projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Merge in crc32c-intel probe
[fio.git] / backend.c
CommitLineData
2e1df07d
JA		 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>
e43606c2 33#include <inttypes.h>
2e1df07d
JA		 34#include <sys/stat.h>
35#include <sys/wait.h>
36#include <sys/ipc.h>
2e1df07d
JA		 37#include <sys/mman.h>
38
39#include "fio.h"
a5e0ee11
O		 40#ifndef FIO_NO_HAVE_SHM_H
41#include <sys/shm.h>
42#endif
2e1df07d
JA		 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"
44404c5a 53#include "lib/getrusage.h"
f2a2ce0e 54#include "idletime.h"
2e1df07d
JA		 55
56static pthread_t disk_util_thread;
9ec7779f 57static struct fio_mutex *disk_thread_mutex;
2e1df07d
JA		 58static struct fio_mutex *startup_mutex;
59static struct fio_mutex *writeout_mutex;
60static struct flist_head *cgroup_list;
61static char *cgroup_mnt;
62static int exit_value;
63static volatile int fio_abort;
64
6eaf09d6 65struct io_log *agg_io_log[DDIR_RWDIR_CNT];
2e1df07d 66
a3efc919
JA		 67int groupid = 0;
68unsigned int thread_number = 0;
108fea77 69unsigned int stat_number = 0;
a3efc919
JA		 70unsigned int nr_process = 0;
71unsigned int nr_thread = 0;
72int shm_id = 0;
73int temp_stall_ts;
74unsigned long done_secs = 0;
27357187 75volatile int disk_util_exit = 0;
a3efc919 76
2e1df07d 77#define PAGE_ALIGN(buf) \
e43606c2 78 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
2e1df07d
JA		 79
80#define JOB_START_TIMEOUT (5 * 1000)
81
82static void sig_int(int sig)
83{
84 if (threads) {
85 if (is_backend)
86 fio_server_got_signal(sig);
87 else {
88 log_info("\nfio: terminating on signal %d\n", sig);
89 fflush(stdout);
90 exit_value = 128;
91 }
92
93 fio_terminate_threads(TERMINATE_ALL);
94 }
95}
96
4c6d91e8
JA		 97static void sig_show_status(int sig)
98{
99 show_running_run_stats();
100}
101
2e1df07d
JA		 102static void set_sig_handlers(void)
103{
104 struct sigaction act;
105
106 memset(&act, 0, sizeof(act));
107 act.sa_handler = sig_int;
108 act.sa_flags = SA_RESTART;
109 sigaction(SIGINT, &act, NULL);
110
111 memset(&act, 0, sizeof(act));
112 act.sa_handler = sig_int;
113 act.sa_flags = SA_RESTART;
114 sigaction(SIGTERM, &act, NULL);
115
2f694507
BC		 116/* Windows uses SIGBREAK as a quit signal from other applications */
117#ifdef WIN32
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGBREAK, &act, NULL);
122#endif
123
4c6d91e8
JA		 124 memset(&act, 0, sizeof(act));
125 act.sa_handler = sig_show_status;
126 act.sa_flags = SA_RESTART;
127 sigaction(SIGUSR1, &act, NULL);
128
2e1df07d
JA		 129 if (is_backend) {
130 memset(&act, 0, sizeof(act));
131 act.sa_handler = sig_int;
132 act.sa_flags = SA_RESTART;
133 sigaction(SIGPIPE, &act, NULL);
134 }
135}
136
137/*
138 * Check if we are above the minimum rate given.
139 */
140static int __check_min_rate(struct thread_data *td, struct timeval *now,
141 enum fio_ddir ddir)
142{
143 unsigned long long bytes = 0;
144 unsigned long iops = 0;
145 unsigned long spent;
146 unsigned long rate;
147 unsigned int ratemin = 0;
148 unsigned int rate_iops = 0;
149 unsigned int rate_iops_min = 0;
150
151 assert(ddir_rw(ddir));
152
153 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
154 return 0;
155
156 /*
157 * allow a 2 second settle period in the beginning
158 */
159 if (mtime_since(&td->start, now) < 2000)
160 return 0;
161
162 iops += td->this_io_blocks[ddir];
163 bytes += td->this_io_bytes[ddir];
164 ratemin += td->o.ratemin[ddir];
165 rate_iops += td->o.rate_iops[ddir];
166 rate_iops_min += td->o.rate_iops_min[ddir];
167
168 /*
169 * if rate blocks is set, sample is running
170 */
171 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
172 spent = mtime_since(&td->lastrate[ddir], now);
173 if (spent < td->o.ratecycle)
174 return 0;
175
176 if (td->o.rate[ddir]) {
177 /*
178 * check bandwidth specified rate
179 */
180 if (bytes < td->rate_bytes[ddir]) {
181 log_err("%s: min rate %u not met\n", td->o.name,
182 ratemin);
183 return 1;
184 } else {
185 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
186 if (rate < ratemin ||
187 bytes < td->rate_bytes[ddir]) {
188 log_err("%s: min rate %u not met, got"
189 " %luKB/sec\n", td->o.name,
190 ratemin, rate);
191 return 1;
192 }
193 }
194 } else {
195 /*
196 * checks iops specified rate
197 */
198 if (iops < rate_iops) {
199 log_err("%s: min iops rate %u not met\n",
200 td->o.name, rate_iops);
201 return 1;
202 } else {
203 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
204 if (rate < rate_iops_min ||
205 iops < td->rate_blocks[ddir]) {
206 log_err("%s: min iops rate %u not met,"
207 " got %lu\n", td->o.name,
208 rate_iops_min, rate);
209 }
210 }
211 }
212 }
213
214 td->rate_bytes[ddir] = bytes;
215 td->rate_blocks[ddir] = iops;
216 memcpy(&td->lastrate[ddir], now, sizeof(*now));
217 return 0;
218}
219
220static int check_min_rate(struct thread_data *td, struct timeval *now,
100f49f1 221 uint64_t *bytes_done)
2e1df07d
JA		 222{
223 int ret = 0;
224
6eaf09d6
SL		 225 if (bytes_done[DDIR_READ])
226 ret |= __check_min_rate(td, now, DDIR_READ);
227 if (bytes_done[DDIR_WRITE])
228 ret |= __check_min_rate(td, now, DDIR_WRITE);
229 if (bytes_done[DDIR_TRIM])
230 ret |= __check_min_rate(td, now, DDIR_TRIM);
2e1df07d
JA		 231
232 return ret;
233}
234
235/*
236 * When job exits, we can cancel the in-flight IO if we are using async
237 * io. Attempt to do so.
238 */
239static void cleanup_pending_aio(struct thread_data *td)
240{
241 struct flist_head *entry, *n;
242 struct io_u *io_u;
243 int r;
244
245 /*
246 * get immediately available events, if any
247 */
248 r = io_u_queued_complete(td, 0, NULL);
249 if (r < 0)
250 return;
251
252 /*
253 * now cancel remaining active events
254 */
255 if (td->io_ops->cancel) {
256 flist_for_each_safe(entry, n, &td->io_u_busylist) {
257 io_u = flist_entry(entry, struct io_u, list);
258
259 /*
260 * if the io_u isn't in flight, then that generally
261 * means someone leaked an io_u. complain but fix
262 * it up, so we don't stall here.
