backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32 #include <pthread.h>
  33
  34 #include "fio.h"
  35 #include "smalloc.h"
  36 #include "verify.h"
  37 #include "diskutil.h"
  38 #include "cgroup.h"
  39 #include "profile.h"
  40 #include "lib/rand.h"
  41 #include "lib/memalign.h"
  42 #include "server.h"
  43 #include "lib/getrusage.h"
  44 #include "idletime.h"
  45 #include "err.h"
  46 #include "workqueue.h"
  47 #include "lib/mountcheck.h"
  48 #include "rate-submit.h"
  49 #include "helper_thread.h"
  50 #include "pshared.h"
  51 #include "zone-dist.h"
  52
  53 static struct fio_sem *startup_sem;
  54 static struct flist_head *cgroup_list;
  55 static struct cgroup_mnt *cgroup_mnt;
  56 static int exit_value;
  57 static volatile bool fio_abort;
  58 static unsigned int nr_process = 0;
  59 static unsigned int nr_thread = 0;
  60
  61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  62
  63 int groupid = 0;
  64 unsigned int thread_number = 0;
  65 unsigned int stat_number = 0;
  66 int shm_id = 0;
  67 int temp_stall_ts;
  68 unsigned long done_secs = 0;
  69 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  70
  71 #define JOB_START_TIMEOUT       (5 * 1000)
  72
  73 static void sig_int(int sig)
  74 {
  75         if (threads) {
  76                 if (is_backend)
  77                         fio_server_got_signal(sig);
  78                 else {
  79                         log_info("\nfio: terminating on signal %d\n", sig);
  80                         log_info_flush();
  81                         exit_value = 128;
  82                 }
  83
  84                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  85         }
  86 }
  87
  88 void sig_show_status(int sig)
  89 {
  90         show_running_run_stats();
  91 }
  92
  93 static void set_sig_handlers(void)
  94 {
  95         struct sigaction act;
  96
  97         memset(&act, 0, sizeof(act));
  98         act.sa_handler = sig_int;
  99         act.sa_flags = SA_RESTART;
 100         sigaction(SIGINT, &act, NULL);
 101
 102         memset(&act, 0, sizeof(act));
 103         act.sa_handler = sig_int;
 104         act.sa_flags = SA_RESTART;
 105         sigaction(SIGTERM, &act, NULL);
 106
 107 /* Windows uses SIGBREAK as a quit signal from other applications */
 108 #ifdef WIN32
 109         memset(&act, 0, sizeof(act));
 110         act.sa_handler = sig_int;
 111         act.sa_flags = SA_RESTART;
 112         sigaction(SIGBREAK, &act, NULL);
 113 #endif
 114
 115         memset(&act, 0, sizeof(act));
 116         act.sa_handler = sig_show_status;
 117         act.sa_flags = SA_RESTART;
 118         sigaction(SIGUSR1, &act, NULL);
 119
 120         if (is_backend) {
 121                 memset(&act, 0, sizeof(act));
 122                 act.sa_handler = sig_int;
 123                 act.sa_flags = SA_RESTART;
 124                 sigaction(SIGPIPE, &act, NULL);
 125         }
 126 }
 127
 128 /*
 129  * Check if we are above the minimum rate given.
 130  */
 131 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 132                              enum fio_ddir ddir)
 133 {
 134         unsigned long long bytes = 0;
 135         unsigned long iops = 0;
 136         unsigned long spent;
 137         unsigned long long rate;
 138         unsigned long long ratemin = 0;
 139         unsigned int rate_iops = 0;
 140         unsigned int rate_iops_min = 0;
 141
 142         assert(ddir_rw(ddir));
 143
 144         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 145                 return false;
 146
 147         /*
 148          * allow a 2 second settle period in the beginning
 149          */
 150         if (mtime_since(&td->start, now) < 2000)
 151                 return false;
 152
 153         iops += td->this_io_blocks[ddir];
 154         bytes += td->this_io_bytes[ddir];
 155         ratemin += td->o.ratemin[ddir];
 156         rate_iops += td->o.rate_iops[ddir];
 157         rate_iops_min += td->o.rate_iops_min[ddir];
 158
 159         /*
 160          * if rate blocks is set, sample is running
 161          */
 162         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 163                 spent = mtime_since(&td->lastrate[ddir], now);
 164                 if (spent < td->o.ratecycle)
 165                         return false;
 166
 167                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 168                         /*
 169                          * check bandwidth specified rate
 170                          */
 171                         if (bytes < td->rate_bytes[ddir]) {
 172                                 log_err("%s: rate_min=%lluB/s not met, only transferred %lluB\n",
 173                                         td->o.name, ratemin, bytes);
 174                                 return true;
 175                         } else {
 176                                 if (spent)
 177                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 178                                 else
 179                                         rate = 0;
 180
 181                                 if (rate < ratemin ||
 182                                     bytes < td->rate_bytes[ddir]) {
 183                                         log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 184                                                 td->o.name, ratemin, rate);
 185                                         return true;
 186                                 }
 187                         }
 188                 } else {
 189                         /*
 190                          * checks iops specified rate
 191                          */
 192                         if (iops < rate_iops) {
 193                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 194                                                 td->o.name, rate_iops, iops);
 195                                 return true;
 196                         } else {
 197                                 if (spent)
 198                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 199                                 else
 200                                         rate = 0;
 201
 202                                 if (rate < rate_iops_min ||
 203                                     iops < td->rate_blocks[ddir]) {
 204                                         log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 205                                                 td->o.name, rate_iops_min, rate);
 206                                         return true;
 207                                 }
 208                         }
 209                 }
 210         }
 211
 212         td->rate_bytes[ddir] = bytes;
 213         td->rate_blocks[ddir] = iops;
 214         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 215         return false;
 216 }
 217
 218 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 219 {
 220         bool ret = false;
 221
 222         if (td->bytes_done[DDIR_READ])
 223                 ret |= __check_min_rate(td, now, DDIR_READ);
 224         if (td->bytes_done[DDIR_WRITE])
 225                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 226         if (td->bytes_done[DDIR_TRIM])
 227                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 228
 229         return ret;
 230 }
 231
 232 /*
 233  * When job exits, we can cancel the in-flight IO if we are using async
 234  * io. Attempt to do so.
 235  */
 236 static void cleanup_pending_aio(struct thread_data *td)
 237 {
 238         int r;
 239
 240         /*
 241          * get immediately available events, if any
 242          */
 243         r = io_u_queued_complete(td, 0);
 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         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 280
 281         if (td_io_prep(td, io_u)) {
 282                 put_io_u(td, io_u);
 283                 return true;
 284         }
 285
 286 requeue:
 287         ret = td_io_queue(td, io_u);
 288         switch (ret) {
 289         case FIO_Q_QUEUED:
 290                 td_io_commit(td);
 291                 if (io_u_queued_complete(td, 1) < 0)
 292                         return true;
 293                 break;
 294         case FIO_Q_COMPLETED:
 295                 if (io_u->error) {
 296                         td_verror(td, io_u->error, "td_io_queue");
 297                         return true;
 298                 }
 299
 300                 if (io_u_sync_complete(td, io_u) < 0)
 301                         return true;
 302                 break;
 303         case FIO_Q_BUSY:
 304                 td_io_commit(td);
 305                 goto requeue;
 306         }
 307
 308         return false;
 309 }
 310
 311 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 312 {
 313         int ret, ret2;
 314
 315         if (fio_file_open(f))
 316                 return fio_io_sync(td, f);
 317
 318         if (td_io_open_file(td, f))
 319                 return 1;
 320
 321         ret = fio_io_sync(td, f);
 322         ret2 = 0;
 323         if (fio_file_open(f))
 324                 ret2 = td_io_close_file(td, f);
 325         return (ret || ret2);
 326 }
 327
 328 static inline void __update_ts_cache(struct thread_data *td)
 329 {
 330         fio_gettime(&td->ts_cache, NULL);
 331 }
 332
 333 static inline void update_ts_cache(struct thread_data *td)
 334 {
 335         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 336                 __update_ts_cache(td);
 337 }
 338
 339 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 340 {
 341         if (in_ramp_time(td))
 342                 return false;
 343         if (!td->o.timeout)
 344                 return false;
 345         if (utime_since(&td->epoch, t) >= td->o.timeout)
 346                 return true;
 347
 348         return false;
 349 }
 350
 351 /*
 352  * We need to update the runtime consistently in ms, but keep a running
 353  * tally of the current elapsed time in microseconds for sub millisecond
 354  * updates.
