projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
[PATCH] Silence bogus gcc 4.x warnings
[fio.git] / fio.c
... / ...
CommitLineData
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;
54static struct timeval genesis;
55
56static void update_io_ticks(void);
57static void disk_util_timer_arm(void);
58static void print_thread_status(void);
59
60/*
61 * thread life cycle
62 */
63enum {
64 TD_NOT_CREATED = 0,
65 TD_CREATED,
66 TD_RUNNING,
67 TD_VERIFYING,
68 TD_EXITED,
69 TD_REAPED,
70};
71
72#define should_fsync(td) (td_write(td) && (!(td)->odirect || (td)->override_sync))
73
74static sem_t startup_sem;
75
76#define TERMINATE_ALL (-1)
77
78static void terminate_threads(int group_id)
79{
80 int i;
81
82 for (i = 0; i < thread_number; i++) {
83 struct thread_data *td = &threads[i];
84
85 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
86 td->terminate = 1;
87 td->start_delay = 0;
88 }
89 }
90}
91
92static void sig_handler(int sig)
93{
94 switch (sig) {
95 case SIGALRM:
96 update_io_ticks();
97 disk_util_timer_arm();
98 print_thread_status();
99 break;
100 default:
101 printf("\nfio: terminating on signal\n");
102 fflush(stdout);
103 terminate_threads(TERMINATE_ALL);
104 break;
105 }
106}
107
108static unsigned long utime_since(struct timeval *s, struct timeval *e)
109{
110 double sec, usec;
111
112 sec = e->tv_sec - s->tv_sec;
113 usec = e->tv_usec - s->tv_usec;
114 if (sec > 0 && usec < 0) {
115 sec--;
116 usec += 1000000;
117 }
118
119 sec *= (double) 1000000;
120
121 return sec + usec;
122}
123
124static unsigned long utime_since_now(struct timeval *s)
125{
126 struct timeval t;
127
128 gettimeofday(&t, NULL);
129 return utime_since(s, &t);
130}
131
132static unsigned long mtime_since(struct timeval *s, struct timeval *e)
133{
134 double sec, usec;
135
136 sec = e->tv_sec - s->tv_sec;
137 usec = e->tv_usec - s->tv_usec;
138 if (sec > 0 && usec < 0) {
139 sec--;
140 usec += 1000000;
141 }
142
143 sec *= (double) 1000;
144 usec /= (double) 1000;
145
146 return sec + usec;
147}
148
149static unsigned long mtime_since_now(struct timeval *s)
150{
151 struct timeval t;
152
153 gettimeofday(&t, NULL);
154 return mtime_since(s, &t);
155}
156
157static inline unsigned long msec_now(struct timeval *s)
158{
159 return s->tv_sec * 1000 + s->tv_usec / 1000;
160}
161
162static unsigned long time_since_now(struct timeval *s)
163{
164 return mtime_since_now(s) / 1000;
165}
166
167static int random_map_free(struct thread_data *td, unsigned long long block)
168{
169 unsigned int idx = RAND_MAP_IDX(td, block);
170 unsigned int bit = RAND_MAP_BIT(td, block);
171
172 return (td->file_map[idx] & (1UL << bit)) == 0;
173}
174
175static int get_next_free_block(struct thread_data *td, unsigned long long *b)
176{
177 int i;
178
179 *b = 0;
180 i = 0;
181 while ((*b) * td->min_bs < td->io_size) {
182 if (td->file_map[i] != -1UL) {
183 *b += ffz(td->file_map[i]);
184 return 0;
185 }
186
187 *b += BLOCKS_PER_MAP;
188 i++;
189 }
190
191 return 1;
192}
193
194static void mark_random_map(struct thread_data *td, struct io_u *io_u)
195{
196 unsigned long block = io_u->offset / td->min_bs;
197 unsigned int blocks = 0;
198
199 while (blocks < (io_u->buflen / td->min_bs)) {
200 unsigned int idx, bit;
201
202 if (!random_map_free(td, block))
203 break;
204
205 idx = RAND_MAP_IDX(td, block);
206 bit = RAND_MAP_BIT(td, block);
207
208 assert(idx < td->num_maps);
209
210 td->file_map[idx] |= (1UL << bit);
211 block++;
212 blocks++;
213 }
214
215 if ((blocks * td->min_bs) < io_u->buflen)
216 io_u->buflen = blocks * td->min_bs;
217}
218
219static inline void add_stat_sample(struct io_stat *is, unsigned long val)
220{
221 if (val > is->max_val)
222 is->max_val = val;
223 if (val < is->min_val)
224 is->min_val = val;
225
226 is->val += val;
227 is->val_sq += val * val;
228 is->samples++;
229}
230
231static void add_log_sample(struct thread_data *td, struct io_log *iolog,
232 unsigned long val, int ddir)
233{
234 if (iolog->nr_samples == iolog->max_samples) {
235 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
236
237 iolog->log = realloc(iolog->log, new_size);
238 iolog->max_samples <<= 1;
239 }
240
241 iolog->log[iolog->nr_samples].val = val;
242 iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
243 iolog->log[iolog->nr_samples].ddir = ddir;
244 iolog->nr_samples++;
245}
246
247static void add_clat_sample(struct thread_data *td, int ddir,unsigned long msec)
248{
249 add_stat_sample(&td->clat_stat[ddir], msec);
250
251 if (td->clat_log)
252 add_log_sample(td, td->clat_log, msec, ddir);
253}
254
255static void add_slat_sample(struct thread_data *td, int ddir,unsigned long msec)
256{
257 add_stat_sample(&td->slat_stat[ddir], msec);
258
259 if (td->slat_log)
260 add_log_sample(td, td->slat_log, msec, ddir);
261}
262
263static void add_bw_sample(struct thread_data *td, int ddir)
264{
265 unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
266 unsigned long rate;
267
268 if (spent < td->bw_avg_time)
269 return;
270
271 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
272 add_stat_sample(&td->bw_stat[ddir], rate);
273
274 if (td->bw_log)
275 add_log_sample(td, td->bw_log, rate, ddir);
276
277 gettimeofday(&td->stat_sample_time[ddir], NULL);
278 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
279}
280
281static int get_next_offset(struct thread_data *td, unsigned long long *offset)
282{
283 unsigned long long b, rb;
284 long r;
285
286 if (!td->sequential) {
287 unsigned long max_blocks = td->io_size / td->min_bs;
288 int loops = 50;
289
290 do {
291 lrand48_r(&td->random_state, &r);
292 b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
293 rb = b + (td->file_offset / td->min_bs);
294 loops--;
295 } while (!random_map_free(td, rb) && loops);
296
297 if (!loops) {
