projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
[PATCH] Kill ->cur_off, always just lseek() in sync ->io_prep
[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 */
21#include <stdio.h>
22#include <stdlib.h>
23#include <unistd.h>
24#include <fcntl.h>
25#include <string.h>
26#include <errno.h>
27#include <signal.h>
28#include <time.h>
29#include <math.h>
30#include <assert.h>
31#include <dirent.h>
32#include <libgen.h>
33#include <sys/types.h>
34#include <sys/stat.h>
35#include <sys/wait.h>
36#include <sys/ipc.h>
37#include <sys/shm.h>
38#include <sys/ioctl.h>
39#include <sys/mman.h>
40
41#include "fio.h"
42#include "os.h"
43
44#define MASK (4095)
45
46#define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
47
48int groupid = 0;
49int thread_number = 0;
50static char run_str[MAX_JOBS + 1];
51int shm_id = 0;
52static LIST_HEAD(disk_list);
53static struct itimerval itimer;
54
55static void update_io_ticks(void);
56static void disk_util_timer_arm(void);
57static void print_thread_status(void);
58
59/*
60 * thread life cycle
61 */
62enum {
63 TD_NOT_CREATED = 0,
64 TD_CREATED,
65 TD_RUNNING,
66 TD_VERIFYING,
67 TD_EXITED,
68 TD_REAPED,
69};
70
71#define should_fsync(td) (td_write(td) && (!(td)->odirect || (td)->override_sync))
72
73static sem_t startup_sem;
74
75#define TERMINATE_ALL (-1)
76
77static void terminate_threads(int group_id)
78{
79 int i;
80
81 for (i = 0; i < thread_number; i++) {
82 struct thread_data *td = &threads[i];
83
84 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
85 td->terminate = 1;
86 td->start_delay = 0;
87 }
88 }
89}
90
91static void sig_handler(int sig)
92{
93 switch (sig) {
94 case SIGALRM:
95 update_io_ticks();
96 disk_util_timer_arm();
97 print_thread_status();
98 break;
99 default:
100 printf("\nfio: terminating on signal\n");
101 fflush(stdout);
102 terminate_threads(TERMINATE_ALL);
103 break;
104 }
105}
106
107static unsigned long utime_since(struct timeval *s, struct timeval *e)
108{
109 double sec, usec;
110
111 sec = e->tv_sec - s->tv_sec;
112 usec = e->tv_usec - s->tv_usec;
113 if (sec > 0 && usec < 0) {
114 sec--;
115 usec += 1000000;
116 }
117
118 sec *= (double) 1000000;
119
120 return sec + usec;
121}
122
123static unsigned long utime_since_now(struct timeval *s)
124{
125 struct timeval t;
126
127 gettimeofday(&t, NULL);
128 return utime_since(s, &t);
129}
130
131static unsigned long mtime_since(struct timeval *s, struct timeval *e)
132{
133 double sec, usec;
134
135 sec = e->tv_sec - s->tv_sec;
136 usec = e->tv_usec - s->tv_usec;
137 if (sec > 0 && usec < 0) {
138 sec--;
139 usec += 1000000;
140 }
141
142 sec *= (double) 1000;
143 usec /= (double) 1000;
144
145 return sec + usec;
146}
147
148static unsigned long mtime_since_now(struct timeval *s)
149{
150 struct timeval t;
151
152 gettimeofday(&t, NULL);
153 return mtime_since(s, &t);
154}
155
156static inline unsigned long msec_now(struct timeval *s)
157{
158 return s->tv_sec * 1000 + s->tv_usec / 1000;
159}
160
161static int random_map_free(struct thread_data *td, unsigned long long block)
162{
163 unsigned int idx = RAND_MAP_IDX(td, block);
164 unsigned int bit = RAND_MAP_BIT(td, block);
165
166 return (td->file_map[idx] & (1UL << bit)) == 0;
167}
168
169static int get_next_free_block(struct thread_data *td, unsigned long long *b)
170{
171 int i;
172
173 *b = 0;
174 i = 0;
175 while ((*b) * td->min_bs < td->io_size) {
176 if (td->file_map[i] != -1UL) {
177 *b += ffz(td->file_map[i]);
178 return 0;
179 }
180
181 *b += BLOCKS_PER_MAP;
182 i++;
183 }
184
185 return 1;
186}
187
188static void mark_random_map(struct thread_data *td, struct io_u *io_u)
189{
190 unsigned long block = io_u->offset / td->min_bs;
191 unsigned int blocks = 0;
192
193 while (blocks < (io_u->buflen / td->min_bs)) {
194 unsigned int idx, bit;
195
196 if (!random_map_free(td, block))
197 break;
198
199 idx = RAND_MAP_IDX(td, block);
200 bit = RAND_MAP_BIT(td, block);
201
202 assert(idx < td->num_maps);
203
204 td->file_map[idx] |= (1UL << bit);
205 block++;
206 blocks++;
207 }
208
209 if ((blocks * td->min_bs) < io_u->buflen)
210 io_u->buflen = blocks * td->min_bs;
211}
212
213static int get_next_offset(struct thread_data *td, unsigned long long *offset)
214{
215 unsigned long long b, rb;
216 long r;
217
218 if (!td->sequential) {
219 unsigned long max_blocks = td->io_size / td->min_bs;
220 int loops = 50;
221
222 do {
223 lrand48_r(&td->random_state, &r);
224 b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
225 rb = b + (td->file_offset / td->min_bs);
226 loops--;
227 } while (!random_map_free(td, rb) && loops);
228
229 if (!loops) {
230 if (get_next_free_block(td, &b))
231 return 1;
232 }
233 } else
234 b = td->last_bytes / td->min_bs;
235
236 *offset = (b * td->min_bs) + td->file_offset;
237 if (*offset > td->file_size)
238 return 1;
239
240 return 0;
241}
242
243static unsigned int get_next_buflen(struct thread_data *td)
244{
245 unsigned int buflen;
246 long r;
247
248 if (td->min_bs == td->max_bs)
249 buflen = td->min_bs;
250 else {
251 lrand48_r(&td->bsrange_state, &r);
252 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
253 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
254 }
255
256 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
257 buflen = td->io_size - td->this_io_bytes[td->ddir];
258
259 return buflen;
260}
261
262static inline void add_stat_sample(struct io_stat *is, unsigned long val)
263{
264 if (val > is->max_val)
265 is->max_val = val;
266 if (val < is->min_val)
267 is->min_val = val;
268
269 is->val += val;
270 is->val_sq += val * val;
271 is->samples++;
272}
273
274static void add_log_sample(struct thread_data *td, struct io_log *iolog,
275 unsigned long val, int ddir)
276{
277 if (iolog->nr_samples == iolog->max_samples) {
278 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
279
280 iolog->log = realloc(iolog->log, new_size);
281 iolog->max_samples <<= 1;
282 }
283
284 iolog->log[iolog->nr_samples].val = val;
285 iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
286 iolog->log[iolog->nr_samples].ddir = ddir;
