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 or number of ios to be twice as big */
 975         if (td_trimwrite(td)) {
 976                 total_bytes += td->total_io_size;
 977                 td->o.number_ios *= 2;
 978         }
 979
 980         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 981                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 982                 td->o.time_based) {
 983                 struct timespec comp_time;
 984                 struct io_u *io_u;
 985                 int full;
 986                 enum fio_ddir ddir;
 987
 988                 check_update_rusage(td);
 989
 990                 if (td->terminate || td->done)
 991                         break;
 992
 993                 update_ts_cache(td);
 994
 995                 if (runtime_exceeded(td, &td->ts_cache)) {
 996                         __update_ts_cache(td);
 997                         if (runtime_exceeded(td, &td->ts_cache)) {
 998                                 fio_mark_td_terminate(td);
 999                                 break;
1000                         }
1001                 }
1002
1003                 if (flow_threshold_exceeded(td))
1004                         continue;
1005
1006                 /*
1007                  * Break if we exceeded the bytes. The exception is time
1008                  * based runs, but we still need to break out of the loop
1009                  * for those to run verification, if enabled.
1010                  * Jobs read from iolog do not use this stop condition.
1011                  */
1012                 if (bytes_issued >= total_bytes &&
1013                     !td->o.read_iolog_file &&
1014                     (!td->o.time_based ||
1015                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
1016                         break;
1017
1018                 io_u = get_io_u(td);
1019                 if (IS_ERR_OR_NULL(io_u)) {
1020                         int err = PTR_ERR(io_u);
1021
1022                         io_u = NULL;
1023                         ddir = DDIR_INVAL;
1024                         if (err == -EBUSY) {
1025                                 ret = FIO_Q_BUSY;
1026                                 goto reap;
1027                         }
1028                         if (td->o.latency_target)
1029                                 goto reap;
1030                         break;
1031                 }
1032
1033                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1034                         populate_verify_io_u(td, io_u);
1035
1036                 ddir = io_u->ddir;
1037
1038                 /*
1039                  * Add verification end_io handler if:
1040                  *      - Asked to verify (!td_rw(td))
1041                  *      - Or the io_u is from our verify list (mixed write/ver)
1042                  */
1043                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1044                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1045
1046                         if (verify_state_should_stop(td, io_u)) {
1047                                 put_io_u(td, io_u);
1048                                 break;
1049                         }
1050
1051                         if (td->o.verify_async)
1052                                 io_u->end_io = verify_io_u_async;
1053                         else
1054                                 io_u->end_io = verify_io_u;
1055                         td_set_runstate(td, TD_VERIFYING);
1056                 } else if (in_ramp_time(td))
1057                         td_set_runstate(td, TD_RAMP);
1058                 else
1059                         td_set_runstate(td, TD_RUNNING);
1060
1061                 /*
1062                  * Always log IO before it's issued, so we know the specific
1063                  * order of it. The logged unit will track when the IO has
1064                  * completed.
1065                  */
1066                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1067                     td->o.do_verify &&
1068                     td->o.verify != VERIFY_NONE &&
1069                     !td->o.experimental_verify)
1070                         log_io_piece(td, io_u);
1071
1072                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1073                         const unsigned long long blen = io_u->xfer_buflen;
1074                         const enum fio_ddir __ddir = acct_ddir(io_u);
1075
1076                         if (td->error)
1077                                 break;
1078
1079                         workqueue_enqueue(&td->io_wq, &io_u->work);
1080                         ret = FIO_Q_QUEUED;
1081
1082                         if (ddir_rw(__ddir)) {
1083                                 td->io_issues[__ddir]++;
1084                                 td->io_issue_bytes[__ddir] += blen;
1085                                 td->rate_io_issue_bytes[__ddir] += blen;
1086                         }
1087
1088                         if (should_check_rate(td)) {
1089                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1090                                 fio_gettime(&comp_time, NULL);
1091                         }
1092
1093                 } else {
1094                         ret = io_u_submit(td, io_u);
1095
1096                         if (should_check_rate(td))
1097                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1098
1099                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1100                                 break;
1101
1102                         /*
1103                          * See if we need to complete some commands. Note that
1104                          * we can get BUSY even without IO queued, if the
1105                          * system is resource starved.
1106                          */
1107 reap:
1108                         full = queue_full(td) ||
1109                                 (ret == FIO_Q_BUSY && td->cur_depth);
1110                         if (full || io_in_polling(td))
1111                                 ret = wait_for_completions(td, &comp_time);
1112                 }
1113                 if (ret < 0)
1114                         break;
1115
1116                 if (ddir_rw(ddir) && td->o.thinktime)
1117                         handle_thinktime(td, ddir, &comp_time);
1118
1119                 if (!ddir_rw_sum(td->bytes_done) &&
1120                     !td_ioengine_flagged(td, FIO_NOIO))
1121                         continue;
1122
1123                 if (!in_ramp_time(td) && should_check_rate(td)) {
1124                         if (check_min_rate(td, &comp_time)) {
1125                                 if (exitall_on_terminate || td->o.exitall_error)
1126                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1127                                 td_verror(td, EIO, "check_min_rate");
1128                                 break;
1129                         }
1130                 }
1131                 if (!in_ramp_time(td) && td->o.latency_target)
1132                         lat_target_check(td);
1133         }
1134
1135         check_update_rusage(td);
1136
1137         if (td->trim_entries)
1138                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1139
1140         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1141                 td->error = 0;
1142                 fio_mark_td_terminate(td);
1143         }
1144         if (!td->error) {
1145                 struct fio_file *f;
1146
1147                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1148                         workqueue_flush(&td->io_wq);
1149                         i = 0;
1150                 } else
1151                         i = td->cur_depth;
1152
1153                 if (i) {
1154                         ret = io_u_queued_complete(td, i);
1155                         if (td->o.fill_device &&
1156                             (td->error == ENOSPC || td->error == EDQUOT))
1157                                 td->error = 0;
1158                 }
1159
1160                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1161                         td_set_runstate(td, TD_FSYNCING);
1162
1163                         for_each_file(td, f, i) {
1164                                 if (!fio_file_fsync(td, f))
1165                                         continue;
1166
