[PATCH] Let thread exit cleanup the small allocs
[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 65
bbfd6b00 66static volatile int startup_sem;
ebac4655
JA
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 {
6dfd46b9 163 r = os_random_long(&td->random_state);
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 {
6dfd46b9 191 r = os_random_long(&td->bsrange_state);
20dc95c4
JA
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) {
6dfd46b9 253 r = os_random_double(&td->verify_state);
ebac4655
JA
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) {
c1ee2ca4 363 int v;
a6ccc7be 364 long r;
3d60d1ed 365
c1ee2ca4
JA
366 r = os_random_long(&td->rwmix_state);
367 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
a6ccc7be
JA
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
32cd46a0
JA
740/*
741 * Main IO worker functions. It retrieves io_u's to process and queues
742 * and reaps them, checking for rate and errors along the way.
743 */
ebac4655
JA
744static void do_io(struct thread_data *td)
745{
746 struct io_completion_data icd;
747 struct timeval s, e;
748 unsigned long usec;
749
750 while (td->this_io_bytes[td->ddir] < td->io_size) {
751 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
752 struct timespec *timeout;
753 int ret, min_evts = 0;
754 struct io_u *io_u;
755
756 if (td->terminate)
757 break;
758
759 io_u = get_io_u(td);
760 if (!io_u)
761 break;
762
763 memcpy(&s, &io_u->start_time, sizeof(s));
764
765 ret = io_u_queue(td, io_u);
766 if (ret) {
767 put_io_u(td, io_u);
768 td_verror(td, ret);
769 break;
770 }
771
772 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
773
774 if (td->cur_depth < td->iodepth) {
775 timeout = &ts;
776 min_evts = 0;
777 } else {
778 timeout = NULL;
779 min_evts = 1;
780 }
781
782 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
783 if (ret < 0) {
784 td_verror(td, ret);
785 break;
786 } else if (!ret)
787 continue;
788
789 icd.nr = ret;
790 ios_completed(td, &icd);
791 if (icd.error) {
792 td_verror(td, icd.error);
793 break;
794 }
795
796 /*
797 * the rate is batched for now, it should work for batches
798 * of completions except the very first one which may look
799 * a little bursty
800 */
801 gettimeofday(&e, NULL);
802 usec = utime_since(&s, &e);
803
804 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
805
806 if (check_min_rate(td, &e)) {
807 td_verror(td, ENOMEM);
808 break;
809 }
810
811 if (runtime_exceeded(td, &e))
812 break;
813
814 if (td->thinktime)
815 usec_sleep(td, td->thinktime);
816
817 if (should_fsync(td) && td->fsync_blocks &&
818 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
819 sync_td(td);
820 }
821
822 if (td->cur_depth)
823 cleanup_pending_aio(td);
824
fc1a4713 825 if (should_fsync(td) && td->end_fsync)
ebac4655
JA
826 sync_td(td);
827}
828
829static void cleanup_io(struct thread_data *td)
830{
831 if (td->io_cleanup)
832 td->io_cleanup(td);
833}
834
835static int init_io(struct thread_data *td)
836{
837 if (td->io_engine == FIO_SYNCIO)
838 return fio_syncio_init(td);
839 else if (td->io_engine == FIO_MMAPIO)
840 return fio_mmapio_init(td);
841 else if (td->io_engine == FIO_LIBAIO)
842 return fio_libaio_init(td);
843 else if (td->io_engine == FIO_POSIXAIO)
844 return fio_posixaio_init(td);
845 else if (td->io_engine == FIO_SGIO)
846 return fio_sgio_init(td);
8756e4d4
JA
847 else if (td->io_engine == FIO_SPLICEIO)
848 return fio_spliceio_init(td);
ebac4655
JA
849 else {
850 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
851 return 1;
852 }
853}
854
855static void cleanup_io_u(struct thread_data *td)
856{
857 struct list_head *entry, *n;
858 struct io_u *io_u;
859
860 list_for_each_safe(entry, n, &td->io_u_freelist) {
861 io_u = list_entry(entry, struct io_u, list);
