[PATCH] More random map fixes
[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
29d610e1
JA
40static unsigned long page_mask;
41#define ALIGN(buf) \
42 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
ebac4655
JA
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;
6ce15a32 50static volatile int fio_abort;
ebac4655 51
bb3884d8
JA
52struct io_log *agg_io_log[2];
53
ebac4655 54#define TERMINATE_ALL (-1)
75154845 55#define JOB_START_TIMEOUT (5 * 1000)
ebac4655 56
6ce15a32
JA
57static inline void td_set_runstate(struct thread_data *td, int runstate)
58{
59 td->runstate = runstate;
60}
61
62static void terminate_threads(int group_id, int forced_kill)
ebac4655 63{
34572e28 64 struct thread_data *td;
ebac4655
JA
65 int i;
66
34572e28 67 for_each_td(td, i) {
ebac4655
JA
68 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
69 td->terminate = 1;
70 td->start_delay = 0;
6ce15a32
JA
71 if (forced_kill)
72 td_set_runstate(td, TD_EXITED);
ebac4655
JA
73 }
74 }
75}
76
77static void sig_handler(int sig)
78{
79 switch (sig) {
80 case SIGALRM:
81 update_io_ticks();
82 disk_util_timer_arm();
83 print_thread_status();
84 break;
85 default:
6ce15a32 86 printf("\nfio: terminating on signal %d\n", sig);
ebac4655 87 fflush(stdout);
6ce15a32 88 terminate_threads(TERMINATE_ALL, 0);
ebac4655
JA
89 break;
90 }
91}
92
906c8d75
JA
93/*
94 * Check if we are above the minimum rate given.
95 */
ebac4655
JA
96static int check_min_rate(struct thread_data *td, struct timeval *now)
97{
98 unsigned long spent;
99 unsigned long rate;
100 int ddir = td->ddir;
101
102 /*
103 * allow a 2 second settle period in the beginning
104 */
105 if (mtime_since(&td->start, now) < 2000)
106 return 0;
107
108 /*
109 * if rate blocks is set, sample is running
110 */
111 if (td->rate_bytes) {
112 spent = mtime_since(&td->lastrate, now);
113 if (spent < td->ratecycle)
114 return 0;
115
116 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
117 if (rate < td->ratemin) {
1e97cce9 118 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
ebac4655
JA
119 return 1;
120 }
121 }
122
123 td->rate_bytes = td->this_io_bytes[ddir];
124 memcpy(&td->lastrate, now, sizeof(*now));
125 return 0;
126}
127
128static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
129{
130 if (!td->timeout)
131 return 0;
132 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
133 return 1;
134
135 return 0;
136}
137
53cdc686
JA
138static struct fio_file *get_next_file(struct thread_data *td)
139{
0ab8db89 140 unsigned int old_next_file = td->next_file;
b2a15192
JA
141 struct fio_file *f;
142
143 do {
144 f = &td->files[td->next_file];
145
146 td->next_file++;
147 if (td->next_file >= td->nr_files)
148 td->next_file = 0;
149
150 if (f->fd != -1)
151 break;
53cdc686 152
b2a15192
JA
153 f = NULL;
154 } while (td->next_file != old_next_file);
53cdc686
JA
155
156 return f;
157}
158
906c8d75
JA
159/*
160 * When job exits, we can cancel the in-flight IO if we are using async
161 * io. Attempt to do so.
162 */
ebac4655
JA
163static void cleanup_pending_aio(struct thread_data *td)
164{
165 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
166 struct list_head *entry, *n;
167 struct io_completion_data icd;
168 struct io_u *io_u;
169 int r;
170
171 /*
172 * get immediately available events, if any
173 */
45bee283 174 r = td_io_getevents(td, 0, td->cur_depth, &ts);
ebac4655
JA
175 if (r > 0) {
176 icd.nr = r;
177 ios_completed(td, &icd);
178 }
179
180 /*
181 * now cancel remaining active events
182 */
2866c82d 183 if (td->io_ops->cancel) {
ebac4655
JA
184 list_for_each_safe(entry, n, &td->io_u_busylist) {
185 io_u = list_entry(entry, struct io_u, list);
186
2866c82d 187 r = td->io_ops->cancel(td, io_u);
ebac4655
JA
188 if (!r)
189 put_io_u(td, io_u);
190 }
191 }
192
193 if (td->cur_depth) {
45bee283 194 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
ebac4655
JA
195 if (r > 0) {
196 icd.nr = r;
197 ios_completed(td, &icd);
198 }
199 }
200}
201
858a3d47
JA
202/*
203 * Helper to handle the final sync of a file. Works just like the normal
204 * io path, just does everything sync.