1167                                 log_err("fio: end_fsync failed for file %s\n",
1168                                                                 f->file_name);
1169                         }
1170                 }
1171         } else {
1172                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1173                         workqueue_flush(&td->io_wq);
1174                 cleanup_pending_aio(td);
1175         }
1176
1177         /*
1178          * stop job if we failed doing any IO
1179          */
1180         if (!ddir_rw_sum(td->this_io_bytes))
1181                 td->done = 1;
1182
1183         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1184                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1185 }
1186
1187 static void free_file_completion_logging(struct thread_data *td)
1188 {
1189         struct fio_file *f;
1190         unsigned int i;
1191
1192         for_each_file(td, f, i) {
1193                 if (!f->last_write_comp)
1194                         break;
1195                 sfree(f->last_write_comp);
1196         }
1197 }
1198
1199 static int init_file_completion_logging(struct thread_data *td,
1200                                         unsigned int depth)
1201 {
1202         struct fio_file *f;
1203         unsigned int i;
1204
1205         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1206                 return 0;
1207
1208         for_each_file(td, f, i) {
1209                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1210                 if (!f->last_write_comp)
1211                         goto cleanup;
1212         }
1213
1214         return 0;
1215
1216 cleanup:
1217         free_file_completion_logging(td);
1218         log_err("fio: failed to alloc write comp data\n");
1219         return 1;
1220 }
1221
1222 static void cleanup_io_u(struct thread_data *td)
1223 {
1224         struct io_u *io_u;
1225
1226         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1227
1228                 if (td->io_ops->io_u_free)
1229                         td->io_ops->io_u_free(td, io_u);
1230
1231                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1232         }
1233
1234         free_io_mem(td);
1235
1236         io_u_rexit(&td->io_u_requeues);
1237         io_u_qexit(&td->io_u_freelist, false);
1238         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1239
1240         free_file_completion_logging(td);
1241 }
1242
1243 static int init_io_u(struct thread_data *td)
1244 {
1245         struct io_u *io_u;
1246         int cl_align, i, max_units;
1247         int err;
1248
1249         max_units = td->o.iodepth;
1250
1251         err = 0;
1252         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1253         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1254         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1255
1256         if (err) {
1257                 log_err("fio: failed setting up IO queues\n");
1258                 return 1;
1259         }
1260
1261         cl_align = os_cache_line_size();
1262
1263         for (i = 0; i < max_units; i++) {
1264                 void *ptr;
1265
1266                 if (td->terminate)
1267                         return 1;
1268
1269                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1270                 if (!ptr) {
1271                         log_err("fio: unable to allocate aligned memory\n");
1272                         return 1;
1273                 }
1274
1275                 io_u = ptr;
1276                 memset(io_u, 0, sizeof(*io_u));
1277                 INIT_FLIST_HEAD(&io_u->verify_list);
1278                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1279
1280                 io_u->index = i;
1281                 io_u->flags = IO_U_F_FREE;
1282                 io_u_qpush(&td->io_u_freelist, io_u);
1283
1284                 /*
1285                  * io_u never leaves this stack, used for iteration of all
1286                  * io_u buffers.
1287                  */
1288                 io_u_qpush(&td->io_u_all, io_u);
1289
1290                 if (td->io_ops->io_u_init) {
1291                         int ret = td->io_ops->io_u_init(td, io_u);
1292
1293                         if (ret) {
1294                                 log_err("fio: failed to init engine data: %d\n", ret);
1295                                 return 1;
1296                         }
1297                 }
1298         }
1299
1300         init_io_u_buffers(td);
1301
1302         if (init_file_completion_logging(td, max_units))
1303                 return 1;
1304
1305         return 0;
1306 }
1307
1308 int init_io_u_buffers(struct thread_data *td)
1309 {
1310         struct io_u *io_u;
1311         unsigned long long max_bs, min_write;
1312         int i, max_units;
1313         int data_xfer = 1;
1314         char *p;
1315
1316         max_units = td->o.iodepth;
1317         max_bs = td_max_bs(td);
1318         min_write = td->o.min_bs[DDIR_WRITE];
1319         td->orig_buffer_size = (unsigned long long) max_bs
1320                                         * (unsigned long long) max_units;
1321
1322         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1323                 data_xfer = 0;
1324
1325         /*
1326          * if we may later need to do address alignment, then add any
1327          * possible adjustment here so that we don't cause a buffer
1328          * overflow later. this adjustment may be too much if we get
1329          * lucky and the allocator gives us an aligned address.
1330          */
1331         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1332             td_ioengine_flagged(td, FIO_RAWIO))
1333                 td->orig_buffer_size += page_mask + td->o.mem_align;
1334
1335         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1336                 unsigned long long bs;
1337
1338                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1339                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1340         }
1341
1342         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1343                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1344                 return 1;
1345         }
1346
1347         if (data_xfer && allocate_io_mem(td))
1348                 return 1;
1349
1350         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1351             td_ioengine_flagged(td, FIO_RAWIO))
1352                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1353         else
1354                 p = td->orig_buffer;
1355
1356         for (i = 0; i < max_units; i++) {
1357                 io_u = td->io_u_all.io_us[i];
1358                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1359
1360                 if (data_xfer) {
1361                         io_u->buf = p;
1362                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1363
1364                         if (td_write(td))
1365                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1366                         if (td_write(td) && td->o.verify_pattern_bytes) {
1367                                 /*
1368                                  * Fill the buffer with the pattern if we are
1369                                  * going to be doing writes.
1370                                  */
1371                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1372                         }
1373                 }
1374                 p += max_bs;
1375         }
1376
1377         return 0;
1378 }
1379
1380 #ifdef FIO_HAVE_IOSCHED_SWITCH
1381 /*
1382  * These functions are Linux specific.
1383  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1384  */
1385 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1386 {
1387         char tmp[256], tmp2[128], *p;
1388         FILE *f;
1389         int ret;
1390
1391         assert(file->du && file->du->sysfs_root);
1392         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1393
1394         f = fopen(tmp, "r+");
1395         if (!f) {
1396                 if (errno == ENOENT) {
1397                         log_err("fio: os or kernel doesn't support IO scheduler"
1398                                 " switching\n");
1399                         return 0;
1400                 }
1401                 td_verror(td, errno, "fopen iosched");
1402                 return 1;
1403         }
1404
1405         /*
1406          * Set io scheduler.
