projects / fio.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
[PATCH] Introduce an extra runstate for monitoring thread startup
[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{
200 unsigned long block = io_u->offset / td->min_bs;
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) {
291 unsigned long max_blocks = td->io_size / td->min_bs;
292 int loops = 50;
293
294 do {
295 lrand48_r(&td->random_state, &r);
296 b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
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
3d60d1ed
JA		 548unsigned int hweight32(unsigned int w)
549{
550 unsigned int res = w - ((w >> 1) & 0x55555555);
551
552 res = (res & 0x33333333) + ((res >> 2) & 0x33333333);
553 res = (res + (res >> 4)) & 0x0F0F0F0F;
554 res = res + (res >> 8);
555
556 return (res + (res >> 16)) & 0x000000FF;
557}
558
559unsigned long hweight64(unsigned long long w)
560{
561#if __WORDSIZE == 32
562 return hweight32((unsigned int)(w >> 32)) + hweight32((unsigned int)w);
563#elif __WORDSIZE == 64
564 unsigned long long v = w - ((w >> 1) & 0x5555555555555555ul);
565
566 v = (v & 0x3333333333333333ul) + ((v >> 2) & 0x3333333333333333ul);
567 v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0Ful;
568 v = v + (v >> 8);
569 v = v + (v >> 16);
570
571 return (v + (v >> 32)) & 0x00000000000000FFul;
572#else
573#error __WORDSIZE not defined
574#endif
575}
576
577static int get_rw_ddir(struct thread_data *td)
578{
579 /*
580 * perhaps cheasy, but use the hamming weight of the position
581 * as a randomizer for data direction.
582 */
583 if (td_rw(td))
584 return hweight64(td->last_pos) & 1;
585 else if (td_read(td))
586 return DDIR_READ;
587 else
588 return DDIR_WRITE;
589}
590
ebac4655
JA		 591/*
592 * fill body of io_u->buf with random data and add a header with the
593 * (eg) sha1sum of that data.
594 */
595static void populate_io_u(struct thread_data *td, struct io_u *io_u)
596{
597 unsigned char *p = (unsigned char *) io_u->buf;
598 struct verify_header hdr;
599
600 hdr.fio_magic = FIO_HDR_MAGIC;
601 hdr.len = io_u->buflen;
602 p += sizeof(hdr);
603 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
604
605 if (td->verify == VERIFY_MD5) {
606 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
607 hdr.verify_type = VERIFY_MD5;
608 } else {
609 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
610 hdr.verify_type = VERIFY_CRC32;
611 }
612
613 memcpy(io_u->buf, &hdr, sizeof(hdr));
614}
615
aea47d44 616static int td_io_prep(struct thread_data *td, struct io_u *io_u)
20dc95c4 617{
20dc95c4
JA		 618 if (td->io_prep && td->io_prep(td, io_u))
619 return 1;
620
621 return 0;
622}
623
b1ff3403 624void put_io_u(struct thread_data *td, struct io_u *io_u)
ebac4655
JA		 625{
626 list_del(&io_u->list);
627 list_add(&io_u->list, &td->io_u_freelist);
628 td->cur_depth--;
629}
630
aea47d44
JA		 631static int fill_io_u(struct thread_data *td, struct io_u *io_u)
632{
633 /*
634 * If using an iolog, grab next piece if any available.
635 */
636 if (td->iolog) {
637 struct io_piece *ipo;
638
639 if (list_empty(&td->io_log_list))
640 return 1;
641
642 ipo = list_entry(td->io_log_list.next, struct io_piece, list);
643 list_del(&ipo->list);
644 io_u->offset = ipo->offset;
645 io_u->buflen = ipo->len;
646 io_u->ddir = ipo->ddir;
647 free(ipo);
648 return 0;
649 }
650
651 /*
652 * No log, let the seq/rand engine retrieve the next position.
653 */
654 if (!get_next_offset(td, &io_u->offset)) {
655 io_u->buflen = get_next_buflen(td);
656
657 if (io_u->buflen) {
658 io_u->ddir = get_rw_ddir(td);
659 return 0;
660 }
661 }
662
663 return 1;
664}
665
ebac4655
JA		 666#define queue_full(td) (list_empty(&(td)->io_u_freelist))
667
b1ff3403 668struct io_u *__get_io_u(struct thread_data *td)
ebac4655
JA		 669{
670 struct io_u *io_u;
671
672 if (queue_full(td))
673 return NULL;
674
675 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
676 io_u->error = 0;
677 io_u->resid = 0;
678 list_del(&io_u->list);
679 list_add(&io_u->list, &td->io_u_busylist);
680 td->cur_depth++;
681 return io_u;
682}
683
ebac4655
JA		 684static struct io_u *get_io_u(struct thread_data *td)
685{
686 struct io_u *io_u;
687
688 io_u = __get_io_u(td);
689 if (!io_u)
690 return NULL;
691
20dc95c4
JA		 692 if (td->zone_bytes >= td->zone_size) {
693 td->zone_bytes = 0;
694 td->last_pos += td->zone_skip;
695 }
696
aea47d44 697 if (fill_io_u(td, io_u)) {
ebac4655
JA		 698 put_io_u(td, io_u);
699 return NULL;
700 }
701
838a3cd3
JA		 702 if (io_u->buflen + io_u->offset > td->real_file_size)
703 io_u->buflen = td->real_file_size - io_u->offset;
ebac4655
JA		 704
705 if (!io_u->buflen) {
706 put_io_u(td, io_u);
707 return NULL;
708 }
709
aea47d44 710 if (!td->iolog && !td->sequential)
ebac4655
JA		 711 mark_random_map(td, io_u);
712
20dc95c4 713 td->last_pos += io_u->buflen;
ebac4655
JA		 714
715 if (td->verify != VERIFY_NONE)
716 populate_io_u(td, io_u);
717
aea47d44 718 if (td_io_prep(td, io_u)) {
ebac4655
JA		 719 put_io_u(td, io_u);
720 return NULL;
721 }
722
723 gettimeofday(&io_u->start_time, NULL);
724 return io_u;
725}
726
727static inline void td_set_runstate(struct thread_data *td, int runstate)
728{
729 td->old_runstate = td->runstate;
730 td->runstate = runstate;
731}
732
aea47d44 733static int get_next_verify(struct thread_data *td, struct io_u *io_u)
ebac4655
JA		 734{
735 struct io_piece *ipo;
736
737 if (list_empty(&td->io_hist_list))
738 return 1;
739
740 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
741 list_del(&ipo->list);
742
aea47d44
JA		 743 io_u->offset = ipo->offset;
744 io_u->buflen = ipo->len;
745 io_u->ddir = DDIR_READ;
ebac4655
JA		 746 free(ipo);
747 return 0;
748}
749
750static void prune_io_piece_log(struct thread_data *td)
751{
752 struct io_piece *ipo;
753
754 while (!list_empty(&td->io_hist_list)) {
755 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
756
757 list_del(&ipo->list);
758 free(ipo);
759 }
760}
761
762/*
763 * log a succesful write, so we can unwind the log for verify
764 */
765static void log_io_piece(struct thread_data *td, struct io_u *io_u)
766{
767 struct io_piece *ipo = malloc(sizeof(struct io_piece));
768 struct list_head *entry;
