projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
[PATCH] Add support for specifying read/write mix workloads better
[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
c04f7ec3
JA		 60extern unsigned long long mlock_size;
61
ebac4655
JA		 62/*
63 * thread life cycle
64 */
65enum {
66 TD_NOT_CREATED = 0,
67 TD_CREATED,
75154845 68 TD_INITIALIZED,
ebac4655
JA		 69 TD_RUNNING,
70 TD_VERIFYING,
71 TD_EXITED,
72 TD_REAPED,
73};
74
3d60d1ed 75#define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
ebac4655
JA		 76
77static sem_t startup_sem;
78
79#define TERMINATE_ALL (-1)
75154845 80#define JOB_START_TIMEOUT (5 * 1000)
ebac4655
JA		 81
82static void terminate_threads(int group_id)
83{
84 int i;
85
86 for (i = 0; i < thread_number; i++) {
87 struct thread_data *td = &threads[i];
88
89 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
90 td->terminate = 1;
91 td->start_delay = 0;
92 }
93 }
94}
95
96static void sig_handler(int sig)
97{
98 switch (sig) {
99 case SIGALRM:
100 update_io_ticks();
101 disk_util_timer_arm();
102 print_thread_status();
103 break;
104 default:
105 printf("\nfio: terminating on signal\n");
106 fflush(stdout);
107 terminate_threads(TERMINATE_ALL);
108 break;
109 }
110}
111
112static unsigned long utime_since(struct timeval *s, struct timeval *e)
113{
114 double sec, usec;
115
116 sec = e->tv_sec - s->tv_sec;
117 usec = e->tv_usec - s->tv_usec;
118 if (sec > 0 && usec < 0) {
119 sec--;
120 usec += 1000000;
121 }
122
123 sec *= (double) 1000000;
124
125 return sec + usec;
126}
127
128static unsigned long utime_since_now(struct timeval *s)
129{
130 struct timeval t;
131
132 gettimeofday(&t, NULL);
133 return utime_since(s, &t);
134}
135
136static unsigned long mtime_since(struct timeval *s, struct timeval *e)
137{
138 double sec, usec;
139
140 sec = e->tv_sec - s->tv_sec;
141 usec = e->tv_usec - s->tv_usec;
142 if (sec > 0 && usec < 0) {
143 sec--;
144 usec += 1000000;
145 }
146
147 sec *= (double) 1000;
148 usec /= (double) 1000;
149
150 return sec + usec;
151}
152
153static unsigned long mtime_since_now(struct timeval *s)
154{
155 struct timeval t;
156
157 gettimeofday(&t, NULL);
158 return mtime_since(s, &t);
159}
160
161static inline unsigned long msec_now(struct timeval *s)
162{
163 return s->tv_sec * 1000 + s->tv_usec / 1000;
164}
165
5289b847
JA		 166static unsigned long time_since_now(struct timeval *s)
167{
168 return mtime_since_now(s) / 1000;
169}
170
ebac4655
JA		 171static int random_map_free(struct thread_data *td, unsigned long long block)
172{
173 unsigned int idx = RAND_MAP_IDX(td, block);
174 unsigned int bit = RAND_MAP_BIT(td, block);
175
176 return (td->file_map[idx] & (1UL << bit)) == 0;
177}
178
179static int get_next_free_block(struct thread_data *td, unsigned long long *b)
180{
181 int i;
182
183 *b = 0;
184 i = 0;
185 while ((*b) * td->min_bs < td->io_size) {
186 if (td->file_map[i] != -1UL) {
187 *b += ffz(td->file_map[i]);
188 return 0;
189 }
190
191 *b += BLOCKS_PER_MAP;
192 i++;
193 }
194
195 return 1;
196}
197
198static void mark_random_map(struct thread_data *td, struct io_u *io_u)
199{
200bc855 200 unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
ebac4655
JA		 201 unsigned int blocks = 0;
202
203 while (blocks < (io_u->buflen / td->min_bs)) {
204 unsigned int idx, bit;
205
206 if (!random_map_free(td, block))
207 break;
208
209 idx = RAND_MAP_IDX(td, block);
210 bit = RAND_MAP_BIT(td, block);
211
212 assert(idx < td->num_maps);
213
214 td->file_map[idx] |= (1UL << bit);
215 block++;
216 blocks++;
217 }
218
219 if ((blocks * td->min_bs) < io_u->buflen)
220 io_u->buflen = blocks * td->min_bs;
221}
222
ebac4655
JA		 223static inline void add_stat_sample(struct io_stat *is, unsigned long val)
224{
225 if (val > is->max_val)
226 is->max_val = val;
227 if (val < is->min_val)
228 is->min_val = val;
229
230 is->val += val;
231 is->val_sq += val * val;
232 is->samples++;
233}
234
235static void add_log_sample(struct thread_data *td, struct io_log *iolog,
236 unsigned long val, int ddir)
237{
238 if (iolog->nr_samples == iolog->max_samples) {
239 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
240
241 iolog->log = realloc(iolog->log, new_size);
242 iolog->max_samples <<= 1;
243 }
244
245 iolog->log[iolog->nr_samples].val = val;
246 iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
247 iolog->log[iolog->nr_samples].ddir = ddir;
248 iolog->nr_samples++;
249}
250
251static void add_clat_sample(struct thread_data *td, int ddir,unsigned long msec)
252{
253 add_stat_sample(&td->clat_stat[ddir], msec);
254
255 if (td->clat_log)
256 add_log_sample(td, td->clat_log, msec, ddir);
257}
258
259static void add_slat_sample(struct thread_data *td, int ddir,unsigned long msec)
260{
261 add_stat_sample(&td->slat_stat[ddir], msec);
262
263 if (td->slat_log)
264 add_log_sample(td, td->slat_log, msec, ddir);
265}
266
267static void add_bw_sample(struct thread_data *td, int ddir)
268{
269 unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
270 unsigned long rate;
271
272 if (spent < td->bw_avg_time)
273 return;
274
275 rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
276 add_stat_sample(&td->bw_stat[ddir], rate);
277
278 if (td->bw_log)
279 add_log_sample(td, td->bw_log, rate, ddir);
280
281 gettimeofday(&td->stat_sample_time[ddir], NULL);
282 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
283}
284
20dc95c4
JA		 285static int get_next_offset(struct thread_data *td, unsigned long long *offset)
286{
287 unsigned long long b, rb;
288 long r;
289
290 if (!td->sequential) {
085227ad 291 unsigned long long max_blocks = td->io_size / td->min_bs;
20dc95c4
JA		 292 int loops = 50;
293
294 do {
295 lrand48_r(&td->random_state, &r);
085227ad 296 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
20dc95c4
JA		 297 rb = b + (td->file_offset / td->min_bs);
298 loops--;
299 } while (!random_map_free(td, rb) && loops);
300
301 if (!loops) {
302 if (get_next_free_block(td, &b))
303 return 1;
304 }
305 } else
306 b = td->last_pos / td->min_bs;
307
308 *offset = (b * td->min_bs) + td->file_offset;
838a3cd3 309 if (*offset > td->real_file_size)
20dc95c4
JA		 310 return 1;
311
312 return 0;
313}
314
315static unsigned int get_next_buflen(struct thread_data *td)
316{
317 unsigned int buflen;
318 long r;
319
320 if (td->min_bs == td->max_bs)
321 buflen = td->min_bs;
322 else {
323 lrand48_r(&td->bsrange_state, &r);
324 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
325 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
326 }
327
328 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
329 buflen = td->io_size - td->this_io_bytes[td->ddir];
330
331 return buflen;
332}
333
ebac4655
JA		 334/*
335 * busy looping version for the last few usec
336 */
337static void __usec_sleep(unsigned int usec)
338{
339 struct timeval start;
340
341 gettimeofday(&start, NULL);
342 while (utime_since_now(&start) < usec)
343 nop;
344}
345
346static void usec_sleep(struct thread_data *td, unsigned long usec)
347{
348 struct timespec req, rem;
349
350 req.tv_sec = usec / 1000000;
