projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
[PATCH] Include ETA for job(s) to finish
[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;
5289b847 54static struct timeval genesis;
ebac4655
JA		 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
5289b847
JA		 162static unsigned long time_since_now(struct timeval *s)
163{
164 return mtime_since_now(s) / 1000;
165}
166
ebac4655
JA		 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 int get_next_offset(struct thread_data *td, unsigned long long *offset)
220{
221 unsigned long long b, rb;
222 long r;
223
224 if (!td->sequential) {
225 unsigned long max_blocks = td->io_size / td->min_bs;
226 int loops = 50;
227
228 do {
229 lrand48_r(&td->random_state, &r);
230 b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
231 rb = b + (td->file_offset / td->min_bs);
232 loops--;
233 } while (!random_map_free(td, rb) && loops);
234
235 if (!loops) {
236 if (get_next_free_block(td, &b))
237 return 1;
238 }
239 } else
240 b = td->last_bytes / td->min_bs;
241
242 *offset = (b * td->min_bs) + td->file_offset;
243 if (*offset > td->file_size)
244 return 1;
245
246 return 0;
247}
248
249static unsigned int get_next_buflen(struct thread_data *td)
250{
251 unsigned int buflen;
252 long r;
253
254 if (td->min_bs == td->max_bs)
255 buflen = td->min_bs;
256 else {
257 lrand48_r(&td->bsrange_state, &r);
258 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
259 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
260 }
261
262 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
263 buflen = td->io_size - td->this_io_bytes[td->ddir];
264
265 return buflen;
266}
267
268static inline void add_stat_sample(struct io_stat *is, unsigned long val)
269{
270 if (val > is->max_val)
271 is->max_val = val;
272 if (val < is->min_val)
273 is->min_val = val;
274
275 is->val += val;
276 is->val_sq += val * val;
277 is->samples++;
278}
279
280static void add_log_sample(struct thread_data *td, struct io_log *iolog,
281 unsigned long val, int ddir)
282{
283 if (iolog->nr_samples == iolog->max_samples) {
284 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
285
286 iolog->log = realloc(iolog->log, new_size);
287 iolog->max_samples <<= 1;
288 }
289
290 iolog->log[iolog->nr_samples].val = val;
291 iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
292 iolog->log[iolog->nr_samples].ddir = ddir;
293 iolog->nr_samples++;
294}
295
296static void add_clat_sample(struct thread_data *td, int ddir,unsigned long msec)
297{
298 add_stat_sample(&td->clat_stat[ddir], msec);
299
300 if (td->clat_log)
301 add_log_sample(td, td->clat_log, msec, ddir);
302}
303
304static void add_slat_sample(struct thread_data *td, int ddir,unsigned long msec)
305{
306 add_stat_sample(&td->slat_stat[ddir], msec);
307
308 if (td->slat_log)
309 add_log_sample(td, td->slat_log, msec, ddir);
310}
311
312static void add_bw_sample(struct thread_data *td, int ddir)
313{
314 unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
315 unsigned long rate;
316
317 if (spent < td->bw_avg_time)
318 return;
319
320 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
321 add_stat_sample(&td->bw_stat[ddir], rate);
322
323 if (td->bw_log)
324 add_log_sample(td, td->bw_log, rate, ddir);
325
326 gettimeofday(&td->stat_sample_time[ddir], NULL);
327 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
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 void put_io_u(struct thread_data *td, struct io_u *io_u)
570{
571 list_del(&io_u->list);
572 list_add(&io_u->list, &td->io_u_freelist);
573 td->cur_depth--;
574}
575
576#define queue_full(td) (list_empty(&(td)->io_u_freelist))
577
578static struct io_u *__get_io_u(struct thread_data *td)
579{
580 struct io_u *io_u;
581
582 if (queue_full(td))
583 return NULL;
584
585 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
586 io_u->error = 0;
587 io_u->resid = 0;
588 list_del(&io_u->list);
589 list_add(&io_u->list, &td->io_u_busylist);
590 td->cur_depth++;
591 return io_u;
592}
593
594static int td_io_prep(struct thread_data *td, struct io_u *io_u, int read)
595{
596 if (read)
597 io_u->ddir = DDIR_READ;
598 else
599 io_u->ddir = DDIR_WRITE;
600
601 if (td->io_prep && td->io_prep(td, io_u))
602 return 1;
603
604 return 0;
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 (get_next_offset(td, &io_u->offset)) {
616 put_io_u(td, io_u);
617 return NULL;
618 }
619
620 io_u->buflen = get_next_buflen(td);
621 if (!io_u->buflen) {
622 put_io_u(td, io_u);
623 return NULL;
624 }
625
626 if (io_u->buflen + io_u->offset > td->file_size)
627 io_u->buflen = td->file_size - io_u->offset;
628
629 if (!io_u->buflen) {
630 put_io_u(td, io_u);
631 return NULL;
632 }
633
634 if (!td->sequential)
635 mark_random_map(td, io_u);
636
637 td->last_bytes += io_u->buflen;
638
639 if (td->verify != VERIFY_NONE)
640 populate_io_u(td, io_u);
641
642 if (td_io_prep(td, io_u, td_read(td))) {
643 put_io_u(td, io_u);
644 return NULL;
645 }
646
647 gettimeofday(&io_u->start_time, NULL);
648 return io_u;
649}
650
651static inline void td_set_runstate(struct thread_data *td, int runstate)
652{
