projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
[PATCH] Guard ioscheduler= option with FIO_HAVE_IOSCHED_SWITCH
[fio.git] / fio.c
CommitLineData
ebac4655
JA		 1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 *
20 */
ebac4655
JA		 21#include <unistd.h>
22#include <fcntl.h>
23#include <string.h>
ebac4655
JA		 24#include <signal.h>
25#include <time.h>
ebac4655 26#include <assert.h>
ebac4655
JA		 27#include <sys/stat.h>
28#include <sys/wait.h>
29#include <sys/ipc.h>
30#include <sys/shm.h>
31#include <sys/ioctl.h>
32#include <sys/mman.h>
33
34#include "fio.h"
35#include "os.h"
36
37#define MASK (4095)
38
39#define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
40
41int groupid = 0;
42int thread_number = 0;
43static char run_str[MAX_JOBS + 1];
44int shm_id = 0;
5289b847 45static struct timeval genesis;
ebac4655 46
ebac4655
JA		 47static void print_thread_status(void);
48
c04f7ec3
JA		 49extern unsigned long long mlock_size;
50
ebac4655
JA		 51/*
52 * thread life cycle
53 */
54enum {
55 TD_NOT_CREATED = 0,
56 TD_CREATED,
75154845 57 TD_INITIALIZED,
ebac4655
JA		 58 TD_RUNNING,
59 TD_VERIFYING,
60 TD_EXITED,
61 TD_REAPED,
62};
63
3d60d1ed 64#define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
ebac4655
JA		 65
66static sem_t startup_sem;
67
68#define TERMINATE_ALL (-1)
75154845 69#define JOB_START_TIMEOUT (5 * 1000)
ebac4655
JA		 70
71static void terminate_threads(int group_id)
72{
73 int i;
74
75 for (i = 0; i < thread_number; i++) {
76 struct thread_data *td = &threads[i];
77
78 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
79 td->terminate = 1;
80 td->start_delay = 0;
81 }
82 }
83}
84
85static void sig_handler(int sig)
86{
87 switch (sig) {
88 case SIGALRM:
89 update_io_ticks();
90 disk_util_timer_arm();
91 print_thread_status();
92 break;
93 default:
94 printf("\nfio: terminating on signal\n");
95 fflush(stdout);
96 terminate_threads(TERMINATE_ALL);
97 break;
98 }
99}
100
ebac4655
JA		 101static int random_map_free(struct thread_data *td, unsigned long long block)
102{
103 unsigned int idx = RAND_MAP_IDX(td, block);
104 unsigned int bit = RAND_MAP_BIT(td, block);
105
106 return (td->file_map[idx] & (1UL << bit)) == 0;
107}
108
109static int get_next_free_block(struct thread_data *td, unsigned long long *b)
110{
111 int i;
112
113 *b = 0;
114 i = 0;
115 while ((*b) * td->min_bs < td->io_size) {
116 if (td->file_map[i] != -1UL) {
117 *b += ffz(td->file_map[i]);
118 return 0;
119 }
120
121 *b += BLOCKS_PER_MAP;
122 i++;
123 }
124
125 return 1;
126}
127
128static void mark_random_map(struct thread_data *td, struct io_u *io_u)
129{
200bc855 130 unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
ebac4655
JA		 131 unsigned int blocks = 0;
132
133 while (blocks < (io_u->buflen / td->min_bs)) {
134 unsigned int idx, bit;
135
136 if (!random_map_free(td, block))
137 break;
138
139 idx = RAND_MAP_IDX(td, block);
140 bit = RAND_MAP_BIT(td, block);
141
142 assert(idx < td->num_maps);
143
144 td->file_map[idx] |= (1UL << bit);
145 block++;
146 blocks++;
147 }
148
149 if ((blocks * td->min_bs) < io_u->buflen)
150 io_u->buflen = blocks * td->min_bs;
151}
152
20dc95c4
JA		 153static int get_next_offset(struct thread_data *td, unsigned long long *offset)
154{
155 unsigned long long b, rb;
156 long r;
157
158 if (!td->sequential) {
085227ad 159 unsigned long long max_blocks = td->io_size / td->min_bs;
20dc95c4
JA		 160 int loops = 50;
161
162 do {
163 lrand48_r(&td->random_state, &r);
085227ad 164 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
20dc95c4
JA		 165 rb = b + (td->file_offset / td->min_bs);
166 loops--;
167 } while (!random_map_free(td, rb) && loops);
168
169 if (!loops) {
170 if (get_next_free_block(td, &b))
171 return 1;
172 }
173 } else
174 b = td->last_pos / td->min_bs;
175
176 *offset = (b * td->min_bs) + td->file_offset;
838a3cd3 177 if (*offset > td->real_file_size)
20dc95c4
JA		 178 return 1;
179
180 return 0;
181}
182
183static unsigned int get_next_buflen(struct thread_data *td)
184{
185 unsigned int buflen;
186 long r;
187
188 if (td->min_bs == td->max_bs)
189 buflen = td->min_bs;
190 else {
191 lrand48_r(&td->bsrange_state, &r);
192 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
193 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
194 }
195
196 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
197 buflen = td->io_size - td->this_io_bytes[td->ddir];
198
199 return buflen;
200}
201
ebac4655
JA		 202static int check_min_rate(struct thread_data *td, struct timeval *now)
203{
204 unsigned long spent;
205 unsigned long rate;
206 int ddir = td->ddir;
207
208 /*
209 * allow a 2 second settle period in the beginning
210 */
211 if (mtime_since(&td->start, now) < 2000)
212 return 0;
213
214 /*
215 * if rate blocks is set, sample is running
216 */
217 if (td->rate_bytes) {
218 spent = mtime_since(&td->lastrate, now);
219 if (spent < td->ratecycle)
220 return 0;
221
222 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
223 if (rate < td->ratemin) {
01452055 224 printf("%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
ebac4655
JA		 225 if (rate_quit)
226 terminate_threads(td->groupid);
227 return 1;
228 }
229 }
230
231 td->rate_bytes = td->this_io_bytes[ddir];
232 memcpy(&td->lastrate, now, sizeof(*now));
233 return 0;
234}
235
236static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