287 iolog->nr_samples++;
288}
289
290static void add_clat_sample(struct thread_data *td, int ddir,unsigned long msec)
291{
292 add_stat_sample(&td->clat_stat[ddir], msec);
293
294 if (td->clat_log)
295 add_log_sample(td, td->clat_log, msec, ddir);
296}
297
298static void add_slat_sample(struct thread_data *td, int ddir,unsigned long msec)
299{
300 add_stat_sample(&td->slat_stat[ddir], msec);
301
302 if (td->slat_log)
303 add_log_sample(td, td->slat_log, msec, ddir);
304}
305
306static void add_bw_sample(struct thread_data *td, int ddir)
307{
308 unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
309 unsigned long rate;
310
311 if (spent < td->bw_avg_time)
312 return;
313
314 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
315 add_stat_sample(&td->bw_stat[ddir], rate);
316
317 if (td->bw_log)
318 add_log_sample(td, td->bw_log, rate, ddir);
319
320 gettimeofday(&td->stat_sample_time[ddir], NULL);
321 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
322}
323
324/*
325 * busy looping version for the last few usec
326 */
327static void __usec_sleep(unsigned int usec)
328{
329 struct timeval start;
330
331 gettimeofday(&start, NULL);
332 while (utime_since_now(&start) < usec)
333 nop;
334}
335
336static void usec_sleep(struct thread_data *td, unsigned long usec)
337{
338 struct timespec req, rem;
339
340 req.tv_sec = usec / 1000000;
341 req.tv_nsec = usec * 1000 - req.tv_sec * 1000000;
342
343 do {
344 if (usec < 5000) {
345 __usec_sleep(usec);
346 break;
347 }
348
349 rem.tv_sec = rem.tv_nsec = 0;
350 if (nanosleep(&req, &rem) < 0)
351 break;
352
353 if ((rem.tv_sec + rem.tv_nsec) == 0)
354 break;
355
356 req.tv_nsec = rem.tv_nsec;
357 req.tv_sec = rem.tv_sec;
358
359 usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000;
360 } while (!td->terminate);
361}
362
363static void rate_throttle(struct thread_data *td, unsigned long time_spent,
364 unsigned int bytes)
365{
366 unsigned long usec_cycle;
367
368 if (!td->rate)
369 return;
370
371 usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs);
372
373 if (time_spent < usec_cycle) {
374 unsigned long s = usec_cycle - time_spent;
375
376 td->rate_pending_usleep += s;
377 if (td->rate_pending_usleep >= 100000) {
378 usec_sleep(td, td->rate_pending_usleep);
379 td->rate_pending_usleep = 0;
380 }
381 } else {
382 long overtime = time_spent - usec_cycle;
383
384 td->rate_pending_usleep -= overtime;
385 }
386}
387
388static int check_min_rate(struct thread_data *td, struct timeval *now)
389{
390 unsigned long spent;
391 unsigned long rate;
392 int ddir = td->ddir;
393
394 /*
395 * allow a 2 second settle period in the beginning
396 */
397 if (mtime_since(&td->start, now) < 2000)
398 return 0;
399
400 /*
401 * if rate blocks is set, sample is running
402 */
403 if (td->rate_bytes) {
404 spent = mtime_since(&td->lastrate, now);
405 if (spent < td->ratecycle)
406 return 0;
407
408 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
409 if (rate < td->ratemin) {
410 printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
411 if (rate_quit)
412 terminate_threads(td->groupid);
413 return 1;
414 }
415 }
416
417 td->rate_bytes = td->this_io_bytes[ddir];
418 memcpy(&td->lastrate, now, sizeof(*now));
419 return 0;
420}
421
422static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
423{
424 if (!td->timeout)
425 return 0;
426 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
427 return 1;
428
429 return 0;
430}
431
432static void fill_random_bytes(struct thread_data *td,
433 unsigned char *p, unsigned int len)
434{
435 unsigned int todo;
436 double r;
437
438 while (len) {
439 drand48_r(&td->verify_state, &r);
440
441 /*
442 * lrand48_r seems to be broken and only fill the bottom
443 * 32-bits, even on 64-bit archs with 64-bit longs
444 */
445 todo = sizeof(r);
446 if (todo > len)
447 todo = len;
448
449 memcpy(p, &r, todo);
450
451 len -= todo;
452 p += todo;
453 }
454}
455
456static void hexdump(void *buffer, int len)
457{
458 unsigned char *p = buffer;
459 int i;
460
461 for (i = 0; i < len; i++)
462 printf("%02x", p[i]);
463 printf("\n");
464}
465
466static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
467{
468 unsigned char *p = (unsigned char *) io_u->buf;
469 unsigned long c;
470 int ret;
471
472 p += sizeof(*hdr);
473 c = crc32(p, hdr->len - sizeof(*hdr));
474 ret = c != hdr->crc32;
475
476 if (ret) {
477 fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
478 fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
479 }
480
481 return ret;
482}
483
484static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
485{
486 unsigned char *p = (unsigned char *) io_u->buf;
487 struct md5_ctx md5_ctx;
488 int ret;
489
490 memset(&md5_ctx, 0, sizeof(md5_ctx));
491 p += sizeof(*hdr);
492 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
493
494 ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
495 if (ret) {
496 fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
497 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
498 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
499 }
500
501 return ret;
502}
503
504static int verify_io_u(struct io_u *io_u)
505{
506 struct verify_header *hdr = (struct verify_header *) io_u->buf;
507 int ret;
508
509 if (hdr->fio_magic != FIO_HDR_MAGIC)
510 return 1;
511
512 if (hdr->verify_type == VERIFY_MD5)
513 ret = verify_io_u_md5(hdr, io_u);
514 else if (hdr->verify_type == VERIFY_CRC32)
515 ret = verify_io_u_crc32(hdr, io_u);
516 else {
517 fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
518 ret = 1;
519 }
520
521 return ret;
522}
523
524static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
525{
526 hdr->crc32 = crc32(p, len);
527}
528
529static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
530{
531 struct md5_ctx md5_ctx;
532
533 memset(&md5_ctx, 0, sizeof(md5_ctx));
534 md5_update(&md5_ctx, p, len);
535 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
536}
537
538/*
539 * fill body of io_u->buf with random data and add a header with the
540 * (eg) sha1sum of that data.