1407          */
1408         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1409         if (ferror(f) || ret != 1) {
1410                 td_verror(td, errno, "fwrite");
1411                 fclose(f);
1412                 return 1;
1413         }
1414
1415         rewind(f);
1416
1417         /*
1418          * Read back and check that the selected scheduler is now the default.
1419          */
1420         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1421         if (ferror(f) || ret < 0) {
1422                 td_verror(td, errno, "fread");
1423                 fclose(f);
1424                 return 1;
1425         }
1426         tmp[ret] = '\0';
1427         /*
1428          * either a list of io schedulers or "none\n" is expected. Strip the
1429          * trailing newline.
1430          */
1431         p = tmp;
1432         strsep(&p, "\n");
1433
1434         /*
1435          * Write to "none" entry doesn't fail, so check the result here.
1436          */
1437         if (!strcmp(tmp, "none")) {
1438                 log_err("fio: io scheduler is not tunable\n");
1439                 fclose(f);
1440                 return 0;
1441         }
1442
1443         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1444         if (!strstr(tmp, tmp2)) {
1445                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1446                 td_verror(td, EINVAL, "iosched_switch");
1447                 fclose(f);
1448                 return 1;
1449         }
1450
1451         fclose(f);
1452         return 0;
1453 }
1454
1455 static int switch_ioscheduler(struct thread_data *td)
1456 {
1457         struct fio_file *f;
1458         unsigned int i;
1459         int ret = 0;
1460
1461         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1462                 return 0;
1463
1464         assert(td->files && td->files[0]);
1465
1466         for_each_file(td, f, i) {
1467
1468                 /* Only consider regular files and block device files */
1469                 switch (f->filetype) {
1470                 case FIO_TYPE_FILE:
1471                 case FIO_TYPE_BLOCK:
1472                         /*
1473                          * Make sure that the device hosting the file could
1474                          * be determined.
1475                          */
1476                         if (!f->du)
1477                                 continue;
1478                         break;
1479                 case FIO_TYPE_CHAR:
1480                 case FIO_TYPE_PIPE:
1481                 default:
1482                         continue;
1483                 }
1484
1485                 ret = set_ioscheduler(td, f);
1486                 if (ret)
1487                         return ret;
1488         }
1489
1490         return 0;
1491 }
1492
1493 #else
1494
1495 static int switch_ioscheduler(struct thread_data *td)
1496 {
1497         return 0;
1498 }
1499
1500 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1501
1502 static bool keep_running(struct thread_data *td)
1503 {
1504         unsigned long long limit;
1505
1506         if (td->done)
1507                 return false;
1508         if (td->terminate)
1509                 return false;
1510         if (td->o.time_based)
1511                 return true;
1512         if (td->o.loops) {
1513                 td->o.loops--;
1514                 return true;
1515         }
1516         if (exceeds_number_ios(td))
1517                 return false;
1518
1519         if (td->o.io_size)
1520                 limit = td->o.io_size;
1521         else
1522                 limit = td->o.size;
1523
1524         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1525                 uint64_t diff;
1526
1527                 /*
1528                  * If the difference is less than the maximum IO size, we
1529                  * are done.
1530                  */
1531                 diff = limit - ddir_rw_sum(td->io_bytes);
1532                 if (diff < td_max_bs(td))
1533                         return false;
1534
1535                 if (fio_files_done(td) && !td->o.io_size)
1536                         return false;
1537
1538                 return true;
1539         }
1540
1541         return false;
1542 }
1543
1544 static int exec_string(struct thread_options *o, const char *string,
1545                        const char *mode)
1546 {
1547         int ret;
1548         char *str;
1549
1550         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1551                 return -1;
1552
1553         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1554                  o->name, mode);
1555         ret = system(str);
1556         if (ret == -1)
1557                 log_err("fio: exec of cmd <%s> failed\n", str);
1558
1559         free(str);
1560         return ret;
1561 }
1562
1563 /*
1564  * Dry run to compute correct state of numberio for verification.
1565  */
1566 static uint64_t do_dry_run(struct thread_data *td)
1567 {
1568         td_set_runstate(td, TD_RUNNING);
1569
1570         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1571                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1572                 struct io_u *io_u;
1573                 int ret;
1574
1575                 if (td->terminate || td->done)
1576                         break;
1577
1578                 io_u = get_io_u(td);
1579                 if (IS_ERR_OR_NULL(io_u))
1580                         break;
1581
1582                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1583                 io_u->error = 0;
1584                 io_u->resid = 0;
1585                 if (ddir_rw(acct_ddir(io_u)))
1586                         td->io_issues[acct_ddir(io_u)]++;
1587                 if (ddir_rw(io_u->ddir)) {
1588                         io_u_mark_depth(td, 1);
1589                         td->ts.total_io_u[io_u->ddir]++;
1590                 }
1591
1592                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1593                     td->o.do_verify &&
1594                     td->o.verify != VERIFY_NONE &&
1595                     !td->o.experimental_verify)
1596                         log_io_piece(td, io_u);
1597
1598                 ret = io_u_sync_complete(td, io_u);
1599                 (void) ret;
1600         }
1601
1602         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1603 }
1604
1605 struct fork_data {
1606         struct thread_data *td;
1607         struct sk_out *sk_out;
1608 };
1609
1610 /*
1611  * Entry point for the thread based jobs. The process based jobs end up
1612  * here as well, after a little setup.
