projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
[PATCH] Woops, forgot to define FIO_HAVE_IOPRIO for Linux
[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
ebac4655
JA		 582static void put_io_u(struct thread_data *td, struct io_u *io_u)
583{
584 list_del(&io_u->list);
585 list_add(&io_u->list, &td->io_u_freelist);
586 td->cur_depth--;
587}
588
589#define queue_full(td) (list_empty(&(td)->io_u_freelist))
590
591static struct io_u *__get_io_u(struct thread_data *td)
592{
593 struct io_u *io_u;
594
595 if (queue_full(td))
596 return NULL;
597
598 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
599 io_u->error = 0;
600 io_u->resid = 0;
601 list_del(&io_u->list);
602 list_add(&io_u->list, &td->io_u_busylist);
603 td->cur_depth++;
604 return io_u;
605}
606
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;
1116 list_add(&io_u->list, &td->io_u_freelist);
1117 }
1118
1119 return 0;
1120}
1121
1122static int create_file(struct thread_data *td, unsigned long long size,
1123 int extend)
1124{
1125 unsigned long long left;
1126 unsigned int bs;
1127 int r, oflags;
1128 char *b;
1129
1130 /*
1131 * unless specifically asked for overwrite, let normal io extend it
1132 */
1133 if (td_write(td) && !td->overwrite)
1134 return 0;
1135
1136 if (!size) {
1137 fprintf(stderr, "Need size for create\n");
1138 td_verror(td, EINVAL);
1139 return 1;
1140 }
1141
1142 if (!extend) {
1143 oflags = O_CREAT | O_TRUNC;
1144 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1145 } else {
1146 oflags = O_APPEND;
1147 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1148 }
1149
1150 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1151 if (td->fd < 0) {
1152 td_verror(td, errno);
1153 return 1;
1154 }
1155
1156 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1157 td_verror(td, errno);
1158 return 1;
1159 }
1160
1161 td->io_size = td->file_size;
1162 b = malloc(td->max_bs);
1163 memset(b, 0, td->max_bs);
1164
1165 left = size;
1166 while (left && !td->terminate) {
1167 bs = td->max_bs;
1168 if (bs > left)
1169 bs = left;
1170
1171 r = write(td->fd, b, bs);
1172
1173 if (r == (int) bs) {
1174 left -= bs;
1175 continue;
1176 } else {
1177 if (r < 0)
1178 td_verror(td, errno);
1179 else
1180 td_verror(td, EIO);
1181
1182 break;
1183 }
1184 }
1185
1186 if (td->terminate)
1187 unlink(td->file_name);
1188 else if (td->create_fsync)
1189 fsync(td->fd);
1190
1191 close(td->fd);
1192 td->fd = -1;
1193 free(b);
1194 return 0;
1195}
1196
1197static int file_size(struct thread_data *td)
1198{
1199 struct stat st;
1200
1201 if (fstat(td->fd, &st) == -1) {
1202 td_verror(td, errno);
1203 return 1;
1204 }
1205
838a3cd3
JA		 1206 td->real_file_size = st.st_size;
1207
1208 if (!td->file_size || td->file_size > td->real_file_size)
1209 td->file_size = td->real_file_size;
ebac4655
JA		 1210
1211 return 0;
1212}
1213
1214static int bdev_size(struct thread_data *td)
1215{
9104f874 1216 unsigned long long bytes;
ebac4655
JA		 1217 int r;
1218
1219 r = blockdev_size(td->fd, &bytes);
1220 if (r) {
1221 td_verror(td, r);
1222 return 1;
1223 }
1224
838a3cd3
JA		 1225 td->real_file_size = bytes;
1226
ebac4655
JA		 1227 /*
1228 * no extend possibilities, so limit size to device size if too large
1229 */
838a3cd3
JA		 1230 if (!td->file_size || td->file_size > td->real_file_size)
1231 td->file_size = td->real_file_size;
ebac4655
JA		 1232
1233 return 0;