653 td->old_runstate = td->runstate;
654 td->runstate = runstate;
655}
656
657static int get_next_verify(struct thread_data *td,
658 unsigned long long *offset, unsigned int *len)
659{
660 struct io_piece *ipo;
661
662 if (list_empty(&td->io_hist_list))
663 return 1;
664
665 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
666 list_del(&ipo->list);
667
668 *offset = ipo->offset;
669 *len = ipo->len;
670 free(ipo);
671 return 0;
672}
673
674static void prune_io_piece_log(struct thread_data *td)
675{
676 struct io_piece *ipo;
677
678 while (!list_empty(&td->io_hist_list)) {
679 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
680
681 list_del(&ipo->list);
682 free(ipo);
683 }
684}
685
686/*
687 * log a succesful write, so we can unwind the log for verify
688 */
689static void log_io_piece(struct thread_data *td, struct io_u *io_u)
690{
691 struct io_piece *ipo = malloc(sizeof(struct io_piece));
692 struct list_head *entry;
693
694 INIT_LIST_HEAD(&ipo->list);
695 ipo->offset = io_u->offset;
696 ipo->len = io_u->buflen;
697
698 /*
699 * for random io where the writes extend the file, it will typically
700 * be laid out with the block scattered as written. it's faster to
701 * read them in in that order again, so don't sort
702 */
703 if (td->sequential || !td->overwrite) {
704 list_add_tail(&ipo->list, &td->io_hist_list);
705 return;
706 }
707
708 /*
709 * for random io, sort the list so verify will run faster
710 */
711 entry = &td->io_hist_list;
712 while ((entry = entry->prev) != &td->io_hist_list) {
713 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
714
715 if (__ipo->offset < ipo->offset)
716 break;
717 }
718
719 list_add(&ipo->list, entry);
720}
721
722static int sync_td(struct thread_data *td)
723{
724 if (td->io_sync)
725 return td->io_sync(td);
726
727 return 0;
728}
729
730static int io_u_getevents(struct thread_data *td, int min, int max,
731 struct timespec *t)
732{
733 return td->io_getevents(td, min, max, t);
734}
735
736static int io_u_queue(struct thread_data *td, struct io_u *io_u)
737{
738 gettimeofday(&io_u->issue_time, NULL);
739
740 return td->io_queue(td, io_u);
741}
742
743#define iocb_time(iocb) ((unsigned long) (iocb)->data)
744
745static void io_completed(struct thread_data *td, struct io_u *io_u,
746 struct io_completion_data *icd)
747{
748 struct timeval e;
749 unsigned long msec;
750
751 gettimeofday(&e, NULL);
752
753 if (!io_u->error) {
754 int idx = io_u->ddir;
755
756 td->io_blocks[idx]++;
757 td->io_bytes[idx] += (io_u->buflen - io_u->resid);
758 td->this_io_bytes[idx] += (io_u->buflen - io_u->resid);
759
760 msec = mtime_since(&io_u->issue_time, &e);
761
762 add_clat_sample(td, io_u->ddir, msec);
763 add_bw_sample(td, io_u->ddir);
764
765 if (td_write(td) && io_u->ddir == DDIR_WRITE)
766 log_io_piece(td, io_u);
767
768 icd->bytes_done[idx] += (io_u->buflen - io_u->resid);
769 } else
770 icd->error = io_u->error;
771}
772
773static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
774{
775 struct io_u *io_u;
776 int i;
777
778 icd->error = 0;
779 icd->bytes_done[0] = icd->bytes_done[1] = 0;
780
781 for (i = 0; i < icd->nr; i++) {
782 io_u = td->io_event(td, i);
783
784 io_completed(td, io_u, icd);
785 put_io_u(td, io_u);
786 }
787}
788
789static void cleanup_pending_aio(struct thread_data *td)
790{
791 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
792 struct list_head *entry, *n;
793 struct io_completion_data icd;
794 struct io_u *io_u;
795 int r;
796
797 /*
798 * get immediately available events, if any
799 */
800 r = io_u_getevents(td, 0, td->cur_depth, &ts);
801 if (r > 0) {
802 icd.nr = r;
803 ios_completed(td, &icd);
804 }
805
806 /*
807 * now cancel remaining active events
808 */
809 if (td->io_cancel) {
810 list_for_each_safe(entry, n, &td->io_u_busylist) {
811 io_u = list_entry(entry, struct io_u, list);
812
813 r = td->io_cancel(td, io_u);
814 if (!r)
815 put_io_u(td, io_u);
816 }
817 }
818
819 if (td->cur_depth) {
820 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
821 if (r > 0) {
822 icd.nr = r;
823 ios_completed(td, &icd);
824 }
825 }
826}
827
828static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
829{
830 struct io_u *v_io_u = *io_u;
831 int ret = 0;
832
833 if (v_io_u) {
834 ret = verify_io_u(v_io_u);
835 put_io_u(td, v_io_u);
836 *io_u = NULL;
837 }
838
839 return ret;
840}
841
842static void do_verify(struct thread_data *td)
843{
844 struct timeval t;
845 struct io_u *io_u, *v_io_u = NULL;
846 struct io_completion_data icd;
847 int ret;
848
849 td_set_runstate(td, TD_VERIFYING);
850
851 do {
852 if (td->terminate)
853 break;
854
855 gettimeofday(&t, NULL);
856 if (runtime_exceeded(td, &t))
857 break;
858
859 io_u = __get_io_u(td);
860 if (!io_u)
861 break;
862
863 if (get_next_verify(td, &io_u->offset, &io_u->buflen)) {
864 put_io_u(td, io_u);
865 break;
866 }
867
868 if (td_io_prep(td, io_u, 1)) {
869 put_io_u(td, io_u);
870 break;
871 }
872
873 ret = io_u_queue(td, io_u);
874 if (ret) {
875 put_io_u(td, io_u);
876 td_verror(td, ret);
877 break;
878 }
879
880 /*
881 * we have one pending to verify, do that while
882 * we are doing io on the next one