769
770 INIT_LIST_HEAD(&ipo->list);
771 ipo->offset = io_u->offset;
772 ipo->len = io_u->buflen;
773
774 /*
775 * for random io where the writes extend the file, it will typically
776 * be laid out with the block scattered as written. it's faster to
777 * read them in in that order again, so don't sort
778 */
779 if (td->sequential || !td->overwrite) {
780 list_add_tail(&ipo->list, &td->io_hist_list);
781 return;
782 }
783
784 /*
785 * for random io, sort the list so verify will run faster
786 */
787 entry = &td->io_hist_list;
788 while ((entry = entry->prev) != &td->io_hist_list) {
789 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
790
791 if (__ipo->offset < ipo->offset)
792 break;
793 }
794
795 list_add(&ipo->list, entry);
796}
797
aea47d44
JA		 798static int init_iolog(struct thread_data *td)
799{
800 unsigned long long offset;
801 unsigned int bytes;
802 char *str, *p;
803 FILE *f;
804 int rw, i, reads, writes;
805
806 if (!td->iolog)
807 return 0;
808
809 f = fopen(td->iolog_file, "r");
810 if (!f) {
811 perror("fopen iolog");
812 return 1;
813 }
814
815 str = malloc(4096);
816 reads = writes = i = 0;
817 while ((p = fgets(str, 4096, f)) != NULL) {
818 struct io_piece *ipo;
819
820 if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
821 fprintf(stderr, "bad iolog: %s\n", p);
822 continue;
823 }
824 if (rw == DDIR_READ)
825 reads++;
826 else if (rw == DDIR_WRITE)
827 writes++;
828 else {
829 fprintf(stderr, "bad ddir: %d\n", rw);
830 continue;
831 }
832
833 ipo = malloc(sizeof(*ipo));
834 INIT_LIST_HEAD(&ipo->list);
835 ipo->offset = offset;
836 ipo->len = bytes;
837 if (bytes > td->max_bs)
838 td->max_bs = bytes;
839 ipo->ddir = rw;
840 list_add_tail(&ipo->list, &td->io_log_list);
841 i++;
842 }
843
844 free(str);
845 fclose(f);
846
847 if (!i)
848 return 1;
849
850 if (reads && !writes)
851 td->ddir = DDIR_READ;
852 else if (!reads && writes)
853 td->ddir = DDIR_READ;
854 else
855 td->iomix = 1;
856
857 return 0;
858}
859
ebac4655
JA		 860static int sync_td(struct thread_data *td)
861{
862 if (td->io_sync)
863 return td->io_sync(td);
864
865 return 0;
866}
867
868static int io_u_getevents(struct thread_data *td, int min, int max,
869 struct timespec *t)
870{
871 return td->io_getevents(td, min, max, t);
872}
873
874static int io_u_queue(struct thread_data *td, struct io_u *io_u)
875{
876 gettimeofday(&io_u->issue_time, NULL);
877
878 return td->io_queue(td, io_u);
879}
880
881#define iocb_time(iocb) ((unsigned long) (iocb)->data)
882
883static void io_completed(struct thread_data *td, struct io_u *io_u,
884 struct io_completion_data *icd)
885{
886 struct timeval e;
887 unsigned long msec;
888
889 gettimeofday(&e, NULL);
890
891 if (!io_u->error) {
20dc95c4 892 unsigned int bytes = io_u->buflen - io_u->resid;
3d60d1ed 893 const int idx = io_u->ddir;
ebac4655
JA		 894
895 td->io_blocks[idx]++;
20dc95c4
JA		 896 td->io_bytes[idx] += bytes;
897 td->zone_bytes += bytes;
898 td->this_io_bytes[idx] += bytes;
ebac4655
JA		 899
900 msec = mtime_since(&io_u->issue_time, &e);
901
3d60d1ed
JA		 902 add_clat_sample(td, idx, msec);
903 add_bw_sample(td, idx);
ebac4655 904
3d60d1ed 905 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
ebac4655
JA		 906 log_io_piece(td, io_u);
907
20dc95c4 908 icd->bytes_done[idx] += bytes;
ebac4655
JA		 909 } else
910 icd->error = io_u->error;
911}
912
913static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
914{
915 struct io_u *io_u;
916 int i;
917
918 icd->error = 0;
919 icd->bytes_done[0] = icd->bytes_done[1] = 0;
920
921 for (i = 0; i < icd->nr; i++) {
922 io_u = td->io_event(td, i);
923
924 io_completed(td, io_u, icd);
925 put_io_u(td, io_u);
926 }
927}
928
929static void cleanup_pending_aio(struct thread_data *td)
930{
931 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
932 struct list_head *entry, *n;
933 struct io_completion_data icd;
934 struct io_u *io_u;
935 int r;
936
937 /*
938 * get immediately available events, if any
939 */
940 r = io_u_getevents(td, 0, td->cur_depth, &ts);
941 if (r > 0) {
942 icd.nr = r;
943 ios_completed(td, &icd);
944 }
945
946 /*
947 * now cancel remaining active events
948 */
949 if (td->io_cancel) {
950 list_for_each_safe(entry, n, &td->io_u_busylist) {
951 io_u = list_entry(entry, struct io_u, list);
952
953 r = td->io_cancel(td, io_u);
954 if (!r)
955 put_io_u(td, io_u);
956 }
957 }
958
959 if (td->cur_depth) {
960 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
961 if (r > 0) {
962 icd.nr = r;
963 ios_completed(td, &icd);
964 }
965 }
966}
967
968static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
969{
970 struct io_u *v_io_u = *io_u;
971 int ret = 0;
972
973 if (v_io_u) {
974 ret = verify_io_u(v_io_u);
975 put_io_u(td, v_io_u);
976 *io_u = NULL;
977 }
978
979 return ret;
980}
981
982static void do_verify(struct thread_data *td)
983{
984 struct timeval t;
985 struct io_u *io_u, *v_io_u = NULL;
986 struct io_completion_data icd;
987 int ret;
988
989 td_set_runstate(td, TD_VERIFYING);
990
991 do {
992 if (td->terminate)
993 break;
994
995 gettimeofday(&t, NULL);
996 if (runtime_exceeded(td, &t))
997 break;
998
999 io_u = __get_io_u(td);
1000 if (!io_u)
1001 break;
1002
aea47d44 1003 if (get_next_verify(td, io_u)) {
ebac4655
JA		 1004 put_io_u(td, io_u);
1005 break;
1006 }
1007
aea47d44 1008 if (td_io_prep(td, io_u)) {
ebac4655
JA		 1009 put_io_u(td, io_u);
1010 break;
1011 }
1012
1013 ret = io_u_queue(td, io_u);
1014 if (ret) {
1015 put_io_u(td, io_u);
1016 td_verror(td, ret);
1017 break;
1018 }
1019
1020 /*
1021 * we have one pending to verify, do that while
1022 * we are doing io on the next one
1023 */
1024 if (do_io_u_verify(td, &v_io_u))
1025 break;
1026
1027 ret = io_u_getevents(td, 1, 1, NULL);
1028 if (ret != 1) {
1029 if (ret < 0)
1030 td_verror(td, ret);
1031 break;
1032 }
1033
1034 v_io_u = td->io_event(td, 0);
1035 icd.nr = 1;
1036 icd.error = 0;
1037 io_completed(td, v_io_u, &icd);
1038
1039 if (icd.error) {
1040 td_verror(td, icd.error);
1041 put_io_u(td, v_io_u);
1042 v_io_u = NULL;
1043 break;
1044 }
1045
ebac4655
JA		 1046 /*
1047 * if we can't submit more io, we need to verify now
1048 */