351 req.tv_nsec = usec * 1000 - req.tv_sec * 1000000;
352
353 do {
354 if (usec < 5000) {
355 __usec_sleep(usec);
356 break;
357 }
358
359 rem.tv_sec = rem.tv_nsec = 0;
360 if (nanosleep(&req, &rem) < 0)
361 break;
362
363 if ((rem.tv_sec + rem.tv_nsec) == 0)
364 break;
365
366 req.tv_nsec = rem.tv_nsec;
367 req.tv_sec = rem.tv_sec;
368
369 usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000;
370 } while (!td->terminate);
371}
372
373static void rate_throttle(struct thread_data *td, unsigned long time_spent,
374 unsigned int bytes)
375{
376 unsigned long usec_cycle;
377
378 if (!td->rate)
379 return;
380
381 usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs);
382
383 if (time_spent < usec_cycle) {
384 unsigned long s = usec_cycle - time_spent;
385
386 td->rate_pending_usleep += s;
387 if (td->rate_pending_usleep >= 100000) {
388 usec_sleep(td, td->rate_pending_usleep);
389 td->rate_pending_usleep = 0;
390 }
391 } else {
392 long overtime = time_spent - usec_cycle;
393
394 td->rate_pending_usleep -= overtime;
395 }
396}
397
398static int check_min_rate(struct thread_data *td, struct timeval *now)
399{
400 unsigned long spent;
401 unsigned long rate;
402 int ddir = td->ddir;
403
404 /*
405 * allow a 2 second settle period in the beginning
406 */
407 if (mtime_since(&td->start, now) < 2000)
408 return 0;
409
410 /*
411 * if rate blocks is set, sample is running
412 */
413 if (td->rate_bytes) {
414 spent = mtime_since(&td->lastrate, now);
415 if (spent < td->ratecycle)
416 return 0;
417
418 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
419 if (rate < td->ratemin) {
420 printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
421 if (rate_quit)
422 terminate_threads(td->groupid);
423 return 1;
424 }
425 }
426
427 td->rate_bytes = td->this_io_bytes[ddir];
428 memcpy(&td->lastrate, now, sizeof(*now));
429 return 0;
430}
431
432static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
433{
434 if (!td->timeout)
435 return 0;
436 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
437 return 1;
438
439 return 0;
440}
441
442static void fill_random_bytes(struct thread_data *td,
443 unsigned char *p, unsigned int len)
444{
445 unsigned int todo;
446 double r;
447
448 while (len) {
449 drand48_r(&td->verify_state, &r);
450
451 /*
452 * lrand48_r seems to be broken and only fill the bottom
453 * 32-bits, even on 64-bit archs with 64-bit longs
454 */
455 todo = sizeof(r);
456 if (todo > len)
457 todo = len;
458
459 memcpy(p, &r, todo);
460
461 len -= todo;
462 p += todo;
463 }
464}
465
466static void hexdump(void *buffer, int len)
467{
468 unsigned char *p = buffer;
469 int i;
470
471 for (i = 0; i < len; i++)
472 printf("%02x", p[i]);
473 printf("\n");
474}
475
476static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
477{
478 unsigned char *p = (unsigned char *) io_u->buf;
479 unsigned long c;
480 int ret;
481
482 p += sizeof(*hdr);
483 c = crc32(p, hdr->len - sizeof(*hdr));
484 ret = c != hdr->crc32;
485
486 if (ret) {
487 fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
488 fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
489 }
490
491 return ret;
492}
493
494static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
495{
496 unsigned char *p = (unsigned char *) io_u->buf;
497 struct md5_ctx md5_ctx;
498 int ret;
499
500 memset(&md5_ctx, 0, sizeof(md5_ctx));
501 p += sizeof(*hdr);
502 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
503
504 ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
505 if (ret) {
506 fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
507 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
508 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
509 }
510
511 return ret;
512}
513
514static int verify_io_u(struct io_u *io_u)
515{
516 struct verify_header *hdr = (struct verify_header *) io_u->buf;
517 int ret;
518
519 if (hdr->fio_magic != FIO_HDR_MAGIC)
520 return 1;
521
522 if (hdr->verify_type == VERIFY_MD5)
523 ret = verify_io_u_md5(hdr, io_u);
524 else if (hdr->verify_type == VERIFY_CRC32)
525 ret = verify_io_u_crc32(hdr, io_u);
526 else {
527 fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
528 ret = 1;
529 }
530
531 return ret;
532}
533
534static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
535{
536 hdr->crc32 = crc32(p, len);
537}
538
539static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
540{
541 struct md5_ctx md5_ctx;
542
543 memset(&md5_ctx, 0, sizeof(md5_ctx));
544 md5_update(&md5_ctx, p, len);
545 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
546}
547
a6ccc7be 548static int get_rw_ddir(struct thread_data *td)
3d60d1ed 549{
a6ccc7be
JA		 550 if (td_rw(td)) {
551 struct timeval now;
552 unsigned long elapsed;
3d60d1ed 553
a6ccc7be
JA		 554 gettimeofday(&now, NULL);
555 elapsed = mtime_since_now(&td->rwmix_switch);
3d60d1ed 556
a6ccc7be
JA		 557 /*
558 * Check if it's time to seed a new data direction.
559 */
560 if (elapsed >= td->rwmixcycle) {
561 unsigned long v;
562 long r;
3d60d1ed 563
a6ccc7be
JA		 564 lrand48_r(&td->random_state, &r);
565 v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
566 if (v < td->rwmixread)
567 td->rwmix_ddir = DDIR_READ;
568 else
569 td->rwmix_ddir = DDIR_WRITE;
570 memcpy(&td->rwmix_switch, &now, sizeof(now));
571 }
572 return td->rwmix_ddir;
573 } else if (td_read(td))
3d60d1ed
JA		 574 return DDIR_READ;
575 else
576 return DDIR_WRITE;
577}
578
ebac4655
JA		 579/*
580 * fill body of io_u->buf with random data and add a header with the
581 * (eg) sha1sum of that data.
582 */
583static void populate_io_u(struct thread_data *td, struct io_u *io_u)
584{
585 unsigned char *p = (unsigned char *) io_u->buf;
586 struct verify_header hdr;
587
588 hdr.fio_magic = FIO_HDR_MAGIC;
589 hdr.len = io_u->buflen;
590 p += sizeof(hdr);
591 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
592
593 if (td->verify == VERIFY_MD5) {
594 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
595 hdr.verify_type = VERIFY_MD5;
596 } else {
597 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
598 hdr.verify_type = VERIFY_CRC32;
599 }
600
601 memcpy(io_u->buf, &hdr, sizeof(hdr));
602}
603
aea47d44 604static int td_io_prep(struct thread_data *td, struct io_u *io_u)
20dc95c4 605{
20dc95c4
JA		 606 if (td->io_prep && td->io_prep(td, io_u))
607 return 1;
608
609 return 0;
610}
611
b1ff3403 612void put_io_u(struct thread_data *td, struct io_u *io_u)
ebac4655
JA		 613{
614 list_del(&io_u->list);
615 list_add(&io_u->list, &td->io_u_freelist);
616 td->cur_depth--;
617}
618
aea47d44
JA		 619static int fill_io_u(struct thread_data *td, struct io_u *io_u)
620{
621 /*
622 * If using an iolog, grab next piece if any available.
623 */
624 if (td->iolog) {
625 struct io_piece *ipo;
626
627 if (list_empty(&td->io_log_list))
628 return 1;
629
630 ipo = list_entry(td->io_log_list.next, struct io_piece, list);
631 list_del(&ipo->list);
632 io_u->offset = ipo->offset;
633 io_u->buflen = ipo->len;
634 io_u->ddir = ipo->ddir;
635 free(ipo);
636 return 0;
637 }
638
639 /*
640 * No log, let the seq/rand engine retrieve the next position.