 355  */
 356 static inline void update_runtime(struct thread_data *td,
 357                                   unsigned long long *elapsed_us,
 358                                   const enum fio_ddir ddir)
 359 {
 360         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 361                 return;
 362
 363         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 364         elapsed_us[ddir] += utime_since_now(&td->start);
 365         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 366 }
 367
 368 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 369                                 int *retptr)
 370 {
 371         int ret = *retptr;
 372
 373         if (ret < 0 || td->error) {
 374                 int err = td->error;
 375                 enum error_type_bit eb;
 376
 377                 if (ret < 0)
 378                         err = -ret;
 379
 380                 eb = td_error_type(ddir, err);
 381                 if (!(td->o.continue_on_error & (1 << eb)))
 382                         return true;
 383
 384                 if (td_non_fatal_error(td, eb, err)) {
 385                         /*
 386                          * Continue with the I/Os in case of
 387                          * a non fatal error.
 388                          */
 389                         update_error_count(td, err);
 390                         td_clear_error(td);
 391                         *retptr = 0;
 392                         return false;
 393                 } else if (td->o.fill_device && err == ENOSPC) {
 394                         /*
 395                          * We expect to hit this error if
 396                          * fill_device option is set.
 397                          */
 398                         td_clear_error(td);
 399                         fio_mark_td_terminate(td);
 400                         return true;
 401                 } else {
 402                         /*
 403                          * Stop the I/O in case of a fatal
 404                          * error.
 405                          */
 406                         update_error_count(td, err);
 407                         return true;
 408                 }
 409         }
 410
 411         return false;
 412 }
 413
 414 static void check_update_rusage(struct thread_data *td)
 415 {
 416         if (td->update_rusage) {
 417                 td->update_rusage = 0;
 418                 update_rusage_stat(td);
 419                 fio_sem_up(td->rusage_sem);
 420         }
 421 }
 422
 423 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 424 {
 425         const int full = queue_full(td);
 426         int min_evts = 0;
 427         int ret;
 428
 429         if (td->flags & TD_F_REGROW_LOGS)
 430                 return io_u_quiesce(td);
 431
 432         /*
 433          * if the queue is full, we MUST reap at least 1 event
 434          */
 435         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 436         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 437                 min_evts = 1;
 438
 439         if (time && __should_check_rate(td))
 440                 fio_gettime(time, NULL);
 441
 442         do {
 443                 ret = io_u_queued_complete(td, min_evts);
 444                 if (ret < 0)
 445                         break;
 446         } while (full && (td->cur_depth > td->o.iodepth_low));
 447
 448         return ret;
 449 }
 450
 451 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 452                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 453                    struct timespec *comp_time)
 454 {
 455         switch (*ret) {
 456         case FIO_Q_COMPLETED:
 457                 if (io_u->error) {
 458                         *ret = -io_u->error;
 459                         clear_io_u(td, io_u);
 460                 } else if (io_u->resid) {
 461                         long long bytes = io_u->xfer_buflen - io_u->resid;
 462                         struct fio_file *f = io_u->file;
 463
 464                         if (bytes_issued)
 465                                 *bytes_issued += bytes;
 466
 467                         if (!from_verify)
 468                                 trim_io_piece(io_u);
 469
 470                         /*
 471                          * zero read, fail
 472                          */
 473                         if (!bytes) {
 474                                 if (!from_verify)
 475                                         unlog_io_piece(td, io_u);
 476                                 td_verror(td, EIO, "full resid");
 477                                 put_io_u(td, io_u);
 478                                 break;
 479                         }
 480
 481                         io_u->xfer_buflen = io_u->resid;
 482                         io_u->xfer_buf += bytes;
 483                         io_u->offset += bytes;
 484
 485                         if (ddir_rw(io_u->ddir))
 486                                 td->ts.short_io_u[io_u->ddir]++;
 487
 488                         if (io_u->offset == f->real_file_size)
 489                                 goto sync_done;
 490
 491                         requeue_io_u(td, &io_u);
 492                 } else {
 493 sync_done:
 494                         if (comp_time && __should_check_rate(td))
 495                                 fio_gettime(comp_time, NULL);
 496
 497                         *ret = io_u_sync_complete(td, io_u);
 498                         if (*ret < 0)
 499                                 break;
 500                 }
 501
 502                 if (td->flags & TD_F_REGROW_LOGS)
 503                         regrow_logs(td);
 504
 505                 /*
 506                  * when doing I/O (not when verifying),
 507                  * check for any errors that are to be ignored
 508                  */
 509                 if (!from_verify)
 510                         break;
 511
 512                 return 0;
 513         case FIO_Q_QUEUED:
 514                 /*
 515                  * if the engine doesn't have a commit hook,
 516                  * the io_u is really queued. if it does have such
 517                  * a hook, it has to call io_u_queued() itself.
 518                  */
 519                 if (td->io_ops->commit == NULL)
 520                         io_u_queued(td, io_u);
 521                 if (bytes_issued)
 522                         *bytes_issued += io_u->xfer_buflen;
 523                 break;
 524         case FIO_Q_BUSY:
 525                 if (!from_verify)
 526                         unlog_io_piece(td, io_u);
 527                 requeue_io_u(td, &io_u);
 528                 td_io_commit(td);
 529                 break;
 530         default:
 531                 assert(*ret < 0);
 532                 td_verror(td, -(*ret), "td_io_queue");
 533                 break;
 534         }
 535
 536         if (break_on_this_error(td, ddir, ret))
 537                 return 1;
 538
 539         return 0;
 540 }
 541
 542 static inline bool io_in_polling(struct thread_data *td)
 543 {
 544         return !td->o.iodepth_batch_complete_min &&
 545                    !td->o.iodepth_batch_complete_max;
 546 }
 547 /*
 548  * Unlinks files from thread data fio_file structure
 549  */
 550 static int unlink_all_files(struct thread_data *td)
 551 {
 552         struct fio_file *f;
 553         unsigned int i;
 554         int ret = 0;
 555
 556         for_each_file(td, f, i) {
 557                 if (f->filetype != FIO_TYPE_FILE)
 558                         continue;
 559                 ret = td_io_unlink_file(td, f);
 560                 if (ret)
 561                         break;
 562         }
 563
 564         if (ret)
 565                 td_verror(td, ret, "unlink_all_files");
 566
 567         return ret;
 568 }
 569
 570 /*
 571  * Check if io_u will overlap an in-flight IO in the queue
 572  */
 573 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 574 {
 575         bool overlap;
 576         struct io_u *check_io_u;
 577         unsigned long long x1, x2, y1, y2;
 578         int i;
 579
 580         x1 = io_u->offset;
 581         x2 = io_u->offset + io_u->buflen;
 582         overlap = false;
 583         io_u_qiter(q, check_io_u, i) {
 584                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 585                         y1 = check_io_u->offset;
 586                         y2 = check_io_u->offset + check_io_u->buflen;
 587
 588                         if (x1 < y2 && y1 < x2) {
 589                                 overlap = true;
 590                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 591                                                 x1, io_u->buflen,
 592                                                 y1, check_io_u->buflen);
 593                                 break;
 594                         }
 595                 }
 596         }
 597
 598         return overlap;
 599 }
 600
 601 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 602 {
 603         /*
 604          * Check for overlap if the user asked us to, and we have
 605          * at least one IO in flight besides this one.