862
863 list_del(&io_u->list);
864 free(io_u);
865 }
866
867 if (td->mem_type == MEM_MALLOC)
868 free(td->orig_buffer);
869 else if (td->mem_type == MEM_SHM) {
870 struct shmid_ds sbuf;
871
872 shmdt(td->orig_buffer);
873 shmctl(td->shm_id, IPC_RMID, &sbuf);
874 } else if (td->mem_type == MEM_MMAP)
875 munmap(td->orig_buffer, td->orig_buffer_size);
876 else
877 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
878
879 td->orig_buffer = NULL;
880}
881
882static int init_io_u(struct thread_data *td)
883{
884 struct io_u *io_u;
885 int i, max_units;
886 char *p;
887
888 if (td->io_engine & FIO_SYNCIO)
889 max_units = 1;
890 else
891 max_units = td->iodepth;
892
893 td->orig_buffer_size = td->max_bs * max_units + MASK;
894
895 if (td->mem_type == MEM_MALLOC)
896 td->orig_buffer = malloc(td->orig_buffer_size);
897 else if (td->mem_type == MEM_SHM) {
898 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
899 if (td->shm_id < 0) {
900 td_verror(td, errno);
901 perror("shmget");
902 return 1;
903 }
904
905 td->orig_buffer = shmat(td->shm_id, NULL, 0);
906 if (td->orig_buffer == (void *) -1) {
907 td_verror(td, errno);
908 perror("shmat");
909 td->orig_buffer = NULL;
910 return 1;
911 }
912 } else if (td->mem_type == MEM_MMAP) {
913 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
914 if (td->orig_buffer == MAP_FAILED) {
915 td_verror(td, errno);
916 perror("mmap");
917 td->orig_buffer = NULL;
918 return 1;
919 }
920 }
921
ebac4655
JA
922 p = ALIGN(td->orig_buffer);
923 for (i = 0; i < max_units; i++) {
924 io_u = malloc(sizeof(*io_u));
925 memset(io_u, 0, sizeof(*io_u));
926 INIT_LIST_HEAD(&io_u->list);
927
928 io_u->buf = p + td->max_bs * i;
b1ff3403 929 io_u->index = i;
ebac4655
JA
930 list_add(&io_u->list, &td->io_u_freelist);
931 }
932
933 return 0;
934}
935
936static int create_file(struct thread_data *td, unsigned long long size,
937 int extend)
938{
939 unsigned long long left;
940 unsigned int bs;
941 int r, oflags;
942 char *b;
943
944 /*
945 * unless specifically asked for overwrite, let normal io extend it
946 */
947 if (td_write(td) && !td->overwrite)
948 return 0;
949
950 if (!size) {
951 fprintf(stderr, "Need size for create\n");
952 td_verror(td, EINVAL);
953 return 1;
954 }
955
956 if (!extend) {
957 oflags = O_CREAT | O_TRUNC;
01452055 958 printf("%s: Laying out IO file (%LuMiB)\n", td->name, size >> 20);
ebac4655
JA
959 } else {
960 oflags = O_APPEND;
01452055 961 printf("%s: Extending IO file (%Lu -> %LuMiB)\n", td->name, (td->file_size - size) >> 20, td->file_size >> 20);
ebac4655
JA
962 }
963
964 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
965 if (td->fd < 0) {
966 td_verror(td, errno);
967 return 1;
968 }
969
970 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
971 td_verror(td, errno);
972 return 1;
973 }
974
975 td->io_size = td->file_size;
976 b = malloc(td->max_bs);
977 memset(b, 0, td->max_bs);
978
979 left = size;
980 while (left && !td->terminate) {
981 bs = td->max_bs;
982 if (bs > left)
983 bs = left;
984
985 r = write(td->fd, b, bs);
986
987 if (r == (int) bs) {
988 left -= bs;
989 continue;
990 } else {
991 if (r < 0)
992 td_verror(td, errno);
993 else
994 td_verror(td, EIO);
995
996 break;
997 }
998 }
999
1000 if (td->terminate)
1001 unlink(td->file_name);
1002 else if (td->create_fsync)
1003 fsync(td->fd);
1004
1005 close(td->fd);
1006 td->fd = -1;
1007 free(b);
1008 return 0;
1009}
1010
1011static int file_size(struct thread_data *td)
1012{
1013 struct stat st;
1014
1015 if (fstat(td->fd, &st) == -1) {
1016 td_verror(td, errno);
1017 return 1;
1018 }
1019
838a3cd3
JA
1020 td->real_file_size = st.st_size;
1021
1022 if (!td->file_size || td->file_size > td->real_file_size)
1023 td->file_size = td->real_file_size;
ebac4655 1024
