[PATCH] Make fio -W clean again
[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
JA
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
11 *
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
16 *
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 *
21 */
ebac4655
JA
22#include <unistd.h>
23#include <fcntl.h>
24#include <string.h>
ebac4655
JA
25#include <signal.h>
26#include <time.h>
ebac4655 27#include <assert.h>
ebac4655
JA
28#include <sys/stat.h>
29#include <sys/wait.h>
30#include <sys/ipc.h>
31#include <sys/shm.h>
32#include <sys/ioctl.h>
33#include <sys/mman.h>
34
35#include "fio.h"
36#include "os.h"
37
38#define MASK (4095)
39
40#define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
41
42int groupid = 0;
43int thread_number = 0;
ebac4655 44int shm_id = 0;
53cdc686 45int temp_stall_ts;
c1d5725e 46char *fio_inst_prefix = _INST_PREFIX;
ebac4655 47
c04f7ec3
JA
48extern unsigned long long mlock_size;
49
3d60d1ed 50#define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
ebac4655 51
bbfd6b00 52static volatile int startup_sem;
ebac4655
JA
53
54#define TERMINATE_ALL (-1)
75154845 55#define JOB_START_TIMEOUT (5 * 1000)
ebac4655
JA
56
57static void terminate_threads(int group_id)
58{
59 int i;
60
61 for (i = 0; i < thread_number; i++) {
62 struct thread_data *td = &threads[i];
63
64 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
65 td->terminate = 1;
66 td->start_delay = 0;
67 }
68 }
69}
70
71static void sig_handler(int sig)
72{
73 switch (sig) {
74 case SIGALRM:
75 update_io_ticks();
76 disk_util_timer_arm();
77 print_thread_status();
78 break;
79 default:
80 printf("\nfio: terminating on signal\n");
81 fflush(stdout);
82 terminate_threads(TERMINATE_ALL);
83 break;
84 }
85}
86
906c8d75
JA
87/*
88 * The ->file_map[] contains a map of blocks we have or have not done io
89 * to yet. Used to make sure we cover the entire range in a fair fashion.
90 */
53cdc686
JA
91static int random_map_free(struct thread_data *td, struct fio_file *f,
92 unsigned long long block)
ebac4655 93{
53cdc686
JA
94 unsigned int idx = RAND_MAP_IDX(td, f, block);
95 unsigned int bit = RAND_MAP_BIT(td, f, block);
ebac4655 96
53cdc686 97 return (f->file_map[idx] & (1UL << bit)) == 0;
ebac4655
JA
98}
99
906c8d75
JA
100/*
101 * Return the next free block in the map.
102 */
53cdc686
JA
103static int get_next_free_block(struct thread_data *td, struct fio_file *f,
104 unsigned long long *b)
ebac4655
JA
105{
106 int i;
107
108 *b = 0;
109 i = 0;
53cdc686
JA
110 while ((*b) * td->min_bs < f->file_size) {
111 if (f->file_map[i] != -1UL) {
112 *b += ffz(f->file_map[i]);
ebac4655
JA
113 return 0;
114 }
115
116 *b += BLOCKS_PER_MAP;
117 i++;
118 }
119
120 return 1;
121}
122
906c8d75
JA
123/*
124 * Mark a given offset as used in the map.
125 */
53cdc686
JA
126static void mark_random_map(struct thread_data *td, struct fio_file *f,
127 struct io_u *io_u)
ebac4655 128{
200bc855 129 unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
ebac4655
JA
130 unsigned int blocks = 0;
131
132 while (blocks < (io_u->buflen / td->min_bs)) {
133 unsigned int idx, bit;
134
53cdc686 135 if (!random_map_free(td, f, block))
ebac4655
JA
136 break;
137
53cdc686
JA
138 idx = RAND_MAP_IDX(td, f, block);
139 bit = RAND_MAP_BIT(td, f, block);
ebac4655 140
53cdc686 141 assert(idx < f->num_maps);
ebac4655 142
53cdc686 143 f->file_map[idx] |= (1UL << bit);
ebac4655
JA
144 block++;
145 blocks++;
146 }
147
148 if ((blocks * td->min_bs) < io_u->buflen)
149 io_u->buflen = blocks * td->min_bs;
150}
151
906c8d75
JA
152/*
153 * For random io, generate a random new block and see if it's used. Repeat
154 * until we find a free one. For sequential io, just return the end of
155 * the last io issued.