237{
238 if (!td->timeout)
239 return 0;
240 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
241 return 1;
242
243 return 0;
244}
245
246static void fill_random_bytes(struct thread_data *td,
247 unsigned char *p, unsigned int len)
248{
249 unsigned int todo;
250 double r;
251
252 while (len) {
253 drand48_r(&td->verify_state, &r);
254
255 /*
256 * lrand48_r seems to be broken and only fill the bottom
257 * 32-bits, even on 64-bit archs with 64-bit longs
258 */
259 todo = sizeof(r);
260 if (todo > len)
261 todo = len;
262
263 memcpy(p, &r, todo);
264
265 len -= todo;
266 p += todo;
267 }
268}
269
270static void hexdump(void *buffer, int len)
271{
272 unsigned char *p = buffer;
273 int i;
274
275 for (i = 0; i < len; i++)
276 printf("%02x", p[i]);
277 printf("\n");
278}
279
280static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
281{
282 unsigned char *p = (unsigned char *) io_u->buf;
283 unsigned long c;
ebac4655
JA		 284
285 p += sizeof(*hdr);
286 c = crc32(p, hdr->len - sizeof(*hdr));
ebac4655 287
22f78b32 288 if (c != hdr->crc32) {
ebac4655
JA		 289 fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
290 fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
22f78b32 291 return 1;
ebac4655
JA		 292 }
293
22f78b32 294 return 0;
ebac4655
JA		 295}
296
297static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
298{
299 unsigned char *p = (unsigned char *) io_u->buf;
300 struct md5_ctx md5_ctx;
ebac4655
JA		 301
302 memset(&md5_ctx, 0, sizeof(md5_ctx));
303 p += sizeof(*hdr);
304 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
305
22f78b32 306 if (memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash))) {
ebac4655
JA		 307 fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
308 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
309 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
22f78b32 310 return 1;
ebac4655
JA		 311 }
312
22f78b32 313 return 0;
ebac4655
JA		 314}
315
316static int verify_io_u(struct io_u *io_u)
317{
318 struct verify_header *hdr = (struct verify_header *) io_u->buf;
319 int ret;
320
321 if (hdr->fio_magic != FIO_HDR_MAGIC)
322 return 1;
323
324 if (hdr->verify_type == VERIFY_MD5)
325 ret = verify_io_u_md5(hdr, io_u);
326 else if (hdr->verify_type == VERIFY_CRC32)
327 ret = verify_io_u_crc32(hdr, io_u);
328 else {
329 fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
330 ret = 1;
331 }
332
333 return ret;
334}
335
336static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
337{
338 hdr->crc32 = crc32(p, len);
339}
340
341static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
342{
343 struct md5_ctx md5_ctx;
344
345 memset(&md5_ctx, 0, sizeof(md5_ctx));
346 md5_update(&md5_ctx, p, len);
347 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
348}
349
a6ccc7be 350static int get_rw_ddir(struct thread_data *td)
3d60d1ed 351{
a6ccc7be
JA		 352 if (td_rw(td)) {
353 struct timeval now;
354 unsigned long elapsed;
3d60d1ed 355
a6ccc7be
JA		 356 gettimeofday(&now, NULL);
357 elapsed = mtime_since_now(&td->rwmix_switch);
3d60d1ed 358
a6ccc7be
JA		 359 /*
360 * Check if it's time to seed a new data direction.
361 */
362 if (elapsed >= td->rwmixcycle) {
363 unsigned long v;
364 long r;
3d60d1ed 365
a6ccc7be
JA		 366 lrand48_r(&td->random_state, &r);
367 v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
368 if (v < td->rwmixread)
369 td->rwmix_ddir = DDIR_READ;
370 else
371 td->rwmix_ddir = DDIR_WRITE;
372 memcpy(&td->rwmix_switch, &now, sizeof(now));
373 }
374 return td->rwmix_ddir;
375 } else if (td_read(td))
3d60d1ed
JA		 376 return DDIR_READ;
377 else
378 return DDIR_WRITE;
379}
380
ebac4655
JA		 381/*
382 * fill body of io_u->buf with random data and add a header with the
22f78b32 383 * crc32 or md5 sum of that data.
ebac4655
JA		 384 */
385static void populate_io_u(struct thread_data *td, struct io_u *io_u)
386{
387 unsigned char *p = (unsigned char *) io_u->buf;
388 struct verify_header hdr;
389
390 hdr.fio_magic = FIO_HDR_MAGIC;
391 hdr.len = io_u->buflen;
392 p += sizeof(hdr);
393 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
394
395 if (td->verify == VERIFY_MD5) {
396 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
397 hdr.verify_type = VERIFY_MD5;
398 } else {
399 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
400 hdr.verify_type = VERIFY_CRC32;
401 }
402
403 memcpy(io_u->buf, &hdr, sizeof(hdr));
404}
405
aea47d44 406static int td_io_prep(struct thread_data *td, struct io_u *io_u)
20dc95c4 407{
20dc95c4
JA		 408 if (td->io_prep && td->io_prep(td, io_u))
409 return 1;
410
411 return 0;
412}
413
b1ff3403 414void put_io_u(struct thread_data *td, struct io_u *io_u)
ebac4655
JA		 415{
416 list_del(&io_u->list);
417 list_add(&io_u->list, &td->io_u_freelist);
418 td->cur_depth--;
419}
420
843a7413
JA		 421static int fill_io_u(struct thread_data *td, struct io_u *io_u)
422{
423 /*
424 * If using an iolog, grab next piece if any available.
425 */
426 if (td->read_iolog)
427 return read_iolog_get(td, io_u);
428
aea47d44
JA		 429 /*
430 * No log, let the seq/rand engine retrieve the next position.
431 */
432 if (!get_next_offset(td, &io_u->offset)) {
433 io_u->buflen = get_next_buflen(td);
434
435 if (io_u->buflen) {
436 io_u->ddir = get_rw_ddir(td);
843a7413
JA		 437
438 /*
439 * If using a write iolog, store this entry.
