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 #include "fio_time.h"
  53
  54 static struct fio_sem *startup_sem;
  55 static struct flist_head *cgroup_list;
  56 static struct cgroup_mnt *cgroup_mnt;
  57 static int exit_value;
  58 static volatile bool fio_abort;
  59 static unsigned int nr_process = 0;
  60 static unsigned int nr_thread = 0;
  61
  62 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  63
  64 int groupid = 0;
  65 unsigned int thread_number = 0;
  66 unsigned int nr_segments = 0;
  67 unsigned int cur_segment = 0;
  68 unsigned int stat_number = 0;
  69 int temp_stall_ts;
  70 unsigned long done_secs = 0;
  71 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
  72 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
  73 #else
  74 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
  75 #endif
  76
  77 #define JOB_START_TIMEOUT       (5 * 1000)
  78
  79 static void sig_int(int sig)
  80 {
  81         if (nr_segments) {
  82                 if (is_backend)
  83                         fio_server_got_signal(sig);
  84                 else {
  85                         log_info("\nfio: terminating on signal %d\n", sig);
  86                         log_info_flush();
  87                         exit_value = 128;
  88                 }
  89
  90                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
  91         }
  92 }
  93
  94 #ifdef WIN32
  95 static void sig_break(int sig)
  96 {
  97         sig_int(sig);
  98
  99         /**
 100          * Windows terminates all job processes on SIGBREAK after the handler
 101          * returns, so give them time to wrap-up and give stats
 102          */
 103         for_each_td(td) {
 104                 while (td->runstate < TD_EXITED)
 105                         sleep(1);
 106         } end_for_each();
 107 }
 108 #endif
 109
 110 void sig_show_status(int sig)
 111 {
 112         show_running_run_stats();
 113 }
 114
 115 static void set_sig_handlers(void)
 116 {
 117         struct sigaction act;
 118
 119         memset(&act, 0, sizeof(act));
 120         act.sa_handler = sig_int;
 121         act.sa_flags = SA_RESTART;
 122         sigaction(SIGINT, &act, NULL);
 123
 124         memset(&act, 0, sizeof(act));
 125         act.sa_handler = sig_int;
 126         act.sa_flags = SA_RESTART;
 127         sigaction(SIGTERM, &act, NULL);
 128
 129 /* Windows uses SIGBREAK as a quit signal from other applications */
 130 #ifdef WIN32
 131         memset(&act, 0, sizeof(act));
 132         act.sa_handler = sig_break;
 133         act.sa_flags = SA_RESTART;
 134         sigaction(SIGBREAK, &act, NULL);
 135 #endif
 136
 137         memset(&act, 0, sizeof(act));
 138         act.sa_handler = sig_show_status;
 139         act.sa_flags = SA_RESTART;
 140         sigaction(SIGUSR1, &act, NULL);
 141
 142         if (is_backend) {
 143                 memset(&act, 0, sizeof(act));
 144                 act.sa_handler = sig_int;
 145                 act.sa_flags = SA_RESTART;
 146                 sigaction(SIGPIPE, &act, NULL);
 147         }
 148 }
 149
 150 /*
 151  * Check if we are above the minimum rate given.
 152  */
 153 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 154                              enum fio_ddir ddir)
 155 {
 156         unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
 157         unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
 158         unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
 159         unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
 160
 161         assert(ddir_rw(ddir));
 162
 163         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 164                 return false;
 165
 166         /*
 167          * allow a 2 second settle period in the beginning
 168          */
 169         if (mtime_since(&td->start, now) < 2000)
 170                 return false;
 171
 172         /*
 173          * if last_rate_check_blocks or last_rate_check_bytes is set,
 174          * we can compute a rate per ratecycle
 175          */
 176         if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
 177                 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
 178                 if (spent < td->o.ratecycle || spent==0)
 179                         return false;
 180
 181                 if (td->o.ratemin[ddir]) {
 182                         /*
 183                          * check bandwidth specified rate
 184                          */
 185                         unsigned long long current_rate_bytes =
 186                                 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
 187                         if (current_rate_bytes < option_rate_bytes_min) {
 188                                 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
 189                                         td->o.name, option_rate_bytes_min, current_rate_bytes);
 190                                 return true;
 191                         }
 192                 } else {
 193                         /*
 194                          * checks iops specified rate
 195                          */
 196                         unsigned long long current_rate_iops =
 197                                 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
 198
 199                         if (current_rate_iops < option_rate_iops_min) {
 200                                 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
 201                                         td->o.name, option_rate_iops_min, current_rate_iops);
 202                                 return true;
 203                         }
 204                 }
 205         }
 206
 207         td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
 208         td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
 209         memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
 210         return false;
 211 }
 212
 213 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 214 {
 215         bool ret = false;
 216
 217         for_each_rw_ddir(ddir) {
 218                 if (td->bytes_done[ddir])
 219                         ret |= __check_min_rate(td, now, ddir);
 220         }
 221
 222         return ret;
 223 }
 224
 225 /*
 226  * When job exits, we can cancel the in-flight IO if we are using async
 227  * io. Attempt to do so.
 228  */
 229 static void cleanup_pending_aio(struct thread_data *td)
 230 {
 231         int r;
 232
 233         /*
 234          * get immediately available events, if any
 235          */
 236         r = io_u_queued_complete(td, 0);
 237
 238         /*
 239          * now cancel remaining active events
 240          */
 241         if (td->io_ops->cancel) {
 242                 struct io_u *io_u;
 243                 int i;
 244
 245                 io_u_qiter(&td->io_u_all, io_u, i) {
 246                         if (io_u->flags & IO_U_F_FLIGHT) {
 247                                 r = td->io_ops->cancel(td, io_u);
 248                                 if (!r)
 249                                         put_io_u(td, io_u);
 250                         }
 251                 }
 252         }
 253
 254         if (td->cur_depth)
 255                 r = io_u_queued_complete(td, td->cur_depth);
 256 }
 257
 258 /*
 259  * Helper to handle the final sync of a file. Works just like the normal
 260  * io path, just does everything sync.
 261  */
 262 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 263 {
 264         struct io_u *io_u = __get_io_u(td);
 265         enum fio_q_status ret;
 266
 267         if (!io_u)
 268                 return true;
 269
 270         io_u->ddir = DDIR_SYNC;
 271         io_u->file = f;
 272         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
 273
 274         if (td_io_prep(td, io_u)) {
 275                 put_io_u(td, io_u);
 276                 return true;
 277         }
 278
 279 requeue:
 280         ret = td_io_queue(td, io_u);
 281         switch (ret) {
 282         case FIO_Q_QUEUED:
 283                 td_io_commit(td);
 284                 if (io_u_queued_complete(td, 1) < 0)
 285                         return true;
 286                 break;
 287         case FIO_Q_COMPLETED:
 288                 if (io_u->error) {
 289                         td_verror(td, io_u->error, "td_io_queue");
 290                         return true;
 291                 }
 292
 293                 if (io_u_sync_complete(td, io_u) < 0)
 294                         return true;
 295                 break;
 296         case FIO_Q_BUSY:
 297                 td_io_commit(td);
 298                 goto requeue;
 299         }
 300
 301         return false;
 302 }
 303
 304 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 305 {
 306         int ret, ret2;
 307
 308         if (fio_file_open(f))
 309                 return fio_io_sync(td, f);
 310
 311         if (td_io_open_file(td, f))
 312                 return 1;
 313
 314         ret = fio_io_sync(td, f);
 315         ret2 = 0;
 316         if (fio_file_open(f))
 317                 ret2 = td_io_close_file(td, f);
 318         return (ret || ret2);
 319 }
 320
 321 static inline void __update_ts_cache(struct thread_data *td)
 322 {
 323         fio_gettime(&td->ts_cache, NULL);
 324 }
 325
 326 static inline void update_ts_cache(struct thread_data *td)
 327 {
 328         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 329                 __update_ts_cache(td);
 330 }
 331
 332 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 333 {
 334         if (in_ramp_time(td))
 335                 return false;
 336         if (!td->o.timeout)
 337                 return false;
 338         if (utime_since(&td->epoch, t) >= td->o.timeout)
 339                 return true;
 340
 341         return false;
 342 }
 343
 344 /*
 345  * We need to update the runtime consistently in ms, but keep a running
 346  * tally of the current elapsed time in microseconds for sub millisecond
 347  * updates.
 348  */
 349 static inline void update_runtime(struct thread_data *td,
 350                                   unsigned long long *elapsed_us,
 351                                   const enum fio_ddir ddir)
 352 {
 353         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 354                 return;
 355
 356         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 357         elapsed_us[ddir] += utime_since_now(&td->start);
 358         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 359 }
 360
 361 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 362                                 int *retptr)
 363 {
 364         int ret = *retptr;
 365
 366         if (ret < 0 || td->error) {
 367                 int err = td->error;
 368                 enum error_type_bit eb;
 369
 370                 if (ret < 0)
 371                         err = -ret;
 372
 373                 eb = td_error_type(ddir, err);
 374                 if (!(td->o.continue_on_error & (1 << eb)))
 375                         return true;
 376
 377                 if (td_non_fatal_error(td, eb, err)) {
 378                         /*
 379                          * Continue with the I/Os in case of
 380                          * a non fatal error.
 381                          */
 382                         update_error_count(td, err);
 383                         td_clear_error(td);
 384                         *retptr = 0;
 385                         return false;
 386                 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
 387                         /*
 388                          * We expect to hit this error if
 389                          * fill_device option is set.
 390                          */
 391                         td_clear_error(td);
 392                         fio_mark_td_terminate(td);
 393                         return true;
 394                 } else {
 395                         /*
 396                          * Stop the I/O in case of a fatal
 397                          * error.
