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 #include "pshared.h"
  62
  63 static struct fio_sem *startup_sem;
  64 static struct flist_head *cgroup_list;
  65 static char *cgroup_mnt;
  66 static int exit_value;
  67 static volatile int fio_abort;
  68 static unsigned int nr_process = 0;
  69 static unsigned int nr_thread = 0;
  70
  71 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  72
  73 int groupid = 0;
  74 unsigned int thread_number = 0;
  75 unsigned int stat_number = 0;
  76 int shm_id = 0;
  77 int temp_stall_ts;
  78 unsigned long done_secs = 0;
  79
  80 #define JOB_START_TIMEOUT       (5 * 1000)
  81
  82 static void sig_int(int sig)
  83 {
  84         if (threads) {
  85                 if (is_backend)
  86                         fio_server_got_signal(sig);
  87                 else {
  88                         log_info("\nfio: terminating on signal %d\n", sig);
  89                         log_info_flush();
  90                         exit_value = 128;
  91                 }
  92
  93                 fio_terminate_threads(TERMINATE_ALL);
  94         }
  95 }
  96
  97 void sig_show_status(int sig)
  98 {
  99         show_running_run_stats();
 100 }
 101
 102 static void set_sig_handlers(void)
 103 {
 104         struct sigaction act;
 105
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGINT, &act, NULL);
 110
 111         memset(&act, 0, sizeof(act));
 112         act.sa_handler = sig_int;
 113         act.sa_flags = SA_RESTART;
 114         sigaction(SIGTERM, &act, NULL);
 115
 116 /* Windows uses SIGBREAK as a quit signal from other applications */
 117 #ifdef WIN32
 118         memset(&act, 0, sizeof(act));
 119         act.sa_handler = sig_int;
 120         act.sa_flags = SA_RESTART;
 121         sigaction(SIGBREAK, &act, NULL);
 122 #endif
 123
 124         memset(&act, 0, sizeof(act));
 125         act.sa_handler = sig_show_status;
 126         act.sa_flags = SA_RESTART;
 127         sigaction(SIGUSR1, &act, NULL);
 128
 129         if (is_backend) {
 130                 memset(&act, 0, sizeof(act));
 131                 act.sa_handler = sig_int;
 132                 act.sa_flags = SA_RESTART;
 133                 sigaction(SIGPIPE, &act, NULL);
 134         }
 135 }
 136
 137 /*
 138  * Check if we are above the minimum rate given.
 139  */
 140 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 141                              enum fio_ddir ddir)
 142 {
 143         unsigned long long bytes = 0;
 144         unsigned long iops = 0;
 145         unsigned long spent;
 146         unsigned long rate;
 147         unsigned int ratemin = 0;
 148         unsigned int rate_iops = 0;
 149         unsigned int rate_iops_min = 0;
 150
 151         assert(ddir_rw(ddir));
 152
 153         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 154                 return false;
 155
 156         /*
 157          * allow a 2 second settle period in the beginning
 158          */
 159         if (mtime_since(&td->start, now) < 2000)
 160                 return false;
 161
 162         iops += td->this_io_blocks[ddir];
 163         bytes += td->this_io_bytes[ddir];
 164         ratemin += td->o.ratemin[ddir];
 165         rate_iops += td->o.rate_iops[ddir];
 166         rate_iops_min += td->o.rate_iops_min[ddir];
 167
 168         /*
 169          * if rate blocks is set, sample is running
 170          */
 171         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 172                 spent = mtime_since(&td->lastrate[ddir], now);
 173                 if (spent < td->o.ratecycle)
 174                         return false;
 175
 176                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 177                         /*
 178                          * check bandwidth specified rate
 179                          */
 180                         if (bytes < td->rate_bytes[ddir]) {
 181                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 182                                         td->o.name, ratemin, bytes);
 183                                 return true;
 184                         } else {
 185                                 if (spent)
 186                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 187                                 else
 188                                         rate = 0;
 189
 190                                 if (rate < ratemin ||
 191                                     bytes < td->rate_bytes[ddir]) {
 192                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 193                                                 td->o.name, ratemin, rate);
 194                                         return true;
 195                                 }
 196                         }
 197                 } else {
 198                         /*
 199                          * checks iops specified rate
 200                          */
 201                         if (iops < rate_iops) {
 202                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 203                                                 td->o.name, rate_iops, iops);
 204                                 return true;
 205                         } else {
 206                                 if (spent)
 207                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 208                                 else
 209                                         rate = 0;
 210
 211                                 if (rate < rate_iops_min ||
 212                                     iops < td->rate_blocks[ddir]) {
 213                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 214                                                 td->o.name, rate_iops_min, rate);
 215                                         return true;
 216                                 }
 217                         }
 218                 }
 219         }
 220
 221         td->rate_bytes[ddir] = bytes;
 222         td->rate_blocks[ddir] = iops;
 223         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 224         return false;
 225 }
 226
 227 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 228 {
 229         bool ret = false;
 230
 231         if (td->bytes_done[DDIR_READ])
 232                 ret |= __check_min_rate(td, now, DDIR_READ);
 233         if (td->bytes_done[DDIR_WRITE])
 234                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 235         if (td->bytes_done[DDIR_TRIM])
 236                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 237
 238         return ret;
 239 }
 240
 241 /*
 242  * When job exits, we can cancel the in-flight IO if we are using async
 243  * io. Attempt to do so.
 244  */
 245 static void cleanup_pending_aio(struct thread_data *td)
 246 {
 247         int r;
 248
 249         /*
 250          * get immediately available events, if any
 251          */
 252         r = io_u_queued_complete(td, 0);
 253         if (r < 0)
 254                 return;
 255
 256         /*
 257          * now cancel remaining active events
 258          */
 259         if (td->io_ops->cancel) {
 260                 struct io_u *io_u;
 261                 int i;
 262
 263                 io_u_qiter(&td->io_u_all, io_u, i) {
 264                         if (io_u->flags & IO_U_F_FLIGHT) {
 265                                 r = td->io_ops->cancel(td, io_u);
 266                                 if (!r)
 267                                         put_io_u(td, io_u);
 268                         }
 269                 }
 270         }
 271
 272         if (td->cur_depth)
 273                 r = io_u_queued_complete(td, td->cur_depth);
 274 }
 275
 276 /*
 277  * Helper to handle the final sync of a file. Works just like the normal
 278  * io path, just does everything sync.
 279  */
 280 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 281 {
 282         struct io_u *io_u = __get_io_u(td);
 283         int ret;
 284
 285         if (!io_u)
 286                 return true;
 287
 288         io_u->ddir = DDIR_SYNC;
 289         io_u->file = f;
 290
 291         if (td_io_prep(td, io_u)) {
 292                 put_io_u(td, io_u);
 293                 return true;
 294         }
 295
 296 requeue:
 297         ret = td_io_queue(td, io_u);
 298         if (ret < 0) {
 299                 td_verror(td, io_u->error, "td_io_queue");
 300                 put_io_u(td, io_u);
 301                 return true;
 302         } else if (ret == FIO_Q_QUEUED) {
 303                 if (td_io_commit(td))
 304                         return true;
 305                 if (io_u_queued_complete(td, 1) < 0)
 306                         return true;
 307         } else if (ret == FIO_Q_COMPLETED) {
 308                 if (io_u->error) {
 309                         td_verror(td, io_u->error, "td_io_queue");
 310                         return true;
 311                 }
 312
 313                 if (io_u_sync_complete(td, io_u) < 0)
 314                         return true;
 315         } else if (ret == FIO_Q_BUSY) {
 316                 if (td_io_commit(td))
 317                         return true;
 318                 goto requeue;
 319         }
 320
 321         return false;
 322 }
 323
 324 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 325 {
 326         int ret;
 327
 328         if (fio_file_open(f))
 329                 return fio_io_sync(td, f);
 330
 331         if (td_io_open_file(td, f))
 332                 return 1;
 333
 334         ret = fio_io_sync(td, f);
 335         td_io_close_file(td, f);
 336         return ret;
 337 }
 338
 339 static inline void __update_ts_cache(struct thread_data *td)
 340 {
 341         fio_gettime(&td->ts_cache, NULL);
 342 }
 343
 344 static inline void update_ts_cache(struct thread_data *td)
 345 {
 346         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 347                 __update_ts_cache(td);
 348 }
 349
 350 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 351 {
 352         if (in_ramp_time(td))
 353                 return false;
 354         if (!td->o.timeout)
 355                 return false;
 356         if (utime_since(&td->epoch, t) >= td->o.timeout)
 357                 return true;
 358
 359         return false;
 360 }
 361
 362 /*
 363  * We need to update the runtime consistently in ms, but keep a running
 364  * tally of the current elapsed time in microseconds for sub millisecond
 365  * updates.
