[PATCH] Time and seek optimizations
[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;
ebac4655 48
bbfd6b00 49static volatile int startup_sem;
ebac4655
JA
50
51#define TERMINATE_ALL (-1)
75154845 52#define JOB_START_TIMEOUT (5 * 1000)
ebac4655
JA
53
54static void terminate_threads(int group_id)
55{
34572e28 56 struct thread_data *td;
ebac4655
JA
57 int i;
58
34572e28 59 for_each_td(td, i) {
ebac4655
JA
60 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
61 td->terminate = 1;
62 td->start_delay = 0;
63 }
64 }
65}
66
67static void sig_handler(int sig)
68{
69 switch (sig) {
70 case SIGALRM:
71 update_io_ticks();
72 disk_util_timer_arm();
73 print_thread_status();
74 break;
75 default:
76 printf("\nfio: terminating on signal\n");
77 fflush(stdout);
78 terminate_threads(TERMINATE_ALL);
79 break;
80 }
81}
82
906c8d75
JA
83/*
84 * Check if we are above the minimum rate given.
85 */
ebac4655
JA
86static int check_min_rate(struct thread_data *td, struct timeval *now)
87{
88 unsigned long spent;
89 unsigned long rate;
90 int ddir = td->ddir;
91
92 /*
93 * allow a 2 second settle period in the beginning
94 */
95 if (mtime_since(&td->start, now) < 2000)
96 return 0;
97
98 /*
99 * if rate blocks is set, sample is running
100 */
101 if (td->rate_bytes) {
102 spent = mtime_since(&td->lastrate, now);
103 if (spent < td->ratecycle)
104 return 0;
105
106 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
107 if (rate < td->ratemin) {
eb8bbf48 108 fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
ebac4655
JA
109 return 1;
110 }
111 }
112
113 td->rate_bytes = td->this_io_bytes[ddir];
114 memcpy(&td->lastrate, now, sizeof(*now));
115 return 0;
116}
117
118static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
119{
120 if (!td->timeout)
121 return 0;
122 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
123 return 1;
124
125 return 0;
126}
127
ebac4655
JA
128static inline void td_set_runstate(struct thread_data *td, int runstate)
129{
ebac4655
JA
130 td->runstate = runstate;
131}
132
53cdc686
JA
133static struct fio_file *get_next_file(struct thread_data *td)
134{
0ab8db89 135 unsigned int old_next_file = td->next_file;
b2a15192
JA
136 struct fio_file *f;
137
138 do {
139 f = &td->files[td->next_file];
140
141 td->next_file++;
142 if (td->next_file >= td->nr_files)
143 td->next_file = 0;
144
145 if (f->fd != -1)
146 break;
53cdc686 147
b2a15192
JA
148 f = NULL;
149 } while (td->next_file != old_next_file);
53cdc686
JA
150
151 return f;
152}
153
906c8d75
JA
154/*
155 * When job exits, we can cancel the in-flight IO if we are using async
156 * io. Attempt to do so.
157 */
ebac4655
JA
158static void cleanup_pending_aio(struct thread_data *td)
159{
160 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
161 struct list_head *entry, *n;
162 struct io_completion_data icd;
163 struct io_u *io_u;
164 int r;
165
166 /*
167 * get immediately available events, if any
168 */
45bee283 169 r = td_io_getevents(td, 0, td->cur_depth, &ts);
ebac4655
JA
170 if (r > 0) {
171 icd.nr = r;
172 ios_completed(td, &icd);
173 }
174
175 /*
176 * now cancel remaining active events
177 */
2866c82d 178 if (td->io_ops->cancel) {
ebac4655
JA
179 list_for_each_safe(entry, n, &td->io_u_busylist) {
180 io_u = list_entry(entry, struct io_u, list);
181
2866c82d 182 r = td->io_ops->cancel(td, io_u);
ebac4655
JA
183 if (!r)
184 put_io_u(td, io_u);
185 }
186 }
187
188 if (td->cur_depth) {
45bee283 189 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
ebac4655
JA
190 if (r > 0) {
191 icd.nr = r;
192 ios_completed(td, &icd);
193 }
194 }
195}
196
858a3d47
JA
197/*
198 * Helper to handle the final sync of a file. Works just like the normal
199 * io path, just does everything sync.