298 if (get_next_free_block(td, &b))
299 return 1;
300 }
301 } else
302 b = td->last_pos / td->min_bs;
303
304 *offset = (b * td->min_bs) + td->file_offset;
305 if (*offset > td->real_file_size)
306 return 1;
307
308 return 0;
309}
310
311static unsigned int get_next_buflen(struct thread_data *td)
312{
313 unsigned int buflen;
314 long r;
315
316 if (td->min_bs == td->max_bs)
317 buflen = td->min_bs;
318 else {
319 lrand48_r(&td->bsrange_state, &r);
320 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
321 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
322 }
323
324 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
325 buflen = td->io_size - td->this_io_bytes[td->ddir];
326
327 return buflen;
328}
329
330/*
331 * busy looping version for the last few usec
332 */
333static void __usec_sleep(unsigned int usec)
334{
335 struct timeval start;
336
337 gettimeofday(&start, NULL);
338 while (utime_since_now(&start) < usec)
339 nop;
340}
341
342static void usec_sleep(struct thread_data *td, unsigned long usec)
343{
344 struct timespec req, rem;
345
346 req.tv_sec = usec / 1000000;
347 req.tv_nsec = usec * 1000 - req.tv_sec * 1000000;
348
349 do {
350 if (usec < 5000) {
351 __usec_sleep(usec);
352 break;
353 }
354
355 rem.tv_sec = rem.tv_nsec = 0;
356 if (nanosleep(&req, &rem) < 0)
357 break;
358
359 if ((rem.tv_sec + rem.tv_nsec) == 0)
360 break;
361
362 req.tv_nsec = rem.tv_nsec;
363 req.tv_sec = rem.tv_sec;
364
365 usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000;
366 } while (!td->terminate);
367}
368
369static void rate_throttle(struct thread_data *td, unsigned long time_spent,
370 unsigned int bytes)
371{
372 unsigned long usec_cycle;
373
374 if (!td->rate)
375 return;
376
377 usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs);
378
379 if (time_spent < usec_cycle) {
380 unsigned long s = usec_cycle - time_spent;
381
382 td->rate_pending_usleep += s;
383 if (td->rate_pending_usleep >= 100000) {
384 usec_sleep(td, td->rate_pending_usleep);
385 td->rate_pending_usleep = 0;
386 }
387 } else {
388 long overtime = time_spent - usec_cycle;
389
390 td->rate_pending_usleep -= overtime;
391 }
392}
393
394static int check_min_rate(struct thread_data *td, struct timeval *now)
395{
396 unsigned long spent;
397 unsigned long rate;
398 int ddir = td->ddir;
399
400 /*
401 * allow a 2 second settle period in the beginning
402 */
403 if (mtime_since(&td->start, now) < 2000)
404 return 0;
405
406 /*
407 * if rate blocks is set, sample is running
408 */
409 if (td->rate_bytes) {
410 spent = mtime_since(&td->lastrate, now);
411 if (spent < td->ratecycle)
412 return 0;
413
414 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
415 if (rate < td->ratemin) {
416 printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
417 if (rate_quit)
418 terminate_threads(td->groupid);
419 return 1;
420 }
421 }
422
423 td->rate_bytes = td->this_io_bytes[ddir];
424 memcpy(&td->lastrate, now, sizeof(*now));
425 return 0;
426}
427
428static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
429{
430 if (!td->timeout)
431 return 0;
432 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
433 return 1;
434
435 return 0;
436}
437
438static void fill_random_bytes(struct thread_data *td,
439 unsigned char *p, unsigned int len)
440{
441 unsigned int todo;
442 double r;
443
444 while (len) {
445 drand48_r(&td->verify_state, &r);
446
447 /*
448 * lrand48_r seems to be broken and only fill the bottom
449 * 32-bits, even on 64-bit archs with 64-bit longs
450 */
451 todo = sizeof(r);
452 if (todo > len)
453 todo = len;
454
455 memcpy(p, &r, todo);
456
457 len -= todo;
458 p += todo;
459 }
460}
461
462static void hexdump(void *buffer, int len)
463{
464 unsigned char *p = buffer;
465 int i;
466
467 for (i = 0; i < len; i++)
468 printf("%02x", p[i]);
469 printf("\n");
470}
471
472static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
473{
474 unsigned char *p = (unsigned char *) io_u->buf;
475 unsigned long c;
476 int ret;
477
478 p += sizeof(*hdr);
479 c = crc32(p, hdr->len - sizeof(*hdr));
480 ret = c != hdr->crc32;
481
482 if (ret) {
483 fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
484 fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
485 }
486
487 return ret;
488}
489
490static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
491{
492 unsigned char *p = (unsigned char *) io_u->buf;
493 struct md5_ctx md5_ctx;
494 int ret;
495
496 memset(&md5_ctx, 0, sizeof(md5_ctx));
497 p += sizeof(*hdr);
498 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
499
500 ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
501 if (ret) {
502 fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
503 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
504 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
505 }
506
507 return ret;
508}
509
510static int verify_io_u(struct io_u *io_u)
511{
512 struct verify_header *hdr = (struct verify_header *) io_u->buf;
513 int ret;
514
515 if (hdr->fio_magic != FIO_HDR_MAGIC)
516 return 1;
517
518 if (hdr->verify_type == VERIFY_MD5)
519 ret = verify_io_u_md5(hdr, io_u);
520 else if (hdr->verify_type == VERIFY_CRC32)
521 ret = verify_io_u_crc32(hdr, io_u);
522 else {
523 fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
524 ret = 1;
525 }
526
527 return ret;
528}
529
530static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
531{
532 hdr->crc32 = crc32(p, len);
533}
534
535static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
536{
537 struct md5_ctx md5_ctx;
538
539 memset(&md5_ctx, 0, sizeof(md5_ctx));
540 md5_update(&md5_ctx, p, len);
541 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
542}
543
544/*
545 * fill body of io_u->buf with random data and add a header with the
546 * (eg) sha1sum of that data.