641 */
642 if (!get_next_offset(td, &io_u->offset)) {
643 io_u->buflen = get_next_buflen(td);
644
645 if (io_u->buflen) {
646 io_u->ddir = get_rw_ddir(td);
647 return 0;
648 }
649 }
650
651 return 1;
652}
653
ebac4655
JA		 654#define queue_full(td) (list_empty(&(td)->io_u_freelist))
655
b1ff3403 656struct io_u *__get_io_u(struct thread_data *td)
ebac4655
JA		 657{
658 struct io_u *io_u;
659
660 if (queue_full(td))
661 return NULL;
662
663 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
664 io_u->error = 0;
665 io_u->resid = 0;
666 list_del(&io_u->list);
667 list_add(&io_u->list, &td->io_u_busylist);
668 td->cur_depth++;
669 return io_u;
670}
671
ebac4655
JA		 672static struct io_u *get_io_u(struct thread_data *td)
673{
674 struct io_u *io_u;
675
676 io_u = __get_io_u(td);
677 if (!io_u)
678 return NULL;
679
20dc95c4
JA		 680 if (td->zone_bytes >= td->zone_size) {
681 td->zone_bytes = 0;
682 td->last_pos += td->zone_skip;
683 }
684
aea47d44 685 if (fill_io_u(td, io_u)) {
ebac4655
JA		 686 put_io_u(td, io_u);
687 return NULL;
688 }
689
838a3cd3
JA		 690 if (io_u->buflen + io_u->offset > td->real_file_size)
691 io_u->buflen = td->real_file_size - io_u->offset;
ebac4655
JA		 692
693 if (!io_u->buflen) {
694 put_io_u(td, io_u);
695 return NULL;
696 }
697
aea47d44 698 if (!td->iolog && !td->sequential)
ebac4655
JA		 699 mark_random_map(td, io_u);
700
20dc95c4 701 td->last_pos += io_u->buflen;
ebac4655
JA		 702
703 if (td->verify != VERIFY_NONE)
704 populate_io_u(td, io_u);
705
aea47d44 706 if (td_io_prep(td, io_u)) {
ebac4655
JA		 707 put_io_u(td, io_u);
708 return NULL;
709 }
710
711 gettimeofday(&io_u->start_time, NULL);
712 return io_u;
713}
714
715static inline void td_set_runstate(struct thread_data *td, int runstate)
716{
717 td->old_runstate = td->runstate;
718 td->runstate = runstate;
719}
720
aea47d44 721static int get_next_verify(struct thread_data *td, struct io_u *io_u)
ebac4655
JA		 722{
723 struct io_piece *ipo;
724
725 if (list_empty(&td->io_hist_list))
726 return 1;
727
728 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
729 list_del(&ipo->list);
730
aea47d44
JA		 731 io_u->offset = ipo->offset;
732 io_u->buflen = ipo->len;
733 io_u->ddir = DDIR_READ;
ebac4655
JA		 734 free(ipo);
735 return 0;
736}
737
738static void prune_io_piece_log(struct thread_data *td)
739{
740 struct io_piece *ipo;
741
742 while (!list_empty(&td->io_hist_list)) {
743 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
744
745 list_del(&ipo->list);
746 free(ipo);
747 }
748}
749
750/*
751 * log a succesful write, so we can unwind the log for verify
752 */
753static void log_io_piece(struct thread_data *td, struct io_u *io_u)
754{
755 struct io_piece *ipo = malloc(sizeof(struct io_piece));
756 struct list_head *entry;
757
758 INIT_LIST_HEAD(&ipo->list);
759 ipo->offset = io_u->offset;
760 ipo->len = io_u->buflen;
761
762 /*
763 * for random io where the writes extend the file, it will typically
764 * be laid out with the block scattered as written. it's faster to
765 * read them in in that order again, so don't sort
766 */
767 if (td->sequential || !td->overwrite) {
768 list_add_tail(&ipo->list, &td->io_hist_list);
769 return;
770 }
771
772 /*
773 * for random io, sort the list so verify will run faster
774 */
775 entry = &td->io_hist_list;
776 while ((entry = entry->prev) != &td->io_hist_list) {
777 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
778
779 if (__ipo->offset < ipo->offset)
780 break;
781 }
782
783 list_add(&ipo->list, entry);
784}
785
aea47d44
JA		 786static int init_iolog(struct thread_data *td)
787{
788 unsigned long long offset;
789 unsigned int bytes;
790 char *str, *p;
791 FILE *f;
792 int rw, i, reads, writes;
793
794 if (!td->iolog)
795 return 0;
796
797 f = fopen(td->iolog_file, "r");
798 if (!f) {
799 perror("fopen iolog");
800 return 1;
801 }
802
803 str = malloc(4096);
804 reads = writes = i = 0;
805 while ((p = fgets(str, 4096, f)) != NULL) {
806 struct io_piece *ipo;
807
808 if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
809 fprintf(stderr, "bad iolog: %s\n", p);
810 continue;
811 }
812 if (rw == DDIR_READ)
813 reads++;
814 else if (rw == DDIR_WRITE)
815 writes++;
816 else {
817 fprintf(stderr, "bad ddir: %d\n", rw);
818 continue;
819 }
820
821 ipo = malloc(sizeof(*ipo));
822 INIT_LIST_HEAD(&ipo->list);
823 ipo->offset = offset;
824 ipo->len = bytes;
825 if (bytes > td->max_bs)
826 td->max_bs = bytes;
827 ipo->ddir = rw;
828 list_add_tail(&ipo->list, &td->io_log_list);
829 i++;
830 }
831
832 free(str);
833 fclose(f);
834
835 if (!i)
836 return 1;
837
838 if (reads && !writes)
839 td->ddir = DDIR_READ;
840 else if (!reads && writes)
841 td->ddir = DDIR_READ;
842 else
843 td->iomix = 1;
844
845 return 0;
846}
847
ebac4655
JA		 848static int sync_td(struct thread_data *td)
849{
850 if (td->io_sync)
851 return td->io_sync(td);
852
853 return 0;
854}
855
856static int io_u_getevents(struct thread_data *td, int min, int max,
857 struct timespec *t)
858{
859 return td->io_getevents(td, min, max, t);
860}
861
862static int io_u_queue(struct thread_data *td, struct io_u *io_u)
863{
864 gettimeofday(&io_u->issue_time, NULL);
865
866 return td->io_queue(td, io_u);
867}
868
869#define iocb_time(iocb) ((unsigned long) (iocb)->data)
870
871static void io_completed(struct thread_data *td, struct io_u *io_u,
872 struct io_completion_data *icd)
873{
874 struct timeval e;
875 unsigned long msec;
876
877 gettimeofday(&e, NULL);
878
879 if (!io_u->error) {
20dc95c4 880 unsigned int bytes = io_u->buflen - io_u->resid;
3d60d1ed 881 const int idx = io_u->ddir;
ebac4655
JA		 882
883 td->io_blocks[idx]++;
20dc95c4
JA		 884 td->io_bytes[idx] += bytes;
885 td->zone_bytes += bytes;
886 td->this_io_bytes[idx] += bytes;
ebac4655
JA		 887
888 msec = mtime_since(&io_u->issue_time, &e);
889
3d60d1ed
JA		 890 add_clat_sample(td, idx, msec);
891 add_bw_sample(td, idx);
ebac4655 892
3d60d1ed 893 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
ebac4655
JA		 894 log_io_piece(td, io_u);
895
20dc95c4 896 icd->bytes_done[idx] += bytes;
ebac4655
JA		 897 } else
898 icd->error = io_u->error;
899}
900
901static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
902{
903 struct io_u *io_u;
904 int i;
905
906 icd->error = 0;
907 icd->bytes_done[0] = icd->bytes_done[1] = 0;
908
909 for (i = 0; i < icd->nr; i++) {
910 io_u = td->io_event(td, i);
911
912 io_completed(td, io_u, icd);
913 put_io_u(td, io_u);
914 }
915}
916
917static void cleanup_pending_aio(struct thread_data *td)
918{
919 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
920 struct list_head *entry, *n;
921 struct io_completion_data icd;
922 struct io_u *io_u;
923 int r;
924
925 /*
926 * get immediately available events, if any
927 */
928 r = io_u_getevents(td, 0, td->cur_depth, &ts);
929 if (r > 0) {
930 icd.nr = r;
931 ios_completed(td, &icd);
932 }
933
934 /*
935 * now cancel remaining active events
936 */
937 if (td->io_cancel) {
938 list_for_each_safe(entry, n, &td->io_u_busylist) {
939 io_u = list_entry(entry, struct io_u, list);
940
941 r = td->io_cancel(td, io_u);
942 if (!r)
943 put_io_u(td, io_u);
944 }
945 }
946
947 if (td->cur_depth) {
948 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
949 if (r > 0) {
950 icd.nr = r;
951 ios_completed(td, &icd);
952 }
953 }
954}
955
956static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
957{
958 struct io_u *v_io_u = *io_u;
959 int ret = 0;
960
961 if (v_io_u) {
962 ret = verify_io_u(v_io_u);
963 put_io_u(td, v_io_u);
964 *io_u = NULL;
965 }
966
967 return ret;
968}
969
970static void do_verify(struct thread_data *td)
971{
972 struct timeval t;
973 struct io_u *io_u, *v_io_u = NULL;
974 struct io_completion_data icd;
975 int ret;
976
977 td_set_runstate(td, TD_VERIFYING);
978
979 do {
980 if (td->terminate)
981 break;
982
983 gettimeofday(&t, NULL);
984 if (runtime_exceeded(td, &t))
985 break;
986
987 io_u = __get_io_u(td);
988 if (!io_u)
989 break;
990
aea47d44 991 if (get_next_verify(td, io_u)) {
ebac4655
JA		 992 put_io_u(td, io_u);
993 break;
994 }
995
aea47d44 996 if (td_io_prep(td, io_u)) {
ebac4655
JA		 997 put_io_u(td, io_u);
998 break;
999 }
1000
1001 ret = io_u_queue(td, io_u);
1002 if (ret) {
1003 put_io_u(td, io_u);
1004 td_verror(td, ret);
1005 break;
1006 }
1007
1008 /*
1009 * we have one pending to verify, do that while
