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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32 #include <pthread.h>
  33
  34 #include "fio.h"
  35 #include "smalloc.h"
  36 #include "verify.h"
  37 #include "diskutil.h"
  38 #include "cgroup.h"
  39 #include "profile.h"
  40 #include "lib/rand.h"
  41 #include "lib/memalign.h"
  42 #include "server.h"
  43 #include "lib/getrusage.h"
  44 #include "idletime.h"
  45 #include "err.h"
  46 #include "workqueue.h"
  47 #include "lib/mountcheck.h"
  48 #include "rate-submit.h"
  49 #include "helper_thread.h"
  50 #include "pshared.h"
  51 #include "zone-dist.h"
  52
  53 static struct fio_sem *startup_sem;
  54 static struct flist_head *cgroup_list;
  55 static struct cgroup_mnt *cgroup_mnt;
  56 static int exit_value;
  57 static volatile bool fio_abort;
  58 static unsigned int nr_process = 0;
  59 static unsigned int nr_thread = 0;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int nr_segments = 0;
  66 unsigned int cur_segment = 0;
  67 unsigned int stat_number = 0;
  68 int temp_stall_ts;
  69 unsigned long done_secs = 0;
  70 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
  71 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
  72 #else
  73 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  74 #endif
  75
  76 #define JOB_START_TIMEOUT       (5 * 1000)
  77
  78 static void sig_int(int sig)
  79 {
  80         if (nr_segments) {
  81                 if (is_backend)
  82                         fio_server_got_signal(sig);
  83                 else {
  84                         log_info("\nfio: terminating on signal %d\n", sig);
  85                         log_info_flush();
  86                         exit_value = 128;
  87                 }
  88
  89                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  90         }
  91 }
  92
  93 void sig_show_status(int sig)
  94 {
  95         show_running_run_stats();
  96 }
  97
  98 static void set_sig_handlers(void)
  99 {
 100         struct sigaction act;
 101
 102         memset(&act, 0, sizeof(act));
 103         act.sa_handler = sig_int;
 104         act.sa_flags = SA_RESTART;
 105         sigaction(SIGINT, &act, NULL);
 106
 107         memset(&act, 0, sizeof(act));
 108         act.sa_handler = sig_int;
 109         act.sa_flags = SA_RESTART;
 110         sigaction(SIGTERM, &act, NULL);
 111
 112 /* Windows uses SIGBREAK as a quit signal from other applications */
 113 #ifdef WIN32
 114         memset(&act, 0, sizeof(act));
 115         act.sa_handler = sig_int;
 116         act.sa_flags = SA_RESTART;
 117         sigaction(SIGBREAK, &act, NULL);
 118 #endif
 119
 120         memset(&act, 0, sizeof(act));
 121         act.sa_handler = sig_show_status;
 122         act.sa_flags = SA_RESTART;
 123         sigaction(SIGUSR1, &act, NULL);
 124
 125         if (is_backend) {
 126                 memset(&act, 0, sizeof(act));
 127                 act.sa_handler = sig_int;
 128                 act.sa_flags = SA_RESTART;
 129                 sigaction(SIGPIPE, &act, NULL);
 130         }
 131 }
 132
 133 /*
 134  * Check if we are above the minimum rate given.
 135  */
 136 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 137                              enum fio_ddir ddir)
 138 {
 139         unsigned long long bytes = 0;
 140         unsigned long iops = 0;
 141         unsigned long spent;
 142         unsigned long long rate;
 143         unsigned long long ratemin = 0;
 144         unsigned int rate_iops = 0;
 145         unsigned int rate_iops_min = 0;
 146
 147         assert(ddir_rw(ddir));
 148
 149         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 150                 return false;
 151
 152         /*
 153          * allow a 2 second settle period in the beginning
 154          */
 155         if (mtime_since(&td->start, now) < 2000)
 156                 return false;
 157
 158         iops += td->this_io_blocks[ddir];
 159         bytes += td->this_io_bytes[ddir];
 160         ratemin += td->o.ratemin[ddir];
 161         rate_iops += td->o.rate_iops[ddir];
 162         rate_iops_min += td->o.rate_iops_min[ddir];
 163
 164         /*
 165          * if rate blocks is set, sample is running
 166          */
 167         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 168                 spent = mtime_since(&td->lastrate[ddir], now);
 169                 if (spent < td->o.ratecycle)
 170                         return false;
 171
 172                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 173                         /*
 174                          * check bandwidth specified rate
 175                          */
 176                         if (bytes < td->rate_bytes[ddir]) {
 177                                 log_err("%s: rate_min=%lluB/s not met, only transferred %lluB\n",
 178                                         td->o.name, ratemin, bytes);
 179                                 return true;
 180                         } else {
 181                                 if (spent)
 182                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 183                                 else
 184                                         rate = 0;
 185
 186                                 if (rate < ratemin ||
 187                                     bytes < td->rate_bytes[ddir]) {
 188                                         log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 189                                                 td->o.name, ratemin, rate);
 190                                         return true;
 191                                 }
 192                         }
 193                 } else {
 194                         /*
 195                          * checks iops specified rate
 196                          */
 197                         if (iops < rate_iops) {
 198                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 199                                                 td->o.name, rate_iops, iops);
 200                                 return true;
 201                         } else {
 202                                 if (spent)
 203                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 204                                 else
 205                                         rate = 0;
 206
 207                                 if (rate < rate_iops_min ||
 208                                     iops < td->rate_blocks[ddir]) {
 209                                         log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 210                                                 td->o.name, rate_iops_min, rate);
 211                                         return true;
 212                                 }
 213                         }
 214                 }
 215         }
 216
 217         td->rate_bytes[ddir] = bytes;
 218         td->rate_blocks[ddir] = iops;
 219         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 220         return false;
 221 }
 222
 223 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 224 {
 225         bool ret = false;
 226
 227         for_each_rw_ddir(ddir) {
 228                 if (td->bytes_done[ddir])
 229                         ret |= __check_min_rate(td, now, ddir);
 230         }
 231
 232         return ret;
 233 }
 234
 235 /*
 236  * When job exits, we can cancel the in-flight IO if we are using async
 237  * io. Attempt to do so.
 238  */
 239 static void cleanup_pending_aio(struct thread_data *td)
 240 {
 241         int r;
 242
 243         /*
 244          * get immediately available events, if any
 245          */
 246         r = io_u_queued_complete(td, 0);
 247
 248         /*
 249          * now cancel remaining active events
 250          */
 251         if (td->io_ops->cancel) {
 252                 struct io_u *io_u;
 253                 int i;
 254
 255                 io_u_qiter(&td->io_u_all, io_u, i) {
 256                         if (io_u->flags & IO_U_F_FLIGHT) {
 257                                 r = td->io_ops->cancel(td, io_u);
 258                                 if (!r)
 259                                         put_io_u(td, io_u);
 260                         }
 261                 }
 262         }
 263
 264         if (td->cur_depth)
 265                 r = io_u_queued_complete(td, td->cur_depth);
 266 }
 267
 268 /*
 269  * Helper to handle the final sync of a file. Works just like the normal
 270  * io path, just does everything sync.
 271  */
 272 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 273 {
 274         struct io_u *io_u = __get_io_u(td);
 275         enum fio_q_status ret;
 276
 277         if (!io_u)
 278                 return true;
 279
 280         io_u->ddir = DDIR_SYNC;
 281         io_u->file = f;
 282         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 283
 284         if (td_io_prep(td, io_u)) {
 285                 put_io_u(td, io_u);
 286                 return true;
 287         }
 288
 289 requeue:
 290         ret = td_io_queue(td, io_u);
 291         switch (ret) {
 292         case FIO_Q_QUEUED:
 293                 td_io_commit(td);
 294                 if (io_u_queued_complete(td, 1) < 0)
 295                         return true;
 296                 break;
 297         case FIO_Q_COMPLETED:
 298                 if (io_u->error) {
 299                         td_verror(td, io_u->error, "td_io_queue");
 300                         return true;
 301                 }
 302
 303                 if (io_u_sync_complete(td, io_u) < 0)
 304                         return true;
 305                 break;
 306         case FIO_Q_BUSY:
 307                 td_io_commit(td);
 308                 goto requeue;
 309         }
 310
 311         return false;
 312 }
 313
 314 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 315 {
 316         int ret, ret2;
 317
 318         if (fio_file_open(f))
 319                 return fio_io_sync(td, f);
 320
 321         if (td_io_open_file(td, f))
 322                 return 1;
 323
 324         ret = fio_io_sync(td, f);
 325         ret2 = 0;
 326         if (fio_file_open(f))
 327                 ret2 = td_io_close_file(td, f);
 328         return (ret || ret2);
 329 }
 330
 331 static inline void __update_ts_cache(struct thread_data *td)
 332 {
 333         fio_gettime(&td->ts_cache, NULL);
 334 }
 335
 336 static inline void update_ts_cache(struct thread_data *td)
 337 {
 338         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 339                 __update_ts_cache(td);
 340 }
 341
 342 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 343 {
 344         if (in_ramp_time(td))
 345                 return false;
 346         if (!td->o.timeout)
 347                 return false;
 348         if (utime_since(&td->epoch, t) >= td->o.timeout)
 349                 return true;
 350
 351         return false;
 352 }
 353
 354 /*
 355  * We need to update the runtime consistently in ms, but keep a running
 356  * tally of the current elapsed time in microseconds for sub millisecond
 357  * updates.