e8e387c1 1025 td->file_size -= td->file_offset;
ebac4655
JA
1026 return 0;
1027}
1028
1029static int bdev_size(struct thread_data *td)
1030{
9104f874 1031 unsigned long long bytes;
ebac4655
JA
1032 int r;
1033
1034 r = blockdev_size(td->fd, &bytes);
1035 if (r) {
1036 td_verror(td, r);
1037 return 1;
1038 }
1039
838a3cd3
JA
1040 td->real_file_size = bytes;
1041
ebac4655
JA
1042 /*
1043 * no extend possibilities, so limit size to device size if too large
1044 */
838a3cd3
JA
1045 if (!td->file_size || td->file_size > td->real_file_size)
1046 td->file_size = td->real_file_size;
ebac4655 1047
e8e387c1 1048 td->file_size -= td->file_offset;
ebac4655
JA
1049 return 0;
1050}
1051
1052static int get_file_size(struct thread_data *td)
1053{
0af7b542 1054 int ret = 0;
ebac4655
JA
1055
1056 if (td->filetype == FIO_TYPE_FILE)
1057 ret = file_size(td);
0af7b542 1058 else if (td->filetype == FIO_TYPE_BD)
ebac4655 1059 ret = bdev_size(td);
0af7b542
JA
1060 else
1061 td->real_file_size = -1;
ebac4655
JA
1062
1063 if (ret)
1064 return ret;
1065
e8e387c1 1066 if (td->file_offset > td->real_file_size) {
01452055 1067 fprintf(stderr, "%s: offset extends end (%Lu > %Lu)\n", td->name, td->file_offset, td->real_file_size);
ebac4655
JA
1068 return 1;
1069 }
1070
838a3cd3 1071 td->io_size = td->file_size;
ebac4655 1072 if (td->io_size == 0) {
01452055 1073 fprintf(stderr, "%s: no io blocks\n", td->name);
ebac4655
JA
1074 td_verror(td, EINVAL);
1075 return 1;
1076 }
1077
20dc95c4
JA
1078 if (!td->zone_size)
1079 td->zone_size = td->io_size;
1080
ebac4655
JA
1081 td->total_io_size = td->io_size * td->loops;
1082 return 0;
1083}
1084
1085static int setup_file_mmap(struct thread_data *td)
1086{
1087 int flags;
1088
3d60d1ed
JA
1089 if (td_rw(td))
1090 flags = PROT_READ | PROT_WRITE;
1091 else if (td_write(td)) {
ebac4655
JA
1092 flags = PROT_WRITE;
1093
1094 if (td->verify != VERIFY_NONE)
1095 flags |= PROT_READ;
3d60d1ed
JA
1096 } else
1097 flags = PROT_READ;
ebac4655
JA
1098
1099 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1100 if (td->mmap == MAP_FAILED) {
1101 td->mmap = NULL;
1102 td_verror(td, errno);
1103 return 1;
1104 }
1105
1106 if (td->invalidate_cache) {
1107 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1108 td_verror(td, errno);
1109 return 1;
1110 }
1111 }
1112
1113 if (td->sequential) {
1114 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1115 td_verror(td, errno);
1116 return 1;
1117 }
1118 } else {
1119 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1120 td_verror(td, errno);
1121 return 1;
1122 }
1123 }
1124
1125 return 0;
1126}
1127
1128static int setup_file_plain(struct thread_data *td)
1129{
1130 if (td->invalidate_cache) {
1131 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1132 td_verror(td, errno);
1133 return 1;
1134 }
1135 }
1136
1137 if (td->sequential) {
1138 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1139 td_verror(td, errno);
1140 return 1;
1141 }
1142 } else {
1143 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1144 td_verror(td, errno);
1145 return 1;
1146 }
1147 }
1148
1149 return 0;
1150}
1151
1152static int setup_file(struct thread_data *td)
1153{
1154 struct stat st;
1155 int flags = 0;
1156
1157 if (stat(td->file_name, &st) == -1) {
1158 if (errno != ENOENT) {
1159 td_verror(td, errno);
1160 return 1;
1161 }
1162 if (!td->create_file) {
1163 td_verror(td, ENOENT);
1164 return 1;
1165 }
1166 if (create_file(td, td->file_size, 0))
1167 return 1;
1168 } else if (td->filetype == FIO_TYPE_FILE) {
6a0106a0 1169 if (st.st_size < (off_t) td->file_size) {
ebac4655
JA
1170 if (create_file(td, td->file_size - st.st_size, 1))
1171 return 1;
1172 }
1173 }
1174
1175 if (td->odirect)
1176 flags |= O_DIRECT;
1177
3d60d1ed 1178 if (td_write(td) || td_rw(td)) {
