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 <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/mman.h>
  38 #include <math.h>
  39
  40 #include "fio.h"
  41 #ifndef FIO_NO_HAVE_SHM_H
  42 #include <sys/shm.h>
  43 #endif
  44 #include "hash.h"
  45 #include "smalloc.h"
  46 #include "verify.h"
  47 #include "trim.h"
  48 #include "diskutil.h"
  49 #include "cgroup.h"
  50 #include "profile.h"
  51 #include "lib/rand.h"
  52 #include "lib/memalign.h"
  53 #include "server.h"
  54 #include "lib/getrusage.h"
  55 #include "idletime.h"
  56 #include "err.h"
  57 #include "workqueue.h"
  58 #include "lib/mountcheck.h"
  59 #include "rate-submit.h"
  60 #include "helper_thread.h"
  61
  62 static struct fio_mutex *startup_mutex;
  63 static struct flist_head *cgroup_list;
  64 static char *cgroup_mnt;
  65 static int exit_value;
  66 static volatile int fio_abort;
  67 static unsigned int nr_process = 0;
  68 static unsigned int nr_thread = 0;
  69
  70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  71
  72 int groupid = 0;
  73 unsigned int thread_number = 0;
  74 unsigned int stat_number = 0;
  75 int shm_id = 0;
  76 int temp_stall_ts;
  77 unsigned long done_secs = 0;
  78
  79 #define JOB_START_TIMEOUT       (5 * 1000)
  80
  81 static void sig_int(int sig)
  82 {
  83         if (threads) {
  84                 if (is_backend)
  85                         fio_server_got_signal(sig);
  86                 else {
  87                         log_info("\nfio: terminating on signal %d\n", sig);
  88                         log_info_flush();
  89                         exit_value = 128;
  90                 }
  91
  92                 fio_terminate_threads(TERMINATE_ALL);
  93         }
  94 }
  95
  96 void sig_show_status(int sig)
  97 {
  98         show_running_run_stats();
  99 }
 100
 101 static void set_sig_handlers(void)
 102 {
 103         struct sigaction act;
 104
 105         memset(&act, 0, sizeof(act));
 106         act.sa_handler = sig_int;
 107         act.sa_flags = SA_RESTART;
 108         sigaction(SIGINT, &act, NULL);
 109
 110         memset(&act, 0, sizeof(act));
 111         act.sa_handler = sig_int;
 112         act.sa_flags = SA_RESTART;
 113         sigaction(SIGTERM, &act, NULL);
 114
 115 /* Windows uses SIGBREAK as a quit signal from other applications */
 116 #ifdef WIN32
 117         memset(&act, 0, sizeof(act));
 118         act.sa_handler = sig_int;
 119         act.sa_flags = SA_RESTART;
 120         sigaction(SIGBREAK, &act, NULL);
 121 #endif
 122
 123         memset(&act, 0, sizeof(act));
 124         act.sa_handler = sig_show_status;
 125         act.sa_flags = SA_RESTART;
 126         sigaction(SIGUSR1, &act, NULL);
 127
 128         if (is_backend) {
 129                 memset(&act, 0, sizeof(act));
 130                 act.sa_handler = sig_int;
 131                 act.sa_flags = SA_RESTART;
 132                 sigaction(SIGPIPE, &act, NULL);
 133         }
 134 }
 135
 136 /*
 137  * Check if we are above the minimum rate given.
 138  */
 139 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 140                              enum fio_ddir ddir)
 141 {
 142         unsigned long long bytes = 0;
 143         unsigned long iops = 0;
 144         unsigned long spent;
 145         unsigned long rate;
 146         unsigned int ratemin = 0;
 147         unsigned int rate_iops = 0;
 148         unsigned int rate_iops_min = 0;
 149
 150         assert(ddir_rw(ddir));
 151
 152         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 153                 return false;
 154
 155         /*
 156          * allow a 2 second settle period in the beginning
 157          */
 158         if (mtime_since(&td->start, now) < 2000)
 159                 return false;
 160
 161         iops += td->this_io_blocks[ddir];
 162         bytes += td->this_io_bytes[ddir];
 163         ratemin += td->o.ratemin[ddir];
 164         rate_iops += td->o.rate_iops[ddir];
 165         rate_iops_min += td->o.rate_iops_min[ddir];
 166
 167         /*
 168          * if rate blocks is set, sample is running
 169          */
 170         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 171                 spent = mtime_since(&td->lastrate[ddir], now);
 172                 if (spent < td->o.ratecycle)
 173                         return false;
 174
 175                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 176                         /*
 177                          * check bandwidth specified rate
 178                          */
 179                         if (bytes < td->rate_bytes[ddir]) {
 180                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 181                                         td->o.name, ratemin, bytes);
 182                                 return true;
 183                         } else {
 184                                 if (spent)
 185                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 186                                 else
 187                                         rate = 0;
 188
 189                                 if (rate < ratemin ||
 190                                     bytes < td->rate_bytes[ddir]) {
 191                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 192                                                 td->o.name, ratemin, rate);
 193                                         return true;
 194                                 }
 195                         }
 196                 } else {
 197                         /*
 198                          * checks iops specified rate
 199                          */
 200                         if (iops < rate_iops) {
 201                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 202                                                 td->o.name, rate_iops, iops);
 203                                 return true;
 204                         } else {
 205                                 if (spent)
 206                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 207                                 else
 208                                         rate = 0;
 209
 210                                 if (rate < rate_iops_min ||
 211                                     iops < td->rate_blocks[ddir]) {
 212                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 213                                                 td->o.name, rate_iops_min, rate);
 214                                         return true;
 215                                 }
 216                         }
 217                 }
 218         }
 219
 220         td->rate_bytes[ddir] = bytes;
 221         td->rate_blocks[ddir] = iops;
 222         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 223         return false;
 224 }
 225
 226 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 227 {
 228         bool ret = false;
 229
 230         if (td->bytes_done[DDIR_READ])
 231                 ret |= __check_min_rate(td, now, DDIR_READ);
 232         if (td->bytes_done[DDIR_WRITE])
 233                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 234         if (td->bytes_done[DDIR_TRIM])
 235                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 236
 237         return ret;
 238 }
 239
 240 /*
 241  * When job exits, we can cancel the in-flight IO if we are using async
 242  * io. Attempt to do so.
 243  */
 244 static void cleanup_pending_aio(struct thread_data *td)
 245 {
 246         int r;
 247
 248         /*
 249          * get immediately available events, if any
 250          */
 251         r = io_u_queued_complete(td, 0);
 252         if (r < 0)
 253                 return;
 254
 255         /*
 256          * now cancel remaining active events
 257          */
 258         if (td->io_ops->cancel) {
 259                 struct io_u *io_u;
 260                 int i;
 261
 262                 io_u_qiter(&td->io_u_all, io_u, i) {
 263                         if (io_u->flags & IO_U_F_FLIGHT) {
 264                                 r = td->io_ops->cancel(td, io_u);
 265                                 if (!r)
 266                                         put_io_u(td, io_u);
 267                         }
 268                 }
 269         }
 270
 271         if (td->cur_depth)
 272                 r = io_u_queued_complete(td, td->cur_depth);
 273 }
 274
 275 /*
 276  * Helper to handle the final sync of a file. Works just like the normal
 277  * io path, just does everything sync.
 278  */
 279 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 280 {
 281         struct io_u *io_u = __get_io_u(td);
 282         int ret;
 283
 284         if (!io_u)
 285                 return true;
 286
 287         io_u->ddir = DDIR_SYNC;
 288         io_u->file = f;
 289
 290         if (td_io_prep(td, io_u)) {
 291                 put_io_u(td, io_u);
 292                 return true;
 293         }
 294
 295 requeue:
 296         ret = td_io_queue(td, io_u);
 297         if (ret < 0) {
 298                 td_verror(td, io_u->error, "td_io_queue");
 299                 put_io_u(td, io_u);
 300                 return true;
 301         } else if (ret == FIO_Q_QUEUED) {
 302                 if (td_io_commit(td))
 303                         return true;
 304                 if (io_u_queued_complete(td, 1) < 0)
 305                         return true;
 306         } else if (ret == FIO_Q_COMPLETED) {
 307                 if (io_u->error) {
 308                         td_verror(td, io_u->error, "td_io_queue");
 309                         return true;
 310                 }
 311
 312                 if (io_u_sync_complete(td, io_u) < 0)
 313                         return true;
 314         } else if (ret == FIO_Q_BUSY) {
 315                 if (td_io_commit(td))
 316                         return true;
 317                 goto requeue;
 318         }
 319
 320         return false;
 321 }
 322
 323 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 324 {
 325         int ret;
 326
 327         if (fio_file_open(f))
 328                 return fio_io_sync(td, f);
 329
 330         if (td_io_open_file(td, f))
 331                 return 1;
 332
 333         ret = fio_io_sync(td, f);
 334         td_io_close_file(td, f);
 335         return ret;
 336 }
 337
 338 static inline void __update_ts_cache(struct thread_data *td)
 339 {
 340         fio_gettime(&td->ts_cache, NULL);
 341 }
 342
 343 static inline void update_ts_cache(struct thread_data *td)
 344 {
 345         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 346                 __update_ts_cache(td);
 347 }
 348
 349 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 350 {
 351         if (in_ramp_time(td))
 352                 return false;
 353         if (!td->o.timeout)
 354                 return false;
 355         if (utime_since(&td->epoch, t) >= td->o.timeout)
 356                 return true;
 357
 358         return false;
 359 }
 360
 361 /*
 362  * We need to update the runtime consistently in ms, but keep a running
 363  * tally of the current elapsed time in microseconds for sub millisecond
 364  * updates.
