them, if we opened them
1779                          * for creation
1780                          */
1781                         for_each_file(td, f, j) {
1782                                 if (fio_file_open(f))
1783                                         td_io_close_file(td, f);
1784                         }
1785                 }
1786         }
1787
1788         /* start idle threads before io threads start to run */
1789         fio_idle_prof_start();
1790
1791         set_genesis_time();
1792
1793         while (todo) {
1794                 struct thread_data *map[REAL_MAX_JOBS];
1795                 struct timeval this_start;
1796                 int this_jobs = 0, left;
1797
1798                 /*
1799                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1800                  */
1801                 for_each_td(td, i) {
1802                         if (td->runstate != TD_NOT_CREATED)
1803                                 continue;
1804
1805                         /*
1806                          * never got a chance to start, killed by other
1807                          * thread for some reason
1808                          */
1809                         if (td->terminate) {
1810                                 todo--;
1811                                 continue;
1812                         }
1813
1814                         if (td->o.start_delay) {
1815                                 spent = utime_since_genesis();
1816
1817                                 if (td->o.start_delay > spent)
1818                                         continue;
1819                         }
1820
1821                         if (td->o.stonewall && (nr_started || nr_running)) {
1822                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1823                                                         td->o.name);
1824                                 break;
1825                         }
1826
1827                         init_disk_util(td);
1828
1829                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1830                         td->update_rusage = 0;
1831
1832                         /*
1833                          * Set state to created. Thread will transition
1834                          * to TD_INITIALIZED when it's done setting up.
1835                          */
1836                         td_set_runstate(td, TD_CREATED);
1837                         map[this_jobs++] = td;
1838                         nr_started++;
1839
1840                         if (td->o.use_thread) {
1841                                 int ret;
1842
1843                                 dprint(FD_PROCESS, "will pthread_create\n");
1844                                 ret = pthread_create(&td->thread, NULL,
1845                                                         thread_main, td);
1846                                 if (ret) {
1847                                         log_err("pthread_create: %s\n",
1848                                                         strerror(ret));
1849                                         nr_started--;
1850                                         break;
1851                                 }
1852                                 ret = pthread_detach(td->thread);
1853                                 if (ret)
1854                                         log_err("pthread_detach: %s",
1855                                                         strerror(ret));
1856                         } else {
1857                                 pid_t pid;
1858                                 dprint(FD_PROCESS, "will fork\n");
1859                                 pid = fork();
1860                                 if (!pid) {
1861                                         int ret = fork_main(shm_id, i);
1862
1863                                         _exit(ret);
1864                                 } else if (i == fio_debug_jobno)
1865                                         *fio_debug_jobp = pid;
1866                         }
1867                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1868                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1869                                 log_err("fio: job startup hung? exiting.\n");
1870                                 fio_terminate_threads(TERMINATE_ALL);
1871                                 fio_abort = 1;
1872                                 nr_started--;
1873                                 break;
1874                         }
1875                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1876                 }
1877
1878                 /*
1879                  * Wait for the started threads to transition to
1880                  * TD_INITIALIZED.