1613  */
1614 static void *thread_main(void *data)
1615 {
1616         struct fork_data *fd = data;
1617         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1618         struct thread_data *td = fd->td;
1619         struct thread_options *o = &td->o;
1620         struct sk_out *sk_out = fd->sk_out;
1621         uint64_t bytes_done[DDIR_RWDIR_CNT];
1622         int deadlock_loop_cnt;
1623         bool clear_state;
1624         int res, ret;
1625
1626         sk_out_assign(sk_out);
1627         free(fd);
1628
1629         if (!o->use_thread) {
1630                 setsid();
1631                 td->pid = getpid();
1632         } else
1633                 td->pid = gettid();
1634
1635         fio_local_clock_init();
1636
1637         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1638
1639         if (is_backend)
1640                 fio_server_send_start(td);
1641
1642         INIT_FLIST_HEAD(&td->io_log_list);
1643         INIT_FLIST_HEAD(&td->io_hist_list);
1644         INIT_FLIST_HEAD(&td->verify_list);
1645         INIT_FLIST_HEAD(&td->trim_list);
1646         td->io_hist_tree = RB_ROOT;
1647
1648         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1649         if (ret) {
1650                 td_verror(td, ret, "mutex_cond_init_pshared");
1651                 goto err;
1652         }
1653         ret = cond_init_pshared(&td->verify_cond);
1654         if (ret) {
1655                 td_verror(td, ret, "mutex_cond_pshared");
1656                 goto err;
1657         }
1658
1659         td_set_runstate(td, TD_INITIALIZED);
1660         dprint(FD_MUTEX, "up startup_sem\n");
1661         fio_sem_up(startup_sem);
1662         dprint(FD_MUTEX, "wait on td->sem\n");
1663         fio_sem_down(td->sem);
1664         dprint(FD_MUTEX, "done waiting on td->sem\n");
1665
1666         /*
1667          * A new gid requires privilege, so we need to do this before setting
1668          * the uid.
1669          */
1670         if (o->gid != -1U && setgid(o->gid)) {
1671                 td_verror(td, errno, "setgid");
1672                 goto err;
1673         }
1674         if (o->uid != -1U && setuid(o->uid)) {
1675                 td_verror(td, errno, "setuid");
1676                 goto err;
1677         }
1678
1679         td_zone_gen_index(td);
1680
1681         /*
1682          * Do this early, we don't want the compress threads to be limited
1683          * to the same CPUs as the IO workers. So do this before we set
1684          * any potential CPU affinity
1685          */
1686         if (iolog_compress_init(td, sk_out))
1687                 goto err;
1688
1689         /*
1690          * If we have a gettimeofday() thread, make sure we exclude that
1691          * thread from this job
1692          */
1693         if (o->gtod_cpu)
1694                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1695
1696         /*
1697          * Set affinity first, in case it has an impact on the memory
1698          * allocations.
1699          */
1700         if (fio_option_is_set(o, cpumask)) {
1701                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1702                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1703                         if (!ret) {
1704                                 log_err("fio: no CPUs set\n");
1705                                 log_err("fio: Try increasing number of available CPUs\n");
1706                                 td_verror(td, EINVAL, "cpus_split");
1707                                 goto err;
1708                         }
1709                 }
1710                 ret = fio_setaffinity(td->pid, o->cpumask);
1711                 if (ret == -1) {
1712                         td_verror(td, errno, "cpu_set_affinity");
1713                         goto err;
1714                 }
1715         }
1716
1717 #ifdef CONFIG_LIBNUMA
1718         /* numa node setup */
1719         if (fio_option_is_set(o, numa_cpunodes) ||
1720             fio_option_is_set(o, numa_memnodes)) {
1721                 struct bitmask *mask;
1722
1723                 if (numa_available() < 0) {
1724                         td_verror(td, errno, "Does not support NUMA API\n");
1725                         goto err;
1726                 }
1727
1728                 if (fio_option_is_set(o, numa_cpunodes)) {
1729                         mask = numa_parse_nodestring(o->numa_cpunodes);
1730                         ret = numa_run_on_node_mask(mask);
1731                         numa_free_nodemask(mask);
1732                         if (ret == -1) {
1733                                 td_verror(td, errno, \
1734                                         "numa_run_on_node_mask failed\n");
1735                                 goto err;
1736                         }
1737                 }
1738
1739                 if (fio_option_is_set(o, numa_memnodes)) {
1740                         mask = NULL;
1741                         if (o->numa_memnodes)
1742                                 mask = numa_parse_nodestring(o->numa_memnodes);
1743
1744                         switch (o->numa_mem_mode) {
1745                         case MPOL_INTERLEAVE:
1746                                 numa_set_interleave_mask(mask);
1747                                 break;
1748                         case MPOL_BIND:
1749                                 numa_set_membind(mask);
1750                                 break;
1751                         case MPOL_LOCAL:
1752                                 numa_set_localalloc();
1753                                 break;
1754                         case MPOL_PREFERRED:
1755                                 numa_set_preferred(o->numa_mem_prefer_node);
1756                                 break;
1757                         case MPOL_DEFAULT:
1758                         default:
1759                                 break;
1760                         }
1761
1762                         if (mask)
1763                                 numa_free_nodemask(mask);
1764
1765                 }
1766         }
1767 #endif
1768
1769         if (fio_pin_memory(td))
1770                 goto err;
1771
1772         /*
1773          * May alter parameters that init_io_u() will use, so we need to
1774          * do this first.