 358  */
 359 static inline void update_runtime(struct thread_data *td,
 360                                   unsigned long long *elapsed_us,
 361                                   const enum fio_ddir ddir)
 362 {
 363         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 364                 return;
 365
 366         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 367         elapsed_us[ddir] += utime_since_now(&td->start);
 368         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 369 }
 370
 371 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 372                                 int *retptr)
 373 {
 374         int ret = *retptr;
 375
 376         if (ret < 0 || td->error) {
 377                 int err = td->error;
 378                 enum error_type_bit eb;
 379
 380                 if (ret < 0)
 381                         err = -ret;
 382
 383                 eb = td_error_type(ddir, err);
 384                 if (!(td->o.continue_on_error & (1 << eb)))
 385                         return true;
 386
 387                 if (td_non_fatal_error(td, eb, err)) {
 388                         /*
 389                          * Continue with the I/Os in case of
 390                          * a non fatal error.
 391                          */
 392                         update_error_count(td, err);
 393                         td_clear_error(td);
 394                         *retptr = 0;
 395                         return false;
 396                 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
 397                         /*
 398                          * We expect to hit this error if
 399                          * fill_device option is set.
 400                          */
 401                         td_clear_error(td);
 402                         fio_mark_td_terminate(td);
 403                         return true;
 404                 } else {
 405                         /*
 406                          * Stop the I/O in case of a fatal
 407                          * error.
 408                          */
 409                         update_error_count(td, err);
 410                         return true;
 411                 }
 412         }
 413
 414         return false;
 415 }
 416
 417 static void check_update_rusage(struct thread_data *td)
 418 {
 419         if (td->update_rusage) {
 420                 td->update_rusage = 0;
 421                 update_rusage_stat(td);
 422                 fio_sem_up(td->rusage_sem);
 423         }
 424 }
 425
 426 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 427 {
 428         const int full = queue_full(td);
 429         int min_evts = 0;
 430         int ret;
 431
 432         if (td->flags & TD_F_REGROW_LOGS)
 433                 return io_u_quiesce(td);
 434
 435         /*
 436          * if the queue is full, we MUST reap at least 1 event
 437          */
 438         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 439         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 440                 min_evts = 1;
 441
 442         if (time && should_check_rate(td))
 443                 fio_gettime(time, NULL);
 444
 445         do {
 446                 ret = io_u_queued_complete(td, min_evts);
 447                 if (ret < 0)
 448                         break;
 449         } while (full && (td->cur_depth > td->o.iodepth_low));
 450
 451         return ret;
 452 }
 453
 454 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 455                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 456                    struct timespec *comp_time)
 457 {
 458         switch (*ret) {
 459         case FIO_Q_COMPLETED:
 460                 if (io_u->error) {
 461                         *ret = -io_u->error;
 462                         clear_io_u(td, io_u);
 463                 } else if (io_u->resid) {
 464                         long long bytes = io_u->xfer_buflen - io_u->resid;
 465                         struct fio_file *f = io_u->file;
 466
 467                         if (bytes_issued)
 468                                 *bytes_issued += bytes;
 469
 470                         if (!from_verify)
 471                                 trim_io_piece(io_u);
 472
 473                         /*
 474                          * zero read, fail
 475                          */
 476                         if (!bytes) {
 477                                 if (!from_verify)
 478                                         unlog_io_piece(td, io_u);
 479                                 td_verror(td, EIO, "full resid");
 480                                 put_io_u(td, io_u);
 481                                 break;
 482                         }
 483
 484                         io_u->xfer_buflen = io_u->resid;
 485                         io_u->xfer_buf += bytes;
 486                         io_u->offset += bytes;
 487
 488                         if (ddir_rw(io_u->ddir))
 489                                 td->ts.short_io_u[io_u->ddir]++;
 490
 491                         if (io_u->offset == f->real_file_size)
 492                                 goto sync_done;
 493
 494                         requeue_io_u(td, &io_u);
 495                 } else {
 496 sync_done:
 497                         if (comp_time && should_check_rate(td))
 498                                 fio_gettime(comp_time, NULL);
 499
 500                         *ret = io_u_sync_complete(td, io_u);
 501                         if (*ret < 0)
 502                                 break;
 503                 }
 504
 505                 if (td->flags & TD_F_REGROW_LOGS)
 506                         regrow_logs(td);
 507
 508                 /*
 509                  * when doing I/O (not when verifying),
 510                  * check for any errors that are to be ignored
 511                  */
 512                 if (!from_verify)
 513                         break;
 514
 515                 return 0;
 516         case FIO_Q_QUEUED:
 517                 /*
 518                  * if the engine doesn't have a commit hook,
 519                  * the io_u is really queued. if it does have such
 520                  * a hook, it has to call io_u_queued() itself.
 521                  */
 522                 if (td->io_ops->commit == NULL)
 523                         io_u_queued(td, io_u);
 524                 if (bytes_issued)
 525                         *bytes_issued += io_u->xfer_buflen;
 526                 break;
 527         case FIO_Q_BUSY:
 528                 if (!from_verify)
 529                         unlog_io_piece(td, io_u);
 530                 requeue_io_u(td, &io_u);
 531                 td_io_commit(td);
 532                 break;
 533         default:
 534                 assert(*ret < 0);
 535                 td_verror(td, -(*ret), "td_io_queue");
 536                 break;
 537         }
 538
 539         if (break_on_this_error(td, ddir, ret))
 540                 return 1;
 541
 542         return 0;
 543 }
 544
 545 static inline bool io_in_polling(struct thread_data *td)
 546 {
 547         return !td->o.iodepth_batch_complete_min &&
 548                    !td->o.iodepth_batch_complete_max;
 549 }
 550 /*
 551  * Unlinks files from thread data fio_file structure
 552  */
 553 static int unlink_all_files(struct thread_data *td)
 554 {
 555         struct fio_file *f;
 556         unsigned int i;
 557         int ret = 0;
 558
 559         for_each_file(td, f, i) {
 560                 if (f->filetype != FIO_TYPE_FILE)
 561                         continue;
 562                 ret = td_io_unlink_file(td, f);
 563                 if (ret)
 564                         break;
 565         }
 566
 567         if (ret)
 568                 td_verror(td, ret, "unlink_all_files");
 569
 570         return ret;
 571 }
 572
 573 /*
 574  * Check if io_u will overlap an in-flight IO in the queue
 575  */
 576 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 577 {
 578         bool overlap;
 579         struct io_u *check_io_u;
 580         unsigned long long x1, x2, y1, y2;
 581         int i;
 582
 583         x1 = io_u->offset;
 584         x2 = io_u->offset + io_u->buflen;
 585         overlap = false;
 586         io_u_qiter(q, check_io_u, i) {
 587                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 588                         y1 = check_io_u->offset;
 589                         y2 = check_io_u->offset + check_io_u->buflen;
 590
 591                         if (x1 < y2 && y1 < x2) {
 592                                 overlap = true;
 593                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 594                                                 x1, io_u->buflen,
 595                                                 y1, check_io_u->buflen);
 596                                 break;
 597                         }
 598                 }
 599         }
 600
 601         return overlap;
 602 }
 603
 604 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 605 {
 606         /*
 607          * Check for overlap if the user asked us to, and we have
 608          * at least one IO in flight besides this one.
 609          */
 610         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 611             in_flight_overlap(&td->io_u_all, io_u))
 612                 return FIO_Q_BUSY;
 613
 614         return td_io_queue(td, io_u);
 615 }
 616
 617 /*
 618  * The main verify engine. Runs over the writes we previously submitted,
 619  * reads the blocks back in, and checks the crc/md5 of the data.
 620  */
 621 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 622 {
 623         struct fio_file *f;
 624         struct io_u *io_u;
 625         int ret, min_events;
 626         unsigned int i;
 627
 628         dprint(FD_VERIFY, "starting loop\n");
 629
 630         /*
 631          * sync io first and invalidate cache, to make sure we really
 632          * read from disk.
 633          */
 634         for_each_file(td, f, i) {
 635                 if (!fio_file_open(f))
 636                         continue;
 637                 if (fio_io_sync(td, f))
 638                         break;
 639                 if (file_invalidate_cache(td, f))
 640                         break;
 641         }
 642
 643         check_update_rusage(td);
 644
 645         if (td->error)
 646                 return;
 647
 648         /*
 649          * verify_state needs to be reset before verification
 650          * proceeds so that expected random seeds match actual
 651          * random seeds in headers. The main loop will reset
 652          * all random number generators if randrepeat is set.
 653          */
 654         if (!td->o.rand_repeatable)
 655                 td_fill_verify_state_seed(td);
 656
 657         td_set_runstate(td, TD_VERIFYING);
 658
 659         io_u = NULL;
 660         while (!td->terminate) {
 661                 enum fio_ddir ddir;
 662                 int full;
 663
 664                 update_ts_cache(td);
 665                 check_update_rusage(td);
 666
 667                 if (runtime_exceeded(td, &td->ts_cache)) {
 668                         __update_ts_cache(td);
 669                         if (runtime_exceeded(td, &td->ts_cache)) {
 670                                 fio_mark_td_terminate(td);
 671                                 break;
 672                         }
 673                 }
 674
 675                 if (flow_threshold_exceeded(td))
 676                         continue;
 677
 678                 if (!td->o.experimental_verify) {
 679                         io_u = __get_io_u(td);
 680                         if (!io_u)
 681                                 break;
 682
 683                         if (get_next_verify(td, io_u)) {
 684                                 put_io_u(td, io_u);
 685                                 break;
 686                         }
 687
 688                         if (td_io_prep(td, io_u)) {
 689                                 put_io_u(td, io_u);
 690                                 break;
 691                         }
 692                 } else {
 693                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 694                                 break;
 695
 696                         while ((io_u = get_io_u(td)) != NULL) {
 697                                 if (IS_ERR_OR_NULL(io_u)) {
 698                                         io_u = NULL;
 699                                         ret = FIO_Q_BUSY;
 700                                         goto reap;
 701                                 }
 702
 703                                 /*
 704                                  * We are only interested in the places where
 705                                  * we wrote or trimmed IOs. Turn those into
 706                                  * reads for verification purposes.