1049 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
1050 break;
1051
1052 } while (1);
1053
1054 do_io_u_verify(td, &v_io_u);
1055
1056 if (td->cur_depth)
1057 cleanup_pending_aio(td);
1058
1059 td_set_runstate(td, TD_RUNNING);
1060}
1061
1062static void do_io(struct thread_data *td)
1063{
1064 struct io_completion_data icd;
1065 struct timeval s, e;
1066 unsigned long usec;
1067
1068 while (td->this_io_bytes[td->ddir] < td->io_size) {
1069 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
1070 struct timespec *timeout;
1071 int ret, min_evts = 0;
1072 struct io_u *io_u;
1073
1074 if (td->terminate)
1075 break;
1076
1077 io_u = get_io_u(td);
1078 if (!io_u)
1079 break;
1080
1081 memcpy(&s, &io_u->start_time, sizeof(s));
1082
1083 ret = io_u_queue(td, io_u);
1084 if (ret) {
1085 put_io_u(td, io_u);
1086 td_verror(td, ret);
1087 break;
1088 }
1089
1090 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
1091
1092 if (td->cur_depth < td->iodepth) {
1093 timeout = &ts;
1094 min_evts = 0;
1095 } else {
1096 timeout = NULL;
1097 min_evts = 1;
1098 }
1099
1100 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
1101 if (ret < 0) {
1102 td_verror(td, ret);
1103 break;
1104 } else if (!ret)
1105 continue;
1106
1107 icd.nr = ret;
1108 ios_completed(td, &icd);
1109 if (icd.error) {
1110 td_verror(td, icd.error);
1111 break;
1112 }
1113
1114 /*
1115 * the rate is batched for now, it should work for batches
1116 * of completions except the very first one which may look
1117 * a little bursty
1118 */
1119 gettimeofday(&e, NULL);
1120 usec = utime_since(&s, &e);
1121
1122 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
1123
1124 if (check_min_rate(td, &e)) {
1125 td_verror(td, ENOMEM);
1126 break;
1127 }
1128
1129 if (runtime_exceeded(td, &e))
1130 break;
1131
1132 if (td->thinktime)
1133 usec_sleep(td, td->thinktime);
1134
1135 if (should_fsync(td) && td->fsync_blocks &&
1136 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
1137 sync_td(td);
1138 }
1139
1140 if (td->cur_depth)
1141 cleanup_pending_aio(td);
1142
fc1a4713 1143 if (should_fsync(td) && td->end_fsync)
ebac4655
JA		 1144 sync_td(td);
1145}
1146
1147static void cleanup_io(struct thread_data *td)
1148{
1149 if (td->io_cleanup)
1150 td->io_cleanup(td);
1151}
1152
1153static int init_io(struct thread_data *td)
1154{
1155 if (td->io_engine == FIO_SYNCIO)
1156 return fio_syncio_init(td);
1157 else if (td->io_engine == FIO_MMAPIO)
1158 return fio_mmapio_init(td);
1159 else if (td->io_engine == FIO_LIBAIO)
1160 return fio_libaio_init(td);
1161 else if (td->io_engine == FIO_POSIXAIO)
1162 return fio_posixaio_init(td);
1163 else if (td->io_engine == FIO_SGIO)
1164 return fio_sgio_init(td);
8756e4d4
JA		 1165 else if (td->io_engine == FIO_SPLICEIO)
1166 return fio_spliceio_init(td);
ebac4655
JA		 1167 else {
1168 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1169 return 1;
1170 }
1171}
1172
1173static void cleanup_io_u(struct thread_data *td)
1174{
1175 struct list_head *entry, *n;
1176 struct io_u *io_u;
1177
1178 list_for_each_safe(entry, n, &td->io_u_freelist) {
1179 io_u = list_entry(entry, struct io_u, list);
1180
1181 list_del(&io_u->list);
1182 free(io_u);
1183 }
1184
1185 if (td->mem_type == MEM_MALLOC)
1186 free(td->orig_buffer);
1187 else if (td->mem_type == MEM_SHM) {
1188 struct shmid_ds sbuf;
1189
1190 shmdt(td->orig_buffer);
1191 shmctl(td->shm_id, IPC_RMID, &sbuf);
1192 } else if (td->mem_type == MEM_MMAP)
1193 munmap(td->orig_buffer, td->orig_buffer_size);
1194 else
1195 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1196
1197 td->orig_buffer = NULL;
1198}
1199
1200static int init_io_u(struct thread_data *td)
1201{
1202 struct io_u *io_u;
1203 int i, max_units;
1204 char *p;
1205
1206 if (td->io_engine & FIO_SYNCIO)
1207 max_units = 1;
1208 else
1209 max_units = td->iodepth;
1210
1211 td->orig_buffer_size = td->max_bs * max_units + MASK;
1212
1213 if (td->mem_type == MEM_MALLOC)
1214 td->orig_buffer = malloc(td->orig_buffer_size);
1215 else if (td->mem_type == MEM_SHM) {
1216 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1217 if (td->shm_id < 0) {
1218 td_verror(td, errno);
1219 perror("shmget");
1220 return 1;
1221 }
1222
1223 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1224 if (td->orig_buffer == (void *) -1) {
1225 td_verror(td, errno);
1226 perror("shmat");
1227 td->orig_buffer = NULL;
1228 return 1;
1229 }
1230 } else if (td->mem_type == MEM_MMAP) {
1231 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1232 if (td->orig_buffer == MAP_FAILED) {
1233 td_verror(td, errno);
1234 perror("mmap");
1235 td->orig_buffer = NULL;
1236 return 1;
1237 }
1238 }
1239
ebac4655
JA		 1240 p = ALIGN(td->orig_buffer);
1241 for (i = 0; i < max_units; i++) {
1242 io_u = malloc(sizeof(*io_u));
1243 memset(io_u, 0, sizeof(*io_u));
1244 INIT_LIST_HEAD(&io_u->list);
1245
1246 io_u->buf = p + td->max_bs * i;
b1ff3403 1247 io_u->index = i;
ebac4655
JA		 1248 list_add(&io_u->list, &td->io_u_freelist);
1249 }
1250
1251 return 0;
1252}
1253
1254static int create_file(struct thread_data *td, unsigned long long size,
1255 int extend)
1256{
1257 unsigned long long left;
1258 unsigned int bs;
1259 int r, oflags;
1260 char *b;
1261
1262 /*
1263 * unless specifically asked for overwrite, let normal io extend it
1264 */
1265 if (td_write(td) && !td->overwrite)
1266 return 0;
1267
1268 if (!size) {
1269 fprintf(stderr, "Need size for create\n");
1270 td_verror(td, EINVAL);
1271 return 1;
1272 }
1273
1274 if (!extend) {
1275 oflags = O_CREAT | O_TRUNC;
1276 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1277 } else {
1278 oflags = O_APPEND;
1279 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1280 }
1281
1282 td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1283 if (td->fd < 0) {
1284 td_verror(td, errno);
1285 return 1;
1286 }
1287
1288 if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1289 td_verror(td, errno);
1290 return 1;
1291 }
1292
1293 td->io_size = td->file_size;
1294 b = malloc(td->max_bs);
1295 memset(b, 0, td->max_bs);
1296
1297 left = size;
1298 while (left && !td->terminate) {
1299 bs = td->max_bs;
1300 if (bs > left)
1301 bs = left;
1302
1303 r = write(td->fd, b, bs);
