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 #ifdef WIN32
  94 static void sig_break(int sig)
  95 {
  96         struct thread_data *td;
  97         int i;
  98
  99         sig_int(sig);
 100
 101         /**
 102          * Windows terminates all job processes on SIGBREAK after the handler
 103          * returns, so give them time to wrap-up and give stats
 104          */
 105         for_each_td(td, i) {
 106                 while (td->runstate < TD_EXITED)
 107                         sleep(1);
 108         }
 109 }
 110 #endif
 111
 112 void sig_show_status(int sig)
 113 {
 114         show_running_run_stats();
 115 }
 116
 117 static void set_sig_handlers(void)
 118 {
 119         struct sigaction act;
 120
 121         memset(&act, 0, sizeof(act));
 122         act.sa_handler = sig_int;
 123         act.sa_flags = SA_RESTART;
 124         sigaction(SIGINT, &act, NULL);
 125
 126         memset(&act, 0, sizeof(act));
 127         act.sa_handler = sig_int;
 128         act.sa_flags = SA_RESTART;
 129         sigaction(SIGTERM, &act, NULL);
 130
 131 /* Windows uses SIGBREAK as a quit signal from other applications */
 132 #ifdef WIN32
 133         memset(&act, 0, sizeof(act));
 134         act.sa_handler = sig_break;
 135         act.sa_flags = SA_RESTART;
 136         sigaction(SIGBREAK, &act, NULL);
 137 #endif
 138
 139         memset(&act, 0, sizeof(act));
 140         act.sa_handler = sig_show_status;
 141         act.sa_flags = SA_RESTART;
 142         sigaction(SIGUSR1, &act, NULL);
 143
 144         if (is_backend) {
 145                 memset(&act, 0, sizeof(act));
 146                 act.sa_handler = sig_int;
 147                 act.sa_flags = SA_RESTART;
 148                 sigaction(SIGPIPE, &act, NULL);
 149         }
 150 }
 151
 152 /*
 153  * Check if we are above the minimum rate given.
 154  */
 155 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 156                              enum fio_ddir ddir)
 157 {
 158         unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
 159         unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
 160         unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
 161         unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
 162
 163         assert(ddir_rw(ddir));
 164
 165         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 166                 return false;
 167
 168         /*
 169          * allow a 2 second settle period in the beginning
 170          */
 171         if (mtime_since(&td->start, now) < 2000)
 172                 return false;
 173
 174         /*
 175          * if last_rate_check_blocks or last_rate_check_bytes is set,
 176          * we can compute a rate per ratecycle
 177          */
 178         if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
 179                 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
 180                 if (spent < td->o.ratecycle || spent==0)
 181                         return false;
 182
 183                 if (td->o.ratemin[ddir]) {
 184                         /*
 185                          * check bandwidth specified rate
 186                          */
 187                         unsigned long long current_rate_bytes =
 188                                 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
 189                         if (current_rate_bytes < option_rate_bytes_min) {
 190                                 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 191                                         td->o.name, option_rate_bytes_min, current_rate_bytes);
 192                                 return true;
 193                         }
 194                 } else {
 195                         /*
 196                          * checks iops specified rate
 197                          */
 198                         unsigned long long current_rate_iops =
 199                                 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
 200
 201                         if (current_rate_iops < option_rate_iops_min) {
 202                                 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 203                                         td->o.name, option_rate_iops_min, current_rate_iops);
 204                                 return true;
 205                         }
 206                 }
 207         }
 208
 209         td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
 210         td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
 211         memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
 212         return false;
 213 }
 214
 215 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 216 {
 217         bool ret = false;
 218
 219         for_each_rw_ddir(ddir) {
 220                 if (td->bytes_done[ddir])
 221                         ret |= __check_min_rate(td, now, ddir);
 222         }
 223
 224         return ret;
 225 }
 226
 227 /*
 228  * When job exits, we can cancel the in-flight IO if we are using async
 229  * io. Attempt to do so.
 230  */
 231 static void cleanup_pending_aio(struct thread_data *td)
 232 {
 233         int r;
 234
 235         /*
 236          * get immediately available events, if any
 237          */
 238         r = io_u_queued_complete(td, 0);
 239
 240         /*
 241          * now cancel remaining active events
 242          */
 243         if (td->io_ops->cancel) {
 244                 struct io_u *io_u;
 245                 int i;
 246
 247                 io_u_qiter(&td->io_u_all, io_u, i) {
 248                         if (io_u->flags & IO_U_F_FLIGHT) {
 249                                 r = td->io_ops->cancel(td, io_u);
 250                                 if (!r)
 251                                         put_io_u(td, io_u);
 252                         }
 253                 }
 254         }
 255
 256         if (td->cur_depth)
 257                 r = io_u_queued_complete(td, td->cur_depth);
 258 }
 259
 260 /*
 261  * Helper to handle the final sync of a file. Works just like the normal
 262  * io path, just does everything sync.
 263  */
 264 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 265 {
 266         struct io_u *io_u = __get_io_u(td);
 267         enum fio_q_status ret;
 268
 269         if (!io_u)
 270                 return true;
 271
 272         io_u->ddir = DDIR_SYNC;
 273         io_u->file = f;
 274         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 275
 276         if (td_io_prep(td, io_u)) {
 277                 put_io_u(td, io_u);
 278                 return true;
 279         }
 280
 281 requeue:
 282         ret = td_io_queue(td, io_u);
 283         switch (ret) {
 284         case FIO_Q_QUEUED:
 285                 td_io_commit(td);
 286                 if (io_u_queued_complete(td, 1) < 0)
 287                         return true;
 288                 break;
 289         case FIO_Q_COMPLETED:
 290                 if (io_u->error) {
 291                         td_verror(td, io_u->error, "td_io_queue");
 292                         return true;
 293                 }
 294
 295                 if (io_u_sync_complete(td, io_u) < 0)
 296                         return true;
 297                 break;
 298         case FIO_Q_BUSY:
 299                 td_io_commit(td);
 300                 goto requeue;
 301         }
 302
 303         return false;
 304 }
 305
 306 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 307 {
 308         int ret, ret2;
 309
 310         if (fio_file_open(f))
 311                 return fio_io_sync(td, f);
 312
 313         if (td_io_open_file(td, f))
 314                 return 1;
 315
 316         ret = fio_io_sync(td, f);
 317         ret2 = 0;
 318         if (fio_file_open(f))
 319                 ret2 = td_io_close_file(td, f);
 320         return (ret || ret2);
 321 }
 322
 323 static inline void __update_ts_cache(struct thread_data *td)
 324 {
 325         fio_gettime(&td->ts_cache, NULL);
 326 }
 327
 328 static inline void update_ts_cache(struct thread_data *td)
 329 {
 330         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 331                 __update_ts_cache(td);
 332 }
 333
 334 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 335 {
 336         if (in_ramp_time(td))
 337                 return false;
 338         if (!td->o.timeout)
 339                 return false;
 340         if (utime_since(&td->epoch, t) >= td->o.timeout)
 341                 return true;
 342
 343         return false;
 344 }
 345
 346 /*
 347  * We need to update the runtime consistently in ms, but keep a running
 348  * tally of the current elapsed time in microseconds for sub millisecond
 349  * updates.
 350  */
 351 static inline void update_runtime(struct thread_data *td,
 352                                   unsigned long long *elapsed_us,
 353                                   const enum fio_ddir ddir)
 354 {
 355         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 356                 return;
 357
 358         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 359         elapsed_us[ddir] += utime_since_now(&td->start);
 360         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 361 }
 362
 363 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 364                                 int *retptr)
 365 {
 366         int ret = *retptr;
 367
 368         if (ret < 0 || td->error) {
 369                 int err = td->error;
 370                 enum error_type_bit eb;
 371
 372                 if (ret < 0)
 373                         err = -ret;
 374
 375                 eb = td_error_type(ddir, err);
 376                 if (!(td->o.continue_on_error & (1 << eb)))
 377                         return true;
 378
 379                 if (td_non_fatal_error(td, eb, err)) {
 380                         /*
 381                          * Continue with the I/Os in case of
 382                          * a non fatal error.
 383                          */
 384                         update_error_count(td, err);
 385                         td_clear_error(td);
 386                         *retptr = 0;
 387                         return false;
 388                 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
 389                         /*
 390                          * We expect to hit this error if
 391                          * fill_device option is set.
 392                          */
 393                         td_clear_error(td);
 394                         fio_mark_td_terminate(td);
 395                         return true;
 396                 } else {
 397                         /*
 398                          * Stop the I/O in case of a fatal
 399                          * error.
