backend.c revision fdb0da8028e156c0da43aca18e1423d1b300bdad
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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA 22 * 23 */ 24#include <unistd.h> 25#include <fcntl.h> 26#include <string.h> 27#include <limits.h> 28#include <signal.h> 29#include <time.h> 30#include <locale.h> 31#include <assert.h> 32#include <time.h> 33#include <inttypes.h> 34#include <sys/stat.h> 35#include <sys/wait.h> 36#include <sys/ipc.h> 37#include <sys/mman.h> 38 39#include "fio.h" 40#ifndef FIO_NO_HAVE_SHM_H 41#include <sys/shm.h> 42#endif 43#include "hash.h" 44#include "smalloc.h" 45#include "verify.h" 46#include "trim.h" 47#include "diskutil.h" 48#include "cgroup.h" 49#include "profile.h" 50#include "lib/rand.h" 51#include "memalign.h" 52#include "server.h" 53#include "lib/getrusage.h" 54#include "idletime.h" 55#include "err.h" 56#include "lib/tp.h" 57 58static pthread_t disk_util_thread; 59static pthread_cond_t du_cond; 60static pthread_mutex_t du_lock; 61 62static struct fio_mutex *startup_mutex; 63static struct flist_head *cgroup_list; 64static char *cgroup_mnt; 65static int exit_value; 66static volatile int fio_abort; 67static unsigned int nr_process = 0; 68static unsigned int nr_thread = 0; 69 70struct io_log *agg_io_log[DDIR_RWDIR_CNT]; 71 72int groupid = 0; 73unsigned int thread_number = 0; 74unsigned int stat_number = 0; 75int shm_id = 0; 76int temp_stall_ts; 77unsigned long done_secs = 0; 78volatile int disk_util_exit = 0; 79 80#define PAGE_ALIGN(buf) \ 81 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask) 82 83#define JOB_START_TIMEOUT (5 * 1000) 84 85static void sig_int(int sig) 86{ 87 if (threads) { 88 if (is_backend) 89 fio_server_got_signal(sig); 90 else { 91 log_info("\nfio: terminating on signal %d\n", sig); 92 log_info_flush(); 93 exit_value = 128; 94 } 95 96 fio_terminate_threads(TERMINATE_ALL); 97 } 98} 99 100static void sig_show_status(int sig) 101{ 102 show_running_run_stats(); 103} 104 105static void set_sig_handlers(void) 106{ 107 struct sigaction act; 108 109 memset(&act, 0, sizeof(act)); 110 act.sa_handler = sig_int; 111 act.sa_flags = SA_RESTART; 112 sigaction(SIGINT, &act, NULL); 113 114 memset(&act, 0, sizeof(act)); 115 act.sa_handler = sig_int; 116 act.sa_flags = SA_RESTART; 117 sigaction(SIGTERM, &act, NULL); 118 119/* Windows uses SIGBREAK as a quit signal from other applications */ 120#ifdef WIN32 121 memset(&act, 0, sizeof(act)); 122 act.sa_handler = sig_int; 123 act.sa_flags = SA_RESTART; 124 sigaction(SIGBREAK, &act, NULL); 125#endif 126 127 memset(&act, 0, sizeof(act)); 128 act.sa_handler = sig_show_status; 129 act.sa_flags = SA_RESTART; 130 sigaction(SIGUSR1, &act, NULL); 131 132 if (is_backend) { 133 memset(&act, 0, sizeof(act)); 134 act.sa_handler = sig_int; 135 act.sa_flags = SA_RESTART; 136 sigaction(SIGPIPE, &act, NULL); 137 } 138} 139 140/* 141 * Check if we are above the minimum rate given. 142 */ 143static int __check_min_rate(struct thread_data *td, struct timeval *now, 144 enum fio_ddir ddir) 145{ 146 unsigned long long bytes = 0; 147 unsigned long iops = 0; 148 unsigned long spent; 149 unsigned long rate; 150 unsigned int ratemin = 0; 151 unsigned int rate_iops = 0; 152 unsigned int rate_iops_min = 0; 153 154 assert(ddir_rw(ddir)); 155 156 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir]) 157 return 0; 158 159 /* 160 * allow a 2 second settle period in the beginning 161 */ 162 if (mtime_since(&td->start, now) < 2000) 163 return 0; 164 165 iops += td->this_io_blocks[ddir]; 166 bytes += td->this_io_bytes[ddir]; 167 ratemin += td->o.ratemin[ddir]; 168 rate_iops += td->o.rate_iops[ddir]; 169 rate_iops_min += td->o.rate_iops_min[ddir]; 170 171 /* 172 * if rate blocks is set, sample is running 173 */ 174 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) { 175 spent = mtime_since(&td->lastrate[ddir], now); 176 if (spent < td->o.ratecycle) 177 return 0; 178 179 if (td->o.rate[ddir]) { 180 /* 181 * check bandwidth specified rate 182 */ 183 if (bytes < td->rate_bytes[ddir]) { 184 log_err("%s: min rate %u not met\n", td->o.name, 185 ratemin); 186 return 1; 187 } else { 188 if (spent) 189 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent; 190 else 191 rate = 0; 192 193 if (rate < ratemin || 194 bytes < td->rate_bytes[ddir]) { 195 log_err("%s: min rate %u not met, got" 196 " %luKB/sec\n", td->o.name, 197 ratemin, rate); 198 return 1; 199 } 200 } 201 } else { 202 /* 203 * checks iops specified rate 204 */ 205 if (iops < rate_iops) { 206 log_err("%s: min iops rate %u not met\n", 207 td->o.name, rate_iops); 208 return 1; 209 } else { 210 if (spent) 211 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent; 212 else 213 rate = 0; 214 215 if (rate < rate_iops_min || 216 iops < td->rate_blocks[ddir]) { 217 log_err("%s: min iops rate %u not met," 218 " got %lu\n", td->o.name, 219 rate_iops_min, rate); 220 } 221 } 222 } 223 } 224 225 td->rate_bytes[ddir] = bytes; 226 td->rate_blocks[ddir] = iops; 227 memcpy(&td->lastrate[ddir], now, sizeof(*now)); 228 return 0; 229} 230 231static int check_min_rate(struct thread_data *td, struct timeval *now, 232 uint64_t *bytes_done) 233{ 234 int ret = 0; 235 236 if (bytes_done[DDIR_READ]) 237 ret |= __check_min_rate(td, now, DDIR_READ); 238 if (bytes_done[DDIR_WRITE]) 239 ret |= __check_min_rate(td, now, DDIR_WRITE); 240 if (bytes_done[DDIR_TRIM]) 241 ret |= __check_min_rate(td, now, DDIR_TRIM); 242 243 return ret; 244} 245 246/* 247 * When job exits, we can cancel the in-flight IO if we are using async 248 * io. Attempt to do so. 249 */ 250static void cleanup_pending_aio(struct thread_data *td) 251{ 252 int r; 253 254 /* 255 * get immediately available events, if any 256 */ 257 r = io_u_queued_complete(td, 0, NULL); 258 if (r < 0) 259 return; 260 261 /* 262 * now cancel remaining active events 263 */ 264 if (td->io_ops->cancel) { 265 struct io_u *io_u; 266 int i; 267 268 io_u_qiter(&td->io_u_all, io_u, i) { 269 if (io_u->flags & IO_U_F_FLIGHT) { 270 r = td->io_ops->cancel(td, io_u); 271 if (!r) 272 put_io_u(td, io_u); 273 } 274 } 275 } 276 277 if (td->cur_depth) 278 r = io_u_queued_complete(td, td->cur_depth, NULL); 279} 280 281/* 282 * Helper to handle the final sync of