156 */
53cdc686
JA
157static int get_next_offset(struct thread_data *td, struct fio_file *f,
158 unsigned long long *offset)
20dc95c4
JA
159{
160 unsigned long long b, rb;
161 long r;
162
163 if (!td->sequential) {
085227ad 164 unsigned long long max_blocks = td->io_size / td->min_bs;
20dc95c4
JA
165 int loops = 50;
166
167 do {
6dfd46b9 168 r = os_random_long(&td->random_state);
085227ad 169 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
53cdc686 170 rb = b + (f->file_offset / td->min_bs);
20dc95c4 171 loops--;
53cdc686 172 } while (!random_map_free(td, f, rb) && loops);
20dc95c4
JA
173
174 if (!loops) {
53cdc686 175 if (get_next_free_block(td, f, &b))
20dc95c4
JA
176 return 1;
177 }
178 } else
53cdc686 179 b = f->last_pos / td->min_bs;
20dc95c4 180
53cdc686
JA
181 *offset = (b * td->min_bs) + f->file_offset;
182 if (*offset > f->file_size)
20dc95c4
JA
183 return 1;
184
185 return 0;
186}
187
188static unsigned int get_next_buflen(struct thread_data *td)
189{
190 unsigned int buflen;
191 long r;
192
193 if (td->min_bs == td->max_bs)
194 buflen = td->min_bs;
195 else {
6dfd46b9 196 r = os_random_long(&td->bsrange_state);
20dc95c4
JA
197 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
198 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
199 }
200
b2a15192
JA
201 if (buflen > td->io_size - td->this_io_bytes[td->ddir]) {
202 /*
203 * if using direct/raw io, we may not be able to
204 * shrink the size. so just fail it.
205 */
206 if (td->io_ops->flags & FIO_RAWIO)
207 return 0;
208
20dc95c4 209 buflen = td->io_size - td->this_io_bytes[td->ddir];
b2a15192 210 }
20dc95c4
JA
211
212 return buflen;
213}
214
906c8d75
JA
215/*
216 * Check if we are above the minimum rate given.
217 */
ebac4655
JA
218static int check_min_rate(struct thread_data *td, struct timeval *now)
219{
220 unsigned long spent;
221 unsigned long rate;
222 int ddir = td->ddir;
223
224 /*
225 * allow a 2 second settle period in the beginning
226 */
227 if (mtime_since(&td->start, now) < 2000)
228 return 0;
229
230 /*
231 * if rate blocks is set, sample is running
232 */
233 if (td->rate_bytes) {
234 spent = mtime_since(&td->lastrate, now);
235 if (spent < td->ratecycle)
236 return 0;
237
238 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
239 if (rate < td->ratemin) {
eb8bbf48 240 fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
ebac4655
JA
241 if (rate_quit)
242 terminate_threads(td->groupid);
243 return 1;
244 }
245 }
246
247 td->rate_bytes = td->this_io_bytes[ddir];
248 memcpy(&td->lastrate, now, sizeof(*now));
249 return 0;
250}
251
252static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
253{
254 if (!td->timeout)
255 return 0;
256 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
257 return 1;
258
259 return 0;
260}
261
906c8d75
JA
262/*
263 * Return the data direction for the next io_u. If the job is a
264 * mixed read/write workload, check the rwmix cycle and switch if
265 * necessary.
266 */
a6ccc7be 267static int get_rw_ddir(struct thread_data *td)
3d60d1ed 268{
a6ccc7be
JA
269 if (td_rw(td)) {
270 struct timeval now;
271 unsigned long elapsed;
3d60d1ed 272
a6ccc7be
JA
273 gettimeofday(&now, NULL);
274 elapsed = mtime_since_now(&td->rwmix_switch);
3d60d1ed 275
a6ccc7be
JA
276 /*
277 * Check if it's time to seed a new data direction.
278 */
279 if (elapsed >= td->rwmixcycle) {
0ab8db89 280 unsigned int v;
a6ccc7be 281 long r;
3d60d1ed 282
c1ee2ca4
JA
283 r = os_random_long(&td->rwmix_state);
284 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
a6ccc7be
JA
285 if (v < td->rwmixread)
286 td->rwmix_ddir = DDIR_READ;
287 else
288 td->rwmix_ddir = DDIR_WRITE;
289 memcpy(&td->rwmix_switch, &now, sizeof(now));
290 }
291 return td->rwmix_ddir;
292 } else if (td_read(td))
3d60d1ed
JA
293 return DDIR_READ;
294 else
295 return DDIR_WRITE;
296}
297
aea47d44 298static int td_io_prep(struct thread_data *td, struct io_u *io_u)
20dc95c4 299{
2866c82d 300 if (td->io_ops->prep && td->io_ops->prep(td, io_u))
20dc95c4
JA
301 return 1;
302
303 return 0;
304}
305
b1ff3403 306void put_io_u(struct thread_data *td, struct io_u *io_u)
ebac4655 307{
53cdc686 308 io_u->file = NULL;
ebac4655
JA
309 list_del(&io_u->list);
310 list_add(&io_u->list, &td->io_u_freelist);
311 td->cur_depth--;
312}
313
53cdc686
JA
314static int fill_io_u(struct thread_data *td, struct fio_file *f,
315 struct io_u *io_u)
843a7413
JA
316{
317 /*
318 * If using an iolog, grab next piece if any available.
319 */
320 if (td->read_iolog)
321 return read_iolog_get(td, io_u);
322
aea47d44
JA
323 /*
324 * No log, let the seq/rand engine retrieve the next position.
325 */
53cdc686 326 if (!get_next_offset(td, f, &io_u->offset)) {
aea47d44
JA
327 io_u->buflen = get_next_buflen(td);
328
329 if (io_u->buflen) {
330 io_u->ddir = get_rw_ddir(td);
843a7413
JA
331
332 /*
333 * If using a write iolog, store this entry.