 400                          */
 401                         update_error_count(td, err);
 402                         return true;
 403                 }
 404         }
 405
 406         return false;
 407 }
 408
 409 static void check_update_rusage(struct thread_data *td)
 410 {
 411         if (td->update_rusage) {
 412                 td->update_rusage = 0;
 413                 update_rusage_stat(td);
 414                 fio_sem_up(td->rusage_sem);
 415         }
 416 }
 417
 418 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 419 {
 420         const int full = queue_full(td);
 421         int min_evts = 0;
 422         int ret;
 423
 424         if (td->flags & TD_F_REGROW_LOGS)
 425                 return io_u_quiesce(td);
 426
 427         /*
 428          * if the queue is full, we MUST reap at least 1 event
 429          */
 430         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 431         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 432                 min_evts = 1;
 433
 434         if (time && should_check_rate(td))
 435                 fio_gettime(time, NULL);
 436
 437         do {
 438                 ret = io_u_queued_complete(td, min_evts);
 439                 if (ret < 0)
 440                         break;
 441         } while (full && (td->cur_depth > td->o.iodepth_low));
 442
 443         return ret;
 444 }
 445
 446 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 447                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 448                    struct timespec *comp_time)
 449 {
 450         switch (*ret) {
 451         case FIO_Q_COMPLETED:
 452                 if (io_u->error) {
 453                         *ret = -io_u->error;
 454                         clear_io_u(td, io_u);
 455                 } else if (io_u->resid) {
 456                         long long bytes = io_u->xfer_buflen - io_u->resid;
 457                         struct fio_file *f = io_u->file;
 458
 459                         if (bytes_issued)
 460                                 *bytes_issued += bytes;
 461
 462                         if (!from_verify)
 463                                 trim_io_piece(io_u);
 464
 465                         /*
 466                          * zero read, fail
 467                          */
 468                         if (!bytes) {
 469                                 if (!from_verify)
 470                                         unlog_io_piece(td, io_u);
 471                                 td_verror(td, EIO, "full resid");
 472                                 put_io_u(td, io_u);
 473                                 break;
 474                         }
 475
 476                         io_u->xfer_buflen = io_u->resid;
 477                         io_u->xfer_buf += bytes;
 478                         io_u->offset += bytes;
 479
 480                         if (ddir_rw(io_u->ddir))
 481                                 td->ts.short_io_u[io_u->ddir]++;
 482
 483                         if (io_u->offset == f->real_file_size)
 484                                 goto sync_done;
 485
 486                         requeue_io_u(td, &io_u);
 487                 } else {
 488 sync_done:
 489                         if (comp_time && should_check_rate(td))
 490                                 fio_gettime(comp_time, NULL);
 491
 492                         *ret = io_u_sync_complete(td, io_u);
 493                         if (*ret < 0)
 494                                 break;
 495                 }
 496
 497                 if (td->flags & TD_F_REGROW_LOGS)
 498                         regrow_logs(td);
 499
 500                 /*
 501                  * when doing I/O (not when verifying),
 502                  * check for any errors that are to be ignored
 503                  */
 504                 if (!from_verify)
 505                         break;
 506
 507                 return 0;
 508         case FIO_Q_QUEUED:
 509                 /*
 510                  * if the engine doesn't have a commit hook,
 511                  * the io_u is really queued. if it does have such
 512                  * a hook, it has to call io_u_queued() itself.
 513                  */
 514                 if (td->io_ops->commit == NULL)
 515                         io_u_queued(td, io_u);
 516                 if (bytes_issued)
 517                         *bytes_issued += io_u->xfer_buflen;
 518                 break;
 519         case FIO_Q_BUSY:
 520                 if (!from_verify)
 521                         unlog_io_piece(td, io_u);
 522                 requeue_io_u(td, &io_u);
 523                 td_io_commit(td);
 524                 break;
 525         default:
 526                 assert(*ret < 0);
 527                 td_verror(td, -(*ret), "td_io_queue");
 528                 break;
 529         }
 530
 531         if (break_on_this_error(td, ddir, ret))
 532                 return 1;
 533
 534         return 0;
 535 }
 536
 537 static inline bool io_in_polling(struct thread_data *td)
 538 {
 539         return !td->o.iodepth_batch_complete_min &&
 540                    !td->o.iodepth_batch_complete_max;
 541 }
 542 /*
 543  * Unlinks files from thread data fio_file structure
 544  */
 545 static int unlink_all_files(struct thread_data *td)
 546 {
 547         struct fio_file *f;
 548         unsigned int i;
 549         int ret = 0;
 550
 551         for_each_file(td, f, i) {
 552                 if (f->filetype != FIO_TYPE_FILE)
 553                         continue;
 554                 ret = td_io_unlink_file(td, f);
 555                 if (ret)
 556                         break;
 557         }
 558
 559         if (ret)
 560                 td_verror(td, ret, "unlink_all_files");
 561
 562         return ret;
 563 }
 564
 565 /*
 566  * Check if io_u will overlap an in-flight IO in the queue
 567  */
 568 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 569 {
 570         bool overlap;
 571         struct io_u *check_io_u;
 572         unsigned long long x1, x2, y1, y2;
 573         int i;
 574
 575         x1 = io_u->offset;
 576         x2 = io_u->offset + io_u->buflen;
 577         overlap = false;
 578         io_u_qiter(q, check_io_u, i) {
 579                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 580                         y1 = check_io_u->offset;
 581                         y2 = check_io_u->offset + check_io_u->buflen;
 582
 583                         if (x1 < y2 && y1 < x2) {
 584                                 overlap = true;
 585                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 586                                                 x1, io_u->buflen,
 587                                                 y1, check_io_u->buflen);
 588                                 break;
 589                         }
 590                 }
 591         }
 592
 593         return overlap;
 594 }
 595
 596 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 597 {
 598         /*
 599          * Check for overlap if the user asked us to, and we have
 600          * at least one IO in flight besides this one.
 601          */
 602         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 603             in_flight_overlap(&td->io_u_all, io_u))
 604                 return FIO_Q_BUSY;
 605
 606         return td_io_queue(td, io_u);
 607 }
 608
 609 /*
 610  * The main verify engine. Runs over the writes we previously submitted,
 611  * reads the blocks back in, and checks the crc/md5 of the data.
 612  */
 613 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 614 {
 615         struct fio_file *f;
 616         struct io_u *io_u;
 617         int ret, min_events;
 618         unsigned int i;
 619
 620         dprint(FD_VERIFY, "starting loop\n");
 621
 622         /*
 623          * sync io first and invalidate cache, to make sure we really
 624          * read from disk.
 625          */
 626         for_each_file(td, f, i) {
 627                 if (!fio_file_open(f))
 628                         continue;
 629                 if (fio_io_sync(td, f))
 630                         break;
 631                 if (file_invalidate_cache(td, f))
 632                         break;
 633         }
 634
 635         check_update_rusage(td);
 636
 637         if (td->error)
 638                 return;
 639
 640         /*
 641          * verify_state needs to be reset before verification
 642          * proceeds so that expected random seeds match actual
 643          * random seeds in headers. The main loop will reset
 644          * all random number generators if randrepeat is set.
 645          */
 646         if (!td->o.rand_repeatable)
 647                 td_fill_verify_state_seed(td);
 648
 649         td_set_runstate(td, TD_VERIFYING);
 650
 651         io_u = NULL;
 652         while (!td->terminate) {
 653                 enum fio_ddir ddir;
 654                 int full;
 655
 656                 update_ts_cache(td);
 657                 check_update_rusage(td);
 658
 659                 if (runtime_exceeded(td, &td->ts_cache)) {
 660                         __update_ts_cache(td);
 661                         if (runtime_exceeded(td, &td->ts_cache)) {
 662                                 fio_mark_td_terminate(td);
 663                                 break;
 664                         }
 665                 }
 666
 667                 if (flow_threshold_exceeded(td))
 668                         continue;
 669
 670                 if (!td->o.experimental_verify) {
 671                         io_u = __get_io_u(td);
 672                         if (!io_u)
 673                                 break;
 674
 675                         if (get_next_verify(td, io_u)) {
 676                                 put_io_u(td, io_u);
 677                                 break;
 678                         }
 679
 680                         if (td_io_prep(td, io_u)) {
 681                                 put_io_u(td, io_u);
 682                                 break;
 683                         }
 684                 } else {
 685                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 686                                 break;
 687
 688                         while ((io_u = get_io_u(td)) != NULL) {
 689                                 if (IS_ERR_OR_NULL(io_u)) {
 690                                         io_u = NULL;
 691                                         ret = FIO_Q_BUSY;
 692                                         goto reap;
 693                                 }
 694
 695                                 /*
 696                                  * We are only interested in the places where
 697                                  * we wrote or trimmed IOs. Turn those into
 698                                  * reads for verification purposes.
 699                                  */
 700                                 if (io_u->ddir == DDIR_READ) {
 701                                         /*
 702                                          * Pretend we issued it for rwmix
 703                                          * accounting
 704                                          */
 705                                         td->io_issues[DDIR_READ]++;
 706                                         put_io_u(td, io_u);
 707                                         continue;
 708                                 } else if (io_u->ddir == DDIR_TRIM) {
 709                                         io_u->ddir = DDIR_READ;
 710                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 711                                         break;
 712                                 } else if (io_u->ddir == DDIR_WRITE) {
 713                                         io_u->ddir = DDIR_READ;
 714                                         populate_verify_io_u(td, io_u);
 715                                         break;
 716                                 } else {
 717                                         put_io_u(td, io_u);
 718                                         continue;
 719                                 }
 720                         }
 721
 722                         if (!io_u)
 723                                 break;
 724                 }
 725
 726                 if (verify_state_should_stop(td, io_u)) {
 727                         put_io_u(td, io_u);
 728                         break;
 729                 }
 730
 731                 if (td->o.verify_async)
 732                         io_u->end_io = verify_io_u_async;
 733                 else
 734                         io_u->end_io = verify_io_u;
 735
 736                 ddir = io_u->ddir;
 737                 if (!td->o.disable_slat)
 738                         fio_gettime(&io_u->start_time, NULL);
 739
 740                 ret = io_u_submit(td, io_u);
 741
 742                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 743                         break;
 744
 745                 /*
 746                  * if we can queue more, do so. but check if there are
 747                  * completed io_u's first. Note that we can get BUSY even
 748                  * without IO queued, if the system is resource starved.