ebac4655
JA
1179 if (td->filetype == FIO_TYPE_FILE) {
1180 if (!td->overwrite)
1181 flags |= O_TRUNC;
1182
1183 flags |= O_CREAT;
1184 }
1185 if (td->sync_io)
1186 flags |= O_SYNC;
1187
1188 flags |= O_RDWR;
1189
1190 td->fd = open(td->file_name, flags, 0600);
3d60d1ed
JA
1191 } else {
1192 if (td->filetype == FIO_TYPE_CHAR)
1193 flags |= O_RDWR;
1194 else
1195 flags |= O_RDONLY;
1196
1197 td->fd = open(td->file_name, flags);
ebac4655
JA
1198 }
1199
1200 if (td->fd == -1) {
1201 td_verror(td, errno);
1202 return 1;
1203 }
1204
1205 if (get_file_size(td))
1206 return 1;
1207
1208 if (td->io_engine != FIO_MMAPIO)
1209 return setup_file_plain(td);
1210 else
1211 return setup_file_mmap(td);
1212}
1213
da86774e
JA
1214static int switch_ioscheduler(struct thread_data *td)
1215{
1216 char tmp[256], tmp2[128];
1217 FILE *f;
1218 int ret;
1219
1220 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1221
1222 f = fopen(tmp, "r+");
1223 if (!f) {
1224 td_verror(td, errno);
1225 return 1;
1226 }
1227
1228 /*
1229 * Set io scheduler.
1230 */
1231 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
1232 if (ferror(f) || ret != 1) {
1233 td_verror(td, errno);
1234 fclose(f);
1235 return 1;
1236 }
1237
1238 rewind(f);
1239
1240 /*
1241 * Read back and check that the selected scheduler is now the default.
1242 */
1243 ret = fread(tmp, 1, sizeof(tmp), f);
1244 if (ferror(f) || ret < 0) {
1245 td_verror(td, errno);
1246 fclose(f);
1247 return 1;
1248 }
1249
1250 sprintf(tmp2, "[%s]", td->ioscheduler);
1251 if (!strstr(tmp, tmp2)) {
1252 fprintf(stderr, "fio: io scheduler %s not found\n", td->ioscheduler);
1253 td_verror(td, EINVAL);
1254 fclose(f);
1255 return 1;
1256 }
1257
1258 fclose(f);
1259 return 0;
1260}
1261
ebac4655
JA
1262static void clear_io_state(struct thread_data *td)
1263{
1264 if (td->io_engine == FIO_SYNCIO)
1265 lseek(td->fd, SEEK_SET, 0);
1266
20dc95c4 1267 td->last_pos = 0;
ebac4655
JA
1268 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1269 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 1270 td->zone_bytes = 0;
ebac4655
JA
1271
1272 if (td->file_map)
1273 memset(td->file_map, 0, td->num_maps * sizeof(long));
1274}
1275
ebac4655
JA
1276static void *thread_main(void *data)
1277{
1278 struct thread_data *td = data;
ebac4655
JA
1279
1280 if (!td->use_thread)
1281 setsid();
1282
1283 td->pid = getpid();
1284
aea47d44
JA
1285 INIT_LIST_HEAD(&td->io_u_freelist);
1286 INIT_LIST_HEAD(&td->io_u_busylist);
1287 INIT_LIST_HEAD(&td->io_hist_list);
1288 INIT_LIST_HEAD(&td->io_log_list);
1289
ebac4655
JA
1290 if (init_io_u(td))
1291 goto err;
1292
1293 if (fio_setaffinity(td) == -1) {
1294 td_verror(td, errno);
1295 goto err;
1296 }
1297
1298 if (init_io(td))
1299 goto err;
1300
aea47d44
JA
1301 if (init_iolog(td))
1302 goto err;
1303
ebac4655
JA
1304 if (td->ioprio) {
1305 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1306 td_verror(td, errno);
1307 goto err;
1308 }
1309 }
1310
b6f4d880
JA
1311 if (nice(td->nice) < 0) {
1312 td_verror(td, errno);
1313 goto err;
1314 }
1315
75154845
JA
1316 if (init_random_state(td))
1317 goto err;
1318
da86774e
JA
1319 if (td->ioscheduler && switch_ioscheduler(td))
1320 goto err;
1321
75154845 1322 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
1323 fio_sem_up(&startup_sem);
1324 fio_sem_down(&td->mutex);
ebac4655
JA
1325
1326 if (!td->create_serialize && setup_file(td))
1327 goto err;
1328
ebac4655
JA
1329 gettimeofday(&td->epoch, NULL);
1330
4e0ba8af
JA
1331 if (td->exec_prerun)
1332 system(td->exec_prerun);
1333
ebac4655
JA
1334 while (td->loops--) {
1335 getrusage(RUSAGE_SELF, &td->ru_start);
1336 gettimeofday(&td->start, NULL);
1337 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1338
1339 if (td->ratemin)