440 */
441 if (td->write_iolog)
442 write_iolog_put(td, io_u);
443
aea47d44
JA		 444 return 0;
445 }
446 }
447
448 return 1;
449}
450
22f78b32 451#define queue_full(td) list_empty(&(td)->io_u_freelist)
ebac4655 452
b1ff3403 453struct io_u *__get_io_u(struct thread_data *td)
ebac4655 454{
22f78b32 455 struct io_u *io_u = NULL;
ebac4655 456
22f78b32
JA		 457 if (!queue_full(td)) {
458 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
459
460 io_u->error = 0;
461 io_u->resid = 0;
462 list_del(&io_u->list);
463 list_add(&io_u->list, &td->io_u_busylist);
464 td->cur_depth++;
465 }
ebac4655 466
ebac4655
JA		 467 return io_u;
468}
469
ebac4655
JA		 470static struct io_u *get_io_u(struct thread_data *td)
471{
472 struct io_u *io_u;
473
474 io_u = __get_io_u(td);
475 if (!io_u)
476 return NULL;
477
20dc95c4
JA		 478 if (td->zone_bytes >= td->zone_size) {
479 td->zone_bytes = 0;
480 td->last_pos += td->zone_skip;
481 }
482
aea47d44 483 if (fill_io_u(td, io_u)) {
ebac4655
JA		 484 put_io_u(td, io_u);
485 return NULL;
486 }
487
838a3cd3
JA		 488 if (io_u->buflen + io_u->offset > td->real_file_size)
489 io_u->buflen = td->real_file_size - io_u->offset;
ebac4655
JA		 490
491 if (!io_u->buflen) {
492 put_io_u(td, io_u);
493 return NULL;
494 }
495
843a7413 496 if (!td->read_iolog && !td->sequential)
ebac4655
JA		 497 mark_random_map(td, io_u);
498
20dc95c4 499 td->last_pos += io_u->buflen;
ebac4655
JA		 500
501 if (td->verify != VERIFY_NONE)
502 populate_io_u(td, io_u);
503
aea47d44 504 if (td_io_prep(td, io_u)) {
ebac4655
JA		 505 put_io_u(td, io_u);
506 return NULL;
507 }
508
509 gettimeofday(&io_u->start_time, NULL);
510 return io_u;
511}
512
513static inline void td_set_runstate(struct thread_data *td, int runstate)
514{
515 td->old_runstate = td->runstate;
516 td->runstate = runstate;
517}
518
aea47d44 519static int get_next_verify(struct thread_data *td, struct io_u *io_u)
ebac4655
JA		 520{
521 struct io_piece *ipo;
522
22f78b32
JA		 523 if (!list_empty(&td->io_hist_list)) {
524 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
ebac4655 525
22f78b32 526 list_del(&ipo->list);
ebac4655 527
22f78b32
JA		 528 io_u->offset = ipo->offset;
529 io_u->buflen = ipo->len;
530 io_u->ddir = DDIR_READ;
531 free(ipo);
532 return 0;
533 }
534
535 return 1;
ebac4655
JA		 536}
537
ebac4655
JA		 538static int sync_td(struct thread_data *td)
539{
540 if (td->io_sync)
541 return td->io_sync(td);
542
543 return 0;
544}
545
546static int io_u_getevents(struct thread_data *td, int min, int max,
547 struct timespec *t)
548{
549 return td->io_getevents(td, min, max, t);
550}
551
552static int io_u_queue(struct thread_data *td, struct io_u *io_u)
553{
554 gettimeofday(&io_u->issue_time, NULL);
555
556 return td->io_queue(td, io_u);
557}
558
559#define iocb_time(iocb) ((unsigned long) (iocb)->data)
560
561static void io_completed(struct thread_data *td, struct io_u *io_u,
562 struct io_completion_data *icd)
563{
564 struct timeval e;
565 unsigned long msec;
566
567 gettimeofday(&e, NULL);
568
569 if (!io_u->error) {
20dc95c4 570 unsigned int bytes = io_u->buflen - io_u->resid;
3d60d1ed 571 const int idx = io_u->ddir;
ebac4655
JA		 572
573 td->io_blocks[idx]++;
20dc95c4
JA		 574 td->io_bytes[idx] += bytes;
575 td->zone_bytes += bytes;
576 td->this_io_bytes[idx] += bytes;
ebac4655
JA		 577
578 msec = mtime_since(&io_u->issue_time, &e);
579
3d60d1ed
JA		 580 add_clat_sample(td, idx, msec);
581 add_bw_sample(td, idx);
ebac4655 582
3d60d1ed 583 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
ebac4655
JA		 584 log_io_piece(td, io_u);
585
20dc95c4 586 icd->bytes_done[idx] += bytes;
ebac4655
JA		 587 } else
588 icd->error = io_u->error;
589}
590
591static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
592{
593 struct io_u *io_u;
594 int i;
595
596 icd->error = 0;
597 icd->bytes_done[0] = icd->bytes_done[1] = 0;
598
599 for (i = 0; i < icd->nr; i++) {
600 io_u = td->io_event(td, i);
601
602 io_completed(td, io_u, icd);
603 put_io_u(td, io_u);
604 }
605}
606
607static void cleanup_pending_aio(struct thread_data *td)
608{
609 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
610 struct list_head *entry, *n;
611 struct io_completion_data icd;
612 struct io_u *io_u;
613 int r;
614
615 /*
616 * get immediately available events, if any
617 */
618 r = io_u_getevents(td, 0, td->cur_depth, &ts);
619 if (r > 0) {
620 icd.nr = r;
621 ios_completed(td, &icd);
622 }
623
624 /*
625 * now cancel remaining active events
626 */
627 if (td->io_cancel) {
628 list_for_each_safe(entry, n, &td->io_u_busylist) {
629 io_u = list_entry(entry, struct io_u, list);
630
631 r = td->io_cancel(td, io_u);
632 if (!r)
633 put_io_u(td, io_u);
634 }
635 }
636
637 if (td->cur_depth) {
638 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
639 if (r > 0) {
640 icd.nr = r;
641 ios_completed(td, &icd);
642 }
643 }
644}
645
646static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
647{
648 struct io_u *v_io_u = *io_u;
649 int ret = 0;
650
651 if (v_io_u) {
652 ret = verify_io_u(v_io_u);
653 put_io_u(td, v_io_u);
654 *io_u = NULL;
655 }
656
657 return ret;
658}
659
660static void do_verify(struct thread_data *td)
661{
662 struct timeval t;
663 struct io_u *io_u, *v_io_u = NULL;
664 struct io_completion_data icd;
665 int ret;
666
667 td_set_runstate(td, TD_VERIFYING);
668
669 do {
670 if (td->terminate)
671 break;
672
673 gettimeofday(&t, NULL);
674 if (runtime_exceeded(td, &t))
675 break;
676
677 io_u = __get_io_u(td);
678 if (!io_u)
679 break;
680
aea47d44 681 if (get_next_verify(td, io_u)) {
ebac4655
JA		 682 put_io_u(td, io_u);
683 break;
684 }
685
aea47d44 686 if (td_io_prep(td, io_u)) {
ebac4655
JA		 687 put_io_u(td, io_u);
688 break;
689 }
690
691 ret = io_u_queue(td, io_u);
692 if (ret) {
693 put_io_u(td, io_u);
694 td_verror(td, ret);
695 break;
696 }
697
698 /*
699 * we have one pending to verify, do that while
700 * we are doing io on the next one
701 */
702 if (do_io_u_verify(td, &v_io_u))
703 break;
704
705 ret = io_u_getevents(td, 1, 1, NULL);
706 if (ret != 1) {
707 if (ret < 0)
708 td_verror(td, ret);
709 break;
710 }
711
712 v_io_u = td->io_event(td, 0);
713 icd.nr = 1;