1010 * we are doing io on the next one
1011 */
1012 if (do_io_u_verify(td, &v_io_u))
1013 break;
1014
1015 ret = io_u_getevents(td, 1, 1, NULL);
1016 if (ret != 1) {
1017 if (ret < 0)
1018 td_verror(td, ret);
1019 break;
1020 }
1021
1022 v_io_u = td->io_event(td, 0);
1023 icd.nr = 1;
1024 icd.error = 0;
1025 io_completed(td, v_io_u, &icd);
1026
1027 if (icd.error) {
1028 td_verror(td, icd.error);
1029 put_io_u(td, v_io_u);
1030 v_io_u = NULL;
1031 break;
1032 }
1033
ebac4655
JA		 1034 /*
1035 * if we can't submit more io, we need to verify now
1036 */
1037 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
1038 break;
1039
1040 } while (1);
1041
1042 do_io_u_verify(td, &v_io_u);
1043
1044 if (td->cur_depth)
1045 cleanup_pending_aio(td);
1046
1047 td_set_runstate(td, TD_RUNNING);
1048}
1049
1050static void do_io(struct thread_data *td)
1051{
1052 struct io_completion_data icd;
1053 struct timeval s, e;
1054 unsigned long usec;
1055
1056 while (td->this_io_bytes[td->ddir] < td->io_size) {
1057 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
1058 struct timespec *timeout;
1059 int ret, min_evts = 0;
1060 struct io_u *io_u;
1061
1062 if (td->terminate)
1063 break;
1064
1065 io_u = get_io_u(td);
1066 if (!io_u)
1067 break;
1068
1069 memcpy(&s, &io_u->start_time, sizeof(s));
1070
1071 ret = io_u_queue(td, io_u);
1072 if (ret) {
1073 put_io_u(td, io_u);
1074 td_verror(td, ret);
1075 break;
1076 }
1077
1078 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
1079
1080 if (td->cur_depth < td->iodepth) {
1081 timeout = &ts;
1082 min_evts = 0;
1083 } else {
1084 timeout = NULL;
1085 min_evts = 1;
1086 }
1087
1088 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
1089 if (ret < 0) {
1090 td_verror(td, ret);
1091 break;
1092 } else if (!ret)
1093 continue;
1094
1095 icd.nr = ret;
1096 ios_completed(td, &icd);
1097 if (icd.error) {
1098 td_verror(td, icd.error);
1099 break;
1100 }
1101
1102 /*
1103 * the rate is batched for now, it should work for batches
1104 * of completions except the very first one which may look
1105 * a little bursty
1106 */
1107 gettimeofday(&e, NULL);
1108 usec = utime_since(&s, &e);
1109
1110 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
1111
1112 if (check_min_rate(td, &e)) {
1113 td_verror(td, ENOMEM);
1114 break;
1115 }
1116
1117 if (runtime_exceeded(td, &e))
1118 break;
1119
1120 if (td->thinktime)
1121 usec_sleep(td, td->thinktime);
1122
1123 if (should_fsync(td) && td->fsync_blocks &&
1124 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
1125 sync_td(td);
1126 }
1127
1128 if (td->cur_depth)
1129 cleanup_pending_aio(td);
1130
fc1a4713 1131 if (should_fsync(td) && td->end_fsync)
ebac4655
JA		 1132 sync_td(td);
1133}
1134
1135static void cleanup_io(struct thread_data *td)
1136{
1137 if (td->io_cleanup)
1138 td->io_cleanup(td);
1139}
1140
1141static int init_io(struct thread_data *td)
1142{
1143 if (td->io_engine == FIO_SYNCIO)
1144 return fio_syncio_init(td);
1145 else if (td->io_engine == FIO_MMAPIO)
1146 return fio_mmapio_init(td);
1147 else if (td->io_engine == FIO_LIBAIO)
1148 return fio_libaio_init(td);
1149 else if (td->io_engine == FIO_POSIXAIO)
1150 return fio_posixaio_init(td);
1151 else if (td->io_engine == FIO_SGIO)
1152 return fio_sgio_init(td);
8756e4d4
JA		 1153 else if (td->io_engine == FIO_SPLICEIO)
1154 return fio_spliceio_init(td);
ebac4655
JA		 1155 else {
1156 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1157 return 1;
1158 }
1159}
1160
1161static void cleanup_io_u(struct thread_data *td)
1162{
1163 struct list_head *entry, *n;
1164 struct io_u *io_u;
1165
1166 list_for_each_safe(entry, n, &td->io_u_freelist) {
1167 io_u = list_entry(entry, struct io_u, list);
1168
1169 list_del(&io_u->list);
1170 free(io_u);
1171 }
1172
1173 if (td->mem_type == MEM_MALLOC)
1174 free(td->orig_buffer);
1175 else if (td->mem_type == MEM_SHM) {
1176 struct shmid_ds sbuf;
1177
1178 shmdt(td->orig_buffer);
1179 shmctl(td->shm_id, IPC_RMID, &sbuf);
1180 } else if (td->mem_type == MEM_MMAP)
1181 munmap(td->orig_buffer, td->orig_buffer_size);
1182 else
1183 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1184
1185 td->orig_buffer = NULL;
1186}
1187
1188static int init_io_u(struct thread_data *td)
1189{
1190 struct io_u *io_u;
1191 int i, max_units;
1192 char *p;
1193
1194 if (td->io_engine & FIO_SYNCIO)
1195 max_units = 1;
1196 else
1197 max_units = td->iodepth;
1198
1199 td->orig_buffer_size = td->max_bs * max_units + MASK;
1200
1201 if (td->mem_type == MEM_MALLOC)
1202 td->orig_buffer = malloc(td->orig_buffer_size);
1203 else if (td->mem_type == MEM_SHM) {
1204 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1205 if (td->shm_id < 0) {
1206 td_verror(td, errno);
1207 perror("shmget");
1208 return 1;
1209 }
1210
1211 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1212 if (td->orig_buffer == (void *) -1) {
1213 td_verror(td, errno);
1214 perror("shmat");
1215 td->orig_buffer = NULL;
1216 return 1;
1217 }
1218 } else if (td->mem_type == MEM_MMAP) {
1219 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1220 if (td->orig_buffer == MAP_FAILED) {
1221 td_verror(td, errno);
1222 perror("mmap");
1223 td->orig_buffer = NULL;
1224 return 1;
1225 }
1226 }
1227
ebac4655
JA		 1228 p = ALIGN(td->orig_buffer);
1229 for (i = 0; i < max_units; i++) {
1230 io_u = malloc(sizeof(*io_u));
1231 memset(io_u, 0, sizeof(*io_u));
1232 INIT_LIST_HEAD(&io_u->list);
1233
1234 io_u->buf = p + td->max_bs * i;
b1ff3403 1235 io_u->index = i;
ebac4655
JA		 1236 list_add(&io_u->list, &td->io_u_freelist);
1237 }
1238
1239 return 0;
1240}
1241
1242static int create_file(struct thread_data *td, unsigned long long size,
1243 int extend)
1244{
1245 unsigned long long left;
1246 unsigned int bs;
1247 int r, oflags;
1248 char *b;
1249
1250 /*
1251 * unless specifically asked for overwrite, let normal io extend it
1252 */
1253 if (td_write(td) && !td->overwrite)
1254 return 0;
1255
1256 if (!size) {
1257 fprintf(stderr, "Need size for create\n");
1258 td_verror(td, EINVAL);
1259 return 1;
1260 }
1261
1262 if (!extend) {
1263 oflags = O_CREAT | O_TRUNC;
1264 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1265 } else {
1266 oflags = O_APPEND;
1267 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1268 }
1269
1270 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1271 if (td->fd < 0) {
1272 td_verror(td, errno);
1273 return 1;
1274 }
1275
1276 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1277 td_verror(td, errno);
1278 return 1;
1279 }
1280
1281 td->io_size = td->file_size;
1282 b = malloc(td->max_bs);
1283 memset(b, 0, td->max_bs);
1284
1285 left = size;
1286 while (left && !td->terminate) {
1287 bs = td->max_bs;
1288 if (bs > left)
1289 bs = left;
1290
1291 r = write(td->fd, b, bs);
1292
1293 if (r == (int) bs) {
1294 left -= bs;
1295 continue;
1296 } else {
1297 if (r < 0)
1298 td_verror(td, errno);
1299 else
1300 td_verror(td, EIO);
1301
1302 break;
1303 }
1304 }
1305
1306 if (td->terminate)
1307 unlink(td->file_name);
1308 else if (td->create_fsync)
1309 fsync(td->fd);
1310
1311 close(td->fd);
1312 td->fd = -1;
1313 free(b);
1314 return 0;
1315}
1316
1317static int file_size(struct thread_data *td)
1318{
1319 struct stat st;
1320
1321 if (fstat(td->fd, &st) == -1) {
1322 td_verror(td, errno);
1323 return 1;
1324 }
1325
838a3cd3
JA		 1326 td->real_file_size = st.st_size;
1327
1328 if (!td->file_size || td->file_size > td->real_file_size)
1329 td->file_size = td->real_file_size;
ebac4655 1330
e8e387c1 1331 td->file_size -= td->file_offset;
ebac4655
JA		 1332 return 0;
1333}
1334
1335static int bdev_size(struct thread_data *td)
1336{
9104f874 1337 unsigned long long bytes;
ebac4655
JA		 1338 int r;
1339
1340 r = blockdev_size(td->fd, &bytes);
1341 if (r) {
1342 td_verror(td, r);
1343 return 1;
1344 }
1345
838a3cd3
JA		 1346 td->real_file_size = bytes;
1347
ebac4655
JA		 1348 /*
1349 * no extend possibilities, so limit size to device size if too large
1350 */
838a3cd3
JA		 1351 if (!td->file_size || td->file_size > td->real_file_size)
1352 td->file_size = td->real_file_size;
ebac4655 1353
e8e387c1 1354 td->file_size -= td->file_offset;
ebac4655
JA		 1355 return 0;
1356}
1357
1358static int get_file_size(struct thread_data *td)
1359{
0af7b542 1360 int ret = 0;
ebac4655
JA		 1361
1362 if (td->filetype == FIO_TYPE_FILE)