 749                  */
 750 reap:
 751                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 752                 if (full || io_in_polling(td))
 753                         ret = wait_for_completions(td, NULL);
 754
 755                 if (ret < 0)
 756                         break;
 757         }
 758
 759         check_update_rusage(td);
 760
 761         if (!td->error) {
 762                 min_events = td->cur_depth;
 763
 764                 if (min_events)
 765                         ret = io_u_queued_complete(td, min_events);
 766         } else
 767                 cleanup_pending_aio(td);
 768
 769         td_set_runstate(td, TD_RUNNING);
 770
 771         dprint(FD_VERIFY, "exiting loop\n");
 772 }
 773
 774 static bool exceeds_number_ios(struct thread_data *td)
 775 {
 776         unsigned long long number_ios;
 777
 778         if (!td->o.number_ios)
 779                 return false;
 780
 781         number_ios = ddir_rw_sum(td->io_blocks);
 782         number_ios += td->io_u_queued + td->io_u_in_flight;
 783
 784         return number_ios >= (td->o.number_ios * td->loops);
 785 }
 786
 787 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 788 {
 789         unsigned long long bytes, limit;
 790
 791         if (td_rw(td))
 792                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 793         else if (td_write(td))
 794                 bytes = this_bytes[DDIR_WRITE];
 795         else if (td_read(td))
 796                 bytes = this_bytes[DDIR_READ];
 797         else
 798                 bytes = this_bytes[DDIR_TRIM];
 799
 800         if (td->o.io_size)
 801                 limit = td->o.io_size;
 802         else
 803                 limit = td->o.size;
 804
 805         limit *= td->loops;
 806         return bytes >= limit || exceeds_number_ios(td);
 807 }
 808
 809 static bool io_issue_bytes_exceeded(struct thread_data *td)
 810 {
 811         return io_bytes_exceeded(td, td->io_issue_bytes);
 812 }
 813
 814 static bool io_complete_bytes_exceeded(struct thread_data *td)
 815 {
 816         return io_bytes_exceeded(td, td->this_io_bytes);
 817 }
 818
 819 /*
 820  * used to calculate the next io time for rate control
 821  *
 822  */
 823 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 824 {
 825         uint64_t bps = td->rate_bps[ddir];
 826
 827         assert(!(td->flags & TD_F_CHILD));
 828
 829         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 830                 uint64_t val, iops;
 831
 832                 iops = bps / td->o.min_bs[ddir];
 833                 val = (int64_t) (1000000 / iops) *
 834                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 835                 if (val) {
 836                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 837                                         (unsigned long long) 1000000 / val,
 838                                         ddir);
 839                 }
 840                 td->last_usec[ddir] += val;
 841                 return td->last_usec[ddir];
 842         } else if (bps) {
 843                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 844                 uint64_t secs = bytes / bps;
 845                 uint64_t remainder = bytes % bps;
 846
 847                 return remainder * 1000000 / bps + secs * 1000000;
 848         }
 849
 850         return 0;
 851 }
 852
 853 static void init_thinktime(struct thread_data *td)
 854 {
 855         if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
 856                 td->thinktime_blocks_counter = td->io_blocks;
 857         else
 858                 td->thinktime_blocks_counter = td->io_issues;
 859         td->last_thinktime = td->epoch;
 860         td->last_thinktime_blocks = 0;
 861 }
 862
 863 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
 864                              struct timespec *time)
 865 {
 866         unsigned long long b;
 867         uint64_t total;
 868         int left;
 869         struct timespec now;
 870         bool stall = false;
 871
 872         if (td->o.thinktime_iotime) {
 873                 fio_gettime(&now, NULL);
 874                 if (utime_since(&td->last_thinktime, &now)
 875                     >= td->o.thinktime_iotime + td->o.thinktime) {
 876                         stall = true;
 877                 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
 878                         /*
 879                          * When thinktime_iotime is set and thinktime_blocks is
 880                          * not set, skip the thinktime_blocks check, since
 881                          * thinktime_blocks default value 1 does not work
 882                          * together with thinktime_iotime.
 883                          */
 884                         return;
 885                 }
 886
 887         }
 888
 889         b = ddir_rw_sum(td->thinktime_blocks_counter);
 890         if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
 891                 stall = true;
 892
 893         if (!stall)
 894                 return;
 895
 896         io_u_quiesce(td);
 897
 898         total = 0;
 899         if (td->o.thinktime_spin)
 900                 total = usec_spin(td->o.thinktime_spin);
 901
 902         left = td->o.thinktime - total;
 903         if (left)
 904                 total += usec_sleep(td, left);
 905
 906         /*
 907          * If we're ignoring thinktime for the rate, add the number of bytes
 908          * we would have done while sleeping, minus one block to ensure we
 909          * start issuing immediately after the sleep.
 910          */
 911         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 912                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 913                 uint64_t bs = td->o.min_bs[ddir];
 914                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 915                 uint64_t over;
 916
 917                 if (usperop <= total)
 918                         over = bs;
 919                 else
 920                         over = (usperop - total) / usperop * -bs;
 921
 922                 td->rate_io_issue_bytes[ddir] += (missed - over);
 923                 /* adjust for rate_process=poisson */
 924                 td->last_usec[ddir] += total;
 925         }
 926
 927         if (time && should_check_rate(td))
 928                 fio_gettime(time, NULL);
 929
 930         td->last_thinktime_blocks = b;
 931         if (td->o.thinktime_iotime)
 932                 td->last_thinktime = now;
 933 }
 934
 935 /*
 936  * Main IO worker function. It retrieves io_u's to process and queues
 937  * and reaps them, checking for rate and errors along the way.
 938  *
 939  * Returns number of bytes written and trimmed.
 940  */
 941 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 942 {
 943         unsigned int i;
 944         int ret = 0;
 945         uint64_t total_bytes, bytes_issued = 0;
 946
 947         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 948                 bytes_done[i] = td->bytes_done[i];
 949
 950         if (in_ramp_time(td))
 951                 td_set_runstate(td, TD_RAMP);
 952         else
 953                 td_set_runstate(td, TD_RUNNING);
 954
 955         lat_target_init(td);
 956
 957         total_bytes = td->o.size;
 958         /*
 959         * Allow random overwrite workloads to write up to io_size
 960         * before starting verification phase as 'size' doesn't apply.
 961         */
 962         if (td_write(td) && td_random(td) && td->o.norandommap)
 963                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 964         /*
 965          * If verify_backlog is enabled, we'll run the verify in this
 966          * handler as well. For that case, we may need up to twice the
 967          * amount of bytes.
 968          */
 969         if (td->o.verify != VERIFY_NONE &&
 970            (td_write(td) && td->o.verify_backlog))
 971                 total_bytes += td->o.size;
 972
 973         /* In trimwrite mode, each byte is trimmed and then written, so
 974          * allow total_bytes to be twice as big */
 975         if (td_trimwrite(td))
 976                 total_bytes += td->total_io_size;
 977
 978         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 979                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 980                 td->o.time_based) {
 981                 struct timespec comp_time;
 982                 struct io_u *io_u;
 983                 int full;
 984                 enum fio_ddir ddir;
 985
 986                 check_update_rusage(td);
 987
 988                 if (td->terminate || td->done)
 989                         break;
 990
 991                 update_ts_cache(td);
 992
 993                 if (runtime_exceeded(td, &td->ts_cache)) {
 994                         __update_ts_cache(td);
 995                         if (runtime_exceeded(td, &td->ts_cache)) {
 996                                 fio_mark_td_terminate(td);
 997                                 break;
 998                         }
 999                 }
1000
1001                 if (flow_threshold_exceeded(td))
1002                         continue;
1003
1004                 /*
1005                  * Break if we exceeded the bytes. The exception is time
1006                  * based runs, but we still need to break out of the loop
1007                  * for those to run verification, if enabled.
1008                  * Jobs read from iolog do not use this stop condition.
1009                  */
1010                 if (bytes_issued >= total_bytes &&
1011                     !td->o.read_iolog_file &&
1012                     (!td->o.time_based ||
1013                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
1014                         break;
1015
1016                 io_u = get_io_u(td);
1017                 if (IS_ERR_OR_NULL(io_u)) {
1018                         int err = PTR_ERR(io_u);
1019
1020                         io_u = NULL;
1021                         ddir = DDIR_INVAL;
1022                         if (err == -EBUSY) {
1023                                 ret = FIO_Q_BUSY;
1024                                 goto reap;
1025                         }
1026                         if (td->o.latency_target)
1027                                 goto reap;
1028                         break;
1029                 }
1030
1031                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1032                         populate_verify_io_u(td, io_u);
1033
1034                 ddir = io_u->ddir;
1035
1036                 /*
1037                  * Add verification end_io handler if:
1038                  *      - Asked to verify (!td_rw(td))
1039                  *      - Or the io_u is from our verify list (mixed write/ver)
1040                  */
1041                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1042                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1043
1044                         if (verify_state_should_stop(td, io_u)) {
1045                                 put_io_u(td, io_u);
1046                                 break;
1047                         }
1048
1049                         if (td->o.verify_async)
1050                                 io_u->end_io = verify_io_u_async;
1051                         else
1052                                 io_u->end_io = verify_io_u;
1053                         td_set_runstate(td, TD_VERIFYING);
1054                 } else if (in_ramp_time(td))
1055                         td_set_runstate(td, TD_RAMP);
1056                 else
1057                         td_set_runstate(td, TD_RUNNING);
1058
1059                 /*
1060                  * Always log IO before it's issued, so we know the specific
1061                  * order of it. The logged unit will track when the IO has
1062                  * completed.