1304
1305 if (r == (int) bs) {
1306 left -= bs;
1307 continue;
1308 } else {
1309 if (r < 0)
1310 td_verror(td, errno);
1311 else
1312 td_verror(td, EIO);
1313
1314 break;
1315 }
1316 }
1317
1318 if (td->terminate)
1319 unlink(td->file_name);
1320 else if (td->create_fsync)
1321 fsync(td->fd);
1322
1323 close(td->fd);
1324 td->fd = -1;
1325 free(b);
1326 return 0;
1327}
1328
1329static int file_size(struct thread_data *td)
1330{
1331 struct stat st;
1332
1333 if (fstat(td->fd, &st) == -1) {
1334 td_verror(td, errno);
1335 return 1;
1336 }
1337
838a3cd3
JA		 1338 td->real_file_size = st.st_size;
1339
1340 if (!td->file_size || td->file_size > td->real_file_size)
1341 td->file_size = td->real_file_size;
ebac4655 1342
e8e387c1 1343 td->file_size -= td->file_offset;
ebac4655
JA		 1344 return 0;
1345}
1346
1347static int bdev_size(struct thread_data *td)
1348{
9104f874 1349 unsigned long long bytes;
ebac4655
JA		 1350 int r;
1351
1352 r = blockdev_size(td->fd, &bytes);
1353 if (r) {
1354 td_verror(td, r);
1355 return 1;
1356 }
1357
838a3cd3
JA		 1358 td->real_file_size = bytes;
1359
ebac4655
JA		 1360 /*
1361 * no extend possibilities, so limit size to device size if too large
1362 */
838a3cd3
JA		 1363 if (!td->file_size || td->file_size > td->real_file_size)
1364 td->file_size = td->real_file_size;
ebac4655 1365
e8e387c1 1366 td->file_size -= td->file_offset;
ebac4655
JA		 1367 return 0;
1368}
1369
1370static int get_file_size(struct thread_data *td)
1371{
0af7b542 1372 int ret = 0;
ebac4655
JA		 1373
1374 if (td->filetype == FIO_TYPE_FILE)
1375 ret = file_size(td);
0af7b542 1376 else if (td->filetype == FIO_TYPE_BD)
ebac4655 1377 ret = bdev_size(td);
0af7b542
JA		 1378 else
1379 td->real_file_size = -1;
ebac4655
JA		 1380
1381 if (ret)
1382 return ret;
1383
e8e387c1
JA		 1384 if (td->file_offset > td->real_file_size) {
1385 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
ebac4655
JA		 1386 return 1;
1387 }
1388
838a3cd3 1389 td->io_size = td->file_size;
ebac4655
JA		 1390 if (td->io_size == 0) {
1391 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1392 td_verror(td, EINVAL);
1393 return 1;
1394 }
1395
20dc95c4
JA		 1396 if (!td->zone_size)
1397 td->zone_size = td->io_size;
1398
ebac4655
JA		 1399 td->total_io_size = td->io_size * td->loops;
1400 return 0;
1401}
1402
1403static int setup_file_mmap(struct thread_data *td)
1404{
1405 int flags;
1406
3d60d1ed
JA		 1407 if (td_rw(td))
1408 flags = PROT_READ | PROT_WRITE;
1409 else if (td_write(td)) {
ebac4655
JA		 1410 flags = PROT_WRITE;
1411
1412 if (td->verify != VERIFY_NONE)
1413 flags |= PROT_READ;
3d60d1ed
JA		 1414 } else
1415 flags = PROT_READ;
ebac4655
JA		 1416
1417 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1418 if (td->mmap == MAP_FAILED) {
1419 td->mmap = NULL;
1420 td_verror(td, errno);
1421 return 1;
1422 }
1423
1424 if (td->invalidate_cache) {
1425 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1426 td_verror(td, errno);
1427 return 1;
1428 }
1429 }
1430
1431 if (td->sequential) {
1432 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1433 td_verror(td, errno);
1434 return 1;
1435 }
1436 } else {
1437 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1438 td_verror(td, errno);
1439 return 1;
1440 }
1441 }
1442
1443 return 0;
1444}
1445
1446static int setup_file_plain(struct thread_data *td)
1447{
1448 if (td->invalidate_cache) {
1449 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1450 td_verror(td, errno);
1451 return 1;
1452 }
1453 }
1454
1455 if (td->sequential) {
1456 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1457 td_verror(td, errno);
1458 return 1;
1459 }
1460 } else {
1461 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1462 td_verror(td, errno);
1463 return 1;
1464 }
1465 }
1466
1467 return 0;
1468}
1469
1470static int setup_file(struct thread_data *td)
1471{
1472 struct stat st;
1473 int flags = 0;
1474
1475 if (stat(td->file_name, &st) == -1) {
1476 if (errno != ENOENT) {
1477 td_verror(td, errno);
1478 return 1;
1479 }
1480 if (!td->create_file) {
1481 td_verror(td, ENOENT);
1482 return 1;
1483 }
1484 if (create_file(td, td->file_size, 0))
1485 return 1;
1486 } else if (td->filetype == FIO_TYPE_FILE) {
6a0106a0 1487 if (st.st_size < (off_t) td->file_size) {
ebac4655
JA		 1488 if (create_file(td, td->file_size - st.st_size, 1))
1489 return 1;
1490 }
1491 }
1492
1493 if (td->odirect)
1494 flags |= O_DIRECT;
1495
3d60d1ed 1496 if (td_write(td) || td_rw(td)) {
ebac4655
JA		 1497 if (td->filetype == FIO_TYPE_FILE) {
1498 if (!td->overwrite)
1499 flags |= O_TRUNC;
1500
1501 flags |= O_CREAT;
1502 }
1503 if (td->sync_io)
1504 flags |= O_SYNC;
1505
1506 flags |= O_RDWR;
1507
1508 td->fd = open(td->file_name, flags, 0600);
3d60d1ed
JA		 1509 } else {
1510 if (td->filetype == FIO_TYPE_CHAR)
1511 flags |= O_RDWR;
1512 else
1513 flags |= O_RDONLY;
1514
1515 td->fd = open(td->file_name, flags);
ebac4655
JA		 1516 }
1517
1518 if (td->fd == -1) {
1519 td_verror(td, errno);
1520 return 1;
1521 }
1522
1523 if (get_file_size(td))
1524 return 1;
1525
1526 if (td->io_engine != FIO_MMAPIO)
1527 return setup_file_plain(td);
1528 else
1529 return setup_file_mmap(td);
1530}
1531
1532static int check_dev_match(dev_t dev, char *path)
1533{
1534 unsigned int major, minor;
1535 char line[256], *p;
1536 FILE *f;
1537
1538 f = fopen(path, "r");
1539 if (!f) {
1540 perror("open path");
1541 return 1;
1542 }
1543
1544 p = fgets(line, sizeof(line), f);
1545 if (!p) {
1546 fclose(f);
1547 return 1;
1548 }
1549
1550 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1551 fclose(f);
1552 return 1;
1553 }
1554
1555 if (((major << 8) | minor) == dev) {
1556 fclose(f);
1557 return 0;
1558 }
1559
1560 fclose(f);
1561 return 1;
1562}
1563
1564static int find_block_dir(dev_t dev, char *path)
1565{
1566 struct dirent *dir;
1567 struct stat st;
1568 int found = 0;
1569 DIR *D;
1570
1571 D = opendir(path);
1572 if (!D)
1573 return 0;
1574
1575 while ((dir = readdir(D)) != NULL) {
1576 char full_path[256];
1577
1578 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1579 continue;
1580 if (!strcmp(dir->d_name, "device"))
1581 continue;
1582
