[PATCH] Add ioscheduler=x option to allow setting io scheduler
[fio.git] / fio.c
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
48 int groupid = 0;
49 int thread_number = 0;
50 static char run_str[MAX_JOBS + 1];
51 int shm_id = 0;
52 static LIST_HEAD(disk_list);
53 static struct itimerval itimer;
54 static struct timeval genesis;
55
56 static void update_io_ticks(void);
57 static void disk_util_timer_arm(void);
58 static void print_thread_status(void);
59
60 extern unsigned long long mlock_size;
61
62 /*
63  * thread life cycle
64  */
65 enum {
66         TD_NOT_CREATED = 0,
67         TD_CREATED,
68         TD_INITIALIZED,
69         TD_RUNNING,
70         TD_VERIFYING,
71         TD_EXITED,
72         TD_REAPED,
73 };
74
75 #define should_fsync(td)        ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
76
77 static sem_t startup_sem;
78
79 #define TERMINATE_ALL           (-1)
80 #define JOB_START_TIMEOUT       (5 * 1000)
81
82 static 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
96 static 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
112 static 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
128 static 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
136 static 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
153 static 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
161 static inline unsigned long msec_now(struct timeval *s)
162 {
163         return s->tv_sec * 1000 + s->tv_usec / 1000;
164 }
165
166 static unsigned long time_since_now(struct timeval *s)
167 {
168         return mtime_since_now(s) / 1000;
169 }
170
171 static 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
179 static 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
198 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
199 {
200         unsigned long long block = io_u->offset / (unsigned long long) 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
223 static 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
235 static 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
251 static 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
259 static 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
267 static 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
285 static 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 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 / (unsigned long long) (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;
309         if (*offset > td->real_file_size)
310                 return 1;
311
312         return 0;
313 }
314
315 static 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
334 /*
335  * busy looping version for the last few usec
336  */
337 static 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
346 static 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
373 static 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
398 static 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
432 static 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
442 static 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
466 static 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
476 static 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
494 static 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
514 static 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
534 static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
535 {
536         hdr->crc32 = crc32(p, len);
537 }
538
539 static 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
548 static int get_rw_ddir(struct thread_data *td)
549 {
550         if (td_rw(td)) {
551                 struct timeval now;
552                 unsigned long elapsed;
553
554                 gettimeofday(&now, NULL);
555                 elapsed = mtime_since_now(&td->rwmix_switch);
556
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;
563
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))
574                 return DDIR_READ;
575         else
576                 return DDIR_WRITE;
577 }
578
579 /*
580  * fill body of io_u->buf with random data and add a header with the
581  * (eg) sha1sum of that data.
582  */
583 static 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
604 static int td_io_prep(struct thread_data *td, struct io_u *io_u)
605 {
606         if (td->io_prep && td->io_prep(td, io_u))
607                 return 1;
608
609         return 0;
610 }
611
612 void put_io_u(struct thread_data *td, struct io_u *io_u)
613 {
614         list_del(&io_u->list);
615         list_add(&io_u->list, &td->io_u_freelist);
616         td->cur_depth--;
617 }
618
619 static void write_iolog_put(struct thread_data *td, struct io_u *io_u)
620 {
621         fprintf(td->iolog_f, "%d,%llu,%u\n", io_u->ddir, io_u->offset, io_u->buflen);
622 }
623
624 static int read_iolog_get(struct thread_data *td, struct io_u *io_u)
625 {
626         struct io_piece *ipo;
627
628         if (!list_empty(&td->io_log_list)) {
629                 ipo = list_entry(td->io_log_list.next, struct io_piece, list);
630                 list_del(&ipo->list);
631                 io_u->offset = ipo->offset;
632                 io_u->buflen = ipo->len;
633                 io_u->ddir = ipo->ddir;
634                 free(ipo);
635                 return 0;
636         }
637
638         return 1;
639 }
640
641 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
642 {
643         /*
644          * If using an iolog, grab next piece if any available.
645          */
646         if (td->read_iolog)
647                 return read_iolog_get(td, io_u);
648
649         /*
650          * No log, let the seq/rand engine retrieve the next position.
651          */
652         if (!get_next_offset(td, &io_u->offset)) {
653                 io_u->buflen = get_next_buflen(td);
654
655                 if (io_u->buflen) {
656                         io_u->ddir = get_rw_ddir(td);
657
658                         /*
659                          * If using a write iolog, store this entry.
660                          */
661                         if (td->write_iolog)
662                                 write_iolog_put(td, io_u);
663
664                         return 0;
665                 }
666         }
667
668         return 1;
669 }
670
671 #define queue_full(td)  (list_empty(&(td)->io_u_freelist))
672
673 struct io_u *__get_io_u(struct thread_data *td)
674 {
675         struct io_u *io_u;
676
677         if (queue_full(td))
678                 return NULL;
679
680         io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
681         io_u->error = 0;
682         io_u->resid = 0;
683         list_del(&io_u->list);
684         list_add(&io_u->list, &td->io_u_busylist);
685         td->cur_depth++;
686         return io_u;
687 }
688
689 static struct io_u *get_io_u(struct thread_data *td)
690 {
691         struct io_u *io_u;
692
693         io_u = __get_io_u(td);
694         if (!io_u)
695                 return NULL;
696
697         if (td->zone_bytes >= td->zone_size) {
698                 td->zone_bytes = 0;
699                 td->last_pos += td->zone_skip;
700         }
701
702         if (fill_io_u(td, io_u)) {
703                 put_io_u(td, io_u);
704                 return NULL;
705         }
706
707         if (io_u->buflen + io_u->offset > td->real_file_size)
708                 io_u->buflen = td->real_file_size - io_u->offset;
709
710         if (!io_u->buflen) {
711                 put_io_u(td, io_u);
712                 return NULL;
713         }
714
715         if (!td->read_iolog && !td->sequential)
716                 mark_random_map(td, io_u);
717
718         td->last_pos += io_u->buflen;
719
720         if (td->verify != VERIFY_NONE)
721                 populate_io_u(td, io_u);
722
723         if (td_io_prep(td, io_u)) {
724                 put_io_u(td, io_u);
725                 return NULL;
726         }
727
728         gettimeofday(&io_u->start_time, NULL);
729         return io_u;
730 }
731
732 static inline void td_set_runstate(struct thread_data *td, int runstate)
733 {
734         td->old_runstate = td->runstate;
735         td->runstate = runstate;
736 }
737
738 static int get_next_verify(struct thread_data *td, struct io_u *io_u)
739 {
740         struct io_piece *ipo;
741
742         if (list_empty(&td->io_hist_list))
743                 return 1;
744
745         ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
746         list_del(&ipo->list);
747
748         io_u->offset = ipo->offset;
749         io_u->buflen = ipo->len;
750         io_u->ddir = DDIR_READ;
751         free(ipo);
752         return 0;
753 }
754
755 static void prune_io_piece_log(struct thread_data *td)
756 {
757         struct io_piece *ipo;
758
759         while (!list_empty(&td->io_hist_list)) {
760                 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
761
762                 list_del(&ipo->list);
763                 free(ipo);
764         }
765 }
766
767 /*
768  * log a succesful write, so we can unwind the log for verify
769  */
770 static void log_io_piece(struct thread_data *td, struct io_u *io_u)
771 {
772         struct io_piece *ipo = malloc(sizeof(struct io_piece));
773         struct list_head *entry;
774
775         INIT_LIST_HEAD(&ipo->list);
776         ipo->offset = io_u->offset;
777         ipo->len = io_u->buflen;
778
779         /*
780          * for random io where the writes extend the file, it will typically
781          * be laid out with the block scattered as written. it's faster to
782          * read them in in that order again, so don't sort
783          */
784         if (td->sequential || !td->overwrite) {
785                 list_add_tail(&ipo->list, &td->io_hist_list);
786                 return;
787         }
788
789         /*
790          * for random io, sort the list so verify will run faster
791          */
792         entry = &td->io_hist_list;
793         while ((entry = entry->prev) != &td->io_hist_list) {
794                 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
795
796                 if (__ipo->offset < ipo->offset)
797                         break;
798         }
799
800         list_add(&ipo->list, entry);
801 }
802
803 static void write_iolog_close(struct thread_data *td)
804 {
805         fflush(td->iolog_f);
806         fclose(td->iolog_f);
807         free(td->iolog_buf);
808 }
809
810 static int init_iolog(struct thread_data *td)
811 {
812         unsigned long long offset;
813         unsigned int bytes;
814         char *str, *p;
815         FILE *f;
816         int rw, i, reads, writes;
817
818         if (!td->read_iolog && !td->write_iolog)
819                 return 0;
820
821         if (td->read_iolog)
822                 f = fopen(td->iolog_file, "r");
823         else
824                 f = fopen(td->iolog_file, "w");
825
826         if (!f) {
827                 perror("fopen iolog");
828                 printf("file %s, %d/%d\n", td->iolog_file, td->read_iolog, td->write_iolog);
829                 return 1;
830         }
831
832         /*
833          * That's it for writing, setup a log buffer and we're done.
