[PATCH] Missing memset and free-on-error in io engines
[fio.git] / fio.c
CommitLineData
ebac4655
JA
1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
aae22ca7 5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
ebac4655 6 *
8e9fe637
JA
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
9 *
ebac4655 10 * This program is free software; you can redistribute it and/or modify
8e9fe637
JA
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
ebac4655
JA
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 */
ebac4655
JA
24#include <unistd.h>
25#include <fcntl.h>
26#include <string.h>
ebac4655
JA
27#include <signal.h>
28#include <time.h>
ebac4655 29#include <assert.h>
ebac4655
JA
30#include <sys/stat.h>
31#include <sys/wait.h>
32#include <sys/ipc.h>
33#include <sys/shm.h>
34#include <sys/ioctl.h>
35#include <sys/mman.h>
36
37#include "fio.h"
38#include "os.h"
39
40#define MASK (4095)
41
42#define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
43
44int groupid = 0;
45int thread_number = 0;
ebac4655 46int shm_id = 0;
53cdc686 47int temp_stall_ts;
3c9b60c1 48const char *fio_inst_prefix = _INST_PREFIX;
ebac4655 49
bbfd6b00 50static volatile int startup_sem;
ebac4655
JA
51
52#define TERMINATE_ALL (-1)
75154845 53#define JOB_START_TIMEOUT (5 * 1000)
ebac4655
JA
54
55static void terminate_threads(int group_id)
56{
34572e28 57 struct thread_data *td;
ebac4655
JA
58 int i;
59
34572e28 60 for_each_td(td, i) {
ebac4655
JA
61 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
62 td->terminate = 1;
63 td->start_delay = 0;
64 }
65 }
66}
67
68static void sig_handler(int sig)
69{
70 switch (sig) {
71 case SIGALRM:
72 update_io_ticks();
73 disk_util_timer_arm();
74 print_thread_status();
75 break;
76 default:
77 printf("\nfio: terminating on signal\n");
78 fflush(stdout);
79 terminate_threads(TERMINATE_ALL);
80 break;
81 }
82}
83
906c8d75
JA
84/*
85 * Check if we are above the minimum rate given.
86 */
ebac4655
JA
87static int check_min_rate(struct thread_data *td, struct timeval *now)
88{
89 unsigned long spent;
90 unsigned long rate;
91 int ddir = td->ddir;
92
93 /*
94 * allow a 2 second settle period in the beginning
95 */
96 if (mtime_since(&td->start, now) < 2000)
97 return 0;
98
99 /*
100 * if rate blocks is set, sample is running
101 */
102 if (td->rate_bytes) {
103 spent = mtime_since(&td->lastrate, now);
104 if (spent < td->ratecycle)
105 return 0;
106
107 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
108 if (rate < td->ratemin) {
eb8bbf48 109 fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
ebac4655
JA
110 return 1;
111 }
112 }
113
114 td->rate_bytes = td->this_io_bytes[ddir];
115 memcpy(&td->lastrate, now, sizeof(*now));
116 return 0;
117}
118
119static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
120{
121 if (!td->timeout)
122 return 0;
123 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
124 return 1;
125
126 return 0;
127}
128
ebac4655
JA
129static inline void td_set_runstate(struct thread_data *td, int runstate)
130{
ebac4655
JA
131 td->runstate = runstate;
132}
133
53cdc686
JA
134static struct fio_file *get_next_file(struct thread_data *td)
135{
0ab8db89 136 unsigned int old_next_file = td->next_file;
b2a15192
JA
137 struct fio_file *f;
138
139 do {
140 f = &td->files[td->next_file];
141
142 td->next_file++;
143 if (td->next_file >= td->nr_files)
144 td->next_file = 0;
145
146 if (f->fd != -1)
147 break;
53cdc686 148
b2a15192
JA
149 f = NULL;
150 } while (td->next_file != old_next_file);
53cdc686
JA
151
152 return f;
153}
154
906c8d75
JA
155/*
156 * When job exits, we can cancel the in-flight IO if we are using async
157 * io. Attempt to do so.
