backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32
  33 #include "fio.h"
  34 #include "smalloc.h"
  35 #include "verify.h"
  36 #include "diskutil.h"
  37 #include "cgroup.h"
  38 #include "profile.h"
  39 #include "lib/rand.h"
  40 #include "lib/memalign.h"
  41 #include "server.h"
  42 #include "lib/getrusage.h"
  43 #include "idletime.h"
  44 #include "err.h"
  45 #include "workqueue.h"
  46 #include "lib/mountcheck.h"
  47 #include "rate-submit.h"
  48 #include "helper_thread.h"
  49 #include "pshared.h"
  50 #include "zone-dist.h"
  51
  52 static struct fio_sem *startup_sem;
  53 static struct flist_head *cgroup_list;
  54 static struct cgroup_mnt *cgroup_mnt;
  55 static int exit_value;
  56 static volatile bool fio_abort;
  57 static unsigned int nr_process = 0;
  58 static unsigned int nr_thread = 0;
  59
  60 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  61
  62 int groupid = 0;
  63 unsigned int thread_number = 0;
  64 unsigned int stat_number = 0;
  65 int shm_id = 0;
  66 int temp_stall_ts;
  67 unsigned long done_secs = 0;
  68
  69 #define JOB_START_TIMEOUT       (5 * 1000)
  70
  71 static void sig_int(int sig)
  72 {
  73         if (threads) {
  74                 if (is_backend)
  75                         fio_server_got_signal(sig);
  76                 else {
  77                         log_info("\nfio: terminating on signal %d\n", sig);
  78                         log_info_flush();
  79                         exit_value = 128;
  80                 }
  81
  82                 fio_terminate_threads(TERMINATE_ALL);
  83         }
  84 }
  85
  86 void sig_show_status(int sig)
  87 {
  88         show_running_run_stats();
  89 }
  90
  91 static void set_sig_handlers(void)
  92 {
  93         struct sigaction act;
  94
  95         memset(&act, 0, sizeof(act));
  96         act.sa_handler = sig_int;
  97         act.sa_flags = SA_RESTART;
  98         sigaction(SIGINT, &act, NULL);
  99
 100         memset(&act, 0, sizeof(act));
 101         act.sa_handler = sig_int;
 102         act.sa_flags = SA_RESTART;
 103         sigaction(SIGTERM, &act, NULL);
 104
 105 /* Windows uses SIGBREAK as a quit signal from other applications */
 106 #ifdef WIN32
 107         memset(&act, 0, sizeof(act));
 108         act.sa_handler = sig_int;
 109         act.sa_flags = SA_RESTART;
 110         sigaction(SIGBREAK, &act, NULL);
 111 #endif
 112
 113         memset(&act, 0, sizeof(act));
 114         act.sa_handler = sig_show_status;
 115         act.sa_flags = SA_RESTART;
 116         sigaction(SIGUSR1, &act, NULL);
 117
 118         if (is_backend) {
 119                 memset(&act, 0, sizeof(act));
 120                 act.sa_handler = sig_int;
 121                 act.sa_flags = SA_RESTART;
 122                 sigaction(SIGPIPE, &act, NULL);
 123         }
 124 }
 125
 126 /*
 127  * Check if we are above the minimum rate given.
 128  */
 129 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 130                              enum fio_ddir ddir)
 131 {
 132         unsigned long long bytes = 0;
 133         unsigned long iops = 0;
 134         unsigned long spent;
 135         unsigned long rate;
 136         unsigned int ratemin = 0;
 137         unsigned int rate_iops = 0;
 138         unsigned int rate_iops_min = 0;
 139
 140         assert(ddir_rw(ddir));
 141
 142         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 143                 return false;
 144
 145         /*
 146          * allow a 2 second settle period in the beginning
 147          */
 148         if (mtime_since(&td->start, now) < 2000)
 149                 return false;
 150
 151         iops += td->this_io_blocks[ddir];
 152         bytes += td->this_io_bytes[ddir];
 153         ratemin += td->o.ratemin[ddir];
 154         rate_iops += td->o.rate_iops[ddir];
 155         rate_iops_min += td->o.rate_iops_min[ddir];
 156
 157         /*
 158          * if rate blocks is set, sample is running
 159          */
 160         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 161                 spent = mtime_since(&td->lastrate[ddir], now);
 162                 if (spent < td->o.ratecycle)
 163                         return false;
 164
 165                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 166                         /*
 167                          * check bandwidth specified rate
 168                          */
 169                         if (bytes < td->rate_bytes[ddir]) {
 170                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 171                                         td->o.name, ratemin, bytes);
 172                                 return true;
 173                         } else {
 174                                 if (spent)
 175                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 176                                 else
 177                                         rate = 0;
 178
 179                                 if (rate < ratemin ||
 180                                     bytes < td->rate_bytes[ddir]) {
 181                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 182                                                 td->o.name, ratemin, rate);
 183                                         return true;
 184                                 }
 185                         }
 186                 } else {
 187                         /*
 188                          * checks iops specified rate
 189                          */
 190                         if (iops < rate_iops) {
 191                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 192                                                 td->o.name, rate_iops, iops);
 193                                 return true;
 194                         } else {
 195                                 if (spent)
 196                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 197                                 else
 198                                         rate = 0;
 199
 200                                 if (rate < rate_iops_min ||
 201                                     iops < td->rate_blocks[ddir]) {
 202                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 203                                                 td->o.name, rate_iops_min, rate);
 204                                         return true;
 205                                 }
 206                         }
 207                 }
 208         }
 209
 210         td->rate_bytes[ddir] = bytes;
 211         td->rate_blocks[ddir] = iops;
 212         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 213         return false;
 214 }
 215
 216 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 217 {
 218         bool ret = false;
 219
 220         if (td->bytes_done[DDIR_READ])
 221                 ret |= __check_min_rate(td, now, DDIR_READ);
 222         if (td->bytes_done[DDIR_WRITE])
 223                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 224         if (td->bytes_done[DDIR_TRIM])
 225                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 226
 227         return ret;
 228 }
 229
 230 /*
 231  * When job exits, we can cancel the in-flight IO if we are using async
 232  * io. Attempt to do so.
 233  */
 234 static void cleanup_pending_aio(struct thread_data *td)
 235 {
 236         int r;
 237
 238         /*
 239          * get immediately available events, if any
 240          */
 241         r = io_u_queued_complete(td, 0);
 242         if (r < 0)
 243                 return;
 244
 245         /*
 246          * now cancel remaining active events
 247          */
 248         if (td->io_ops->cancel) {
 249                 struct io_u *io_u;
 250                 int i;
 251
 252                 io_u_qiter(&td->io_u_all, io_u, i) {
 253                         if (io_u->flags & IO_U_F_FLIGHT) {
 254                                 r = td->io_ops->cancel(td, io_u);
 255                                 if (!r)
 256                                         put_io_u(td, io_u);
 257                         }
 258                 }
 259         }
 260
 261         if (td->cur_depth)
 262                 r = io_u_queued_complete(td, td->cur_depth);
 263 }
 264
 265 /*
 266  * Helper to handle the final sync of a file. Works just like the normal
 267  * io path, just does everything sync.