1363 ret = file_size(td);
0af7b542 1364 else if (td->filetype == FIO_TYPE_BD)
ebac4655 1365 ret = bdev_size(td);
0af7b542
JA		 1366 else
1367 td->real_file_size = -1;
ebac4655
JA		 1368
1369 if (ret)
1370 return ret;
1371
e8e387c1
JA		 1372 if (td->file_offset > td->real_file_size) {
1373 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
ebac4655
JA		 1374 return 1;
1375 }
1376
838a3cd3 1377 td->io_size = td->file_size;
ebac4655
JA		 1378 if (td->io_size == 0) {
1379 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1380 td_verror(td, EINVAL);
1381 return 1;
1382 }
1383
20dc95c4
JA		 1384 if (!td->zone_size)
1385 td->zone_size = td->io_size;
1386
ebac4655
JA		 1387 td->total_io_size = td->io_size * td->loops;
1388 return 0;
1389}
1390
1391static int setup_file_mmap(struct thread_data *td)
1392{
1393 int flags;
1394
3d60d1ed
JA		 1395 if (td_rw(td))
1396 flags = PROT_READ | PROT_WRITE;
1397 else if (td_write(td)) {
ebac4655
JA		 1398 flags = PROT_WRITE;
1399
1400 if (td->verify != VERIFY_NONE)
1401 flags |= PROT_READ;
3d60d1ed
JA		 1402 } else
1403 flags = PROT_READ;
ebac4655
JA		 1404
1405 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1406 if (td->mmap == MAP_FAILED) {
1407 td->mmap = NULL;
1408 td_verror(td, errno);
1409 return 1;
1410 }
1411
1412 if (td->invalidate_cache) {
1413 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1414 td_verror(td, errno);
1415 return 1;
1416 }
1417 }
1418
1419 if (td->sequential) {
1420 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1421 td_verror(td, errno);
1422 return 1;
1423 }
1424 } else {
1425 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1426 td_verror(td, errno);
1427 return 1;
1428 }
1429 }
1430
1431 return 0;
1432}
1433
1434static int setup_file_plain(struct thread_data *td)
1435{
1436 if (td->invalidate_cache) {
1437 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1438 td_verror(td, errno);
1439 return 1;
1440 }
1441 }
1442
1443 if (td->sequential) {
1444 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1445 td_verror(td, errno);
1446 return 1;
1447 }
1448 } else {
1449 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1450 td_verror(td, errno);
1451 return 1;
1452 }
1453 }
1454
1455 return 0;
1456}
1457
1458static int setup_file(struct thread_data *td)
1459{
1460 struct stat st;
1461 int flags = 0;
1462
1463 if (stat(td->file_name, &st) == -1) {
1464 if (errno != ENOENT) {
1465 td_verror(td, errno);
1466 return 1;
1467 }
1468 if (!td->create_file) {
1469 td_verror(td, ENOENT);
1470 return 1;
1471 }
1472 if (create_file(td, td->file_size, 0))
1473 return 1;
1474 } else if (td->filetype == FIO_TYPE_FILE) {
6a0106a0 1475 if (st.st_size < (off_t) td->file_size) {
ebac4655
JA		 1476 if (create_file(td, td->file_size - st.st_size, 1))
1477 return 1;
1478 }
1479 }
1480
1481 if (td->odirect)
1482 flags |= O_DIRECT;
1483
3d60d1ed 1484 if (td_write(td) || td_rw(td)) {
ebac4655
JA		 1485 if (td->filetype == FIO_TYPE_FILE) {
1486 if (!td->overwrite)
1487 flags |= O_TRUNC;
1488
1489 flags |= O_CREAT;
1490 }
1491 if (td->sync_io)
1492 flags |= O_SYNC;
1493
1494 flags |= O_RDWR;
1495
1496 td->fd = open(td->file_name, flags, 0600);
3d60d1ed
JA		 1497 } else {
1498 if (td->filetype == FIO_TYPE_CHAR)
1499 flags |= O_RDWR;
1500 else
1501 flags |= O_RDONLY;
1502
1503 td->fd = open(td->file_name, flags);
ebac4655
JA		 1504 }
1505
1506 if (td->fd == -1) {
1507 td_verror(td, errno);
1508 return 1;
1509 }
1510
1511 if (get_file_size(td))
1512 return 1;
1513
1514 if (td->io_engine != FIO_MMAPIO)
1515 return setup_file_plain(td);
1516 else
1517 return setup_file_mmap(td);
1518}
1519
1520static int check_dev_match(dev_t dev, char *path)
1521{
1522 unsigned int major, minor;
1523 char line[256], *p;
1524 FILE *f;
1525
1526 f = fopen(path, "r");
1527 if (!f) {
1528 perror("open path");
1529 return 1;
1530 }
1531
1532 p = fgets(line, sizeof(line), f);
1533 if (!p) {
1534 fclose(f);
1535 return 1;
1536 }
1537
1538 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1539 fclose(f);
1540 return 1;
1541 }
1542
1543 if (((major << 8) | minor) == dev) {
1544 fclose(f);
1545 return 0;
1546 }
1547
1548 fclose(f);
1549 return 1;
1550}
1551
1552static int find_block_dir(dev_t dev, char *path)
1553{
1554 struct dirent *dir;
1555 struct stat st;
1556 int found = 0;
1557 DIR *D;
1558
1559 D = opendir(path);
1560 if (!D)
1561 return 0;
1562
1563 while ((dir = readdir(D)) != NULL) {
1564 char full_path[256];
1565
1566 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1567 continue;
1568 if (!strcmp(dir->d_name, "device"))
1569 continue;
1570
1571 sprintf(full_path, "%s/%s", path, dir->d_name);
1572
1573 if (!strcmp(dir->d_name, "dev")) {
1574 if (!check_dev_match(dev, full_path)) {
1575 found = 1;
1576 break;
1577 }
1578 }
1579
1580 if (stat(full_path, &st) == -1) {
1581 perror("stat");
1582 break;
1583 }
1584
1585 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1586 continue;
1587
1588 found = find_block_dir(dev, full_path);
1589 if (found) {
1590 strcpy(path, full_path);
1591 break;
1592 }
1593 }
1594
1595 closedir(D);
1596 return found;
1597}
1598
1599static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1600{
1601 unsigned in_flight;
1602 char line[256];
1603 FILE *f;
1604 char *p;
1605
1606 f = fopen(du->path, "r");
1607 if (!f)
1608 return 1;
1609
1610 p = fgets(line, sizeof(line), f);
1611 if (!p) {
1612 fclose(f);
1613 return 1;
1614 }
1615
1616 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) {
1617 fclose(f);
1618 return 1;
1619 }
1620
1621 fclose(f);
1622 return 0;
1623}
1624
1625static void update_io_tick_disk(struct disk_util *du)
1626{
1627 struct disk_util_stat __dus, *dus, *ldus;
1628 struct timeval t;
1629
1630 if (get_io_ticks(du, &__dus))
1631 return;
1632
1633 dus = &du->dus;
1634 ldus = &du->last_dus;
1635
1636 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1637 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1638 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1639 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1640 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1641 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1642 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1643 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1644 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1645 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1646
1647 gettimeofday(&t, NULL);
1648 du->msec += mtime_since(&du->time, &t);
1649 memcpy(&du->time, &t, sizeof(t));
1650 memcpy(ldus, &__dus, sizeof(__dus));
1651}
1652
1653static void update_io_ticks(void)
1654{
1655 struct list_head *entry;
1656 struct disk_util *du;
1657
1658 list_for_each(entry, &disk_list) {
1659 du = list_entry(entry, struct disk_util, list);
1660 update_io_tick_disk(du);
1661 }
1662}
1663
1664static int disk_util_exists(dev_t dev)
1665{
1666 struct list_head *entry;
1667 struct disk_util *du;
1668
1669 list_for_each(entry, &disk_list) {
1670 du = list_entry(entry, struct disk_util, list);
1671
1672 if (du->dev == dev)
1673 return 1;
1674 }
1675
1676 return 0;
1677}
1678
1679static void disk_util_add(dev_t dev, char *path)
1680{
1681 struct disk_util *du = malloc(sizeof(*du));
1682
1683 memset(du, 0, sizeof(*du));
1684 INIT_LIST_HEAD(&du->list);
1685 sprintf(du->path, "%s/stat", path);
1686 du->name = strdup(basename(path));
1687 du->dev = dev;
1688
1689 gettimeofday(&du->time, NULL);
1690 get_io_ticks(du, &du->last_dus);
1691
1692 list_add_tail(&du->list, &disk_list);
1693}
1694
1695static void init_disk_util(struct thread_data *td)
1696{
1697 struct stat st;
1698 char foo[256], tmp[256];
1699 dev_t dev;
1700 char *p;
1701
1702 if (!td->do_disk_util)
1703 return;
1704
1705 if (!stat(td->file_name, &st)) {
1706 if (S_ISBLK(st.st_mode))
1707 dev = st.st_rdev;
1708 else
1709 dev = st.st_dev;
1710 } else {
1711 /*
1712 * must be a file, open "." in that path
1713 */
1714 strcpy(foo, td->file_name);
1715 p = dirname(foo);
1716 if (stat(p, &st)) {
1717 perror("disk util stat");
1718 return;
1719 }
1720
1721 dev = st.st_dev;
1722 }
1723
1724 if (disk_util_exists(dev))
1725 return;
1726
1727 sprintf(foo, "/sys/block");
1728 if (!find_block_dir(dev, foo))
1729 return;
1730
1731 /*
d07f439e
JA		 1732 * If there's a ../queue/ directory there, we are inside a partition.