 398                          */
 399                         update_error_count(td, err);
 400                         return true;
 401                 }
 402         }
 403
 404         return false;
 405 }
 406
 407 static void check_update_rusage(struct thread_data *td)
 408 {
 409         if (td->update_rusage) {
 410                 td->update_rusage = 0;
 411                 update_rusage_stat(td);
 412                 fio_sem_up(td->rusage_sem);
 413         }
 414 }
 415
 416 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 417 {
 418         const int full = queue_full(td);
 419         int min_evts = 0;
 420         int ret;
 421
 422         if (td->flags & TD_F_REGROW_LOGS)
 423                 return io_u_quiesce(td);
 424
 425         /*
 426          * if the queue is full, we MUST reap at least 1 event
 427          */
 428         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 429         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 430                 min_evts = 1;
 431
 432         if (time && should_check_rate(td))
 433                 fio_gettime(time, NULL);
 434
 435         do {
 436                 ret = io_u_queued_complete(td, min_evts);
 437                 if (ret < 0)
 438                         break;
 439         } while (full && (td->cur_depth > td->o.iodepth_low));
 440
 441         return ret;
 442 }
 443
 444 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 445                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 446                    struct timespec *comp_time)
 447 {
 448         switch (*ret) {
 449         case FIO_Q_COMPLETED:
 450                 if (io_u->error) {
 451                         *ret = -io_u->error;
 452                         clear_io_u(td, io_u);
 453                 } else if (io_u->resid) {
 454                         long long bytes = io_u->xfer_buflen - io_u->resid;
 455                         struct fio_file *f = io_u->file;
 456
 457                         if (bytes_issued)
 458                                 *bytes_issued += bytes;
 459
 460                         if (!from_verify)
 461                                 trim_io_piece(io_u);
 462
 463                         /*
 464                          * zero read, fail
 465                          */
 466                         if (!bytes) {
 467                                 if (!from_verify)
 468                                         unlog_io_piece(td, io_u);
 469                                 td_verror(td, EIO, "full resid");
 470                                 clear_io_u(td, io_u);
 471                                 break;
 472                         }
 473
 474                         io_u->xfer_buflen = io_u->resid;
 475                         io_u->xfer_buf += bytes;
 476                         io_u->offset += bytes;
 477
 478                         if (ddir_rw(io_u->ddir))
 479                                 td->ts.short_io_u[io_u->ddir]++;
 480
 481                         if (io_u->offset == f->real_file_size)
 482                                 goto sync_done;
 483
 484                         requeue_io_u(td, &io_u);
 485                 } else {
 486 sync_done:
 487                         if (comp_time && should_check_rate(td))
 488                                 fio_gettime(comp_time, NULL);
 489
 490                         *ret = io_u_sync_complete(td, io_u);
 491                         if (*ret < 0)
 492                                 break;
 493                 }
 494
 495                 if (td->flags & TD_F_REGROW_LOGS)
 496                         regrow_logs(td);
 497
 498                 /*
 499                  * when doing I/O (not when verifying),
 500                  * check for any errors that are to be ignored
 501                  */
 502                 if (!from_verify)
 503                         break;
 504
 505                 return 0;
 506         case FIO_Q_QUEUED:
 507                 /*
 508                  * if the engine doesn't have a commit hook,
 509                  * the io_u is really queued. if it does have such
 510                  * a hook, it has to call io_u_queued() itself.
 511                  */
 512                 if (td->io_ops->commit == NULL)
 513                         io_u_queued(td, io_u);
 514                 if (bytes_issued)
 515                         *bytes_issued += io_u->xfer_buflen;
 516                 break;
 517         case FIO_Q_BUSY:
 518                 if (!from_verify)
 519                         unlog_io_piece(td, io_u);
 520                 requeue_io_u(td, &io_u);
 521                 td_io_commit(td);
 522                 break;
 523         default:
 524                 assert(*ret < 0);
 525                 td_verror(td, -(*ret), "td_io_queue");
 526                 break;
 527         }
 528
 529         if (break_on_this_error(td, ddir, ret))
 530                 return 1;
 531
 532         return 0;
 533 }
 534
 535 static inline bool io_in_polling(struct thread_data *td)
 536 {
 537         return !td->o.iodepth_batch_complete_min &&
 538                    !td->o.iodepth_batch_complete_max;
 539 }
 540 /*
 541  * Unlinks files from thread data fio_file structure
 542  */
 543 static int unlink_all_files(struct thread_data *td)
 544 {
 545         struct fio_file *f;
 546         unsigned int i;
 547         int ret = 0;
 548
 549         for_each_file(td, f, i) {
 550                 if (f->filetype != FIO_TYPE_FILE)
 551                         continue;
 552                 ret = td_io_unlink_file(td, f);
 553                 if (ret)
 554                         break;
 555         }
 556
 557         if (ret)
 558                 td_verror(td, ret, "unlink_all_files");
 559
 560         return ret;
 561 }
 562
 563 /*
 564  * Check if io_u will overlap an in-flight IO in the queue
 565  */
 566 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 567 {
 568         bool overlap;
 569         struct io_u *check_io_u;
 570         unsigned long long x1, x2, y1, y2;
 571         int i;
 572
 573         x1 = io_u->offset;
 574         x2 = io_u->offset + io_u->buflen;
 575         overlap = false;
 576         io_u_qiter(q, check_io_u, i) {
 577                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 578                         y1 = check_io_u->offset;
 579                         y2 = check_io_u->offset + check_io_u->buflen;
 580
 581                         if (x1 < y2 && y1 < x2) {
 582                                 overlap = true;
 583                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 584                                                 x1, io_u->buflen,
 585                                                 y1, check_io_u->buflen);
 586                                 break;
 587                         }
 588                 }
 589         }
 590
 591         return overlap;
 592 }
 593
 594 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 595 {
 596         /*
 597          * Check for overlap if the user asked us to, and we have
 598          * at least one IO in flight besides this one.
 599          */
 600         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 601             in_flight_overlap(&td->io_u_all, io_u))
 602                 return FIO_Q_BUSY;
 603
 604         return td_io_queue(td, io_u);
 605 }
 606
 607 /*
 608  * The main verify engine. Runs over the writes we previously submitted,
 609  * reads the blocks back in, and checks the crc/md5 of the data.
 610  */
 611 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 612 {
 613         struct fio_file *f;
 614         struct io_u *io_u;
 615         int ret, min_events;
 616         unsigned int i;
 617
 618         dprint(FD_VERIFY, "starting loop\n");
 619
 620         /*
 621          * sync io first and invalidate cache, to make sure we really
 622          * read from disk.
 623          */
 624         for_each_file(td, f, i) {
 625                 if (!fio_file_open(f))
 626                         continue;
 627                 if (fio_io_sync(td, f))
 628                         break;
 629                 if (file_invalidate_cache(td, f))
 630                         break;
 631         }
 632
 633         check_update_rusage(td);
 634
 635         if (td->error)
 636                 return;
 637
 638         td_set_runstate(td, TD_VERIFYING);
 639
 640         io_u = NULL;
 641         while (!td->terminate) {
 642                 enum fio_ddir ddir;
 643                 int full;
 644
 645                 update_ts_cache(td);
 646                 check_update_rusage(td);
 647
 648                 if (runtime_exceeded(td, &td->ts_cache)) {
 649                         __update_ts_cache(td);
 650                         if (runtime_exceeded(td, &td->ts_cache)) {
 651                                 fio_mark_td_terminate(td);
 652                                 break;
 653                         }
 654                 }
 655
 656                 if (flow_threshold_exceeded(td))
 657                         continue;
 658
 659                 if (!td->o.experimental_verify) {
 660                         io_u = __get_io_u(td);
 661                         if (!io_u)
 662                                 break;
 663
 664                         if (get_next_verify(td, io_u)) {
 665                                 put_io_u(td, io_u);
 666                                 break;
 667                         }
 668
 669                         if (td_io_prep(td, io_u)) {
 670                                 put_io_u(td, io_u);
 671                                 break;
 672                         }
 673                 } else {
 674                         if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
 675                                 break;
 676
 677                         while ((io_u = get_io_u(td)) != NULL) {
 678                                 if (IS_ERR_OR_NULL(io_u)) {
 679                                         io_u = NULL;
 680                                         ret = FIO_Q_BUSY;
 681                                         goto reap;
 682                                 }
 683
 684                                 /*
 685                                  * We are only interested in the places where
 686                                  * we wrote or trimmed IOs. Turn those into
 687                                  * reads for verification purposes.
 688                                  */
 689                                 if (io_u->ddir == DDIR_READ) {
 690                                         /*
 691                                          * Pretend we issued it for rwmix
 692                                          * accounting
 693                                          */
 694                                         td->io_issues[DDIR_READ]++;
 695                                         put_io_u(td, io_u);
 696                                         continue;
 697                                 } else if (io_u->ddir == DDIR_TRIM) {
 698                                         io_u->ddir = DDIR_READ;
 699                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 700                                         break;
 701                                 } else if (io_u->ddir == DDIR_WRITE) {
 702                                         io_u->ddir = DDIR_READ;
 703                                         io_u->numberio = td->verify_read_issues;
 704                                         td->verify_read_issues++;
 705                                         populate_verify_io_u(td, io_u);
 706                                         break;
 707                                 } else {
 708                                         put_io_u(td, io_u);
 709                                         continue;
 710                                 }
 711                         }
 712
 713                         if (!io_u)
 714                                 break;
 715                 }
 716
 717                 if (verify_state_should_stop(td, io_u)) {
 718                         put_io_u(td, io_u);
 719                         break;
 720                 }
 721
 722                 if (td->o.verify_async)
 723                         io_u->end_io = verify_io_u_async;
 724                 else
 725                         io_u->end_io = verify_io_u;
 726
 727                 ddir = io_u->ddir;
 728                 if (!td->o.disable_slat)
 729                         fio_gettime(&io_u->start_time, NULL);
 730
 731                 ret = io_u_submit(td, io_u);
 732
 733                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 734                         break;
 735
 736                 /*
 737                  * if we can queue more, do so. but check if there are
 738                  * completed io_u's first. Note that we can get BUSY even
 739                  * without IO queued, if the system is resource starved.