 268  */
 269 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 270 {
 271         struct io_u *io_u = __get_io_u(td);
 272         enum fio_q_status ret;
 273
 274         if (!io_u)
 275                 return true;
 276
 277         io_u->ddir = DDIR_SYNC;
 278         io_u->file = f;
 279
 280         if (td_io_prep(td, io_u)) {
 281                 put_io_u(td, io_u);
 282                 return true;
 283         }
 284
 285 requeue:
 286         ret = td_io_queue(td, io_u);
 287         switch (ret) {
 288         case FIO_Q_QUEUED:
 289                 td_io_commit(td);
 290                 if (io_u_queued_complete(td, 1) < 0)
 291                         return true;
 292                 break;
 293         case FIO_Q_COMPLETED:
 294                 if (io_u->error) {
 295                         td_verror(td, io_u->error, "td_io_queue");
 296                         return true;
 297                 }
 298
 299                 if (io_u_sync_complete(td, io_u) < 0)
 300                         return true;
 301                 break;
 302         case FIO_Q_BUSY:
 303                 td_io_commit(td);
 304                 goto requeue;
 305         }
 306
 307         return false;
 308 }
 309
 310 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 311 {
 312         int ret;
 313
 314         if (fio_file_open(f))
 315                 return fio_io_sync(td, f);
 316
 317         if (td_io_open_file(td, f))
 318                 return 1;
 319
 320         ret = fio_io_sync(td, f);
 321         td_io_close_file(td, f);
 322         return ret;
 323 }
 324
 325 static inline void __update_ts_cache(struct thread_data *td)
 326 {
 327         fio_gettime(&td->ts_cache, NULL);
 328 }
 329
 330 static inline void update_ts_cache(struct thread_data *td)
 331 {
 332         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 333                 __update_ts_cache(td);
 334 }
 335
 336 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 337 {
 338         if (in_ramp_time(td))
 339                 return false;
 340         if (!td->o.timeout)
 341                 return false;
 342         if (utime_since(&td->epoch, t) >= td->o.timeout)
 343                 return true;
 344
 345         return false;
 346 }
 347
 348 /*
 349  * We need to update the runtime consistently in ms, but keep a running
 350  * tally of the current elapsed time in microseconds for sub millisecond
 351  * updates.
 352  */
 353 static inline void update_runtime(struct thread_data *td,
 354                                   unsigned long long *elapsed_us,
 355                                   const enum fio_ddir ddir)
 356 {
 357         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 358                 return;
 359
 360         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 361         elapsed_us[ddir] += utime_since_now(&td->start);
 362         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 363 }
 364
 365 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 366                                 int *retptr)
 367 {
 368         int ret = *retptr;
 369
 370         if (ret < 0 || td->error) {
 371                 int err = td->error;
 372                 enum error_type_bit eb;
 373
 374                 if (ret < 0)
 375                         err = -ret;
 376
 377                 eb = td_error_type(ddir, err);
 378                 if (!(td->o.continue_on_error & (1 << eb)))
 379                         return true;
 380
 381                 if (td_non_fatal_error(td, eb, err)) {
 382                         /*
 383                          * Continue with the I/Os in case of
 384                          * a non fatal error.
 385                          */
 386                         update_error_count(td, err);
 387                         td_clear_error(td);
 388                         *retptr = 0;
 389                         return false;
 390                 } else if (td->o.fill_device && err == ENOSPC) {
 391                         /*
 392                          * We expect to hit this error if
 393                          * fill_device option is set.
 394                          */
 395                         td_clear_error(td);
 396                         fio_mark_td_terminate(td);
 397                         return true;
 398                 } else {
 399                         /*
 400                          * Stop the I/O in case of a fatal
 401                          * error.
 402                          */
 403                         update_error_count(td, err);
 404                         return true;
 405                 }
 406         }
 407
 408         return false;
 409 }
 410
 411 static void check_update_rusage(struct thread_data *td)
 412 {
 413         if (td->update_rusage) {
 414                 td->update_rusage = 0;
 415                 update_rusage_stat(td);
 416                 fio_sem_up(td->rusage_sem);
 417         }
 418 }
 419
 420 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 421 {
 422         const int full = queue_full(td);
 423         int min_evts = 0;
 424         int ret;
 425
 426         if (td->flags & TD_F_REGROW_LOGS)
 427                 return io_u_quiesce(td);
 428
 429         /*
 430          * if the queue is full, we MUST reap at least 1 event
 431          */
 432         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 433         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 434                 min_evts = 1;
 435
 436         if (time && __should_check_rate(td))
 437                 fio_gettime(time, NULL);
 438
 439         do {
 440                 ret = io_u_queued_complete(td, min_evts);
 441                 if (ret < 0)
 442                         break;
 443         } while (full && (td->cur_depth > td->o.iodepth_low));
 444
 445         return ret;
 446 }
 447
 448 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 449                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 450                    struct timespec *comp_time)
 451 {
 452         switch (*ret) {
 453         case FIO_Q_COMPLETED:
 454                 if (io_u->error) {
 455                         *ret = -io_u->error;
 456                         clear_io_u(td, io_u);
 457                 } else if (io_u->resid) {
 458                         long long bytes = io_u->xfer_buflen - io_u->resid;
 459                         struct fio_file *f = io_u->file;
 460
 461                         if (bytes_issued)
 462                                 *bytes_issued += bytes;
 463
 464                         if (!from_verify)
 465                                 trim_io_piece(io_u);
 466
 467                         /*
 468                          * zero read, fail
 469                          */
 470                         if (!bytes) {
 471                                 if (!from_verify)
 472                                         unlog_io_piece(td, io_u);
 473                                 td_verror(td, EIO, "full resid");
 474                                 put_io_u(td, io_u);
 475                                 break;
 476                         }
 477
 478                         io_u->xfer_buflen = io_u->resid;
 479                         io_u->xfer_buf += bytes;
 480                         io_u->offset += bytes;
 481
 482                         if (ddir_rw(io_u->ddir))
 483                                 td->ts.short_io_u[io_u->ddir]++;
 484
 485                         if (io_u->offset == f->real_file_size)
 486                                 goto sync_done;
 487
 488                         requeue_io_u(td, &io_u);
 489                 } else {
 490 sync_done:
 491                         if (comp_time && __should_check_rate(td))
 492                                 fio_gettime(comp_time, NULL);
 493
 494                         *ret = io_u_sync_complete(td, io_u);
 495                         if (*ret < 0)
 496                                 break;
 497                 }
 498
 499                 if (td->flags & TD_F_REGROW_LOGS)
 500                         regrow_logs(td);
 501
 502                 /*
 503                  * when doing I/O (not when verifying),
 504                  * check for any errors that are to be ignored
 505                  */
 506                 if (!from_verify)
 507                         break;
 508
 509                 return 0;
 510         case FIO_Q_QUEUED:
 511                 /*
 512                  * if the engine doesn't have a commit hook,
 513                  * the io_u is really queued. if it does have such
 514                  * a hook, it has to call io_u_queued() itself.
 515                  */
 516                 if (td->io_ops->commit == NULL)
 517                         io_u_queued(td, io_u);
 518                 if (bytes_issued)
 519                         *bytes_issued += io_u->xfer_buflen;
 520                 break;
 521         case FIO_Q_BUSY:
 522                 if (!from_verify)
 523                         unlog_io_piece(td, io_u);
 524                 requeue_io_u(td, &io_u);
 525                 td_io_commit(td);
 526                 break;
 527         default:
 528                 assert(*ret < 0);
 529                 td_verror(td, -(*ret), "td_io_queue");
 530                 break;
 531         }
 532
 533         if (break_on_this_error(td, ddir, ret))
 534                 return 1;
 535
 536         return 0;
 537 }
 538
 539 static inline bool io_in_polling(struct thread_data *td)
 540 {
 541         return !td->o.iodepth_batch_complete_min &&
 542                    !td->o.iodepth_batch_complete_max;
 543 }
 544 /*
 545  * Unlinks files from thread data fio_file structure
 546  */
 547 static int unlink_all_files(struct thread_data *td)
 548 {
 549         struct fio_file *f;
 550         unsigned int i;
 551         int ret = 0;
 552
 553         for_each_file(td, f, i) {
 554                 if (f->filetype != FIO_TYPE_FILE)
 555                         continue;
 556                 ret = td_io_unlink_file(td, f);
 557                 if (ret)
 558                         break;
 559         }
 560
 561         if (ret)
 562                 td_verror(td, ret, "unlink_all_files");
 563
 564         return ret;
 565 }
 566
 567 /*
 568  * Check if io_u will overlap an in-flight IO in the queue
 569  */
 570 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 571 {
 572         bool overlap;
 573         struct io_u *check_io_u;
 574         unsigned long long x1, x2, y1, y2;
 575         int i;
 576
 577         x1 = io_u->offset;
 578         x2 = io_u->offset + io_u->buflen;
 579         overlap = false;
 580         io_u_qiter(q, check_io_u, i) {
 581                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 582                         y1 = check_io_u->offset;
 583                         y2 = check_io_u->offset + check_io_u->buflen;
 584
 585                         if (x1 < y2 && y1 < x2) {
 586                                 overlap = true;
 587                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 588                                                 x1, io_u->buflen,
 589                                                 y1, check_io_u->buflen);
 590                                 break;
 591                         }
 592                 }
 593         }
 594
 595         return overlap;
 596 }
 597
 598 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 599 {
 600         /*
 601          * Check for overlap if the user asked us to, and we have
 602          * at least one IO in flight besides this one.