1775          */
1776         if (!init_iolog(td))
1777                 goto err;
1778
1779         /* ioprio_set() has to be done before td_io_init() */
1780         if (fio_option_is_set(o, ioprio) ||
1781             fio_option_is_set(o, ioprio_class)) {
1782                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1783                 if (ret == -1) {
1784                         td_verror(td, errno, "ioprio_set");
1785                         goto err;
1786                 }
1787                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1788                 td->ts.ioprio = td->ioprio;
1789         }
1790
1791         if (td_io_init(td))
1792                 goto err;
1793
1794         if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1) {
1795                 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1796                          "are selected, queue depth will be capped at 1\n");
1797         }
1798
1799         if (init_io_u(td))
1800                 goto err;
1801
1802         if (td->io_ops->post_init && td->io_ops->post_init(td))
1803                 goto err;
1804
1805         if (o->verify_async && verify_async_init(td))
1806                 goto err;
1807
1808         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1809                 goto err;
1810
1811         errno = 0;
1812         if (nice(o->nice) == -1 && errno != 0) {
1813                 td_verror(td, errno, "nice");
1814                 goto err;
1815         }
1816
1817         if (o->ioscheduler && switch_ioscheduler(td))
1818                 goto err;
1819
1820         if (!o->create_serialize && setup_files(td))
1821                 goto err;
1822
1823         if (!init_random_map(td))
1824                 goto err;
1825
1826         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1827                 goto err;
1828
1829         if (o->pre_read && !pre_read_files(td))
1830                 goto err;
1831
1832         fio_verify_init(td);
1833
1834         if (rate_submit_init(td, sk_out))
1835                 goto err;
1836
1837         set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1838         fio_getrusage(&td->ru_start);
1839         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1840         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1841         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1842
1843         init_thinktime(td);
1844
1845         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1846                         o->ratemin[DDIR_TRIM]) {
1847                 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1848                                         sizeof(td->bw_sample_time));
1849                 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1850                                         sizeof(td->bw_sample_time));
1851                 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1852                                         sizeof(td->bw_sample_time));
1853         }
1854
1855         memset(bytes_done, 0, sizeof(bytes_done));
1856         clear_state = false;
1857
1858         while (keep_running(td)) {
1859                 uint64_t verify_bytes;
1860
1861                 fio_gettime(&td->start, NULL);
1862                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1863
1864                 if (clear_state) {
1865                         clear_io_state(td, 0);
1866
1867                         if (o->unlink_each_loop && unlink_all_files(td))
1868                                 break;
1869                 }
1870
1871                 prune_io_piece_log(td);
1872
1873                 if (td->o.verify_only && td_write(td))
1874                         verify_bytes = do_dry_run(td);
1875                 else {
1876                         do_io(td, bytes_done);
1877
1878                         if (!ddir_rw_sum(bytes_done)) {
1879                                 fio_mark_td_terminate(td);
1880                                 verify_bytes = 0;
1881                         } else {
1882                                 verify_bytes = bytes_done[DDIR_WRITE] +
1883                                                 bytes_done[DDIR_TRIM];
1884                         }
1885                 }
1886
1887                 /*
1888                  * If we took too long to shut down, the main thread could
1889                  * already consider us reaped/exited. If that happens, break
1890                  * out and clean up.
1891                  */
1892                 if (td->runstate >= TD_EXITED)
1893                         break;
1894
1895                 clear_state = true;
1896
1897                 /*
1898                  * Make sure we've successfully updated the rusage stats
1899                  * before waiting on the stat mutex. Otherwise we could have
1900                  * the stat thread holding stat mutex and waiting for
1901                  * the rusage_sem, which would never get upped because
1902                  * this thread is waiting for the stat mutex.
1903                  */
1904                 deadlock_loop_cnt = 0;
1905                 do {
1906                         check_update_rusage(td);
1907                         if (!fio_sem_down_trylock(stat_sem))
1908                                 break;
1909                         usleep(1000);
1910                         if (deadlock_loop_cnt++ > 5000) {
1911                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1912                                 td->error = EDEADLK;
1913                                 goto err;
1914                         }
1915                 } while (1);
1916
1917                 if (td_read(td) && td->io_bytes[DDIR_READ])
1918                         update_runtime(td, elapsed_us, DDIR_READ);
1919                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1920                         update_runtime(td, elapsed_us, DDIR_WRITE);
1921                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1922                         update_runtime(td, elapsed_us, DDIR_TRIM);
1923                 fio_gettime(&td->start, NULL);
1924                 fio_sem_up(stat_sem);
1925
1926                 if (td->error || td->terminate)
1927                         break;
1928
1929                 if (!o->do_verify ||
1930                     o->verify == VERIFY_NONE ||
1931                     td_ioengine_flagged(td, FIO_UNIDIR))
1932                         continue;
1933
1934                 clear_io_state(td, 0);
1935
1936                 fio_gettime(&td->start, NULL);
1937
1938                 do_verify(td, verify_bytes);
1939
1940                 /*
1941                  * See comment further up for why this is done here.
1942                  */
1943                 check_update_rusage(td);
1944
1945                 fio_sem_down(stat_sem);
1946                 update_runtime(td, elapsed_us, DDIR_READ);
1947                 fio_gettime(&td->start, NULL);
1948                 fio_sem_up(stat_sem);
1949
1950                 if (td->error || td->terminate)
1951                         break;
1952         }
1953
1954         /*
1955          * Acquire this lock if we were doing overlap checking in
1956          * offload mode so that we don't clean up this job while
1957          * another thread is checking its io_u's for overlap
1958          */
1959         if (td_offload_overlap(td)) {
1960                 int res = pthread_mutex_lock(&overlap_check);
1961                 assert(res == 0);
1962         }
1963         td_set_runstate(td, TD_FINISHING);
1964         if (td_offload_overlap(td)) {
1965                 res = pthread_mutex_unlock(&overlap_check);
1966                 assert(res == 0);
1967         }
1968
1969         update_rusage_stat(td);
1970         td->ts.total_run_time = mtime_since_now(&td->epoch);
1971         for_each_rw_ddir(ddir) {
1972                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1973         }
1974
1975         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1976             (td->o.verify != VERIFY_NONE && td_write(td)))
1977                 verify_save_state(td->thread_number);
1978
1979         fio_unpin_memory(td);
1980
1981         td_writeout_logs(td, true);
1982
1983         iolog_compress_exit(td);
1984         rate_submit_exit(td);
1985
1986         if (o->exec_postrun)
1987                 exec_string(o, o->exec_postrun, "postrun");
1988
1989         if (exitall_on_terminate || (o->exitall_error && td->error))
1990                 fio_terminate_threads(td->groupid, td->o.exit_what);
1991
1992 err:
1993         if (td->error)
1994                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1995                                                         td->verror);
1996
1997         if (o->verify_async)
1998                 verify_async_exit(td);
1999
2000         close_and_free_files(td);
2001         cleanup_io_u(td);
2002         close_ioengine(td);
2003         cgroup_shutdown(td, cgroup_mnt);
2004         verify_free_state(td);
2005         td_zone_free_index(td);
2006
2007         if (fio_option_is_set(o, cpumask)) {
2008                 ret = fio_cpuset_exit(&o->cpumask);
2009                 if (ret)
2010                         td_verror(td, ret, "fio_cpuset_exit");
2011         }
2012
2013         /*
2014          * do this very late, it will log file closing as well
2015          */
2016         if (o->write_iolog_file)
2017                 write_iolog_close(td);
2018         if (td->io_log_rfile)
2019                 fclose(td->io_log_rfile);
2020
2021         td_set_runstate(td, TD_EXITED);
2022
2023         /*
2024          * Do this last after setting our runstate to exited, so we
2025          * know that the stat thread is signaled.
