backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <string.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <inttypes.h>
  29 #include <sys/stat.h>
  30 #include <sys/wait.h>
  31 #include <math.h>
  32
  33 #include "fio.h"
  34 #include "smalloc.h"
  35 #include "verify.h"
  36 #include "diskutil.h"
  37 #include "cgroup.h"
  38 #include "profile.h"
  39 #include "lib/rand.h"
  40 #include "lib/memalign.h"
  41 #include "server.h"
  42 #include "lib/getrusage.h"
  43 #include "idletime.h"
  44 #include "err.h"
  45 #include "workqueue.h"
  46 #include "lib/mountcheck.h"
  47 #include "rate-submit.h"
  48 #include "helper_thread.h"
  49 #include "pshared.h"
  50
  51 static struct fio_sem *startup_sem;
  52 static struct flist_head *cgroup_list;
  53 static struct cgroup_mnt *cgroup_mnt;
  54 static int exit_value;
  55 static volatile int fio_abort;
  56 static unsigned int nr_process = 0;
  57 static unsigned int nr_thread = 0;
  58
  59 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  60
  61 int groupid = 0;
  62 unsigned int thread_number = 0;
  63 unsigned int stat_number = 0;
  64 int shm_id = 0;
  65 int temp_stall_ts;
  66 unsigned long done_secs = 0;
  67
  68 #define JOB_START_TIMEOUT       (5 * 1000)
  69
  70 static void sig_int(int sig)
  71 {
  72         if (threads) {
  73                 if (is_backend)
  74                         fio_server_got_signal(sig);
  75                 else {
  76                         log_info("\nfio: terminating on signal %d\n", sig);
  77                         log_info_flush();
  78                         exit_value = 128;
  79                 }
  80
  81                 fio_terminate_threads(TERMINATE_ALL);
  82         }
  83 }
  84
  85 void sig_show_status(int sig)
  86 {
  87         show_running_run_stats();
  88 }
  89
  90 static void set_sig_handlers(void)
  91 {
  92         struct sigaction act;
  93
  94         memset(&act, 0, sizeof(act));
  95         act.sa_handler = sig_int;
  96         act.sa_flags = SA_RESTART;
  97         sigaction(SIGINT, &act, NULL);
  98
  99         memset(&act, 0, sizeof(act));
 100         act.sa_handler = sig_int;
 101         act.sa_flags = SA_RESTART;
 102         sigaction(SIGTERM, &act, NULL);
 103
 104 /* Windows uses SIGBREAK as a quit signal from other applications */
 105 #ifdef WIN32
 106         memset(&act, 0, sizeof(act));
 107         act.sa_handler = sig_int;
 108         act.sa_flags = SA_RESTART;
 109         sigaction(SIGBREAK, &act, NULL);
 110 #endif
 111
 112         memset(&act, 0, sizeof(act));
 113         act.sa_handler = sig_show_status;
 114         act.sa_flags = SA_RESTART;
 115         sigaction(SIGUSR1, &act, NULL);
 116
 117         if (is_backend) {
 118                 memset(&act, 0, sizeof(act));
 119                 act.sa_handler = sig_int;
 120                 act.sa_flags = SA_RESTART;
 121                 sigaction(SIGPIPE, &act, NULL);
 122         }
 123 }
 124
 125 /*
 126  * Check if we are above the minimum rate given.
 127  */
 128 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
 129                              enum fio_ddir ddir)
 130 {
 131         unsigned long long bytes = 0;
 132         unsigned long iops = 0;
 133         unsigned long spent;
 134         unsigned long rate;
 135         unsigned int ratemin = 0;
 136         unsigned int rate_iops = 0;
 137         unsigned int rate_iops_min = 0;
 138
 139         assert(ddir_rw(ddir));
 140
 141         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 142                 return false;
 143
 144         /*
 145          * allow a 2 second settle period in the beginning
 146          */
 147         if (mtime_since(&td->start, now) < 2000)
 148                 return false;
 149
 150         iops += td->this_io_blocks[ddir];
 151         bytes += td->this_io_bytes[ddir];
 152         ratemin += td->o.ratemin[ddir];
 153         rate_iops += td->o.rate_iops[ddir];
 154         rate_iops_min += td->o.rate_iops_min[ddir];
 155
 156         /*
 157          * if rate blocks is set, sample is running
 158          */
 159         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 160                 spent = mtime_since(&td->lastrate[ddir], now);
 161                 if (spent < td->o.ratecycle)
 162                         return false;
 163
 164                 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
 165                         /*
 166                          * check bandwidth specified rate
 167                          */
 168                         if (bytes < td->rate_bytes[ddir]) {
 169                                 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
 170                                         td->o.name, ratemin, bytes);
 171                                 return true;
 172                         } else {
 173                                 if (spent)
 174                                         rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 175                                 else
 176                                         rate = 0;
 177
 178                                 if (rate < ratemin ||
 179                                     bytes < td->rate_bytes[ddir]) {
 180                                         log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
 181                                                 td->o.name, ratemin, rate);
 182                                         return true;
 183                                 }
 184                         }
 185                 } else {
 186                         /*
 187                          * checks iops specified rate
 188                          */
 189                         if (iops < rate_iops) {
 190                                 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
 191                                                 td->o.name, rate_iops, iops);
 192                                 return true;
 193                         } else {
 194                                 if (spent)
 195                                         rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 196                                 else
 197                                         rate = 0;
 198
 199                                 if (rate < rate_iops_min ||
 200                                     iops < td->rate_blocks[ddir]) {
 201                                         log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
 202                                                 td->o.name, rate_iops_min, rate);
 203                                         return true;
 204                                 }
 205                         }
 206                 }
 207         }
 208
 209         td->rate_bytes[ddir] = bytes;
 210         td->rate_blocks[ddir] = iops;
 211         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 212         return false;
 213 }
 214
 215 static bool check_min_rate(struct thread_data *td, struct timespec *now)
 216 {
 217         bool ret = false;
 218
 219         if (td->bytes_done[DDIR_READ])
 220                 ret |= __check_min_rate(td, now, DDIR_READ);
 221         if (td->bytes_done[DDIR_WRITE])
 222                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 223         if (td->bytes_done[DDIR_TRIM])
 224                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 225
 226         return ret;
 227 }
 228
 229 /*
 230  * When job exits, we can cancel the in-flight IO if we are using async
 231  * io. Attempt to do so.
 232  */
 233 static void cleanup_pending_aio(struct thread_data *td)
 234 {
 235         int r;
 236
 237         /*
 238          * get immediately available events, if any
 239          */
 240         r = io_u_queued_complete(td, 0);
 241         if (r < 0)
 242                 return;
 243
 244         /*
 245          * now cancel remaining active events
 246          */
 247         if (td->io_ops->cancel) {
 248                 struct io_u *io_u;
 249                 int i;
 250
 251                 io_u_qiter(&td->io_u_all, io_u, i) {
 252                         if (io_u->flags & IO_U_F_FLIGHT) {
 253                                 r = td->io_ops->cancel(td, io_u);
 254                                 if (!r)
 255                                         put_io_u(td, io_u);
 256                         }
 257                 }
 258         }
 259
 260         if (td->cur_depth)
 261                 r = io_u_queued_complete(td, td->cur_depth);
 262 }
 263
 264 /*
 265  * Helper to handle the final sync of a file. Works just like the normal
 266  * io path, just does everything sync.
 267  */
 268 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 269 {
 270         struct io_u *io_u = __get_io_u(td);
 271         enum fio_q_status ret;
 272
 273         if (!io_u)
 274                 return true;
 275
 276         io_u->ddir = DDIR_SYNC;
 277         io_u->file = f;
 278
 279         if (td_io_prep(td, io_u)) {
 280                 put_io_u(td, io_u);
 281                 return true;
 282         }
 283
 284 requeue:
 285         ret = td_io_queue(td, io_u);
 286         switch (ret) {
 287         case FIO_Q_QUEUED:
 288                 td_io_commit(td);
 289                 if (io_u_queued_complete(td, 1) < 0)
 290                         return true;
 291                 break;
 292         case FIO_Q_COMPLETED:
 293                 if (io_u->error) {
 294                         td_verror(td, io_u->error, "td_io_queue");
 295                         return true;
 296                 }
 297
 298                 if (io_u_sync_complete(td, io_u) < 0)
 299                         return true;
 300                 break;
 301         case FIO_Q_BUSY:
 302                 td_io_commit(td);
 303                 goto requeue;
 304         }
 305
 306         return false;
 307 }
 308
 309 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
 310 {
 311         int ret;
 312
 313         if (fio_file_open(f))
 314                 return fio_io_sync(td, f);
 315
 316         if (td_io_open_file(td, f))
 317                 return 1;
 318
 319         ret = fio_io_sync(td, f);
 320         td_io_close_file(td, f);
 321         return ret;
 322 }
 323
 324 static inline void __update_ts_cache(struct thread_data *td)
 325 {
 326         fio_gettime(&td->ts_cache, NULL);
 327 }
 328
 329 static inline void update_ts_cache(struct thread_data *td)
 330 {
 331         if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
 332                 __update_ts_cache(td);
 333 }
 334
 335 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 336 {
 337         if (in_ramp_time(td))
 338                 return false;
 339         if (!td->o.timeout)
 340                 return false;
 341         if (utime_since(&td->epoch, t) >= td->o.timeout)
 342                 return true;
 343
 344         return false;
 345 }
 346
 347 /*
 348  * We need to update the runtime consistently in ms, but keep a running
 349  * tally of the current elapsed time in microseconds for sub millisecond
 350  * updates.