 603          */
 604         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 605             in_flight_overlap(&td->io_u_all, io_u))
 606                 return FIO_Q_BUSY;
 607
 608         return td_io_queue(td, io_u);
 609 }
 610
 611 /*
 612  * The main verify engine. Runs over the writes we previously submitted,
 613  * reads the blocks back in, and checks the crc/md5 of the data.
 614  */
 615 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 616 {
 617         struct fio_file *f;
 618         struct io_u *io_u;
 619         int ret, min_events;
 620         unsigned int i;
 621
 622         dprint(FD_VERIFY, "starting loop\n");
 623
 624         /*
 625          * sync io first and invalidate cache, to make sure we really
 626          * read from disk.
 627          */
 628         for_each_file(td, f, i) {
 629                 if (!fio_file_open(f))
 630                         continue;
 631                 if (fio_io_sync(td, f))
 632                         break;
 633                 if (file_invalidate_cache(td, f))
 634                         break;
 635         }
 636
 637         check_update_rusage(td);
 638
 639         if (td->error)
 640                 return;
 641
 642         /*
 643          * verify_state needs to be reset before verification
 644          * proceeds so that expected random seeds match actual
 645          * random seeds in headers. The main loop will reset
 646          * all random number generators if randrepeat is set.
 647          */
 648         if (!td->o.rand_repeatable)
 649                 td_fill_verify_state_seed(td);
 650
 651         td_set_runstate(td, TD_VERIFYING);
 652
 653         io_u = NULL;
 654         while (!td->terminate) {
 655                 enum fio_ddir ddir;
 656                 int full;
 657
 658                 update_ts_cache(td);
 659                 check_update_rusage(td);
 660
 661                 if (runtime_exceeded(td, &td->ts_cache)) {
 662                         __update_ts_cache(td);
 663                         if (runtime_exceeded(td, &td->ts_cache)) {
 664                                 fio_mark_td_terminate(td);
 665                                 break;
 666                         }
 667                 }
 668
 669                 if (flow_threshold_exceeded(td))
 670                         continue;
 671
 672                 if (!td->o.experimental_verify) {
 673                         io_u = __get_io_u(td);
 674                         if (!io_u)
 675                                 break;
 676
 677                         if (get_next_verify(td, io_u)) {
 678                                 put_io_u(td, io_u);
 679                                 break;
 680                         }
 681
 682                         if (td_io_prep(td, io_u)) {
 683                                 put_io_u(td, io_u);
 684                                 break;
 685                         }
 686                 } else {
 687                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 688                                 break;
 689
 690                         while ((io_u = get_io_u(td)) != NULL) {
 691                                 if (IS_ERR_OR_NULL(io_u)) {
 692                                         io_u = NULL;
 693                                         ret = FIO_Q_BUSY;
 694                                         goto reap;
 695                                 }
 696
 697                                 /*
 698                                  * We are only interested in the places where
 699                                  * we wrote or trimmed IOs. Turn those into
 700                                  * reads for verification purposes.
 701                                  */
 702                                 if (io_u->ddir == DDIR_READ) {
 703                                         /*
 704                                          * Pretend we issued it for rwmix
 705                                          * accounting
 706                                          */
 707                                         td->io_issues[DDIR_READ]++;
 708                                         put_io_u(td, io_u);
 709                                         continue;
 710                                 } else if (io_u->ddir == DDIR_TRIM) {
 711                                         io_u->ddir = DDIR_READ;
 712                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 713                                         break;
 714                                 } else if (io_u->ddir == DDIR_WRITE) {
 715                                         io_u->ddir = DDIR_READ;
 716                                         populate_verify_io_u(td, io_u);
 717                                         break;
 718                                 } else {
 719                                         put_io_u(td, io_u);
 720                                         continue;
 721                                 }
 722                         }
 723
 724                         if (!io_u)
 725                                 break;
 726                 }
 727
 728                 if (verify_state_should_stop(td, io_u)) {
 729                         put_io_u(td, io_u);
 730                         break;
 731                 }
 732
 733                 if (td->o.verify_async)
 734                         io_u->end_io = verify_io_u_async;
 735                 else
 736                         io_u->end_io = verify_io_u;
 737
 738                 ddir = io_u->ddir;
 739                 if (!td->o.disable_slat)
 740                         fio_gettime(&io_u->start_time, NULL);
 741
 742                 ret = io_u_submit(td, io_u);
 743
 744                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 745                         break;
 746
 747                 /*
 748                  * if we can queue more, do so. but check if there are
 749                  * completed io_u's first. Note that we can get BUSY even
 750                  * without IO queued, if the system is resource starved.
 751                  */
 752 reap:
 753                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 754                 if (full || io_in_polling(td))
 755                         ret = wait_for_completions(td, NULL);
 756
 757                 if (ret < 0)
 758                         break;
 759         }
 760
 761         check_update_rusage(td);
 762
 763         if (!td->error) {
 764                 min_events = td->cur_depth;
 765
 766                 if (min_events)
 767                         ret = io_u_queued_complete(td, min_events);
 768         } else
 769                 cleanup_pending_aio(td);
 770
 771         td_set_runstate(td, TD_RUNNING);
 772
 773         dprint(FD_VERIFY, "exiting loop\n");
 774 }
 775
 776 static bool exceeds_number_ios(struct thread_data *td)
 777 {
 778         unsigned long long number_ios;
 779
 780         if (!td->o.number_ios)
 781                 return false;
 782
 783         number_ios = ddir_rw_sum(td->io_blocks);
 784         number_ios += td->io_u_queued + td->io_u_in_flight;
 785
 786         return number_ios >= (td->o.number_ios * td->loops);
 787 }
 788
 789 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 790 {
 791         unsigned long long bytes, limit;
 792
 793         if (td_rw(td))
 794                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 795         else if (td_write(td))
 796                 bytes = this_bytes[DDIR_WRITE];
 797         else if (td_read(td))
 798                 bytes = this_bytes[DDIR_READ];
 799         else
 800                 bytes = this_bytes[DDIR_TRIM];
 801
 802         if (td->o.io_size)
 803                 limit = td->o.io_size;
 804         else
 805                 limit = td->o.size;
 806
 807         limit *= td->loops;
 808         return bytes >= limit || exceeds_number_ios(td);
 809 }
 810
 811 static bool io_issue_bytes_exceeded(struct thread_data *td)
 812 {
 813         return io_bytes_exceeded(td, td->io_issue_bytes);
 814 }
 815
 816 static bool io_complete_bytes_exceeded(struct thread_data *td)
 817 {
 818         return io_bytes_exceeded(td, td->this_io_bytes);
 819 }
 820
 821 /*
 822  * used to calculate the next io time for rate control
 823  *
 824  */
 825 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 826 {
 827         uint64_t bps = td->rate_bps[ddir];
 828
 829         assert(!(td->flags & TD_F_CHILD));
 830
 831         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 832                 uint64_t val, iops;
 833
 834                 iops = bps / td->o.bs[ddir];
 835                 val = (int64_t) (1000000 / iops) *
 836                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 837                 if (val) {
 838                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 839                                         (unsigned long long) 1000000 / val,
 840                                         ddir);
 841                 }
 842                 td->last_usec[ddir] += val;
 843                 return td->last_usec[ddir];
 844         } else if (bps) {
 845                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 846                 uint64_t secs = bytes / bps;
 847                 uint64_t remainder = bytes % bps;
 848
 849                 return remainder * 1000000 / bps + secs * 1000000;
 850         }
 851
 852         return 0;
 853 }
 854
 855 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 856 {
 857         unsigned long long b;
 858         uint64_t total;
 859         int left;
 860
 861         b = ddir_rw_sum(td->io_blocks);
 862         if (b % td->o.thinktime_blocks)
 863                 return;
 864
 865         io_u_quiesce(td);
 866
 867         total = 0;
 868         if (td->o.thinktime_spin)
 869                 total = usec_spin(td->o.thinktime_spin);
 870
 871         left = td->o.thinktime - total;
 872         if (left)
 873                 total += usec_sleep(td, left);
 874
 875         /*
 876          * If we're ignoring thinktime for the rate, add the number of bytes
 877          * we would have done while sleeping, minus one block to ensure we
 878          * start issuing immediately after the sleep.