883 */
884 if (do_io_u_verify(td, &v_io_u))
885 break;
886
887 ret = io_u_getevents(td, 1, 1, NULL);
888 if (ret != 1) {
889 if (ret < 0)
890 td_verror(td, ret);
891 break;
892 }
893
894 v_io_u = td->io_event(td, 0);
895 icd.nr = 1;
896 icd.error = 0;
897 io_completed(td, v_io_u, &icd);
898
899 if (icd.error) {
900 td_verror(td, icd.error);
901 put_io_u(td, v_io_u);
902 v_io_u = NULL;
903 break;
904 }
905
ebac4655
JA		 906 /*
907 * if we can't submit more io, we need to verify now
908 */
909 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
910 break;
911
912 } while (1);
913
914 do_io_u_verify(td, &v_io_u);
915
916 if (td->cur_depth)
917 cleanup_pending_aio(td);
918
919 td_set_runstate(td, TD_RUNNING);
920}
921
922static void do_io(struct thread_data *td)
923{
924 struct io_completion_data icd;
925 struct timeval s, e;
926 unsigned long usec;
927
928 while (td->this_io_bytes[td->ddir] < td->io_size) {
929 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
930 struct timespec *timeout;
931 int ret, min_evts = 0;
932 struct io_u *io_u;
933
934 if (td->terminate)
935 break;
936
937 io_u = get_io_u(td);
938 if (!io_u)
939 break;
940
941 memcpy(&s, &io_u->start_time, sizeof(s));
942
943 ret = io_u_queue(td, io_u);
944 if (ret) {
945 put_io_u(td, io_u);
946 td_verror(td, ret);
947 break;
948 }
949
950 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
951
952 if (td->cur_depth < td->iodepth) {
953 timeout = &ts;
954 min_evts = 0;
955 } else {
956 timeout = NULL;
957 min_evts = 1;
958 }
959
960 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
961 if (ret < 0) {
962 td_verror(td, ret);
963 break;
964 } else if (!ret)
965 continue;
966
967 icd.nr = ret;
968 ios_completed(td, &icd);
969 if (icd.error) {
970 td_verror(td, icd.error);
971 break;
972 }
973
974 /*
975 * the rate is batched for now, it should work for batches
976 * of completions except the very first one which may look
977 * a little bursty
978 */
979 gettimeofday(&e, NULL);
980 usec = utime_since(&s, &e);
981
982 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
983
984 if (check_min_rate(td, &e)) {
985 td_verror(td, ENOMEM);
986 break;
987 }
988
989 if (runtime_exceeded(td, &e))
990 break;
991
992 if (td->thinktime)
993 usec_sleep(td, td->thinktime);
994
995 if (should_fsync(td) && td->fsync_blocks &&
996 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
997 sync_td(td);
998 }
999
1000 if (td->cur_depth)
1001 cleanup_pending_aio(td);
1002
1003 if (should_fsync(td))
1004 sync_td(td);
1005}
1006
1007static void cleanup_io(struct thread_data *td)
1008{
1009 if (td->io_cleanup)
1010 td->io_cleanup(td);
1011}
1012
1013static int init_io(struct thread_data *td)
1014{
1015 if (td->io_engine == FIO_SYNCIO)
1016 return fio_syncio_init(td);
1017 else if (td->io_engine == FIO_MMAPIO)
1018 return fio_mmapio_init(td);
1019 else if (td->io_engine == FIO_LIBAIO)
1020 return fio_libaio_init(td);
1021 else if (td->io_engine == FIO_POSIXAIO)
1022 return fio_posixaio_init(td);
1023 else if (td->io_engine == FIO_SGIO)
1024 return fio_sgio_init(td);
1025 else {
1026 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1027 return 1;
1028 }
1029}
1030
1031static void cleanup_io_u(struct thread_data *td)
1032{
1033 struct list_head *entry, *n;
1034 struct io_u *io_u;
1035
1036 list_for_each_safe(entry, n, &td->io_u_freelist) {
1037 io_u = list_entry(entry, struct io_u, list);
1038
1039 list_del(&io_u->list);
1040 free(io_u);
1041 }
1042
1043 if (td->mem_type == MEM_MALLOC)
1044 free(td->orig_buffer);
1045 else if (td->mem_type == MEM_SHM) {
1046 struct shmid_ds sbuf;
1047
1048 shmdt(td->orig_buffer);
1049 shmctl(td->shm_id, IPC_RMID, &sbuf);
1050 } else if (td->mem_type == MEM_MMAP)
1051 munmap(td->orig_buffer, td->orig_buffer_size);
1052 else
1053 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1054
1055 td->orig_buffer = NULL;
1056}
1057
1058static int init_io_u(struct thread_data *td)
1059{
1060 struct io_u *io_u;
1061 int i, max_units;
1062 char *p;
1063
1064 if (td->io_engine & FIO_SYNCIO)
1065 max_units = 1;
1066 else
1067 max_units = td->iodepth;
1068
1069 td->orig_buffer_size = td->max_bs * max_units + MASK;
1070
1071 if (td->mem_type == MEM_MALLOC)
1072 td->orig_buffer = malloc(td->orig_buffer_size);
1073 else if (td->mem_type == MEM_SHM) {
1074 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1075 if (td->shm_id < 0) {
1076 td_verror(td, errno);
1077 perror("shmget");
1078 return 1;
1079 }
1080
1081 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1082 if (td->orig_buffer == (void *) -1) {
1083 td_verror(td, errno);
1084 perror("shmat");
1085 td->orig_buffer = NULL;
1086 return 1;
1087 }
1088 } else if (td->mem_type == MEM_MMAP) {
1089 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1090 if (td->orig_buffer == MAP_FAILED) {
1091 td_verror(td, errno);
1092 perror("mmap");
1093 td->orig_buffer = NULL;
1094 return 1;
1095 }
1096 }
1097
1098 INIT_LIST_HEAD(&td->io_u_freelist);
1099 INIT_LIST_HEAD(&td->io_u_busylist);
1100 INIT_LIST_HEAD(&td->io_hist_list);
1101
1102 p = ALIGN(td->orig_buffer);
1103 for (i = 0; i < max_units; i++) {