 351  */
 352 static inline void update_runtime(struct thread_data *td,
 353                                   unsigned long long *elapsed_us,
 354                                   const enum fio_ddir ddir)
 355 {
 356         if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
 357                 return;
 358
 359         td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
 360         elapsed_us[ddir] += utime_since_now(&td->start);
 361         td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
 362 }
 363
 364 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 365                                 int *retptr)
 366 {
 367         int ret = *retptr;
 368
 369         if (ret < 0 || td->error) {
 370                 int err = td->error;
 371                 enum error_type_bit eb;
 372
 373                 if (ret < 0)
 374                         err = -ret;
 375
 376                 eb = td_error_type(ddir, err);
 377                 if (!(td->o.continue_on_error & (1 << eb)))
 378                         return true;
 379
 380                 if (td_non_fatal_error(td, eb, err)) {
 381                         /*
 382                          * Continue with the I/Os in case of
 383                          * a non fatal error.
 384                          */
 385                         update_error_count(td, err);
 386                         td_clear_error(td);
 387                         *retptr = 0;
 388                         return false;
 389                 } else if (td->o.fill_device && err == ENOSPC) {
 390                         /*
 391                          * We expect to hit this error if
 392                          * fill_device option is set.
 393                          */
 394                         td_clear_error(td);
 395                         fio_mark_td_terminate(td);
 396                         return true;
 397                 } else {
 398                         /*
 399                          * Stop the I/O in case of a fatal
 400                          * error.
 401                          */
 402                         update_error_count(td, err);
 403                         return true;
 404                 }
 405         }
 406
 407         return false;
 408 }
 409
 410 static void check_update_rusage(struct thread_data *td)
 411 {
 412         if (td->update_rusage) {
 413                 td->update_rusage = 0;
 414                 update_rusage_stat(td);
 415                 fio_sem_up(td->rusage_sem);
 416         }
 417 }
 418
 419 static int wait_for_completions(struct thread_data *td, struct timespec *time)
 420 {
 421         const int full = queue_full(td);
 422         int min_evts = 0;
 423         int ret;
 424
 425         if (td->flags & TD_F_REGROW_LOGS)
 426                 return io_u_quiesce(td);
 427
 428         /*
 429          * if the queue is full, we MUST reap at least 1 event
 430          */
 431         min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
 432         if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
 433                 min_evts = 1;
 434
 435         if (time && __should_check_rate(td))
 436                 fio_gettime(time, NULL);
 437
 438         do {
 439                 ret = io_u_queued_complete(td, min_evts);
 440                 if (ret < 0)
 441                         break;
 442         } while (full && (td->cur_depth > td->o.iodepth_low));
 443
 444         return ret;
 445 }
 446
 447 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
 448                    enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
 449                    struct timespec *comp_time)
 450 {
 451         switch (*ret) {
 452         case FIO_Q_COMPLETED:
 453                 if (io_u->error) {
 454                         *ret = -io_u->error;
 455                         clear_io_u(td, io_u);
 456                 } else if (io_u->resid) {
 457                         long long bytes = io_u->xfer_buflen - io_u->resid;
 458                         struct fio_file *f = io_u->file;
 459
 460                         if (bytes_issued)
 461                                 *bytes_issued += bytes;
 462
 463                         if (!from_verify)
 464                                 trim_io_piece(io_u);
 465
 466                         /*
 467                          * zero read, fail
 468                          */
 469                         if (!bytes) {
 470                                 if (!from_verify)
 471                                         unlog_io_piece(td, io_u);
 472                                 td_verror(td, EIO, "full resid");
 473                                 put_io_u(td, io_u);
 474                                 break;
 475                         }
 476
 477                         io_u->xfer_buflen = io_u->resid;
 478                         io_u->xfer_buf += bytes;
 479                         io_u->offset += bytes;
 480
 481                         if (ddir_rw(io_u->ddir))
 482                                 td->ts.short_io_u[io_u->ddir]++;
 483
 484                         if (io_u->offset == f->real_file_size)
 485                                 goto sync_done;
 486
 487                         requeue_io_u(td, &io_u);
 488                 } else {
 489 sync_done:
 490                         if (comp_time && __should_check_rate(td))
 491                                 fio_gettime(comp_time, NULL);
 492
 493                         *ret = io_u_sync_complete(td, io_u);
 494                         if (*ret < 0)
 495                                 break;
 496                 }
 497
 498                 if (td->flags & TD_F_REGROW_LOGS)
 499                         regrow_logs(td);
 500
 501                 /*
 502                  * when doing I/O (not when verifying),
 503                  * check for any errors that are to be ignored
 504                  */
 505                 if (!from_verify)
 506                         break;
 507
 508                 return 0;
 509         case FIO_Q_QUEUED:
 510                 /*
 511                  * if the engine doesn't have a commit hook,
 512                  * the io_u is really queued. if it does have such
 513                  * a hook, it has to call io_u_queued() itself.
 514                  */
 515                 if (td->io_ops->commit == NULL)
 516                         io_u_queued(td, io_u);
 517                 if (bytes_issued)
 518                         *bytes_issued += io_u->xfer_buflen;
 519                 break;
 520         case FIO_Q_BUSY:
 521                 if (!from_verify)
 522                         unlog_io_piece(td, io_u);
 523                 requeue_io_u(td, &io_u);
 524                 td_io_commit(td);
 525                 break;
 526         default:
 527                 assert(*ret < 0);
 528                 td_verror(td, -(*ret), "td_io_queue");
 529                 break;
 530         }
 531
 532         if (break_on_this_error(td, ddir, ret))
 533                 return 1;
 534
 535         return 0;
 536 }
 537
 538 static inline bool io_in_polling(struct thread_data *td)
 539 {
 540         return !td->o.iodepth_batch_complete_min &&
 541                    !td->o.iodepth_batch_complete_max;
 542 }
 543 /*
 544  * Unlinks files from thread data fio_file structure
 545  */
 546 static int unlink_all_files(struct thread_data *td)
 547 {
 548         struct fio_file *f;
 549         unsigned int i;
 550         int ret = 0;
 551
 552         for_each_file(td, f, i) {
 553                 if (f->filetype != FIO_TYPE_FILE)
 554                         continue;
 555                 ret = td_io_unlink_file(td, f);
 556                 if (ret)
 557                         break;
 558         }
 559
 560         if (ret)
 561                 td_verror(td, ret, "unlink_all_files");
 562
 563         return ret;
 564 }
 565
 566 /*
 567  * Check if io_u will overlap an in-flight IO in the queue
 568  */
 569 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 570 {
 571         bool overlap;
 572         struct io_u *check_io_u;
 573         unsigned long long x1, x2, y1, y2;
 574         int i;
 575
 576         x1 = io_u->offset;
 577         x2 = io_u->offset + io_u->buflen;
 578         overlap = false;
 579         io_u_qiter(q, check_io_u, i) {
 580                 if (check_io_u->flags & IO_U_F_FLIGHT) {
 581                         y1 = check_io_u->offset;
 582                         y2 = check_io_u->offset + check_io_u->buflen;
 583
 584                         if (x1 < y2 && y1 < x2) {
 585                                 overlap = true;
 586                                 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
 587                                                 x1, io_u->buflen,
 588                                                 y1, check_io_u->buflen);
 589                                 break;
 590                         }
 591                 }
 592         }
 593
 594         return overlap;
 595 }
 596
 597 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
 598 {
 599         /*
 600          * Check for overlap if the user asked us to, and we have
 601          * at least one IO in flight besides this one.
 602          */
 603         if (td->o.serialize_overlap && td->cur_depth > 1 &&
 604             in_flight_overlap(&td->io_u_all, io_u))
 605                 return FIO_Q_BUSY;
 606
 607         return td_io_queue(td, io_u);
 608 }
 609
 610 /*
 611  * The main verify engine. Runs over the writes we previously submitted,
 612  * reads the blocks back in, and checks the crc/md5 of the data.
 613  */
 614 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 615 {
 616         struct fio_file *f;
 617         struct io_u *io_u;
 618         int ret, min_events;
 619         unsigned int i;
 620
 621         dprint(FD_VERIFY, "starting loop\n");
 622
 623         /*
 624          * sync io first and invalidate cache, to make sure we really
 625          * read from disk.
 626          */
 627         for_each_file(td, f, i) {
 628                 if (!fio_file_open(f))
 629                         continue;
 630                 if (fio_io_sync(td, f))
 631                         break;
 632                 if (file_invalidate_cache(td, f))
 633                         break;
 634         }
 635
 636         check_update_rusage(td);
 637
 638         if (td->error)
 639                 return;
 640
 641         /*
 642          * verify_state needs to be reset before verification
 643          * proceeds so that expected random seeds match actual
 644          * random seeds in headers. The main loop will reset
 645          * all random number generators if randrepeat is set.