1583 sprintf(full_path, "%s/%s", path, dir->d_name);
1584
1585 if (!strcmp(dir->d_name, "dev")) {
1586 if (!check_dev_match(dev, full_path)) {
1587 found = 1;
1588 break;
1589 }
1590 }
1591
1592 if (stat(full_path, &st) == -1) {
1593 perror("stat");
1594 break;
1595 }
1596
1597 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1598 continue;
1599
1600 found = find_block_dir(dev, full_path);
1601 if (found) {
1602 strcpy(path, full_path);
1603 break;
1604 }
1605 }
1606
1607 closedir(D);
1608 return found;
1609}
1610
1611static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1612{
1613 unsigned in_flight;
1614 char line[256];
1615 FILE *f;
1616 char *p;
1617
1618 f = fopen(du->path, "r");
1619 if (!f)
1620 return 1;
1621
1622 p = fgets(line, sizeof(line), f);
1623 if (!p) {
1624 fclose(f);
1625 return 1;
1626 }
1627
1628 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) {
1629 fclose(f);
1630 return 1;
1631 }
1632
1633 fclose(f);
1634 return 0;
1635}
1636
1637static void update_io_tick_disk(struct disk_util *du)
1638{
1639 struct disk_util_stat __dus, *dus, *ldus;
1640 struct timeval t;
1641
1642 if (get_io_ticks(du, &__dus))
1643 return;
1644
1645 dus = &du->dus;
1646 ldus = &du->last_dus;
1647
1648 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1649 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1650 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1651 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1652 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1653 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1654 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1655 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1656 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1657 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1658
1659 gettimeofday(&t, NULL);
1660 du->msec += mtime_since(&du->time, &t);
1661 memcpy(&du->time, &t, sizeof(t));
1662 memcpy(ldus, &__dus, sizeof(__dus));
1663}
1664
1665static void update_io_ticks(void)
1666{
1667 struct list_head *entry;
1668 struct disk_util *du;
1669
1670 list_for_each(entry, &disk_list) {
1671 du = list_entry(entry, struct disk_util, list);
1672 update_io_tick_disk(du);
1673 }
1674}
1675
1676static int disk_util_exists(dev_t dev)
1677{
1678 struct list_head *entry;
1679 struct disk_util *du;
1680
1681 list_for_each(entry, &disk_list) {
1682 du = list_entry(entry, struct disk_util, list);
1683
1684 if (du->dev == dev)
1685 return 1;
1686 }
1687
1688 return 0;
1689}
1690
1691static void disk_util_add(dev_t dev, char *path)
1692{
1693 struct disk_util *du = malloc(sizeof(*du));
1694
1695 memset(du, 0, sizeof(*du));
1696 INIT_LIST_HEAD(&du->list);
1697 sprintf(du->path, "%s/stat", path);
1698 du->name = strdup(basename(path));
1699 du->dev = dev;
1700
1701 gettimeofday(&du->time, NULL);
1702 get_io_ticks(du, &du->last_dus);
1703
1704 list_add_tail(&du->list, &disk_list);
1705}
1706
1707static void init_disk_util(struct thread_data *td)
1708{
1709 struct stat st;
1710 char foo[256], tmp[256];
1711 dev_t dev;
1712 char *p;
1713
1714 if (!td->do_disk_util)
1715 return;
1716
1717 if (!stat(td->file_name, &st)) {
1718 if (S_ISBLK(st.st_mode))
1719 dev = st.st_rdev;
1720 else
1721 dev = st.st_dev;
1722 } else {
1723 /*
1724 * must be a file, open "." in that path
1725 */
1726 strcpy(foo, td->file_name);
1727 p = dirname(foo);
1728 if (stat(p, &st)) {
1729 perror("disk util stat");
1730 return;
1731 }
1732
1733 dev = st.st_dev;
1734 }
1735
1736 if (disk_util_exists(dev))
1737 return;
1738
1739 sprintf(foo, "/sys/block");
1740 if (!find_block_dir(dev, foo))
1741 return;
1742
1743 /*
d07f439e
JA		 1744 * If there's a ../queue/ directory there, we are inside a partition.
1745 * Check if that is the case and jump back. For loop/md/dm etc we
1746 * are already in the right spot.
ebac4655 1747 */
d07f439e
JA		 1748 sprintf(tmp, "%s/../queue", foo);
1749 if (!stat(tmp, &st)) {
1750 p = dirname(foo);
1751 sprintf(tmp, "%s/queue", p);
ebac4655 1752 if (stat(tmp, &st)) {
d07f439e
JA		 1753 fprintf(stderr, "unknown sysfs layout\n");
1754 return;
ebac4655 1755 }
d07f439e 1756 sprintf(foo, "%s", p);
ebac4655
JA		 1757 }
1758
1759 disk_util_add(dev, foo);
1760}
1761
1762static void disk_util_timer_arm(void)
1763{
1764 itimer.it_value.tv_sec = 0;
1765 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1766 setitimer(ITIMER_REAL, &itimer, NULL);
1767}
1768
1769static void clear_io_state(struct thread_data *td)
1770{
1771 if (td->io_engine == FIO_SYNCIO)
1772 lseek(td->fd, SEEK_SET, 0);
1773
20dc95c4 1774 td->last_pos = 0;
ebac4655
JA		 1775 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1776 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 1777 td->zone_bytes = 0;
ebac4655
JA		 1778
1779 if (td->file_map)
1780 memset(td->file_map, 0, td->num_maps * sizeof(long));
1781}
1782
1783static void update_rusage_stat(struct thread_data *td)
1784{
1785 if (!(td->runtime[0] + td->runtime[1]))
1786 return;
1787
1788 getrusage(RUSAGE_SELF, &td->ru_end);
1789
1790 td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1791 td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1792 td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1793
1794
1795 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1796}
1797
1798static void *thread_main(void *data)
1799{
1800 struct thread_data *td = data;
1801 int ret = 1;
1802
1803 if (!td->use_thread)
1804 setsid();
1805
1806 td->pid = getpid();
1807
aea47d44
JA		 1808 INIT_LIST_HEAD(&td->io_u_freelist);
1809 INIT_LIST_HEAD(&td->io_u_busylist);
1810 INIT_LIST_HEAD(&td->io_hist_list);
1811 INIT_LIST_HEAD(&td->io_log_list);
1812
ebac4655
JA		 1813 if (init_io_u(td))
1814 goto err;
1815
1816 if (fio_setaffinity(td) == -1) {
1817 td_verror(td, errno);
1818 goto err;
1819 }
1820
1821 if (init_io(td))
1822 goto err;
1823
aea47d44
JA		 1824 if (init_iolog(td))
1825 goto err;
1826
ebac4655
JA		 1827 if (td->ioprio) {
1828 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1829 td_verror(td, errno);
1830 goto err;
1831 }
1832 }
1833
75154845
JA		 1834 if (init_random_state(td))
1835 goto err;
1836
1837 td_set_runstate(td, TD_INITIALIZED);
ebac4655