541 */
542static void populate_io_u(struct thread_data *td, struct io_u *io_u)
543{
544 unsigned char *p = (unsigned char *) io_u->buf;
545 struct verify_header hdr;
546
547 hdr.fio_magic = FIO_HDR_MAGIC;
548 hdr.len = io_u->buflen;
549 p += sizeof(hdr);
550 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
551
552 if (td->verify == VERIFY_MD5) {
553 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
554 hdr.verify_type = VERIFY_MD5;
555 } else {
556 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
557 hdr.verify_type = VERIFY_CRC32;
558 }
559
560 memcpy(io_u->buf, &hdr, sizeof(hdr));
561}
562
563static void put_io_u(struct thread_data *td, struct io_u *io_u)
564{
565 list_del(&io_u->list);
566 list_add(&io_u->list, &td->io_u_freelist);
567 td->cur_depth--;
568}
569
570#define queue_full(td) (list_empty(&(td)->io_u_freelist))
571
572static struct io_u *__get_io_u(struct thread_data *td)
573{
574 struct io_u *io_u;
575
576 if (queue_full(td))
577 return NULL;
578
579 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
580 io_u->error = 0;
581 io_u->resid = 0;
582 list_del(&io_u->list);
583 list_add(&io_u->list, &td->io_u_busylist);
584 td->cur_depth++;
585 return io_u;
586}
587
588static int td_io_prep(struct thread_data *td, struct io_u *io_u, int read)
589{
590 if (read)
591 io_u->ddir = DDIR_READ;
592 else
593 io_u->ddir = DDIR_WRITE;
594
595 if (td->io_prep && td->io_prep(td, io_u))
596 return 1;
597
598 return 0;
599}
600
601static struct io_u *get_io_u(struct thread_data *td)
602{
603 struct io_u *io_u;
604
605 io_u = __get_io_u(td);
606 if (!io_u)
607 return NULL;
608
609 if (get_next_offset(td, &io_u->offset)) {
610 put_io_u(td, io_u);
611 return NULL;
612 }
613
614 io_u->buflen = get_next_buflen(td);
615 if (!io_u->buflen) {
616 put_io_u(td, io_u);
617 return NULL;
618 }
619
620 if (io_u->buflen + io_u->offset > td->file_size)
621 io_u->buflen = td->file_size - io_u->offset;
622
623 if (!io_u->buflen) {
624 put_io_u(td, io_u);
625 return NULL;
626 }
627
628 if (!td->sequential)
629 mark_random_map(td, io_u);
630
631 td->last_bytes += io_u->buflen;
632
633 if (td->verify != VERIFY_NONE)
634 populate_io_u(td, io_u);
635
636 if (td_io_prep(td, io_u, td_read(td))) {
637 put_io_u(td, io_u);
638 return NULL;
639 }
640
641 gettimeofday(&io_u->start_time, NULL);
642 return io_u;
643}
644
645static inline void td_set_runstate(struct thread_data *td, int runstate)
646{
647 td->old_runstate = td->runstate;
648 td->runstate = runstate;
649}
650
651static int get_next_verify(struct thread_data *td,
652 unsigned long long *offset, unsigned int *len)
653{
654 struct io_piece *ipo;
655
656 if (list_empty(&td->io_hist_list))
657 return 1;
658
659 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
660 list_del(&ipo->list);
661
662 *offset = ipo->offset;
663 *len = ipo->len;
664 free(ipo);
665 return 0;
666}
667
668static void prune_io_piece_log(struct thread_data *td)
669{
670 struct io_piece *ipo;
671
672 while (!list_empty(&td->io_hist_list)) {
673 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
674
675 list_del(&ipo->list);
676 free(ipo);
677 }
678}
679
680/*
681 * log a succesful write, so we can unwind the log for verify
682 */
683static void log_io_piece(struct thread_data *td, struct io_u *io_u)
684{
685 struct io_piece *ipo = malloc(sizeof(struct io_piece));
686 struct list_head *entry;
687
688 INIT_LIST_HEAD(&ipo->list);
689 ipo->offset = io_u->offset;
690 ipo->len = io_u->buflen;
691
692 /*
693 * for random io where the writes extend the file, it will typically
694 * be laid out with the block scattered as written. it's faster to
695 * read them in in that order again, so don't sort
696 */
697 if (td->sequential || !td->overwrite) {
698 list_add_tail(&ipo->list, &td->io_hist_list);
699 return;
700 }
701
702 /*
703 * for random io, sort the list so verify will run faster
704 */
705 entry = &td->io_hist_list;
706 while ((entry = entry->prev) != &td->io_hist_list) {
707 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
708
709 if (__ipo->offset < ipo->offset)
710 break;
711 }
712
713 list_add(&ipo->list, entry);
714}
715
716static int sync_td(struct thread_data *td)
717{
718 if (td->io_sync)
719 return td->io_sync(td);
720
721 return 0;
722}
723
724static int io_u_getevents(struct thread_data *td, int min, int max,
725 struct timespec *t)
726{
727 return td->io_getevents(td, min, max, t);
728}
729
730static int io_u_queue(struct thread_data *td, struct io_u *io_u)
731{
732 gettimeofday(&io_u->issue_time, NULL);
733
734 return td->io_queue(td, io_u);
735}
736
737#define iocb_time(iocb) ((unsigned long) (iocb)->data)
738
739static void io_completed(struct thread_data *td, struct io_u *io_u,
740 struct io_completion_data *icd)
741{
742 struct timeval e;
743 unsigned long msec;
744
745 gettimeofday(&e, NULL);
746
747 if (!io_u->error) {
748 int idx = io_u->ddir;
749
750 td->io_blocks[idx]++;
751 td->io_bytes[idx] += (io_u->buflen - io_u->resid);
752 td->this_io_bytes[idx] += (io_u->buflen - io_u->resid);
753
754 msec = mtime_since(&io_u->issue_time, &e);
755
756 add_clat_sample(td, io_u->ddir, msec);
757 add_bw_sample(td, io_u->ddir);
758
759 if (td_write(td) && io_u->ddir == DDIR_WRITE)
760 log_io_piece(td, io_u);
761
762 icd->bytes_done[idx] += (io_u->buflen - io_u->resid);
763 } else
764 icd->error = io_u->error;
765}
766
767static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
768{
769 struct io_u *io_u;
770 int i;
771
772 icd->error = 0;
773 icd->bytes_done[0] = icd->bytes_done[1] = 0;
774
775 for (i = 0; i < icd->nr; i++) {
776 io_u = td->io_event(td, i);
777
778 io_completed(td, io_u, icd);
779 put_io_u(td, io_u);
780 }
781}
782
783static void cleanup_pending_aio(struct thread_data *td)
784{
785 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
786 struct list_head *entry, *n;
787 struct io_completion_data icd;
788 struct io_u *io_u;
789 int r;
790
791 /*
792 * get immediately available events, if any
793 */
794 r = io_u_getevents(td, 0, td->cur_depth, &ts);
795 if (r > 0) {
796 icd.nr = r;
797 ios_completed(td, &icd);
798 }
799
800 /*
801 * now cancel remaining active events
802 */
803 if (td->io_cancel) {
804 list_for_each_safe(entry, n, &td->io_u_busylist) {