263 */
264 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
265 log_err("fio: non-busy IO on busy list\n");
266 put_io_u(td, io_u);
267 } else {
268 r = td->io_ops->cancel(td, io_u);
269 if (!r)
270 put_io_u(td, io_u);
271 }
272 }
273 }
274
275 if (td->cur_depth)
276 r = io_u_queued_complete(td, td->cur_depth, NULL);
277}
278
279/*
280 * Helper to handle the final sync of a file. Works just like the normal
281 * io path, just does everything sync.
282 */
283static int fio_io_sync(struct thread_data *td, struct fio_file *f)
284{
285 struct io_u *io_u = __get_io_u(td);
286 int ret;
287
288 if (!io_u)
289 return 1;
290
291 io_u->ddir = DDIR_SYNC;
292 io_u->file = f;
293
294 if (td_io_prep(td, io_u)) {
295 put_io_u(td, io_u);
296 return 1;
297 }
298
299requeue:
300 ret = td_io_queue(td, io_u);
301 if (ret < 0) {
302 td_verror(td, io_u->error, "td_io_queue");
303 put_io_u(td, io_u);
304 return 1;
305 } else if (ret == FIO_Q_QUEUED) {
306 if (io_u_queued_complete(td, 1, NULL) < 0)
307 return 1;
308 } else if (ret == FIO_Q_COMPLETED) {
309 if (io_u->error) {
310 td_verror(td, io_u->error, "td_io_queue");
311 return 1;
312 }
313
314 if (io_u_sync_complete(td, io_u, NULL) < 0)
315 return 1;
316 } else if (ret == FIO_Q_BUSY) {
317 if (td_io_commit(td))
318 return 1;
319 goto requeue;
320 }
321
322 return 0;
323}
a3efc919 324
61ee0f86
JA		 325static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
326{
327 int ret;
328
329 if (fio_file_open(f))
330 return fio_io_sync(td, f);
331
332 if (td_io_open_file(td, f))
333 return 1;
334
335 ret = fio_io_sync(td, f);
336 td_io_close_file(td, f);
337 return ret;
338}
339
2e1df07d
JA		 340static inline void __update_tv_cache(struct thread_data *td)
341{
342 fio_gettime(&td->tv_cache, NULL);
343}
344
345static inline void update_tv_cache(struct thread_data *td)
346{
347 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
348 __update_tv_cache(td);
349}
350
351static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
352{
353 if (in_ramp_time(td))
354 return 0;
355 if (!td->o.timeout)
356 return 0;
357 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
358 return 1;
359
360 return 0;
361}
362
363static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
364 int *retptr)
365{
366 int ret = *retptr;
367
368 if (ret < 0 || td->error) {
8b28bd41
DM		 369 int err = td->error;
370 enum error_type_bit eb;
2e1df07d
JA		 371
372 if (ret < 0)
373 err = -ret;
2e1df07d 374
8b28bd41
DM		 375 eb = td_error_type(ddir, err);
376 if (!(td->o.continue_on_error & (1 << eb)))
2e1df07d
JA		 377 return 1;
378
8b28bd41 379 if (td_non_fatal_error(td, eb, err)) {
2e1df07d
JA		 380 /*
381 * Continue with the I/Os in case of
382 * a non fatal error.
383 */
384 update_error_count(td, err);
385 td_clear_error(td);
386 *retptr = 0;
387 return 0;
388 } else if (td->o.fill_device && err == ENOSPC) {
389 /*
390 * We expect to hit this error if
391 * fill_device option is set.
392 */
393 td_clear_error(td);
394 td->terminate = 1;
395 return 1;
396 } else {
397 /*
398 * Stop the I/O in case of a fatal
399 * error.
400 */
401 update_error_count(td, err);
402 return 1;
403 }
404 }
405
406 return 0;
407}
408
c97f1ad6
JA		 409static void check_update_rusage(struct thread_data *td)
410{
411 if (td->update_rusage) {
412 td->update_rusage = 0;
413 update_rusage_stat(td);
414 fio_mutex_up(td->rusage_sem);
415 }
416}
417
2e1df07d
JA		 418/*
419 * The main verify engine. Runs over the writes we previously submitted,
420 * reads the blocks back in, and checks the crc/md5 of the data.
421 */
100f49f1 422static void do_verify(struct thread_data *td, uint64_t verify_bytes)
2e1df07d 423{
100f49f1 424 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
2e1df07d
JA		 425 struct fio_file *f;
426 struct io_u *io_u;
427 int ret, min_events;
428 unsigned int i;
429
430 dprint(FD_VERIFY, "starting loop\n");
431
432 /*
433 * sync io first and invalidate cache, to make sure we really
434 * read from disk.
435 */
436 for_each_file(td, f, i) {
437 if (!fio_file_open(f))
438 continue;
439 if (fio_io_sync(td, f))
440 break;
441 if (file_invalidate_cache(td, f))
442 break;
443 }
444
c97f1ad6
JA		 445 check_update_rusage(td);
446
2e1df07d
JA		 447 if (td->error)
448 return;
449
450 td_set_runstate(td, TD_VERIFYING);
451
452 io_u = NULL;
453 while (!td->terminate) {
fbccf46c 454 enum fio_ddir ddir;
2e1df07d
JA		 455 int ret2, full;
456
457 update_tv_cache(td);
c97f1ad6 458 check_update_rusage(td);
2e1df07d
JA		 459
460 if (runtime_exceeded(td, &td->tv_cache)) {
461 __update_tv_cache(td);
462 if (runtime_exceeded(td, &td->tv_cache)) {
463 td->terminate = 1;
464 break;
465 }
466 }
467
9e684a49
DE		 468 if (flow_threshold_exceeded(td))
469 continue;
470
44cbc6da
JA		 471 if (!td->o.experimental_verify) {
472 io_u = __get_io_u(td);
473 if (!io_u)
474 break;
2e1df07d 475
44cbc6da
JA		 476 if (get_next_verify(td, io_u)) {
477 put_io_u(td, io_u);
478 break;
479 }
2e1df07d 480
44cbc6da
JA		 481 if (td_io_prep(td, io_u)) {
482 put_io_u(td, io_u);
483 break;
484 }
485 } else {
100f49f1
JA		 486 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
487 break;
488
bcd5abfa
JA		 489 while ((io_u = get_io_u(td)) != NULL) {
490 /*
491 * We are only interested in the places where
492 * we wrote or trimmed IOs. Turn those into
493 * reads for verification purposes.