 646          */
 647         if (!td->o.rand_repeatable)
 648                 td_fill_verify_state_seed(td);
 649
 650         td_set_runstate(td, TD_VERIFYING);
 651
 652         io_u = NULL;
 653         while (!td->terminate) {
 654                 enum fio_ddir ddir;
 655                 int full;
 656
 657                 update_ts_cache(td);
 658                 check_update_rusage(td);
 659
 660                 if (runtime_exceeded(td, &td->ts_cache)) {
 661                         __update_ts_cache(td);
 662                         if (runtime_exceeded(td, &td->ts_cache)) {
 663                                 fio_mark_td_terminate(td);
 664                                 break;
 665                         }
 666                 }
 667
 668                 if (flow_threshold_exceeded(td))
 669                         continue;
 670
 671                 if (!td->o.experimental_verify) {
 672                         io_u = __get_io_u(td);
 673                         if (!io_u)
 674                                 break;
 675
 676                         if (get_next_verify(td, io_u)) {
 677                                 put_io_u(td, io_u);
 678                                 break;
 679                         }
 680
 681                         if (td_io_prep(td, io_u)) {
 682                                 put_io_u(td, io_u);
 683                                 break;
 684                         }
 685                 } else {
 686                         if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
 687                                 break;
 688
 689                         while ((io_u = get_io_u(td)) != NULL) {
 690                                 if (IS_ERR_OR_NULL(io_u)) {
 691                                         io_u = NULL;
 692                                         ret = FIO_Q_BUSY;
 693                                         goto reap;
 694                                 }
 695
 696                                 /*
 697                                  * We are only interested in the places where
 698                                  * we wrote or trimmed IOs. Turn those into
 699                                  * reads for verification purposes.
 700                                  */
 701                                 if (io_u->ddir == DDIR_READ) {
 702                                         /*
 703                                          * Pretend we issued it for rwmix
 704                                          * accounting
 705                                          */
 706                                         td->io_issues[DDIR_READ]++;
 707                                         put_io_u(td, io_u);
 708                                         continue;
 709                                 } else if (io_u->ddir == DDIR_TRIM) {
 710                                         io_u->ddir = DDIR_READ;
 711                                         io_u_set(td, io_u, IO_U_F_TRIMMED);
 712                                         break;
 713                                 } else if (io_u->ddir == DDIR_WRITE) {
 714                                         io_u->ddir = DDIR_READ;
 715                                         populate_verify_io_u(td, io_u);
 716                                         break;
 717                                 } else {
 718                                         put_io_u(td, io_u);
 719                                         continue;
 720                                 }
 721                         }
 722
 723                         if (!io_u)
 724                                 break;
 725                 }
 726
 727                 if (verify_state_should_stop(td, io_u)) {
 728                         put_io_u(td, io_u);
 729                         break;
 730                 }
 731
 732                 if (td->o.verify_async)
 733                         io_u->end_io = verify_io_u_async;
 734                 else
 735                         io_u->end_io = verify_io_u;
 736
 737                 ddir = io_u->ddir;
 738                 if (!td->o.disable_slat)
 739                         fio_gettime(&io_u->start_time, NULL);
 740
 741                 ret = io_u_submit(td, io_u);
 742
 743                 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
 744                         break;
 745
 746                 /*
 747                  * if we can queue more, do so. but check if there are
 748                  * completed io_u's first. Note that we can get BUSY even
 749                  * without IO queued, if the system is resource starved.
 750                  */
 751 reap:
 752                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 753                 if (full || io_in_polling(td))
 754                         ret = wait_for_completions(td, NULL);
 755
 756                 if (ret < 0)
 757                         break;
 758         }
 759
 760         check_update_rusage(td);
 761
 762         if (!td->error) {
 763                 min_events = td->cur_depth;
 764
 765                 if (min_events)
 766                         ret = io_u_queued_complete(td, min_events);
 767         } else
 768                 cleanup_pending_aio(td);
 769
 770         td_set_runstate(td, TD_RUNNING);
 771
 772         dprint(FD_VERIFY, "exiting loop\n");
 773 }
 774
 775 static bool exceeds_number_ios(struct thread_data *td)
 776 {
 777         unsigned long long number_ios;
 778
 779         if (!td->o.number_ios)
 780                 return false;
 781
 782         number_ios = ddir_rw_sum(td->io_blocks);
 783         number_ios += td->io_u_queued + td->io_u_in_flight;
 784
 785         return number_ios >= (td->o.number_ios * td->loops);
 786 }
 787
 788 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 789 {
 790         unsigned long long bytes, limit;
 791
 792         if (td_rw(td))
 793                 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
 794         else if (td_write(td))
 795                 bytes = this_bytes[DDIR_WRITE];
 796         else if (td_read(td))
 797                 bytes = this_bytes[DDIR_READ];
 798         else
 799                 bytes = this_bytes[DDIR_TRIM];
 800
 801         if (td->o.io_size)
 802                 limit = td->o.io_size;
 803         else
 804                 limit = td->o.size;
 805
 806         limit *= td->loops;
 807         return bytes >= limit || exceeds_number_ios(td);
 808 }
 809
 810 static bool io_issue_bytes_exceeded(struct thread_data *td)
 811 {
 812         return io_bytes_exceeded(td, td->io_issue_bytes);
 813 }
 814
 815 static bool io_complete_bytes_exceeded(struct thread_data *td)
 816 {
 817         return io_bytes_exceeded(td, td->this_io_bytes);
 818 }
 819
 820 /*
 821  * used to calculate the next io time for rate control
 822  *
 823  */
 824 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 825 {
 826         uint64_t bps = td->rate_bps[ddir];
 827
 828         assert(!(td->flags & TD_F_CHILD));
 829
 830         if (td->o.rate_process == RATE_PROCESS_POISSON) {
 831                 uint64_t val, iops;
 832
 833                 iops = bps / td->o.bs[ddir];
 834                 val = (int64_t) (1000000 / iops) *
 835                                 -logf(__rand_0_1(&td->poisson_state[ddir]));
 836                 if (val) {
 837                         dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
 838                                         (unsigned long long) 1000000 / val,
 839                                         ddir);
 840                 }
 841                 td->last_usec[ddir] += val;
 842                 return td->last_usec[ddir];
 843         } else if (bps) {
 844                 uint64_t bytes = td->rate_io_issue_bytes[ddir];
 845                 uint64_t secs = bytes / bps;
 846                 uint64_t remainder = bytes % bps;
 847
 848                 return remainder * 1000000 / bps + secs * 1000000;
 849         }
 850
 851         return 0;
 852 }
 853
 854 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
 855 {
 856         unsigned long long b;
 857         uint64_t total;
 858         int left;
 859
 860         b = ddir_rw_sum(td->io_blocks);
 861         if (b % td->o.thinktime_blocks)
 862                 return;
 863
 864         io_u_quiesce(td);
 865
 866         total = 0;
 867         if (td->o.thinktime_spin)
 868                 total = usec_spin(td->o.thinktime_spin);
 869
 870         left = td->o.thinktime - total;
 871         if (left)
 872                 total += usec_sleep(td, left);
 873
 874         /*
 875          * If we're ignoring thinktime for the rate, add the number of bytes
 876          * we would have done while sleeping, minus one block to ensure we
 877          * start issuing immediately after the sleep.
 878          */
 879         if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
 880                 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
 881                 uint64_t bs = td->o.min_bs[ddir];
 882                 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
 883                 uint64_t over;
 884
 885                 if (usperop <= total)
 886                         over = bs;
 887                 else
 888                         over = (usperop - total) / usperop * -bs;
 889
 890                 td->rate_io_issue_bytes[ddir] += (missed - over);
 891                 /* adjust for rate_process=poisson */
 892                 td->last_usec[ddir] += total;
 893         }
 894 }
 895
 896 /*
 897  * Main IO worker function. It retrieves io_u's to process and queues
 898  * and reaps them, checking for rate and errors along the way.
 899  *
 900  * Returns number of bytes written and trimmed.
 901  */
 902 static void do_io(struct thread_data *td, uint64_t *bytes_done)
 903 {
 904         unsigned int i;
 905         int ret = 0;
 906         uint64_t total_bytes, bytes_issued = 0;
 907
 908         for (i = 0; i < DDIR_RWDIR_CNT; i++)
 909                 bytes_done[i] = td->bytes_done[i];
 910
 911         if (in_ramp_time(td))
 912                 td_set_runstate(td, TD_RAMP);
 913         else
 914                 td_set_runstate(td, TD_RUNNING);
 915
 916         lat_target_init(td);
 917
 918         total_bytes = td->o.size;
 919         /*
 920         * Allow random overwrite workloads to write up to io_size
 921         * before starting verification phase as 'size' doesn't apply.
 922         */
 923         if (td_write(td) && td_random(td) && td->o.norandommap)
 924                 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
 925         /*
 926          * If verify_backlog is enabled, we'll run the verify in this
 927          * handler as well. For that case, we may need up to twice the
 928          * amount of bytes.
