projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
[PATCH] cleanup sg ioctl vs read approach
[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
ebac4655
JA		 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
20dc95c4
JA		 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;
838a3cd3 305 if (*offset > td->real_file_size)
20dc95c4
JA		 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
ebac4655
JA		 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
20dc95c4
JA		 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
b1ff3403 582void put_io_u(struct thread_data *td, struct io_u *io_u)
ebac4655
JA		 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
b1ff3403 591struct io_u *__get_io_u(struct thread_data *td)
ebac4655
JA		 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
ebac4655
JA		 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
20dc95c4
JA		 615 if (td->zone_bytes >= td->zone_size) {
616 td->zone_bytes = 0;
617 td->last_pos += td->zone_skip;
618 }
619
ebac4655
JA		 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
838a3cd3
JA		 631 if (io_u->buflen + io_u->offset > td->real_file_size)
632 io_u->buflen = td->real_file_size - io_u->offset;
ebac4655
JA		 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
20dc95c4 642 td->last_pos += io_u->buflen;
ebac4655
JA		 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) {
20dc95c4 759 unsigned int bytes = io_u->buflen - io_u->resid;
ebac4655
JA		 760 int idx = io_u->ddir;
761
762 td->io_blocks[idx]++;
20dc95c4
JA		 763 td->io_bytes[idx] += bytes;
764 td->zone_bytes += bytes;
765 td->this_io_bytes[idx] += bytes;
ebac4655
JA		 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
20dc95c4 775 icd->bytes_done[idx] += bytes;
ebac4655
JA		 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
ebac4655
JA		 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;
b1ff3403 1116 io_u->index = i;
ebac4655
JA		 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
838a3cd3
JA		 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;
ebac4655 1211
e8e387c1 1212 td->file_size -= td->file_offset;
ebac4655
JA		 1213 return 0;
1214}
1215
1216static int bdev_size(struct thread_data *td)
1217{
9104f874 1218 unsigned long long bytes;
ebac4655
JA		 1219 int r;
1220
1221 r = blockdev_size(td->fd, &bytes);
1222 if (r) {
1223 td_verror(td, r);
1224 return 1;
1225 }
1226
838a3cd3
JA		 1227 td->real_file_size = bytes;
1228
ebac4655
JA		 1229 /*
1230 * no extend possibilities, so limit size to device size if too large
1231 */
838a3cd3
JA		 1232 if (!td->file_size || td->file_size > td->real_file_size)
1233 td->file_size = td->real_file_size;
ebac4655 1234
e8e387c1 1235 td->file_size -= td->file_offset;
ebac4655
JA		 1236 return 0;
1237}
1238
1239static int get_file_size(struct thread_data *td)
1240{
0af7b542 1241 int ret = 0;
ebac4655
JA		 1242
1243 if (td->filetype == FIO_TYPE_FILE)
1244 ret = file_size(td);
0af7b542 1245 else if (td->filetype == FIO_TYPE_BD)
ebac4655 1246 ret = bdev_size(td);
0af7b542
JA		 1247 else
1248 td->real_file_size = -1;
ebac4655
JA		 1249
1250 if (ret)
1251 return ret;
1252
e8e387c1
JA		 1253 if (td->file_offset > td->real_file_size) {
1254 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
ebac4655
JA		 1255 return 1;
1256 }
1257
838a3cd3 1258 td->io_size = td->file_size;
ebac4655
JA		 1259 if (td->io_size == 0) {
1260 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1261 td_verror(td, EINVAL);
1262 return 1;
1263 }
1264
20dc95c4
JA		 1265 if (!td->zone_size)