834           */
835         if (td->write_iolog) {
836                 td->iolog_f = f;
837                 td->iolog_buf = malloc(8192);
838                 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
839                 return 0;
840         }
841
842         /*
843          * Read in the read iolog and store it, reuse the infrastructure
844          * for doing verifications.
845          */
846         str = malloc(4096);
847         reads = writes = i = 0;
848         while ((p = fgets(str, 4096, f)) != NULL) {
849                 struct io_piece *ipo;
850
851                 if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
852                         fprintf(stderr, "bad iolog: %s\n", p);
853                         continue;
854                 }
855                 if (rw == DDIR_READ)
856                         reads++;
857                 else if (rw == DDIR_WRITE)
858                         writes++;
859                 else {
860                         fprintf(stderr, "bad ddir: %d\n", rw);
861                         continue;
862                 }
863
864                 ipo = malloc(sizeof(*ipo));
865                 INIT_LIST_HEAD(&ipo->list);
866                 ipo->offset = offset;
867                 ipo->len = bytes;
868                 if (bytes > td->max_bs)
869                         td->max_bs = bytes;
870                 ipo->ddir = rw;
871                 list_add_tail(&ipo->list, &td->io_log_list);
872                 i++;
873         }
874
875         free(str);
876         fclose(f);
877
878         if (!i)
879                 return 1;
880
881         if (reads && !writes)
882                 td->ddir = DDIR_READ;
883         else if (!reads && writes)
884                 td->ddir = DDIR_READ;
885         else
886                 td->iomix = 1;
887
888         return 0;
889 }
890
891 static int sync_td(struct thread_data *td)
892 {
893         if (td->io_sync)
894                 return td->io_sync(td);
895
896         return 0;
897 }
898
899 static int io_u_getevents(struct thread_data *td, int min, int max,
900                           struct timespec *t)
901 {
902         return td->io_getevents(td, min, max, t);
903 }
904
905 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
906 {
907         gettimeofday(&io_u->issue_time, NULL);
908
909         return td->io_queue(td, io_u);
910 }
911
912 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
913
914 static void io_completed(struct thread_data *td, struct io_u *io_u,
915                          struct io_completion_data *icd)
916 {
917         struct timeval e;
918         unsigned long msec;
919
920         gettimeofday(&e, NULL);
921
922         if (!io_u->error) {
923                 unsigned int bytes = io_u->buflen - io_u->resid;
924                 const int idx = io_u->ddir;
925
926                 td->io_blocks[idx]++;
927                 td->io_bytes[idx] += bytes;
928                 td->zone_bytes += bytes;
929                 td->this_io_bytes[idx] += bytes;
930
931                 msec = mtime_since(&io_u->issue_time, &e);
932
933                 add_clat_sample(td, idx, msec);
934                 add_bw_sample(td, idx);
935
936                 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
937                         log_io_piece(td, io_u);
938
939                 icd->bytes_done[idx] += bytes;
940         } else
941                 icd->error = io_u->error;
942 }
943
944 static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
945 {
946         struct io_u *io_u;
947         int i;
948
949         icd->error = 0;
950         icd->bytes_done[0] = icd->bytes_done[1] = 0;
951
952         for (i = 0; i < icd->nr; i++) {
953                 io_u = td->io_event(td, i);
954
955                 io_completed(td, io_u, icd);
956                 put_io_u(td, io_u);
957         }
958 }
959
960 static void cleanup_pending_aio(struct thread_data *td)
961 {
962         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
963         struct list_head *entry, *n;
964         struct io_completion_data icd;
965         struct io_u *io_u;
966         int r;
967
968         /*
969          * get immediately available events, if any
970          */
971         r = io_u_getevents(td, 0, td->cur_depth, &ts);
972         if (r > 0) {
973                 icd.nr = r;
974                 ios_completed(td, &icd);
975         }
976
977         /*
978          * now cancel remaining active events
979          */
980         if (td->io_cancel) {
981                 list_for_each_safe(entry, n, &td->io_u_busylist) {
982                         io_u = list_entry(entry, struct io_u, list);
983
984                         r = td->io_cancel(td, io_u);
985                         if (!r)
986                                 put_io_u(td, io_u);
987                 }
988         }
989
990         if (td->cur_depth) {
991                 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
992                 if (r > 0) {
993                         icd.nr = r;
994                         ios_completed(td, &icd);
995                 }
996         }
997 }
998
999 static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
1000 {
1001         struct io_u *v_io_u = *io_u;
1002         int ret = 0;
1003
1004         if (v_io_u) {
1005                 ret = verify_io_u(v_io_u);
1006                 put_io_u(td, v_io_u);
1007                 *io_u = NULL;
1008         }
1009
1010         return ret;
1011 }
1012
1013 static void do_verify(struct thread_data *td)
1014 {
1015         struct timeval t;
1016         struct io_u *io_u, *v_io_u = NULL;
1017         struct io_completion_data icd;
1018         int ret;
1019
1020         td_set_runstate(td, TD_VERIFYING);
1021
1022         do {
1023                 if (td->terminate)
1024                         break;
1025
1026                 gettimeofday(&t, NULL);
1027                 if (runtime_exceeded(td, &t))
1028                         break;
1029
1030                 io_u = __get_io_u(td);
1031                 if (!io_u)
1032                         break;
1033
1034                 if (get_next_verify(td, io_u)) {
1035                         put_io_u(td, io_u);
1036                         break;
1037                 }
1038
1039                 if (td_io_prep(td, io_u)) {
1040                         put_io_u(td, io_u);
1041                         break;
1042                 }
1043
1044                 ret = io_u_queue(td, io_u);
1045                 if (ret) {
1046                         put_io_u(td, io_u);
1047                         td_verror(td, ret);
1048                         break;
1049                 }
1050
1051                 /*
1052                  * we have one pending to verify, do that while
1053                  * we are doing io on the next one
1054                  */
1055                 if (do_io_u_verify(td, &v_io_u))
1056                         break;
1057
1058                 ret = io_u_getevents(td, 1, 1, NULL);
1059                 if (ret != 1) {
1060                         if (ret < 0)
1061                                 td_verror(td, ret);
1062                         break;
1063                 }
1064
1065                 v_io_u = td->io_event(td, 0);
1066                 icd.nr = 1;
1067                 icd.error = 0;
1068                 io_completed(td, v_io_u, &icd);
1069
1070                 if (icd.error) {
1071                         td_verror(td, icd.error);
1072                         put_io_u(td, v_io_u);
1073                         v_io_u = NULL;
1074                         break;
1075                 }
1076
1077                 /*
1078                  * if we can't submit more io, we need to verify now
1079                  */
1080                 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
1081                         break;
1082
1083         } while (1);
1084
1085         do_io_u_verify(td, &v_io_u);