1733 * Check if that is the case and jump back. For loop/md/dm etc we
1734 * are already in the right spot.
ebac4655 1735 */
d07f439e
JA		 1736 sprintf(tmp, "%s/../queue", foo);
1737 if (!stat(tmp, &st)) {
1738 p = dirname(foo);
1739 sprintf(tmp, "%s/queue", p);
ebac4655 1740 if (stat(tmp, &st)) {
d07f439e
JA		 1741 fprintf(stderr, "unknown sysfs layout\n");
1742 return;
ebac4655 1743 }
d07f439e 1744 sprintf(foo, "%s", p);
ebac4655
JA		 1745 }
1746
1747 disk_util_add(dev, foo);
1748}
1749
1750static void disk_util_timer_arm(void)
1751{
1752 itimer.it_value.tv_sec = 0;
1753 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1754 setitimer(ITIMER_REAL, &itimer, NULL);
1755}
1756
1757static void clear_io_state(struct thread_data *td)
1758{
1759 if (td->io_engine == FIO_SYNCIO)
1760 lseek(td->fd, SEEK_SET, 0);
1761
20dc95c4 1762 td->last_pos = 0;
ebac4655
JA		 1763 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1764 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 1765 td->zone_bytes = 0;
ebac4655
JA		 1766
1767 if (td->file_map)
1768 memset(td->file_map, 0, td->num_maps * sizeof(long));
1769}
1770
1771static void update_rusage_stat(struct thread_data *td)
1772{
1773 if (!(td->runtime[0] + td->runtime[1]))
1774 return;
1775
1776 getrusage(RUSAGE_SELF, &td->ru_end);
1777
1778 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1779 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1780 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1781
1782
1783 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1784}
1785
1786static void *thread_main(void *data)
1787{
1788 struct thread_data *td = data;
1789 int ret = 1;
1790
1791 if (!td->use_thread)
1792 setsid();
1793
1794 td->pid = getpid();
1795
aea47d44
JA		 1796 INIT_LIST_HEAD(&td->io_u_freelist);
1797 INIT_LIST_HEAD(&td->io_u_busylist);
1798 INIT_LIST_HEAD(&td->io_hist_list);
1799 INIT_LIST_HEAD(&td->io_log_list);
1800
ebac4655
JA		 1801 if (init_io_u(td))
1802 goto err;
1803
1804 if (fio_setaffinity(td) == -1) {
1805 td_verror(td, errno);
1806 goto err;
1807 }
1808
1809 if (init_io(td))
1810 goto err;
1811
aea47d44
JA		 1812 if (init_iolog(td))
1813 goto err;
1814
ebac4655
JA		 1815 if (td->ioprio) {
1816 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1817 td_verror(td, errno);
1818 goto err;
1819 }
1820 }
1821
75154845
JA		 1822 if (init_random_state(td))
1823 goto err;
1824
1825 td_set_runstate(td, TD_INITIALIZED);
ebac4655
JA		 1826 sem_post(&startup_sem);
1827 sem_wait(&td->mutex);
1828
1829 if (!td->create_serialize && setup_file(td))
1830 goto err;
1831
ebac4655
JA		 1832 gettimeofday(&td->epoch, NULL);
1833
1834 while (td->loops--) {
1835 getrusage(RUSAGE_SELF, &td->ru_start);
1836 gettimeofday(&td->start, NULL);
1837 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1838
1839 if (td->ratemin)
1840 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1841
1842 clear_io_state(td);
1843 prune_io_piece_log(td);
1844
1845 do_io(td);
1846
1847 td->runtime[td->ddir] += mtime_since_now(&td->start);
aea47d44 1848 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
3d60d1ed
JA		 1849 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1850
ebac4655
JA		 1851 update_rusage_stat(td);
1852
1853 if (td->error || td->terminate)
1854 break;
1855
1856 if (td->verify == VERIFY_NONE)
1857 continue;
1858
1859 clear_io_state(td);
1860 gettimeofday(&td->start, NULL);
1861
1862 do_verify(td);
1863
1864 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1865
1866 if (td->error || td->terminate)
1867 break;
1868 }
1869
1870 ret = 0;
1871
1872 if (td->bw_log)
1873 finish_log(td, td->bw_log, "bw");
1874 if (td->slat_log)
1875 finish_log(td, td->slat_log, "slat");
1876 if (td->clat_log)
1877 finish_log(td, td->clat_log, "clat");
1878
1879 if (exitall_on_terminate)
1880 terminate_threads(td->groupid);
1881
1882err:
1883 if (td->fd != -1) {
1884 close(td->fd);
1885 td->fd = -1;
1886 }
1887 if (td->mmap)
1888 munmap(td->mmap, td->file_size);
1889 cleanup_io(td);
1890 cleanup_io_u(td);
1891 if (ret) {
1892 sem_post(&startup_sem);
1893 sem_wait(&td->mutex);
1894 }
1895 td_set_runstate(td, TD_EXITED);
1896 return NULL;
1897
1898}
1899
1900static void *fork_main(int shmid, int offset)
1901{
1902 struct thread_data *td;
1903 void *data;
1904
1905 data = shmat(shmid, NULL, 0);
1906 if (data == (void *) -1) {
1907 perror("shmat");
1908 return NULL;
1909 }
1910
1911 td = data + offset * sizeof(struct thread_data);
1912 thread_main(td);
1913 shmdt(data);
1914 return NULL;
1915}
1916
1917static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1918 double *mean, double *dev)
1919{
1920 double n;
1921
1922 if (is->samples == 0)
1923 return 0;
1924
1925 *min = is->min_val;
1926 *max = is->max_val;
1927
1928 n = (double) is->samples;
1929 *mean = (double) is->val / n;
1930 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1931 if (!(*min + *max) && !(*mean + *dev))
1932 return 0;
1933
1934 return 1;
1935}
1936
1937static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1938 int ddir)
1939{
1940 char *ddir_str[] = { "read ", "write" };
200bc855
JA		 1941 unsigned long min, max;
1942 unsigned long long bw;
ebac4655
JA		 1943 double mean, dev;
1944
1945 if (!td->runtime[ddir])
1946 return;
1947
1948 bw = td->io_bytes[ddir] / td->runtime[ddir];
200bc855 1949 printf(" %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
ebac4655
JA		 1950
1951 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1952 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1953
1954 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1955 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1956
1957 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1958 double p_of_agg;
1959
1960 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1961 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);
1962 }
1963}
1964
1965static void show_thread_status(struct thread_data *td,
1966 struct group_run_stats *rs)
1967{
1968 double usr_cpu, sys_cpu;
1969
1970 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1971 return;
1972
1973 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1974
1975 show_ddir_status(td, rs, td->ddir);
aea47d44
JA		 1976 if (td->io_bytes[td->ddir ^ 1])
1977 show_ddir_status(td, rs, td->ddir ^ 1);
ebac4655
JA		 1978
1979 if (td->runtime[0] + td->runtime[1]) {
1980 double runt = td->runtime[0] + td->runtime[1];
1981
1982 usr_cpu = (double) td->usr_time * 100 / runt;
1983 sys_cpu = (double) td->sys_time * 100 / runt;
1984 } else {
1985 usr_cpu = 0;
1986 sys_cpu = 0;
1987 }
1988
1989 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
1990}
1991
1992static void check_str_update(struct thread_data *td)
1993{
1994 char c = run_str[td->thread_number - 1];
1995
1996 if (td->runstate == td->old_runstate)
1997 return;
1998
1999 switch (td->runstate) {
2000 case TD_REAPED:
2001 c = '_';
2002 break;
2003 case TD_EXITED:
2004 c = 'E';
2005 break;
2006 case TD_RUNNING:
3d60d1ed
JA		 2007 if (td_rw(td)) {
2008 if (td->sequential)
2009 c = 'M';
2010 else
2011 c = 'm';
2012 } else if (td_read(td)) {
ebac4655
JA		 2013 if (td->sequential)
2014 c = 'R';
2015 else
2016 c = 'r';
2017 } else {
2018 if (td->sequential)
2019 c = 'W';
2020 else
2021 c = 'w';
2022 }
2023 break;
2024 case TD_VERIFYING:
2025 c = 'V';
2026 break;
2027 case TD_CREATED:
2028 c = 'C';
2029 break;
75154845
JA		 2030 case TD_INITIALIZED:
2031 c = 'I';
2032 break;
ebac4655
JA		 2033 case TD_NOT_CREATED:
2034 c = 'P';
2035 break;
2036 default:
2037 printf("state %d\n", td->runstate);
2038 }
2039
2040 run_str[td->thread_number - 1] = c;
2041 td->old_runstate = td->runstate;
2042}
2043
5289b847
JA		 2044static void eta_to_str(char *str, int eta_sec)
2045{
2046 unsigned int d, h, m, s;
2047 static int always_d, always_h;
2048
2049 d = h = m = s = 0;
2050
2051 s = eta_sec % 60;
2052 eta_sec /= 60;
2053 m = eta_sec % 60;
2054 eta_sec /= 60;
2055 h = eta_sec % 24;
2056 eta_sec /= 24;
2057 d = eta_sec;
2058
2059 if (d || always_d) {
2060 always_d = 1;
2061 str += sprintf(str, "%02dd:", d);
2062 }
2063 if (h || always_h) {
2064 always_h = 1;
2065 str += sprintf(str, "%02dh:", h);
2066 }
2067
2068 str += sprintf(str, "%02dm:", m);
2069 str += sprintf(str, "%02ds", s);
2070}
2071
6a0106a0
JA		 2072static int thread_eta(struct thread_data *td, unsigned long elapsed)
2073{
2074 unsigned long long bytes_total, bytes_done;
2075 unsigned int eta_sec = 0;
2076
2077 bytes_total = td->total_io_size;
3d60d1ed
JA		 2078
2079 /*
2080 * if writing, bytes_total will be twice the size. If mixing,
2081 * assume a 50/50 split and thus bytes_total will be 50% larger.