1266 td->zone_size = td->io_size;
1267
ebac4655
JA		 1268 td->total_io_size = td->io_size * td->loops;
1269 return 0;
1270}
1271
1272static int setup_file_mmap(struct thread_data *td)
1273{
1274 int flags;
1275
1276 if (td_read(td))
1277 flags = PROT_READ;
1278 else {
1279 flags = PROT_WRITE;
1280
1281 if (td->verify != VERIFY_NONE)
1282 flags |= PROT_READ;
1283 }
1284
1285 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1286 if (td->mmap == MAP_FAILED) {
1287 td->mmap = NULL;
1288 td_verror(td, errno);
1289 return 1;
1290 }
1291
1292 if (td->invalidate_cache) {
1293 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1294 td_verror(td, errno);
1295 return 1;
1296 }
1297 }
1298
1299 if (td->sequential) {
1300 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1301 td_verror(td, errno);
1302 return 1;
1303 }
1304 } else {
1305 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1306 td_verror(td, errno);
1307 return 1;
1308 }
1309 }
1310
1311 return 0;
1312}
1313
1314static int setup_file_plain(struct thread_data *td)
1315{
1316 if (td->invalidate_cache) {
1317 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1318 td_verror(td, errno);
1319 return 1;
1320 }
1321 }
1322
1323 if (td->sequential) {
1324 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1325 td_verror(td, errno);
1326 return 1;
1327 }
1328 } else {
1329 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1330 td_verror(td, errno);
1331 return 1;
1332 }
1333 }
1334
1335 return 0;
1336}
1337
1338static int setup_file(struct thread_data *td)
1339{
1340 struct stat st;
1341 int flags = 0;
1342
1343 if (stat(td->file_name, &st) == -1) {
1344 if (errno != ENOENT) {
1345 td_verror(td, errno);
1346 return 1;
1347 }
1348 if (!td->create_file) {
1349 td_verror(td, ENOENT);
1350 return 1;
1351 }
1352 if (create_file(td, td->file_size, 0))
1353 return 1;
1354 } else if (td->filetype == FIO_TYPE_FILE) {
6a0106a0 1355 if (st.st_size < (off_t) td->file_size) {
ebac4655
JA		 1356 if (create_file(td, td->file_size - st.st_size, 1))
1357 return 1;
1358 }
1359 }
1360
1361 if (td->odirect)
1362 flags |= O_DIRECT;
1363
0af7b542
JA		 1364 if (td_read(td)) {
1365 if (td->filetype == FIO_TYPE_CHAR)
1366 flags |= O_RDWR;
1367 else
1368 flags |= O_RDONLY;
1369
1370 td->fd = open(td->file_name, flags);
1371 } else {
ebac4655
JA		 1372 if (td->filetype == FIO_TYPE_FILE) {
1373 if (!td->overwrite)
1374 flags |= O_TRUNC;
1375
1376 flags |= O_CREAT;
1377 }
1378 if (td->sync_io)
1379 flags |= O_SYNC;
1380
1381 flags |= O_RDWR;
1382
1383 td->fd = open(td->file_name, flags, 0600);
1384 }
1385
1386 if (td->fd == -1) {
1387 td_verror(td, errno);
1388 return 1;
1389 }
1390
1391 if (get_file_size(td))
1392 return 1;
1393
1394 if (td->io_engine != FIO_MMAPIO)
1395 return setup_file_plain(td);
1396 else
1397 return setup_file_mmap(td);
1398}
1399
1400static int check_dev_match(dev_t dev, char *path)
1401{
1402 unsigned int major, minor;
1403 char line[256], *p;
1404 FILE *f;
1405
1406 f = fopen(path, "r");
1407 if (!f) {
1408 perror("open path");
1409 return 1;
1410 }
1411
1412 p = fgets(line, sizeof(line), f);
1413 if (!p) {
1414 fclose(f);
1415 return 1;
1416 }
1417
1418 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1419 fclose(f);
1420 return 1;
1421 }
1422
1423 if (((major << 8) | minor) == dev) {
1424 fclose(f);
1425 return 0;
1426 }
1427
1428 fclose(f);
1429 return 1;
1430}
1431
1432static int find_block_dir(dev_t dev, char *path)
1433{
1434 struct dirent *dir;
1435 struct stat st;
1436 int found = 0;
1437 DIR *D;
1438