1086
1087         if (td->cur_depth)
1088                 cleanup_pending_aio(td);
1089
1090         td_set_runstate(td, TD_RUNNING);
1091 }
1092
1093 static void do_io(struct thread_data *td)
1094 {
1095         struct io_completion_data icd;
1096         struct timeval s, e;
1097         unsigned long usec;
1098
1099         while (td->this_io_bytes[td->ddir] < td->io_size) {
1100                 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
1101                 struct timespec *timeout;
1102                 int ret, min_evts = 0;
1103                 struct io_u *io_u;
1104
1105                 if (td->terminate)
1106                         break;
1107
1108                 io_u = get_io_u(td);
1109                 if (!io_u)
1110                         break;
1111
1112                 memcpy(&s, &io_u->start_time, sizeof(s));
1113
1114                 ret = io_u_queue(td, io_u);
1115                 if (ret) {
1116                         put_io_u(td, io_u);
1117                         td_verror(td, ret);
1118                         break;
1119                 }
1120
1121                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
1122
1123                 if (td->cur_depth < td->iodepth) {
1124                         timeout = &ts;
1125                         min_evts = 0;
1126                 } else {
1127                         timeout = NULL;
1128                         min_evts = 1;
1129                 }
1130
1131                 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
1132                 if (ret < 0) {
1133                         td_verror(td, ret);
1134                         break;
1135                 } else if (!ret)
1136                         continue;
1137
1138                 icd.nr = ret;
1139                 ios_completed(td, &icd);
1140                 if (icd.error) {
1141                         td_verror(td, icd.error);
1142                         break;
1143                 }
1144
1145                 /*
1146                  * the rate is batched for now, it should work for batches
1147                  * of completions except the very first one which may look
1148                  * a little bursty
1149                  */
1150                 gettimeofday(&e, NULL);
1151                 usec = utime_since(&s, &e);
1152
1153                 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
1154
1155                 if (check_min_rate(td, &e)) {
1156                         td_verror(td, ENOMEM);
1157                         break;
1158                 }
1159
1160                 if (runtime_exceeded(td, &e))
1161                         break;
1162
1163                 if (td->thinktime)
1164                         usec_sleep(td, td->thinktime);
1165
1166                 if (should_fsync(td) && td->fsync_blocks &&
1167                     (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
1168                         sync_td(td);
1169         }
1170
1171         if (td->cur_depth)
1172                 cleanup_pending_aio(td);
1173
1174         if (should_fsync(td) && td->end_fsync)
1175                 sync_td(td);
1176 }
1177
1178 static void cleanup_io(struct thread_data *td)
1179 {
1180         if (td->io_cleanup)
1181                 td->io_cleanup(td);
1182 }
1183
1184 static int init_io(struct thread_data *td)
1185 {
1186         if (td->io_engine == FIO_SYNCIO)
1187                 return fio_syncio_init(td);
1188         else if (td->io_engine == FIO_MMAPIO)
1189                 return fio_mmapio_init(td);
1190         else if (td->io_engine == FIO_LIBAIO)
1191                 return fio_libaio_init(td);
1192         else if (td->io_engine == FIO_POSIXAIO)
1193                 return fio_posixaio_init(td);
1194         else if (td->io_engine == FIO_SGIO)
1195                 return fio_sgio_init(td);
1196         else if (td->io_engine == FIO_SPLICEIO)
1197                 return fio_spliceio_init(td);
1198         else {
1199                 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1200                 return 1;
1201         }
1202 }
1203
1204 static void cleanup_io_u(struct thread_data *td)
1205 {
1206         struct list_head *entry, *n;
1207         struct io_u *io_u;
1208
1209         list_for_each_safe(entry, n, &td->io_u_freelist) {
1210                 io_u = list_entry(entry, struct io_u, list);
1211
1212                 list_del(&io_u->list);
1213                 free(io_u);
1214         }
1215
1216         if (td->mem_type == MEM_MALLOC)
1217                 free(td->orig_buffer);
1218         else if (td->mem_type == MEM_SHM) {
1219                 struct shmid_ds sbuf;
1220
1221                 shmdt(td->orig_buffer);
1222                 shmctl(td->shm_id, IPC_RMID, &sbuf);
1223         } else if (td->mem_type == MEM_MMAP)
1224                 munmap(td->orig_buffer, td->orig_buffer_size);
1225         else
1226                 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1227
1228         td->orig_buffer = NULL;
1229 }
1230
1231 static int init_io_u(struct thread_data *td)
1232 {
1233         struct io_u *io_u;
1234         int i, max_units;
1235         char *p;
1236
1237         if (td->io_engine & FIO_SYNCIO)
1238                 max_units = 1;
1239         else
1240                 max_units = td->iodepth;
1241
1242         td->orig_buffer_size = td->max_bs * max_units + MASK;
1243
1244         if (td->mem_type == MEM_MALLOC)
1245                 td->orig_buffer = malloc(td->orig_buffer_size);
1246         else if (td->mem_type == MEM_SHM) {
1247                 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1248                 if (td->shm_id < 0) {
1249                         td_verror(td, errno);
1250                         perror("shmget");
1251                         return 1;
1252                 }
1253
1254                 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1255                 if (td->orig_buffer == (void *) -1) {
1256                         td_verror(td, errno);
1257                         perror("shmat");
1258                         td->orig_buffer = NULL;
1259                         return 1;
1260                 }
1261         } else if (td->mem_type == MEM_MMAP) {
1262                 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1263                 if (td->orig_buffer == MAP_FAILED) {
1264                         td_verror(td, errno);
1265                         perror("mmap");
1266                         td->orig_buffer = NULL;
1267                         return 1;
1268                 }
1269         }
1270
1271         p = ALIGN(td->orig_buffer);
1272         for (i = 0; i < max_units; i++) {
1273                 io_u = malloc(sizeof(*io_u));
1274                 memset(io_u, 0, sizeof(*io_u));
1275                 INIT_LIST_HEAD(&io_u->list);
1276
1277                 io_u->buf = p + td->max_bs * i;
1278                 io_u->index = i;
1279                 list_add(&io_u->list, &td->io_u_freelist);
1280         }
1281
1282         return 0;
1283 }
1284
1285 static int create_file(struct thread_data *td, unsigned long long size,
1286                        int extend)
1287 {
1288         unsigned long long left;
1289         unsigned int bs;
1290         int r, oflags;
1291         char *b;
1292
1293         /*
1294          * unless specifically asked for overwrite, let normal io extend it
1295          */
1296         if (td_write(td) && !td->overwrite)
1297                 return 0;
1298
1299         if (!size) {
1300                 fprintf(stderr, "Need size for create\n");
1301                 td_verror(td, EINVAL);
1302                 return 1;
1303         }
1304
1305         if (!extend) {
1306                 oflags = O_CREAT | O_TRUNC;
1307                 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1308         } else {
1309                 oflags = O_APPEND;