494 */
495 if (io_u->ddir == DDIR_READ) {
496 /*
497 * Pretend we issued it for rwmix
498 * accounting
499 */
500 td->io_issues[DDIR_READ]++;
501 put_io_u(td, io_u);
502 continue;
503 } else if (io_u->ddir == DDIR_TRIM) {
504 io_u->ddir = DDIR_READ;
505 io_u->flags |= IO_U_F_TRIMMED;
506 break;
507 } else if (io_u->ddir == DDIR_WRITE) {
508 io_u->ddir = DDIR_READ;
509 break;
510 } else {
511 put_io_u(td, io_u);
512 continue;
513 }
514 }
44cbc6da 515
bcd5abfa 516 if (!io_u)
44cbc6da 517 break;
2e1df07d
JA		 518 }
519
520 if (td->o.verify_async)
521 io_u->end_io = verify_io_u_async;
522 else
523 io_u->end_io = verify_io_u;
524
fbccf46c
JA		 525 ddir = io_u->ddir;
526
2e1df07d
JA		 527 ret = td_io_queue(td, io_u);
528 switch (ret) {
529 case FIO_Q_COMPLETED:
530 if (io_u->error) {
531 ret = -io_u->error;
532 clear_io_u(td, io_u);
533 } else if (io_u->resid) {
534 int bytes = io_u->xfer_buflen - io_u->resid;
535
536 /*
537 * zero read, fail
538 */
539 if (!bytes) {
540 td_verror(td, EIO, "full resid");
541 put_io_u(td, io_u);
542 break;
543 }
544
545 io_u->xfer_buflen = io_u->resid;
546 io_u->xfer_buf += bytes;
547 io_u->offset += bytes;
548
549 if (ddir_rw(io_u->ddir))
550 td->ts.short_io_u[io_u->ddir]++;
551
552 f = io_u->file;
553 if (io_u->offset == f->real_file_size)
554 goto sync_done;
555
556 requeue_io_u(td, &io_u);
557 } else {
558sync_done:
100f49f1 559 ret = io_u_sync_complete(td, io_u, bytes_done);
2e1df07d
JA		 560 if (ret < 0)
561 break;
562 }
563 continue;
564 case FIO_Q_QUEUED:
565 break;
566 case FIO_Q_BUSY:
567 requeue_io_u(td, &io_u);
568 ret2 = td_io_commit(td);
569 if (ret2 < 0)
570 ret = ret2;
571 break;
572 default:
573 assert(ret < 0);
574 td_verror(td, -ret, "td_io_queue");
575 break;
576 }
577
fbccf46c 578 if (break_on_this_error(td, ddir, &ret))
2e1df07d
JA		 579 break;
580
581 /*
582 * if we can queue more, do so. but check if there are
583 * completed io_u's first. Note that we can get BUSY even
584 * without IO queued, if the system is resource starved.
585 */
586 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
587 if (full || !td->o.iodepth_batch_complete) {
588 min_events = min(td->o.iodepth_batch_complete,
589 td->cur_depth);
8a74b56d
JA		 590 /*
591 * if the queue is full, we MUST reap at least 1 event
592 */
593 if (full && !min_events)
2e1df07d
JA		 594 min_events = 1;
595
596 do {
597 /*
598 * Reap required number of io units, if any,
599 * and do the verification on them through
600 * the callback handler
601 */
100f49f1 602 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
2e1df07d
JA		 603 ret = -1;
604 break;
605 }
606 } while (full && (td->cur_depth > td->o.iodepth_low));
607 }
608 if (ret < 0)
609 break;
610 }
611
c97f1ad6
JA		 612 check_update_rusage(td);
613
2e1df07d
JA		 614 if (!td->error) {
615 min_events = td->cur_depth;
616
617 if (min_events)
618 ret = io_u_queued_complete(td, min_events, NULL);
619 } else
620 cleanup_pending_aio(td);
621
622 td_set_runstate(td, TD_RUNNING);
623
624 dprint(FD_VERIFY, "exiting loop\n");
625}
626
f7078f7b
JA		 627static int io_bytes_exceeded(struct thread_data *td)
628{
629 unsigned long long bytes;
630
631 if (td_rw(td))
6eaf09d6 632 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
f7078f7b 633 else if (td_write(td))
6eaf09d6
SL		 634 bytes = td->this_io_bytes[DDIR_WRITE];
635 else if (td_read(td))
636 bytes = td->this_io_bytes[DDIR_READ];
f7078f7b 637 else
6eaf09d6 638 bytes = td->this_io_bytes[DDIR_TRIM];
f7078f7b
JA		 639
640 return bytes >= td->o.size;
641}
642
2e1df07d
JA		 643/*
644 * Main IO worker function. It retrieves io_u's to process and queues
645 * and reaps them, checking for rate and errors along the way.
100f49f1
JA		 646 *
647 * Returns number of bytes written and trimmed.
2e1df07d 648 */
100f49f1 649static uint64_t do_io(struct thread_data *td)
2e1df07d 650{
100f49f1 651 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
2e1df07d
JA		 652 unsigned int i;
653 int ret = 0;
654
655 if (in_ramp_time(td))
656 td_set_runstate(td, TD_RAMP);
657 else
658 td_set_runstate(td, TD_RUNNING);
659
f7078f7b 660 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
c04e4661
DE		 661 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
662 td->o.time_based) {
2e1df07d 663 struct timeval comp_time;
2e1df07d
JA		 664 int min_evts = 0;
665 struct io_u *io_u;
666 int ret2, full;
667 enum fio_ddir ddir;
668
c97f1ad6
JA		 669 check_update_rusage(td);
670
7d7803fa 671 if (td->terminate || td->done)
2e1df07d
JA		 672 break;
673
674 update_tv_cache(td);
675
676 if (runtime_exceeded(td, &td->tv_cache)) {
677 __update_tv_cache(td);
678 if (runtime_exceeded(td, &td->tv_cache)) {
679 td->terminate = 1;
680 break;
681 }
682 }
683
9e684a49
DE		 684 if (flow_threshold_exceeded(td))
685 continue;
686
2e1df07d
JA		 687 io_u = get_io_u(td);
688 if (!io_u)
689 break;
690
691 ddir = io_u->ddir;
692
693 /*
82af2a7c
JA		 694 * Add verification end_io handler if:
695 * - Asked to verify (!td_rw(td))
696 * - Or the io_u is from our verify list (mixed write/ver)
2e1df07d
JA		 697 */
698 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
82af2a7c 699 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
2e1df07d
JA		 700 if (td->o.verify_async)
701 io_u->end_io = verify_io_u_async;
702 else
703 io_u->end_io = verify_io_u;
704 td_set_runstate(td, TD_VERIFYING);
705 } else if (in_ramp_time(td))
706 td_set_runstate(td, TD_RAMP);
707 else
708 td_set_runstate(td, TD_RUNNING);
709
710 ret = td_io_queue(td, io_u);
711 switch (ret) {
712 case FIO_Q_COMPLETED:
713 if (io_u->error) {
714 ret = -io_u->error;
715 clear_io_u(td, io_u);
716 } else if (io_u->resid) {
717 int bytes = io_u->xfer_buflen - io_u->resid;
718 struct fio_file *f = io_u->file;
719
720 /*
721 * zero read, fail
722 */
723 if (!bytes) {
724 td_verror(td, EIO, "full resid");
725 put_io_u(td, io_u);
726 break;
727 }
728
729 io_u->xfer_buflen = io_u->resid;
730 io_u->xfer_buf += bytes;
731 io_u->offset += bytes;
732
733 if (ddir_rw(io_u->ddir))
734 td->ts.short_io_u[io_u->ddir]++;
735
736 if (io_u->offset == f->real_file_size)
737 goto sync_done;
738
739 requeue_io_u(td, &io_u);
740 } else {
741sync_done:
6eaf09d6
SL		 742 if (__should_check_rate(td, DDIR_READ) ||
743 __should_check_rate(td, DDIR_WRITE) ||
744 __should_check_rate(td, DDIR_TRIM))
2e1df07d
JA		 745 fio_gettime(&comp_time, NULL);
746
747 ret = io_u_sync_complete(td, io_u, bytes_done);
748 if (ret < 0)
749 break;
750 }
751 break;
752 case FIO_Q_QUEUED:
753 /*
754 * if the engine doesn't have a commit hook,
755 * the io_u is really queued. if it does have such
756 * a hook, it has to call io_u_queued() itself.