2026          */
2027         check_update_rusage(td);
2028
2029         sk_out_drop();
2030         return (void *) (uintptr_t) td->error;
2031 }
2032
2033 /*
2034  * Run over the job map and reap the threads that have exited, if any.
2035  */
2036 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2037                          uint64_t *m_rate)
2038 {
2039         struct thread_data *td;
2040         unsigned int cputhreads, realthreads, pending;
2041         int i, status, ret;
2042
2043         /*
2044          * reap exited threads (TD_EXITED -> TD_REAPED)
2045          */
2046         realthreads = pending = cputhreads = 0;
2047         for_each_td(td, i) {
2048                 int flags = 0;
2049
2050                 if (!strcmp(td->o.ioengine, "cpuio"))
2051                         cputhreads++;
2052                 else
2053                         realthreads++;
2054
2055                 if (!td->pid) {
2056                         pending++;
2057                         continue;
2058                 }
2059                 if (td->runstate == TD_REAPED)
2060                         continue;
2061                 if (td->o.use_thread) {
2062                         if (td->runstate == TD_EXITED) {
2063                                 td_set_runstate(td, TD_REAPED);
2064                                 goto reaped;
2065                         }
2066                         continue;
2067                 }
2068
2069                 flags = WNOHANG;
2070                 if (td->runstate == TD_EXITED)
2071                         flags = 0;
2072
2073                 /*
2074                  * check if someone quit or got killed in an unusual way
2075                  */
2076                 ret = waitpid(td->pid, &status, flags);
2077                 if (ret < 0) {
2078                         if (errno == ECHILD) {
2079                                 log_err("fio: pid=%d disappeared %d\n",
2080                                                 (int) td->pid, td->runstate);
2081                                 td->sig = ECHILD;
2082                                 td_set_runstate(td, TD_REAPED);
2083                                 goto reaped;
2084                         }
2085                         perror("waitpid");
2086                 } else if (ret == td->pid) {
2087                         if (WIFSIGNALED(status)) {
2088                                 int sig = WTERMSIG(status);
2089
2090                                 if (sig != SIGTERM && sig != SIGUSR2)
2091                                         log_err("fio: pid=%d, got signal=%d\n",
2092                                                         (int) td->pid, sig);
2093                                 td->sig = sig;
2094                                 td_set_runstate(td, TD_REAPED);
2095                                 goto reaped;
2096                         }
2097                         if (WIFEXITED(status)) {
2098                                 if (WEXITSTATUS(status) && !td->error)
2099                                         td->error = WEXITSTATUS(status);
2100
2101                                 td_set_runstate(td, TD_REAPED);
2102                                 goto reaped;
2103                         }
2104                 }
2105
2106                 /*
2107                  * If the job is stuck, do a forceful timeout of it and
2108                  * move on.
2109                  */
2110                 if (td->terminate &&
2111                     td->runstate < TD_FSYNCING &&
2112                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2113                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2114                                 "%lu seconds, it appears to be stuck. Doing "
2115                                 "forceful exit of this job.\n",
2116                                 td->o.name, td->runstate,
2117                                 (unsigned long) time_since_now(&td->terminate_time));
2118                         td_set_runstate(td, TD_REAPED);
2119                         goto reaped;
2120                 }
2121
2122                 /*
2123                  * thread is not dead, continue
2124                  */
2125                 pending++;
2126                 continue;
2127 reaped:
2128                 (*nr_running)--;
2129                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2130                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2131                 if (!td->pid)
2132                         pending--;
2133
2134                 if (td->error)
2135                         exit_value++;
2136
2137                 done_secs += mtime_since_now(&td->epoch) / 1000;
2138                 profile_td_exit(td);
2139                 flow_exit_job(td);
2140         }
2141
2142         if (*nr_running == cputhreads && !pending && realthreads)
2143                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2144 }
2145
2146 static bool __check_trigger_file(void)
2147 {
2148         struct stat sb;
2149
2150         if (!trigger_file)
2151                 return false;
2152
2153         if (stat(trigger_file, &sb))
2154                 return false;
2155
2156         if (unlink(trigger_file) < 0)
2157                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2158                                                         strerror(errno));
2159
2160         return true;
2161 }
2162
2163 static bool trigger_timedout(void)
2164 {
2165         if (trigger_timeout)
2166                 if (time_since_genesis() >= trigger_timeout) {
2167                         trigger_timeout = 0;
2168                         return true;
2169                 }
2170
2171         return false;
2172 }
2173
2174 void exec_trigger(const char *cmd)
2175 {
2176         int ret;
2177
2178         if (!cmd || cmd[0] == '\0')
2179                 return;
2180
2181         ret = system(cmd);
2182         if (ret == -1)
2183                 log_err("fio: failed executing %s trigger\n", cmd);
2184 }
2185
2186 void check_trigger_file(void)
2187 {
2188         if (__check_trigger_file() || trigger_timedout()) {
2189                 if (nr_clients)
2190                         fio_clients_send_trigger(trigger_remote_cmd);
2191                 else {
2192                         verify_save_state(IO_LIST_ALL);
2193                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2194                         exec_trigger(trigger_cmd);
2195                 }
2196         }
2197 }
2198
2199 static int fio_verify_load_state(struct thread_data *td)
2200 {
2201         int ret;
2202
2203         if (!td->o.verify_state)
2204                 return 0;
2205
2206         if (is_backend) {
2207                 void *data;
2208
2209                 ret = fio_server_get_verify_state(td->o.name,
2210                                         td->thread_number - 1, &data);
2211                 if (!ret)
2212                         verify_assign_state(td, data);
2213         } else {
2214                 char prefix[PATH_MAX];
2215
2216                 if (aux_path)
2217                         sprintf(prefix, "%s%clocal", aux_path,
2218                                         FIO_OS_PATH_SEPARATOR);
2219                 else
2220                         strcpy(prefix, "local");
2221                 ret = verify_load_state(td, prefix);
2222         }
2223
2224         return ret;
2225 }
2226
2227 static void do_usleep(unsigned int usecs)
2228 {
2229         check_for_running_stats();
2230         check_trigger_file();
2231         usleep(usecs);
2232 }
2233
2234 static bool check_mount_writes(struct thread_data *td)
2235 {
2236         struct fio_file *f;
2237         unsigned int i;
2238
2239         if (!td_write(td) || td->o.allow_mounted_write)
2240                 return false;
2241
2242         /*
2243          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2244          * are mkfs'd and mounted.