1310                 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1311         }
1312
1313         td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1314         if (td->fd < 0) {
1315                 td_verror(td, errno);
1316                 return 1;
1317         }
1318
1319         if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1320                 td_verror(td, errno);
1321                 return 1;
1322         }
1323
1324         td->io_size = td->file_size;
1325         b = malloc(td->max_bs);
1326         memset(b, 0, td->max_bs);
1327
1328         left = size;
1329         while (left && !td->terminate) {
1330                 bs = td->max_bs;
1331                 if (bs > left)
1332                         bs = left;
1333
1334                 r = write(td->fd, b, bs);
1335
1336                 if (r == (int) bs) {
1337                         left -= bs;
1338                         continue;
1339                 } else {
1340                         if (r < 0)
1341                                 td_verror(td, errno);
1342                         else
1343                                 td_verror(td, EIO);
1344
1345                         break;
1346                 }
1347         }
1348
1349         if (td->terminate)
1350                 unlink(td->file_name);
1351         else if (td->create_fsync)
1352                 fsync(td->fd);
1353
1354         close(td->fd);
1355         td->fd = -1;
1356         free(b);
1357         return 0;
1358 }
1359
1360 static int file_size(struct thread_data *td)
1361 {
1362         struct stat st;
1363
1364         if (fstat(td->fd, &st) == -1) {
1365                 td_verror(td, errno);
1366                 return 1;
1367         }
1368
1369         td->real_file_size = st.st_size;
1370
1371         if (!td->file_size || td->file_size > td->real_file_size)
1372                 td->file_size = td->real_file_size;
1373
1374         td->file_size -= td->file_offset;
1375         return 0;
1376 }
1377
1378 static int bdev_size(struct thread_data *td)
1379 {
1380         unsigned long long bytes;
1381         int r;
1382
1383         r = blockdev_size(td->fd, &bytes);
1384         if (r) {
1385                 td_verror(td, r);
1386                 return 1;
1387         }
1388
1389         td->real_file_size = bytes;
1390
1391         /*
1392          * no extend possibilities, so limit size to device size if too large
1393          */
1394         if (!td->file_size || td->file_size > td->real_file_size)
1395                 td->file_size = td->real_file_size;
1396
1397         td->file_size -= td->file_offset;
1398         return 0;
1399 }
1400
1401 static int get_file_size(struct thread_data *td)
1402 {
1403         int ret = 0;
1404
1405         if (td->filetype == FIO_TYPE_FILE)
1406                 ret = file_size(td);
1407         else if (td->filetype == FIO_TYPE_BD)
1408                 ret = bdev_size(td);
1409         else
1410                 td->real_file_size = -1;
1411
1412         if (ret)
1413                 return ret;
1414
1415         if (td->file_offset > td->real_file_size) {
1416                 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
1417                 return 1;
1418         }
1419
1420         td->io_size = td->file_size;
1421         if (td->io_size == 0) {
1422                 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1423                 td_verror(td, EINVAL);
1424                 return 1;
1425         }
1426
1427         if (!td->zone_size)
1428                 td->zone_size = td->io_size;
1429
1430         td->total_io_size = td->io_size * td->loops;
1431         return 0;
1432 }
1433
1434 static int setup_file_mmap(struct thread_data *td)
1435 {
1436         int flags;
1437
1438         if (td_rw(td))
1439                 flags = PROT_READ | PROT_WRITE;
1440         else if (td_write(td)) {
1441                 flags = PROT_WRITE;
1442
1443                 if (td->verify != VERIFY_NONE)
1444                         flags |= PROT_READ;
1445         } else
1446                 flags = PROT_READ;
1447
1448         td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1449         if (td->mmap == MAP_FAILED) {
1450                 td->mmap = NULL;
1451                 td_verror(td, errno);
1452                 return 1;
1453         }
1454
1455         if (td->invalidate_cache) {
1456                 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1457                         td_verror(td, errno);
1458                         return 1;
1459                 }
1460         }
1461
1462         if (td->sequential) {
1463                 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1464                         td_verror(td, errno);
1465                         return 1;
1466                 }
1467         } else {
1468                 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1469                         td_verror(td, errno);
1470                         return 1;
1471                 }
1472         }
1473
1474         return 0;
1475 }
1476
1477 static int setup_file_plain(struct thread_data *td)
1478 {
1479         if (td->invalidate_cache) {
1480                 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1481                         td_verror(td, errno);
1482                         return 1;
1483                 }
1484         }
1485
1486         if (td->sequential) {
1487                 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1488                         td_verror(td, errno);
1489                         return 1;
1490                 }
1491         } else {
1492                 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1493                         td_verror(td, errno);
1494                         return 1;
1495                 }
1496         }
1497
1498         return 0;
1499 }
1500
1501 static int setup_file(struct thread_data *td)
1502 {
1503         struct stat st;
1504         int flags = 0;
1505
1506         if (stat(td->file_name, &st) == -1) {
1507                 if (errno != ENOENT) {
1508                         td_verror(td, errno);
1509                         return 1;
1510                 }
1511                 if (!td->create_file) {
1512                         td_verror(td, ENOENT);
1513                         return 1;
1514                 }
1515                 if (create_file(td, td->file_size, 0))
1516                         return 1;
1517         } else if (td->filetype == FIO_TYPE_FILE) {
1518                 if (st.st_size < (off_t) td->file_size) {
1519                         if (create_file(td, td->file_size - st.st_size, 1))
1520                                 return 1;
1521                 }
1522         }
1523
1524         if (td->odirect)
1525                 flags |= O_DIRECT;
1526
1527         if (td_write(td) || td_rw(td)) {
1528                 if (td->filetype == FIO_TYPE_FILE) {
1529                         if (!td->overwrite)
1530                                 flags |= O_TRUNC;
1531
1532                         flags |= O_CREAT;
1533                 }
1534                 if (td->sync_io)
1535                         flags |= O_SYNC;
1536
1537                 flags |= O_RDWR;
1538
1539                 td->fd = open(td->file_name, flags, 0600);
1540         } else {
1541                 if (td->filetype == FIO_TYPE_CHAR)
1542                         flags |= O_RDWR;
1543                 else
1544                         flags |= O_RDONLY;
1545
1546                 td->fd = open(td->file_name, flags);
1547         }
1548
1549         if (td->fd == -1) {
1550                 td_verror(td, errno);
1551                 return 1;
1552         }
1553
1554         if (get_file_size(td))
1555                 return 1;
1556
1557         if (td->io_engine != FIO_MMAPIO)
1558                 return setup_file_plain(td);
1559         else
1560                 return setup_file_mmap(td);
1561 }
1562