158 */
ebac4655
JA
159static void cleanup_pending_aio(struct thread_data *td)
160{
161 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
162 struct list_head *entry, *n;
163 struct io_completion_data icd;
164 struct io_u *io_u;
165 int r;
166
167 /*
168 * get immediately available events, if any
169 */
45bee283 170 r = td_io_getevents(td, 0, td->cur_depth, &ts);
ebac4655
JA
171 if (r > 0) {
172 icd.nr = r;
173 ios_completed(td, &icd);
174 }
175
176 /*
177 * now cancel remaining active events
178 */
2866c82d 179 if (td->io_ops->cancel) {
ebac4655
JA
180 list_for_each_safe(entry, n, &td->io_u_busylist) {
181 io_u = list_entry(entry, struct io_u, list);
182
2866c82d 183 r = td->io_ops->cancel(td, io_u);
ebac4655
JA
184 if (!r)
185 put_io_u(td, io_u);
186 }
187 }
188
189 if (td->cur_depth) {
45bee283 190 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
ebac4655
JA
191 if (r > 0) {
192 icd.nr = r;
193 ios_completed(td, &icd);
194 }
195 }
196}
197
858a3d47
JA
198/*
199 * Helper to handle the final sync of a file. Works just like the normal
200 * io path, just does everything sync.
201 */
202static int fio_io_sync(struct thread_data *td, struct fio_file *f)
203{
204 struct io_u *io_u = __get_io_u(td);
205 struct io_completion_data icd;
206 int ret;
207
208 if (!io_u)
209 return 1;
210
211 io_u->ddir = DDIR_SYNC;
212 io_u->file = f;
213
214 if (td_io_prep(td, io_u)) {
215 put_io_u(td, io_u);
216 return 1;
217 }
218
219 ret = td_io_queue(td, io_u);
220 if (ret) {
353a7e0e 221 td_verror(td, io_u->error);
858a3d47 222 put_io_u(td, io_u);
858a3d47
JA
223 return 1;
224 }
225
226 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
227 if (ret < 0) {
353a7e0e 228 td_verror(td, ret);
858a3d47
JA
229 return 1;
230 }
231
232 icd.nr = ret;
233 ios_completed(td, &icd);
234 if (icd.error) {
235 td_verror(td, icd.error);
236 return 1;
237 }
238
239 return 0;
240}
241
906c8d75
JA
242/*
243 * The main verify engine. Runs over the writes we previusly submitted,
244 * reads the blocks back in, and checks the crc/md5 of the data.
245 */
a9619d44 246void do_verify(struct thread_data *td)
ebac4655
JA
247{
248 struct timeval t;
249 struct io_u *io_u, *v_io_u = NULL;
250 struct io_completion_data icd;
53cdc686 251 struct fio_file *f;
e5b401d4
JA
252 int ret, i;
253
254 /*
255 * sync io first and invalidate cache, to make sure we really
256 * read from disk.