JA		 1838 sem_post(&startup_sem);
1839 sem_wait(&td->mutex);
1840
1841 if (!td->create_serialize && setup_file(td))
1842 goto err;
1843
ebac4655
JA		 1844 gettimeofday(&td->epoch, NULL);
1845
1846 while (td->loops--) {
1847 getrusage(RUSAGE_SELF, &td->ru_start);
1848 gettimeofday(&td->start, NULL);
1849 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1850
1851 if (td->ratemin)
1852 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1853
1854 clear_io_state(td);
1855 prune_io_piece_log(td);
1856
1857 do_io(td);
1858
1859 td->runtime[td->ddir] += mtime_since_now(&td->start);
aea47d44 1860 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
3d60d1ed
JA		 1861 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1862
ebac4655
JA		 1863 update_rusage_stat(td);
1864
1865 if (td->error || td->terminate)
1866 break;
1867
1868 if (td->verify == VERIFY_NONE)
1869 continue;
1870
1871 clear_io_state(td);
1872 gettimeofday(&td->start, NULL);
1873
1874 do_verify(td);
1875
1876 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1877
1878 if (td->error || td->terminate)
1879 break;
1880 }
1881
1882 ret = 0;
1883
1884 if (td->bw_log)
1885 finish_log(td, td->bw_log, "bw");
1886 if (td->slat_log)
1887 finish_log(td, td->slat_log, "slat");
1888 if (td->clat_log)
1889 finish_log(td, td->clat_log, "clat");
1890
1891 if (exitall_on_terminate)
1892 terminate_threads(td->groupid);
1893
1894err:
1895 if (td->fd != -1) {
1896 close(td->fd);
1897 td->fd = -1;
1898 }
1899 if (td->mmap)
1900 munmap(td->mmap, td->file_size);
1901 cleanup_io(td);
1902 cleanup_io_u(td);
1903 if (ret) {
1904 sem_post(&startup_sem);
1905 sem_wait(&td->mutex);
1906 }
1907 td_set_runstate(td, TD_EXITED);
1908 return NULL;
1909
1910}
1911
1912static void *fork_main(int shmid, int offset)
1913{
1914 struct thread_data *td;
1915 void *data;
1916
1917 data = shmat(shmid, NULL, 0);
1918 if (data == (void *) -1) {
1919 perror("shmat");
1920 return NULL;
1921 }
1922
1923 td = data + offset * sizeof(struct thread_data);
1924 thread_main(td);
1925 shmdt(data);
1926 return NULL;
1927}
1928
1929static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1930 double *mean, double *dev)
1931{
1932 double n;
1933
1934 if (is->samples == 0)
1935 return 0;
1936
1937 *min = is->min_val;
1938 *max = is->max_val;
1939
1940 n = (double) is->samples;
1941 *mean = (double) is->val / n;
1942 *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1943 if (!(*min + *max) && !(*mean + *dev))
1944 return 0;
1945
1946 return 1;
1947}
1948
1949static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
1950 int ddir)
1951{
1952 char *ddir_str[] = { "read ", "write" };
1953 unsigned long min, max, bw;
1954 double mean, dev;
1955
1956 if (!td->runtime[ddir])
1957 return;
1958
1959 bw = td->io_bytes[ddir] / td->runtime[ddir];
9104f874 1960 printf(" %s: io=%6lluMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
ebac4655
JA		 1961
1962 if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
1963 printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1964
1965 if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
1966 printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1967
1968 if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
1969 double p_of_agg;
1970
1971 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1972 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);
1973 }
1974}
1975
1976static void show_thread_status(struct thread_data *td,
1977 struct group_run_stats *rs)
1978{
1979 double usr_cpu, sys_cpu;
1980
1981 if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
1982 return;
1983
1984 printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
1985
1986 show_ddir_status(td, rs, td->ddir);
aea47d44
JA		 1987 if (td->io_bytes[td->ddir ^ 1])
1988 show_ddir_status(td, rs, td->ddir ^ 1);
ebac4655
JA		 1989
1990 if (td->runtime[0] + td->runtime[1]) {
1991 double runt = td->runtime[0] + td->runtime[1];
1992
1993 usr_cpu = (double) td->usr_time * 100 / runt;
1994 sys_cpu = (double) td->sys_time * 100 / runt;
1995 } else {
1996 usr_cpu = 0;
1997 sys_cpu = 0;
1998 }
1999
2000 printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
2001}
2002
2003static void check_str_update(struct thread_data *td)
2004{
2005 char c = run_str[td->thread_number - 1];
2006
2007 if (td->runstate == td->old_runstate)
2008 return;
2009
2010 switch (td->runstate) {
2011 case TD_REAPED:
2012 c = '_';
2013 break;
2014 case TD_EXITED:
2015 c = 'E';
2016 break;
2017 case TD_RUNNING:
3d60d1ed
JA		 2018 if (td_rw(td)) {
2019 if (td->sequential)
2020 c = 'M';
2021 else
2022 c = 'm';
2023 } else if (td_read(td)) {
ebac4655
JA		 2024 if (td->sequential)
2025 c = 'R';
2026 else
2027 c = 'r';
2028 } else {
2029 if (td->sequential)
2030 c = 'W';
2031 else
2032 c = 'w';
2033 }
2034 break;
2035 case TD_VERIFYING:
2036 c = 'V';
2037 break;
2038 case TD_CREATED:
2039 c = 'C';
2040 break;
75154845
JA		 2041 case TD_INITIALIZED:
2042 c = 'I';
2043 break;
ebac4655
JA		 2044 case TD_NOT_CREATED:
2045 c = 'P';
2046 break;
2047 default:
2048 printf("state %d\n", td->runstate);
2049 }
2050
2051 run_str[td->thread_number - 1] = c;
2052 td->old_runstate = td->runstate;
2053}
2054
5289b847
JA		 2055static void eta_to_str(char *str, int eta_sec)
2056{
2057 unsigned int d, h, m, s;
2058 static int always_d, always_h;
2059
2060 d = h = m = s = 0;
2061
2062 s = eta_sec % 60;
2063 eta_sec /= 60;
2064 m = eta_sec % 60;
2065 eta_sec /= 60;
2066 h = eta_sec % 24;
2067 eta_sec /= 24;
2068 d = eta_sec;
2069
2070 if (d || always_d) {
2071 always_d = 1;
2072 str += sprintf(str, "%02dd:", d);
2073 }
2074 if (h || always_h) {
2075 always_h = 1;
2076 str += sprintf(str, "%02dh:", h);
2077 }
2078
2079 str += sprintf(str, "%02dm:", m);
2080 str += sprintf(str, "%02ds", s);
2081}
2082
6a0106a0
JA		 2083static int thread_eta(struct thread_data *td, unsigned long elapsed)
2084{
2085 unsigned long long bytes_total, bytes_done;
2086 unsigned int eta_sec = 0;
2087
2088 bytes_total = td->total_io_size;
3d60d1ed
JA		 2089
2090 /*
2091 * if writing, bytes_total will be twice the size. If mixing,
2092 * assume a 50/50 split and thus bytes_total will be 50% larger.