334 */
335 if (td->write_iolog)
336 write_iolog_put(td, io_u);
337
53cdc686 338 io_u->file = f;
aea47d44
JA
339 return 0;
340 }
341 }
342
343 return 1;
344}
345
22f78b32 346#define queue_full(td) list_empty(&(td)->io_u_freelist)
ebac4655 347
b1ff3403 348struct io_u *__get_io_u(struct thread_data *td)
ebac4655 349{
22f78b32 350 struct io_u *io_u = NULL;
ebac4655 351
22f78b32
JA
352 if (!queue_full(td)) {
353 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
354
355 io_u->error = 0;
356 io_u->resid = 0;
357 list_del(&io_u->list);
358 list_add(&io_u->list, &td->io_u_busylist);
359 td->cur_depth++;
360 }
ebac4655 361
ebac4655
JA
362 return io_u;
363}
364
906c8d75
JA
365/*
366 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
367 * etc. The returned io_u is fully ready to be prepped and submitted.
368 */
53cdc686 369static struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
ebac4655
JA
370{
371 struct io_u *io_u;
372
373 io_u = __get_io_u(td);
374 if (!io_u)
375 return NULL;
376
20dc95c4
JA
377 if (td->zone_bytes >= td->zone_size) {
378 td->zone_bytes = 0;
53cdc686 379 f->last_pos += td->zone_skip;
20dc95c4
JA
380 }
381
53cdc686 382 if (fill_io_u(td, f, io_u)) {
ebac4655
JA
383 put_io_u(td, io_u);
384 return NULL;
385 }
386
b2a15192
JA
387 if (io_u->buflen + io_u->offset > f->file_size) {
388 if (td->io_ops->flags & FIO_RAWIO) {
389 put_io_u(td, io_u);
390 return NULL;
391 }
392
53cdc686 393 io_u->buflen = f->file_size - io_u->offset;
b2a15192 394 }
ebac4655
JA
395
396 if (!io_u->buflen) {
397 put_io_u(td, io_u);
398 return NULL;
399 }
400
843a7413 401 if (!td->read_iolog && !td->sequential)
53cdc686 402 mark_random_map(td, f, io_u);
ebac4655 403
53cdc686 404 f->last_pos += io_u->buflen;
ebac4655
JA
405
406 if (td->verify != VERIFY_NONE)
e29d1b70 407 populate_verify_io_u(td, io_u);
ebac4655 408
aea47d44 409 if (td_io_prep(td, io_u)) {
ebac4655
JA
410 put_io_u(td, io_u);
411 return NULL;
412 }
413
414 gettimeofday(&io_u->start_time, NULL);
415 return io_u;
416}
417
418static inline void td_set_runstate(struct thread_data *td, int runstate)
419{
ebac4655
JA
420 td->runstate = runstate;
421}
422
53cdc686
JA
423static struct fio_file *get_next_file(struct thread_data *td)
424{
0ab8db89 425 unsigned int old_next_file = td->next_file;
b2a15192
JA
426 struct fio_file *f;
427
428 do {
429 f = &td->files[td->next_file];
430
431 td->next_file++;
432 if (td->next_file >= td->nr_files)
433 td->next_file = 0;
434
435 if (f->fd != -1)
436 break;
53cdc686 437
b2a15192
JA
438 f = NULL;
439 } while (td->next_file != old_next_file);
53cdc686
JA
440
441 return f;
442}
443
444static int td_io_sync(struct thread_data *td, struct fio_file *f)
ebac4655 445{
2866c82d 446 if (td->io_ops->sync)
53cdc686 447 return td->io_ops->sync(td, f);
ebac4655
JA
448
449 return 0;
450}
451
45bee283 452static int td_io_getevents(struct thread_data *td, int min, int max,
ebac4655
JA
453 struct timespec *t)
454{
2866c82d 455 return td->io_ops->getevents(td, min, max, t);
ebac4655
JA
456}
457
45bee283 458static int td_io_queue(struct thread_data *td, struct io_u *io_u)
ebac4655
JA
459{
460 gettimeofday(&io_u->issue_time, NULL);
461
2866c82d 462 return td->io_ops->queue(td, io_u);
ebac4655
JA
463}
464
465#define iocb_time(iocb) ((unsigned long) (iocb)->data)
466
467static void io_completed(struct thread_data *td, struct io_u *io_u,
468 struct io_completion_data *icd)
469{
470 struct timeval e;
471 unsigned long msec;
472
473 gettimeofday(&e, NULL);
474
475 if (!io_u->error) {
20dc95c4 476 unsigned int bytes = io_u->buflen - io_u->resid;
3d60d1ed 477 const int idx = io_u->ddir;
ebac4655
JA
478
479 td->io_blocks[idx]++;
20dc95c4
JA
480 td->io_bytes[idx] += bytes;
481 td->zone_bytes += bytes;
482 td->this_io_bytes[idx] += bytes;
ebac4655
JA
483
484 msec = mtime_since(&io_u->issue_time, &e);
485
3d60d1ed
JA
486 add_clat_sample(td, idx, msec);
487 add_bw_sample(td, idx);
ebac4655 488
3d60d1ed 489 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
ebac4655
JA
490 log_io_piece(td, io_u);
491
20dc95c4 492 icd->bytes_done[idx] += bytes;
ebac4655
JA
493 } else
494 icd->error = io_u->error;
495}
496
497static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
498{
499 struct io_u *io_u;
500 int i;
501
502 icd->error = 0;
503 icd->bytes_done[0] = icd->bytes_done[1] = 0;
504
505 for (i = 0; i < icd->nr; i++) {
2866c82d 506 io_u = td->io_ops->event(td, i);
ebac4655
JA
507
508 io_completed(td, io_u, icd);
509 put_io_u(td, io_u);
510 }
511}
512
906c8d75
JA
513/*
514 * When job exits, we can cancel the in-flight IO if we are using async
515 * io. Attempt to do so.