1063                  */
1064                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1065                     td->o.do_verify &&
1066                     td->o.verify != VERIFY_NONE &&
1067                     !td->o.experimental_verify)
1068                         log_io_piece(td, io_u);
1069
1070                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1071                         const unsigned long long blen = io_u->xfer_buflen;
1072                         const enum fio_ddir __ddir = acct_ddir(io_u);
1073
1074                         if (td->error)
1075                                 break;
1076
1077                         workqueue_enqueue(&td->io_wq, &io_u->work);
1078                         ret = FIO_Q_QUEUED;
1079
1080                         if (ddir_rw(__ddir)) {
1081                                 td->io_issues[__ddir]++;
1082                                 td->io_issue_bytes[__ddir] += blen;
1083                                 td->rate_io_issue_bytes[__ddir] += blen;
1084                         }
1085
1086                         if (should_check_rate(td)) {
1087                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1088                                 fio_gettime(&comp_time, NULL);
1089                         }
1090
1091                 } else {
1092                         ret = io_u_submit(td, io_u);
1093
1094                         if (should_check_rate(td))
1095                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1096
1097                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1098                                 break;
1099
1100                         /*
1101                          * See if we need to complete some commands. Note that
1102                          * we can get BUSY even without IO queued, if the
1103                          * system is resource starved.
1104                          */
1105 reap:
1106                         full = queue_full(td) ||
1107                                 (ret == FIO_Q_BUSY && td->cur_depth);
1108                         if (full || io_in_polling(td))
1109                                 ret = wait_for_completions(td, &comp_time);
1110                 }
1111                 if (ret < 0)
1112                         break;
1113
1114                 if (ddir_rw(ddir) && td->o.thinktime)
1115                         handle_thinktime(td, ddir, &comp_time);
1116
1117                 if (!ddir_rw_sum(td->bytes_done) &&
1118                     !td_ioengine_flagged(td, FIO_NOIO))
1119                         continue;
1120
1121                 if (!in_ramp_time(td) && should_check_rate(td)) {
1122                         if (check_min_rate(td, &comp_time)) {
1123                                 if (exitall_on_terminate || td->o.exitall_error)
1124                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1125                                 td_verror(td, EIO, "check_min_rate");
1126                                 break;
1127                         }
1128                 }
1129                 if (!in_ramp_time(td) && td->o.latency_target)
1130                         lat_target_check(td);
1131         }
1132
1133         check_update_rusage(td);
1134
1135         if (td->trim_entries)
1136                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1137
1138         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1139                 td->error = 0;
1140                 fio_mark_td_terminate(td);
1141         }
1142         if (!td->error) {
1143                 struct fio_file *f;
1144
1145                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1146                         workqueue_flush(&td->io_wq);
1147                         i = 0;
1148                 } else
1149                         i = td->cur_depth;
1150
1151                 if (i) {
1152                         ret = io_u_queued_complete(td, i);
1153                         if (td->o.fill_device &&
1154                             (td->error == ENOSPC || td->error == EDQUOT))
1155                                 td->error = 0;
1156                 }
1157
1158                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1159                         td_set_runstate(td, TD_FSYNCING);
1160
1161                         for_each_file(td, f, i) {
1162                                 if (!fio_file_fsync(td, f))
1163                                         continue;
1164
1165                                 log_err("fio: end_fsync failed for file %s\n",
1166                                                                 f->file_name);
1167                         }
1168                 }
1169         } else {
1170                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1171                         workqueue_flush(&td->io_wq);
1172                 cleanup_pending_aio(td);
1173         }
1174
1175         /*
1176          * stop job if we failed doing any IO
1177          */
1178         if (!ddir_rw_sum(td->this_io_bytes))
1179                 td->done = 1;
1180
1181         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1182                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1183 }
1184
1185 static void free_file_completion_logging(struct thread_data *td)
1186 {
1187         struct fio_file *f;
1188         unsigned int i;
1189
1190         for_each_file(td, f, i) {
1191                 if (!f->last_write_comp)
1192                         break;
1193                 sfree(f->last_write_comp);
1194         }
1195 }
1196
1197 static int init_file_completion_logging(struct thread_data *td,
1198                                         unsigned int depth)
1199 {
1200         struct fio_file *f;
1201         unsigned int i;
1202
1203         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1204                 return 0;
1205
1206         for_each_file(td, f, i) {
1207                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1208                 if (!f->last_write_comp)
1209                         goto cleanup;
1210         }
1211
1212         return 0;
1213
1214 cleanup:
1215         free_file_completion_logging(td);
1216         log_err("fio: failed to alloc write comp data\n");
1217         return 1;
1218 }
1219
1220 static void cleanup_io_u(struct thread_data *td)
1221 {
1222         struct io_u *io_u;
1223
1224         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1225
1226                 if (td->io_ops->io_u_free)
1227                         td->io_ops->io_u_free(td, io_u);
1228
1229                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1230         }
1231
1232         free_io_mem(td);
1233
1234         io_u_rexit(&td->io_u_requeues);
1235         io_u_qexit(&td->io_u_freelist, false);
1236         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1237
1238         free_file_completion_logging(td);
1239 }
1240
1241 static int init_io_u(struct thread_data *td)
1242 {
1243         struct io_u *io_u;
1244         int cl_align, i, max_units;
1245         int err;
1246
1247         max_units = td->o.iodepth;
1248
1249         err = 0;
1250         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1251         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1252         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1253
1254         if (err) {
1255                 log_err("fio: failed setting up IO queues\n");
1256                 return 1;
1257         }
1258
1259         cl_align = os_cache_line_size();
1260
1261         for (i = 0; i < max_units; i++) {
1262                 void *ptr;
1263
1264                 if (td->terminate)
1265                         return 1;
1266
1267                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1268                 if (!ptr) {
1269                         log_err("fio: unable to allocate aligned memory\n");
1270                         return 1;
1271                 }
1272
1273                 io_u = ptr;
1274                 memset(io_u, 0, sizeof(*io_u));
1275                 INIT_FLIST_HEAD(&io_u->verify_list);
1276                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1277
1278                 io_u->index = i;
1279                 io_u->flags = IO_U_F_FREE;
1280                 io_u_qpush(&td->io_u_freelist, io_u);
1281
1282                 /*
1283                  * io_u never leaves this stack, used for iteration of all
1284                  * io_u buffers.
1285                  */
1286                 io_u_qpush(&td->io_u_all, io_u);
1287
1288                 if (td->io_ops->io_u_init) {
1289                         int ret = td->io_ops->io_u_init(td, io_u);
1290
1291                         if (ret) {
1292                                 log_err("fio: failed to init engine data: %d\n", ret);
1293                                 return 1;
1294                         }
1295                 }
1296         }
1297
1298         init_io_u_buffers(td);
1299
1300         if (init_file_completion_logging(td, max_units))
1301                 return 1;
1302
1303         return 0;
1304 }
1305
1306 int init_io_u_buffers(struct thread_data *td)
1307 {
1308         struct io_u *io_u;
1309         unsigned long long max_bs, min_write;
1310         int i, max_units;
1311         int data_xfer = 1;
1312         char *p;
1313
1314         max_units = td->o.iodepth;
1315         max_bs = td_max_bs(td);
1316         min_write = td->o.min_bs[DDIR_WRITE];
1317         td->orig_buffer_size = (unsigned long long) max_bs
1318                                         * (unsigned long long) max_units;
1319
1320         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1321                 data_xfer = 0;
1322
1323         /*
1324          * if we may later need to do address alignment, then add any
1325          * possible adjustment here so that we don't cause a buffer
1326          * overflow later. this adjustment may be too much if we get
1327          * lucky and the allocator gives us an aligned address.
1328          */
1329         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1330             td_ioengine_flagged(td, FIO_RAWIO))
1331                 td->orig_buffer_size += page_mask + td->o.mem_align;
1332
1333         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1334                 unsigned long long bs;
1335
1336                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1337                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1338         }
1339
1340         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1341                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1342                 return 1;
1343         }
1344
1345         if (data_xfer && allocate_io_mem(td))
1346                 return 1;
1347
1348         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1349             td_ioengine_flagged(td, FIO_RAWIO))
1350                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1351         else
1352                 p = td->orig_buffer;
1353
1354         for (i = 0; i < max_units; i++) {
1355                 io_u = td->io_u_all.io_us[i];
1356                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1357
1358                 if (data_xfer) {
1359                         io_u->buf = p;
1360                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1361
1362                         if (td_write(td))
1363                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1364                         if (td_write(td) && td->o.verify_pattern_bytes) {
1365                                 /*
1366                                  * Fill the buffer with the pattern if we are
1367                                  * going to be doing writes.
1368                                  */
1369                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1370                         }
1371                 }
1372                 p += max_bs;
1373         }
1374
1375         return 0;
1376 }
1377
1378 #ifdef FIO_HAVE_IOSCHED_SWITCH
1379 /*
1380  * These functions are Linux specific.