 879          */
 880         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 881                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 882                 uint64_t bs = td->o.min_bs[ddir];
 883                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 884                 uint64_t over;
 885
 886                 if (usperop <= total)
 887                         over = bs;
 888                 else
 889                         over = (usperop - total) / usperop * -bs;
 890
 891                 td->rate_io_issue_bytes[ddir] += (missed - over);
 892                 /* adjust for rate_process=poisson */
 893                 td->last_usec[ddir] += total;
 894         }
 895 }
 896
 897 /*
 898  * Main IO worker function. It retrieves io_u's to process and queues
 899  * and reaps them, checking for rate and errors along the way.
 900  *
 901  * Returns number of bytes written and trimmed.
 902  */
 903 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 904 {
 905         unsigned int i;
 906         int ret = 0;
 907         uint64_t total_bytes, bytes_issued = 0;
 908
 909         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 910                 bytes_done[i] = td->bytes_done[i];
 911
 912         if (in_ramp_time(td))
 913                 td_set_runstate(td, TD_RAMP);
 914         else
 915                 td_set_runstate(td, TD_RUNNING);
 916
 917         lat_target_init(td);
 918
 919         total_bytes = td->o.size;
 920         /*
 921         * Allow random overwrite workloads to write up to io_size
 922         * before starting verification phase as 'size' doesn't apply.
 923         */
 924         if (td_write(td) && td_random(td) && td->o.norandommap)
 925                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 926         /*
 927          * If verify_backlog is enabled, we'll run the verify in this
 928          * handler as well. For that case, we may need up to twice the
 929          * amount of bytes.
 930          */
 931         if (td->o.verify != VERIFY_NONE &&
 932            (td_write(td) && td->o.verify_backlog))
 933                 total_bytes += td->o.size;
 934
 935         /* In trimwrite mode, each byte is trimmed and then written, so
 936          * allow total_bytes to be twice as big */
 937         if (td_trimwrite(td))
 938                 total_bytes += td->total_io_size;
 939
 940         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 941                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 942                 td->o.time_based) {
 943                 struct timespec comp_time;
 944                 struct io_u *io_u;
 945                 int full;
 946                 enum fio_ddir ddir;
 947
 948                 check_update_rusage(td);
 949
 950                 if (td->terminate || td->done)
 951                         break;
 952
 953                 update_ts_cache(td);
 954
 955                 if (runtime_exceeded(td, &td->ts_cache)) {
 956                         __update_ts_cache(td);
 957                         if (runtime_exceeded(td, &td->ts_cache)) {
 958                                 fio_mark_td_terminate(td);
 959                                 break;
 960                         }
 961                 }
 962
 963                 if (flow_threshold_exceeded(td))
 964                         continue;
 965
 966                 /*
 967                  * Break if we exceeded the bytes. The exception is time
 968                  * based runs, but we still need to break out of the loop
 969                  * for those to run verification, if enabled.
 970                  * Jobs read from iolog do not use this stop condition.
 971                  */
 972                 if (bytes_issued >= total_bytes &&
 973                     !td->o.read_iolog_file &&
 974                     (!td->o.time_based ||
 975                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 976                         break;
 977
 978                 io_u = get_io_u(td);
 979                 if (IS_ERR_OR_NULL(io_u)) {
 980                         int err = PTR_ERR(io_u);
 981
 982                         io_u = NULL;
 983                         ddir = DDIR_INVAL;
 984                         if (err == -EBUSY) {
 985                                 ret = FIO_Q_BUSY;
 986                                 goto reap;
 987                         }
 988                         if (td->o.latency_target)
 989                                 goto reap;
 990                         break;
 991                 }
 992
 993                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
 994                         populate_verify_io_u(td, io_u);
 995
 996                 ddir = io_u->ddir;
 997
 998                 /*
 999                  * Add verification end_io handler if:
1000                  *      - Asked to verify (!td_rw(td))
1001                  *      - Or the io_u is from our verify list (mixed write/ver)
1002                  */
1003                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1004                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1005
1006                         if (!td->o.verify_pattern_bytes) {
1007                                 io_u->rand_seed = __rand(&td->verify_state);
1008                                 if (sizeof(int) != sizeof(long *))
1009                                         io_u->rand_seed *= __rand(&td->verify_state);
1010                         }
1011
1012                         if (verify_state_should_stop(td, io_u)) {
1013                                 put_io_u(td, io_u);
1014                                 break;
1015                         }
1016
1017                         if (td->o.verify_async)
1018                                 io_u->end_io = verify_io_u_async;
1019                         else
1020                                 io_u->end_io = verify_io_u;
1021                         td_set_runstate(td, TD_VERIFYING);
1022                 } else if (in_ramp_time(td))
1023                         td_set_runstate(td, TD_RAMP);
1024                 else
1025                         td_set_runstate(td, TD_RUNNING);
1026
1027                 /*
1028                  * Always log IO before it's issued, so we know the specific
1029                  * order of it. The logged unit will track when the IO has
1030                  * completed.
1031                  */
1032                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1033                     td->o.do_verify &&
1034                     td->o.verify != VERIFY_NONE &&
1035                     !td->o.experimental_verify)
1036                         log_io_piece(td, io_u);
1037
1038                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1039                         const unsigned long long blen = io_u->xfer_buflen;
1040                         const enum fio_ddir __ddir = acct_ddir(io_u);
1041
1042                         if (td->error)
1043                                 break;
1044
1045                         workqueue_enqueue(&td->io_wq, &io_u->work);
1046                         ret = FIO_Q_QUEUED;
1047
1048                         if (ddir_rw(__ddir)) {
1049                                 td->io_issues[__ddir]++;
1050                                 td->io_issue_bytes[__ddir] += blen;
1051                                 td->rate_io_issue_bytes[__ddir] += blen;
1052                         }
1053
1054                         if (should_check_rate(td))
1055                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1056
1057                 } else {
1058                         ret = io_u_submit(td, io_u);
1059
1060                         if (should_check_rate(td))
1061                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1062
1063                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1064                                 break;
1065
1066                         /*
1067                          * See if we need to complete some commands. Note that
1068                          * we can get BUSY even without IO queued, if the
1069                          * system is resource starved.