a file. Works just like the normal 283 * io path, just does everything sync. 284 */ 285static int fio_io_sync(struct thread_data *td, struct fio_file *f) 286{ 287 struct io_u *io_u = __get_io_u(td); 288 int ret; 289 290 if (!io_u) 291 return 1; 292 293 io_u->ddir = DDIR_SYNC; 294 io_u->file = f; 295 296 if (td_io_prep(td, io_u)) { 297 put_io_u(td, io_u); 298 return 1; 299 } 300 301requeue: 302 ret = td_io_queue(td, io_u); 303 if (ret < 0) { 304 td_verror(td, io_u->error, "td_io_queue"); 305 put_io_u(td, io_u); 306 return 1; 307 } else if (ret == FIO_Q_QUEUED) { 308 if (io_u_queued_complete(td, 1, NULL) < 0) 309 return 1; 310 } else if (ret == FIO_Q_COMPLETED) { 311 if (io_u->error) { 312 td_verror(td, io_u->error, "td_io_queue"); 313 return 1; 314 } 315 316 if (io_u_sync_complete(td, io_u, NULL) < 0) 317 return 1; 318 } else if (ret == FIO_Q_BUSY) { 319 if (td_io_commit(td)) 320 return 1; 321 goto requeue; 322 } 323 324 return 0; 325} 326 327static int fio_file_fsync(struct thread_data *td, struct fio_file *f) 328{ 329 int ret; 330 331 if (fio_file_open(f)) 332 return fio_io_sync(td, f); 333 334 if (td_io_open_file(td, f)) 335 return 1; 336 337 ret = fio_io_sync(td, f); 338 td_io_close_file(td, f); 339 return ret; 340} 341 342static inline void __update_tv_cache(struct thread_data *td) 343{ 344 fio_gettime(&td->tv_cache, NULL); 345} 346 347static inline void update_tv_cache(struct thread_data *td) 348{ 349 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask) 350 __update_tv_cache(td); 351} 352 353static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) 354{ 355 if (in_ramp_time(td)) 356 return 0; 357 if (!td->o.timeout) 358 return 0; 359 if (utime_since(&td->epoch, t) >= td->o.timeout) 360 return 1; 361 362 return 0; 363} 364 365static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir, 366 int *retptr) 367{ 368 int ret = *retptr; 369 370 if (ret < 0 || td->error) { 371 int err = td->error; 372 enum error_type_bit eb; 373 374 if (ret < 0) 375 err = -ret; 376 377 eb = td_error_type(ddir, err); 378 if (!(td->o.continue_on_error & (1 << eb))) 379 return 1; 380 381 if (td_non_fatal_error(td, eb, err)) { 382 /* 383 * Continue with the I/Os in case of 384 * a non fatal error. 385 */ 386 update_error_count(td, err); 387 td_clear_error(td); 388 *retptr = 0; 389 return 0; 390 } else if (td->o.fill_device && err == ENOSPC) { 391 /* 392 * We expect to hit this error if 393 * fill_device option is set. 394 */ 395 td_clear_error(td); 396 fio_mark_td_terminate(td); 397 return 1; 398 } else { 399 /* 400 * Stop the I/O in case of a fatal 401 * error. 402 */ 403 update_error_count(td, err); 404 return 1; 405 } 406 } 407 408 return 0; 409} 410 411static void check_update_rusage(struct thread_data *td) 412{ 413 if (td->update_rusage) { 414 td->update_rusage = 0; 415 update_rusage_stat(td); 416 fio_mutex_up(td->rusage_sem); 417 } 418} 419 420/* 421 * The main verify engine. Runs over the writes we previously submitted, 422 * reads the blocks back in, and checks the crc/md5 of the data. 423 */ 424static void do_verify(struct thread_data *td, uint64_t verify_bytes) 425{ 426 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 }; 427 struct fio_file *f; 428 struct io_u *io_u; 429 int ret, min_events; 430 unsigned int i; 431 432 dprint(FD_VERIFY, "starting loop\n"); 433 434 /* 435 * sync io first and invalidate cache, to make sure we really 436 * read from disk. 437 */ 438 for_each_file(td, f, i) { 439 if (!fio_file_open(f)) 440 continue; 441 if (fio_io_sync(td, f)) 442 break; 443 if (file_invalidate_cache(td, f)) 444 break; 445 } 446 447 check_update_rusage(td); 448 449 if (td->error) 450 return; 451 452 td_set_runstate(td, TD_VERIFYING); 453 454 io_u = NULL; 455 while (!td->terminate) { 456 enum fio_ddir ddir; 457 int ret2, full; 458 459 update_tv_cache(td); 460 check_update_rusage(td); 461 462 if (runtime_exceeded(td, &td->tv_cache)) { 463 __update_tv_cache(td); 464 if (runtime_exceeded(td, &td->tv_cache)) { 465 fio_mark_td_terminate(td); 466 break; 467 } 468 } 469 470 if (flow_threshold_exceeded(td)) 471 continue; 472 473 if (!td->o.experimental_verify) { 474 io_u = __get_io_u(td); 475 if (!io_u) 476 break; 477 478 if (get_next_verify(td, io_u)) { 479 put_io_u(td, io_u); 480 break; 481 } 482 483 if (td_io_prep(td, io_u)) { 484 put_io_u(td, io_u); 485 break; 486 } 487 } else { 488 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes) 489 break; 490 491 while ((io_u = get_io_u(td)) != NULL) { 492 if (IS_ERR(io_u)) { 493 io_u = NULL; 494 ret = FIO_Q_BUSY; 495 goto reap; 496 } 497 498 /* 499 * We are only interested in the places where 500 * we wrote or trimmed IOs. Turn those into 501 * reads for verification purposes. 502 */ 503 if (io_u->ddir == DDIR_READ) { 504 /* 505 * Pretend we issued it for rwmix 506 * accounting 507 */ 508 td->io_issues[DDIR_READ]++; 509 put_io_u(td, io_u); 510 continue; 511 } else if (io_u->ddir == DDIR_TRIM) { 512 io_u->ddir = DDIR_READ; 513 io_u->flags |= IO_U_F_TRIMMED; 514 break; 515 } else if (io_u->ddir == DDIR_WRITE) { 516 io_u->ddir = DDIR_READ; 517 break; 518 } else { 519 put_io_u(td, io_u); 520 continue; 521 } 522 } 523 524 if (!io_u) 525 break; 526 } 527 528 if (td->o.verify_async) 529 io_u->end_io = verify_io_u_async; 530 else 531 io_u->end_io = verify_io_u; 532 533 ddir = io_u->ddir; 534 535 ret = td_io_queue(td, io_u); 536 switch (ret) { 537 case FIO_Q_COMPLETED: 538 if (io_u->error) { 539 ret = -io_u->error; 540 clear_io_u(td, io_u); 541 } else if (io_u->resid) { 542 int bytes = io_u->xfer_buflen - io_u->resid; 543 544 /* 545 * zero read, fail 546 */ 547 if (!bytes) { 548 td_verror(td, EIO, "full resid"); 549 put_io_u(td, io_u); 550 break; 551 } 552 553 io_u->xfer_buflen = io_u->resid; 554 io_u->xfer_buf += bytes; 555 io_u->offset += bytes; 556 557 if (ddir_rw(io_u->ddir)) 558 td->ts.short_io_u[io_u->ddir]++; 559 560 f = io_u->file; 561 if (io_u->offset == f->real_file_size) 562 goto sync_done; 563 564 requeue_io_u(td, &io_u); 565 } else { 566sync_done: 567 ret = io_u_sync_complete(td, io_u, bytes_done); 568 if (ret < 0) 569 break; 570 } 571 continue; 572 case FIO_Q_QUEUED: 573 break; 574 case FIO_Q_BUSY: 575 requeue_io_u(td, &io_u); 576 ret2 = td_io_commit(td); 577 if (ret2 < 0) 578 ret = ret2; 579 break; 580 default: 581 assert(ret < 0); 582 td_verror(td, -ret, "td_io_queue"); 583 break; 584 } 585 586 if (break_on_this_error(td, ddir, &ret)) 587 break; 588 589 /* 590 * if we can queue more, do so. but check if there are 591 * completed io_u's first. Note that we can get BUSY even 592 * without IO queued, if the system is resource starved. 