205 */
206static int fio_io_sync(struct thread_data *td, struct fio_file *f)
207{
208 struct io_u *io_u = __get_io_u(td);
209 struct io_completion_data icd;
210 int ret;
211
212 if (!io_u)
213 return 1;
214
215 io_u->ddir = DDIR_SYNC;
216 io_u->file = f;
217
218 if (td_io_prep(td, io_u)) {
219 put_io_u(td, io_u);
220 return 1;
221 }
222
223 ret = td_io_queue(td, io_u);
224 if (ret) {
353a7e0e 225 td_verror(td, io_u->error);
858a3d47 226 put_io_u(td, io_u);
858a3d47
JA
227 return 1;
228 }
229
230 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
231 if (ret < 0) {
353a7e0e 232 td_verror(td, ret);
858a3d47
JA
233 return 1;
234 }
235
236 icd.nr = ret;
237 ios_completed(td, &icd);
238 if (icd.error) {
239 td_verror(td, icd.error);
240 return 1;
241 }
242
243 return 0;
244}
245
906c8d75
JA
246/*
247 * The main verify engine. Runs over the writes we previusly submitted,
248 * reads the blocks back in, and checks the crc/md5 of the data.
249 */
1e97cce9 250static void do_verify(struct thread_data *td)
ebac4655 251{
ebac4655
JA
252 struct io_u *io_u, *v_io_u = NULL;
253 struct io_completion_data icd;
53cdc686 254 struct fio_file *f;
e5b401d4
JA
255 int ret, i;
256
257 /*
258 * sync io first and invalidate cache, to make sure we really
259 * read from disk.
260 */
261 for_each_file(td, f, i) {
858a3d47 262 fio_io_sync(td, f);
e5b401d4
JA
263 file_invalidate_cache(td, f);
264 }
ebac4655
JA
265
266 td_set_runstate(td, TD_VERIFYING);
267
268 do {
269 if (td->terminate)
270 break;
271
ebac4655
JA
272 io_u = __get_io_u(td);
273 if (!io_u)
274 break;
275
02bcaa8c
JA
276 if (runtime_exceeded(td, &io_u->start_time)) {
277 put_io_u(td, io_u);
278 break;
279 }
280
aea47d44 281 if (get_next_verify(td, io_u)) {
ebac4655
JA
282 put_io_u(td, io_u);
283 break;
284 }
285
53cdc686
JA
286 f = get_next_file(td);
287 if (!f)
288 break;
289
290 io_u->file = f;
291
aea47d44 292 if (td_io_prep(td, io_u)) {
ebac4655
JA
293 put_io_u(td, io_u);
294 break;
295 }
296
45bee283 297 ret = td_io_queue(td, io_u);
ebac4655 298 if (ret) {
353a7e0e 299 td_verror(td, io_u->error);
ebac4655 300 put_io_u(td, io_u);
ebac4655
JA
301 break;
302 }
303
304 /*
305 * we have one pending to verify, do that while
306 * we are doing io on the next one
307 */
308 if (do_io_u_verify(td, &v_io_u))
309 break;
310
45bee283 311 ret = td_io_getevents(td, 1, 1, NULL);
ebac4655
JA
312 if (ret != 1) {
313 if (ret < 0)
314 td_verror(td, ret);
315 break;
316 }
317
2866c82d 318 v_io_u = td->io_ops->event(td, 0);
ebac4655
JA
319 icd.nr = 1;
320 icd.error = 0;
02bcaa8c 321 fio_gettime(&icd.time, NULL);
ebac4655
JA
322 io_completed(td, v_io_u, &icd);
323
324 if (icd.error) {
325 td_verror(td, icd.error);
326 put_io_u(td, v_io_u);
327 v_io_u = NULL;
328 break;
329 }
330
ebac4655
JA
331 /*
332 * if we can't submit more io, we need to verify now
333 */
334 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
335 break;
336
337 } while (1);
338
339 do_io_u_verify(td, &v_io_u);
340
341 if (td->cur_depth)
342 cleanup_pending_aio(td);
343
344 td_set_runstate(td, TD_RUNNING);
345}
346
b990b5c0
JA
347/*
348 * Not really an io thread, all it does is burn CPU cycles in the specified
349 * manner.