805 io_u = list_entry(entry, struct io_u, list);
806
807 r = td->io_cancel(td, io_u);
808 if (!r)
809 put_io_u(td, io_u);
810 }
811 }
812
813 if (td->cur_depth) {
814 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
815 if (r > 0) {
816 icd.nr = r;
817 ios_completed(td, &icd);
818 }
819 }
820}
821
822static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
823{
824 struct io_u *v_io_u = *io_u;
825 int ret = 0;
826
827 if (v_io_u) {
828 ret = verify_io_u(v_io_u);
829 put_io_u(td, v_io_u);
830 *io_u = NULL;
831 }
832
833 return ret;
834}
835
836static void do_verify(struct thread_data *td)
837{
838 struct timeval t;
839 struct io_u *io_u, *v_io_u = NULL;
840 struct io_completion_data icd;
841 int ret;
842
843 td_set_runstate(td, TD_VERIFYING);
844
845 do {
846 if (td->terminate)
847 break;
848
849 gettimeofday(&t, NULL);
850 if (runtime_exceeded(td, &t))
851 break;
852
853 io_u = __get_io_u(td);
854 if (!io_u)
855 break;
856
857 if (get_next_verify(td, &io_u->offset, &io_u->buflen)) {
858 put_io_u(td, io_u);
859 break;
860 }
861
862 if (td_io_prep(td, io_u, 1)) {
863 put_io_u(td, io_u);
864 break;
865 }
866
867 ret = io_u_queue(td, io_u);
868 if (ret) {
869 put_io_u(td, io_u);
870 td_verror(td, ret);
871 break;
872 }
873
874 /*
875 * we have one pending to verify, do that while
876 * we are doing io on the next one
877 */
878 if (do_io_u_verify(td, &v_io_u))
879 break;
880
881 ret = io_u_getevents(td, 1, 1, NULL);
882 if (ret != 1) {
883 if (ret < 0)
884 td_verror(td, ret);
885 break;
886 }
887
888 v_io_u = td->io_event(td, 0);
889 icd.nr = 1;
890 icd.error = 0;
891 io_completed(td, v_io_u, &icd);
892
893 if (icd.error) {
894 td_verror(td, icd.error);
895 put_io_u(td, v_io_u);
896 v_io_u = NULL;
897 break;
898 }
899
ebac4655
JA		 900 /*
901 * if we can't submit more io, we need to verify now
902 */
903 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
904 break;
905
906 } while (1);
907
908 do_io_u_verify(td, &v_io_u);
909
910 if (td->cur_depth)
911 cleanup_pending_aio(td);
912
913 td_set_runstate(td, TD_RUNNING);
914}
915
916static void do_io(struct thread_data *td)
917{
918 struct io_completion_data icd;
919 struct timeval s, e;
920 unsigned long usec;
921
922 while (td->this_io_bytes[td->ddir] < td->io_size) {
923 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
924 struct timespec *timeout;
925 int ret, min_evts = 0;
926 struct io_u *io_u;
927
928 if (td->terminate)
929 break;
930
931 io_u = get_io_u(td);
932 if (!io_u)
933 break;
934
935 memcpy(&s, &io_u->start_time, sizeof(s));
936
937 ret = io_u_queue(td, io_u);
938 if (ret) {
939 put_io_u(td, io_u);
940 td_verror(td, ret);
941 break;
942 }
943
944 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
945
946 if (td->cur_depth < td->iodepth) {
947 timeout = &ts;
948 min_evts = 0;
949 } else {
950 timeout = NULL;
951 min_evts = 1;
952 }
953
954 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
955 if (ret < 0) {
956 td_verror(td, ret);
957 break;
958 } else if (!ret)
959 continue;
960
961 icd.nr = ret;
962 ios_completed(td, &icd);
963 if (icd.error) {
964 td_verror(td, icd.error);
965 break;
966 }
967
968 /*
969 * the rate is batched for now, it should work for batches
970 * of completions except the very first one which may look
971 * a little bursty
972 */
973 gettimeofday(&e, NULL);
974 usec = utime_since(&s, &e);
975
976 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
977
978 if (check_min_rate(td, &e)) {
979 td_verror(td, ENOMEM);
980 break;
981 }
982
983 if (runtime_exceeded(td, &e))
984 break;
985
986 if (td->thinktime)
987 usec_sleep(td, td->thinktime);
988
989 if (should_fsync(td) && td->fsync_blocks &&
990 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
991 sync_td(td);
992 }
993
994 if (td->cur_depth)
995 cleanup_pending_aio(td);
996
997 if (should_fsync(td))
998 sync_td(td);
999}
1000
1001static void cleanup_io(struct thread_data *td)
1002{
1003 if (td->io_cleanup)
1004 td->io_cleanup(td);
1005}
1006
1007static int init_io(struct thread_data *td)
1008{
1009 if (td->io_engine == FIO_SYNCIO)
1010 return fio_syncio_init(td);
1011 else if (td->io_engine == FIO_MMAPIO)
1012 return fio_mmapio_init(td);
1013 else if (td->io_engine == FIO_LIBAIO)
1014 return fio_libaio_init(td);
1015 else if (td->io_engine == FIO_POSIXAIO)
1016 return fio_posixaio_init(td);
1017 else if (td->io_engine == FIO_SGIO)
1018 return fio_sgio_init(td);
1019 else {
1020 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1021 return 1;
1022 }
1023}
1024
1025static void cleanup_io_u(struct thread_data *td)
1026{
1027 struct list_head *entry, *n;
1028 struct io_u *io_u;
1029
1030 list_for_each_safe(entry, n, &td->io_u_freelist) {
1031 io_u = list_entry(entry, struct io_u, list);
1032
1033 list_del(&io_u->list);
1034 free(io_u);
1035 }
1036
1037 if (td->mem_type == MEM_MALLOC)
1038 free(td->orig_buffer);
1039 else if (td->mem_type == MEM_SHM) {
1040 struct shmid_ds sbuf;
1041
1042 shmdt(td->orig_buffer);
1043 shmctl(td->shm_id, IPC_RMID, &sbuf);
1044 } else if (td->mem_type == MEM_MMAP)
1045 munmap(td->orig_buffer, td->orig_buffer_size);
1046 else
1047 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1048
1049 td->orig_buffer = NULL;
1050}
1051
1052static int init_io_u(struct thread_data *td)
1053{
1054 struct io_u *io_u;
1055 int i, max_units;
1056 char *p;
1057
1058 if (td->io_engine & FIO_SYNCIO)
1059 max_units = 1;
1060 else
1061 max_units = td->iodepth;
1062
1063 td->orig_buffer_size = td->max_bs * max_units + MASK;
1064
1065 if (td->mem_type == MEM_MALLOC)
1066 td->orig_buffer = malloc(td->orig_buffer_size);
1067 else if (td->mem_type == MEM_SHM) {
1068 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1069 if (td->shm_id < 0) {
1070 td_verror(td, errno);
1071 perror("shmget");
1072 return 1;
1073 }
1074
1075 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1076 if (td->orig_buffer == (void *) -1) {
1077 td_verror(td, errno);
1078 perror("shmat");
1079 td->orig_buffer = NULL;
1080 return 1;
1081 }
1082 } else if (td->mem_type == MEM_MMAP) {
1083 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1084 if (td->orig_buffer == MAP_FAILED) {
1085 td_verror(td, errno);
1086 perror("mmap");