 366  */
 367 static inline void update_runtime(struct thread_data *td,
 368                                   unsigned long long *elapsed_us,
 369                                   const enum fio_ddir ddir)
 370 {
 371         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 372                 return;
 373
 374         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 375         elapsed_us[ddir] += utime_since_now(&td->start);
 376         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 377 }
 378
 379 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 380                                 int *retptr)
 381 {
 382         int ret = *retptr;
 383
 384         if (ret < 0 || td->error) {
 385                 int err = td->error;
 386                 enum error_type_bit eb;
 387
 388                 if (ret < 0)
 389                         err = -ret;
 390
 391                 eb = td_error_type(ddir, err);
 392                 if (!(td->o.continue_on_error & (1 << eb)))
 393                         return true;
 394
 395                 if (td_non_fatal_error(td, eb, err)) {
 396                         /*
 397                          * Continue with the I/Os in case of
 398                          * a non fatal error.
 399                          */
 400                         update_error_count(td, err);
 401                         td_clear_error(td);
 402                         *retptr = 0;
 403                         return false;
 404                 } else if (td->o.fill_device && err == ENOSPC) {
 405                         /*
 406                          * We expect to hit this error if
 407                          * fill_device option is set.
 408                          */
 409                         td_clear_error(td);
 410                         fio_mark_td_terminate(td);
 411                         return true;
 412                 } else {
 413                         /*
 414                          * Stop the I/O in case of a fatal
 415                          * error.
 416                          */
 417                         update_error_count(td, err);
 418                         return true;
 419                 }
 420         }
 421
 422         return false;
 423 }
 424
 425 static void check_update_rusage(struct thread_data *td)
 426 {
 427         if (td->update_rusage) {
 428                 td->update_rusage = 0;
 429                 update_rusage_stat(td);
 430                 fio_sem_up(td->rusage_sem);
 431         }
 432 }
 433
 434 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 435 {
 436         const int full = queue_full(td);
 437         int min_evts = 0;
 438         int ret;
 439
 440         if (td->flags & TD_F_REGROW_LOGS)
 441                 return io_u_quiesce(td);
 442
 443         /*
 444          * if the queue is full, we MUST reap at least 1 event
 445          */
 446         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 447         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 448                 min_evts = 1;
 449
 450         if (time && (__should_check_rate(td, DDIR_READ) ||
 451             __should_check_rate(td, DDIR_WRITE) ||
 452             __should_check_rate(td, DDIR_TRIM)))
 453                 fio_gettime(time, NULL);
 454
 455         do {
 456                 ret = io_u_queued_complete(td, min_evts);
 457                 if (ret < 0)
 458                         break;
 459         } while (full && (td->cur_depth > td->o.iodepth_low));
 460
 461         return ret;
 462 }
 463
 464 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 465                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 466                    struct timespec *comp_time)
 467 {
 468         int ret2;
 469
 470         switch (*ret) {
 471         case FIO_Q_COMPLETED:
 472                 if (io_u->error) {
 473                         *ret = -io_u->error;
 474                         clear_io_u(td, io_u);
 475                 } else if (io_u->resid) {
 476                         int bytes = io_u->xfer_buflen - io_u->resid;
 477                         struct fio_file *f = io_u->file;
 478
 479                         if (bytes_issued)
 480                                 *bytes_issued += bytes;
 481
 482                         if (!from_verify)
 483                                 trim_io_piece(td, io_u);
 484
 485                         /*
 486                          * zero read, fail
 487                          */
 488                         if (!bytes) {
 489                                 if (!from_verify)
 490                                         unlog_io_piece(td, io_u);
 491                                 td_verror(td, EIO, "full resid");
 492                                 put_io_u(td, io_u);
 493                                 break;
 494                         }
 495
 496                         io_u->xfer_buflen = io_u->resid;
 497                         io_u->xfer_buf += bytes;
 498                         io_u->offset += bytes;
 499
 500                         if (ddir_rw(io_u->ddir))
 501                                 td->ts.short_io_u[io_u->ddir]++;
 502
 503                         if (io_u->offset == f->real_file_size)
 504                                 goto sync_done;
 505
 506                         requeue_io_u(td, &io_u);
 507                 } else {
 508 sync_done:
 509                         if (comp_time && (__should_check_rate(td, DDIR_READ) ||
 510                             __should_check_rate(td, DDIR_WRITE) ||
 511                             __should_check_rate(td, DDIR_TRIM)))
 512                                 fio_gettime(comp_time, NULL);
 513
 514                         *ret = io_u_sync_complete(td, io_u);
 515                         if (*ret < 0)
 516                                 break;
 517                 }
 518
 519                 if (td->flags & TD_F_REGROW_LOGS)
 520                         regrow_logs(td);
 521
 522                 /*
 523                  * when doing I/O (not when verifying),
 524                  * check for any errors that are to be ignored
 525                  */
 526                 if (!from_verify)
 527                         break;
 528
 529                 return 0;
 530         case FIO_Q_QUEUED:
 531                 /*
 532                  * if the engine doesn't have a commit hook,
 533                  * the io_u is really queued. if it does have such
 534                  * a hook, it has to call io_u_queued() itself.
 535                  */
 536                 if (td->io_ops->commit == NULL)
 537                         io_u_queued(td, io_u);
 538                 if (bytes_issued)
 539                         *bytes_issued += io_u->xfer_buflen;
 540                 break;
 541         case FIO_Q_BUSY:
 542                 if (!from_verify)
 543                         unlog_io_piece(td, io_u);
 544                 requeue_io_u(td, &io_u);
 545                 ret2 = td_io_commit(td);
 546                 if (ret2 < 0)
 547                         *ret = ret2;
 548                 break;
 549         default:
 550                 assert(*ret < 0);
 551                 td_verror(td, -(*ret), "td_io_queue");
 552                 break;
 553         }
 554
 555         if (break_on_this_error(td, ddir, ret))
 556                 return 1;
 557
 558         return 0;
 559 }
 560
 561 static inline bool io_in_polling(struct thread_data *td)
 562 {
 563         return !td->o.iodepth_batch_complete_min &&
 564                    !td->o.iodepth_batch_complete_max;
 565 }
 566 /*
 567  * Unlinks files from thread data fio_file structure
 568  */
 569 static int unlink_all_files(struct thread_data *td)
 570 {
 571         struct fio_file *f;
 572         unsigned int i;
 573         int ret = 0;
 574
 575         for_each_file(td, f, i) {
 576                 if (f->filetype != FIO_TYPE_FILE)
 577                         continue;
 578                 ret = td_io_unlink_file(td, f);
 579                 if (ret)
 580                         break;
 581         }
 582
 583         if (ret)
 584                 td_verror(td, ret, "unlink_all_files");
 585
 586         return ret;
 587 }
 588
 589 /*
 590  * Check if io_u will overlap an in-flight IO in the queue
 591  */
 592 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 593 {
 594         bool overlap;
 595         struct io_u *check_io_u;
 596         unsigned long long x1, x2, y1, y2;
 597         int i;
 598
 599         x1 = io_u->offset;
 600         x2 = io_u->offset + io_u->buflen;
 601         overlap = false;
 602         io_u_qiter(q, check_io_u, i) {
 603                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 604                         y1 = check_io_u->offset;
 605                         y2 = check_io_u->offset + check_io_u->buflen;
 606
 607                         if (x1 < y2 && y1 < x2) {
 608                                 overlap = true;
 609                                 dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
 610                                                 x1, io_u->buflen,
 611                                                 y1, check_io_u->buflen);
 612                                 break;
 613                         }
 614                 }
 615         }
 616
 617         return overlap;
 618 }
 619
 620 static int io_u_submit(struct thread_data *td, struct io_u *io_u)
 621 {
 622         /*
 623          * Check for overlap if the user asked us to, and we have
 624          * at least one IO in flight besides this one.
 625          */
 626         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 627             in_flight_overlap(&td->io_u_all, io_u))
 628                 return FIO_Q_BUSY;
 629
 630         return td_io_queue(td, io_u);
 631 }
 632
 633 /*
 634  * The main verify engine. Runs over the writes we previously submitted,
 635  * reads the blocks back in, and checks the crc/md5 of the data.
 636  */
 637 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 638 {
 639         struct fio_file *f;
 640         struct io_u *io_u;
 641         int ret, min_events;
 642         unsigned int i;
 643
 644         dprint(FD_VERIFY, "starting loop\n");
 645
 646         /*
 647          * sync io first and invalidate cache, to make sure we really
 648          * read from disk.
 649          */
 650         for_each_file(td, f, i) {
 651                 if (!fio_file_open(f))
 652                         continue;
 653                 if (fio_io_sync(td, f))
 654                         break;
 655                 if (file_invalidate_cache(td, f))
 656                         break;
 657         }
 658
 659         check_update_rusage(td);
 660
 661         if (td->error)
 662                 return;
 663
 664         /*
 665          * verify_state needs to be reset before verification
 666          * proceeds so that expected random seeds match actual
 667          * random seeds in headers. The main loop will reset
 668          * all random number generators if randrepeat is set.