 707                                  */
 708                                 if (io_u->ddir == DDIR_READ) {
 709                                         /*
 710                                          * Pretend we issued it for rwmix
 711                                          * accounting
 712                                          */
 713                                         td->io_issues[DDIR_READ]++;
 714                                         put_io_u(td, io_u);
 715                                         continue;
 716                                 } else if (io_u->ddir == DDIR_TRIM) {
 717                                         io_u->ddir = DDIR_READ;
 718                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 719                                         break;
 720                                 } else if (io_u->ddir == DDIR_WRITE) {
 721                                         io_u->ddir = DDIR_READ;
 722                                         populate_verify_io_u(td, io_u);
 723                                         break;
 724                                 } else {
 725                                         put_io_u(td, io_u);
 726                                         continue;
 727                                 }
 728                         }
 729
 730                         if (!io_u)
 731                                 break;
 732                 }
 733
 734                 if (verify_state_should_stop(td, io_u)) {
 735                         put_io_u(td, io_u);
 736                         break;
 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
 744                 ddir = io_u->ddir;
 745                 if (!td->o.disable_slat)
 746                         fio_gettime(&io_u->start_time, NULL);
 747
 748                 ret = io_u_submit(td, io_u);
 749
 750                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 751                         break;
 752
 753                 /*
 754                  * if we can queue more, do so. but check if there are
 755                  * completed io_u's first. Note that we can get BUSY even
 756                  * without IO queued, if the system is resource starved.
 757                  */
 758 reap:
 759                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 760                 if (full || io_in_polling(td))
 761                         ret = wait_for_completions(td, NULL);
 762
 763                 if (ret < 0)
 764                         break;
 765         }
 766
 767         check_update_rusage(td);
 768
 769         if (!td->error) {
 770                 min_events = td->cur_depth;
 771
 772                 if (min_events)
 773                         ret = io_u_queued_complete(td, min_events);
 774         } else
 775                 cleanup_pending_aio(td);
 776
 777         td_set_runstate(td, TD_RUNNING);
 778
 779         dprint(FD_VERIFY, "exiting loop\n");
 780 }
 781
 782 static bool exceeds_number_ios(struct thread_data *td)
 783 {
 784         unsigned long long number_ios;
 785
 786         if (!td->o.number_ios)
 787                 return false;
 788
 789         number_ios = ddir_rw_sum(td->io_blocks);
 790         number_ios += td->io_u_queued + td->io_u_in_flight;
 791
 792         return number_ios >= (td->o.number_ios * td->loops);
 793 }
 794
 795 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 796 {
 797         unsigned long long bytes, limit;
 798
 799         if (td_rw(td))
 800                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 801         else if (td_write(td))
 802                 bytes = this_bytes[DDIR_WRITE];
 803         else if (td_read(td))
 804                 bytes = this_bytes[DDIR_READ];
 805         else
 806                 bytes = this_bytes[DDIR_TRIM];
 807
 808         if (td->o.io_size)
 809                 limit = td->o.io_size;
 810         else
 811                 limit = td->o.size;
 812
 813         limit *= td->loops;
 814         return bytes >= limit || exceeds_number_ios(td);
 815 }
 816
 817 static bool io_issue_bytes_exceeded(struct thread_data *td)
 818 {
 819         return io_bytes_exceeded(td, td->io_issue_bytes);
 820 }
 821
 822 static bool io_complete_bytes_exceeded(struct thread_data *td)
 823 {
 824         return io_bytes_exceeded(td, td->this_io_bytes);
 825 }
 826
 827 /*
 828  * used to calculate the next io time for rate control
 829  *
 830  */
 831 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 832 {
 833         uint64_t bps = td->rate_bps[ddir];
 834
 835         assert(!(td->flags & TD_F_CHILD));
 836
 837         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 838                 uint64_t val, iops;
 839
 840                 iops = bps / td->o.min_bs[ddir];
 841                 val = (int64_t) (1000000 / iops) *
 842                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 843                 if (val) {
 844                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 845                                         (unsigned long long) 1000000 / val,
 846                                         ddir);
 847                 }
 848                 td->last_usec[ddir] += val;
 849                 return td->last_usec[ddir];
 850         } else if (bps) {
 851                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 852                 uint64_t secs = bytes / bps;
 853                 uint64_t remainder = bytes % bps;
 854
 855                 return remainder * 1000000 / bps + secs * 1000000;
 856         }
 857
 858         return 0;
 859 }
 860
 861 static void init_thinktime(struct thread_data *td)
 862 {
 863         if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
 864                 td->thinktime_blocks_counter = td->io_blocks;
 865         else
 866                 td->thinktime_blocks_counter = td->io_issues;
 867         td->last_thinktime = td->epoch;
 868         td->last_thinktime_blocks = 0;
 869 }
 870
 871 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
 872                              struct timespec *time)
 873 {
 874         unsigned long long b;
 875         uint64_t total;
 876         int left;
 877         struct timespec now;
 878         bool stall = false;
 879
 880         if (td->o.thinktime_iotime) {
 881                 fio_gettime(&now, NULL);
 882                 if (utime_since(&td->last_thinktime, &now)
 883                     >= td->o.thinktime_iotime + td->o.thinktime) {
 884                         stall = true;
 885                 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
 886                         /*
 887                          * When thinktime_iotime is set and thinktime_blocks is
 888                          * not set, skip the thinktime_blocks check, since
 889                          * thinktime_blocks default value 1 does not work
 890                          * together with thinktime_iotime.
 891                          */
 892                         return;
 893                 }
 894
 895         }
 896
 897         b = ddir_rw_sum(td->thinktime_blocks_counter);
 898         if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
 899                 stall = true;
 900
 901         if (!stall)
 902                 return;
 903
 904         io_u_quiesce(td);
 905
 906         total = 0;
 907         if (td->o.thinktime_spin)
 908                 total = usec_spin(td->o.thinktime_spin);
 909
 910         left = td->o.thinktime - total;
 911         if (left)
 912                 total += usec_sleep(td, left);
 913
 914         /*
 915          * If we're ignoring thinktime for the rate, add the number of bytes
 916          * we would have done while sleeping, minus one block to ensure we
 917          * start issuing immediately after the sleep.
 918          */
 919         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 920                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 921                 uint64_t bs = td->o.min_bs[ddir];
 922                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 923                 uint64_t over;
 924
 925                 if (usperop <= total)
 926                         over = bs;
 927                 else
 928                         over = (usperop - total) / usperop * -bs;
 929
 930                 td->rate_io_issue_bytes[ddir] += (missed - over);
 931                 /* adjust for rate_process=poisson */
 932                 td->last_usec[ddir] += total;
 933         }
 934
 935         if (time && should_check_rate(td))
 936                 fio_gettime(time, NULL);
 937
 938         td->last_thinktime_blocks = b;
 939         if (td->o.thinktime_iotime)
 940                 td->last_thinktime = now;
 941 }
 942
 943 /*
 944  * Main IO worker function. It retrieves io_u's to process and queues
 945  * and reaps them, checking for rate and errors along the way.
 946  *
 947  * Returns number of bytes written and trimmed.
 948  */
 949 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 950 {
 951         unsigned int i;
 952         int ret = 0;
 953         uint64_t total_bytes, bytes_issued = 0;
 954
 955         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 956                 bytes_done[i] = td->bytes_done[i];
 957
 958         if (in_ramp_time(td))
 959                 td_set_runstate(td, TD_RAMP);
 960         else
 961                 td_set_runstate(td, TD_RUNNING);
 962
 963         lat_target_init(td);
 964
 965         total_bytes = td->o.size;
 966         /*
 967         * Allow random overwrite workloads to write up to io_size
 968         * before starting verification phase as 'size' doesn't apply.
 969         */
 970         if (td_write(td) && td_random(td) && td->o.norandommap)
 971                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 972         /*
 973          * If verify_backlog is enabled, we'll run the verify in this
 974          * handler as well. For that case, we may need up to twice the
 975          * amount of bytes.
 976          */
 977         if (td->o.verify != VERIFY_NONE &&
 978            (td_write(td) && td->o.verify_backlog))
 979                 total_bytes += td->o.size;
 980
 981         /* In trimwrite mode, each byte is trimmed and then written, so
 982          * allow total_bytes to be twice as big */
 983         if (td_trimwrite(td))
 984                 total_bytes += td->total_io_size;
 985
 986         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 987                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 988                 td->o.time_based) {
 989                 struct timespec comp_time;
 990                 struct io_u *io_u;
 991                 int full;
 992                 enum fio_ddir ddir;
 993
 994                 check_update_rusage(td);
 995
 996                 if (td->terminate || td->done)
 997                         break;
 998
 999                 update_ts_cache(td);
1000
1001                 if (runtime_exceeded(td, &td->ts_cache)) {
1002                         __update_ts_cache(td);
1003                         if (runtime_exceeded(td, &td->ts_cache)) {
1004                                 fio_mark_td_terminate(td);
1005                                 break;
1006                         }
1007                 }
1008
1009                 if (flow_threshold_exceeded(td))
1010                         continue;
1011
1012                 /*
1013                  * Break if we exceeded the bytes. The exception is time
1014                  * based runs, but we still need to break out of the loop
1015                  * for those to run verification, if enabled.