1340 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1341
1342 clear_io_state(td);
1343 prune_io_piece_log(td);
1344
1345 do_io(td);
1346
1347 td->runtime[td->ddir] += mtime_since_now(&td->start);
aea47d44 1348 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
3d60d1ed
JA
1349 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1350
ebac4655
JA
1351 update_rusage_stat(td);
1352
1353 if (td->error || td->terminate)
1354 break;
1355
1356 if (td->verify == VERIFY_NONE)
1357 continue;
1358
1359 clear_io_state(td);
1360 gettimeofday(&td->start, NULL);
1361
1362 do_verify(td);
1363
1364 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1365
1366 if (td->error || td->terminate)
1367 break;
1368 }
1369
ebac4655
JA
1370 if (td->bw_log)
1371 finish_log(td, td->bw_log, "bw");
1372 if (td->slat_log)
1373 finish_log(td, td->slat_log, "slat");
1374 if (td->clat_log)
1375 finish_log(td, td->clat_log, "clat");
843a7413
JA
1376 if (td->write_iolog)
1377 write_iolog_close(td);
4e0ba8af
JA
1378 if (td->exec_postrun)
1379 system(td->exec_postrun);
ebac4655
JA
1380
1381 if (exitall_on_terminate)
1382 terminate_threads(td->groupid);
1383
1384err:
1385 if (td->fd != -1) {
1386 close(td->fd);
1387 td->fd = -1;
1388 }
1389 if (td->mmap)
1390 munmap(td->mmap, td->file_size);
1391 cleanup_io(td);
1392 cleanup_io_u(td);
ebac4655
JA
1393 td_set_runstate(td, TD_EXITED);
1394 return NULL;
1395
1396}
1397
1398static void *fork_main(int shmid, int offset)
1399{
1400 struct thread_data *td;
1401 void *data;
1402
1403 data = shmat(shmid, NULL, 0);
1404 if (data == (void *) -1) {
1405 perror("shmat");
1406 return NULL;
1407 }
1408
1409 td = data + offset * sizeof(struct thread_data);
1410 thread_main(td);
1411 shmdt(data);
1412 return NULL;
1413}
1414
ebac4655
JA
1415static void check_str_update(struct thread_data *td)
1416{
1417 char c = run_str[td->thread_number - 1];
1418
1419 if (td->runstate == td->old_runstate)
1420 return;
1421
1422 switch (td->runstate) {
1423 case TD_REAPED:
1424 c = '_';
1425 break;
1426 case TD_EXITED:
1427 c = 'E';
1428 break;
1429 case TD_RUNNING:
3d60d1ed
JA
1430 if (td_rw(td)) {
1431 if (td->sequential)
1432 c = 'M';
1433 else
1434 c = 'm';
1435 } else if (td_read(td)) {
ebac4655
JA
1436 if (td->sequential)
1437 c = 'R';
1438 else
1439 c = 'r';
1440 } else {
1441 if (td->sequential)
1442 c = 'W';
1443 else
1444 c = 'w';
1445 }
1446 break;
1447 case TD_VERIFYING:
1448 c = 'V';
1449 break;
1450 case TD_CREATED:
1451 c = 'C';
1452 break;
75154845
JA
1453 case TD_INITIALIZED:
1454 c = 'I';
1455 break;
ebac4655
JA
1456 case TD_NOT_CREATED:
1457 c = 'P';
1458 break;
1459 default:
1460 printf("state %d\n", td->runstate);
1461 }
1462
1463 run_str[td->thread_number - 1] = c;
1464 td->old_runstate = td->runstate;
1465}
1466
5289b847
JA
1467static void eta_to_str(char *str, int eta_sec)
1468{
1469 unsigned int d, h, m, s;
1470 static int always_d, always_h;
1471
1472 d = h = m = s = 0;
1473
1474 s = eta_sec % 60;
1475 eta_sec /= 60;
1476 m = eta_sec % 60;
1477 eta_sec /= 60;
1478 h = eta_sec % 24;
1479 eta_sec /= 24;
1480 d = eta_sec;
1481
1482 if (d || always_d) {
1483 always_d = 1;
1484 str += sprintf(str, "%02dd:", d);
1485 }
1486 if (h || always_h) {
1487 always_h = 1;
1488 str += sprintf(str, "%02dh:", h);
1489 }
1490
1491 str += sprintf(str, "%02dm:", m);
1492 str += sprintf(str, "%02ds", s);
1493}
1494
6a0106a0
JA
1495static int thread_eta(struct thread_data *td, unsigned long elapsed)
1496{
1497 unsigned long long bytes_total, bytes_done;
1498 unsigned int eta_sec = 0;
1499
1500 bytes_total = td->total_io_size;
3d60d1ed
JA
1501
1502 /*
1503 * if writing, bytes_total will be twice the size. If mixing,
1504 * assume a 50/50 split and thus bytes_total will be 50% larger.
