[PATCH] fio: make random io writes verifiable as well
[disktools.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 <ctype.h>
30 #include <sched.h>
31 #include <libaio.h>
32 #include <math.h>
33 #include <limits.h>
34 #include <sys/time.h>
35 #include <sys/types.h>
36 #include <sys/stat.h>
37 #include <sys/wait.h>
38 #include <semaphore.h>
39 #include <sys/ipc.h>
40 #include <sys/shm.h>
41 #include <asm/unistd.h>
42 #include <asm/types.h>
43
44 #include "list.h"
45 #include "md5.h"
46
47 #define MAX_JOBS        (1024)
48
49 /*
50  * assume we don't have _get either, if _set isn't defined
51  */
52 #ifndef __NR_ioprio_set
53 #if defined(__i386__)
54 #define __NR_ioprio_set         289
55 #define __NR_ioprio_get         290
56 #elif defined(__powerpc__) || defined(__powerpc64__)
57 #define __NR_ioprio_set         273
58 #define __NR_ioprio_get         274
59 #elif defined(__x86_64__)
60 #define __NR_ioprio_set         251
61 #define __NR_ioprio_get         252
62 #elif defined(__ia64__)
63 #define __NR_ioprio_set         1274
64 #define __NR_ioprio_get         1275
65 #elif defined(__alpha__)
66 #define __NR_ioprio_set         442
67 #define __NR_ioprio_get         443
68 #elif defined(__s390x__) || defined(__s390__)
69 #define __NR_ioprio_set         282
70 #define __NR_ioprio_get         283
71 #else
72 #error "Unsupported arch"
73 #endif
74 #endif
75
76 #ifndef __NR_fadvise64
77 #if defined(__i386__)
78 #define __NR_fadvise64          250
79 #elif defined(__powerpc__) || defined(__powerpc64__)
80 #define __NR_fadvise64          233
81 #elif defined(__x86_64__)
82 #define __NR_fadvise64          221
83 #elif defined(__ia64__)
84 #define __NR_fadvise64          1234
85 #elif defined(__alpha__)
86 #define __NR_fadvise64          413
87 #elif defined(__s390x__) || defined(__s390__)
88 #define __NR_fadvise64          253
89 #else
90 #error "Unsupported arch"
91 #endif
92 #endif
93
94 static int ioprio_set(int which, int who, int ioprio)
95 {
96         return syscall(__NR_ioprio_set, which, who, ioprio);
97 }
98
99 /*
100  * we want fadvise64 really, but it's so tangled... later
101  */
102 static int fadvise(int fd, loff_t offset, size_t len, int advice)
103 {
104 #if 0
105         return syscall(__NR_fadvise64, fd, offset, offset >> 32, len, advice);
106 #else
107         return posix_fadvise(fd, (off_t) offset, len, advice);
108 #endif
109 }
110
111 enum {
112         IOPRIO_WHO_PROCESS = 1,
113         IOPRIO_WHO_PGRP,
114         IOPRIO_WHO_USER,
115 };
116
117 #define IOPRIO_CLASS_SHIFT      13
118
119 #define MASK    (4095)
120
121 #define DEF_BS          (4096)
122 #define DEF_TIMEOUT     (30)
123 #define DEF_RATE_CYCLE  (1000)
124 #define DEF_ODIRECT     (1)
125 #define DEF_SEQUENTIAL  (1)
126 #define DEF_RAND_REPEAT (1)
127 #define DEF_OVERWRITE   (0)
128 #define DEF_CREATE      (1)
129 #define DEF_INVALIDATE  (1)
130 #define DEF_SYNCIO      (0)
131 #define DEF_RANDSEED    (0xb1899bedUL)
132 #define DEF_BWAVGTIME   (500)
133 #define DEF_CREATE_SER  (1)
134 #define DEF_CREATE_FSYNC        (1)
135 #define DEF_LOOPS       (1)
136 #define DEF_VERIFY      (0)
137
138 #define ALIGN(buf)      (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
139
140 static int repeatable = DEF_RAND_REPEAT;
141 static int rate_quit = 1;
142 static int write_lat_log;
143 static int write_bw_log;
144 static int exitall_on_terminate;
145
146 static int thread_number;
147 static char *ini_file;
148
149 static int max_jobs = MAX_JOBS;
150
151 static char run_str[MAX_JOBS + 1];
152
153 static int shm_id;
154
155 enum {
156         DDIR_READ = 0,
157         DDIR_WRITE,
158 };
159
160 /*
161  * thread life cycle
162  */
163 enum {
164         TD_NOT_CREATED = 0,
165         TD_CREATED,
166         TD_RUNNING,
167         TD_VERIFYING,
168         TD_EXITED,
169         TD_REAPED,
170 };
171
172 enum {
173         MEM_MALLOC,
174         MEM_SHM,
175 };
176
177 /*
178  * The io unit
179  */
180 struct io_u {
181         struct iocb iocb;
182         struct timeval start_time;
183         struct timeval issue_time;
184
185         char *buf;
186         unsigned int buflen;
187         unsigned long long offset;
188
189         struct list_head list;
190 };
191
192 struct io_stat {
193         unsigned long val;
194         unsigned long val_sq;
195         unsigned long max_val;
196         unsigned long min_val;
197         unsigned long samples;
198 };
199
200 struct io_sample {
201         unsigned long time;
202         unsigned long val;
203 };
204
205 struct io_log {
206         unsigned long nr_samples;
207         unsigned long max_samples;
208         struct io_sample *log;
209 };
210
211 struct io_piece {
212         struct list_head list;
213         unsigned long long offset;
214         unsigned int len;
215 };
216
217 #define FIO_HDR_MAGIC   0xf00baaef
218
219 struct verify_header {
220         unsigned int fio_magic;
221         unsigned int len;
222         char md5_digest[MD5_HASH_WORDS * 4];
223 };
224
225 #define td_read(td)             ((td)->ddir == DDIR_READ)
226 #define should_fsync(td)        (!td_read(td) && !(td)->odirect)
227
228 struct thread_data {
229         char file_name[256];
230         int thread_number;
231         int error;
232         int fd;
233         pid_t pid;
234         char *orig_buffer;
235         volatile int terminate;
236         volatile int runstate;
237         volatile int old_runstate;
238         unsigned int ddir;
239         unsigned int ioprio;
240         unsigned int sequential;
241         unsigned int bs;
242         unsigned int min_bs;
243         unsigned int max_bs;
244         unsigned int odirect;
245         unsigned int thinktime;
246         unsigned int fsync_blocks;
247         unsigned int start_delay;
248         unsigned int timeout;
249         unsigned int use_aio;
250         unsigned int create_file;
251         unsigned int overwrite;
252         unsigned int invalidate_cache;
253         unsigned int bw_avg_time;
254         unsigned int create_serialize;
255         unsigned int create_fsync;
256         unsigned int loops;
257         unsigned long long file_size;
258         unsigned long long file_offset;
259         unsigned int sync_io;
260         unsigned int mem_type;
261         unsigned int verify;
262         cpu_set_t cpumask;
263
264         struct drand48_data bsrange_state;
265         struct drand48_data verify_state;
266
267         int shm_id;
268
269         off_t cur_off;
270
271         io_context_t aio_ctx;
272         unsigned int aio_depth;
273         struct io_event *aio_events;
274
275         unsigned int cur_depth;
276         struct list_head io_u_freelist;
277         struct list_head io_u_busylist;
278
279         unsigned int rate;
280         unsigned int ratemin;
281         unsigned int ratecycle;
282         unsigned long rate_usec_cycle;
283         long rate_pending_usleep;
284         unsigned long rate_kb;
285         struct timeval lastrate;
286
287         unsigned long runtime;          /* sec */
288         unsigned long kb;
289         unsigned long io_blocks;
290         unsigned long io_kb;
291         unsigned long this_io_kb;
292         unsigned long last_kb;
293         sem_t mutex;
294         struct drand48_data random_state;
295
296         /*
297          * bandwidth and latency stats
298          */
299         struct io_stat clat_stat;               /* completion latency */
300         struct io_stat slat_stat;               /* submission latency */
301
302         struct io_stat bw_stat;                 /* bandwidth stats */
303         unsigned long stat_io_kb;
304         struct timeval stat_sample_time;
305
306         struct io_log *lat_log;
307         struct io_log *bw_log;
308
309         struct timeval start;
310
311         struct list_head io_hist_list;
312 };
313
314 static struct thread_data *threads;
315 static struct thread_data def_thread;
316
317 static sem_t startup_sem;
318
319 static void sig_handler(int sig)
320 {
321         int i;
322
323         for (i = 0; i < thread_number; i++) {
324                 struct thread_data *td = &threads[i];
325
326                 td->terminate = 1;
327                 td->start_delay = 0;
328         }
329 }
330
331 static int init_random_state(struct thread_data *td)
332 {
333         unsigned long seed;
334         int fd;
335
336         fd = open("/dev/random", O_RDONLY);
337         if (fd == -1) {
338                 td->error = errno;
339                 return 1;
340         }
341
342         if (read(fd, &seed, sizeof(seed)) < (int) sizeof(seed)) {
343                 td->error = EIO;