 669          */
 670         if (!td->o.rand_repeatable)
 671                 td_fill_verify_state_seed(td);
 672
 673         td_set_runstate(td, TD_VERIFYING);
 674
 675         io_u = NULL;
 676         while (!td->terminate) {
 677                 enum fio_ddir ddir;
 678                 int full;
 679
 680                 update_ts_cache(td);
 681                 check_update_rusage(td);
 682
 683                 if (runtime_exceeded(td, &td->ts_cache)) {
 684                         __update_ts_cache(td);
 685                         if (runtime_exceeded(td, &td->ts_cache)) {
 686                                 fio_mark_td_terminate(td);
 687                                 break;
 688                         }
 689                 }
 690
 691                 if (flow_threshold_exceeded(td))
 692                         continue;
 693
 694                 if (!td->o.experimental_verify) {
 695                         io_u = __get_io_u(td);
 696                         if (!io_u)
 697                                 break;
 698
 699                         if (get_next_verify(td, io_u)) {
 700                                 put_io_u(td, io_u);
 701                                 break;
 702                         }
 703
 704                         if (td_io_prep(td, io_u)) {
 705                                 put_io_u(td, io_u);
 706                                 break;
 707                         }
 708                 } else {
 709                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 710                                 break;
 711
 712                         while ((io_u = get_io_u(td)) != NULL) {
 713                                 if (IS_ERR_OR_NULL(io_u)) {
 714                                         io_u = NULL;
 715                                         ret = FIO_Q_BUSY;
 716                                         goto reap;
 717                                 }
 718
 719                                 /*
 720                                  * We are only interested in the places where
 721                                  * we wrote or trimmed IOs. Turn those into
 722                                  * reads for verification purposes.
 723                                  */
 724                                 if (io_u->ddir == DDIR_READ) {
 725                                         /*
 726                                          * Pretend we issued it for rwmix
 727                                          * accounting
 728                                          */
 729                                         td->io_issues[DDIR_READ]++;
 730                                         put_io_u(td, io_u);
 731                                         continue;
 732                                 } else if (io_u->ddir == DDIR_TRIM) {
 733                                         io_u->ddir = DDIR_READ;
 734                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 735                                         break;
 736                                 } else if (io_u->ddir == DDIR_WRITE) {
 737                                         io_u->ddir = DDIR_READ;
 738                                         break;
 739                                 } else {
 740                                         put_io_u(td, io_u);
 741                                         continue;
 742                                 }
 743                         }
 744
 745                         if (!io_u)
 746                                 break;
 747                 }
 748
 749                 if (verify_state_should_stop(td, io_u)) {
 750                         put_io_u(td, io_u);
 751                         break;
 752                 }
 753
 754                 if (td->o.verify_async)
 755                         io_u->end_io = verify_io_u_async;
 756                 else
 757                         io_u->end_io = verify_io_u;
 758
 759                 ddir = io_u->ddir;
 760                 if (!td->o.disable_slat)
 761                         fio_gettime(&io_u->start_time, NULL);
 762
 763                 ret = io_u_submit(td, io_u);
 764
 765                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 766                         break;
 767
 768                 /*
 769                  * if we can queue more, do so. but check if there are
 770                  * completed io_u's first. Note that we can get BUSY even
 771                  * without IO queued, if the system is resource starved.
 772                  */
 773 reap:
 774                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 775                 if (full || io_in_polling(td))
 776                         ret = wait_for_completions(td, NULL);
 777
 778                 if (ret < 0)
 779                         break;
 780         }
 781
 782         check_update_rusage(td);
 783
 784         if (!td->error) {
 785                 min_events = td->cur_depth;
 786
 787                 if (min_events)
 788                         ret = io_u_queued_complete(td, min_events);
 789         } else
 790                 cleanup_pending_aio(td);
 791
 792         td_set_runstate(td, TD_RUNNING);
 793
 794         dprint(FD_VERIFY, "exiting loop\n");
 795 }
 796
 797 static bool exceeds_number_ios(struct thread_data *td)
 798 {
 799         unsigned long long number_ios;
 800
 801         if (!td->o.number_ios)
 802                 return false;
 803
 804         number_ios = ddir_rw_sum(td->io_blocks);
 805         number_ios += td->io_u_queued + td->io_u_in_flight;
 806
 807         return number_ios >= (td->o.number_ios * td->loops);
 808 }
 809
 810 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 811 {
 812         unsigned long long bytes, limit;
 813
 814         if (td_rw(td))
 815                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 816         else if (td_write(td))
 817                 bytes = this_bytes[DDIR_WRITE];
 818         else if (td_read(td))
 819                 bytes = this_bytes[DDIR_READ];
 820         else
 821                 bytes = this_bytes[DDIR_TRIM];
 822
 823         if (td->o.io_size)
 824                 limit = td->o.io_size;
 825         else
 826                 limit = td->o.size;
 827
 828         limit *= td->loops;
 829         return bytes >= limit || exceeds_number_ios(td);
 830 }
 831
 832 static bool io_issue_bytes_exceeded(struct thread_data *td)
 833 {
 834         return io_bytes_exceeded(td, td->io_issue_bytes);
 835 }
 836
 837 static bool io_complete_bytes_exceeded(struct thread_data *td)
 838 {
 839         return io_bytes_exceeded(td, td->this_io_bytes);
 840 }
 841
 842 /*
 843  * used to calculate the next io time for rate control
 844  *
 845  */
 846 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 847 {
 848         uint64_t bps = td->rate_bps[ddir];
 849
 850         assert(!(td->flags & TD_F_CHILD));
 851
 852         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 853                 uint64_t val, iops;
 854
 855                 iops = bps / td->o.bs[ddir];
 856                 val = (int64_t) (1000000 / iops) *
 857                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 858                 if (val) {
 859                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 860                                         (unsigned long long) 1000000 / val,
 861                                         ddir);
 862                 }
 863                 td->last_usec[ddir] += val;
 864                 return td->last_usec[ddir];
 865         } else if (bps) {
 866                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 867                 uint64_t secs = bytes / bps;
 868                 uint64_t remainder = bytes % bps;
 869
 870                 return remainder * 1000000 / bps + secs * 1000000;
 871         }
 872
 873         return 0;
 874 }
 875
 876 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 877 {
 878         unsigned long long b;
 879         uint64_t total;
 880         int left;
 881
 882         b = ddir_rw_sum(td->io_blocks);
 883         if (b % td->o.thinktime_blocks)
 884                 return;
 885
 886         io_u_quiesce(td);
 887
 888         total = 0;
 889         if (td->o.thinktime_spin)
 890                 total = usec_spin(td->o.thinktime_spin);
 891
 892         left = td->o.thinktime - total;
 893         if (left)
 894                 total += usec_sleep(td, left);
 895
 896         /*
 897          * If we're ignoring thinktime for the rate, add the number of bytes
 898          * we would have done while sleeping, minus one block to ensure we
 899          * start issuing immediately after the sleep.
 900          */
 901         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 902                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 903                 uint64_t bs = td->o.min_bs[ddir];
 904                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 905                 uint64_t over;
 906
 907                 if (usperop <= total)
 908                         over = bs;
 909                 else
 910                         over = (usperop - total) / usperop * -bs;
 911
 912                 td->rate_io_issue_bytes[ddir] += (missed - over);
 913         }
 914 }
 915
 916 /*
 917  * Main IO worker function. It retrieves io_u's to process and queues
 918  * and reaps them, checking for rate and errors along the way.
 919  *
 920  * Returns number of bytes written and trimmed.
 921  */
 922 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 923 {
 924         unsigned int i;
 925         int ret = 0;
 926         uint64_t total_bytes, bytes_issued = 0;
 927
 928         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 929                 bytes_done[i] = td->bytes_done[i];
 930
 931         if (in_ramp_time(td))
 932                 td_set_runstate(td, TD_RAMP);
 933         else
 934                 td_set_runstate(td, TD_RUNNING);
 935
 936         lat_target_init(td);
 937
 938         total_bytes = td->o.size;
 939         /*
 940         * Allow random overwrite workloads to write up to io_size
 941         * before starting verification phase as 'size' doesn't apply.
 942         */
 943         if (td_write(td) && td_random(td) && td->o.norandommap)
 944                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 945         /*
 946          * If verify_backlog is enabled, we'll run the verify in this
 947          * handler as well. For that case, we may need up to twice the
 948          * amount of bytes.
 949          */
 950         if (td->o.verify != VERIFY_NONE &&
 951            (td_write(td) && td->o.verify_backlog))
 952                 total_bytes += td->o.size;
 953
 954         /* In trimwrite mode, each byte is trimmed and then written, so
 955          * allow total_bytes to be twice as big */
 956         if (td_trimwrite(td))
 957                 total_bytes += td->total_io_size;
 958
 959         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 960                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 961                 td->o.time_based) {
 962                 struct timespec comp_time;
 963                 struct io_u *io_u;
 964                 int full;
 965                 enum fio_ddir ddir;
 966
 967                 check_update_rusage(td);
 968
 969                 if (td->terminate || td->done)
 970                         break;
 971
 972                 update_ts_cache(td);
 973
 974                 if (runtime_exceeded(td, &td->ts_cache)) {
 975                         __update_ts_cache(td);
 976                         if (runtime_exceeded(td, &td->ts_cache)) {
 977                                 fio_mark_td_terminate(td);
 978                                 break;
 979                         }
 980                 }
 981
 982                 if (flow_threshold_exceeded(td))
 983                         continue;
 984
 985                 /*
 986                  * Break if we exceeded the bytes. The exception is time
 987                  * based runs, but we still need to break out of the loop
 988                  * for those to run verification, if enabled.