1563 static int check_dev_match(dev_t dev, char *path)
1564 {
1565         unsigned int major, minor;
1566         char line[256], *p;
1567         FILE *f;
1568
1569         f = fopen(path, "r");
1570         if (!f) {
1571                 perror("open path");
1572                 return 1;
1573         }
1574
1575         p = fgets(line, sizeof(line), f);
1576         if (!p) {
1577                 fclose(f);
1578                 return 1;
1579         }
1580
1581         if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1582                 fclose(f);
1583                 return 1;
1584         }
1585
1586         if (((major << 8) | minor) == dev) {
1587                 fclose(f);
1588                 return 0;
1589         }
1590
1591         fclose(f);
1592         return 1;
1593 }
1594
1595 static int find_block_dir(dev_t dev, char *path)
1596 {
1597         struct dirent *dir;
1598         struct stat st;
1599         int found = 0;
1600         DIR *D;
1601
1602         D = opendir(path);
1603         if (!D)
1604                 return 0;
1605
1606         while ((dir = readdir(D)) != NULL) {
1607                 char full_path[256];
1608
1609                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1610                         continue;
1611                 if (!strcmp(dir->d_name, "device"))
1612                         continue;
1613
1614                 sprintf(full_path, "%s/%s", path, dir->d_name);
1615
1616                 if (!strcmp(dir->d_name, "dev")) {
1617                         if (!check_dev_match(dev, full_path)) {
1618                                 found = 1;
1619                                 break;
1620                         }
1621                 }
1622
1623                 if (stat(full_path, &st) == -1) {
1624                         perror("stat");
1625                         break;
1626                 }
1627
1628                 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1629                         continue;
1630
1631                 found = find_block_dir(dev, full_path);
1632                 if (found) {
1633                         strcpy(path, full_path);
1634                         break;
1635                 }
1636         }
1637
1638         closedir(D);
1639         return found;
1640 }
1641
1642 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1643 {
1644         unsigned in_flight;
1645         char line[256];
1646         FILE *f;
1647         char *p;
1648
1649         f = fopen(du->path, "r");
1650         if (!f)
1651                 return 1;
1652
1653         p = fgets(line, sizeof(line), f);
1654         if (!p) {
1655                 fclose(f);
1656                 return 1;
1657         }
1658
1659         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) {
1660                 fclose(f);
1661                 return 1;
1662         }
1663
1664         fclose(f);
1665         return 0;
1666 }
1667
1668 static void update_io_tick_disk(struct disk_util *du)
1669 {
1670         struct disk_util_stat __dus, *dus, *ldus;
1671         struct timeval t;
1672
1673         if (get_io_ticks(du, &__dus))
1674                 return;
1675
1676         dus = &du->dus;
1677         ldus = &du->last_dus;
1678
1679         dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1680         dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1681         dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1682         dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1683         dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1684         dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1685         dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1686         dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1687         dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1688         dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1689
1690         gettimeofday(&t, NULL);
1691         du->msec += mtime_since(&du->time, &t);
1692         memcpy(&du->time, &t, sizeof(t));
1693         memcpy(ldus, &__dus, sizeof(__dus));
1694 }
1695
1696 static void update_io_ticks(void)
1697 {
1698         struct list_head *entry;
1699         struct disk_util *du;
1700
1701         list_for_each(entry, &disk_list) {
1702                 du = list_entry(entry, struct disk_util, list);
1703                 update_io_tick_disk(du);
1704         }
1705 }
1706
1707 static int disk_util_exists(dev_t dev)
1708 {
1709         struct list_head *entry;
1710         struct disk_util *du;
1711
1712         list_for_each(entry, &disk_list) {
1713                 du = list_entry(entry, struct disk_util, list);
1714
1715                 if (du->dev == dev)
1716                         return 1;
1717         }
1718
1719         return 0;
1720 }
1721
1722 static void disk_util_add(dev_t dev, char *path)
1723 {
1724         struct disk_util *du = malloc(sizeof(*du));
1725
1726         memset(du, 0, sizeof(*du));
1727         INIT_LIST_HEAD(&du->list);
1728         sprintf(du->path, "%s/stat", path);
1729         du->name = strdup(basename(path));
1730         du->dev = dev;
1731
1732         gettimeofday(&du->time, NULL);
1733         get_io_ticks(du, &du->last_dus);
1734
1735         list_add_tail(&du->list, &disk_list);
1736 }
1737
1738 static void init_disk_util(struct thread_data *td)
1739 {
1740         struct stat st;
1741         char foo[256], tmp[256];
1742         dev_t dev;
1743         char *p;
1744
1745         if (!td->do_disk_util)
1746                 return;
1747
1748         if (!stat(td->file_name, &st)) {
1749                 if (S_ISBLK(st.st_mode))
1750                         dev = st.st_rdev;
1751                 else
1752                         dev = st.st_dev;
1753         } else {
1754                 /*
1755                  * must be a file, open "." in that path
1756                  */
1757                 strcpy(foo, td->file_name);
1758                 p = dirname(foo);
1759                 if (stat(p, &st)) {
1760                         perror("disk util stat");
1761                         return;
1762                 }
1763
1764                 dev = st.st_dev;
1765         }
1766
1767         if (disk_util_exists(dev))
1768                 return;
1769                 
1770         sprintf(foo, "/sys/block");
1771         if (!find_block_dir(dev, foo))
1772                 return;
1773
1774         /*
1775          * If there's a ../queue/ directory there, we are inside a partition.
1776          * Check if that is the case and jump back. For loop/md/dm etc we
1777          * are already in the right spot.
1778          */
1779         sprintf(tmp, "%s/../queue", foo);
1780         if (!stat(tmp, &st)) {
1781                 p = dirname(foo);
1782                 sprintf(tmp, "%s/queue", p);
1783                 if (stat(tmp, &st)) {
1784                         fprintf(stderr, "unknown sysfs layout\n");
1785                         return;
1786                 }
1787                 sprintf(foo, "%s", p);
1788         }
1789
1790         td->sysfs_root = strdup(foo);
1791         disk_util_add(dev, foo);
1792 }
1793
1794 static int switch_ioscheduler(struct thread_data *td)
1795 {
1796         char tmp[256], tmp2[128];
1797         FILE *f;
1798         int ret;
1799
1800         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1801
1802         f = fopen(tmp, "r+");
1803         if (!f) {
1804                 td_verror(td, errno);
1805                 return 1;
1806         }
1807
1808         /*
1809          * Set io scheduler.
1810          */
1811         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
1812         if (ferror(f) || ret != 1) {
1813                 td_verror(td, errno);
1814                 fclose(f);
1815                 return 1;
1816         }
1817
1818         rewind(f);
1819
1820         /*
1821          * Read back and check that the selected scheduler is now the default.
