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                  * Jobs read from iolog do not use this stop condition.
 970                  */
 971                 if (bytes_issued >= total_bytes &&
 972                     !td->o.read_iolog_file &&
 973                     (!td->o.time_based ||
 974                      (td->o.time_based && td->o.verify != VERIFY_NONE)))
 975                         break;
 976
 977                 io_u = get_io_u(td);
 978                 if (IS_ERR_OR_NULL(io_u)) {
 979                         int err = PTR_ERR(io_u);
 980
 981                         io_u = NULL;
 982                         ddir = DDIR_INVAL;
 983                         if (err == -EBUSY) {
 984                                 ret = FIO_Q_BUSY;
 985                                 goto reap;
 986                         }
 987                         if (td->o.latency_target)
 988                                 goto reap;
 989                         break;
 990                 }
 991
 992                 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
 993                         populate_verify_io_u(td, io_u);
 994
 995                 ddir = io_u->ddir;
 996
 997                 /*
 998                  * Add verification end_io handler if:
 999                  *      - Asked to verify (!td_rw(td))
1000                  *      - Or the io_u is from our verify list (mixed write/ver)
1001                  */
1002                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1003                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1004
1005                         if (!td->o.verify_pattern_bytes) {
1006                                 io_u->rand_seed = __rand(&td->verify_state);
1007                                 if (sizeof(int) != sizeof(long *))
1008                                         io_u->rand_seed *= __rand(&td->verify_state);
1009                         }
1010
1011                         if (verify_state_should_stop(td, io_u)) {
1012                                 put_io_u(td, io_u);
1013                                 break;
1014                         }
1015
1016                         if (td->o.verify_async)
1017                                 io_u->end_io = verify_io_u_async;
1018                         else
1019                                 io_u->end_io = verify_io_u;
1020                         td_set_runstate(td, TD_VERIFYING);
1021                 } else if (in_ramp_time(td))
1022                         td_set_runstate(td, TD_RAMP);
1023                 else
1024                         td_set_runstate(td, TD_RUNNING);
1025
1026                 /*
1027                  * Always log IO before it's issued, so we know the specific
1028                  * order of it. The logged unit will track when the IO has
1029                  * completed.
1030                  */
1031                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1032                     td->o.do_verify &&
1033                     td->o.verify != VERIFY_NONE &&
1034                     !td->o.experimental_verify)
1035                         log_io_piece(td, io_u);
1036
1037                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1038                         const unsigned long long blen = io_u->xfer_buflen;
1039                         const enum fio_ddir __ddir = acct_ddir(io_u);
1040
1041                         if (td->error)
1042                                 break;
1043
1044                         workqueue_enqueue(&td->io_wq, &io_u->work);
1045                         ret = FIO_Q_QUEUED;
1046
1047                         if (ddir_rw(__ddir)) {
1048                                 td->io_issues[__ddir]++;
1049                                 td->io_issue_bytes[__ddir] += blen;
1050                                 td->rate_io_issue_bytes[__ddir] += blen;
1051                         }
1052
1053                         if (should_check_rate(td))
1054                                 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1055
1056                 } else {
1057                         ret = io_u_submit(td, io_u);
1058
1059                         if (should_check_rate(td))
1060                                 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1061
1062                         if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1063                                 break;
1064
1065                         /*
1066                          * See if we need to complete some commands. Note that
1067                          * we can get BUSY even without IO queued, if the
1068                          * system is resource starved.
1069                          */
1070 reap:
1071                         full = queue_full(td) ||
1072                                 (ret == FIO_Q_BUSY && td->cur_depth);
1073                         if (full || io_in_polling(td))
1074                                 ret = wait_for_completions(td, &comp_time);
1075                 }
1076                 if (ret < 0)
1077                         break;
1078                 if (!ddir_rw_sum(td->bytes_done) &&
1079                     !td_ioengine_flagged(td, FIO_NOIO))
1080                         continue;
1081
1082                 if (!in_ramp_time(td) && should_check_rate(td)) {
1083                         if (check_min_rate(td, &comp_time)) {
1084                                 if (exitall_on_terminate || td->o.exitall_error)
1085                                         fio_terminate_threads(td->groupid);
1086                                 td_verror(td, EIO, "check_min_rate");
1087                                 break;
1088                         }
1089                 }
1090                 if (!in_ramp_time(td) && td->o.latency_target)
1091                         lat_target_check(td);
1092
1093                 if (ddir_rw(ddir) && td->o.thinktime)
1094                         handle_thinktime(td, ddir);
1095         }
1096
1097         check_update_rusage(td);
1098
1099         if (td->trim_entries)
1100                 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1101
1102         if (td->o.fill_device && td->error == ENOSPC) {
1103                 td->error = 0;
1104                 fio_mark_td_terminate(td);
1105         }
1106         if (!td->error) {
1107                 struct fio_file *f;
1108
1109                 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1110                         workqueue_flush(&td->io_wq);
1111                         i = 0;
1112                 } else
1113                         i = td->cur_depth;
1114
1115                 if (i) {
1116                         ret = io_u_queued_complete(td, i);
1117                         if (td->o.fill_device && td->error == ENOSPC)
1118                                 td->error = 0;
1119                 }
1120
1121                 if (should_fsync(td) && td->o.end_fsync) {
1122                         td_set_runstate(td, TD_FSYNCING);
1123
1124                         for_each_file(td, f, i) {
1125                                 if (!fio_file_fsync(td, f))
1126                                         continue;
1127
1128                                 log_err("fio: end_fsync failed for file %s\n",
1129                                                                 f->file_name);
1130                         }
1131                 }
1132         } else
1133                 cleanup_pending_aio(td);
1134
1135         /*
1136          * stop job if we failed doing any IO
1137          */
1138         if (!ddir_rw_sum(td->this_io_bytes))
1139                 td->done = 1;
1140
1141         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1142                 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1143 }