 929          */
 930         if (td->o.verify != VERIFY_NONE &&
 931            (td_write(td) && td->o.verify_backlog))
 932                 total_bytes += td->o.size;
 933
 934         /* In trimwrite mode, each byte is trimmed and then written, so
 935          * allow total_bytes to be twice as big */
 936         if (td_trimwrite(td))
 937                 total_bytes += td->total_io_size;
 938
 939         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 940                 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
 941                 td->o.time_based) {
 942                 struct timespec comp_time;
 943                 struct io_u *io_u;
 944                 int full;
 945                 enum fio_ddir ddir;
 946
 947                 check_update_rusage(td);
 948
 949                 if (td->terminate || td->done)
 950                         break;
 951
 952                 update_ts_cache(td);
 953
 954                 if (runtime_exceeded(td, &td->ts_cache)) {
 955                         __update_ts_cache(td);
 956                         if (runtime_exceeded(td, &td->ts_cache)) {
 957                                 fio_mark_td_terminate(td);
 958                                 break;
 959                         }
 960                 }
 961
 962                 if (flow_threshold_exceeded(td))
 963                         continue;
 964
 965                 /*
 966                  * Break if we exceeded the bytes. The exception is time
 967                  * based runs, but we still need to break out of the loop
 968                  * for those to run verification, if enabled.
 969                  */
 970                 if (bytes_issued >= total_bytes &&
 971                     (!td->o.time_based ||
 972                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 973                         break;
 974
 975                 io_u = get_io_u(td);
 976                 if (IS_ERR_OR_NULL(io_u)) {
 977                         int err = PTR_ERR(io_u);
 978
 979                         io_u = NULL;
 980                         ddir = DDIR_INVAL;
 981                         if (err == -EBUSY) {
 982                                 ret = FIO_Q_BUSY;
 983                                 goto reap;
 984                         }
 985                         if (td->o.latency_target)
 986                                 goto reap;
 987                         break;
 988                 }
 989
 990                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
 991                         populate_verify_io_u(td, io_u);
 992
 993                 ddir = io_u->ddir;
 994
 995                 /*
 996                  * Add verification end_io handler if:
 997                  *      - Asked to verify (!td_rw(td))
 998                  *      - Or the io_u is from our verify list (mixed write/ver)
 999                  */
1000                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1001                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1002
1003                         if (!td->o.verify_pattern_bytes) {
1004                                 io_u->rand_seed = __rand(&td->verify_state);
1005                                 if (sizeof(int) != sizeof(long *))
1006                                         io_u->rand_seed *= __rand(&td->verify_state);
1007                         }
1008
1009                         if (verify_state_should_stop(td, io_u)) {
1010                                 put_io_u(td, io_u);
1011                                 break;
1012                         }
1013
1014                         if (td->o.verify_async)
1015                                 io_u->end_io = verify_io_u_async;
1016                         else
1017                                 io_u->end_io = verify_io_u;
1018                         td_set_runstate(td, TD_VERIFYING);
1019                 } else if (in_ramp_time(td))
1020                         td_set_runstate(td, TD_RAMP);
1021                 else
1022                         td_set_runstate(td, TD_RUNNING);
1023
1024                 /*
1025                  * Always log IO before it's issued, so we know the specific
1026                  * order of it. The logged unit will track when the IO has
1027                  * completed.
1028                  */
1029                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1030                     td->o.do_verify &&
1031                     td->o.verify != VERIFY_NONE &&
1032                     !td->o.experimental_verify)
1033                         log_io_piece(td, io_u);
1034
1035                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1036                         const unsigned long long blen = io_u->xfer_buflen;
1037                         const enum fio_ddir __ddir = acct_ddir(io_u);
1038
1039                         if (td->error)
1040                                 break;
1041
1042                         workqueue_enqueue(&td->io_wq, &io_u->work);
1043                         ret = FIO_Q_QUEUED;
1044
1045                         if (ddir_rw(__ddir)) {
1046                                 td->io_issues[__ddir]++;
1047                                 td->io_issue_bytes[__ddir] += blen;
1048                                 td->rate_io_issue_bytes[__ddir] += blen;
1049                         }
1050
1051                         if (should_check_rate(td))
1052                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1053
1054                 } else {
1055                         ret = io_u_submit(td, io_u);
1056
1057                         if (should_check_rate(td))
1058                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1059
1060                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1061                                 break;
1062
1063                         /*
1064                          * See if we need to complete some commands. Note that
1065                          * we can get BUSY even without IO queued, if the
1066                          * system is resource starved.
1067                          */
1068 reap:
1069                         full = queue_full(td) ||
1070                                 (ret == FIO_Q_BUSY && td->cur_depth);
1071                         if (full || io_in_polling(td))
1072                                 ret = wait_for_completions(td, &comp_time);
1073                 }
1074                 if (ret < 0)
1075                         break;
1076                 if (!ddir_rw_sum(td->bytes_done) &&
1077                     !td_ioengine_flagged(td, FIO_NOIO))
1078                         continue;
1079
1080                 if (!in_ramp_time(td) && should_check_rate(td)) {
1081                         if (check_min_rate(td, &comp_time)) {
1082                                 if (exitall_on_terminate || td->o.exitall_error)
1083                                         fio_terminate_threads(td->groupid);
1084                                 td_verror(td, EIO, "check_min_rate");
1085                                 break;
1086                         }
1087                 }
1088                 if (!in_ramp_time(td) && td->o.latency_target)
1089                         lat_target_check(td);
1090
1091                 if (ddir_rw(ddir) && td->o.thinktime)
1092                         handle_thinktime(td, ddir);
1093         }
1094
1095         check_update_rusage(td);
1096
1097         if (td->trim_entries)
1098                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1099
1100         if (td->o.fill_device && td->error == ENOSPC) {
1101                 td->error = 0;
1102                 fio_mark_td_terminate(td);
1103         }
1104         if (!td->error) {
1105                 struct fio_file *f;
1106
1107                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1108                         workqueue_flush(&td->io_wq);
1109                         i = 0;
1110                 } else
1111                         i = td->cur_depth;
1112
1113                 if (i) {
1114                         ret = io_u_queued_complete(td, i);
1115                         if (td->o.fill_device && td->error == ENOSPC)
1116                                 td->error = 0;
1117                 }
1118
1119                 if (should_fsync(td) && td->o.end_fsync) {
1120                         td_set_runstate(td, TD_FSYNCING);
1121
1122                         for_each_file(td, f, i) {
1123                                 if (!fio_file_fsync(td, f))
1124                                         continue;
1125
1126                                 log_err("fio: end_fsync failed for file %s\n",
1127                                                                 f->file_name);
1128                         }
1129                 }
1130         } else
1131                 cleanup_pending_aio(td);
1132
1133         /*
1134          * stop job if we failed doing any IO
1135          */
1136         if (!ddir_rw_sum(td->this_io_bytes))
1137                 td->done = 1;
1138
1139         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1140                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1141 }
1142
1143 static void free_file_completion_logging(struct thread_data *td)
1144 {
1145         struct fio_file *f;
1146         unsigned int i;
1147
1148         for_each_file(td, f, i) {
1149                 if (!f->last_write_comp)
1150                         break;
1151                 sfree(f->last_write_comp);
1152         }
1153 }
1154
1155 static int init_file_completion_logging(struct thread_data *td,
1156                                         unsigned int depth)
1157 {
1158         struct fio_file *f;
1159         unsigned int i;
1160
1161         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1162                 return 0;
1163
1164         for_each_file(td, f, i) {
1165                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1166                 if (!f->last_write_comp)
1167                         goto cleanup;
1168         }
1169
1170         return 0;
1171
1172 cleanup:
1173         free_file_completion_logging(td);
1174         log_err("fio: failed to alloc write comp data\n");
1175         return 1;
1176 }
1177
1178 static void cleanup_io_u(struct thread_data *td)
1179 {
1180         struct io_u *io_u;
1181
1182         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1183
1184                 if (td->io_ops->io_u_free)
1185                         td->io_ops->io_u_free(td, io_u);
1186
1187                 fio_memfree(io_u, sizeof(*io_u));
1188         }
1189
1190         free_io_mem(td);
1191
1192         io_u_rexit(&td->io_u_requeues);
1193         io_u_qexit(&td->io_u_freelist);
1194         io_u_qexit(&td->io_u_all);
1195
1196         free_file_completion_logging(td);
1197 }
1198
1199 static int init_io_u(struct thread_data *td)
1200 {
1201         struct io_u *io_u;
1202         unsigned long long max_bs, min_write;
1203         int cl_align, i, max_units;
1204         int data_xfer = 1, err;
1205         char *p;
1206
1207         max_units = td->o.iodepth;
1208         max_bs = td_max_bs(td);
1209         min_write = td->o.min_bs[DDIR_WRITE];
1210         td->orig_buffer_size = (unsigned long long) max_bs
1211                                         * (unsigned long long) max_units;
1212
1213         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1214                 data_xfer = 0;
1215
1216         err = 0;
1217         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1218         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1219         err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1220
1221         if (err) {
1222                 log_err("fio: failed setting up IO queues\n");
1223                 return 1;
1224         }
1225
1226         /*
1227          * if we may later need to do address alignment, then add any
1228          * possible adjustment here so that we don't cause a buffer
1229          * overflow later. this adjustment may be too much if we get
1230          * lucky and the allocator gives us an aligned address.