757 */
758 if (td->io_ops->commit == NULL)
759 io_u_queued(td, io_u);
760 break;
761 case FIO_Q_BUSY:
762 requeue_io_u(td, &io_u);
763 ret2 = td_io_commit(td);
764 if (ret2 < 0)
765 ret = ret2;
766 break;
767 default:
768 assert(ret < 0);
769 put_io_u(td, io_u);
770 break;
771 }
772
773 if (break_on_this_error(td, ddir, &ret))
774 break;
775
776 /*
777 * See if we need to complete some commands. Note that we
778 * can get BUSY even without IO queued, if the system is
779 * resource starved.
780 */
781 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
782 if (full || !td->o.iodepth_batch_complete) {
783 min_evts = min(td->o.iodepth_batch_complete,
784 td->cur_depth);
8a74b56d
JA		 785 /*
786 * if the queue is full, we MUST reap at least 1 event
787 */
788 if (full && !min_evts)
2e1df07d
JA		 789 min_evts = 1;
790
6eaf09d6
SL		 791 if (__should_check_rate(td, DDIR_READ) ||
792 __should_check_rate(td, DDIR_WRITE) ||
793 __should_check_rate(td, DDIR_TRIM))
2e1df07d
JA		 794 fio_gettime(&comp_time, NULL);
795
796 do {
797 ret = io_u_queued_complete(td, min_evts, bytes_done);
798 if (ret < 0)
799 break;
800
801 } while (full && (td->cur_depth > td->o.iodepth_low));
802 }
803
804 if (ret < 0)
805 break;
d5abee06 806 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
2e1df07d
JA		 807 continue;
808
809 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
810 if (check_min_rate(td, &comp_time, bytes_done)) {
811 if (exitall_on_terminate)
812 fio_terminate_threads(td->groupid);
813 td_verror(td, EIO, "check_min_rate");
814 break;
815 }
816 }
817
818 if (td->o.thinktime) {
819 unsigned long long b;
820
342f4be4 821 b = ddir_rw_sum(td->io_blocks);
2e1df07d
JA		 822 if (!(b % td->o.thinktime_blocks)) {
823 int left;
824
825 if (td->o.thinktime_spin)
826 usec_spin(td->o.thinktime_spin);
827
828 left = td->o.thinktime - td->o.thinktime_spin;
829 if (left)
830 usec_sleep(td, left);
831 }
832 }
833 }
834
c97f1ad6
JA		 835 check_update_rusage(td);
836
2e1df07d
JA		 837 if (td->trim_entries)
838 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
839
840 if (td->o.fill_device && td->error == ENOSPC) {
841 td->error = 0;
842 td->terminate = 1;
843 }
844 if (!td->error) {
845 struct fio_file *f;
846
847 i = td->cur_depth;
848 if (i) {
5bd5f71a 849 ret = io_u_queued_complete(td, i, bytes_done);
2e1df07d
JA		 850 if (td->o.fill_device && td->error == ENOSPC)
851 td->error = 0;
852 }
853
854 if (should_fsync(td) && td->o.end_fsync) {
855 td_set_runstate(td, TD_FSYNCING);
856
857 for_each_file(td, f, i) {
61ee0f86 858 if (!fio_file_fsync(td, f))
2e1df07d 859 continue;
61ee0f86
JA		 860
861 log_err("fio: end_fsync failed for file %s\n",
862 f->file_name);
2e1df07d
JA		 863 }
864 }
865 } else
866 cleanup_pending_aio(td);
867
868 /*
869 * stop job if we failed doing any IO
870 */
342f4be4 871 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 872 td->done = 1;
100f49f1
JA		 873
874 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
2e1df07d
JA		 875}
876
877static void cleanup_io_u(struct thread_data *td)
878{
879 struct flist_head *entry, *n;
880 struct io_u *io_u;
881
882 flist_for_each_safe(entry, n, &td->io_u_freelist) {
883 io_u = flist_entry(entry, struct io_u, list);
884
885 flist_del(&io_u->list);
c73ed246
JA		 886
887 if (td->io_ops->io_u_free)
888 td->io_ops->io_u_free(td, io_u);
889
2e1df07d
JA		 890 fio_memfree(io_u, sizeof(*io_u));
891 }
892
893 free_io_mem(td);
894}
895
896static int init_io_u(struct thread_data *td)
897{
898 struct io_u *io_u;
9c42684e 899 unsigned int max_bs, min_write;
2e1df07d 900 int cl_align, i, max_units;
59d8d0f5 901 int data_xfer = 1;
2e1df07d
JA		 902 char *p;
903
904 max_units = td->o.iodepth;
74f4b020 905 max_bs = td_max_bs(td);
9c42684e 906 min_write = td->o.min_bs[DDIR_WRITE];
2e1df07d
JA		 907 td->orig_buffer_size = (unsigned long long) max_bs
908 * (unsigned long long) max_units;
909
88045e04 910 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
59d8d0f5
JA		 911 data_xfer = 0;
912
2e1df07d
JA		 913 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
914 unsigned long bs;
915
916 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
917 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
918 }
919
920 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
921 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
922 return 1;
923 }
924
59d8d0f5 925 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA		 926 return 1;
927
928 if (td->o.odirect || td->o.mem_align ||
929 (td->io_ops->flags & FIO_RAWIO))
930 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
931 else
932 p = td->orig_buffer;
933
934 cl_align = os_cache_line_size();
935
936 for (i = 0; i < max_units; i++) {
937 void *ptr;
938
939 if (td->terminate)
940 return 1;
941
942 ptr = fio_memalign(cl_align, sizeof(*io_u));
943 if (!ptr) {
944 log_err("fio: unable to allocate aligned memory\n");
945 break;
946 }
947
948 io_u = ptr;
949 memset(io_u, 0, sizeof(*io_u));
950 INIT_FLIST_HEAD(&io_u->list);
951 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
952
59d8d0f5 953 if (data_xfer) {
2e1df07d
JA		 954 io_u->buf = p;
955 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
956
957 if (td_write(td))
9c42684e 958 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA		 959 if (td_write(td) && td->o.verify_pattern_bytes) {
960 /*
961 * Fill the buffer with the pattern if we are
962 * going to be doing writes.