 740                  */
 741 reap:
 742                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 743                 if (full || io_in_polling(td))
 744                         ret = wait_for_completions(td, NULL);
 745
 746                 if (ret < 0)
 747                         break;
 748         }
 749
 750         check_update_rusage(td);
 751
 752         if (!td->error) {
 753                 min_events = td->cur_depth;
 754
 755                 if (min_events)
 756                         ret = io_u_queued_complete(td, min_events);
 757         } else
 758                 cleanup_pending_aio(td);
 759
 760         td_set_runstate(td, TD_RUNNING);
 761
 762         dprint(FD_VERIFY, "exiting loop\n");
 763 }
 764
 765 static bool exceeds_number_ios(struct thread_data *td)
 766 {
 767         unsigned long long number_ios;
 768
 769         if (!td->o.number_ios)
 770                 return false;
 771
 772         number_ios = ddir_rw_sum(td->io_blocks);
 773         number_ios += td->io_u_queued + td->io_u_in_flight;
 774
 775         return number_ios >= (td->o.number_ios * td->loops);
 776 }
 777
 778 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 779 {
 780         unsigned long long bytes, limit;
 781
 782         if (td_rw(td))
 783                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 784         else if (td_write(td))
 785                 bytes = this_bytes[DDIR_WRITE];
 786         else if (td_read(td))
 787                 bytes = this_bytes[DDIR_READ];
 788         else
 789                 bytes = this_bytes[DDIR_TRIM];
 790
 791         if (td->o.io_size)
 792                 limit = td->o.io_size;
 793         else
 794                 limit = td->o.size;
 795
 796         limit *= td->loops;
 797         return bytes >= limit || exceeds_number_ios(td);
 798 }
 799
 800 static bool io_issue_bytes_exceeded(struct thread_data *td)
 801 {
 802         return io_bytes_exceeded(td, td->io_issue_bytes);
 803 }
 804
 805 static bool io_complete_bytes_exceeded(struct thread_data *td)
 806 {
 807         return io_bytes_exceeded(td, td->this_io_bytes);
 808 }
 809
 810 /*
 811  * used to calculate the next io time for rate control
 812  *
 813  */
 814 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 815 {
 816         uint64_t bps = td->rate_bps[ddir];
 817
 818         assert(!(td->flags & TD_F_CHILD));
 819
 820         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 821                 uint64_t val, iops;
 822
 823                 iops = bps / td->o.min_bs[ddir];
 824                 val = (int64_t) (1000000 / iops) *
 825                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 826                 if (val) {
 827                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 828                                         (unsigned long long) 1000000 / val,
 829                                         ddir);
 830                 }
 831                 td->last_usec[ddir] += val;
 832                 return td->last_usec[ddir];
 833         } else if (bps) {
 834                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 835                 uint64_t secs = bytes / bps;
 836                 uint64_t remainder = bytes % bps;
 837
 838                 return remainder * 1000000 / bps + secs * 1000000;
 839         }
 840
 841         return 0;
 842 }
 843
 844 static void init_thinktime(struct thread_data *td)
 845 {
 846         if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
 847                 td->thinktime_blocks_counter = td->io_blocks;
 848         else
 849                 td->thinktime_blocks_counter = td->io_issues;
 850         td->last_thinktime = td->epoch;
 851         td->last_thinktime_blocks = 0;
 852 }
 853
 854 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
 855                              struct timespec *time)
 856 {
 857         unsigned long long b;
 858         unsigned long long runtime_left;
 859         uint64_t total;
 860         int left;
 861         struct timespec now;
 862         bool stall = false;
 863
 864         if (td->o.thinktime_iotime) {
 865                 fio_gettime(&now, NULL);
 866                 if (utime_since(&td->last_thinktime, &now)
 867                     >= td->o.thinktime_iotime) {
 868                         stall = true;
 869                 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
 870                         /*
 871                          * When thinktime_iotime is set and thinktime_blocks is
 872                          * not set, skip the thinktime_blocks check, since
 873                          * thinktime_blocks default value 1 does not work
 874                          * together with thinktime_iotime.
 875                          */
 876                         return;
 877                 }
 878
 879         }
 880
 881         b = ddir_rw_sum(td->thinktime_blocks_counter);
 882         if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
 883                 stall = true;
 884
 885         if (!stall)
 886                 return;
 887
 888         io_u_quiesce(td);
 889
 890         left = td->o.thinktime_spin;
 891         if (td->o.timeout) {
 892                 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
 893                 if (runtime_left < (unsigned long long)left)
 894                         left = runtime_left;
 895         }
 896
 897         total = 0;
 898         if (left)
 899                 total = usec_spin(left);
 900
 901         /*
 902          * usec_spin() might run for slightly longer than intended in a VM
 903          * where the vCPU could get descheduled or the hypervisor could steal
 904          * CPU time. Ensure "left" doesn't become negative.
 905          */
 906         if (total < td->o.thinktime)
 907                 left = td->o.thinktime - total;
 908         else
 909                 left = 0;
 910
 911         if (td->o.timeout) {
 912                 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
 913                 if (runtime_left < (unsigned long long)left)
 914                         left = runtime_left;
 915         }
 916
 917         if (left)
 918                 total += usec_sleep(td, left);
 919
 920         /*
 921          * If we're ignoring thinktime for the rate, add the number of bytes
 922          * we would have done while sleeping, minus one block to ensure we
 923          * start issuing immediately after the sleep.
 924          */
 925         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 926                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 927                 uint64_t bs = td->o.min_bs[ddir];
 928                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 929                 uint64_t over;
 930
 931                 if (usperop <= total)
 932                         over = bs;
 933                 else
 934                         over = (usperop - total) / usperop * -bs;
 935
 936                 td->rate_io_issue_bytes[ddir] += (missed - over);
 937                 /* adjust for rate_process=poisson */
 938                 td->last_usec[ddir] += total;
 939         }
 940
 941         if (time && should_check_rate(td))
 942                 fio_gettime(time, NULL);
 943
 944         td->last_thinktime_blocks = b;
 945         if (td->o.thinktime_iotime) {
 946                 fio_gettime(&now, NULL);
 947                 td->last_thinktime = now;
 948         }
 949 }
 950
 951 /*
 952  * Main IO worker function. It retrieves io_u's to process and queues
 953  * and reaps them, checking for rate and errors along the way.
 954  *
 955  * Returns number of bytes written and trimmed.
 956  */
 957 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 958 {
 959         unsigned int i;
 960         int ret = 0;
 961         uint64_t total_bytes, bytes_issued = 0;
 962
 963         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 964                 bytes_done[i] = td->bytes_done[i];
 965
 966         if (in_ramp_time(td))
 967                 td_set_runstate(td, TD_RAMP);
 968         else
 969                 td_set_runstate(td, TD_RUNNING);
 970
 971         lat_target_init(td);
 972
 973         total_bytes = td->o.size;
 974         /*
 975         * Allow random overwrite workloads to write up to io_size
 976         * before starting verification phase as 'size' doesn't apply.
 977         */
 978         if (td_write(td) && td_random(td) && td->o.norandommap)
 979                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 980
 981         /* Don't break too early if io_size > size */
 982         if (td_rw(td) && !td_random(td))
 983                 total_bytes = max(total_bytes, (uint64_t)td->o.io_size);
 984
 985         /*
 986          * If verify_backlog is enabled, we'll run the verify in this
 987          * handler as well. For that case, we may need up to twice the
 988          * amount of bytes.
 989          */
 990         if (td->o.verify != VERIFY_NONE &&
 991            (td_write(td) && td->o.verify_backlog))
 992                 total_bytes += td->o.size;
 993
 994         /* In trimwrite mode, each byte is trimmed and then written, so
 995          * allow total_bytes or number of ios to be twice as big */
 996         if (td_trimwrite(td)) {
 997                 total_bytes += td->total_io_size;
 998                 td->o.number_ios *= 2;
 999         }
1000
1001         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1002                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
1003                 td->o.time_based) {
1004                 struct timespec comp_time;
1005                 struct io_u *io_u;
1006                 int full;
1007                 enum fio_ddir ddir;
1008
1009                 check_update_rusage(td);
1010
1011                 if (td->terminate || td->done)
1012                         break;
1013
1014                 update_ts_cache(td);
1015
1016                 if (runtime_exceeded(td, &td->ts_cache)) {
1017                         __update_ts_cache(td);
1018                         if (runtime_exceeded(td, &td->ts_cache)) {
1019                                 fio_mark_td_terminate(td);
1020                                 break;
1021                         }
1022                 }
1023
1024                 if (flow_threshold_exceeded(td))
1025                         continue;
1026
1027                 /*
1028                  * Break if we exceeded the bytes. The exception is time
1029                  * based runs, but we still need to break out of the loop
1030                  * for those to run verification, if enabled.
1031                  * Jobs read from iolog do not use this stop condition.