1231          */
1232         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1233             td_ioengine_flagged(td, FIO_RAWIO))
1234                 td->orig_buffer_size += page_mask + td->o.mem_align;
1235
1236         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1237                 unsigned long long bs;
1238
1239                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1240                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1241         }
1242
1243         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1244                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1245                 return 1;
1246         }
1247
1248         if (data_xfer && allocate_io_mem(td))
1249                 return 1;
1250
1251         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1252             td_ioengine_flagged(td, FIO_RAWIO))
1253                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1254         else
1255                 p = td->orig_buffer;
1256
1257         cl_align = os_cache_line_size();
1258
1259         for (i = 0; i < max_units; i++) {
1260                 void *ptr;
1261
1262                 if (td->terminate)
1263                         return 1;
1264
1265                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1266                 if (!ptr) {
1267                         log_err("fio: unable to allocate aligned memory\n");
1268                         break;
1269                 }
1270
1271                 io_u = ptr;
1272                 memset(io_u, 0, sizeof(*io_u));
1273                 INIT_FLIST_HEAD(&io_u->verify_list);
1274                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1275
1276                 if (data_xfer) {
1277                         io_u->buf = p;
1278                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1279
1280                         if (td_write(td))
1281                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1282                         if (td_write(td) && td->o.verify_pattern_bytes) {
1283                                 /*
1284                                  * Fill the buffer with the pattern if we are
1285                                  * going to be doing writes.
1286                                  */
1287                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1288                         }
1289                 }
1290
1291                 io_u->index = i;
1292                 io_u->flags = IO_U_F_FREE;
1293                 io_u_qpush(&td->io_u_freelist, io_u);
1294
1295                 /*
1296                  * io_u never leaves this stack, used for iteration of all
1297                  * io_u buffers.
1298                  */
1299                 io_u_qpush(&td->io_u_all, io_u);
1300
1301                 if (td->io_ops->io_u_init) {
1302                         int ret = td->io_ops->io_u_init(td, io_u);
1303
1304                         if (ret) {
1305                                 log_err("fio: failed to init engine data: %d\n", ret);
1306                                 return 1;
1307                         }
1308                 }
1309
1310                 p += max_bs;
1311         }
1312
1313         if (init_file_completion_logging(td, max_units))
1314                 return 1;
1315
1316         return 0;
1317 }
1318
1319 /*
1320  * This function is Linux specific.
1321  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1322  */
1323 static int switch_ioscheduler(struct thread_data *td)
1324 {
1325 #ifdef FIO_HAVE_IOSCHED_SWITCH
1326         char tmp[256], tmp2[128], *p;
1327         FILE *f;
1328         int ret;
1329
1330         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1331                 return 0;
1332
1333         assert(td->files && td->files[0]);
1334         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1335
1336         f = fopen(tmp, "r+");
1337         if (!f) {
1338                 if (errno == ENOENT) {
1339                         log_err("fio: os or kernel doesn't support IO scheduler"
1340                                 " switching\n");
1341                         return 0;
1342                 }
1343                 td_verror(td, errno, "fopen iosched");
1344                 return 1;
1345         }
1346
1347         /*
1348          * Set io scheduler.
1349          */
1350         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1351         if (ferror(f) || ret != 1) {
1352                 td_verror(td, errno, "fwrite");
1353                 fclose(f);
1354                 return 1;
1355         }
1356
1357         rewind(f);
1358
1359         /*
1360          * Read back and check that the selected scheduler is now the default.
1361          */
1362         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1363         if (ferror(f) || ret < 0) {
1364                 td_verror(td, errno, "fread");
1365                 fclose(f);
1366                 return 1;
1367         }
1368         tmp[ret] = '\0';
1369         /*
1370          * either a list of io schedulers or "none\n" is expected. Strip the
1371          * trailing newline.
1372          */
1373         p = tmp;
1374         strsep(&p, "\n");
1375
1376         /*
1377          * Write to "none" entry doesn't fail, so check the result here.
1378          */
1379         if (!strcmp(tmp, "none")) {
1380                 log_err("fio: io scheduler is not tunable\n");
1381                 fclose(f);
1382                 return 0;
1383         }
1384
1385         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1386         if (!strstr(tmp, tmp2)) {
1387                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1388                 td_verror(td, EINVAL, "iosched_switch");
1389                 fclose(f);
1390                 return 1;
1391         }
1392
1393         fclose(f);
1394         return 0;
1395 #else
1396         return 0;
1397 #endif
1398 }
1399
1400 static bool keep_running(struct thread_data *td)
1401 {
1402         unsigned long long limit;
1403
1404         if (td->done)
1405                 return false;
1406         if (td->terminate)
1407                 return false;
1408         if (td->o.time_based)
1409                 return true;
1410         if (td->o.loops) {
1411                 td->o.loops--;
1412                 return true;
1413         }
1414         if (exceeds_number_ios(td))
1415                 return false;
1416
1417         if (td->o.io_size)
1418                 limit = td->o.io_size;
1419         else
1420                 limit = td->o.size;
1421
1422         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1423                 uint64_t diff;
1424
1425                 /*
1426                  * If the difference is less than the maximum IO size, we
1427                  * are done.
1428                  */
1429                 diff = limit - ddir_rw_sum(td->io_bytes);
1430                 if (diff < td_max_bs(td))
1431                         return false;
1432
1433                 if (fio_files_done(td) && !td->o.io_size)
1434                         return false;
1435
1436                 return true;
1437         }
1438
1439         return false;
1440 }
1441
1442 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1443 {
1444         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1445         int ret;
1446         char *str;
1447
1448         str = malloc(newlen);
1449         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1450
1451         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1452         ret = system(str);
1453         if (ret == -1)
1454                 log_err("fio: exec of cmd <%s> failed\n", str);
1455
1456         free(str);
1457         return ret;
1458 }
1459
1460 /*
1461  * Dry run to compute correct state of numberio for verification.
1462  */
1463 static uint64_t do_dry_run(struct thread_data *td)
1464 {
1465         td_set_runstate(td, TD_RUNNING);
1466
1467         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1468                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1469                 struct io_u *io_u;
1470                 int ret;
1471
1472                 if (td->terminate || td->done)
1473                         break;
1474
1475                 io_u = get_io_u(td);
1476                 if (IS_ERR_OR_NULL(io_u))
1477                         break;
1478
1479                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1480                 io_u->error = 0;
1481                 io_u->resid = 0;
1482                 if (ddir_rw(acct_ddir(io_u)))
1483                         td->io_issues[acct_ddir(io_u)]++;
1484                 if (ddir_rw(io_u->ddir)) {
1485                         io_u_mark_depth(td, 1);
1486                         td->ts.total_io_u[io_u->ddir]++;
1487                 }
1488
1489                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1490                     td->o.do_verify &&
1491                     td->o.verify != VERIFY_NONE &&
1492                     !td->o.experimental_verify)
1493                         log_io_piece(td, io_u);
1494
1495                 ret = io_u_sync_complete(td, io_u);
1496                 (void) ret;
1497         }
1498
1499         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1500 }
1501
1502 struct fork_data {
1503         struct thread_data *td;
1504         struct sk_out *sk_out;
1505 };
1506
1507 /*
1508  * Entry point for the thread based jobs. The process based jobs end up
1509  * here as well, after a little setup.
1510  */
1511 static void *thread_main(void *data)
1512 {
1513         struct fork_data *fd = data;
1514         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1515         struct thread_data *td = fd->td;
1516         struct thread_options *o = &td->o;
1517         struct sk_out *sk_out = fd->sk_out;
1518         uint64_t bytes_done[DDIR_RWDIR_CNT];
1519         int deadlock_loop_cnt;
1520         bool clear_state, did_some_io;
1521         int ret;
1522
1523         sk_out_assign(sk_out);
1524         free(fd);
1525
1526         if (!o->use_thread) {
1527                 setsid();
1528                 td->pid = getpid();
1529         } else
1530                 td->pid = gettid();
1531
1532         fio_local_clock_init();
1533
1534         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1535
1536         if (is_backend)
1537                 fio_server_send_start(td);
1538
1539         INIT_FLIST_HEAD(&td->io_log_list);
1540         INIT_FLIST_HEAD(&td->io_hist_list);
1541         INIT_FLIST_HEAD(&td->verify_list);
1542         INIT_FLIST_HEAD(&td->trim_list);
1543         td->io_hist_tree = RB_ROOT;
1544
1545         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1546         if (ret) {
1547                 td_verror(td, ret, "mutex_cond_init_pshared");
1548                 goto err;
1549         }
1550         ret = cond_init_pshared(&td->verify_cond);
1551         if (ret) {
1552                 td_verror(td, ret, "mutex_cond_pshared");
1553                 goto err;
1554         }
1555
1556         td_set_runstate(td, TD_INITIALIZED);
1557         dprint(FD_MUTEX, "up startup_sem\n");
1558         fio_sem_up(startup_sem);
1559         dprint(FD_MUTEX, "wait on td->sem\n");
1560         fio_sem_down(td->sem);
1561         dprint(FD_MUTEX, "done waiting on td->sem\n");
1562
1563         /*
1564          * A new gid requires privilege, so we need to do this before setting
1565          * the uid.
1566          */
1567         if (o->gid != -1U && setgid(o->gid)) {
1568                 td_verror(td, errno, "setgid");
1569                 goto err;
1570         }
1571         if (o->uid != -1U && setuid(o->uid)) {
1572                 td_verror(td, errno, "setuid");
1573                 goto err;
1574         }
1575
1576         /*
1577          * Do this early, we don't want the compress threads to be limited
1578          * to the same CPUs as the IO workers. So do this before we set
1579          * any potential CPU affinity
1580          */
1581         if (iolog_compress_init(td, sk_out))
1582                 goto err;
1583
1584         /*
1585          * If we have a gettimeofday() thread, make sure we exclude that
1586          * thread from this job
1587          */
1588         if (o->gtod_cpu)
1589                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1590
1591         /*
1592          * Set affinity first, in case it has an impact on the memory
1593          * allocations.