200 */
201static int fio_io_sync(struct thread_data *td, struct fio_file *f)
202{
203 struct io_u *io_u = __get_io_u(td);
204 struct io_completion_data icd;
205 int ret;
206
207 if (!io_u)
208 return 1;
209
210 io_u->ddir = DDIR_SYNC;
211 io_u->file = f;
212
213 if (td_io_prep(td, io_u)) {
214 put_io_u(td, io_u);
215 return 1;
216 }
217
218 ret = td_io_queue(td, io_u);
219 if (ret) {
353a7e0e 220 td_verror(td, io_u->error);
858a3d47 221 put_io_u(td, io_u);
858a3d47
JA
222 return 1;
223 }
224
225 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
226 if (ret < 0) {
353a7e0e 227 td_verror(td, ret);
858a3d47
JA
228 return 1;
229 }
230
231 icd.nr = ret;
232 ios_completed(td, &icd);
233 if (icd.error) {
234 td_verror(td, icd.error);
235 return 1;
236 }
237
238 return 0;
239}
240
906c8d75
JA
241/*
242 * The main verify engine. Runs over the writes we previusly submitted,
243 * reads the blocks back in, and checks the crc/md5 of the data.
244 */
a9619d44 245void do_verify(struct thread_data *td)
ebac4655 246{
ebac4655
JA
247 struct io_u *io_u, *v_io_u = NULL;
248 struct io_completion_data icd;
53cdc686 249 struct fio_file *f;
e5b401d4
JA
250 int ret, i;
251
252 /*
253 * sync io first and invalidate cache, to make sure we really
254 * read from disk.
255 */
256 for_each_file(td, f, i) {
858a3d47 257 fio_io_sync(td, f);
e5b401d4
JA
258 file_invalidate_cache(td, f);
259 }
ebac4655
JA
260
261 td_set_runstate(td, TD_VERIFYING);
262
263 do {
264 if (td->terminate)
265 break;
266
ebac4655
JA
267 io_u = __get_io_u(td);
268 if (!io_u)
269 break;
270
02bcaa8c
JA
271 if (runtime_exceeded(td, &io_u->start_time)) {
272 put_io_u(td, io_u);
273 break;
274 }
275
aea47d44 276 if (get_next_verify(td, io_u)) {
ebac4655
JA
277 put_io_u(td, io_u);
278 break;
279 }
280
53cdc686
JA
281 f = get_next_file(td);
282 if (!f)
283 break;
284
285 io_u->file = f;
286
aea47d44 287 if (td_io_prep(td, io_u)) {
ebac4655
JA
288 put_io_u(td, io_u);
289 break;
290 }
291
45bee283 292 ret = td_io_queue(td, io_u);
ebac4655 293 if (ret) {
353a7e0e 294 td_verror(td, io_u->error);
ebac4655 295 put_io_u(td, io_u);
ebac4655
JA
296 break;
297 }
298
299 /*
300 * we have one pending to verify, do that while
301 * we are doing io on the next one
302 */
303 if (do_io_u_verify(td, &v_io_u))
304 break;
305
45bee283 306 ret = td_io_getevents(td, 1, 1, NULL);
ebac4655
JA
307 if (ret != 1) {
308 if (ret < 0)
309 td_verror(td, ret);
310 break;
311 }
312
2866c82d 313 v_io_u = td->io_ops->event(td, 0);
ebac4655
JA
314 icd.nr = 1;
315 icd.error = 0;
02bcaa8c 316 fio_gettime(&icd.time, NULL);
ebac4655
JA
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) {
02bcaa8c 353 fio_gettime(&e, NULL);
b990b5c0
JA
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;
02bcaa8c 374 struct timeval s;
ebac4655 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 */
02bcaa8c 436 usec = utime_since(&s, &icd.time);
ebac4655 437
a00735e6 438 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
ebac4655 439