516 */
ebac4655
JA
517static void cleanup_pending_aio(struct thread_data *td)
518{
519 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
520 struct list_head *entry, *n;
521 struct io_completion_data icd;
522 struct io_u *io_u;
523 int r;
524
525 /*
526 * get immediately available events, if any
527 */
45bee283 528 r = td_io_getevents(td, 0, td->cur_depth, &ts);
ebac4655
JA
529 if (r > 0) {
530 icd.nr = r;
531 ios_completed(td, &icd);
532 }
533
534 /*
535 * now cancel remaining active events
536 */
2866c82d 537 if (td->io_ops->cancel) {
ebac4655
JA
538 list_for_each_safe(entry, n, &td->io_u_busylist) {
539 io_u = list_entry(entry, struct io_u, list);
540
2866c82d 541 r = td->io_ops->cancel(td, io_u);
ebac4655
JA
542 if (!r)
543 put_io_u(td, io_u);
544 }
545 }
546
547 if (td->cur_depth) {
45bee283 548 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
ebac4655
JA
549 if (r > 0) {
550 icd.nr = r;
551 ios_completed(td, &icd);
552 }
553 }
554}
555
906c8d75
JA
556/*
557 * The main verify engine. Runs over the writes we previusly submitted,
558 * reads the blocks back in, and checks the crc/md5 of the data.
559 */
a9619d44 560void do_verify(struct thread_data *td)
ebac4655
JA
561{
562 struct timeval t;
563 struct io_u *io_u, *v_io_u = NULL;
564 struct io_completion_data icd;
53cdc686 565 struct fio_file *f;
ebac4655
JA
566 int ret;
567
568 td_set_runstate(td, TD_VERIFYING);
569
570 do {
571 if (td->terminate)
572 break;
573
574 gettimeofday(&t, NULL);
575 if (runtime_exceeded(td, &t))
576 break;
577
578 io_u = __get_io_u(td);
579 if (!io_u)
580 break;
581
aea47d44 582 if (get_next_verify(td, io_u)) {
ebac4655
JA
583 put_io_u(td, io_u);
584 break;
585 }
586
53cdc686
JA
587 f = get_next_file(td);
588 if (!f)
589 break;
590
591 io_u->file = f;
592
aea47d44 593 if (td_io_prep(td, io_u)) {
ebac4655
JA
594 put_io_u(td, io_u);
595 break;
596 }
597
45bee283 598 ret = td_io_queue(td, io_u);
ebac4655
JA
599 if (ret) {
600 put_io_u(td, io_u);
601 td_verror(td, ret);
602 break;
603 }
604
605 /*
606 * we have one pending to verify, do that while
607 * we are doing io on the next one
608 */
609 if (do_io_u_verify(td, &v_io_u))
610 break;
611
45bee283 612 ret = td_io_getevents(td, 1, 1, NULL);
ebac4655
JA
613 if (ret != 1) {
614 if (ret < 0)
615 td_verror(td, ret);
616 break;
617 }
618
2866c82d 619 v_io_u = td->io_ops->event(td, 0);
ebac4655
JA
620 icd.nr = 1;
621 icd.error = 0;
622 io_completed(td, v_io_u, &icd);
623
624 if (icd.error) {
625 td_verror(td, icd.error);
626 put_io_u(td, v_io_u);
627 v_io_u = NULL;
628 break;
629 }
630
ebac4655
JA
631 /*
632 * if we can't submit more io, we need to verify now
633 */
634 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
635 break;
636
637 } while (1);
638
639 do_io_u_verify(td, &v_io_u);
640
641 if (td->cur_depth)
642 cleanup_pending_aio(td);
643
644 td_set_runstate(td, TD_RUNNING);
645}
646
b990b5c0
JA
647/*
648 * Not really an io thread, all it does is burn CPU cycles in the specified
649 * manner.
650 */
651static void do_cpuio(struct thread_data *td)
652{
653 struct timeval e;
654 int split = 100 / td->cpuload;
655 int i = 0;
656
657 while (!td->terminate) {
658 gettimeofday(&e, NULL);
659
660 if (runtime_exceeded(td, &e))
661 break;
662
663 if (!(i % split))
664 __usec_sleep(10000);
665 else
666 usec_sleep(td, 10000);
667
668 i++;
669 }
670}
671
32cd46a0 672/*
906c8d75 673 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
674 * and reaps them, checking for rate and errors along the way.