 365  */
 366 static inline void update_runtime(struct thread_data *td,
 367                                   unsigned long long *elapsed_us,
 368                                   const enum fio_ddir ddir)
 369 {
 370         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 371                 return;
 372
 373         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 374         elapsed_us[ddir] += utime_since_now(&td->start);
 375         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 376 }
 377
 378 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 379                                 int *retptr)
 380 {
 381         int ret = *retptr;
 382
 383         if (ret < 0 || td->error) {
 384                 int err = td->error;
 385                 enum error_type_bit eb;
 386
 387                 if (ret < 0)
 388                         err = -ret;
 389
 390                 eb = td_error_type(ddir, err);
 391                 if (!(td->o.continue_on_error & (1 << eb)))
 392                         return true;
 393
 394                 if (td_non_fatal_error(td, eb, err)) {
 395                         /*
 396                          * Continue with the I/Os in case of
 397                          * a non fatal error.
 398                          */
 399                         update_error_count(td, err);
 400                         td_clear_error(td);
 401                         *retptr = 0;
 402                         return false;
 403                 } else if (td->o.fill_device && err == ENOSPC) {
 404                         /*
 405                          * We expect to hit this error if
 406                          * fill_device option is set.
 407                          */
 408                         td_clear_error(td);
 409                         fio_mark_td_terminate(td);
 410                         return true;
 411                 } else {
 412                         /*
 413                          * Stop the I/O in case of a fatal
 414                          * error.
 415                          */
 416                         update_error_count(td, err);
 417                         return true;
 418                 }
 419         }
 420
 421         return false;
 422 }
 423
 424 static void check_update_rusage(struct thread_data *td)
 425 {
 426         if (td->update_rusage) {
 427                 td->update_rusage = 0;
 428                 update_rusage_stat(td);
 429                 fio_mutex_up(td->rusage_sem);
 430         }
 431 }
 432
 433 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 434 {
 435         const int full = queue_full(td);
 436         int min_evts = 0;
 437         int ret;
 438
 439         if (td->flags & TD_F_REGROW_LOGS)
 440                 return io_u_quiesce(td);
 441
 442         /*
 443          * if the queue is full, we MUST reap at least 1 event
 444          */
 445         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 446         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 447                 min_evts = 1;
 448
 449         if (time && (__should_check_rate(td, DDIR_READ) ||
 450             __should_check_rate(td, DDIR_WRITE) ||
 451             __should_check_rate(td, DDIR_TRIM)))
 452                 fio_gettime(time, NULL);
 453
 454         do {
 455                 ret = io_u_queued_complete(td, min_evts);
 456                 if (ret < 0)
 457                         break;
 458         } while (full && (td->cur_depth > td->o.iodepth_low));
 459
 460         return ret;
 461 }
 462
 463 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 464                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 465                    struct timespec *comp_time)
 466 {
 467         int ret2;
 468
 469         switch (*ret) {
 470         case FIO_Q_COMPLETED:
 471                 if (io_u->error) {
 472                         *ret = -io_u->error;
 473                         clear_io_u(td, io_u);
 474                 } else if (io_u->resid) {
 475                         int bytes = io_u->xfer_buflen - io_u->resid;
 476                         struct fio_file *f = io_u->file;
 477
 478                         if (bytes_issued)
 479                                 *bytes_issued += bytes;
 480
 481                         if (!from_verify)
 482                                 trim_io_piece(td, io_u);
 483
 484                         /*
 485                          * zero read, fail
 486                          */
 487                         if (!bytes) {
 488                                 if (!from_verify)
 489                                         unlog_io_piece(td, io_u);
 490                                 td_verror(td, EIO, "full resid");
 491                                 put_io_u(td, io_u);
 492                                 break;
 493                         }
 494
 495                         io_u->xfer_buflen = io_u->resid;
 496                         io_u->xfer_buf += bytes;
 497                         io_u->offset += bytes;
 498
 499                         if (ddir_rw(io_u->ddir))
 500                                 td->ts.short_io_u[io_u->ddir]++;
 501
 502                         if (io_u->offset == f->real_file_size)
 503                                 goto sync_done;
 504
 505                         requeue_io_u(td, &io_u);
 506                 } else {
 507 sync_done:
 508                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 509                             __should_check_rate(td, DDIR_WRITE) ||
 510                             __should_check_rate(td, DDIR_TRIM)))
 511                                 fio_gettime(comp_time, NULL);
 512
 513                         *ret = io_u_sync_complete(td, io_u);
 514                         if (*ret < 0)
 515                                 break;
 516                 }
 517
 518                 if (td->flags & TD_F_REGROW_LOGS)
 519                         regrow_logs(td);
 520
 521                 /*
 522                  * when doing I/O (not when verifying),
 523                  * check for any errors that are to be ignored
 524                  */
 525                 if (!from_verify)
 526                         break;
 527
 528                 return 0;
 529         case FIO_Q_QUEUED:
 530                 /*
 531                  * if the engine doesn't have a commit hook,
 532                  * the io_u is really queued. if it does have such
 533                  * a hook, it has to call io_u_queued() itself.
 534                  */
 535                 if (td->io_ops->commit == NULL)
 536                         io_u_queued(td, io_u);
 537                 if (bytes_issued)
 538                         *bytes_issued += io_u->xfer_buflen;
 539                 break;
 540         case FIO_Q_BUSY:
 541                 if (!from_verify)
 542                         unlog_io_piece(td, io_u);
 543                 requeue_io_u(td, &io_u);
 544                 ret2 = td_io_commit(td);
 545                 if (ret2 < 0)
 546                         *ret = ret2;
 547                 break;
 548         default:
 549                 assert(*ret < 0);
 550                 td_verror(td, -(*ret), "td_io_queue");
 551                 break;
 552         }
 553
 554         if (break_on_this_error(td, ddir, ret))
 555                 return 1;
 556
 557         return 0;
 558 }
 559
 560 static inline bool io_in_polling(struct thread_data *td)
 561 {
 562         return !td->o.iodepth_batch_complete_min &&
 563                    !td->o.iodepth_batch_complete_max;
 564 }
 565 /*
 566  * Unlinks files from thread data fio_file structure
 567  */
 568 static int unlink_all_files(struct thread_data *td)
 569 {
 570         struct fio_file *f;
 571         unsigned int i;
 572         int ret = 0;
 573
 574         for_each_file(td, f, i) {
 575                 if (f->filetype != FIO_TYPE_FILE)
 576                         continue;
 577                 ret = td_io_unlink_file(td, f);
 578                 if (ret)
 579                         break;
 580         }
 581
 582         if (ret)
 583                 td_verror(td, ret, "unlink_all_files");
 584
 585         return ret;
 586 }
 587
 588 /*
 589  * Check if io_u will overlap an in-flight IO in the queue
 590  */
 591 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 592 {
 593         bool overlap;
 594         struct io_u *check_io_u;
 595         unsigned long long x1, x2, y1, y2;
 596         int i;
 597
 598         x1 = io_u->offset;
 599         x2 = io_u->offset + io_u->buflen;
 600         overlap = false;
 601         io_u_qiter(q, check_io_u, i) {
 602                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 603                         y1 = check_io_u->offset;
 604                         y2 = check_io_u->offset + check_io_u->buflen;
 605
 606                         if (x1 < y2 && y1 < x2) {
 607                                 overlap = true;
 608                                 dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
 609                                                 x1, io_u->buflen,
 610                                                 y1, check_io_u->buflen);
 611                                 break;
 612                         }
 613                 }
 614         }
 615
 616         return overlap;
 617 }
 618
 619 static int io_u_submit(struct thread_data *td, struct io_u *io_u)
 620 {
 621         /*
 622          * Check for overlap if the user asked us to, and we have
 623          * at least one IO in flight besides this one.
 624          */
 625         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 626             in_flight_overlap(&td->io_u_all, io_u))
 627                 return FIO_Q_BUSY;
 628
 629         return td_io_queue(td, io_u);
 630 }
 631
 632 /*
 633  * The main verify engine. Runs over the writes we previously submitted,
 634  * reads the blocks back in, and checks the crc/md5 of the data.
 635  */
 636 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 637 {
 638         struct fio_file *f;
 639         struct io_u *io_u;
 640         int ret, min_events;
 641         unsigned int i;
 642
 643         dprint(FD_VERIFY, "starting loop\n");
 644
 645         /*
 646          * sync io first and invalidate cache, to make sure we really
 647          * read from disk.
 648          */
 649         for_each_file(td, f, i) {
 650                 if (!fio_file_open(f))
 651                         continue;
 652                 if (fio_io_sync(td, f))
 653                         break;
 654                 if (file_invalidate_cache(td, f))
 655                         break;
 656         }
 657
 658         check_update_rusage(td);
 659
 660         if (td->error)
 661                 return;
 662
 663         /*
 664          * verify_state needs to be reset before verification
 665          * proceeds so that expected random seeds match actual
 666          * random seeds in headers. The main loop will reset
 667          * all random number generators if randrepeat is set.