02bcaa8c 440 if (check_min_rate(td, &icd.time)) {
98aa62d8 441 if (exitall_on_terminate)
2f9ade3c 442 terminate_threads(td->groupid);
ebac4655
JA
443 td_verror(td, ENOMEM);
444 break;
445 }
446
02bcaa8c 447 if (runtime_exceeded(td, &icd.time))
ebac4655
JA
448 break;
449
450 if (td->thinktime)
451 usec_sleep(td, td->thinktime);
ebac4655
JA
452 }
453
4d2413c6 454 if (!td->error) {
84585003
JA
455 if (td->cur_depth)
456 cleanup_pending_aio(td);
ebac4655 457
84585003
JA
458 if (should_fsync(td) && td->end_fsync) {
459 td_set_runstate(td, TD_FSYNCING);
460 for_each_file(td, f, i)
858a3d47 461 fio_io_sync(td, f);
84585003 462 }
5853e5a8 463 }
ebac4655
JA
464}
465
ebac4655
JA
466static void cleanup_io_u(struct thread_data *td)
467{
468 struct list_head *entry, *n;
469 struct io_u *io_u;
470
471 list_for_each_safe(entry, n, &td->io_u_freelist) {
472 io_u = list_entry(entry, struct io_u, list);
473
474 list_del(&io_u->list);
475 free(io_u);
476 }
477
2f9ade3c 478 free_io_mem(td);
ebac4655
JA
479}
480
6b9cea23
JA
481/*
482 * "randomly" fill the buffer contents
483 */
66eeb296 484static void fill_rand_buf(struct io_u *io_u, int max_bs)
6b9cea23 485{
66eeb296 486 int *ptr = io_u->buf;
6b9cea23
JA
487
488 while ((void *) ptr - io_u->buf < max_bs) {
489 *ptr = rand() * 0x9e370001;
490 ptr++;
491 }
492}
493
ebac4655
JA
494static int init_io_u(struct thread_data *td)
495{
496 struct io_u *io_u;
a00735e6 497 unsigned int max_bs;
ebac4655
JA
498 int i, max_units;
499 char *p;
500
2866c82d 501 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
502 return 0;
503
2866c82d 504 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
505 max_units = 1;
506 else
507 max_units = td->iodepth;
508
a00735e6
JA
509 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
510 td->orig_buffer_size = max_bs * max_units + MASK;
ebac4655 511
2f9ade3c
JA
512 if (allocate_io_mem(td))
513 return 1;
ebac4655 514
ebac4655
JA
515 p = ALIGN(td->orig_buffer);
516 for (i = 0; i < max_units; i++) {
517 io_u = malloc(sizeof(*io_u));
518 memset(io_u, 0, sizeof(*io_u));
519 INIT_LIST_HEAD(&io_u->list);
520
a00735e6 521 io_u->buf = p + max_bs * i;
6b9cea23 522 if (td_write(td) || td_rw(td))
a00735e6 523 fill_rand_buf(io_u, max_bs);
6b9cea23 524
b1ff3403 525 io_u->index = i;
ebac4655
JA
526 list_add(&io_u->list, &td->io_u_freelist);
527 }
528
529 return 0;
530}
531
da86774e
JA
532static int switch_ioscheduler(struct thread_data *td)
533{
534 char tmp[256], tmp2[128];
535 FILE *f;
536 int ret;
537
f48b467c
JA
538 if (td->io_ops->flags & FIO_CPUIO)
539 return 0;
540
da86774e
JA
541 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
542
543 f = fopen(tmp, "r+");
544 if (!f) {
545 td_verror(td, errno);
546 return 1;
547 }
548
549 /*
550 * Set io scheduler.
551 */
552 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
553 if (ferror(f) || ret != 1) {
554 td_verror(td, errno);
555 fclose(f);
556 return 1;
557 }
558
559 rewind(f);
560
561 /*
562 * Read back and check that the selected scheduler is now the default.