714 icd.error = 0;
715 io_completed(td, v_io_u, &icd);
716
717 if (icd.error) {
718 td_verror(td, icd.error);
719 put_io_u(td, v_io_u);
720 v_io_u = NULL;
721 break;
722 }
723
ebac4655
JA		 724 /*
725 * if we can't submit more io, we need to verify now
726 */
727 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
728 break;
729
730 } while (1);
731
732 do_io_u_verify(td, &v_io_u);
733
734 if (td->cur_depth)
735 cleanup_pending_aio(td);
736
737 td_set_runstate(td, TD_RUNNING);
738}
739
740static void do_io(struct thread_data *td)
741{
742 struct io_completion_data icd;
743 struct timeval s, e;
744 unsigned long usec;
745
746 while (td->this_io_bytes[td->ddir] < td->io_size) {
747 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
748 struct timespec *timeout;
749 int ret, min_evts = 0;
750 struct io_u *io_u;
751
752 if (td->terminate)
753 break;
754
755 io_u = get_io_u(td);
756 if (!io_u)
757 break;
758
759 memcpy(&s, &io_u->start_time, sizeof(s));
760
761 ret = io_u_queue(td, io_u);
762 if (ret) {
763 put_io_u(td, io_u);
764 td_verror(td, ret);
765 break;
766 }
767
768 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
769
770 if (td->cur_depth < td->iodepth) {
771 timeout = &ts;
772 min_evts = 0;
773 } else {
774 timeout = NULL;
775 min_evts = 1;
776 }
777
778 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
779 if (ret < 0) {
780 td_verror(td, ret);
781 break;
782 } else if (!ret)
783 continue;
784
785 icd.nr = ret;
786 ios_completed(td, &icd);
787 if (icd.error) {
788 td_verror(td, icd.error);
789 break;
790 }
791
792 /*
793 * the rate is batched for now, it should work for batches
794 * of completions except the very first one which may look
795 * a little bursty
796 */
797 gettimeofday(&e, NULL);
798 usec = utime_since(&s, &e);
799
800 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
801
802 if (check_min_rate(td, &e)) {
803 td_verror(td, ENOMEM);
804 break;
805 }
806
807 if (runtime_exceeded(td, &e))
808 break;
809
810 if (td->thinktime)
811 usec_sleep(td, td->thinktime);
812
813 if (should_fsync(td) && td->fsync_blocks &&
814 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
815 sync_td(td);
816 }
817
818 if (td->cur_depth)
819 cleanup_pending_aio(td);
820
fc1a4713 821 if (should_fsync(td) && td->end_fsync)
ebac4655
JA		 822 sync_td(td);
823}
824
825static void cleanup_io(struct thread_data *td)
826{
827 if (td->io_cleanup)
828 td->io_cleanup(td);
829}
830
831static int init_io(struct thread_data *td)
832{
833 if (td->io_engine == FIO_SYNCIO)
834 return fio_syncio_init(td);
835 else if (td->io_engine == FIO_MMAPIO)
836 return fio_mmapio_init(td);
837 else if (td->io_engine == FIO_LIBAIO)
838 return fio_libaio_init(td);
839 else if (td->io_engine == FIO_POSIXAIO)
840 return fio_posixaio_init(td);
841 else if (td->io_engine == FIO_SGIO)
842 return fio_sgio_init(td);
8756e4d4
JA		 843 else if (td->io_engine == FIO_SPLICEIO)
844 return fio_spliceio_init(td);
ebac4655
JA		 845 else {
846 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
847 return 1;
848 }
849}
850
851static void cleanup_io_u(struct thread_data *td)
852{
853 struct list_head *entry, *n;
854 struct io_u *io_u;
855
856 list_for_each_safe(entry, n, &td->io_u_freelist) {
857 io_u = list_entry(entry, struct io_u, list);
858
859 list_del(&io_u->list);
860 free(io_u);
861 }
862
863 if (td->mem_type == MEM_MALLOC)
864 free(td->orig_buffer);
865 else if (td->mem_type == MEM_SHM) {
866 struct shmid_ds sbuf;
867
868 shmdt(td->orig_buffer);
869 shmctl(td->shm_id, IPC_RMID, &sbuf);
870 } else if (td->mem_type == MEM_MMAP)
871 munmap(td->orig_buffer, td->orig_buffer_size);
872 else
873 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
874
875 td->orig_buffer = NULL;
876}
877
878static int init_io_u(struct thread_data *td)
879{
880 struct io_u *io_u;
881 int i, max_units;
882 char *p;
883
884 if (td->io_engine & FIO_SYNCIO)
885 max_units = 1;
886 else
887 max_units = td->iodepth;
888
889 td->orig_buffer_size = td->max_bs * max_units + MASK;
890
891 if (td->mem_type == MEM_MALLOC)
892 td->orig_buffer = malloc(td->orig_buffer_size);
893 else if (td->mem_type == MEM_SHM) {
894 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
895 if (td->shm_id < 0) {
896 td_verror(td, errno);
897 perror("shmget");
898 return 1;
899 }
900
901 td->orig_buffer = shmat(td->shm_id, NULL, 0);
902 if (td->orig_buffer == (void *) -1) {
903 td_verror(td, errno);
904 perror("shmat");
905 td->orig_buffer = NULL;
906 return 1;
907 }
908 } else if (td->mem_type == MEM_MMAP) {
909 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
910 if (td->orig_buffer == MAP_FAILED) {
911 td_verror(td, errno);
912 perror("mmap");
913 td->orig_buffer = NULL;
914 return 1;
915 }
916 }
917
ebac4655
JA		 918 p = ALIGN(td->orig_buffer);
919 for (i = 0; i < max_units; i++) {
920 io_u = malloc(sizeof(*io_u));
921 memset(io_u, 0, sizeof(*io_u));
922 INIT_LIST_HEAD(&io_u->list);
923
924 io_u->buf = p + td->max_bs * i;
b1ff3403 925 io_u->index = i;
ebac4655
JA		 926 list_add(&io_u->list, &td->io_u_freelist);
927 }
928
929 return 0;
930}
931
4b5bd14c
JA		 932static void cleanup_allocs(struct thread_data *td)
933{
934 if (td->directory)
935 free(td->directory);
936 if (td->iolog_file)
937 free(td->iolog_file);
938 if (td->exec_prerun)
939 free(td->exec_prerun);
940 if (td->exec_postrun)
941 free(td->exec_postrun);
942 if (td->ioscheduler)
943 free(td->ioscheduler);
944 if (td->sysfs_root)
945 free(td->sysfs_root);
946}
947
ebac4655
JA		 948static int create_file(struct thread_data *td, unsigned long long size,
949 int extend)
950{
951 unsigned long long left;
952 unsigned int bs;
953 int r, oflags;
954 char *b;
955
956 /*
957 * unless specifically asked for overwrite, let normal io extend it
958 */
959 if (td_write(td) && !td->overwrite)
960 return 0;
961
962 if (!size) {
963 fprintf(stderr, "Need size for create\n");
964 td_verror(td, EINVAL);
965 return 1;
966 }
967
968 if (!extend) {
969 oflags = O_CREAT | O_TRUNC;
01452055 970 printf("%s: Laying out IO file (%LuMiB)\n", td->name, size >> 20);
ebac4655
JA		 971 } else {
972 oflags = O_APPEND;