350 */
351static void do_cpuio(struct thread_data *td)
352{
353 struct timeval e;
354 int split = 100 / td->cpuload;
355 int i = 0;
356
357 while (!td->terminate) {
02bcaa8c 358 fio_gettime(&e, NULL);
b990b5c0
JA
359
360 if (runtime_exceeded(td, &e))
361 break;
362
363 if (!(i % split))
364 __usec_sleep(10000);
365 else
366 usec_sleep(td, 10000);
367
368 i++;
369 }
370}
371
32cd46a0 372/*
906c8d75 373 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
374 * and reaps them, checking for rate and errors along the way.
375 */
ebac4655
JA
376static void do_io(struct thread_data *td)
377{
378 struct io_completion_data icd;
02bcaa8c 379 struct timeval s;
ebac4655 380 unsigned long usec;
53cdc686 381 struct fio_file *f;
84585003 382 int i, ret = 0;
ebac4655 383
5853e5a8
JA
384 td_set_runstate(td, TD_RUNNING);
385
ebac4655 386 while (td->this_io_bytes[td->ddir] < td->io_size) {
ebac4655 387 struct timespec *timeout;
84585003 388 int min_evts = 0;
ebac4655
JA
389 struct io_u *io_u;
390
391 if (td->terminate)
392 break;
393
53cdc686
JA
394 f = get_next_file(td);
395 if (!f)
396 break;
397
398 io_u = get_io_u(td, f);
ebac4655
JA
399 if (!io_u)
400 break;
401
402 memcpy(&s, &io_u->start_time, sizeof(s));
403
cec6b55d 404requeue:
45bee283 405 ret = td_io_queue(td, io_u);
ebac4655 406 if (ret) {
cec6b55d
JA
407 if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
408 io_u->resid) {
409 /*
410 * short read/write. requeue.
411 */
412 io_u->xfer_buflen = io_u->resid;
413 io_u->xfer_buf += ret;
414 goto requeue;
415 } else {
416 td_verror(td, io_u->error);
417 put_io_u(td, io_u);
418 break;
419 }
ebac4655
JA
420 }
421
422 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
423
424 if (td->cur_depth < td->iodepth) {
8d7ad8d9
JA
425 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
426
ebac4655
JA
427 timeout = &ts;
428 min_evts = 0;
429 } else {
430 timeout = NULL;
431 min_evts = 1;
432 }
433
45bee283 434 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
ebac4655 435 if (ret < 0) {
87dc1ab1 436 td_verror(td, ret);
ebac4655
JA
437 break;
438 } else if (!ret)
439 continue;
440
441 icd.nr = ret;
442 ios_completed(td, &icd);
443 if (icd.error) {
444 td_verror(td, icd.error);
445 break;
446 }
447
448 /*
449 * the rate is batched for now, it should work for batches
450 * of completions except the very first one which may look
451 * a little bursty
452 */
02bcaa8c 453 usec = utime_since(&s, &icd.time);
ebac4655 454
a00735e6 455 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
ebac4655 456
02bcaa8c 457 if (check_min_rate(td, &icd.time)) {
98aa62d8 458 if (exitall_on_terminate)
6ce15a32 459 terminate_threads(td->groupid, 0);
fd841467 460 td_verror(td, ENODATA);
ebac4655
JA
461 break;
462 }
463
02bcaa8c 464 if (runtime_exceeded(td, &icd.time))
ebac4655
JA
465 break;
466
9c1f7434
JA
467 if (td->thinktime) {
468 unsigned long long b;
469
470 b = td->io_blocks[0] + td->io_blocks[1];
25309a33 471 if (!(b % td->thinktime_blocks))
9c1f7434
JA
472 usec_sleep(td, td->thinktime);
473 }
ebac4655
JA
474 }
475
4d2413c6 476 if (!td->error) {
84585003
JA
477 if (td->cur_depth)
478 cleanup_pending_aio(td);
ebac4655 479
84585003
JA
480 if (should_fsync(td) && td->end_fsync) {
481 td_set_runstate(td, TD_FSYNCING);
482 for_each_file(td, f, i)
858a3d47 483 fio_io_sync(td, f);
84585003 484 }
5853e5a8 485 }
ebac4655
JA
486}
487
ebac4655
JA
488static void cleanup_io_u(struct thread_data *td)
489{
490 struct list_head *entry, *n;
491 struct io_u *io_u;
492