1822          */
1823         ret = fread(tmp, 1, sizeof(tmp), f);
1824         if (ferror(f) || ret < 0) {
1825                 td_verror(td, errno);
1826                 fclose(f);
1827                 return 1;
1828         }
1829
1830         sprintf(tmp2, "[%s]", td->ioscheduler);
1831         if (!strstr(tmp, tmp2)) {
1832                 fprintf(stderr, "fio: io scheduler %s not found\n", td->ioscheduler);
1833                 td_verror(td, EINVAL);
1834                 fclose(f);
1835                 return 1;
1836         }
1837
1838         fclose(f);
1839         return 0;
1840 }
1841
1842 static void disk_util_timer_arm(void)
1843 {
1844         itimer.it_value.tv_sec = 0;
1845         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1846         setitimer(ITIMER_REAL, &itimer, NULL);
1847 }
1848
1849 static void clear_io_state(struct thread_data *td)
1850 {
1851         if (td->io_engine == FIO_SYNCIO)
1852                 lseek(td->fd, SEEK_SET, 0);
1853
1854         td->last_pos = 0;
1855         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1856         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1857         td->zone_bytes = 0;
1858
1859         if (td->file_map)
1860                 memset(td->file_map, 0, td->num_maps * sizeof(long));
1861 }
1862
1863 static void update_rusage_stat(struct thread_data *td)
1864 {
1865         if (!(td->runtime[0] + td->runtime[1]))
1866                 return;
1867
1868         getrusage(RUSAGE_SELF, &td->ru_end);
1869
1870         td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1871         td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1872         td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1873
1874         
1875         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1876 }
1877
1878 static void *thread_main(void *data)
1879 {
1880         struct thread_data *td = data;
1881
1882         if (!td->use_thread)
1883                 setsid();
1884
1885         td->pid = getpid();
1886
1887         INIT_LIST_HEAD(&td->io_u_freelist);
1888         INIT_LIST_HEAD(&td->io_u_busylist);
1889         INIT_LIST_HEAD(&td->io_hist_list);
1890         INIT_LIST_HEAD(&td->io_log_list);
1891
1892         if (init_io_u(td))
1893                 goto err;
1894
1895         if (fio_setaffinity(td) == -1) {
1896                 td_verror(td, errno);
1897                 goto err;
1898         }
1899
1900         if (init_io(td))
1901                 goto err;
1902
1903         if (init_iolog(td))
1904                 goto err;
1905
1906         if (td->ioprio) {
1907                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1908                         td_verror(td, errno);
1909                         goto err;
1910                 }
1911         }
1912
1913         if (nice(td->nice) < 0) {
1914                 td_verror(td, errno);
1915                 goto err;
1916         }
1917
1918         if (init_random_state(td))
1919                 goto err;
1920
1921         if (td->ioscheduler && switch_ioscheduler(td))
1922                 goto err;
1923
1924         td_set_runstate(td, TD_INITIALIZED);
1925         sem_post(&startup_sem);
1926         sem_wait(&td->mutex);
1927
1928         if (!td->create_serialize && setup_file(td))
1929                 goto err;
1930
1931         gettimeofday(&td->epoch, NULL);
1932
1933         if (td->exec_prerun)
1934                 system(td->exec_prerun);
1935
1936         while (td->loops--) {
1937                 getrusage(RUSAGE_SELF, &td->ru_start);
1938                 gettimeofday(&td->start, NULL);
1939                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1940
1941                 if (td->ratemin)
1942                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1943
1944                 clear_io_state(td);
1945                 prune_io_piece_log(td);
1946
1947                 do_io(td);
1948
1949                 td->runtime[td->ddir] += mtime_since_now(&td->start);
1950                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
1951                         td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1952
1953                 update_rusage_stat(td);
1954
1955                 if (td->error || td->terminate)
1956                         break;
1957
1958                 if (td->verify == VERIFY_NONE)
1959                         continue;
1960
1961                 clear_io_state(td);
1962                 gettimeofday(&td->start, NULL);
1963
1964                 do_verify(td);
1965
1966                 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1967
1968                 if (td->error || td->terminate)
1969                         break;
1970         }
1971
1972         if (td->bw_log)
1973                 finish_log(td, td->bw_log, "bw");
1974         if (td->slat_log)
1975                 finish_log(td, td->slat_log, "slat");
1976         if (td->clat_log)
1977                 finish_log(td, td->clat_log, "clat");
1978         if (td->write_iolog)
1979                 write_iolog_close(td);
1980         if (td->exec_postrun)
1981                 system(td->exec_postrun);
1982
1983         if (exitall_on_terminate)
1984                 terminate_threads(td->groupid);
1985
1986 err:
1987         if (td->fd != -1) {
1988                 close(td->fd);
1989                 td->fd = -1;
1990         }
1991         if (td->mmap)
1992                 munmap(td->mmap, td->file_size);
1993         if (td->directory)
1994                 free(td->directory);
1995         if (td->iolog_file)
1996                 free(td->iolog_file);
1997         if (td->exec_prerun)
1998                 free(td->exec_prerun);
1999         if (td->exec_postrun)
2000                 free(td->exec_postrun);
2001         if (td->ioscheduler)
2002                 free(td->ioscheduler);
2003         if (td->sysfs_root)
2004                 free(td->sysfs_root);
2005         cleanup_io(td);
2006         cleanup_io_u(td);
2007         td_set_runstate(td, TD_EXITED);
2008         return NULL;
2009
2010 }
2011
2012 static void *fork_main(int shmid, int offset)
2013 {
2014         struct thread_data *td;
2015         void *data;
2016
2017         data = shmat(shmid, NULL, 0);
2018         if (data == (void *) -1) {
2019                 perror("shmat");
2020                 return NULL;
2021         }
2022
2023         td = data + offset * sizeof(struct thread_data);
2024         thread_main(td);
2025         shmdt(data);
2026         return NULL;
2027 }
2028
2029 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
2030                     double *mean, double *dev)
2031 {
2032         double n;
2033
2034         if (is->samples == 0)
2035                 return 0;
2036
2037         *min = is->min_val;
2038         *max = is->max_val;
2039
2040         n = (double) is->samples;
2041         *mean = (double) is->val / n;
2042         *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
2043         if (!(*min + *max) && !(*mean + *dev))
2044                 return 0;
2045
2046         return 1;
2047 }
2048
2049 static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
2050                              int ddir)
2051 {
2052         char *ddir_str[] = { "read ", "write" };
2053         unsigned long min, max;
2054         unsigned long long bw;
2055         double mean, dev;
2056
2057         if (!td->runtime[ddir])
2058                 return;
2059
2060         bw = td->io_bytes[ddir] / td->runtime[ddir];
2061         printf("  %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
2062
2063         if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
2064                 printf("    slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
2065
2066         if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
2067                 printf("    clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
2068
2069         if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
2070                 double p_of_agg;
2071
2072                 p_of_agg = mean * 100 / (double) rs->agg[ddir];
2073                 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);
2074         }
2075 }
2076
2077 static void show_thread_status(struct thread_data *td,
2078                                struct group_run_stats *rs)
2079 {
2080         double usr_cpu, sys_cpu;
2081
2082         if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
2083                 return;
2084
2085         printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
2086
2087         show_ddir_status(td, rs, td->ddir);
2088         if (td->io_bytes[td->ddir ^ 1])
2089                 show_ddir_status(td, rs, td->ddir ^ 1);
2090
2091         if (td->runtime[0] + td->runtime[1]) {
2092                 double runt = td->runtime[0] + td->runtime[1];
2093
2094                 usr_cpu = (double) td->usr_time * 100 / runt;
2095                 sys_cpu = (double) td->sys_time * 100 / runt;
2096         } else {
2097                 usr_cpu = 0;
2098                 sys_cpu = 0;
2099         }
2100
2101         printf("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
2102 }
2103
2104 static void check_str_update(struct thread_data *td)
2105 {
2106         char c = run_str[td->thread_number - 1];
2107
2108         if (td->runstate == td->old_runstate)
2109                 return;
2110
2111         switch (td->runstate) {
2112                 case TD_REAPED:
2113                         c = '_';
2114                         break;
2115                 case TD_EXITED:
2116                         c = 'E';
2117                         break;
2118                 case TD_RUNNING:
2119                         if (td_rw(td)) {
2120                                 if (td->sequential)
2121                                         c = 'M';
2122                                 else
2123                                         c = 'm';
2124                         } else if (td_read(td)) {
2125                                 if (td->sequential)
2126                                         c = 'R';
2127                                 else
2128                                         c = 'r';
2129                         } else {
2130                                 if (td->sequential)
2131                                         c = 'W';
2132                                 else
2133                                         c = 'w';
2134                         }
2135                         break;
2136                 case TD_VERIFYING:
2137                         c = 'V';
2138                         break;
2139                 case TD_CREATED:
2140                         c = 'C';
2141                         break;
2142                 case TD_INITIALIZED:
2143                         c = 'I';
2144                         break;
2145                 case TD_NOT_CREATED:
2146                         c = 'P';
2147                         break;
2148                 default:
2149                         printf("state %d\n", td->runstate);
2150         }
2151
2152         run_str[td->thread_number - 1] = c;
2153         td->old_runstate = td->runstate;
2154 }
2155
2156 static void eta_to_str(char *str, int eta_sec)
2157 {
2158         unsigned int d, h, m, s;
2159         static int always_d, always_h;
2160
2161         d = h = m = s = 0;
2162
2163         s = eta_sec % 60;
2164         eta_sec /= 60;
2165         m = eta_sec % 60;
2166         eta_sec /= 60;
2167         h = eta_sec % 24;
2168         eta_sec /= 24;
2169         d = eta_sec;
2170
2171         if (d || always_d) {
2172                 always_d = 1;
2173                 str += sprintf(str, "%02dd:", d);
2174         }
2175         if (h || always_h) {
2176                 always_h = 1;
2177                 str += sprintf(str, "%02dh:", h);
2178         }
2179
2180         str += sprintf(str, "%02dm:", m);
2181         str += sprintf(str, "%02ds", s);
2182 }
2183
2184 static int thread_eta(struct thread_data *td, unsigned long elapsed)
2185 {
2186         unsigned long long bytes_total, bytes_done;
2187         unsigned int eta_sec = 0;
2188
2189         bytes_total = td->total_io_size;
2190
2191         /*
2192          * if writing, bytes_total will be twice the size. If mixing,
2193          * assume a 50/50 split and thus bytes_total will be 50% larger.