 668          */
 669         if (!td->o.rand_repeatable)
 670                 td_fill_verify_state_seed(td);
 671
 672         td_set_runstate(td, TD_VERIFYING);
 673
 674         io_u = NULL;
 675         while (!td->terminate) {
 676                 enum fio_ddir ddir;
 677                 int full;
 678
 679                 update_ts_cache(td);
 680                 check_update_rusage(td);
 681
 682                 if (runtime_exceeded(td, &td->ts_cache)) {
 683                         __update_ts_cache(td);
 684                         if (runtime_exceeded(td, &td->ts_cache)) {
 685                                 fio_mark_td_terminate(td);
 686                                 break;
 687                         }
 688                 }
 689
 690                 if (flow_threshold_exceeded(td))
 691                         continue;
 692
 693                 if (!td->o.experimental_verify) {
 694                         io_u = __get_io_u(td);
 695                         if (!io_u)
 696                                 break;
 697
 698                         if (get_next_verify(td, io_u)) {
 699                                 put_io_u(td, io_u);
 700                                 break;
 701                         }
 702
 703                         if (td_io_prep(td, io_u)) {
 704                                 put_io_u(td, io_u);
 705                                 break;
 706                         }
 707                 } else {
 708                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 709                                 break;
 710
 711                         while ((io_u = get_io_u(td)) != NULL) {
 712                                 if (IS_ERR_OR_NULL(io_u)) {
 713                                         io_u = NULL;
 714                                         ret = FIO_Q_BUSY;
 715                                         goto reap;
 716                                 }
 717
 718                                 /*
 719                                  * We are only interested in the places where
 720                                  * we wrote or trimmed IOs. Turn those into
 721                                  * reads for verification purposes.
 722                                  */
 723                                 if (io_u->ddir == DDIR_READ) {
 724                                         /*
 725                                          * Pretend we issued it for rwmix
 726                                          * accounting
 727                                          */
 728                                         td->io_issues[DDIR_READ]++;
 729                                         put_io_u(td, io_u);
 730                                         continue;
 731                                 } else if (io_u->ddir == DDIR_TRIM) {
 732                                         io_u->ddir = DDIR_READ;
 733                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 734                                         break;
 735                                 } else if (io_u->ddir == DDIR_WRITE) {
 736                                         io_u->ddir = DDIR_READ;
 737                                         break;
 738                                 } else {
 739                                         put_io_u(td, io_u);
 740                                         continue;
 741                                 }
 742                         }
 743
 744                         if (!io_u)
 745                                 break;
 746                 }
 747
 748                 if (verify_state_should_stop(td, io_u)) {
 749                         put_io_u(td, io_u);
 750                         break;
 751                 }
 752
 753                 if (td->o.verify_async)
 754                         io_u->end_io = verify_io_u_async;
 755                 else
 756                         io_u->end_io = verify_io_u;
 757
 758                 ddir = io_u->ddir;
 759                 if (!td->o.disable_slat)
 760                         fio_gettime(&io_u->start_time, NULL);
 761
 762                 ret = io_u_submit(td, io_u);
 763
 764                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 765                         break;
 766
 767                 /*
 768                  * if we can queue more, do so. but check if there are
 769                  * completed io_u's first. Note that we can get BUSY even
 770                  * without IO queued, if the system is resource starved.
 771                  */
 772 reap:
 773                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 774                 if (full || io_in_polling(td))
 775                         ret = wait_for_completions(td, NULL);
 776
 777                 if (ret < 0)
 778                         break;
 779         }
 780
 781         check_update_rusage(td);
 782
 783         if (!td->error) {
 784                 min_events = td->cur_depth;
 785
 786                 if (min_events)
 787                         ret = io_u_queued_complete(td, min_events);
 788         } else
 789                 cleanup_pending_aio(td);
 790
 791         td_set_runstate(td, TD_RUNNING);
 792
 793         dprint(FD_VERIFY, "exiting loop\n");
 794 }
 795
 796 static bool exceeds_number_ios(struct thread_data *td)
 797 {
 798         unsigned long long number_ios;
 799
 800         if (!td->o.number_ios)
 801                 return false;
 802
 803         number_ios = ddir_rw_sum(td->io_blocks);
 804         number_ios += td->io_u_queued + td->io_u_in_flight;
 805
 806         return number_ios >= (td->o.number_ios * td->loops);
 807 }
 808
 809 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 810 {
 811         unsigned long long bytes, limit;
 812
 813         if (td_rw(td))
 814                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 815         else if (td_write(td))
 816                 bytes = this_bytes[DDIR_WRITE];
 817         else if (td_read(td))
 818                 bytes = this_bytes[DDIR_READ];
 819         else
 820                 bytes = this_bytes[DDIR_TRIM];
 821
 822         if (td->o.io_size)
 823                 limit = td->o.io_size;
 824         else
 825                 limit = td->o.size;
 826
 827         limit *= td->loops;
 828         return bytes >= limit || exceeds_number_ios(td);
 829 }
 830
 831 static bool io_issue_bytes_exceeded(struct thread_data *td)
 832 {
 833         return io_bytes_exceeded(td, td->io_issue_bytes);
 834 }
 835
 836 static bool io_complete_bytes_exceeded(struct thread_data *td)
 837 {
 838         return io_bytes_exceeded(td, td->this_io_bytes);
 839 }
 840
 841 /*
 842  * used to calculate the next io time for rate control
 843  *
 844  */
 845 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 846 {
 847         uint64_t bps = td->rate_bps[ddir];
 848
 849         assert(!(td->flags & TD_F_CHILD));
 850
 851         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 852                 uint64_t val, iops;
 853
 854                 iops = bps / td->o.bs[ddir];
 855                 val = (int64_t) (1000000 / iops) *
 856                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 857                 if (val) {
 858                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 859                                         (unsigned long long) 1000000 / val,
 860                                         ddir);
 861                 }
 862                 td->last_usec[ddir] += val;
 863                 return td->last_usec[ddir];
 864         } else if (bps) {
 865                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 866                 uint64_t secs = bytes / bps;
 867                 uint64_t remainder = bytes % bps;
 868
 869                 return remainder * 1000000 / bps + secs * 1000000;
 870         }
 871
 872         return 0;
 873 }
 874
 875 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 876 {
 877         unsigned long long b;
 878         uint64_t total;
 879         int left;
 880
 881         b = ddir_rw_sum(td->io_blocks);
 882         if (b % td->o.thinktime_blocks)
 883                 return;
 884
 885         io_u_quiesce(td);
 886
 887         total = 0;
 888         if (td->o.thinktime_spin)
 889                 total = usec_spin(td->o.thinktime_spin);
 890
 891         left = td->o.thinktime - total;
 892         if (left)
 893                 total += usec_sleep(td, left);
 894
 895         /*
 896          * If we're ignoring thinktime for the rate, add the number of bytes
 897          * we would have done while sleeping, minus one block to ensure we
 898          * start issuing immediately after the sleep.
 899          */
 900         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 901                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 902                 uint64_t bs = td->o.min_bs[ddir];
 903                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 904                 uint64_t over;
 905
 906                 if (usperop <= total)
 907                         over = bs;
 908                 else
 909                         over = (usperop - total) / usperop * -bs;
 910
 911                 td->rate_io_issue_bytes[ddir] += (missed - over);
 912         }
 913 }
 914
 915 /*
 916  * Main IO worker function. It retrieves io_u's to process and queues
 917  * and reaps them, checking for rate and errors along the way.
 918  *
 919  * Returns number of bytes written and trimmed.
 920  */
 921 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 922 {
 923         unsigned int i;
 924         int ret = 0;
 925         uint64_t total_bytes, bytes_issued = 0;
 926
 927         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 928                 bytes_done[i] = td->bytes_done[i];
 929
 930         if (in_ramp_time(td))
 931                 td_set_runstate(td, TD_RAMP);
 932         else
 933                 td_set_runstate(td, TD_RUNNING);
 934
 935         lat_target_init(td);
 936
 937         total_bytes = td->o.size;
 938         /*
 939         * Allow random overwrite workloads to write up to io_size
 940         * before starting verification phase as 'size' doesn't apply.
 941         */
 942         if (td_write(td) && td_random(td) && td->o.norandommap)
 943                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 944         /*
 945          * If verify_backlog is enabled, we'll run the verify in this
 946          * handler as well. For that case, we may need up to twice the
 947          * amount of bytes.
 948          */
 949         if (td->o.verify != VERIFY_NONE &&
 950            (td_write(td) && td->o.verify_backlog))
 951                 total_bytes += td->o.size;
 952
 953         /* In trimwrite mode, each byte is trimmed and then written, so
 954          * allow total_bytes to be twice as big */
 955         if (td_trimwrite(td))
 956                 total_bytes += td->total_io_size;
 957
 958         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 959                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 960                 td->o.time_based) {
 961                 struct timespec comp_time;
 962                 struct io_u *io_u;
 963                 int full;
 964                 enum fio_ddir ddir;
 965
 966                 check_update_rusage(td);
 967
 968                 if (td->terminate || td->done)
 969                         break;
 970
 971                 update_ts_cache(td);
 972
 973                 if (runtime_exceeded(td, &td->ts_cache)) {
 974                         __update_ts_cache(td);
 975                         if (runtime_exceeded(td, &td->ts_cache)) {
 976                                 fio_mark_td_terminate(td);
 977                                 break;
 978                         }
 979                 }
 980
 981                 if (flow_threshold_exceeded(td))
 982                         continue;
 983
 984                 /*
 985                  * Break if we exceeded the bytes. The exception is time
 986                  * based runs, but we still need to break out of the loop
 987                  * for those to run verification, if enabled.