2093 */
2094 if (td->verify) {
2095 if (td_rw(td))
2096 bytes_total = bytes_total * 3 / 2;
2097 else
2098 bytes_total <<= 1;
2099 }
6a0106a0
JA		 2100 if (td->zone_size && td->zone_skip)
2101 bytes_total /= (td->zone_skip / td->zone_size);
2102
2103 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
2104 double perc;
8b611c34 2105
6a0106a0
JA		 2106 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
2107 perc = (double) bytes_done / (double) bytes_total;
2108 if (perc > 1.0)
2109 perc = 1.0;
2110
2111 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
2112
8b611c34 2113 if (td->timeout && eta_sec > (td->timeout - elapsed))
6a0106a0 2114 eta_sec = td->timeout - elapsed;
75154845
JA		 2115 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
2116 || td->runstate == TD_INITIALIZED) {
6a0106a0
JA		 2117 int t_eta = 0, r_eta = 0;
2118
2119 /*
2120 * We can only guess - assume it'll run the full timeout
2121 * if given, otherwise assume it'll run at the specified rate.
2122 */
2123 if (td->timeout)
2124 t_eta = td->timeout + td->start_delay - elapsed;
2125 if (td->rate) {
2126 r_eta = (bytes_total / 1024) / td->rate;
2127 r_eta += td->start_delay - elapsed;
2128 }
2129
2130 if (r_eta && t_eta)
2131 eta_sec = min(r_eta, t_eta);
2132 else if (r_eta)
2133 eta_sec = r_eta;
2134 else if (t_eta)
2135 eta_sec = t_eta;
2136 else
2137 eta_sec = INT_MAX;
2138 } else {
2139 /*
2140 * thread is already done
2141 */
2142 eta_sec = 0;
2143 }
2144
2145 return eta_sec;
2146}
2147
ebac4655
JA		 2148static void print_thread_status(void)
2149{
6a0106a0
JA		 2150 unsigned long elapsed = time_since_now(&genesis);
2151 int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
5289b847 2152 char eta_str[32];
6a0106a0
JA		 2153 double perc = 0.0;
2154
2155 eta_secs = malloc(thread_number * sizeof(int));
2156 memset(eta_secs, 0, thread_number * sizeof(int));
ebac4655 2157
ebac4655
JA		 2158 nr_running = t_rate = m_rate = 0;
2159 for (i = 0; i < thread_number; i++) {
2160 struct thread_data *td = &threads[i];
2161
2162 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
2163 nr_running++;
2164 t_rate += td->rate;
2165 m_rate += td->ratemin;
2166 }
2167
6a0106a0
JA		 2168 if (elapsed >= 3)
2169 eta_secs[i] = thread_eta(td, elapsed);
8b611c34
JA		 2170 else
2171 eta_secs[i] = INT_MAX;
ebac4655
JA		 2172
2173 check_str_update(td);
2174 }
2175
6a0106a0
JA		 2176 if (exitall_on_terminate)
2177 eta_sec = INT_MAX;
2178 else
2179 eta_sec = 0;
5289b847 2180
6a0106a0
JA		 2181 for (i = 0; i < thread_number; i++) {
2182 if (exitall_on_terminate) {
2183 if (eta_secs[i] < eta_sec)
2184 eta_sec = eta_secs[i];
2185 } else {
2186 if (eta_secs[i] > eta_sec)
2187 eta_sec = eta_secs[i];
5289b847 2188 }
6a0106a0 2189 }
5289b847 2190
8b611c34 2191 if (eta_sec != INT_MAX && elapsed) {
6a0106a0
JA		 2192 perc = (double) elapsed / (double) (elapsed + eta_sec);
2193 eta_to_str(eta_str, eta_sec);
ebac4655
JA		 2194 }
2195
5289b847 2196 printf("Threads now running (%d)", nr_running);
ebac4655
JA		 2197 if (m_rate || t_rate)
2198 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
8b611c34 2199 if (eta_sec != INT_MAX) {
6a0106a0
JA		 2200 perc *= 100.0;
2201 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
2202 }
5289b847 2203 printf("\r");
ebac4655 2204 fflush(stdout);
6a0106a0 2205 free(eta_secs);
ebac4655
JA		 2206}
2207
2208static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2209{
2210 int i;
2211
2212 /*
2213 * reap exited threads (TD_EXITED -> TD_REAPED)
2214 */
2215 for (i = 0; i < thread_number; i++) {
2216 struct thread_data *td = &threads[i];
2217
2218 if (td->runstate != TD_EXITED)
2219 continue;
2220
2221 td_set_runstate(td, TD_REAPED);
2222
2223 if (td->use_thread) {
2224 long ret;
2225
2226 if (pthread_join(td->thread, (void *) &ret))
2227 perror("thread_join");
2228 } else
2229 waitpid(td->pid, NULL, 0);
2230
2231 (*nr_running)--;
2232 (*m_rate) -= td->ratemin;
2233 (*t_rate) -= td->rate;
2234 }
2235}
2236
fcb6ade2
JA		 2237static void fio_unpin_memory(void *pinned)
2238{
2239 if (pinned) {
2240 if (munlock(pinned, mlock_size) < 0)
2241 perror("munlock");
2242 munmap(pinned, mlock_size);
2243 }
2244}
2245
2246static void *fio_pin_memory(void)
2247{
2248 long pagesize, pages;
2249 void *ptr;
2250
2251 if (!mlock_size)
2252 return NULL;
2253
2254 /*
2255 * Don't allow mlock of more than real_mem-128MB
2256 */
2257 pagesize = sysconf(_SC_PAGESIZE);
2258 pages = sysconf(_SC_PHYS_PAGES);
2259 if (pages != -1 && pagesize != -1) {
2260 unsigned long long real_mem = pages * pagesize;
2261
2262 if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
2263 mlock_size = real_mem - 128 * 1024 * 1024;
2264 printf("fio: limiting mlocked memory to %lluMiB\n",
2265 mlock_size >> 20);
2266 }
2267 }
2268
2269 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
2270 if (!ptr) {
2271 perror("malloc locked mem");
2272 return NULL;
2273 }
2274 if (mlock(ptr, mlock_size) < 0) {
2275 munmap(ptr, mlock_size);
2276 perror("mlock");
2277 return NULL;
2278 }
2279
2280 return ptr;
2281}
2282
ebac4655
JA		 2283static void run_threads(void)
2284{
ebac4655
JA		 2285 struct thread_data *td;
2286 unsigned long spent;
2287 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2
JA		 2288 void *mlocked_mem;
2289
2290 mlocked_mem = fio_pin_memory();
ebac4655
JA		 2291
2292 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
fcb6ade2 2293 fflush(stdout);
c04f7ec3 2294
4efa970e
JA		 2295 signal(SIGINT, sig_handler);
2296 signal(SIGALRM, sig_handler);
2297
ebac4655
JA		 2298 todo = thread_number;
2299 nr_running = 0;
2300 nr_started = 0;
2301 m_rate = t_rate = 0;
2302
2303 for (i = 0; i < thread_number; i++) {
2304 td = &threads[i];
2305
2306 run_str[td->thread_number - 1] = 'P';
2307
2308 init_disk_util(td);
2309
2310 if (!td->create_serialize)
2311 continue;
2312
2313 /*
2314 * do file setup here so it happens sequentially,
2315 * we don't want X number of threads getting their
2316 * client data interspersed on disk
2317 */
2318 if (setup_file(td)) {
2319 td_set_runstate(td, TD_REAPED);
2320 todo--;
2321 }
2322 }
2323
2324 gettimeofday(&genesis, NULL);
2325
2326 while (todo) {
75154845
JA		 2327 struct thread_data *map[MAX_JOBS];
2328 struct timeval this_start;
2329 int this_jobs = 0, left;
2330
ebac4655
JA		 2331 /*
2332 * create threads (TD_NOT_CREATED -> TD_CREATED)
2333 */
2334 for (i = 0; i < thread_number; i++) {
2335 td = &threads[i];
2336
2337 if (td->runstate != TD_NOT_CREATED)
2338 continue;
2339
2340 /*
2341 * never got a chance to start, killed by other
2342 * thread for some reason
2343 */
2344 if (td->terminate) {
2345 todo--;
2346 continue;
2347 }
2348
2349 if (td->start_delay) {
2350 spent = mtime_since_now(&genesis);
2351
2352 if (td->start_delay * 1000 > spent)
2353 continue;
2354 }
2355
2356 if (td->stonewall && (nr_started || nr_running))
2357 break;
2358
75154845
JA		 2359 /*
2360 * Set state to created. Thread will transition
2361 * to TD_INITIALIZED when it's done setting up.