1070                          */
1071 reap:
1072                         full = queue_full(td) ||
1073                                 (ret == FIO_Q_BUSY && td->cur_depth);
1074                         if (full || io_in_polling(td))
1075                                 ret = wait_for_completions(td, &comp_time);
1076                 }
1077                 if (ret < 0)
1078                         break;
1079                 if (!ddir_rw_sum(td->bytes_done) &&
1080                     !td_ioengine_flagged(td, FIO_NOIO))
1081                         continue;
1082
1083                 if (!in_ramp_time(td) && should_check_rate(td)) {
1084                         if (check_min_rate(td, &comp_time)) {
1085                                 if (exitall_on_terminate || td->o.exitall_error)
1086                                         fio_terminate_threads(td->groupid);
1087                                 td_verror(td, EIO, "check_min_rate");
1088                                 break;
1089                         }
1090                 }
1091                 if (!in_ramp_time(td) && td->o.latency_target)
1092                         lat_target_check(td);
1093
1094                 if (ddir_rw(ddir) && td->o.thinktime)
1095                         handle_thinktime(td, ddir);
1096         }
1097
1098         check_update_rusage(td);
1099
1100         if (td->trim_entries)
1101                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1102
1103         if (td->o.fill_device && td->error == ENOSPC) {
1104                 td->error = 0;
1105                 fio_mark_td_terminate(td);
1106         }
1107         if (!td->error) {
1108                 struct fio_file *f;
1109
1110                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1111                         workqueue_flush(&td->io_wq);
1112                         i = 0;
1113                 } else
1114                         i = td->cur_depth;
1115
1116                 if (i) {
1117                         ret = io_u_queued_complete(td, i);
1118                         if (td->o.fill_device && td->error == ENOSPC)
1119                                 td->error = 0;
1120                 }
1121
1122                 if (should_fsync(td) && td->o.end_fsync) {
1123                         td_set_runstate(td, TD_FSYNCING);
1124
1125                         for_each_file(td, f, i) {
1126                                 if (!fio_file_fsync(td, f))
1127                                         continue;
1128
1129                                 log_err("fio: end_fsync failed for file %s\n",
1130                                                                 f->file_name);
1131                         }
1132                 }
1133         } else
1134                 cleanup_pending_aio(td);
1135
1136         /*
1137          * stop job if we failed doing any IO
1138          */
1139         if (!ddir_rw_sum(td->this_io_bytes))
1140                 td->done = 1;
1141
1142         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1143                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1144 }
1145
1146 static void free_file_completion_logging(struct thread_data *td)
1147 {
1148         struct fio_file *f;
1149         unsigned int i;
1150
1151         for_each_file(td, f, i) {
1152                 if (!f->last_write_comp)
1153                         break;
1154                 sfree(f->last_write_comp);
1155         }
1156 }
1157
1158 static int init_file_completion_logging(struct thread_data *td,
1159                                         unsigned int depth)
1160 {
1161         struct fio_file *f;
1162         unsigned int i;
1163
1164         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1165                 return 0;
1166
1167         for_each_file(td, f, i) {
1168                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1169                 if (!f->last_write_comp)
1170                         goto cleanup;
1171         }
1172
1173         return 0;
1174
1175 cleanup:
1176         free_file_completion_logging(td);
1177         log_err("fio: failed to alloc write comp data\n");
1178         return 1;
1179 }
1180
1181 static void cleanup_io_u(struct thread_data *td)
1182 {
1183         struct io_u *io_u;
1184
1185         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1186
1187                 if (td->io_ops->io_u_free)
1188                         td->io_ops->io_u_free(td, io_u);
1189
1190                 fio_memfree(io_u, sizeof(*io_u));
1191         }
1192
1193         free_io_mem(td);
1194
1195         io_u_rexit(&td->io_u_requeues);
1196         io_u_qexit(&td->io_u_freelist);
1197         io_u_qexit(&td->io_u_all);
1198
1199         free_file_completion_logging(td);
1200 }
1201
1202 static int init_io_u(struct thread_data *td)
1203 {
1204         struct io_u *io_u;
1205         int cl_align, i, max_units;
1206         int err;
1207
1208         max_units = td->o.iodepth;
1209
1210         err = 0;
1211         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1212         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1213         err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1214
1215         if (err) {
1216                 log_err("fio: failed setting up IO queues\n");
1217                 return 1;
1218         }
1219
1220         cl_align = os_cache_line_size();
1221
1222         for (i = 0; i < max_units; i++) {
1223                 void *ptr;
1224
1225                 if (td->terminate)
1226                         return 1;
1227
1228                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1229                 if (!ptr) {
1230                         log_err("fio: unable to allocate aligned memory\n");
1231                         break;
1232                 }
1233
1234                 io_u = ptr;
1235                 memset(io_u, 0, sizeof(*io_u));
1236                 INIT_FLIST_HEAD(&io_u->verify_list);
1237                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1238
1239                 io_u->index = i;
1240                 io_u->flags = IO_U_F_FREE;
1241                 io_u_qpush(&td->io_u_freelist, io_u);
1242
1243                 /*
1244                  * io_u never leaves this stack, used for iteration of all
1245                  * io_u buffers.
1246                  */
1247                 io_u_qpush(&td->io_u_all, io_u);
1248
1249                 if (td->io_ops->io_u_init) {
1250                         int ret = td->io_ops->io_u_init(td, io_u);
1251
1252                         if (ret) {
1253                                 log_err("fio: failed to init engine data: %d\n", ret);
1254                                 return 1;
1255                         }
1256                 }
1257         }
1258
1259         init_io_u_buffers(td);
1260
1261         if (init_file_completion_logging(td, max_units))
1262                 return 1;
1263
1264         return 0;
1265 }
1266
1267 int init_io_u_buffers(struct thread_data *td)
1268 {
1269         struct io_u *io_u;
1270         unsigned long long max_bs, min_write;
1271         int i, max_units;
1272         int data_xfer = 1;
1273         char *p;
1274
1275         max_units = td->o.iodepth;
1276         max_bs = td_max_bs(td);
1277         min_write = td->o.min_bs[DDIR_WRITE];
1278         td->orig_buffer_size = (unsigned long long) max_bs
1279                                         * (unsigned long long) max_units;
1280
1281         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1282                 data_xfer = 0;
1283
1284         /*
1285          * if we may later need to do address alignment, then add any
1286          * possible adjustment here so that we don't cause a buffer
1287          * overflow later. this adjustment may be too much if we get
1288          * lucky and the allocator gives us an aligned address.
1289          */
1290         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1291             td_ioengine_flagged(td, FIO_RAWIO))
1292                 td->orig_buffer_size += page_mask + td->o.mem_align;
1293
1294         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1295                 unsigned long long bs;
1296
1297                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1298                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1299         }
1300
1301         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1302                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1303                 return 1;
1304         }
1305
1306         if (data_xfer && allocate_io_mem(td))
1307                 return 1;
1308
1309         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1310             td_ioengine_flagged(td, FIO_RAWIO))
1311                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1312         else
1313                 p = td->orig_buffer;
1314
1315         for (i = 0; i < max_units; i++) {
1316                 io_u = td->io_u_all.io_us[i];
1317                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1318
1319                 if (data_xfer) {
1320                         io_u->buf = p;
1321                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1322
1323                         if (td_write(td))
1324                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1325                         if (td_write(td) && td->o.verify_pattern_bytes) {
1326                                 /*
1327                                  * Fill the buffer with the pattern if we are
1328                                  * going to be doing writes.
1329                                  */
1330                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1331                         }
1332                 }
1333                 p += max_bs;
1334         }
1335
1336         return 0;
1337 }
1338
1339 /*
1340  * This function is Linux specific.
1341  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1342  */
1343 static int switch_ioscheduler(struct thread_data *td)
1344 {
1345 #ifdef FIO_HAVE_IOSCHED_SWITCH
1346         char tmp[256], tmp2[128], *p;
1347         FILE *f;
1348         int ret;
1349
1350         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1351                 return 0;
1352
1353         assert(td->files && td->files[0]);
1354         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1355
1356         f = fopen(tmp, "r+");
1357         if (!f) {
1358                 if (errno == ENOENT) {
1359                         log_err("fio: os or kernel doesn't support IO scheduler"
1360                                 " switching\n");
1361                         return 0;
1362                 }
1363                 td_verror(td, errno, "fopen iosched");
1364                 return 1;
1365         }
1366
1367         /*
1368          * Set io scheduler.
1369          */
1370         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1371         if (ferror(f) || ret != 1) {
1372                 td_verror(td, errno, "fwrite");
1373                 fclose(f);
1374                 return 1;
1375         }
1376
1377         rewind(f);
1378
1379         /*
1380          * Read back and check that the selected scheduler is now the default.
1381          */
1382         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1383         if (ferror(f) || ret < 0) {
1384                 td_verror(td, errno, "fread");
1385                 fclose(f);
1386                 return 1;
1387         }
1388         tmp[ret] = '\0';
1389         /*
1390          * either a list of io schedulers or "none\n" is expected. Strip the
1391          * trailing newline.
1392          */
1393         p = tmp;
1394         strsep(&p, "\n");
1395
1396         /*
1397          * Write to "none" entry doesn't fail, so check the result here.
1398          */
1399         if (!strcmp(tmp, "none")) {
1400                 log_err("fio: io scheduler is not tunable\n");
1401                 fclose(f);
1402                 return 0;
1403         }
1404
1405         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1406         if (!strstr(tmp, tmp2)) {
1407                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1408                 td_verror(td, EINVAL, "iosched_switch");
1409                 fclose(f);
1410                 return 1;
1411         }
1412
1413         fclose(f);
1414         return 0;
1415 #else
1416         return 0;
1417 #endif
1418 }
1419
1420 static bool keep_running(struct thread_data *td)
1421 {
1422         unsigned long long limit;
1423
1424         if (td->done)
1425                 return false;
1426         if (td->terminate)
1427                 return false;
1428         if (td->o.time_based)
1429                 return true;
1430         if (td->o.loops) {
1431                 td->o.loops--;
1432                 return true;
1433         }
1434         if (exceeds_number_ios(td))
1435                 return false;
1436
1437         if (td->o.io_size)
1438                 limit = td->o.io_size;
1439         else
1440                 limit = td->o.size;
1441
1442         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1443                 uint64_t diff;
1444
1445                 /*
1446                  * If the difference is less than the maximum IO size, we
1447                  * are done.