1144
1145 static void free_file_completion_logging(struct thread_data *td)
1146 {
1147         struct fio_file *f;
1148         unsigned int i;
1149
1150         for_each_file(td, f, i) {
1151                 if (!f->last_write_comp)
1152                         break;
1153                 sfree(f->last_write_comp);
1154         }
1155 }
1156
1157 static int init_file_completion_logging(struct thread_data *td,
1158                                         unsigned int depth)
1159 {
1160         struct fio_file *f;
1161         unsigned int i;
1162
1163         if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1164                 return 0;
1165
1166         for_each_file(td, f, i) {
1167                 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1168                 if (!f->last_write_comp)
1169                         goto cleanup;
1170         }
1171
1172         return 0;
1173
1174 cleanup:
1175         free_file_completion_logging(td);
1176         log_err("fio: failed to alloc write comp data\n");
1177         return 1;
1178 }
1179
1180 static void cleanup_io_u(struct thread_data *td)
1181 {
1182         struct io_u *io_u;
1183
1184         while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1185
1186                 if (td->io_ops->io_u_free)
1187                         td->io_ops->io_u_free(td, io_u);
1188
1189                 fio_memfree(io_u, sizeof(*io_u));
1190         }
1191
1192         free_io_mem(td);
1193
1194         io_u_rexit(&td->io_u_requeues);
1195         io_u_qexit(&td->io_u_freelist);
1196         io_u_qexit(&td->io_u_all);
1197
1198         free_file_completion_logging(td);
1199 }
1200
1201 static int init_io_u(struct thread_data *td)
1202 {
1203         struct io_u *io_u;
1204         int cl_align, i, max_units;
1205         int err;
1206
1207         max_units = td->o.iodepth;
1208
1209         err = 0;
1210         err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1211         err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1212         err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1213
1214         if (err) {
1215                 log_err("fio: failed setting up IO queues\n");
1216                 return 1;
1217         }
1218
1219         cl_align = os_cache_line_size();
1220
1221         for (i = 0; i < max_units; i++) {
1222                 void *ptr;
1223
1224                 if (td->terminate)
1225                         return 1;
1226
1227                 ptr = fio_memalign(cl_align, sizeof(*io_u));
1228                 if (!ptr) {
1229                         log_err("fio: unable to allocate aligned memory\n");
1230                         break;
1231                 }
1232
1233                 io_u = ptr;
1234                 memset(io_u, 0, sizeof(*io_u));
1235                 INIT_FLIST_HEAD(&io_u->verify_list);
1236                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1237
1238                 io_u->index = i;
1239                 io_u->flags = IO_U_F_FREE;
1240                 io_u_qpush(&td->io_u_freelist, io_u);
1241
1242                 /*
1243                  * io_u never leaves this stack, used for iteration of all
1244                  * io_u buffers.
1245                  */
1246                 io_u_qpush(&td->io_u_all, io_u);
1247
1248                 if (td->io_ops->io_u_init) {
1249                         int ret = td->io_ops->io_u_init(td, io_u);
1250
1251                         if (ret) {
1252                                 log_err("fio: failed to init engine data: %d\n", ret);
1253                                 return 1;
1254                         }
1255                 }
1256         }
1257
1258         init_io_u_buffers(td);
1259
1260         if (init_file_completion_logging(td, max_units))
1261                 return 1;
1262
1263         return 0;
1264 }
1265
1266 int init_io_u_buffers(struct thread_data *td)
1267 {
1268         struct io_u *io_u;
1269         unsigned long long max_bs, min_write;
1270         int i, max_units;
1271         int data_xfer = 1;
1272         char *p;
1273
1274         max_units = td->o.iodepth;
1275         max_bs = td_max_bs(td);
1276         min_write = td->o.min_bs[DDIR_WRITE];
1277         td->orig_buffer_size = (unsigned long long) max_bs
1278                                         * (unsigned long long) max_units;
1279
1280         if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1281                 data_xfer = 0;
1282
1283         /*
1284          * if we may later need to do address alignment, then add any
1285          * possible adjustment here so that we don't cause a buffer
1286          * overflow later. this adjustment may be too much if we get
1287          * lucky and the allocator gives us an aligned address.
1288          */
1289         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1290             td_ioengine_flagged(td, FIO_RAWIO))
1291                 td->orig_buffer_size += page_mask + td->o.mem_align;
1292
1293         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1294                 unsigned long long bs;
1295
1296                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1297                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1298         }
1299
1300         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1301                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1302                 return 1;
1303         }
1304
1305         if (data_xfer && allocate_io_mem(td))
1306                 return 1;
1307
1308         if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1309             td_ioengine_flagged(td, FIO_RAWIO))
1310                 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1311         else
1312                 p = td->orig_buffer;
1313
1314         for (i = 0; i < max_units; i++) {
1315                 io_u = td->io_u_all.io_us[i];
1316                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1317
1318                 if (data_xfer) {
1319                         io_u->buf = p;
1320                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1321
1322                         if (td_write(td))
1323                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
1324                         if (td_write(td) && td->o.verify_pattern_bytes) {
1325                                 /*
1326                                  * Fill the buffer with the pattern if we are
1327                                  * going to be doing writes.
1328                                  */
1329                                 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1330                         }
1331                 }
1332                 p += max_bs;
1333         }
1334
1335         return 0;
1336 }
1337
1338 /*
1339  * This function is Linux specific.
1340  * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1341  */
1342 static int switch_ioscheduler(struct thread_data *td)
1343 {
1344 #ifdef FIO_HAVE_IOSCHED_SWITCH
1345         char tmp[256], tmp2[128], *p;
1346         FILE *f;
1347         int ret;
1348
1349         if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1350                 return 0;
1351
1352         assert(td->files && td->files[0]);
1353         sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1354
1355         f = fopen(tmp, "r+");
1356         if (!f) {
1357                 if (errno == ENOENT) {
1358                         log_err("fio: os or kernel doesn't support IO scheduler"
1359                                 " switching\n");
1360                         return 0;
1361                 }
1362                 td_verror(td, errno, "fopen iosched");
1363                 return 1;
1364         }
1365
1366         /*
1367          * Set io scheduler.