2194          */
2195         if (td->verify) {
2196                 if (td_rw(td))
2197                         bytes_total = bytes_total * 3 / 2;
2198                 else
2199                         bytes_total <<= 1;
2200         }
2201         if (td->zone_size && td->zone_skip)
2202                 bytes_total /= (td->zone_skip / td->zone_size);
2203
2204         if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
2205                 double perc;
2206
2207                 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
2208                 perc = (double) bytes_done / (double) bytes_total;
2209                 if (perc > 1.0)
2210                         perc = 1.0;
2211
2212                 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
2213
2214                 if (td->timeout && eta_sec > (td->timeout - elapsed))
2215                         eta_sec = td->timeout - elapsed;
2216         } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
2217                         || td->runstate == TD_INITIALIZED) {
2218                 int t_eta = 0, r_eta = 0;
2219
2220                 /*
2221                  * We can only guess - assume it'll run the full timeout
2222                  * if given, otherwise assume it'll run at the specified rate.
2223                  */
2224                 if (td->timeout)
2225                         t_eta = td->timeout + td->start_delay - elapsed;
2226                 if (td->rate) {
2227                         r_eta = (bytes_total / 1024) / td->rate;
2228                         r_eta += td->start_delay - elapsed;
2229                 }
2230
2231                 if (r_eta && t_eta)
2232                         eta_sec = min(r_eta, t_eta);
2233                 else if (r_eta)
2234                         eta_sec = r_eta;
2235                 else if (t_eta)
2236                         eta_sec = t_eta;
2237                 else
2238                         eta_sec = INT_MAX;
2239         } else {
2240                 /*
2241                  * thread is already done
2242                  */
2243                 eta_sec = 0;
2244         }
2245
2246         return eta_sec;
2247 }
2248
2249 static void print_thread_status(void)
2250 {
2251         unsigned long elapsed = time_since_now(&genesis);
2252         int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
2253         char eta_str[32];
2254         double perc = 0.0;
2255
2256         eta_secs = malloc(thread_number * sizeof(int));
2257         memset(eta_secs, 0, thread_number * sizeof(int));
2258
2259         nr_running = t_rate = m_rate = 0;
2260         for (i = 0; i < thread_number; i++) {
2261                 struct thread_data *td = &threads[i];
2262
2263                 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
2264                         nr_running++;
2265                         t_rate += td->rate;
2266                         m_rate += td->ratemin;
2267                 }
2268
2269                 if (elapsed >= 3)
2270                         eta_secs[i] = thread_eta(td, elapsed);
2271                 else
2272                         eta_secs[i] = INT_MAX;
2273
2274                 check_str_update(td);
2275         }
2276
2277         if (exitall_on_terminate)
2278                 eta_sec = INT_MAX;
2279         else
2280                 eta_sec = 0;
2281
2282         for (i = 0; i < thread_number; i++) {
2283                 if (exitall_on_terminate) {
2284                         if (eta_secs[i] < eta_sec)
2285                                 eta_sec = eta_secs[i];
2286                 } else {
2287                         if (eta_secs[i] > eta_sec)
2288                                 eta_sec = eta_secs[i];
2289                 }
2290         }
2291
2292         if (eta_sec != INT_MAX && elapsed) {
2293                 perc = (double) elapsed / (double) (elapsed + eta_sec);
2294                 eta_to_str(eta_str, eta_sec);
2295         }
2296
2297         printf("Threads now running (%d)", nr_running);
2298         if (m_rate || t_rate)
2299                 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
2300         if (eta_sec != INT_MAX) {
2301                 perc *= 100.0;
2302                 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
2303         }
2304         printf("\r");
2305         fflush(stdout);
2306         free(eta_secs);
2307 }
2308
2309 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2310 {
2311         int i;
2312
2313         /*
2314          * reap exited threads (TD_EXITED -> TD_REAPED)
2315          */
2316         for (i = 0; i < thread_number; i++) {
2317                 struct thread_data *td = &threads[i];
2318
2319                 if (td->runstate != TD_EXITED)
2320                         continue;
2321
2322                 td_set_runstate(td, TD_REAPED);
2323
2324                 if (td->use_thread) {
2325                         long ret;
2326
2327                         if (pthread_join(td->thread, (void *) &ret))
2328                                 perror("thread_join");
2329                 } else
2330                         waitpid(td->pid, NULL, 0);
2331
2332                 (*nr_running)--;
2333                 (*m_rate) -= td->ratemin;
2334                 (*t_rate) -= td->rate;
2335         }
2336 }
2337
2338 static void fio_unpin_memory(void *pinned)
2339 {
2340         if (pinned) {
2341                 if (munlock(pinned, mlock_size) < 0)
2342                         perror("munlock");
2343                 munmap(pinned, mlock_size);
2344         }
2345 }
2346
2347 static void *fio_pin_memory(void)
2348 {
2349         long pagesize, pages;
2350         void *ptr;
2351
2352         if (!mlock_size)
2353                 return NULL;
2354
2355         /*
2356          * Don't allow mlock of more than real_mem-128MB
2357          */
2358         pagesize = sysconf(_SC_PAGESIZE);
2359         pages = sysconf(_SC_PHYS_PAGES);
2360         if (pages != -1 && pagesize != -1) {
2361                 unsigned long long real_mem = pages * pagesize;
2362
2363                 if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
2364                         mlock_size = real_mem - 128 * 1024 * 1024;
2365                         printf("fio: limiting mlocked memory to %lluMiB\n",
2366                                                         mlock_size >> 20);
2367                 }
2368         }
2369
2370         ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
2371         if (!ptr) {
2372                 perror("malloc locked mem");
2373                 return NULL;
2374         }
2375         if (mlock(ptr, mlock_size) < 0) {
2376                 munmap(ptr, mlock_size);
2377                 perror("mlock");
2378                 return NULL;
2379         }
2380
2381         return ptr;
2382 }
2383
2384 static void run_threads(void)
2385 {
2386         struct thread_data *td;
2387         unsigned long spent;
2388         int i, todo, nr_running, m_rate, t_rate, nr_started;
2389         void *mlocked_mem;
2390
2391         mlocked_mem = fio_pin_memory();
2392
2393         printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
2394         fflush(stdout);
2395
2396         signal(SIGINT, sig_handler);
2397         signal(SIGALRM, sig_handler);
2398
2399         todo = thread_number;
2400         nr_running = 0;
2401         nr_started = 0;
2402         m_rate = t_rate = 0;
2403
2404         for (i = 0; i < thread_number; i++) {
2405                 td = &threads[i];
2406
2407                 run_str[td->thread_number - 1] = 'P';
2408
2409                 init_disk_util(td);
2410
2411                 if (!td->create_serialize)
2412                         continue;
2413
2414                 /*
2415                  * do file setup here so it happens sequentially,
2416                  * we don't want X number of threads getting their
2417                  * client data interspersed on disk
2418                  */
2419                 if (setup_file(td)) {
2420                         td_set_runstate(td, TD_REAPED);
2421                         todo--;
2422                 }
2423         }
2424
2425         gettimeofday(&genesis, NULL);
2426
2427         while (todo) {
2428                 struct thread_data *map[MAX_JOBS];
2429                 struct timeval this_start;
2430                 int this_jobs = 0, left;
2431
2432                 /*
2433                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2434                  */
2435                 for (i = 0; i < thread_number; i++) {
2436                         td = &threads[i];
2437
2438                         if (td->runstate != TD_NOT_CREATED)
2439                                 continue;
2440
2441                         /*
2442                          * never got a chance to start, killed by other
2443                          * thread for some reason
2444                          */
2445                         if (td->terminate) {
2446                                 todo--;
2447                                 continue;
2448                         }
2449
2450                         if (td->start_delay) {
2451                                 spent = mtime_since_now(&genesis);
2452
2453                                 if (td->start_delay * 1000 > spent)
2454                                         continue;
2455                         }
2456
2457                         if (td->stonewall && (nr_started || nr_running))
2458                                 break;
2459
2460                         /*
2461                          * Set state to created. Thread will transition
2462                          * to TD_INITIALIZED when it's done setting up.