1016                  * Jobs read from iolog do not use this stop condition.
1017                  */
1018                 if (bytes_issued >= total_bytes &&
1019                     !td->o.read_iolog_file &&
1020                     (!td->o.time_based ||
1021                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
1022                         break;
1023
1024                 io_u = get_io_u(td);
1025                 if (IS_ERR_OR_NULL(io_u)) {
1026                         int err = PTR_ERR(io_u);
1027
1028                         io_u = NULL;
1029                         ddir = DDIR_INVAL;
1030                         if (err == -EBUSY) {
1031                                 ret = FIO_Q_BUSY;
1032                                 goto reap;
1033                         }
1034                         if (td->o.latency_target)
1035                                 goto reap;
1036                         break;
1037                 }
1038
1039                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1040                         populate_verify_io_u(td, io_u);
1041
1042                 ddir = io_u->ddir;
1043
1044                 /*
1045                  * Add verification end_io handler if:
1046                  *      - Asked to verify (!td_rw(td))
1047                  *      - Or the io_u is from our verify list (mixed write/ver)
1048                  */
1049                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1050                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1051
1052                         if (verify_state_should_stop(td, io_u)) {
1053                                 put_io_u(td, io_u);
1054                                 break;
1055                         }
1056
1057                         if (td->o.verify_async)
1058                                 io_u->end_io = verify_io_u_async;
1059                         else
1060                                 io_u->end_io = verify_io_u;
1061                         td_set_runstate(td, TD_VERIFYING);
1062                 } else if (in_ramp_time(td))
1063                         td_set_runstate(td, TD_RAMP);
1064                 else
1065                         td_set_runstate(td, TD_RUNNING);
1066
1067                 /*
1068                  * Always log IO before it's issued, so we know the specific
1069                  * order of it. The logged unit will track when the IO has
1070                  * completed.
1071                  */
1072                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1073                     td->o.do_verify &&
1074                     td->o.verify != VERIFY_NONE &&
1075                     !td->o.experimental_verify)
1076                         log_io_piece(td, io_u);
1077
1078                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1079                         const unsigned long long blen = io_u->xfer_buflen;
1080                         const enum fio_ddir __ddir = acct_ddir(io_u);
1081
1082                         if (td->error)
1083                                 break;
1084
1085                         workqueue_enqueue(&td->io_wq, &io_u->work);
1086                         ret = FIO_Q_QUEUED;
1087
1088                         if (ddir_rw(__ddir)) {
1089                                 td->io_issues[__ddir]++;
1090                                 td->io_issue_bytes[__ddir] += blen;
1091                                 td->rate_io_issue_bytes[__ddir] += blen;
1092                         }
1093
1094                         if (should_check_rate(td)) {
1095                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1096                                 fio_gettime(&comp_time, NULL);
1097                         }
1098
1099                 } else {
1100                         ret = io_u_submit(td, io_u);
1101
1102                         if (should_check_rate(td))
1103                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1104
1105                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1106                                 break;
1107
1108                         /*
1109                          * See if we need to complete some commands. Note that
1110                          * we can get BUSY even without IO queued, if the
1111                          * system is resource starved.
1112                          */
1113 reap:
1114                         full = queue_full(td) ||
1115                                 (ret == FIO_Q_BUSY && td->cur_depth);
1116                         if (full || io_in_polling(td))
1117                                 ret = wait_for_completions(td, &comp_time);
1118                 }
1119                 if (ret < 0)
1120                         break;
1121
1122                 if (ddir_rw(ddir) && td->o.thinktime)
1123                         handle_thinktime(td, ddir, &comp_time);
1124
1125                 if (!ddir_rw_sum(td->bytes_done) &&
1126                     !td_ioengine_flagged(td, FIO_NOIO))
1127                         continue;
1128
1129                 if (!in_ramp_time(td) && should_check_rate(td)) {
1130                         if (check_min_rate(td, &comp_time)) {
1131                                 if (exitall_on_terminate || td->o.exitall_error)
1132                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1133                                 td_verror(td, EIO, "check_min_rate");
1134                                 break;
1135                         }
1136                 }
1137                 if (!in_ramp_time(td) && td->o.latency_target)
1138                         lat_target_check(td);
1139         }
1140
1141         check_update_rusage(td);
1142
1143         if (td->trim_entries)
1144                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1145
1146         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1147                 td->error = 0;
1148                 fio_mark_td_terminate(td);
1149         }
1150         if (!td->error) {
1151                 struct fio_file *f;
1152
1153                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1154                         workqueue_flush(&td->io_wq);
1155                         i = 0;
1156                 } else
1157                         i = td->cur_depth;
1158
1159                 if (i) {
1160                         ret = io_u_queued_complete(td, i);
1161                         if (td->o.fill_device &&
1162                             (td->error == ENOSPC || td->error == EDQUOT))
1163                                 td->error = 0;
1164                 }
1165
1166                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1167                         td_set_runstate(td, TD_FSYNCING);
1168
1169                         for_each_file(td, f, i) {
1170                                 if (!fio_file_fsync(td, f))
1171                                         continue;
1172
1173                                 log_err("fio: end_fsync failed for file %s\n",
1174                                                                 f->file_name);
1175                         }
1176                 }
1177         } else {
1178                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1179                         workqueue_flush(&td->io_wq);
1180                 cleanup_pending_aio(td);
1181         }
1182
1183         /*
1184          * stop job if we failed doing any IO
1185          */
1186         if (!ddir_rw_sum(td->this_io_bytes))
1187                 td->done = 1;
1188
1189         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1190                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1191 }
1192
1193 static void free_file_completion_logging(struct thread_data *td)
1194 {
1195         struct fio_file *f;
1196         unsigned int i;
1197
1198         for_each_file(td, f, i) {
1199                 if (!f->last_write_comp)
1200                         break;
1201                 sfree(f->last_write_comp);
1202         }
1203 }
1204
1205 static int init_file_completion_logging(struct thread_data *td,
1206                                         unsigned int depth)
1207 {
1208         struct fio_file *f;
1209         unsigned int i;
1210
1211         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1212                 return 0;
1213
1214         for_each_file(td, f, i) {
1215                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1216                 if (!f->last_write_comp)
1217                         goto cleanup;
1218         }
1219
1220         return 0;
1221
1222 cleanup:
1223         free_file_completion_logging(td);
1224         log_err("fio: failed to alloc write comp data\n");
1225         return 1;
1226 }
1227
1228 static void cleanup_io_u(struct thread_data *td)
1229 {
1230         struct io_u *io_u;
1231
1232         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1233
1234                 if (td->io_ops->io_u_free)
1235                         td->io_ops->io_u_free(td, io_u);
1236
1237                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1238         }
1239
1240         free_io_mem(td);
1241
1242         io_u_rexit(&td->io_u_requeues);
1243         io_u_qexit(&td->io_u_freelist, false);
1244         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1245
1246         free_file_completion_logging(td);
1247 }
1248
1249 static int init_io_u(struct thread_data *td)
1250 {
1251         struct io_u *io_u;
1252         int cl_align, i, max_units;
1253         int err;
1254
1255         max_units = td->o.iodepth;
1256
1257         err = 0;
1258         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1259         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1260         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1261
1262         if (err) {
1263                 log_err("fio: failed setting up IO queues\n");
1264                 return 1;
1265         }
1266
1267         cl_align = os_cache_line_size();
1268
1269         for (i = 0; i < max_units; i++) {
1270                 void *ptr;
1271
1272                 if (td->terminate)
1273                         return 1;
1274
1275                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1276                 if (!ptr) {
1277                         log_err("fio: unable to allocate aligned memory\n");
1278                         return 1;
1279                 }
1280
1281                 io_u = ptr;
1282                 memset(io_u, 0, sizeof(*io_u));
1283                 INIT_FLIST_HEAD(&io_u->verify_list);
1284                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1285
1286                 io_u->index = i;
1287                 io_u->flags = IO_U_F_FREE;
1288                 io_u_qpush(&td->io_u_freelist, io_u);
1289
1290                 /*
1291                  * io_u never leaves this stack, used for iteration of all
1292                  * io_u buffers.
1293                  */
1294                 io_u_qpush(&td->io_u_all, io_u);
1295
1296                 if (td->io_ops->io_u_init) {
1297                         int ret = td->io_ops->io_u_init(td, io_u);
1298
1299                         if (ret) {
1300                                 log_err("fio: failed to init engine data: %d\n", ret);
1301                                 return 1;
1302                         }
1303                 }
1304         }
1305
1306         init_io_u_buffers(td);
1307
1308         if (init_file_completion_logging(td, max_units))
1309                 return 1;
1310
1311         return 0;
1312 }
1313
1314 int init_io_u_buffers(struct thread_data *td)
1315 {
1316         struct io_u *io_u;
1317         unsigned long long max_bs, min_write;
1318         int i, max_units;
1319         int data_xfer = 1;
1320         char *p;
1321
1322         max_units = td->o.iodepth;
1323         max_bs = td_max_bs(td);
1324         min_write = td->o.min_bs[DDIR_WRITE];
1325         td->orig_buffer_size = (unsigned long long) max_bs
1326                                         * (unsigned long long) max_units;
1327
1328         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1329                 data_xfer = 0;
1330
1331         /*
1332          * if we may later need to do address alignment, then add any
1333          * possible adjustment here so that we don't cause a buffer
1334          * overflow later. this adjustment may be too much if we get
1335          * lucky and the allocator gives us an aligned address.