344                 close(fd);
345                 return 1;
346         }
347
348         close(fd);
349
350         srand48_r(seed, &td->bsrange_state);
351         srand48_r(seed, &td->verify_state);
352
353         if (td->sequential)
354                 return 0;
355
356         if (repeatable)
357                 seed = DEF_RANDSEED;
358
359         srand48_r(seed, &td->random_state);
360         return 0;
361 }
362
363 static unsigned long utime_since(struct timeval *s, struct timeval *e)
364 {
365         double sec, usec;
366
367         sec = e->tv_sec - s->tv_sec;
368         usec = e->tv_usec - s->tv_usec;
369         if (sec > 0 && usec < 0) {
370                 sec--;
371                 usec += 1000000;
372         }
373
374         sec *= (double) 1000000;
375
376         return sec + usec;
377 }
378
379 static unsigned long mtime_since(struct timeval *s, struct timeval *e)
380 {
381         double sec, usec;
382
383         sec = e->tv_sec - s->tv_sec;
384         usec = e->tv_usec - s->tv_usec;
385         if (sec > 0 && usec < 0) {
386                 sec--;
387                 usec += 1000000;
388         }
389
390         sec *= (double) 1000;
391         usec /= (double) 1000;
392
393         return sec + usec;
394 }
395
396 static unsigned long mtime_since_now(struct timeval *s)
397 {
398         struct timeval t;
399
400         gettimeofday(&t, NULL);
401         return mtime_since(s, &t);
402 }
403
404 static inline unsigned long msec_now(struct timeval *s)
405 {
406         return s->tv_sec * 1000 + s->tv_usec / 1000;
407 }
408
409 static unsigned long long get_next_offset(struct thread_data *td)
410 {
411         unsigned long long kb;
412         long r;
413
414         if (!td->sequential) {
415                 int min_bs_kb = td->min_bs >> 10;
416
417                 lrand48_r(&td->random_state, &r);
418                 kb = (1+(double) (td->kb-1) * r / (RAND_MAX+1.0));
419                 kb = (kb + min_bs_kb - 1) & ~(min_bs_kb - 1);
420         } else
421                 kb = td->last_kb;
422
423         return (kb << 10) + td->file_offset;
424 }
425
426 static unsigned int get_next_buflen(struct thread_data *td)
427 {
428         unsigned int buflen;
429         long r;
430
431         if (td->min_bs == td->max_bs)
432                 buflen = td->min_bs;
433         else {
434                 lrand48_r(&td->bsrange_state, &r);
435                 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
436                 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
437         }
438
439         if (buflen > ((td->kb - td->this_io_kb) << 10))
440                 buflen = (td->kb - td->this_io_kb) << 10;
441
442         td->last_kb += buflen >> 10;
443         return buflen;
444 }
445
446 static inline void add_stat_sample(struct thread_data *td, struct io_stat *is,
447                                    unsigned long val)
448 {
449         if (val > is->max_val)
450                 is->max_val = val;
451         if (val < is->min_val)
452                 is->min_val = val;
453
454         is->val += val;
455         is->val_sq += val * val;
456         is->samples++;
457 }
458
459 static void add_log_sample(struct thread_data *td, struct io_log *log,
460                            unsigned long val)
461 {
462         if (log->nr_samples == log->max_samples) {
463                 int new_size = sizeof(struct io_sample) * log->max_samples * 2;
464
465                 log->log = realloc(log->log, new_size);
466                 log->max_samples <<= 1;
467         }
468
469         log->log[log->nr_samples].val = val;
470         log->log[log->nr_samples].time = mtime_since_now(&td->start);
471         log->nr_samples++;
472 }
473
474 static void add_clat_sample(struct thread_data *td, unsigned long msec)
475 {
476         add_stat_sample(td, &td->clat_stat, msec);
477
478         if (td->lat_log)
479                 add_log_sample(td, td->lat_log, msec);
480 }
481
482 static void add_slat_sample(struct thread_data *td, unsigned long msec)
483 {
484         add_stat_sample(td, &td->slat_stat, msec);
485 }
486
487 static void add_bw_sample(struct thread_data *td)
488 {
489         unsigned long spent = mtime_since_now(&td->stat_sample_time);
490         unsigned long rate;
491
492         if (spent < td->bw_avg_time)
493                 return;
494
495         rate = ((td->this_io_kb - td->stat_io_kb) * 1024) / spent;
496         add_stat_sample(td, &td->bw_stat, rate);
497
498         if (td->bw_log)
499                 add_log_sample(td, td->bw_log, rate);
500
501         gettimeofday(&td->stat_sample_time, NULL);
502         td->stat_io_kb = td->this_io_kb;
503 }
504
505 static void usec_sleep(int usec)
506 {
507         struct timespec req = { .tv_sec = 0, .tv_nsec = usec * 1000 };
508         struct timespec rem;
509
510         do {
511                 rem.tv_sec = rem.tv_nsec = 0;
512                 nanosleep(&req, &rem);
513                 if (!rem.tv_nsec)
514                         break;
515
516                 req.tv_nsec = rem.tv_nsec;
517         } while (1);
518 }
519
520 static void rate_throttle(struct thread_data *td, unsigned long time_spent,
521                           unsigned int bytes)
522 {
523         unsigned long usec_cycle;
524
525         if (!td->rate)
526                 return;
527
528         usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs);
529
530         if (time_spent < usec_cycle) {
531                 unsigned long s = usec_cycle - time_spent;
532
533                 td->rate_pending_usleep += s;
534                 if (td->rate_pending_usleep >= 100000) {
535                         usec_sleep(td->rate_pending_usleep);
536                         td->rate_pending_usleep = 0;
537                 }
538         } else {
539                 long overtime = time_spent - usec_cycle;
540
541                 td->rate_pending_usleep -= overtime;
542         }
543 }
544
545 static int check_min_rate(struct thread_data *td, struct timeval *now)
546 {
547         unsigned long spent;
548         unsigned long rate;
549
550         /*
551          * allow a 2 second settle period in the beginning
552          */
553         if (mtime_since(&td->start, now) < 2000)
554                 return 0;
555
556         /*
557          * if rate blocks is set, sample is running
558          */
559         if (td->rate_kb) {
560                 spent = mtime_since(&td->lastrate, now);
561                 if (spent < td->ratecycle)
562                         return 0;
563
564                 rate = ((td->this_io_kb - td->rate_kb) * 1024) / spent;
565                 if (rate < td->ratemin) {
566                         printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
567                         if (rate_quit)
568                                 sig_handler(0);
569                         return 1;
570                 }
571         }
572
573         td->rate_kb = td->this_io_kb;
574         memcpy(&td->lastrate, now, sizeof(*now));
575         return 0;
576 }
577
578 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
579 {
580         if (mtime_since(&td->start, t) >= td->timeout * 1000)
581                 return 1;
582
583         return 0;
584 }
585
586 static void fill_random_bytes(struct thread_data *td,
587                               unsigned char *p, unsigned int len)
588 {
589         unsigned int todo;
590         double r;
591
592         while (len) {
593                 drand48_r(&td->verify_state, &r);
594
595                 /*
596                  * lrand48_r seems to be broken and only fill the bottom
597                  * 32-bits, even on 64-bit archs with 64-bit longs
598                  */
599                 todo = sizeof(r);
600                 if (todo > len)
601                         todo = len;
602
603                 memcpy(p, &r, todo);
604
605                 len -= todo;
606                 p += todo;
607         }
608 }
609
610 static int verify_io_u(struct io_u *io_u)
611 {
612         struct verify_header *hdr = (struct verify_header *) io_u->buf;
613         unsigned char *p = (unsigned char *) io_u->buf;
614         struct md5_ctx md5_ctx;
615
616         if (hdr->fio_magic != FIO_HDR_MAGIC)
617                 return 1;
618
619         memset(&md5_ctx, 0, sizeof(md5_ctx));
620         p += sizeof(*hdr);
621         md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
622
623         return memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
624 }
625
626 /*
627  * fill body of io_u->buf with random data and add a header with the
628  * (eg) sha1sum of that data.