1234}
1235
1236static int get_file_size(struct thread_data *td)
1237{
1238 int ret;
1239
1240 if (td->filetype == FIO_TYPE_FILE)
1241 ret = file_size(td);
1242 else
1243 ret = bdev_size(td);
1244
1245 if (ret)
1246 return ret;
1247
838a3cd3
JA		 1248 if (td->file_offset + td->file_size > td->real_file_size) {
1249 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset + td->file_size, td->real_file_size);
ebac4655
JA		 1250 return 1;
1251 }
1252
838a3cd3 1253 td->io_size = td->file_size;
ebac4655
JA		 1254 if (td->io_size == 0) {
1255 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1256 td_verror(td, EINVAL);
1257 return 1;
1258 }
1259
20dc95c4
JA		 1260 if (!td->zone_size)
1261 td->zone_size = td->io_size;
1262
ebac4655
JA		 1263 td->total_io_size = td->io_size * td->loops;
1264 return 0;
1265}
1266
1267static int setup_file_mmap(struct thread_data *td)
1268{
1269 int flags;
1270
1271 if (td_read(td))
1272 flags = PROT_READ;
1273 else {
1274 flags = PROT_WRITE;
1275
1276 if (td->verify != VERIFY_NONE)
1277 flags |= PROT_READ;
1278 }
1279
1280 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1281 if (td->mmap == MAP_FAILED) {
1282 td->mmap = NULL;
1283 td_verror(td, errno);
1284 return 1;
1285 }
1286
1287 if (td->invalidate_cache) {
1288 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1289 td_verror(td, errno);
1290 return 1;
1291 }
1292 }
1293
1294 if (td->sequential) {
1295 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1296 td_verror(td, errno);
1297 return 1;
1298 }
1299 } else {
1300 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1301 td_verror(td, errno);
1302 return 1;
1303 }
1304 }
1305
1306 return 0;
1307}
1308
1309static int setup_file_plain(struct thread_data *td)
1310{
1311 if (td->invalidate_cache) {
1312 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1313 td_verror(td, errno);
1314 return 1;
1315 }
1316 }
1317
1318 if (td->sequential) {
1319 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1320 td_verror(td, errno);
1321 return 1;
1322 }
1323 } else {
1324 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1325 td_verror(td, errno);
1326 return 1;
1327 }
1328 }
1329
1330 return 0;
1331}
1332
1333static int setup_file(struct thread_data *td)
1334{
1335 struct stat st;
1336 int flags = 0;
1337
1338 if (stat(td->file_name, &st) == -1) {
1339 if (errno != ENOENT) {
1340 td_verror(td, errno);
1341 return 1;
1342 }
1343 if (!td->create_file) {
1344 td_verror(td, ENOENT);
1345 return 1;
1346 }
1347 if (create_file(td, td->file_size, 0))
1348 return 1;
1349 } else if (td->filetype == FIO_TYPE_FILE) {
1350 if (st.st_size < td->file_size) {
1351 if (create_file(td, td->file_size - st.st_size, 1))
1352 return 1;
1353 }
1354 }
1355
1356 if (td->odirect)
1357 flags |= O_DIRECT;
1358
1359 if (td_read(td))
1360 td->fd = open(td->file_name, flags | O_RDONLY);
1361 else {
1362 if (td->filetype == FIO_TYPE_FILE) {
1363 if (!td->overwrite)
1364 flags |= O_TRUNC;
1365
1366 flags |= O_CREAT;
1367 }
1368 if (td->sync_io)
1369 flags |= O_SYNC;
1370
1371 flags |= O_RDWR;
1372
1373 td->fd = open(td->file_name, flags, 0600);
1374 }
1375
1376 if (td->fd == -1) {
1377 td_verror(td, errno);
1378 return 1;
1379 }
1380
1381 if (get_file_size(td))
1382 return 1;
1383
1384 if (td->io_engine != FIO_MMAPIO)
1385 return setup_file_plain(td);
1386 else
1387 return setup_file_mmap(td);
1388}
1389