1368          */
1369         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1370         if (ferror(f) || ret != 1) {
1371                 td_verror(td, errno, "fwrite");
1372                 fclose(f);
1373                 return 1;
1374         }
1375
1376         rewind(f);
1377
1378         /*
1379          * Read back and check that the selected scheduler is now the default.
1380          */
1381         ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1382         if (ferror(f) || ret < 0) {
1383                 td_verror(td, errno, "fread");
1384                 fclose(f);
1385                 return 1;
1386         }
1387         tmp[ret] = '\0';
1388         /*
1389          * either a list of io schedulers or "none\n" is expected. Strip the
1390          * trailing newline.
1391          */
1392         p = tmp;
1393         strsep(&p, "\n");
1394
1395         /*
1396          * Write to "none" entry doesn't fail, so check the result here.
1397          */
1398         if (!strcmp(tmp, "none")) {
1399                 log_err("fio: io scheduler is not tunable\n");
1400                 fclose(f);
1401                 return 0;
1402         }
1403
1404         sprintf(tmp2, "[%s]", td->o.ioscheduler);
1405         if (!strstr(tmp, tmp2)) {
1406                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1407                 td_verror(td, EINVAL, "iosched_switch");
1408                 fclose(f);
1409                 return 1;
1410         }
1411
1412         fclose(f);
1413         return 0;
1414 #else
1415         return 0;
1416 #endif
1417 }
1418
1419 static bool keep_running(struct thread_data *td)
1420 {
1421         unsigned long long limit;
1422
1423         if (td->done)
1424                 return false;
1425         if (td->terminate)
1426                 return false;
1427         if (td->o.time_based)
1428                 return true;
1429         if (td->o.loops) {
1430                 td->o.loops--;
1431                 return true;
1432         }
1433         if (exceeds_number_ios(td))
1434                 return false;
1435
1436         if (td->o.io_size)
1437                 limit = td->o.io_size;
1438         else
1439                 limit = td->o.size;
1440
1441         if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1442                 uint64_t diff;
1443
1444                 /*
1445                  * If the difference is less than the maximum IO size, we
1446                  * are done.
1447                  */
1448                 diff = limit - ddir_rw_sum(td->io_bytes);
1449                 if (diff < td_max_bs(td))
1450                         return false;
1451
1452                 if (fio_files_done(td) && !td->o.io_size)
1453                         return false;
1454
1455                 return true;
1456         }
1457
1458         return false;
1459 }
1460
1461 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1462 {
1463         size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1464         int ret;
1465         char *str;
1466
1467         str = malloc(newlen);
1468         sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1469
1470         log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1471         ret = system(str);
1472         if (ret == -1)
1473                 log_err("fio: exec of cmd <%s> failed\n", str);
1474
1475         free(str);
1476         return ret;
1477 }
1478
1479 /*
1480  * Dry run to compute correct state of numberio for verification.
1481  */
1482 static uint64_t do_dry_run(struct thread_data *td)
1483 {
1484         td_set_runstate(td, TD_RUNNING);
1485
1486         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1487                 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1488                 struct io_u *io_u;
1489                 int ret;
1490
1491                 if (td->terminate || td->done)
1492                         break;
1493
1494                 io_u = get_io_u(td);
1495                 if (IS_ERR_OR_NULL(io_u))
1496                         break;
1497
1498                 io_u_set(td, io_u, IO_U_F_FLIGHT);
1499                 io_u->error = 0;
1500                 io_u->resid = 0;
1501                 if (ddir_rw(acct_ddir(io_u)))
1502                         td->io_issues[acct_ddir(io_u)]++;
1503                 if (ddir_rw(io_u->ddir)) {
1504                         io_u_mark_depth(td, 1);
1505                         td->ts.total_io_u[io_u->ddir]++;
1506                 }
1507
1508                 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1509                     td->o.do_verify &&
1510                     td->o.verify != VERIFY_NONE &&
1511                     !td->o.experimental_verify)
1512                         log_io_piece(td, io_u);
1513
1514                 ret = io_u_sync_complete(td, io_u);
1515                 (void) ret;
1516         }
1517
1518         return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1519 }
1520
1521 struct fork_data {
1522         struct thread_data *td;
1523         struct sk_out *sk_out;
1524 };
1525
1526 /*
1527  * Entry point for the thread based jobs. The process based jobs end up
1528  * here as well, after a little setup.
1529  */
1530 static void *thread_main(void *data)
1531 {
1532         struct fork_data *fd = data;
1533         unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1534         struct thread_data *td = fd->td;
1535         struct thread_options *o = &td->o;
1536         struct sk_out *sk_out = fd->sk_out;
1537         uint64_t bytes_done[DDIR_RWDIR_CNT];
1538         int deadlock_loop_cnt;
1539         bool clear_state, did_some_io;
1540         int ret;
1541
1542         sk_out_assign(sk_out);
1543         free(fd);
1544
1545         if (!o->use_thread) {
1546                 setsid();
1547                 td->pid = getpid();
1548         } else
1549                 td->pid = gettid();
1550
1551         fio_local_clock_init();
1552
1553         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1554
1555         if (is_backend)
1556                 fio_server_send_start(td);
1557
1558         INIT_FLIST_HEAD(&td->io_log_list);
1559         INIT_FLIST_HEAD(&td->io_hist_list);
1560         INIT_FLIST_HEAD(&td->verify_list);
1561         INIT_FLIST_HEAD(&td->trim_list);
1562         td->io_hist_tree = RB_ROOT;
1563
1564         ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1565         if (ret) {
1566                 td_verror(td, ret, "mutex_cond_init_pshared");
1567                 goto err;
1568         }
1569         ret = cond_init_pshared(&td->verify_cond);
1570         if (ret) {
1571                 td_verror(td, ret, "mutex_cond_pshared");
1572                 goto err;
1573         }
1574
1575         td_set_runstate(td, TD_INITIALIZED);
1576         dprint(FD_MUTEX, "up startup_sem\n");
1577         fio_sem_up(startup_sem);
1578         dprint(FD_MUTEX, "wait on td->sem\n");
1579         fio_sem_down(td->sem);
1580         dprint(FD_MUTEX, "done waiting on td->sem\n");
1581
1582         /*
1583          * A new gid requires privilege, so we need to do this before setting
1584          * the uid.