629  */
630 static void populate_io_u(struct thread_data *td, struct io_u *io_u)
631 {
632         struct md5_ctx md5_ctx;
633         struct verify_header hdr;
634         unsigned char *p = (unsigned char *) io_u->buf;
635
636         hdr.fio_magic = FIO_HDR_MAGIC;
637         hdr.len = io_u->buflen;
638         p += sizeof(hdr);
639         fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
640
641         memset(&md5_ctx, 0, sizeof(md5_ctx));
642         md5_update(&md5_ctx, p, io_u->buflen - sizeof(hdr));
643         memcpy(hdr.md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
644         memcpy(io_u->buf, &hdr, sizeof(hdr));
645 }
646
647 static void put_io_u(struct thread_data *td, struct io_u *io_u)
648 {
649         list_del(&io_u->list);
650         list_add(&io_u->list, &td->io_u_freelist);
651         td->cur_depth--;
652 }
653
654 #define queue_full(td)  (list_empty(&(td)->io_u_freelist))
655
656 static struct io_u *__get_io_u(struct thread_data *td)
657 {
658         struct io_u *io_u;
659
660         if (queue_full(td))
661                 return NULL;
662
663         io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
664         list_del(&io_u->list);
665         list_add(&io_u->list, &td->io_u_busylist);
666         td->cur_depth++;
667         return io_u;
668 }
669
670 static struct io_u *get_io_u(struct thread_data *td)
671 {
672         struct io_u *io_u;
673
674         io_u = __get_io_u(td);
675         if (!io_u)
676                 return NULL;
677
678         io_u->offset = get_next_offset(td);
679         io_u->buflen = get_next_buflen(td);
680         if (!io_u->buflen) {
681                 put_io_u(td, io_u);
682                 return NULL;
683         }
684
685         if (!td_read(td) && td->verify)
686                 populate_io_u(td, io_u);
687
688         if (td->use_aio) {
689                 if (td_read(td))
690                         io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
691                 else
692                         io_prep_pwrite(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
693         }
694
695         gettimeofday(&io_u->start_time, NULL);
696         return io_u;
697 }
698
699 static inline void td_set_runstate(struct thread_data *td, int runstate)
700 {
701         td->old_runstate = td->runstate;
702         td->runstate = runstate;
703 }
704
705 static int get_next_verify(struct thread_data *td,
706                            unsigned long long *offset, unsigned int *len)
707 {
708         struct io_piece *ipo;
709
710         if (list_empty(&td->io_hist_list))
711                 return 1;
712
713         ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
714         list_del(&ipo->list);
715
716         *offset = ipo->offset;
717         *len = ipo->len;
718         free(ipo);
719         return 0;
720 }
721
722 static void do_sync_verify(struct thread_data *td)
723 {
724         struct timeval t;
725         struct io_u *io_u = NULL;
726         int ret;
727
728         td_set_runstate(td, TD_VERIFYING);
729
730         io_u = __get_io_u(td);
731
732         if (!td->odirect) {
733                 unsigned long size = td->kb << 10;
734
735                 if (fadvise(td->fd, 0, size, POSIX_FADV_DONTNEED) < 0) {
736                         td->error = errno;
737                         goto out;
738                 }
739         }
740
741         do {
742                 if (td->terminate)
743                         break;
744
745                 gettimeofday(&t, NULL);
746                 if (runtime_exceeded(td, &t))
747                         break;
748
749                 if (get_next_verify(td, &io_u->offset, &io_u->buflen))
750                         break;
751
752                 if (td->cur_off != io_u->offset) {
753                         if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
754                                 td->error = errno;
755                                 break;
756                         }
757                 }
758
759                 ret = read(td->fd, io_u->buf, io_u->buflen);
760                 if (ret < (int) io_u->buflen) {
761                         if (ret == -1) {
762                                 td->error = errno;
763                                 break;
764                         } else if (!ret)
765                                 break;
766                         else
767                                 io_u->buflen = ret;
768                 }
769
770                 if (verify_io_u(io_u))
771                         break;
772
773                 td->cur_off = io_u->offset + io_u->buflen;
774         } while (1);
775
776 out:
777         td_set_runstate(td, TD_RUNNING);
778         put_io_u(td, io_u);
779 }
780
781 /*
782  * log a succesful write, so we can unwind the log for verify
783  */
784 static void log_io_piece(struct thread_data *td, struct io_u *io_u)
785 {
786         struct io_piece *ipo = malloc(sizeof(*ipo));
787         struct list_head *entry;
788
789         INIT_LIST_HEAD(&ipo->list);
790         ipo->offset = io_u->offset;
791         ipo->len = io_u->buflen;
792
793         if (td->sequential) {
794                 list_add_tail(&ipo->list, &td->io_hist_list);
795                 return;
796         }
797
798         /*
799          * for random io, sort the list so verify will run faster
800          */
801         entry = &td->io_hist_list;
802         while ((entry = entry->prev) != &td->io_hist_list) {
803                 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
804
805                 if (__ipo->offset == ipo->offset &&
806                     __ipo->len == ipo->len) {
807                         free(ipo);
808                         ipo = NULL;
809                         break;
810                 } else if (__ipo->offset < ipo->offset)
811                         break;
812         }
813
814         if (ipo)
815                 list_add(&ipo->list, entry);
816 }
817
818 static void do_sync_io(struct thread_data *td)
819 {
820         unsigned long msec, usec;
821         struct io_u *io_u = NULL;
822         struct timeval e;
823
824         while (td->this_io_kb < td->kb) {
825                 int ret;
826
827                 if (td->terminate)
828                         break;
829
830                 io_u = get_io_u(td);
831                 if (!io_u)
832                         break;
833
834                 if (td->cur_off != io_u->offset) {
835                         if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
836                                 td->error = errno;
837                                 break;
838                         }
839                 }
840
841                 if (td_read(td))
842                         ret = read(td->fd, io_u->buf, io_u->buflen);
843                 else
844                         ret = write(td->fd, io_u->buf, io_u->buflen);
845
846                 if (ret < (int) io_u->buflen) {
847                         if (ret == -1)
848                                 td->error = errno;
849                         break;
850                 }
851
852                 if (!td_read(td))
853                         log_io_piece(td, io_u);
854
855                 td->io_blocks++;
856                 td->io_kb += io_u->buflen >> 10;
857                 td->this_io_kb += io_u->buflen >> 10;
858                 td->cur_off = io_u->offset + io_u->buflen;
859
860                 gettimeofday(&e, NULL);
861
862                 usec = utime_since(&io_u->start_time, &e);
863
864                 rate_throttle(td, usec, io_u->buflen);
865
866                 if (check_min_rate(td, &e)) {
867                         td->error = ENODATA;
868                         break;
869                 }
870
871                 msec = usec / 1000;
872                 add_clat_sample(td, msec);
873                 add_bw_sample(td);
874
875                 if (runtime_exceeded(td, &e))
876                         break;
877
878                 put_io_u(td, io_u);
879                 io_u = NULL;
880
881                 if (td->thinktime)
882                         usec_sleep(td->thinktime);
883
884                 if (should_fsync(td) && td->fsync_blocks &&
885                     (td->io_blocks % td->fsync_blocks) == 0)
886                         fsync(td->fd);
887         }
888
889         if (io_u)
890                 put_io_u(td, io_u);
891
892         if (should_fsync(td))
893                 fsync(td->fd);
894 }
895
896 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
897 {
898         struct iocb *iocb = &io_u->iocb;
899         int ret;
900
901         do {
902                 ret = io_submit(td->aio_ctx, 1, &iocb);
903                 if (ret == 1)
904                         return 0;
905                 else if (ret == EAGAIN)
906                         usleep(100);
907                 else if (ret == EINTR)
908                         continue;
909                 else
910                         break;
911         } while (1);
912
913         return ret;
914 }
915
916 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
917 #define ev_to_iou(ev)   (struct io_u *) ((unsigned long) (ev)->obj)
918
919 static int ios_completed(struct thread_data *td, int nr)
920 {
921         unsigned long msec;
922         struct io_u *io_u;
923         struct timeval e;
924         int i, bytes_done;
925
926         gettimeofday(&e, NULL);
927
928         for (i = 0, bytes_done = 0; i < nr; i++) {
929                 io_u = ev_to_iou(td->aio_events + i);
930
931                 td->io_blocks++;
932                 td->io_kb += io_u->buflen >> 10;
933                 td->this_io_kb += io_u->buflen >> 10;
934
935                 msec = mtime_since(&io_u->issue_time, &e);
936
937                 add_clat_sample(td, msec);
938                 add_bw_sample(td);
939
940                 if (!td_read(td))
941                         log_io_piece(td, io_u);
942
943                 bytes_done += io_u->buflen;
944                 put_io_u(td, io_u);
945         }
946
947         return bytes_done;
948 }