1390static int check_dev_match(dev_t dev, char *path)
1391{
1392 unsigned int major, minor;
1393 char line[256], *p;
1394 FILE *f;
1395
1396 f = fopen(path, "r");
1397 if (!f) {
1398 perror("open path");
1399 return 1;
1400 }
1401
1402 p = fgets(line, sizeof(line), f);
1403 if (!p) {
1404 fclose(f);
1405 return 1;
1406 }
1407
1408 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1409 fclose(f);
1410 return 1;
1411 }
1412
1413 if (((major << 8) | minor) == dev) {
1414 fclose(f);
1415 return 0;
1416 }
1417
1418 fclose(f);
1419 return 1;
1420}
1421
1422static int find_block_dir(dev_t dev, char *path)
1423{
1424 struct dirent *dir;
1425 struct stat st;
1426 int found = 0;
1427 DIR *D;
1428
1429 D = opendir(path);
1430 if (!D)
1431 return 0;
1432
1433 while ((dir = readdir(D)) != NULL) {
1434 char full_path[256];
1435
1436 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1437 continue;
1438 if (!strcmp(dir->d_name, "device"))
1439 continue;
1440
1441 sprintf(full_path, "%s/%s", path, dir->d_name);
1442
1443 if (!strcmp(dir->d_name, "dev")) {
1444 if (!check_dev_match(dev, full_path)) {
1445 found = 1;
1446 break;
1447 }
1448 }
1449
1450 if (stat(full_path, &st) == -1) {
1451 perror("stat");
1452 break;
1453 }
1454
1455 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1456 continue;
1457
1458 found = find_block_dir(dev, full_path);
1459 if (found) {
1460 strcpy(path, full_path);
1461 break;
1462 }
1463 }
1464
1465 closedir(D);
1466 return found;
1467}
1468
1469static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1470{
1471 unsigned in_flight;
1472 char line[256];
1473 FILE *f;
1474 char *p;
1475
1476 f = fopen(du->path, "r");
1477 if (!f)
1478 return 1;
1479
1480 p = fgets(line, sizeof(line), f);
1481 if (!p) {
1482 fclose(f);
1483 return 1;
1484 }
1485
1486 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) {
1487 fclose(f);
1488 return 1;
1489 }
1490
1491 fclose(f);
1492 return 0;
1493}
1494
1495static void update_io_tick_disk(struct disk_util *du)
1496{
1497 struct disk_util_stat __dus, *dus, *ldus;
1498 struct timeval t;
1499
1500 if (get_io_ticks(du, &__dus))
1501 return;
1502
1503 dus = &du->dus;
1504 ldus = &du->last_dus;
1505
1506 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1507 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1508 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1509 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1510 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1511 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1512 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1513 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1514 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1515 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1516
1517 gettimeofday(&t, NULL);
1518 du->msec += mtime_since(&du->time, &t);
1519 memcpy(&du->time, &t, sizeof(t));
1520 memcpy(ldus, &__dus, sizeof(__dus));
1521}
1522
1523static void update_io_ticks(void)
1524{
1525 struct list_head *entry;
1526 struct disk_util *du;
1527
1528 list_for_each(entry, &disk_list) {
1529 du = list_entry(entry, struct disk_util, list);
1530 update_io_tick_disk(du);
1531 }
1532}
1533
1534static int disk_util_exists(dev_t dev)
1535{
1536 struct list_head *entry;
1537 struct disk_util *du;
1538
1539 list_for_each(entry, &disk_list) {
1540 du = list_entry(entry, struct disk_util, list);