593 */ 594reap: 595 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); 596 if (full || !td->o.iodepth_batch_complete) { 597 min_events = min(td->o.iodepth_batch_complete, 598 td->cur_depth); 599 /* 600 * if the queue is full, we MUST reap at least 1 event 601 */ 602 if (full && !min_events) 603 min_events = 1; 604 605 do { 606 /* 607 * Reap required number of io units, if any, 608 * and do the verification on them through 609 * the callback handler 610 */ 611 if (io_u_queued_complete(td, min_events, bytes_done) < 0) { 612 ret = -1; 613 break; 614 } 615 } while (full && (td->cur_depth > td->o.iodepth_low)); 616 } 617 if (ret < 0) 618 break; 619 } 620 621 check_update_rusage(td); 622 623 if (!td->error) { 624 min_events = td->cur_depth; 625 626 if (min_events) 627 ret = io_u_queued_complete(td, min_events, NULL); 628 } else 629 cleanup_pending_aio(td); 630 631 td_set_runstate(td, TD_RUNNING); 632 633 dprint(FD_VERIFY, "exiting loop\n"); 634} 635 636static unsigned int exceeds_number_ios(struct thread_data *td) 637{ 638 unsigned long long number_ios; 639 640 if (!td->o.number_ios) 641 return 0; 642 643 number_ios = ddir_rw_sum(td->this_io_blocks); 644 number_ios += td->io_u_queued + td->io_u_in_flight; 645 646 return number_ios >= td->o.number_ios; 647} 648 649static int io_bytes_exceeded(struct thread_data *td) 650{ 651 unsigned long long bytes, limit; 652 653 if (td_rw(td)) 654 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE]; 655 else if (td_write(td)) 656 bytes = td->this_io_bytes[DDIR_WRITE]; 657 else if (td_read(td)) 658 bytes = td->this_io_bytes[DDIR_READ]; 659 else 660 bytes = td->this_io_bytes[DDIR_TRIM]; 661 662 if (td->o.io_limit) 663 limit = td->o.io_limit; 664 else 665 limit = td->o.size; 666 667 return bytes >= limit || exceeds_number_ios(td); 668} 669 670/* 671 * Main IO worker function. It retrieves io_u's to process and queues 672 * and reaps them, checking for rate and errors along the way. 673 * 674 * Returns number of bytes written and trimmed. 675 */ 676static uint64_t do_io(struct thread_data *td) 677{ 678 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 }; 679 unsigned int i; 680 int ret = 0; 681 uint64_t total_bytes, bytes_issued = 0; 682 683 if (in_ramp_time(td)) 684 td_set_runstate(td, TD_RAMP); 685 else 686 td_set_runstate(td, TD_RUNNING); 687 688 lat_target_init(td); 689 690 /* 691 * If verify_backlog is enabled, we'll run the verify in this 692 * handler as well. For that case, we may need up to twice the 693 * amount of bytes. 694 */ 695 total_bytes = td->o.size; 696 if (td->o.verify != VERIFY_NONE && 697 (td_write(td) && td->o.verify_backlog)) 698 total_bytes += td->o.size; 699 700 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || 701 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) || 702 td->o.time_based) { 703 struct timeval comp_time; 704 int min_evts = 0; 705 struct io_u *io_u; 706 int ret2, full; 707 enum fio_ddir ddir; 708 709 check_update_rusage(td); 710 711 if (td->terminate || td->done) 712 break; 713 714 update_tv_cache(td); 715 716 if (runtime_exceeded(td, &td->tv_cache)) { 717 __update_tv_cache(td); 718 if (runtime_exceeded(td, &td->tv_cache)) { 719 fio_mark_td_terminate(td); 720 break; 721 } 722 } 723 724 if (flow_threshold_exceeded(td)) 725 continue; 726 727 if (bytes_issued >= total_bytes) 728 break; 729 730 io_u = get_io_u(td); 731 if (IS_ERR_OR_NULL(io_u)) { 732 int err = PTR_ERR(io_u); 733 734 io_u = NULL; 735 if (err == -EBUSY) { 736 ret = FIO_Q_BUSY; 737 goto reap; 738 } 739 if (td->o.latency_target) 740 goto reap; 741 break; 742 } 743 744 ddir = io_u->ddir; 745 746 /* 747 * Add verification end_io handler if: 748 * - Asked to verify (!td_rw(td)) 749 * - Or the io_u is from our verify list (mixed write/ver) 750 */ 751 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ && 752 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) { 753 754 if (!td->o.verify_pattern_bytes) { 755 io_u->rand_seed = __rand(&td->__verify_state); 756 if (sizeof(int) != sizeof(long *)) 757 io_u->rand_seed *= __rand(&td->__verify_state); 758 } 759 760 if (td->o.verify_async) 761 io_u->end_io = verify_io_u_async; 762 else 763 io_u->end_io = verify_io_u; 764 td_set_runstate(td, TD_VERIFYING); 765 } else if (in_ramp_time(td)) 766 td_set_runstate(td, TD_RAMP); 767 else 768 td_set_runstate(td, TD_RUNNING); 769 770 /* 771 * Always log IO before it's issued, so we know the specific 772 * order of it. The logged unit will track when the IO has 773 * completed. 774 */ 775 if (td_write(td) && io_u->ddir == DDIR_WRITE && 776 td->o.do_verify && 777 td->o.verify != VERIFY_NONE && 778 !td->o.experimental_verify) 779 log_io_piece(td, io_u); 780 781 ret = td_io_queue(td, io_u); 782 switch (ret) { 783 case FIO_Q_COMPLETED: 784 if (io_u->error) { 785 ret = -io_u->error; 786 unlog_io_piece(td, io_u); 787 clear_io_u(td, io_u); 788 } else if (io_u->resid) { 789 int bytes = io_u->xfer_buflen - io_u->resid; 790 struct fio_file *f = io_u->file; 791 792 bytes_issued += bytes; 793 794 trim_io_piece(td, io_u); 795 796 /* 797 * zero read, fail 798 */ 799 if (!bytes) { 800 unlog_io_piece(td, io_u); 801 td_verror(td, EIO, "full resid"); 802 put_io_u(td, io_u); 803 break; 804 } 805 806 io_u->xfer_buflen = io_u->resid; 807 io_u->xfer_buf += bytes; 808 io_u->offset += bytes; 809 810 if (ddir_rw(io_u->ddir)) 811 td->ts.short_io_u[io_u->ddir]++; 812 813 if (io_u->offset == f->real_file_size) 814 goto sync_done; 815 816 requeue_io_u(td, &io_u); 817 } else { 818sync_done: 819 if (__should_check_rate(td, DDIR_READ) || 820 __should_check_rate(td, DDIR_WRITE) || 821 __should_check_rate(td, DDIR_TRIM)) 822 fio_gettime(&comp_time, NULL); 823 824 ret = io_u_sync_complete(td, io_u, bytes_done); 825 if (ret < 0) 826 break; 827 bytes_issued += io_u->xfer_buflen; 828 } 829 break; 830 case FIO_Q_QUEUED: 831 /* 832 * if the engine doesn't have a commit hook, 833 * the io_u is really queued. if it does have such 834 * a hook, it has to call io_u_queued() itself. 