1881                  */
1882                 fio_gettime(&this_start, NULL);
1883                 left = this_jobs;
1884                 while (left && !fio_abort) {
1885                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1886                                 break;
1887
1888                         do_usleep(100000);
1889
1890                         for (i = 0; i < this_jobs; i++) {
1891                                 td = map[i];
1892                                 if (!td)
1893                                         continue;
1894                                 if (td->runstate == TD_INITIALIZED) {
1895                                         map[i] = NULL;
1896                                         left--;
1897                                 } else if (td->runstate >= TD_EXITED) {
1898                                         map[i] = NULL;
1899                                         left--;
1900                                         todo--;
1901                                         nr_running++; /* work-around... */
1902                                 }
1903                         }
1904                 }
1905
1906                 if (left) {
1907                         log_err("fio: %d job%s failed to start\n", left,
1908                                         left > 1 ? "s" : "");
1909                         for (i = 0; i < this_jobs; i++) {
1910                                 td = map[i];
1911                                 if (!td)
1912                                         continue;
1913                                 kill(td->pid, SIGTERM);
1914                         }
1915                         break;
1916                 }
1917
1918                 /*
1919                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1920                  */
1921                 for_each_td(td, i) {
1922                         if (td->runstate != TD_INITIALIZED)
1923                                 continue;
1924
1925                         if (in_ramp_time(td))
1926                                 td_set_runstate(td, TD_RAMP);
1927                         else
1928                                 td_set_runstate(td, TD_RUNNING);
1929                         nr_running++;
1930                         nr_started--;
1931                         m_rate += ddir_rw_sum(td->o.ratemin);
1932                         t_rate += ddir_rw_sum(td->o.rate);
1933                         todo--;
1934                         fio_mutex_up(td->mutex);
1935                 }
1936
1937                 reap_threads(&nr_running, &t_rate, &m_rate);
1938
1939                 if (todo)
1940                         do_usleep(100000);
1941         }
1942
1943         while (nr_running) {
1944                 reap_threads(&nr_running, &t_rate, &m_rate);
1945                 do_usleep(10000);
1946         }
1947
1948         fio_idle_prof_stop();
1949
1950         update_io_ticks();
1951 }
1952
1953 void wait_for_disk_thread_exit(void)
1954 {
1955         fio_mutex_down(disk_thread_mutex);
1956 }
1957
1958 static void free_disk_util(void)
1959 {
1960         disk_util_start_exit();
1961         wait_for_disk_thread_exit();
1962         disk_util_prune_entries();
1963 }
1964
1965 static void *disk_thread_main(void *data)
1966 {
1967         int ret = 0;
1968
1969         fio_mutex_up(startup_mutex);
1970
1971         while (threads && !ret) {
1972                 usleep(DISK_UTIL_MSEC * 1000);
1973                 if (!threads)
1974                         break;
1975                 ret = update_io_ticks();
1976
1977                 if (!is_backend)
1978                         print_thread_status();
1979         }
1980
1981         fio_mutex_up(disk_thread_mutex);
1982         return NULL;
1983 }
1984
1985 static int create_disk_util_thread(void)
1986 {
1987         int ret;
1988
1989         setup_disk_util();
1990
1991         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1992
1993         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1994         if (ret) {
1995                 fio_mutex_remove(disk_thread_mutex);
1996                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1997                 return 1;
1998         }
1999
2000         ret = pthread_detach(disk_util_thread);
2001         if (ret) {
2002                 fio_mutex_remove(disk_thread_mutex);
2003                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
2004                 return 1;
2005         }
2006
2007         dprint(FD_MUTEX, "wait on startup_mutex\n");
2008         fio_mutex_down(startup_mutex);
2009         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2010         return 0;
2011 }
2012
2013 int fio_backend(void)
2014 {
2015         struct thread_data *td;
2016         int i;
2017
2018         if (exec_profile) {
2019                 if (load_profile(exec_profile))
2020                         return 1;
2021                 free(exec_profile);
2022                 exec_profile = NULL;
2023         }
2024         if (!thread_number)
2025                 return 0;
2026
2027         if (write_bw_log) {
2028                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
2029                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
2030                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2031         }
2032
2033         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2034         if (startup_mutex == NULL)
2035                 return 1;
2036         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2037         if (writeout_mutex == NULL)
2038                 return 1;
2039
2040         set_genesis_time();
2041         stat_init();
2042         create_disk_util_thread();
2043
2044         cgroup_list = smalloc(sizeof(*cgroup_list));
2045         INIT_FLIST_HEAD(cgroup_list);
2046
2047         run_threads();
2048
2049         if (!fio_abort) {
2050                 show_run_stats();
2051                 if (write_bw_log) {
2052                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2053                         __finish_log(agg_io_log[DDIR_WRITE],
2054                                         "agg-write_bw.log");
2055                         __finish_log(agg_io_log[DDIR_TRIM],
2056                                         "agg-write_bw.log");
2057                 }
2058         }
2059
2060         for_each_td(td, i)
2061                 fio_options_free(td);
2062
2063         free_disk_util();
2064         cgroup_kill(cgroup_list);
2065         sfree(cgroup_list);
2066         sfree(cgroup_mnt);
2067
2068         fio_mutex_remove(startup_mutex);
2069         fio_mutex_remove(writeout_mutex);
2070         fio_mutex_remove(disk_thread_mutex);
2071         stat_exit();
2072         return exit_value;
2073 }