1585          */
1586         if (o->gid != -1U && setgid(o->gid)) {
1587                 td_verror(td, errno, "setgid");
1588                 goto err;
1589         }
1590         if (o->uid != -1U && setuid(o->uid)) {
1591                 td_verror(td, errno, "setuid");
1592                 goto err;
1593         }
1594
1595         /*
1596          * Do this early, we don't want the compress threads to be limited
1597          * to the same CPUs as the IO workers. So do this before we set
1598          * any potential CPU affinity
1599          */
1600         if (iolog_compress_init(td, sk_out))
1601                 goto err;
1602
1603         /*
1604          * If we have a gettimeofday() thread, make sure we exclude that
1605          * thread from this job
1606          */
1607         if (o->gtod_cpu)
1608                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1609
1610         /*
1611          * Set affinity first, in case it has an impact on the memory
1612          * allocations.
1613          */
1614         if (fio_option_is_set(o, cpumask)) {
1615                 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1616                         ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1617                         if (!ret) {
1618                                 log_err("fio: no CPUs set\n");
1619                                 log_err("fio: Try increasing number of available CPUs\n");
1620                                 td_verror(td, EINVAL, "cpus_split");
1621                                 goto err;
1622                         }
1623                 }
1624                 ret = fio_setaffinity(td->pid, o->cpumask);
1625                 if (ret == -1) {
1626                         td_verror(td, errno, "cpu_set_affinity");
1627                         goto err;
1628                 }
1629         }
1630
1631 #ifdef CONFIG_LIBNUMA
1632         /* numa node setup */
1633         if (fio_option_is_set(o, numa_cpunodes) ||
1634             fio_option_is_set(o, numa_memnodes)) {
1635                 struct bitmask *mask;
1636
1637                 if (numa_available() < 0) {
1638                         td_verror(td, errno, "Does not support NUMA API\n");
1639                         goto err;
1640                 }
1641
1642                 if (fio_option_is_set(o, numa_cpunodes)) {
1643                         mask = numa_parse_nodestring(o->numa_cpunodes);
1644                         ret = numa_run_on_node_mask(mask);
1645                         numa_free_nodemask(mask);
1646                         if (ret == -1) {
1647                                 td_verror(td, errno, \
1648                                         "numa_run_on_node_mask failed\n");
1649                                 goto err;
1650                         }
1651                 }
1652
1653                 if (fio_option_is_set(o, numa_memnodes)) {
1654                         mask = NULL;
1655                         if (o->numa_memnodes)
1656                                 mask = numa_parse_nodestring(o->numa_memnodes);
1657
1658                         switch (o->numa_mem_mode) {
1659                         case MPOL_INTERLEAVE:
1660                                 numa_set_interleave_mask(mask);
1661                                 break;
1662                         case MPOL_BIND:
1663                                 numa_set_membind(mask);
1664                                 break;
1665                         case MPOL_LOCAL:
1666                                 numa_set_localalloc();
1667                                 break;
1668                         case MPOL_PREFERRED:
1669                                 numa_set_preferred(o->numa_mem_prefer_node);
1670                                 break;
1671                         case MPOL_DEFAULT:
1672                         default:
1673                                 break;
1674                         }
1675
1676                         if (mask)
1677                                 numa_free_nodemask(mask);
1678
1679                 }
1680         }
1681 #endif
1682
1683         if (fio_pin_memory(td))
1684                 goto err;
1685
1686         /*
1687          * May alter parameters that init_io_u() will use, so we need to
1688          * do this first.
1689          */
1690         if (!init_iolog(td))
1691                 goto err;
1692
1693         if (init_io_u(td))
1694                 goto err;
1695
1696         if (o->verify_async && verify_async_init(td))
1697                 goto err;
1698
1699         if (fio_option_is_set(o, ioprio) ||
1700             fio_option_is_set(o, ioprio_class)) {
1701                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1702                 if (ret == -1) {
1703                         td_verror(td, errno, "ioprio_set");
1704                         goto err;
1705                 }
1706         }
1707
1708         if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1709                 goto err;
1710
1711         errno = 0;
1712         if (nice(o->nice) == -1 && errno != 0) {
1713                 td_verror(td, errno, "nice");
1714                 goto err;
1715         }
1716
1717         if (o->ioscheduler && switch_ioscheduler(td))
1718                 goto err;
1719
1720         if (!o->create_serialize && setup_files(td))
1721                 goto err;
1722
1723         if (td_io_init(td))
1724                 goto err;
1725
1726         if (!init_random_map(td))
1727                 goto err;
1728
1729         if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1730                 goto err;
1731
1732         if (o->pre_read && !pre_read_files(td))
1733                 goto err;
1734
1735         fio_verify_init(td);
1736
1737         if (rate_submit_init(td, sk_out))
1738                 goto err;
1739
1740         set_epoch_time(td, o->log_unix_epoch);
1741         fio_getrusage(&td->ru_start);
1742         memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1743         memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1744         memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1745
1746         if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1747                         o->ratemin[DDIR_TRIM]) {
1748                 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1749                                         sizeof(td->bw_sample_time));
1750                 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1751                                         sizeof(td->bw_sample_time));
1752                 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1753                                         sizeof(td->bw_sample_time));
1754         }
1755
1756         memset(bytes_done, 0, sizeof(bytes_done));
1757         clear_state = false;
1758         did_some_io = false;
1759
1760         while (keep_running(td)) {
1761                 uint64_t verify_bytes;
1762
1763                 fio_gettime(&td->start, NULL);
1764                 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1765
1766                 if (clear_state) {
1767                         clear_io_state(td, 0);
1768
1769                         if (o->unlink_each_loop && unlink_all_files(td))
1770                                 break;
1771                 }
1772
1773                 prune_io_piece_log(td);
1774
1775                 if (td->o.verify_only && td_write(td))
1776                         verify_bytes = do_dry_run(td);
1777                 else {
1778                         do_io(td, bytes_done);
1779
1780                         if (!ddir_rw_sum(bytes_done)) {
1781                                 fio_mark_td_terminate(td);
1782                                 verify_bytes = 0;
1783                         } else {
1784                                 verify_bytes = bytes_done[DDIR_WRITE] +
1785                                                 bytes_done[DDIR_TRIM];
1786                         }
1787                 }
1788
1789                 /*
1790                  * If we took too long to shut down, the main thread could
1791                  * already consider us reaped/exited. If that happens, break
1792                  * out and clean up.