1505 */
1506 if (td->verify) {
1507 if (td_rw(td))
1508 bytes_total = bytes_total * 3 / 2;
1509 else
1510 bytes_total <<= 1;
1511 }
6a0106a0
JA
1512 if (td->zone_size && td->zone_skip)
1513 bytes_total /= (td->zone_skip / td->zone_size);
1514
1515 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
1516 double perc;
8b611c34 1517
6a0106a0
JA
1518 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
1519 perc = (double) bytes_done / (double) bytes_total;
1520 if (perc > 1.0)
1521 perc = 1.0;
1522
1523 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
1524
8b611c34 1525 if (td->timeout && eta_sec > (td->timeout - elapsed))
6a0106a0 1526 eta_sec = td->timeout - elapsed;
75154845
JA
1527 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
1528 || td->runstate == TD_INITIALIZED) {
6a0106a0
JA
1529 int t_eta = 0, r_eta = 0;
1530
1531 /*
1532 * We can only guess - assume it'll run the full timeout
1533 * if given, otherwise assume it'll run at the specified rate.
1534 */
1535 if (td->timeout)
1536 t_eta = td->timeout + td->start_delay - elapsed;
1537 if (td->rate) {
1538 r_eta = (bytes_total / 1024) / td->rate;
1539 r_eta += td->start_delay - elapsed;
1540 }
1541
1542 if (r_eta && t_eta)
1543 eta_sec = min(r_eta, t_eta);
1544 else if (r_eta)
1545 eta_sec = r_eta;
1546 else if (t_eta)
1547 eta_sec = t_eta;
1548 else
1549 eta_sec = INT_MAX;
1550 } else {
1551 /*
1552 * thread is already done
1553 */
1554 eta_sec = 0;
1555 }
1556
1557 return eta_sec;
1558}
1559
ebac4655
JA
1560static void print_thread_status(void)
1561{
6a0106a0
JA
1562 unsigned long elapsed = time_since_now(&genesis);
1563 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
5289b847 1564 char eta_str[32];
6a0106a0
JA
1565 double perc = 0.0;
1566
1567 eta_secs = malloc(thread_number * sizeof(int));
1568 memset(eta_secs, 0, thread_number * sizeof(int));
ebac4655 1569
ebac4655
JA
1570 nr_running = t_rate = m_rate = 0;
1571 for (i = 0; i < thread_number; i++) {
1572 struct thread_data *td = &threads[i];
1573
1574 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
1575 nr_running++;
1576 t_rate += td->rate;
1577 m_rate += td->ratemin;
1578 }
1579
6a0106a0
JA
1580 if (elapsed >= 3)
1581 eta_secs[i] = thread_eta(td, elapsed);
8b611c34
JA
1582 else
1583 eta_secs[i] = INT_MAX;
ebac4655
JA
1584
1585 check_str_update(td);
1586 }
1587
6a0106a0
JA
1588 if (exitall_on_terminate)
1589 eta_sec = INT_MAX;
1590 else
1591 eta_sec = 0;
5289b847 1592
6a0106a0
JA
1593 for (i = 0; i < thread_number; i++) {
1594 if (exitall_on_terminate) {
1595 if (eta_secs[i] < eta_sec)
1596 eta_sec = eta_secs[i];
1597 } else {
1598 if (eta_secs[i] > eta_sec)
1599 eta_sec = eta_secs[i];
5289b847 1600 }
6a0106a0 1601 }
5289b847 1602
8b611c34 1603 if (eta_sec != INT_MAX && elapsed) {
6a0106a0
JA
1604 perc = (double) elapsed / (double) (elapsed + eta_sec);
1605 eta_to_str(eta_str, eta_sec);
ebac4655
JA
1606 }
1607
5289b847 1608 printf("Threads now running (%d)", nr_running);
ebac4655
JA
1609 if (m_rate || t_rate)
1610 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
8b611c34 1611 if (eta_sec != INT_MAX) {
6a0106a0
JA
1612 perc *= 100.0;
1613 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
1614 }
5289b847 1615 printf("\r");
ebac4655 1616 fflush(stdout);
6a0106a0 1617 free(eta_secs);
ebac4655
JA
1618}
1619
1620static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1621{
1622 int i;
1623
1624 /*