01452055 973 printf("%s: Extending IO file (%Lu -> %LuMiB)\n", td->name, (td->file_size - size) >> 20, td->file_size >> 20);
ebac4655
JA		 974 }
975
976 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
977 if (td->fd < 0) {
978 td_verror(td, errno);
979 return 1;
980 }
981
982 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
983 td_verror(td, errno);
984 return 1;
985 }
986
987 td->io_size = td->file_size;
988 b = malloc(td->max_bs);
989 memset(b, 0, td->max_bs);
990
991 left = size;
992 while (left && !td->terminate) {
993 bs = td->max_bs;
994 if (bs > left)
995 bs = left;
996
997 r = write(td->fd, b, bs);
998
999 if (r == (int) bs) {
1000 left -= bs;
1001 continue;
1002 } else {
1003 if (r < 0)
1004 td_verror(td, errno);
1005 else
1006 td_verror(td, EIO);
1007
1008 break;
1009 }
1010 }
1011
1012 if (td->terminate)
1013 unlink(td->file_name);
1014 else if (td->create_fsync)
1015 fsync(td->fd);
1016
1017 close(td->fd);
1018 td->fd = -1;
1019 free(b);
1020 return 0;
1021}
1022
1023static int file_size(struct thread_data *td)
1024{
1025 struct stat st;
1026
1027 if (fstat(td->fd, &st) == -1) {
1028 td_verror(td, errno);
1029 return 1;
1030 }
1031
838a3cd3
JA		 1032 td->real_file_size = st.st_size;
1033
1034 if (!td->file_size || td->file_size > td->real_file_size)
1035 td->file_size = td->real_file_size;
ebac4655 1036
e8e387c1 1037 td->file_size -= td->file_offset;
ebac4655
JA		 1038 return 0;
1039}
1040
1041static int bdev_size(struct thread_data *td)
1042{
9104f874 1043 unsigned long long bytes;
ebac4655
JA		 1044 int r;
1045
1046 r = blockdev_size(td->fd, &bytes);
1047 if (r) {
1048 td_verror(td, r);
1049 return 1;
1050 }
1051
838a3cd3
JA		 1052 td->real_file_size = bytes;
1053
ebac4655
JA		 1054 /*
1055 * no extend possibilities, so limit size to device size if too large
1056 */
838a3cd3
JA		 1057 if (!td->file_size || td->file_size > td->real_file_size)
1058 td->file_size = td->real_file_size;
ebac4655 1059
e8e387c1 1060 td->file_size -= td->file_offset;
ebac4655
JA		 1061 return 0;
1062}
1063
1064static int get_file_size(struct thread_data *td)
1065{
0af7b542 1066 int ret = 0;
ebac4655
JA		 1067
1068 if (td->filetype == FIO_TYPE_FILE)
1069 ret = file_size(td);
0af7b542 1070 else if (td->filetype == FIO_TYPE_BD)
ebac4655 1071 ret = bdev_size(td);
0af7b542
JA		 1072 else
1073 td->real_file_size = -1;
ebac4655
JA		 1074
1075 if (ret)
1076 return ret;
1077
e8e387c1 1078 if (td->file_offset > td->real_file_size) {
01452055 1079 fprintf(stderr, "%s: offset extends end (%Lu > %Lu)\n", td->name, td->file_offset, td->real_file_size);
ebac4655
JA		 1080 return 1;
1081 }
1082
838a3cd3 1083 td->io_size = td->file_size;
ebac4655 1084 if (td->io_size == 0) {
01452055 1085 fprintf(stderr, "%s: no io blocks\n", td->name);
ebac4655
JA		 1086 td_verror(td, EINVAL);
1087 return 1;
1088 }
1089
20dc95c4
JA		 1090 if (!td->zone_size)
1091 td->zone_size = td->io_size;
1092
ebac4655
JA		 1093 td->total_io_size = td->io_size * td->loops;
1094 return 0;
1095}
1096
1097static int setup_file_mmap(struct thread_data *td)
1098{
1099 int flags;
1100
3d60d1ed
JA		 1101 if (td_rw(td))
1102 flags = PROT_READ | PROT_WRITE;
1103 else if (td_write(td)) {
ebac4655
JA		 1104 flags = PROT_WRITE;
1105
1106 if (td->verify != VERIFY_NONE)
1107 flags |= PROT_READ;
3d60d1ed
JA		 1108 } else
1109 flags = PROT_READ;
ebac4655
JA		 1110
1111 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1112 if (td->mmap == MAP_FAILED) {
1113 td->mmap = NULL;
1114 td_verror(td, errno);
1115 return 1;
1116 }
1117
1118 if (td->invalidate_cache) {
1119 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1120 td_verror(td, errno);
1121 return 1;
1122 }
1123 }
1124
1125 if (td->sequential) {
1126 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1127 td_verror(td, errno);
1128 return 1;
1129 }
1130 } else {
1131 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1132 td_verror(td, errno);
1133 return 1;
1134 }
1135 }
1136
1137 return 0;
1138}
1139
1140static int setup_file_plain(struct thread_data *td)
1141{
1142 if (td->invalidate_cache) {
1143 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1144 td_verror(td, errno);
1145 return 1;
1146 }
1147 }
1148
1149 if (td->sequential) {
1150 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1151 td_verror(td, errno);
1152 return 1;
1153 }
1154 } else {
1155 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1156 td_verror(td, errno);
1157 return 1;
1158 }
1159 }
1160
1161 return 0;
1162}
1163
1164static int setup_file(struct thread_data *td)
1165{
1166 struct stat st;
1167 int flags = 0;
1168
1169 if (stat(td->file_name, &st) == -1) {
1170 if (errno != ENOENT) {
1171 td_verror(td, errno);
1172 return 1;
1173 }
1174 if (!td->create_file) {
1175 td_verror(td, ENOENT);
1176 return 1;
1177 }
1178 if (create_file(td, td->file_size, 0))
1179 return 1;
1180 } else if (td->filetype == FIO_TYPE_FILE) {
6a0106a0 1181 if (st.st_size < (off_t) td->file_size) {
ebac4655
JA		 1182 if (create_file(td, td->file_size - st.st_size, 1))
1183 return 1;
1184 }
1185 }
1186
1187 if (td->odirect)
1188 flags |= O_DIRECT;
1189
3d60d1ed 1190 if (td_write(td) || td_rw(td)) {
ebac4655
JA		 1191 if (td->filetype == FIO_TYPE_FILE) {
1192 if (!td->overwrite)
1193 flags |= O_TRUNC;
1194
1195 flags |= O_CREAT;
1196 }
1197 if (td->sync_io)
1198 flags |= O_SYNC;
1199
1200 flags |= O_RDWR;
1201
1202 td->fd = open(td->file_name, flags, 0600);
3d60d1ed
JA		 1203 } else {
1204 if (td->filetype == FIO_TYPE_CHAR)
1205 flags |= O_RDWR;
1206 else
1207 flags |= O_RDONLY;
1208
1209 td->fd = open(td->file_name, flags);
ebac4655
JA		 1210 }
1211
1212 if (td->fd == -1) {
1213 td_verror(td, errno);
1214 return 1;
1215 }
1216
1217 if (get_file_size(td))
1218 return 1;
1219
1220 if (td->io_engine != FIO_MMAPIO)
1221 return setup_file_plain(td);
1222 else
1223 return setup_file_mmap(td);
1224}
1225
da86774e
JA		 1226static int switch_ioscheduler(struct thread_data *td)
1227{
1228 char tmp[256], tmp2[128];
1229 FILE *f;
1230 int ret;
1231
1232 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1233
1234 f = fopen(tmp, "r+");
1235 if (!f) {
1236 td_verror(td, errno);
1237 return 1;
1238 }
1239
1240 /*
1241 * Set io scheduler.