835 */ 836 if (td->io_ops->commit == NULL) 837 io_u_queued(td, io_u); 838 bytes_issued += io_u->xfer_buflen; 839 break; 840 case FIO_Q_BUSY: 841 unlog_io_piece(td, io_u); 842 requeue_io_u(td, &io_u); 843 ret2 = td_io_commit(td); 844 if (ret2 < 0) 845 ret = ret2; 846 break; 847 default: 848 assert(ret < 0); 849 put_io_u(td, io_u); 850 break; 851 } 852 853 if (break_on_this_error(td, ddir, &ret)) 854 break; 855 856 /* 857 * See if we need to complete some commands. Note that we 858 * can get BUSY even without IO queued, if the system is 859 * resource starved. 860 */ 861reap: 862 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); 863 if (full || !td->o.iodepth_batch_complete) { 864 min_evts = min(td->o.iodepth_batch_complete, 865 td->cur_depth); 866 /* 867 * if the queue is full, we MUST reap at least 1 event 868 */ 869 if (full && !min_evts) 870 min_evts = 1; 871 872 if (__should_check_rate(td, DDIR_READ) || 873 __should_check_rate(td, DDIR_WRITE) || 874 __should_check_rate(td, DDIR_TRIM)) 875 fio_gettime(&comp_time, NULL); 876 877 do { 878 ret = io_u_queued_complete(td, min_evts, bytes_done); 879 if (ret < 0) 880 break; 881 882 } while (full && (td->cur_depth > td->o.iodepth_low)); 883 } 884 885 if (ret < 0) 886 break; 887 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO)) 888 continue; 889 890 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) { 891 if (check_min_rate(td, &comp_time, bytes_done)) { 892 if (exitall_on_terminate) 893 fio_terminate_threads(td->groupid); 894 td_verror(td, EIO, "check_min_rate"); 895 break; 896 } 897 } 898 if (!in_ramp_time(td) && td->o.latency_target) 899 lat_target_check(td); 900 901 if (td->o.thinktime) { 902 unsigned long long b; 903 904 b = ddir_rw_sum(td->io_blocks); 905 if (!(b % td->o.thinktime_blocks)) { 906 int left; 907 908 io_u_quiesce(td); 909 910 if (td->o.thinktime_spin) 911 usec_spin(td->o.thinktime_spin); 912 913 left = td->o.thinktime - td->o.thinktime_spin; 914 if (left) 915 usec_sleep(td, left); 916 } 917 } 918 } 919 920 check_update_rusage(td); 921 922 if (td->trim_entries) 923 log_err("fio: %lu trim entries leaked?\n", td->trim_entries); 924 925 if (td->o.fill_device && td->error == ENOSPC) { 926 td->error = 0; 927 fio_mark_td_terminate(td); 928 } 929 if (!td->error) { 930 struct fio_file *f; 931 932 i = td->cur_depth; 933 if (i) { 934 ret = io_u_queued_complete(td, i, bytes_done); 935 if (td->o.fill_device && td->error == ENOSPC) 936 td->error = 0; 937 } 938 939 if (should_fsync(td) && td->o.end_fsync) { 940 td_set_runstate(td, TD_FSYNCING); 941 942 for_each_file(td, f, i) { 943 if (!fio_file_fsync(td, f)) 944 continue; 945 946 log_err("fio: end_fsync failed for file %s\n", 947 f->file_name); 948 } 949 } 950 } else 951 cleanup_pending_aio(td); 952 953 /* 954 * stop job if we failed doing any IO 955 */ 956 if (!ddir_rw_sum(td->this_io_bytes)) 957 td->done = 1; 958 959 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM]; 960} 961 962static void cleanup_io_u(struct thread_data *td) 963{ 964 struct io_u *io_u; 965 966 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) { 967 968 if (td->io_ops->io_u_free) 969 td->io_ops->io_u_free(td, io_u); 970 971 fio_memfree(io_u, sizeof(*io_u)); 972 } 973 974 free_io_mem(td); 975 976 io_u_rexit(&td->io_u_requeues); 977 io_u_qexit(&td->io_u_freelist); 978 io_u_qexit(&td->io_u_all); 979} 980 981static int init_io_u(struct thread_data *td) 982{ 983 struct io_u *io_u; 984 unsigned int max_bs, min_write; 985 int cl_align, i, max_units; 986 int data_xfer = 1, err; 987 char *p; 988 989 max_units = td->o.iodepth; 990 max_bs = td_max_bs(td); 991 min_write = td->o.min_bs[DDIR_WRITE]; 992 td->orig_buffer_size = (unsigned long long) max_bs 993 * (unsigned long long) max_units; 994 995 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td))) 996 data_xfer = 0; 997 998 err = 0; 999 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth); 1000 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth); 1001 err += io_u_qinit(&td->io_u_all, td->o.iodepth); 1002 1003 if (err) { 1004 log_err("fio: failed setting up IO queues\n"); 1005 return 1; 1006 } 1007 1008 /* 1009 * if we may later need to do address alignment, then add any 1010 * possible adjustment here so that we don't cause a buffer 1011 * overflow later. this adjustment may be too much if we get 1012 * lucky and the allocator gives us an aligned address. 1013 */ 1014 if (td->o.odirect || td->o.mem_align || td->o.oatomic || 1015 (td->io_ops->flags & FIO_RAWIO)) 1016 td->orig_buffer_size += page_mask + td->o.mem_align; 1017 1018 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) { 1019 unsigned long bs; 1020 1021 bs = td->orig_buffer_size + td->o.hugepage_size - 1; 1022 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1); 1023 } 1024 1025 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) { 1026 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n"); 1027 return 1; 1028 } 1029 1030 if (data_xfer && allocate_io_mem(td)) 1031 return 1; 1032 1033 if (td->o.odirect || td->o.mem_align || td->o.oatomic || 1034 (td->io_ops->flags & FIO_RAWIO)) 1035 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align; 1036 else 1037 p = td->orig_buffer; 1038 1039 cl_align = os_cache_line_size(); 1040 1041 for (i = 0; i < max_units; i++) { 1042 void *ptr; 1043 1044 if (td->terminate) 1045 return 1; 1046 1047 ptr = fio_memalign(cl_align, sizeof(*io_u)); 1048 if (!ptr) { 1049 log_err("fio: unable to allocate aligned memory\n"); 1050 break; 1051 } 1052 1053 io_u = ptr; 1054 memset(io_u, 0, sizeof(*io_u)); 1055 INIT_FLIST_HEAD(&io_u->verify_list); 1056 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i); 1057 1058 if (data_xfer) { 1059 io_u->buf = p; 1060 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf); 1061 1062 if (td_write(td)) 1063 io_u_fill_buffer(td, io_u, min_write, max_bs); 1064 if (td_write(td) && td->o.verify_pattern_bytes) { 1065 /* 1066 * Fill the buffer with the pattern if we are 1067 * going to be doing writes. 