547 */
548static void populate_io_u(struct thread_data *td, struct io_u *io_u)
549{
550 unsigned char *p = (unsigned char *) io_u->buf;
551 struct verify_header hdr;
552
553 hdr.fio_magic = FIO_HDR_MAGIC;
554 hdr.len = io_u->buflen;
555 p += sizeof(hdr);
556 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
557
558 if (td->verify == VERIFY_MD5) {
559 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
560 hdr.verify_type = VERIFY_MD5;
561 } else {
562 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
563 hdr.verify_type = VERIFY_CRC32;
564 }
565
566 memcpy(io_u->buf, &hdr, sizeof(hdr));
567}
568
569static int td_io_prep(struct thread_data *td, struct io_u *io_u, int read)
570{
571 if (read)
572 io_u->ddir = DDIR_READ;
573 else
574 io_u->ddir = DDIR_WRITE;
575
576 if (td->io_prep && td->io_prep(td, io_u))
577 return 1;
578
579 return 0;
580}
581
582void put_io_u(struct thread_data *td, struct io_u *io_u)
583{
584 list_del(&io_u->list);
585 list_add(&io_u->list, &td->io_u_freelist);
586 td->cur_depth--;
587}
588
589#define queue_full(td) (list_empty(&(td)->io_u_freelist))
590
591struct io_u *__get_io_u(struct thread_data *td)
592{
593 struct io_u *io_u;
594
595 if (queue_full(td))
596 return NULL;
597
598 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
599 io_u->error = 0;
600 io_u->resid = 0;
601 list_del(&io_u->list);
602 list_add(&io_u->list, &td->io_u_busylist);
603 td->cur_depth++;
604 return io_u;
605}
606
607static struct io_u *get_io_u(struct thread_data *td)
608{
609 struct io_u *io_u;
610
611 io_u = __get_io_u(td);
612 if (!io_u)
613 return NULL;
614
615 if (td->zone_bytes >= td->zone_size) {
616 td->zone_bytes = 0;
617 td->last_pos += td->zone_skip;
618 }
619
620 if (get_next_offset(td, &io_u->offset)) {
621 put_io_u(td, io_u);
622 return NULL;
623 }
624
625 io_u->buflen = get_next_buflen(td);
626 if (!io_u->buflen) {
627 put_io_u(td, io_u);
628 return NULL;
629 }
630
631 if (io_u->buflen + io_u->offset > td->real_file_size)
632 io_u->buflen = td->real_file_size - io_u->offset;
633
634 if (!io_u->buflen) {
635 put_io_u(td, io_u);
636 return NULL;
637 }
638
639 if (!td->sequential)
640 mark_random_map(td, io_u);
641
642 td->last_pos += io_u->buflen;
643
644 if (td->verify != VERIFY_NONE)
645 populate_io_u(td, io_u);
646
647 if (td_io_prep(td, io_u, td_read(td))) {
648 put_io_u(td, io_u);
649 return NULL;
650 }
651
652 gettimeofday(&io_u->start_time, NULL);
653 return io_u;
654}
655
656static inline void td_set_runstate(struct thread_data *td, int runstate)
657{
658 td->old_runstate = td->runstate;
659 td->runstate = runstate;
660}
661
662static int get_next_verify(struct thread_data *td,
663 unsigned long long *offset, unsigned int *len)
664{
665 struct io_piece *ipo;
666
667 if (list_empty(&td->io_hist_list))
668 return 1;
669
670 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
671 list_del(&ipo->list);
672
673 *offset = ipo->offset;
674 *len = ipo->len;
675 free(ipo);
676 return 0;
677}
678
679static void prune_io_piece_log(struct thread_data *td)
680{
681 struct io_piece *ipo;
682
683 while (!list_empty(&td->io_hist_list)) {
684 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
685
686 list_del(&ipo->list);
687 free(ipo);
688 }
689}
690
691/*
692 * log a succesful write, so we can unwind the log for verify
693 */
694static void log_io_piece(struct thread_data *td, struct io_u *io_u)
695{
696 struct io_piece *ipo = malloc(sizeof(struct io_piece));
697 struct list_head *entry;
698
699 INIT_LIST_HEAD(&ipo->list);
700 ipo->offset = io_u->offset;
701 ipo->len = io_u->buflen;
702
703 /*
704 * for random io where the writes extend the file, it will typically
705 * be laid out with the block scattered as written. it's faster to
706 * read them in in that order again, so don't sort
707 */
708 if (td->sequential || !td->overwrite) {
709 list_add_tail(&ipo->list, &td->io_hist_list);
710 return;
711 }
712
713 /*
714 * for random io, sort the list so verify will run faster
715 */
716 entry = &td->io_hist_list;
717 while ((entry = entry->prev) != &td->io_hist_list) {
718 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
719
720 if (__ipo->offset < ipo->offset)
721 break;
722 }
723
724 list_add(&ipo->list, entry);
725}
726
727static int sync_td(struct thread_data *td)
728{
729 if (td->io_sync)
730 return td->io_sync(td);
731
732 return 0;
733}
734
735static int io_u_getevents(struct thread_data *td, int min, int max,
736 struct timespec *t)
737{
738 return td->io_getevents(td, min, max, t);
739}
740
741static int io_u_queue(struct thread_data *td, struct io_u *io_u)
742{
743 gettimeofday(&io_u->issue_time, NULL);
744
745 return td->io_queue(td, io_u);
746}
747
748#define iocb_time(iocb) ((unsigned long) (iocb)->data)
749
750static void io_completed(struct thread_data *td, struct io_u *io_u,
751 struct io_completion_data *icd)
752{
753 struct timeval e;
754 unsigned long msec;
755
756 gettimeofday(&e, NULL);
757
758 if (!io_u->error) {
759 unsigned int bytes = io_u->buflen - io_u->resid;
760 int idx = io_u->ddir;
761
762 td->io_blocks[idx]++;
763 td->io_bytes[idx] += bytes;
764 td->zone_bytes += bytes;
765 td->this_io_bytes[idx] += bytes;
766
767 msec = mtime_since(&io_u->issue_time, &e);
768
769 add_clat_sample(td, io_u->ddir, msec);
770 add_bw_sample(td, io_u->ddir);
771
772 if (td_write(td) && io_u->ddir == DDIR_WRITE)
773 log_io_piece(td, io_u);
774
775 icd->bytes_done[idx] += bytes;
776 } else
777 icd->error = io_u->error;
778}
779
780static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
781{
782 struct io_u *io_u;
783 int i;
784
785 icd->error = 0;
786 icd->bytes_done[0] = icd->bytes_done[1] = 0;
787
788 for (i = 0; i < icd->nr; i++) {
789 io_u = td->io_event(td, i);
790
791 io_completed(td, io_u, icd);
792 put_io_u(td, io_u);
793 }
794}
795
796static void cleanup_pending_aio(struct thread_data *td)
797{
798 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
799 struct list_head *entry, *n;
800 struct io_completion_data icd;
801 struct io_u *io_u;
802 int r;
803
804 /*
805 * get immediately available events, if any
806 */
807 r = io_u_getevents(td, 0, td->cur_depth, &ts);
808 if (r > 0) {
809 icd.nr = r;
810 ios_completed(td, &icd);
811 }
812
813 /*
814 * now cancel remaining active events
815 */
816 if (td->io_cancel) {
817 list_for_each_safe(entry, n, &td->io_u_busylist) {
818 io_u = list_entry(entry, struct io_u, list);
819
820 r = td->io_cancel(td, io_u);
821 if (!r)
822 put_io_u(td, io_u);
823 }
824 }
825