1242 */
1243 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
1244 if (ferror(f) || ret != 1) {
1245 td_verror(td, errno);
1246 fclose(f);
1247 return 1;
1248 }
1249
1250 rewind(f);
1251
1252 /*
1253 * Read back and check that the selected scheduler is now the default.
1254 */
1255 ret = fread(tmp, 1, sizeof(tmp), f);
1256 if (ferror(f) || ret < 0) {
1257 td_verror(td, errno);
1258 fclose(f);
1259 return 1;
1260 }
1261
1262 sprintf(tmp2, "[%s]", td->ioscheduler);
1263 if (!strstr(tmp, tmp2)) {
1264 fprintf(stderr, "fio: io scheduler %s not found\n", td->ioscheduler);
1265 td_verror(td, EINVAL);
1266 fclose(f);
1267 return 1;
1268 }
1269
1270 fclose(f);
1271 return 0;
1272}
1273
ebac4655
JA		 1274static void clear_io_state(struct thread_data *td)
1275{
1276 if (td->io_engine == FIO_SYNCIO)
1277 lseek(td->fd, SEEK_SET, 0);
1278
20dc95c4 1279 td->last_pos = 0;
ebac4655
JA		 1280 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1281 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 1282 td->zone_bytes = 0;
ebac4655
JA		 1283
1284 if (td->file_map)
1285 memset(td->file_map, 0, td->num_maps * sizeof(long));
1286}
1287
ebac4655
JA		 1288static void *thread_main(void *data)
1289{
1290 struct thread_data *td = data;
ebac4655
JA		 1291
1292 if (!td->use_thread)
1293 setsid();
1294
1295 td->pid = getpid();
1296
aea47d44
JA		 1297 INIT_LIST_HEAD(&td->io_u_freelist);
1298 INIT_LIST_HEAD(&td->io_u_busylist);
1299 INIT_LIST_HEAD(&td->io_hist_list);
1300 INIT_LIST_HEAD(&td->io_log_list);
1301
ebac4655
JA		 1302 if (init_io_u(td))
1303 goto err;
1304
1305 if (fio_setaffinity(td) == -1) {
1306 td_verror(td, errno);
1307 goto err;
1308 }
1309
1310 if (init_io(td))
1311 goto err;
1312
aea47d44
JA		 1313 if (init_iolog(td))
1314 goto err;
1315
ebac4655
JA		 1316 if (td->ioprio) {
1317 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1318 td_verror(td, errno);
1319 goto err;
1320 }
1321 }
1322
b6f4d880
JA		 1323 if (nice(td->nice) < 0) {
1324 td_verror(td, errno);
1325 goto err;
1326 }
1327
75154845
JA		 1328 if (init_random_state(td))
1329 goto err;
1330
da86774e
JA		 1331 if (td->ioscheduler && switch_ioscheduler(td))
1332 goto err;
1333
75154845 1334 td_set_runstate(td, TD_INITIALIZED);
ebac4655
JA		 1335 sem_post(&startup_sem);
1336 sem_wait(&td->mutex);
1337
1338 if (!td->create_serialize && setup_file(td))
1339 goto err;
1340
ebac4655
JA		 1341 gettimeofday(&td->epoch, NULL);
1342
4e0ba8af
JA		 1343 if (td->exec_prerun)
1344 system(td->exec_prerun);
1345
ebac4655
JA		 1346 while (td->loops--) {
1347 getrusage(RUSAGE_SELF, &td->ru_start);
1348 gettimeofday(&td->start, NULL);
1349 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1350
1351 if (td->ratemin)
1352 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1353
1354 clear_io_state(td);
1355 prune_io_piece_log(td);
1356
1357 do_io(td);
1358
1359 td->runtime[td->ddir] += mtime_since_now(&td->start);
aea47d44 1360 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
3d60d1ed
JA		 1361 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1362
ebac4655
JA		 1363 update_rusage_stat(td);
1364
1365 if (td->error || td->terminate)
1366 break;
1367
1368 if (td->verify == VERIFY_NONE)
1369 continue;
1370
1371 clear_io_state(td);
1372 gettimeofday(&td->start, NULL);
1373
1374 do_verify(td);
1375
1376 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1377
1378 if (td->error || td->terminate)
1379 break;
1380 }
1381
ebac4655
JA		 1382 if (td->bw_log)
1383 finish_log(td, td->bw_log, "bw");
1384 if (td->slat_log)
1385 finish_log(td, td->slat_log, "slat");
1386 if (td->clat_log)
1387 finish_log(td, td->clat_log, "clat");
843a7413
JA		 1388 if (td->write_iolog)
1389 write_iolog_close(td);
4e0ba8af
JA		 1390 if (td->exec_postrun)
1391 system(td->exec_postrun);
ebac4655
JA		 1392
1393 if (exitall_on_terminate)
1394 terminate_threads(td->groupid);
1395
1396err:
1397 if (td->fd != -1) {
1398 close(td->fd);
1399 td->fd = -1;
1400 }
1401 if (td->mmap)
1402 munmap(td->mmap, td->file_size);
4b5bd14c 1403 cleanup_allocs(td);
ebac4655
JA		 1404 cleanup_io(td);
1405 cleanup_io_u(td);
ebac4655
JA		 1406 td_set_runstate(td, TD_EXITED);
1407 return NULL;
1408
1409}
1410
1411static void *fork_main(int shmid, int offset)
1412{
1413 struct thread_data *td;
1414 void *data;
1415
1416 data = shmat(shmid, NULL, 0);
1417 if (data == (void *) -1) {
1418 perror("shmat");
1419 return NULL;
1420 }
1421
1422 td = data + offset * sizeof(struct thread_data);
1423 thread_main(td);
1424 shmdt(data);
1425 return NULL;
1426}
1427
ebac4655
JA		 1428static void check_str_update(struct thread_data *td)
1429{
1430 char c = run_str[td->thread_number - 1];
1431
1432 if (td->runstate == td->old_runstate)
1433 return;
1434
1435 switch (td->runstate) {
1436 case TD_REAPED:
1437 c = '_';
1438 break;
1439 case TD_EXITED:
1440 c = 'E';
1441 break;
1442 case TD_RUNNING:
3d60d1ed
JA		 1443 if (td_rw(td)) {
1444 if (td->sequential)
1445 c = 'M';
1446 else
1447 c = 'm';
1448 } else if (td_read(td)) {
ebac4655
JA		 1449 if (td->sequential)
1450 c = 'R';
1451 else
1452 c = 'r';
1453 } else {
1454 if (td->sequential)
1455 c = 'W';
1456 else
1457 c = 'w';
1458 }
1459 break;
1460 case TD_VERIFYING:
1461 c = 'V';
1462 break;
1463 case TD_CREATED:
1464 c = 'C';
1465 break;
75154845
JA		 1466 case TD_INITIALIZED:
1467 c = 'I';
1468 break;
ebac4655
JA		 1469 case TD_NOT_CREATED:
1470 c = 'P';
1471 break;
1472 default:
1473 printf("state %d\n", td->runstate);
1474 }
1475
1476 run_str[td->thread_number - 1] = c;
1477 td->old_runstate = td->runstate;
1478}
1479
5289b847
JA		 1480static void eta_to_str(char *str, int eta_sec)
1481{
1482 unsigned int d, h, m, s;
1483 static int always_d, always_h;
1484
1485 d = h = m = s = 0;
1486
1487 s = eta_sec % 60;
1488 eta_sec /= 60;
1489 m = eta_sec % 60;
1490 eta_sec /= 60;
1491 h = eta_sec % 24;
1492 eta_sec /= 24;
1493 d = eta_sec;
1494
1495 if (d || always_d) {
1496 always_d = 1;
1497 str += sprintf(str, "%02dd:", d);
1498 }
1499 if (h || always_h) {
1500 always_h = 1;
1501 str += sprintf(str, "%02dh:", h);
1502 }
1503
1504 str += sprintf(str, "%02dm:", m);
1505 str += sprintf(str, "%02ds", s);
1506}
1507
6a0106a0
JA		 1508static int thread_eta(struct thread_data *td, unsigned long elapsed)
1509{
1510 unsigned long long bytes_total, bytes_done;
1511 unsigned int eta_sec = 0;
1512
1513 bytes_total = td->total_io_size;
3d60d1ed
JA		 1514
1515 /*
1516 * if writing, bytes_total will be twice the size. If mixing,
1517 * assume a 50/50 split and thus bytes_total will be 50% larger.