1032                  */
1033                 if (bytes_issued >= total_bytes &&
1034                     !td->o.read_iolog_file &&
1035                     (!td->o.time_based ||
1036                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
1037                         break;
1038
1039                 io_u = get_io_u(td);
1040                 if (IS_ERR_OR_NULL(io_u)) {
1041                         int err = PTR_ERR(io_u);
1042
1043                         io_u = NULL;
1044                         ddir = DDIR_INVAL;
1045                         if (err == -EBUSY) {
1046                                 ret = FIO_Q_BUSY;
1047                                 goto reap;
1048                         }
1049                         if (td->o.latency_target)
1050                                 goto reap;
1051                         break;
1052                 }
1053
1054                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
1055                         if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
1056                                 io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
1057                                 io_u->numberio = td->io_issues[io_u->ddir];
1058                                 populate_verify_io_u(td, io_u);
1059                         }
1060                 }
1061
1062                 ddir = io_u->ddir;
1063
1064                 /*
1065                  * Add verification end_io handler if:
1066                  *      - Asked to verify (!td_rw(td))
1067                  *      - Or the io_u is from our verify list (mixed write/ver)
1068                  */
1069                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1070                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1071
1072                         if (verify_state_should_stop(td, io_u)) {
1073                                 put_io_u(td, io_u);
1074                                 break;
1075                         }
1076
1077                         if (td->o.verify_async)
1078                                 io_u->end_io = verify_io_u_async;
1079                         else
1080                                 io_u->end_io = verify_io_u;
1081                         td_set_runstate(td, TD_VERIFYING);
1082                 } else if (in_ramp_time(td))
1083                         td_set_runstate(td, TD_RAMP);
1084                 else
1085                         td_set_runstate(td, TD_RUNNING);
1086
1087                 /*
1088                  * Always log IO before it's issued, so we know the specific
1089                  * order of it. The logged unit will track when the IO has
1090                  * completed.
1091                  */
1092                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1093                     td->o.do_verify &&
1094                     td->o.verify != VERIFY_NONE &&
1095                     !td->o.experimental_verify)
1096                         log_io_piece(td, io_u);
1097
1098                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1099                         const unsigned long long blen = io_u->xfer_buflen;
1100                         const enum fio_ddir __ddir = acct_ddir(io_u);
1101
1102                         if (td->error)
1103                                 break;
1104
1105                         workqueue_enqueue(&td->io_wq, &io_u->work);
1106                         ret = FIO_Q_QUEUED;
1107
1108                         if (ddir_rw(__ddir)) {
1109                                 td->io_issues[__ddir]++;
1110                                 td->io_issue_bytes[__ddir] += blen;
1111                                 td->rate_io_issue_bytes[__ddir] += blen;
1112                         }
1113
1114                         if (should_check_rate(td)) {
1115                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1116                                 fio_gettime(&comp_time, NULL);
1117                         }
1118
1119                 } else {
1120                         ret = io_u_submit(td, io_u);
1121
1122                         if (should_check_rate(td))
1123                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1124
1125                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1126                                 break;
1127
1128                         /*
1129                          * See if we need to complete some commands. Note that
1130                          * we can get BUSY even without IO queued, if the
1131                          * system is resource starved.
1132                          */
1133 reap:
1134                         full = queue_full(td) ||
1135                                 (ret == FIO_Q_BUSY && td->cur_depth);
1136                         if (full || io_in_polling(td))
1137                                 ret = wait_for_completions(td, &comp_time);
1138                 }
1139                 if (ret < 0)
1140                         break;
1141
1142                 if (ddir_rw(ddir) && td->o.thinkcycles)
1143                         cycles_spin(td->o.thinkcycles);
1144
1145                 if (ddir_rw(ddir) && td->o.thinktime)
1146                         handle_thinktime(td, ddir, &comp_time);
1147
1148                 if (!ddir_rw_sum(td->bytes_done) &&
1149                     !td_ioengine_flagged(td, FIO_NOIO))
1150                         continue;
1151
1152                 if (!in_ramp_time(td) && should_check_rate(td)) {
1153                         if (check_min_rate(td, &comp_time)) {
1154                                 if (exitall_on_terminate || td->o.exitall_error)
1155                                         fio_terminate_threads(td->groupid, td->o.exit_what);
1156                                 td_verror(td, EIO, "check_min_rate");
1157                                 break;
1158                         }
1159                 }
1160                 if (!in_ramp_time(td) && td->o.latency_target)
1161                         lat_target_check(td);
1162         }
1163
1164         check_update_rusage(td);
1165
1166         if (td->trim_entries)
1167                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1168
1169         if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1170                 td->error = 0;
1171                 fio_mark_td_terminate(td);
1172         }
1173         if (!td->error) {
1174                 struct fio_file *f;
1175
1176                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1177                         workqueue_flush(&td->io_wq);
1178                         i = 0;
1179                 } else
1180                         i = td->cur_depth;
1181
1182                 if (i) {
1183                         ret = io_u_queued_complete(td, i);
1184                         if (td->o.fill_device &&
1185                             (td->error == ENOSPC || td->error == EDQUOT))
1186                                 td->error = 0;
1187                 }
1188
1189                 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1190                         td_set_runstate(td, TD_FSYNCING);
1191
1192                         for_each_file(td, f, i) {
1193                                 if (!fio_file_fsync(td, f))
1194                                         continue;
1195
1196                                 log_err("fio: end_fsync failed for file %s\n",
1197                                                                 f->file_name);
1198                         }
1199                 }
1200         } else {
1201                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1202                         workqueue_flush(&td->io_wq);
1203                 cleanup_pending_aio(td);
1204         }
1205
1206         /*
1207          * stop job if we failed doing any IO
1208          */
1209         if (!ddir_rw_sum(td->this_io_bytes))
1210                 td->done = 1;
1211
1212         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1213                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1214 }
1215
1216 static void free_file_completion_logging(struct thread_data *td)
1217 {
1218         struct fio_file *f;
1219         unsigned int i;
1220
1221         for_each_file(td, f, i) {
1222                 if (!f->last_write_comp)
1223                         break;
1224                 sfree(f->last_write_comp);
1225         }
1226 }
1227
1228 static int init_file_completion_logging(struct thread_data *td,
1229                                         unsigned int depth)
1230 {
1231         struct fio_file *f;
1232         unsigned int i;
1233
1234         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1235                 return 0;
1236
1237         for_each_file(td, f, i) {
1238                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1239                 if (!f->last_write_comp)
1240                         goto cleanup;
1241         }
1242
1243         return 0;
1244
1245 cleanup:
1246         free_file_completion_logging(td);
1247         log_err("fio: failed to alloc write comp data\n");
1248         return 1;
1249 }
1250
1251 static void cleanup_io_u(struct thread_data *td)
1252 {
1253         struct io_u *io_u;
1254
1255         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1256
1257                 if (td->io_ops->io_u_free)
1258                         td->io_ops->io_u_free(td, io_u);
1259
1260                 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1261         }
1262
1263         free_io_mem(td);
1264
1265         io_u_rexit(&td->io_u_requeues);
1266         io_u_qexit(&td->io_u_freelist, false);
1267         io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1268
1269         free_file_completion_logging(td);
1270 }
1271
1272 static int init_io_u(struct thread_data *td)
1273 {
1274         struct io_u *io_u;
1275         int cl_align, i, max_units;
1276         int err;
1277
1278         max_units = td->o.iodepth;
1279
1280         err = 0;
1281         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1282         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1283         err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1284
1285         if (err) {
1286                 log_err("fio: failed setting up IO queues\n");
1287                 return 1;
1288         }
1289
1290         cl_align = os_cache_line_size();
1291
1292         for (i = 0; i < max_units; i++) {
1293                 void *ptr;
1294
1295                 if (td->terminate)
1296                         return 1;
1297
1298                 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1299                 if (!ptr) {
1300                         log_err("fio: unable to allocate aligned memory\n");
1301                         return 1;
1302                 }
1303
1304                 io_u = ptr;
1305                 memset(io_u, 0, sizeof(*io_u));
1306                 INIT_FLIST_HEAD(&io_u->verify_list);
1307                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1308
1309                 io_u->index = i;
1310                 io_u->flags = IO_U_F_FREE;
1311                 io_u_qpush(&td->io_u_freelist, io_u);
1312
1313                 /*
1314                  * io_u never leaves this stack, used for iteration of all
1315                  * io_u buffers.
1316                  */
1317                 io_u_qpush(&td->io_u_all, io_u);
1318
1319                 if (td->io_ops->io_u_init) {
1320                         int ret = td->io_ops->io_u_init(td, io_u);
1321
1322                         if (ret) {
1323                                 log_err("fio: failed to init engine data: %d\n", ret);
1324                                 return 1;
1325                         }
1326                 }
1327         }
1328
1329         if (init_io_u_buffers(td))
1330                 return 1;
1331
1332         if (init_file_completion_logging(td, max_units))
1333                 return 1;
1334
1335         return 0;
1336 }
1337
1338 int init_io_u_buffers(struct thread_data *td)
1339 {
1340         struct io_u *io_u;
1341         unsigned long long max_bs, min_write, trim_bs = 0;
1342         int i, max_units;
1343         int data_xfer = 1;
1344         char *p;
1345
1346         max_units = td->o.iodepth;
1347         max_bs = td_max_bs(td);
1348         min_write = td->o.min_bs[DDIR_WRITE];
1349         td->orig_buffer_size = (unsigned long long) max_bs
1350                                         * (unsigned long long) max_units;
1351
1352         if (td_trim(td) && td->o.num_range > 1) {
1353                 trim_bs = td->o.num_range * sizeof(struct trim_range);
1354                 td->orig_buffer_size = trim_bs
1355                                         * (unsigned long long) max_units;
1356         }
1357
1358         /*
1359          * For reads, writes, and multi-range trim operations we need a
1360          * data buffer
1361          */
1362         if (td_ioengine_flagged(td, FIO_NOIO) ||
1363             !(td_read(td) || td_write(td) || (td_trim(td) && td->o.num_range > 1)))
1364                 data_xfer = 0;
1365
1366         /*
1367          * if we may later need to do address alignment, then add any
1368          * possible adjustment here so that we don't cause a buffer
1369          * overflow later. this adjustment may be too much if we get
1370          * lucky and the allocator gives us an aligned address.