949
950 static void cleanup_pending_aio(struct thread_data *td)
951 {
952         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
953         struct list_head *entry, *n;
954         struct io_u *io_u;
955         int r;
956
957         /*
958          * get immediately available events, if any
959          */
960         r = io_getevents(td->aio_ctx, 0, td->cur_depth, td->aio_events, &ts);
961         if (r > 0)
962                 ios_completed(td, r);
963
964         /*
965          * now cancel remaining active events
966          */
967         list_for_each_safe(entry, n, &td->io_u_busylist) {
968                 io_u = list_entry(entry, struct io_u, list);
969
970                 r = io_cancel(td->aio_ctx, &io_u->iocb, td->aio_events);
971                 if (!r)
972                         put_io_u(td, io_u);
973         }
974
975         if (td->cur_depth) {
976                 r = io_getevents(td->aio_ctx, td->cur_depth, td->cur_depth, td->aio_events, NULL);
977                 if (r > 0)
978                         ios_completed(td, r);
979         }
980 }
981
982 static int async_do_verify(struct thread_data *td, struct io_u **io_u)
983 {
984         struct io_u *v_io_u = *io_u;
985         int ret = 0;
986
987         if (v_io_u) {
988                 ret = verify_io_u(v_io_u);
989                 put_io_u(td, v_io_u);
990                 *io_u = NULL;
991         }
992
993         return ret;
994 }
995
996 static void do_async_verify(struct thread_data *td)
997 {
998         struct timeval t;
999         struct io_u *io_u, *v_io_u = NULL;
1000         int ret;
1001
1002         td_set_runstate(td, TD_VERIFYING);
1003
1004         do {
1005                 if (td->terminate)
1006                         break;
1007
1008                 gettimeofday(&t, NULL);
1009                 if (runtime_exceeded(td, &t))
1010                         break;
1011
1012                 io_u = __get_io_u(td);
1013                 if (!io_u)
1014                         break;
1015
1016                 if (get_next_verify(td, &io_u->offset, &io_u->buflen)) {
1017                         put_io_u(td, io_u);
1018                         break;
1019                 }
1020
1021                 io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
1022                 ret = io_u_queue(td, io_u);
1023                 if (ret) {
1024                         put_io_u(td, io_u);
1025                         td->error = ret;
1026                         break;
1027                 }
1028
1029                 /*
1030                  * we have one pending to verify, do that while the next
1031                  * we are doing io on the next one
1032                  */
1033                 if (async_do_verify(td, &v_io_u))
1034                         break;
1035
1036                 ret = io_getevents(td->aio_ctx, 1, 1, td->aio_events, NULL);
1037                 if (ret != 1) {
1038                         if (ret < 0)
1039                                 td->error = ret;
1040                         break;
1041                 }
1042
1043                 v_io_u = ev_to_iou(td->aio_events);
1044
1045                 td->cur_off = v_io_u->offset + v_io_u->buflen;
1046
1047                 /*
1048                  * if we can't submit more io, we need to verify now
1049                  */
1050                 if (queue_full(td) && async_do_verify(td, &v_io_u))
1051                         break;
1052
1053         } while (1);
1054
1055         async_do_verify(td, &v_io_u);
1056
1057         if (td->cur_depth)
1058                 cleanup_pending_aio(td);
1059
1060         td_set_runstate(td, TD_RUNNING);
1061 }
1062
1063 static void do_async_io(struct thread_data *td)
1064 {
1065         struct timeval s, e;
1066         unsigned long usec;
1067
1068         while (td->this_io_kb < td->kb) {
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                 unsigned int bytes_done;
1074
1075                 if (td->terminate)
1076                         break;
1077
1078                 io_u = get_io_u(td);
1079                 if (!io_u)
1080                         break;
1081
1082                 memcpy(&s, &io_u->start_time, sizeof(s));
1083
1084                 ret = io_u_queue(td, io_u);
1085                 if (ret) {
1086                         put_io_u(td, io_u);
1087                         td->error = ret;
1088                         break;
1089                 }
1090
1091                 gettimeofday(&io_u->issue_time, NULL);
1092                 add_slat_sample(td, mtime_since(&io_u->start_time, &io_u->issue_time));
1093                 if (td->cur_depth < td->aio_depth) {
1094                         timeout = &ts;
1095                         min_evts = 0;
1096                 } else {
1097                         timeout = NULL;
1098                         min_evts = 1;
1099                 }
1100
1101                 ret = io_getevents(td->aio_ctx, min_evts, td->cur_depth, td->aio_events, timeout);
1102                 if (ret < 0) {
1103                         td->error = errno;
1104                         break;
1105                 } else if (!ret)
1106                         continue;
1107
1108                 bytes_done = ios_completed(td, ret);
1109
1110                 /*
1111                  * the rate is batched for now, it should work for batches
1112                  * of completions except the very first one which may look
1113                  * a little bursty
1114                  */
1115                 gettimeofday(&e, NULL);
1116                 usec = utime_since(&s, &e);
1117
1118                 rate_throttle(td, usec, bytes_done);
1119
1120                 if (check_min_rate(td, &e)) {
1121                         td->error = ENODATA;
1122                         break;
1123                 }
1124
1125                 if (runtime_exceeded(td, &e))
1126                         break;
1127
1128                 if (td->thinktime)
1129                         usec_sleep(td->thinktime);
1130
1131                 if (should_fsync(td) && td->fsync_blocks &&
1132                     (td->io_blocks % td->fsync_blocks) == 0)
1133                         fsync(td->fd);
1134         }
1135
1136         if (td->cur_depth)
1137                 cleanup_pending_aio(td);
1138
1139         if (should_fsync(td))
1140                 fsync(td->fd);
1141 }
1142
1143 static void cleanup_aio(struct thread_data *td)
1144 {
1145         io_destroy(td->aio_ctx);
1146
1147         if (td->aio_events)
1148                 free(td->aio_events);
1149 }
1150
1151 static int init_aio(struct thread_data *td)
1152 {
1153         if (io_queue_init(td->aio_depth, &td->aio_ctx)) {
1154                 td->error = errno;
1155                 return 1;
1156         }
1157
1158         td->aio_events = malloc(td->aio_depth * sizeof(struct io_event));
1159         return 0;
1160 }
1161
1162 static void cleanup_io_u(struct thread_data *td)
1163 {
1164         struct list_head *entry, *n;
1165         struct io_u *io_u;
1166
1167         list_for_each_safe(entry, n, &td->io_u_freelist) {
1168                 io_u = list_entry(entry, struct io_u, list);
1169
1170                 list_del(&io_u->list);
1171                 free(io_u);
1172         }
1173
1174         if (td->mem_type == MEM_MALLOC)
1175                 free(td->orig_buffer);
1176         else if (td->mem_type == MEM_SHM) {
1177                 struct shmid_ds sbuf;
1178
1179                 shmdt(td->orig_buffer);
1180                 shmctl(td->shm_id, IPC_RMID, &sbuf);
1181         }
1182 }
1183
1184 static int init_io_u(struct thread_data *td)
1185 {
1186         struct io_u *io_u;
1187         int i, max_units, mem_size;
1188         char *p;
1189
1190         if (!td->use_aio)
1191                 max_units = 1;
1192         else
1193                 max_units = td->aio_depth;
1194
1195         mem_size = td->max_bs * max_units + MASK;
1196
1197         if (td->mem_type == MEM_MALLOC)
1198                 td->orig_buffer = malloc(mem_size);
1199         else if (td->mem_type == MEM_SHM) {
1200                 td->shm_id = shmget(IPC_PRIVATE, mem_size, IPC_CREAT | 0600);
1201                 if (td->shm_id < 0) {
1202                         td->error = errno;
1203                         perror("shmget");
1204                         return 1;
1205                 }
1206
1207                 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1208                 if (td->orig_buffer == (void *) -1) {
1209                         td->error = errno;
1210                         perror("shmat");
1211                         return 1;
1212                 }
1213         }
1214
1215         INIT_LIST_HEAD(&td->io_u_freelist);
1216         INIT_LIST_HEAD(&td->io_u_busylist);
1217         INIT_LIST_HEAD(&td->io_hist_list);
1218
1219         p = ALIGN(td->orig_buffer);
1220         for (i = 0; i < max_units; i++) {
1221                 io_u = malloc(sizeof(*io_u));
1222                 memset(io_u, 0, sizeof(*io_u));
1223                 INIT_LIST_HEAD(&io_u->list);
1224
1225                 io_u->buf = p + td->max_bs * i;
1226                 list_add(&io_u->list, &td->io_u_freelist);
1227         }
1228
1229         return 0;
1230 }
1231
1232 static void setup_log(struct io_log **log)
1233 {
1234         struct io_log *l = malloc(sizeof(*l));
1235
1236         l->nr_samples = 0;
1237         l->max_samples = 1024;
1238         l->log = malloc(l->max_samples * sizeof(struct io_sample));
1239         *log = l;
1240 }
1241
1242 static void finish_log(struct thread_data *td, struct io_log *log, char *name)
1243 {
1244         char file_name[128];
1245         FILE *f;
1246         unsigned int i;
1247
1248         sprintf(file_name, "client%d_%s.log", td->thread_number, name);
1249         f = fopen(file_name, "w");
1250         if (!f) {
1251                 perror("fopen log");
1252                 return;
1253         }
1254
1255         for (i = 0; i < log->nr_samples; i++)
1256                 fprintf(f, "%lu, %lu\n", log->log[i].time, log->log[i].val);
1257