 988                  */
 989                 if (bytes_issued >= total_bytes &&
 990                     (!td->o.time_based ||
 991                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 992                         break;
 993
 994                 io_u = get_io_u(td);
 995                 if (IS_ERR_OR_NULL(io_u)) {
 996                         int err = PTR_ERR(io_u);
 997
 998                         io_u = NULL;
 999                         ddir = DDIR_INVAL;
1000                         if (err == -EBUSY) {
1001                                 ret = FIO_Q_BUSY;
1002                                 goto reap;
1003                         }
1004                         if (td->o.latency_target)
1005                                 goto reap;
1006                         break;
1007                 }
1008
1009                 ddir = io_u->ddir;
1010
1011                 /*
1012                  * Add verification end_io handler if:
1013                  *      - Asked to verify (!td_rw(td))
1014                  *      - Or the io_u is from our verify list (mixed write/ver)
1015                  */
1016                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1017                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1018
1019                         if (!td->o.verify_pattern_bytes) {
1020                                 io_u->rand_seed = __rand(&td->verify_state);
1021                                 if (sizeof(int) != sizeof(long *))
1022                                         io_u->rand_seed *= __rand(&td->verify_state);
1023                         }
1024
1025                         if (verify_state_should_stop(td, io_u)) {
1026                                 put_io_u(td, io_u);
1027                                 break;
1028                         }
1029
1030                         if (td->o.verify_async)
1031                                 io_u->end_io = verify_io_u_async;
1032                         else
1033                                 io_u->end_io = verify_io_u;
1034                         td_set_runstate(td, TD_VERIFYING);
1035                 } else if (in_ramp_time(td))
1036                         td_set_runstate(td, TD_RAMP);
1037                 else
1038                         td_set_runstate(td, TD_RUNNING);
1039
1040                 /*
1041                  * Always log IO before it's issued, so we know the specific
1042                  * order of it. The logged unit will track when the IO has
1043                  * completed.
1044                  */
1045                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1046                     td->o.do_verify &&
1047                     td->o.verify != VERIFY_NONE &&
1048                     !td->o.experimental_verify)
1049                         log_io_piece(td, io_u);
1050
1051                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1052                         const unsigned long blen = io_u->xfer_buflen;
1053                         const enum fio_ddir ddir = acct_ddir(io_u);
1054
1055                         if (td->error)
1056                                 break;
1057
1058                         workqueue_enqueue(&td->io_wq, &io_u->work);
1059                         ret = FIO_Q_QUEUED;
1060
1061                         if (ddir_rw(ddir)) {
1062                                 td->io_issues[ddir]++;
1063                                 td->io_issue_bytes[ddir] += blen;
1064                                 td->rate_io_issue_bytes[ddir] += blen;
1065                         }
1066
1067                         if (should_check_rate(td))
1068                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1069
1070                 } else {
1071                         ret = io_u_submit(td, io_u);
1072
1073                         if (should_check_rate(td))
1074                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1075
1076                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1077                                 break;
1078
1079                         /*
1080                          * See if we need to complete some commands. Note that
1081                          * we can get BUSY even without IO queued, if the
1082                          * system is resource starved.
1083                          */
1084 reap:
1085                         full = queue_full(td) ||
1086                                 (ret == FIO_Q_BUSY && td->cur_depth);
1087                         if (full || io_in_polling(td))
1088                                 ret = wait_for_completions(td, &comp_time);
1089                 }
1090                 if (ret < 0)
1091                         break;
1092                 if (!ddir_rw_sum(td->bytes_done) &&
1093                     !td_ioengine_flagged(td, FIO_NOIO))
1094                         continue;
1095
1096                 if (!in_ramp_time(td) && should_check_rate(td)) {
1097                         if (check_min_rate(td, &comp_time)) {
1098                                 if (exitall_on_terminate || td->o.exitall_error)
1099                                         fio_terminate_threads(td->groupid);
1100                                 td_verror(td, EIO, "check_min_rate");
1101                                 break;
1102                         }
1103                 }
1104                 if (!in_ramp_time(td) && td->o.latency_target)
1105                         lat_target_check(td);
1106
1107                 if (ddir_rw(ddir) && td->o.thinktime)
1108                         handle_thinktime(td, ddir);
1109         }
1110
1111         check_update_rusage(td);
1112
1113         if (td->trim_entries)
1114                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1115
1116         if (td->o.fill_device && td->error == ENOSPC) {
1117                 td->error = 0;
1118                 fio_mark_td_terminate(td);
1119         }
1120         if (!td->error) {
1121                 struct fio_file *f;
1122
1123                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1124                         workqueue_flush(&td->io_wq);
1125                         i = 0;
1126                 } else
1127                         i = td->cur_depth;
1128
1129                 if (i) {
1130                         ret = io_u_queued_complete(td, i);
1131                         if (td->o.fill_device && td->error == ENOSPC)
1132                                 td->error = 0;
1133                 }
1134
1135                 if (should_fsync(td) && td->o.end_fsync) {
1136                         td_set_runstate(td, TD_FSYNCING);
1137
1138                         for_each_file(td, f, i) {
1139                                 if (!fio_file_fsync(td, f))
1140                                         continue;
1141
1142                                 log_err("fio: end_fsync failed for file %s\n",
1143                                                                 f->file_name);
1144                         }
1145                 }
1146         } else
1147                 cleanup_pending_aio(td);
1148
1149         /*
1150          * stop job if we failed doing any IO
1151          */
1152         if (!ddir_rw_sum(td->this_io_bytes))
1153                 td->done = 1;
1154
1155         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1156                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1157 }
1158
1159 static void free_file_completion_logging(struct thread_data *td)
1160 {
1161         struct fio_file *f;
1162         unsigned int i;
1163
1164         for_each_file(td, f, i) {
1165                 if (!f->last_write_comp)
1166                         break;
1167                 sfree(f->last_write_comp);
1168         }
1169 }
1170
1171 static int init_file_completion_logging(struct thread_data *td,
1172                                         unsigned int depth)
1173 {
1174         struct fio_file *f;
1175         unsigned int i;
1176
1177         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1178                 return 0;
1179
1180         for_each_file(td, f, i) {
1181                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1182                 if (!f->last_write_comp)
1183                         goto cleanup;
1184         }
1185
1186         return 0;
1187
1188 cleanup:
1189         free_file_completion_logging(td);
1190         log_err("fio: failed to alloc write comp data\n");
1191         return 1;
1192 }
1193
1194 static void cleanup_io_u(struct thread_data *td)
1195 {
1196         struct io_u *io_u;
1197
1198         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1199
1200                 if (td->io_ops->io_u_free)
1201                         td->io_ops->io_u_free(td, io_u);
1202
1203                 fio_memfree(io_u, sizeof(*io_u));
1204         }
1205
1206         free_io_mem(td);
1207
1208         io_u_rexit(&td->io_u_requeues);
1209         io_u_qexit(&td->io_u_freelist);
1210         io_u_qexit(&td->io_u_all);
1211
1212         free_file_completion_logging(td);
1213 }
1214
1215 static int init_io_u(struct thread_data *td)
1216 {
1217         struct io_u *io_u;
1218         unsigned int max_bs, min_write;
1219         int cl_align, i, max_units;
1220         int data_xfer = 1, err;
1221         char *p;
1222
1223         max_units = td->o.iodepth;
1224         max_bs = td_max_bs(td);
1225         min_write = td->o.min_bs[DDIR_WRITE];
1226         td->orig_buffer_size = (unsigned long long) max_bs
1227                                         * (unsigned long long) max_units;
1228
1229         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1230                 data_xfer = 0;
1231
1232         err = 0;
1233         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1234         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1235         err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1236
1237         if (err) {
1238                 log_err("fio: failed setting up IO queues\n");
1239                 return 1;
1240         }
1241
1242         /*
1243          * if we may later need to do address alignment, then add any
1244          * possible adjustment here so that we don't cause a buffer
1245          * overflow later. this adjustment may be too much if we get
1246          * lucky and the allocator gives us an aligned address.