1336          */
1337         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1338             td_ioengine_flagged(td, FIO_RAWIO))
1339                 td->orig_buffer_size += page_mask + td->o.mem_align;
1340
1341         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1342                 unsigned long long bs;
1343
1344                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1345                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1346         }
1347
1348         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1349                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1350                 return 1;
1351         }
1352
1353         if (data_xfer && allocate_io_mem(td))
1354                 return 1;
1355
1356         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1357             td_ioengine_flagged(td, FIO_RAWIO))
1358                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1359         else
1360                 p = td->orig_buffer;
1361
1362         for (i = 0; i < max_units; i++) {
1363                 io_u = td->io_u_all.io_us[i];
1364                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1365
1366                 if (data_xfer) {
1367                         io_u->buf = p;
1368                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1369
1370                         if (td_write(td))
1371                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1372                         if (td_write(td) && td->o.verify_pattern_bytes) {
1373                                 /*
1374                                  * Fill the buffer with the pattern if we are
1375                                  * going to be doing writes.
1376                                  */
1377                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1378                         }
1379                 }
1380                 p += max_bs;
1381         }
1382
1383         return 0;
1384 }
1385
1386 #ifdef FIO_HAVE_IOSCHED_SWITCH
1387 /*
1388  * These functions are Linux specific.
1389  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1390  */
1391 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1392 {
1393         char tmp[256], tmp2[128], *p;
1394         FILE *f;
1395         int ret;
1396
1397         assert(file->du && file->du->sysfs_root);
1398         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1399
1400         f = fopen(tmp, "r+");
1401         if (!f) {
1402                 if (errno == ENOENT) {
1403                         log_err("fio: os or kernel doesn't support IO scheduler"
1404                                 " switching\n");
1405                         return 0;
1406                 }
1407                 td_verror(td, errno, "fopen iosched");
1408                 return 1;
1409         }
1410
1411         /*
1412          * Set io scheduler.
1413          */
1414         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1415         if (ferror(f) || ret != 1) {
1416                 td_verror(td, errno, "fwrite");
1417                 fclose(f);
1418                 return 1;
1419         }
1420
1421         rewind(f);
1422
1423         /*
1424          * Read back and check that the selected scheduler is now the default.
1425          */
1426         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1427         if (ferror(f) || ret < 0) {
1428                 td_verror(td, errno, "fread");
1429                 fclose(f);
1430                 return 1;
1431         }
1432         tmp[ret] = '\0';
1433         /*
1434          * either a list of io schedulers or "none\n" is expected. Strip the
1435          * trailing newline.
1436          */
1437         p = tmp;
1438         strsep(&p, "\n");
1439
1440         /*
1441          * Write to "none" entry doesn't fail, so check the result here.
1442          */
1443         if (!strcmp(tmp, "none")) {
1444                 log_err("fio: io scheduler is not tunable\n");
1445                 fclose(f);
1446                 return 0;
1447         }
1448
1449         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1450         if (!strstr(tmp, tmp2)) {
1451                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1452                 td_verror(td, EINVAL, "iosched_switch");
1453                 fclose(f);
1454                 return 1;
1455         }
1456
1457         fclose(f);
1458         return 0;
1459 }
1460
1461 static int switch_ioscheduler(struct thread_data *td)
1462 {
1463         struct fio_file *f;
1464         unsigned int i;
1465         int ret = 0;
1466
1467         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1468                 return 0;
1469
1470         assert(td->files && td->files[0]);
1471
1472         for_each_file(td, f, i) {
1473
1474                 /* Only consider regular files and block device files */
1475                 switch (f->filetype) {
1476                 case FIO_TYPE_FILE:
1477                 case FIO_TYPE_BLOCK:
1478                         /*
1479                          * Make sure that the device hosting the file could
1480                          * be determined.
1481                          */
1482                         if (!f->du)
1483                                 continue;
1484                         break;
1485                 case FIO_TYPE_CHAR:
1486                 case FIO_TYPE_PIPE:
1487                 default:
1488                         continue;
1489                 }
1490
1491                 ret = set_ioscheduler(td, f);
1492                 if (ret)
1493                         return ret;
1494         }
1495
1496         return 0;
1497 }
1498
1499 #else
1500
1501 static int switch_ioscheduler(struct thread_data *td)
1502 {
1503         return 0;
1504 }
1505
1506 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1507
1508 static bool keep_running(struct thread_data *td)
1509 {
1510         unsigned long long limit;
1511
1512         if (td->done)
1513                 return false;
1514         if (td->terminate)
1515                 return false;
1516         if (td->o.time_based)
1517                 return true;
1518         if (td->o.loops) {
1519                 td->o.loops--;
1520                 return true;
1521         }
1522         if (exceeds_number_ios(td))
1523                 return false;
1524
1525         if (td->o.io_size)
1526                 limit = td->o.io_size;
1527         else
1528                 limit = td->o.size;
1529
1530         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1531                 uint64_t diff;
1532
1533                 /*
1534                  * If the difference is less than the maximum IO size, we
1535                  * are done.
1536                  */
1537                 diff = limit - ddir_rw_sum(td->io_bytes);
1538                 if (diff < td_max_bs(td))
1539                         return false;
1540
1541                 if (fio_files_done(td) && !td->o.io_size)
1542                         return false;
1543
1544                 return true;
1545         }
1546
1547         return false;
1548 }
1549
1550 static int exec_string(struct thread_options *o, const char *string,
1551                        const char *mode)
1552 {
1553         int ret;
1554         char *str;
1555
1556         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1557                 return -1;
1558
1559         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1560                  o->name, mode);
1561         ret = system(str);
1562         if (ret == -1)
1563                 log_err("fio: exec of cmd <%s> failed\n", str);
1564
1565         free(str);
1566         return ret;
1567 }
1568
1569 /*
1570  * Dry run to compute correct state of numberio for verification.
1571  */
1572 static uint64_t do_dry_run(struct thread_data *td)
1573 {
1574         td_set_runstate(td, TD_RUNNING);
1575
1576         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1577                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1578                 struct io_u *io_u;
1579                 int ret;
1580
1581                 if (td->terminate || td->done)
1582                         break;
1583
1584                 io_u = get_io_u(td);
1585                 if (IS_ERR_OR_NULL(io_u))
1586                         break;
1587
1588                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1589                 io_u->error = 0;
1590                 io_u->resid = 0;
1591                 if (ddir_rw(acct_ddir(io_u)))
1592                         td->io_issues[acct_ddir(io_u)]++;
1593                 if (ddir_rw(io_u->ddir)) {
1594                         io_u_mark_depth(td, 1);
1595                         td->ts.total_io_u[io_u->ddir]++;
1596                 }
1597
1598                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1599                     td->o.do_verify &&
1600                     td->o.verify != VERIFY_NONE &&
1601                     !td->o.experimental_verify)
1602                         log_io_piece(td, io_u);
1603
1604                 ret = io_u_sync_complete(td, io_u);
1605                 (void) ret;
1606         }
1607
1608         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1609 }
1610
1611 struct fork_data {
1612         struct thread_data *td;
1613         struct sk_out *sk_out;
1614 };
1615
1616 /*
1617  * Entry point for the thread based jobs. The process based jobs end up
1618  * here as well, after a little setup.
1619  */
1620 static void *thread_main(void *data)
1621 {
1622         struct fork_data *fd = data;
1623         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1624         struct thread_data *td = fd->td;
1625         struct thread_options *o = &td->o;
1626         struct sk_out *sk_out = fd->sk_out;
1627         uint64_t bytes_done[DDIR_RWDIR_CNT];
1628         int deadlock_loop_cnt;
1629         bool clear_state;
1630         int res, ret;
1631
1632         sk_out_assign(sk_out);
1633         free(fd);
1634
1635         if (!o->use_thread) {
1636                 setsid();
1637                 td->pid = getpid();
1638         } else
1639                 td->pid = gettid();
1640
1641         fio_local_clock_init();
1642
1643         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1644
1645         if (is_backend)
1646                 fio_server_send_start(td);
1647
1648         INIT_FLIST_HEAD(&td->io_log_list);
1649         INIT_FLIST_HEAD(&td->io_hist_list);
1650         INIT_FLIST_HEAD(&td->verify_list);
1651         INIT_FLIST_HEAD(&td->trim_list);
1652         td->io_hist_tree = RB_ROOT;
1653
1654         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1655         if (ret) {
1656                 td_verror(td, ret, "mutex_cond_init_pshared");
1657                 goto err;
1658         }
1659         ret = cond_init_pshared(&td->verify_cond);
1660         if (ret) {
1661                 td_verror(td, ret, "mutex_cond_pshared");
1662                 goto err;
1663         }
1664
1665         td_set_runstate(td, TD_INITIALIZED);
1666         dprint(FD_MUTEX, "up startup_sem\n");
1667         fio_sem_up(startup_sem);
1668         dprint(FD_MUTEX, "wait on td->sem\n");
1669         fio_sem_down(td->sem);
1670         dprint(FD_MUTEX, "done waiting on td->sem\n");
1671
1672         /*
1673          * A new gid requires privilege, so we need to do this before setting
1674          * the uid.