1258         fclose(f);
1259         free(log->log);
1260         free(log);
1261 }
1262
1263 static int create_file(struct thread_data *td)
1264 {
1265         unsigned long long left;
1266         unsigned int bs;
1267         char *b;
1268         int r;
1269
1270         /*
1271          * unless specifically asked for overwrite, let normal io extend it
1272          */
1273         if (!td_read(td) && !td->overwrite)
1274                 return 0;
1275
1276         if (!td->file_size) {
1277                 fprintf(stderr, "Need size for create\n");
1278                 td->error = EINVAL;
1279                 return 1;
1280         }
1281
1282         printf("Client%d: Laying out IO file\n", td->thread_number);
1283
1284         td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
1285         if (td->fd < 0) {
1286                 td->error = errno;
1287                 return 1;
1288         }
1289
1290         td->kb = td->file_size >> 10;
1291         b = malloc(td->max_bs);
1292         memset(b, 0, td->max_bs);
1293
1294         left = td->file_size;
1295         while (left) {
1296                 bs = td->max_bs;
1297                 if (bs > left)
1298                         bs = left;
1299
1300                 r = write(td->fd, b, bs);
1301
1302                 if (r == (int) bs) {
1303                         left -= bs;
1304                         continue;
1305                 } else {
1306                         if (r < 0)
1307                                 td->error = errno;
1308                         else
1309                                 td->error = EIO;
1310
1311                         break;
1312                 }
1313         }
1314
1315         if (td->create_fsync)
1316                 fsync(td->fd);
1317
1318         close(td->fd);
1319         td->fd = -1;
1320         free(b);
1321         return 0;
1322 }
1323
1324 static int file_exists(struct thread_data *td)
1325 {
1326         struct stat st;
1327
1328         if (stat(td->file_name, &st) != -1)
1329                 return 1;
1330
1331         return errno != ENOENT;
1332 }
1333
1334 static int setup_file(struct thread_data *td)
1335 {
1336         struct stat st;
1337         int flags = 0;
1338
1339         if (!file_exists(td)) {
1340                 if (!td->create_file) {
1341                         td->error = ENOENT;
1342                         return 1;
1343                 }
1344                 if (create_file(td))
1345                         return 1;
1346         }
1347
1348         if (td->odirect)
1349                 flags |= O_DIRECT;
1350
1351         if (td_read(td))
1352                 td->fd = open(td->file_name, flags | O_RDONLY);
1353         else {
1354                 if (!td->overwrite)
1355                         flags |= O_TRUNC;
1356                 if (td->sync_io)
1357                         flags |= O_SYNC;
1358                 if (td->verify)
1359                         flags |= O_RDWR;
1360                 else
1361                         flags |= O_WRONLY;
1362
1363                 td->fd = open(td->file_name, flags | O_CREAT, 0600);
1364         }
1365
1366         if (td->fd == -1) {
1367                 td->error = errno;
1368                 return 1;
1369         }
1370
1371         if (td_read(td)) {
1372                 if (fstat(td->fd, &st) == -1) {
1373                         td->error = errno;
1374                         return 1;
1375                 }
1376
1377                 if (td->file_size > st.st_size)
1378                         st.st_size = td->file_size;
1379         } else {
1380                 if (!td->file_size)
1381                         td->file_size = 1024 * 1024 * 1024;
1382
1383                 st.st_size = td->file_size;
1384         }
1385
1386         td->kb = (st.st_size - td->file_offset) / 1024;
1387         if (!td->kb) {
1388                 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1389                 td->error = EINVAL;
1390                 return 1;
1391         }
1392
1393         if (td->invalidate_cache) {
1394                 if (fadvise(td->fd, 0, st.st_size, POSIX_FADV_DONTNEED) < 0) {
1395                         td->error = errno;
1396                         return 1;
1397                 }
1398         }
1399
1400         return 0;
1401 }
1402
1403 static void clear_io_state(struct thread_data *td)
1404 {
1405         td->cur_off = 0;
1406         td->last_kb = 0;
1407         td->stat_io_kb = 0;
1408         td->this_io_kb = 0;
1409 }
1410
1411 static void *thread_main(int shm_id, int offset, char *argv[])
1412 {
1413         struct thread_data *td;
1414         int ret = 1;
1415         void *data;
1416
1417         setsid();
1418
1419         data = shmat(shm_id, NULL, 0);
1420         if (data == (void *) -1) {
1421                 perror("shmat");
1422                 return NULL;
1423         }
1424
1425         td = data + offset * sizeof(struct thread_data);
1426         td->pid = getpid();
1427
1428         if (init_io_u(td))
1429                 goto err;
1430
1431         if (sched_setaffinity(td->pid, sizeof(td->cpumask), &td->cpumask) == -1) {
1432                 td->error = errno;
1433                 goto err;
1434         }
1435
1436         sprintf(argv[0], "fio%d", offset);
1437
1438         if (td->use_aio && init_aio(td))
1439                 goto err;
1440
1441         if (init_random_state(td))
1442                 goto err;
1443
1444         if (td->ioprio) {
1445                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1446                         td->error = errno;
1447                         goto err;
1448                 }
1449         }
1450
1451         sem_post(&startup_sem);
1452         sem_wait(&td->mutex);
1453
1454         if (!td->create_serialize && setup_file(td))
1455                 goto err;
1456
1457         gettimeofday(&td->start, NULL);
1458
1459         while (td->loops--) {
1460                 gettimeofday(&td->stat_sample_time, NULL);
1461
1462                 if (td->ratemin)
1463                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1464
1465                 clear_io_state(td);
1466
1467                 if (!td->use_aio)
1468                         do_sync_io(td);
1469                 else
1470                         do_async_io(td);
1471
1472                 if (td->error)
1473                         break;
1474
1475                 if (!td->verify)
1476                         continue;
1477
1478                 clear_io_state(td);
1479
1480                 if (!td->use_aio)
1481                         do_sync_verify(td);
1482                 else
1483                         do_async_verify(td);
1484
1485                 if (td->error)
1486                         break;
1487         }
1488
1489         td->runtime = mtime_since_now(&td->start);
1490         ret = 0;
1491
1492         if (td->bw_log)
1493                 finish_log(td, td->bw_log, "bw");
1494         if (td->lat_log)
1495                 finish_log(td, td->lat_log, "lat");
1496
1497         if (exitall_on_terminate)
1498                 sig_handler(0);
1499
1500 err:
1501         if (td->fd != -1) {
1502                 close(td->fd);
1503                 td->fd = -1;
1504         }
1505         if (td->use_aio)
1506                 cleanup_aio(td);
1507         cleanup_io_u(td);
1508         if (ret) {
1509                 sem_post(&startup_sem);
1510                 sem_wait(&td->mutex);
1511         }
1512         td_set_runstate(td, TD_EXITED);
1513         shmdt(data);
1514         return NULL;
1515 }
1516
1517 static void free_shm(void)
1518 {
1519         struct shmid_ds sbuf;
1520
1521         if (threads) {
1522                 shmdt(threads);
1523                 threads = NULL;
1524                 shmctl(shm_id, IPC_RMID, &sbuf);
1525         }
1526 }
1527
1528 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
1529                     double *mean, double *dev)
1530 {
1531         double n;
1532
1533         if (is->samples == 0)
1534                 return 0;
1535
1536         *min = is->min_val;
1537         *max = is->max_val;
1538
1539         n = (double) is->samples;
1540         *mean = (double) is->val / n;
1541         *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
1542         return 1;
1543 }
1544
1545 static void show_thread_status(struct thread_data *td)
1546 {
1547         int prio, prio_class;
1548         unsigned long min, max, bw = 0;
1549         double mean, dev;
1550
1551         if (!td->io_kb && !td->error)
1552                 return;
1553
1554         if (td->runtime)
1555                 bw = td->io_kb * 1024 / td->runtime;
1556
1557         prio = td->ioprio & 0xff;
1558         prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT;
1559
1560         printf("Client%d: err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->error, td->io_kb >> 10, bw, td->runtime);
1561
1562         if (calc_lat(&td->slat_stat, &min, &max, &mean, &dev))
1563                 printf("  slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1564         if (calc_lat(&td->clat_stat, &min, &max, &mean, &dev))
1565                 printf("  clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1566         if (calc_lat(&td->bw_stat, &min, &max, &mean, &dev))
1567                 printf("  bw (KiB/s) : min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1568 }
1569
1570 static int setup_rate(struct thread_data *td)
1571 {
1572         int nr_reads_per_sec;
1573
1574         if (!td->rate)
1575                 return 0;
1576
1577         if (td->rate < td->ratemin) {
1578                 fprintf(stderr, "min rate larger than nominal rate\n");
1579                 return -1;
1580         }
1581
1582         nr_reads_per_sec = td->rate * 1024 / td->min_bs;
1583         td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
1584         td->rate_pending_usleep = 0;
1585         return 0;
1586 }
1587
1588 static struct thread_data *get_new_job(int global)
1589 {
1590         struct thread_data *td;
1591
1592         if (global)