2245          */
2246         for_each_file(td, f, i) {
2247 #ifdef FIO_HAVE_CHARDEV_SIZE
2248                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2249 #else
2250                 if (f->filetype != FIO_TYPE_BLOCK)
2251 #endif
2252                         continue;
2253                 if (device_is_mounted(f->file_name))
2254                         goto mounted;
2255         }
2256
2257         return false;
2258 mounted:
2259         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2260         return true;
2261 }
2262
2263 static bool waitee_running(struct thread_data *me)
2264 {
2265         const char *waitee = me->o.wait_for;
2266         const char *self = me->o.name;
2267         struct thread_data *td;
2268         int i;
2269
2270         if (!waitee)
2271                 return false;
2272
2273         for_each_td(td, i) {
2274                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2275                         continue;
2276
2277                 if (td->runstate < TD_EXITED) {
2278                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2279                                         self, td->o.name,
2280                                         runstate_to_name(td->runstate));
2281                         return true;
2282                 }
2283         }
2284
2285         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2286         return false;
2287 }
2288
2289 /*
2290  * Main function for kicking off and reaping jobs, as needed.
2291  */
2292 static void run_threads(struct sk_out *sk_out)
2293 {
2294         struct thread_data *td;
2295         unsigned int i, todo, nr_running, nr_started;
2296         uint64_t m_rate, t_rate;
2297         uint64_t spent;
2298
2299         if (fio_gtod_offload && fio_start_gtod_thread())
2300                 return;
2301
2302         fio_idle_prof_init();
2303
2304         set_sig_handlers();
2305
2306         nr_thread = nr_process = 0;
2307         for_each_td(td, i) {
2308                 if (check_mount_writes(td))
2309                         return;
2310                 if (td->o.use_thread)
2311                         nr_thread++;
2312                 else
2313                         nr_process++;
2314         }
2315
2316         if (output_format & FIO_OUTPUT_NORMAL) {
2317                 struct buf_output out;
2318
2319                 buf_output_init(&out);
2320                 __log_buf(&out, "Starting ");
2321                 if (nr_thread)
2322                         __log_buf(&out, "%d thread%s", nr_thread,
2323                                                 nr_thread > 1 ? "s" : "");
2324                 if (nr_process) {
2325                         if (nr_thread)
2326                                 __log_buf(&out, " and ");
2327                         __log_buf(&out, "%d process%s", nr_process,
2328                                                 nr_process > 1 ? "es" : "");
2329                 }
2330                 __log_buf(&out, "\n");
2331                 log_info_buf(out.buf, out.buflen);
2332                 buf_output_free(&out);
2333         }
2334
2335         todo = thread_number;
2336         nr_running = 0;
2337         nr_started = 0;
2338         m_rate = t_rate = 0;
2339
2340         for_each_td(td, i) {
2341                 print_status_init(td->thread_number - 1);
2342
2343                 if (!td->o.create_serialize)
2344                         continue;
2345
2346                 if (fio_verify_load_state(td))
2347                         goto reap;
2348
2349                 /*
2350                  * do file setup here so it happens sequentially,
2351                  * we don't want X number of threads getting their
2352                  * client data interspersed on disk
2353                  */
2354                 if (setup_files(td)) {
2355 reap:
2356                         exit_value++;
2357                         if (td->error)
2358                                 log_err("fio: pid=%d, err=%d/%s\n",
2359                                         (int) td->pid, td->error, td->verror);
2360                         td_set_runstate(td, TD_REAPED);
2361                         todo--;
2362                 } else {
2363                         struct fio_file *f;
2364                         unsigned int j;
2365
2366                         /*
2367                          * for sharing to work, each job must always open
2368                          * its own files. so close them, if we opened them
2369                          * for creation
2370                          */
2371                         for_each_file(td, f, j) {
2372                                 if (fio_file_open(f))
2373                                         td_io_close_file(td, f);
2374                         }
2375                 }
2376         }
2377
2378         /* start idle threads before io threads start to run */
2379         fio_idle_prof_start();
2380
2381         set_genesis_time();
2382
2383         while (todo) {
2384                 struct thread_data *map[REAL_MAX_JOBS];
2385                 struct timespec this_start;
2386                 int this_jobs = 0, left;
2387                 struct fork_data *fd;
2388
2389                 /*
2390                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2391                  */
2392                 for_each_td(td, i) {
2393                         if (td->runstate != TD_NOT_CREATED)
2394                                 continue;
2395
2396                         /*
2397                          * never got a chance to start, killed by other
2398                          * thread for some reason
2399                          */
2400                         if (td->terminate) {
2401                                 todo--;
2402                                 continue;
2403                         }
2404
2405                         if (td->o.start_delay) {
2406                                 spent = utime_since_genesis();
2407
2408                                 if (td->o.start_delay > spent)
2409                                         continue;
2410                         }
2411
2412                         if (td->o.stonewall && (nr_started || nr_running)) {
2413                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2414                                                         td->o.name);
2415                                 break;
2416                         }
2417
2418                         if (waitee_running(td)) {
2419                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2420                                                 td->o.name, td->o.wait_for);
2421                                 continue;
2422                         }
2423
2424                         init_disk_util(td);
2425
2426                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2427                         td->update_rusage = 0;
2428
2429                         /*
2430                          * Set state to created. Thread will transition
2431                          * to TD_INITIALIZED when it's done setting up.