563 */
564 ret = fread(tmp, 1, sizeof(tmp), f);
565 if (ferror(f) || ret < 0) {
566 td_verror(td, errno);
567 fclose(f);
568 return 1;
569 }
570
571 sprintf(tmp2, "[%s]", td->ioscheduler);
572 if (!strstr(tmp, tmp2)) {
3b70d7e5 573 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
da86774e
JA
574 td_verror(td, EINVAL);
575 fclose(f);
576 return 1;
577 }
578
579 fclose(f);
580 return 0;
581}
582
ebac4655
JA
583static void clear_io_state(struct thread_data *td)
584{
53cdc686
JA
585 struct fio_file *f;
586 int i;
ebac4655 587
ebac4655
JA
588 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
589 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 590 td->zone_bytes = 0;
ebac4655 591
53cdc686
JA
592 for_each_file(td, f, i) {
593 f->last_pos = 0;
594 if (td->io_ops->flags & FIO_SYNCIO)
595 lseek(f->fd, SEEK_SET, 0);
596
597 if (f->file_map)
598 memset(f->file_map, 0, f->num_maps * sizeof(long));
599 }
ebac4655
JA
600}
601
906c8d75
JA
602/*
603 * Entry point for the thread based jobs. The process based jobs end up
604 * here as well, after a little setup.
605 */
ebac4655
JA
606static void *thread_main(void *data)
607{
608 struct thread_data *td = data;
ebac4655
JA
609
610 if (!td->use_thread)
611 setsid();
612
613 td->pid = getpid();
614
aea47d44
JA
615 INIT_LIST_HEAD(&td->io_u_freelist);
616 INIT_LIST_HEAD(&td->io_u_busylist);
617 INIT_LIST_HEAD(&td->io_hist_list);
618 INIT_LIST_HEAD(&td->io_log_list);
619
ebac4655
JA
620 if (init_io_u(td))
621 goto err;
622
623 if (fio_setaffinity(td) == -1) {
624 td_verror(td, errno);
625 goto err;
626 }
627
45bee283 628 if (td_io_init(td))
ebac4655
JA
629 goto err;
630
aea47d44
JA
631 if (init_iolog(td))
632 goto err;
633
ebac4655
JA
634 if (td->ioprio) {
635 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
636 td_verror(td, errno);
637 goto err;
638 }
639 }
640
1056eaad 641 if (nice(td->nice) == -1) {
b6f4d880
JA
642 td_verror(td, errno);
643 goto err;
644 }
645
75154845
JA
646 if (init_random_state(td))
647 goto err;
648
56f9498d
JA
649 if (td->ioscheduler && switch_ioscheduler(td))
650 goto err;
da86774e 651
75154845 652 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
653 fio_sem_up(&startup_sem);
654 fio_sem_down(&td->mutex);
ebac4655 655
53cdc686 656 if (!td->create_serialize && setup_files(td))
ebac4655 657 goto err;
21972cde
JA
658 if (open_files(td))
659 goto err;
ebac4655 660
02bcaa8c 661 fio_gettime(&td->epoch, NULL);
ebac4655 662
56f9498d 663 if (td->exec_prerun)
4e0ba8af
JA
664 system(td->exec_prerun);
665
ebac4655
JA
666 while (td->loops--) {
667 getrusage(RUSAGE_SELF, &td->ru_start);
02bcaa8c 668 fio_gettime(&td->start, NULL);
ebac4655
JA
669 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
670
671 if (td->ratemin)
672 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
673
674 clear_io_state(td);
675 prune_io_piece_log(td);
676
2866c82d 677 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
678 do_cpuio(td);
679 else
680 do_io(td);
ebac4655
JA
681
682 td->runtime[td->ddir] += mtime_since_now(&td->start);
aea47d44 683 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
3d60d1ed
JA
684 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
685
ebac4655
JA
686 update_rusage_stat(td);
687
688 if (td->error || td->terminate)
689 break;
690
691 if (td->verify == VERIFY_NONE)
692 continue;
693
694 clear_io_state(td);
02bcaa8c 695 fio_gettime(&td->start, NULL);
ebac4655
JA
696
697 do_verify(td);
698
699 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
700
701 if (td->error || td->terminate)
702 break;
703 }
704
ebac4655
JA
705 if (td->bw_log)
706 finish_log(td, td->bw_log, "bw");
707 if (td->slat_log)
708 finish_log(td, td->slat_log, "slat");
709 if (td->clat_log)
710 finish_log(td, td->clat_log, "clat");
076efc7c 711 if (td->write_iolog_file)
843a7413 712 write_iolog_close(td);
56f9498d 713 if (td->exec_postrun)
4e0ba8af 714 system(td->exec_postrun);
ebac4655
JA
715
716 if (exitall_on_terminate)
717 terminate_threads(td->groupid);
718
719err:
53cdc686 720 close_files(td);
2866c82d 721 close_ioengine(td);
ebac4655 722 cleanup_io_u(td);
ebac4655
JA
723 td_set_runstate(td, TD_EXITED);
724 return NULL;
725
726}
727
906c8d75
JA
728/*
729 * We cannot pass the td data into a forked process, so attach the td and
730 * pass it to the thread worker.