675 */
ebac4655
JA
676static void do_io(struct thread_data *td)
677{
678 struct io_completion_data icd;
679 struct timeval s, e;
680 unsigned long usec;
53cdc686
JA
681 struct fio_file *f;
682 int i;
ebac4655 683
5853e5a8
JA
684 td_set_runstate(td, TD_RUNNING);
685
ebac4655
JA
686 while (td->this_io_bytes[td->ddir] < td->io_size) {
687 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
688 struct timespec *timeout;
689 int ret, min_evts = 0;
690 struct io_u *io_u;
691
692 if (td->terminate)
693 break;
694
53cdc686
JA
695 f = get_next_file(td);
696 if (!f)
697 break;
698
699 io_u = get_io_u(td, f);
ebac4655
JA
700 if (!io_u)
701 break;
702
703 memcpy(&s, &io_u->start_time, sizeof(s));
704
45bee283 705 ret = td_io_queue(td, io_u);
ebac4655
JA
706 if (ret) {
707 put_io_u(td, io_u);
708 td_verror(td, ret);
709 break;
710 }
711
712 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
713
714 if (td->cur_depth < td->iodepth) {
715 timeout = &ts;
716 min_evts = 0;
717 } else {
718 timeout = NULL;
719 min_evts = 1;
720 }
721
45bee283 722 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
ebac4655
JA
723 if (ret < 0) {
724 td_verror(td, ret);
725 break;
726 } else if (!ret)
727 continue;
728
729 icd.nr = ret;
730 ios_completed(td, &icd);
731 if (icd.error) {
732 td_verror(td, icd.error);
733 break;
734 }
735
736 /*
737 * the rate is batched for now, it should work for batches
738 * of completions except the very first one which may look
739 * a little bursty
740 */
741 gettimeofday(&e, NULL);
742 usec = utime_since(&s, &e);
743
744 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
745
746 if (check_min_rate(td, &e)) {
747 td_verror(td, ENOMEM);
748 break;
749 }
750
751 if (runtime_exceeded(td, &e))
752 break;
753
754 if (td->thinktime)
755 usec_sleep(td, td->thinktime);
756
757 if (should_fsync(td) && td->fsync_blocks &&
758 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
53cdc686 759 td_io_sync(td, f);
ebac4655
JA
760 }
761
762 if (td->cur_depth)
763 cleanup_pending_aio(td);
764
5853e5a8
JA
765 if (should_fsync(td) && td->end_fsync) {
766 td_set_runstate(td, TD_FSYNCING);
53cdc686
JA
767 for_each_file(td, f, i)
768 td_io_sync(td, f);
5853e5a8 769 }
ebac4655
JA
770}
771
45bee283 772static int td_io_init(struct thread_data *td)
ebac4655 773{
2866c82d
JA
774 if (td->io_ops->init)
775 return td->io_ops->init(td);
776
777 return 0;
ebac4655
JA
778}
779
780static void cleanup_io_u(struct thread_data *td)
781{
782 struct list_head *entry, *n;
783 struct io_u *io_u;
784
785 list_for_each_safe(entry, n, &td->io_u_freelist) {
786 io_u = list_entry(entry, struct io_u, list);
787
788 list_del(&io_u->list);
789 free(io_u);
790 }
791
792 if (td->mem_type == MEM_MALLOC)
793 free(td->orig_buffer);
794 else if (td->mem_type == MEM_SHM) {
795 struct shmid_ds sbuf;
796
797 shmdt(td->orig_buffer);
798 shmctl(td->shm_id, IPC_RMID, &sbuf);
799 } else if (td->mem_type == MEM_MMAP)
800 munmap(td->orig_buffer, td->orig_buffer_size);
801 else
3b70d7e5 802 log_err("Bad memory type %d\n", td->mem_type);
ebac4655
JA
803
804 td->orig_buffer = NULL;
805}
806
807static int init_io_u(struct thread_data *td)
808{
809 struct io_u *io_u;
810 int i, max_units;
811 char *p;
812
2866c82d 813 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
814 return 0;
815
2866c82d 816 if (td->io_ops->flags & FIO_SYNCIO)
ebac4655
JA
817 max_units = 1;
818 else
819 max_units = td->iodepth;
820
821 td->orig_buffer_size = td->max_bs * max_units + MASK;
822
823 if (td->mem_type == MEM_MALLOC)
824 td->orig_buffer = malloc(td->orig_buffer_size);
825 else if (td->mem_type == MEM_SHM) {
826 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
827 if (td->shm_id < 0) {
828 td_verror(td, errno);
829 perror("shmget");
830 return 1;
831 }
832
833 td->orig_buffer = shmat(td->shm_id, NULL, 0);
834 if (td->orig_buffer == (void *) -1) {
835 td_verror(td, errno);
836 perror("shmat");
837 td->orig_buffer = NULL;
838 return 1;
839 }
840 } else if (td->mem_type == MEM_MMAP) {
841 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
842 if (td->orig_buffer == MAP_FAILED) {
843 td_verror(td, errno);
844 perror("mmap");
845 td->orig_buffer = NULL;
846 return 1;
847 }
848 }
849
ebac4655
JA
850 p = ALIGN(td->orig_buffer);
851 for (i = 0; i < max_units; i++) {
852 io_u = malloc(sizeof(*io_u));
853 memset(io_u, 0, sizeof(*io_u));
854 INIT_LIST_HEAD(&io_u->list);
855
856 io_u->buf = p + td->max_bs * i;
b1ff3403 857 io_u->index = i;
ebac4655
JA
858 list_add(&io_u->list, &td->io_u_freelist);
859 }
860
861 return 0;
862}
863
da86774e
JA
864static int switch_ioscheduler(struct thread_data *td)
865{
866 char tmp[256], tmp2[128];
867 FILE *f;
868 int ret;
869
870 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
871
872 f = fopen(tmp, "r+");
873 if (!f) {
874 td_verror(td, errno);
875 return 1;
876 }
877
878 /*
879 * Set io scheduler.