2082 */
2083 if (td->verify) {
2084 if (td_rw(td))
2085 bytes_total = bytes_total * 3 / 2;
2086 else
2087 bytes_total <<= 1;
2088 }
6a0106a0
JA		 2089 if (td->zone_size && td->zone_skip)
2090 bytes_total /= (td->zone_skip / td->zone_size);
2091
2092 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
2093 double perc;
8b611c34 2094
6a0106a0
JA		 2095 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
2096 perc = (double) bytes_done / (double) bytes_total;
2097 if (perc > 1.0)
2098 perc = 1.0;
2099
2100 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
2101
8b611c34 2102 if (td->timeout && eta_sec > (td->timeout - elapsed))
6a0106a0 2103 eta_sec = td->timeout - elapsed;
75154845
JA		 2104 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
2105 || td->runstate == TD_INITIALIZED) {
6a0106a0
JA		 2106 int t_eta = 0, r_eta = 0;
2107
2108 /*
2109 * We can only guess - assume it'll run the full timeout
2110 * if given, otherwise assume it'll run at the specified rate.
2111 */
2112 if (td->timeout)
2113 t_eta = td->timeout + td->start_delay - elapsed;
2114 if (td->rate) {
2115 r_eta = (bytes_total / 1024) / td->rate;
2116 r_eta += td->start_delay - elapsed;
2117 }
2118
2119 if (r_eta && t_eta)
2120 eta_sec = min(r_eta, t_eta);
2121 else if (r_eta)
2122 eta_sec = r_eta;
2123 else if (t_eta)
2124 eta_sec = t_eta;
2125 else
2126 eta_sec = INT_MAX;
2127 } else {
2128 /*
2129 * thread is already done
2130 */
2131 eta_sec = 0;
2132 }
2133
2134 return eta_sec;
2135}
2136
ebac4655
JA		 2137static void print_thread_status(void)
2138{
6a0106a0
JA		 2139 unsigned long elapsed = time_since_now(&genesis);
2140 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
5289b847 2141 char eta_str[32];
6a0106a0
JA		 2142 double perc = 0.0;
2143
2144 eta_secs = malloc(thread_number * sizeof(int));
2145 memset(eta_secs, 0, thread_number * sizeof(int));
ebac4655 2146
ebac4655
JA		 2147 nr_running = t_rate = m_rate = 0;
2148 for (i = 0; i < thread_number; i++) {
2149 struct thread_data *td = &threads[i];
2150
2151 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
2152 nr_running++;
2153 t_rate += td->rate;
2154 m_rate += td->ratemin;
2155 }
2156
6a0106a0
JA		 2157 if (elapsed >= 3)
2158 eta_secs[i] = thread_eta(td, elapsed);
8b611c34
JA		 2159 else
2160 eta_secs[i] = INT_MAX;
ebac4655
JA		 2161
2162 check_str_update(td);
2163 }
2164
6a0106a0
JA		 2165 if (exitall_on_terminate)
2166 eta_sec = INT_MAX;
2167 else
2168 eta_sec = 0;
5289b847 2169
6a0106a0
JA		 2170 for (i = 0; i < thread_number; i++) {
2171 if (exitall_on_terminate) {
2172 if (eta_secs[i] < eta_sec)
2173 eta_sec = eta_secs[i];
2174 } else {
2175 if (eta_secs[i] > eta_sec)
2176 eta_sec = eta_secs[i];
5289b847 2177 }
6a0106a0 2178 }
5289b847 2179
8b611c34 2180 if (eta_sec != INT_MAX && elapsed) {
6a0106a0
JA		 2181 perc = (double) elapsed / (double) (elapsed + eta_sec);
2182 eta_to_str(eta_str, eta_sec);
ebac4655
JA		 2183 }
2184
5289b847 2185 printf("Threads now running (%d)", nr_running);
ebac4655
JA		 2186 if (m_rate || t_rate)
2187 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
8b611c34 2188 if (eta_sec != INT_MAX) {
6a0106a0
JA		 2189 perc *= 100.0;
2190 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
2191 }
5289b847 2192 printf("\r");
ebac4655 2193 fflush(stdout);
6a0106a0 2194 free(eta_secs);
ebac4655
JA		 2195}
2196
2197static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2198{
2199 int i;
2200
2201 /*
2202 * reap exited threads (TD_EXITED -> TD_REAPED)
2203 */
2204 for (i = 0; i < thread_number; i++) {
2205 struct thread_data *td = &threads[i];
2206
2207 if (td->runstate != TD_EXITED)
2208 continue;
2209
2210 td_set_runstate(td, TD_REAPED);
2211
2212 if (td->use_thread) {
2213 long ret;
2214
2215 if (pthread_join(td->thread, (void *) &ret))
2216 perror("thread_join");
2217 } else
2218 waitpid(td->pid, NULL, 0);
2219
2220 (*nr_running)--;
2221 (*m_rate) -= td->ratemin;
2222 (*t_rate) -= td->rate;
2223 }
2224}
2225
fcb6ade2
JA		 2226static void fio_unpin_memory(void *pinned)
2227{
2228 if (pinned) {
2229 if (munlock(pinned, mlock_size) < 0)
2230 perror("munlock");
2231 munmap(pinned, mlock_size);
2232 }
2233}
2234
2235static void *fio_pin_memory(void)
2236{
2237 long pagesize, pages;
2238 void *ptr;
2239
2240 if (!mlock_size)
2241 return NULL;
2242
2243 /*
2244 * Don't allow mlock of more than real_mem-128MB
2245 */
2246 pagesize = sysconf(_SC_PAGESIZE);
2247 pages = sysconf(_SC_PHYS_PAGES);
2248 if (pages != -1 && pagesize != -1) {
2249 unsigned long long real_mem = pages * pagesize;
2250
2251 if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
2252 mlock_size = real_mem - 128 * 1024 * 1024;
2253 printf("fio: limiting mlocked memory to %lluMiB\n",
2254 mlock_size >> 20);
2255 }
2256 }
2257
2258 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
2259 if (!ptr) {
2260 perror("malloc locked mem");
2261 return NULL;
2262 }
2263 if (mlock(ptr, mlock_size) < 0) {
2264 munmap(ptr, mlock_size);
2265 perror("mlock");
2266 return NULL;
2267 }
2268
2269 return ptr;
2270}
2271
ebac4655
JA		 2272static void run_threads(void)
2273{
ebac4655
JA		 2274 struct thread_data *td;
2275 unsigned long spent;
2276 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2
JA		 2277 void *mlocked_mem;
2278
2279 mlocked_mem = fio_pin_memory();
ebac4655
JA		 2280
2281 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
fcb6ade2 2282 fflush(stdout);
c04f7ec3 2283
4efa970e
JA		 2284 signal(SIGINT, sig_handler);
2285 signal(SIGALRM, sig_handler);
2286
ebac4655
JA		 2287 todo = thread_number;
2288 nr_running = 0;
2289 nr_started = 0;
2290 m_rate = t_rate = 0;
2291
2292 for (i = 0; i < thread_number; i++) {
2293 td = &threads[i];
2294
2295 run_str[td->thread_number - 1] = 'P';
2296
2297 init_disk_util(td);
2298
2299 if (!td->create_serialize)
2300 continue;
2301
2302 /*
2303 * do file setup here so it happens sequentially,
2304 * we don't want X number of threads getting their
2305 * client data interspersed on disk
2306 */
2307 if (setup_file(td)) {
2308 td_set_runstate(td, TD_REAPED);
2309 todo--;
2310 }
2311 }
2312
2313 gettimeofday(&genesis, NULL);
2314
2315 while (todo) {
75154845
JA		 2316 struct thread_data *map[MAX_JOBS];
2317 struct timeval this_start;
2318 int this_jobs = 0, left;
2319
ebac4655
JA		 2320 /*
2321 * create threads (TD_NOT_CREATED -> TD_CREATED)
2322 */
2323 for (i = 0; i < thread_number; i++) {
2324 td = &threads[i];
2325
2326 if (td->runstate != TD_NOT_CREATED)
2327 continue;
2328
2329 /*
2330 * never got a chance to start, killed by other
2331 * thread for some reason
2332 */
2333 if (td->terminate) {
2334 todo--;
2335 continue;
2336 }
2337
2338 if (td->start_delay) {
2339 spent = mtime_since_now(&genesis);
2340
2341 if (td->start_delay * 1000 > spent)
2342 continue;
2343 }
2344
2345 if (td->stonewall && (nr_started || nr_running))
2346 break;
2347
75154845
JA		 2348 /*
2349 * Set state to created. Thread will transition
2350 * to TD_INITIALIZED when it's done setting up.