1448                  */
1449                 diff = limit - ddir_rw_sum(td->io_bytes);
1450                 if (diff < td_max_bs(td))
1451                         return false;
1452
1453                 if (fio_files_done(td) && !td->o.io_size)
1454                         return false;
1455
1456                 return true;
1457         }
1458
1459         return false;
1460 }
1461
1462 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1463 {
1464         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1465         int ret;
1466         char *str;
1467
1468         str = malloc(newlen);
1469         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1470
1471         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1472         ret = system(str);
1473         if (ret == -1)
1474                 log_err("fio: exec of cmd <%s> failed\n", str);
1475
1476         free(str);
1477         return ret;
1478 }
1479
1480 /*
1481  * Dry run to compute correct state of numberio for verification.
1482  */
1483 static uint64_t do_dry_run(struct thread_data *td)
1484 {
1485         td_set_runstate(td, TD_RUNNING);
1486
1487         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1488                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1489                 struct io_u *io_u;
1490                 int ret;
1491
1492                 if (td->terminate || td->done)
1493                         break;
1494
1495                 io_u = get_io_u(td);
1496                 if (IS_ERR_OR_NULL(io_u))
1497                         break;
1498
1499                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1500                 io_u->error = 0;
1501                 io_u->resid = 0;
1502                 if (ddir_rw(acct_ddir(io_u)))
1503                         td->io_issues[acct_ddir(io_u)]++;
1504                 if (ddir_rw(io_u->ddir)) {
1505                         io_u_mark_depth(td, 1);
1506                         td->ts.total_io_u[io_u->ddir]++;
1507                 }
1508
1509                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1510                     td->o.do_verify &&
1511                     td->o.verify != VERIFY_NONE &&
1512                     !td->o.experimental_verify)
1513                         log_io_piece(td, io_u);
1514
1515                 ret = io_u_sync_complete(td, io_u);
1516                 (void) ret;
1517         }
1518
1519         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1520 }
1521
1522 struct fork_data {
1523         struct thread_data *td;
1524         struct sk_out *sk_out;
1525 };
1526
1527 /*
1528  * Entry point for the thread based jobs. The process based jobs end up
1529  * here as well, after a little setup.
1530  */
1531 static void *thread_main(void *data)
1532 {
1533         struct fork_data *fd = data;
1534         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1535         struct thread_data *td = fd->td;
1536         struct thread_options *o = &td->o;
1537         struct sk_out *sk_out = fd->sk_out;
1538         uint64_t bytes_done[DDIR_RWDIR_CNT];
1539         int deadlock_loop_cnt;
1540         bool clear_state, did_some_io;
1541         int ret;
1542
1543         sk_out_assign(sk_out);
1544         free(fd);
1545
1546         if (!o->use_thread) {
1547                 setsid();
1548                 td->pid = getpid();
1549         } else
1550                 td->pid = gettid();
1551
1552         fio_local_clock_init();
1553
1554         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1555
1556         if (is_backend)
1557                 fio_server_send_start(td);
1558
1559         INIT_FLIST_HEAD(&td->io_log_list);
1560         INIT_FLIST_HEAD(&td->io_hist_list);
1561         INIT_FLIST_HEAD(&td->verify_list);
1562         INIT_FLIST_HEAD(&td->trim_list);
1563         td->io_hist_tree = RB_ROOT;
1564
1565         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1566         if (ret) {
1567                 td_verror(td, ret, "mutex_cond_init_pshared");
1568                 goto err;
1569         }
1570         ret = cond_init_pshared(&td->verify_cond);
1571         if (ret) {
1572                 td_verror(td, ret, "mutex_cond_pshared");
1573                 goto err;
1574         }
1575
1576         td_set_runstate(td, TD_INITIALIZED);
1577         dprint(FD_MUTEX, "up startup_sem\n");
1578         fio_sem_up(startup_sem);
1579         dprint(FD_MUTEX, "wait on td->sem\n");
1580         fio_sem_down(td->sem);
1581         dprint(FD_MUTEX, "done waiting on td->sem\n");
1582
1583         /*
1584          * A new gid requires privilege, so we need to do this before setting
1585          * the uid.
1586          */
1587         if (o->gid != -1U && setgid(o->gid)) {
1588                 td_verror(td, errno, "setgid");
1589                 goto err;
1590         }
1591         if (o->uid != -1U && setuid(o->uid)) {
1592                 td_verror(td, errno, "setuid");
1593                 goto err;
1594         }
1595
1596         td_zone_gen_index(td);
1597
1598         /*
1599          * Do this early, we don't want the compress threads to be limited
1600          * to the same CPUs as the IO workers. So do this before we set
1601          * any potential CPU affinity
1602          */
1603         if (iolog_compress_init(td, sk_out))
1604                 goto err;
1605
1606         /*
1607          * If we have a gettimeofday() thread, make sure we exclude that
1608          * thread from this job
1609          */
1610         if (o->gtod_cpu)
1611                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1612
1613         /*
1614          * Set affinity first, in case it has an impact on the memory
1615          * allocations.
1616          */
1617         if (fio_option_is_set(o, cpumask)) {
1618                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1619                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1620                         if (!ret) {
1621                                 log_err("fio: no CPUs set\n");
1622                                 log_err("fio: Try increasing number of available CPUs\n");
1623                                 td_verror(td, EINVAL, "cpus_split");
1624                                 goto err;
1625                         }
1626                 }
1627                 ret = fio_setaffinity(td->pid, o->cpumask);
1628                 if (ret == -1) {
1629                         td_verror(td, errno, "cpu_set_affinity");
1630                         goto err;
1631                 }
1632         }
1633
1634 #ifdef CONFIG_LIBNUMA
1635         /* numa node setup */
1636         if (fio_option_is_set(o, numa_cpunodes) ||
1637             fio_option_is_set(o, numa_memnodes)) {
1638                 struct bitmask *mask;
1639
1640                 if (numa_available() < 0) {
1641                         td_verror(td, errno, "Does not support NUMA API\n");
1642                         goto err;
1643                 }
1644
1645                 if (fio_option_is_set(o, numa_cpunodes)) {
1646                         mask = numa_parse_nodestring(o->numa_cpunodes);
1647                         ret = numa_run_on_node_mask(mask);
1648                         numa_free_nodemask(mask);
1649                         if (ret == -1) {
1650                                 td_verror(td, errno, \
1651                                         "numa_run_on_node_mask failed\n");
1652                                 goto err;
1653                         }
1654                 }
1655
1656                 if (fio_option_is_set(o, numa_memnodes)) {
1657                         mask = NULL;
1658                         if (o->numa_memnodes)
1659                                 mask = numa_parse_nodestring(o->numa_memnodes);
1660
1661                         switch (o->numa_mem_mode) {
1662                         case MPOL_INTERLEAVE:
1663                                 numa_set_interleave_mask(mask);
1664                                 break;
1665                         case MPOL_BIND:
1666                                 numa_set_membind(mask);
1667                                 break;
1668                         case MPOL_LOCAL:
1669                                 numa_set_localalloc();
1670                                 break;
1671                         case MPOL_PREFERRED:
1672                                 numa_set_preferred(o->numa_mem_prefer_node);
1673                                 break;
1674                         case MPOL_DEFAULT:
1675                         default:
1676                                 break;
1677                         }
1678
1679                         if (mask)
1680                                 numa_free_nodemask(mask);
1681
1682                 }
1683         }
1684 #endif
1685
1686         if (fio_pin_memory(td))
1687                 goto err;
1688
1689         /*
1690          * May alter parameters that init_io_u() will use, so we need to
1691          * do this first.