963 */
964 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
965 }
966 }
967
968 io_u->index = i;
969 io_u->flags = IO_U_F_FREE;
970 flist_add(&io_u->list, &td->io_u_freelist);
c73ed246
JA		 971
972 if (td->io_ops->io_u_init) {
973 int ret = td->io_ops->io_u_init(td, io_u);
974
975 if (ret) {
976 log_err("fio: failed to init engine data: %d\n", ret);
977 return 1;
978 }
979 }
980
2e1df07d
JA		 981 p += max_bs;
982 }
983
984 return 0;
985}
986
987static int switch_ioscheduler(struct thread_data *td)
988{
989 char tmp[256], tmp2[128];
990 FILE *f;
991 int ret;
992
993 if (td->io_ops->flags & FIO_DISKLESSIO)
994 return 0;
995
996 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
997
998 f = fopen(tmp, "r+");
999 if (!f) {
1000 if (errno == ENOENT) {
1001 log_err("fio: os or kernel doesn't support IO scheduler"
1002 " switching\n");
1003 return 0;
1004 }
1005 td_verror(td, errno, "fopen iosched");
1006 return 1;
1007 }
1008
1009 /*
1010 * Set io scheduler.
1011 */
1012 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1013 if (ferror(f) || ret != 1) {
1014 td_verror(td, errno, "fwrite");
1015 fclose(f);
1016 return 1;
1017 }
1018
1019 rewind(f);
1020
1021 /*
1022 * Read back and check that the selected scheduler is now the default.
1023 */
1024 ret = fread(tmp, 1, sizeof(tmp), f);
1025 if (ferror(f) || ret < 0) {
1026 td_verror(td, errno, "fread");
1027 fclose(f);
1028 return 1;
1029 }
1030
1031 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1032 if (!strstr(tmp, tmp2)) {
1033 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1034 td_verror(td, EINVAL, "iosched_switch");
1035 fclose(f);
1036 return 1;
1037 }
1038
1039 fclose(f);
1040 return 0;
1041}
1042
1043static int keep_running(struct thread_data *td)
1044{
2e1df07d
JA		 1045 if (td->done)
1046 return 0;
1047 if (td->o.time_based)
1048 return 1;
1049 if (td->o.loops) {
1050 td->o.loops--;
1051 return 1;
1052 }
1053
ee98e234 1054 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
5bd5f71a
JA		 1055 uint64_t diff;
1056
1057 /*
1058 * If the difference is less than the minimum IO size, we
1059 * are done.
1060 */
1061 diff = td->o.size - ddir_rw_sum(td->io_bytes);
74f4b020 1062 if (diff < td_max_bs(td))
5bd5f71a
JA		 1063 return 0;
1064
2e1df07d 1065 return 1;
5bd5f71a 1066 }
2e1df07d
JA		 1067
1068 return 0;
1069}
1070
1071static int exec_string(const char *string)
1072{
1073 int ret, newlen = strlen(string) + 1 + 8;
1074 char *str;
1075
1076 str = malloc(newlen);
1077 sprintf(str, "sh -c %s", string);
1078
1079 ret = system(str);
1080 if (ret == -1)
1081 log_err("fio: exec of cmd <%s> failed\n", str);
1082
1083 free(str);
1084 return ret;
1085}
1086
1087/*
1088 * Entry point for the thread based jobs. The process based jobs end up
1089 * here as well, after a little setup.
1090 */
1091static void *thread_main(void *data)
1092{
1093 unsigned long long elapsed;
1094 struct thread_data *td = data;
1095 pthread_condattr_t attr;
1096 int clear_state;
1097
1098 if (!td->o.use_thread) {
1099 setsid();
1100 td->pid = getpid();
1101 } else
1102 td->pid = gettid();
1103
5d879392
JA		 1104 fio_local_clock_init(td->o.use_thread);
1105
2e1df07d
JA		 1106 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1107
1108 INIT_FLIST_HEAD(&td->io_u_freelist);
1109 INIT_FLIST_HEAD(&td->io_u_busylist);
1110 INIT_FLIST_HEAD(&td->io_u_requeues);
1111 INIT_FLIST_HEAD(&td->io_log_list);
1112 INIT_FLIST_HEAD(&td->io_hist_list);
1113 INIT_FLIST_HEAD(&td->verify_list);
1114 INIT_FLIST_HEAD(&td->trim_list);
1ae83d45 1115 INIT_FLIST_HEAD(&td->next_rand_list);
2e1df07d
JA		 1116 pthread_mutex_init(&td->io_u_lock, NULL);
1117 td->io_hist_tree = RB_ROOT;
1118
1119 pthread_condattr_init(&attr);
1120 pthread_cond_init(&td->verify_cond, &attr);
1121 pthread_cond_init(&td->free_cond, &attr);
1122
1123 td_set_runstate(td, TD_INITIALIZED);
1124 dprint(FD_MUTEX, "up startup_mutex\n");
1125 fio_mutex_up(startup_mutex);
1126 dprint(FD_MUTEX, "wait on td->mutex\n");
1127 fio_mutex_down(td->mutex);
1128 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1129
1130 /*
1131 * the ->mutex mutex is now no longer used, close it to avoid
1132 * eating a file descriptor
1133 */
1134 fio_mutex_remove(td->mutex);
1135
1136 /*
1137 * A new gid requires privilege, so we need to do this before setting
1138 * the uid.
1139 */
1140 if (td->o.gid != -1U && setgid(td->o.gid)) {
1141 td_verror(td, errno, "setgid");
1142 goto err;
1143 }
1144 if (td->o.uid != -1U && setuid(td->o.uid)) {
1145 td_verror(td, errno, "setuid");
1146 goto err;
1147 }
1148
1149 /*
1150 * If we have a gettimeofday() thread, make sure we exclude that
1151 * thread from this job
1152 */
1153 if (td->o.gtod_cpu)
1154 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1155
1156 /*
1157 * Set affinity first, in case it has an impact on the memory
1158 * allocations.
1159 */
1160 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1161 td_verror(td, errno, "cpu_set_affinity");
1162 goto err;
1163 }
1164
67bf9823 1165#ifdef CONFIG_LIBNUMA
d0b937ed
YR		 1166 /* numa node setup */
1167 if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1168 int ret;
1169
1170 if (numa_available() < 0) {
1171 td_verror(td, errno, "Does not support NUMA API\n");
1172 goto err;
1173 }
1174
1175 if (td->o.numa_cpumask_set) {
1176 ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1177 if (ret == -1) {
1178 td_verror(td, errno, \
1179 "numa_run_on_node_mask failed\n");
1180 goto err;
1181 }
1182 }
1183
1184 if (td->o.numa_memmask_set) {
1185
1186 switch (td->o.numa_mem_mode) {
1187 case MPOL_INTERLEAVE:
1188 numa_set_interleave_mask(td->o.numa_memnodesmask);
1189 break;
1190 case MPOL_BIND:
1191 numa_set_membind(td->o.numa_memnodesmask);
1192 break;
1193 case MPOL_LOCAL:
1194 numa_set_localalloc();
1195 break;
1196 case MPOL_PREFERRED:
1197 numa_set_preferred(td->o.numa_mem_prefer_node);
1198 break;
1199 case MPOL_DEFAULT:
1200 default:
1201 break;
1202 }
1203
1204 }
1205 }
1206#endif
1207
2e1df07d
JA		 1208 /*
1209 * May alter parameters that init_io_u() will use, so we need to
1210 * do this first.