826 if (td->cur_depth) {
827 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
828 if (r > 0) {
829 icd.nr = r;
830 ios_completed(td, &icd);
831 }
832 }
833}
834
835static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
836{
837 struct io_u *v_io_u = *io_u;
838 int ret = 0;
839
840 if (v_io_u) {
841 ret = verify_io_u(v_io_u);
842 put_io_u(td, v_io_u);
843 *io_u = NULL;
844 }
845
846 return ret;
847}
848
849static void do_verify(struct thread_data *td)
850{
851 struct timeval t;
852 struct io_u *io_u, *v_io_u = NULL;
853 struct io_completion_data icd;
854 int ret;
855
856 td_set_runstate(td, TD_VERIFYING);
857
858 do {
859 if (td->terminate)
860 break;
861
862 gettimeofday(&t, NULL);
863 if (runtime_exceeded(td, &t))
864 break;
865
866 io_u = __get_io_u(td);
867 if (!io_u)
868 break;
869
870 if (get_next_verify(td, &io_u->offset, &io_u->buflen)) {
871 put_io_u(td, io_u);
872 break;
873 }
874
875 if (td_io_prep(td, io_u, 1)) {
876 put_io_u(td, io_u);
877 break;
878 }
879
880 ret = io_u_queue(td, io_u);
881 if (ret) {
882 put_io_u(td, io_u);
883 td_verror(td, ret);
884 break;
885 }
886
887 /*
888 * we have one pending to verify, do that while
889 * we are doing io on the next one
890 */
891 if (do_io_u_verify(td, &v_io_u))
892 break;
893
894 ret = io_u_getevents(td, 1, 1, NULL);
895 if (ret != 1) {
896 if (ret < 0)
897 td_verror(td, ret);
898 break;
899 }
900
901 v_io_u = td->io_event(td, 0);
902 icd.nr = 1;
903 icd.error = 0;
904 io_completed(td, v_io_u, &icd);
905
906 if (icd.error) {
907 td_verror(td, icd.error);
908 put_io_u(td, v_io_u);
909 v_io_u = NULL;
910 break;
911 }
912
913 /*
914 * if we can't submit more io, we need to verify now
915 */
916 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
917 break;
918
919 } while (1);
920
921 do_io_u_verify(td, &v_io_u);
922
923 if (td->cur_depth)
924 cleanup_pending_aio(td);
925
926 td_set_runstate(td, TD_RUNNING);
927}
928
929static void do_io(struct thread_data *td)
930{
931 struct io_completion_data icd;
932 struct timeval s, e;
933 unsigned long usec;
934
935 while (td->this_io_bytes[td->ddir] < td->io_size) {
936 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
937 struct timespec *timeout;
938 int ret, min_evts = 0;
939 struct io_u *io_u;
940
941 if (td->terminate)
942 break;
943
944 io_u = get_io_u(td);
945 if (!io_u)
946 break;
947
948 memcpy(&s, &io_u->start_time, sizeof(s));
949
950 ret = io_u_queue(td, io_u);
951 if (ret) {
952 put_io_u(td, io_u);
953 td_verror(td, ret);
954 break;
955 }
956
957 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
958
959 if (td->cur_depth < td->iodepth) {
960 timeout = &ts;
961 min_evts = 0;
962 } else {
963 timeout = NULL;
964 min_evts = 1;
965 }
966
967 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
968 if (ret < 0) {
969 td_verror(td, ret);
970 break;
971 } else if (!ret)
972 continue;
973
974 icd.nr = ret;
975 ios_completed(td, &icd);
976 if (icd.error) {
977 td_verror(td, icd.error);
978 break;
979 }
980
981 /*
982 * the rate is batched for now, it should work for batches
983 * of completions except the very first one which may look
984 * a little bursty
985 */
986 gettimeofday(&e, NULL);
987 usec = utime_since(&s, &e);
988
989 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
990
991 if (check_min_rate(td, &e)) {
992 td_verror(td, ENOMEM);
993 break;
994 }
995
996 if (runtime_exceeded(td, &e))
997 break;
998
999 if (td->thinktime)
1000 usec_sleep(td, td->thinktime);
1001
1002 if (should_fsync(td) && td->fsync_blocks &&
1003 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
1004 sync_td(td);
1005 }
1006
1007 if (td->cur_depth)
1008 cleanup_pending_aio(td);
1009
1010 if (should_fsync(td))
1011 sync_td(td);
1012}
1013
1014static void cleanup_io(struct thread_data *td)
1015{
1016 if (td->io_cleanup)
1017 td->io_cleanup(td);
1018}
1019
1020static int init_io(struct thread_data *td)
1021{
1022 if (td->io_engine == FIO_SYNCIO)
1023 return fio_syncio_init(td);
1024 else if (td->io_engine == FIO_MMAPIO)
1025 return fio_mmapio_init(td);
1026 else if (td->io_engine == FIO_LIBAIO)
1027 return fio_libaio_init(td);
1028 else if (td->io_engine == FIO_POSIXAIO)
1029 return fio_posixaio_init(td);
1030 else if (td->io_engine == FIO_SGIO)
1031 return fio_sgio_init(td);
1032 else {
1033 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1034 return 1;
1035 }
1036}
1037
1038static void cleanup_io_u(struct thread_data *td)
1039{
1040 struct list_head *entry, *n;
1041 struct io_u *io_u;
1042
1043 list_for_each_safe(entry, n, &td->io_u_freelist) {
1044 io_u = list_entry(entry, struct io_u, list);
1045
1046 list_del(&io_u->list);
1047 free(io_u);
1048 }
1049
1050 if (td->mem_type == MEM_MALLOC)
1051 free(td->orig_buffer);
1052 else if (td->mem_type == MEM_SHM) {
1053 struct shmid_ds sbuf;
1054
1055 shmdt(td->orig_buffer);
1056 shmctl(td->shm_id, IPC_RMID, &sbuf);
1057 } else if (td->mem_type == MEM_MMAP)
1058 munmap(td->orig_buffer, td->orig_buffer_size);
1059 else
1060 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1061
1062 td->orig_buffer = NULL;
1063}
1064
1065static int init_io_u(struct thread_data *td)
1066{
1067 struct io_u *io_u;
1068 int i, max_units;
1069 char *p;
1070
1071 if (td->io_engine & FIO_SYNCIO)
1072 max_units = 1;
1073 else
1074 max_units = td->iodepth;
1075
1076 td->orig_buffer_size = td->max_bs * max_units + MASK;
1077
1078 if (td->mem_type == MEM_MALLOC)
1079 td->orig_buffer = malloc(td->orig_buffer_size);
1080 else if (td->mem_type == MEM_SHM) {
1081 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1082 if (td->shm_id < 0) {
1083 td_verror(td, errno);
1084 perror("shmget");
1085 return 1;
1086 }
1087
1088 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1089 if (td->orig_buffer == (void *) -1) {
1090 td_verror(td, errno);
1091 perror("shmat");
1092 td->orig_buffer = NULL;
1093 return 1;
1094 }
1095 } else if (td->mem_type == MEM_MMAP) {
1096 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1097 if (td->orig_buffer == MAP_FAILED) {
1098 td_verror(td, errno);
1099 perror("mmap");
1100 td->orig_buffer = NULL;
1101 return 1;
1102 }
1103 }
1104
1105 INIT_LIST_HEAD(&td->io_u_freelist);
1106 INIT_LIST_HEAD(&td->io_u_busylist);
1107 INIT_LIST_HEAD(&td->io_hist_list);
1108
1109 p = ALIGN(td->orig_buffer);
1110 for (i = 0; i < max_units; i++) {
1111 io_u = malloc(sizeof(*io_u));
1112 memset(io_u, 0, sizeof(*io_u));
1113 INIT_LIST_HEAD(&io_u->list);
1114
1115 io_u->buf = p + td->max_bs * i;
1116 io_u->index = i;