1439 D = opendir(path);
1440 if (!D)
1441 return 0;
1442
1443 while ((dir = readdir(D)) != NULL) {
1444 char full_path[256];
1445
1446 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1447 continue;
1448 if (!strcmp(dir->d_name, "device"))
1449 continue;
1450
1451 sprintf(full_path, "%s/%s", path, dir->d_name);
1452
1453 if (!strcmp(dir->d_name, "dev")) {
1454 if (!check_dev_match(dev, full_path)) {
1455 found = 1;
1456 break;
1457 }
1458 }
1459
1460 if (stat(full_path, &st) == -1) {
1461 perror("stat");
1462 break;
1463 }
1464
1465 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1466 continue;
1467
1468 found = find_block_dir(dev, full_path);
1469 if (found) {
1470 strcpy(path, full_path);
1471 break;
1472 }
1473 }
1474
1475 closedir(D);
1476 return found;
1477}
1478
1479static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1480{
1481 unsigned in_flight;
1482 char line[256];
1483 FILE *f;
1484 char *p;
1485
1486 f = fopen(du->path, "r");
1487 if (!f)
1488 return 1;
1489
1490 p = fgets(line, sizeof(line), f);
1491 if (!p) {
1492 fclose(f);
1493 return 1;
1494 }
1495
1496 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) {
1497 fclose(f);
1498 return 1;
1499 }
1500
1501 fclose(f);
1502 return 0;
1503}
1504
1505static void update_io_tick_disk(struct disk_util *du)
1506{
1507 struct disk_util_stat __dus, *dus, *ldus;
1508 struct timeval t;
1509
1510 if (get_io_ticks(du, &__dus))
1511 return;
1512
1513 dus = &du->dus;
1514 ldus = &du->last_dus;
1515
1516 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1517 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1518 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1519 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1520 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1521 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1522 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1523 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1524 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1525 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1526
1527 gettimeofday(&t, NULL);
1528 du->msec += mtime_since(&du->time, &t);
1529 memcpy(&du->time, &t, sizeof(t));
1530 memcpy(ldus, &__dus, sizeof(__dus));
1531}
1532
1533static void update_io_ticks(void)
1534{
1535 struct list_head *entry;
1536 struct disk_util *du;
1537
1538 list_for_each(entry, &disk_list) {
1539 du = list_entry(entry, struct disk_util, list);
1540 update_io_tick_disk(du);
1541 }
1542}
1543
1544static int disk_util_exists(dev_t dev)
1545{
1546 struct list_head *entry;
1547 struct disk_util *du;
1548
1549 list_for_each(entry, &disk_list) {
1550 du = list_entry(entry, struct disk_util, list);
1551
1552 if (du->dev == dev)
1553 return 1;
1554 }
1555
1556 return 0;
1557}
1558
1559static void disk_util_add(dev_t dev, char *path)
1560{
1561 struct disk_util *du = malloc(sizeof(*du));
1562
1563 memset(du, 0, sizeof(*du));
1564 INIT_LIST_HEAD(&du->list);
1565 sprintf(du->path, "%s/stat", path);
1566 du->name = strdup(basename(path));
1567 du->dev = dev;
1568
1569 gettimeofday(&du->time, NULL);
1570 get_io_ticks(du, &du->last_dus);
1571
1572 list_add_tail(&du->list, &disk_list);
1573}
1574
1575static void init_disk_util(struct thread_data *td)
1576{
1577 struct stat st;
1578 char foo[256], tmp[256];
1579 dev_t dev;
1580 char *p;
1581
1582 if (!td->do_disk_util)
1583 return;
1584
1585 if (!stat(td->file_name, &st)) {
1586 if (S_ISBLK(st.st_mode))
1587 dev = st.st_rdev;
1588 else
1589 dev = st.st_dev;
1590 } else {
1591 /*