1675          */
1676         if (o->gid != -1U && setgid(o->gid)) {
1677                 td_verror(td, errno, "setgid");
1678                 goto err;
1679         }
1680         if (o->uid != -1U && setuid(o->uid)) {
1681                 td_verror(td, errno, "setuid");
1682                 goto err;
1683         }
1684
1685         td_zone_gen_index(td);
1686
1687         /*
1688          * Do this early, we don't want the compress threads to be limited
1689          * to the same CPUs as the IO workers. So do this before we set
1690          * any potential CPU affinity
1691          */
1692         if (iolog_compress_init(td, sk_out))
1693                 goto err;
1694
1695         /*
1696          * If we have a gettimeofday() thread, make sure we exclude that
1697          * thread from this job
1698          */
1699         if (o->gtod_cpu)
1700                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1701
1702         /*
1703          * Set affinity first, in case it has an impact on the memory
1704          * allocations.
1705          */
1706         if (fio_option_is_set(o, cpumask)) {
1707                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1708                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1709                         if (!ret) {
1710                                 log_err("fio: no CPUs set\n");
1711                                 log_err("fio: Try increasing number of available CPUs\n");
1712                                 td_verror(td, EINVAL, "cpus_split");
1713                                 goto err;
1714                         }
1715                 }
1716                 ret = fio_setaffinity(td->pid, o->cpumask);
1717                 if (ret == -1) {
1718                         td_verror(td, errno, "cpu_set_affinity");
1719                         goto err;
1720                 }
1721         }
1722
1723 #ifdef CONFIG_LIBNUMA
1724         /* numa node setup */
1725         if (fio_option_is_set(o, numa_cpunodes) ||
1726             fio_option_is_set(o, numa_memnodes)) {
1727                 struct bitmask *mask;
1728
1729                 if (numa_available() < 0) {
1730                         td_verror(td, errno, "Does not support NUMA API\n");
1731                         goto err;
1732                 }
1733
1734                 if (fio_option_is_set(o, numa_cpunodes)) {
1735                         mask = numa_parse_nodestring(o->numa_cpunodes);
1736                         ret = numa_run_on_node_mask(mask);
1737                         numa_free_nodemask(mask);
1738                         if (ret == -1) {
1739                                 td_verror(td, errno, \
1740                                         "numa_run_on_node_mask failed\n");
1741                                 goto err;
1742                         }
1743                 }
1744
1745                 if (fio_option_is_set(o, numa_memnodes)) {
1746                         mask = NULL;
1747                         if (o->numa_memnodes)
1748                                 mask = numa_parse_nodestring(o->numa_memnodes);
1749
1750                         switch (o->numa_mem_mode) {
1751                         case MPOL_INTERLEAVE:
1752                                 numa_set_interleave_mask(mask);
1753                                 break;
1754                         case MPOL_BIND:
1755                                 numa_set_membind(mask);
1756                                 break;
1757                         case MPOL_LOCAL:
1758                                 numa_set_localalloc();
1759                                 break;
1760                         case MPOL_PREFERRED:
1761                                 numa_set_preferred(o->numa_mem_prefer_node);
1762                                 break;
1763                         case MPOL_DEFAULT:
1764                         default:
1765                                 break;
1766                         }
1767
1768                         if (mask)
1769                                 numa_free_nodemask(mask);
1770
1771                 }
1772         }
1773 #endif
1774
1775         if (fio_pin_memory(td))
1776                 goto err;
1777
1778         /*
1779          * May alter parameters that init_io_u() will use, so we need to
1780          * do this first.
1781          */
1782         if (!init_iolog(td))
1783                 goto err;
1784
1785         /* ioprio_set() has to be done before td_io_init() */
1786         if (fio_option_is_set(o, ioprio) ||
1787             fio_option_is_set(o, ioprio_class)) {
1788                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1789                 if (ret == -1) {
1790                         td_verror(td, errno, "ioprio_set");
1791                         goto err;
1792                 }
1793                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1794                 td->ts.ioprio = td->ioprio;
1795         }
1796
1797         if (td_io_init(td))
1798                 goto err;
1799
1800         if (init_io_u(td))
1801                 goto err;
1802
1803         if (td->io_ops->post_init && td->io_ops->post_init(td))
1804                 goto err;
1805
1806         if (o->verify_async && verify_async_init(td))
1807                 goto err;
1808
1809         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1810                 goto err;
1811
1812         errno = 0;
1813         if (nice(o->nice) == -1 && errno != 0) {
1814                 td_verror(td, errno, "nice");
1815                 goto err;
1816         }
1817
1818         if (o->ioscheduler && switch_ioscheduler(td))
1819                 goto err;
1820
1821         if (!o->create_serialize && setup_files(td))
1822                 goto err;
1823
1824         if (!init_random_map(td))
1825                 goto err;
1826
1827         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1828                 goto err;
1829
1830         if (o->pre_read && !pre_read_files(td))
1831                 goto err;
1832
1833         fio_verify_init(td);
1834
1835         if (rate_submit_init(td, sk_out))
1836                 goto err;
1837
1838         set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1839         fio_getrusage(&td->ru_start);
1840         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1841         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1842         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1843
1844         init_thinktime(td);
1845
1846         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1847                         o->ratemin[DDIR_TRIM]) {
1848                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1849                                         sizeof(td->bw_sample_time));
1850                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1851                                         sizeof(td->bw_sample_time));
1852                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1853                                         sizeof(td->bw_sample_time));
1854         }
1855
1856         memset(bytes_done, 0, sizeof(bytes_done));
1857         clear_state = false;
1858
1859         while (keep_running(td)) {
1860                 uint64_t verify_bytes;
1861
1862                 fio_gettime(&td->start, NULL);
1863                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1864
1865                 if (clear_state) {
1866                         clear_io_state(td, 0);
1867
1868                         if (o->unlink_each_loop && unlink_all_files(td))
1869                                 break;
1870                 }
1871
1872                 prune_io_piece_log(td);
1873
1874                 if (td->o.verify_only && td_write(td))
1875                         verify_bytes = do_dry_run(td);
1876                 else {
1877                         do_io(td, bytes_done);
1878
1879                         if (!ddir_rw_sum(bytes_done)) {
1880                                 fio_mark_td_terminate(td);
1881                                 verify_bytes = 0;
1882                         } else {
1883                                 verify_bytes = bytes_done[DDIR_WRITE] +
1884                                                 bytes_done[DDIR_TRIM];
1885                         }
1886                 }
1887
1888                 /*
1889                  * If we took too long to shut down, the main thread could
1890                  * already consider us reaped/exited. If that happens, break
1891                  * out and clean up.
1892                  */
1893                 if (td->runstate >= TD_EXITED)
1894                         break;
1895
1896                 clear_state = true;
1897
1898                 /*
1899                  * Make sure we've successfully updated the rusage stats
1900                  * before waiting on the stat mutex. Otherwise we could have
1901                  * the stat thread holding stat mutex and waiting for
1902                  * the rusage_sem, which would never get upped because
1903                  * this thread is waiting for the stat mutex.
1904                  */
1905                 deadlock_loop_cnt = 0;
1906                 do {
1907                         check_update_rusage(td);
1908                         if (!fio_sem_down_trylock(stat_sem))
1909                                 break;
1910                         usleep(1000);
1911                         if (deadlock_loop_cnt++ > 5000) {
1912                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1913                                 td->error = EDEADLK;
1914                                 goto err;
1915                         }
1916                 } while (1);
1917
1918                 if (td_read(td) && td->io_bytes[DDIR_READ])
1919                         update_runtime(td, elapsed_us, DDIR_READ);
1920                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1921                         update_runtime(td, elapsed_us, DDIR_WRITE);
1922                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1923                         update_runtime(td, elapsed_us, DDIR_TRIM);
1924                 fio_gettime(&td->start, NULL);
1925                 fio_sem_up(stat_sem);
1926
1927                 if (td->error || td->terminate)
1928                         break;
1929
1930                 if (!o->do_verify ||
1931                     o->verify == VERIFY_NONE ||
1932                     td_ioengine_flagged(td, FIO_UNIDIR))
1933                         continue;
1934
1935                 clear_io_state(td, 0);
1936
1937                 fio_gettime(&td->start, NULL);
1938
1939                 do_verify(td, verify_bytes);
1940
1941                 /*
1942                  * See comment further up for why this is done here.
1943                  */
1944                 check_update_rusage(td);
1945
1946                 fio_sem_down(stat_sem);
1947                 update_runtime(td, elapsed_us, DDIR_READ);
1948                 fio_gettime(&td->start, NULL);
1949                 fio_sem_up(stat_sem);
1950
1951                 if (td->error || td->terminate)
1952                         break;
1953         }
1954
1955         /*
1956          * Acquire this lock if we were doing overlap checking in
1957          * offload mode so that we don't clean up this job while
1958          * another thread is checking its io_u's for overlap
1959          */
1960         if (td_offload_overlap(td)) {
1961                 int res = pthread_mutex_lock(&overlap_check);
1962                 assert(res == 0);
1963         }
1964         td_set_runstate(td, TD_FINISHING);
1965         if (td_offload_overlap(td)) {
1966                 res = pthread_mutex_unlock(&overlap_check);
1967                 assert(res == 0);
1968         }
1969
1970         update_rusage_stat(td);
1971         td->ts.total_run_time = mtime_since_now(&td->epoch);
1972         for_each_rw_ddir(ddir) {
1973                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1974         }
1975
1976         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1977             (td->o.verify != VERIFY_NONE && td_write(td)))
1978                 verify_save_state(td->thread_number);
1979
1980         fio_unpin_memory(td);
1981
1982         td_writeout_logs(td, true);
1983
1984         iolog_compress_exit(td);
1985         rate_submit_exit(td);
1986
1987         if (o->exec_postrun)
1988                 exec_string(o, o->exec_postrun, "postrun");
1989
1990         if (exitall_on_terminate || (o->exitall_error && td->error))
1991                 fio_terminate_threads(td->groupid, td->o.exit_what);
1992
1993 err:
1994         if (td->error)
1995                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1996                                                         td->verror);
1997
1998         if (o->verify_async)
1999                 verify_async_exit(td);
2000
2001         close_and_free_files(td);
2002         cleanup_io_u(td);
2003         close_ioengine(td);
2004         cgroup_shutdown(td, cgroup_mnt);
2005         verify_free_state(td);
2006         td_zone_free_index(td);
2007
2008         if (fio_option_is_set(o, cpumask)) {
2009                 ret = fio_cpuset_exit(&o->cpumask);
2010                 if (ret)
2011                         td_verror(td, ret, "fio_cpuset_exit");
2012         }
2013
2014         /*
2015          * do this very late, it will log file closing as well
2016          */
2017         if (o->write_iolog_file)
2018                 write_iolog_close(td);
2019         if (td->io_log_rfile)
2020                 fclose(td->io_log_rfile);
2021
2022         td_set_runstate(td, TD_EXITED);
2023
2024         /*
2025          * Do this last after setting our runstate to exited, so we
2026          * know that the stat thread is signaled.