1104 io_u = malloc(sizeof(*io_u));
1105 memset(io_u, 0, sizeof(*io_u));
1106 INIT_LIST_HEAD(&io_u->list);
1107
1108 io_u->buf = p + td->max_bs * i;
1109 list_add(&io_u->list, &td->io_u_freelist);
1110 }
1111
1112 return 0;
1113}
1114
1115static int create_file(struct thread_data *td, unsigned long long size,
1116 int extend)
1117{
1118 unsigned long long left;
1119 unsigned int bs;
1120 int r, oflags;
1121 char *b;
1122
1123 /*
1124 * unless specifically asked for overwrite, let normal io extend it
1125 */
1126 if (td_write(td) && !td->overwrite)
1127 return 0;
1128
1129 if (!size) {
1130 fprintf(stderr, "Need size for create\n");
1131 td_verror(td, EINVAL);
1132 return 1;
1133 }
1134
1135 if (!extend) {
1136 oflags = O_CREAT | O_TRUNC;
1137 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1138 } else {
1139 oflags = O_APPEND;
1140 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1141 }
1142
1143 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1144 if (td->fd < 0) {
1145 td_verror(td, errno);
1146 return 1;
1147 }
1148
1149 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1150 td_verror(td, errno);
1151 return 1;
1152 }
1153
1154 td->io_size = td->file_size;
1155 b = malloc(td->max_bs);
1156 memset(b, 0, td->max_bs);
1157
1158 left = size;
1159 while (left && !td->terminate) {
1160 bs = td->max_bs;
1161 if (bs > left)
1162 bs = left;
1163
1164 r = write(td->fd, b, bs);
1165
1166 if (r == (int) bs) {
1167 left -= bs;
1168 continue;
1169 } else {
1170 if (r < 0)
1171 td_verror(td, errno);
1172 else
1173 td_verror(td, EIO);
1174
1175 break;
1176 }
1177 }
1178
1179 if (td->terminate)
1180 unlink(td->file_name);
1181 else if (td->create_fsync)
1182 fsync(td->fd);
1183
1184 close(td->fd);
1185 td->fd = -1;
1186 free(b);
1187 return 0;
1188}
1189
1190static int file_size(struct thread_data *td)
1191{
1192 struct stat st;
1193
1194 if (fstat(td->fd, &st) == -1) {
1195 td_verror(td, errno);
1196 return 1;
1197 }
1198
1199 if (!td->file_size)
1200 td->file_size = st.st_size;
1201
1202 return 0;
1203}
1204
1205static int bdev_size(struct thread_data *td)
1206{
1207 size_t bytes;
1208 int r;
1209
1210 r = blockdev_size(td->fd, &bytes);
1211 if (r) {
1212 td_verror(td, r);
1213 return 1;
1214 }
1215
1216 /*
1217 * no extend possibilities, so limit size to device size if too large
1218 */
1219 if (!td->file_size || td->file_size > bytes)
1220 td->file_size = bytes;
1221
1222 return 0;
1223}
1224
1225static int get_file_size(struct thread_data *td)
1226{
1227 int ret;
1228
1229 if (td->filetype == FIO_TYPE_FILE)
1230 ret = file_size(td);
1231 else
1232 ret = bdev_size(td);
1233
1234 if (ret)
1235 return ret;
1236
1237 if (td->file_offset > td->file_size) {
1238 fprintf(stderr, "Client%d: offset larger than length (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->file_size);
1239 return 1;
1240 }
1241
1242 td->io_size = td->file_size - td->file_offset;
1243 if (td->io_size == 0) {
1244 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1245 td_verror(td, EINVAL);
1246 return 1;
1247 }
1248
1249 td->total_io_size = td->io_size * td->loops;
1250 return 0;
1251}
1252
1253static int setup_file_mmap(struct thread_data *td)
1254{
1255 int flags;
1256
1257 if (td_read(td))
1258 flags = PROT_READ;
1259 else {
1260 flags = PROT_WRITE;
1261
1262 if (td->verify != VERIFY_NONE)
1263 flags |= PROT_READ;
1264 }
1265
1266 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1267 if (td->mmap == MAP_FAILED) {
1268 td->mmap = NULL;
1269 td_verror(td, errno);
1270 return 1;
1271 }
1272
1273 if (td->invalidate_cache) {
1274 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1275 td_verror(td, errno);
1276 return 1;
1277 }
1278 }
1279
1280 if (td->sequential) {
1281 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1282 td_verror(td, errno);
1283 return 1;
1284 }
1285 } else {
1286 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1287 td_verror(td, errno);
1288 return 1;
1289 }
1290 }
1291
1292 return 0;
1293}
1294
1295static int setup_file_plain(struct thread_data *td)
1296{
1297 if (td->invalidate_cache) {
1298 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1299 td_verror(td, errno);
1300 return 1;
1301 }
1302 }
1303
1304 if (td->sequential) {
1305 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1306 td_verror(td, errno);
1307 return 1;
1308 }
1309 } else {
1310 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1311 td_verror(td, errno);
1312 return 1;
1313 }
1314 }
1315
1316 return 0;
1317}
1318
1319static int setup_file(struct thread_data *td)
1320{
1321 struct stat st;
1322 int flags = 0;
1323
1324 if (stat(td->file_name, &st) == -1) {
1325 if (errno != ENOENT) {
1326 td_verror(td, errno);
1327 return 1;
1328 }
1329 if (!td->create_file) {
1330 td_verror(td, ENOENT);
1331 return 1;
1332 }
1333 if (create_file(td, td->file_size, 0))
1334 return 1;
1335 } else if (td->filetype == FIO_TYPE_FILE) {
1336 if (st.st_size < td->file_size) {
1337 if (create_file(td, td->file_size - st.st_size, 1))
1338 return 1;
1339 }
1340 }
1341
1342 if (td->odirect)
1343 flags |= O_DIRECT;
1344