1371          */
1372         if (td->o.odirect || td->o.mem_align ||
1373             td_ioengine_flagged(td, FIO_RAWIO))
1374                 td->orig_buffer_size += page_mask + td->o.mem_align;
1375
1376         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1377                 unsigned long long bs;
1378
1379                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1380                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1381         }
1382
1383         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1384                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1385                 return 1;
1386         }
1387
1388         if (data_xfer && allocate_io_mem(td))
1389                 return 1;
1390
1391         if (td->o.odirect || td->o.mem_align ||
1392             td_ioengine_flagged(td, FIO_RAWIO))
1393                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1394         else
1395                 p = td->orig_buffer;
1396
1397         for (i = 0; i < max_units; i++) {
1398                 io_u = td->io_u_all.io_us[i];
1399                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1400
1401                 if (data_xfer) {
1402                         io_u->buf = p;
1403                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1404
1405                         if (td_write(td))
1406                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1407                         if (td_write(td) && td->o.verify_pattern_bytes) {
1408                                 /*
1409                                  * Fill the buffer with the pattern if we are
1410                                  * going to be doing writes.
1411                                  */
1412                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1413                         }
1414                 }
1415                 if (td_trim(td) && td->o.num_range > 1)
1416                         p += trim_bs;
1417                 else
1418                         p += max_bs;
1419         }
1420
1421         return 0;
1422 }
1423
1424 #ifdef FIO_HAVE_IOSCHED_SWITCH
1425 /*
1426  * These functions are Linux specific.
1427  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1428  */
1429 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1430 {
1431         char tmp[256], tmp2[128], *p;
1432         FILE *f;
1433         int ret;
1434
1435         assert(file->du && file->du->sysfs_root);
1436         sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1437
1438         f = fopen(tmp, "r+");
1439         if (!f) {
1440                 if (errno == ENOENT) {
1441                         log_err("fio: os or kernel doesn't support IO scheduler"
1442                                 " switching\n");
1443                         return 0;
1444                 }
1445                 td_verror(td, errno, "fopen iosched");
1446                 return 1;
1447         }
1448
1449         /*
1450          * Set io scheduler.
1451          */
1452         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1453         if (ferror(f) || ret != 1) {
1454                 td_verror(td, errno, "fwrite");
1455                 fclose(f);
1456                 return 1;
1457         }
1458
1459         rewind(f);
1460
1461         /*
1462          * Read back and check that the selected scheduler is now the default.
1463          */
1464         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1465         if (ferror(f) || ret < 0) {
1466                 td_verror(td, errno, "fread");
1467                 fclose(f);
1468                 return 1;
1469         }
1470         tmp[ret] = '\0';
1471         /*
1472          * either a list of io schedulers or "none\n" is expected. Strip the
1473          * trailing newline.
1474          */
1475         p = tmp;
1476         strsep(&p, "\n");
1477
1478         /*
1479          * Write to "none" entry doesn't fail, so check the result here.
1480          */
1481         if (!strcmp(tmp, "none")) {
1482                 log_err("fio: io scheduler is not tunable\n");
1483                 fclose(f);
1484                 return 0;
1485         }
1486
1487         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1488         if (!strstr(tmp, tmp2)) {
1489                 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1490                 td_verror(td, EINVAL, "iosched_switch");
1491                 fclose(f);
1492                 return 1;
1493         }
1494
1495         fclose(f);
1496         return 0;
1497 }
1498
1499 static int switch_ioscheduler(struct thread_data *td)
1500 {
1501         struct fio_file *f;
1502         unsigned int i;
1503         int ret = 0;
1504
1505         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1506                 return 0;
1507
1508         assert(td->files && td->files[0]);
1509
1510         for_each_file(td, f, i) {
1511
1512                 /* Only consider regular files and block device files */
1513                 switch (f->filetype) {
1514                 case FIO_TYPE_FILE:
1515                 case FIO_TYPE_BLOCK:
1516                         /*
1517                          * Make sure that the device hosting the file could
1518                          * be determined.
1519                          */
1520                         if (!f->du)
1521                                 continue;
1522                         break;
1523                 case FIO_TYPE_CHAR:
1524                 case FIO_TYPE_PIPE:
1525                 default:
1526                         continue;
1527                 }
1528
1529                 ret = set_ioscheduler(td, f);
1530                 if (ret)
1531                         return ret;
1532         }
1533
1534         return 0;
1535 }
1536
1537 #else
1538
1539 static int switch_ioscheduler(struct thread_data *td)
1540 {
1541         return 0;
1542 }
1543
1544 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1545
1546 static bool keep_running(struct thread_data *td)
1547 {
1548         unsigned long long limit;
1549
1550         if (td->done)
1551                 return false;
1552         if (td->terminate)
1553                 return false;
1554         if (td->o.time_based)
1555                 return true;
1556         if (td->o.loops) {
1557                 td->o.loops--;
1558                 return true;
1559         }
1560         if (exceeds_number_ios(td))
1561                 return false;
1562
1563         if (td->o.io_size)
1564                 limit = td->o.io_size;
1565         else
1566                 limit = td->o.size;
1567
1568         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1569                 uint64_t diff;
1570
1571                 /*
1572                  * If the difference is less than the maximum IO size, we
1573                  * are done.
1574                  */
1575                 diff = limit - ddir_rw_sum(td->io_bytes);
1576                 if (diff < td_max_bs(td))
1577                         return false;
1578
1579                 if (fio_files_done(td) && !td->o.io_size)
1580                         return false;
1581
1582                 return true;
1583         }
1584
1585         return false;
1586 }
1587
1588 static int exec_string(struct thread_options *o, const char *string,
1589                        const char *mode)
1590 {
1591         int ret;
1592         char *str;
1593
1594         if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1595                 return -1;
1596
1597         log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1598                  o->name, mode);
1599         ret = system(str);
1600         if (ret == -1)
1601                 log_err("fio: exec of cmd <%s> failed\n", str);
1602
1603         free(str);
1604         return ret;
1605 }
1606
1607 /*
1608  * Dry run to compute correct state of numberio for verification.
1609  */
1610 static uint64_t do_dry_run(struct thread_data *td)
1611 {
1612         td_set_runstate(td, TD_RUNNING);
1613
1614         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1615                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1616                 struct io_u *io_u;
1617                 int ret;
1618
1619                 if (td->terminate || td->done)
1620                         break;
1621
1622                 io_u = get_io_u(td);
1623                 if (IS_ERR_OR_NULL(io_u))
1624                         break;
1625
1626                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1627                 io_u->error = 0;
1628                 io_u->resid = 0;
1629                 if (ddir_rw(acct_ddir(io_u)))
1630                         td->io_issues[acct_ddir(io_u)]++;
1631                 if (ddir_rw(io_u->ddir)) {
1632                         io_u_mark_depth(td, 1);
1633                         td->ts.total_io_u[io_u->ddir]++;
1634                 }
1635
1636                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1637                     td->o.do_verify &&
1638                     td->o.verify != VERIFY_NONE &&
1639                     !td->o.experimental_verify)
1640                         log_io_piece(td, io_u);
1641
1642                 ret = io_u_sync_complete(td, io_u);
1643                 (void) ret;
1644         }
1645
1646         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1647 }
1648
1649 struct fork_data {
1650         struct thread_data *td;
1651         struct sk_out *sk_out;
1652 };
1653
1654 /*
1655  * Entry point for the thread based jobs. The process based jobs end up
1656  * here as well, after a little setup.
1657  */
1658 static void *thread_main(void *data)
1659 {
1660         struct fork_data *fd = data;
1661         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1662         struct thread_data *td = fd->td;
1663         struct thread_options *o = &td->o;
1664         struct sk_out *sk_out = fd->sk_out;
1665         uint64_t bytes_done[DDIR_RWDIR_CNT];
1666         int deadlock_loop_cnt;
1667         bool clear_state;
1668         int ret;
1669
1670         sk_out_assign(sk_out);
1671         free(fd);
1672
1673         if (!o->use_thread) {
1674                 setsid();
1675                 td->pid = getpid();
1676         } else
1677                 td->pid = gettid();
1678
1679         fio_local_clock_init();
1680
1681         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1682
1683         if (is_backend)
1684                 fio_server_send_start(td);
1685
1686         INIT_FLIST_HEAD(&td->io_log_list);
1687         INIT_FLIST_HEAD(&td->io_hist_list);
1688         INIT_FLIST_HEAD(&td->verify_list);
1689         INIT_FLIST_HEAD(&td->trim_list);
1690         td->io_hist_tree = RB_ROOT;
1691
1692         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1693         if (ret) {
1694                 td_verror(td, ret, "mutex_cond_init_pshared");
1695                 goto err;
1696         }
1697         ret = cond_init_pshared(&td->verify_cond);
1698         if (ret) {
1699                 td_verror(td, ret, "mutex_cond_pshared");
1700                 goto err;
1701         }
1702
1703         td_set_runstate(td, TD_INITIALIZED);
1704         dprint(FD_MUTEX, "up startup_sem\n");
1705         fio_sem_up(startup_sem);
1706         dprint(FD_MUTEX, "wait on td->sem\n");
1707         fio_sem_down(td->sem);
1708         dprint(FD_MUTEX, "done waiting on td->sem\n");
1709
1710         /*
1711          * A new gid requires privilege, so we need to do this before setting
1712          * the uid.
1713          */
1714         if (o->gid != -1U && setgid(o->gid)) {
1715                 td_verror(td, errno, "setgid");
1716                 goto err;
1717         }
1718         if (o->uid != -1U && setuid(o->uid)) {
1719                 td_verror(td, errno, "setuid");
1720                 goto err;
1721         }
1722
1723         td_zone_gen_index(td);
1724
1725         /*
1726          * Do this early, we don't want the compress threads to be limited
1727          * to the same CPUs as the IO workers. So do this before we set
1728          * any potential CPU affinity
1729          */
1730         if (iolog_compress_init(td, sk_out))
1731                 goto err;
1732
1733         /*
1734          * If we have a gettimeofday() thread, make sure we exclude that
1735          * thread from this job
1736          */
1737         if (o->gtod_cpu)
1738                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1739
1740         /*
1741          * Set affinity first, in case it has an impact on the memory
1742          * allocations.