1117 list_add(&io_u->list, &td->io_u_freelist);
1118 }
1119
1120 return 0;
1121}
1122
1123static int create_file(struct thread_data *td, unsigned long long size,
1124 int extend)
1125{
1126 unsigned long long left;
1127 unsigned int bs;
1128 int r, oflags;
1129 char *b;
1130
1131 /*
1132 * unless specifically asked for overwrite, let normal io extend it
1133 */
1134 if (td_write(td) && !td->overwrite)
1135 return 0;
1136
1137 if (!size) {
1138 fprintf(stderr, "Need size for create\n");
1139 td_verror(td, EINVAL);
1140 return 1;
1141 }
1142
1143 if (!extend) {
1144 oflags = O_CREAT | O_TRUNC;
1145 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1146 } else {
1147 oflags = O_APPEND;
1148 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1149 }
1150
1151 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1152 if (td->fd < 0) {
1153 td_verror(td, errno);
1154 return 1;
1155 }
1156
1157 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1158 td_verror(td, errno);
1159 return 1;
1160 }
1161
1162 td->io_size = td->file_size;
1163 b = malloc(td->max_bs);
1164 memset(b, 0, td->max_bs);
1165
1166 left = size;
1167 while (left && !td->terminate) {
1168 bs = td->max_bs;
1169 if (bs > left)
1170 bs = left;
1171
1172 r = write(td->fd, b, bs);
1173
1174 if (r == (int) bs) {
1175 left -= bs;
1176 continue;
1177 } else {
1178 if (r < 0)
1179 td_verror(td, errno);
1180 else
1181 td_verror(td, EIO);
1182
1183 break;
1184 }
1185 }
1186
1187 if (td->terminate)
1188 unlink(td->file_name);
1189 else if (td->create_fsync)
1190 fsync(td->fd);
1191
1192 close(td->fd);
1193 td->fd = -1;
1194 free(b);
1195 return 0;
1196}
1197
1198static int file_size(struct thread_data *td)
1199{
1200 struct stat st;
1201
1202 if (fstat(td->fd, &st) == -1) {
1203 td_verror(td, errno);
1204 return 1;
1205 }
1206
1207 td->real_file_size = st.st_size;
1208
1209 if (!td->file_size || td->file_size > td->real_file_size)
1210 td->file_size = td->real_file_size;
1211
1212 return 0;
1213}
1214
1215static int bdev_size(struct thread_data *td)
1216{
1217 unsigned long long bytes;
1218 int r;
1219
1220 r = blockdev_size(td->fd, &bytes);
1221 if (r) {
1222 td_verror(td, r);
1223 return 1;
1224 }
1225
1226 td->real_file_size = bytes;
1227
1228 /*
1229 * no extend possibilities, so limit size to device size if too large
1230 */
1231 if (!td->file_size || td->file_size > td->real_file_size)
1232 td->file_size = td->real_file_size;
1233
1234 return 0;
1235}
1236
1237static int get_file_size(struct thread_data *td)
1238{
1239 int ret = 0;
1240
1241 if (td->filetype == FIO_TYPE_FILE)
1242 ret = file_size(td);
1243 else if (td->filetype == FIO_TYPE_BD)
1244 ret = bdev_size(td);
1245 else
1246 td->real_file_size = -1;
1247
1248 if (ret)
1249 return ret;
1250
1251 if (td->file_offset + td->file_size > td->real_file_size) {
1252 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset + td->file_size, td->real_file_size);
1253 return 1;
1254 }
1255
1256 td->io_size = td->file_size;
1257 if (td->io_size == 0) {
1258 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1259 td_verror(td, EINVAL);
1260 return 1;
1261 }
1262
1263 if (!td->zone_size)
1264 td->zone_size = td->io_size;
1265
1266 td->total_io_size = td->io_size * td->loops;
1267 return 0;
1268}
1269
1270static int setup_file_mmap(struct thread_data *td)
1271{
1272 int flags;
1273
1274 if (td_read(td))
1275 flags = PROT_READ;
1276 else {
1277 flags = PROT_WRITE;
1278
1279 if (td->verify != VERIFY_NONE)
1280 flags |= PROT_READ;
1281 }
1282
1283 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1284 if (td->mmap == MAP_FAILED) {
1285 td->mmap = NULL;
1286 td_verror(td, errno);
1287 return 1;
1288 }
1289
1290 if (td->invalidate_cache) {
1291 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1292 td_verror(td, errno);
1293 return 1;
1294 }
1295 }
1296
1297 if (td->sequential) {
1298 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1299 td_verror(td, errno);
1300 return 1;
1301 }
1302 } else {
1303 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1304 td_verror(td, errno);
1305 return 1;
1306 }
1307 }
1308
1309 return 0;
1310}
1311
1312static int setup_file_plain(struct thread_data *td)
1313{
1314 if (td->invalidate_cache) {
1315 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1316 td_verror(td, errno);
1317 return 1;
1318 }
1319 }
1320
1321 if (td->sequential) {
1322 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1323 td_verror(td, errno);
1324 return 1;
1325 }
1326 } else {
1327 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1328 td_verror(td, errno);
1329 return 1;
1330 }
1331 }
1332
1333 return 0;
1334}
1335
1336static int setup_file(struct thread_data *td)
1337{
1338 struct stat st;
1339 int flags = 0;
1340
1341 if (stat(td->file_name, &st) == -1) {
1342 if (errno != ENOENT) {
1343 td_verror(td, errno);
1344 return 1;
1345 }
1346 if (!td->create_file) {
1347 td_verror(td, ENOENT);
1348 return 1;
1349 }
1350 if (create_file(td, td->file_size, 0))
1351 return 1;
1352 } else if (td->filetype == FIO_TYPE_FILE) {
1353 if (st.st_size < td->file_size) {
1354 if (create_file(td, td->file_size - st.st_size, 1))
1355 return 1;
1356 }
1357 }
1358
1359 if (td->odirect)
1360 flags |= O_DIRECT;
1361
1362 if (td_read(td)) {
1363 if (td->filetype == FIO_TYPE_CHAR)
1364 flags |= O_RDWR;
1365 else
1366 flags |= O_RDONLY;
1367
1368 td->fd = open(td->file_name, flags);
1369 } else {
1370 if (td->filetype == FIO_TYPE_FILE) {
1371 if (!td->overwrite)
1372 flags |= O_TRUNC;
1373
1374 flags |= O_CREAT;
1375 }
1376 if (td->sync_io)
1377 flags |= O_SYNC;
1378
1379 flags |= O_RDWR;
1380
1381 td->fd = open(td->file_name, flags, 0600);
1382 }
1383
1384 if (td->fd == -1) {
1385 td_verror(td, errno);
1386 return 1;
1387 }
1388
1389 if (get_file_size(td))
1390 return 1;
1391
1392 if (td->io_engine != FIO_MMAPIO)
1393 return setup_file_plain(td);
1394 else
1395 return setup_file_mmap(td);
1396}
1397
1398static int check_dev_match(dev_t dev, char *path)
1399{
1400 unsigned int major, minor;
1401 char line[256], *p;
1402 FILE *f;
1403
1404 f = fopen(path, "r");
1405 if (!f) {
1406 perror("open path");
1407 return 1;
1408 }
1409
1410 p = fgets(line, sizeof(line), f);
1411 if (!p) {
1412 fclose(f);
1413 return 1;
1414 }
1415
1416 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1417 fclose(f);
1418 return 1;
1419 }
1420
1421 if (((major << 8) | minor) == dev) {
1422 fclose(f);
1423 return 0;
1424 }
1425
1426 fclose(f);
1427 return 1;
1428}
1429
