[PATCH] Change O_DIRECT vs buffered setup
[fio.git] / fio.h
CommitLineData
ebac4655
JA
1#ifndef FIO_H
2#define FIO_H
3
4#include <sched.h>
5#include <limits.h>
6#include <pthread.h>
7#include <sys/time.h>
8#include <sys/resource.h>
3c39a379
JA
9#include <errno.h>
10#include <stdlib.h>
11#include <stdio.h>
6d6f031f 12#include <unistd.h>
34cfcdaf 13#include <string.h>
ebac4655
JA
14
15#include "list.h"
16#include "md5.h"
17#include "crc32.h"
18#include "arch.h"
19#include "os.h"
20
1e97cce9
JA
21enum fio_ddir {
22 DDIR_READ = 0,
23 DDIR_WRITE,
24 DDIR_SYNC,
25};
26
ebac4655 27struct io_stat {
ebac4655
JA
28 unsigned long max_val;
29 unsigned long min_val;
30 unsigned long samples;
68704084
JA
31
32 double mean;
33 double S;
ebac4655
JA
34};
35
36struct io_sample {
37 unsigned long time;
38 unsigned long val;
1e97cce9 39 enum fio_ddir ddir;
ebac4655
JA
40};
41
42struct io_log {
43 unsigned long nr_samples;
44 unsigned long max_samples;
45 struct io_sample *log;
46};
47
48struct io_piece {
49 struct list_head list;
53cdc686 50 struct fio_file *file;
ebac4655
JA
51 unsigned long long offset;
52 unsigned int len;
1e97cce9 53 enum fio_ddir ddir;
ebac4655
JA
54};
55
56/*
57 * The io unit
58 */
59struct io_u {
60 union {
61#ifdef FIO_HAVE_LIBAIO
62 struct iocb iocb;
63#endif
64#ifdef FIO_HAVE_POSIXAIO
65 struct aiocb aiocb;
66#endif
67#ifdef FIO_HAVE_SGIO
68 struct sg_io_hdr hdr;
69#endif
70 };
71 struct timeval start_time;
72 struct timeval issue_time;
73
6b9cea23 74 void *buf;
ebac4655
JA
75 unsigned int buflen;
76 unsigned long long offset;
77
78 unsigned int resid;
79 unsigned int error;
80
1e97cce9 81 enum fio_ddir ddir;
ebac4655 82
dfd7bc2c
JA
83 /*
84 * io engine private data
85 */
86 union {
87 unsigned int index;
88 unsigned int seen;
89 };
90
53cdc686
JA
91 struct fio_file *file;
92
ebac4655
JA
93 struct list_head list;
94};
95
96#define FIO_HDR_MAGIC 0xf00baaef
97
98enum {
99 VERIFY_NONE = 0,
100 VERIFY_MD5,
101 VERIFY_CRC32,
102};
103
104struct verify_header {
105 unsigned int fio_magic;
106 unsigned int len;
107 unsigned int verify_type;
108 union {
109 char md5_digest[MD5_HASH_WORDS * 4];
110 unsigned long crc32;
111 };
112};
113
114struct group_run_stats {
9104f874
JA
115 unsigned long long max_run[2], min_run[2];
116 unsigned long long max_bw[2], min_bw[2];
e9b2a3fa 117 unsigned long long io_kb[2];
9104f874 118 unsigned long long agg[2];
ebac4655
JA
119};
120
e9c047a0
JA
121/*
122 * What type of allocation to use for io buffers
123 */
124enum fio_memtype {
125 MEM_MALLOC = 0, /* ordinary malloc */
126 MEM_SHM, /* use shared memory segments */
74b025b0 127 MEM_SHMHUGE, /* use shared memory segments with huge pages */
e9c047a0 128 MEM_MMAP, /* use anonynomous mmap */
d0bdaf49 129 MEM_MMAPHUGE, /* memory mapped huge file */
e9c047a0
JA
130};
131
132/*
133 * The type of object we are working on
134 */
135enum fio_filetype {
136 FIO_TYPE_FILE = 1,
137 FIO_TYPE_BD,
138 FIO_TYPE_CHAR,
139};
140
2866c82d 141enum fio_ioengine_flags {
e9c047a0 142 FIO_SYNCIO = 1 << 0,
2866c82d
JA
143 FIO_CPUIO = 1 << 1,
144 FIO_MMAPIO = 1 << 2,
b2a15192 145 FIO_RAWIO = 1 << 3,
e9c047a0
JA
146};
147
53cdc686
JA
148struct fio_file {
149 /*
150 * A file may not be a file descriptor, let the io engine decide
151 */
152 union {
153 unsigned long file_data;
154 int fd;
155 };
156 char *file_name;
157 void *mmap;
158 unsigned long long file_size;
159 unsigned long long real_file_size;
160 unsigned long long file_offset;
161 unsigned long long last_pos;
02bcaa8c 162 unsigned long long last_completed_pos;
53cdc686
JA
163
164 unsigned long *file_map;
165 unsigned int num_maps;
9b031dc8
JA
166
167 unsigned int unlink;
53cdc686
JA
168};
169
e9c047a0
JA
170/*
171 * This describes a single thread/process executing a fio job.