1518 */
1519 if (td->verify) {
1520 if (td_rw(td))
1521 bytes_total = bytes_total * 3 / 2;
1522 else
1523 bytes_total <<= 1;
1524 }
6a0106a0
JA		 1525 if (td->zone_size && td->zone_skip)
1526 bytes_total /= (td->zone_skip / td->zone_size);
1527
1528 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
1529 double perc;
8b611c34 1530
6a0106a0
JA		 1531 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
1532 perc = (double) bytes_done / (double) bytes_total;
1533 if (perc > 1.0)
1534 perc = 1.0;
1535
1536 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
1537
8b611c34 1538 if (td->timeout && eta_sec > (td->timeout - elapsed))
6a0106a0 1539 eta_sec = td->timeout - elapsed;
75154845
JA		 1540 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
1541 || td->runstate == TD_INITIALIZED) {
6a0106a0
JA		 1542 int t_eta = 0, r_eta = 0;
1543
1544 /*
1545 * We can only guess - assume it'll run the full timeout
1546 * if given, otherwise assume it'll run at the specified rate.
1547 */
1548 if (td->timeout)
1549 t_eta = td->timeout + td->start_delay - elapsed;
1550 if (td->rate) {
1551 r_eta = (bytes_total / 1024) / td->rate;
1552 r_eta += td->start_delay - elapsed;
1553 }
1554
1555 if (r_eta && t_eta)
1556 eta_sec = min(r_eta, t_eta);
1557 else if (r_eta)
1558 eta_sec = r_eta;
1559 else if (t_eta)
1560 eta_sec = t_eta;
1561 else
1562 eta_sec = INT_MAX;
1563 } else {
1564 /*
1565 * thread is already done
1566 */
1567 eta_sec = 0;
1568 }
1569
1570 return eta_sec;
1571}
1572
ebac4655
JA		 1573static void print_thread_status(void)
1574{
6a0106a0
JA		 1575 unsigned long elapsed = time_since_now(&genesis);
1576 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
5289b847 1577 char eta_str[32];
6a0106a0
JA		 1578 double perc = 0.0;
1579
1580 eta_secs = malloc(thread_number * sizeof(int));
1581 memset(eta_secs, 0, thread_number * sizeof(int));
ebac4655 1582
ebac4655
JA		 1583 nr_running = t_rate = m_rate = 0;
1584 for (i = 0; i < thread_number; i++) {
1585 struct thread_data *td = &threads[i];
1586
1587 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
1588 nr_running++;
1589 t_rate += td->rate;
1590 m_rate += td->ratemin;
1591 }
1592
6a0106a0
JA		 1593 if (elapsed >= 3)
1594 eta_secs[i] = thread_eta(td, elapsed);
8b611c34
JA		 1595 else
1596 eta_secs[i] = INT_MAX;
ebac4655
JA		 1597
1598 check_str_update(td);
1599 }
1600
6a0106a0
JA		 1601 if (exitall_on_terminate)
1602 eta_sec = INT_MAX;
1603 else
1604 eta_sec = 0;
5289b847 1605
6a0106a0
JA		 1606 for (i = 0; i < thread_number; i++) {
1607 if (exitall_on_terminate) {
1608 if (eta_secs[i] < eta_sec)
1609 eta_sec = eta_secs[i];
1610 } else {
1611 if (eta_secs[i] > eta_sec)
1612 eta_sec = eta_secs[i];
5289b847 1613 }
6a0106a0 1614 }
5289b847 1615
8b611c34 1616 if (eta_sec != INT_MAX && elapsed) {
6a0106a0
JA		 1617 perc = (double) elapsed / (double) (elapsed + eta_sec);
1618 eta_to_str(eta_str, eta_sec);
ebac4655
JA		 1619 }
1620
5289b847 1621 printf("Threads now running (%d)", nr_running);
ebac4655
JA		 1622 if (m_rate || t_rate)
1623 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
8b611c34 1624 if (eta_sec != INT_MAX) {
6a0106a0
JA		 1625 perc *= 100.0;
1626 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
1627 }
5289b847 1628 printf("\r");
ebac4655 1629 fflush(stdout);
6a0106a0 1630 free(eta_secs);
ebac4655
JA		 1631}
1632
1633static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1634{
1635 int i;
1636
1637 /*
1638 * reap exited threads (TD_EXITED -> TD_REAPED)
1639 */
1640 for (i = 0; i < thread_number; i++) {
1641 struct thread_data *td = &threads[i];
1642
1643 if (td->runstate != TD_EXITED)
1644 continue;
1645
1646 td_set_runstate(td, TD_REAPED);
1647
1648 if (td->use_thread) {
1649 long ret;
1650
1651 if (pthread_join(td->thread, (void *) &ret))
1652 perror("thread_join");
1653 } else
1654 waitpid(td->pid, NULL, 0);
1655
1656 (*nr_running)--;
1657 (*m_rate) -= td->ratemin;
1658 (*t_rate) -= td->rate;
1659 }
1660}
1661
fcb6ade2
JA		 1662static void fio_unpin_memory(void *pinned)
1663{
1664 if (pinned) {
1665 if (munlock(pinned, mlock_size) < 0)
1666 perror("munlock");
1667 munmap(pinned, mlock_size);
1668 }
1669}
1670
1671static void *fio_pin_memory(void)
1672{
1673 long pagesize, pages;
1674 void *ptr;
1675
1676 if (!mlock_size)
1677 return NULL;
1678
1679 /*
1680 * Don't allow mlock of more than real_mem-128MB
1681 */
1682 pagesize = sysconf(_SC_PAGESIZE);
1683 pages = sysconf(_SC_PHYS_PAGES);
1684 if (pages != -1 && pagesize != -1) {
1685 unsigned long long real_mem = pages * pagesize;
1686
1687 if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
1688 mlock_size = real_mem - 128 * 1024 * 1024;
1689 printf("fio: limiting mlocked memory to %lluMiB\n",
1690 mlock_size >> 20);
1691 }
1692 }
1693