1541
1542 if (du->dev == dev)
1543 return 1;
1544 }
1545
1546 return 0;
1547}
1548
1549static void disk_util_add(dev_t dev, char *path)
1550{
1551 struct disk_util *du = malloc(sizeof(*du));
1552
1553 memset(du, 0, sizeof(*du));
1554 INIT_LIST_HEAD(&du->list);
1555 sprintf(du->path, "%s/stat", path);
1556 du->name = strdup(basename(path));
1557 du->dev = dev;
1558
1559 gettimeofday(&du->time, NULL);
1560 get_io_ticks(du, &du->last_dus);
1561
1562 list_add_tail(&du->list, &disk_list);
1563}
1564
1565static void init_disk_util(struct thread_data *td)
1566{
1567 struct stat st;
1568 char foo[256], tmp[256];
1569 dev_t dev;
1570 char *p;
1571
1572 if (!td->do_disk_util)
1573 return;
1574
1575 if (!stat(td->file_name, &st)) {
1576 if (S_ISBLK(st.st_mode))
1577 dev = st.st_rdev;
1578 else
1579 dev = st.st_dev;
1580 } else {
1581 /*
1582 * must be a file, open "." in that path
1583 */
1584 strcpy(foo, td->file_name);
1585 p = dirname(foo);
1586 if (stat(p, &st)) {
1587 perror("disk util stat");
1588 return;
1589 }
1590
1591 dev = st.st_dev;
1592 }
1593
1594 if (disk_util_exists(dev))
1595 return;
1596
1597 sprintf(foo, "/sys/block");
1598 if (!find_block_dir(dev, foo))
1599 return;
1600
1601 /*
64cb3e10 1602 * for md/dm, there's no queue dir. we already have the right place
ebac4655 1603 */
64cb3e10 1604 sprintf(tmp, "%s/stat", foo);
ebac4655 1605 if (stat(tmp, &st)) {
64cb3e10
JA		 1606 /*
1607 * if this is inside a partition dir, jump back to parent
1608 */
1609 sprintf(tmp, "%s/queue", foo);
ebac4655 1610 if (stat(tmp, &st)) {
64cb3e10
JA		 1611 p = dirname(foo);
1612 sprintf(tmp, "%s/queue", p);
1613 if (stat(tmp, &st)) {
1614 fprintf(stderr, "unknown sysfs layout\n");
1615 return;
1616 }
1617 sprintf(foo, "%s", p);
ebac4655 1618 }
ebac4655
JA		 1619 }
1620
1621 disk_util_add(dev, foo);
1622}
1623
1624static void disk_util_timer_arm(void)
1625{
1626 itimer.it_value.tv_sec = 0;
1627 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1628 setitimer(ITIMER_REAL, &itimer, NULL);
1629}
1630
1631static void clear_io_state(struct thread_data *td)
1632{
1633 if (td->io_engine == FIO_SYNCIO)
1634 lseek(td->fd, SEEK_SET, 0);
1635
20dc95c4 1636 td->last_pos = 0;
ebac4655
JA		 1637 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1638 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 1639 td->zone_bytes = 0;
ebac4655
JA		 1640
1641 if (td->file_map)
1642 memset(td->file_map, 0, td->num_maps * sizeof(long));
1643}
1644
1645static void update_rusage_stat(struct thread_data *td)
1646{
1647 if (!(td->runtime[0] + td->runtime[1]))
1648 return;
1649
1650 getrusage(RUSAGE_SELF, &td->ru_end);
1651
1652 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1653 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1654 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1655
1656
1657 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1658}
1659
1660static void *thread_main(void *data)
1661{
1662 struct thread_data *td = data;
1663 int ret = 1;
1664
1665 if (!td->use_thread)
1666 setsid();
1667
1668 td->pid = getpid();
1669
1670 if (init_io_u(td))
1671 goto err;
1672
1673 if (fio_setaffinity(td) == -1) {
1674 td_verror(td, errno);
1675 goto err;
1676 }
1677
1678 if (init_io(td))
1679 goto err;
1680
1681 if (td->ioprio) {
1682 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1683 td_verror(td, errno);
1684 goto err;
1685 }
1686 }