1692          */
1693         if (!init_iolog(td))
1694                 goto err;
1695
1696         if (init_io_u(td))
1697                 goto err;
1698
1699         if (o->verify_async && verify_async_init(td))
1700                 goto err;
1701
1702         if (fio_option_is_set(o, ioprio) ||
1703             fio_option_is_set(o, ioprio_class)) {
1704                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1705                 if (ret == -1) {
1706                         td_verror(td, errno, "ioprio_set");
1707                         goto err;
1708                 }
1709         }
1710
1711         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1712                 goto err;
1713
1714         errno = 0;
1715         if (nice(o->nice) == -1 && errno != 0) {
1716                 td_verror(td, errno, "nice");
1717                 goto err;
1718         }
1719
1720         if (o->ioscheduler && switch_ioscheduler(td))
1721                 goto err;
1722
1723         if (!o->create_serialize && setup_files(td))
1724                 goto err;
1725
1726         if (td_io_init(td))
1727                 goto err;
1728
1729         if (!init_random_map(td))
1730                 goto err;
1731
1732         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1733                 goto err;
1734
1735         if (o->pre_read && !pre_read_files(td))
1736                 goto err;
1737
1738         fio_verify_init(td);
1739
1740         if (rate_submit_init(td, sk_out))
1741                 goto err;
1742
1743         set_epoch_time(td, o->log_unix_epoch);
1744         fio_getrusage(&td->ru_start);
1745         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1746         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1747         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1748
1749         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1750                         o->ratemin[DDIR_TRIM]) {
1751                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1752                                         sizeof(td->bw_sample_time));
1753                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1754                                         sizeof(td->bw_sample_time));
1755                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1756                                         sizeof(td->bw_sample_time));
1757         }
1758
1759         memset(bytes_done, 0, sizeof(bytes_done));
1760         clear_state = false;
1761         did_some_io = false;
1762
1763         while (keep_running(td)) {
1764                 uint64_t verify_bytes;
1765
1766                 fio_gettime(&td->start, NULL);
1767                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1768
1769                 if (clear_state) {
1770                         clear_io_state(td, 0);
1771
1772                         if (o->unlink_each_loop && unlink_all_files(td))
1773                                 break;
1774                 }
1775
1776                 prune_io_piece_log(td);
1777
1778                 if (td->o.verify_only && td_write(td))
1779                         verify_bytes = do_dry_run(td);
1780                 else {
1781                         do_io(td, bytes_done);
1782
1783                         if (!ddir_rw_sum(bytes_done)) {
1784                                 fio_mark_td_terminate(td);
1785                                 verify_bytes = 0;
1786                         } else {
1787                                 verify_bytes = bytes_done[DDIR_WRITE] +
1788                                                 bytes_done[DDIR_TRIM];
1789                         }
1790                 }
1791
1792                 /*
1793                  * If we took too long to shut down, the main thread could
1794                  * already consider us reaped/exited. If that happens, break
1795                  * out and clean up.
1796                  */
1797                 if (td->runstate >= TD_EXITED)
1798                         break;
1799
1800                 clear_state = true;
1801
1802                 /*
1803                  * Make sure we've successfully updated the rusage stats
1804                  * before waiting on the stat mutex. Otherwise we could have
1805                  * the stat thread holding stat mutex and waiting for
1806                  * the rusage_sem, which would never get upped because
1807                  * this thread is waiting for the stat mutex.
1808                  */
1809                 deadlock_loop_cnt = 0;
1810                 do {
1811                         check_update_rusage(td);
1812                         if (!fio_sem_down_trylock(stat_sem))
1813                                 break;
1814                         usleep(1000);
1815                         if (deadlock_loop_cnt++ > 5000) {
1816                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1817                                 td->error = EDEADLK;
1818                                 goto err;
1819                         }
1820                 } while (1);
1821
1822                 if (td_read(td) && td->io_bytes[DDIR_READ])
1823                         update_runtime(td, elapsed_us, DDIR_READ);
1824                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1825                         update_runtime(td, elapsed_us, DDIR_WRITE);
1826                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1827                         update_runtime(td, elapsed_us, DDIR_TRIM);
1828                 fio_gettime(&td->start, NULL);
1829                 fio_sem_up(stat_sem);
1830
1831                 if (td->error || td->terminate)
1832                         break;
1833
1834                 if (!o->do_verify ||
1835                     o->verify == VERIFY_NONE ||
1836                     td_ioengine_flagged(td, FIO_UNIDIR))
1837                         continue;
1838
1839                 if (ddir_rw_sum(bytes_done))
1840                         did_some_io = true;
1841
1842                 clear_io_state(td, 0);
1843
1844                 fio_gettime(&td->start, NULL);
1845
1846                 do_verify(td, verify_bytes);
1847
1848                 /*
1849                  * See comment further up for why this is done here.
1850                  */
1851                 check_update_rusage(td);
1852
1853                 fio_sem_down(stat_sem);
1854                 update_runtime(td, elapsed_us, DDIR_READ);
1855                 fio_gettime(&td->start, NULL);
1856                 fio_sem_up(stat_sem);
1857
1858                 if (td->error || td->terminate)
1859                         break;
1860         }
1861
1862         /*
1863          * If td ended up with no I/O when it should have had,
1864          * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1865          * (Are we not missing other flags that can be ignored ?)
1866          */
1867         if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1868             !did_some_io && !td->o.create_only &&
1869             !(td_ioengine_flagged(td, FIO_NOIO) ||
1870               td_ioengine_flagged(td, FIO_DISKLESSIO)))
1871                 log_err("%s: No I/O performed by %s, "
1872                          "perhaps try --debug=io option for details?\n",
1873                          td->o.name, td->io_ops->name);
1874
1875         td_set_runstate(td, TD_FINISHING);
1876
1877         update_rusage_stat(td);
1878         td->ts.total_run_time = mtime_since_now(&td->epoch);
1879         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1880         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1881         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1882
1883         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1884             (td->o.verify != VERIFY_NONE && td_write(td)))
1885                 verify_save_state(td->thread_number);
1886
1887         fio_unpin_memory(td);
1888
1889         td_writeout_logs(td, true);
1890
1891         iolog_compress_exit(td);
1892         rate_submit_exit(td);
1893
1894         if (o->exec_postrun)
1895                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1896
1897         if (exitall_on_terminate || (o->exitall_error && td->error))
1898                 fio_terminate_threads(td->groupid);
1899
1900 err:
1901         if (td->error)
1902                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1903                                                         td->verror);
1904
1905         if (o->verify_async)
1906                 verify_async_exit(td);
1907
1908         close_and_free_files(td);
1909         cleanup_io_u(td);
1910         close_ioengine(td);
1911         cgroup_shutdown(td, cgroup_mnt);
1912         verify_free_state(td);
1913         td_zone_free_index(td);
1914
1915         if (fio_option_is_set(o, cpumask)) {
1916                 ret = fio_cpuset_exit(&o->cpumask);
1917                 if (ret)
1918                         td_verror(td, ret, "fio_cpuset_exit");
1919         }
1920
1921         /*
1922          * do this very late, it will log file closing as well
1923          */
1924         if (o->write_iolog_file)
1925                 write_iolog_close(td);
1926         if (td->io_log_rfile)
1927                 fclose(td->io_log_rfile);
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                 struct buf_output out;
2217
2218                 buf_output_init(&out);
2219                 __log_buf(&out, "Starting ");
2220                 if (nr_thread)
2221                         __log_buf(&out, "%d thread%s", nr_thread,
2222                                                 nr_thread > 1 ? "s" : "");
2223                 if (nr_process) {
2224                         if (nr_thread)
2225                                 __log_buf(&out, " and ");
2226                         __log_buf(&out, "%d process%s", nr_process,
2227                                                 nr_process > 1 ? "es" : "");
2228                 }
2229                 __log_buf(&out, "\n");
2230                 log_info_buf(out.buf, out.buflen);
2231                 buf_output_free(&out);
2232         }
2233
2234         todo = thread_number;
2235         nr_running = 0;
2236         nr_started = 0;
2237         m_rate = t_rate = 0;
2238
2239         for_each_td(td, i) {
2240                 print_status_init(td->thread_number - 1);
2241
2242                 if (!td->o.create_serialize)
2243                         continue;
2244
2245                 if (fio_verify_load_state(td))
2246                         goto reap;
2247
2248                 /*
2249                  * do file setup here so it happens sequentially,
2250                  * we don't want X number of threads getting their
2251                  * client data interspersed on disk
2252                  */
2253                 if (setup_files(td)) {
2254 reap:
2255                         exit_value++;
2256                         if (td->error)
2257                                 log_err("fio: pid=%d, err=%d/%s\n",
2258                                         (int) td->pid, td->error, td->verror);
2259                         td_set_runstate(td, TD_REAPED);
2260                         todo--;
2261                 } else {
2262                         struct fio_file *f;
2263                         unsigned int j;
2264
2265                         /*
2266                          * for sharing to work, each job must always open
2267                          * its own files. so close them, if we opened them
2268                          * for creation
2269                          */
2270                         for_each_file(td, f, j) {
2271                                 if (fio_file_open(f))
2272                                         td_io_close_file(td, f);
2273                         }
2274                 }
2275         }
2276
2277         /* start idle threads before io threads start to run */
2278         fio_idle_prof_start();
2279
2280         set_genesis_time();
2281
2282         while (todo) {
2283                 struct thread_data *map[REAL_MAX_JOBS];
2284                 struct timespec this_start;
2285                 int this_jobs = 0, left;
2286                 struct fork_data *fd;
2287
2288                 /*
2289                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2290                  */
2291                 for_each_td(td, i) {
2292                         if (td->runstate != TD_NOT_CREATED)
2293                                 continue;
2294
2295                         /*
2296                          * never got a chance to start, killed by other
2297                          * thread for some reason
2298                          */
2299                         if (td->terminate) {
2300                                 todo--;
2301                                 continue;
2302                         }
2303
2304                         if (td->o.start_delay) {
2305                                 spent = utime_since_genesis();
2306
2307                                 if (td->o.start_delay > spent)
2308                                         continue;
2309                         }
2310
2311                         if (td->o.stonewall && (nr_started || nr_running)) {
2312                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2313                                                         td->o.name);
2314                                 break;
2315                         }
2316
2317                         if (waitee_running(td)) {
2318                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2319                                                 td->o.name, td->o.wait_for);
2320                                 continue;
2321                         }
2322
2323                         init_disk_util(td);
2324
2325                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2326                         td->update_rusage = 0;
2327
2328                         /*
2329                          * Set state to created. Thread will transition
2330                          * to TD_INITIALIZED when it's done setting up.