1625 * reap exited threads (TD_EXITED -> TD_REAPED)
1626 */
1627 for (i = 0; i < thread_number; i++) {
1628 struct thread_data *td = &threads[i];
1629
1630 if (td->runstate != TD_EXITED)
1631 continue;
1632
1633 td_set_runstate(td, TD_REAPED);
1634
1635 if (td->use_thread) {
1636 long ret;
1637
1638 if (pthread_join(td->thread, (void *) &ret))
1639 perror("thread_join");
1640 } else
1641 waitpid(td->pid, NULL, 0);
1642
1643 (*nr_running)--;
1644 (*m_rate) -= td->ratemin;
1645 (*t_rate) -= td->rate;
1646 }
1647}
1648
fcb6ade2
JA
1649static void fio_unpin_memory(void *pinned)
1650{
1651 if (pinned) {
1652 if (munlock(pinned, mlock_size) < 0)
1653 perror("munlock");
1654 munmap(pinned, mlock_size);
1655 }
1656}
1657
1658static void *fio_pin_memory(void)
1659{
32cd46a0 1660 unsigned long long phys_mem;
fcb6ade2
JA
1661 void *ptr;
1662
1663 if (!mlock_size)
1664 return NULL;
1665
1666 /*
1667 * Don't allow mlock of more than real_mem-128MB
1668 */
32cd46a0
JA
1669 phys_mem = os_phys_mem();
1670 if (phys_mem) {
1671 if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
1672 mlock_size = phys_mem - 128 * 1024 * 1024;
fcb6ade2
JA
1673 printf("fio: limiting mlocked memory to %lluMiB\n",
1674 mlock_size >> 20);
1675 }
1676 }
1677
1678 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1679 if (!ptr) {
1680 perror("malloc locked mem");
1681 return NULL;
1682 }
1683 if (mlock(ptr, mlock_size) < 0) {
1684 munmap(ptr, mlock_size);
1685 perror("mlock");
1686 return NULL;
1687 }
1688
1689 return ptr;
1690}
1691
ebac4655
JA
1692static void run_threads(void)
1693{
ebac4655
JA
1694 struct thread_data *td;
1695 unsigned long spent;
1696 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2
JA
1697 void *mlocked_mem;
1698
1699 mlocked_mem = fio_pin_memory();
ebac4655
JA
1700
1701 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
fcb6ade2 1702 fflush(stdout);
c04f7ec3 1703
4efa970e
JA
1704 signal(SIGINT, sig_handler);
1705 signal(SIGALRM, sig_handler);
1706
ebac4655
JA
1707 todo = thread_number;
1708 nr_running = 0;
1709 nr_started = 0;
1710 m_rate = t_rate = 0;
1711
1712 for (i = 0; i < thread_number; i++) {
1713 td = &threads[i];
1714
1715 run_str[td->thread_number - 1] = 'P';
1716
1717 init_disk_util(td);
1718
1719 if (!td->create_serialize)
1720 continue;
1721
1722 /*
1723 * do file setup here so it happens sequentially,
1724 * we don't want X number of threads getting their
1725 * client data interspersed on disk
1726 */
1727 if (setup_file(td)) {
1728 td_set_runstate(td, TD_REAPED);
1729 todo--;
1730 }
1731 }
1732
1733 gettimeofday(&genesis, NULL);
1734
1735 while (todo) {
75154845
JA
1736 struct thread_data *map[MAX_JOBS];
1737 struct timeval this_start;
1738 int this_jobs = 0, left;
1739
ebac4655
JA
1740 /*
1741 * create threads (TD_NOT_CREATED -> TD_CREATED)
1742 */
1743 for (i = 0; i < thread_number; i++) {
1744 td = &threads[i];
1745
1746 if (td->runstate != TD_NOT_CREATED)
1747 continue;
1748
1749 /*
1750 * never got a chance to start, killed by other
1751 * thread for some reason
1752 */
1753 if (td->terminate) {
1754 todo--;
1755 continue;
1756 }
1757
1758 if (td->start_delay) {
1759 spent = mtime_since_now(&genesis);
1760
1761 if (td->start_delay * 1000 > spent)
1762 continue;
1763 }
1764
1765 if (td->stonewall && (nr_started || nr_running))
1766 break;
1767
75154845
JA
1768 /*
1769 * Set state to created. Thread will transition
1770 * to TD_INITIALIZED when it's done setting up.