1743          */
1744         if (fio_option_is_set(o, cpumask)) {
1745                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1746                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1747                         if (!ret) {
1748                                 log_err("fio: no CPUs set\n");
1749                                 log_err("fio: Try increasing number of available CPUs\n");
1750                                 td_verror(td, EINVAL, "cpus_split");
1751                                 goto err;
1752                         }
1753                 }
1754                 ret = fio_setaffinity(td->pid, o->cpumask);
1755                 if (ret == -1) {
1756                         td_verror(td, errno, "cpu_set_affinity");
1757                         goto err;
1758                 }
1759         }
1760
1761 #ifdef CONFIG_LIBNUMA
1762         /* numa node setup */
1763         if (fio_option_is_set(o, numa_cpunodes) ||
1764             fio_option_is_set(o, numa_memnodes)) {
1765                 struct bitmask *mask;
1766
1767                 if (numa_available() < 0) {
1768                         td_verror(td, errno, "Does not support NUMA API\n");
1769                         goto err;
1770                 }
1771
1772                 if (fio_option_is_set(o, numa_cpunodes)) {
1773                         mask = numa_parse_nodestring(o->numa_cpunodes);
1774                         ret = numa_run_on_node_mask(mask);
1775                         numa_free_nodemask(mask);
1776                         if (ret == -1) {
1777                                 td_verror(td, errno, \
1778                                         "numa_run_on_node_mask failed\n");
1779                                 goto err;
1780                         }
1781                 }
1782
1783                 if (fio_option_is_set(o, numa_memnodes)) {
1784                         mask = NULL;
1785                         if (o->numa_memnodes)
1786                                 mask = numa_parse_nodestring(o->numa_memnodes);
1787
1788                         switch (o->numa_mem_mode) {
1789                         case MPOL_INTERLEAVE:
1790                                 numa_set_interleave_mask(mask);
1791                                 break;
1792                         case MPOL_BIND:
1793                                 numa_set_membind(mask);
1794                                 break;
1795                         case MPOL_LOCAL:
1796                                 numa_set_localalloc();
1797                                 break;
1798                         case MPOL_PREFERRED:
1799                                 numa_set_preferred(o->numa_mem_prefer_node);
1800                                 break;
1801                         case MPOL_DEFAULT:
1802                         default:
1803                                 break;
1804                         }
1805
1806                         if (mask)
1807                                 numa_free_nodemask(mask);
1808
1809                 }
1810         }
1811 #endif
1812
1813         if (fio_pin_memory(td))
1814                 goto err;
1815
1816         /*
1817          * May alter parameters that init_io_u() will use, so we need to
1818          * do this first.
1819          */
1820         if (!init_iolog(td))
1821                 goto err;
1822
1823         /* ioprio_set() has to be done before td_io_init() */
1824         if (fio_option_is_set(o, ioprio) ||
1825             fio_option_is_set(o, ioprio_class) ||
1826             fio_option_is_set(o, ioprio_hint)) {
1827                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class,
1828                                  o->ioprio, o->ioprio_hint);
1829                 if (ret == -1) {
1830                         td_verror(td, errno, "ioprio_set");
1831                         goto err;
1832                 }
1833                 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio,
1834                                           o->ioprio_hint);
1835                 td->ts.ioprio = td->ioprio;
1836         }
1837
1838         if (td_io_init(td))
1839                 goto err;
1840
1841         if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
1842                 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1843                          "are selected, queue depth will be capped at 1\n");
1844         }
1845
1846         if (init_io_u(td))
1847                 goto err;
1848
1849         if (td->io_ops->post_init && td->io_ops->post_init(td))
1850                 goto err;
1851
1852         if (o->verify_async && verify_async_init(td))
1853                 goto err;
1854
1855         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1856                 goto err;
1857
1858         errno = 0;
1859         if (nice(o->nice) == -1 && errno != 0) {
1860                 td_verror(td, errno, "nice");
1861                 goto err;
1862         }
1863
1864         if (o->ioscheduler && switch_ioscheduler(td))
1865                 goto err;
1866
1867         if (!o->create_serialize && setup_files(td))
1868                 goto err;
1869
1870         if (!init_random_map(td))
1871                 goto err;
1872
1873         if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1874                 goto err;
1875
1876         if (o->pre_read && !pre_read_files(td))
1877                 goto err;
1878
1879         fio_verify_init(td);
1880
1881         if (rate_submit_init(td, sk_out))
1882                 goto err;
1883
1884         set_epoch_time(td, o->log_alternate_epoch_clock_id, o->job_start_clock_id);
1885         fio_getrusage(&td->ru_start);
1886         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1887         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1888         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1889
1890         init_thinktime(td);
1891
1892         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1893                         o->ratemin[DDIR_TRIM]) {
1894                 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1895                                         sizeof(td->bw_sample_time));
1896                 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1897                                         sizeof(td->bw_sample_time));
1898                 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1899                                         sizeof(td->bw_sample_time));
1900         }
1901
1902         memset(bytes_done, 0, sizeof(bytes_done));
1903         clear_state = false;
1904
1905         while (keep_running(td)) {
1906                 uint64_t verify_bytes;
1907
1908                 fio_gettime(&td->start, NULL);
1909                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1910
1911                 if (clear_state) {
1912                         clear_io_state(td, 0);
1913
1914                         if (o->unlink_each_loop && unlink_all_files(td))
1915                                 break;
1916                 }
1917
1918                 prune_io_piece_log(td);
1919
1920                 if (td->o.verify_only && td_write(td))
1921                         verify_bytes = do_dry_run(td);
1922                 else {
1923                         if (!td->o.rand_repeatable)
1924                                 /* save verify rand state to replay hdr seeds later at verify */
1925                                 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
1926                         do_io(td, bytes_done);
1927                         if (!td->o.rand_repeatable)
1928                                 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
1929                         if (!ddir_rw_sum(bytes_done)) {
1930                                 fio_mark_td_terminate(td);
1931                                 verify_bytes = 0;
1932                         } else {
1933                                 verify_bytes = bytes_done[DDIR_WRITE] +
1934                                                 bytes_done[DDIR_TRIM];
1935                         }
1936                 }
1937
1938                 /*
1939                  * If we took too long to shut down, the main thread could
1940                  * already consider us reaped/exited. If that happens, break
1941                  * out and clean up.
1942                  */
1943                 if (td->runstate >= TD_EXITED)
1944                         break;
1945
1946                 clear_state = true;
1947
1948                 /*
1949                  * Make sure we've successfully updated the rusage stats
1950                  * before waiting on the stat mutex. Otherwise we could have
1951                  * the stat thread holding stat mutex and waiting for
1952                  * the rusage_sem, which would never get upped because
1953                  * this thread is waiting for the stat mutex.
1954                  */
1955                 deadlock_loop_cnt = 0;
1956                 do {
1957                         check_update_rusage(td);
1958                         if (!fio_sem_down_trylock(stat_sem))
1959                                 break;
1960                         usleep(1000);
1961                         if (deadlock_loop_cnt++ > 5000) {
1962                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1963                                 td->error = EDEADLK;
1964                                 goto err;
1965                         }
1966                 } while (1);
1967
1968                 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1969                         ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
1970                         update_runtime(td, elapsed_us, DDIR_READ);
1971                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1972                         update_runtime(td, elapsed_us, DDIR_WRITE);
1973                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1974                         update_runtime(td, elapsed_us, DDIR_TRIM);
1975                 fio_gettime(&td->start, NULL);
1976                 fio_sem_up(stat_sem);
1977
1978                 if (td->error || td->terminate)
1979                         break;
1980
1981                 if (!o->do_verify ||
1982                     o->verify == VERIFY_NONE ||
1983                     td_ioengine_flagged(td, FIO_UNIDIR))
1984                         continue;
1985
1986                 clear_io_state(td, 0);
1987
1988                 fio_gettime(&td->start, NULL);
1989
1990                 do_verify(td, verify_bytes);
1991
1992                 /*
1993                  * See comment further up for why this is done here.
1994                  */
1995                 check_update_rusage(td);
1996
1997                 fio_sem_down(stat_sem);
1998                 update_runtime(td, elapsed_us, DDIR_READ);
1999                 fio_gettime(&td->start, NULL);
2000                 fio_sem_up(stat_sem);
2001
2002                 if (td->error || td->terminate)
2003                         break;
2004         }
2005
2006         /*
2007          * Acquire this lock if we were doing overlap checking in
2008          * offload mode so that we don't clean up this job while
2009          * another thread is checking its io_u's for overlap
2010          */
2011         if (td_offload_overlap(td)) {
2012                 int res;
2013
2014                 res = pthread_mutex_lock(&overlap_check);
2015                 if (res) {
2016                         td->error = errno;
2017                         goto err;
2018                 }
2019         }
2020         td_set_runstate(td, TD_FINISHING);
2021         if (td_offload_overlap(td)) {
2022                 int res;
2023
2024                 res = pthread_mutex_unlock(&overlap_check);
2025                 if (res) {
2026                         td->error = errno;
2027                         goto err;
2028                 }
2029         }
2030
2031         update_rusage_stat(td);
2032         td->ts.total_run_time = mtime_since_now(&td->epoch);
2033         for_each_rw_ddir(ddir) {
2034                 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2035         }
2036
2037         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
2038             (td->o.verify != VERIFY_NONE && td_write(td)))
2039                 verify_save_state(td->thread_number);
2040
2041         fio_unpin_memory(td);
2042
2043         td_writeout_logs(td, true);
2044
2045         iolog_compress_exit(td);
2046         rate_submit_exit(td);
2047
2048         if (o->exec_postrun)
2049                 exec_string(o, o->exec_postrun, "postrun");
2050
2051         if (exitall_on_terminate || (o->exitall_error && td->error))
2052                 fio_terminate_threads(td->groupid, td->o.exit_what);
2053
2054 err:
2055         if (td->error)
2056                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
2057                                                         td->verror);
2058
2059         if (o->verify_async)
2060                 verify_async_exit(td);
2061
2062         close_and_free_files(td);
2063         cleanup_io_u(td);
2064         close_ioengine(td);
2065         cgroup_shutdown(td, cgroup_mnt);
2066         verify_free_state(td);
2067         td_zone_free_index(td);
2068
2069         if (fio_option_is_set(o, cpumask)) {
2070                 ret = fio_cpuset_exit(&o->cpumask);
2071                 if (ret)
2072                         td_verror(td, ret, "fio_cpuset_exit");
2073         }
2074
2075         /*
2076          * do this very late, it will log file closing as well
2077          */
2078         if (o->write_iolog_file)
2079                 write_iolog_close(td);
2080         if (td->io_log_rfile)
2081                 fclose(td->io_log_rfile);
2082
2083         td_set_runstate(td, TD_EXITED);
2084
2085         /*
2086          * Do this last after setting our runstate to exited, so we
2087          * know that the stat thread is signaled.