493 list_for_each_safe(entry, n, &td->io_u_freelist) {
494 io_u = list_entry(entry, struct io_u, list);
495
496 list_del(&io_u->list);
497 free(io_u);
498 }
499
2f9ade3c 500 free_io_mem(td);
ebac4655
JA
501}
502
6b9cea23
JA
503/*
504 * "randomly" fill the buffer contents
505 */
66eeb296 506static void fill_rand_buf(struct io_u *io_u, int max_bs)
6b9cea23 507{
66eeb296 508 int *ptr = io_u->buf;
6b9cea23
JA
509
510 while ((void *) ptr - io_u->buf < max_bs) {
511 *ptr = rand() * 0x9e370001;
512 ptr++;
513 }
514}
515
ebac4655
JA
516static int init_io_u(struct thread_data *td)
517{
518 struct io_u *io_u;
a00735e6 519 unsigned int max_bs;
ebac4655
JA
520 int i, max_units;
521 char *p;
522
2866c82d 523 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
524 return 0;
525
2866c82d 526 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
527 max_units = 1;
528 else
529 max_units = td->iodepth;
530
a00735e6 531 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
74b025b0
JA
532 td->orig_buffer_size = max_bs * max_units;
533
d0bdaf49 534 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
56bb17f2 535 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
74b025b0 536 else
29d610e1 537 td->orig_buffer_size += page_mask;
ebac4655 538
2f9ade3c
JA
539 if (allocate_io_mem(td))
540 return 1;
ebac4655 541
ebac4655
JA
542 p = ALIGN(td->orig_buffer);
543 for (i = 0; i < max_units; i++) {
544 io_u = malloc(sizeof(*io_u));
545 memset(io_u, 0, sizeof(*io_u));
546 INIT_LIST_HEAD(&io_u->list);
547
a00735e6 548 io_u->buf = p + max_bs * i;
6b9cea23 549 if (td_write(td) || td_rw(td))
a00735e6 550 fill_rand_buf(io_u, max_bs);
6b9cea23 551
b1ff3403 552 io_u->index = i;
ebac4655
JA
553 list_add(&io_u->list, &td->io_u_freelist);
554 }
555
556 return 0;
557}
558
da86774e
JA
559static int switch_ioscheduler(struct thread_data *td)
560{
561 char tmp[256], tmp2[128];
562 FILE *f;
563 int ret;
564
f48b467c
JA
565 if (td->io_ops->flags & FIO_CPUIO)
566 return 0;
567
da86774e
JA
568 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
569
570 f = fopen(tmp, "r+");
571 if (!f) {
572 td_verror(td, errno);
573 return 1;
574 }
575
576 /*
577 * Set io scheduler.
578 */
579 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
580 if (ferror(f) || ret != 1) {
581 td_verror(td, errno);
582 fclose(f);
583 return 1;
584 }
585
586 rewind(f);
587
588 /*
589 * Read back and check that the selected scheduler is now the default.
590 */
591 ret = fread(tmp, 1, sizeof(tmp), f);
592 if (ferror(f) || ret < 0) {
593 td_verror(td, errno);
594 fclose(f);
595 return 1;
596 }
597
598 sprintf(tmp2, "[%s]", td->ioscheduler);
599 if (!strstr(tmp, tmp2)) {
3b70d7e5 600 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
da86774e
JA
601 td_verror(td, EINVAL);
602 fclose(f);
603 return 1;
604 }
605
606 fclose(f);
607 return 0;
608}
609
ebac4655
JA
610static void clear_io_state(struct thread_data *td)
611{
53cdc686
JA
612 struct fio_file *f;
613 int i;
ebac4655 614
ebac4655
JA
615 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
616 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 617 td->zone_bytes = 0;
ebac4655 618
53cdc686
JA
619 for_each_file(td, f, i) {
620 f->last_pos = 0;
621 if (td->io_ops->flags & FIO_SYNCIO)
622 lseek(f->fd, SEEK_SET, 0);
623
624 if (f->file_map)
625 memset(f->file_map, 0, f->num_maps * sizeof(long));
626 }
ebac4655
JA
627}
628
906c8d75
JA
629/*
630 * Entry point for the thread based jobs. The process based jobs end up
631 * here as well, after a little setup.