257 */
258 for_each_file(td, f, i) {
858a3d47 259 fio_io_sync(td, f);
e5b401d4
JA
260 file_invalidate_cache(td, f);
261 }
ebac4655
JA
262
263 td_set_runstate(td, TD_VERIFYING);
264
265 do {
266 if (td->terminate)
267 break;
268
269 gettimeofday(&t, NULL);
270 if (runtime_exceeded(td, &t))
271 break;
272
273 io_u = __get_io_u(td);
274 if (!io_u)
275 break;
276
aea47d44 277 if (get_next_verify(td, io_u)) {
ebac4655
JA
278 put_io_u(td, io_u);
279 break;
280 }
281
53cdc686
JA
282 f = get_next_file(td);
283 if (!f)
284 break;
285
286 io_u->file = f;
287
aea47d44 288 if (td_io_prep(td, io_u)) {
ebac4655
JA
289 put_io_u(td, io_u);
290 break;
291 }
292
45bee283 293 ret = td_io_queue(td, io_u);
ebac4655 294 if (ret) {
353a7e0e 295 td_verror(td, io_u->error);
ebac4655 296 put_io_u(td, io_u);
ebac4655
JA
297 break;
298 }
299
300 /*
301 * we have one pending to verify, do that while
302 * we are doing io on the next one
303 */
304 if (do_io_u_verify(td, &v_io_u))
305 break;
306
45bee283 307 ret = td_io_getevents(td, 1, 1, NULL);
ebac4655
JA
308 if (ret != 1) {
309 if (ret < 0)
310 td_verror(td, ret);
311 break;
312 }
313
2866c82d 314 v_io_u = td->io_ops->event(td, 0);
ebac4655
JA
315 icd.nr = 1;
316 icd.error = 0;
317 io_completed(td, v_io_u, &icd);
318
319 if (icd.error) {
320 td_verror(td, icd.error);
321 put_io_u(td, v_io_u);
322 v_io_u = NULL;
323 break;
324 }
325
ebac4655
JA
326 /*
327 * if we can't submit more io, we need to verify now
328 */
329 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
330 break;
331
332 } while (1);
333
334 do_io_u_verify(td, &v_io_u);
335
336 if (td->cur_depth)
337 cleanup_pending_aio(td);
338
339 td_set_runstate(td, TD_RUNNING);
340}
341
b990b5c0
JA
342/*
343 * Not really an io thread, all it does is burn CPU cycles in the specified
344 * manner.
345 */
346static void do_cpuio(struct thread_data *td)
347{
348 struct timeval e;
349 int split = 100 / td->cpuload;
350 int i = 0;
351
352 while (!td->terminate) {
353 gettimeofday(&e, NULL);
354
355 if (runtime_exceeded(td, &e))
356 break;
357
358 if (!(i % split))
359 __usec_sleep(10000);
360 else
361 usec_sleep(td, 10000);
362
363 i++;
364 }
365}
366
32cd46a0 367/*
906c8d75 368 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
369 * and reaps them, checking for rate and errors along the way.
370 */
ebac4655
JA
371static void do_io(struct thread_data *td)
372{
373 struct io_completion_data icd;
374 struct timeval s, e;
375 unsigned long usec;
53cdc686 376 struct fio_file *f;
84585003 377 int i, ret = 0;
ebac4655 378
5853e5a8
JA
379 td_set_runstate(td, TD_RUNNING);
380
ebac4655
JA
381 while (td->this_io_bytes[td->ddir] < td->io_size) {
382 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
383 struct timespec *timeout;
84585003 384 int min_evts = 0;
ebac4655
JA
385 struct io_u *io_u;
386
387 if (td->terminate)
388 break;
389
53cdc686
JA
390 f = get_next_file(td);
391 if (!f)
392 break;
393
394 io_u = get_io_u(td, f);
ebac4655
JA
395 if (!io_u)
396 break;
397
398 memcpy(&s, &io_u->start_time, sizeof(s));
399
45bee283 400 ret = td_io_queue(td, io_u);
ebac4655 401 if (ret) {
353a7e0e 402 td_verror(td, io_u->error);
ebac4655 403 put_io_u(td, io_u);
ebac4655
JA
404 break;
405 }
406
407 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
408
409 if (td->cur_depth < td->iodepth) {
410 timeout = &ts;
411 min_evts = 0;
412 } else {
413 timeout = NULL;
414 min_evts = 1;
415 }
416
45bee283 417 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