1345 if (td_read(td))
1346 td->fd = open(td->file_name, flags | O_RDONLY);
1347 else {
1348 if (td->filetype == FIO_TYPE_FILE) {
1349 if (!td->overwrite)
1350 flags |= O_TRUNC;
1351
1352 flags |= O_CREAT;
1353 }
1354 if (td->sync_io)
1355 flags |= O_SYNC;
1356
1357 flags |= O_RDWR;
1358
1359 td->fd = open(td->file_name, flags, 0600);
1360 }
1361
1362 if (td->fd == -1) {
1363 td_verror(td, errno);
1364 return 1;
1365 }
1366
1367 if (get_file_size(td))
1368 return 1;
1369
1370 if (td->io_engine != FIO_MMAPIO)
1371 return setup_file_plain(td);
1372 else
1373 return setup_file_mmap(td);
1374}
1375
1376static int check_dev_match(dev_t dev, char *path)
1377{
1378 unsigned int major, minor;
1379 char line[256], *p;
1380 FILE *f;
1381
1382 f = fopen(path, "r");
1383 if (!f) {
1384 perror("open path");
1385 return 1;
1386 }
1387
1388 p = fgets(line, sizeof(line), f);
1389 if (!p) {
1390 fclose(f);
1391 return 1;
1392 }
1393
1394 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1395 fclose(f);
1396 return 1;
1397 }
1398
1399 if (((major << 8) | minor) == dev) {
1400 fclose(f);
1401 return 0;
1402 }
1403
1404 fclose(f);
1405 return 1;
1406}
1407
1408static int find_block_dir(dev_t dev, char *path)
1409{
1410 struct dirent *dir;
1411 struct stat st;
1412 int found = 0;
1413 DIR *D;
1414
1415 D = opendir(path);
1416 if (!D)
1417 return 0;
1418
1419 while ((dir = readdir(D)) != NULL) {
1420 char full_path[256];
1421
1422 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1423 continue;
1424 if (!strcmp(dir->d_name, "device"))
1425 continue;
1426
1427 sprintf(full_path, "%s/%s", path, dir->d_name);
1428
1429 if (!strcmp(dir->d_name, "dev")) {
1430 if (!check_dev_match(dev, full_path)) {
1431 found = 1;
1432 break;
1433 }
1434 }
1435
1436 if (stat(full_path, &st) == -1) {
1437 perror("stat");
1438 break;
1439 }
1440
1441 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1442 continue;
1443
1444 found = find_block_dir(dev, full_path);
1445 if (found) {
1446 strcpy(path, full_path);
1447 break;
1448 }
1449 }
1450
1451 closedir(D);
1452 return found;
1453}
1454
1455static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1456{
1457 unsigned in_flight;
1458 char line[256];
1459 FILE *f;
1460 char *p;
1461
1462 f = fopen(du->path, "r");
1463 if (!f)
1464 return 1;
1465
1466 p = fgets(line, sizeof(line), f);
1467 if (!p) {
1468 fclose(f);
1469 return 1;
1470 }
1471
1472 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) {
1473 fclose(f);
1474 return 1;
1475 }
1476
1477 fclose(f);
1478 return 0;
1479}
1480
1481static void update_io_tick_disk(struct disk_util *du)
1482{
1483 struct disk_util_stat __dus, *dus, *ldus;
1484 struct timeval t;
1485
1486 if (get_io_ticks(du, &__dus))
1487 return;
1488
1489 dus = &du->dus;
1490 ldus = &du->last_dus;
1491
1492 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1493 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1494 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1495 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1496 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1497 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1498 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1499 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1500 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1501 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1502
1503 gettimeofday(&t, NULL);
1504 du->msec += mtime_since(&du->time, &t);
1505 memcpy(&du->time, &t, sizeof(t));
1506 memcpy(ldus, &__dus, sizeof(__dus));
1507}
1508
1509static void update_io_ticks(void)
1510{
1511 struct list_head *entry;
1512 struct disk_util *du;
1513
1514 list_for_each(entry, &disk_list) {
1515 du = list_entry(entry, struct disk_util, list);
1516 update_io_tick_disk(du);
1517 }
1518}
1519
1520static int disk_util_exists(dev_t dev)
1521{
1522 struct list_head *entry;
1523 struct disk_util *du;
1524
1525 list_for_each(entry, &disk_list) {
1526 du = list_entry(entry, struct disk_util, list);
1527
1528 if (du->dev == dev)
1529 return 1;
1530 }
1531
1532 return 0;
1533}
1534
1535static void disk_util_add(dev_t dev, char *path)
1536{
1537 struct disk_util *du = malloc(sizeof(*du));
1538
1539 memset(du, 0, sizeof(*du));
1540 INIT_LIST_HEAD(&du->list);
1541 sprintf(du->path, "%s/stat", path);
1542 du->name = strdup(basename(path));
1543 du->dev = dev;
1544
1545 gettimeofday(&du->time, NULL);
1546 get_io_ticks(du, &du->last_dus);
1547
1548 list_add_tail(&du->list, &disk_list);
1549}
1550
1551static void init_disk_util(struct thread_data *td)
1552{
1553 struct stat st;
1554 char foo[256], tmp[256];
1555 dev_t dev;
1556 char *p;
1557
1558 if (!td->do_disk_util)
1559 return;
1560
1561 if (!stat(td->file_name, &st)) {
1562 if (S_ISBLK(st.st_mode))
1563 dev = st.st_rdev;
1564 else
1565 dev = st.st_dev;
1566 } else {
1567 /*
1568 * must be a file, open "." in that path
1569 */
1570 strcpy(foo, td->file_name);
1571 p = dirname(foo);
1572 if (stat(p, &st)) {
1573 perror("disk util stat");
1574 return;
1575 }
1576
1577 dev = st.st_dev;
1578 }
1579
1580 if (disk_util_exists(dev))
1581 return;
1582