 606          */
 607         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 608             in_flight_overlap(&td->io_u_all, io_u))
 609                 return FIO_Q_BUSY;
 610
 611         return td_io_queue(td, io_u);
 612 }
 613
 614 /*
 615  * The main verify engine. Runs over the writes we previously submitted,
 616  * reads the blocks back in, and checks the crc/md5 of the data.
 617  */
 618 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 619 {
 620         struct fio_file *f;
 621         struct io_u *io_u;
 622         int ret, min_events;
 623         unsigned int i;
 624
 625         dprint(FD_VERIFY, "starting loop\n");
 626
 627         /*
 628          * sync io first and invalidate cache, to make sure we really
 629          * read from disk.
 630          */
 631         for_each_file(td, f, i) {
 632                 if (!fio_file_open(f))
 633                         continue;
 634                 if (fio_io_sync(td, f))
 635                         break;
 636                 if (file_invalidate_cache(td, f))
 637                         break;
 638         }
 639
 640         check_update_rusage(td);
 641
 642         if (td->error)
 643                 return;
 644
 645         /*
 646          * verify_state needs to be reset before verification
 647          * proceeds so that expected random seeds match actual
 648          * random seeds in headers. The main loop will reset
 649          * all random number generators if randrepeat is set.
 650          */
 651         if (!td->o.rand_repeatable)
 652                 td_fill_verify_state_seed(td);
 653
 654         td_set_runstate(td, TD_VERIFYING);
 655
 656         io_u = NULL;
 657         while (!td->terminate) {
 658                 enum fio_ddir ddir;
 659                 int full;
 660
 661                 update_ts_cache(td);
 662                 check_update_rusage(td);
 663
 664                 if (runtime_exceeded(td, &td->ts_cache)) {
 665                         __update_ts_cache(td);
 666                         if (runtime_exceeded(td, &td->ts_cache)) {
 667                                 fio_mark_td_terminate(td);
 668                                 break;
 669                         }
 670                 }
 671
 672                 if (flow_threshold_exceeded(td))
 673                         continue;
 674
 675                 if (!td->o.experimental_verify) {
 676                         io_u = __get_io_u(td);
 677                         if (!io_u)
 678                                 break;
 679
 680                         if (get_next_verify(td, io_u)) {
 681                                 put_io_u(td, io_u);
 682                                 break;
 683                         }
 684
 685                         if (td_io_prep(td, io_u)) {
 686                                 put_io_u(td, io_u);
 687                                 break;
 688                         }
 689                 } else {
 690                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 691                                 break;
 692
 693                         while ((io_u = get_io_u(td)) != NULL) {
 694                                 if (IS_ERR_OR_NULL(io_u)) {
 695                                         io_u = NULL;
 696                                         ret = FIO_Q_BUSY;
 697                                         goto reap;
 698                                 }
 699
 700                                 /*
 701                                  * We are only interested in the places where
 702                                  * we wrote or trimmed IOs. Turn those into
 703                                  * reads for verification purposes.
 704                                  */
 705                                 if (io_u->ddir == DDIR_READ) {
 706                                         /*
 707                                          * Pretend we issued it for rwmix
 708                                          * accounting
 709                                          */
 710                                         td->io_issues[DDIR_READ]++;
 711                                         put_io_u(td, io_u);
 712                                         continue;
 713                                 } else if (io_u->ddir == DDIR_TRIM) {
 714                                         io_u->ddir = DDIR_READ;
 715                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 716                                         break;
 717                                 } else if (io_u->ddir == DDIR_WRITE) {
 718                                         io_u->ddir = DDIR_READ;
 719                                         populate_verify_io_u(td, io_u);
 720                                         break;
 721                                 } else {
 722                                         put_io_u(td, io_u);
 723                                         continue;
 724                                 }
 725                         }
 726
 727                         if (!io_u)
 728                                 break;
 729                 }
 730
 731                 if (verify_state_should_stop(td, io_u)) {
 732                         put_io_u(td, io_u);
 733                         break;
 734                 }
 735
 736                 if (td->o.verify_async)
 737                         io_u->end_io = verify_io_u_async;
 738                 else
 739                         io_u->end_io = verify_io_u;
 740
 741                 ddir = io_u->ddir;
 742                 if (!td->o.disable_slat)
 743                         fio_gettime(&io_u->start_time, NULL);
 744
 745                 ret = io_u_submit(td, io_u);
 746
 747                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 748                         break;
 749
 750                 /*
 751                  * if we can queue more, do so. but check if there are
 752                  * completed io_u's first. Note that we can get BUSY even
 753                  * without IO queued, if the system is resource starved.
 754                  */
 755 reap:
 756                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 757                 if (full || io_in_polling(td))
 758                         ret = wait_for_completions(td, NULL);
 759
 760                 if (ret < 0)
 761                         break;
 762         }
 763
 764         check_update_rusage(td);
 765
 766         if (!td->error) {
 767                 min_events = td->cur_depth;
 768
 769                 if (min_events)
 770                         ret = io_u_queued_complete(td, min_events);
 771         } else
 772                 cleanup_pending_aio(td);
 773
 774         td_set_runstate(td, TD_RUNNING);
 775
 776         dprint(FD_VERIFY, "exiting loop\n");
 777 }
 778
 779 static bool exceeds_number_ios(struct thread_data *td)
 780 {
 781         unsigned long long number_ios;
 782
 783         if (!td->o.number_ios)
 784                 return false;
 785
 786         number_ios = ddir_rw_sum(td->io_blocks);
 787         number_ios += td->io_u_queued + td->io_u_in_flight;
 788
 789         return number_ios >= (td->o.number_ios * td->loops);
 790 }
 791
 792 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 793 {
 794         unsigned long long bytes, limit;
 795
 796         if (td_rw(td))
 797                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 798         else if (td_write(td))
 799                 bytes = this_bytes[DDIR_WRITE];
 800         else if (td_read(td))
 801                 bytes = this_bytes[DDIR_READ];
 802         else
 803                 bytes = this_bytes[DDIR_TRIM];
 804
 805         if (td->o.io_size)
 806                 limit = td->o.io_size;
 807         else
 808                 limit = td->o.size;
 809
 810         limit *= td->loops;
 811         return bytes >= limit || exceeds_number_ios(td);
 812 }
 813
 814 static bool io_issue_bytes_exceeded(struct thread_data *td)
 815 {
 816         return io_bytes_exceeded(td, td->io_issue_bytes);
 817 }
 818
 819 static bool io_complete_bytes_exceeded(struct thread_data *td)
 820 {
 821         return io_bytes_exceeded(td, td->this_io_bytes);
 822 }
 823
 824 /*
 825  * used to calculate the next io time for rate control
 826  *
 827  */
 828 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 829 {
 830         uint64_t bps = td->rate_bps[ddir];
 831
 832         assert(!(td->flags & TD_F_CHILD));
 833
 834         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 835                 uint64_t val, iops;
 836
 837                 iops = bps / td->o.bs[ddir];
 838                 val = (int64_t) (1000000 / iops) *
 839                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 840                 if (val) {
 841                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 842                                         (unsigned long long) 1000000 / val,
 843                                         ddir);
 844                 }
 845                 td->last_usec[ddir] += val;
 846                 return td->last_usec[ddir];
 847         } else if (bps) {
 848                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 849                 uint64_t secs = bytes / bps;
 850                 uint64_t remainder = bytes % bps;
 851
 852                 return remainder * 1000000 / bps + secs * 1000000;
 853         }
 854
 855         return 0;
 856 }
 857
 858 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 859 {
 860         unsigned long long b;
 861         uint64_t total;
 862         int left;
 863
 864         b = ddir_rw_sum(td->io_blocks);
 865         if (b % td->o.thinktime_blocks)
 866                 return;
 867
 868         io_u_quiesce(td);
 869
 870         total = 0;
 871         if (td->o.thinktime_spin)
 872                 total = usec_spin(td->o.thinktime_spin);
 873
 874         left = td->o.thinktime - total;
 875         if (left)
 876                 total += usec_sleep(td, left);
 877
 878         /*
 879          * If we're ignoring thinktime for the rate, add the number of bytes
 880          * we would have done while sleeping, minus one block to ensure we
 881          * start issuing immediately after the sleep.