1687
1688 sem_post(&startup_sem);
1689 sem_wait(&td->mutex);
1690
1691 if (!td->create_serialize && setup_file(td))
1692 goto err;
1693
1694 if (init_random_state(td))
1695 goto err;
1696
1697 gettimeofday(&td->epoch, NULL);
1698
1699 while (td->loops--) {
1700 getrusage(RUSAGE_SELF, &td->ru_start);
1701 gettimeofday(&td->start, NULL);
1702 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1703
1704 if (td->ratemin)
1705 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1706
1707 clear_io_state(td);
1708 prune_io_piece_log(td);
1709
1710 do_io(td);
1711
1712 td->runtime[td->ddir] += mtime_since_now(&td->start);
1713 update_rusage_stat(td);
1714
1715 if (td->error || td->terminate)
1716 break;
1717
1718 if (td->verify == VERIFY_NONE)
1719 continue;
1720
1721 clear_io_state(td);
1722 gettimeofday(&td->start, NULL);
1723
1724 do_verify(td);
1725
1726 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1727
1728 if (td->error || td->terminate)
1729 break;
1730 }
1731
1732 ret = 0;
1733
1734 if (td->bw_log)
1735 finish_log(td, td->bw_log, "bw");
1736 if (td->slat_log)
1737 finish_log(td, td->slat_log, "slat");
1738 if (td->clat_log)
1739 finish_log(td, td->clat_log, "clat");
1740
1741 if (exitall_on_terminate)
1742 terminate_threads(td->groupid);
1743
1744err:
1745 if (td->fd != -1) {
1746 close(td->fd);
1747 td->fd = -1;
1748 }
1749 if (td->mmap)
1750 munmap(td->mmap, td->file_size);
1751 cleanup_io(td);
1752 cleanup_io_u(td);
1753 if (ret) {
1754 sem_post(&startup_sem);
1755 sem_wait(&td->mutex);
1756 }
1757 td_set_runstate(td, TD_EXITED);
1758 return NULL;
1759
1760}
1761
1762static void *fork_main(int shmid, int offset)
1763{
1764 struct thread_data *td;
1765 void *data;
1766
1767 data = shmat(shmid, NULL, 0);
1768 if (data == (void *) -1) {
1769 perror("shmat");
1770 return NULL;
1771 }
1772
1773 td = data + offset * sizeof(struct thread_data);
1774 thread_main(td);
1775 shmdt(data);
1776 return NULL;
1777}
1778
1779static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1780 double *mean, double *dev)
1781{
1782 double n;
1783
1784 if (is->samples == 0)
1785 return 0;
1786
1787 *min = is->min_val;
1788 *max = is->max_val;
1789
1790 n = (double) is->samples;
1791 *mean = (double) is->val / n;
1792 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1793 if (!(*min + *max) && !(*mean + *dev))
1794 return 0;
1795
1796 return 1;
1797}
1798
1799static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1800 int ddir)
1801{
1802 char *ddir_str[] = { "read ", "write" };
1803 unsigned long min, max, bw;
1804 double mean, dev;
1805
1806 if (!td->runtime[ddir])
1807 return;
1808
1809 bw = td->io_bytes[ddir] / td->runtime[ddir];
9104f874 1810 printf(" %s: io=%6lluMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
ebac4655
JA		 1811
1812 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1813 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1814
1815 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1816 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1817
1818 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1819 double p_of_agg;
1820
1821 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1822 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);
1823 }
1824}
1825
1826static void show_thread_status(struct thread_data *td,
1827 struct group_run_stats *rs)
1828{
1829 double usr_cpu, sys_cpu;
1830