2027          */
2028         check_update_rusage(td);
2029
2030         sk_out_drop();
2031         return (void *) (uintptr_t) td->error;
2032 }
2033
2034 /*
2035  * Run over the job map and reap the threads that have exited, if any.
2036  */
2037 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2038                          uint64_t *m_rate)
2039 {
2040         struct thread_data *td;
2041         unsigned int cputhreads, realthreads, pending;
2042         int i, status, ret;
2043
2044         /*
2045          * reap exited threads (TD_EXITED -> TD_REAPED)
2046          */
2047         realthreads = pending = cputhreads = 0;
2048         for_each_td(td, i) {
2049                 int flags = 0;
2050
2051                  if (!strcmp(td->o.ioengine, "cpuio"))
2052                         cputhreads++;
2053                 else
2054                         realthreads++;
2055
2056                 if (!td->pid) {
2057                         pending++;
2058                         continue;
2059                 }
2060                 if (td->runstate == TD_REAPED)
2061                         continue;
2062                 if (td->o.use_thread) {
2063                         if (td->runstate == TD_EXITED) {
2064                                 td_set_runstate(td, TD_REAPED);
2065                                 goto reaped;
2066                         }
2067                         continue;
2068                 }
2069
2070                 flags = WNOHANG;
2071                 if (td->runstate == TD_EXITED)
2072                         flags = 0;
2073
2074                 /*
2075                  * check if someone quit or got killed in an unusual way
2076                  */
2077                 ret = waitpid(td->pid, &status, flags);
2078                 if (ret < 0) {
2079                         if (errno == ECHILD) {
2080                                 log_err("fio: pid=%d disappeared %d\n",
2081                                                 (int) td->pid, td->runstate);
2082                                 td->sig = ECHILD;
2083                                 td_set_runstate(td, TD_REAPED);
2084                                 goto reaped;
2085                         }
2086                         perror("waitpid");
2087                 } else if (ret == td->pid) {
2088                         if (WIFSIGNALED(status)) {
2089                                 int sig = WTERMSIG(status);
2090
2091                                 if (sig != SIGTERM && sig != SIGUSR2)
2092                                         log_err("fio: pid=%d, got signal=%d\n",
2093                                                         (int) td->pid, sig);
2094                                 td->sig = sig;
2095                                 td_set_runstate(td, TD_REAPED);
2096                                 goto reaped;
2097                         }
2098                         if (WIFEXITED(status)) {
2099                                 if (WEXITSTATUS(status) && !td->error)
2100                                         td->error = WEXITSTATUS(status);
2101
2102                                 td_set_runstate(td, TD_REAPED);
2103                                 goto reaped;
2104                         }
2105                 }
2106
2107                 /*
2108                  * If the job is stuck, do a forceful timeout of it and
2109                  * move on.
2110                  */
2111                 if (td->terminate &&
2112                     td->runstate < TD_FSYNCING &&
2113                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2114                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2115                                 "%lu seconds, it appears to be stuck. Doing "
2116                                 "forceful exit of this job.\n",
2117                                 td->o.name, td->runstate,
2118                                 (unsigned long) time_since_now(&td->terminate_time));
2119                         td_set_runstate(td, TD_REAPED);
2120                         goto reaped;
2121                 }
2122
2123                 /*
2124                  * thread is not dead, continue
2125                  */
2126                 pending++;
2127                 continue;
2128 reaped:
2129                 (*nr_running)--;
2130                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2131                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2132                 if (!td->pid)
2133                         pending--;
2134
2135                 if (td->error)
2136                         exit_value++;
2137
2138                 done_secs += mtime_since_now(&td->epoch) / 1000;
2139                 profile_td_exit(td);
2140                 flow_exit_job(td);
2141         }
2142
2143         if (*nr_running == cputhreads && !pending && realthreads)
2144                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2145 }
2146
2147 static bool __check_trigger_file(void)
2148 {
2149         struct stat sb;
2150
2151         if (!trigger_file)
2152                 return false;
2153
2154         if (stat(trigger_file, &sb))
2155                 return false;
2156
2157         if (unlink(trigger_file) < 0)
2158                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2159                                                         strerror(errno));
2160
2161         return true;
2162 }
2163
2164 static bool trigger_timedout(void)
2165 {
2166         if (trigger_timeout)
2167                 if (time_since_genesis() >= trigger_timeout) {
2168                         trigger_timeout = 0;
2169                         return true;
2170                 }
2171
2172         return false;
2173 }
2174
2175 void exec_trigger(const char *cmd)
2176 {
2177         int ret;
2178
2179         if (!cmd || cmd[0] == '\0')
2180                 return;
2181
2182         ret = system(cmd);
2183         if (ret == -1)
2184                 log_err("fio: failed executing %s trigger\n", cmd);
2185 }
2186
2187 void check_trigger_file(void)
2188 {
2189         if (__check_trigger_file() || trigger_timedout()) {
2190                 if (nr_clients)
2191                         fio_clients_send_trigger(trigger_remote_cmd);
2192                 else {
2193                         verify_save_state(IO_LIST_ALL);
2194                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2195                         exec_trigger(trigger_cmd);
2196                 }
2197         }
2198 }
2199
2200 static int fio_verify_load_state(struct thread_data *td)
2201 {
2202         int ret;
2203
2204         if (!td->o.verify_state)
2205                 return 0;
2206
2207         if (is_backend) {
2208                 void *data;
2209
2210                 ret = fio_server_get_verify_state(td->o.name,
2211                                         td->thread_number - 1, &data);
2212                 if (!ret)
2213                         verify_assign_state(td, data);
2214         } else {
2215                 char prefix[PATH_MAX];
2216
2217                 if (aux_path)
2218                         sprintf(prefix, "%s%clocal", aux_path,
2219                                         FIO_OS_PATH_SEPARATOR);
2220                 else
2221                         strcpy(prefix, "local");
2222                 ret = verify_load_state(td, prefix);
2223         }
2224
2225         return ret;
2226 }
2227
2228 static void do_usleep(unsigned int usecs)
2229 {
2230         check_for_running_stats();
2231         check_trigger_file();
2232         usleep(usecs);
2233 }
2234
2235 static bool check_mount_writes(struct thread_data *td)
2236 {
2237         struct fio_file *f;
2238         unsigned int i;
2239
2240         if (!td_write(td) || td->o.allow_mounted_write)
2241                 return false;
2242
2243         /*
2244          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2245          * are mkfs'd and mounted.
2246          */
2247         for_each_file(td, f, i) {
2248 #ifdef FIO_HAVE_CHARDEV_SIZE
2249                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2250 #else
2251                 if (f->filetype != FIO_TYPE_BLOCK)
2252 #endif
2253                         continue;
2254                 if (device_is_mounted(f->file_name))
2255                         goto mounted;
2256         }
2257
2258         return false;
2259 mounted:
2260         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2261         return true;
2262 }
2263
2264 static bool waitee_running(struct thread_data *me)
2265 {
2266         const char *waitee = me->o.wait_for;
2267         const char *self = me->o.name;
2268         struct thread_data *td;
2269         int i;
2270
2271         if (!waitee)
2272                 return false;
2273
2274         for_each_td(td, i) {
2275                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2276                         continue;
2277
2278                 if (td->runstate < TD_EXITED) {
2279                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2280                                         self, td->o.name,
2281                                         runstate_to_name(td->runstate));
2282                         return true;
2283                 }
2284         }
2285
2286         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2287         return false;
2288 }
2289
2290 /*
2291  * Main function for kicking off and reaping jobs, as needed.