 989                  */
 990                 if (bytes_issued >= total_bytes &&
 991                     (!td->o.time_based ||
 992                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 993                         break;
 994
 995                 io_u = get_io_u(td);
 996                 if (IS_ERR_OR_NULL(io_u)) {
 997                         int err = PTR_ERR(io_u);
 998
 999                         io_u = NULL;
1000                         ddir = DDIR_INVAL;
1001                         if (err == -EBUSY) {
1002                                 ret = FIO_Q_BUSY;
1003                                 goto reap;
1004                         }
1005                         if (td->o.latency_target)
1006                                 goto reap;
1007                         break;
1008                 }
1009
1010                 ddir = io_u->ddir;
1011
1012                 /*
1013                  * Add verification end_io handler if:
1014                  *      - Asked to verify (!td_rw(td))
1015                  *      - Or the io_u is from our verify list (mixed write/ver)
1016                  */
1017                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1018                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1019
1020                         if (!td->o.verify_pattern_bytes) {
1021                                 io_u->rand_seed = __rand(&td->verify_state);
1022                                 if (sizeof(int) != sizeof(long *))
1023                                         io_u->rand_seed *= __rand(&td->verify_state);
1024                         }
1025
1026                         if (verify_state_should_stop(td, io_u)) {
1027                                 put_io_u(td, io_u);
1028                                 break;
1029                         }
1030
1031                         if (td->o.verify_async)
1032                                 io_u->end_io = verify_io_u_async;
1033                         else
1034                                 io_u->end_io = verify_io_u;
1035                         td_set_runstate(td, TD_VERIFYING);
1036                 } else if (in_ramp_time(td))
1037                         td_set_runstate(td, TD_RAMP);
1038                 else
1039                         td_set_runstate(td, TD_RUNNING);
1040
1041                 /*
1042                  * Always log IO before it's issued, so we know the specific
1043                  * order of it. The logged unit will track when the IO has
1044                  * completed.
1045                  */
1046                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1047                     td->o.do_verify &&
1048                     td->o.verify != VERIFY_NONE &&
1049                     !td->o.experimental_verify)
1050                         log_io_piece(td, io_u);
1051
1052                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1053                         const unsigned long blen = io_u->xfer_buflen;
1054                         const enum fio_ddir ddir = acct_ddir(io_u);
1055
1056                         if (td->error)
1057                                 break;
1058
1059                         workqueue_enqueue(&td->io_wq, &io_u->work);
1060                         ret = FIO_Q_QUEUED;
1061
1062                         if (ddir_rw(ddir)) {
1063                                 td->io_issues[ddir]++;
1064                                 td->io_issue_bytes[ddir] += blen;
1065                                 td->rate_io_issue_bytes[ddir] += blen;
1066                         }
1067
1068                         if (should_check_rate(td))
1069                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1070
1071                 } else {
1072                         ret = io_u_submit(td, io_u);
1073
1074                         if (should_check_rate(td))
1075                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1076
1077                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1078                                 break;
1079
1080                         /*
1081                          * See if we need to complete some commands. Note that
1082                          * we can get BUSY even without IO queued, if the
1083                          * system is resource starved.
1084                          */
1085 reap:
1086                         full = queue_full(td) ||
1087                                 (ret == FIO_Q_BUSY && td->cur_depth);
1088                         if (full || io_in_polling(td))
1089                                 ret = wait_for_completions(td, &comp_time);
1090                 }
1091                 if (ret < 0)
1092                         break;
1093                 if (!ddir_rw_sum(td->bytes_done) &&
1094                     !td_ioengine_flagged(td, FIO_NOIO))
1095                         continue;
1096
1097                 if (!in_ramp_time(td) && should_check_rate(td)) {
1098                         if (check_min_rate(td, &comp_time)) {
1099                                 if (exitall_on_terminate || td->o.exitall_error)
1100                                         fio_terminate_threads(td->groupid);
1101                                 td_verror(td, EIO, "check_min_rate");
1102                                 break;
1103                         }
1104                 }
1105                 if (!in_ramp_time(td) && td->o.latency_target)
1106                         lat_target_check(td);
1107
1108                 if (ddir_rw(ddir) && td->o.thinktime)
1109                         handle_thinktime(td, ddir);
1110         }
1111
1112         check_update_rusage(td);
1113
1114         if (td->trim_entries)
1115                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1116
1117         if (td->o.fill_device && td->error == ENOSPC) {
1118                 td->error = 0;
1119                 fio_mark_td_terminate(td);
1120         }
1121         if (!td->error) {
1122                 struct fio_file *f;
1123
1124                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1125                         workqueue_flush(&td->io_wq);
1126                         i = 0;
1127                 } else
1128                         i = td->cur_depth;
1129
1130                 if (i) {
1131                         ret = io_u_queued_complete(td, i);
1132                         if (td->o.fill_device && td->error == ENOSPC)
1133                                 td->error = 0;
1134                 }
1135
1136                 if (should_fsync(td) && td->o.end_fsync) {
1137                         td_set_runstate(td, TD_FSYNCING);
1138
1139                         for_each_file(td, f, i) {
1140                                 if (!fio_file_fsync(td, f))
1141                                         continue;
1142
1143                                 log_err("fio: end_fsync failed for file %s\n",
1144                                                                 f->file_name);
1145                         }
1146                 }
1147         } else
1148                 cleanup_pending_aio(td);
1149
1150         /*
1151          * stop job if we failed doing any IO
1152          */
1153         if (!ddir_rw_sum(td->this_io_bytes))
1154                 td->done = 1;
1155
1156         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1157                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1158 }
1159
1160 static void free_file_completion_logging(struct thread_data *td)
1161 {
1162         struct fio_file *f;
1163         unsigned int i;
1164
1165         for_each_file(td, f, i) {
1166                 if (!f->last_write_comp)
1167                         break;
1168                 sfree(f->last_write_comp);
1169         }
1170 }
1171
1172 static int init_file_completion_logging(struct thread_data *td,
1173                                         unsigned int depth)
1174 {
1175         struct fio_file *f;
1176         unsigned int i;
1177
1178         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1179                 return 0;
1180
1181         for_each_file(td, f, i) {
1182                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1183                 if (!f->last_write_comp)
1184                         goto cleanup;
1185         }
1186
1187         return 0;
1188
1189 cleanup:
1190         free_file_completion_logging(td);
1191         log_err("fio: failed to alloc write comp data\n");
1192         return 1;
1193 }
1194
1195 static void cleanup_io_u(struct thread_data *td)
1196 {
1197         struct io_u *io_u;
1198
1199         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1200
1201                 if (td->io_ops->io_u_free)
1202                         td->io_ops->io_u_free(td, io_u);
1203
1204                 fio_memfree(io_u, sizeof(*io_u));
1205         }
1206
1207         free_io_mem(td);
1208
1209         io_u_rexit(&td->io_u_requeues);
1210         io_u_qexit(&td->io_u_freelist);
1211         io_u_qexit(&td->io_u_all);
1212
1213         free_file_completion_logging(td);
1214 }
1215
1216 static int init_io_u(struct thread_data *td)
1217 {
1218         struct io_u *io_u;
1219         unsigned int max_bs, min_write;
1220         int cl_align, i, max_units;
1221         int data_xfer = 1, err;
1222         char *p;
1223
1224         max_units = td->o.iodepth;
1225         max_bs = td_max_bs(td);
1226         min_write = td->o.min_bs[DDIR_WRITE];
1227         td->orig_buffer_size = (unsigned long long) max_bs
1228                                         * (unsigned long long) max_units;
1229
1230         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1231                 data_xfer = 0;
1232
1233         err = 0;
1234         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1235         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1236         err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1237
1238         if (err) {
1239                 log_err("fio: failed setting up IO queues\n");
1240                 return 1;
1241         }
1242
1243         /*
1244          * if we may later need to do address alignment, then add any
1245          * possible adjustment here so that we don't cause a buffer
1246          * overflow later. this adjustment may be too much if we get
1247          * lucky and the allocator gives us an aligned address.
1248          */
1249         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1250             td_ioengine_flagged(td, FIO_RAWIO))
1251                 td->orig_buffer_size += page_mask + td->o.mem_align;
1252
1253         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1254                 unsigned long bs;
1255
1256                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1257                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1258         }
1259
1260         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1261                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1262                 return 1;
1263         }
1264
1265         if (data_xfer && allocate_io_mem(td))
1266                 return 1;
1267
1268         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1269             td_ioengine_flagged(td, FIO_RAWIO))
1270                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1271         else
1272                 p = td->orig_buffer;
1273
1274         cl_align = os_cache_line_size();
1275
1276         for (i = 0; i < max_units; i++) {
1277                 void *ptr;
1278
1279                 if (td->terminate)
1280                         return 1;
1281
1282                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1283                 if (!ptr) {
1284                         log_err("fio: unable to allocate aligned memory\n");
1285                         break;
1286                 }
1287
1288                 io_u = ptr;
1289                 memset(io_u, 0, sizeof(*io_u));
1290                 INIT_FLIST_HEAD(&io_u->verify_list);
1291                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1292
1293                 if (data_xfer) {
1294                         io_u->buf = p;
1295                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1296
1297                         if (td_write(td))
1298                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1299                         if (td_write(td) && td->o.verify_pattern_bytes) {
1300                                 /*
1301                                  * Fill the buffer with the pattern if we are
1302                                  * going to be doing writes.