1211 */
1212 if (init_iolog(td))
1213 goto err;
1214
1215 if (init_io_u(td))
1216 goto err;
1217
1218 if (td->o.verify_async && verify_async_init(td))
1219 goto err;
1220
1221 if (td->ioprio_set) {
1222 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1223 td_verror(td, errno, "ioprio_set");
1224 goto err;
1225 }
1226 }
1227
5d89ff79 1228 if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA		 1229 goto err;
1230
649c10c5
BC		 1231 errno = 0;
1232 if (nice(td->o.nice) == -1 && errno != 0) {
2e1df07d
JA		 1233 td_verror(td, errno, "nice");
1234 goto err;
1235 }
1236
1237 if (td->o.ioscheduler && switch_ioscheduler(td))
1238 goto err;
1239
1240 if (!td->o.create_serialize && setup_files(td))
1241 goto err;
1242
1243 if (td_io_init(td))
1244 goto err;
1245
1246 if (init_random_map(td))
1247 goto err;
1248
1249 if (td->o.exec_prerun) {
1250 if (exec_string(td->o.exec_prerun))
1251 goto err;
1252 }
1253
1254 if (td->o.pre_read) {
1255 if (pre_read_files(td) < 0)
1256 goto err;
1257 }
1258
dc5bfbb2
JA		 1259 fio_verify_init(td);
1260
2e1df07d 1261 fio_gettime(&td->epoch, NULL);
44404c5a 1262 fio_getrusage(&td->ru_start);
2e1df07d
JA		 1263 clear_state = 0;
1264 while (keep_running(td)) {
100f49f1
JA		 1265 uint64_t verify_bytes;
1266
2e1df07d
JA		 1267 fio_gettime(&td->start, NULL);
1268 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1269 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1270 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1271
6eaf09d6
SL		 1272 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1273 td->o.ratemin[DDIR_TRIM]) {
1274 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
2e1df07d 1275 sizeof(td->bw_sample_time));
6eaf09d6
SL		 1276 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1277 sizeof(td->bw_sample_time));
1278 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
2e1df07d
JA		 1279 sizeof(td->bw_sample_time));
1280 }
1281
1282 if (clear_state)
1283 clear_io_state(td);
1284
1285 prune_io_piece_log(td);
1286
100f49f1 1287 verify_bytes = do_io(td);
2e1df07d
JA		 1288
1289 clear_state = 1;
1290
1291 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1292 elapsed = utime_since_now(&td->start);
1293 td->ts.runtime[DDIR_READ] += elapsed;
1294 }
1295 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1296 elapsed = utime_since_now(&td->start);
1297 td->ts.runtime[DDIR_WRITE] += elapsed;
1298 }
6eaf09d6
SL		 1299 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1300 elapsed = utime_since_now(&td->start);
1301 td->ts.runtime[DDIR_TRIM] += elapsed;
1302 }
2e1df07d
JA		 1303
1304 if (td->error || td->terminate)
1305 break;
1306
1307 if (!td->o.do_verify ||
1308 td->o.verify == VERIFY_NONE ||
1309 (td->io_ops->flags & FIO_UNIDIR))
1310 continue;
1311
1312 clear_io_state(td);
1313
1314 fio_gettime(&td->start, NULL);
1315
100f49f1 1316 do_verify(td, verify_bytes);
2e1df07d
JA		 1317
1318 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1319
1320 if (td->error || td->terminate)
1321 break;
1322 }
1323
1324 update_rusage_stat(td);
6eaf09d6
SL		 1325 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1326 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1327 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
2e1df07d 1328 td->ts.total_run_time = mtime_since_now(&td->epoch);
6eaf09d6
SL		 1329 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1330 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1331 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2e1df07d
JA		 1332
1333 fio_mutex_down(writeout_mutex);
1334 if (td->bw_log) {
1335 if (td->o.bw_log_file) {
1336 finish_log_named(td, td->bw_log,
1337 td->o.bw_log_file, "bw");
1338 } else
1339 finish_log(td, td->bw_log, "bw");
1340 }
1341 if (td->lat_log) {
1342 if (td->o.lat_log_file) {
1343 finish_log_named(td, td->lat_log,
1344 td->o.lat_log_file, "lat");
1345 } else
1346 finish_log(td, td->lat_log, "lat");
1347 }
1348 if (td->slat_log) {
1349 if (td->o.lat_log_file) {
1350 finish_log_named(td, td->slat_log,
1351 td->o.lat_log_file, "slat");
1352 } else
1353 finish_log(td, td->slat_log, "slat");
1354 }
1355 if (td->clat_log) {
1356 if (td->o.lat_log_file) {
1357 finish_log_named(td, td->clat_log,
1358 td->o.lat_log_file, "clat");
1359 } else
1360 finish_log(td, td->clat_log, "clat");
1361 }
1362 if (td->iops_log) {
1363 if (td->o.iops_log_file) {
1364 finish_log_named(td, td->iops_log,
1365 td->o.iops_log_file, "iops");
1366 } else
1367 finish_log(td, td->iops_log, "iops");
1368 }
1369
1370 fio_mutex_up(writeout_mutex);
1371 if (td->o.exec_postrun)
1372 exec_string(td->o.exec_postrun);
1373
1374 if (exitall_on_terminate)
1375 fio_terminate_threads(td->groupid);
1376
1377err:
1378 if (td->error)
1379 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1380 td->verror);
1381
1382 if (td->o.verify_async)
1383 verify_async_exit(td);
1384
1385 close_and_free_files(td);
2e1df07d 1386 cleanup_io_u(td);
32dbca2c 1387 close_ioengine(td);
2e1df07d
JA		 1388 cgroup_shutdown(td, &cgroup_mnt);
1389
1390 if (td->o.cpumask_set) {
1391 int ret = fio_cpuset_exit(&td->o.cpumask);
1392
1393 td_verror(td, ret, "fio_cpuset_exit");
1394 }
1395
1396 /*
1397 * do this very late, it will log file closing as well
1398 */
1399 if (td->o.write_iolog_file)
1400 write_iolog_close(td);
1401
c97f1ad6
JA		 1402 fio_mutex_remove(td->rusage_sem);
1403 td->rusage_sem = NULL;
1404
2e1df07d 1405 td_set_runstate(td, TD_EXITED);
e43606c2 1406 return (void *) (uintptr_t) td->error;
2e1df07d
JA		 1407}
1408
1409
1410/*
1411 * We cannot pass the td data into a forked process, so attach the td and
1412 * pass it to the thread worker.
1413 */
1414static int fork_main(int shmid, int offset)
1415{
1416 struct thread_data *td;
1417 void *data, *ret;
1418
1419#ifndef __hpux
1420 data = shmat(shmid, NULL, 0);
1421 if (data == (void *) -1) {
1422 int __err = errno;
1423
1424 perror("shmat");
1425 return __err;
1426 }
1427#else
1428 /*
1429 * HP-UX inherits shm mappings?
1430 */
1431 data = threads;
1432#endif
1433
1434 td = data + offset * sizeof(struct thread_data);
1435 ret = thread_main(td);
1436 shmdt(data);
e43606c2 1437 return (int) (uintptr_t) ret;
2e1df07d
JA		 1438}
1439
1440/*
1441 * Run over the job map and reap the threads that have exited, if any.