1793                  */
1794                 if (td->runstate >= TD_EXITED)
1795                         break;
1796
1797                 clear_state = true;
1798
1799                 /*
1800                  * Make sure we've successfully updated the rusage stats
1801                  * before waiting on the stat mutex. Otherwise we could have
1802                  * the stat thread holding stat mutex and waiting for
1803                  * the rusage_sem, which would never get upped because
1804                  * this thread is waiting for the stat mutex.
1805                  */
1806                 deadlock_loop_cnt = 0;
1807                 do {
1808                         check_update_rusage(td);
1809                         if (!fio_sem_down_trylock(stat_sem))
1810                                 break;
1811                         usleep(1000);
1812                         if (deadlock_loop_cnt++ > 5000) {
1813                                 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1814                                 td->error = EDEADLK;
1815                                 goto err;
1816                         }
1817                 } while (1);
1818
1819                 if (td_read(td) && td->io_bytes[DDIR_READ])
1820                         update_runtime(td, elapsed_us, DDIR_READ);
1821                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1822                         update_runtime(td, elapsed_us, DDIR_WRITE);
1823                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1824                         update_runtime(td, elapsed_us, DDIR_TRIM);
1825                 fio_gettime(&td->start, NULL);
1826                 fio_sem_up(stat_sem);
1827
1828                 if (td->error || td->terminate)
1829                         break;
1830
1831                 if (!o->do_verify ||
1832                     o->verify == VERIFY_NONE ||
1833                     td_ioengine_flagged(td, FIO_UNIDIR))
1834                         continue;
1835
1836                 if (ddir_rw_sum(bytes_done))
1837                         did_some_io = true;
1838
1839                 clear_io_state(td, 0);
1840
1841                 fio_gettime(&td->start, NULL);
1842
1843                 do_verify(td, verify_bytes);
1844
1845                 /*
1846                  * See comment further up for why this is done here.
1847                  */
1848                 check_update_rusage(td);
1849
1850                 fio_sem_down(stat_sem);
1851                 update_runtime(td, elapsed_us, DDIR_READ);
1852                 fio_gettime(&td->start, NULL);
1853                 fio_sem_up(stat_sem);
1854
1855                 if (td->error || td->terminate)
1856                         break;
1857         }
1858
1859         /*
1860          * If td ended up with no I/O when it should have had,
1861          * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1862          * (Are we not missing other flags that can be ignored ?)
1863          */
1864         if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1865             !did_some_io && !td->o.create_only &&
1866             !(td_ioengine_flagged(td, FIO_NOIO) ||
1867               td_ioengine_flagged(td, FIO_DISKLESSIO)))
1868                 log_err("%s: No I/O performed by %s, "
1869                          "perhaps try --debug=io option for details?\n",
1870                          td->o.name, td->io_ops->name);
1871
1872         td_set_runstate(td, TD_FINISHING);
1873
1874         update_rusage_stat(td);
1875         td->ts.total_run_time = mtime_since_now(&td->epoch);
1876         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1877         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1878         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1879
1880         if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1881             (td->o.verify != VERIFY_NONE && td_write(td)))
1882                 verify_save_state(td->thread_number);
1883
1884         fio_unpin_memory(td);
1885
1886         td_writeout_logs(td, true);
1887
1888         iolog_compress_exit(td);
1889         rate_submit_exit(td);
1890
1891         if (o->exec_postrun)
1892                 exec_string(o, o->exec_postrun, (const char *)"postrun");
1893
1894         if (exitall_on_terminate || (o->exitall_error && td->error))
1895                 fio_terminate_threads(td->groupid);
1896
1897 err:
1898         if (td->error)
1899                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1900                                                         td->verror);
1901
1902         if (o->verify_async)
1903                 verify_async_exit(td);
1904
1905         close_and_free_files(td);
1906         cleanup_io_u(td);
1907         close_ioengine(td);
1908         cgroup_shutdown(td, cgroup_mnt);
1909         verify_free_state(td);
1910
1911         if (td->zone_state_index) {
1912                 int i;
1913
1914                 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1915                         free(td->zone_state_index[i]);
1916                 free(td->zone_state_index);
1917                 td->zone_state_index = NULL;
1918         }
1919
1920         if (fio_option_is_set(o, cpumask)) {
1921                 ret = fio_cpuset_exit(&o->cpumask);
1922                 if (ret)
1923                         td_verror(td, ret, "fio_cpuset_exit");
1924         }
1925
1926         /*
1927          * do this very late, it will log file closing as well
1928          */
1929         if (o->write_iolog_file)
1930                 write_iolog_close(td);
1931         if (td->io_log_rfile)
1932                 fclose(td->io_log_rfile);
1933
1934         td_set_runstate(td, TD_EXITED);
1935
1936         /*
1937          * Do this last after setting our runstate to exited, so we
1938          * know that the stat thread is signaled.