1303                                  */
1304                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1305                         }
1306                 }
1307
1308                 io_u->index = i;
1309                 io_u->flags = IO_U_F_FREE;
1310                 io_u_qpush(&td->io_u_freelist, io_u);
1311
1312                 /*
1313                  * io_u never leaves this stack, used for iteration of all
1314                  * io_u buffers.
1315                  */
1316                 io_u_qpush(&td->io_u_all, io_u);
1317
1318                 if (td->io_ops->io_u_init) {
1319                         int ret = td->io_ops->io_u_init(td, io_u);
1320
1321                         if (ret) {
1322                                 log_err("fio: failed to init engine data: %d\n", ret);
1323                                 return 1;
1324                         }
1325                 }
1326
1327                 p += max_bs;
1328         }
1329
1330         if (init_file_completion_logging(td, max_units))
1331                 return 1;
1332
1333         return 0;
1334 }
1335
1336 /*
1337  * This function is Linux specific.
1338  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1339  */
1340 static int switch_ioscheduler(struct thread_data *td)
1341 {
1342 #ifdef FIO_HAVE_IOSCHED_SWITCH
1343         char tmp[256], tmp2[128];
1344         FILE *f;
1345         int ret;
1346
1347         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1348                 return 0;
1349
1350         assert(td->files && td->files[0]);
1351         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1352
1353         f = fopen(tmp, "r+");
1354         if (!f) {
1355                 if (errno == ENOENT) {
1356                         log_err("fio: os or kernel doesn't support IO scheduler"
1357                                 " switching\n");
1358                         return 0;
1359                 }
1360                 td_verror(td, errno, "fopen iosched");
1361                 return 1;
1362         }
1363
1364         /*
1365          * Set io scheduler.
1366          */
1367         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1368         if (ferror(f) || ret != 1) {
1369                 td_verror(td, errno, "fwrite");
1370                 fclose(f);
1371                 return 1;
1372         }
1373
1374         rewind(f);
1375
1376         /*
1377          * Read back and check that the selected scheduler is now the default.
1378          */
1379         memset(tmp, 0, sizeof(tmp));
1380         ret = fread(tmp, sizeof(tmp), 1, f);
1381         if (ferror(f) || ret < 0) {
1382                 td_verror(td, errno, "fread");
1383                 fclose(f);
1384                 return 1;
1385         }
1386         /*
1387          * either a list of io schedulers or "none\n" is expected.
1388          */
1389         tmp[strlen(tmp) - 1] = '\0';
1390
1391         /*
1392          * Write to "none" entry doesn't fail, so check the result here.
1393          */
1394         if (!strcmp(tmp, "none")) {
1395                 log_err("fio: io scheduler is not tunable\n");
1396                 fclose(f);
1397                 return 0;
1398         }
1399
1400         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1401         if (!strstr(tmp, tmp2)) {
1402                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1403                 td_verror(td, EINVAL, "iosched_switch");
1404                 fclose(f);
1405                 return 1;
1406         }
1407
1408         fclose(f);
1409         return 0;
1410 #else
1411         return 0;
1412 #endif
1413 }
1414
1415 static bool keep_running(struct thread_data *td)
1416 {
1417         unsigned long long limit;
1418
1419         if (td->done)
1420                 return false;
1421         if (td->terminate)
1422                 return false;
1423         if (td->o.time_based)
1424                 return true;
1425         if (td->o.loops) {
1426                 td->o.loops--;
1427                 return true;
1428         }
1429         if (exceeds_number_ios(td))
1430                 return false;
1431
1432         if (td->o.io_size)
1433                 limit = td->o.io_size;
1434         else
1435                 limit = td->o.size;
1436
1437         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1438                 uint64_t diff;
1439
1440                 /*
1441                  * If the difference is less than the maximum IO size, we
1442                  * are done.
1443                  */
1444                 diff = limit - ddir_rw_sum(td->io_bytes);
1445                 if (diff < td_max_bs(td))
1446                         return false;
1447
1448                 if (fio_files_done(td) && !td->o.io_size)
1449                         return false;
1450
1451                 return true;
1452         }
1453
1454         return false;
1455 }
1456
1457 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1458 {
1459         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1460         int ret;
1461         char *str;
1462
1463         str = malloc(newlen);
1464         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1465
1466         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1467         ret = system(str);
1468         if (ret == -1)
1469                 log_err("fio: exec of cmd <%s> failed\n", str);
1470
1471         free(str);
1472         return ret;
1473 }
1474
1475 /*
1476  * Dry run to compute correct state of numberio for verification.
1477  */
1478 static uint64_t do_dry_run(struct thread_data *td)
1479 {
1480         td_set_runstate(td, TD_RUNNING);
1481
1482         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1483                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1484                 struct io_u *io_u;
1485                 int ret;
1486
1487                 if (td->terminate || td->done)
1488                         break;
1489
1490                 io_u = get_io_u(td);
1491                 if (IS_ERR_OR_NULL(io_u))
1492                         break;
1493
1494                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1495                 io_u->error = 0;
1496                 io_u->resid = 0;
1497                 if (ddir_rw(acct_ddir(io_u)))
1498                         td->io_issues[acct_ddir(io_u)]++;
1499                 if (ddir_rw(io_u->ddir)) {
1500                         io_u_mark_depth(td, 1);
1501                         td->ts.total_io_u[io_u->ddir]++;
1502                 }
1503
1504                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1505                     td->o.do_verify &&
1506                     td->o.verify != VERIFY_NONE &&
1507                     !td->o.experimental_verify)
1508                         log_io_piece(td, io_u);
1509
1510                 ret = io_u_sync_complete(td, io_u);
1511                 (void) ret;
1512         }
1513
1514         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1515 }
1516
1517 struct fork_data {
1518         struct thread_data *td;
1519         struct sk_out *sk_out;
1520 };
1521
1522 /*
1523  * Entry point for the thread based jobs. The process based jobs end up
1524  * here as well, after a little setup.
1525  */
1526 static void *thread_main(void *data)
1527 {
1528         struct fork_data *fd = data;
1529         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1530         struct thread_data *td = fd->td;
1531         struct thread_options *o = &td->o;
1532         struct sk_out *sk_out = fd->sk_out;
1533         uint64_t bytes_done[DDIR_RWDIR_CNT];
1534         int deadlock_loop_cnt;
1535         bool clear_state, did_some_io;
1536         int ret;
1537
1538         sk_out_assign(sk_out);
1539         free(fd);
1540
1541         if (!o->use_thread) {
1542                 setsid();
1543                 td->pid = getpid();
1544         } else
1545                 td->pid = gettid();
1546
1547         fio_local_clock_init(o->use_thread);
1548
1549         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1550
1551         if (is_backend)
1552                 fio_server_send_start(td);
1553
1554         INIT_FLIST_HEAD(&td->io_log_list);
1555         INIT_FLIST_HEAD(&td->io_hist_list);
1556         INIT_FLIST_HEAD(&td->verify_list);
1557         INIT_FLIST_HEAD(&td->trim_list);
1558         INIT_FLIST_HEAD(&td->next_rand_list);
1559         td->io_hist_tree = RB_ROOT;
1560
1561         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1562         if (ret) {
1563                 td_verror(td, ret, "mutex_cond_init_pshared");
1564                 goto err;
1565         }
1566         ret = cond_init_pshared(&td->verify_cond);
1567         if (ret) {
1568                 td_verror(td, ret, "mutex_cond_pshared");
1569                 goto err;
1570         }
1571
1572         td_set_runstate(td, TD_INITIALIZED);
1573         dprint(FD_MUTEX, "up startup_sem\n");
1574         fio_sem_up(startup_sem);
1575         dprint(FD_MUTEX, "wait on td->sem\n");
1576         fio_sem_down(td->sem);
1577         dprint(FD_MUTEX, "done waiting on td->sem\n");
1578
1579         /*
1580          * A new gid requires privilege, so we need to do this before setting
1581          * the uid.
1582          */
1583         if (o->gid != -1U && setgid(o->gid)) {
1584                 td_verror(td, errno, "setgid");
1585                 goto err;
1586         }
1587         if (o->uid != -1U && setuid(o->uid)) {
1588                 td_verror(td, errno, "setuid");
1589                 goto err;
1590         }
1591
1592         /*
1593          * Do this early, we don't want the compress threads to be limited
1594          * to the same CPUs as the IO workers. So do this before we set
1595          * any potential CPU affinity
1596          */
1597         if (iolog_compress_init(td, sk_out))
1598                 goto err;
1599
1600         /*
1601          * If we have a gettimeofday() thread, make sure we exclude that
1602          * thread from this job
1603          */
1604         if (o->gtod_cpu)
1605                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1606
1607         /*
1608          * Set affinity first, in case it has an impact on the memory
1609          * allocations.