1592 * must be a file, open "." in that path
1593 */
1594 strcpy(foo, td->file_name);
1595 p = dirname(foo);
1596 if (stat(p, &st)) {
1597 perror("disk util stat");
1598 return;
1599 }
1600
1601 dev = st.st_dev;
1602 }
1603
1604 if (disk_util_exists(dev))
1605 return;
1606
1607 sprintf(foo, "/sys/block");
1608 if (!find_block_dir(dev, foo))
1609 return;
1610
1611 /*
64cb3e10 1612 * for md/dm, there's no queue dir. we already have the right place
ebac4655 1613 */
64cb3e10 1614 sprintf(tmp, "%s/stat", foo);
ebac4655 1615 if (stat(tmp, &st)) {
64cb3e10
JA		 1616 /*
1617 * if this is inside a partition dir, jump back to parent
1618 */
1619 sprintf(tmp, "%s/queue", foo);
ebac4655 1620 if (stat(tmp, &st)) {
64cb3e10
JA		 1621 p = dirname(foo);
1622 sprintf(tmp, "%s/queue", p);
1623 if (stat(tmp, &st)) {
1624 fprintf(stderr, "unknown sysfs layout\n");
1625 return;
1626 }
1627 sprintf(foo, "%s", p);
ebac4655 1628 }
ebac4655
JA		 1629 }
1630
1631 disk_util_add(dev, foo);
1632}
1633
1634static void disk_util_timer_arm(void)
1635{
1636 itimer.it_value.tv_sec = 0;
1637 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1638 setitimer(ITIMER_REAL, &itimer, NULL);
1639}
1640
1641static void clear_io_state(struct thread_data *td)
1642{
1643 if (td->io_engine == FIO_SYNCIO)
1644 lseek(td->fd, SEEK_SET, 0);
1645
20dc95c4 1646 td->last_pos = 0;
ebac4655
JA		 1647 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1648 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 1649 td->zone_bytes = 0;
ebac4655
JA		 1650
1651 if (td->file_map)
1652 memset(td->file_map, 0, td->num_maps * sizeof(long));
1653}
1654
1655static void update_rusage_stat(struct thread_data *td)
1656{
1657 if (!(td->runtime[0] + td->runtime[1]))
1658 return;
1659
1660 getrusage(RUSAGE_SELF, &td->ru_end);
1661
1662 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1663 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1664 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1665
1666
1667 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1668}
1669
1670static void *thread_main(void *data)
1671{
1672 struct thread_data *td = data;
1673 int ret = 1;
1674
1675 if (!td->use_thread)
1676 setsid();
1677
1678 td->pid = getpid();
1679
1680 if (init_io_u(td))
1681 goto err;
1682
1683 if (fio_setaffinity(td) == -1) {
1684 td_verror(td, errno);
1685 goto err;
1686 }
1687
1688 if (init_io(td))
1689 goto err;
1690
1691 if (td->ioprio) {
1692 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1693 td_verror(td, errno);
1694 goto err;
1695 }
1696 }
1697
1698 sem_post(&startup_sem);
1699 sem_wait(&td->mutex);
1700
1701 if (!td->create_serialize && setup_file(td))
1702 goto err;
1703
1704 if (init_random_state(td))
1705 goto err;
1706
1707 gettimeofday(&td->epoch, NULL);
1708
1709 while (td->loops--) {
1710 getrusage(RUSAGE_SELF, &td->ru_start);
1711 gettimeofday(&td->start, NULL);
1712 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1713
1714 if (td->ratemin)
1715 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1716
1717 clear_io_state(td);
1718 prune_io_piece_log(td);
1719
1720 do_io(td);
1721
1722 td->runtime[td->ddir] += mtime_since_now(&td->start);
1723 update_rusage_stat(td);
1724
1725 if (td->error || td->terminate)
1726 break;
1727
1728 if (td->verify == VERIFY_NONE)
1729 continue;
1730
1731 clear_io_state(td);
1732 gettimeofday(&td->start, NULL);
1733
1734 do_verify(td);
1735
1736 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1737
1738 if (td->error || td->terminate)
1739 break;
1740 }
1741
1742 ret = 0;