1381  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1382  */
1383 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1384 {
1385         char tmp[256], tmp2[128], *p;
1386         FILE *f;
1387         int ret;
1388
1389         assert(file->du && file->du->sysfs_root);
1390         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1391
1392         f = fopen(tmp, "r+");
1393         if (!f) {
1394                 if (errno == ENOENT) {
1395                         log_err("fio: os or kernel doesn't support IO scheduler"
1396                                 " switching\n");
1397                         return 0;
1398                 }
1399                 td_verror(td, errno, "fopen iosched");
1400                 return 1;
1401         }
1402
1403         /*
1404          * Set io scheduler.
1405          */
1406         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1407         if (ferror(f) || ret != 1) {
1408                 td_verror(td, errno, "fwrite");
1409                 fclose(f);
1410                 return 1;
1411         }
1412
1413         rewind(f);
1414
1415         /*
1416          * Read back and check that the selected scheduler is now the default.
1417          */
1418         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1419         if (ferror(f) || ret < 0) {
1420                 td_verror(td, errno, "fread");
1421                 fclose(f);
1422                 return 1;
1423         }
1424         tmp[ret] = '\0';
1425         /*
1426          * either a list of io schedulers or "none\n" is expected. Strip the
1427          * trailing newline.
1428          */
1429         p = tmp;
1430         strsep(&p, "\n");
1431
1432         /*
1433          * Write to "none" entry doesn't fail, so check the result here.
1434          */
1435         if (!strcmp(tmp, "none")) {
1436                 log_err("fio: io scheduler is not tunable\n");
1437                 fclose(f);
1438                 return 0;
1439         }
1440
1441         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1442         if (!strstr(tmp, tmp2)) {
1443                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1444                 td_verror(td, EINVAL, "iosched_switch");
1445                 fclose(f);
1446                 return 1;
1447         }
1448
1449         fclose(f);
1450         return 0;
1451 }
1452
1453 static int switch_ioscheduler(struct thread_data *td)
1454 {
1455         struct fio_file *f;
1456         unsigned int i;
1457         int ret = 0;
1458
1459         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1460                 return 0;
1461
1462         assert(td->files && td->files[0]);
1463
1464         for_each_file(td, f, i) {
1465
1466                 /* Only consider regular files and block device files */
1467                 switch (f->filetype) {
1468                 case FIO_TYPE_FILE:
1469                 case FIO_TYPE_BLOCK:
1470                         /*
1471                          * Make sure that the device hosting the file could
1472                          * be determined.
1473                          */
1474                         if (!f->du)
1475                                 continue;
1476                         break;
1477                 case FIO_TYPE_CHAR:
1478                 case FIO_TYPE_PIPE:
1479                 default:
1480                         continue;
1481                 }
1482
1483                 ret = set_ioscheduler(td, f);
1484                 if (ret)
1485                         return ret;
1486         }
1487
1488         return 0;
1489 }
1490
1491 #else
1492
1493 static int switch_ioscheduler(struct thread_data *td)
1494 {
1495         return 0;
1496 }
1497
1498 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1499
1500 static bool keep_running(struct thread_data *td)
1501 {
1502         unsigned long long limit;
1503
1504         if (td->done)
1505                 return false;
1506         if (td->terminate)
1507                 return false;
1508         if (td->o.time_based)
1509                 return true;
1510         if (td->o.loops) {
1511                 td->o.loops--;
1512                 return true;
1513         }
1514         if (exceeds_number_ios(td))
1515                 return false;
1516
1517         if (td->o.io_size)
1518                 limit = td->o.io_size;
1519         else
1520                 limit = td->o.size;
1521
1522         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1523                 uint64_t diff;
1524
1525                 /*
1526                  * If the difference is less than the maximum IO size, we
1527                  * are done.
1528                  */
1529                 diff = limit - ddir_rw_sum(td->io_bytes);
1530                 if (diff < td_max_bs(td))
1531                         return false;
1532
1533                 if (fio_files_done(td) && !td->o.io_size)
1534                         return false;
1535
1536                 return true;
1537         }
1538
1539         return false;
1540 }
1541
1542 static int exec_string(struct thread_options *o, const char *string,
1543                        const char *mode)
1544 {
1545         int ret;
1546         char *str;
1547
1548         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1549                 return -1;
1550
1551         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1552                  o->name, mode);
1553         ret = system(str);
1554         if (ret == -1)
1555                 log_err("fio: exec of cmd <%s> failed\n", str);
1556
1557         free(str);
1558         return ret;
1559 }
1560
1561 /*
1562  * Dry run to compute correct state of numberio for verification.
1563  */
1564 static uint64_t do_dry_run(struct thread_data *td)
1565 {
1566         td_set_runstate(td, TD_RUNNING);
1567
1568         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1569                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1570                 struct io_u *io_u;
1571                 int ret;
1572
1573                 if (td->terminate || td->done)
1574                         break;
1575
1576                 io_u = get_io_u(td);
1577                 if (IS_ERR_OR_NULL(io_u))
1578                         break;
1579
1580                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1581                 io_u->error = 0;
1582                 io_u->resid = 0;
1583                 if (ddir_rw(acct_ddir(io_u)))
1584                         td->io_issues[acct_ddir(io_u)]++;
1585                 if (ddir_rw(io_u->ddir)) {
1586                         io_u_mark_depth(td, 1);
1587                         td->ts.total_io_u[io_u->ddir]++;
1588                 }
1589
1590                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1591                     td->o.do_verify &&
1592                     td->o.verify != VERIFY_NONE &&
1593                     !td->o.experimental_verify)
1594                         log_io_piece(td, io_u);
1595
1596                 ret = io_u_sync_complete(td, io_u);
1597                 (void) ret;
1598         }
1599
1600         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1601 }
1602
1603 struct fork_data {
1604         struct thread_data *td;
1605         struct sk_out *sk_out;
1606 };
1607
1608 /*
1609  * Entry point for the thread based jobs. The process based jobs end up
1610  * here as well, after a little setup.
1611  */
1612 static void *thread_main(void *data)
1613 {
1614         struct fork_data *fd = data;
1615         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1616         struct thread_data *td = fd->td;
1617         struct thread_options *o = &td->o;
1618         struct sk_out *sk_out = fd->sk_out;
1619         uint64_t bytes_done[DDIR_RWDIR_CNT];
1620         int deadlock_loop_cnt;
1621         bool clear_state;
1622         int res, ret;
1623
1624         sk_out_assign(sk_out);
1625         free(fd);
1626
1627         if (!o->use_thread) {
1628                 setsid();
1629                 td->pid = getpid();
1630         } else
1631                 td->pid = gettid();
1632
1633         fio_local_clock_init();
1634
1635         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1636
1637         if (is_backend)
1638                 fio_server_send_start(td);
1639
1640         INIT_FLIST_HEAD(&td->io_log_list);
1641         INIT_FLIST_HEAD(&td->io_hist_list);
1642         INIT_FLIST_HEAD(&td->verify_list);
1643         INIT_FLIST_HEAD(&td->trim_list);
1644         td->io_hist_tree = RB_ROOT;
1645
1646         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1647         if (ret) {
1648                 td_verror(td, ret, "mutex_cond_init_pshared");
1649                 goto err;
1650         }
1651         ret = cond_init_pshared(&td->verify_cond);
1652         if (ret) {
1653                 td_verror(td, ret, "mutex_cond_pshared");
1654                 goto err;
1655         }
1656
1657         td_set_runstate(td, TD_INITIALIZED);
1658         dprint(FD_MUTEX, "up startup_sem\n");
1659         fio_sem_up(startup_sem);
1660         dprint(FD_MUTEX, "wait on td->sem\n");
1661         fio_sem_down(td->sem);
1662         dprint(FD_MUTEX, "done waiting on td->sem\n");
1663
1664         /*
1665          * A new gid requires privilege, so we need to do this before setting
1666          * the uid.
1667          */
1668         if (o->gid != -1U && setgid(o->gid)) {
1669                 td_verror(td, errno, "setgid");
1670                 goto err;
1671         }
1672         if (o->uid != -1U && setuid(o->uid)) {
1673                 td_verror(td, errno, "setuid");
1674                 goto err;
1675         }
1676
1677         td_zone_gen_index(td);
1678
1679         /*
1680          * Do this early, we don't want the compress threads to be limited
1681          * to the same CPUs as the IO workers. So do this before we set
1682          * any potential CPU affinity
1683          */
1684         if (iolog_compress_init(td, sk_out))
1685                 goto err;
1686
1687         /*
1688          * If we have a gettimeofday() thread, make sure we exclude that
1689          * thread from this job
1690          */
1691         if (o->gtod_cpu)
1692                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1693
1694         /*
1695          * Set affinity first, in case it has an impact on the memory
1696          * allocations.