1594          */
1595         if (fio_option_is_set(o, cpumask)) {
1596                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1597                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1598                         if (!ret) {
1599                                 log_err("fio: no CPUs set\n");
1600                                 log_err("fio: Try increasing number of available CPUs\n");
1601                                 td_verror(td, EINVAL, "cpus_split");
1602                                 goto err;
1603                         }
1604                 }
1605                 ret = fio_setaffinity(td->pid, o->cpumask);
1606                 if (ret == -1) {
1607                         td_verror(td, errno, "cpu_set_affinity");
1608                         goto err;
1609                 }
1610         }
1611
1612 #ifdef CONFIG_LIBNUMA
1613         /* numa node setup */
1614         if (fio_option_is_set(o, numa_cpunodes) ||
1615             fio_option_is_set(o, numa_memnodes)) {
1616                 struct bitmask *mask;
1617
1618                 if (numa_available() < 0) {
1619                         td_verror(td, errno, "Does not support NUMA API\n");
1620                         goto err;
1621                 }
1622
1623                 if (fio_option_is_set(o, numa_cpunodes)) {
1624                         mask = numa_parse_nodestring(o->numa_cpunodes);
1625                         ret = numa_run_on_node_mask(mask);
1626                         numa_free_nodemask(mask);
1627                         if (ret == -1) {
1628                                 td_verror(td, errno, \
1629                                         "numa_run_on_node_mask failed\n");
1630                                 goto err;
1631                         }
1632                 }
1633
1634                 if (fio_option_is_set(o, numa_memnodes)) {
1635                         mask = NULL;
1636                         if (o->numa_memnodes)
1637                                 mask = numa_parse_nodestring(o->numa_memnodes);
1638
1639                         switch (o->numa_mem_mode) {
1640                         case MPOL_INTERLEAVE:
1641                                 numa_set_interleave_mask(mask);
1642                                 break;
1643                         case MPOL_BIND:
1644                                 numa_set_membind(mask);
1645                                 break;
1646                         case MPOL_LOCAL:
1647                                 numa_set_localalloc();
1648                                 break;
1649                         case MPOL_PREFERRED:
1650                                 numa_set_preferred(o->numa_mem_prefer_node);
1651                                 break;
1652                         case MPOL_DEFAULT:
1653                         default:
1654                                 break;
1655                         }
1656
1657                         if (mask)
1658                                 numa_free_nodemask(mask);
1659
1660                 }
1661         }
1662 #endif
1663
1664         if (fio_pin_memory(td))
1665                 goto err;
1666
1667         /*
1668          * May alter parameters that init_io_u() will use, so we need to
1669          * do this first.
1670          */
1671         if (!init_iolog(td))
1672                 goto err;
1673
1674         if (init_io_u(td))
1675                 goto err;
1676
1677         if (o->verify_async && verify_async_init(td))
1678                 goto err;
1679
1680         if (fio_option_is_set(o, ioprio) ||
1681             fio_option_is_set(o, ioprio_class)) {
1682                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1683                 if (ret == -1) {
1684                         td_verror(td, errno, "ioprio_set");
1685                         goto err;
1686                 }
1687         }
1688
1689         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1690                 goto err;
1691
1692         errno = 0;
1693         if (nice(o->nice) == -1 && errno != 0) {
1694                 td_verror(td, errno, "nice");
1695                 goto err;
1696         }
1697
1698         if (o->ioscheduler && switch_ioscheduler(td))
1699                 goto err;
1700
1701         if (!o->create_serialize && setup_files(td))
1702                 goto err;
1703
1704         if (td_io_init(td))
1705                 goto err;
1706
1707         if (!init_random_map(td))
1708                 goto err;
1709
1710         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1711                 goto err;
1712
1713         if (o->pre_read && !pre_read_files(td))
1714                 goto err;
1715
1716         fio_verify_init(td);
1717
1718         if (rate_submit_init(td, sk_out))
1719                 goto err;
1720
1721         set_epoch_time(td, o->log_unix_epoch);
1722         fio_getrusage(&td->ru_start);
1723         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1724         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1725         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1726
1727         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1728                         o->ratemin[DDIR_TRIM]) {
1729                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1730                                         sizeof(td->bw_sample_time));
1731                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1732                                         sizeof(td->bw_sample_time));
1733                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1734                                         sizeof(td->bw_sample_time));
1735         }
1736
1737         memset(bytes_done, 0, sizeof(bytes_done));
1738         clear_state = false;
1739         did_some_io = false;
1740
1741         while (keep_running(td)) {
1742                 uint64_t verify_bytes;
1743
1744                 fio_gettime(&td->start, NULL);
1745                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1746
1747                 if (clear_state) {
1748                         clear_io_state(td, 0);
1749
1750                         if (o->unlink_each_loop && unlink_all_files(td))
1751                                 break;
1752                 }
1753
1754                 prune_io_piece_log(td);
1755
1756                 if (td->o.verify_only && td_write(td))
1757                         verify_bytes = do_dry_run(td);
1758                 else {
1759                         do_io(td, bytes_done);
1760
1761                         if (!ddir_rw_sum(bytes_done)) {
1762                                 fio_mark_td_terminate(td);
1763                                 verify_bytes = 0;
1764                         } else {
1765                                 verify_bytes = bytes_done[DDIR_WRITE] +
1766                                                 bytes_done[DDIR_TRIM];
1767                         }
1768                 }
1769
1770                 /*
1771                  * If we took too long to shut down, the main thread could
1772                  * already consider us reaped/exited. If that happens, break
1773                  * out and clean up.
1774                  */
1775                 if (td->runstate >= TD_EXITED)
1776                         break;
1777
1778                 clear_state = true;
1779
1780                 /*
1781                  * Make sure we've successfully updated the rusage stats
1782                  * before waiting on the stat mutex. Otherwise we could have
1783                  * the stat thread holding stat mutex and waiting for
1784                  * the rusage_sem, which would never get upped because
1785                  * this thread is waiting for the stat mutex.
1786                  */
1787                 deadlock_loop_cnt = 0;
1788                 do {
1789                         check_update_rusage(td);
1790                         if (!fio_sem_down_trylock(stat_sem))
1791                                 break;
1792                         usleep(1000);
1793                         if (deadlock_loop_cnt++ > 5000) {
1794                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1795                                 td->error = EDEADLK;
1796                                 goto err;
1797                         }
1798                 } while (1);
1799
1800                 if (td_read(td) && td->io_bytes[DDIR_READ])
1801                         update_runtime(td, elapsed_us, DDIR_READ);
1802                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1803                         update_runtime(td, elapsed_us, DDIR_WRITE);
1804                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1805                         update_runtime(td, elapsed_us, DDIR_TRIM);
1806                 fio_gettime(&td->start, NULL);
1807                 fio_sem_up(stat_sem);
1808
1809                 if (td->error || td->terminate)
1810                         break;
1811
1812                 if (!o->do_verify ||
1813                     o->verify == VERIFY_NONE ||
1814                     td_ioengine_flagged(td, FIO_UNIDIR))
1815                         continue;
1816
1817                 if (ddir_rw_sum(bytes_done))
1818                         did_some_io = true;
1819
1820                 clear_io_state(td, 0);
1821
1822                 fio_gettime(&td->start, NULL);
1823
1824                 do_verify(td, verify_bytes);
1825
1826                 /*
1827                  * See comment further up for why this is done here.
1828                  */
1829                 check_update_rusage(td);
1830
1831                 fio_sem_down(stat_sem);
1832                 update_runtime(td, elapsed_us, DDIR_READ);
1833                 fio_gettime(&td->start, NULL);
1834                 fio_sem_up(stat_sem);
1835
1836                 if (td->error || td->terminate)
1837                         break;
1838         }
1839
1840         /*
1841          * If td ended up with no I/O when it should have had,
1842          * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1843          * (Are we not missing other flags that can be ignored ?)