 882          */
 883         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 884                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 885                 uint64_t bs = td->o.min_bs[ddir];
 886                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 887                 uint64_t over;
 888
 889                 if (usperop <= total)
 890                         over = bs;
 891                 else
 892                         over = (usperop - total) / usperop * -bs;
 893
 894                 td->rate_io_issue_bytes[ddir] += (missed - over);
 895                 /* adjust for rate_process=poisson */
 896                 td->last_usec[ddir] += total;
 897         }
 898 }
 899
 900 /*
 901  * Main IO worker function. It retrieves io_u's to process and queues
 902  * and reaps them, checking for rate and errors along the way.
 903  *
 904  * Returns number of bytes written and trimmed.
 905  */
 906 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 907 {
 908         unsigned int i;
 909         int ret = 0;
 910         uint64_t total_bytes, bytes_issued = 0;
 911
 912         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 913                 bytes_done[i] = td->bytes_done[i];
 914
 915         if (in_ramp_time(td))
 916                 td_set_runstate(td, TD_RAMP);
 917         else
 918                 td_set_runstate(td, TD_RUNNING);
 919
 920         lat_target_init(td);
 921
 922         total_bytes = td->o.size;
 923         /*
 924         * Allow random overwrite workloads to write up to io_size
 925         * before starting verification phase as 'size' doesn't apply.
 926         */
 927         if (td_write(td) && td_random(td) && td->o.norandommap)
 928                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 929         /*
 930          * If verify_backlog is enabled, we'll run the verify in this
 931          * handler as well. For that case, we may need up to twice the
 932          * amount of bytes.
 933          */
 934         if (td->o.verify != VERIFY_NONE &&
 935            (td_write(td) && td->o.verify_backlog))
 936                 total_bytes += td->o.size;
 937
 938         /* In trimwrite mode, each byte is trimmed and then written, so
 939          * allow total_bytes to be twice as big */
 940         if (td_trimwrite(td))
 941                 total_bytes += td->total_io_size;
 942
 943         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 944                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 945                 td->o.time_based) {
 946                 struct timespec comp_time;
 947                 struct io_u *io_u;
 948                 int full;
 949                 enum fio_ddir ddir;
 950
 951                 check_update_rusage(td);
 952
 953                 if (td->terminate || td->done)
 954                         break;
 955
 956                 update_ts_cache(td);
 957
 958                 if (runtime_exceeded(td, &td->ts_cache)) {
 959                         __update_ts_cache(td);
 960                         if (runtime_exceeded(td, &td->ts_cache)) {
 961                                 fio_mark_td_terminate(td);
 962                                 break;
 963                         }
 964                 }
 965
 966                 if (flow_threshold_exceeded(td))
 967                         continue;
 968
 969                 /*
 970                  * Break if we exceeded the bytes. The exception is time
 971                  * based runs, but we still need to break out of the loop
 972                  * for those to run verification, if enabled.
 973                  * Jobs read from iolog do not use this stop condition.
 974                  */
 975                 if (bytes_issued >= total_bytes &&
 976                     !td->o.read_iolog_file &&
 977                     (!td->o.time_based ||
 978                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 979                         break;
 980
 981                 io_u = get_io_u(td);
 982                 if (IS_ERR_OR_NULL(io_u)) {
 983                         int err = PTR_ERR(io_u);
 984
 985                         io_u = NULL;
 986                         ddir = DDIR_INVAL;
 987                         if (err == -EBUSY) {
 988                                 ret = FIO_Q_BUSY;
 989                                 goto reap;
 990                         }
 991                         if (td->o.latency_target)
 992                                 goto reap;
 993                         break;
 994                 }
 995
 996                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
 997                         populate_verify_io_u(td, io_u);
 998
 999                 ddir = io_u->ddir;
1000
1001                 /*
1002                  * Add verification end_io handler if:
1003                  *      - Asked to verify (!td_rw(td))
1004                  *      - Or the io_u is from our verify list (mixed write/ver)
1005                  */
1006                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1007                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1008
1009                         if (verify_state_should_stop(td, io_u)) {
1010                                 put_io_u(td, io_u);
1011                                 break;
1012                         }
1013
1014                         if (td->o.verify_async)
1015                                 io_u->end_io = verify_io_u_async;
1016                         else
1017                                 io_u->end_io = verify_io_u;
1018                         td_set_runstate(td, TD_VERIFYING);
1019                 } else if (in_ramp_time(td))
1020                         td_set_runstate(td, TD_RAMP);
1021                 else
1022                         td_set_runstate(td, TD_RUNNING);
1023
1024                 /*
1025                  * Always log IO before it's issued, so we know the specific
1026                  * order of it. The logged unit will track when the IO has
1027                  * completed.
1028                  */
1029                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1030                     td->o.do_verify &&
1031                     td->o.verify != VERIFY_NONE &&
1032                     !td->o.experimental_verify)
1033                         log_io_piece(td, io_u);
1034
1035                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1036                         const unsigned long long blen = io_u->xfer_buflen;
1037                         const enum fio_ddir __ddir = acct_ddir(io_u);
1038
1039                         if (td->error)
1040                                 break;
1041
1042                         workqueue_enqueue(&td->io_wq, &io_u->work);
1043                         ret = FIO_Q_QUEUED;
1044
1045                         if (ddir_rw(__ddir)) {
1046                                 td->io_issues[__ddir]++;
1047                                 td->io_issue_bytes[__ddir] += blen;
1048                                 td->rate_io_issue_bytes[__ddir] += blen;
1049                         }
1050
1051                         if (should_check_rate(td))
1052                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1053
1054                 } else {
1055                         ret = io_u_submit(td, io_u);
1056
1057                         if (should_check_rate(td))
1058                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1059
1060                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1061                                 break;
1062
1063                         /*
1064                          * See if we need to complete some commands. Note that
1065                          * we can get BUSY even without IO queued, if the
1066                          * system is resource starved.