1430static int find_block_dir(dev_t dev, char *path)
1431{
1432 struct dirent *dir;
1433 struct stat st;
1434 int found = 0;
1435 DIR *D;
1436
1437 D = opendir(path);
1438 if (!D)
1439 return 0;
1440
1441 while ((dir = readdir(D)) != NULL) {
1442 char full_path[256];
1443
1444 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1445 continue;
1446 if (!strcmp(dir->d_name, "device"))
1447 continue;
1448
1449 sprintf(full_path, "%s/%s", path, dir->d_name);
1450
1451 if (!strcmp(dir->d_name, "dev")) {
1452 if (!check_dev_match(dev, full_path)) {
1453 found = 1;
1454 break;
1455 }
1456 }
1457
1458 if (stat(full_path, &st) == -1) {
1459 perror("stat");
1460 break;
1461 }
1462
1463 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1464 continue;
1465
1466 found = find_block_dir(dev, full_path);
1467 if (found) {
1468 strcpy(path, full_path);
1469 break;
1470 }
1471 }
1472
1473 closedir(D);
1474 return found;
1475}
1476
1477static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1478{
1479 unsigned in_flight;
1480 char line[256];
1481 FILE *f;
1482 char *p;
1483
1484 f = fopen(du->path, "r");
1485 if (!f)
1486 return 1;
1487
1488 p = fgets(line, sizeof(line), f);
1489 if (!p) {
1490 fclose(f);
1491 return 1;
1492 }
1493
1494 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) {
1495 fclose(f);
1496 return 1;
1497 }
1498
1499 fclose(f);
1500 return 0;
1501}
1502
1503static void update_io_tick_disk(struct disk_util *du)
1504{
1505 struct disk_util_stat __dus, *dus, *ldus;
1506 struct timeval t;
1507
1508 if (get_io_ticks(du, &__dus))
1509 return;
1510
1511 dus = &du->dus;
1512 ldus = &du->last_dus;
1513
1514 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1515 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1516 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1517 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1518 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1519 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1520 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1521 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1522 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1523 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1524
1525 gettimeofday(&t, NULL);
1526 du->msec += mtime_since(&du->time, &t);
1527 memcpy(&du->time, &t, sizeof(t));
1528 memcpy(ldus, &__dus, sizeof(__dus));
1529}
1530
1531static void update_io_ticks(void)
1532{
1533 struct list_head *entry;
1534 struct disk_util *du;
1535
1536 list_for_each(entry, &disk_list) {
1537 du = list_entry(entry, struct disk_util, list);
1538 update_io_tick_disk(du);
1539 }
1540}
1541
1542static int disk_util_exists(dev_t dev)
1543{
1544 struct list_head *entry;
1545 struct disk_util *du;
1546
1547 list_for_each(entry, &disk_list) {
1548 du = list_entry(entry, struct disk_util, list);
1549
1550 if (du->dev == dev)
1551 return 1;
1552 }
1553
1554 return 0;
1555}
1556
1557static void disk_util_add(dev_t dev, char *path)
1558{
1559 struct disk_util *du = malloc(sizeof(*du));
1560
1561 memset(du, 0, sizeof(*du));
1562 INIT_LIST_HEAD(&du->list);
1563 sprintf(du->path, "%s/stat", path);
1564 du->name = strdup(basename(path));
1565 du->dev = dev;
1566
1567 gettimeofday(&du->time, NULL);
1568 get_io_ticks(du, &du->last_dus);
1569
1570 list_add_tail(&du->list, &disk_list);
1571}
1572
1573static void init_disk_util(struct thread_data *td)
1574{
1575 struct stat st;
1576 char foo[256], tmp[256];
1577 dev_t dev;
1578 char *p;
1579
1580 if (!td->do_disk_util)
1581 return;
1582
1583 if (!stat(td->file_name, &st)) {
1584 if (S_ISBLK(st.st_mode))
1585 dev = st.st_rdev;
1586 else
1587 dev = st.st_dev;
1588 } else {
1589 /*
1590 * must be a file, open "." in that path
1591 */
1592 strcpy(foo, td->file_name);
1593 p = dirname(foo);
1594 if (stat(p, &st)) {
1595 perror("disk util stat");
1596 return;
1597 }
1598
1599 dev = st.st_dev;
1600 }
1601
1602 if (disk_util_exists(dev))
1603 return;
1604
1605 sprintf(foo, "/sys/block");
1606 if (!find_block_dir(dev, foo))
1607 return;
1608
1609 /*
1610 * for md/dm, there's no queue dir. we already have the right place
1611 */
1612 sprintf(tmp, "%s/stat", foo);
1613 if (stat(tmp, &st)) {
1614 /*
1615 * if this is inside a partition dir, jump back to parent
1616 */
1617 sprintf(tmp, "%s/queue", foo);
1618 if (stat(tmp, &st)) {
1619 p = dirname(foo);
1620 sprintf(tmp, "%s/queue", p);
1621 if (stat(tmp, &st)) {
1622 fprintf(stderr, "unknown sysfs layout\n");
1623 return;
1624 }
1625 sprintf(foo, "%s", p);
1626 }
1627 }
1628
1629 disk_util_add(dev, foo);
1630}
1631
1632static void disk_util_timer_arm(void)
1633{
1634 itimer.it_value.tv_sec = 0;
1635 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1636 setitimer(ITIMER_REAL, &itimer, NULL);
1637}
1638
1639static void clear_io_state(struct thread_data *td)
1640{
1641 if (td->io_engine == FIO_SYNCIO)
1642 lseek(td->fd, SEEK_SET, 0);
1643
1644 td->last_pos = 0;
1645 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1646 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1647 td->zone_bytes = 0;
1648
1649 if (td->file_map)
1650 memset(td->file_map, 0, td->num_maps * sizeof(long));
1651}
1652
1653static void update_rusage_stat(struct thread_data *td)
1654{
1655 if (!(td->runtime[0] + td->runtime[1]))
1656 return;
1657
1658 getrusage(RUSAGE_SELF, &td->ru_end);
1659
1660 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1661 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1662 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1663
1664
1665 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1666}
1667
1668static void *thread_main(void *data)
1669{
1670 struct thread_data *td = data;
1671 int ret = 1;
1672
1673 if (!td->use_thread)
1674 setsid();
1675
1676 td->pid = getpid();
1677
1678 if (init_io_u(td))
1679 goto err;
1680
1681 if (fio_setaffinity(td) == -1) {
1682 td_verror(td, errno);
1683 goto err;
1684 }
1685
1686 if (init_io(td))
1687 goto err;
1688
1689 if (td->ioprio) {
1690 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1691 td_verror(td, errno);
1692 goto err;
1693 }
1694 }
1695
1696 sem_post(&startup_sem);
1697 sem_wait(&td->mutex);
1698
1699 if (!td->create_serialize && setup_file(td))
1700 goto err;
1701
1702 if (init_random_state(td))
1703 goto err;
1704
1705 gettimeofday(&td->epoch, NULL);
1706
1707 while (td->loops--) {
1708 getrusage(RUSAGE_SELF, &td->ru_start);
1709 gettimeofday(&td->start, NULL);
1710 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1711
1712 if (td->ratemin)