1844          */
1845         if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1846             !did_some_io && !td->o.create_only &&
1847             !(td_ioengine_flagged(td, FIO_NOIO) ||
1848               td_ioengine_flagged(td, FIO_DISKLESSIO)))
1849                 log_err("%s: No I/O performed by %s, "
1850                          "perhaps try --debug=io option for details?\n",
1851                          td->o.name, td->io_ops->name);
1852
1853         td_set_runstate(td, TD_FINISHING);
1854
1855         update_rusage_stat(td);
1856         td->ts.total_run_time = mtime_since_now(&td->epoch);
1857         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1858         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1859         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1860
1861         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1862             (td->o.verify != VERIFY_NONE && td_write(td)))
1863                 verify_save_state(td->thread_number);
1864
1865         fio_unpin_memory(td);
1866
1867         td_writeout_logs(td, true);
1868
1869         iolog_compress_exit(td);
1870         rate_submit_exit(td);
1871
1872         if (o->exec_postrun)
1873                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1874
1875         if (exitall_on_terminate || (o->exitall_error && td->error))
1876                 fio_terminate_threads(td->groupid);
1877
1878 err:
1879         if (td->error)
1880                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1881                                                         td->verror);
1882
1883         if (o->verify_async)
1884                 verify_async_exit(td);
1885
1886         close_and_free_files(td);
1887         cleanup_io_u(td);
1888         close_ioengine(td);
1889         cgroup_shutdown(td, cgroup_mnt);
1890         verify_free_state(td);
1891
1892         if (td->zone_state_index) {
1893                 int i;
1894
1895                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1896                         free(td->zone_state_index[i]);
1897                 free(td->zone_state_index);
1898                 td->zone_state_index = NULL;
1899         }
1900
1901         if (fio_option_is_set(o, cpumask)) {
1902                 ret = fio_cpuset_exit(&o->cpumask);
1903                 if (ret)
1904                         td_verror(td, ret, "fio_cpuset_exit");
1905         }
1906
1907         /*
1908          * do this very late, it will log file closing as well
1909          */
1910         if (o->write_iolog_file)
1911                 write_iolog_close(td);
1912
1913         td_set_runstate(td, TD_EXITED);
1914
1915         /*
1916          * Do this last after setting our runstate to exited, so we
1917          * know that the stat thread is signaled.
1918          */
1919         check_update_rusage(td);
1920
1921         sk_out_drop();
1922         return (void *) (uintptr_t) td->error;
1923 }
1924
1925 /*
1926  * Run over the job map and reap the threads that have exited, if any.
1927  */
1928 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1929                          uint64_t *m_rate)
1930 {
1931         struct thread_data *td;
1932         unsigned int cputhreads, realthreads, pending;
1933         int i, status, ret;
1934
1935         /*
1936          * reap exited threads (TD_EXITED -> TD_REAPED)
1937          */
1938         realthreads = pending = cputhreads = 0;
1939         for_each_td(td, i) {
1940                 int flags = 0;
1941
1942                  if (!strcmp(td->o.ioengine, "cpuio"))
1943                         cputhreads++;
1944                 else
1945                         realthreads++;
1946
1947                 if (!td->pid) {
1948                         pending++;
1949                         continue;
1950                 }
1951                 if (td->runstate == TD_REAPED)
1952                         continue;
1953                 if (td->o.use_thread) {
1954                         if (td->runstate == TD_EXITED) {
1955                                 td_set_runstate(td, TD_REAPED);
1956                                 goto reaped;
1957                         }
1958                         continue;
1959                 }
1960
1961                 flags = WNOHANG;
1962                 if (td->runstate == TD_EXITED)
1963                         flags = 0;
1964
1965                 /*
1966                  * check if someone quit or got killed in an unusual way
1967                  */
1968                 ret = waitpid(td->pid, &status, flags);
1969                 if (ret < 0) {
1970                         if (errno == ECHILD) {
1971                                 log_err("fio: pid=%d disappeared %d\n",
1972                                                 (int) td->pid, td->runstate);
1973                                 td->sig = ECHILD;
1974                                 td_set_runstate(td, TD_REAPED);
1975                                 goto reaped;
1976                         }
1977                         perror("waitpid");
1978                 } else if (ret == td->pid) {
1979                         if (WIFSIGNALED(status)) {
1980                                 int sig = WTERMSIG(status);
1981
1982                                 if (sig != SIGTERM && sig != SIGUSR2)
1983                                         log_err("fio: pid=%d, got signal=%d\n",
1984                                                         (int) td->pid, sig);
1985                                 td->sig = sig;
1986                                 td_set_runstate(td, TD_REAPED);
1987                                 goto reaped;
1988                         }
1989                         if (WIFEXITED(status)) {
1990                                 if (WEXITSTATUS(status) && !td->error)
1991                                         td->error = WEXITSTATUS(status);
1992
1993                                 td_set_runstate(td, TD_REAPED);
1994                                 goto reaped;
1995                         }
1996                 }
1997
1998                 /*
1999                  * If the job is stuck, do a forceful timeout of it and
2000                  * move on.
2001                  */
2002                 if (td->terminate &&
2003                     td->runstate < TD_FSYNCING &&
2004                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2005                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2006                                 "%lu seconds, it appears to be stuck. Doing "
2007                                 "forceful exit of this job.\n",
2008                                 td->o.name, td->runstate,
2009                                 (unsigned long) time_since_now(&td->terminate_time));
2010                         td_set_runstate(td, TD_REAPED);
2011                         goto reaped;
2012                 }
2013
2014                 /*
2015                  * thread is not dead, continue
2016                  */
2017                 pending++;
2018                 continue;
2019 reaped:
2020                 (*nr_running)--;
2021                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2022                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2023                 if (!td->pid)
2024                         pending--;
2025
2026                 if (td->error)
2027                         exit_value++;
2028
2029                 done_secs += mtime_since_now(&td->epoch) / 1000;
2030                 profile_td_exit(td);
2031         }
2032
2033         if (*nr_running == cputhreads && !pending && realthreads)
2034                 fio_terminate_threads(TERMINATE_ALL);
2035 }
2036
2037 static bool __check_trigger_file(void)
2038 {
2039         struct stat sb;
2040
2041         if (!trigger_file)
2042                 return false;
2043
2044         if (stat(trigger_file, &sb))
2045                 return false;
2046
2047         if (unlink(trigger_file) < 0)
2048                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2049                                                         strerror(errno));
2050
2051         return true;
2052 }
2053
2054 static bool trigger_timedout(void)
2055 {
2056         if (trigger_timeout)
2057                 if (time_since_genesis() >= trigger_timeout) {
2058                         trigger_timeout = 0;
2059                         return true;
2060                 }
2061
2062         return false;
2063 }
2064
2065 void exec_trigger(const char *cmd)
2066 {
2067         int ret;
2068
2069         if (!cmd || cmd[0] == '\0')
2070                 return;
2071
2072         ret = system(cmd);
2073         if (ret == -1)
2074                 log_err("fio: failed executing %s trigger\n", cmd);
2075 }
2076
2077 void check_trigger_file(void)
2078 {
2079         if (__check_trigger_file() || trigger_timedout()) {
2080                 if (nr_clients)
2081                         fio_clients_send_trigger(trigger_remote_cmd);
2082                 else {
2083                         verify_save_state(IO_LIST_ALL);
2084                         fio_terminate_threads(TERMINATE_ALL);
2085                         exec_trigger(trigger_cmd);
2086                 }
2087         }
2088 }
2089
2090 static int fio_verify_load_state(struct thread_data *td)
2091 {
2092         int ret;
2093
2094         if (!td->o.verify_state)
2095                 return 0;
2096
2097         if (is_backend) {
2098                 void *data;
2099
2100                 ret = fio_server_get_verify_state(td->o.name,
2101                                         td->thread_number - 1, &data);
2102                 if (!ret)
2103                         verify_assign_state(td, data);
2104         } else
2105                 ret = verify_load_state(td, "local");
2106
2107         return ret;
2108 }
2109
2110 static void do_usleep(unsigned int usecs)
2111 {
2112         check_for_running_stats();
2113         check_trigger_file();
2114         usleep(usecs);
2115 }
2116
2117 static bool check_mount_writes(struct thread_data *td)
2118 {
2119         struct fio_file *f;
2120         unsigned int i;
2121
2122         if (!td_write(td) || td->o.allow_mounted_write)
2123                 return false;
2124
2125         /*
2126          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2127          * are mkfs'd and mounted.
2128          */
2129         for_each_file(td, f, i) {
2130 #ifdef FIO_HAVE_CHARDEV_SIZE
2131                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2132 #else
2133                 if (f->filetype != FIO_TYPE_BLOCK)
2134 #endif
2135                         continue;
2136                 if (device_is_mounted(f->file_name))
2137                         goto mounted;
2138         }
2139
2140         return false;
2141 mounted:
2142         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2143         return true;
2144 }
2145
2146 static bool waitee_running(struct thread_data *me)
2147 {
2148         const char *waitee = me->o.wait_for;
2149         const char *self = me->o.name;
2150         struct thread_data *td;
2151         int i;
2152
2153         if (!waitee)
2154                 return false;
2155
2156         for_each_td(td, i) {
2157                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2158                         continue;
2159
2160                 if (td->runstate < TD_EXITED) {
2161                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2162                                         self, td->o.name,
2163                                         runstate_to_name(td->runstate));
2164                         return true;
2165                 }
2166         }
2167
2168         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2169         return false;
2170 }
2171
2172 /*
2173  * Main function for kicking off and reaping jobs, as needed.