1067                          */
1068 reap:
1069                         full = queue_full(td) ||
1070                                 (ret == FIO_Q_BUSY && td->cur_depth);
1071                         if (full || io_in_polling(td))
1072                                 ret = wait_for_completions(td, &comp_time);
1073                 }
1074                 if (ret < 0)
1075                         break;
1076                 if (!ddir_rw_sum(td->bytes_done) &&
1077                     !td_ioengine_flagged(td, FIO_NOIO))
1078                         continue;
1079
1080                 if (!in_ramp_time(td) && should_check_rate(td)) {
1081                         if (check_min_rate(td, &comp_time)) {
1082                                 if (exitall_on_terminate || td->o.exitall_error)
1083                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1084                                 td_verror(td, EIO, "check_min_rate");
1085                                 break;
1086                         }
1087                 }
1088                 if (!in_ramp_time(td) && td->o.latency_target)
1089                         lat_target_check(td);
1090
1091                 if (ddir_rw(ddir) && td->o.thinktime)
1092                         handle_thinktime(td, ddir);
1093         }
1094
1095         check_update_rusage(td);
1096
1097         if (td->trim_entries)
1098                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1099
1100         if (td->o.fill_device && td->error == ENOSPC) {
1101                 td->error = 0;
1102                 fio_mark_td_terminate(td);
1103         }
1104         if (!td->error) {
1105                 struct fio_file *f;
1106
1107                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1108                         workqueue_flush(&td->io_wq);
1109                         i = 0;
1110                 } else
1111                         i = td->cur_depth;
1112
1113                 if (i) {
1114                         ret = io_u_queued_complete(td, i);
1115                         if (td->o.fill_device && td->error == ENOSPC)
1116                                 td->error = 0;
1117                 }
1118
1119                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1120                         td_set_runstate(td, TD_FSYNCING);
1121
1122                         for_each_file(td, f, i) {
1123                                 if (!fio_file_fsync(td, f))
1124                                         continue;
1125
1126                                 log_err("fio: end_fsync failed for file %s\n",
1127                                                                 f->file_name);
1128                         }
1129                 }
1130         } else
1131                 cleanup_pending_aio(td);
1132
1133         /*
1134          * stop job if we failed doing any IO
1135          */
1136         if (!ddir_rw_sum(td->this_io_bytes))
1137                 td->done = 1;
1138
1139         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1140                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1141 }
1142
1143 static void free_file_completion_logging(struct thread_data *td)
1144 {
1145         struct fio_file *f;
1146         unsigned int i;
1147
1148         for_each_file(td, f, i) {
1149                 if (!f->last_write_comp)
1150                         break;
1151                 sfree(f->last_write_comp);
1152         }
1153 }
1154
1155 static int init_file_completion_logging(struct thread_data *td,
1156                                         unsigned int depth)
1157 {
1158         struct fio_file *f;
1159         unsigned int i;
1160
1161         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1162                 return 0;
1163
1164         for_each_file(td, f, i) {
1165                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1166                 if (!f->last_write_comp)
1167                         goto cleanup;
1168         }
1169
1170         return 0;
1171
1172 cleanup:
1173         free_file_completion_logging(td);
1174         log_err("fio: failed to alloc write comp data\n");
1175         return 1;
1176 }
1177
1178 static void cleanup_io_u(struct thread_data *td)
1179 {
1180         struct io_u *io_u;
1181
1182         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1183
1184                 if (td->io_ops->io_u_free)
1185                         td->io_ops->io_u_free(td, io_u);
1186
1187                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1188         }
1189
1190         free_io_mem(td);
1191
1192         io_u_rexit(&td->io_u_requeues);
1193         io_u_qexit(&td->io_u_freelist, false);
1194         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1195
1196         free_file_completion_logging(td);
1197 }
1198
1199 static int init_io_u(struct thread_data *td)
1200 {
1201         struct io_u *io_u;
1202         int cl_align, i, max_units;
1203         int err;
1204
1205         max_units = td->o.iodepth;
1206
1207         err = 0;
1208         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1209         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1210         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1211
1212         if (err) {
1213                 log_err("fio: failed setting up IO queues\n");
1214                 return 1;
1215         }
1216
1217         cl_align = os_cache_line_size();
1218
1219         for (i = 0; i < max_units; i++) {
1220                 void *ptr;
1221
1222                 if (td->terminate)
1223                         return 1;
1224
1225                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1226                 if (!ptr) {
1227                         log_err("fio: unable to allocate aligned memory\n");
1228                         return 1;
1229                 }
1230
1231                 io_u = ptr;
1232                 memset(io_u, 0, sizeof(*io_u));
1233                 INIT_FLIST_HEAD(&io_u->verify_list);
1234                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1235
1236                 io_u->index = i;
1237                 io_u->flags = IO_U_F_FREE;
1238                 io_u_qpush(&td->io_u_freelist, io_u);
1239
1240                 /*
1241                  * io_u never leaves this stack, used for iteration of all
1242                  * io_u buffers.
1243                  */
1244                 io_u_qpush(&td->io_u_all, io_u);
1245
1246                 if (td->io_ops->io_u_init) {
1247                         int ret = td->io_ops->io_u_init(td, io_u);
1248
1249                         if (ret) {
1250                                 log_err("fio: failed to init engine data: %d\n", ret);
1251                                 return 1;
1252                         }
1253                 }
1254         }
1255
1256         init_io_u_buffers(td);
1257
1258         if (init_file_completion_logging(td, max_units))
1259                 return 1;
1260
1261         return 0;
1262 }
1263
1264 int init_io_u_buffers(struct thread_data *td)
1265 {
1266         struct io_u *io_u;
1267         unsigned long long max_bs, min_write;
1268         int i, max_units;
1269         int data_xfer = 1;
1270         char *p;
1271
1272         max_units = td->o.iodepth;
1273         max_bs = td_max_bs(td);
1274         min_write = td->o.min_bs[DDIR_WRITE];
1275         td->orig_buffer_size = (unsigned long long) max_bs
1276                                         * (unsigned long long) max_units;
1277
1278         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1279                 data_xfer = 0;
1280
1281         /*
1282          * if we may later need to do address alignment, then add any
1283          * possible adjustment here so that we don't cause a buffer
1284          * overflow later. this adjustment may be too much if we get
1285          * lucky and the allocator gives us an aligned address.
1286          */
1287         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1288             td_ioengine_flagged(td, FIO_RAWIO))
1289                 td->orig_buffer_size += page_mask + td->o.mem_align;
1290
1291         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1292                 unsigned long long bs;
1293
1294                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1295                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1296         }
1297
1298         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1299                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1300                 return 1;
1301         }
1302
1303         if (data_xfer && allocate_io_mem(td))
1304                 return 1;
1305
1306         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1307             td_ioengine_flagged(td, FIO_RAWIO))
1308                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1309         else
1310                 p = td->orig_buffer;
1311
1312         for (i = 0; i < max_units; i++) {
1313                 io_u = td->io_u_all.io_us[i];
1314                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1315
1316                 if (data_xfer) {
1317                         io_u->buf = p;
1318                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1319
1320                         if (td_write(td))
1321                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1322                         if (td_write(td) && td->o.verify_pattern_bytes) {
1323                                 /*
1324                                  * Fill the buffer with the pattern if we are
1325                                  * going to be doing writes.
1326                                  */
1327                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1328                         }
1329                 }
1330                 p += max_bs;
1331         }
1332
1333         return 0;
1334 }
1335
1336 /*
1337  * This function is Linux specific.
1338  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1339  */
1340 static int switch_ioscheduler(struct thread_data *td)
1341 {
1342 #ifdef FIO_HAVE_IOSCHED_SWITCH
1343         char tmp[256], tmp2[128], *p;
1344         FILE *f;
1345         int ret;
1346
1347         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1348                 return 0;
1349
1350         assert(td->files && td->files[0]);
1351         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1352
1353         f = fopen(tmp, "r+");
1354         if (!f) {
1355                 if (errno == ENOENT) {
1356                         log_err("fio: os or kernel doesn't support IO scheduler"
1357                                 " switching\n");
1358                         return 0;
1359                 }
1360                 td_verror(td, errno, "fopen iosched");
1361                 return 1;
1362         }
1363
1364         /*
1365          * Set io scheduler.
1366          */
1367         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1368         if (ferror(f) || ret != 1) {
1369                 td_verror(td, errno, "fwrite");
1370                 fclose(f);
1371                 return 1;
1372         }
1373
1374         rewind(f);
1375
1376         /*
1377          * Read back and check that the selected scheduler is now the default.