1713 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1714
1715 clear_io_state(td);
1716 prune_io_piece_log(td);
1717
1718 do_io(td);
1719
1720 td->runtime[td->ddir] += mtime_since_now(&td->start);
1721 update_rusage_stat(td);
1722
1723 if (td->error || td->terminate)
1724 break;
1725
1726 if (td->verify == VERIFY_NONE)
1727 continue;
1728
1729 clear_io_state(td);
1730 gettimeofday(&td->start, NULL);
1731
1732 do_verify(td);
1733
1734 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1735
1736 if (td->error || td->terminate)
1737 break;
1738 }
1739
1740 ret = 0;
1741
1742 if (td->bw_log)
1743 finish_log(td, td->bw_log, "bw");
1744 if (td->slat_log)
1745 finish_log(td, td->slat_log, "slat");
1746 if (td->clat_log)
1747 finish_log(td, td->clat_log, "clat");
1748
1749 if (exitall_on_terminate)
1750 terminate_threads(td->groupid);
1751
1752err:
1753 if (td->fd != -1) {
1754 close(td->fd);
1755 td->fd = -1;
1756 }
1757 if (td->mmap)
1758 munmap(td->mmap, td->file_size);
1759 cleanup_io(td);
1760 cleanup_io_u(td);
1761 if (ret) {
1762 sem_post(&startup_sem);
1763 sem_wait(&td->mutex);
1764 }
1765 td_set_runstate(td, TD_EXITED);
1766 return NULL;
1767
1768}
1769
1770static void *fork_main(int shmid, int offset)
1771{
1772 struct thread_data *td;
1773 void *data;
1774
1775 data = shmat(shmid, NULL, 0);
1776 if (data == (void *) -1) {
1777 perror("shmat");
1778 return NULL;
1779 }
1780
1781 td = data + offset * sizeof(struct thread_data);
1782 thread_main(td);
1783 shmdt(data);
1784 return NULL;
1785}
1786
1787static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1788 double *mean, double *dev)
1789{
1790 double n;
1791
1792 if (is->samples == 0)
1793 return 0;
1794
1795 *min = is->min_val;
1796 *max = is->max_val;
1797
1798 n = (double) is->samples;
1799 *mean = (double) is->val / n;
1800 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1801 if (!(*min + *max) && !(*mean + *dev))
1802 return 0;
1803
1804 return 1;
1805}
1806
1807static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1808 int ddir)
1809{
1810 char *ddir_str[] = { "read ", "write" };
1811 unsigned long min, max, bw;
1812 double mean, dev;
1813
1814 if (!td->runtime[ddir])
1815 return;
1816
1817 bw = td->io_bytes[ddir] / td->runtime[ddir];
1818 printf(" %s: io=%6lluMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
1819
1820 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1821 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1822
1823 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1824 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1825
1826 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1827 double p_of_agg;
1828
1829 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1830 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);
1831 }
1832}
1833
1834static void show_thread_status(struct thread_data *td,
1835 struct group_run_stats *rs)
1836{
1837 double usr_cpu, sys_cpu;
1838
1839 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1840 return;
1841
1842 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1843
1844 show_ddir_status(td, rs, td->ddir);
1845 show_ddir_status(td, rs, td->ddir ^ 1);
1846
1847 if (td->runtime[0] + td->runtime[1]) {
1848 double runt = td->runtime[0] + td->runtime[1];
1849
1850 usr_cpu = (double) td->usr_time * 100 / runt;
1851 sys_cpu = (double) td->sys_time * 100 / runt;
1852 } else {
1853 usr_cpu = 0;
1854 sys_cpu = 0;
1855 }
1856
1857 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
1858}
1859
1860static void check_str_update(struct thread_data *td)
1861{
1862 char c = run_str[td->thread_number - 1];
1863
1864 if (td->runstate == td->old_runstate)
1865 return;
1866
1867 switch (td->runstate) {
1868 case TD_REAPED:
1869 c = '_';
1870 break;
1871 case TD_EXITED:
1872 c = 'E';
1873 break;
1874 case TD_RUNNING:
1875 if (td_read(td)) {
1876 if (td->sequential)
1877 c = 'R';
1878 else
1879 c = 'r';
1880 } else {
1881 if (td->sequential)
1882 c = 'W';
1883 else
1884 c = 'w';
1885 }
1886 break;
1887 case TD_VERIFYING:
1888 c = 'V';
1889 break;
1890 case TD_CREATED:
1891 c = 'C';
1892 break;
1893 case TD_NOT_CREATED:
1894 c = 'P';
1895 break;
1896 default:
1897 printf("state %d\n", td->runstate);
1898 }
1899
1900 run_str[td->thread_number - 1] = c;
1901 td->old_runstate = td->runstate;
1902}
1903
1904static void eta_to_str(char *str, int eta_sec)
1905{
1906 unsigned int d, h, m, s;
1907 static int always_d, always_h;
1908
1909 d = h = m = s = 0;
1910
1911 s = eta_sec % 60;
1912 eta_sec /= 60;
1913 m = eta_sec % 60;
1914 eta_sec /= 60;
1915 h = eta_sec % 24;
1916 eta_sec /= 24;
1917 d = eta_sec;
1918
1919 if (d || always_d) {
1920 always_d = 1;
1921 str += sprintf(str, "%02dd:", d);
1922 }
1923 if (h || always_h) {
1924 always_h = 1;
1925 str += sprintf(str, "%02dh:", h);
1926 }
1927
1928 str += sprintf(str, "%02dm:", m);
1929 str += sprintf(str, "%02ds", s);
1930}
1931
1932static void print_thread_status(void)
1933{
1934 unsigned long long bytes_done, bytes_total;
1935 int i, nr_running, t_rate, m_rate, eta_sec;
1936 char eta_str[32];
1937 double perc;
1938
1939 bytes_done = bytes_total = 0;
1940 nr_running = t_rate = m_rate = 0;
1941 for (i = 0; i < thread_number; i++) {
1942 struct thread_data *td = &threads[i];
1943
1944 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
1945 nr_running++;
1946 t_rate += td->rate;
1947 m_rate += td->ratemin;
1948 }
1949
1950 bytes_total += td->total_io_size;
1951 if (td->verify)
1952 bytes_total += td->total_io_size;
1953
1954 if (td->zone_size && td->zone_skip)
1955 bytes_total /= (td->zone_skip / td->zone_size);
1956
1957 bytes_done += td->io_bytes[DDIR_READ] +td->io_bytes[DDIR_WRITE];
1958
1959 check_str_update(td);
1960 }
1961
1962 perc = 0;
1963 eta_sec = 0;
1964 if (bytes_total && bytes_done) {
1965 unsigned long runtime;
1966
1967 perc = (double) bytes_done / (double) bytes_total;
1968 if (perc > 1.0)
1969 perc = 1.0;
1970
1971 runtime = time_since_now(&genesis);
1972 if (runtime >= 5) {
1973 memset(eta_str, 0, sizeof(eta_str));
1974 eta_sec = (runtime * (1.0 / perc)) - runtime;
1975 eta_to_str(eta_str, eta_sec);
1976 }
1977
1978 perc *= 100.0;
1979 }
1980
1981 printf("Threads now running (%d)", nr_running);
1982 if (m_rate || t_rate)
1983 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
1984 printf(": [%s] [%3.2f%% done]", run_str, perc);
1985 if (eta_sec)
1986 printf(" [eta %s]", eta_str);
1987 printf("\r");
1988 fflush(stdout);
1989}
1990