2088          */
2089         check_update_rusage(td);
2090
2091         sk_out_drop();
2092         return (void *) (uintptr_t) td->error;
2093 }
2094
2095 /*
2096  * Run over the job map and reap the threads that have exited, if any.
2097  */
2098 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2099                          uint64_t *m_rate)
2100 {
2101         unsigned int cputhreads, realthreads, pending;
2102         int ret;
2103
2104         /*
2105          * reap exited threads (TD_EXITED -> TD_REAPED)
2106          */
2107         realthreads = pending = cputhreads = 0;
2108         for_each_td(td) {
2109                 int flags = 0, status;
2110
2111                 if (!strcmp(td->o.ioengine, "cpuio"))
2112                         cputhreads++;
2113                 else
2114                         realthreads++;
2115
2116                 if (!td->pid) {
2117                         pending++;
2118                         continue;
2119                 }
2120                 if (td->runstate == TD_REAPED)
2121                         continue;
2122                 if (td->o.use_thread) {
2123                         if (td->runstate == TD_EXITED) {
2124                                 td_set_runstate(td, TD_REAPED);
2125                                 goto reaped;
2126                         }
2127                         continue;
2128                 }
2129
2130                 flags = WNOHANG;
2131                 if (td->runstate == TD_EXITED)
2132                         flags = 0;
2133
2134                 /*
2135                  * check if someone quit or got killed in an unusual way
2136                  */
2137                 ret = waitpid(td->pid, &status, flags);
2138                 if (ret < 0) {
2139                         if (errno == ECHILD) {
2140                                 log_err("fio: pid=%d disappeared %d\n",
2141                                                 (int) td->pid, td->runstate);
2142                                 td->sig = ECHILD;
2143                                 td_set_runstate(td, TD_REAPED);
2144                                 goto reaped;
2145                         }
2146                         perror("waitpid");
2147                 } else if (ret == td->pid) {
2148                         if (WIFSIGNALED(status)) {
2149                                 int sig = WTERMSIG(status);
2150
2151                                 if (sig != SIGTERM && sig != SIGUSR2)
2152                                         log_err("fio: pid=%d, got signal=%d\n",
2153                                                         (int) td->pid, sig);
2154                                 td->sig = sig;
2155                                 td_set_runstate(td, TD_REAPED);
2156                                 goto reaped;
2157                         }
2158                         if (WIFEXITED(status)) {
2159                                 if (WEXITSTATUS(status) && !td->error)
2160                                         td->error = WEXITSTATUS(status);
2161
2162                                 td_set_runstate(td, TD_REAPED);
2163                                 goto reaped;
2164                         }
2165                 }
2166
2167                 /*
2168                  * If the job is stuck, do a forceful timeout of it and
2169                  * move on.
2170                  */
2171                 if (td->terminate &&
2172                     td->runstate < TD_FSYNCING &&
2173                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2174                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2175                                 "%lu seconds, it appears to be stuck. Doing "
2176                                 "forceful exit of this job.\n",
2177                                 td->o.name, td->runstate,
2178                                 (unsigned long) time_since_now(&td->terminate_time));
2179                         td_set_runstate(td, TD_REAPED);
2180                         goto reaped;
2181                 }
2182
2183                 /*
2184                  * thread is not dead, continue
2185                  */
2186                 pending++;
2187                 continue;
2188 reaped:
2189                 (*nr_running)--;
2190                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2191                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2192                 if (!td->pid)
2193                         pending--;
2194
2195                 if (td->error)
2196                         exit_value++;
2197
2198                 done_secs += mtime_since_now(&td->epoch) / 1000;
2199                 profile_td_exit(td);
2200                 flow_exit_job(td);
2201         } end_for_each();
2202
2203         if (*nr_running == cputhreads && !pending && realthreads)
2204                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2205 }
2206
2207 static bool __check_trigger_file(void)
2208 {
2209         struct stat sb;
2210
2211         if (!trigger_file)
2212                 return false;
2213
2214         if (stat(trigger_file, &sb))
2215                 return false;
2216
2217         if (unlink(trigger_file) < 0)
2218                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2219                                                         strerror(errno));
2220
2221         return true;
2222 }
2223
2224 static bool trigger_timedout(void)
2225 {
2226         if (trigger_timeout)
2227                 if (time_since_genesis() >= trigger_timeout) {
2228                         trigger_timeout = 0;
2229                         return true;
2230                 }
2231
2232         return false;
2233 }
2234
2235 void exec_trigger(const char *cmd)
2236 {
2237         int ret;
2238
2239         if (!cmd || cmd[0] == '\0')
2240                 return;
2241
2242         ret = system(cmd);
2243         if (ret == -1)
2244                 log_err("fio: failed executing %s trigger\n", cmd);
2245 }
2246
2247 void check_trigger_file(void)
2248 {
2249         if (__check_trigger_file() || trigger_timedout()) {
2250                 if (nr_clients)
2251                         fio_clients_send_trigger(trigger_remote_cmd);
2252                 else {
2253                         verify_save_state(IO_LIST_ALL);
2254                         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2255                         exec_trigger(trigger_cmd);
2256                 }
2257         }
2258 }
2259
2260 static int fio_verify_load_state(struct thread_data *td)
2261 {
2262         int ret;
2263
2264         if (!td->o.verify_state)
2265                 return 0;
2266
2267         if (is_backend) {
2268                 void *data;
2269
2270                 ret = fio_server_get_verify_state(td->o.name,
2271                                         td->thread_number - 1, &data);
2272                 if (!ret)
2273                         verify_assign_state(td, data);
2274         } else {
2275                 char prefix[PATH_MAX];
2276
2277                 if (aux_path)
2278                         sprintf(prefix, "%s%clocal", aux_path,
2279                                         FIO_OS_PATH_SEPARATOR);
2280                 else
2281                         strcpy(prefix, "local");
2282                 ret = verify_load_state(td, prefix);
2283         }
2284
2285         return ret;
2286 }
2287
2288 static void do_usleep(unsigned int usecs)
2289 {
2290         check_for_running_stats();
2291         check_trigger_file();
2292         usleep(usecs);
2293 }
2294
2295 static bool check_mount_writes(struct thread_data *td)
2296 {
2297         struct fio_file *f;
2298         unsigned int i;
2299
2300         if (!td_write(td) || td->o.allow_mounted_write)
2301                 return false;
2302
2303         /*
2304          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2305          * are mkfs'd and mounted.
2306          */
2307         for_each_file(td, f, i) {
2308 #ifdef FIO_HAVE_CHARDEV_SIZE
2309                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2310 #else
2311                 if (f->filetype != FIO_TYPE_BLOCK)
2312 #endif
2313                         continue;
2314                 if (device_is_mounted(f->file_name))
2315                         goto mounted;
2316         }
2317
2318         return false;
2319 mounted:
2320         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2321         return true;
2322 }
2323
2324 static bool waitee_running(struct thread_data *me)
2325 {
2326         const char *waitee = me->o.wait_for;
2327         const char *self = me->o.name;
2328
2329         if (!waitee)
2330                 return false;
2331
2332         for_each_td(td) {
2333                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2334                         continue;
2335
2336                 if (td->runstate < TD_EXITED) {
2337                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2338                                         self, td->o.name,
2339                                         runstate_to_name(td->runstate));
2340                         return true;
2341                 }
2342         } end_for_each();
2343
2344         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2345         return false;
2346 }
2347
2348 /*
2349  * Main function for kicking off and reaping jobs, as needed.