1087 td->orig_buffer = NULL;
1088 return 1;
1089 }
1090 }
1091
1092 INIT_LIST_HEAD(&td->io_u_freelist);
1093 INIT_LIST_HEAD(&td->io_u_busylist);
1094 INIT_LIST_HEAD(&td->io_hist_list);
1095
1096 p = ALIGN(td->orig_buffer);
1097 for (i = 0; i < max_units; i++) {
1098 io_u = malloc(sizeof(*io_u));
1099 memset(io_u, 0, sizeof(*io_u));
1100 INIT_LIST_HEAD(&io_u->list);
1101
1102 io_u->buf = p + td->max_bs * i;
1103 list_add(&io_u->list, &td->io_u_freelist);
1104 }
1105
1106 return 0;
1107}
1108
1109static int create_file(struct thread_data *td, unsigned long long size,
1110 int extend)
1111{
1112 unsigned long long left;
1113 unsigned int bs;
1114 int r, oflags;
1115 char *b;
1116
1117 /*
1118 * unless specifically asked for overwrite, let normal io extend it
1119 */
1120 if (td_write(td) && !td->overwrite)
1121 return 0;
1122
1123 if (!size) {
1124 fprintf(stderr, "Need size for create\n");
1125 td_verror(td, EINVAL);
1126 return 1;
1127 }
1128
1129 if (!extend) {
1130 oflags = O_CREAT | O_TRUNC;
1131 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1132 } else {
1133 oflags = O_APPEND;
1134 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1135 }
1136
1137 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1138 if (td->fd < 0) {
1139 td_verror(td, errno);
1140 return 1;
1141 }
1142
1143 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1144 td_verror(td, errno);
1145 return 1;
1146 }
1147
1148 td->io_size = td->file_size;
1149 b = malloc(td->max_bs);
1150 memset(b, 0, td->max_bs);
1151
1152 left = size;
1153 while (left && !td->terminate) {
1154 bs = td->max_bs;
1155 if (bs > left)
1156 bs = left;
1157
1158 r = write(td->fd, b, bs);
1159
1160 if (r == (int) bs) {
1161 left -= bs;
1162 continue;
1163 } else {
1164 if (r < 0)
1165 td_verror(td, errno);
1166 else
1167 td_verror(td, EIO);
1168
1169 break;
1170 }
1171 }
1172
1173 if (td->terminate)
1174 unlink(td->file_name);
1175 else if (td->create_fsync)
1176 fsync(td->fd);
1177
1178 close(td->fd);
1179 td->fd = -1;
1180 free(b);
1181 return 0;
1182}
1183
1184static int file_size(struct thread_data *td)
1185{
1186 struct stat st;
1187
1188 if (fstat(td->fd, &st) == -1) {
1189 td_verror(td, errno);
1190 return 1;
1191 }
1192
1193 if (!td->file_size)
1194 td->file_size = st.st_size;
1195
1196 return 0;
1197}
1198
1199static int bdev_size(struct thread_data *td)
1200{
1201 size_t bytes;
1202 int r;
1203
1204 r = blockdev_size(td->fd, &bytes);
1205 if (r) {
1206 td_verror(td, r);
1207 return 1;
1208 }
1209
1210 /*
1211 * no extend possibilities, so limit size to device size if too large
1212 */
1213 if (!td->file_size || td->file_size > bytes)
1214 td->file_size = bytes;
1215
1216 return 0;
1217}
1218
1219static int get_file_size(struct thread_data *td)
1220{
1221 int ret;
1222
1223 if (td->filetype == FIO_TYPE_FILE)
1224 ret = file_size(td);
1225 else
1226 ret = bdev_size(td);
1227
1228 if (ret)
1229 return ret;
1230
1231 if (td->file_offset > td->file_size) {
1232 fprintf(stderr, "Client%d: offset larger than length (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->file_size);
1233 return 1;
1234 }
1235
1236 td->io_size = td->file_size - td->file_offset;
1237 if (td->io_size == 0) {
1238 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1239 td_verror(td, EINVAL);
1240 return 1;
1241 }
1242
1243 td->total_io_size = td->io_size * td->loops;
1244 return 0;
1245}
1246
1247static int setup_file_mmap(struct thread_data *td)
1248{
1249 int flags;
1250
1251 if (td_read(td))
1252 flags = PROT_READ;
1253 else {
1254 flags = PROT_WRITE;
1255
1256 if (td->verify != VERIFY_NONE)
1257 flags |= PROT_READ;
1258 }
1259
1260 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1261 if (td->mmap == MAP_FAILED) {
1262 td->mmap = NULL;
1263 td_verror(td, errno);
1264 return 1;
1265 }
1266
1267 if (td->invalidate_cache) {
1268 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1269 td_verror(td, errno);
1270 return 1;
1271 }
1272 }
1273
1274 if (td->sequential) {
1275 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1276 td_verror(td, errno);
1277 return 1;
1278 }
1279 } else {
1280 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1281 td_verror(td, errno);
1282 return 1;
1283 }
1284 }
1285
1286 return 0;
1287}
1288
1289static int setup_file_plain(struct thread_data *td)
1290{
1291 if (td->invalidate_cache) {
1292 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1293 td_verror(td, errno);
1294 return 1;
1295 }
1296 }
1297
1298 if (td->sequential) {
1299 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1300 td_verror(td, errno);
1301 return 1;
1302 }
1303 } else {
1304 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1305 td_verror(td, errno);
1306 return 1;
1307 }
1308 }
1309
1310 return 0;
1311}
1312
1313static int setup_file(struct thread_data *td)
1314{
1315 struct stat st;
1316 int flags = 0;
1317
1318 if (stat(td->file_name, &st) == -1) {
1319 if (errno != ENOENT) {
1320 td_verror(td, errno);
1321 return 1;
1322 }
1323 if (!td->create_file) {
1324 td_verror(td, ENOENT);
1325 return 1;
1326 }
1327 if (create_file(td, td->file_size, 0))
1328 return 1;
1329 } else if (td->filetype == FIO_TYPE_FILE) {
1330 if (st.st_size < td->file_size) {
1331 if (create_file(td, td->file_size - st.st_size, 1))
1332 return 1;
1333 }
1334 }
1335
1336 if (td->odirect)
1337 flags |= O_DIRECT;
1338
1339 if (td_read(td))
1340 td->fd = open(td->file_name, flags | O_RDONLY);
1341 else {
1342 if (td->filetype == FIO_TYPE_FILE) {
1343 if (!td->overwrite)
1344 flags |= O_TRUNC;
1345
1346 flags |= O_CREAT;
1347 }
1348 if (td->sync_io)
1349 flags |= O_SYNC;
1350
1351 flags |= O_RDWR;
1352
1353 td->fd = open(td->file_name, flags, 0600);
1354 }
1355
1356 if (td->fd == -1) {
1357 td_verror(td, errno);
1358 return 1;
1359 }
1360
1361 if (get_file_size(td))
1362 return 1;
1363
1364 if (td->io_engine != FIO_MMAPIO)
1365 return setup_file_plain(td);
1366 else
1367 return setup_file_mmap(td);
1368}
1369
1370static int check_dev_match(dev_t dev, char *path)
1371{
1372 unsigned int major, minor;
1373 char line[256], *p;
1374 FILE *f;
1375
1376 f = fopen(path, "r");
1377 if (!f) {
1378 perror("open path");
1379 return 1;
1380 }
1381