1068 */ 1069 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0); 1070 } 1071 } 1072 1073 io_u->index = i; 1074 io_u->flags = IO_U_F_FREE; 1075 io_u_qpush(&td->io_u_freelist, io_u); 1076 1077 /* 1078 * io_u never leaves this stack, used for iteration of all 1079 * io_u buffers. 1080 */ 1081 io_u_qpush(&td->io_u_all, io_u); 1082 1083 if (td->io_ops->io_u_init) { 1084 int ret = td->io_ops->io_u_init(td, io_u); 1085 1086 if (ret) { 1087 log_err("fio: failed to init engine data: %d\n", ret); 1088 return 1; 1089 } 1090 } 1091 1092 p += max_bs; 1093 } 1094 1095 return 0; 1096} 1097 1098static int switch_ioscheduler(struct thread_data *td) 1099{ 1100 char tmp[256], tmp2[128]; 1101 FILE *f; 1102 int ret; 1103 1104 if (td->io_ops->flags & FIO_DISKLESSIO) 1105 return 0; 1106 1107 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root); 1108 1109 f = fopen(tmp, "r+"); 1110 if (!f) { 1111 if (errno == ENOENT) { 1112 log_err("fio: os or kernel doesn't support IO scheduler" 1113 " switching\n"); 1114 return 0; 1115 } 1116 td_verror(td, errno, "fopen iosched"); 1117 return 1; 1118 } 1119 1120 /* 1121 * Set io scheduler. 1122 */ 1123 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f); 1124 if (ferror(f) || ret != 1) { 1125 td_verror(td, errno, "fwrite"); 1126 fclose(f); 1127 return 1; 1128 } 1129 1130 rewind(f); 1131 1132 /* 1133 * Read back and check that the selected scheduler is now the default. 1134 */ 1135 ret = fread(tmp, sizeof(tmp), 1, f); 1136 if (ferror(f) || ret < 0) { 1137 td_verror(td, errno, "fread"); 1138 fclose(f); 1139 return 1; 1140 } 1141 tmp[sizeof(tmp) - 1] = '\0'; 1142 1143 1144 sprintf(tmp2, "[%s]", td->o.ioscheduler); 1145 if (!strstr(tmp, tmp2)) { 1146 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler); 1147 td_verror(td, EINVAL, "iosched_switch"); 1148 fclose(f); 1149 return 1; 1150 } 1151 1152 fclose(f); 1153 return 0; 1154} 1155 1156static int keep_running(struct thread_data *td) 1157{ 1158 unsigned long long limit; 1159 1160 if (td->done) 1161 return 0; 1162 if (td->o.time_based) 1163 return 1; 1164 if (td->o.loops) { 1165 td->o.loops--; 1166 return 1; 1167 } 1168 if (exceeds_number_ios(td)) 1169 return 0; 1170 1171 if (td->o.io_limit) 1172 limit = td->o.io_limit; 1173 else 1174 limit = td->o.size; 1175 1176 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) { 1177 uint64_t diff; 1178 1179 /* 1180 * If the difference is less than the minimum IO size, we 1181 * are done. 1182 */ 1183 diff = limit - ddir_rw_sum(td->io_bytes); 1184 if (diff < td_max_bs(td)) 1185 return 0; 1186 1187 if (fio_files_done(td)) 1188 return 0; 1189 1190 return 1; 1191 } 1192 1193 return 0; 1194} 1195 1196static int exec_string(struct thread_options *o, const char *string, const char *mode) 1197{ 1198 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1; 1199 char *str; 1200 1201 str = malloc(newlen); 1202 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode); 1203 1204 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode); 1205 ret = system(str); 1206 if (ret == -1) 1207 log_err("fio: exec of cmd <%s> failed\n", str); 1208 1209 free(str); 1210 return ret; 1211} 1212 1213/* 1214 * Dry run to compute correct state of numberio for verification. 1215 */ 1216static uint64_t do_dry_run(struct thread_data *td) 1217{ 1218 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 }; 1219 1220 td_set_runstate(td, TD_RUNNING); 1221 1222 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || 1223 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) { 1224 struct io_u *io_u; 1225 int ret; 1226 1227 if (td->terminate || td->done) 1228 break; 1229 1230 io_u = get_io_u(td); 1231 if (!io_u) 1232 break; 1233 1234 io_u->flags |= IO_U_F_FLIGHT; 1235 io_u->error = 0; 1236 io_u->resid = 0; 1237 if (ddir_rw(acct_ddir(io_u))) 1238 td->io_issues[acct_ddir(io_u)]++; 1239 if (ddir_rw(io_u->ddir)) { 1240 io_u_mark_depth(td, 1); 1241 td->ts.total_io_u[io_u->ddir]++; 1242 } 1243 1244 if (td_write(td) && io_u->ddir == DDIR_WRITE && 1245 td->o.do_verify && 1246 td->o.verify != VERIFY_NONE && 1247 !td->o.experimental_verify) 1248 log_io_piece(td, io_u); 1249 1250 ret = io_u_sync_complete(td, io_u, bytes_done); 1251 (void) ret; 1252 } 1253 1254 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM]; 1255} 1256 1257/* 1258 * Entry point for the thread based jobs. The process based jobs end up 1259 * here as well, after a little setup. 1260 */ 1261static void *thread_main(void *data) 1262{ 1263 unsigned long long elapsed; 1264 struct thread_data *td = data; 1265 struct thread_options *o = &td->o; 1266 pthread_condattr_t attr; 1267 int clear_state; 1268 int ret; 1269 1270 if (!o->use_thread) { 1271 setsid(); 1272 td->pid = getpid(); 1273 } else 1274 td->pid = gettid(); 1275 1276 fio_local_clock_init(o->use_thread); 1277 1278 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid); 1279 1280 if (is_backend) 1281 fio_server_send_start(td); 1282 1283 INIT_FLIST_HEAD(&td->io_log_list); 1284 INIT_FLIST_HEAD(&td->io_hist_list); 1285 INIT_FLIST_HEAD(&td->verify_list); 1286 INIT_FLIST_HEAD(&td->trim_list); 1287 INIT_FLIST_HEAD(&td->next_rand_list); 1288 pthread_mutex_init(&td->io_u_lock, NULL); 1289 td->io_hist_tree = RB_ROOT; 1290 1291 pthread_condattr_init(&attr); 1292 pthread_cond_init(&td->verify_cond, &attr); 1293 pthread_cond_init(&td->free_cond, &attr); 1294 1295 td_set_runstate(td, TD_INITIALIZED); 1296 dprint(FD_MUTEX, "up startup_mutex\n"); 1297 fio_mutex_up(startup_mutex); 1298 dprint(FD_MUTEX, "wait on td->mutex\n"); 1299 fio_mutex_down(td->mutex); 1300 dprint(FD_MUTEX, "done waiting on td->mutex\n"); 1301 1302 /* 1303 * A new gid requires privilege, so we need to do this before setting 1304 * the uid. 1305 */ 1306 if (o->gid != -1U && setgid(o->gid)) { 1307 td_verror(td, errno, "setgid"); 1308 goto err; 1309 } 1310 if (o->uid != -1U && setuid(o->uid)) { 1311 td_verror(td, errno, "setuid"); 1312 goto err; 1313 } 1314 1315 /* 1316 * If we have a gettimeofday() thread, make sure we exclude that 1317 * thread from this job 1318 */ 1319 if (o->gtod_cpu) 1320 fio_cpu_clear(&o->cpumask, o->gtod_cpu); 1321 1322 /* 1323 * Set affinity first, in case it has an impact on the memory 1324 * allocations. 