2351 */
ebac4655 2352 td_set_runstate(td, TD_CREATED);
75154845 2353 map[this_jobs++] = td;
ebac4655 2354 sem_init(&startup_sem, 0, 1);
ebac4655
JA		 2355 nr_started++;
2356
2357 if (td->use_thread) {
2358 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2359 perror("thread_create");
2360 nr_started--;
2361 }
2362 } else {
2363 if (fork())
2364 sem_wait(&startup_sem);
2365 else {
2366 fork_main(shm_id, i);
2367 exit(0);
2368 }
2369 }
2370 }
2371
2372 /*
75154845
JA		 2373 * Wait for the started threads to transition to
2374 * TD_INITIALIZED.
ebac4655 2375 */
75154845
JA		 2376 printf("fio: Waiting for threads to initialize...\n");
2377 gettimeofday(&this_start, NULL);
2378 left = this_jobs;
2379 while (left) {
2380 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2381 break;
2382
2383 usleep(100000);
2384
2385 for (i = 0; i < this_jobs; i++) {
2386 td = map[i];
2387 if (!td)
2388 continue;
2389 if (td->runstate == TD_INITIALIZED ||
2390 td->runstate >= TD_EXITED) {
2391 map[i] = NULL;
2392 left--;
2393 continue;
2394 }
2395 }
2396 }
2397
2398 if (left) {
2399 fprintf(stderr, "fio: %d jobs failed to start\n", left);
2400 for (i = 0; i < this_jobs; i++) {
2401 td = map[i];
2402 if (!td)
2403 continue;
2404 kill(td->pid, SIGTERM);
2405 }
2406 break;
2407 }
2408
2409 /*
2410 * start created threads (TD_INITIALIZED -> TD_RUNNING)
2411 */
2412 printf("fio: Go for launch\n");
ebac4655
JA		 2413 for (i = 0; i < thread_number; i++) {
2414 td = &threads[i];
2415
75154845 2416 if (td->runstate != TD_INITIALIZED)
ebac4655
JA		 2417 continue;
2418
2419 td_set_runstate(td, TD_RUNNING);
2420 nr_running++;
2421 nr_started--;
2422 m_rate += td->ratemin;
2423 t_rate += td->rate;
75154845 2424 todo--;
ebac4655
JA		 2425 sem_post(&td->mutex);
2426 }
2427
2428 reap_threads(&nr_running, &t_rate, &m_rate);
2429
2430 if (todo)
2431 usleep(100000);
2432 }
2433
2434 while (nr_running) {
2435 reap_threads(&nr_running, &t_rate, &m_rate);
2436 usleep(10000);
2437 }
2438
2439 update_io_ticks();
fcb6ade2 2440 fio_unpin_memory(mlocked_mem);
ebac4655
JA		 2441}
2442
2443static void show_group_stats(struct group_run_stats *rs, int id)
2444{
2445 printf("\nRun status group %d (all jobs):\n", id);
2446
2447 if (rs->max_run[DDIR_READ])
9104f874 2448 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 2449 if (rs->max_run[DDIR_WRITE])
9104f874 2450 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		 2451}
2452
2453static void show_disk_util(void)
2454{
2455 struct disk_util_stat *dus;
2456 struct list_head *entry;
2457 struct disk_util *du;
2458 double util;
2459
2460 printf("\nDisk stats (read/write):\n");
2461
2462 list_for_each(entry, &disk_list) {
2463 du = list_entry(entry, struct disk_util, list);
2464 dus = &du->dus;
2465
2466 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2467 if (util > 100.0)
2468 util = 100.0;
2469
2470 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);
2471 }
2472}
2473
2474static void show_run_stats(void)
2475{
2476 struct group_run_stats *runstats, *rs;
2477 struct thread_data *td;
2478 int i;
2479
2480 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2481
2482 for (i = 0; i < groupid + 1; i++) {
2483 rs = &runstats[i];
2484
2485 memset(rs, 0, sizeof(*rs));
2486 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2487 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2488 }
2489
2490 for (i = 0; i < thread_number; i++) {
200bc855 2491 unsigned long long rbw, wbw;
ebac4655
JA		 2492
2493 td = &threads[i];
2494
2495 if (td->error) {
2496 printf("Client%d: %s\n", td->thread_number, td->verror);
2497 continue;
2498 }
2499
2500 rs = &runstats[td->groupid];
2501
2502 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2503 rs->min_run[0] = td->runtime[0];
2504 if (td->runtime[0] > rs->max_run[0])
2505 rs->max_run[0] = td->runtime[0];
2506 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2507 rs->min_run[1] = td->runtime[1];
2508 if (td->runtime[1] > rs->max_run[1])
2509 rs->max_run[1] = td->runtime[1];
2510
2511 rbw = wbw = 0;
2512 if (td->runtime[0])
200bc855 2513 rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
ebac4655 2514 if (td->runtime[1])
200bc855 2515 wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
ebac4655
JA		 2516
2517 if (rbw < rs->min_bw[0])
2518 rs->min_bw[0] = rbw;
2519 if (wbw < rs->min_bw[1])
2520 rs->min_bw[1] = wbw;
2521 if (rbw > rs->max_bw[0])
2522 rs->max_bw[0] = rbw;
2523 if (wbw > rs->max_bw[1])
2524 rs->max_bw[1] = wbw;
2525
2526 rs->io_mb[0] += td->io_bytes[0] >> 20;
2527 rs->io_mb[1] += td->io_bytes[1] >> 20;
2528 }
2529
2530 for (i = 0; i < groupid + 1; i++) {
2531 rs = &runstats[i];
2532
2533 if (rs->max_run[0])
2534 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2535 if (rs->max_run[1])
2536 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2537 }
2538
2539 /*
2540 * don't overwrite last signal output
2541 */
2542 printf("\n");
2543
2544 for (i = 0; i < thread_number; i++) {
2545 td = &threads[i];
2546 rs = &runstats[td->groupid];
2547
2548 show_thread_status(td, rs);
2549 }
2550
2551 for (i = 0; i < groupid + 1; i++)
2552 show_group_stats(&runstats[i], i);
2553
2554 show_disk_util();
2555}
2556
2557int main(int argc, char *argv[])
2558{
2559 if (parse_options(argc, argv))
2560 return 1;
2561
2562 if (!thread_number) {
2563 printf("Nothing to do\n");
2564 return 1;
2565 }
2566
2567 disk_util_timer_arm();
2568
2569 run_threads();
2570 show_run_stats();
2571
2572 return 0;
2573}
fio - Flexible IO Tester