1694 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1695 if (!ptr) {
1696 perror("malloc locked mem");
1697 return NULL;
1698 }
1699 if (mlock(ptr, mlock_size) < 0) {
1700 munmap(ptr, mlock_size);
1701 perror("mlock");
1702 return NULL;
1703 }
1704
1705 return ptr;
1706}
1707
ebac4655
JA		 1708static void run_threads(void)
1709{
ebac4655
JA		 1710 struct thread_data *td;
1711 unsigned long spent;
1712 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2
JA		 1713 void *mlocked_mem;
1714
1715 mlocked_mem = fio_pin_memory();
ebac4655
JA		 1716
1717 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
fcb6ade2 1718 fflush(stdout);
c04f7ec3 1719
4efa970e
JA		 1720 signal(SIGINT, sig_handler);
1721 signal(SIGALRM, sig_handler);
1722
ebac4655
JA		 1723 todo = thread_number;
1724 nr_running = 0;
1725 nr_started = 0;
1726 m_rate = t_rate = 0;
1727
1728 for (i = 0; i < thread_number; i++) {
1729 td = &threads[i];
1730
1731 run_str[td->thread_number - 1] = 'P';
1732
1733 init_disk_util(td);
1734
1735 if (!td->create_serialize)
1736 continue;
1737
1738 /*
1739 * do file setup here so it happens sequentially,
1740 * we don't want X number of threads getting their
1741 * client data interspersed on disk
1742 */
1743 if (setup_file(td)) {
1744 td_set_runstate(td, TD_REAPED);
1745 todo--;
1746 }
1747 }
1748
1749 gettimeofday(&genesis, NULL);
1750
1751 while (todo) {
75154845
JA		 1752 struct thread_data *map[MAX_JOBS];
1753 struct timeval this_start;
1754 int this_jobs = 0, left;
1755
ebac4655
JA		 1756 /*
1757 * create threads (TD_NOT_CREATED -> TD_CREATED)
1758 */
1759 for (i = 0; i < thread_number; i++) {
1760 td = &threads[i];
1761
1762 if (td->runstate != TD_NOT_CREATED)
1763 continue;
1764
1765 /*
1766 * never got a chance to start, killed by other
1767 * thread for some reason
1768 */
1769 if (td->terminate) {
1770 todo--;
1771 continue;
1772 }
1773
1774 if (td->start_delay) {
1775 spent = mtime_since_now(&genesis);
1776
1777 if (td->start_delay * 1000 > spent)
1778 continue;
1779 }
1780
1781 if (td->stonewall && (nr_started || nr_running))
1782 break;
1783
75154845
JA		 1784 /*
1785 * Set state to created. Thread will transition
1786 * to TD_INITIALIZED when it's done setting up.
1787 */
ebac4655 1788 td_set_runstate(td, TD_CREATED);
75154845 1789 map[this_jobs++] = td;
ebac4655 1790 sem_init(&startup_sem, 0, 1);
ebac4655
JA		 1791 nr_started++;
1792
1793 if (td->use_thread) {
1794 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1795 perror("thread_create");
1796 nr_started--;
1797 }
1798 } else {
1799 if (fork())
1800 sem_wait(&startup_sem);
1801 else {
1802 fork_main(shm_id, i);
1803 exit(0);
1804 }
1805 }
1806 }
1807
1808 /*
75154845
JA		 1809 * Wait for the started threads to transition to
1810 * TD_INITIALIZED.
ebac4655 1811 */
75154845
JA		 1812 printf("fio: Waiting for threads to initialize...\n");
1813 gettimeofday(&this_start, NULL);
1814 left = this_jobs;
1815 while (left) {
1816 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1817 break;
1818
1819 usleep(100000);
1820
1821 for (i = 0; i < this_jobs; i++) {
1822 td = map[i];
1823 if (!td)
1824 continue;
b6f4d880 1825 if (td->runstate == TD_INITIALIZED) {
75154845
JA		 1826 map[i] = NULL;
1827 left--;
b6f4d880
JA		 1828 } else if (td->runstate >= TD_EXITED) {
1829 map[i] = NULL;
1830 left--;
1831 todo--;
1832 nr_running++; /* work-around... */
75154845
JA		 1833 }
1834 }
1835 }
1836
1837 if (left) {
1838 fprintf(stderr, "fio: %d jobs failed to start\n", left);
1839 for (i = 0; i < this_jobs; i++) {
1840 td = map[i];
1841 if (!td)
1842 continue;
1843 kill(td->pid, SIGTERM);
1844 }
1845 break;
1846 }
1847
1848 /*
b6f4d880 1849 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845
JA		 1850 */
1851 printf("fio: Go for launch\n");
ebac4655
JA		 1852 for (i = 0; i < thread_number; i++) {
1853 td = &threads[i];
1854
75154845 1855 if (td->runstate != TD_INITIALIZED)
ebac4655
JA		 1856 continue;
1857
1858 td_set_runstate(td, TD_RUNNING);
1859 nr_running++;
1860 nr_started--;
1861 m_rate += td->ratemin;
1862 t_rate += td->rate;
75154845 1863 todo--;
ebac4655
JA		 1864 sem_post(&td->mutex);
1865 }
1866
1867 reap_threads(&nr_running, &t_rate, &m_rate);
1868
1869 if (todo)
1870 usleep(100000);
1871 }
1872
1873 while (nr_running) {
1874 reap_threads(&nr_running, &t_rate, &m_rate);
1875 usleep(10000);
1876 }
1877
1878 update_io_ticks();
fcb6ade2 1879 fio_unpin_memory(mlocked_mem);
ebac4655
JA		 1880}
1881
ebac4655
JA		 1882int main(int argc, char *argv[])
1883{
1884 if (parse_options(argc, argv))
1885 return 1;
1886
1887 if (!thread_number) {
1888 printf("Nothing to do\n");
1889 return 1;
1890 }
1891
1892 disk_util_timer_arm();
1893
1894 run_threads();
1895 show_run_stats();
1896
1897 return 0;
1898}
fio - Flexible IO Tester