2174  */
2175 static void run_threads(struct sk_out *sk_out)
2176 {
2177         struct thread_data *td;
2178         unsigned int i, todo, nr_running, nr_started;
2179         uint64_t m_rate, t_rate;
2180         uint64_t spent;
2181
2182         if (fio_gtod_offload && fio_start_gtod_thread())
2183                 return;
2184
2185         fio_idle_prof_init();
2186
2187         set_sig_handlers();
2188
2189         nr_thread = nr_process = 0;
2190         for_each_td(td, i) {
2191                 if (check_mount_writes(td))
2192                         return;
2193                 if (td->o.use_thread)
2194                         nr_thread++;
2195                 else
2196                         nr_process++;
2197         }
2198
2199         if (output_format & FIO_OUTPUT_NORMAL) {
2200                 log_info("Starting ");
2201                 if (nr_thread)
2202                         log_info("%d thread%s", nr_thread,
2203                                                 nr_thread > 1 ? "s" : "");
2204                 if (nr_process) {
2205                         if (nr_thread)
2206                                 log_info(" and ");
2207                         log_info("%d process%s", nr_process,
2208                                                 nr_process > 1 ? "es" : "");
2209                 }
2210                 log_info("\n");
2211                 log_info_flush();
2212         }
2213
2214         todo = thread_number;
2215         nr_running = 0;
2216         nr_started = 0;
2217         m_rate = t_rate = 0;
2218
2219         for_each_td(td, i) {
2220                 print_status_init(td->thread_number - 1);
2221
2222                 if (!td->o.create_serialize)
2223                         continue;
2224
2225                 if (fio_verify_load_state(td))
2226                         goto reap;
2227
2228                 /*
2229                  * do file setup here so it happens sequentially,
2230                  * we don't want X number of threads getting their
2231                  * client data interspersed on disk
2232                  */
2233                 if (setup_files(td)) {
2234 reap:
2235                         exit_value++;
2236                         if (td->error)
2237                                 log_err("fio: pid=%d, err=%d/%s\n",
2238                                         (int) td->pid, td->error, td->verror);
2239                         td_set_runstate(td, TD_REAPED);
2240                         todo--;
2241                 } else {
2242                         struct fio_file *f;
2243                         unsigned int j;
2244
2245                         /*
2246                          * for sharing to work, each job must always open
2247                          * its own files. so close them, if we opened them
2248                          * for creation
2249                          */
2250                         for_each_file(td, f, j) {
2251                                 if (fio_file_open(f))
2252                                         td_io_close_file(td, f);
2253                         }
2254                 }
2255         }
2256
2257         /* start idle threads before io threads start to run */
2258         fio_idle_prof_start();
2259
2260         set_genesis_time();
2261
2262         while (todo) {
2263                 struct thread_data *map[REAL_MAX_JOBS];
2264                 struct timespec this_start;
2265                 int this_jobs = 0, left;
2266                 struct fork_data *fd;
2267
2268                 /*
2269                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2270                  */
2271                 for_each_td(td, i) {
2272                         if (td->runstate != TD_NOT_CREATED)
2273                                 continue;
2274
2275                         /*
2276                          * never got a chance to start, killed by other
2277                          * thread for some reason
2278                          */
2279                         if (td->terminate) {
2280                                 todo--;
2281                                 continue;
2282                         }
2283
2284                         if (td->o.start_delay) {
2285                                 spent = utime_since_genesis();
2286
2287                                 if (td->o.start_delay > spent)
2288                                         continue;
2289                         }
2290
2291                         if (td->o.stonewall && (nr_started || nr_running)) {
2292                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2293                                                         td->o.name);
2294                                 break;
2295                         }
2296
2297                         if (waitee_running(td)) {
2298                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2299                                                 td->o.name, td->o.wait_for);
2300                                 continue;
2301                         }
2302
2303                         init_disk_util(td);
2304
2305                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2306                         td->update_rusage = 0;
2307
2308                         /*
2309                          * Set state to created. Thread will transition
2310                          * to TD_INITIALIZED when it's done setting up.
2311                          */
2312                         td_set_runstate(td, TD_CREATED);
2313                         map[this_jobs++] = td;
2314                         nr_started++;
2315
2316                         fd = calloc(1, sizeof(*fd));
2317                         fd->td = td;
2318                         fd->sk_out = sk_out;
2319
2320                         if (td->o.use_thread) {
2321                                 int ret;
2322
2323                                 dprint(FD_PROCESS, "will pthread_create\n");
2324                                 ret = pthread_create(&td->thread, NULL,
2325                                                         thread_main, fd);
2326                                 if (ret) {
2327                                         log_err("pthread_create: %s\n",
2328                                                         strerror(ret));
2329                                         free(fd);
2330                                         nr_started--;
2331                                         break;
2332                                 }
2333                                 fd = NULL;
2334                                 ret = pthread_detach(td->thread);
2335                                 if (ret)
2336                                         log_err("pthread_detach: %s",
2337                                                         strerror(ret));
2338                         } else {
2339                                 pid_t pid;
2340                                 dprint(FD_PROCESS, "will fork\n");
2341                                 pid = fork();
2342                                 if (!pid) {
2343                                         int ret;
2344
2345                                         ret = (int)(uintptr_t)thread_main(fd);
2346                                         _exit(ret);
2347                                 } else if (i == fio_debug_jobno)
2348                                         *fio_debug_jobp = pid;
2349                         }
2350                         dprint(FD_MUTEX, "wait on startup_sem\n");
2351                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2352                                 log_err("fio: job startup hung? exiting.\n");
2353                                 fio_terminate_threads(TERMINATE_ALL);
2354                                 fio_abort = 1;
2355                                 nr_started--;
2356                                 free(fd);
2357                                 break;
2358                         }
2359                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2360                 }
2361
2362                 /*
2363                  * Wait for the started threads to transition to
2364                  * TD_INITIALIZED.
2365                  */
2366                 fio_gettime(&this_start, NULL);
2367                 left = this_jobs;
2368                 while (left && !fio_abort) {
2369                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2370                                 break;
2371
2372                         do_usleep(100000);
2373
2374                         for (i = 0; i < this_jobs; i++) {
2375                                 td = map[i];
2376                                 if (!td)
2377                                         continue;
2378                                 if (td->runstate == TD_INITIALIZED) {
2379                                         map[i] = NULL;
2380                                         left--;
2381                                 } else if (td->runstate >= TD_EXITED) {
2382                                         map[i] = NULL;
2383                                         left--;
2384                                         todo--;
2385                                         nr_running++; /* work-around... */
2386                                 }
2387                         }
2388                 }
2389
2390                 if (left) {
2391                         log_err("fio: %d job%s failed to start\n", left,
2392                                         left > 1 ? "s" : "");
2393                         for (i = 0; i < this_jobs; i++) {
2394                                 td = map[i];
2395                                 if (!td)
2396                                         continue;
2397                                 kill(td->pid, SIGTERM);
2398                         }
2399                         break;
2400                 }
2401
2402                 /*
2403                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2404                  */
2405                 for_each_td(td, i) {
2406                         if (td->runstate != TD_INITIALIZED)
2407                                 continue;
2408
2409                         if (in_ramp_time(td))
2410                                 td_set_runstate(td, TD_RAMP);
2411                         else
2412                                 td_set_runstate(td, TD_RUNNING);
2413                         nr_running++;
2414                         nr_started--;
2415                         m_rate += ddir_rw_sum(td->o.ratemin);
2416                         t_rate += ddir_rw_sum(td->o.rate);
2417                         todo--;
2418                         fio_sem_up(td->sem);
2419                 }
2420
2421                 reap_threads(&nr_running, &t_rate, &m_rate);
2422
2423                 if (todo)
2424                         do_usleep(100000);
2425         }
2426
2427         while (nr_running) {
2428                 reap_threads(&nr_running, &t_rate, &m_rate);
2429                 do_usleep(10000);
2430         }
2431
2432         fio_idle_prof_stop();
2433
2434         update_io_ticks();
2435 }
2436
2437 static void free_disk_util(void)
2438 {
2439         disk_util_prune_entries();
2440         helper_thread_destroy();
2441 }
2442
2443 int fio_backend(struct sk_out *sk_out)
2444 {
2445         struct thread_data *td;
2446         int i;
2447
2448         if (exec_profile) {
2449                 if (load_profile(exec_profile))
2450                         return 1;
2451                 free(exec_profile);
2452                 exec_profile = NULL;
2453         }
2454         if (!thread_number)
2455                 return 0;
2456
2457         if (write_bw_log) {
2458                 struct log_params p = {
2459                         .log_type = IO_LOG_TYPE_BW,
2460                 };
2461
2462                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2463                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2464                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2465         }
2466
2467         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2468         if (startup_sem == NULL)
2469                 return 1;
2470
2471         set_genesis_time();
2472         stat_init();
2473         helper_thread_create(startup_sem, sk_out);
2474
2475         cgroup_list = smalloc(sizeof(*cgroup_list));
2476         if (cgroup_list)
2477                 INIT_FLIST_HEAD(cgroup_list);
2478
2479         run_threads(sk_out);
2480
2481         helper_thread_exit();
2482
2483         if (!fio_abort) {
2484                 __show_run_stats();
2485                 if (write_bw_log) {
2486                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2487                                 struct io_log *log = agg_io_log[i];
2488
2489                                 flush_log(log, false);
2490                                 free_log(log);
2491                         }
2492                 }
2493         }
2494
2495         for_each_td(td, i) {
2496                 steadystate_free(td);
2497                 fio_options_free(td);
2498                 if (td->rusage_sem) {
2499                         fio_sem_remove(td->rusage_sem);
2500                         td->rusage_sem = NULL;
2501                 }
2502                 fio_sem_remove(td->sem);
2503                 td->sem = NULL;
2504         }
2505
2506         free_disk_util();
2507         if (cgroup_list) {
2508                 cgroup_kill(cgroup_list);
2509                 sfree(cgroup_list);
2510         }
2511
2512         fio_sem_remove(startup_sem);
2513         stat_exit();
2514         return exit_value;
2515 }