1378          */
1379         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1380         if (ferror(f) || ret < 0) {
1381                 td_verror(td, errno, "fread");
1382                 fclose(f);
1383                 return 1;
1384         }
1385         tmp[ret] = '\0';
1386         /*
1387          * either a list of io schedulers or "none\n" is expected. Strip the
1388          * trailing newline.
1389          */
1390         p = tmp;
1391         strsep(&p, "\n");
1392
1393         /*
1394          * Write to "none" entry doesn't fail, so check the result here.
1395          */
1396         if (!strcmp(tmp, "none")) {
1397                 log_err("fio: io scheduler is not tunable\n");
1398                 fclose(f);
1399                 return 0;
1400         }
1401
1402         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1403         if (!strstr(tmp, tmp2)) {
1404                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1405                 td_verror(td, EINVAL, "iosched_switch");
1406                 fclose(f);
1407                 return 1;
1408         }
1409
1410         fclose(f);
1411         return 0;
1412 #else
1413         return 0;
1414 #endif
1415 }
1416
1417 static bool keep_running(struct thread_data *td)
1418 {
1419         unsigned long long limit;
1420
1421         if (td->done)
1422                 return false;
1423         if (td->terminate)
1424                 return false;
1425         if (td->o.time_based)
1426                 return true;
1427         if (td->o.loops) {
1428                 td->o.loops--;
1429                 return true;
1430         }
1431         if (exceeds_number_ios(td))
1432                 return false;
1433
1434         if (td->o.io_size)
1435                 limit = td->o.io_size;
1436         else
1437                 limit = td->o.size;
1438
1439         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1440                 uint64_t diff;
1441
1442                 /*
1443                  * If the difference is less than the maximum IO size, we
1444                  * are done.
1445                  */
1446                 diff = limit - ddir_rw_sum(td->io_bytes);
1447                 if (diff < td_max_bs(td))
1448                         return false;
1449
1450                 if (fio_files_done(td) && !td->o.io_size)
1451                         return false;
1452
1453                 return true;
1454         }
1455
1456         return false;
1457 }
1458
1459 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1460 {
1461         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 13 + 1;
1462         int ret;
1463         char *str;
1464
1465         str = malloc(newlen);
1466         sprintf(str, "%s > %s.%s.txt 2>&1", string, o->name, mode);
1467
1468         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1469         ret = system(str);
1470         if (ret == -1)
1471                 log_err("fio: exec of cmd <%s> failed\n", str);
1472
1473         free(str);
1474         return ret;
1475 }
1476
1477 /*
1478  * Dry run to compute correct state of numberio for verification.
1479  */
1480 static uint64_t do_dry_run(struct thread_data *td)
1481 {
1482         td_set_runstate(td, TD_RUNNING);
1483
1484         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1485                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1486                 struct io_u *io_u;
1487                 int ret;
1488
1489                 if (td->terminate || td->done)
1490                         break;
1491
1492                 io_u = get_io_u(td);
1493                 if (IS_ERR_OR_NULL(io_u))
1494                         break;
1495
1496                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1497                 io_u->error = 0;
1498                 io_u->resid = 0;
1499                 if (ddir_rw(acct_ddir(io_u)))
1500                         td->io_issues[acct_ddir(io_u)]++;
1501                 if (ddir_rw(io_u->ddir)) {
1502                         io_u_mark_depth(td, 1);
1503                         td->ts.total_io_u[io_u->ddir]++;
1504                 }
1505
1506                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1507                     td->o.do_verify &&
1508                     td->o.verify != VERIFY_NONE &&
1509                     !td->o.experimental_verify)
1510                         log_io_piece(td, io_u);
1511
1512                 ret = io_u_sync_complete(td, io_u);
1513                 (void) ret;
1514         }
1515
1516         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1517 }
1518
1519 struct fork_data {
1520         struct thread_data *td;
1521         struct sk_out *sk_out;
1522 };
1523
1524 /*
1525  * Entry point for the thread based jobs. The process based jobs end up
1526  * here as well, after a little setup.
1527  */
1528 static void *thread_main(void *data)
1529 {
1530         struct fork_data *fd = data;
1531         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1532         struct thread_data *td = fd->td;
1533         struct thread_options *o = &td->o;
1534         struct sk_out *sk_out = fd->sk_out;
1535         uint64_t bytes_done[DDIR_RWDIR_CNT];
1536         int deadlock_loop_cnt;
1537         bool clear_state;
1538         int ret;
1539
1540         sk_out_assign(sk_out);
1541         free(fd);
1542
1543         if (!o->use_thread) {
1544                 setsid();
1545                 td->pid = getpid();
1546         } else
1547                 td->pid = gettid();
1548
1549         fio_local_clock_init();
1550
1551         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1552
1553         if (is_backend)
1554                 fio_server_send_start(td);
1555
1556         INIT_FLIST_HEAD(&td->io_log_list);
1557         INIT_FLIST_HEAD(&td->io_hist_list);
1558         INIT_FLIST_HEAD(&td->verify_list);
1559         INIT_FLIST_HEAD(&td->trim_list);
1560         td->io_hist_tree = RB_ROOT;
1561
1562         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1563         if (ret) {
1564                 td_verror(td, ret, "mutex_cond_init_pshared");
1565                 goto err;
1566         }
1567         ret = cond_init_pshared(&td->verify_cond);
1568         if (ret) {
1569                 td_verror(td, ret, "mutex_cond_pshared");
1570                 goto err;
1571         }
1572
1573         td_set_runstate(td, TD_INITIALIZED);
1574         dprint(FD_MUTEX, "up startup_sem\n");
1575         fio_sem_up(startup_sem);
1576         dprint(FD_MUTEX, "wait on td->sem\n");
1577         fio_sem_down(td->sem);
1578         dprint(FD_MUTEX, "done waiting on td->sem\n");
1579
1580         /*
1581          * A new gid requires privilege, so we need to do this before setting
1582          * the uid.
1583          */
1584         if (o->gid != -1U && setgid(o->gid)) {
1585                 td_verror(td, errno, "setgid");
1586                 goto err;
1587         }
1588         if (o->uid != -1U && setuid(o->uid)) {
1589                 td_verror(td, errno, "setuid");
1590                 goto err;
1591         }
1592
1593         td_zone_gen_index(td);
1594
1595         /*
1596          * Do this early, we don't want the compress threads to be limited
1597          * to the same CPUs as the IO workers. So do this before we set
1598          * any potential CPU affinity
1599          */
1600         if (iolog_compress_init(td, sk_out))
1601                 goto err;
1602
1603         /*
1604          * If we have a gettimeofday() thread, make sure we exclude that
1605          * thread from this job
1606          */
1607         if (o->gtod_cpu)
1608                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1609
1610         /*
1611          * Set affinity first, in case it has an impact on the memory
1612          * allocations.