1771 */
ebac4655 1772 td_set_runstate(td, TD_CREATED);
75154845 1773 map[this_jobs++] = td;
bbfd6b00 1774 fio_sem_init(&startup_sem, 1);
ebac4655
JA
1775 nr_started++;
1776
1777 if (td->use_thread) {
1778 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1779 perror("thread_create");
1780 nr_started--;
1781 }
1782 } else {
1783 if (fork())
bbfd6b00 1784 fio_sem_down(&startup_sem);
ebac4655
JA
1785 else {
1786 fork_main(shm_id, i);
1787 exit(0);
1788 }
1789 }
1790 }
1791
1792 /*
75154845
JA
1793 * Wait for the started threads to transition to
1794 * TD_INITIALIZED.
ebac4655 1795 */
75154845
JA
1796 printf("fio: Waiting for threads to initialize...\n");
1797 gettimeofday(&this_start, NULL);
1798 left = this_jobs;
1799 while (left) {
1800 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1801 break;
1802
1803 usleep(100000);
1804
1805 for (i = 0; i < this_jobs; i++) {
1806 td = map[i];
1807 if (!td)
1808 continue;
b6f4d880 1809 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1810 map[i] = NULL;
1811 left--;
b6f4d880
JA
1812 } else if (td->runstate >= TD_EXITED) {
1813 map[i] = NULL;
1814 left--;
1815 todo--;
1816 nr_running++; /* work-around... */
75154845
JA
1817 }
1818 }
1819 }
1820
1821 if (left) {
1822 fprintf(stderr, "fio: %d jobs failed to start\n", left);
1823 for (i = 0; i < this_jobs; i++) {
1824 td = map[i];
1825 if (!td)
1826 continue;
1827 kill(td->pid, SIGTERM);
1828 }
1829 break;
1830 }
1831
1832 /*
b6f4d880 1833 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845
JA
1834 */
1835 printf("fio: Go for launch\n");
ebac4655
JA
1836 for (i = 0; i < thread_number; i++) {
1837 td = &threads[i];
1838
75154845 1839 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1840 continue;
1841
1842 td_set_runstate(td, TD_RUNNING);
1843 nr_running++;
1844 nr_started--;
1845 m_rate += td->ratemin;
1846 t_rate += td->rate;
75154845 1847 todo--;
bbfd6b00 1848 fio_sem_up(&td->mutex);
ebac4655
JA
1849 }
1850
1851 reap_threads(&nr_running, &t_rate, &m_rate);
1852
1853 if (todo)
1854 usleep(100000);
1855 }
1856
1857 while (nr_running) {
1858 reap_threads(&nr_running, &t_rate, &m_rate);
1859 usleep(10000);
1860 }
1861
1862 update_io_ticks();
fcb6ade2 1863 fio_unpin_memory(mlocked_mem);
ebac4655
JA
1864}
1865
ebac4655
JA
1866int main(int argc, char *argv[])
1867{
1868 if (parse_options(argc, argv))
1869 return 1;
1870
1871 if (!thread_number) {
1872 printf("Nothing to do\n");
1873 return 1;
1874 }
1875
1876 disk_util_timer_arm();
1877
1878 run_threads();
1879 show_run_stats();
1880
1881 return 0;
1882}