632 */
ebac4655
JA
633static void *thread_main(void *data)
634{
69008999 635 unsigned long long runtime[2];
ebac4655 636 struct thread_data *td = data;
ebac4655
JA
637
638 if (!td->use_thread)
639 setsid();
640
641 td->pid = getpid();
642
aea47d44
JA
643 INIT_LIST_HEAD(&td->io_u_freelist);
644 INIT_LIST_HEAD(&td->io_u_busylist);
645 INIT_LIST_HEAD(&td->io_hist_list);
646 INIT_LIST_HEAD(&td->io_log_list);
647
ebac4655
JA
648 if (init_io_u(td))
649 goto err;
650
651 if (fio_setaffinity(td) == -1) {
652 td_verror(td, errno);
653 goto err;
654 }
655
aea47d44
JA
656 if (init_iolog(td))
657 goto err;
658
ebac4655
JA
659 if (td->ioprio) {
660 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
661 td_verror(td, errno);
662 goto err;
663 }
664 }
665
1056eaad 666 if (nice(td->nice) == -1) {
b6f4d880
JA
667 td_verror(td, errno);
668 goto err;
669 }
670
75154845
JA
671 if (init_random_state(td))
672 goto err;
673
56f9498d
JA
674 if (td->ioscheduler && switch_ioscheduler(td))
675 goto err;
da86774e 676
75154845 677 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
678 fio_sem_up(&startup_sem);
679 fio_sem_down(&td->mutex);
ebac4655 680
53cdc686 681 if (!td->create_serialize && setup_files(td))
ebac4655 682 goto err;
21972cde
JA
683 if (open_files(td))
684 goto err;
ebac4655 685
7d6c5283
JA
686 /*
687 * Do this late, as some IO engines would like to have the
688 * files setup prior to initializing structures.
689 */
690 if (td_io_init(td))
691 goto err;
692
69cfd7e0
JA
693 if (td->exec_prerun) {
694 if (system(td->exec_prerun) < 0)
695 goto err;
696 }
4e0ba8af 697
69008999 698 fio_gettime(&td->epoch, NULL);
36dff966 699 getrusage(RUSAGE_SELF, &td->ru_start);
69008999
JA
700
701 runtime[0] = runtime[1] = 0;
ebac4655 702 while (td->loops--) {
02bcaa8c 703 fio_gettime(&td->start, NULL);
ebac4655
JA
704 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
705
706 if (td->ratemin)
707 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
708
709 clear_io_state(td);
710 prune_io_piece_log(td);
711
2866c82d 712 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
713 do_cpuio(td);
714 else
715 do_io(td);
ebac4655 716
69008999 717 runtime[td->ddir] += utime_since_now(&td->start);
aea47d44 718 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
69008999 719 runtime[td->ddir ^ 1] = runtime[td->ddir];
3d60d1ed 720
ebac4655
JA
721 if (td->error || td->terminate)
722 break;
723
724 if (td->verify == VERIFY_NONE)
725 continue;
726
727 clear_io_state(td);
02bcaa8c 728 fio_gettime(&td->start, NULL);
ebac4655
JA
729
730 do_verify(td);
731
69008999 732 runtime[DDIR_READ] += utime_since_now(&td->start);
ebac4655
JA
733
734 if (td->error || td->terminate)
735 break;
736 }
737
36dff966 738 update_rusage_stat(td);
69008999
JA
739 fio_gettime(&td->end_time, NULL);
740 td->runtime[0] = runtime[0] / 1000;
741 td->runtime[1] = runtime[1] / 1000;
742
ebac4655
JA
743 if (td->bw_log)
744 finish_log(td, td->bw_log, "bw");
745 if (td->slat_log)
746 finish_log(td, td->slat_log, "slat");
747 if (td->clat_log)
748 finish_log(td, td->clat_log, "clat");
076efc7c 749 if (td->write_iolog_file)
843a7413 750 write_iolog_close(td);
69cfd7e0
JA
751 if (td->exec_postrun) {
752 if (system(td->exec_postrun) < 0)
753 log_err("fio: postrun %s failed\n", td->exec_postrun);
754 }
ebac4655
JA
755
756 if (exitall_on_terminate)
6ce15a32 757 terminate_threads(td->groupid, 0);
ebac4655
JA
758
759err:
53cdc686 760 close_files(td);
2866c82d 761 close_ioengine(td);
ebac4655 762 cleanup_io_u(td);
ebac4655
JA
763 td_set_runstate(td, TD_EXITED);
764 return NULL;
765
766}
767
906c8d75
JA
768/*
769 * We cannot pass the td data into a forked process, so attach the td and
770 * pass it to the thread worker.