1610          */
1611         if (fio_option_is_set(o, cpumask)) {
1612                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1613                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1614                         if (!ret) {
1615                                 log_err("fio: no CPUs set\n");
1616                                 log_err("fio: Try increasing number of available CPUs\n");
1617                                 td_verror(td, EINVAL, "cpus_split");
1618                                 goto err;
1619                         }
1620                 }
1621                 ret = fio_setaffinity(td->pid, o->cpumask);
1622                 if (ret == -1) {
1623                         td_verror(td, errno, "cpu_set_affinity");
1624                         goto err;
1625                 }
1626         }
1627
1628 #ifdef CONFIG_LIBNUMA
1629         /* numa node setup */
1630         if (fio_option_is_set(o, numa_cpunodes) ||
1631             fio_option_is_set(o, numa_memnodes)) {
1632                 struct bitmask *mask;
1633
1634                 if (numa_available() < 0) {
1635                         td_verror(td, errno, "Does not support NUMA API\n");
1636                         goto err;
1637                 }
1638
1639                 if (fio_option_is_set(o, numa_cpunodes)) {
1640                         mask = numa_parse_nodestring(o->numa_cpunodes);
1641                         ret = numa_run_on_node_mask(mask);
1642                         numa_free_nodemask(mask);
1643                         if (ret == -1) {
1644                                 td_verror(td, errno, \
1645                                         "numa_run_on_node_mask failed\n");
1646                                 goto err;
1647                         }
1648                 }
1649
1650                 if (fio_option_is_set(o, numa_memnodes)) {
1651                         mask = NULL;
1652                         if (o->numa_memnodes)
1653                                 mask = numa_parse_nodestring(o->numa_memnodes);
1654
1655                         switch (o->numa_mem_mode) {
1656                         case MPOL_INTERLEAVE:
1657                                 numa_set_interleave_mask(mask);
1658                                 break;
1659                         case MPOL_BIND:
1660                                 numa_set_membind(mask);
1661                                 break;
1662                         case MPOL_LOCAL:
1663                                 numa_set_localalloc();
1664                                 break;
1665                         case MPOL_PREFERRED:
1666                                 numa_set_preferred(o->numa_mem_prefer_node);
1667                                 break;
1668                         case MPOL_DEFAULT:
1669                         default:
1670                                 break;
1671                         }
1672
1673                         if (mask)
1674                                 numa_free_nodemask(mask);
1675
1676                 }
1677         }
1678 #endif
1679
1680         if (fio_pin_memory(td))
1681                 goto err;
1682
1683         /*
1684          * May alter parameters that init_io_u() will use, so we need to
1685          * do this first.
1686          */
1687         if (init_iolog(td))
1688                 goto err;
1689
1690         if (init_io_u(td))
1691                 goto err;
1692
1693         if (o->verify_async && verify_async_init(td))
1694                 goto err;
1695
1696         if (fio_option_is_set(o, ioprio) ||
1697             fio_option_is_set(o, ioprio_class)) {
1698                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1699                 if (ret == -1) {
1700                         td_verror(td, errno, "ioprio_set");
1701                         goto err;
1702                 }
1703         }
1704
1705         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1706                 goto err;
1707
1708         errno = 0;
1709         if (nice(o->nice) == -1 && errno != 0) {
1710                 td_verror(td, errno, "nice");
1711                 goto err;
1712         }
1713
1714         if (o->ioscheduler && switch_ioscheduler(td))
1715                 goto err;
1716
1717         if (!o->create_serialize && setup_files(td))
1718                 goto err;
1719
1720         if (td_io_init(td))
1721                 goto err;
1722
1723         if (!init_random_map(td))
1724                 goto err;
1725
1726         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1727                 goto err;
1728
1729         if (o->pre_read && !pre_read_files(td))
1730                 goto err;
1731
1732         fio_verify_init(td);
1733
1734         if (rate_submit_init(td, sk_out))
1735                 goto err;
1736
1737         set_epoch_time(td, o->log_unix_epoch);
1738         fio_getrusage(&td->ru_start);
1739         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1740         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1741         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1742
1743         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1744                         o->ratemin[DDIR_TRIM]) {
1745                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1746                                         sizeof(td->bw_sample_time));
1747                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1748                                         sizeof(td->bw_sample_time));
1749                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1750                                         sizeof(td->bw_sample_time));
1751         }
1752
1753         memset(bytes_done, 0, sizeof(bytes_done));
1754         clear_state = false;
1755         did_some_io = false;
1756
1757         while (keep_running(td)) {
1758                 uint64_t verify_bytes;
1759
1760                 fio_gettime(&td->start, NULL);
1761                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1762
1763                 if (clear_state) {
1764                         clear_io_state(td, 0);
1765
1766                         if (o->unlink_each_loop && unlink_all_files(td))
1767                                 break;
1768                 }
1769
1770                 prune_io_piece_log(td);
1771
1772                 if (td->o.verify_only && td_write(td))
1773                         verify_bytes = do_dry_run(td);
1774                 else {
1775                         do_io(td, bytes_done);
1776
1777                         if (!ddir_rw_sum(bytes_done)) {
1778                                 fio_mark_td_terminate(td);
1779                                 verify_bytes = 0;
1780                         } else {
1781                                 verify_bytes = bytes_done[DDIR_WRITE] +
1782                                                 bytes_done[DDIR_TRIM];
1783                         }
1784                 }
1785
1786                 /*
1787                  * If we took too long to shut down, the main thread could
1788                  * already consider us reaped/exited. If that happens, break
1789                  * out and clean up.
1790                  */
1791                 if (td->runstate >= TD_EXITED)
1792                         break;
1793
1794                 clear_state = true;
1795
1796                 /*
1797                  * Make sure we've successfully updated the rusage stats
1798                  * before waiting on the stat mutex. Otherwise we could have
1799                  * the stat thread holding stat mutex and waiting for
1800                  * the rusage_sem, which would never get upped because
1801                  * this thread is waiting for the stat mutex.
1802                  */
1803                 deadlock_loop_cnt = 0;
1804                 do {
1805                         check_update_rusage(td);
1806                         if (!fio_sem_down_trylock(stat_sem))
1807                                 break;
1808                         usleep(1000);
1809                         if (deadlock_loop_cnt++ > 5000) {
1810                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1811                                 td->error = EDEADLK;
1812                                 goto err;
1813                         }
1814                 } while (1);
1815
1816                 if (td_read(td) && td->io_bytes[DDIR_READ])
1817                         update_runtime(td, elapsed_us, DDIR_READ);
1818                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1819                         update_runtime(td, elapsed_us, DDIR_WRITE);
1820                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1821                         update_runtime(td, elapsed_us, DDIR_TRIM);
1822                 fio_gettime(&td->start, NULL);
1823                 fio_sem_up(stat_sem);
1824
1825                 if (td->error || td->terminate)
1826                         break;
1827
1828                 if (!o->do_verify ||
1829                     o->verify == VERIFY_NONE ||
1830                     td_ioengine_flagged(td, FIO_UNIDIR))
1831                         continue;
1832
1833                 if (ddir_rw_sum(bytes_done))
1834                         did_some_io = true;
1835
1836                 clear_io_state(td, 0);
1837
1838                 fio_gettime(&td->start, NULL);
1839
1840                 do_verify(td, verify_bytes);
1841
1842                 /*
1843                  * See comment further up for why this is done here.
1844                  */
1845                 check_update_rusage(td);
1846
1847                 fio_sem_down(stat_sem);
1848                 update_runtime(td, elapsed_us, DDIR_READ);
1849                 fio_gettime(&td->start, NULL);
1850                 fio_sem_up(stat_sem);
1851
1852                 if (td->error || td->terminate)
1853                         break;
1854         }
1855
1856         /*
1857          * If td ended up with no I/O when it should have had,
1858          * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1859          * (Are we not missing other flags that can be ignored ?)
1860          */
1861         if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1862             !did_some_io && !td->o.create_only &&
1863             !(td_ioengine_flagged(td, FIO_NOIO) ||
1864               td_ioengine_flagged(td, FIO_DISKLESSIO)))
1865                 log_err("%s: No I/O performed by %s, "
1866                          "perhaps try --debug=io option for details?\n",
1867                          td->o.name, td->io_ops->name);
1868
1869         td_set_runstate(td, TD_FINISHING);
1870
1871         update_rusage_stat(td);
1872         td->ts.total_run_time = mtime_since_now(&td->epoch);
1873         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1874         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1875         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1876
1877         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1878             (td->o.verify != VERIFY_NONE && td_write(td)))
1879                 verify_save_state(td->thread_number);
1880
1881         fio_unpin_memory(td);
1882
1883         td_writeout_logs(td, true);
1884
1885         iolog_compress_exit(td);
1886         rate_submit_exit(td);
1887
1888         if (o->exec_postrun)
1889                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1890
1891         if (exitall_on_terminate || (o->exitall_error && td->error))
1892                 fio_terminate_threads(td->groupid);
1893
1894 err:
1895         if (td->error)
1896                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1897                                                         td->verror);
1898
1899         if (o->verify_async)
1900                 verify_async_exit(td);
1901
1902         close_and_free_files(td);
1903         cleanup_io_u(td);
1904         close_ioengine(td);
1905         cgroup_shutdown(td, &cgroup_mnt);
1906         verify_free_state(td);
1907
1908         if (td->zone_state_index) {
1909                 int i;
1910
1911                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1912                         free(td->zone_state_index[i]);
1913                 free(td->zone_state_index);
1914                 td->zone_state_index = NULL;
1915         }
1916
1917         if (fio_option_is_set(o, cpumask)) {
1918                 ret = fio_cpuset_exit(&o->cpumask);
1919                 if (ret)
1920                         td_verror(td, ret, "fio_cpuset_exit");
1921         }
1922
1923         /*
1924          * do this very late, it will log file closing as well
1925          */
1926         if (o->write_iolog_file)
1927                 write_iolog_close(td);
1928
1929         td_set_runstate(td, TD_EXITED);
1930
1931         /*
1932          * Do this last after setting our runstate to exited, so we
1933          * know that the stat thread is signaled.