1583 sprintf(foo, "/sys/block");
1584 if (!find_block_dir(dev, foo))
1585 return;
1586
1587 /*
1588 * if this is inside a partition dir, jump back to parent
1589 */
1590 sprintf(tmp, "%s/queue", foo);
1591 if (stat(tmp, &st)) {
1592 p = dirname(foo);
1593 sprintf(tmp, "%s/queue", p);
1594 if (stat(tmp, &st)) {
1595 fprintf(stderr, "unknown sysfs layout\n");
1596 return;
1597 }
1598 sprintf(foo, "%s", p);
1599 }
1600
1601 disk_util_add(dev, foo);
1602}
1603
1604static void disk_util_timer_arm(void)
1605{
1606 itimer.it_value.tv_sec = 0;
1607 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1608 setitimer(ITIMER_REAL, &itimer, NULL);
1609}
1610
1611static void clear_io_state(struct thread_data *td)
1612{
1613 if (td->io_engine == FIO_SYNCIO)
1614 lseek(td->fd, SEEK_SET, 0);
1615
ebac4655
JA		 1616 td->last_bytes = 0;
1617 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1618 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1619
1620 if (td->file_map)
1621 memset(td->file_map, 0, td->num_maps * sizeof(long));
1622}
1623
1624static void update_rusage_stat(struct thread_data *td)
1625{
1626 if (!(td->runtime[0] + td->runtime[1]))
1627 return;
1628
1629 getrusage(RUSAGE_SELF, &td->ru_end);
1630
1631 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1632 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1633 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1634
1635
1636 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1637}
1638
1639static void *thread_main(void *data)
1640{
1641 struct thread_data *td = data;
1642 int ret = 1;
1643
1644 if (!td->use_thread)
1645 setsid();
1646
1647 td->pid = getpid();
1648
1649 if (init_io_u(td))
1650 goto err;
1651
1652 if (fio_setaffinity(td) == -1) {
1653 td_verror(td, errno);
1654 goto err;
1655 }
1656
1657 if (init_io(td))
1658 goto err;
1659
1660 if (td->ioprio) {
1661 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1662 td_verror(td, errno);
1663 goto err;
1664 }
1665 }
1666
1667 sem_post(&startup_sem);
1668 sem_wait(&td->mutex);
1669
1670 if (!td->create_serialize && setup_file(td))
1671 goto err;
1672
1673 if (init_random_state(td))
1674 goto err;
1675
1676 gettimeofday(&td->epoch, NULL);
1677
1678 while (td->loops--) {
1679 getrusage(RUSAGE_SELF, &td->ru_start);
1680 gettimeofday(&td->start, NULL);
1681 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1682
1683 if (td->ratemin)
1684 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1685
1686 clear_io_state(td);
1687 prune_io_piece_log(td);
1688
1689 do_io(td);
1690
1691 td->runtime[td->ddir] += mtime_since_now(&td->start);
1692 update_rusage_stat(td);
1693
1694 if (td->error || td->terminate)
1695 break;
1696
1697 if (td->verify == VERIFY_NONE)
1698 continue;
1699
1700 clear_io_state(td);
1701 gettimeofday(&td->start, NULL);
1702
1703 do_verify(td);
1704
1705 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1706
1707 if (td->error || td->terminate)
1708 break;
1709 }
1710
1711 ret = 0;
1712
1713 if (td->bw_log)
1714 finish_log(td, td->bw_log, "bw");
1715 if (td->slat_log)
1716 finish_log(td, td->slat_log, "slat");
1717 if (td->clat_log)
1718 finish_log(td, td->clat_log, "clat");
1719
1720 if (exitall_on_terminate)
1721 terminate_threads(td->groupid);
1722
1723err:
1724 if (td->fd != -1) {
1725 close(td->fd);
1726 td->fd = -1;
1727 }
1728 if (td->mmap)
1729 munmap(td->mmap, td->file_size);
1730 cleanup_io(td);
1731 cleanup_io_u(td);
1732 if (ret) {
1733 sem_post(&startup_sem);
1734 sem_wait(&td->mutex);
1735 }
1736 td_set_runstate(td, TD_EXITED);
1737 return NULL;
1738
1739}
1740
1741static void *fork_main(int shmid, int offset)
1742{
1743 struct thread_data *td;
1744 void *data;
1745
1746 data = shmat(shmid, NULL, 0);
1747 if (data == (void *) -1) {
1748 perror("shmat");
1749 return NULL;
1750 }
1751
1752 td = data + offset * sizeof(struct thread_data);
1753 thread_main(td);
1754 shmdt(data);
1755 return NULL;
1756}
1757
1758static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1759 double *mean, double *dev)
1760{
1761 double n;
1762
1763 if (is->samples == 0)
1764 return 0;
1765
1766 *min = is->min_val;
1767 *max = is->max_val;
1768
1769 n = (double) is->samples;
1770 *mean = (double) is->val / n;
1771 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1772 if (!(*min + *max) && !(*mean + *dev))
1773 return 0;
1774
1775 return 1;
1776}
1777
1778static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1779 int ddir)
1780{
1781 char *ddir_str[] = { "read ", "write" };
1782 unsigned long min, max, bw;
1783 double mean, dev;
1784
1785 if (!td->runtime[ddir])
1786 return;
1787
1788 bw = td->io_bytes[ddir] / td->runtime[ddir];
1789 printf(" %s: io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
1790
1791 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1792 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1793
1794 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1795 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1796
1797 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1798 double p_of_agg;
1799
1800 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1801 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);