2362 */
ebac4655 2363 td_set_runstate(td, TD_CREATED);
75154845 2364 map[this_jobs++] = td;
ebac4655 2365 sem_init(&startup_sem, 0, 1);
ebac4655
JA		 2366 nr_started++;
2367
2368 if (td->use_thread) {
2369 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2370 perror("thread_create");
2371 nr_started--;
2372 }
2373 } else {
2374 if (fork())
2375 sem_wait(&startup_sem);
2376 else {
2377 fork_main(shm_id, i);
2378 exit(0);
2379 }
2380 }
2381 }
2382
2383 /*
75154845
JA		 2384 * Wait for the started threads to transition to
2385 * TD_INITIALIZED.
ebac4655 2386 */
75154845
JA		 2387 printf("fio: Waiting for threads to initialize...\n");
2388 gettimeofday(&this_start, NULL);
2389 left = this_jobs;
2390 while (left) {
2391 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2392 break;
2393
2394 usleep(100000);
2395
2396 for (i = 0; i < this_jobs; i++) {
2397 td = map[i];
2398 if (!td)
2399 continue;
2400 if (td->runstate == TD_INITIALIZED ||
2401 td->runstate >= TD_EXITED) {
2402 map[i] = NULL;
2403 left--;
2404 continue;
2405 }
2406 }
2407 }
2408
2409 if (left) {
2410 fprintf(stderr, "fio: %d jobs failed to start\n", left);
2411 for (i = 0; i < this_jobs; i++) {
2412 td = map[i];
2413 if (!td)
2414 continue;
2415 kill(td->pid, SIGTERM);
2416 }
2417 break;
2418 }
2419
2420 /*
2421 * start created threads (TD_INITIALIZED -> TD_RUNNING)
2422 */
2423 printf("fio: Go for launch\n");
ebac4655
JA		 2424 for (i = 0; i < thread_number; i++) {
2425 td = &threads[i];
2426
75154845 2427 if (td->runstate != TD_INITIALIZED)
ebac4655
JA		 2428 continue;
2429
2430 td_set_runstate(td, TD_RUNNING);
2431 nr_running++;
2432 nr_started--;
2433 m_rate += td->ratemin;
2434 t_rate += td->rate;
75154845 2435 todo--;
ebac4655
JA		 2436 sem_post(&td->mutex);
2437 }
2438
2439 reap_threads(&nr_running, &t_rate, &m_rate);
2440
2441 if (todo)
2442 usleep(100000);
2443 }
2444
2445 while (nr_running) {
2446 reap_threads(&nr_running, &t_rate, &m_rate);
2447 usleep(10000);
2448 }
2449
2450 update_io_ticks();
fcb6ade2 2451 fio_unpin_memory(mlocked_mem);
ebac4655
JA		 2452}
2453
2454static void show_group_stats(struct group_run_stats *rs, int id)
2455{
2456 printf("\nRun status group %d (all jobs):\n", id);
2457
2458 if (rs->max_run[DDIR_READ])
9104f874 2459 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 2460 if (rs->max_run[DDIR_WRITE])
9104f874 2461 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		 2462}
2463
2464static void show_disk_util(void)
2465{
2466 struct disk_util_stat *dus;
2467 struct list_head *entry;
2468 struct disk_util *du;
2469 double util;
2470
2471 printf("\nDisk stats (read/write):\n");
2472
2473 list_for_each(entry, &disk_list) {
2474 du = list_entry(entry, struct disk_util, list);
2475 dus = &du->dus;
2476
2477 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2478 if (util > 100.0)
2479 util = 100.0;
2480
2481 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);
2482 }
2483}
2484
2485static void show_run_stats(void)
2486{
2487 struct group_run_stats *runstats, *rs;
2488 struct thread_data *td;
2489 int i;
2490
2491 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2492
2493 for (i = 0; i < groupid + 1; i++) {
2494 rs = &runstats[i];
2495
2496 memset(rs, 0, sizeof(*rs));
2497 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2498 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2499 }
2500
2501 for (i = 0; i < thread_number; i++) {
2502 unsigned long rbw, wbw;
2503
2504 td = &threads[i];
2505
2506 if (td->error) {
2507 printf("Client%d: %s\n", td->thread_number, td->verror);
2508 continue;
2509 }
2510
2511 rs = &runstats[td->groupid];
2512
2513 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2514 rs->min_run[0] = td->runtime[0];
2515 if (td->runtime[0] > rs->max_run[0])
2516 rs->max_run[0] = td->runtime[0];
2517 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2518 rs->min_run[1] = td->runtime[1];
2519 if (td->runtime[1] > rs->max_run[1])
2520 rs->max_run[1] = td->runtime[1];
2521
2522 rbw = wbw = 0;
2523 if (td->runtime[0])
2524 rbw = td->io_bytes[0] / td->runtime[0];
2525 if (td->runtime[1])
2526 wbw = td->io_bytes[1] / td->runtime[1];
2527
2528 if (rbw < rs->min_bw[0])
2529 rs->min_bw[0] = rbw;
2530 if (wbw < rs->min_bw[1])
2531 rs->min_bw[1] = wbw;
2532 if (rbw > rs->max_bw[0])
2533 rs->max_bw[0] = rbw;
2534 if (wbw > rs->max_bw[1])
2535 rs->max_bw[1] = wbw;
2536
2537 rs->io_mb[0] += td->io_bytes[0] >> 20;
2538 rs->io_mb[1] += td->io_bytes[1] >> 20;
2539 }
2540
2541 for (i = 0; i < groupid + 1; i++) {
2542 rs = &runstats[i];
2543
2544 if (rs->max_run[0])
2545 rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2546 if (rs->max_run[1])
2547 rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2548 }
2549
2550 /*
2551 * don't overwrite last signal output
2552 */
2553 printf("\n");
2554
2555 for (i = 0; i < thread_number; i++) {
2556 td = &threads[i];
2557 rs = &runstats[td->groupid];
2558
2559 show_thread_status(td, rs);
2560 }
2561
2562 for (i = 0; i < groupid + 1; i++)
2563 show_group_stats(&runstats[i], i);
2564
2565 show_disk_util();
2566}
2567
2568int main(int argc, char *argv[])
2569{
2570 if (parse_options(argc, argv))
2571 return 1;
2572
2573 if (!thread_number) {
2574 printf("Nothing to do\n");
2575 return 1;
2576 }
2577
2578 disk_util_timer_arm();
2579
2580 run_threads();
2581 show_run_stats();
2582
2583 return 0;
2584}
fio - Flexible IO Tester