2350  */
2351 static void run_threads(struct sk_out *sk_out)
2352 {
2353         struct thread_data *td;
2354         unsigned int i, todo, nr_running, nr_started;
2355         uint64_t m_rate, t_rate;
2356         uint64_t spent;
2357
2358         if (fio_gtod_offload && fio_start_gtod_thread())
2359                 return;
2360
2361         fio_idle_prof_init();
2362
2363         set_sig_handlers();
2364
2365         nr_thread = nr_process = 0;
2366         for_each_td(td) {
2367                 if (check_mount_writes(td))
2368                         return;
2369                 if (td->o.use_thread)
2370                         nr_thread++;
2371                 else
2372                         nr_process++;
2373         } end_for_each();
2374
2375         if (output_format & FIO_OUTPUT_NORMAL) {
2376                 struct buf_output out;
2377
2378                 buf_output_init(&out);
2379                 __log_buf(&out, "Starting ");
2380                 if (nr_thread)
2381                         __log_buf(&out, "%d thread%s", nr_thread,
2382                                                 nr_thread > 1 ? "s" : "");
2383                 if (nr_process) {
2384                         if (nr_thread)
2385                                 __log_buf(&out, " and ");
2386                         __log_buf(&out, "%d process%s", nr_process,
2387                                                 nr_process > 1 ? "es" : "");
2388                 }
2389                 __log_buf(&out, "\n");
2390                 log_info_buf(out.buf, out.buflen);
2391                 buf_output_free(&out);
2392         }
2393
2394         todo = thread_number;
2395         nr_running = 0;
2396         nr_started = 0;
2397         m_rate = t_rate = 0;
2398
2399         for_each_td(td) {
2400                 print_status_init(td->thread_number - 1);
2401
2402                 if (!td->o.create_serialize)
2403                         continue;
2404
2405                 if (fio_verify_load_state(td))
2406                         goto reap;
2407
2408                 /*
2409                  * do file setup here so it happens sequentially,
2410                  * we don't want X number of threads getting their
2411                  * client data interspersed on disk
2412                  */
2413                 if (setup_files(td)) {
2414 reap:
2415                         exit_value++;
2416                         if (td->error)
2417                                 log_err("fio: pid=%d, err=%d/%s\n",
2418                                         (int) td->pid, td->error, td->verror);
2419                         td_set_runstate(td, TD_REAPED);
2420                         todo--;
2421                 } else {
2422                         struct fio_file *f;
2423                         unsigned int j;
2424
2425                         /*
2426                          * for sharing to work, each job must always open
2427                          * its own files. so close them, if we opened them
2428                          * for creation
2429                          */
2430                         for_each_file(td, f, j) {
2431                                 if (fio_file_open(f))
2432                                         td_io_close_file(td, f);
2433                         }
2434                 }
2435         } end_for_each();
2436
2437         /* start idle threads before io threads start to run */
2438         fio_idle_prof_start();
2439
2440         set_genesis_time();
2441
2442         while (todo) {
2443                 struct thread_data *map[REAL_MAX_JOBS];
2444                 struct timespec this_start;
2445                 int this_jobs = 0, left;
2446                 struct fork_data *fd;
2447
2448                 /*
2449                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2450                  */
2451                 for_each_td(td) {
2452                         if (td->runstate != TD_NOT_CREATED)
2453                                 continue;
2454
2455                         /*
2456                          * never got a chance to start, killed by other
2457                          * thread for some reason
2458                          */
2459                         if (td->terminate) {
2460                                 todo--;
2461                                 continue;
2462                         }
2463
2464                         if (td->o.start_delay) {
2465                                 spent = utime_since_genesis();
2466
2467                                 if (td->o.start_delay > spent)
2468                                         continue;
2469                         }
2470
2471                         if (td->o.stonewall && (nr_started || nr_running)) {
2472                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2473                                                         td->o.name);
2474                                 break;
2475                         }
2476
2477                         if (waitee_running(td)) {
2478                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2479                                                 td->o.name, td->o.wait_for);
2480                                 continue;
2481                         }
2482
2483                         init_disk_util(td);
2484
2485                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2486                         td->update_rusage = 0;
2487
2488                         /*
2489                          * Set state to created. Thread will transition
2490                          * to TD_INITIALIZED when it's done setting up.
2491                          */
2492                         td_set_runstate(td, TD_CREATED);
2493                         map[this_jobs++] = td;
2494                         nr_started++;
2495
2496                         fd = calloc(1, sizeof(*fd));
2497                         fd->td = td;
2498                         fd->sk_out = sk_out;
2499
2500                         if (td->o.use_thread) {
2501                                 int ret;
2502
2503                                 dprint(FD_PROCESS, "will pthread_create\n");
2504                                 ret = pthread_create(&td->thread, NULL,
2505                                                         thread_main, fd);
2506                                 if (ret) {
2507                                         log_err("pthread_create: %s\n",
2508                                                         strerror(ret));
2509                                         free(fd);
2510                                         nr_started--;
2511                                         break;
2512                                 }
2513                                 fd = NULL;
2514                                 ret = pthread_detach(td->thread);
2515                                 if (ret)
2516                                         log_err("pthread_detach: %s",
2517                                                         strerror(ret));
2518                         } else {
2519                                 pid_t pid;
2520                                 void *eo;
2521                                 dprint(FD_PROCESS, "will fork\n");
2522                                 eo = td->eo;
2523                                 read_barrier();
2524                                 pid = fork();
2525                                 if (!pid) {
2526                                         int ret;
2527
2528                                         ret = (int)(uintptr_t)thread_main(fd);
2529                                         _exit(ret);
2530                                 } else if (__td_index == fio_debug_jobno)
2531                                         *fio_debug_jobp = pid;
2532                                 free(eo);
2533                                 free(fd);
2534                                 fd = NULL;
2535                         }
2536                         dprint(FD_MUTEX, "wait on startup_sem\n");
2537                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2538                                 log_err("fio: job startup hung? exiting.\n");
2539                                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2540                                 fio_abort = true;
2541                                 nr_started--;
2542                                 free(fd);
2543                                 break;
2544                         }
2545                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2546                 } end_for_each();
2547
2548                 /*
2549                  * Wait for the started threads to transition to
2550                  * TD_INITIALIZED.
2551                  */
2552                 fio_gettime(&this_start, NULL);
2553                 left = this_jobs;
2554                 while (left && !fio_abort) {
2555                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2556                                 break;
2557
2558                         do_usleep(100000);
2559
2560                         for (i = 0; i < this_jobs; i++) {
2561                                 td = map[i];
2562                                 if (!td)
2563                                         continue;
2564                                 if (td->runstate == TD_INITIALIZED) {
2565                                         map[i] = NULL;
2566                                         left--;
2567                                 } else if (td->runstate >= TD_EXITED) {
2568                                         map[i] = NULL;
2569                                         left--;
2570                                         todo--;
2571                                         nr_running++; /* work-around... */
2572                                 }
2573                         }
2574                 }
2575
2576                 if (left) {
2577                         log_err("fio: %d job%s failed to start\n", left,
2578                                         left > 1 ? "s" : "");
2579                         for (i = 0; i < this_jobs; i++) {
2580                                 td = map[i];
2581                                 if (!td)
2582                                         continue;
2583                                 kill(td->pid, SIGTERM);
2584                         }
2585                         break;
2586                 }
2587
2588                 /*
2589                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2590                  */
2591                 for_each_td(td) {
2592                         if (td->runstate != TD_INITIALIZED)
2593                                 continue;
2594
2595                         if (in_ramp_time(td))
2596                                 td_set_runstate(td, TD_RAMP);
2597                         else
2598                                 td_set_runstate(td, TD_RUNNING);
2599                         nr_running++;
2600                         nr_started--;
2601                         m_rate += ddir_rw_sum(td->o.ratemin);
2602                         t_rate += ddir_rw_sum(td->o.rate);
2603                         todo--;
2604                         fio_sem_up(td->sem);
2605                 } end_for_each();
2606
2607                 reap_threads(&nr_running, &t_rate, &m_rate);
2608
2609                 if (todo)
2610                         do_usleep(100000);
2611         }
2612
2613         while (nr_running) {
2614                 reap_threads(&nr_running, &t_rate, &m_rate);
2615                 do_usleep(10000);
2616         }
2617
2618         fio_idle_prof_stop();
2619
2620         update_io_ticks();
2621 }
2622
2623 static void free_disk_util(void)
2624 {
2625         disk_util_prune_entries();
2626         helper_thread_destroy();
2627 }
2628
2629 int fio_backend(struct sk_out *sk_out)
2630 {
2631         int i;
2632         if (exec_profile) {
2633                 if (load_profile(exec_profile))
2634                         return 1;
2635                 free(exec_profile);
2636                 exec_profile = NULL;
2637         }
2638         if (!thread_number)
2639                 return 0;
2640
2641         if (write_bw_log) {
2642                 struct log_params p = {
2643                         .log_type = IO_LOG_TYPE_BW,
2644                 };
2645
2646                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2647                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2648                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2649         }
2650
2651         if (init_global_dedupe_working_set_seeds()) {
2652                 log_err("fio: failed to initialize global dedupe working set\n");
2653                 return 1;
2654         }
2655
2656         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2657         if (!sk_out)
2658                 is_local_backend = true;
2659         if (startup_sem == NULL)
2660                 return 1;
2661
2662         set_genesis_time();
2663         stat_init();
2664         if (helper_thread_create(startup_sem, sk_out))
2665                 log_err("fio: failed to create helper thread\n");
2666
2667         cgroup_list = smalloc(sizeof(*cgroup_list));
2668         if (cgroup_list)
2669                 INIT_FLIST_HEAD(cgroup_list);
2670
2671         run_threads(sk_out);
2672
2673         helper_thread_exit();
2674
2675         if (!fio_abort) {
2676                 __show_run_stats();
2677                 if (write_bw_log) {
2678                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2679                                 struct io_log *log = agg_io_log[i];
2680
2681                                 flush_log(log, false);
2682                                 free_log(log);
2683                         }
2684                 }
2685         }
2686
2687         for_each_td(td) {
2688                 struct thread_stat *ts = &td->ts;
2689
2690                 free_clat_prio_stats(ts);
2691                 steadystate_free(td);
2692                 fio_options_free(td);
2693                 fio_dump_options_free(td);
2694                 if (td->rusage_sem) {
2695                         fio_sem_remove(td->rusage_sem);
2696                         td->rusage_sem = NULL;
2697                 }
2698                 fio_sem_remove(td->sem);
2699                 td->sem = NULL;
2700         } end_for_each();
2701
2702         free_disk_util();
2703         if (cgroup_list) {
2704                 cgroup_kill(cgroup_list);
2705                 sfree(cgroup_list);
2706         }
2707
2708         fio_sem_remove(startup_sem);
2709         stat_exit();
2710         return exit_value;
2711 }