1802 }
1803}
1804
1805static void show_thread_status(struct thread_data *td,
1806 struct group_run_stats *rs)
1807{
1808 double usr_cpu, sys_cpu;
1809
1810 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1811 return;
1812
1813 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1814
1815 show_ddir_status(td, rs, td->ddir);
1816 show_ddir_status(td, rs, td->ddir ^ 1);
1817
1818 if (td->runtime[0] + td->runtime[1]) {
1819 double runt = td->runtime[0] + td->runtime[1];
1820
1821 usr_cpu = (double) td->usr_time * 100 / runt;
1822 sys_cpu = (double) td->sys_time * 100 / runt;
1823 } else {
1824 usr_cpu = 0;
1825 sys_cpu = 0;
1826 }
1827
1828 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
1829}
1830
1831static void check_str_update(struct thread_data *td)
1832{
1833 char c = run_str[td->thread_number - 1];
1834
1835 if (td->runstate == td->old_runstate)
1836 return;
1837
1838 switch (td->runstate) {
1839 case TD_REAPED:
1840 c = '_';
1841 break;
1842 case TD_EXITED:
1843 c = 'E';
1844 break;
1845 case TD_RUNNING:
1846 if (td_read(td)) {
1847 if (td->sequential)
1848 c = 'R';
1849 else
1850 c = 'r';
1851 } else {
1852 if (td->sequential)
1853 c = 'W';
1854 else
1855 c = 'w';
1856 }
1857 break;
1858 case TD_VERIFYING:
1859 c = 'V';
1860 break;
1861 case TD_CREATED:
1862 c = 'C';
1863 break;
1864 case TD_NOT_CREATED:
1865 c = 'P';
1866 break;
1867 default:
1868 printf("state %d\n", td->runstate);
1869 }
1870
1871 run_str[td->thread_number - 1] = c;
1872 td->old_runstate = td->runstate;
1873}
1874
5289b847
JA		 1875static void eta_to_str(char *str, int eta_sec)
1876{
1877 unsigned int d, h, m, s;
1878 static int always_d, always_h;
1879
1880 d = h = m = s = 0;
1881
1882 s = eta_sec % 60;
1883 eta_sec /= 60;
1884 m = eta_sec % 60;
1885 eta_sec /= 60;
1886 h = eta_sec % 24;
1887 eta_sec /= 24;
1888 d = eta_sec;
1889
1890 if (d || always_d) {
1891 always_d = 1;
1892 str += sprintf(str, "%02dd:", d);
1893 }
1894 if (h || always_h) {
1895 always_h = 1;
1896 str += sprintf(str, "%02dh:", h);
1897 }
1898
1899 str += sprintf(str, "%02dm:", m);
1900 str += sprintf(str, "%02ds", s);
1901}
1902
ebac4655
JA		 1903static void print_thread_status(void)
1904{
1905 unsigned long long bytes_done, bytes_total;
5289b847
JA		 1906 int i, nr_running, t_rate, m_rate, eta_sec;
1907 char eta_str[32];
ebac4655
JA		 1908 double perc;
1909
1910 bytes_done = bytes_total = 0;
1911 nr_running = t_rate = m_rate = 0;
1912 for (i = 0; i < thread_number; i++) {
1913 struct thread_data *td = &threads[i];
1914
1915 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
1916 nr_running++;
1917 t_rate += td->rate;
1918 m_rate += td->ratemin;
1919 }
1920
1921 bytes_total += td->total_io_size;
1922 if (td->verify)
1923 bytes_total += td->total_io_size;
1924
1925 bytes_done += td->io_bytes[DDIR_READ] +td->io_bytes[DDIR_WRITE];
1926
1927 check_str_update(td);
1928 }
1929
1930 perc = 0;
5289b847 1931 eta_sec = 0;
ebac4655 1932 if (bytes_total && bytes_done) {
5289b847
JA		 1933 unsigned long runtime;
1934
1935 perc = (double) bytes_done / (double) bytes_total;
1936 if (perc > 1.0)
1937 perc = 1.0;
1938
1939 runtime = time_since_now(&genesis);
1940 if (runtime >= 5) {
1941 memset(eta_str, 0, sizeof(eta_str));
1942 eta_sec = (runtime * (1.0 / perc)) - runtime;
1943 eta_to_str(eta_str, eta_sec);
1944 }
1945
1946 perc *= 100.0;
ebac4655
JA		 1947 }
1948
5289b847 1949 printf("Threads now running (%d)", nr_running);
ebac4655
JA		 1950 if (m_rate || t_rate)
1951 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
5289b847
JA		 1952 printf(": [%s] [%3.2f%% done]", run_str, perc);
1953 if (eta_sec)
1954 printf(" [eta %s]", eta_str);
1955 printf("\r");
ebac4655
JA		 1956 fflush(stdout);
1957}
1958
1959static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1960{
1961 int i;
1962
1963 /*
1964 * reap exited threads (TD_EXITED -> TD_REAPED)
1965 */
1966 for (i = 0; i < thread_number; i++) {
1967 struct thread_data *td = &threads[i];
1968
1969 if (td->runstate != TD_EXITED)
1970 continue;
1971
1972 td_set_runstate(td, TD_REAPED);
1973
1974 if (td->use_thread) {
1975 long ret;
1976
1977 if (pthread_join(td->thread, (void *) &ret))
1978 perror("thread_join");
1979 } else
1980 waitpid(td->pid, NULL, 0);
1981
1982 (*nr_running)--;
1983 (*m_rate) -= td->ratemin;
1984 (*t_rate) -= td->rate;
1985 }
1986}
1987
1988static void run_threads(void)
1989{
ebac4655
JA		 1990 struct thread_data *td;
1991 unsigned long spent;
1992 int i, todo, nr_running, m_rate, t_rate, nr_started;
1993
1994 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
1995 fflush(stdout);
1996
1997 signal(SIGINT, sig_handler);
1998 signal(SIGALRM, sig_handler);
1999
2000 todo = thread_number;
2001 nr_running = 0;
2002 nr_started = 0;
2003 m_rate = t_rate = 0;
2004
2005 for (i = 0; i < thread_number; i++) {
2006 td = &threads[i];
2007
2008 run_str[td->thread_number - 1] = 'P';
2009
2010 init_disk_util(td);
2011
2012 if (!td->create_serialize)
2013 continue;
2014
2015 /*
2016 * do file setup here so it happens sequentially,