1325 */ 1326 if (o->cpumask_set) { 1327 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) { 1328 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1); 1329 if (!ret) { 1330 log_err("fio: no CPUs set\n"); 1331 log_err("fio: Try increasing number of available CPUs\n"); 1332 td_verror(td, EINVAL, "cpus_split"); 1333 goto err; 1334 } 1335 } 1336 ret = fio_setaffinity(td->pid, o->cpumask); 1337 if (ret == -1) { 1338 td_verror(td, errno, "cpu_set_affinity"); 1339 goto err; 1340 } 1341 } 1342 1343#ifdef CONFIG_LIBNUMA 1344 /* numa node setup */ 1345 if (o->numa_cpumask_set || o->numa_memmask_set) { 1346 struct bitmask *mask; 1347 int ret; 1348 1349 if (numa_available() < 0) { 1350 td_verror(td, errno, "Does not support NUMA API\n"); 1351 goto err; 1352 } 1353 1354 if (o->numa_cpumask_set) { 1355 mask = numa_parse_nodestring(o->numa_cpunodes); 1356 ret = numa_run_on_node_mask(mask); 1357 numa_free_nodemask(mask); 1358 if (ret == -1) { 1359 td_verror(td, errno, \ 1360 "numa_run_on_node_mask failed\n"); 1361 goto err; 1362 } 1363 } 1364 1365 if (o->numa_memmask_set) { 1366 1367 mask = NULL; 1368 if (o->numa_memnodes) 1369 mask = numa_parse_nodestring(o->numa_memnodes); 1370 1371 switch (o->numa_mem_mode) { 1372 case MPOL_INTERLEAVE: 1373 numa_set_interleave_mask(mask); 1374 break; 1375 case MPOL_BIND: 1376 numa_set_membind(mask); 1377 break; 1378 case MPOL_LOCAL: 1379 numa_set_localalloc(); 1380 break; 1381 case MPOL_PREFERRED: 1382 numa_set_preferred(o->numa_mem_prefer_node); 1383 break; 1384 case MPOL_DEFAULT: 1385 default: 1386 break; 1387 } 1388 1389 if (mask) 1390 numa_free_nodemask(mask); 1391 1392 } 1393 } 1394#endif 1395 1396 if (fio_pin_memory(td)) 1397 goto err; 1398 1399 /* 1400 * May alter parameters that init_io_u() will use, so we need to 1401 * do this first. 1402 */ 1403 if (init_iolog(td)) 1404 goto err; 1405 1406 if (init_io_u(td)) 1407 goto err; 1408 1409 if (o->verify_async && verify_async_init(td)) 1410 goto err; 1411 1412 if (o->ioprio) { 1413 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio); 1414 if (ret == -1) { 1415 td_verror(td, errno, "ioprio_set"); 1416 goto err; 1417 } 1418 } 1419 1420 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt)) 1421 goto err; 1422 1423 errno = 0; 1424 if (nice(o->nice) == -1 && errno != 0) { 1425 td_verror(td, errno, "nice"); 1426 goto err; 1427 } 1428 1429 if (o->ioscheduler && switch_ioscheduler(td)) 1430 goto err; 1431 1432 if (!o->create_serialize && setup_files(td)) 1433 goto err; 1434 1435 if (td_io_init(td)) 1436 goto err; 1437 1438 if (init_random_map(td)) 1439 goto err; 1440 1441 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun")) 1442 goto err; 1443 1444 if (o->pre_read) { 1445 if (pre_read_files(td) < 0) 1446 goto err; 1447 } 1448 1449 if (td->flags & TD_F_COMPRESS_LOG) 1450 tp_init(&td->tp_data); 1451 1452 fio_verify_init(td); 1453 1454 fio_gettime(&td->epoch, NULL); 1455 fio_getrusage(&td->ru_start); 1456 clear_state = 0; 1457 while (keep_running(td)) { 1458 uint64_t verify_bytes; 1459 1460 fio_gettime(&td->start, NULL); 1461 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start)); 1462 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start)); 1463 memcpy(&td->tv_cache, &td->start, sizeof(td->start)); 1464 1465 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] || 1466 o->ratemin[DDIR_TRIM]) { 1467 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time, 1468 sizeof(td->bw_sample_time)); 1469 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time, 1470 sizeof(td->bw_sample_time)); 1471 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time, 1472 sizeof(td->bw_sample_time)); 1473 } 1474 1475 if (clear_state) 1476 clear_io_state(td); 1477 1478 prune_io_piece_log(td); 1479 1480 if (td->o.verify_only && (td_write(td) || td_rw(td))) 1481 verify_bytes = do_dry_run(td); 1482 else 1483 verify_bytes = do_io(td); 1484 1485 clear_state = 1; 1486 1487 if (td_read(td) && td->io_bytes[DDIR_READ]) { 1488 elapsed = utime_since_now(&td->start); 1489 td->ts.runtime[DDIR_READ] += elapsed; 1490 } 1491 if (td_write(td) && td->io_bytes[DDIR_WRITE]) { 1492 elapsed = utime_since_now(&td->start); 1493 td->ts.runtime[DDIR_WRITE] += elapsed; 1494 } 1495 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) { 1496 elapsed = utime_since_now(&td->start); 1497 td->ts.runtime[DDIR_TRIM] += elapsed; 1498 } 1499 1500 if (td->error || td->terminate) 1501 break; 1502 1503 if (!o->do_verify || 1504 o->verify == VERIFY_NONE || 1505 (td->io_ops->flags & FIO_UNIDIR)) 1506 continue; 1507 1508 clear_io_state(td); 1509 1510 fio_gettime(&td->start, NULL); 1511 1512 do_verify(td, verify_bytes); 1513 1514 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start); 1515 1516 if (td->error || td->terminate) 1517 break; 1518 } 1519 1520 update_rusage_stat(td); 1521 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000; 1522 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000; 1523 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000; 1524 td->ts.total_run_time = mtime_since_now(&td->epoch); 1525 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; 1526 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; 1527 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; 1528 1529 fio_unpin_memory(td); 1530 1531 fio_writeout_logs(td); 1532 1533 if (td->flags & TD_F_COMPRESS_LOG) 1534 tp_exit(&td->tp_data); 1535 1536 if (o->exec_postrun) 1537 exec_string(o, o->exec_postrun, (const char *)"postrun"); 1538 1539 if (exitall_on_terminate) 1540 fio_terminate_threads(td->groupid); 1541 1542err: 1543 if (td->error) 1544 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error, 1545 td->verror); 1546 1547 if (o->verify_async) 1548 verify_async_exit(td); 1549 1550 close_and_free_files(td); 1551 cleanup_io_u(td); 1552 close_ioengine(td); 1553 cgroup_shutdown(td, &cgroup_mnt); 1554 1555 if (o->cpumask_set) { 1556 int ret = fio_cpuset_exit(&o->cpumask); 1557 1558 td_verror(td, ret, "fio_cpuset_exit"); 1559 } 1560 1561 /* 1562 * do this very late, it will log file closing as well 1563 */ 1564 if (o->write_iolog_file) 1565 write_iolog_close(td); 1566 1567 fio_mutex_remove(td->rusage_sem); 1568 td->rusage_sem = NULL; 1569 1570 fio_mutex_remove(td->mutex); 1571 td->mutex = NULL; 1572 1573 td_set_runstate(td, TD_EXITED); 1574 return (void *) (uintptr_t) td->error; 1575} 1576 1577 1578/* 1579 * We cannot pass the td data into a forked process, so attach the td and 1580 * pass it to the thread worker. 