ebac4655 418 if (ret < 0) {
87dc1ab1 419 td_verror(td, ret);
ebac4655
JA
420 break;
421 } else if (!ret)
422 continue;
423
424 icd.nr = ret;
425 ios_completed(td, &icd);
426 if (icd.error) {
427 td_verror(td, icd.error);
428 break;
429 }
430
431 /*
432 * the rate is batched for now, it should work for batches
433 * of completions except the very first one which may look
434 * a little bursty
435 */
436 gettimeofday(&e, NULL);
437 usec = utime_since(&s, &e);
438
a00735e6 439 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
ebac4655
JA
440
441 if (check_min_rate(td, &e)) {
2f9ade3c
JA
442 if (rate_quit)
443 terminate_threads(td->groupid);
ebac4655
JA
444 td_verror(td, ENOMEM);
445 break;
446 }
447
448 if (runtime_exceeded(td, &e))
449 break;
450
451 if (td->thinktime)
452 usec_sleep(td, td->thinktime);
ebac4655
JA
453 }
454
84585003
JA
455 if (!ret) {
456 if (td->cur_depth)
457 cleanup_pending_aio(td);
ebac4655 458
84585003
JA
459 if (should_fsync(td) && td->end_fsync) {
460 td_set_runstate(td, TD_FSYNCING);
461 for_each_file(td, f, i)
858a3d47 462 fio_io_sync(td, f);
84585003 463 }
5853e5a8 464 }
ebac4655
JA
465}
466
ebac4655
JA
467static void cleanup_io_u(struct thread_data *td)
468{
469 struct list_head *entry, *n;
470 struct io_u *io_u;
471
472 list_for_each_safe(entry, n, &td->io_u_freelist) {
473 io_u = list_entry(entry, struct io_u, list);
474
475 list_del(&io_u->list);
476 free(io_u);
477 }
478
2f9ade3c 479 free_io_mem(td);
ebac4655
JA
480}
481
6b9cea23
JA
482/*
483 * "randomly" fill the buffer contents
484 */
66eeb296 485static void fill_rand_buf(struct io_u *io_u, int max_bs)
6b9cea23 486{
66eeb296 487 int *ptr = io_u->buf;
6b9cea23
JA
488
489 while ((void *) ptr - io_u->buf < max_bs) {
490 *ptr = rand() * 0x9e370001;
491 ptr++;
492 }
493}
494
ebac4655
JA
495static int init_io_u(struct thread_data *td)
496{
497 struct io_u *io_u;
a00735e6 498 unsigned int max_bs;
ebac4655
JA
499 int i, max_units;
500 char *p;
501
2866c82d 502 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
503 return 0;
504
2866c82d 505 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
506 max_units = 1;
507 else
508 max_units = td->iodepth;
509
a00735e6
JA
510 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
511 td->orig_buffer_size = max_bs * max_units + MASK;
ebac4655 512
2f9ade3c
JA
513 if (allocate_io_mem(td))
514 return 1;
ebac4655 515
ebac4655
JA
516 p = ALIGN(td->orig_buffer);
517 for (i = 0; i < max_units; i++) {
518 io_u = malloc(sizeof(*io_u));
519 memset(io_u, 0, sizeof(*io_u));
520 INIT_LIST_HEAD(&io_u->list);
521
a00735e6 522 io_u->buf = p + max_bs * i;
6b9cea23 523 if (td_write(td) || td_rw(td))
a00735e6 524 fill_rand_buf(io_u, max_bs);
6b9cea23 525
b1ff3403 526 io_u->index = i;
ebac4655
JA
527 list_add(&io_u->list, &td->io_u_freelist);
528 }
529
530 return 0;
531}
532
da86774e
JA
533static int switch_ioscheduler(struct thread_data *td)
534{
535 char tmp[256], tmp2[128];
536 FILE *f;
537 int ret;
538
f48b467c
JA
539 if (td->io_ops->flags & FIO_CPUIO)
540 return 0;
541
da86774e
JA
542 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
543
544 f = fopen(tmp, "r+");
545 if (!f) {
546 td_verror(td, errno);
547 return 1;
548 }
549
550 /*
551 * Set io scheduler.
552 */
553 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
554 if (ferror(f) || ret != 1) {
555 td_verror(td, errno);
556 fclose(f);
557 return 1;
558 }
559
560 rewind(f);
561
562 /*
563 * Read back and check that the selected scheduler is now the default.