1247          */
1248         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1249             td_ioengine_flagged(td, FIO_RAWIO))
1250                 td->orig_buffer_size += page_mask + td->o.mem_align;
1251
1252         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1253                 unsigned long bs;
1254
1255                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1256                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1257         }
1258
1259         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1260                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1261                 return 1;
1262         }
1263
1264         if (data_xfer && allocate_io_mem(td))
1265                 return 1;
1266
1267         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1268             td_ioengine_flagged(td, FIO_RAWIO))
1269                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1270         else
1271                 p = td->orig_buffer;
1272
1273         cl_align = os_cache_line_size();
1274
1275         for (i = 0; i < max_units; i++) {
1276                 void *ptr;
1277
1278                 if (td->terminate)
1279                         return 1;
1280
1281                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1282                 if (!ptr) {
1283                         log_err("fio: unable to allocate aligned memory\n");
1284                         break;
1285                 }
1286
1287                 io_u = ptr;
1288                 memset(io_u, 0, sizeof(*io_u));
1289                 INIT_FLIST_HEAD(&io_u->verify_list);
1290                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1291
1292                 if (data_xfer) {
1293                         io_u->buf = p;
1294                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1295
1296                         if (td_write(td))
1297                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1298                         if (td_write(td) && td->o.verify_pattern_bytes) {
1299                                 /*
1300                                  * Fill the buffer with the pattern if we are
1301                                  * going to be doing writes.
1302                                  */
1303                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1304                         }
1305                 }
1306
1307                 io_u->index = i;
1308                 io_u->flags = IO_U_F_FREE;
1309                 io_u_qpush(&td->io_u_freelist, io_u);
1310
1311                 /*
1312                  * io_u never leaves this stack, used for iteration of all
1313                  * io_u buffers.
1314                  */
1315                 io_u_qpush(&td->io_u_all, io_u);
1316
1317                 if (td->io_ops->io_u_init) {
1318                         int ret = td->io_ops->io_u_init(td, io_u);
1319
1320                         if (ret) {
1321                                 log_err("fio: failed to init engine data: %d\n", ret);
1322                                 return 1;
1323                         }
1324                 }
1325
1326                 p += max_bs;
1327         }
1328
1329         if (init_file_completion_logging(td, max_units))
1330                 return 1;
1331
1332         return 0;
1333 }
1334
1335 /*
1336  * This function is Linux specific.
1337  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1338  */
1339 static int switch_ioscheduler(struct thread_data *td)
1340 {
1341 #ifdef FIO_HAVE_IOSCHED_SWITCH
1342         char tmp[256], tmp2[128];
1343         FILE *f;
1344         int ret;
1345
1346         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1347                 return 0;
1348
1349         assert(td->files && td->files[0]);
1350         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1351
1352         f = fopen(tmp, "r+");
1353         if (!f) {
1354                 if (errno == ENOENT) {
1355                         log_err("fio: os or kernel doesn't support IO scheduler"
1356                                 " switching\n");
1357                         return 0;
1358                 }
1359                 td_verror(td, errno, "fopen iosched");
1360                 return 1;
1361         }
1362
1363         /*
1364          * Set io scheduler.
1365          */
1366         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1367         if (ferror(f) || ret != 1) {
1368                 td_verror(td, errno, "fwrite");
1369                 fclose(f);
1370                 return 1;
1371         }
1372
1373         rewind(f);
1374
1375         /*
1376          * Read back and check that the selected scheduler is now the default.
1377          */
1378         memset(tmp, 0, sizeof(tmp));
1379         ret = fread(tmp, sizeof(tmp), 1, f);
1380         if (ferror(f) || ret < 0) {
1381                 td_verror(td, errno, "fread");
1382                 fclose(f);
1383                 return 1;
1384         }
1385         /*
1386          * either a list of io schedulers or "none\n" is expected.
1387          */
1388         tmp[strlen(tmp) - 1] = '\0';
1389
1390         /*
1391          * Write to "none" entry doesn't fail, so check the result here.
1392          */
1393         if (!strcmp(tmp, "none")) {
1394                 log_err("fio: io scheduler is not tunable\n");
1395                 fclose(f);
1396                 return 0;
1397         }
1398
1399         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1400         if (!strstr(tmp, tmp2)) {
1401                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1402                 td_verror(td, EINVAL, "iosched_switch");
1403                 fclose(f);
1404                 return 1;
1405         }
1406
1407         fclose(f);
1408         return 0;
1409 #else
1410         return 0;
1411 #endif
1412 }
1413
1414 static bool keep_running(struct thread_data *td)
1415 {
1416         unsigned long long limit;
1417
1418         if (td->done)
1419                 return false;
1420         if (td->terminate)
1421                 return false;
1422         if (td->o.time_based)
1423                 return true;
1424         if (td->o.loops) {
1425                 td->o.loops--;
1426                 return true;
1427         }
1428         if (exceeds_number_ios(td))
1429                 return false;
1430
1431         if (td->o.io_size)
1432                 limit = td->o.io_size;
1433         else
1434                 limit = td->o.size;
1435
1436         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1437                 uint64_t diff;
1438
1439                 /*
1440                  * If the difference is less than the maximum IO size, we
1441                  * are done.
1442                  */
1443                 diff = limit - ddir_rw_sum(td->io_bytes);
1444                 if (diff < td_max_bs(td))
1445                         return false;
1446
1447                 if (fio_files_done(td) && !td->o.io_size)
1448                         return false;
1449
1450                 return true;
1451         }
1452
1453         return false;
1454 }
1455
1456 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1457 {
1458         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1459         int ret;
1460         char *str;
1461
1462         str = malloc(newlen);
1463         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1464
1465         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1466         ret = system(str);
1467         if (ret == -1)
1468                 log_err("fio: exec of cmd <%s> failed\n", str);
1469
1470         free(str);
1471         return ret;
1472 }
1473
1474 /*
1475  * Dry run to compute correct state of numberio for verification.
1476  */
1477 static uint64_t do_dry_run(struct thread_data *td)
1478 {
1479         td_set_runstate(td, TD_RUNNING);
1480
1481         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1482                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1483                 struct io_u *io_u;
1484                 int ret;
1485
1486                 if (td->terminate || td->done)
1487                         break;
1488
1489                 io_u = get_io_u(td);
1490                 if (IS_ERR_OR_NULL(io_u))
1491                         break;
1492
1493                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1494                 io_u->error = 0;
1495                 io_u->resid = 0;
1496                 if (ddir_rw(acct_ddir(io_u)))
1497                         td->io_issues[acct_ddir(io_u)]++;
1498                 if (ddir_rw(io_u->ddir)) {
1499                         io_u_mark_depth(td, 1);
1500                         td->ts.total_io_u[io_u->ddir]++;
1501                 }
1502
1503                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1504                     td->o.do_verify &&
1505                     td->o.verify != VERIFY_NONE &&
1506                     !td->o.experimental_verify)
1507                         log_io_piece(td, io_u);
1508
1509                 ret = io_u_sync_complete(td, io_u);
1510                 (void) ret;
1511         }
1512
1513         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1514 }
1515
1516 struct fork_data {
1517         struct thread_data *td;
1518         struct sk_out *sk_out;
1519 };
1520
1521 /*
1522  * Entry point for the thread based jobs. The process based jobs end up
1523  * here as well, after a little setup.
1524  */
1525 static void *thread_main(void *data)
1526 {
1527         struct fork_data *fd = data;
1528         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1529         struct thread_data *td = fd->td;
1530         struct thread_options *o = &td->o;
1531         struct sk_out *sk_out = fd->sk_out;
1532         uint64_t bytes_done[DDIR_RWDIR_CNT];
1533         int deadlock_loop_cnt;
1534         bool clear_state, did_some_io;
1535         int ret;
1536
1537         sk_out_assign(sk_out);
1538         free(fd);
1539
1540         if (!o->use_thread) {
1541                 setsid();
1542                 td->pid = getpid();
1543         } else
1544                 td->pid = gettid();
1545
1546         fio_local_clock_init(o->use_thread);
1547
1548         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1549
1550         if (is_backend)
1551                 fio_server_send_start(td);
1552
1553         INIT_FLIST_HEAD(&td->io_log_list);
1554         INIT_FLIST_HEAD(&td->io_hist_list);
1555         INIT_FLIST_HEAD(&td->verify_list);
1556         INIT_FLIST_HEAD(&td->trim_list);
1557         INIT_FLIST_HEAD(&td->next_rand_list);
1558         td->io_hist_tree = RB_ROOT;
1559
1560         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1561         if (ret) {
1562                 td_verror(td, ret, "mutex_cond_init_pshared");
1563                 goto err;
1564         }
1565         ret = cond_init_pshared(&td->verify_cond);
1566         if (ret) {
1567                 td_verror(td, ret, "mutex_cond_pshared");
1568                 goto err;
1569         }
1570
1571         td_set_runstate(td, TD_INITIALIZED);
1572         dprint(FD_MUTEX, "up startup_mutex\n");
1573         fio_mutex_up(startup_mutex);
1574         dprint(FD_MUTEX, "wait on td->mutex\n");
1575         fio_mutex_down(td->mutex);
1576         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1577
1578         /*
1579          * A new gid requires privilege, so we need to do this before setting
1580          * the uid.
1581          */
1582         if (o->gid != -1U && setgid(o->gid)) {
1583                 td_verror(td, errno, "setgid");
1584                 goto err;
1585         }
1586         if (o->uid != -1U && setuid(o->uid)) {
1587                 td_verror(td, errno, "setuid");
1588                 goto err;
1589         }
1590
1591         /*
1592          * Do this early, we don't want the compress threads to be limited
1593          * to the same CPUs as the IO workers. So do this before we set
1594          * any potential CPU affinity
1595          */
1596         if (iolog_compress_init(td, sk_out))
1597                 goto err;
1598
1599         /*
1600          * If we have a gettimeofday() thread, make sure we exclude that
1601          * thread from this job
1602          */
1603         if (o->gtod_cpu)
1604                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1605
1606         /*
1607          * Set affinity first, in case it has an impact on the memory
1608          * allocations.