1743
1744 if (td->bw_log)
1745 finish_log(td, td->bw_log, "bw");
1746 if (td->slat_log)
1747 finish_log(td, td->slat_log, "slat");
1748 if (td->clat_log)
1749 finish_log(td, td->clat_log, "clat");
1750
1751 if (exitall_on_terminate)
1752 terminate_threads(td->groupid);
1753
1754err:
1755 if (td->fd != -1) {
1756 close(td->fd);
1757 td->fd = -1;
1758 }
1759 if (td->mmap)
1760 munmap(td->mmap, td->file_size);
1761 cleanup_io(td);
1762 cleanup_io_u(td);
1763 if (ret) {
1764 sem_post(&startup_sem);
1765 sem_wait(&td->mutex);
1766 }
1767 td_set_runstate(td, TD_EXITED);
1768 return NULL;
1769
1770}
1771
1772static void *fork_main(int shmid, int offset)
1773{
1774 struct thread_data *td;
1775 void *data;
1776
1777 data = shmat(shmid, NULL, 0);
1778 if (data == (void *) -1) {
1779 perror("shmat");
1780 return NULL;
1781 }
1782
1783 td = data + offset * sizeof(struct thread_data);
1784 thread_main(td);
1785 shmdt(data);
1786 return NULL;
1787}
1788
1789static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1790 double *mean, double *dev)
1791{
1792 double n;
1793
1794 if (is->samples == 0)
1795 return 0;
1796
1797 *min = is->min_val;
1798 *max = is->max_val;
1799
1800 n = (double) is->samples;
1801 *mean = (double) is->val / n;
1802 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1803 if (!(*min + *max) && !(*mean + *dev))
1804 return 0;
1805
1806 return 1;
1807}
1808
1809static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1810 int ddir)
1811{
1812 char *ddir_str[] = { "read ", "write" };
1813 unsigned long min, max, bw;
1814 double mean, dev;
1815
1816 if (!td->runtime[ddir])
1817 return;
1818
1819 bw = td->io_bytes[ddir] / td->runtime[ddir];
9104f874 1820 printf(" %s: io=%6lluMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
ebac4655
JA		 1821
1822 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1823 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1824
1825 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1826 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1827
1828 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1829 double p_of_agg;
1830
1831 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1832 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);
1833 }
1834}
1835
1836static void show_thread_status(struct thread_data *td,
1837 struct group_run_stats *rs)
1838{
1839 double usr_cpu, sys_cpu;
1840
1841 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1842 return;
1843
1844 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1845
1846 show_ddir_status(td, rs, td->ddir);
1847 show_ddir_status(td, rs, td->ddir ^ 1);
1848
1849 if (td->runtime[0] + td->runtime[1]) {
1850 double runt = td->runtime[0] + td->runtime[1];
1851
1852 usr_cpu = (double) td->usr_time * 100 / runt;
1853 sys_cpu = (double) td->sys_time * 100 / runt;
1854 } else {
1855 usr_cpu = 0;
1856 sys_cpu = 0;
1857 }
1858
1859 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
1860}
1861
1862static void check_str_update(struct thread_data *td)
1863{
1864 char c = run_str[td->thread_number - 1];
1865
1866 if (td->runstate == td->old_runstate)
1867 return;
1868
1869 switch (td->runstate) {
1870 case TD_REAPED:
1871 c = '_';
1872 break;
1873 case TD_EXITED:
1874 c = 'E';
1875 break;
1876 case TD_RUNNING:
1877 if (td_read(td)) {
1878 if (td->sequential)
1879 c = 'R';
1880 else
1881 c = 'r';
1882 } else {
1883 if (td->sequential)
1884 c = 'W';
1885 else
1886 c = 'w';
1887 }
1888 break;
1889 case TD_VERIFYING:
1890 c = 'V';
1891 break;
1892 case TD_CREATED:
1893 c = 'C';
1894 break;
1895 case TD_NOT_CREATED:
1896 c = 'P';
1897 break;
1898 default:
1899 printf("state %d\n", td->runstate);
1900 }
1901
1902 run_str[td->thread_number - 1] = c;
1903 td->old_runstate = td->runstate;
1904}
1905
5289b847
JA		 1906static void eta_to_str(char *str, int eta_sec)
1907{
1908 unsigned int d, h, m, s;
1909 static int always_d, always_h;
1910
1911 d = h = m = s = 0;
1912
1913 s = eta_sec % 60;
1914 eta_sec /= 60;
1915 m = eta_sec % 60;
1916 eta_sec /= 60;
1917 h = eta_sec % 24;
1918 eta_sec /= 24;
1919 d = eta_sec;
1920
1921 if (d || always_d) {
1922 always_d = 1;
1923 str += sprintf(str, "%02dd:", d);
1924 }
1925 if (h || always_h) {
1926 always_h = 1;
1927 str += sprintf(str, "%02dh:", h);
1928 }
1929
1930 str += sprintf(str, "%02dm:", m);
1931 str += sprintf(str, "%02ds", s);
1932}
1933
6a0106a0
JA		 1934static int thread_eta(struct thread_data *td, unsigned long elapsed)
1935{
1936 unsigned long long bytes_total, bytes_done;
1937 unsigned int eta_sec = 0;
1938
1939 bytes_total = td->total_io_size;
1940 if (td->verify)
1941 bytes_total <<= 1;
1942 if (td->zone_size && td->zone_skip)
1943 bytes_total /= (td->zone_skip / td->zone_size);
1944
1945 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
1946 double perc;
1947 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
1948 perc = (double) bytes_done / (double) bytes_total;
1949 if (perc > 1.0)
1950 perc = 1.0;
1951
1952 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
1953
1954 if (eta_sec > (td->timeout - elapsed))
1955 eta_sec = td->timeout - elapsed;
1956 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED) {
1957 int t_eta = 0, r_eta = 0;
1958
1959 /*
1960 * We can only guess - assume it'll run the full timeout
1961 * if given, otherwise assume it'll run at the specified rate.
1962 */
1963 if (td->timeout)
1964 t_eta = td->timeout + td->start_delay - elapsed;
1965 if (td->rate) {
1966 r_eta = (bytes_total / 1024) / td->rate;
1967 r_eta += td->start_delay - elapsed;
1968 }
1969
1970 if (r_eta && t_eta)
1971 eta_sec = min(r_eta, t_eta);
1972 else if (r_eta)
1973 eta_sec = r_eta;
1974 else if (t_eta)
1975 eta_sec = t_eta;
1976 else
1977 eta_sec = INT_MAX;
1978 } else {
1979 /*
1980 * thread is already done
1981 */
1982 eta_sec = 0;
1983 }
1984
1985 return eta_sec;
1986}
1987
ebac4655
JA		 1988static void print_thread_status(void)
1989{
6a0106a0
JA		 1990 unsigned long elapsed = time_since_now(&genesis);
1991 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
5289b847 1992 char eta_str[32];
6a0106a0
JA		 1993 double perc = 0.0;
1994
1995 eta_secs = malloc(thread_number * sizeof(int));
1996 memset(eta_secs, 0, thread_number * sizeof(int));
ebac4655 1997
ebac4655
JA		 1998 nr_running = t_rate = m_rate = 0;
1999 for (i = 0; i < thread_number; i++) {
2000 struct thread_data *td = &threads[i];
2001
2002 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
2003 nr_running++;
2004 t_rate += td->rate;
2005 m_rate += td->ratemin;
2006 }
2007
6a0106a0
JA		 2008 if (elapsed >= 3)
2009 eta_secs[i] = thread_eta(td, elapsed);
ebac4655
JA		 2010
2011 check_str_update(td);
2012 }
2013
6a0106a0
JA		 2014 if (exitall_on_terminate)
2015 eta_sec = INT_MAX;
2016 else
2017 eta_sec = 0;
5289b847 2018
6a0106a0
JA		 2019 for (i = 0; i < thread_number; i++) {
2020 if (exitall_on_terminate) {
2021 if (eta_secs[i] < eta_sec)
2022 eta_sec = eta_secs[i];
2023 } else {
2024 if (eta_secs[i] > eta_sec)
2025 eta_sec = eta_secs[i];
5289b847 2026 }
6a0106a0 2027 }
5289b847 2028
6a0106a0