2463                          */
2464                         td_set_runstate(td, TD_CREATED);
2465                         map[this_jobs++] = td;
2466                         sem_init(&startup_sem, 0, 1);
2467                         nr_started++;
2468
2469                         if (td->use_thread) {
2470                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2471                                         perror("thread_create");
2472                                         nr_started--;
2473                                 }
2474                         } else {
2475                                 if (fork())
2476                                         sem_wait(&startup_sem);
2477                                 else {
2478                                         fork_main(shm_id, i);
2479                                         exit(0);
2480                                 }
2481                         }
2482                 }
2483
2484                 /*
2485                  * Wait for the started threads to transition to
2486                  * TD_INITIALIZED.
2487                  */
2488                 printf("fio: Waiting for threads to initialize...\n");
2489                 gettimeofday(&this_start, NULL);
2490                 left = this_jobs;
2491                 while (left) {
2492                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2493                                 break;
2494
2495                         usleep(100000);
2496
2497                         for (i = 0; i < this_jobs; i++) {
2498                                 td = map[i];
2499                                 if (!td)
2500                                         continue;
2501                                 if (td->runstate == TD_INITIALIZED) {
2502                                         map[i] = NULL;
2503                                         left--;
2504                                 } else if (td->runstate >= TD_EXITED) {
2505                                         map[i] = NULL;
2506                                         left--;
2507                                         todo--;
2508                                         nr_running++; /* work-around... */
2509                                 }
2510                         }
2511                 }
2512
2513                 if (left) {
2514                         fprintf(stderr, "fio: %d jobs failed to start\n", left);
2515                         for (i = 0; i < this_jobs; i++) {
2516                                 td = map[i];
2517                                 if (!td)
2518                                         continue;
2519                                 kill(td->pid, SIGTERM);
2520                         }
2521                         break;
2522                 }
2523
2524                 /*
2525                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2526                  */
2527                 printf("fio: Go for launch\n");
2528                 for (i = 0; i < thread_number; i++) {
2529                         td = &threads[i];
2530
2531                         if (td->runstate != TD_INITIALIZED)
2532                                 continue;
2533
2534                         td_set_runstate(td, TD_RUNNING);
2535                         nr_running++;
2536                         nr_started--;
2537                         m_rate += td->ratemin;
2538                         t_rate += td->rate;
2539                         todo--;
2540                         sem_post(&td->mutex);
2541                 }
2542
2543                 reap_threads(&nr_running, &t_rate, &m_rate);
2544
2545                 if (todo)
2546                         usleep(100000);
2547         }
2548
2549         while (nr_running) {
2550                 reap_threads(&nr_running, &t_rate, &m_rate);
2551                 usleep(10000);
2552         }
2553
2554         update_io_ticks();
2555         fio_unpin_memory(mlocked_mem);
2556 }
2557
2558 static void show_group_stats(struct group_run_stats *rs, int id)
2559 {
2560         printf("\nRun status group %d (all jobs):\n", id);
2561
2562         if (rs->max_run[DDIR_READ])
2563                 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]);
2564         if (rs->max_run[DDIR_WRITE])
2565                 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]);
2566 }
2567
2568 static void show_disk_util(void)
2569 {
2570         struct disk_util_stat *dus;
2571         struct list_head *entry;
2572         struct disk_util *du;
2573         double util;
2574
2575         printf("\nDisk stats (read/write):\n");
2576
2577         list_for_each(entry, &disk_list) {
2578                 du = list_entry(entry, struct disk_util, list);
2579                 dus = &du->dus;
2580
2581                 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2582                 if (util > 100.0)
2583                         util = 100.0;
2584
2585                 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);
2586         }
2587 }
2588
2589 static void show_run_stats(void)
2590 {
2591         struct group_run_stats *runstats, *rs;
2592         struct thread_data *td;
2593         int i;
2594
2595         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2596
2597         for (i = 0; i < groupid + 1; i++) {
2598                 rs = &runstats[i];
2599
2600                 memset(rs, 0, sizeof(*rs));
2601                 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2602                 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2603         }
2604
2605         for (i = 0; i < thread_number; i++) {
2606                 unsigned long long rbw, wbw;
2607
2608                 td = &threads[i];
2609
2610                 if (td->error) {
2611                         printf("Client%d: %s\n", td->thread_number, td->verror);
2612                         continue;
2613                 }
2614
2615                 rs = &runstats[td->groupid];
2616
2617                 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2618                         rs->min_run[0] = td->runtime[0];
2619                 if (td->runtime[0] > rs->max_run[0])
2620                         rs->max_run[0] = td->runtime[0];
2621                 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2622                         rs->min_run[1] = td->runtime[1];
2623                 if (td->runtime[1] > rs->max_run[1])
2624                         rs->max_run[1] = td->runtime[1];
2625
2626                 rbw = wbw = 0;
2627                 if (td->runtime[0])
2628                         rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
2629                 if (td->runtime[1])
2630                         wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
2631
2632                 if (rbw < rs->min_bw[0])
2633                         rs->min_bw[0] = rbw;
2634                 if (wbw < rs->min_bw[1])
2635                         rs->min_bw[1] = wbw;
2636                 if (rbw > rs->max_bw[0])
2637                         rs->max_bw[0] = rbw;
2638                 if (wbw > rs->max_bw[1])
2639                         rs->max_bw[1] = wbw;
2640
2641                 rs->io_mb[0] += td->io_bytes[0] >> 20;
2642                 rs->io_mb[1] += td->io_bytes[1] >> 20;
2643         }
2644
2645         for (i = 0; i < groupid + 1; i++) {
2646                 rs = &runstats[i];
2647
2648                 if (rs->max_run[0])
2649                         rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2650                 if (rs->max_run[1])
2651                         rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2652         }
2653
2654         /*
2655          * don't overwrite last signal output
2656          */
2657         printf("\n");
2658
2659         for (i = 0; i < thread_number; i++) {
2660                 td = &threads[i];
2661                 rs = &runstats[td->groupid];
2662
2663                 show_thread_status(td, rs);
2664         }
2665
2666         for (i = 0; i < groupid + 1; i++)
2667                 show_group_stats(&runstats[i], i);
2668
2669         show_disk_util();
2670 }
2671
2672 int main(int argc, char *argv[])
2673 {
2674         if (parse_options(argc, argv))
2675                 return 1;
2676
2677         if (!thread_number) {
2678                 printf("Nothing to do\n");
2679                 return 1;
2680         }
2681
2682         disk_util_timer_arm();
2683
2684         run_threads();
2685         show_run_stats();
2686
2687         return 0;
2688 }