1609          */
1610         if (fio_option_is_set(o, cpumask)) {
1611                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1612                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1613                         if (!ret) {
1614                                 log_err("fio: no CPUs set\n");
1615                                 log_err("fio: Try increasing number of available CPUs\n");
1616                                 td_verror(td, EINVAL, "cpus_split");
1617                                 goto err;
1618                         }
1619                 }
1620                 ret = fio_setaffinity(td->pid, o->cpumask);
1621                 if (ret == -1) {
1622                         td_verror(td, errno, "cpu_set_affinity");
1623                         goto err;
1624                 }
1625         }
1626
1627 #ifdef CONFIG_LIBNUMA
1628         /* numa node setup */
1629         if (fio_option_is_set(o, numa_cpunodes) ||
1630             fio_option_is_set(o, numa_memnodes)) {
1631                 struct bitmask *mask;
1632
1633                 if (numa_available() < 0) {
1634                         td_verror(td, errno, "Does not support NUMA API\n");
1635                         goto err;
1636                 }
1637
1638                 if (fio_option_is_set(o, numa_cpunodes)) {
1639                         mask = numa_parse_nodestring(o->numa_cpunodes);
1640                         ret = numa_run_on_node_mask(mask);
1641                         numa_free_nodemask(mask);
1642                         if (ret == -1) {
1643                                 td_verror(td, errno, \
1644                                         "numa_run_on_node_mask failed\n");
1645                                 goto err;
1646                         }
1647                 }
1648
1649                 if (fio_option_is_set(o, numa_memnodes)) {
1650                         mask = NULL;
1651                         if (o->numa_memnodes)
1652                                 mask = numa_parse_nodestring(o->numa_memnodes);
1653
1654                         switch (o->numa_mem_mode) {
1655                         case MPOL_INTERLEAVE:
1656                                 numa_set_interleave_mask(mask);
1657                                 break;
1658                         case MPOL_BIND:
1659                                 numa_set_membind(mask);
1660                                 break;
1661                         case MPOL_LOCAL:
1662                                 numa_set_localalloc();
1663                                 break;
1664                         case MPOL_PREFERRED:
1665                                 numa_set_preferred(o->numa_mem_prefer_node);
1666                                 break;
1667                         case MPOL_DEFAULT:
1668                         default:
1669                                 break;
1670                         }
1671
1672                         if (mask)
1673                                 numa_free_nodemask(mask);
1674
1675                 }
1676         }
1677 #endif
1678
1679         if (fio_pin_memory(td))
1680                 goto err;
1681
1682         /*
1683          * May alter parameters that init_io_u() will use, so we need to
1684          * do this first.
1685          */
1686         if (init_iolog(td))
1687                 goto err;
1688
1689         if (init_io_u(td))
1690                 goto err;
1691
1692         if (o->verify_async && verify_async_init(td))
1693                 goto err;
1694
1695         if (fio_option_is_set(o, ioprio) ||
1696             fio_option_is_set(o, ioprio_class)) {
1697                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1698                 if (ret == -1) {
1699                         td_verror(td, errno, "ioprio_set");
1700                         goto err;
1701                 }
1702         }
1703
1704         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1705                 goto err;
1706
1707         errno = 0;
1708         if (nice(o->nice) == -1 && errno != 0) {
1709                 td_verror(td, errno, "nice");
1710                 goto err;
1711         }
1712
1713         if (o->ioscheduler && switch_ioscheduler(td))
1714                 goto err;
1715
1716         if (!o->create_serialize && setup_files(td))
1717                 goto err;
1718
1719         if (td_io_init(td))
1720                 goto err;
1721
1722         if (!init_random_map(td))
1723                 goto err;
1724
1725         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1726                 goto err;
1727
1728         if (o->pre_read && !pre_read_files(td))
1729                 goto err;
1730
1731         fio_verify_init(td);
1732
1733         if (rate_submit_init(td, sk_out))
1734                 goto err;
1735
1736         set_epoch_time(td, o->log_unix_epoch);
1737         fio_getrusage(&td->ru_start);
1738         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1739         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1740         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1741
1742         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1743                         o->ratemin[DDIR_TRIM]) {
1744                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1745                                         sizeof(td->bw_sample_time));
1746                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1747                                         sizeof(td->bw_sample_time));
1748                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1749                                         sizeof(td->bw_sample_time));
1750         }
1751
1752         memset(bytes_done, 0, sizeof(bytes_done));
1753         clear_state = false;
1754         did_some_io = false;
1755
1756         while (keep_running(td)) {
1757                 uint64_t verify_bytes;
1758
1759                 fio_gettime(&td->start, NULL);
1760                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1761
1762                 if (clear_state) {
1763                         clear_io_state(td, 0);
1764
1765                         if (o->unlink_each_loop && unlink_all_files(td))
1766                                 break;
1767                 }
1768
1769                 prune_io_piece_log(td);
1770
1771                 if (td->o.verify_only && td_write(td))
1772                         verify_bytes = do_dry_run(td);
1773                 else {
1774                         do_io(td, bytes_done);
1775
1776                         if (!ddir_rw_sum(bytes_done)) {
1777                                 fio_mark_td_terminate(td);
1778                                 verify_bytes = 0;
1779                         } else {
1780                                 verify_bytes = bytes_done[DDIR_WRITE] +
1781                                                 bytes_done[DDIR_TRIM];
1782                         }
1783                 }
1784
1785                 /*
1786                  * If we took too long to shut down, the main thread could
1787                  * already consider us reaped/exited. If that happens, break
1788                  * out and clean up.
1789                  */
1790                 if (td->runstate >= TD_EXITED)
1791                         break;
1792
1793                 clear_state = true;
1794
1795                 /*
1796                  * Make sure we've successfully updated the rusage stats
1797                  * before waiting on the stat mutex. Otherwise we could have
1798                  * the stat thread holding stat mutex and waiting for
1799                  * the rusage_sem, which would never get upped because
1800                  * this thread is waiting for the stat mutex.
1801                  */
1802                 deadlock_loop_cnt = 0;
1803                 do {
1804                         check_update_rusage(td);
1805                         if (!fio_mutex_down_trylock(stat_mutex))
1806                                 break;
1807                         usleep(1000);
1808                         if (deadlock_loop_cnt++ > 5000) {
1809                                 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1810                                 td->error = EDEADLK;
1811                                 goto err;
1812                         }
1813                 } while (1);
1814
1815                 if (td_read(td) && td->io_bytes[DDIR_READ])
1816                         update_runtime(td, elapsed_us, DDIR_READ);
1817                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1818                         update_runtime(td, elapsed_us, DDIR_WRITE);
1819                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1820                         update_runtime(td, elapsed_us, DDIR_TRIM);
1821                 fio_gettime(&td->start, NULL);
1822                 fio_mutex_up(stat_mutex);
1823
1824                 if (td->error || td->terminate)
1825                         break;
1826
1827                 if (!o->do_verify ||
1828                     o->verify == VERIFY_NONE ||
1829                     td_ioengine_flagged(td, FIO_UNIDIR))
1830                         continue;
1831
1832                 if (ddir_rw_sum(bytes_done))
1833                         did_some_io = true;
1834
1835                 clear_io_state(td, 0);
1836
1837                 fio_gettime(&td->start, NULL);
1838
1839                 do_verify(td, verify_bytes);
1840
1841                 /*
1842                  * See comment further up for why this is done here.
1843                  */
1844                 check_update_rusage(td);
1845
1846                 fio_mutex_down(stat_mutex);
1847                 update_runtime(td, elapsed_us, DDIR_READ);
1848                 fio_gettime(&td->start, NULL);
1849                 fio_mutex_up(stat_mutex);
1850
1851                 if (td->error || td->terminate)
1852                         break;
1853         }
1854
1855         /*
1856          * If td ended up with no I/O when it should have had,
1857          * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1858          * (Are we not missing other flags that can be ignored ?)
1859          */
1860         if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1861             !did_some_io && !td->o.create_only &&
1862             !(td_ioengine_flagged(td, FIO_NOIO) ||
1863               td_ioengine_flagged(td, FIO_DISKLESSIO)))
1864                 log_err("%s: No I/O performed by %s, "
1865                          "perhaps try --debug=io option for details?\n",
1866                          td->o.name, td->io_ops->name);
1867
1868         td_set_runstate(td, TD_FINISHING);
1869
1870         update_rusage_stat(td);
1871         td->ts.total_run_time = mtime_since_now(&td->epoch);
1872         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1873         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1874         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1875
1876         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1877             (td->o.verify != VERIFY_NONE && td_write(td)))
1878                 verify_save_state(td->thread_number);
1879
1880         fio_unpin_memory(td);
1881
1882         td_writeout_logs(td, true);
1883
1884         iolog_compress_exit(td);
1885         rate_submit_exit(td);
1886
1887         if (o->exec_postrun)
1888                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1889
1890         if (exitall_on_terminate || (o->exitall_error && td->error))
1891                 fio_terminate_threads(td->groupid);
1892
1893 err:
1894         if (td->error)
1895                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1896                                                         td->verror);
1897
1898         if (o->verify_async)
1899                 verify_async_exit(td);
1900
1901         close_and_free_files(td);
1902         cleanup_io_u(td);
1903         close_ioengine(td);
1904         cgroup_shutdown(td, &cgroup_mnt);
1905         verify_free_state(td);
1906
1907         if (td->zone_state_index) {
1908                 int i;
1909
1910                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1911                         free(td->zone_state_index[i]);
1912                 free(td->zone_state_index);
1913                 td->zone_state_index = NULL;
1914         }
1915
1916         if (fio_option_is_set(o, cpumask)) {
1917                 ret = fio_cpuset_exit(&o->cpumask);
1918                 if (ret)
1919                         td_verror(td, ret, "fio_cpuset_exit");
1920         }
1921
1922         /*
1923          * do this very late, it will log file closing as well
1924          */
1925         if (o->write_iolog_file)
1926                 write_iolog_close(td);
1927
1928         td_set_runstate(td, TD_EXITED);
1929
1930         /*
1931          * Do this last after setting our runstate to exited, so we
1932          * know that the stat thread is signaled.