1697          */
1698         if (fio_option_is_set(o, cpumask)) {
1699                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1700                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1701                         if (!ret) {
1702                                 log_err("fio: no CPUs set\n");
1703                                 log_err("fio: Try increasing number of available CPUs\n");
1704                                 td_verror(td, EINVAL, "cpus_split");
1705                                 goto err;
1706                         }
1707                 }
1708                 ret = fio_setaffinity(td->pid, o->cpumask);
1709                 if (ret == -1) {
1710                         td_verror(td, errno, "cpu_set_affinity");
1711                         goto err;
1712                 }
1713         }
1714
1715 #ifdef CONFIG_LIBNUMA
1716         /* numa node setup */
1717         if (fio_option_is_set(o, numa_cpunodes) ||
1718             fio_option_is_set(o, numa_memnodes)) {
1719                 struct bitmask *mask;
1720
1721                 if (numa_available() < 0) {
1722                         td_verror(td, errno, "Does not support NUMA API\n");
1723                         goto err;
1724                 }
1725
1726                 if (fio_option_is_set(o, numa_cpunodes)) {
1727                         mask = numa_parse_nodestring(o->numa_cpunodes);
1728                         ret = numa_run_on_node_mask(mask);
1729                         numa_free_nodemask(mask);
1730                         if (ret == -1) {
1731                                 td_verror(td, errno, \
1732                                         "numa_run_on_node_mask failed\n");
1733                                 goto err;
1734                         }
1735                 }
1736
1737                 if (fio_option_is_set(o, numa_memnodes)) {
1738                         mask = NULL;
1739                         if (o->numa_memnodes)
1740                                 mask = numa_parse_nodestring(o->numa_memnodes);
1741
1742                         switch (o->numa_mem_mode) {
1743                         case MPOL_INTERLEAVE:
1744                                 numa_set_interleave_mask(mask);
1745                                 break;
1746                         case MPOL_BIND:
1747                                 numa_set_membind(mask);
1748                                 break;
1749                         case MPOL_LOCAL:
1750                                 numa_set_localalloc();
1751                                 break;
1752                         case MPOL_PREFERRED:
1753                                 numa_set_preferred(o->numa_mem_prefer_node);
1754                                 break;
1755                         case MPOL_DEFAULT:
1756                         default:
1757                                 break;
1758                         }
1759
1760                         if (mask)
1761                                 numa_free_nodemask(mask);
1762
1763                 }
1764         }
1765 #endif
1766
1767         if (fio_pin_memory(td))
1768                 goto err;
1769
1770         /*
1771          * May alter parameters that init_io_u() will use, so we need to
1772          * do this first.
1773          */
1774         if (!init_iolog(td))
1775                 goto err;
1776
1777         /* ioprio_set() has to be done before td_io_init() */
1778         if (fio_option_is_set(o, ioprio) ||
1779             fio_option_is_set(o, ioprio_class)) {
1780                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1781                 if (ret == -1) {
1782                         td_verror(td, errno, "ioprio_set");
1783                         goto err;
1784                 }
1785                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1786                 td->ts.ioprio = td->ioprio;
1787         }
1788
1789         if (td_io_init(td))
1790                 goto err;
1791
1792         if (init_io_u(td))
1793                 goto err;
1794
1795         if (td->io_ops->post_init && td->io_ops->post_init(td))
1796                 goto err;
1797
1798         if (o->verify_async && verify_async_init(td))
1799                 goto err;
1800
1801         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1802                 goto err;
1803
1804         errno = 0;
1805         if (nice(o->nice) == -1 && errno != 0) {
1806                 td_verror(td, errno, "nice");
1807                 goto err;
1808         }
1809
1810         if (o->ioscheduler && switch_ioscheduler(td))
1811                 goto err;
1812
1813         if (!o->create_serialize && setup_files(td))
1814                 goto err;
1815
1816         if (!init_random_map(td))
1817                 goto err;
1818
1819         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1820                 goto err;
1821
1822         if (o->pre_read && !pre_read_files(td))
1823                 goto err;
1824
1825         fio_verify_init(td);
1826
1827         if (rate_submit_init(td, sk_out))
1828                 goto err;
1829
1830         set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1831         fio_getrusage(&td->ru_start);
1832         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1833         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1834         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1835
1836         init_thinktime(td);
1837
1838         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1839                         o->ratemin[DDIR_TRIM]) {
1840                 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1841                                         sizeof(td->bw_sample_time));
1842                 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1843                                         sizeof(td->bw_sample_time));
1844                 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1845                                         sizeof(td->bw_sample_time));
1846         }
1847
1848         memset(bytes_done, 0, sizeof(bytes_done));
1849         clear_state = false;
1850
1851         while (keep_running(td)) {
1852                 uint64_t verify_bytes;
1853
1854                 fio_gettime(&td->start, NULL);
1855                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1856
1857                 if (clear_state) {
1858                         clear_io_state(td, 0);
1859
1860                         if (o->unlink_each_loop && unlink_all_files(td))
1861                                 break;
1862                 }
1863
1864                 prune_io_piece_log(td);
1865
1866                 if (td->o.verify_only && td_write(td))
1867                         verify_bytes = do_dry_run(td);
1868                 else {
1869                         do_io(td, bytes_done);
1870
1871                         if (!ddir_rw_sum(bytes_done)) {
1872                                 fio_mark_td_terminate(td);
1873                                 verify_bytes = 0;
1874                         } else {
1875                                 verify_bytes = bytes_done[DDIR_WRITE] +
1876                                                 bytes_done[DDIR_TRIM];
1877                         }
1878                 }
1879
1880                 /*
1881                  * If we took too long to shut down, the main thread could
1882                  * already consider us reaped/exited. If that happens, break
1883                  * out and clean up.
1884                  */
1885                 if (td->runstate >= TD_EXITED)
1886                         break;
1887
1888                 clear_state = true;
1889
1890                 /*
1891                  * Make sure we've successfully updated the rusage stats
1892                  * before waiting on the stat mutex. Otherwise we could have
1893                  * the stat thread holding stat mutex and waiting for
1894                  * the rusage_sem, which would never get upped because
1895                  * this thread is waiting for the stat mutex.
1896                  */
1897                 deadlock_loop_cnt = 0;
1898                 do {
1899                         check_update_rusage(td);
1900                         if (!fio_sem_down_trylock(stat_sem))
1901                                 break;
1902                         usleep(1000);
1903                         if (deadlock_loop_cnt++ > 5000) {
1904                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1905                                 td->error = EDEADLK;
1906                                 goto err;
1907                         }
1908                 } while (1);
1909
1910                 if (td_read(td) && td->io_bytes[DDIR_READ])
1911                         update_runtime(td, elapsed_us, DDIR_READ);
1912                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1913                         update_runtime(td, elapsed_us, DDIR_WRITE);
1914                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1915                         update_runtime(td, elapsed_us, DDIR_TRIM);
1916                 fio_gettime(&td->start, NULL);
1917                 fio_sem_up(stat_sem);
1918
1919                 if (td->error || td->terminate)
1920                         break;
1921
1922                 if (!o->do_verify ||
1923                     o->verify == VERIFY_NONE ||
1924                     td_ioengine_flagged(td, FIO_UNIDIR))
1925                         continue;
1926
1927                 clear_io_state(td, 0);
1928
1929                 fio_gettime(&td->start, NULL);
1930
1931                 do_verify(td, verify_bytes);
1932
1933                 /*
1934                  * See comment further up for why this is done here.
1935                  */
1936                 check_update_rusage(td);
1937
1938                 fio_sem_down(stat_sem);
1939                 update_runtime(td, elapsed_us, DDIR_READ);
1940                 fio_gettime(&td->start, NULL);
1941                 fio_sem_up(stat_sem);
1942
1943                 if (td->error || td->terminate)
1944                         break;
1945         }
1946
1947         /*
1948          * Acquire this lock if we were doing overlap checking in
1949          * offload mode so that we don't clean up this job while
1950          * another thread is checking its io_u's for overlap
1951          */
1952         if (td_offload_overlap(td)) {
1953                 int res = pthread_mutex_lock(&overlap_check);
1954                 assert(res == 0);
1955         }
1956         td_set_runstate(td, TD_FINISHING);
1957         if (td_offload_overlap(td)) {
1958                 res = pthread_mutex_unlock(&overlap_check);
1959                 assert(res == 0);
1960         }
1961
1962         update_rusage_stat(td);
1963         td->ts.total_run_time = mtime_since_now(&td->epoch);
1964         for_each_rw_ddir(ddir) {
1965                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1966         }
1967
1968         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1969             (td->o.verify != VERIFY_NONE && td_write(td)))
1970                 verify_save_state(td->thread_number);
1971
1972         fio_unpin_memory(td);
1973
1974         td_writeout_logs(td, true);
1975
1976         iolog_compress_exit(td);
1977         rate_submit_exit(td);
1978
1979         if (o->exec_postrun)
1980                 exec_string(o, o->exec_postrun, "postrun");
1981
1982         if (exitall_on_terminate || (o->exitall_error && td->error))
1983                 fio_terminate_threads(td->groupid, td->o.exit_what);
1984
1985 err:
1986         if (td->error)
1987                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1988                                                         td->verror);
1989
1990         if (o->verify_async)
1991                 verify_async_exit(td);
1992
1993         close_and_free_files(td);
1994         cleanup_io_u(td);
1995         close_ioengine(td);
1996         cgroup_shutdown(td, cgroup_mnt);
1997         verify_free_state(td);
1998         td_zone_free_index(td);
1999
2000         if (fio_option_is_set(o, cpumask)) {
2001                 ret = fio_cpuset_exit(&o->cpumask);
2002                 if (ret)
2003                         td_verror(td, ret, "fio_cpuset_exit");
2004         }
2005
2006         /*
2007          * do this very late, it will log file closing as well
2008          */
2009         if (o->write_iolog_file)
2010                 write_iolog_close(td);
2011         if (td->io_log_rfile)
2012                 fclose(td->io_log_rfile);
2013
2014         td_set_runstate(td, TD_EXITED);
2015
2016         /*
2017          * Do this last after setting our runstate to exited, so we
2018          * know that the stat thread is signaled.