1939          */
1940         check_update_rusage(td);
1941
1942         sk_out_drop();
1943         return (void *) (uintptr_t) td->error;
1944 }
1945
1946 /*
1947  * Run over the job map and reap the threads that have exited, if any.
1948  */
1949 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1950                          uint64_t *m_rate)
1951 {
1952         struct thread_data *td;
1953         unsigned int cputhreads, realthreads, pending;
1954         int i, status, ret;
1955
1956         /*
1957          * reap exited threads (TD_EXITED -> TD_REAPED)
1958          */
1959         realthreads = pending = cputhreads = 0;
1960         for_each_td(td, i) {
1961                 int flags = 0;
1962
1963                  if (!strcmp(td->o.ioengine, "cpuio"))
1964                         cputhreads++;
1965                 else
1966                         realthreads++;
1967
1968                 if (!td->pid) {
1969                         pending++;
1970                         continue;
1971                 }
1972                 if (td->runstate == TD_REAPED)
1973                         continue;
1974                 if (td->o.use_thread) {
1975                         if (td->runstate == TD_EXITED) {
1976                                 td_set_runstate(td, TD_REAPED);
1977                                 goto reaped;
1978                         }
1979                         continue;
1980                 }
1981
1982                 flags = WNOHANG;
1983                 if (td->runstate == TD_EXITED)
1984                         flags = 0;
1985
1986                 /*
1987                  * check if someone quit or got killed in an unusual way
1988                  */
1989                 ret = waitpid(td->pid, &status, flags);
1990                 if (ret < 0) {
1991                         if (errno == ECHILD) {
1992                                 log_err("fio: pid=%d disappeared %d\n",
1993                                                 (int) td->pid, td->runstate);
1994                                 td->sig = ECHILD;
1995                                 td_set_runstate(td, TD_REAPED);
1996                                 goto reaped;
1997                         }
1998                         perror("waitpid");
1999                 } else if (ret == td->pid) {
2000                         if (WIFSIGNALED(status)) {
2001                                 int sig = WTERMSIG(status);
2002
2003                                 if (sig != SIGTERM && sig != SIGUSR2)
2004                                         log_err("fio: pid=%d, got signal=%d\n",
2005                                                         (int) td->pid, sig);
2006                                 td->sig = sig;
2007                                 td_set_runstate(td, TD_REAPED);
2008                                 goto reaped;
2009                         }
2010                         if (WIFEXITED(status)) {
2011                                 if (WEXITSTATUS(status) && !td->error)
2012                                         td->error = WEXITSTATUS(status);
2013
2014                                 td_set_runstate(td, TD_REAPED);
2015                                 goto reaped;
2016                         }
2017                 }
2018
2019                 /*
2020                  * If the job is stuck, do a forceful timeout of it and
2021                  * move on.
2022                  */
2023                 if (td->terminate &&
2024                     td->runstate < TD_FSYNCING &&
2025                     time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2026                         log_err("fio: job '%s' (state=%d) hasn't exited in "
2027                                 "%lu seconds, it appears to be stuck. Doing "
2028                                 "forceful exit of this job.\n",
2029                                 td->o.name, td->runstate,
2030                                 (unsigned long) time_since_now(&td->terminate_time));
2031                         td_set_runstate(td, TD_REAPED);
2032                         goto reaped;
2033                 }
2034
2035                 /*
2036                  * thread is not dead, continue
2037                  */
2038                 pending++;
2039                 continue;
2040 reaped:
2041                 (*nr_running)--;
2042                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2043                 (*t_rate) -= ddir_rw_sum(td->o.rate);
2044                 if (!td->pid)
2045                         pending--;
2046
2047                 if (td->error)
2048                         exit_value++;
2049
2050                 done_secs += mtime_since_now(&td->epoch) / 1000;
2051                 profile_td_exit(td);
2052         }
2053
2054         if (*nr_running == cputhreads && !pending && realthreads)
2055                 fio_terminate_threads(TERMINATE_ALL);
2056 }
2057
2058 static bool __check_trigger_file(void)
2059 {
2060         struct stat sb;
2061
2062         if (!trigger_file)
2063                 return false;
2064
2065         if (stat(trigger_file, &sb))
2066                 return false;
2067
2068         if (unlink(trigger_file) < 0)
2069                 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2070                                                         strerror(errno));
2071
2072         return true;
2073 }
2074
2075 static bool trigger_timedout(void)
2076 {
2077         if (trigger_timeout)
2078                 if (time_since_genesis() >= trigger_timeout) {
2079                         trigger_timeout = 0;
2080                         return true;
2081                 }
2082
2083         return false;
2084 }
2085
2086 void exec_trigger(const char *cmd)
2087 {
2088         int ret;
2089
2090         if (!cmd || cmd[0] == '\0')
2091                 return;
2092
2093         ret = system(cmd);
2094         if (ret == -1)
2095                 log_err("fio: failed executing %s trigger\n", cmd);
2096 }
2097
2098 void check_trigger_file(void)
2099 {
2100         if (__check_trigger_file() || trigger_timedout()) {
2101                 if (nr_clients)
2102                         fio_clients_send_trigger(trigger_remote_cmd);
2103                 else {
2104                         verify_save_state(IO_LIST_ALL);
2105                         fio_terminate_threads(TERMINATE_ALL);
2106                         exec_trigger(trigger_cmd);
2107                 }
2108         }
2109 }
2110
2111 static int fio_verify_load_state(struct thread_data *td)
2112 {
2113         int ret;
2114
2115         if (!td->o.verify_state)
2116                 return 0;
2117
2118         if (is_backend) {
2119                 void *data;
2120
2121                 ret = fio_server_get_verify_state(td->o.name,
2122                                         td->thread_number - 1, &data);
2123                 if (!ret)
2124                         verify_assign_state(td, data);
2125         } else
2126                 ret = verify_load_state(td, "local");
2127
2128         return ret;
2129 }
2130
2131 static void do_usleep(unsigned int usecs)
2132 {
2133         check_for_running_stats();
2134         check_trigger_file();
2135         usleep(usecs);
2136 }
2137
2138 static bool check_mount_writes(struct thread_data *td)
2139 {
2140         struct fio_file *f;
2141         unsigned int i;
2142
2143         if (!td_write(td) || td->o.allow_mounted_write)
2144                 return false;
2145
2146         /*
2147          * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2148          * are mkfs'd and mounted.