1593                 return &def_thread;
1594         if (thread_number >= max_jobs)
1595                 return NULL;
1596
1597         td = &threads[thread_number++];
1598         memset(td, 0, sizeof(*td));
1599
1600         td->fd = -1;
1601         td->thread_number = thread_number;
1602
1603         td->ddir = def_thread.ddir;
1604         td->ioprio = def_thread.ioprio;
1605         td->sequential = def_thread.sequential;
1606         td->bs = def_thread.bs;
1607         td->min_bs = def_thread.min_bs;
1608         td->max_bs = def_thread.max_bs;
1609         td->odirect = def_thread.odirect;
1610         td->thinktime = def_thread.thinktime;
1611         td->fsync_blocks = def_thread.fsync_blocks;
1612         td->start_delay = def_thread.start_delay;
1613         td->timeout = def_thread.timeout;
1614         td->use_aio = def_thread.use_aio;
1615         td->create_file = def_thread.create_file;
1616         td->overwrite = def_thread.overwrite;
1617         td->invalidate_cache = def_thread.invalidate_cache;
1618         td->file_size = def_thread.file_size;
1619         td->file_offset = def_thread.file_offset;
1620         td->rate = def_thread.rate;
1621         td->ratemin = def_thread.ratemin;
1622         td->ratecycle = def_thread.ratecycle;
1623         td->aio_depth = def_thread.aio_depth;
1624         td->sync_io = def_thread.sync_io;
1625         td->mem_type = def_thread.mem_type;
1626         td->bw_avg_time = def_thread.bw_avg_time;
1627         td->create_serialize = def_thread.create_serialize;
1628         td->create_fsync = def_thread.create_fsync;
1629         td->loops = def_thread.loops;
1630         td->verify = def_thread.verify;
1631         memcpy(&td->cpumask, &def_thread.cpumask, sizeof(td->cpumask));
1632
1633         return td;
1634 }
1635
1636 static void put_job(struct thread_data *td)
1637 {
1638         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
1639         thread_number--;
1640 }
1641
1642 static int add_job(struct thread_data *td, const char *filename, int prioclass,
1643                    int prio)
1644 {
1645         if (td == &def_thread)
1646                 return 0;
1647
1648         strcpy(td->file_name, filename);
1649         sem_init(&td->mutex, 1, 0);
1650         td->ioprio = (prioclass << IOPRIO_CLASS_SHIFT) | prio;
1651
1652         td->clat_stat.min_val = ULONG_MAX;
1653         td->slat_stat.min_val = ULONG_MAX;
1654         td->bw_stat.min_val = ULONG_MAX;
1655
1656         run_str[td->thread_number - 1] = 'P';
1657
1658         if (td->use_aio && !td->aio_depth)
1659                 td->aio_depth = 1;
1660
1661         if (td->min_bs == -1U)
1662                 td->min_bs = td->bs;
1663         if (td->max_bs == -1U)
1664                 td->max_bs = td->bs;
1665         if (td_read(td))
1666                 td->verify = 0;
1667
1668         if (setup_rate(td))
1669                 return -1;
1670
1671         if (write_lat_log)
1672                 setup_log(&td->lat_log);
1673         if (write_bw_log)
1674                 setup_log(&td->bw_log);
1675
1676         printf("Client%d: file=%s, rw=%d, prio=%d/%d, seq=%d, odir=%d, bs=%d-%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, filename, td->ddir, prioclass, prio, td->sequential, td->odirect, td->min_bs, td->max_bs, td->rate, td->use_aio, td->aio_depth);
1677         return 0;
1678 }
1679
1680 static void fill_cpu_mask(cpu_set_t cpumask, int cpu)
1681 {
1682         unsigned int i;
1683
1684         CPU_ZERO(&cpumask);
1685
1686         for (i = 0; i < sizeof(int) * 8; i++) {
1687                 if ((1 << i) & cpu)
1688                         CPU_SET(i, &cpumask);
1689         }
1690 }
1691
1692 unsigned long get_mult(char c)
1693 {
1694         switch (c) {
1695                 case 'k':
1696                 case 'K':
1697                         return 1024;
1698                 case 'm':
1699                 case 'M':
1700                         return 1024 * 1024;
1701                 case 'g':
1702                 case 'G':
1703                         return 1024 * 1024 * 1024;
1704                 default:
1705                         return 1;
1706         }
1707 }
1708
1709 /*
1710  * convert string after '=' into decimal value, noting any size suffix
1711  */
1712 static int str_cnv(char *p, unsigned long long *val)
1713 {
1714         char *str;
1715         int len;
1716
1717         str = strstr(p, "=");
1718         if (!str)
1719                 return 1;
1720
1721         str++;
1722         len = strlen(str);
1723
1724         *val = strtoul(str, NULL, 10);
1725         if (*val == ULONG_MAX && errno == ERANGE)
1726                 return 1;
1727
1728         *val *= get_mult(str[len - 2]);
1729         return 0;
1730 }
1731
1732 static int check_strcnv(char *p, char *name, unsigned long long *val)
1733 {
1734         if (!strstr(p, name))
1735                 return 1;
1736
1737         return str_cnv(p, val);
1738 }
1739
1740 static int check_str(char *p, char *name, char *option)
1741 {
1742         char *s = strstr(p, name);
1743
1744         if (!s)
1745                 return 1;
1746
1747         s += strlen(name);
1748         if (strstr(s, option))
1749                 return 0;
1750
1751         return 1;
1752 }
1753
1754 static int check_range(char *p, char *name, unsigned long *s, unsigned long *e)
1755 {
1756         char str[128];
1757         char s1, s2;
1758
1759         sprintf(str, "%s=%%lu%%c-%%lu%%c", name);
1760         if (sscanf(p, str, s, &s1, e, &s2) == 4) {
1761                 *s *= get_mult(s1);
1762                 *e *= get_mult(s2);
1763                 return 0;
1764         }
1765
1766         sprintf(str, "%s = %%lu%%c-%%lu%%c", name);
1767         if (sscanf(p, str, s, &s1, e, &s2) == 4) {
1768                 *s *= get_mult(s1);
1769                 *e *= get_mult(s2);
1770                 return 0;
1771         }
1772
1773         sprintf(str, "%s=%%lu-%%lu", name);
1774         if (sscanf(p, str, s, e) == 2)
1775                 return 0;
1776
1777         sprintf(str, "%s = %%lu-%%lu", name);
1778         if (sscanf(p, str, s, e) == 2)
1779                 return 0;
1780
1781         return 1;
1782
1783 }
1784
1785 static int check_int(char *p, char *name, unsigned int *val)
1786 {
1787         char str[128];
1788
1789         sprintf(str, "%s=%%d", name);
1790         if (sscanf(p, str, val) == 1)
1791                 return 0;
1792
1793         sprintf(str, "%s = %%d", name);
1794         if (sscanf(p, str, val) == 1)
1795                 return 0;
1796
1797         return 1;
1798 }
1799
1800 static int is_empty_or_comment(char *line)
1801 {
1802         unsigned int i;
1803
1804         for (i = 0; i < strlen(line); i++) {
1805                 if (line[i] == ';')
1806                         return 1;
1807                 if (!isspace(line[i]) && !iscntrl(line[i]))
1808                         return 0;
1809         }
1810
1811         return 1;
1812 }
1813
1814 static int parse_jobs_ini(char *file)
1815 {
1816         unsigned int prioclass, prio, cpu, global;
1817         unsigned long long ull;
1818         unsigned long ul1, ul2;
1819         struct thread_data *td;
1820         char *string, *name;
1821         fpos_t off;
1822         FILE *f;
1823         char *p;
1824
1825         f = fopen(file, "r");
1826         if (!f) {
1827                 perror("fopen");
1828                 return 1;
1829         }
1830
1831         string = malloc(4096);
1832         name = malloc(256);
1833
1834         while ((p = fgets(string, 4096, f)) != NULL) {
1835                 if (is_empty_or_comment(p))
1836                         continue;
1837                 if (sscanf(p, "[%s]", name) != 1)
1838                         continue;
1839
1840                 global = !strncmp(name, "global", 6);
1841
1842                 name[strlen(name) - 1] = '\0';
1843
1844                 td = get_new_job(global);
1845                 if (!td)
1846                         break;
1847
1848                 prioclass = 2;
1849                 prio = 4;
1850
1851                 fgetpos(f, &off);
1852                 while ((p = fgets(string, 4096, f)) != NULL) {
1853                         if (is_empty_or_comment(p))
1854                                 continue;
1855                         if (strstr(p, "["))
1856                                 break;
1857                         if (!check_int(p, "rw", &td->ddir)) {
1858                                 fgetpos(f, &off);
1859                                 continue;
1860                         }
1861                         if (!check_int(p, "prio", &prio)) {
1862                                 fgetpos(f, &off);
1863                                 continue;
1864                         }
1865                         if (!check_int(p, "prioclass", &prioclass)) {
1866                                 fgetpos(f, &off);
1867                                 continue;
1868                         }
1869                         if (!check_int(p, "direct", &td->odirect)) {
1870                                 fgetpos(f, &off);
1871                                 continue;
1872                         }
1873                         if (!check_int(p, "rate", &td->rate)) {
1874                                 fgetpos(f, &off);
1875                                 continue;
1876                         }
1877                         if (!check_int(p, "ratemin", &td->ratemin)) {
1878                                 fgetpos(f, &off);
1879                                 continue;
1880                         }
1881                         if (!check_int(p, "ratecycle", &td->ratecycle)) {
1882                                 fgetpos(f, &off);
1883                                 continue;
1884                         }
1885                         if (!check_int(p, "thinktime", &td->thinktime)) {
1886                                 fgetpos(f, &off);
1887                                 continue;
1888                         }
1889                         if (!check_int(p, "cpumask", &cpu)) {