2017 * we don't want X number of threads getting their
2018 * client data interspersed on disk
2019 */
2020 if (setup_file(td)) {
2021 td_set_runstate(td, TD_REAPED);
2022 todo--;
2023 }
2024 }
2025
2026 gettimeofday(&genesis, NULL);
2027
2028 while (todo) {
2029 /*
2030 * create threads (TD_NOT_CREATED -> TD_CREATED)
2031 */
2032 for (i = 0; i < thread_number; i++) {
2033 td = &threads[i];
2034
2035 if (td->runstate != TD_NOT_CREATED)
2036 continue;
2037
2038 /*
2039 * never got a chance to start, killed by other
2040 * thread for some reason
2041 */
2042 if (td->terminate) {
2043 todo--;
2044 continue;
2045 }
2046
2047 if (td->start_delay) {
2048 spent = mtime_since_now(&genesis);
2049
2050 if (td->start_delay * 1000 > spent)
2051 continue;
2052 }
2053
2054 if (td->stonewall && (nr_started || nr_running))
2055 break;
2056
2057 td_set_runstate(td, TD_CREATED);
2058 sem_init(&startup_sem, 0, 1);
2059 todo--;
2060 nr_started++;
2061
2062 if (td->use_thread) {
2063 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2064 perror("thread_create");
2065 nr_started--;
2066 }
2067 } else {
2068 if (fork())
2069 sem_wait(&startup_sem);
2070 else {
2071 fork_main(shm_id, i);
2072 exit(0);
2073 }
2074 }
2075 }
2076
2077 /*
2078 * start created threads (TD_CREATED -> TD_RUNNING)
2079 */
2080 for (i = 0; i < thread_number; i++) {
2081 td = &threads[i];
2082
2083 if (td->runstate != TD_CREATED)
2084 continue;
2085
2086 td_set_runstate(td, TD_RUNNING);
2087 nr_running++;
2088 nr_started--;
2089 m_rate += td->ratemin;
2090 t_rate += td->rate;
2091 sem_post(&td->mutex);
2092 }
2093
2094 reap_threads(&nr_running, &t_rate, &m_rate);
2095
2096 if (todo)
2097 usleep(100000);
2098 }
2099
2100 while (nr_running) {
2101 reap_threads(&nr_running, &t_rate, &m_rate);
2102 usleep(10000);
2103 }
2104
2105 update_io_ticks();
2106}
2107
2108static void show_group_stats(struct group_run_stats *rs, int id)
2109{
2110 printf("\nRun status group %d (all jobs):\n", id);
2111
2112 if (rs->max_run[DDIR_READ])
2113 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]);
2114 if (rs->max_run[DDIR_WRITE])
2115 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]);
2116}
2117
2118static void show_disk_util(void)
2119{
2120 struct disk_util_stat *dus;
2121 struct list_head *entry;
2122 struct disk_util *du;
2123 double util;
2124
2125 printf("\nDisk stats (read/write):\n");
2126
2127 list_for_each(entry, &disk_list) {
2128 du = list_entry(entry, struct disk_util, list);
2129 dus = &du->dus;
2130
2131 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2132 if (util > 100.0)
2133 util = 100.0;
2134
2135 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);
2136 }
2137}
2138
2139static void show_run_stats(void)
2140{
2141 struct group_run_stats *runstats, *rs;
2142 struct thread_data *td;
2143 int i;
2144
2145 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2146
2147 for (i = 0; i < groupid + 1; i++) {
2148 rs = &runstats[i];
2149
2150 memset(rs, 0, sizeof(*rs));
2151 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2152 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2153 }
2154
2155 for (i = 0; i < thread_number; i++) {
2156 unsigned long rbw, wbw;
2157
2158 td = &threads[i];
2159
2160 if (td->error) {
2161 printf("Client%d: %s\n", td->thread_number, td->verror);
2162 continue;
2163 }
2164
2165 rs = &runstats[td->groupid];
2166
2167 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2168 rs->min_run[0] = td->runtime[0];
2169 if (td->runtime[0] > rs->max_run[0])
2170 rs->max_run[0] = td->runtime[0];
2171 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2172 rs->min_run[1] = td->runtime[1];
2173 if (td->runtime[1] > rs->max_run[1])
2174 rs->max_run[1] = td->runtime[1];
2175
2176 rbw = wbw = 0;
2177 if (td->runtime[0])
2178 rbw = td->io_bytes[0] / td->runtime[0];
2179 if (td->runtime[1])
2180 wbw = td->io_bytes[1] / td->runtime[1];
2181
2182 if (rbw < rs->min_bw[0])
2183 rs->min_bw[0] = rbw;
2184 if (wbw < rs->min_bw[1])
2185 rs->min_bw[1] = wbw;
2186 if (rbw > rs->max_bw[0])
2187 rs->max_bw[0] = rbw;
2188 if (wbw > rs->max_bw[1])
2189 rs->max_bw[1] = wbw;
2190
2191 rs->io_mb[0] += td->io_bytes[0] >> 20;
2192 rs->io_mb[1] += td->io_bytes[1] >> 20;
2193 }
2194
2195 for (i = 0; i < groupid + 1; i++) {
2196 rs = &runstats[i];
2197
2198 if (rs->max_run[0])
2199 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2200 if (rs->max_run[1])
2201 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2202 }
2203
2204 /*
2205 * don't overwrite last signal output
2206 */
2207 printf("\n");
2208
2209 for (i = 0; i < thread_number; i++) {
2210 td = &threads[i];
2211 rs = &runstats[td->groupid];
2212
2213 show_thread_status(td, rs);
2214 }
2215
2216 for (i = 0; i < groupid + 1; i++)
2217 show_group_stats(&runstats[i], i);
2218
2219 show_disk_util();
2220}
2221
2222int main(int argc, char *argv[])
2223{
2224 if (parse_options(argc, argv))
2225 return 1;
2226
2227 if (!thread_number) {
2228 printf("Nothing to do\n");
2229 return 1;
2230 }
2231
2232 disk_util_timer_arm();
2233
2234 run_threads();
2235 show_run_stats();
2236
2237 return 0;
2238}
fio - Flexible IO Tester