564 */
565 ret = fread(tmp, 1, sizeof(tmp), f);
566 if (ferror(f) || ret < 0) {
567 td_verror(td, errno);
568 fclose(f);
569 return 1;
570 }
571
572 sprintf(tmp2, "[%s]", td->ioscheduler);
573 if (!strstr(tmp, tmp2)) {
3b70d7e5 574 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
da86774e
JA
575 td_verror(td, EINVAL);
576 fclose(f);
577 return 1;
578 }
579
580 fclose(f);
581 return 0;
582}
583
ebac4655
JA
584static void clear_io_state(struct thread_data *td)
585{
53cdc686
JA
586 struct fio_file *f;
587 int i;
ebac4655 588
ebac4655
JA
589 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
590 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 591 td->zone_bytes = 0;
ebac4655 592
53cdc686
JA
593 for_each_file(td, f, i) {
594 f->last_pos = 0;
595 if (td->io_ops->flags & FIO_SYNCIO)
596 lseek(f->fd, SEEK_SET, 0);
597
598 if (f->file_map)
599 memset(f->file_map, 0, f->num_maps * sizeof(long));
600 }
ebac4655
JA
601}
602
906c8d75
JA
603/*
604 * Entry point for the thread based jobs. The process based jobs end up
605 * here as well, after a little setup.
606 */
ebac4655
JA
607static void *thread_main(void *data)
608{
609 struct thread_data *td = data;
ebac4655
JA
610
611 if (!td->use_thread)
612 setsid();
613
614 td->pid = getpid();
615
aea47d44
JA
616 INIT_LIST_HEAD(&td->io_u_freelist);
617 INIT_LIST_HEAD(&td->io_u_busylist);
618 INIT_LIST_HEAD(&td->io_hist_list);
619 INIT_LIST_HEAD(&td->io_log_list);
620
ebac4655
JA
621 if (init_io_u(td))
622 goto err;
623
624 if (fio_setaffinity(td) == -1) {
625 td_verror(td, errno);
626 goto err;
627 }
628
45bee283 629 if (td_io_init(td))
ebac4655
JA
630 goto err;
631
aea47d44
JA
632 if (init_iolog(td))
633 goto err;
634
ebac4655
JA
635 if (td->ioprio) {
636 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
637 td_verror(td, errno);
638 goto err;
639 }
640 }
641
1056eaad 642 if (nice(td->nice) == -1) {
b6f4d880
JA
643 td_verror(td, errno);
644 goto err;
645 }
646
75154845
JA
647 if (init_random_state(td))
648 goto err;
649
56f9498d
JA
650 if (td->ioscheduler && switch_ioscheduler(td))
651 goto err;
da86774e 652
75154845 653 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
654 fio_sem_up(&startup_sem);
655 fio_sem_down(&td->mutex);
ebac4655 656
53cdc686 657 if (!td->create_serialize && setup_files(td))
ebac4655
JA
658 goto err;
659
ebac4655
JA
660 gettimeofday(&td->epoch, NULL);
661
56f9498d 662 if (td->exec_prerun)
4e0ba8af
JA
663 system(td->exec_prerun);
664
ebac4655
JA
665 while (td->loops--) {
666 getrusage(RUSAGE_SELF, &td->ru_start);
667 gettimeofday(&td->start, NULL);
668 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
669
670 if (td->ratemin)
671 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
672
673 clear_io_state(td);
674 prune_io_piece_log(td);
675
2866c82d 676 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
677 do_cpuio(td);
678 else
679 do_io(td);
ebac4655
JA
680
681 td->runtime[td->ddir] += mtime_since_now(&td->start);
aea47d44 682 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
3d60d1ed
JA
683 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
684
ebac4655
JA
685 update_rusage_stat(td);
686
687 if (td->error || td->terminate)
688 break;
689
690 if (td->verify == VERIFY_NONE)
691 continue;
692
693 clear_io_state(td);
694 gettimeofday(&td->start, NULL);
695
696 do_verify(td);
697
698 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
699
700 if (td->error || td->terminate)
701 break;
702 }
703
ebac4655
JA
704 if (td->bw_log)
705 finish_log(td, td->bw_log, "bw");
706 if (td->slat_log)
707 finish_log(td, td->slat_log, "slat");
708 if (td->clat_log)
709 finish_log(td, td->clat_log, "clat");
076efc7c 710 if (td->write_iolog_file)
843a7413 711 write_iolog_close(td);
56f9498d 712 if (td->exec_postrun)
4e0ba8af 713 system(td->exec_postrun);
ebac4655
JA
714
715 if (exitall_on_terminate)
716 terminate_threads(td->groupid);
717
718err:
53cdc686 719 close_files(td);
2866c82d 720 close_ioengine(td);
ebac4655 721 cleanup_io_u(td);
ebac4655
JA
722 td_set_runstate(td, TD_EXITED);
723 return NULL;
724
725}
726
906c8d75
JA
727/*
728 * We cannot pass the td data into a forked process, so attach the td and
729 * pass it to the thread worker.