1991static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1992{
1993 int i;
1994
1995 /*
1996 * reap exited threads (TD_EXITED -> TD_REAPED)
1997 */
1998 for (i = 0; i < thread_number; i++) {
1999 struct thread_data *td = &threads[i];
2000
2001 if (td->runstate != TD_EXITED)
2002 continue;
2003
2004 td_set_runstate(td, TD_REAPED);
2005
2006 if (td->use_thread) {
2007 long ret;
2008
2009 if (pthread_join(td->thread, (void *) &ret))
2010 perror("thread_join");
2011 } else
2012 waitpid(td->pid, NULL, 0);
2013
2014 (*nr_running)--;
2015 (*m_rate) -= td->ratemin;
2016 (*t_rate) -= td->rate;
2017 }
2018}
2019
2020static void run_threads(void)
2021{
2022 struct thread_data *td;
2023 unsigned long spent;
2024 int i, todo, nr_running, m_rate, t_rate, nr_started;
2025
2026 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
2027 fflush(stdout);
2028
2029 signal(SIGINT, sig_handler);
2030 signal(SIGALRM, sig_handler);
2031
2032 todo = thread_number;
2033 nr_running = 0;
2034 nr_started = 0;
2035 m_rate = t_rate = 0;
2036
2037 for (i = 0; i < thread_number; i++) {
2038 td = &threads[i];
2039
2040 run_str[td->thread_number - 1] = 'P';
2041
2042 init_disk_util(td);
2043
2044 if (!td->create_serialize)
2045 continue;
2046
2047 /*
2048 * do file setup here so it happens sequentially,
2049 * we don't want X number of threads getting their
2050 * client data interspersed on disk
2051 */
2052 if (setup_file(td)) {
2053 td_set_runstate(td, TD_REAPED);
2054 todo--;
2055 }
2056 }
2057
2058 gettimeofday(&genesis, NULL);
2059
2060 while (todo) {
2061 /*
2062 * create threads (TD_NOT_CREATED -> TD_CREATED)
2063 */
2064 for (i = 0; i < thread_number; i++) {
2065 td = &threads[i];
2066
2067 if (td->runstate != TD_NOT_CREATED)
2068 continue;
2069
2070 /*
2071 * never got a chance to start, killed by other
2072 * thread for some reason
2073 */
2074 if (td->terminate) {
2075 todo--;
2076 continue;
2077 }
2078
2079 if (td->start_delay) {
2080 spent = mtime_since_now(&genesis);
2081
2082 if (td->start_delay * 1000 > spent)
2083 continue;
2084 }
2085
2086 if (td->stonewall && (nr_started || nr_running))
2087 break;
2088
2089 td_set_runstate(td, TD_CREATED);
2090 sem_init(&startup_sem, 0, 1);
2091 todo--;
2092 nr_started++;
2093
2094 if (td->use_thread) {
2095 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2096 perror("thread_create");
2097 nr_started--;
2098 }
2099 } else {
2100 if (fork())
2101 sem_wait(&startup_sem);
2102 else {
2103 fork_main(shm_id, i);
2104 exit(0);
2105 }
2106 }
2107 }
2108
2109 /*
2110 * start created threads (TD_CREATED -> TD_RUNNING)
2111 */
2112 for (i = 0; i < thread_number; i++) {
2113 td = &threads[i];
2114
2115 if (td->runstate != TD_CREATED)
2116 continue;
2117
2118 td_set_runstate(td, TD_RUNNING);
2119 nr_running++;
2120 nr_started--;
2121 m_rate += td->ratemin;
2122 t_rate += td->rate;
2123 sem_post(&td->mutex);
2124 }
2125
2126 reap_threads(&nr_running, &t_rate, &m_rate);
2127
2128 if (todo)
2129 usleep(100000);
2130 }
2131
2132 while (nr_running) {
2133 reap_threads(&nr_running, &t_rate, &m_rate);
2134 usleep(10000);
2135 }
2136
2137 update_io_ticks();
2138}
2139
2140static void show_group_stats(struct group_run_stats *rs, int id)
2141{
2142 printf("\nRun status group %d (all jobs):\n", id);
2143
2144 if (rs->max_run[DDIR_READ])
2145 printf(" READ: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_mb[0], rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]);
2146 if (rs->max_run[DDIR_WRITE])
2147 printf(" WRITE: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
2148}
2149
2150static void show_disk_util(void)
2151{
2152 struct disk_util_stat *dus;
2153 struct list_head *entry;
2154 struct disk_util *du;
2155 double util;
2156
2157 printf("\nDisk stats (read/write):\n");
2158
2159 list_for_each(entry, &disk_list) {
2160 du = list_entry(entry, struct disk_util, list);
2161 dus = &du->dus;
2162
2163 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2164 if (util > 100.0)
2165 util = 100.0;
2166
2167 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);
2168 }
2169}
2170
2171static void show_run_stats(void)
2172{
2173 struct group_run_stats *runstats, *rs;
2174 struct thread_data *td;
2175 int i;
2176
2177 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2178
2179 for (i = 0; i < groupid + 1; i++) {
2180 rs = &runstats[i];
2181
2182 memset(rs, 0, sizeof(*rs));
2183 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2184 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2185 }
2186
2187 for (i = 0; i < thread_number; i++) {
2188 unsigned long rbw, wbw;
2189
2190 td = &threads[i];
2191
2192 if (td->error) {
2193 printf("Client%d: %s\n", td->thread_number, td->verror);
2194 continue;
2195 }
2196
2197 rs = &runstats[td->groupid];
2198
2199 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2200 rs->min_run[0] = td->runtime[0];
2201 if (td->runtime[0] > rs->max_run[0])
2202 rs->max_run[0] = td->runtime[0];
2203 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2204 rs->min_run[1] = td->runtime[1];
2205 if (td->runtime[1] > rs->max_run[1])
2206 rs->max_run[1] = td->runtime[1];
2207
2208 rbw = wbw = 0;
2209 if (td->runtime[0])
2210 rbw = td->io_bytes[0] / td->runtime[0];
2211 if (td->runtime[1])
2212 wbw = td->io_bytes[1] / td->runtime[1];
2213
2214 if (rbw < rs->min_bw[0])
2215 rs->min_bw[0] = rbw;
2216 if (wbw < rs->min_bw[1])
2217 rs->min_bw[1] = wbw;
2218 if (rbw > rs->max_bw[0])
2219 rs->max_bw[0] = rbw;
2220 if (wbw > rs->max_bw[1])
2221 rs->max_bw[1] = wbw;
2222
2223 rs->io_mb[0] += td->io_bytes[0] >> 20;
2224 rs->io_mb[1] += td->io_bytes[1] >> 20;
2225 }
2226
2227 for (i = 0; i < groupid + 1; i++) {
2228 rs = &runstats[i];
2229
2230 if (rs->max_run[0])
2231 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2232 if (rs->max_run[1])
2233 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2234 }
2235
2236 /*
2237 * don't overwrite last signal output
2238 */
2239 printf("\n");
2240
2241 for (i = 0; i < thread_number; i++) {
2242 td = &threads[i];
2243 rs = &runstats[td->groupid];
2244
2245 show_thread_status(td, rs);
2246 }
2247
2248 for (i = 0; i < groupid + 1; i++)
2249 show_group_stats(&runstats[i], i);
2250
2251 show_disk_util();
2252}
2253
2254int main(int argc, char *argv[])
2255{
2256 if (parse_options(argc, argv))
2257 return 1;
2258
2259 if (!thread_number) {
2260 printf("Nothing to do\n");
2261 return 1;
2262 }
2263
2264 disk_util_timer_arm();
2265
2266 run_threads();
2267 show_run_stats();
2268
2269 return 0;
2270}
fio - Flexible IO Tester