1890                                 fill_cpu_mask(td->cpumask, cpu);
1891                                 fgetpos(f, &off);
1892                                 continue;
1893                         }
1894                         if (!check_int(p, "fsync", &td->fsync_blocks)) {
1895                                 fgetpos(f, &off);
1896                                 continue;
1897                         }
1898                         if (!check_int(p, "startdelay", &td->start_delay)) {
1899                                 fgetpos(f, &off);
1900                                 continue;
1901                         }
1902                         if (!check_int(p, "timeout", &td->timeout)) {
1903                                 fgetpos(f, &off);
1904                                 continue;
1905                         }
1906                         if (!check_int(p, "invalidate",&td->invalidate_cache)) {
1907                                 fgetpos(f, &off);
1908                                 continue;
1909                         }
1910                         if (!check_int(p, "aio_depth", &td->aio_depth)) {
1911                                 fgetpos(f, &off);
1912                                 continue;
1913                         }
1914                         if (!check_int(p, "sync", &td->sync_io)) {
1915                                 fgetpos(f, &off);
1916                                 continue;
1917                         }
1918                         if (!check_int(p, "bwavgtime", &td->bw_avg_time)) {
1919                                 fgetpos(f, &off);
1920                                 continue;
1921                         }
1922                         if (!check_int(p, "create_serialize", &td->create_serialize)) {
1923                                 fgetpos(f, &off);
1924                                 continue;
1925                         }
1926                         if (!check_int(p, "create_fsync", &td->create_fsync)) {
1927                                 fgetpos(f, &off);
1928                                 continue;
1929                         }
1930                         if (!check_int(p, "loops", &td->loops)) {
1931                                 fgetpos(f, &off);
1932                                 continue;
1933                         }
1934                         if (!check_int(p, "verify", &td->verify)) {
1935                                 fgetpos(f, &off);
1936                                 continue;
1937                         }
1938                         if (!check_range(p, "bsrange", &ul1, &ul2)) {
1939                                 td->min_bs = ul1;
1940                                 td->max_bs = ul2;
1941                                 fgetpos(f, &off);
1942                                 continue;
1943                         }
1944                         if (!check_strcnv(p, "bs", &ull)) {
1945                                 td->bs = ull;
1946                                 fgetpos(f, &off);
1947                                 continue;
1948                         }
1949                         if (!check_strcnv(p, "size", &td->file_size)) {
1950                                 fgetpos(f, &off);
1951                                 continue;
1952                         }
1953                         if (!check_strcnv(p, "offset", &td->file_offset)) {
1954                                 fgetpos(f, &off);
1955                                 continue;
1956                         }
1957                         if (!check_str(p, "mem", "malloc")) {
1958                                 td->mem_type = MEM_MALLOC;
1959                                 fgetpos(f, &off);
1960                                 continue;
1961                         }
1962                         if (!check_str(p, "mem", "shm")) {
1963                                 td->mem_type = MEM_SHM;
1964                                 fgetpos(f, &off);
1965                                 continue;
1966                         }
1967                         if (!strncmp(p, "sequential", 10)) {
1968                                 td->sequential = 1;
1969                                 fgetpos(f, &off);
1970                                 continue;
1971                         }
1972                         if (!strncmp(p, "random", 6)) {
1973                                 td->sequential = 0;
1974                                 fgetpos(f, &off);
1975                                 continue;
1976                         }
1977                         if (!strncmp(p, "aio", 3)) {
1978                                 td->use_aio = 1;
1979                                 fgetpos(f, &off);
1980                                 continue;
1981                         }
1982                         if (!strncmp(p, "create", 6)) {
1983                                 td->create_file = 1;
1984                                 fgetpos(f, &off);
1985                                 continue;
1986                         }
1987                         if (!strncmp(p, "overwrite", 9)) {
1988                                 td->overwrite = 1;
1989                                 fgetpos(f, &off);
1990                                 continue;
1991                         }
1992                         if (!strncmp(p, "exitall", 7)) {
1993                                 exitall_on_terminate = 1;
1994                                 fgetpos(f, &off);
1995                                 continue;
1996                         }
1997                         printf("Client%d: bad option %s\n",td->thread_number,p);
1998                 }
1999                 fsetpos(f, &off);
2000
2001                 if (add_job(td, name, prioclass, prio))
2002                         put_job(td);
2003         }
2004
2005         free(string);
2006         free(name);
2007         fclose(f);
2008         return 0;
2009 }
2010
2011 static int parse_options(int argc, char *argv[])
2012 {
2013         int i;
2014
2015         for (i = 1; i < argc; i++) {
2016                 char *parm = argv[i];
2017
2018                 if (parm[0] != '-')
2019                         break;
2020
2021                 parm++;
2022                 switch (*parm) {
2023                         case 's':
2024                                 parm++;
2025                                 def_thread.sequential = !!atoi(parm);
2026                                 break;
2027                         case 'b':
2028                                 parm++;
2029                                 def_thread.bs = atoi(parm);
2030                                 def_thread.bs <<= 10;
2031                                 if (!def_thread.bs) {
2032                                         printf("bad block size\n");
2033                                         def_thread.bs = DEF_BS;
2034                                 }
2035                                 break;
2036                         case 't':
2037                                 parm++;
2038                                 def_thread.timeout = atoi(parm);
2039                                 break;
2040                         case 'r':
2041                                 parm++;
2042                                 repeatable = !!atoi(parm);
2043                                 break;
2044                         case 'R':
2045                                 parm++;
2046                                 rate_quit = !!atoi(parm);
2047                                 break;
2048                         case 'o':
2049                                 parm++;
2050                                 def_thread.odirect = !!atoi(parm);
2051                                 break;
2052                         case 'f':
2053                                 if (i + 1 >= argc) {
2054                                         printf("-f needs file as arg\n");
2055                                         break;
2056                                 }
2057                                 ini_file = strdup(argv[i+1]);
2058                                 i++;
2059                                 break;
2060                         case 'l':
2061                                 write_lat_log = 1;
2062                                 break;
2063                         case 'w':
2064                                 write_bw_log = 1;
2065                                 break;
2066                         default:
2067                                 printf("bad option %s\n", argv[i]);
2068                                 break;
2069                 }
2070         }
2071
2072         return i;
2073 }
2074
2075 static void print_thread_status(struct thread_data *td, int nr_running,
2076                                 int t_rate, int m_rate)
2077 {
2078         printf("Threads now running: %d", nr_running);
2079         if (m_rate || t_rate)
2080                 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
2081         printf(" : [%s]\r", run_str);
2082         fflush(stdout);
2083 }
2084
2085 static void check_str_update(struct thread_data *td, int n, int t, int m)
2086 {
2087         char c = run_str[td->thread_number - 1];
2088
2089         if (td->runstate == td->old_runstate)
2090                 return;
2091
2092         switch (td->runstate) {
2093                 case TD_REAPED:
2094                         c = '_';
2095                         break;
2096                 case TD_EXITED:
2097                         c = 'E';
2098                         break;
2099                 case TD_RUNNING:
2100                         c = '+';
2101                         break;
2102                 case TD_VERIFYING:
2103                         c = 'V';
2104                         break;
2105                 case TD_CREATED:
2106                         c = 'C';
2107                         break;
2108                 case TD_NOT_CREATED:
2109                         c = 'P';
2110                         break;
2111                 default:
2112                         printf("state %d\n", td->runstate);