JA		 2029 if (eta_sec == INT_MAX)
2030 eta_sec = 0;
2031
2032 if (eta_sec) {
2033 perc = (double) elapsed / (double) (elapsed + eta_sec);
2034 eta_to_str(eta_str, eta_sec);
ebac4655
JA		 2035 }
2036
5289b847 2037 printf("Threads now running (%d)", nr_running);
ebac4655
JA		 2038 if (m_rate || t_rate)
2039 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
6a0106a0
JA		 2040 if (eta_sec) {
2041 perc *= 100.0;
2042 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
2043 }
5289b847 2044 printf("\r");
ebac4655 2045 fflush(stdout);
6a0106a0 2046 free(eta_secs);
ebac4655
JA		 2047}
2048
2049static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2050{
2051 int i;
2052
2053 /*
2054 * reap exited threads (TD_EXITED -> TD_REAPED)
2055 */
2056 for (i = 0; i < thread_number; i++) {
2057 struct thread_data *td = &threads[i];
2058
2059 if (td->runstate != TD_EXITED)
2060 continue;
2061
2062 td_set_runstate(td, TD_REAPED);
2063
2064 if (td->use_thread) {
2065 long ret;
2066
2067 if (pthread_join(td->thread, (void *) &ret))
2068 perror("thread_join");
2069 } else
2070 waitpid(td->pid, NULL, 0);
2071
2072 (*nr_running)--;
2073 (*m_rate) -= td->ratemin;
2074 (*t_rate) -= td->rate;
2075 }
2076}
2077
2078static void run_threads(void)
2079{
ebac4655
JA		 2080 struct thread_data *td;
2081 unsigned long spent;
2082 int i, todo, nr_running, m_rate, t_rate, nr_started;
2083
2084 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
2085 fflush(stdout);
2086
2087 signal(SIGINT, sig_handler);
2088 signal(SIGALRM, sig_handler);
2089
2090 todo = thread_number;
2091 nr_running = 0;
2092 nr_started = 0;
2093 m_rate = t_rate = 0;
2094
2095 for (i = 0; i < thread_number; i++) {
2096 td = &threads[i];
2097
2098 run_str[td->thread_number - 1] = 'P';
2099
2100 init_disk_util(td);
2101
2102 if (!td->create_serialize)
2103 continue;
2104
2105 /*
2106 * do file setup here so it happens sequentially,
2107 * we don't want X number of threads getting their
2108 * client data interspersed on disk
2109 */
2110 if (setup_file(td)) {
2111 td_set_runstate(td, TD_REAPED);
2112 todo--;
2113 }
2114 }
2115
2116 gettimeofday(&genesis, NULL);
2117
2118 while (todo) {
2119 /*
2120 * create threads (TD_NOT_CREATED -> TD_CREATED)
2121 */
2122 for (i = 0; i < thread_number; i++) {
2123 td = &threads[i];
2124
2125 if (td->runstate != TD_NOT_CREATED)
2126 continue;
2127
2128 /*
2129 * never got a chance to start, killed by other
2130 * thread for some reason
2131 */
2132 if (td->terminate) {
2133 todo--;
2134 continue;
2135 }
2136
2137 if (td->start_delay) {
2138 spent = mtime_since_now(&genesis);
2139
2140 if (td->start_delay * 1000 > spent)
2141 continue;
2142 }
2143
2144 if (td->stonewall && (nr_started || nr_running))
2145 break;
2146
2147 td_set_runstate(td, TD_CREATED);
2148 sem_init(&startup_sem, 0, 1);
2149 todo--;
2150 nr_started++;
2151
2152 if (td->use_thread) {
2153 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2154 perror("thread_create");
2155 nr_started--;
2156 }
2157 } else {
2158 if (fork())
2159 sem_wait(&startup_sem);
2160 else {
2161 fork_main(shm_id, i);
2162 exit(0);
2163 }
2164 }
2165 }
2166
2167 /*
2168 * start created threads (TD_CREATED -> TD_RUNNING)
2169 */
2170 for (i = 0; i < thread_number; i++) {
2171 td = &threads[i];
2172
2173 if (td->runstate != TD_CREATED)
2174 continue;
2175
2176 td_set_runstate(td, TD_RUNNING);
2177 nr_running++;
2178 nr_started--;
2179 m_rate += td->ratemin;
2180 t_rate += td->rate;
2181 sem_post(&td->mutex);
2182 }
2183
2184 reap_threads(&nr_running, &t_rate, &m_rate);