1382 p = fgets(line, sizeof(line), f);
1383 if (!p) {
1384 fclose(f);
1385 return 1;
1386 }
1387
1388 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1389 fclose(f);
1390 return 1;
1391 }
1392
1393 if (((major << 8) | minor) == dev) {
1394 fclose(f);
1395 return 0;
1396 }
1397
1398 fclose(f);
1399 return 1;
1400}
1401
1402static int find_block_dir(dev_t dev, char *path)
1403{
1404 struct dirent *dir;
1405 struct stat st;
1406 int found = 0;
1407 DIR *D;
1408
1409 D = opendir(path);
1410 if (!D)
1411 return 0;
1412
1413 while ((dir = readdir(D)) != NULL) {
1414 char full_path[256];
1415
1416 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1417 continue;
1418 if (!strcmp(dir->d_name, "device"))
1419 continue;
1420
1421 sprintf(full_path, "%s/%s", path, dir->d_name);
1422
1423 if (!strcmp(dir->d_name, "dev")) {
1424 if (!check_dev_match(dev, full_path)) {
1425 found = 1;
1426 break;
1427 }
1428 }
1429
1430 if (stat(full_path, &st) == -1) {
1431 perror("stat");
1432 break;
1433 }
1434
1435 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1436 continue;
1437
1438 found = find_block_dir(dev, full_path);
1439 if (found) {
1440 strcpy(path, full_path);
1441 break;
1442 }
1443 }
1444
1445 closedir(D);
1446 return found;
1447}
1448
1449static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1450{
1451 unsigned in_flight;
1452 char line[256];
1453 FILE *f;
1454 char *p;
1455
1456 f = fopen(du->path, "r");
1457 if (!f)
1458 return 1;
1459
1460 p = fgets(line, sizeof(line), f);
1461 if (!p) {
1462 fclose(f);
1463 return 1;
1464 }
1465
1466 if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
1467 fclose(f);
1468 return 1;
1469 }
1470
1471 fclose(f);
1472 return 0;
1473}
1474
1475static void update_io_tick_disk(struct disk_util *du)
1476{
1477 struct disk_util_stat __dus, *dus, *ldus;
1478 struct timeval t;
1479
1480 if (get_io_ticks(du, &__dus))
1481 return;
1482
1483 dus = &du->dus;
1484 ldus = &du->last_dus;
1485
1486 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1487 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1488 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1489 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1490 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1491 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1492 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1493 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1494 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1495 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1496
1497 gettimeofday(&t, NULL);
1498 du->msec += mtime_since(&du->time, &t);
1499 memcpy(&du->time, &t, sizeof(t));
1500 memcpy(ldus, &__dus, sizeof(__dus));
1501}
1502
1503static void update_io_ticks(void)
1504{
1505 struct list_head *entry;
1506 struct disk_util *du;
1507
1508 list_for_each(entry, &disk_list) {
1509 du = list_entry(entry, struct disk_util, list);
1510 update_io_tick_disk(du);
1511 }
1512}
1513
1514static int disk_util_exists(dev_t dev)
1515{
1516 struct list_head *entry;
1517 struct disk_util *du;
1518
1519 list_for_each(entry, &disk_list) {
1520 du = list_entry(entry, struct disk_util, list);
1521
1522 if (du->dev == dev)
1523 return 1;
1524 }
1525
1526 return 0;
1527}
1528
1529static void disk_util_add(dev_t dev, char *path)
1530{
1531 struct disk_util *du = malloc(sizeof(*du));
1532
1533 memset(du, 0, sizeof(*du));
1534 INIT_LIST_HEAD(&du->list);
1535 sprintf(du->path, "%s/stat", path);
1536 du->name = strdup(basename(path));
1537 du->dev = dev;
1538
1539 gettimeofday(&du->time, NULL);
1540 get_io_ticks(du, &du->last_dus);
1541
1542 list_add_tail(&du->list, &disk_list);
1543}
1544
1545static void init_disk_util(struct thread_data *td)
1546{
1547 struct stat st;
1548 char foo[256], tmp[256];
1549 dev_t dev;
1550 char *p;
1551
1552 if (!td->do_disk_util)
1553 return;
1554
1555 if (!stat(td->file_name, &st)) {
1556 if (S_ISBLK(st.st_mode))
1557 dev = st.st_rdev;
1558 else
1559 dev = st.st_dev;
1560 } else {
1561 /*
1562 * must be a file, open "." in that path
1563 */
1564 strcpy(foo, td->file_name);
1565 p = dirname(foo);
1566 if (stat(p, &st)) {
1567 perror("disk util stat");
1568 return;
1569 }
1570
1571 dev = st.st_dev;
1572 }
1573
1574 if (disk_util_exists(dev))
1575 return;
1576
1577 sprintf(foo, "/sys/block");
1578 if (!find_block_dir(dev, foo))
1579 return;
1580
1581 /*
1582 * if this is inside a partition dir, jump back to parent
1583 */
1584 sprintf(tmp, "%s/queue", foo);
1585 if (stat(tmp, &st)) {
1586 p = dirname(foo);
1587 sprintf(tmp, "%s/queue", p);
1588 if (stat(tmp, &st)) {
1589 fprintf(stderr, "unknown sysfs layout\n");
1590 return;
1591 }
1592 sprintf(foo, "%s", p);
1593 }
1594
1595 disk_util_add(dev, foo);
1596}
1597
1598static void disk_util_timer_arm(void)
1599{
1600 itimer.it_value.tv_sec = 0;
1601 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1602 setitimer(ITIMER_REAL, &itimer, NULL);
1603}
1604
1605static void clear_io_state(struct thread_data *td)
1606{
1607 if (td->io_engine == FIO_SYNCIO)
1608 lseek(td->fd, SEEK_SET, 0);
1609
ebac4655
JA		 1610 td->last_bytes = 0;
1611 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1612 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1613
1614 if (td->file_map)
1615 memset(td->file_map, 0, td->num_maps * sizeof(long));
1616}
1617
1618static void update_rusage_stat(struct thread_data *td)
1619{
1620 if (!(td->runtime[0] + td->runtime[1]))
1621 return;
1622
1623 getrusage(RUSAGE_SELF, &td->ru_end);
1624
1625 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1626 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1627 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1628
1629
1630 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1631}
1632
1633static void *thread_main(void *data)
1634{
1635 struct thread_data *td = data;
1636 int ret = 1;
1637
1638 if (!td->use_thread)
1639 setsid();
1640
1641 td->pid = getpid();
1642
1643 if (init_io_u(td))
1644 goto err;
1645
1646 if (fio_setaffinity(td) == -1) {
1647 td_verror(td, errno);
1648 goto err;
1649 }
1650
1651 if (init_io(td))
1652 goto err;
1653
1654 if (td->ioprio) {
1655 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1656 td_verror(td, errno);