730 */
ebac4655
JA
731static void *fork_main(int shmid, int offset)
732{
733 struct thread_data *td;
734 void *data;
735
736 data = shmat(shmid, NULL, 0);
737 if (data == (void *) -1) {
738 perror("shmat");
739 return NULL;
740 }
741
742 td = data + offset * sizeof(struct thread_data);
743 thread_main(td);
744 shmdt(data);
745 return NULL;
746}
747
906c8d75
JA
748/*
749 * Run over the job map and reap the threads that have exited, if any.
750 */
ebac4655
JA
751static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
752{
34572e28 753 struct thread_data *td;
b990b5c0 754 int i, cputhreads;
ebac4655
JA
755
756 /*
757 * reap exited threads (TD_EXITED -> TD_REAPED)
758 */
34572e28
JA
759 cputhreads = 0;
760 for_each_td(td, i) {
84585003
JA
761 /*
762 * ->io_ops is NULL for a thread that has closed its
763 * io engine
764 */
765 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
766 cputhreads++;
767
ebac4655
JA
768 if (td->runstate != TD_EXITED)
769 continue;
770
771 td_set_runstate(td, TD_REAPED);
772
773 if (td->use_thread) {
774 long ret;
775
776 if (pthread_join(td->thread, (void *) &ret))
777 perror("thread_join");
778 } else
779 waitpid(td->pid, NULL, 0);
780
781 (*nr_running)--;
782 (*m_rate) -= td->ratemin;
783 (*t_rate) -= td->rate;
784 }
b990b5c0
JA
785
786 if (*nr_running == cputhreads)
787 terminate_threads(TERMINATE_ALL);
ebac4655
JA
788}
789
906c8d75
JA
790/*
791 * Main function for kicking off and reaping jobs, as needed.
792 */
ebac4655
JA
793static void run_threads(void)
794{
ebac4655
JA
795 struct thread_data *td;
796 unsigned long spent;
797 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 798
2f9ade3c
JA
799 if (fio_pin_memory())
800 return;
ebac4655 801
c6ae0a5b
JA
802 if (!terse_output) {
803 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
804 fflush(stdout);
805 }
c04f7ec3 806
4efa970e
JA
807 signal(SIGINT, sig_handler);
808 signal(SIGALRM, sig_handler);
809
ebac4655
JA
810 todo = thread_number;
811 nr_running = 0;
812 nr_started = 0;
813 m_rate = t_rate = 0;
814
34572e28 815 for_each_td(td, i) {
263e529f 816 print_status_init(td->thread_number - 1);
ebac4655
JA
817
818 init_disk_util(td);
819
820 if (!td->create_serialize)
821 continue;
822
823 /*
824 * do file setup here so it happens sequentially,
825 * we don't want X number of threads getting their
826 * client data interspersed on disk
827 */
53cdc686 828 if (setup_files(td)) {
ebac4655
JA
829 td_set_runstate(td, TD_REAPED);
830 todo--;
831 }
832 }
833
263e529f 834 time_init();
ebac4655
JA
835
836 while (todo) {
75154845
JA
837 struct thread_data *map[MAX_JOBS];
838 struct timeval this_start;
839 int this_jobs = 0, left;
840
ebac4655
JA
841 /*
842 * create threads (TD_NOT_CREATED -> TD_CREATED)
843 */
34572e28 844 for_each_td(td, i) {
ebac4655
JA
845 if (td->runstate != TD_NOT_CREATED)
846 continue;
847
848 /*
849 * never got a chance to start, killed by other
850 * thread for some reason
851 */
852 if (td->terminate) {
853 todo--;
854 continue;
855 }
856
857 if (td->start_delay) {
263e529f 858 spent = mtime_since_genesis();
ebac4655
JA
859
860 if (td->start_delay * 1000 > spent)
861 continue;
862 }
863
864 if (td->stonewall && (nr_started || nr_running))
865 break;
866
75154845
JA
867 /*
868 * Set state to created. Thread will transition
869 * to TD_INITIALIZED when it's done setting up.