2019          */
2020         check_update_rusage(td);
2021
2022         sk_out_drop();
2023         return (void *) (uintptr_t) td->error;
2024 }
2025
2026 /*
2027  * Run over the job map and reap the threads that have exited, if any.
2028  */
2029 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2030                          uint64_t *m_rate)
2031 {
2032         struct thread_data *td;
2033         unsigned int cputhreads, realthreads, pending;
2034         int i, status, ret;
2035
2036         /*
2037          * reap exited threads (TD_EXITED -> TD_REAPED)
2038          */
2039         realthreads = pending = cputhreads = 0;
2040         for_each_td(td, i) {
2041                 int flags = 0;
2042
2043                 if (!strcmp(td->o.ioengine, "cpuio"))
2044                         cputhreads++;
2045                 else
2046                         realthreads++;
2047
2048                 if (!td->pid) {
2049                         pending++;
2050                         continue;
2051                 }
2052                 if (td->runstate == TD_REAPED)
2053                         continue;
2054                 if (td->o.use_thread) {
2055                         if (td->runstate == TD_EXITED) {
2056                                 td_set_runstate(td, TD_REAPED);
2057                                 goto reaped;
2058                         }
2059                         continue;
2060                 }
2061
2062                 flags = WNOHANG;
2063                 if (td->runstate == TD_EXITED)
2064                         flags = 0;
2065
2066                 /*
2067                  * check if someone quit or got killed in an unusual way
2068                  */
2069                 ret = waitpid(td->pid, &status, flags);
2070                 if (ret < 0) {
2071                         if (errno == ECHILD) {
2072                                 log_err("fio: pid=%d disappeared %d\n",
2073                                                 (int) td->pid, td->runstate);
2074                                 td->sig = ECHILD;
2075                                 td_set_runstate(td, TD_REAPED);
2076                                 goto reaped;
2077                         }
2078                         perror("waitpid");
2079                 } else if (ret == td->pid) {
2080                         if (WIFSIGNALED(status)) {
2081                                 int sig = WTERMSIG(status);
2082
2083                                 if (sig != SIGTERM && sig != SIGUSR2)
2084                                         log_err("fio: pid=%d, got signal=%d\n",
2085                                                         (int) td->pid, sig);
2086                                 td->sig = sig;
2087                                 td_set_runstate(td, TD_REAPED);
2088                                 goto reaped;
2089                         }
2090                         if (WIFEXITED(status)) {
2091                                 if (WEXITSTATUS(status) && !td->error)
2092                                         td->error = WEXITSTATUS(status);
2093
2094                                 td_set_runstate(td, TD_REAPED);
2095                                 goto reaped;
2096                         }
2097                 }
2098
2099                 /*
2100                  * If the job is stuck, do a forceful timeout of it and
2101                  * move on.
2102                  */
2103                 if (td->terminate &&
2104                     td->runstate < TD_FSYNCING &&
2105                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2106                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2107                                 "%lu seconds, it appears to be stuck. Doing "
2108                                 "forceful exit of this job.\n",
2109                                 td->o.name, td->runstate,
2110                                 (unsigned long) time_since_now(&td->terminate_time));
2111                         td_set_runstate(td, TD_REAPED);
2112                         goto reaped;
2113                 }
2114
2115                 /*
2116                  * thread is not dead, continue
2117                  */
2118                 pending++;
2119                 continue;
2120 reaped:
2121                 (*nr_running)--;
2122                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2123                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2124                 if (!td->pid)
2125                         pending--;
2126
2127                 if (td->error)
2128                         exit_value++;
2129
2130                 done_secs += mtime_since_now(&td->epoch) / 1000;
2131                 profile_td_exit(td);
2132                 flow_exit_job(td);
2133         }
2134
2135         if (*nr_running == cputhreads && !pending && realthreads)
2136                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2137 }
2138
2139 static bool __check_trigger_file(void)
2140 {
2141         struct stat sb;
2142
2143         if (!trigger_file)
2144                 return false;
2145
2146         if (stat(trigger_file, &sb))
2147                 return false;
2148
2149         if (unlink(trigger_file) < 0)
2150                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2151                                                         strerror(errno));
2152
2153         return true;
2154 }
2155
2156 static bool trigger_timedout(void)
2157 {
2158         if (trigger_timeout)
2159                 if (time_since_genesis() >= trigger_timeout) {
2160                         trigger_timeout = 0;
2161                         return true;
2162                 }
2163
2164         return false;
2165 }
2166
2167 void exec_trigger(const char *cmd)
2168 {
2169         int ret;
2170
2171         if (!cmd || cmd[0] == '\0')
2172                 return;
2173
2174         ret = system(cmd);
2175         if (ret == -1)
2176                 log_err("fio: failed executing %s trigger\n", cmd);
2177 }
2178
2179 void check_trigger_file(void)
2180 {
2181         if (__check_trigger_file() || trigger_timedout()) {
2182                 if (nr_clients)
2183                         fio_clients_send_trigger(trigger_remote_cmd);
2184                 else {
2185                         verify_save_state(IO_LIST_ALL);
2186                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2187                         exec_trigger(trigger_cmd);
2188                 }
2189         }
2190 }
2191
2192 static int fio_verify_load_state(struct thread_data *td)
2193 {
2194         int ret;
2195
2196         if (!td->o.verify_state)
2197                 return 0;
2198
2199         if (is_backend) {
2200                 void *data;
2201
2202                 ret = fio_server_get_verify_state(td->o.name,
2203                                         td->thread_number - 1, &data);
2204                 if (!ret)
2205                         verify_assign_state(td, data);
2206         } else {
2207                 char prefix[PATH_MAX];
2208
2209                 if (aux_path)
2210                         sprintf(prefix, "%s%clocal", aux_path,
2211                                         FIO_OS_PATH_SEPARATOR);
2212                 else
2213                         strcpy(prefix, "local");
2214                 ret = verify_load_state(td, prefix);
2215         }
2216
2217         return ret;
2218 }
2219
2220 static void do_usleep(unsigned int usecs)
2221 {
2222         check_for_running_stats();
2223         check_trigger_file();
2224         usleep(usecs);
2225 }
2226
2227 static bool check_mount_writes(struct thread_data *td)
2228 {
2229         struct fio_file *f;
2230         unsigned int i;
2231
2232         if (!td_write(td) || td->o.allow_mounted_write)
2233                 return false;
2234
2235         /*
2236          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2237          * are mkfs'd and mounted.
2238          */
2239         for_each_file(td, f, i) {
2240 #ifdef FIO_HAVE_CHARDEV_SIZE
2241                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2242 #else
2243                 if (f->filetype != FIO_TYPE_BLOCK)
2244 #endif
2245                         continue;
2246                 if (device_is_mounted(f->file_name))
2247                         goto mounted;
2248         }
2249
2250         return false;
2251 mounted:
2252         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2253         return true;
2254 }
2255
2256 static bool waitee_running(struct thread_data *me)
2257 {
2258         const char *waitee = me->o.wait_for;
2259         const char *self = me->o.name;
2260         struct thread_data *td;
2261         int i;
2262
2263         if (!waitee)
2264                 return false;
2265
2266         for_each_td(td, i) {
2267                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2268                         continue;
2269
2270                 if (td->runstate < TD_EXITED) {
2271                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2272                                         self, td->o.name,
2273                                         runstate_to_name(td->runstate));
2274                         return true;
2275                 }
2276         }
2277
2278         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2279         return false;
2280 }
2281
2282 /*
2283  * Main function for kicking off and reaping jobs, as needed.