1933          */
1934         check_update_rusage(td);
1935
1936         sk_out_drop();
1937         return (void *) (uintptr_t) td->error;
1938 }
1939
1940 /*
1941  * Run over the job map and reap the threads that have exited, if any.
1942  */
1943 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1944                          uint64_t *m_rate)
1945 {
1946         struct thread_data *td;
1947         unsigned int cputhreads, realthreads, pending;
1948         int i, status, ret;
1949
1950         /*
1951          * reap exited threads (TD_EXITED -> TD_REAPED)
1952          */
1953         realthreads = pending = cputhreads = 0;
1954         for_each_td(td, i) {
1955                 int flags = 0;
1956
1957                  if (!strcmp(td->o.ioengine, "cpuio"))
1958                         cputhreads++;
1959                 else
1960                         realthreads++;
1961
1962                 if (!td->pid) {
1963                         pending++;
1964                         continue;
1965                 }
1966                 if (td->runstate == TD_REAPED)
1967                         continue;
1968                 if (td->o.use_thread) {
1969                         if (td->runstate == TD_EXITED) {
1970                                 td_set_runstate(td, TD_REAPED);
1971                                 goto reaped;
1972                         }
1973                         continue;
1974                 }
1975
1976                 flags = WNOHANG;
1977                 if (td->runstate == TD_EXITED)
1978                         flags = 0;
1979
1980                 /*
1981                  * check if someone quit or got killed in an unusual way
1982                  */
1983                 ret = waitpid(td->pid, &status, flags);
1984                 if (ret < 0) {
1985                         if (errno == ECHILD) {
1986                                 log_err("fio: pid=%d disappeared %d\n",
1987                                                 (int) td->pid, td->runstate);
1988                                 td->sig = ECHILD;
1989                                 td_set_runstate(td, TD_REAPED);
1990                                 goto reaped;
1991                         }
1992                         perror("waitpid");
1993                 } else if (ret == td->pid) {
1994                         if (WIFSIGNALED(status)) {
1995                                 int sig = WTERMSIG(status);
1996
1997                                 if (sig != SIGTERM && sig != SIGUSR2)
1998                                         log_err("fio: pid=%d, got signal=%d\n",
1999                                                         (int) td->pid, sig);
2000                                 td->sig = sig;
2001                                 td_set_runstate(td, TD_REAPED);
2002                                 goto reaped;
2003                         }
2004                         if (WIFEXITED(status)) {
2005                                 if (WEXITSTATUS(status) && !td->error)
2006                                         td->error = WEXITSTATUS(status);
2007
2008                                 td_set_runstate(td, TD_REAPED);
2009                                 goto reaped;
2010                         }
2011                 }
2012
2013                 /*
2014                  * If the job is stuck, do a forceful timeout of it and
2015                  * move on.
2016                  */
2017                 if (td->terminate &&
2018                     td->runstate < TD_FSYNCING &&
2019                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2020                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2021                                 "%lu seconds, it appears to be stuck. Doing "
2022                                 "forceful exit of this job.\n",
2023                                 td->o.name, td->runstate,
2024                                 (unsigned long) time_since_now(&td->terminate_time));
2025                         td_set_runstate(td, TD_REAPED);
2026                         goto reaped;
2027                 }
2028
2029                 /*
2030                  * thread is not dead, continue
2031                  */
2032                 pending++;
2033                 continue;
2034 reaped:
2035                 (*nr_running)--;
2036                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2037                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2038                 if (!td->pid)
2039                         pending--;
2040
2041                 if (td->error)
2042                         exit_value++;
2043
2044                 done_secs += mtime_since_now(&td->epoch) / 1000;
2045                 profile_td_exit(td);
2046         }
2047
2048         if (*nr_running == cputhreads && !pending && realthreads)
2049                 fio_terminate_threads(TERMINATE_ALL);
2050 }
2051
2052 static bool __check_trigger_file(void)
2053 {
2054         struct stat sb;
2055
2056         if (!trigger_file)
2057                 return false;
2058
2059         if (stat(trigger_file, &sb))
2060                 return false;
2061
2062         if (unlink(trigger_file) < 0)
2063                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2064                                                         strerror(errno));
2065
2066         return true;
2067 }
2068
2069 static bool trigger_timedout(void)
2070 {
2071         if (trigger_timeout)
2072                 if (time_since_genesis() >= trigger_timeout) {
2073                         trigger_timeout = 0;
2074                         return true;
2075                 }
2076
2077         return false;
2078 }
2079
2080 void exec_trigger(const char *cmd)
2081 {
2082         int ret;
2083
2084         if (!cmd || cmd[0] == '\0')
2085                 return;
2086
2087         ret = system(cmd);
2088         if (ret == -1)
2089                 log_err("fio: failed executing %s trigger\n", cmd);
2090 }
2091
2092 void check_trigger_file(void)
2093 {
2094         if (__check_trigger_file() || trigger_timedout()) {
2095                 if (nr_clients)
2096                         fio_clients_send_trigger(trigger_remote_cmd);
2097                 else {
2098                         verify_save_state(IO_LIST_ALL);
2099                         fio_terminate_threads(TERMINATE_ALL);
2100                         exec_trigger(trigger_cmd);
2101                 }
2102         }
2103 }
2104
2105 static int fio_verify_load_state(struct thread_data *td)
2106 {
2107         int ret;
2108
2109         if (!td->o.verify_state)
2110                 return 0;
2111
2112         if (is_backend) {
2113                 void *data;
2114
2115                 ret = fio_server_get_verify_state(td->o.name,
2116                                         td->thread_number - 1, &data);
2117                 if (!ret)
2118                         verify_assign_state(td, data);
2119         } else
2120                 ret = verify_load_state(td, "local");
2121
2122         return ret;
2123 }
2124
2125 static void do_usleep(unsigned int usecs)
2126 {
2127         check_for_running_stats();
2128         check_trigger_file();
2129         usleep(usecs);
2130 }
2131
2132 static bool check_mount_writes(struct thread_data *td)
2133 {
2134         struct fio_file *f;
2135         unsigned int i;
2136
2137         if (!td_write(td) || td->o.allow_mounted_write)
2138                 return false;
2139
2140         /*
2141          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2142          * are mkfs'd and mounted.
2143          */
2144         for_each_file(td, f, i) {
2145 #ifdef FIO_HAVE_CHARDEV_SIZE
2146                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2147 #else
2148                 if (f->filetype != FIO_TYPE_BLOCK)
2149 #endif
2150                         continue;
2151                 if (device_is_mounted(f->file_name))
2152                         goto mounted;
2153         }
2154
2155         return false;
2156 mounted:
2157         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2158         return true;
2159 }
2160
2161 static bool waitee_running(struct thread_data *me)
2162 {
2163         const char *waitee = me->o.wait_for;
2164         const char *self = me->o.name;
2165         struct thread_data *td;
2166         int i;
2167
2168         if (!waitee)
2169                 return false;
2170
2171         for_each_td(td, i) {
2172                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2173                         continue;
2174
2175                 if (td->runstate < TD_EXITED) {
2176                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2177                                         self, td->o.name,
2178                                         runstate_to_name(td->runstate));
2179                         return true;
2180                 }
2181         }
2182
2183         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2184         return false;
2185 }
2186
2187 /*
2188  * Main function for kicking off and reaping jobs, as needed.