731 */
ebac4655
JA
732static void *fork_main(int shmid, int offset)
733{
734 struct thread_data *td;
735 void *data;
736
737 data = shmat(shmid, NULL, 0);
738 if (data == (void *) -1) {
739 perror("shmat");
740 return NULL;
741 }
742
743 td = data + offset * sizeof(struct thread_data);
744 thread_main(td);
745 shmdt(data);
746 return NULL;
747}
748
906c8d75
JA
749/*
750 * Run over the job map and reap the threads that have exited, if any.
751 */
ebac4655
JA
752static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
753{
34572e28 754 struct thread_data *td;
4d2413c6 755 int i, cputhreads, pending;
ebac4655
JA
756
757 /*
758 * reap exited threads (TD_EXITED -> TD_REAPED)
759 */
4d2413c6 760 pending = cputhreads = 0;
34572e28 761 for_each_td(td, i) {
84585003
JA
762 /*
763 * ->io_ops is NULL for a thread that has closed its
764 * io engine
765 */
766 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
767 cputhreads++;
768
4d2413c6
JA
769 if (td->runstate != TD_EXITED) {
770 if (td->runstate < TD_RUNNING)
771 pending++;
772
ebac4655 773 continue;
4d2413c6 774 }
ebac4655
JA
775
776 td_set_runstate(td, TD_REAPED);
777
778 if (td->use_thread) {
779 long ret;
780
781 if (pthread_join(td->thread, (void *) &ret))
782 perror("thread_join");
783 } else
784 waitpid(td->pid, NULL, 0);
785
786 (*nr_running)--;
787 (*m_rate) -= td->ratemin;
788 (*t_rate) -= td->rate;
789 }
b990b5c0 790
4d2413c6 791 if (*nr_running == cputhreads && !pending)
b990b5c0 792 terminate_threads(TERMINATE_ALL);
ebac4655
JA
793}
794
906c8d75
JA
795/*
796 * Main function for kicking off and reaping jobs, as needed.