2149          */
2150         for_each_file(td, f, i) {
2151 #ifdef FIO_HAVE_CHARDEV_SIZE
2152                 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2153 #else
2154                 if (f->filetype != FIO_TYPE_BLOCK)
2155 #endif
2156                         continue;
2157                 if (device_is_mounted(f->file_name))
2158                         goto mounted;
2159         }
2160
2161         return false;
2162 mounted:
2163         log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2164         return true;
2165 }
2166
2167 static bool waitee_running(struct thread_data *me)
2168 {
2169         const char *waitee = me->o.wait_for;
2170         const char *self = me->o.name;
2171         struct thread_data *td;
2172         int i;
2173
2174         if (!waitee)
2175                 return false;
2176
2177         for_each_td(td, i) {
2178                 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2179                         continue;
2180
2181                 if (td->runstate < TD_EXITED) {
2182                         dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2183                                         self, td->o.name,
2184                                         runstate_to_name(td->runstate));
2185                         return true;
2186                 }
2187         }
2188
2189         dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2190         return false;
2191 }
2192
2193 /*
2194  * Main function for kicking off and reaping jobs, as needed.
2195  */
2196 static void run_threads(struct sk_out *sk_out)
2197 {
2198         struct thread_data *td;
2199         unsigned int i, todo, nr_running, nr_started;
2200         uint64_t m_rate, t_rate;
2201         uint64_t spent;
2202
2203         if (fio_gtod_offload && fio_start_gtod_thread())
2204                 return;
2205
2206         fio_idle_prof_init();
2207
2208         set_sig_handlers();
2209
2210         nr_thread = nr_process = 0;
2211         for_each_td(td, i) {
2212                 if (check_mount_writes(td))
2213                         return;
2214                 if (td->o.use_thread)
2215                         nr_thread++;
2216                 else
2217                         nr_process++;
2218         }
2219
2220         if (output_format & FIO_OUTPUT_NORMAL) {
2221                 log_info("Starting ");
2222                 if (nr_thread)
2223                         log_info("%d thread%s", nr_thread,
2224                                                 nr_thread > 1 ? "s" : "");
2225                 if (nr_process) {
2226                         if (nr_thread)
2227                                 log_info(" and ");
2228                         log_info("%d process%s", nr_process,
2229                                                 nr_process > 1 ? "es" : "");
2230                 }
2231                 log_info("\n");
2232                 log_info_flush();
2233         }
2234
2235         todo = thread_number;
2236         nr_running = 0;
2237         nr_started = 0;
2238         m_rate = t_rate = 0;
2239
2240         for_each_td(td, i) {
2241                 print_status_init(td->thread_number - 1);
2242
2243                 if (!td->o.create_serialize)
2244                         continue;
2245
2246                 if (fio_verify_load_state(td))
2247                         goto reap;
2248
2249                 /*
2250                  * do file setup here so it happens sequentially,
2251                  * we don't want X number of threads getting their
2252                  * client data interspersed on disk
2253                  */
2254                 if (setup_files(td)) {
2255 reap:
2256                         exit_value++;
2257                         if (td->error)
2258                                 log_err("fio: pid=%d, err=%d/%s\n",
2259                                         (int) td->pid, td->error, td->verror);
2260                         td_set_runstate(td, TD_REAPED);
2261                         todo--;
2262                 } else {
2263                         struct fio_file *f;
2264                         unsigned int j;
2265
2266                         /*
2267                          * for sharing to work, each job must always open
2268                          * its own files. so close them, if we opened them
2269                          * for creation
2270                          */
2271                         for_each_file(td, f, j) {
2272                                 if (fio_file_open(f))
2273                                         td_io_close_file(td, f);
2274                         }
2275                 }
2276         }
2277
2278         /* start idle threads before io threads start to run */
2279         fio_idle_prof_start();
2280
2281         set_genesis_time();
2282
2283         while (todo) {
2284                 struct thread_data *map[REAL_MAX_JOBS];
2285                 struct timespec this_start;
2286                 int this_jobs = 0, left;
2287                 struct fork_data *fd;
2288
2289                 /*
2290                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2291                  */
2292                 for_each_td(td, i) {
2293                         if (td->runstate != TD_NOT_CREATED)
2294                                 continue;
2295
2296                         /*
2297                          * never got a chance to start, killed by other
2298                          * thread for some reason
2299                          */
2300                         if (td->terminate) {
2301                                 todo--;
2302                                 continue;
2303                         }
2304
2305                         if (td->o.start_delay) {
2306                                 spent = utime_since_genesis();
2307
2308                                 if (td->o.start_delay > spent)
2309                                         continue;
2310                         }
2311
2312                         if (td->o.stonewall && (nr_started || nr_running)) {
2313                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
2314                                                         td->o.name);
2315                                 break;
2316                         }
2317
2318                         if (waitee_running(td)) {
2319                                 dprint(FD_PROCESS, "%s: waiting for %s\n",
2320                                                 td->o.name, td->o.wait_for);
2321                                 continue;
2322                         }
2323
2324                         init_disk_util(td);
2325
2326                         td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2327                         td->update_rusage = 0;
2328
2329                         /*
2330                          * Set state to created. Thread will transition
2331                          * to TD_INITIALIZED when it's done setting up.