2185
2186 if (todo)
2187 usleep(100000);
2188 }
2189
2190 while (nr_running) {
2191 reap_threads(&nr_running, &t_rate, &m_rate);
2192 usleep(10000);
2193 }
2194
2195 update_io_ticks();
2196}
2197
2198static void show_group_stats(struct group_run_stats *rs, int id)
2199{
2200 printf("\nRun status group %d (all jobs):\n", id);
2201
2202 if (rs->max_run[DDIR_READ])
9104f874 2203 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]);
ebac4655 2204 if (rs->max_run[DDIR_WRITE])
9104f874 2205 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]);
ebac4655
JA		 2206}
2207
2208static void show_disk_util(void)
2209{
2210 struct disk_util_stat *dus;
2211 struct list_head *entry;
2212 struct disk_util *du;
2213 double util;
2214
2215 printf("\nDisk stats (read/write):\n");
2216
2217 list_for_each(entry, &disk_list) {
2218 du = list_entry(entry, struct disk_util, list);
2219 dus = &du->dus;
2220
2221 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2222 if (util > 100.0)
2223 util = 100.0;
2224
2225 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);
2226 }
2227}
2228
2229static void show_run_stats(void)
2230{
2231 struct group_run_stats *runstats, *rs;
2232 struct thread_data *td;
2233 int i;
2234
2235 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2236
2237 for (i = 0; i < groupid + 1; i++) {
2238 rs = &runstats[i];
2239
2240 memset(rs, 0, sizeof(*rs));
2241 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2242 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2243 }
2244
2245 for (i = 0; i < thread_number; i++) {
2246 unsigned long rbw, wbw;
2247
2248 td = &threads[i];
2249
2250 if (td->error) {
2251 printf("Client%d: %s\n", td->thread_number, td->verror);
2252 continue;
2253 }
2254
2255 rs = &runstats[td->groupid];
2256
2257 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2258 rs->min_run[0] = td->runtime[0];
2259 if (td->runtime[0] > rs->max_run[0])
2260 rs->max_run[0] = td->runtime[0];
2261 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2262 rs->min_run[1] = td->runtime[1];
2263 if (td->runtime[1] > rs->max_run[1])
2264 rs->max_run[1] = td->runtime[1];
2265
2266 rbw = wbw = 0;
2267 if (td->runtime[0])
2268 rbw = td->io_bytes[0] / td->runtime[0];
2269 if (td->runtime[1])
2270 wbw = td->io_bytes[1] / td->runtime[1];
2271
2272 if (rbw < rs->min_bw[0])
2273 rs->min_bw[0] = rbw;
2274 if (wbw < rs->min_bw[1])
2275 rs->min_bw[1] = wbw;
2276 if (rbw > rs->max_bw[0])
2277 rs->max_bw[0] = rbw;
2278 if (wbw > rs->max_bw[1])
2279 rs->max_bw[1] = wbw;
2280
2281 rs->io_mb[0] += td->io_bytes[0] >> 20;
2282 rs->io_mb[1] += td->io_bytes[1] >> 20;
2283 }
2284
2285 for (i = 0; i < groupid + 1; i++) {
2286 rs = &runstats[i];
2287
2288 if (rs->max_run[0])
2289 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2290 if (rs->max_run[1])
2291 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2292 }
2293
2294 /*
2295 * don't overwrite last signal output
2296 */
2297 printf("\n");
2298
2299 for (i = 0; i < thread_number; i++) {
2300 td = &threads[i];
2301 rs = &runstats[td->groupid];
2302
2303 show_thread_status(td, rs);
2304 }
2305
2306 for (i = 0; i < groupid + 1; i++)
2307 show_group_stats(&runstats[i], i);
2308
2309 show_disk_util();
2310}
2311
2312int main(int argc, char *argv[])
2313{
2314 if (parse_options(argc, argv))
2315 return 1;
2316
2317 if (!thread_number) {
2318 printf("Nothing to do\n");
2319 return 1;
2320 }
2321
2322 disk_util_timer_arm();
2323
2324 run_threads();
2325 show_run_stats();
2326
2327 return 0;
2328}
fio - Flexible IO Tester