1657 goto err;
1658 }
1659 }
1660
1661 sem_post(&startup_sem);
1662 sem_wait(&td->mutex);
1663
1664 if (!td->create_serialize && setup_file(td))
1665 goto err;
1666
1667 if (init_random_state(td))
1668 goto err;
1669
1670 gettimeofday(&td->epoch, NULL);
1671
1672 while (td->loops--) {
1673 getrusage(RUSAGE_SELF, &td->ru_start);
1674 gettimeofday(&td->start, NULL);
1675 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1676
1677 if (td->ratemin)
1678 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1679
1680 clear_io_state(td);
1681 prune_io_piece_log(td);
1682
1683 do_io(td);
1684
1685 td->runtime[td->ddir] += mtime_since_now(&td->start);
1686 update_rusage_stat(td);
1687
1688 if (td->error || td->terminate)
1689 break;
1690
1691 if (td->verify == VERIFY_NONE)
1692 continue;
1693
1694 clear_io_state(td);
1695 gettimeofday(&td->start, NULL);
1696
1697 do_verify(td);
1698
1699 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1700
1701 if (td->error || td->terminate)
1702 break;
1703 }
1704
1705 ret = 0;
1706
1707 if (td->bw_log)
1708 finish_log(td, td->bw_log, "bw");
1709 if (td->slat_log)
1710 finish_log(td, td->slat_log, "slat");
1711 if (td->clat_log)
1712 finish_log(td, td->clat_log, "clat");
1713
1714 if (exitall_on_terminate)
1715 terminate_threads(td->groupid);
1716
1717err:
1718 if (td->fd != -1) {
1719 close(td->fd);
1720 td->fd = -1;
1721 }
1722 if (td->mmap)
1723 munmap(td->mmap, td->file_size);
1724 cleanup_io(td);
1725 cleanup_io_u(td);
1726 if (ret) {
1727 sem_post(&startup_sem);
1728 sem_wait(&td->mutex);
1729 }
1730 td_set_runstate(td, TD_EXITED);
1731 return NULL;
1732
1733}
1734
1735static void *fork_main(int shmid, int offset)
1736{
1737 struct thread_data *td;
1738 void *data;
1739
1740 data = shmat(shmid, NULL, 0);
1741 if (data == (void *) -1) {
1742 perror("shmat");
1743 return NULL;
1744 }
1745
1746 td = data + offset * sizeof(struct thread_data);
1747 thread_main(td);
1748 shmdt(data);
1749 return NULL;
1750}
1751
1752static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1753 double *mean, double *dev)
1754{
1755 double n;
1756
1757 if (is->samples == 0)
1758 return 0;
1759
1760 *min = is->min_val;
1761 *max = is->max_val;
1762
1763 n = (double) is->samples;
1764 *mean = (double) is->val / n;
1765 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1766 if (!(*min + *max) && !(*mean + *dev))
1767 return 0;
1768
1769 return 1;
1770}
1771
1772static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1773 int ddir)
1774{
1775 char *ddir_str[] = { "read ", "write" };
1776 unsigned long min, max, bw;
1777 double mean, dev;
1778
1779 if (!td->runtime[ddir])
1780 return;
1781
1782 bw = td->io_bytes[ddir] / td->runtime[ddir];
1783 printf(" %s: io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
1784
1785 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1786 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1787
1788 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1789 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1790
1791 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1792 double p_of_agg;
1793
1794 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1795 printf(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev);
1796 }
1797}
1798
1799static void show_thread_status(struct thread_data *td,
1800 struct group_run_stats *rs)
1801{
1802 double usr_cpu, sys_cpu;
1803
1804 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1805 return;
1806
1807 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1808
1809 show_ddir_status(td, rs, td->ddir);
1810 show_ddir_status(td, rs, td->ddir ^ 1);
1811
1812 if (td->runtime[0] + td->runtime[1]) {
1813 double runt = td->runtime[0] + td->runtime[1];
1814
1815 usr_cpu = (double) td->usr_time * 100 / runt;
1816 sys_cpu = (double) td->sys_time * 100 / runt;
1817 } else {
1818 usr_cpu = 0;
1819 sys_cpu = 0;
1820 }
1821
1822 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
1823}
1824
1825static void check_str_update(struct thread_data *td)
1826{
1827 char c = run_str[td->thread_number - 1];
1828
1829 if (td->runstate == td->old_runstate)
1830 return;
1831
1832 switch (td->runstate) {
1833 case TD_REAPED:
1834 c = '_';
1835 break;
1836 case TD_EXITED:
1837 c = 'E';
1838 break;
1839 case TD_RUNNING:
1840 if (td_read(td)) {
1841 if (td->sequential)
1842 c = 'R';
1843 else
1844 c = 'r';
1845 } else {
1846 if (td->sequential)
1847 c = 'W';
1848 else
1849 c = 'w';
1850 }
1851 break;
1852 case TD_VERIFYING:
1853 c = 'V';
1854 break;
1855 case TD_CREATED:
1856 c = 'C';
1857 break;
1858 case TD_NOT_CREATED:
1859 c = 'P';
1860 break;
1861 default:
1862 printf("state %d\n", td->runstate);
1863 }
1864
1865 run_str[td->thread_number - 1] = c;
1866 td->old_runstate = td->runstate;
1867}
1868
1869static void print_thread_status(void)
1870{
1871 unsigned long long bytes_done, bytes_total;
1872 int i, nr_running, t_rate, m_rate;
1873 double perc;
1874
1875 bytes_done = bytes_total = 0;
1876 nr_running = t_rate = m_rate = 0;
1877 for (i = 0; i < thread_number; i++) {
1878 struct thread_data *td = &threads[i];
1879
1880 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
1881 nr_running++;
1882 t_rate += td->rate;
1883 m_rate += td->ratemin;
1884 }
1885
1886 bytes_total += td->total_io_size;
1887 if (td->verify)
1888 bytes_total += td->total_io_size;
1889
1890 bytes_done += td->io_bytes[DDIR_READ] +td->io_bytes[DDIR_WRITE];
1891
1892 check_str_update(td);
1893 }
1894
1895 perc = 0;
1896 if (bytes_total && bytes_done) {
1897 perc = (double) 100 * bytes_done / (double) bytes_total;
1898 if (perc > 100.0)
1899 perc = 100.0;
1900 }
1901
1902 printf("Threads now running: %d", nr_running);
1903 if (m_rate || t_rate)
1904 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
1905 printf(" : [%s] [%3.2f%% done]\r", run_str, perc);
1906 fflush(stdout);
1907}
1908
1909static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1910{
1911 int i;
1912
1913 /*
1914 * reap exited threads (TD_EXITED -> TD_REAPED)
1915 */
1916 for (i = 0; i < thread_number; i++) {
1917 struct thread_data *td = &threads[i];
1918
1919 if (td->runstate != TD_EXITED)
1920 continue;
1921
1922 td_set_runstate(td, TD_REAPED);
1923