771 */
ebac4655
JA
772static void *fork_main(int shmid, int offset)
773{
774 struct thread_data *td;
775 void *data;
776
777 data = shmat(shmid, NULL, 0);
778 if (data == (void *) -1) {
779 perror("shmat");
780 return NULL;
781 }
782
783 td = data + offset * sizeof(struct thread_data);
784 thread_main(td);
785 shmdt(data);
786 return NULL;
787}
788
906c8d75
JA
789/*
790 * Run over the job map and reap the threads that have exited, if any.
791 */
ebac4655
JA
792static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
793{
34572e28 794 struct thread_data *td;
4d2413c6 795 int i, cputhreads, pending;
ebac4655
JA
796
797 /*
798 * reap exited threads (TD_EXITED -> TD_REAPED)
799 */
4d2413c6 800 pending = cputhreads = 0;
34572e28 801 for_each_td(td, i) {
84585003
JA
802 /*
803 * ->io_ops is NULL for a thread that has closed its
804 * io engine
805 */
806 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
807 cputhreads++;
808
4d2413c6
JA
809 if (td->runstate != TD_EXITED) {
810 if (td->runstate < TD_RUNNING)
811 pending++;
812
ebac4655 813 continue;
4d2413c6 814 }
ebac4655
JA
815
816 td_set_runstate(td, TD_REAPED);
817
818 if (td->use_thread) {
819 long ret;
820
821 if (pthread_join(td->thread, (void *) &ret))
822 perror("thread_join");
823 } else
824 waitpid(td->pid, NULL, 0);
825
826 (*nr_running)--;
827 (*m_rate) -= td->ratemin;
828 (*t_rate) -= td->rate;
829 }
b990b5c0 830
4d2413c6 831 if (*nr_running == cputhreads && !pending)
6ce15a32 832 terminate_threads(TERMINATE_ALL, 0);
ebac4655
JA
833}
834
906c8d75
JA
835/*
836 * Main function for kicking off and reaping jobs, as needed.
837 */
ebac4655
JA
838static void run_threads(void)
839{
ebac4655
JA
840 struct thread_data *td;
841 unsigned long spent;
842 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2 843
2f9ade3c
JA
844 if (fio_pin_memory())
845 return;
ebac4655 846
c6ae0a5b
JA
847 if (!terse_output) {
848 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
849 fflush(stdout);
850 }
c04f7ec3 851
4efa970e
JA
852 signal(SIGINT, sig_handler);
853 signal(SIGALRM, sig_handler);
854
ebac4655
JA
855 todo = thread_number;
856 nr_running = 0;
857 nr_started = 0;
858 m_rate = t_rate = 0;
859
34572e28 860 for_each_td(td, i) {
263e529f 861 print_status_init(td->thread_number - 1);
ebac4655 862
380cf265
JA
863 if (!td->create_serialize) {
864 init_disk_util(td);
ebac4655 865 continue;
380cf265 866 }
ebac4655
JA
867
868 /*
869 * do file setup here so it happens sequentially,
870 * we don't want X number of threads getting their
871 * client data interspersed on disk
872 */
53cdc686 873 if (setup_files(td)) {
ebac4655
JA
874 td_set_runstate(td, TD_REAPED);
875 todo--;
876 }
380cf265
JA
877
878 init_disk_util(td);
ebac4655
JA
879 }
880
ebac4655 881 while (todo) {
75154845
JA
882 struct thread_data *map[MAX_JOBS];
883 struct timeval this_start;
884 int this_jobs = 0, left;
885
ebac4655
JA
886 /*
887 * create threads (TD_NOT_CREATED -> TD_CREATED)
888 */
34572e28 889 for_each_td(td, i) {
ebac4655
JA
890 if (td->runstate != TD_NOT_CREATED)
891 continue;
892
893 /*
894 * never got a chance to start, killed by other
895 * thread for some reason
896 */
897 if (td->terminate) {
898 todo--;
899 continue;
900 }
901
902 if (td->start_delay) {
263e529f 903 spent = mtime_since_genesis();
ebac4655
JA
904
905 if (td->start_delay * 1000 > spent)
906 continue;
907 }
908
909 if (td->stonewall && (nr_started || nr_running))
910 break;
911
75154845
JA
912 /*
913 * Set state to created. Thread will transition
914 * to TD_INITIALIZED when it's done setting up.