880 */
881 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
882 if (ferror(f) || ret != 1) {
883 td_verror(td, errno);
884 fclose(f);
885 return 1;
886 }
887
888 rewind(f);
889
890 /*
891 * Read back and check that the selected scheduler is now the default.
892 */
893 ret = fread(tmp, 1, sizeof(tmp), f);
894 if (ferror(f) || ret < 0) {
895 td_verror(td, errno);
896 fclose(f);
897 return 1;
898 }
899
900 sprintf(tmp2, "[%s]", td->ioscheduler);
901 if (!strstr(tmp, tmp2)) {
3b70d7e5 902 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
da86774e
JA
903 td_verror(td, EINVAL);
904 fclose(f);
905 return 1;
906 }
907
908 fclose(f);
909 return 0;
910}
911
ebac4655
JA
912static void clear_io_state(struct thread_data *td)
913{
53cdc686
JA
914 struct fio_file *f;
915 int i;
ebac4655 916
ebac4655
JA
917 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
918 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 919 td->zone_bytes = 0;
ebac4655 920
53cdc686
JA
921 for_each_file(td, f, i) {
922 f->last_pos = 0;
923 if (td->io_ops->flags & FIO_SYNCIO)
924 lseek(f->fd, SEEK_SET, 0);
925
926 if (f->file_map)
927 memset(f->file_map, 0, f->num_maps * sizeof(long));
928 }
ebac4655
JA
929}
930
906c8d75
JA
931/*
932 * Entry point for the thread based jobs. The process based jobs end up
933 * here as well, after a little setup.
934 */
ebac4655
JA
935static void *thread_main(void *data)
936{
937 struct thread_data *td = data;
ebac4655
JA
938
939 if (!td->use_thread)
940 setsid();
941
942 td->pid = getpid();
943
aea47d44
JA
944 INIT_LIST_HEAD(&td->io_u_freelist);
945 INIT_LIST_HEAD(&td->io_u_busylist);
946 INIT_LIST_HEAD(&td->io_hist_list);
947 INIT_LIST_HEAD(&td->io_log_list);
948
ebac4655
JA
949 if (init_io_u(td))
950 goto err;
951
952 if (fio_setaffinity(td) == -1) {
953 td_verror(td, errno);
954 goto err;
955 }
956
45bee283 957 if (td_io_init(td))
ebac4655
JA
958 goto err;
959
aea47d44
JA
960 if (init_iolog(td))
961 goto err;
962
ebac4655
JA
963 if (td->ioprio) {
964 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
965 td_verror(td, errno);
966 goto err;
967 }
968 }
969
1056eaad 970 if (nice(td->nice) == -1) {
b6f4d880
JA
971 td_verror(td, errno);
972 goto err;
973 }
974
75154845
JA
975 if (init_random_state(td))
976 goto err;
977
da86774e
JA
978 if (td->ioscheduler && switch_ioscheduler(td))
979 goto err;
980
75154845 981 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
982 fio_sem_up(&startup_sem);
983 fio_sem_down(&td->mutex);
ebac4655 984
53cdc686 985 if (!td->create_serialize && setup_files(td))
ebac4655
JA
986 goto err;
987
ebac4655
JA
988 gettimeofday(&td->epoch, NULL);
989
4e0ba8af
JA
990 if (td->exec_prerun)
991 system(td->exec_prerun);
992
ebac4655
JA
993 while (td->loops--) {
994 getrusage(RUSAGE_SELF, &td->ru_start);
995 gettimeofday(&td->start, NULL);
996 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
997
998 if (td->ratemin)
999 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1000
1001 clear_io_state(td);
1002 prune_io_piece_log(td);
1003
2866c82d 1004 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
1005 do_cpuio(td);
1006 else
1007 do_io(td);
ebac4655
JA
1008
1009 td->runtime[td->ddir] += mtime_since_now(&td->start);
aea47d44 1010 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
3d60d1ed
JA
1011 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1012
ebac4655
JA
1013 update_rusage_stat(td);
1014
1015 if (td->error || td->terminate)
1016 break;
1017
1018 if (td->verify == VERIFY_NONE)
1019 continue;
1020
1021 clear_io_state(td);
1022 gettimeofday(&td->start, NULL);
1023
1024 do_verify(td);
1025
1026 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1027
1028 if (td->error || td->terminate)
1029 break;
1030 }
1031
ebac4655
JA
1032 if (td->bw_log)
1033 finish_log(td, td->bw_log, "bw");
1034 if (td->slat_log)
1035 finish_log(td, td->slat_log, "slat");
1036 if (td->clat_log)
1037 finish_log(td, td->clat_log, "clat");
843a7413
JA
1038 if (td->write_iolog)
1039 write_iolog_close(td);
4e0ba8af
JA
1040 if (td->exec_postrun)
1041 system(td->exec_postrun);
ebac4655
JA
1042
1043 if (exitall_on_terminate)
1044 terminate_threads(td->groupid);
1045
1046err:
53cdc686 1047 close_files(td);
2866c82d 1048 close_ioengine(td);
ebac4655 1049 cleanup_io_u(td);
ebac4655
JA
1050 td_set_runstate(td, TD_EXITED);
1051 return NULL;
1052
1053}
1054
906c8d75
JA
1055/*
1056 * We cannot pass the td data into a forked process, so attach the td and
1057 * pass it to the thread worker.