2284  */
2285 static void run_threads(struct sk_out *sk_out)
2286 {
2287         struct thread_data *td;
2288         unsigned int i, todo, nr_running, nr_started;
2289         uint64_t m_rate, t_rate;
2290         uint64_t spent;
2291
2292         if (fio_gtod_offload && fio_start_gtod_thread())
2293                 return;
2294
2295         fio_idle_prof_init();
2296
2297         set_sig_handlers();
2298
2299         nr_thread = nr_process = 0;
2300         for_each_td(td, i) {
2301                 if (check_mount_writes(td))
2302                         return;
2303                 if (td->o.use_thread)
2304                         nr_thread++;
2305                 else
2306                         nr_process++;
2307         }
2308
2309         if (output_format & FIO_OUTPUT_NORMAL) {
2310                 struct buf_output out;
2311
2312                 buf_output_init(&out);
2313                 __log_buf(&out, "Starting ");
2314                 if (nr_thread)
2315                         __log_buf(&out, "%d thread%s", nr_thread,
2316                                                 nr_thread > 1 ? "s" : "");
2317                 if (nr_process) {
2318                         if (nr_thread)
2319                                 __log_buf(&out, " and ");
2320                         __log_buf(&out, "%d process%s", nr_process,
2321                                                 nr_process > 1 ? "es" : "");
2322                 }
2323                 __log_buf(&out, "\n");
2324                 log_info_buf(out.buf, out.buflen);
2325                 buf_output_free(&out);
2326         }
2327
2328         todo = thread_number;
2329         nr_running = 0;
2330         nr_started = 0;
2331         m_rate = t_rate = 0;
2332
2333         for_each_td(td, i) {
2334                 print_status_init(td->thread_number - 1);
2335
2336                 if (!td->o.create_serialize)
2337                         continue;
2338
2339                 if (fio_verify_load_state(td))
2340                         goto reap;
2341
2342                 /*
2343                  * do file setup here so it happens sequentially,
2344                  * we don't want X number of threads getting their
2345                  * client data interspersed on disk
2346                  */
2347                 if (setup_files(td)) {
2348 reap:
2349                         exit_value++;
2350                         if (td->error)
2351                                 log_err("fio: pid=%d, err=%d/%s\n",
2352                                         (int) td->pid, td->error, td->verror);
2353                         td_set_runstate(td, TD_REAPED);
2354                         todo--;
2355                 } else {
2356                         struct fio_file *f;
2357                         unsigned int j;
2358
2359                         /*
2360                          * for sharing to work, each job must always open
2361                          * its own files. so close them, if we opened them
2362                          * for creation
2363                          */
2364                         for_each_file(td, f, j) {
2365                                 if (fio_file_open(f))
2366                                         td_io_close_file(td, f);
2367                         }
2368                 }
2369         }
2370
2371         /* start idle threads before io threads start to run */
2372         fio_idle_prof_start();
2373
2374         set_genesis_time();
2375
2376         while (todo) {
2377                 struct thread_data *map[REAL_MAX_JOBS];
2378                 struct timespec this_start;
2379                 int this_jobs = 0, left;
2380                 struct fork_data *fd;
2381
2382                 /*
2383                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2384                  */
2385                 for_each_td(td, i) {
2386                         if (td->runstate != TD_NOT_CREATED)
2387                                 continue;
2388
2389                         /*
2390                          * never got a chance to start, killed by other
2391                          * thread for some reason
2392                          */
2393                         if (td->terminate) {
2394                                 todo--;
2395                                 continue;
2396                         }
2397
2398                         if (td->o.start_delay) {
2399                                 spent = utime_since_genesis();
2400
2401                                 if (td->o.start_delay > spent)
2402                                         continue;
2403                         }
2404
2405                         if (td->o.stonewall && (nr_started || nr_running)) {
2406                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2407                                                         td->o.name);
2408                                 break;
2409                         }
2410
2411                         if (waitee_running(td)) {
2412                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2413                                                 td->o.name, td->o.wait_for);
2414                                 continue;
2415                         }
2416
2417                         init_disk_util(td);
2418
2419                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2420                         td->update_rusage = 0;
2421
2422                         /*
2423                          * Set state to created. Thread will transition
2424                          * to TD_INITIALIZED when it's done setting up.
2425                          */
2426                         td_set_runstate(td, TD_CREATED);
2427                         map[this_jobs++] = td;
2428                         nr_started++;
2429
2430                         fd = calloc(1, sizeof(*fd));
2431                         fd->td = td;
2432                         fd->sk_out = sk_out;
2433
2434                         if (td->o.use_thread) {
2435                                 int ret;
2436
2437                                 dprint(FD_PROCESS, "will pthread_create\n");
2438                                 ret = pthread_create(&td->thread, NULL,
2439                                                         thread_main, fd);
2440                                 if (ret) {
2441                                         log_err("pthread_create: %s\n",
2442                                                         strerror(ret));
2443                                         free(fd);
2444                                         nr_started--;
2445                                         break;
2446                                 }
2447                                 fd = NULL;
2448                                 ret = pthread_detach(td->thread);
2449                                 if (ret)
2450                                         log_err("pthread_detach: %s",
2451                                                         strerror(ret));
2452                         } else {
2453                                 pid_t pid;
2454                                 struct fio_file **files;
2455                                 void *eo;
2456                                 dprint(FD_PROCESS, "will fork\n");
2457                                 files = td->files;
2458                                 eo = td->eo;
2459                                 read_barrier();
2460                                 pid = fork();
2461                                 if (!pid) {
2462                                         int ret;
2463
2464                                         ret = (int)(uintptr_t)thread_main(fd);
2465                                         _exit(ret);
2466                                 } else if (i == fio_debug_jobno)
2467                                         *fio_debug_jobp = pid;
2468                                 // freeing previously allocated memory for files
2469                                 // this memory freed MUST NOT be shared between processes, only the pointer itself may be shared within TD
2470                                 free(files);
2471                                 free(eo);
2472                                 free(fd);
2473                                 fd = NULL;
2474                         }
2475                         dprint(FD_MUTEX, "wait on startup_sem\n");
2476                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2477                                 log_err("fio: job startup hung? exiting.\n");
2478                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2479                                 fio_abort = true;
2480                                 nr_started--;
2481                                 free(fd);
2482                                 break;
2483                         }
2484                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2485                 }
2486
2487                 /*
2488                  * Wait for the started threads to transition to
2489                  * TD_INITIALIZED.
2490                  */
2491                 fio_gettime(&this_start, NULL);
2492                 left = this_jobs;
2493                 while (left && !fio_abort) {
2494                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2495                                 break;
2496
2497                         do_usleep(100000);
2498
2499                         for (i = 0; i < this_jobs; i++) {
2500                                 td = map[i];
2501                                 if (!td)
2502                                         continue;
2503                                 if (td->runstate == TD_INITIALIZED) {
2504                                         map[i] = NULL;
2505                                         left--;
2506                                 } else if (td->runstate >= TD_EXITED) {
2507                                         map[i] = NULL;
2508                                         left--;
2509                                         todo--;
2510                                         nr_running++; /* work-around... */
2511                                 }
2512                         }
2513                 }
2514
2515                 if (left) {
2516                         log_err("fio: %d job%s failed to start\n", left,
2517                                         left > 1 ? "s" : "");
2518                         for (i = 0; i < this_jobs; i++) {
2519                                 td = map[i];
2520                                 if (!td)
2521                                         continue;
2522                                 kill(td->pid, SIGTERM);
2523                         }
2524                         break;
2525                 }
2526
2527                 /*
2528                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2529                  */
2530                 for_each_td(td, i) {
2531                         if (td->runstate != TD_INITIALIZED)
2532                                 continue;
2533
2534                         if (in_ramp_time(td))
2535                                 td_set_runstate(td, TD_RAMP);
2536                         else
2537                                 td_set_runstate(td, TD_RUNNING);
2538                         nr_running++;
2539                         nr_started--;
2540                         m_rate += ddir_rw_sum(td->o.ratemin);
2541                         t_rate += ddir_rw_sum(td->o.rate);
2542                         todo--;
2543                         fio_sem_up(td->sem);
2544                 }
2545
2546                 reap_threads(&nr_running, &t_rate, &m_rate);
2547
2548                 if (todo)
2549                         do_usleep(100000);
2550         }
2551
2552         while (nr_running) {
2553                 reap_threads(&nr_running, &t_rate, &m_rate);
2554                 do_usleep(10000);
2555         }
2556
2557         fio_idle_prof_stop();
2558
2559         update_io_ticks();
2560 }
2561
2562 static void free_disk_util(void)
2563 {
2564         disk_util_prune_entries();
2565         helper_thread_destroy();
2566 }
2567
2568 int fio_backend(struct sk_out *sk_out)
2569 {
2570         struct thread_data *td;
2571         int i;
2572
2573         if (exec_profile) {
2574                 if (load_profile(exec_profile))
2575                         return 1;
2576                 free(exec_profile);
2577                 exec_profile = NULL;
2578         }
2579         if (!thread_number)
2580                 return 0;
2581
2582         if (write_bw_log) {
2583                 struct log_params p = {
2584                         .log_type = IO_LOG_TYPE_BW,
2585                 };
2586
2587                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2588                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2589                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2590         }
2591
2592         if (init_global_dedupe_working_set_seeds()) {
2593                 log_err("fio: failed to initialize global dedupe working set\n");
2594                 return 1;
2595         }
2596
2597         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2598         if (!sk_out)
2599                 is_local_backend = true;
2600         if (startup_sem == NULL)
2601                 return 1;
2602
2603         set_genesis_time();
2604         stat_init();
2605         if (helper_thread_create(startup_sem, sk_out))
2606                 log_err("fio: failed to create helper thread\n");
2607
2608         cgroup_list = smalloc(sizeof(*cgroup_list));
2609         if (cgroup_list)
2610                 INIT_FLIST_HEAD(cgroup_list);
2611
2612         run_threads(sk_out);
2613
2614         helper_thread_exit();
2615
2616         if (!fio_abort) {
2617                 __show_run_stats();
2618                 if (write_bw_log) {
2619                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2620                                 struct io_log *log = agg_io_log[i];
2621
2622                                 flush_log(log, false);
2623                                 free_log(log);
2624                         }
2625                 }
2626         }
2627
2628         for_each_td(td, i) {
2629                 struct thread_stat *ts = &td->ts;
2630
2631                 free_clat_prio_stats(ts);
2632                 steadystate_free(td);
2633                 fio_options_free(td);
2634                 fio_dump_options_free(td);
2635                 if (td->rusage_sem) {
2636                         fio_sem_remove(td->rusage_sem);
2637                         td->rusage_sem = NULL;
2638                 }
2639                 fio_sem_remove(td->sem);
2640                 td->sem = NULL;
2641         }
2642
2643         free_disk_util();
2644         if (cgroup_list) {
2645                 cgroup_kill(cgroup_list);
2646                 sfree(cgroup_list);
2647         }
2648
2649         fio_sem_remove(startup_sem);
2650         stat_exit();
2651         return exit_value;
2652 }