915 */
ebac4655 916 td_set_runstate(td, TD_CREATED);
75154845 917 map[this_jobs++] = td;
bbfd6b00 918 fio_sem_init(&startup_sem, 1);
ebac4655
JA
919 nr_started++;
920
921 if (td->use_thread) {
922 if (pthread_create(&td->thread, NULL, thread_main, td)) {
923 perror("thread_create");
924 nr_started--;
925 }
926 } else {
927 if (fork())
bbfd6b00 928 fio_sem_down(&startup_sem);
ebac4655
JA
929 else {
930 fork_main(shm_id, i);
931 exit(0);
932 }
933 }
934 }
935
936 /*
75154845
JA
937 * Wait for the started threads to transition to
938 * TD_INITIALIZED.
ebac4655 939 */
02bcaa8c 940 fio_gettime(&this_start, NULL);
75154845 941 left = this_jobs;
6ce15a32 942 while (left && !fio_abort) {
75154845
JA
943 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
944 break;
945
946 usleep(100000);
947
948 for (i = 0; i < this_jobs; i++) {
949 td = map[i];
950 if (!td)
951 continue;
b6f4d880 952 if (td->runstate == TD_INITIALIZED) {
75154845
JA
953 map[i] = NULL;
954 left--;
b6f4d880
JA
955 } else if (td->runstate >= TD_EXITED) {
956 map[i] = NULL;
957 left--;
958 todo--;
959 nr_running++; /* work-around... */
75154845
JA
960 }
961 }
962 }
963
964 if (left) {
3b70d7e5 965 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
966 for (i = 0; i < this_jobs; i++) {
967 td = map[i];
968 if (!td)
969 continue;
970 kill(td->pid, SIGTERM);
971 }
972 break;
973 }
974
975 /*
b6f4d880 976 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 977 */
34572e28 978 for_each_td(td, i) {
75154845 979 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
980 continue;
981
982 td_set_runstate(td, TD_RUNNING);
983 nr_running++;
984 nr_started--;
985 m_rate += td->ratemin;
986 t_rate += td->rate;
75154845 987 todo--;
bbfd6b00 988 fio_sem_up(&td->mutex);
ebac4655
JA
989 }
990
991 reap_threads(&nr_running, &t_rate, &m_rate);
992
993 if (todo)
994 usleep(100000);
995 }
996
997 while (nr_running) {
998 reap_threads(&nr_running, &t_rate, &m_rate);
999 usleep(10000);
1000 }
1001
1002 update_io_ticks();
2f9ade3c 1003 fio_unpin_memory();
ebac4655
JA
1004}
1005
ebac4655
JA
1006int main(int argc, char *argv[])
1007{
29d610e1
JA
1008 long ps;
1009
ebac4655
JA
1010 if (parse_options(argc, argv))
1011 return 1;
1012
1013 if (!thread_number) {
3b70d7e5 1014 log_err("Nothing to do\n");
ebac4655
JA
1015 return 1;
1016 }
1017
29d610e1
JA
1018 ps = sysconf(_SC_PAGESIZE);
1019 if (ps < 0) {
1020 log_err("Failed to get page size\n");
1021 return 1;
1022 }
1023
1024 page_mask = ps - 1;
1025
bb3884d8
JA
1026 if (write_bw_log) {
1027 setup_log(&agg_io_log[DDIR_READ]);
1028 setup_log(&agg_io_log[DDIR_WRITE]);
1029 }
1030
ebac4655
JA
1031 disk_util_timer_arm();
1032
1033 run_threads();
6ce15a32 1034
bb3884d8 1035 if (!fio_abort) {
6ce15a32 1036 show_run_stats();
bb3884d8
JA
1037 if (write_bw_log) {
1038 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1039 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1040 }
1041 }
ebac4655
JA
1042
1043 return 0;
1044}