1831 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1832 return;
1833
1834 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1835
1836 show_ddir_status(td, rs, td->ddir);
1837 show_ddir_status(td, rs, td->ddir ^ 1);
1838
1839 if (td->runtime[0] + td->runtime[1]) {
1840 double runt = td->runtime[0] + td->runtime[1];
1841
1842 usr_cpu = (double) td->usr_time * 100 / runt;
1843 sys_cpu = (double) td->sys_time * 100 / runt;
1844 } else {
1845 usr_cpu = 0;
1846 sys_cpu = 0;
1847 }
1848
1849 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
1850}
1851
1852static void check_str_update(struct thread_data *td)
1853{
1854 char c = run_str[td->thread_number - 1];
1855
1856 if (td->runstate == td->old_runstate)
1857 return;
1858
1859 switch (td->runstate) {
1860 case TD_REAPED:
1861 c = '_';
1862 break;
1863 case TD_EXITED:
1864 c = 'E';
1865 break;
1866 case TD_RUNNING:
1867 if (td_read(td)) {
1868 if (td->sequential)
1869 c = 'R';
1870 else
1871 c = 'r';
1872 } else {
1873 if (td->sequential)
1874 c = 'W';
1875 else
1876 c = 'w';
1877 }
1878 break;
1879 case TD_VERIFYING:
1880 c = 'V';
1881 break;
1882 case TD_CREATED:
1883 c = 'C';
1884 break;
1885 case TD_NOT_CREATED:
1886 c = 'P';
1887 break;
1888 default:
1889 printf("state %d\n", td->runstate);
1890 }
1891
1892 run_str[td->thread_number - 1] = c;
1893 td->old_runstate = td->runstate;
1894}
1895
5289b847
JA		 1896static void eta_to_str(char *str, int eta_sec)
1897{
1898 unsigned int d, h, m, s;
1899 static int always_d, always_h;
1900
1901 d = h = m = s = 0;
1902
1903 s = eta_sec % 60;
1904 eta_sec /= 60;
1905 m = eta_sec % 60;
1906 eta_sec /= 60;
1907 h = eta_sec % 24;
1908 eta_sec /= 24;
1909 d = eta_sec;
1910
1911 if (d || always_d) {
1912 always_d = 1;
1913 str += sprintf(str, "%02dd:", d);
1914 }
1915 if (h || always_h) {
1916 always_h = 1;
1917 str += sprintf(str, "%02dh:", h);
1918 }
1919
1920 str += sprintf(str, "%02dm:", m);
1921 str += sprintf(str, "%02ds", s);
1922}
1923
ebac4655
JA		 1924static void print_thread_status(void)
1925{
1926 unsigned long long bytes_done, bytes_total;
5289b847
JA		 1927 int i, nr_running, t_rate, m_rate, eta_sec;
1928 char eta_str[32];
ebac4655
JA		 1929 double perc;
1930
1931 bytes_done = bytes_total = 0;
1932 nr_running = t_rate = m_rate = 0;
1933 for (i = 0; i < thread_number; i++) {
1934 struct thread_data *td = &threads[i];
1935
1936 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
1937 nr_running++;
1938 t_rate += td->rate;
1939 m_rate += td->ratemin;
1940 }
1941
1942 bytes_total += td->total_io_size;
1943 if (td->verify)
1944 bytes_total += td->total_io_size;
1945
0d98aadc 1946 if (td->zone_size && td->zone_skip)
20dc95c4
JA		 1947 bytes_total /= (td->zone_skip / td->zone_size);
1948
ebac4655
JA		 1949 bytes_done += td->io_bytes[DDIR_READ] +td->io_bytes[DDIR_WRITE];
1950
1951 check_str_update(td);
1952 }
1953
1954 perc = 0;
5289b847 1955 eta_sec = 0;
ebac4655 1956 if (bytes_total && bytes_done) {
5289b847
JA		 1957 unsigned long runtime;
1958
1959 perc = (double) bytes_done / (double) bytes_total;
1960 if (perc > 1.0)
1961 perc = 1.0;
1962
1963 runtime = time_since_now(&genesis);
1964 if (runtime >= 5) {
1965 memset(eta_str, 0, sizeof(eta_str));
1966 eta_sec = (runtime * (1.0 / perc)) - runtime;
1967 eta_to_str(eta_str, eta_sec);
1968 }
1969
1970 perc *= 100.0;
ebac4655
JA		 1971 }
1972