1442 */
1443static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1444 unsigned int *m_rate)
1445{
1446 struct thread_data *td;
1447 unsigned int cputhreads, realthreads, pending;
1448 int i, status, ret;
1449
1450 /*
1451 * reap exited threads (TD_EXITED -> TD_REAPED)
1452 */
1453 realthreads = pending = cputhreads = 0;
1454 for_each_td(td, i) {
1455 int flags = 0;
1456
1457 /*
1458 * ->io_ops is NULL for a thread that has closed its
1459 * io engine
1460 */
1461 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1462 cputhreads++;
1463 else
1464 realthreads++;
1465
1466 if (!td->pid) {
1467 pending++;
1468 continue;
1469 }
1470 if (td->runstate == TD_REAPED)
1471 continue;
1472 if (td->o.use_thread) {
1473 if (td->runstate == TD_EXITED) {
1474 td_set_runstate(td, TD_REAPED);
1475 goto reaped;
1476 }
1477 continue;
1478 }
1479
1480 flags = WNOHANG;
1481 if (td->runstate == TD_EXITED)
1482 flags = 0;
1483
1484 /*
1485 * check if someone quit or got killed in an unusual way
1486 */
1487 ret = waitpid(td->pid, &status, flags);
1488 if (ret < 0) {
1489 if (errno == ECHILD) {
1490 log_err("fio: pid=%d disappeared %d\n",
1491 (int) td->pid, td->runstate);
a5e371a6 1492 td->sig = ECHILD;
2e1df07d
JA		 1493 td_set_runstate(td, TD_REAPED);
1494 goto reaped;
1495 }
1496 perror("waitpid");
1497 } else if (ret == td->pid) {
1498 if (WIFSIGNALED(status)) {
1499 int sig = WTERMSIG(status);
1500
36d80bc7 1501 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA		 1502 log_err("fio: pid=%d, got signal=%d\n",
1503 (int) td->pid, sig);
a5e371a6 1504 td->sig = sig;
2e1df07d
JA		 1505 td_set_runstate(td, TD_REAPED);
1506 goto reaped;
1507 }
1508 if (WIFEXITED(status)) {
1509 if (WEXITSTATUS(status) && !td->error)
1510 td->error = WEXITSTATUS(status);
1511
1512 td_set_runstate(td, TD_REAPED);
1513 goto reaped;
1514 }
1515 }
1516
1517 /*
1518 * thread is not dead, continue
1519 */
1520 pending++;
1521 continue;
1522reaped:
1523 (*nr_running)--;
342f4be4
JA		 1524 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1525 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA		 1526 if (!td->pid)
1527 pending--;
1528
1529 if (td->error)
1530 exit_value++;
1531
1532 done_secs += mtime_since_now(&td->epoch) / 1000;
1533 }
1534
1535 if (*nr_running == cputhreads && !pending && realthreads)
1536 fio_terminate_threads(TERMINATE_ALL);
1537}
1538
2e1df07d
JA		 1539/*
1540 * Main function for kicking off and reaping jobs, as needed.
1541 */
1542static void run_threads(void)
1543{
1544 struct thread_data *td;
1545 unsigned long spent;
1546 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1547
1548 if (fio_pin_memory())
1549 return;
1550
1551 if (fio_gtod_offload && fio_start_gtod_thread())
1552 return;
f2a2ce0e
HL		 1553
1554 fio_idle_prof_init();
2e1df07d
JA		 1555
1556 set_sig_handlers();
1557
f3afa57e 1558 if (output_format == FIO_OUTPUT_NORMAL) {
2e1df07d
JA		 1559 log_info("Starting ");
1560 if (nr_thread)
1561 log_info("%d thread%s", nr_thread,
1562 nr_thread > 1 ? "s" : "");
1563 if (nr_process) {
1564 if (nr_thread)
1565 log_info(" and ");
1566 log_info("%d process%s", nr_process,
1567 nr_process > 1 ? "es" : "");
1568 }
1569 log_info("\n");
1570 fflush(stdout);
1571 }
1572
1573 todo = thread_number;
1574 nr_running = 0;
1575 nr_started = 0;
1576 m_rate = t_rate = 0;
1577
1578 for_each_td(td, i) {
1579 print_status_init(td->thread_number - 1);
1580
1581 if (!td->o.create_serialize)
1582 continue;
1583
1584 /*
1585 * do file setup here so it happens sequentially,
1586 * we don't want X number of threads getting their
1587 * client data interspersed on disk
1588 */
1589 if (setup_files(td)) {
1590 exit_value++;
1591 if (td->error)
1592 log_err("fio: pid=%d, err=%d/%s\n",
1593 (int) td->pid, td->error, td->verror);
1594 td_set_runstate(td, TD_REAPED);
1595 todo--;
1596 } else {
1597 struct fio_file *f;
1598 unsigned int j;
1599
1600 /*
1601 * for sharing to work, each job must always open
1602 * its own files. so close them, if we opened them
1603 * for creation
1604 */
1605 for_each_file(td, f, j) {
1606 if (fio_file_open(f))
1607 td_io_close_file(td, f);
1608 }
1609 }
1610 }
1611
f2a2ce0e
HL		 1612 /* start idle threads before io threads start to run */
1613 fio_idle_prof_start();
1614
2e1df07d
JA		 1615 set_genesis_time();
1616
1617 while (todo) {
1618 struct thread_data *map[REAL_MAX_JOBS];
1619 struct timeval this_start;
1620 int this_jobs = 0, left;
1621
1622 /*
1623 * create threads (TD_NOT_CREATED -> TD_CREATED)
1624 */
1625 for_each_td(td, i) {
1626 if (td->runstate != TD_NOT_CREATED)
1627 continue;
1628
1629 /*
1630 * never got a chance to start, killed by other
1631 * thread for some reason
1632 */
1633 if (td->terminate) {
1634 todo--;
1635 continue;
1636 }
1637
1638 if (td->o.start_delay) {
1639 spent = mtime_since_genesis();
1640
1641 if (td->o.start_delay * 1000 > spent)
1642 continue;
1643 }
1644
1645 if (td->o.stonewall && (nr_started || nr_running)) {
1646 dprint(FD_PROCESS, "%s: stonewall wait\n",
1647 td->o.name);
1648 break;
1649 }
1650
1651 init_disk_util(td);
1652
c97f1ad6
JA		 1653 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1654 td->update_rusage = 0;
1655
2e1df07d
JA		 1656 /*
1657 * Set state to created. Thread will transition
1658 * to TD_INITIALIZED when it's done setting up.