1581 */ 1582static int fork_main(int shmid, int offset) 1583{ 1584 struct thread_data *td; 1585 void *data, *ret; 1586 1587#if !defined(__hpux) && !defined(CONFIG_NO_SHM) 1588 data = shmat(shmid, NULL, 0); 1589 if (data == (void *) -1) { 1590 int __err = errno; 1591 1592 perror("shmat"); 1593 return __err; 1594 } 1595#else 1596 /* 1597 * HP-UX inherits shm mappings? 1598 */ 1599 data = threads; 1600#endif 1601 1602 td = data + offset * sizeof(struct thread_data); 1603 ret = thread_main(td); 1604 shmdt(data); 1605 return (int) (uintptr_t) ret; 1606} 1607 1608static void dump_td_info(struct thread_data *td) 1609{ 1610 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to " 1611 "be stuck. Doing forceful exit of this job.\n", td->o.name, 1612 (unsigned long) time_since_now(&td->terminate_time)); 1613} 1614 1615/* 1616 * Run over the job map and reap the threads that have exited, if any. 1617 */ 1618static void reap_threads(unsigned int *nr_running, unsigned int *t_rate, 1619 unsigned int *m_rate) 1620{ 1621 struct thread_data *td; 1622 unsigned int cputhreads, realthreads, pending; 1623 int i, status, ret; 1624 1625 /* 1626 * reap exited threads (TD_EXITED -> TD_REAPED) 1627 */ 1628 realthreads = pending = cputhreads = 0; 1629 for_each_td(td, i) { 1630 int flags = 0; 1631 1632 /* 1633 * ->io_ops is NULL for a thread that has closed its 1634 * io engine 1635 */ 1636 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio")) 1637 cputhreads++; 1638 else 1639 realthreads++; 1640 1641 if (!td->pid) { 1642 pending++; 1643 continue; 1644 } 1645 if (td->runstate == TD_REAPED) 1646 continue; 1647 if (td->o.use_thread) { 1648 if (td->runstate == TD_EXITED) { 1649 td_set_runstate(td, TD_REAPED); 1650 goto reaped; 1651 } 1652 continue; 1653 } 1654 1655 flags = WNOHANG; 1656 if (td->runstate == TD_EXITED) 1657 flags = 0; 1658 1659 /* 1660 * check if someone quit or got killed in an unusual way 1661 */ 1662 ret = waitpid(td->pid, &status, flags); 1663 if (ret < 0) { 1664 if (errno == ECHILD) { 1665 log_err("fio: pid=%d disappeared %d\n", 1666 (int) td->pid, td->runstate); 1667 td->sig = ECHILD; 1668 td_set_runstate(td, TD_REAPED); 1669 goto reaped; 1670 } 1671 perror("waitpid"); 1672 } else if (ret == td->pid) { 1673 if (WIFSIGNALED(status)) { 1674 int sig = WTERMSIG(status); 1675 1676 if (sig != SIGTERM && sig != SIGUSR2) 1677 log_err("fio: pid=%d, got signal=%d\n", 1678 (int) td->pid, sig); 1679 td->sig = sig; 1680 td_set_runstate(td, TD_REAPED); 1681 goto reaped; 1682 } 1683 if (WIFEXITED(status)) { 1684 if (WEXITSTATUS(status) && !td->error) 1685 td->error = WEXITSTATUS(status); 1686 1687 td_set_runstate(td, TD_REAPED); 1688 goto reaped; 1689 } 1690 } 1691 1692 /* 1693 * If the job is stuck, do a forceful timeout of it and 1694 * move on. 1695 */ 1696 if (td->terminate && 1697 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) { 1698 dump_td_info(td); 1699 td_set_runstate(td, TD_REAPED); 1700 goto reaped; 1701 } 1702 1703 /* 1704 * thread is not dead, continue 1705 */ 1706 pending++; 1707 continue; 1708reaped: 1709 (*nr_running)--; 1710 (*m_rate) -= ddir_rw_sum(td->o.ratemin); 1711 (*t_rate) -= ddir_rw_sum(td->o.rate); 1712 if (!td->pid) 1713 pending--; 1714 1715 if (td->error) 1716 exit_value++; 1717 1718 done_secs += mtime_since_now(&td->epoch) / 1000; 1719 profile_td_exit(td); 1720 } 1721 1722 if (*nr_running == cputhreads && !pending && realthreads) 1723 fio_terminate_threads(TERMINATE_ALL); 1724} 1725 1726static void do_usleep(unsigned int usecs) 1727{ 1728 check_for_running_stats(); 1729 usleep(usecs); 1730} 1731 1732/* 1733 * Main function for kicking off and reaping jobs, as needed. 1734 */ 1735static void run_threads(void) 1736{ 1737 struct thread_data *td; 1738 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started; 1739 uint64_t spent; 1740 1741 if (fio_gtod_offload && fio_start_gtod_thread()) 1742 return; 1743 1744 fio_idle_prof_init(); 1745 1746 set_sig_handlers(); 1747 1748 nr_thread = nr_process = 0; 1749 for_each_td(td, i) { 1750 if (td->o.use_thread) 1751 nr_thread++; 1752 else 1753 nr_process++; 1754 } 1755 1756 if (output_format == FIO_OUTPUT_NORMAL) { 1757 log_info("Starting "); 1758 if (nr_thread) 1759 log_info("%d thread%s", nr_thread, 1760 nr_thread > 1 ? "s" : ""); 1761 if (nr_process) { 1762 if (nr_thread) 1763 log_info(" and "); 1764 log_info("%d process%s", nr_process, 1765 nr_process > 1 ? "es" : ""); 1766 } 1767 log_info("\n"); 1768 log_info_flush(); 1769 } 1770 1771 todo = thread_number; 1772 nr_running = 0; 1773 nr_started = 0; 1774 m_rate = t_rate = 0; 1775 1776 for_each_td(td, i) { 1777 print_status_init(td->thread_number - 1); 1778 1779 if (!td->o.create_serialize) 1780 continue; 1781 1782 /* 1783 * do file setup here so it happens sequentially, 1784 * we don't want X number of threads getting their 1785 * client data interspersed on disk 1786 */ 1787 if (setup_files(td)) { 1788 exit_value++; 1789 if (td->error) 1790 log_err("fio: pid=%d, err=%d/%s\n", 1791 (int) td->pid, td->error, td->verror); 1792 td_set_runstate(td, TD_REAPED); 1793 todo--; 1794 } else { 1795 struct fio_file *f; 1796 unsigned int j; 1797 1798 /* 1799 * for sharing to work, each job must always open 1800 * its own files. so close them, if we opened them 1801 * for creation 1802 */ 1803 for_each_file(td, f, j) { 1804 if (fio_file_open(f)) 1805 td_io_close_file(td, f); 1806 } 1807 } 1808 } 1809 1810 /* start idle threads before io threads start to run */ 1811 fio_idle_prof_start(); 1812 1813 set_genesis_time(); 1814 1815 while (todo) { 1816 struct thread_data *map[REAL_MAX_JOBS]; 1817 struct timeval this_start; 1818 int this_jobs = 0, left; 1819 1820 /* 1821 * create threads (TD_NOT_CREATED -> TD_CREATED) 1822 */ 1823 for_each_td(td, i) { 1824 if (td->runstate != TD_NOT_CREATED) 1825 continue; 1826 1827 /* 1828 * never got a chance to start, killed by other 1829 * thread for some reason 1830 */ 1831 if (td->terminate) { 1832 todo--; 1833 continue; 1834 } 1835 1836 if (td->o.start_delay) { 1837 