1934          */
1935         check_update_rusage(td);
1936
1937         sk_out_drop();
1938         return (void *) (uintptr_t) td->error;
1939 }
1940
1941 /*
1942  * Run over the job map and reap the threads that have exited, if any.
1943  */
1944 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1945                          uint64_t *m_rate)
1946 {
1947         struct thread_data *td;
1948         unsigned int cputhreads, realthreads, pending;
1949         int i, status, ret;
1950
1951         /*
1952          * reap exited threads (TD_EXITED -> TD_REAPED)
1953          */
1954         realthreads = pending = cputhreads = 0;
1955         for_each_td(td, i) {
1956                 int flags = 0;
1957
1958                  if (!strcmp(td->o.ioengine, "cpuio"))
1959                         cputhreads++;
1960                 else
1961                         realthreads++;
1962
1963                 if (!td->pid) {
1964                         pending++;
1965                         continue;
1966                 }
1967                 if (td->runstate == TD_REAPED)
1968                         continue;
1969                 if (td->o.use_thread) {
1970                         if (td->runstate == TD_EXITED) {
1971                                 td_set_runstate(td, TD_REAPED);
1972                                 goto reaped;
1973                         }
1974                         continue;
1975                 }
1976
1977                 flags = WNOHANG;
1978                 if (td->runstate == TD_EXITED)
1979                         flags = 0;
1980
1981                 /*
1982                  * check if someone quit or got killed in an unusual way
1983                  */
1984                 ret = waitpid(td->pid, &status, flags);
1985                 if (ret < 0) {
1986                         if (errno == ECHILD) {
1987                                 log_err("fio: pid=%d disappeared %d\n",
1988                                                 (int) td->pid, td->runstate);
1989                                 td->sig = ECHILD;
1990                                 td_set_runstate(td, TD_REAPED);
1991                                 goto reaped;
1992                         }
1993                         perror("waitpid");
1994                 } else if (ret == td->pid) {
1995                         if (WIFSIGNALED(status)) {
1996                                 int sig = WTERMSIG(status);
1997
1998                                 if (sig != SIGTERM && sig != SIGUSR2)
1999                                         log_err("fio: pid=%d, got signal=%d\n",
2000                                                         (int) td->pid, sig);
2001                                 td->sig = sig;
2002                                 td_set_runstate(td, TD_REAPED);
2003                                 goto reaped;
2004                         }
2005                         if (WIFEXITED(status)) {
2006                                 if (WEXITSTATUS(status) && !td->error)
2007                                         td->error = WEXITSTATUS(status);
2008
2009                                 td_set_runstate(td, TD_REAPED);
2010                                 goto reaped;
2011                         }
2012                 }
2013
2014                 /*
2015                  * If the job is stuck, do a forceful timeout of it and
2016                  * move on.
2017                  */
2018                 if (td->terminate &&
2019                     td->runstate < TD_FSYNCING &&
2020                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2021                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2022                                 "%lu seconds, it appears to be stuck. Doing "
2023                                 "forceful exit of this job.\n",
2024                                 td->o.name, td->runstate,
2025                                 (unsigned long) time_since_now(&td->terminate_time));
2026                         td_set_runstate(td, TD_REAPED);
2027                         goto reaped;
2028                 }
2029
2030                 /*
2031                  * thread is not dead, continue
2032                  */
2033                 pending++;
2034                 continue;
2035 reaped:
2036                 (*nr_running)--;
2037                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2038                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2039                 if (!td->pid)
2040                         pending--;
2041
2042                 if (td->error)
2043                         exit_value++;
2044
2045                 done_secs += mtime_since_now(&td->epoch) / 1000;
2046                 profile_td_exit(td);
2047         }
2048
2049         if (*nr_running == cputhreads && !pending && realthreads)
2050                 fio_terminate_threads(TERMINATE_ALL);
2051 }
2052
2053 static bool __check_trigger_file(void)
2054 {
2055         struct stat sb;
2056
2057         if (!trigger_file)
2058                 return false;
2059
2060         if (stat(trigger_file, &sb))
2061                 return false;
2062
2063         if (unlink(trigger_file) < 0)
2064                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2065                                                         strerror(errno));
2066
2067         return true;
2068 }
2069
2070 static bool trigger_timedout(void)
2071 {
2072         if (trigger_timeout)
2073                 if (time_since_genesis() >= trigger_timeout) {
2074                         trigger_timeout = 0;
2075                         return true;
2076                 }
2077
2078         return false;
2079 }
2080
2081 void exec_trigger(const char *cmd)
2082 {
2083         int ret;
2084
2085         if (!cmd || cmd[0] == '\0')
2086                 return;
2087
2088         ret = system(cmd);
2089         if (ret == -1)
2090                 log_err("fio: failed executing %s trigger\n", cmd);
2091 }
2092
2093 void check_trigger_file(void)
2094 {
2095         if (__check_trigger_file() || trigger_timedout()) {
2096                 if (nr_clients)
2097                         fio_clients_send_trigger(trigger_remote_cmd);
2098                 else {
2099                         verify_save_state(IO_LIST_ALL);
2100                         fio_terminate_threads(TERMINATE_ALL);
2101                         exec_trigger(trigger_cmd);
2102                 }
2103         }
2104 }
2105
2106 static int fio_verify_load_state(struct thread_data *td)
2107 {
2108         int ret;
2109
2110         if (!td->o.verify_state)
2111                 return 0;
2112
2113         if (is_backend) {
2114                 void *data;
2115
2116                 ret = fio_server_get_verify_state(td->o.name,
2117                                         td->thread_number - 1, &data);
2118                 if (!ret)
2119                         verify_assign_state(td, data);
2120         } else
2121                 ret = verify_load_state(td, "local");
2122
2123         return ret;
2124 }
2125
2126 static void do_usleep(unsigned int usecs)
2127 {
2128         check_for_running_stats();
2129         check_trigger_file();
2130         usleep(usecs);
2131 }
2132
2133 static bool check_mount_writes(struct thread_data *td)
2134 {
2135         struct fio_file *f;
2136         unsigned int i;
2137
2138         if (!td_write(td) || td->o.allow_mounted_write)
2139                 return false;
2140
2141         /*
2142          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2143          * are mkfs'd and mounted.
2144          */
2145         for_each_file(td, f, i) {
2146 #ifdef FIO_HAVE_CHARDEV_SIZE
2147                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2148 #else
2149                 if (f->filetype != FIO_TYPE_BLOCK)
2150 #endif
2151                         continue;
2152                 if (device_is_mounted(f->file_name))
2153                         goto mounted;
2154         }
2155
2156         return false;
2157 mounted:
2158         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2159         return true;
2160 }
2161
2162 static bool waitee_running(struct thread_data *me)
2163 {
2164         const char *waitee = me->o.wait_for;
2165         const char *self = me->o.name;
2166         struct thread_data *td;
2167         int i;
2168
2169         if (!waitee)
2170                 return false;
2171
2172         for_each_td(td, i) {
2173                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2174                         continue;
2175
2176                 if (td->runstate < TD_EXITED) {
2177                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2178                                         self, td->o.name,
2179                                         runstate_to_name(td->runstate));
2180                         return true;
2181                 }
2182         }
2183
2184         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2185         return false;
2186 }
2187
2188 /*
2189  * Main function for kicking off and reaping jobs, as needed.