1613          */
1614         if (fio_option_is_set(o, cpumask)) {
1615                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1616                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1617                         if (!ret) {
1618                                 log_err("fio: no CPUs set\n");
1619                                 log_err("fio: Try increasing number of available CPUs\n");
1620                                 td_verror(td, EINVAL, "cpus_split");
1621                                 goto err;
1622                         }
1623                 }
1624                 ret = fio_setaffinity(td->pid, o->cpumask);
1625                 if (ret == -1) {
1626                         td_verror(td, errno, "cpu_set_affinity");
1627                         goto err;
1628                 }
1629         }
1630
1631 #ifdef CONFIG_LIBNUMA
1632         /* numa node setup */
1633         if (fio_option_is_set(o, numa_cpunodes) ||
1634             fio_option_is_set(o, numa_memnodes)) {
1635                 struct bitmask *mask;
1636
1637                 if (numa_available() < 0) {
1638                         td_verror(td, errno, "Does not support NUMA API\n");
1639                         goto err;
1640                 }
1641
1642                 if (fio_option_is_set(o, numa_cpunodes)) {
1643                         mask = numa_parse_nodestring(o->numa_cpunodes);
1644                         ret = numa_run_on_node_mask(mask);
1645                         numa_free_nodemask(mask);
1646                         if (ret == -1) {
1647                                 td_verror(td, errno, \
1648                                         "numa_run_on_node_mask failed\n");
1649                                 goto err;
1650                         }
1651                 }
1652
1653                 if (fio_option_is_set(o, numa_memnodes)) {
1654                         mask = NULL;
1655                         if (o->numa_memnodes)
1656                                 mask = numa_parse_nodestring(o->numa_memnodes);
1657
1658                         switch (o->numa_mem_mode) {
1659                         case MPOL_INTERLEAVE:
1660                                 numa_set_interleave_mask(mask);
1661                                 break;
1662                         case MPOL_BIND:
1663                                 numa_set_membind(mask);
1664                                 break;
1665                         case MPOL_LOCAL:
1666                                 numa_set_localalloc();
1667                                 break;
1668                         case MPOL_PREFERRED:
1669                                 numa_set_preferred(o->numa_mem_prefer_node);
1670                                 break;
1671                         case MPOL_DEFAULT:
1672                         default:
1673                                 break;
1674                         }
1675
1676                         if (mask)
1677                                 numa_free_nodemask(mask);
1678
1679                 }
1680         }
1681 #endif
1682
1683         if (fio_pin_memory(td))
1684                 goto err;
1685
1686         /*
1687          * May alter parameters that init_io_u() will use, so we need to
1688          * do this first.
1689          */
1690         if (!init_iolog(td))
1691                 goto err;
1692
1693         if (td_io_init(td))
1694                 goto err;
1695
1696         if (init_io_u(td))
1697                 goto err;
1698
1699         if (td->io_ops->post_init && td->io_ops->post_init(td))
1700                 goto err;
1701
1702         if (o->verify_async && verify_async_init(td))
1703                 goto err;
1704
1705         if (fio_option_is_set(o, ioprio) ||
1706             fio_option_is_set(o, ioprio_class)) {
1707                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1708                 if (ret == -1) {
1709                         td_verror(td, errno, "ioprio_set");
1710                         goto err;
1711                 }
1712         }
1713
1714         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1715                 goto err;
1716
1717         errno = 0;
1718         if (nice(o->nice) == -1 && errno != 0) {
1719                 td_verror(td, errno, "nice");
1720                 goto err;
1721         }
1722
1723         if (o->ioscheduler && switch_ioscheduler(td))
1724                 goto err;
1725
1726         if (!o->create_serialize && setup_files(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
1762         while (keep_running(td)) {
1763                 uint64_t verify_bytes;
1764
1765                 fio_gettime(&td->start, NULL);
1766                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1767
1768                 if (clear_state) {
1769                         clear_io_state(td, 0);
1770
1771                         if (o->unlink_each_loop && unlink_all_files(td))
1772                                 break;
1773                 }
1774
1775                 prune_io_piece_log(td);
1776
1777                 if (td->o.verify_only && td_write(td))
1778                         verify_bytes = do_dry_run(td);
1779                 else {
1780                         do_io(td, bytes_done);
1781
1782                         if (!ddir_rw_sum(bytes_done)) {
1783                                 fio_mark_td_terminate(td);
1784                                 verify_bytes = 0;
1785                         } else {
1786                                 verify_bytes = bytes_done[DDIR_WRITE] +
1787                                                 bytes_done[DDIR_TRIM];
1788                         }
1789                 }
1790
1791                 /*
1792                  * If we took too long to shut down, the main thread could
1793                  * already consider us reaped/exited. If that happens, break
1794                  * out and clean up.
1795                  */
1796                 if (td->runstate >= TD_EXITED)
1797                         break;
1798
1799                 clear_state = true;
1800
1801                 /*
1802                  * Make sure we've successfully updated the rusage stats
1803                  * before waiting on the stat mutex. Otherwise we could have
1804                  * the stat thread holding stat mutex and waiting for
1805                  * the rusage_sem, which would never get upped because
1806                  * this thread is waiting for the stat mutex.
1807                  */
1808                 deadlock_loop_cnt = 0;
1809                 do {
1810                         check_update_rusage(td);
1811                         if (!fio_sem_down_trylock(stat_sem))
1812                                 break;
1813                         usleep(1000);
1814                         if (deadlock_loop_cnt++ > 5000) {
1815                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1816                                 td->error = EDEADLK;
1817                                 goto err;
1818                         }
1819                 } while (1);
1820
1821                 if (td_read(td) && td->io_bytes[DDIR_READ])
1822                         update_runtime(td, elapsed_us, DDIR_READ);
1823                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1824                         update_runtime(td, elapsed_us, DDIR_WRITE);
1825                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1826                         update_runtime(td, elapsed_us, DDIR_TRIM);
1827                 fio_gettime(&td->start, NULL);
1828                 fio_sem_up(stat_sem);
1829
1830                 if (td->error || td->terminate)
1831                         break;
1832
1833                 if (!o->do_verify ||
1834                     o->verify == VERIFY_NONE ||
1835                     td_ioengine_flagged(td, FIO_UNIDIR))
1836                         continue;
1837
1838                 clear_io_state(td, 0);
1839
1840                 fio_gettime(&td->start, NULL);
1841
1842                 do_verify(td, verify_bytes);
1843
1844                 /*
1845                  * See comment further up for why this is done here.
1846                  */
1847                 check_update_rusage(td);
1848
1849                 fio_sem_down(stat_sem);
1850                 update_runtime(td, elapsed_us, DDIR_READ);
1851                 fio_gettime(&td->start, NULL);
1852                 fio_sem_up(stat_sem);
1853
1854                 if (td->error || td->terminate)
1855                         break;
1856         }
1857
1858         /*
1859          * Acquire this lock if we were doing overlap checking in
1860          * offload mode so that we don't clean up this job while
1861          * another thread is checking its io_u's for overlap
1862          */
1863         if (td_offload_overlap(td))
1864                 pthread_mutex_lock(&overlap_check);
1865         td_set_runstate(td, TD_FINISHING);
1866         if (td_offload_overlap(td))
1867                 pthread_mutex_unlock(&overlap_check);
1868
1869         update_rusage_stat(td);
1870         td->ts.total_run_time = mtime_since_now(&td->epoch);
1871         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1872         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1873         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1874
1875         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1876             (td->o.verify != VERIFY_NONE && td_write(td)))
1877                 verify_save_state(td->thread_number);
1878
1879         fio_unpin_memory(td);
1880
1881         td_writeout_logs(td, true);
1882
1883         iolog_compress_exit(td);
1884         rate_submit_exit(td);
1885
1886         if (o->exec_postrun)
1887                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1888
1889         if (exitall_on_terminate || (o->exitall_error && td->error))
1890                 fio_terminate_threads(td->groupid, td->o.exit_what);
1891
1892 err:
1893         if (td->error)
1894                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1895                                                         td->verror);
1896
1897         if (o->verify_async)
1898                 verify_async_exit(td);
1899
1900         close_and_free_files(td);
1901         cleanup_io_u(td);
1902         close_ioengine(td);
1903         cgroup_shutdown(td, cgroup_mnt);
1904         verify_free_state(td);
1905         td_zone_free_index(td);
1906
1907         if (fio_option_is_set(o, cpumask)) {
1908                 ret = fio_cpuset_exit(&o->cpumask);
1909                 if (ret)
1910                         td_verror(td, ret, "fio_cpuset_exit");
1911         }
1912
1913         /*
1914          * do this very late, it will log file closing as well
1915          */
1916         if (o->write_iolog_file)
1917                 write_iolog_close(td);
1918         if (td->io_log_rfile)
1919                 fclose(td->io_log_rfile);
1920
1921         td_set_runstate(td, TD_EXITED);
1922
1923         /*
1924          * Do this last after setting our runstate to exited, so we
1925          * know that the stat thread is signaled.