5289b847 1973 printf("Threads now running (%d)", nr_running);
ebac4655
JA		 1974 if (m_rate || t_rate)
1975 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
5289b847
JA		 1976 printf(": [%s] [%3.2f%% done]", run_str, perc);
1977 if (eta_sec)
1978 printf(" [eta %s]", eta_str);
1979 printf("\r");
ebac4655
JA		 1980 fflush(stdout);
1981}
1982
1983static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1984{
1985 int i;
1986
1987 /*
1988 * reap exited threads (TD_EXITED -> TD_REAPED)
1989 */
1990 for (i = 0; i < thread_number; i++) {
1991 struct thread_data *td = &threads[i];
1992
1993 if (td->runstate != TD_EXITED)
1994 continue;
1995
1996 td_set_runstate(td, TD_REAPED);
1997
1998 if (td->use_thread) {
1999 long ret;
2000
2001 if (pthread_join(td->thread, (void *) &ret))
2002 perror("thread_join");
2003 } else
2004 waitpid(td->pid, NULL, 0);
2005
2006 (*nr_running)--;
2007 (*m_rate) -= td->ratemin;
2008 (*t_rate) -= td->rate;
2009 }
2010}
2011
2012static void run_threads(void)
2013{
ebac4655
JA		 2014 struct thread_data *td;
2015 unsigned long spent;
2016 int i, todo, nr_running, m_rate, t_rate, nr_started;
2017
2018 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
2019 fflush(stdout);
2020
2021 signal(SIGINT, sig_handler);
2022 signal(SIGALRM, sig_handler);
2023
2024 todo = thread_number;
2025 nr_running = 0;
2026 nr_started = 0;
2027 m_rate = t_rate = 0;
2028
2029 for (i = 0; i < thread_number; i++) {
2030 td = &threads[i];
2031
2032 run_str[td->thread_number - 1] = 'P';
2033
2034 init_disk_util(td);
2035
2036 if (!td->create_serialize)
2037 continue;
2038
2039 /*
2040 * do file setup here so it happens sequentially,
2041 * we don't want X number of threads getting their
2042 * client data interspersed on disk
2043 */
2044 if (setup_file(td)) {
2045 td_set_runstate(td, TD_REAPED);
2046 todo--;
2047 }
2048 }
2049
2050 gettimeofday(&genesis, NULL);
2051
2052 while (todo) {
2053 /*
2054 * create threads (TD_NOT_CREATED -> TD_CREATED)
2055 */
2056 for (i = 0; i < thread_number; i++) {
2057 td = &threads[i];
2058
2059 if (td->runstate != TD_NOT_CREATED)
2060 continue;
2061
2062 /*
2063 * never got a chance to start, killed by other
2064 * thread for some reason
2065 */
2066 if (td->terminate) {
2067 todo--;
2068 continue;
2069 }
2070
2071 if (td->start_delay) {
2072 spent = mtime_since_now(&genesis);
2073
2074 if (td->start_delay * 1000 > spent)
2075 continue;
2076 }
2077
2078 if (td->stonewall && (nr_started || nr_running))
2079 break;
2080
2081 td_set_runstate(td, TD_CREATED);
2082 sem_init(&startup_sem, 0, 1);
2083 todo--;
2084 nr_started++;
2085
2086 if (td->use_thread) {
2087 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2088 perror("thread_create");
2089 nr_started--;
2090 }
2091 } else {
2092 if (fork())
2093 sem_wait(&startup_sem);
2094 else {
2095 fork_main(shm_id, i);
2096 exit(0);
2097 }
2098 }
2099 }
2100
2101 /*
2102 * start created threads (TD_CREATED -> TD_RUNNING)
2103 */
2104 for (i = 0; i < thread_number; i++) {
2105 td = &threads[i];
2106
2107 if (td->runstate != TD_CREATED)
2108 continue;
2109
2110 td_set_runstate(td, TD_RUNNING);
2111 nr_running++;
2112 nr_started--;
2113 m_rate += td->ratemin;
2114 t_rate += td->rate;
2115 sem_post(&td->mutex);
2116 }
2117
2118 reap_threads(&nr_running, &t_rate, &m_rate);
2119
2120 if (todo)
2121 usleep(100000);
2122 }
2123
2124 while (nr_running) {
2125 reap_threads(&nr_running, &t_rate, &m_rate);