1058 */
ebac4655
JA
1059static void *fork_main(int shmid, int offset)
1060{
1061 struct thread_data *td;
1062 void *data;
1063
1064 data = shmat(shmid, NULL, 0);
1065 if (data == (void *) -1) {
1066 perror("shmat");
1067 return NULL;
1068 }
1069
1070 td = data + offset * sizeof(struct thread_data);
1071 thread_main(td);
1072 shmdt(data);
1073 return NULL;
1074}
1075
906c8d75
JA
1076/*
1077 * Run over the job map and reap the threads that have exited, if any.
1078 */
ebac4655
JA
1079static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1080{
b990b5c0 1081 int i, cputhreads;
ebac4655
JA
1082
1083 /*
1084 * reap exited threads (TD_EXITED -> TD_REAPED)
1085 */
b990b5c0 1086 for (i = 0, cputhreads = 0; i < thread_number; i++) {
ebac4655
JA
1087 struct thread_data *td = &threads[i];
1088
2866c82d 1089 if (td->io_ops->flags & FIO_CPUIO)
b990b5c0
JA
1090 cputhreads++;
1091
ebac4655
JA
1092 if (td->runstate != TD_EXITED)
1093 continue;
1094
1095 td_set_runstate(td, TD_REAPED);
1096
1097 if (td->use_thread) {
1098 long ret;
1099
1100 if (pthread_join(td->thread, (void *) &ret))
1101 perror("thread_join");
1102 } else
1103 waitpid(td->pid, NULL, 0);
1104
1105 (*nr_running)--;
1106 (*m_rate) -= td->ratemin;
1107 (*t_rate) -= td->rate;
1108 }
b990b5c0
JA
1109
1110 if (*nr_running == cputhreads)
1111 terminate_threads(TERMINATE_ALL);
ebac4655
JA
1112}
1113
fcb6ade2
JA
1114static void fio_unpin_memory(void *pinned)
1115{
1116 if (pinned) {
1117 if (munlock(pinned, mlock_size) < 0)
1118 perror("munlock");
1119 munmap(pinned, mlock_size);
1120 }
1121}
1122
1123static void *fio_pin_memory(void)
1124{
32cd46a0 1125 unsigned long long phys_mem;
fcb6ade2
JA
1126 void *ptr;
1127
1128 if (!mlock_size)
1129 return NULL;
1130
1131 /*
1132 * Don't allow mlock of more than real_mem-128MB
1133 */
32cd46a0
JA
1134 phys_mem = os_phys_mem();
1135 if (phys_mem) {
1136 if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
1137 mlock_size = phys_mem - 128 * 1024 * 1024;
eb8bbf48 1138 fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20);
fcb6ade2
JA
1139 }
1140 }
1141
1142 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1143 if (!ptr) {
1144 perror("malloc locked mem");
1145 return NULL;
1146 }
1147 if (mlock(ptr, mlock_size) < 0) {
1148 munmap(ptr, mlock_size);
1149 perror("mlock");
1150 return NULL;
1151 }
1152
1153 return ptr;
1154}
1155
906c8d75
JA
1156/*
1157 * Main function for kicking off and reaping jobs, as needed.