1659 */
1660 td_set_runstate(td, TD_CREATED);
1661 map[this_jobs++] = td;
1662 nr_started++;
1663
1664 if (td->o.use_thread) {
1665 int ret;
1666
1667 dprint(FD_PROCESS, "will pthread_create\n");
1668 ret = pthread_create(&td->thread, NULL,
1669 thread_main, td);
1670 if (ret) {
1671 log_err("pthread_create: %s\n",
1672 strerror(ret));
1673 nr_started--;
1674 break;
1675 }
1676 ret = pthread_detach(td->thread);
1677 if (ret)
1678 log_err("pthread_detach: %s",
1679 strerror(ret));
1680 } else {
1681 pid_t pid;
1682 dprint(FD_PROCESS, "will fork\n");
1683 pid = fork();
1684 if (!pid) {
1685 int ret = fork_main(shm_id, i);
1686
1687 _exit(ret);
1688 } else if (i == fio_debug_jobno)
1689 *fio_debug_jobp = pid;
1690 }
1691 dprint(FD_MUTEX, "wait on startup_mutex\n");
1692 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1693 log_err("fio: job startup hung? exiting.\n");
1694 fio_terminate_threads(TERMINATE_ALL);
1695 fio_abort = 1;
1696 nr_started--;
1697 break;
1698 }
1699 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1700 }
1701
1702 /*
1703 * Wait for the started threads to transition to
1704 * TD_INITIALIZED.
1705 */
1706 fio_gettime(&this_start, NULL);
1707 left = this_jobs;
1708 while (left && !fio_abort) {
1709 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1710 break;
1711
1712 usleep(100000);
1713
1714 for (i = 0; i < this_jobs; i++) {
1715 td = map[i];
1716 if (!td)
1717 continue;
1718 if (td->runstate == TD_INITIALIZED) {
1719 map[i] = NULL;
1720 left--;
1721 } else if (td->runstate >= TD_EXITED) {
1722 map[i] = NULL;
1723 left--;
1724 todo--;
1725 nr_running++; /* work-around... */
1726 }
1727 }
1728 }
1729
1730 if (left) {
4e87c37a
JA		 1731 log_err("fio: %d job%s failed to start\n", left,
1732 left > 1 ? "s" : "");
2e1df07d
JA		 1733 for (i = 0; i < this_jobs; i++) {
1734 td = map[i];
1735 if (!td)
1736 continue;
1737 kill(td->pid, SIGTERM);
1738 }
1739 break;
1740 }
1741
1742 /*
1743 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1744 */
1745 for_each_td(td, i) {
1746 if (td->runstate != TD_INITIALIZED)
1747 continue;
1748
1749 if (in_ramp_time(td))
1750 td_set_runstate(td, TD_RAMP);
1751 else
1752 td_set_runstate(td, TD_RUNNING);
1753 nr_running++;
1754 nr_started--;
342f4be4
JA		 1755 m_rate += ddir_rw_sum(td->o.ratemin);
1756 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d
JA		 1757 todo--;
1758 fio_mutex_up(td->mutex);
1759 }
1760
1761 reap_threads(&nr_running, &t_rate, &m_rate);
1762
1763 if (todo) {
1764 if (is_backend)
1765 fio_server_idle_loop();
1766 else
1767 usleep(100000);
1768 }
1769 }
1770
1771 while (nr_running) {
1772 reap_threads(&nr_running, &t_rate, &m_rate);
1773
1774 if (is_backend)
1775 fio_server_idle_loop();
1776 else
1777 usleep(10000);
1778 }
1779
f2a2ce0e
HL		 1780 fio_idle_prof_stop();
1781
2e1df07d
JA		 1782 update_io_ticks();
1783 fio_unpin_memory();
1784}
1785
9ec7779f
JA		 1786void wait_for_disk_thread_exit(void)
1787{
1788 fio_mutex_down(disk_thread_mutex);
1789}
1790
27357187
JA		 1791static void free_disk_util(void)
1792{
1793 disk_util_start_exit();
1794 wait_for_disk_thread_exit();
1795 disk_util_prune_entries();
1796}
1797
2e1df07d
JA		 1798static void *disk_thread_main(void *data)
1799{
9ec7779f
JA		 1800 int ret = 0;
1801
2e1df07d
JA		 1802 fio_mutex_up(startup_mutex);
1803
9ec7779f 1804 while (threads && !ret) {
2e1df07d
JA		 1805 usleep(DISK_UTIL_MSEC * 1000);
1806 if (!threads)
1807 break;
9ec7779f 1808 ret = update_io_ticks();
2e1df07d
JA		 1809
1810 if (!is_backend)
1811 print_thread_status();
1812 }
1813
9ec7779f 1814 fio_mutex_up(disk_thread_mutex);
2e1df07d
JA		 1815 return NULL;
1816}
1817
1818static int create_disk_util_thread(void)
1819{
1820 int ret;
1821
9ec7779f
JA		 1822 setup_disk_util();
1823
521da527 1824 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
9ec7779f 1825
2e1df07d
JA		 1826 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1827 if (ret) {
9ec7779f 1828 fio_mutex_remove(disk_thread_mutex);
2e1df07d
JA		 1829 log_err("Can't create disk util thread: %s\n", strerror(ret));
1830 return 1;
1831 }
1832
1833 ret = pthread_detach(disk_util_thread);
1834 if (ret) {
9ec7779f 1835 fio_mutex_remove(disk_thread_mutex);
2e1df07d
JA		 1836 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1837 return 1;
1838 }
1839
1840 dprint(FD_MUTEX, "wait on startup_mutex\n");
1841 fio_mutex_down(startup_mutex);
1842 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1843 return 0;
1844}
1845
2e1df07d
JA		 1846int fio_backend(void)
1847{
1848 struct thread_data *td;
1849 int i;
1850
1851 if (exec_profile) {
1852 if (load_profile(exec_profile))
1853 return 1;
1854 free(exec_profile);
1855 exec_profile = NULL;
1856 }
1857 if (!thread_number)
1858 return 0;
1859
1860 if (write_bw_log) {
1861 setup_log(&agg_io_log[DDIR_READ], 0);
1862 setup_log(&agg_io_log[DDIR_WRITE], 0);
6eaf09d6 1863 setup_log(&agg_io_log[DDIR_TRIM], 0);
2e1df07d
JA		 1864 }
1865
521da527 1866 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2e1df07d
JA		 1867 if (startup_mutex == NULL)
1868 return 1;
521da527 1869 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2e1df07d
JA		 1870 if (writeout_mutex == NULL)
1871 return 1;
1872
1873 set_genesis_time();
1874 create_disk_util_thread();
1875
1876 cgroup_list = smalloc(sizeof(*cgroup_list));
1877 INIT_FLIST_HEAD(cgroup_list);
1878
1879 run_threads();
1880
1881 if (!fio_abort) {
1882 show_run_stats();
1883 if (write_bw_log) {
1884 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1885 __finish_log(agg_io_log[DDIR_WRITE],
1886 "agg-write_bw.log");
6eaf09d6
SL		 1887 __finish_log(agg_io_log[DDIR_TRIM],
1888 "agg-write_bw.log");
2e1df07d
JA		 1889 }
1890 }
1891
1892 for_each_td(td, i)
1893 fio_options_free(td);
1894
a462baef 1895 free_disk_util();
2e1df07d
JA		 1896 cgroup_kill(cgroup_list);
1897 sfree(cgroup_list);
1898 sfree(cgroup_mnt);
1899
1900 fio_mutex_remove(startup_mutex);
1901 fio_mutex_remove(writeout_mutex);
9ec7779f 1902 fio_mutex_remove(disk_thread_mutex);
2e1df07d
JA		 1903 return exit_value;
1904}
fio - Flexible IO Tester