2126 usleep(10000);
2127 }
2128
2129 update_io_ticks();
2130}
2131
2132static void show_group_stats(struct group_run_stats *rs, int id)
2133{
2134 printf("\nRun status group %d (all jobs):\n", id);
2135
2136 if (rs->max_run[DDIR_READ])
9104f874 2137 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 2138 if (rs->max_run[DDIR_WRITE])
9104f874 2139 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		 2140}
2141
2142static void show_disk_util(void)
2143{
2144 struct disk_util_stat *dus;
2145 struct list_head *entry;
2146 struct disk_util *du;
2147 double util;
2148
2149 printf("\nDisk stats (read/write):\n");
2150
2151 list_for_each(entry, &disk_list) {
2152 du = list_entry(entry, struct disk_util, list);
2153 dus = &du->dus;
2154
2155 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2156 if (util > 100.0)
2157 util = 100.0;
2158
2159 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);
2160 }
2161}
2162
2163static void show_run_stats(void)
2164{
2165 struct group_run_stats *runstats, *rs;
2166 struct thread_data *td;
2167 int i;
2168
2169 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2170
2171 for (i = 0; i < groupid + 1; i++) {
2172 rs = &runstats[i];
2173
2174 memset(rs, 0, sizeof(*rs));
2175 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2176 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2177 }
2178
2179 for (i = 0; i < thread_number; i++) {
2180 unsigned long rbw, wbw;
2181
2182 td = &threads[i];
2183
2184 if (td->error) {
2185 printf("Client%d: %s\n", td->thread_number, td->verror);
2186 continue;
2187 }
2188
2189 rs = &runstats[td->groupid];
2190
2191 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2192 rs->min_run[0] = td->runtime[0];
2193 if (td->runtime[0] > rs->max_run[0])
2194 rs->max_run[0] = td->runtime[0];
2195 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2196 rs->min_run[1] = td->runtime[1];
2197 if (td->runtime[1] > rs->max_run[1])
2198 rs->max_run[1] = td->runtime[1];
2199
2200 rbw = wbw = 0;
2201 if (td->runtime[0])
2202 rbw = td->io_bytes[0] / td->runtime[0];
2203 if (td->runtime[1])
2204 wbw = td->io_bytes[1] / td->runtime[1];
2205
2206 if (rbw < rs->min_bw[0])
2207 rs->min_bw[0] = rbw;
2208 if (wbw < rs->min_bw[1])
2209 rs->min_bw[1] = wbw;
2210 if (rbw > rs->max_bw[0])
2211 rs->max_bw[0] = rbw;
2212 if (wbw > rs->max_bw[1])
2213 rs->max_bw[1] = wbw;
2214
2215 rs->io_mb[0] += td->io_bytes[0] >> 20;
2216 rs->io_mb[1] += td->io_bytes[1] >> 20;
2217 }
2218
2219 for (i = 0; i < groupid + 1; i++) {
2220 rs = &runstats[i];
2221
2222 if (rs->max_run[0])
2223 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2224 if (rs->max_run[1])
2225 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2226 }
2227
2228 /*
2229 * don't overwrite last signal output
2230 */
2231 printf("\n");
2232
2233 for (i = 0; i < thread_number; i++) {
2234 td = &threads[i];
2235 rs = &runstats[td->groupid];
2236
2237 show_thread_status(td, rs);
2238 }
2239
2240 for (i = 0; i < groupid + 1; i++)
2241 show_group_stats(&runstats[i], i);
2242
2243 show_disk_util();
2244}
2245
2246int main(int argc, char *argv[])
2247{
2248 if (parse_options(argc, argv))
2249 return 1;
2250
2251 if (!thread_number) {
2252 printf("Nothing to do\n");
2253 return 1;
2254 }
2255
2256 disk_util_timer_arm();
2257
2258 run_threads();
2259 show_run_stats();
2260
2261 return 0;
2262}
fio - Flexible IO Tester