2190  */
2191 static void run_threads(struct sk_out *sk_out)
2192 {
2193         struct thread_data *td;
2194         unsigned int i, todo, nr_running, nr_started;
2195         uint64_t m_rate, t_rate;
2196         uint64_t spent;
2197
2198         if (fio_gtod_offload && fio_start_gtod_thread())
2199                 return;
2200
2201         fio_idle_prof_init();
2202
2203         set_sig_handlers();
2204
2205         nr_thread = nr_process = 0;
2206         for_each_td(td, i) {
2207                 if (check_mount_writes(td))
2208                         return;
2209                 if (td->o.use_thread)
2210                         nr_thread++;
2211                 else
2212                         nr_process++;
2213         }
2214
2215         if (output_format & FIO_OUTPUT_NORMAL) {
2216                 log_info("Starting ");
2217                 if (nr_thread)
2218                         log_info("%d thread%s", nr_thread,
2219                                                 nr_thread > 1 ? "s" : "");
2220                 if (nr_process) {
2221                         if (nr_thread)
2222                                 log_info(" and ");
2223                         log_info("%d process%s", nr_process,
2224                                                 nr_process > 1 ? "es" : "");
2225                 }
2226                 log_info("\n");
2227                 log_info_flush();
2228         }
2229
2230         todo = thread_number;
2231         nr_running = 0;
2232         nr_started = 0;
2233         m_rate = t_rate = 0;
2234
2235         for_each_td(td, i) {
2236                 print_status_init(td->thread_number - 1);
2237
2238                 if (!td->o.create_serialize)
2239                         continue;
2240
2241                 if (fio_verify_load_state(td))
2242                         goto reap;
2243
2244                 /*
2245                  * do file setup here so it happens sequentially,
2246                  * we don't want X number of threads getting their
2247                  * client data interspersed on disk
2248                  */
2249                 if (setup_files(td)) {
2250 reap:
2251                         exit_value++;
2252                         if (td->error)
2253                                 log_err("fio: pid=%d, err=%d/%s\n",
2254                                         (int) td->pid, td->error, td->verror);
2255                         td_set_runstate(td, TD_REAPED);
2256                         todo--;
2257                 } else {
2258                         struct fio_file *f;
2259                         unsigned int j;
2260
2261                         /*
2262                          * for sharing to work, each job must always open
2263                          * its own files. so close them, if we opened them
2264                          * for creation
2265                          */
2266                         for_each_file(td, f, j) {
2267                                 if (fio_file_open(f))
2268                                         td_io_close_file(td, f);
2269                         }
2270                 }
2271         }
2272
2273         /* start idle threads before io threads start to run */
2274         fio_idle_prof_start();
2275
2276         set_genesis_time();
2277
2278         while (todo) {
2279                 struct thread_data *map[REAL_MAX_JOBS];
2280                 struct timespec this_start;
2281                 int this_jobs = 0, left;
2282                 struct fork_data *fd;
2283
2284                 /*
2285                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2286                  */
2287                 for_each_td(td, i) {
2288                         if (td->runstate != TD_NOT_CREATED)
2289                                 continue;
2290
2291                         /*
2292                          * never got a chance to start, killed by other
2293                          * thread for some reason
2294                          */
2295                         if (td->terminate) {
2296                                 todo--;
2297                                 continue;
2298                         }
2299
2300                         if (td->o.start_delay) {
2301                                 spent = utime_since_genesis();
2302
2303                                 if (td->o.start_delay > spent)
2304                                         continue;
2305                         }
2306
2307                         if (td->o.stonewall && (nr_started || nr_running)) {
2308                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2309                                                         td->o.name);
2310                                 break;
2311                         }
2312
2313                         if (waitee_running(td)) {
2314                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2315                                                 td->o.name, td->o.wait_for);
2316                                 continue;
2317                         }
2318
2319                         init_disk_util(td);
2320
2321                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2322                         td->update_rusage = 0;
2323
2324                         /*
2325                          * Set state to created. Thread will transition
2326                          * to TD_INITIALIZED when it's done setting up.
2327                          */
2328                         td_set_runstate(td, TD_CREATED);
2329                         map[this_jobs++] = td;
2330                         nr_started++;
2331
2332                         fd = calloc(1, sizeof(*fd));
2333                         fd->td = td;
2334                         fd->sk_out = sk_out;
2335
2336                         if (td->o.use_thread) {
2337                                 int ret;
2338
2339                                 dprint(FD_PROCESS, "will pthread_create\n");
2340                                 ret = pthread_create(&td->thread, NULL,
2341                                                         thread_main, fd);
2342                                 if (ret) {
2343                                         log_err("pthread_create: %s\n",
2344                                                         strerror(ret));
2345                                         free(fd);
2346                                         nr_started--;
2347                                         break;
2348                                 }
2349                                 fd = NULL;
2350                                 ret = pthread_detach(td->thread);
2351                                 if (ret)
2352                                         log_err("pthread_detach: %s",
2353                                                         strerror(ret));
2354                         } else {
2355                                 pid_t pid;
2356                                 dprint(FD_PROCESS, "will fork\n");
2357                                 pid = fork();
2358                                 if (!pid) {
2359                                         int ret;
2360
2361                                         ret = (int)(uintptr_t)thread_main(fd);
2362                                         _exit(ret);
2363                                 } else if (i == fio_debug_jobno)
2364                                         *fio_debug_jobp = pid;
2365                         }
2366                         dprint(FD_MUTEX, "wait on startup_sem\n");
2367                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2368                                 log_err("fio: job startup hung? exiting.\n");
2369                                 fio_terminate_threads(TERMINATE_ALL);
2370                                 fio_abort = 1;
2371                                 nr_started--;
2372                                 free(fd);
2373                                 break;
2374                         }
2375                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2376                 }
2377
2378                 /*
2379                  * Wait for the started threads to transition to
2380                  * TD_INITIALIZED.
2381                  */
2382                 fio_gettime(&this_start, NULL);
2383                 left = this_jobs;
2384                 while (left && !fio_abort) {
2385                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2386                                 break;
2387
2388                         do_usleep(100000);
2389
2390                         for (i = 0; i < this_jobs; i++) {
2391                                 td = map[i];
2392                                 if (!td)
2393                                         continue;
2394                                 if (td->runstate == TD_INITIALIZED) {
2395                                         map[i] = NULL;
2396                                         left--;
2397                                 } else if (td->runstate >= TD_EXITED) {
2398                                         map[i] = NULL;
2399                                         left--;
2400                                         todo--;
2401                                         nr_running++; /* work-around... */
2402                                 }
2403                         }
2404                 }
2405
2406                 if (left) {
2407                         log_err("fio: %d job%s failed to start\n", left,
2408                                         left > 1 ? "s" : "");
2409                         for (i = 0; i < this_jobs; i++) {
2410                                 td = map[i];
2411                                 if (!td)
2412                                         continue;
2413                                 kill(td->pid, SIGTERM);
2414                         }
2415                         break;
2416                 }
2417
2418                 /*
2419                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2420                  */
2421                 for_each_td(td, i) {
2422                         if (td->runstate != TD_INITIALIZED)
2423                                 continue;
2424
2425                         if (in_ramp_time(td))
2426                                 td_set_runstate(td, TD_RAMP);
2427                         else
2428                                 td_set_runstate(td, TD_RUNNING);
2429                         nr_running++;
2430                         nr_started--;
2431                         m_rate += ddir_rw_sum(td->o.ratemin);
2432                         t_rate += ddir_rw_sum(td->o.rate);
2433                         todo--;
2434                         fio_sem_up(td->sem);
2435                 }
2436
2437                 reap_threads(&nr_running, &t_rate, &m_rate);
2438
2439                 if (todo)
2440                         do_usleep(100000);
2441         }
2442
2443         while (nr_running) {
2444                 reap_threads(&nr_running, &t_rate, &m_rate);
2445                 do_usleep(10000);
2446         }
2447
2448         fio_idle_prof_stop();
2449
2450         update_io_ticks();
2451 }
2452
2453 static void free_disk_util(void)
2454 {
2455         disk_util_prune_entries();
2456         helper_thread_destroy();
2457 }
2458
2459 int fio_backend(struct sk_out *sk_out)
2460 {
2461         struct thread_data *td;
2462         int i;
2463
2464         if (exec_profile) {
2465                 if (load_profile(exec_profile))
2466                         return 1;
2467                 free(exec_profile);
2468                 exec_profile = NULL;
2469         }
2470         if (!thread_number)
2471                 return 0;
2472
2473         if (write_bw_log) {
2474                 struct log_params p = {
2475                         .log_type = IO_LOG_TYPE_BW,
2476                 };
2477
2478                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2479                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2480                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2481         }
2482
2483         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2484         if (startup_sem == NULL)
2485                 return 1;
2486
2487         set_genesis_time();
2488         stat_init();
2489         helper_thread_create(startup_sem, sk_out);
2490
2491         cgroup_list = smalloc(sizeof(*cgroup_list));
2492         INIT_FLIST_HEAD(cgroup_list);
2493
2494         run_threads(sk_out);
2495
2496         helper_thread_exit();
2497
2498         if (!fio_abort) {
2499                 __show_run_stats();
2500                 if (write_bw_log) {
2501                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2502                                 struct io_log *log = agg_io_log[i];
2503
2504                                 flush_log(log, false);
2505                                 free_log(log);
2506                         }
2507                 }
2508         }
2509
2510         for_each_td(td, i) {
2511                 steadystate_free(td);
2512                 fio_options_free(td);
2513                 if (td->rusage_sem) {
2514                         fio_sem_remove(td->rusage_sem);
2515                         td->rusage_sem = NULL;
2516                 }
2517                 fio_sem_remove(td->sem);
2518                 td->sem = NULL;
2519         }
2520
2521         free_disk_util();
2522         cgroup_kill(cgroup_list);
2523         sfree(cgroup_list);
2524         sfree(cgroup_mnt);
2525
2526         fio_sem_remove(startup_sem);
2527         stat_exit();
2528         return exit_value;
2529 }