870 */
ebac4655 871 td_set_runstate(td, TD_CREATED);
75154845 872 map[this_jobs++] = td;
bbfd6b00 873 fio_sem_init(&startup_sem, 1);
ebac4655
JA
874 nr_started++;
875
876 if (td->use_thread) {
877 if (pthread_create(&td->thread, NULL, thread_main, td)) {
878 perror("thread_create");
879 nr_started--;
880 }
881 } else {
882 if (fork())
bbfd6b00 883 fio_sem_down(&startup_sem);
ebac4655
JA
884 else {
885 fork_main(shm_id, i);
886 exit(0);
887 }
888 }
889 }
890
891 /*
75154845
JA
892 * Wait for the started threads to transition to
893 * TD_INITIALIZED.
ebac4655 894 */
75154845
JA
895 gettimeofday(&this_start, NULL);
896 left = this_jobs;
897 while (left) {
898 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
899 break;
900
901 usleep(100000);
902
903 for (i = 0; i < this_jobs; i++) {
904 td = map[i];
905 if (!td)
906 continue;
b6f4d880 907 if (td->runstate == TD_INITIALIZED) {
75154845
JA
908 map[i] = NULL;
909 left--;
b6f4d880
JA
910 } else if (td->runstate >= TD_EXITED) {
911 map[i] = NULL;
912 left--;
913 todo--;
914 nr_running++; /* work-around... */
75154845
JA
915 }
916 }
917 }
918
919 if (left) {
3b70d7e5 920 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
921 for (i = 0; i < this_jobs; i++) {
922 td = map[i];
923 if (!td)
924 continue;
925 kill(td->pid, SIGTERM);
926 }
927 break;
928 }
929
930 /*
b6f4d880 931 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 932 */
34572e28 933 for_each_td(td, i) {
75154845 934 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
935 continue;
936
937 td_set_runstate(td, TD_RUNNING);
938 nr_running++;
939 nr_started--;
940 m_rate += td->ratemin;
941 t_rate += td->rate;
75154845 942 todo--;
bbfd6b00 943 fio_sem_up(&td->mutex);
ebac4655
JA
944 }
945
946 reap_threads(&nr_running, &t_rate, &m_rate);
947
948 if (todo)
949 usleep(100000);
950 }
951
952 while (nr_running) {
953 reap_threads(&nr_running, &t_rate, &m_rate);
954 usleep(10000);
955 }
956
957 update_io_ticks();
2f9ade3c 958 fio_unpin_memory();
ebac4655
JA
959}
960
ebac4655
JA
961int main(int argc, char *argv[])
962{
963 if (parse_options(argc, argv))
964 return 1;
965
966 if (!thread_number) {
3b70d7e5 967 log_err("Nothing to do\n");
ebac4655
JA
968 return 1;
969 }
970
971 disk_util_timer_arm();
972
973 run_threads();
974 show_run_stats();
975
976 return 0;
977}