1926          */
1927         check_update_rusage(td);
1928
1929         sk_out_drop();
1930         return (void *) (uintptr_t) td->error;
1931 }
1932
1933 /*
1934  * Run over the job map and reap the threads that have exited, if any.
1935  */
1936 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1937                          uint64_t *m_rate)
1938 {
1939         struct thread_data *td;
1940         unsigned int cputhreads, realthreads, pending;
1941         int i, status, ret;
1942
1943         /*
1944          * reap exited threads (TD_EXITED -> TD_REAPED)
1945          */
1946         realthreads = pending = cputhreads = 0;
1947         for_each_td(td, i) {
1948                 int flags = 0;
1949
1950                  if (!strcmp(td->o.ioengine, "cpuio"))
1951                         cputhreads++;
1952                 else
1953                         realthreads++;
1954
1955                 if (!td->pid) {
1956                         pending++;
1957                         continue;
1958                 }
1959                 if (td->runstate == TD_REAPED)
1960                         continue;
1961                 if (td->o.use_thread) {
1962                         if (td->runstate == TD_EXITED) {
1963                                 td_set_runstate(td, TD_REAPED);
1964                                 goto reaped;
1965                         }
1966                         continue;
1967                 }
1968
1969                 flags = WNOHANG;
1970                 if (td->runstate == TD_EXITED)
1971                         flags = 0;
1972
1973                 /*
1974                  * check if someone quit or got killed in an unusual way
1975                  */
1976                 ret = waitpid(td->pid, &status, flags);
1977                 if (ret < 0) {
1978                         if (errno == ECHILD) {
1979                                 log_err("fio: pid=%d disappeared %d\n",
1980                                                 (int) td->pid, td->runstate);
1981                                 td->sig = ECHILD;
1982                                 td_set_runstate(td, TD_REAPED);
1983                                 goto reaped;
1984                         }
1985                         perror("waitpid");
1986                 } else if (ret == td->pid) {
1987                         if (WIFSIGNALED(status)) {
1988                                 int sig = WTERMSIG(status);
1989
1990                                 if (sig != SIGTERM && sig != SIGUSR2)
1991                                         log_err("fio: pid=%d, got signal=%d\n",
1992                                                         (int) td->pid, sig);
1993                                 td->sig = sig;
1994                                 td_set_runstate(td, TD_REAPED);
1995                                 goto reaped;
1996                         }
1997                         if (WIFEXITED(status)) {
1998                                 if (WEXITSTATUS(status) && !td->error)
1999                                         td->error = WEXITSTATUS(status);
2000
2001                                 td_set_runstate(td, TD_REAPED);
2002                                 goto reaped;
2003                         }
2004                 }
2005
2006                 /*
2007                  * If the job is stuck, do a forceful timeout of it and
2008                  * move on.
2009                  */
2010                 if (td->terminate &&
2011                     td->runstate < TD_FSYNCING &&
2012                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2013                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2014                                 "%lu seconds, it appears to be stuck. Doing "
2015                                 "forceful exit of this job.\n",
2016                                 td->o.name, td->runstate,
2017                                 (unsigned long) time_since_now(&td->terminate_time));
2018                         td_set_runstate(td, TD_REAPED);
2019                         goto reaped;
2020                 }
2021
2022                 /*
2023                  * thread is not dead, continue
2024                  */
2025                 pending++;
2026                 continue;
2027 reaped:
2028                 (*nr_running)--;
2029                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2030                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2031                 if (!td->pid)
2032                         pending--;
2033
2034                 if (td->error)
2035                         exit_value++;
2036
2037                 done_secs += mtime_since_now(&td->epoch) / 1000;
2038                 profile_td_exit(td);
2039         }
2040
2041         if (*nr_running == cputhreads && !pending && realthreads)
2042                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2043 }
2044
2045 static bool __check_trigger_file(void)
2046 {
2047         struct stat sb;
2048
2049         if (!trigger_file)
2050                 return false;
2051
2052         if (stat(trigger_file, &sb))
2053                 return false;
2054
2055         if (unlink(trigger_file) < 0)
2056                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2057                                                         strerror(errno));
2058
2059         return true;
2060 }
2061
2062 static bool trigger_timedout(void)
2063 {
2064         if (trigger_timeout)
2065                 if (time_since_genesis() >= trigger_timeout) {
2066                         trigger_timeout = 0;
2067                         return true;
2068                 }
2069
2070         return false;
2071 }
2072
2073 void exec_trigger(const char *cmd)
2074 {
2075         int ret;
2076
2077         if (!cmd || cmd[0] == '\0')
2078                 return;
2079
2080         ret = system(cmd);
2081         if (ret == -1)
2082                 log_err("fio: failed executing %s trigger\n", cmd);
2083 }
2084
2085 void check_trigger_file(void)
2086 {
2087         if (__check_trigger_file() || trigger_timedout()) {
2088                 if (nr_clients)
2089                         fio_clients_send_trigger(trigger_remote_cmd);
2090                 else {
2091                         verify_save_state(IO_LIST_ALL);
2092                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2093                         exec_trigger(trigger_cmd);
2094                 }
2095         }
2096 }
2097
2098 static int fio_verify_load_state(struct thread_data *td)
2099 {
2100         int ret;
2101
2102         if (!td->o.verify_state)
2103                 return 0;
2104
2105         if (is_backend) {
2106                 void *data;
2107
2108                 ret = fio_server_get_verify_state(td->o.name,
2109                                         td->thread_number - 1, &data);
2110                 if (!ret)
2111                         verify_assign_state(td, data);
2112         } else {
2113                 char prefix[PATH_MAX];
2114
2115                 if (aux_path)
2116                         sprintf(prefix, "%s%clocal", aux_path,
2117                                         FIO_OS_PATH_SEPARATOR);
2118                 else
2119                         strcpy(prefix, "local");
2120                 ret = verify_load_state(td, prefix);
2121         }
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, 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         if (helper_thread_create(startup_sem, sk_out))
2496                 log_err("fio: failed to create helper thread\n");
2497
2498         cgroup_list = smalloc(sizeof(*cgroup_list));
2499         if (cgroup_list)
2500                 INIT_FLIST_HEAD(cgroup_list);
2501
2502         run_threads(sk_out);
2503
2504         helper_thread_exit();
2505
2506         if (!fio_abort) {
2507                 __show_run_stats();
2508                 if (write_bw_log) {
2509                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2510                                 struct io_log *log = agg_io_log[i];
2511
2512                                 flush_log(log, false);
2513                                 free_log(log);
2514                         }
2515                 }
2516         }
2517
2518         for_each_td(td, i) {
2519                 steadystate_free(td);
2520                 fio_options_free(td);
2521                 if (td->rusage_sem) {
2522                         fio_sem_remove(td->rusage_sem);
2523                         td->rusage_sem = NULL;
2524                 }
2525                 fio_sem_remove(td->sem);
2526                 td->sem = NULL;
2527         }
2528
2529         free_disk_util();
2530         if (cgroup_list) {
2531                 cgroup_kill(cgroup_list);
2532                 sfree(cgroup_list);
2533         }
2534
2535         fio_sem_remove(startup_sem);
2536         stat_exit();
2537         return exit_value;
2538 }