1924 if (td->use_thread) {
1925 long ret;
1926
1927 if (pthread_join(td->thread, (void *) &ret))
1928 perror("thread_join");
1929 } else
1930 waitpid(td->pid, NULL, 0);
1931
1932 (*nr_running)--;
1933 (*m_rate) -= td->ratemin;
1934 (*t_rate) -= td->rate;
1935 }
1936}
1937
1938static void run_threads(void)
1939{
1940 struct timeval genesis;
1941 struct thread_data *td;
1942 unsigned long spent;
1943 int i, todo, nr_running, m_rate, t_rate, nr_started;
1944
1945 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
1946 fflush(stdout);
1947
1948 signal(SIGINT, sig_handler);
1949 signal(SIGALRM, sig_handler);
1950
1951 todo = thread_number;
1952 nr_running = 0;
1953 nr_started = 0;
1954 m_rate = t_rate = 0;
1955
1956 for (i = 0; i < thread_number; i++) {
1957 td = &threads[i];
1958
1959 run_str[td->thread_number - 1] = 'P';
1960
1961 init_disk_util(td);
1962
1963 if (!td->create_serialize)
1964 continue;
1965
1966 /*
1967 * do file setup here so it happens sequentially,
1968 * we don't want X number of threads getting their
1969 * client data interspersed on disk
1970 */
1971 if (setup_file(td)) {
1972 td_set_runstate(td, TD_REAPED);
1973 todo--;
1974 }
1975 }
1976
1977 gettimeofday(&genesis, NULL);
1978
1979 while (todo) {
1980 /*
1981 * create threads (TD_NOT_CREATED -> TD_CREATED)
1982 */
1983 for (i = 0; i < thread_number; i++) {
1984 td = &threads[i];
1985
1986 if (td->runstate != TD_NOT_CREATED)
1987 continue;
1988
1989 /*
1990 * never got a chance to start, killed by other
1991 * thread for some reason
1992 */
1993 if (td->terminate) {
1994 todo--;
1995 continue;
1996 }
1997
1998 if (td->start_delay) {
1999 spent = mtime_since_now(&genesis);
2000
2001 if (td->start_delay * 1000 > spent)
2002 continue;
2003 }
2004
2005 if (td->stonewall && (nr_started || nr_running))
2006 break;
2007
2008 td_set_runstate(td, TD_CREATED);
2009 sem_init(&startup_sem, 0, 1);
2010 todo--;
2011 nr_started++;
2012
2013 if (td->use_thread) {
2014 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2015 perror("thread_create");
2016 nr_started--;
2017 }
2018 } else {
2019 if (fork())
2020 sem_wait(&startup_sem);
2021 else {
2022 fork_main(shm_id, i);
2023 exit(0);
2024 }
2025 }
2026 }
2027
2028 /*
2029 * start created threads (TD_CREATED -> TD_RUNNING)
2030 */
2031 for (i = 0; i < thread_number; i++) {
2032 td = &threads[i];
2033
2034 if (td->runstate != TD_CREATED)
2035 continue;
2036
2037 td_set_runstate(td, TD_RUNNING);
2038 nr_running++;
2039 nr_started--;
2040 m_rate += td->ratemin;
2041 t_rate += td->rate;
2042 sem_post(&td->mutex);
2043 }
2044
2045 reap_threads(&nr_running, &t_rate, &m_rate);
2046
2047 if (todo)
2048 usleep(100000);
2049 }
2050
2051 while (nr_running) {
2052 reap_threads(&nr_running, &t_rate, &m_rate);
2053 usleep(10000);
2054 }
2055
2056 update_io_ticks();
2057}
2058
2059static void show_group_stats(struct group_run_stats *rs, int id)
2060{
2061 printf("\nRun status group %d (all jobs):\n", id);
2062
2063 if (rs->max_run[DDIR_READ])
2064 printf(" READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[0], rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]);
2065 if (rs->max_run[DDIR_WRITE])
2066 printf(" WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
2067}
2068
2069static void show_disk_util(void)
2070{
2071 struct disk_util_stat *dus;
2072 struct list_head *entry;
2073 struct disk_util *du;
2074 double util;
2075
2076 printf("\nDisk stats (read/write):\n");
2077
2078 list_for_each(entry, &disk_list) {
2079 du = list_entry(entry, struct disk_util, list);
2080 dus = &du->dus;
2081
2082 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2083 if (util > 100.0)
2084 util = 100.0;
2085
2086 printf(" %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
2087 }
2088}
2089
2090static void show_run_stats(void)
2091{
2092 struct group_run_stats *runstats, *rs;
2093 struct thread_data *td;
2094 int i;
2095
2096 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2097
2098 for (i = 0; i < groupid + 1; i++) {
2099 rs = &runstats[i];
2100
2101 memset(rs, 0, sizeof(*rs));
2102 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2103 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2104 }
2105
2106 for (i = 0; i < thread_number; i++) {
2107 unsigned long rbw, wbw;
2108
2109 td = &threads[i];
2110
2111 if (td->error) {
2112 printf("Client%d: %s\n", td->thread_number, td->verror);
2113 continue;
2114 }
2115
2116 rs = &runstats[td->groupid];
2117
2118 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2119 rs->min_run[0] = td->runtime[0];
2120 if (td->runtime[0] > rs->max_run[0])
2121 rs->max_run[0] = td->runtime[0];
2122 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2123 rs->min_run[1] = td->runtime[1];
2124 if (td->runtime[1] > rs->max_run[1])
2125 rs->max_run[1] = td->runtime[1];
2126
2127 rbw = wbw = 0;
2128 if (td->runtime[0])
2129 rbw = td->io_bytes[0] / td->runtime[0];
2130 if (td->runtime[1])
2131 wbw = td->io_bytes[1] / td->runtime[1];
2132
2133 if (rbw < rs->min_bw[0])
2134 rs->min_bw[0] = rbw;
2135 if (wbw < rs->min_bw[1])
2136 rs->min_bw[1] = wbw;
2137 if (rbw > rs->max_bw[0])
2138 rs->max_bw[0] = rbw;
2139 if (wbw > rs->max_bw[1])
2140 rs->max_bw[1] = wbw;
2141
2142 rs->io_mb[0] += td->io_bytes[0] >> 20;
2143 rs->io_mb[1] += td->io_bytes[1] >> 20;
2144 }
2145
2146 for (i = 0; i < groupid + 1; i++) {
2147 rs = &runstats[i];
2148
2149 if (rs->max_run[0])
2150 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2151 if (rs->max_run[1])
2152 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2153 }
2154
2155 /*
2156 * don't overwrite last signal output
2157 */
2158 printf("\n");
2159
2160 for (i = 0; i < thread_number; i++) {
2161 td = &threads[i];
2162 rs = &runstats[td->groupid];
2163
2164 show_thread_status(td, rs);
2165 }
2166
2167 for (i = 0; i < groupid + 1; i++)
2168 show_group_stats(&runstats[i], i);
2169
2170 show_disk_util();
2171}
2172
2173int main(int argc, char *argv[])
2174{
2175 if (parse_options(argc, argv))
2176 return 1;
2177
2178 if (!thread_number) {
2179 printf("Nothing to do\n");
2180 return 1;
2181 }
2182
2183 disk_util_timer_arm();
2184
2185 run_threads();
2186 show_run_stats();
2187
2188 return 0;
2189}
fio - Flexible IO Tester