2332                          */
2333                         td_set_runstate(td, TD_CREATED);
2334                         map[this_jobs++] = td;
2335                         nr_started++;
2336
2337                         fd = calloc(1, sizeof(*fd));
2338                         fd->td = td;
2339                         fd->sk_out = sk_out;
2340
2341                         if (td->o.use_thread) {
2342                                 int ret;
2343
2344                                 dprint(FD_PROCESS, "will pthread_create\n");
2345                                 ret = pthread_create(&td->thread, NULL,
2346                                                         thread_main, fd);
2347                                 if (ret) {
2348                                         log_err("pthread_create: %s\n",
2349                                                         strerror(ret));
2350                                         free(fd);
2351                                         nr_started--;
2352                                         break;
2353                                 }
2354                                 fd = NULL;
2355                                 ret = pthread_detach(td->thread);
2356                                 if (ret)
2357                                         log_err("pthread_detach: %s",
2358                                                         strerror(ret));
2359                         } else {
2360                                 pid_t pid;
2361                                 dprint(FD_PROCESS, "will fork\n");
2362                                 pid = fork();
2363                                 if (!pid) {
2364                                         int ret;
2365
2366                                         ret = (int)(uintptr_t)thread_main(fd);
2367                                         _exit(ret);
2368                                 } else if (i == fio_debug_jobno)
2369                                         *fio_debug_jobp = pid;
2370                         }
2371                         dprint(FD_MUTEX, "wait on startup_sem\n");
2372                         if (fio_sem_down_timeout(startup_sem, 10000)) {
2373                                 log_err("fio: job startup hung? exiting.\n");
2374                                 fio_terminate_threads(TERMINATE_ALL);
2375                                 fio_abort = 1;
2376                                 nr_started--;
2377                                 free(fd);
2378                                 break;
2379                         }
2380                         dprint(FD_MUTEX, "done waiting on startup_sem\n");
2381                 }
2382
2383                 /*
2384                  * Wait for the started threads to transition to
2385                  * TD_INITIALIZED.
2386                  */
2387                 fio_gettime(&this_start, NULL);
2388                 left = this_jobs;
2389                 while (left && !fio_abort) {
2390                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2391                                 break;
2392
2393                         do_usleep(100000);
2394
2395                         for (i = 0; i < this_jobs; i++) {
2396                                 td = map[i];
2397                                 if (!td)
2398                                         continue;
2399                                 if (td->runstate == TD_INITIALIZED) {
2400                                         map[i] = NULL;
2401                                         left--;
2402                                 } else if (td->runstate >= TD_EXITED) {
2403                                         map[i] = NULL;
2404                                         left--;
2405                                         todo--;
2406                                         nr_running++; /* work-around... */
2407                                 }
2408                         }
2409                 }
2410
2411                 if (left) {
2412                         log_err("fio: %d job%s failed to start\n", left,
2413                                         left > 1 ? "s" : "");
2414                         for (i = 0; i < this_jobs; i++) {
2415                                 td = map[i];
2416                                 if (!td)
2417                                         continue;
2418                                 kill(td->pid, SIGTERM);
2419                         }
2420                         break;
2421                 }
2422
2423                 /*
2424                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2425                  */
2426                 for_each_td(td, i) {
2427                         if (td->runstate != TD_INITIALIZED)
2428                                 continue;
2429
2430                         if (in_ramp_time(td))
2431                                 td_set_runstate(td, TD_RAMP);
2432                         else
2433                                 td_set_runstate(td, TD_RUNNING);
2434                         nr_running++;
2435                         nr_started--;
2436                         m_rate += ddir_rw_sum(td->o.ratemin);
2437                         t_rate += ddir_rw_sum(td->o.rate);
2438                         todo--;
2439                         fio_sem_up(td->sem);
2440                 }
2441
2442                 reap_threads(&nr_running, &t_rate, &m_rate);
2443
2444                 if (todo)
2445                         do_usleep(100000);
2446         }
2447
2448         while (nr_running) {
2449                 reap_threads(&nr_running, &t_rate, &m_rate);
2450                 do_usleep(10000);
2451         }
2452
2453         fio_idle_prof_stop();
2454
2455         update_io_ticks();
2456 }
2457
2458 static void free_disk_util(void)
2459 {
2460         disk_util_prune_entries();
2461         helper_thread_destroy();
2462 }
2463
2464 int fio_backend(struct sk_out *sk_out)
2465 {
2466         struct thread_data *td;
2467         int i;
2468
2469         if (exec_profile) {
2470                 if (load_profile(exec_profile))
2471                         return 1;
2472                 free(exec_profile);
2473                 exec_profile = NULL;
2474         }
2475         if (!thread_number)
2476                 return 0;
2477
2478         if (write_bw_log) {
2479                 struct log_params p = {
2480                         .log_type = IO_LOG_TYPE_BW,
2481                 };
2482
2483                 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2484                 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2485                 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2486         }
2487
2488         startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2489         if (startup_sem == NULL)
2490                 return 1;
2491
2492         set_genesis_time();
2493         stat_init();
2494         helper_thread_create(startup_sem, sk_out);
2495
2496         cgroup_list = smalloc(sizeof(*cgroup_list));
2497         if (cgroup_list)
2498                 INIT_FLIST_HEAD(cgroup_list);
2499
2500         run_threads(sk_out);
2501
2502         helper_thread_exit();
2503
2504         if (!fio_abort) {
2505                 __show_run_stats();
2506                 if (write_bw_log) {
2507                         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2508                                 struct io_log *log = agg_io_log[i];
2509
2510                                 flush_log(log, false);
2511                                 free_log(log);
2512                         }
2513                 }
2514         }
2515
2516         for_each_td(td, i) {
2517                 steadystate_free(td);
2518                 fio_options_free(td);
2519                 if (td->rusage_sem) {
2520                         fio_sem_remove(td->rusage_sem);
2521                         td->rusage_sem = NULL;
2522                 }
2523                 fio_sem_remove(td->sem);
2524                 td->sem = NULL;
2525         }
2526
2527         free_disk_util();
2528         if (cgroup_list) {
2529                 cgroup_kill(cgroup_list);
2530                 sfree(cgroup_list);
2531         }
2532
2533         fio_sem_remove(startup_sem);
2534         stat_exit();
2535         return exit_value;
2536 }