spent = utime_since_genesis(); 1838 1839 if (td->o.start_delay > spent) 1840 continue; 1841 } 1842 1843 if (td->o.stonewall && (nr_started || nr_running)) { 1844 dprint(FD_PROCESS, "%s: stonewall wait\n", 1845 td->o.name); 1846 break; 1847 } 1848 1849 init_disk_util(td); 1850 1851 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED); 1852 td->update_rusage = 0; 1853 1854 /* 1855 * Set state to created. Thread will transition 1856 * to TD_INITIALIZED when it's done setting up. 1857 */ 1858 td_set_runstate(td, TD_CREATED); 1859 map[this_jobs++] = td; 1860 nr_started++; 1861 1862 if (td->o.use_thread) { 1863 int ret; 1864 1865 dprint(FD_PROCESS, "will pthread_create\n"); 1866 ret = pthread_create(&td->thread, NULL, 1867 thread_main, td); 1868 if (ret) { 1869 log_err("pthread_create: %s\n", 1870 strerror(ret)); 1871 nr_started--; 1872 break; 1873 } 1874 ret = pthread_detach(td->thread); 1875 if (ret) 1876 log_err("pthread_detach: %s", 1877 strerror(ret)); 1878 } else { 1879 pid_t pid; 1880 dprint(FD_PROCESS, "will fork\n"); 1881 pid = fork(); 1882 if (!pid) { 1883 int ret = fork_main(shm_id, i); 1884 1885 _exit(ret); 1886 } else if (i == fio_debug_jobno) 1887 *fio_debug_jobp = pid; 1888 } 1889 dprint(FD_MUTEX, "wait on startup_mutex\n"); 1890 if (fio_mutex_down_timeout(startup_mutex, 10)) { 1891 log_err("fio: job startup hung? exiting.\n"); 1892 fio_terminate_threads(TERMINATE_ALL); 1893 fio_abort = 1; 1894 nr_started--; 1895 break; 1896 } 1897 dprint(FD_MUTEX, "done waiting on startup_mutex\n"); 1898 } 1899 1900 /* 1901 * Wait for the started threads to transition to 1902 * TD_INITIALIZED. 1903 */ 1904 fio_gettime(&this_start, NULL); 1905 left = this_jobs; 1906 while (left && !fio_abort) { 1907 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT) 1908 break; 1909 1910 do_usleep(100000); 1911 1912 for (i = 0; i < this_jobs; i++) { 1913 td = map[i]; 1914 if (!td) 1915 continue; 1916 if (td->runstate == TD_INITIALIZED) { 1917 map[i] = NULL; 1918 left--; 1919 } else if (td->runstate >= TD_EXITED) { 1920 map[i] = NULL; 1921 left--; 1922 todo--; 1923 nr_running++; /* work-around... */ 1924 } 1925 } 1926 } 1927 1928 if (left) { 1929 log_err("fio: %d job%s failed to start\n", left, 1930 left > 1 ? "s" : ""); 1931 for (i = 0; i < this_jobs; i++) { 1932 td = map[i]; 1933 if (!td) 1934 continue; 1935 kill(td->pid, SIGTERM); 1936 } 1937 break; 1938 } 1939 1940 /* 1941 * start created threads (TD_INITIALIZED -> TD_RUNNING). 1942 */ 1943 for_each_td(td, i) { 1944 if (td->runstate != TD_INITIALIZED) 1945 continue; 1946 1947 if (in_ramp_time(td)) 1948 td_set_runstate(td, TD_RAMP); 1949 else 1950 td_set_runstate(td, TD_RUNNING); 1951 nr_running++; 1952 nr_started--; 1953 m_rate += ddir_rw_sum(td->o.ratemin); 1954 t_rate += ddir_rw_sum(td->o.rate); 1955 todo--; 1956 fio_mutex_up(td->mutex); 1957 } 1958 1959 reap_threads(&nr_running, &t_rate, &m_rate); 1960 1961 if (todo) 1962 do_usleep(100000); 1963 } 1964 1965 while (nr_running) { 1966 reap_threads(&nr_running, &t_rate, &m_rate); 1967 do_usleep(10000); 1968 } 1969 1970 fio_idle_prof_stop(); 1971 1972 update_io_ticks(); 1973} 1974 1975static void wait_for_disk_thread_exit(void) 1976{ 1977 void *ret; 1978 1979 disk_util_start_exit(); 1980 pthread_cond_signal(&du_cond); 1981 pthread_join(disk_util_thread, &ret); 1982} 1983 1984static void free_disk_util(void) 1985{ 1986 disk_util_prune_entries(); 1987 1988 pthread_cond_destroy(&du_cond); 1989} 1990 1991static void *disk_thread_main(void *data) 1992{ 1993 int ret = 0; 1994 1995 fio_mutex_up(startup_mutex); 1996 1997 while (!ret) { 1998 uint64_t sec = DISK_UTIL_MSEC / 1000; 1999 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000; 2000 struct timespec ts; 2001 struct timeval tv; 2002 2003 gettimeofday(&tv, NULL); 2004 ts.tv_sec = tv.tv_sec + sec; 2005 ts.tv_nsec = (tv.tv_usec * 1000) + nsec; 2006 if (ts.tv_nsec > 1000000000ULL) { 2007 ts.tv_nsec -= 1000000000ULL; 2008 ts.tv_sec++; 2009 } 2010 2011 ret = pthread_cond_timedwait(&du_cond, &du_lock, &ts); 2012 if (ret != ETIMEDOUT) 2013 break; 2014 2015 ret = update_io_ticks(); 2016 2017 if (!is_backend) 2018 print_thread_status(); 2019 } 2020 2021 return NULL; 2022} 2023 2024static int create_disk_util_thread(void) 2025{ 2026 int ret; 2027 2028 setup_disk_util(); 2029 2030 pthread_cond_init(&du_cond, NULL); 2031 pthread_mutex_init(&du_lock, NULL); 2032 2033 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL); 2034 if (ret) { 2035 log_err("Can't create disk util thread: %s\n", strerror(ret)); 2036 return 1; 2037 } 2038 2039 dprint(FD_MUTEX, "wait on startup_mutex\n"); 2040 fio_mutex_down(startup_mutex); 2041 dprint(FD_MUTEX, "done waiting on startup_mutex\n"); 2042 return 0; 2043} 2044 2045int fio_backend(void) 2046{ 2047 struct thread_data *td; 2048 int i; 2049 2050 if (exec_profile) { 2051 if (load_profile(exec_profile)) 2052 return 1; 2053 free(exec_profile); 2054 exec_profile = NULL; 2055 } 2056 if (!thread_number) 2057 return 0; 2058 2059 if (write_bw_log) { 2060 struct log_params p = { 2061 .log_type = IO_LOG_TYPE_BW, 2062 }; 2063 2064 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log"); 2065 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log"); 2066 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log"); 2067 } 2068 2069 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED); 2070 if (startup_mutex == NULL) 2071 return 1; 2072 2073 set_genesis_time(); 2074 stat_init(); 2075 create_disk_util_thread(); 2076 create_status_interval_thread(); 2077 2078 cgroup_list = smalloc(sizeof(*cgroup_list)); 2079 INIT_FLIST_HEAD(cgroup_list); 2080 2081 run_threads(); 2082 2083 wait_for_disk_thread_exit(); 2084 wait_for_status_interval_thread_exit(); 2085 2086 if (!fio_abort) { 2087 __show_run_stats(); 2088 if (write_bw_log) { 2089 int i; 2090 2091 for (i = 0; i < DDIR_RWDIR_CNT; i++) { 2092 struct io_log *log = agg_io_log[i]; 2093 2094 flush_log(log); 2095 free_log(log); 2096 } 2097 } 2098 } 2099 2100 for_each_td(td, i) 2101 fio_options_free(td); 2102 2103 free_disk_util(); 2104 cgroup_kill(cgroup_list); 2105 sfree(cgroup_list); 2106 sfree(cgroup_mnt); 2107 2108 fio_mutex_remove(startup_mutex); 2109 stat_exit(); 2110 return exit_value; 2111} 2112