2432                          */
2433                         td_set_runstate(td, TD_CREATED);
2434                         map[this_jobs++] = td;
2435                         nr_started++;
2436
2437                         fd = calloc(1, sizeof(*fd));
2438                         fd->td = td;
2439                         fd->sk_out = sk_out;
2440
2441                         if (td->o.use_thread) {
2442                                 int ret;
2443
2444                                 dprint(FD_PROCESS, "will pthread_create\n");
2445                                 ret = pthread_create(&td->thread, NULL,
2446                                                         thread_main, fd);
2447                                 if (ret) {
2448                                         log_err("pthread_create: %s\n",
2449                                                         strerror(ret));
2450                                         free(fd);
2451                                         nr_started--;
2452                                         break;
2453                                 }
2454                                 fd = NULL;
2455                                 ret = pthread_detach(td->thread);
2456                                 if (ret)
2457                                         log_err("pthread_detach: %s",
2458                                                         strerror(ret));
2459                         } else {
2460                                 pid_t pid;
2461                                 void *eo;
2462                                 dprint(FD_PROCESS, "will fork\n");
2463                                 eo = td->eo;
2464                                 read_barrier();
2465                                 pid = fork();
2466                                 if (!pid) {
2467                                         int ret;
2468
2469                                         ret = (int)(uintptr_t)thread_main(fd);
2470                                         _exit(ret);
2471                                 } else if (i == fio_debug_jobno)
2472                                         *fio_debug_jobp = pid;
2473                                 free(eo);
2474                                 free(fd);
2475                                 fd = NULL;
2476                         }
2477                         dprint(FD_MUTEX, "wait on startup_sem\n");
2478                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2479                                 log_err("fio: job startup hung? exiting.\n");
2480                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2481                                 fio_abort = true;
2482                                 nr_started--;
2483                                 free(fd);
2484                                 break;
2485                         }
2486                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2487                 }
2488
2489                 /*
2490                  * Wait for the started threads to transition to
2491                  * TD_INITIALIZED.
2492                  */
2493                 fio_gettime(&this_start, NULL);
2494                 left = this_jobs;
2495                 while (left && !fio_abort) {
2496                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2497                                 break;
2498
2499                         do_usleep(100000);
2500
2501                         for (i = 0; i < this_jobs; i++) {
2502                                 td = map[i];
2503                                 if (!td)
2504                                         continue;
2505                                 if (td->runstate == TD_INITIALIZED) {
2506                                         map[i] = NULL;
2507                                         left--;
2508                                 } else if (td->runstate >= TD_EXITED) {
2509                                         map[i] = NULL;
2510                                         left--;
2511                                         todo--;
2512                                         nr_running++; /* work-around... */
2513                                 }
2514                         }
2515                 }
2516
2517                 if (left) {
2518                         log_err("fio: %d job%s failed to start\n", left,
2519                                         left > 1 ? "s" : "");
2520                         for (i = 0; i < this_jobs; i++) {
2521                                 td = map[i];
2522                                 if (!td)
2523                                         continue;
2524                                 kill(td->pid, SIGTERM);
2525                         }
2526                         break;
2527                 }
2528
2529                 /*
2530                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2531                  */
2532                 for_each_td(td, i) {
2533                         if (td->runstate != TD_INITIALIZED)
2534                                 continue;
2535
2536                         if (in_ramp_time(td))
2537                                 td_set_runstate(td, TD_RAMP);
2538                         else
2539                                 td_set_runstate(td, TD_RUNNING);
2540                         nr_running++;
2541                         nr_started--;
2542                         m_rate += ddir_rw_sum(td->o.ratemin);
2543                         t_rate += ddir_rw_sum(td->o.rate);
2544                         todo--;
2545                         fio_sem_up(td->sem);
2546                 }
2547
2548                 reap_threads(&nr_running, &t_rate, &m_rate);
2549
2550                 if (todo)
2551                         do_usleep(100000);
2552         }
2553
2554         while (nr_running) {
2555                 reap_threads(&nr_running, &t_rate, &m_rate);
2556                 do_usleep(10000);
2557         }
2558
2559         fio_idle_prof_stop();
2560
2561         update_io_ticks();
2562 }
2563
2564 static void free_disk_util(void)
2565 {
2566         disk_util_prune_entries();
2567         helper_thread_destroy();
2568 }
2569
2570 int fio_backend(struct sk_out *sk_out)
2571 {
2572         struct thread_data *td;
2573         int i;
2574
2575         if (exec_profile) {
2576                 if (load_profile(exec_profile))
2577                         return 1;
2578                 free(exec_profile);
2579                 exec_profile = NULL;
2580         }
2581         if (!thread_number)
2582                 return 0;
2583
2584         if (write_bw_log) {
2585                 struct log_params p = {
2586                         .log_type = IO_LOG_TYPE_BW,
2587                 };
2588
2589                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2590                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2591                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2592         }
2593
2594         if (init_global_dedupe_working_set_seeds()) {
2595                 log_err("fio: failed to initialize global dedupe working set\n");
2596                 return 1;
2597         }
2598
2599         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2600         if (!sk_out)
2601                 is_local_backend = true;
2602         if (startup_sem == NULL)
2603                 return 1;
2604
2605         set_genesis_time();
2606         stat_init();
2607         if (helper_thread_create(startup_sem, sk_out))
2608                 log_err("fio: failed to create helper thread\n");
2609
2610         cgroup_list = smalloc(sizeof(*cgroup_list));
2611         if (cgroup_list)
2612                 INIT_FLIST_HEAD(cgroup_list);
2613
2614         run_threads(sk_out);
2615
2616         helper_thread_exit();
2617
2618         if (!fio_abort) {
2619                 __show_run_stats();
2620                 if (write_bw_log) {
2621                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2622                                 struct io_log *log = agg_io_log[i];
2623
2624                                 flush_log(log, false);
2625                                 free_log(log);
2626                         }
2627                 }
2628         }
2629
2630         for_each_td(td, i) {
2631                 struct thread_stat *ts = &td->ts;
2632
2633                 free_clat_prio_stats(ts);
2634                 steadystate_free(td);
2635                 fio_options_free(td);
2636                 fio_dump_options_free(td);
2637                 if (td->rusage_sem) {
2638                         fio_sem_remove(td->rusage_sem);
2639                         td->rusage_sem = NULL;
2640                 }
2641                 fio_sem_remove(td->sem);
2642                 td->sem = NULL;
2643         }
2644
2645         free_disk_util();
2646         if (cgroup_list) {
2647                 cgroup_kill(cgroup_list);
2648                 sfree(cgroup_list);
2649         }
2650
2651         fio_sem_remove(startup_sem);
2652         stat_exit();
2653         return exit_value;
2654 }