2331                          */
2332                         td_set_runstate(td, TD_CREATED);
2333                         map[this_jobs++] = td;
2334                         nr_started++;
2335
2336                         fd = calloc(1, sizeof(*fd));
2337                         fd->td = td;
2338                         fd->sk_out = sk_out;
2339
2340                         if (td->o.use_thread) {
2341                                 int ret;
2342
2343                                 dprint(FD_PROCESS, "will pthread_create\n");
2344                                 ret = pthread_create(&td->thread, NULL,
2345                                                         thread_main, fd);
2346                                 if (ret) {
2347                                         log_err("pthread_create: %s\n",
2348                                                         strerror(ret));
2349                                         free(fd);
2350                                         nr_started--;
2351                                         break;
2352                                 }
2353                                 fd = NULL;
2354                                 ret = pthread_detach(td->thread);
2355                                 if (ret)
2356                                         log_err("pthread_detach: %s",
2357                                                         strerror(ret));
2358                         } else {
2359                                 pid_t pid;
2360                                 dprint(FD_PROCESS, "will fork\n");
2361                                 pid = fork();
2362                                 if (!pid) {
2363                                         int ret;
2364
2365                                         ret = (int)(uintptr_t)thread_main(fd);
2366                                         _exit(ret);
2367                                 } else if (i == fio_debug_jobno)
2368                                         *fio_debug_jobp = pid;
2369                         }
2370                         dprint(FD_MUTEX, "wait on startup_sem\n");
2371                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2372                                 log_err("fio: job startup hung? exiting.\n");
2373                                 fio_terminate_threads(TERMINATE_ALL);
2374                                 fio_abort = true;
2375                                 nr_started--;
2376                                 free(fd);
2377                                 break;
2378                         }
2379                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2380                 }
2381
2382                 /*
2383                  * Wait for the started threads to transition to
2384                  * TD_INITIALIZED.
2385                  */
2386                 fio_gettime(&this_start, NULL);
2387                 left = this_jobs;
2388                 while (left && !fio_abort) {
2389                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2390                                 break;
2391
2392                         do_usleep(100000);
2393
2394                         for (i = 0; i < this_jobs; i++) {
2395                                 td = map[i];
2396                                 if (!td)
2397                                         continue;
2398                                 if (td->runstate == TD_INITIALIZED) {
2399                                         map[i] = NULL;
2400                                         left--;
2401                                 } else if (td->runstate >= TD_EXITED) {
2402                                         map[i] = NULL;
2403                                         left--;
2404                                         todo--;
2405                                         nr_running++; /* work-around... */
2406                                 }
2407                         }
2408                 }
2409
2410                 if (left) {
2411                         log_err("fio: %d job%s failed to start\n", left,
2412                                         left > 1 ? "s" : "");
2413                         for (i = 0; i < this_jobs; i++) {
2414                                 td = map[i];
2415                                 if (!td)
2416                                         continue;
2417                                 kill(td->pid, SIGTERM);
2418                         }
2419                         break;
2420                 }
2421
2422                 /*
2423                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2424                  */
2425                 for_each_td(td, i) {
2426                         if (td->runstate != TD_INITIALIZED)
2427                                 continue;
2428
2429                         if (in_ramp_time(td))
2430                                 td_set_runstate(td, TD_RAMP);
2431                         else
2432                                 td_set_runstate(td, TD_RUNNING);
2433                         nr_running++;
2434                         nr_started--;
2435                         m_rate += ddir_rw_sum(td->o.ratemin);
2436                         t_rate += ddir_rw_sum(td->o.rate);
2437                         todo--;
2438                         fio_sem_up(td->sem);
2439                 }
2440
2441                 reap_threads(&nr_running, &t_rate, &m_rate);
2442
2443                 if (todo)
2444                         do_usleep(100000);
2445         }
2446
2447         while (nr_running) {
2448                 reap_threads(&nr_running, &t_rate, &m_rate);
2449                 do_usleep(10000);
2450         }
2451
2452         fio_idle_prof_stop();
2453
2454         update_io_ticks();
2455 }
2456
2457 static void free_disk_util(void)
2458 {
2459         disk_util_prune_entries();
2460         helper_thread_destroy();
2461 }
2462
2463 int fio_backend(struct sk_out *sk_out)
2464 {
2465         struct thread_data *td;
2466         int i;
2467
2468         if (exec_profile) {
2469                 if (load_profile(exec_profile))
2470                         return 1;
2471                 free(exec_profile);
2472                 exec_profile = NULL;
2473         }
2474         if (!thread_number)
2475                 return 0;
2476
2477         if (write_bw_log) {
2478                 struct log_params p = {
2479                         .log_type = IO_LOG_TYPE_BW,
2480                 };
2481
2482                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2483                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2484                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2485         }
2486
2487         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2488         if (!sk_out)
2489                 is_local_backend = true;
2490         if (startup_sem == NULL)
2491                 return 1;
2492
2493         set_genesis_time();
2494         stat_init();
2495         helper_thread_create(startup_sem, sk_out);
2496
2497         cgroup_list = smalloc(sizeof(*cgroup_list));
2498         if (cgroup_list)
2499                 INIT_FLIST_HEAD(cgroup_list);
2500
2501         run_threads(sk_out);
2502
2503         helper_thread_exit();
2504
2505         if (!fio_abort) {
2506                 __show_run_stats();
2507                 if (write_bw_log) {
2508                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2509                                 struct io_log *log = agg_io_log[i];
2510
2511                                 flush_log(log, false);
2512                                 free_log(log);
2513                         }
2514                 }
2515         }
2516
2517         for_each_td(td, i) {
2518                 steadystate_free(td);
2519                 fio_options_free(td);
2520                 if (td->rusage_sem) {
2521                         fio_sem_remove(td->rusage_sem);
2522                         td->rusage_sem = NULL;
2523                 }
2524                 fio_sem_remove(td->sem);
2525                 td->sem = NULL;
2526         }
2527
2528         free_disk_util();
2529         if (cgroup_list) {
2530                 cgroup_kill(cgroup_list);
2531                 sfree(cgroup_list);
2532         }
2533
2534         fio_sem_remove(startup_sem);
2535         stat_exit();
2536         return exit_value;
2537 }