797 */
ebac4655
JA
798static void run_threads(void)
799{
ebac4655
JA
800 struct thread_data *td;
801 unsigned long spent;
802 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 803
2f9ade3c
JA
804 if (fio_pin_memory())
805 return;
ebac4655 806
c6ae0a5b
JA
807 if (!terse_output) {
808 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
809 fflush(stdout);
810 }
c04f7ec3 811
4efa970e
JA
812 signal(SIGINT, sig_handler);
813 signal(SIGALRM, sig_handler);
814
ebac4655
JA
815 todo = thread_number;
816 nr_running = 0;
817 nr_started = 0;
818 m_rate = t_rate = 0;
819
34572e28 820 for_each_td(td, i) {
263e529f 821 print_status_init(td->thread_number - 1);
ebac4655
JA
822
823 init_disk_util(td);
824
825 if (!td->create_serialize)
826 continue;
827
828 /*
829 * do file setup here so it happens sequentially,
830 * we don't want X number of threads getting their
831 * client data interspersed on disk
832 */
53cdc686 833 if (setup_files(td)) {
ebac4655
JA
834 td_set_runstate(td, TD_REAPED);
835 todo--;
836 }
837 }
838
ebac4655 839 while (todo) {
75154845
JA
840 struct thread_data *map[MAX_JOBS];
841 struct timeval this_start;
842 int this_jobs = 0, left;
843
ebac4655
JA
844 /*
845 * create threads (TD_NOT_CREATED -> TD_CREATED)
846 */
34572e28 847 for_each_td(td, i) {
ebac4655
JA
848 if (td->runstate != TD_NOT_CREATED)
849 continue;
850
851 /*
852 * never got a chance to start, killed by other
853 * thread for some reason
854 */
855 if (td->terminate) {
856 todo--;
857 continue;
858 }
859
860 if (td->start_delay) {
263e529f 861 spent = mtime_since_genesis();
ebac4655
JA
862
863 if (td->start_delay * 1000 > spent)
864 continue;
865 }
866
867 if (td->stonewall && (nr_started || nr_running))
868 break;
869
75154845
JA
870 /*
871 * Set state to created. Thread will transition
872 * to TD_INITIALIZED when it's done setting up.
873 */
ebac4655 874 td_set_runstate(td, TD_CREATED);
75154845 875 map[this_jobs++] = td;
bbfd6b00 876 fio_sem_init(&startup_sem, 1);
ebac4655
JA
877 nr_started++;
878
879 if (td->use_thread) {
880 if (pthread_create(&td->thread, NULL, thread_main, td)) {
881 perror("thread_create");
882 nr_started--;
883 }
884 } else {
885 if (fork())
bbfd6b00 886 fio_sem_down(&startup_sem);
ebac4655
JA
887 else {
888 fork_main(shm_id, i);
889 exit(0);
890 }
891 }
892 }
893
894 /*
75154845
JA
895 * Wait for the started threads to transition to
896 * TD_INITIALIZED.
ebac4655 897 */
02bcaa8c 898 fio_gettime(&this_start, NULL);
75154845
JA
899 left = this_jobs;
900 while (left) {
901 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
902 break;
903
904 usleep(100000);
905
906 for (i = 0; i < this_jobs; i++) {
907 td = map[i];
908 if (!td)
909 continue;
b6f4d880 910 if (td->runstate == TD_INITIALIZED) {
75154845
JA
911 map[i] = NULL;
912 left--;
b6f4d880
JA
913 } else if (td->runstate >= TD_EXITED) {
914 map[i] = NULL;
915 left--;
916 todo--;
917 nr_running++; /* work-around... */
75154845
JA
918 }
919 }
920 }
921
922 if (left) {
3b70d7e5 923 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
924 for (i = 0; i < this_jobs; i++) {
925 td = map[i];
926 if (!td)
927 continue;
928 kill(td->pid, SIGTERM);
929 }
930 break;
931 }
932
933 /*
b6f4d880 934 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 935 */
34572e28 936 for_each_td(td, i) {
75154845 937 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
938 continue;
939
940 td_set_runstate(td, TD_RUNNING);
941 nr_running++;
942 nr_started--;
943 m_rate += td->ratemin;
944 t_rate += td->rate;
75154845 945 todo--;
bbfd6b00 946 fio_sem_up(&td->mutex);
ebac4655
JA
947 }
948
949 reap_threads(&nr_running, &t_rate, &m_rate);
950
951 if (todo)
952 usleep(100000);
953 }
954
955 while (nr_running) {
956 reap_threads(&nr_running, &t_rate, &m_rate);
957 usleep(10000);
958 }
959
960 update_io_ticks();
2f9ade3c 961 fio_unpin_memory();
ebac4655
JA
962}
963
ebac4655
JA
964int main(int argc, char *argv[])
965{
966 if (parse_options(argc, argv))
967 return 1;
968
969 if (!thread_number) {
3b70d7e5 970 log_err("Nothing to do\n");
ebac4655
JA
971 return 1;
972 }
973
974 disk_util_timer_arm();
975
976 run_threads();
977 show_run_stats();
978
979 return 0;
980}