2189  */
2190 static void run_threads(struct sk_out *sk_out)
2191 {
2192         struct thread_data *td;
2193         unsigned int i, todo, nr_running, nr_started;
2194         uint64_t m_rate, t_rate;
2195         uint64_t spent;
2196
2197         if (fio_gtod_offload && fio_start_gtod_thread())
2198                 return;
2199
2200         fio_idle_prof_init();
2201
2202         set_sig_handlers();
2203
2204         nr_thread = nr_process = 0;
2205         for_each_td(td, i) {
2206                 if (check_mount_writes(td))
2207                         return;
2208                 if (td->o.use_thread)
2209                         nr_thread++;
2210                 else
2211                         nr_process++;
2212         }
2213
2214         if (output_format & FIO_OUTPUT_NORMAL) {
2215                 log_info("Starting ");
2216                 if (nr_thread)
2217                         log_info("%d thread%s", nr_thread,
2218                                                 nr_thread > 1 ? "s" : "");
2219                 if (nr_process) {
2220                         if (nr_thread)
2221                                 log_info(" and ");
2222                         log_info("%d process%s", nr_process,
2223                                                 nr_process > 1 ? "es" : "");
2224                 }
2225                 log_info("\n");
2226                 log_info_flush();
2227         }
2228
2229         todo = thread_number;
2230         nr_running = 0;
2231         nr_started = 0;
2232         m_rate = t_rate = 0;
2233
2234         for_each_td(td, i) {
2235                 print_status_init(td->thread_number - 1);
2236
2237                 if (!td->o.create_serialize)
2238                         continue;
2239
2240                 if (fio_verify_load_state(td))
2241                         goto reap;
2242
2243                 /*
2244                  * do file setup here so it happens sequentially,
2245                  * we don't want X number of threads getting their
2246                  * client data interspersed on disk
2247                  */
2248                 if (setup_files(td)) {
2249 reap:
2250                         exit_value++;
2251                         if (td->error)
2252                                 log_err("fio: pid=%d, err=%d/%s\n",
2253                                         (int) td->pid, td->error, td->verror);
2254                         td_set_runstate(td, TD_REAPED);
2255                         todo--;
2256                 } else {
2257                         struct fio_file *f;
2258                         unsigned int j;
2259
2260                         /*
2261                          * for sharing to work, each job must always open
2262                          * its own files. so close them, if we opened them
2263                          * for creation
2264                          */
2265                         for_each_file(td, f, j) {
2266                                 if (fio_file_open(f))
2267                                         td_io_close_file(td, f);
2268                         }
2269                 }
2270         }
2271
2272         /* start idle threads before io threads start to run */
2273         fio_idle_prof_start();
2274
2275         set_genesis_time();
2276
2277         while (todo) {
2278                 struct thread_data *map[REAL_MAX_JOBS];
2279                 struct timespec this_start;
2280                 int this_jobs = 0, left;
2281                 struct fork_data *fd;
2282
2283                 /*
2284                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2285                  */
2286                 for_each_td(td, i) {
2287                         if (td->runstate != TD_NOT_CREATED)
2288                                 continue;
2289
2290                         /*
2291                          * never got a chance to start, killed by other
2292                          * thread for some reason
2293                          */
2294                         if (td->terminate) {
2295                                 todo--;
2296                                 continue;
2297                         }
2298
2299                         if (td->o.start_delay) {
2300                                 spent = utime_since_genesis();
2301
2302                                 if (td->o.start_delay > spent)
2303                                         continue;
2304                         }
2305
2306                         if (td->o.stonewall && (nr_started || nr_running)) {
2307                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2308                                                         td->o.name);
2309                                 break;
2310                         }
2311
2312                         if (waitee_running(td)) {
2313                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2314                                                 td->o.name, td->o.wait_for);
2315                                 continue;
2316                         }
2317
2318                         init_disk_util(td);
2319
2320                         td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2321                         td->update_rusage = 0;
2322
2323                         /*
2324                          * Set state to created. Thread will transition
2325                          * to TD_INITIALIZED when it's done setting up.
2326                          */
2327                         td_set_runstate(td, TD_CREATED);
2328                         map[this_jobs++] = td;
2329                         nr_started++;
2330
2331                         fd = calloc(1, sizeof(*fd));
2332                         fd->td = td;
2333                         fd->sk_out = sk_out;
2334
2335                         if (td->o.use_thread) {
2336                                 int ret;
2337
2338                                 dprint(FD_PROCESS, "will pthread_create\n");
2339                                 ret = pthread_create(&td->thread, NULL,
2340                                                         thread_main, fd);
2341                                 if (ret) {
2342                                         log_err("pthread_create: %s\n",
2343                                                         strerror(ret));
2344                                         free(fd);
2345                                         nr_started--;
2346                                         break;
2347                                 }
2348                                 fd = NULL;
2349                                 ret = pthread_detach(td->thread);
2350                                 if (ret)
2351                                         log_err("pthread_detach: %s",
2352                                                         strerror(ret));
2353                         } else {
2354                                 pid_t pid;
2355                                 dprint(FD_PROCESS, "will fork\n");
2356                                 pid = fork();
2357                                 if (!pid) {
2358                                         int ret;
2359
2360                                         ret = (int)(uintptr_t)thread_main(fd);
2361                                         _exit(ret);
2362                                 } else if (i == fio_debug_jobno)
2363                                         *fio_debug_jobp = pid;
2364                         }
2365                         dprint(FD_MUTEX, "wait on startup_mutex\n");
2366                         if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2367                                 log_err("fio: job startup hung? exiting.\n");
2368                                 fio_terminate_threads(TERMINATE_ALL);
2369                                 fio_abort = 1;
2370                                 nr_started--;
2371                                 free(fd);
2372                                 break;
2373                         }
2374                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2375                 }
2376
2377                 /*
2378                  * Wait for the started threads to transition to
2379                  * TD_INITIALIZED.
2380                  */
2381                 fio_gettime(&this_start, NULL);
2382                 left = this_jobs;
2383                 while (left && !fio_abort) {
2384                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2385                                 break;
2386
2387                         do_usleep(100000);
2388
2389                         for (i = 0; i < this_jobs; i++) {
2390                                 td = map[i];
2391                                 if (!td)
2392                                         continue;
2393                                 if (td->runstate == TD_INITIALIZED) {
2394                                         map[i] = NULL;
2395                                         left--;
2396                                 } else if (td->runstate >= TD_EXITED) {
2397                                         map[i] = NULL;
2398                                         left--;
2399                                         todo--;
2400                                         nr_running++; /* work-around... */
2401                                 }
2402                         }
2403                 }
2404
2405                 if (left) {
2406                         log_err("fio: %d job%s failed to start\n", left,
2407                                         left > 1 ? "s" : "");
2408                         for (i = 0; i < this_jobs; i++) {
2409                                 td = map[i];
2410                                 if (!td)
2411                                         continue;
2412                                 kill(td->pid, SIGTERM);
2413                         }
2414                         break;
2415                 }
2416
2417                 /*
2418                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2419                  */
2420                 for_each_td(td, i) {
2421                         if (td->runstate != TD_INITIALIZED)
2422                                 continue;
2423
2424                         if (in_ramp_time(td))
2425                                 td_set_runstate(td, TD_RAMP);
2426                         else
2427                                 td_set_runstate(td, TD_RUNNING);
2428                         nr_running++;
2429                         nr_started--;
2430                         m_rate += ddir_rw_sum(td->o.ratemin);
2431                         t_rate += ddir_rw_sum(td->o.rate);
2432                         todo--;
2433                         fio_mutex_up(td->mutex);
2434                 }
2435
2436                 reap_threads(&nr_running, &t_rate, &m_rate);
2437
2438                 if (todo)
2439                         do_usleep(100000);
2440         }
2441
2442         while (nr_running) {
2443                 reap_threads(&nr_running, &t_rate, &m_rate);
2444                 do_usleep(10000);
2445         }
2446
2447         fio_idle_prof_stop();
2448
2449         update_io_ticks();
2450 }
2451
2452 static void free_disk_util(void)
2453 {
2454         disk_util_prune_entries();
2455         helper_thread_destroy();
2456 }
2457
2458 int fio_backend(struct sk_out *sk_out)
2459 {
2460         struct thread_data *td;
2461         int i;
2462
2463         if (exec_profile) {
2464                 if (load_profile(exec_profile))
2465                         return 1;
2466                 free(exec_profile);
2467                 exec_profile = NULL;
2468         }
2469         if (!thread_number)
2470                 return 0;
2471
2472         if (write_bw_log) {
2473                 struct log_params p = {
2474                         .log_type = IO_LOG_TYPE_BW,
2475                 };
2476
2477                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2478                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2479                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2480         }
2481
2482         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2483         if (startup_mutex == NULL)
2484                 return 1;
2485
2486         set_genesis_time();
2487         stat_init();
2488         helper_thread_create(startup_mutex, sk_out);
2489
2490         cgroup_list = smalloc(sizeof(*cgroup_list));
2491         INIT_FLIST_HEAD(cgroup_list);
2492
2493         run_threads(sk_out);
2494
2495         helper_thread_exit();
2496
2497         if (!fio_abort) {
2498                 __show_run_stats();
2499                 if (write_bw_log) {
2500                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2501                                 struct io_log *log = agg_io_log[i];
2502
2503                                 flush_log(log, false);
2504                                 free_log(log);
2505                         }
2506                 }
2507         }
2508
2509         for_each_td(td, i) {
2510                 steadystate_free(td);
2511                 fio_options_free(td);
2512                 if (td->rusage_sem) {
2513                         fio_mutex_remove(td->rusage_sem);
2514                         td->rusage_sem = NULL;
2515                 }
2516                 fio_mutex_remove(td->mutex);
2517                 td->mutex = NULL;
2518         }
2519
2520         free_disk_util();
2521         cgroup_kill(cgroup_list);
2522         sfree(cgroup_list);
2523         sfree(cgroup_mnt);
2524
2525         fio_mutex_remove(startup_mutex);
2526         stat_exit();
2527         return exit_value;
2528 }