2292  */
2293 static void run_threads(struct sk_out *sk_out)
2294 {
2295         struct thread_data *td;
2296         unsigned int i, todo, nr_running, nr_started;
2297         uint64_t m_rate, t_rate;
2298         uint64_t spent;
2299
2300         if (fio_gtod_offload && fio_start_gtod_thread())
2301                 return;
2302
2303         fio_idle_prof_init();
2304
2305         set_sig_handlers();
2306
2307         nr_thread = nr_process = 0;
2308         for_each_td(td, i) {
2309                 if (check_mount_writes(td))
2310                         return;
2311                 if (td->o.use_thread)
2312                         nr_thread++;
2313                 else
2314                         nr_process++;
2315         }
2316
2317         if (output_format & FIO_OUTPUT_NORMAL) {
2318                 struct buf_output out;
2319
2320                 buf_output_init(&out);
2321                 __log_buf(&out, "Starting ");
2322                 if (nr_thread)
2323                         __log_buf(&out, "%d thread%s", nr_thread,
2324                                                 nr_thread > 1 ? "s" : "");
2325                 if (nr_process) {
2326                         if (nr_thread)
2327                                 __log_buf(&out, " and ");
2328                         __log_buf(&out, "%d process%s", nr_process,
2329                                                 nr_process > 1 ? "es" : "");
2330                 }
2331                 __log_buf(&out, "\n");
2332                 log_info_buf(out.buf, out.buflen);
2333                 buf_output_free(&out);
2334         }
2335
2336         todo = thread_number;
2337         nr_running = 0;
2338         nr_started = 0;
2339         m_rate = t_rate = 0;
2340
2341         for_each_td(td, i) {
2342                 print_status_init(td->thread_number - 1);
2343
2344                 if (!td->o.create_serialize)
2345                         continue;
2346
2347                 if (fio_verify_load_state(td))
2348                         goto reap;
2349
2350                 /*
2351                  * do file setup here so it happens sequentially,
2352                  * we don't want X number of threads getting their
2353                  * client data interspersed on disk
2354                  */
2355                 if (setup_files(td)) {
2356 reap:
2357                         exit_value++;
2358                         if (td->error)
2359                                 log_err("fio: pid=%d, err=%d/%s\n",
2360                                         (int) td->pid, td->error, td->verror);
2361                         td_set_runstate(td, TD_REAPED);
2362                         todo--;
2363                 } else {
2364                         struct fio_file *f;
2365                         unsigned int j;
2366
2367                         /*
2368                          * for sharing to work, each job must always open
2369                          * its own files. so close them, if we opened them
2370                          * for creation
2371                          */
2372                         for_each_file(td, f, j) {
2373                                 if (fio_file_open(f))
2374                                         td_io_close_file(td, f);
2375                         }
2376                 }
2377         }
2378
2379         /* start idle threads before io threads start to run */
2380         fio_idle_prof_start();
2381
2382         set_genesis_time();
2383
2384         while (todo) {
2385                 struct thread_data *map[REAL_MAX_JOBS];
2386                 struct timespec this_start;
2387                 int this_jobs = 0, left;
2388                 struct fork_data *fd;
2389
2390                 /*
2391                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2392                  */
2393                 for_each_td(td, i) {
2394                         if (td->runstate != TD_NOT_CREATED)
2395                                 continue;
2396
2397                         /*
2398                          * never got a chance to start, killed by other
2399                          * thread for some reason
2400                          */
2401                         if (td->terminate) {
2402                                 todo--;
2403                                 continue;
2404                         }
2405
2406                         if (td->o.start_delay) {
2407                                 spent = utime_since_genesis();
2408
2409                                 if (td->o.start_delay > spent)
2410                                         continue;
2411                         }
2412
2413                         if (td->o.stonewall && (nr_started || nr_running)) {
2414                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2415                                                         td->o.name);
2416                                 break;
2417                         }
2418
2419                         if (waitee_running(td)) {
2420                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2421                                                 td->o.name, td->o.wait_for);
2422                                 continue;
2423                         }
2424
2425                         init_disk_util(td);
2426
2427                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2428                         td->update_rusage = 0;
2429
2430                         /*
2431                          * Set state to created. Thread will transition
2432                          * to TD_INITIALIZED when it's done setting up.
2433                          */
2434                         td_set_runstate(td, TD_CREATED);
2435                         map[this_jobs++] = td;
2436                         nr_started++;
2437
2438                         fd = calloc(1, sizeof(*fd));
2439                         fd->td = td;
2440                         fd->sk_out = sk_out;
2441
2442                         if (td->o.use_thread) {
2443                                 int ret;
2444
2445                                 dprint(FD_PROCESS, "will pthread_create\n");
2446                                 ret = pthread_create(&td->thread, NULL,
2447                                                         thread_main, fd);
2448                                 if (ret) {
2449                                         log_err("pthread_create: %s\n",
2450                                                         strerror(ret));
2451                                         free(fd);
2452                                         nr_started--;
2453                                         break;
2454                                 }
2455                                 fd = NULL;
2456                                 ret = pthread_detach(td->thread);
2457                                 if (ret)
2458                                         log_err("pthread_detach: %s",
2459                                                         strerror(ret));
2460                         } else {
2461                                 pid_t pid;
2462                                 dprint(FD_PROCESS, "will fork\n");
2463                                 pid = fork();
2464                                 if (!pid) {
2465                                         int ret;
2466
2467                                         ret = (int)(uintptr_t)thread_main(fd);
2468                                         _exit(ret);
2469                                 } else if (i == fio_debug_jobno)
2470                                         *fio_debug_jobp = pid;
2471                         }
2472                         dprint(FD_MUTEX, "wait on startup_sem\n");
2473                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2474                                 log_err("fio: job startup hung? exiting.\n");
2475                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2476                                 fio_abort = true;
2477                                 nr_started--;
2478                                 free(fd);
2479                                 break;
2480                         }
2481                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2482                 }
2483
2484                 /*
2485                  * Wait for the started threads to transition to
2486                  * TD_INITIALIZED.
2487                  */
2488                 fio_gettime(&this_start, NULL);
2489                 left = this_jobs;
2490                 while (left && !fio_abort) {
2491                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2492                                 break;
2493
2494                         do_usleep(100000);
2495
2496                         for (i = 0; i < this_jobs; i++) {
2497                                 td = map[i];
2498                                 if (!td)
2499                                         continue;
2500                                 if (td->runstate == TD_INITIALIZED) {
2501                                         map[i] = NULL;
2502                                         left--;
2503                                 } else if (td->runstate >= TD_EXITED) {
2504                                         map[i] = NULL;
2505                                         left--;
2506                                         todo--;
2507                                         nr_running++; /* work-around... */
2508                                 }
2509                         }
2510                 }
2511
2512                 if (left) {
2513                         log_err("fio: %d job%s failed to start\n", left,
2514                                         left > 1 ? "s" : "");
2515                         for (i = 0; i < this_jobs; i++) {
2516                                 td = map[i];
2517                                 if (!td)
2518                                         continue;
2519                                 kill(td->pid, SIGTERM);
2520                         }
2521                         break;
2522                 }
2523
2524                 /*
2525                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2526                  */
2527                 for_each_td(td, i) {
2528                         if (td->runstate != TD_INITIALIZED)
2529                                 continue;
2530
2531                         if (in_ramp_time(td))
2532                                 td_set_runstate(td, TD_RAMP);
2533                         else
2534                                 td_set_runstate(td, TD_RUNNING);
2535                         nr_running++;
2536                         nr_started--;
2537                         m_rate += ddir_rw_sum(td->o.ratemin);
2538                         t_rate += ddir_rw_sum(td->o.rate);
2539                         todo--;
2540                         fio_sem_up(td->sem);
2541                 }
2542
2543                 reap_threads(&nr_running, &t_rate, &m_rate);
2544
2545                 if (todo)
2546                         do_usleep(100000);
2547         }
2548
2549         while (nr_running) {
2550                 reap_threads(&nr_running, &t_rate, &m_rate);
2551                 do_usleep(10000);
2552         }
2553
2554         fio_idle_prof_stop();
2555
2556         update_io_ticks();
2557 }
2558
2559 static void free_disk_util(void)
2560 {
2561         disk_util_prune_entries();
2562         helper_thread_destroy();
2563 }
2564
2565 int fio_backend(struct sk_out *sk_out)
2566 {
2567         struct thread_data *td;
2568         int i;
2569
2570         if (exec_profile) {
2571                 if (load_profile(exec_profile))
2572                         return 1;
2573                 free(exec_profile);
2574                 exec_profile = NULL;
2575         }
2576         if (!thread_number)
2577                 return 0;
2578
2579         if (write_bw_log) {
2580                 struct log_params p = {
2581                         .log_type = IO_LOG_TYPE_BW,
2582                 };
2583
2584                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2585                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2586                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2587         }
2588
2589         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2590         if (!sk_out)
2591                 is_local_backend = true;
2592         if (startup_sem == NULL)
2593                 return 1;
2594
2595         set_genesis_time();
2596         stat_init();
2597         if (helper_thread_create(startup_sem, sk_out))
2598                 log_err("fio: failed to create helper thread\n");
2599
2600         cgroup_list = smalloc(sizeof(*cgroup_list));
2601         if (cgroup_list)
2602                 INIT_FLIST_HEAD(cgroup_list);
2603
2604         run_threads(sk_out);
2605
2606         helper_thread_exit();
2607
2608         if (!fio_abort) {
2609                 __show_run_stats();
2610                 if (write_bw_log) {
2611                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2612                                 struct io_log *log = agg_io_log[i];
2613
2614                                 flush_log(log, false);
2615                                 free_log(log);
2616                         }
2617                 }
2618         }
2619
2620         for_each_td(td, i) {
2621                 struct thread_stat *ts = &td->ts;
2622
2623                 free_clat_prio_stats(ts);
2624                 steadystate_free(td);
2625                 fio_options_free(td);
2626                 fio_dump_options_free(td);
2627                 if (td->rusage_sem) {
2628                         fio_sem_remove(td->rusage_sem);
2629                         td->rusage_sem = NULL;
2630                 }
2631                 fio_sem_remove(td->sem);
2632                 td->sem = NULL;
2633         }
2634
2635         free_disk_util();
2636         if (cgroup_list) {
2637                 cgroup_kill(cgroup_list);
2638                 sfree(cgroup_list);
2639         }
2640
2641         fio_sem_remove(startup_sem);
2642         stat_exit();
2643         return exit_value;
2644 }