1158 */
ebac4655
JA
1159static void run_threads(void)
1160{
ebac4655
JA
1161 struct thread_data *td;
1162 unsigned long spent;
1163 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2
JA
1164 void *mlocked_mem;
1165
1166 mlocked_mem = fio_pin_memory();
ebac4655 1167
c6ae0a5b
JA
1168 if (!terse_output) {
1169 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
1170 fflush(stdout);
1171 }
c04f7ec3 1172
4efa970e
JA
1173 signal(SIGINT, sig_handler);
1174 signal(SIGALRM, sig_handler);
1175
ebac4655
JA
1176 todo = thread_number;
1177 nr_running = 0;
1178 nr_started = 0;
1179 m_rate = t_rate = 0;
1180
1181 for (i = 0; i < thread_number; i++) {
1182 td = &threads[i];
1183
263e529f 1184 print_status_init(td->thread_number - 1);
ebac4655
JA
1185
1186 init_disk_util(td);
1187
1188 if (!td->create_serialize)
1189 continue;
1190
1191 /*
1192 * do file setup here so it happens sequentially,
1193 * we don't want X number of threads getting their
1194 * client data interspersed on disk
1195 */
53cdc686 1196 if (setup_files(td)) {
ebac4655
JA
1197 td_set_runstate(td, TD_REAPED);
1198 todo--;
1199 }
1200 }
1201
263e529f 1202 time_init();
ebac4655
JA
1203
1204 while (todo) {
75154845
JA
1205 struct thread_data *map[MAX_JOBS];
1206 struct timeval this_start;
1207 int this_jobs = 0, left;
1208
ebac4655
JA
1209 /*
1210 * create threads (TD_NOT_CREATED -> TD_CREATED)
1211 */
1212 for (i = 0; i < thread_number; i++) {
1213 td = &threads[i];
1214
1215 if (td->runstate != TD_NOT_CREATED)
1216 continue;
1217
1218 /*
1219 * never got a chance to start, killed by other
1220 * thread for some reason
1221 */
1222 if (td->terminate) {
1223 todo--;
1224 continue;
1225 }
1226
1227 if (td->start_delay) {
263e529f 1228 spent = mtime_since_genesis();
ebac4655
JA
1229
1230 if (td->start_delay * 1000 > spent)
1231 continue;
1232 }
1233
1234 if (td->stonewall && (nr_started || nr_running))
1235 break;
1236
75154845
JA
1237 /*
1238 * Set state to created. Thread will transition
1239 * to TD_INITIALIZED when it's done setting up.
1240 */
ebac4655 1241 td_set_runstate(td, TD_CREATED);
75154845 1242 map[this_jobs++] = td;
bbfd6b00 1243 fio_sem_init(&startup_sem, 1);
ebac4655
JA
1244 nr_started++;
1245
1246 if (td->use_thread) {
1247 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1248 perror("thread_create");
1249 nr_started--;
1250 }
1251 } else {
1252 if (fork())
bbfd6b00 1253 fio_sem_down(&startup_sem);
ebac4655
JA
1254 else {
1255 fork_main(shm_id, i);
1256 exit(0);
1257 }
1258 }
1259 }
1260
1261 /*
75154845
JA
1262 * Wait for the started threads to transition to
1263 * TD_INITIALIZED.
ebac4655 1264 */
75154845
JA
1265 gettimeofday(&this_start, NULL);
1266 left = this_jobs;
1267 while (left) {
1268 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1269 break;
1270
1271 usleep(100000);
1272
1273 for (i = 0; i < this_jobs; i++) {
1274 td = map[i];
1275 if (!td)
1276 continue;
b6f4d880 1277 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1278 map[i] = NULL;
1279 left--;
b6f4d880
JA
1280 } else if (td->runstate >= TD_EXITED) {
1281 map[i] = NULL;
1282 left--;
1283 todo--;
1284 nr_running++; /* work-around... */
75154845
JA
1285 }
1286 }
1287 }
1288
1289 if (left) {
3b70d7e5 1290 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1291 for (i = 0; i < this_jobs; i++) {
1292 td = map[i];
1293 if (!td)
1294 continue;
1295 kill(td->pid, SIGTERM);
1296 }
1297 break;
1298 }
1299
1300 /*
b6f4d880 1301 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1302 */
ebac4655
JA
1303 for (i = 0; i < thread_number; i++) {
1304 td = &threads[i];
1305
75154845 1306 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1307 continue;
1308
1309 td_set_runstate(td, TD_RUNNING);
1310 nr_running++;
1311 nr_started--;
1312 m_rate += td->ratemin;
1313 t_rate += td->rate;
75154845 1314 todo--;
bbfd6b00 1315 fio_sem_up(&td->mutex);
ebac4655
JA
1316 }
1317
1318 reap_threads(&nr_running, &t_rate, &m_rate);
1319
1320 if (todo)
1321 usleep(100000);
1322 }
1323
1324 while (nr_running) {
1325 reap_threads(&nr_running, &t_rate, &m_rate);
1326 usleep(10000);
1327 }
1328
1329 update_io_ticks();
fcb6ade2 1330 fio_unpin_memory(mlocked_mem);
ebac4655
JA
1331}
1332
ebac4655
JA
1333int main(int argc, char *argv[])
1334{
1335 if (parse_options(argc, argv))
1336 return 1;
1337
1338 if (!thread_number) {
3b70d7e5 1339 log_err("Nothing to do\n");
ebac4655
JA
1340 return 1;
1341 }
1342
1343 disk_util_timer_arm();
1344
1345 run_threads();
1346 show_run_stats();
1347
1348 return 0;
1349}