2113         }
2114
2115         run_str[td->thread_number - 1] = c;
2116         print_thread_status(td, n, t, m);
2117         td->old_runstate = td->runstate;
2118 }
2119
2120 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2121 {
2122         int i;
2123
2124         /*
2125          * reap exited threads (TD_EXITED -> TD_REAPED)
2126          */
2127         for (i = 0; i < thread_number; i++) {
2128                 struct thread_data *td = &threads[i];
2129
2130                 check_str_update(td, *nr_running, *t_rate, *m_rate);
2131
2132                 if (td->runstate != TD_EXITED)
2133                         continue;
2134
2135                 td_set_runstate(td, TD_REAPED);
2136                 waitpid(td->pid, NULL, 0);
2137                 (*nr_running)--;
2138                 (*m_rate) -= td->ratemin;
2139                 (*t_rate) -= td->rate;
2140                 check_str_update(td, *nr_running, *t_rate, *m_rate);
2141
2142                 if (td->terminate)
2143                         continue;
2144         }
2145 }
2146
2147 static void run_threads(char *argv[])
2148 {
2149         struct timeval genesis;
2150         struct thread_data *td;
2151         unsigned long spent;
2152         int i, todo, nr_running, m_rate, t_rate;
2153
2154         printf("Starting %d threads\n", thread_number);
2155         fflush(stdout);
2156
2157         signal(SIGINT, sig_handler);
2158
2159         todo = thread_number;
2160         nr_running = 0;
2161         m_rate = t_rate = 0;
2162
2163         for (i = 0; i < thread_number; i++) {
2164                 td = &threads[i];
2165
2166                 if (!td->create_serialize)
2167                         continue;
2168
2169                 /*
2170                  * do file setup here so it happens sequentially,
2171                  * we don't want X number of threads getting their
2172                  * client data interspersed on disk
2173                  */
2174                 if (setup_file(td)) {
2175                         td_set_runstate(td, TD_REAPED);
2176                         todo--;
2177                 }
2178         }
2179
2180         gettimeofday(&genesis, NULL);
2181
2182         while (todo) {
2183                 /*
2184                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2185                  */
2186                 for (i = 0; i < thread_number; i++) {
2187                         td = &threads[i];
2188
2189                         if (td->runstate != TD_NOT_CREATED)
2190                                 continue;
2191
2192                         /*
2193                          * never got a chance to start, killed by other
2194                          * thread for some reason
2195                          */
2196                         if (td->terminate) {
2197                                 todo--;
2198                                 continue;
2199                         }
2200
2201                         if (td->start_delay) {
2202                                 spent = mtime_since_now(&genesis);
2203
2204                                 if (td->start_delay * 1000 > spent)
2205                                         continue;
2206                         }
2207
2208                         td_set_runstate(td, TD_CREATED);
2209                         check_str_update(td, nr_running, t_rate, m_rate);
2210                         sem_init(&startup_sem, 1, 1);
2211                         todo--;
2212
2213                         if (fork())
2214                                 sem_wait(&startup_sem);
2215                         else {
2216                                 thread_main(shm_id, i, argv);
2217                                 exit(0);
2218                         }
2219                 }
2220
2221                 /*
2222                  * start created threads (TD_CREATED -> TD_RUNNING)
2223                  */
2224                 for (i = 0; i < thread_number; i++) {
2225                         struct thread_data *td = &threads[i];
2226
2227                         if (td->runstate != TD_CREATED)
2228                                 continue;
2229
2230                         td_set_runstate(td, TD_RUNNING);
2231                         nr_running++;
2232                         m_rate += td->ratemin;
2233                         t_rate += td->rate;
2234                         check_str_update(td, nr_running, t_rate, m_rate);
2235                         sem_post(&td->mutex);
2236                 }
2237
2238                 for (i = 0; i < thread_number; i++) {
2239                         struct thread_data *td = &threads[i];
2240
2241                         if (td->runstate == TD_RUNNING)
2242                                 run_str[td->thread_number - 1] = '+';
2243                         else if (td->runstate == TD_VERIFYING)
2244                                 run_str[td->thread_number - 1] = 'V';
2245                         else
2246                                 continue;
2247
2248                         check_str_update(td, nr_running, t_rate, m_rate);
2249                 }
2250
2251                 reap_threads(&nr_running, &t_rate, &m_rate);
2252
2253                 if (todo)
2254                         usleep(100000);
2255         }
2256
2257         while (nr_running) {
2258                 reap_threads(&nr_running, &t_rate, &m_rate);
2259                 usleep(10000);
2260         }
2261 }
2262
2263 int setup_thread_area(void)
2264 {
2265         /*
2266          * 1024 is too much on some machines, scale max_jobs if
2267          * we get a failure that looks like too large a shm segment
2268          */
2269         do {
2270                 int s = max_jobs * sizeof(struct thread_data);
2271
2272                 shm_id = shmget(0, s, IPC_CREAT | 0600);
2273                 if (shm_id != -1)
2274                         break;
2275                 if (errno != EINVAL) {
2276                         perror("shmget");
2277                         break;
2278                 }
2279
2280                 max_jobs >>= 1;
2281         } while (max_jobs);
2282
2283         if (shm_id == -1)
2284                 return 1;
2285
2286         threads = shmat(shm_id, NULL, 0);
2287         if (threads == (void *) -1) {
2288                 perror("shmat");
2289                 return 1;
2290         }
2291
2292         atexit(free_shm);
2293         return 0;
2294 }
2295
2296 int main(int argc, char *argv[])
2297 {
2298         static unsigned long max_run[2], min_run[2];
2299         static unsigned long max_bw[2], min_bw[2];
2300         static unsigned long io_mb[2], agg[2];
2301         int i;
2302
2303         if (setup_thread_area())
2304                 return 1;
2305
2306         if (sched_getaffinity(getpid(), sizeof(cpu_set_t), &def_thread.cpumask) == -1) {
2307                 perror("sched_getaffinity");
2308                 return 1;
2309         }
2310
2311         /*
2312          * fill globals
2313          */
2314         def_thread.ddir = DDIR_READ;
2315         def_thread.bs = DEF_BS;
2316         def_thread.min_bs = -1;
2317         def_thread.max_bs = -1;
2318         def_thread.odirect = DEF_ODIRECT;
2319         def_thread.ratecycle = DEF_RATE_CYCLE;
2320         def_thread.sequential = DEF_SEQUENTIAL;
2321         def_thread.timeout = DEF_TIMEOUT;
2322         def_thread.create_file = DEF_CREATE;
2323         def_thread.overwrite = DEF_OVERWRITE;
2324         def_thread.invalidate_cache = DEF_INVALIDATE;
2325         def_thread.sync_io = DEF_SYNCIO;
2326         def_thread.mem_type = MEM_MALLOC;
2327         def_thread.bw_avg_time = DEF_BWAVGTIME;
2328         def_thread.create_serialize = DEF_CREATE_SER;
2329         def_thread.create_fsync = DEF_CREATE_FSYNC;
2330         def_thread.loops = DEF_LOOPS;
2331         def_thread.verify = DEF_VERIFY;
2332
2333         i = parse_options(argc, argv);
2334
2335         if (!ini_file) {
2336                 printf("Need job file\n");
2337                 return 1;
2338         }
2339
2340         if (parse_jobs_ini(ini_file))
2341                 return 1;
2342
2343         if (!thread_number) {
2344                 printf("Nothing to do\n");
2345                 return 1;
2346         }
2347
2348         run_threads(argv);
2349
2350         min_bw[0] = min_run[0] = ~0UL;
2351         min_bw[1] = min_run[1] = ~0UL;
2352         io_mb[0] = io_mb[1] = 0;
2353         agg[0] = agg[1] = 0;
2354         for (i = 0; i < thread_number; i++) {
2355                 struct thread_data *td = &threads[i];
2356                 unsigned long bw = 0;
2357
2358                 if (!td->error) {
2359                         if (td->runtime < min_run[td->ddir])
2360                                 min_run[td->ddir] = td->runtime;
2361                         if (td->runtime > max_run[td->ddir])
2362                                 max_run[td->ddir] = td->runtime;
2363
2364                         if (td->runtime)
2365                                 bw = td->io_kb * 1024 / td->runtime;
2366                         if (bw < min_bw[td->ddir])
2367                                 min_bw[td->ddir] = bw;
2368                         if (bw > max_bw[td->ddir])
2369                                 max_bw[td->ddir] = bw;
2370
2371                         io_mb[td->ddir] += td->io_kb >> 10;
2372                 }
2373
2374                 show_thread_status(td);
2375         }
2376         
2377         if (max_run[0])
2378                 agg[0] = io_mb[0] * 1024 * 1000 / max_run[0];
2379         if (max_run[1])
2380                 agg[1] = io_mb[1] * 1024 * 1000 / max_run[1];
2381
2382         printf("\nRun status:\n");
2383         if (max_run[DDIR_READ])
2384                 printf("   READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", io_mb[0], agg[0], min_bw[0], max_bw[0], min_run[0], max_run[0]);
2385         if (max_run[DDIR_WRITE])
2386                 printf("  WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", io_mb[1], agg[1], min_bw[1], max_bw[1], min_run[1], max_run[1]);
2387
2388         return 0;
2389 }