172 */
ebac4655 173struct thread_data {
b4692828 174 char *name;
ef899b63 175 char *directory;
13f8e2d2 176 char *filename;
ebac4655
JA
177 char verror[80];
178 pthread_t thread;
179 int thread_number;
180 int groupid;
e9c047a0 181 enum fio_filetype filetype;
53cdc686
JA
182 struct fio_file *files;
183 unsigned int nr_files;
13f8e2d2 184 unsigned int nr_uniq_files;
53cdc686 185 unsigned int next_file;
ebac4655 186 int error;
ebac4655
JA
187 pid_t pid;
188 char *orig_buffer;
189 size_t orig_buffer_size;
190 volatile int terminate;
191 volatile int runstate;
e9c047a0 192 enum fio_ddir ddir;
3d60d1ed 193 unsigned int iomix;
ebac4655 194 unsigned int ioprio;
87dc1ab1 195 unsigned int last_was_sync;
e9c047a0 196
9158d2f7
JA
197 unsigned int sequential;
198 unsigned int odirect;
199 unsigned int invalidate_cache;
200 unsigned int create_serialize;
201 unsigned int create_fsync;
202 unsigned int end_fsync;
203 unsigned int sync_io;
204 unsigned int verify;
205 unsigned int use_thread;
206 unsigned int unlink;
207 unsigned int do_disk_util;
208 unsigned int override_sync;
209 unsigned int rand_repeatable;
210 unsigned int write_lat_log;
211 unsigned int write_bw_log;
bb8895e0 212 unsigned int norandommap;
690adba3 213 unsigned int bs_unaligned;
e9c047a0 214
a00735e6
JA
215 unsigned int bs[2];
216 unsigned int min_bs[2];
217 unsigned int max_bs[2];
56bb17f2 218 unsigned int hugepage_size;
a00735e6 219 unsigned int rw_min_bs;
ebac4655 220 unsigned int thinktime;
9c1f7434 221 unsigned int thinktime_blocks;
ebac4655
JA
222 unsigned int fsync_blocks;
223 unsigned int start_delay;
906c8d75 224 unsigned long timeout;
ebac4655 225 unsigned int overwrite;
ebac4655 226 unsigned int bw_avg_time;
ebac4655 227 unsigned int loops;
20dc95c4
JA
228 unsigned long long zone_size;
229 unsigned long long zone_skip;
e9c047a0 230 enum fio_memtype mem_type;
313cb206
JA
231 char *mmapfile;
232 int mmapfd;
ebac4655
JA
233 unsigned int stonewall;
234 unsigned int numjobs;
ebac4655
JA
235 unsigned int iodepth;
236 os_cpu_mask_t cpumask;
aea47d44 237 unsigned int iolog;
843a7413 238 unsigned int read_iolog;
a6ccc7be
JA
239 unsigned int rwmixcycle;
240 unsigned int rwmixread;
e1f36503 241 unsigned int rwmixwrite;
b6f4d880 242 unsigned int nice;
aea47d44 243
076efc7c
JA
244 char *read_iolog_file;
245 char *write_iolog_file;
843a7413
JA
246 void *iolog_buf;
247 FILE *iolog_f;
ebac4655 248
da86774e 249 char *sysfs_root;
da86774e
JA
250 char *ioscheduler;
251
6dfd46b9
JA
252 os_random_state_t bsrange_state;
253 os_random_state_t verify_state;
ebac4655
JA
254
255 int shm_id;
256
e9c047a0
JA
257 /*
258 * IO engine hooks, contains everything needed to submit an io_u
259 * to any of the available IO engines.
260 */
2866c82d 261 struct ioengine_ops *io_ops;
ebac4655 262
e9c047a0
JA
263 /*
264 * Current IO depth and list of free and busy io_u's.
265 */
ebac4655
JA
266 unsigned int cur_depth;
267 struct list_head io_u_freelist;
268 struct list_head io_u_busylist;
269
e9c047a0
JA
270 /*
271 * Rate state
272 */
ebac4655
JA
273 unsigned int rate;
274 unsigned int ratemin;
275 unsigned int ratecycle;
276 unsigned long rate_usec_cycle;
277 long rate_pending_usleep;
278 unsigned long rate_bytes;
279 struct timeval lastrate;
280
281 unsigned long runtime[2]; /* msec */
282 unsigned long long io_size;
53cdc686
JA
283 unsigned long long total_file_size;
284 unsigned long long start_offset;
ebac4655
JA
285 unsigned long long total_io_size;
286
9104f874
JA
287 unsigned long long io_blocks[2];
288 unsigned long long io_bytes[2];
289 unsigned long long zone_bytes;
290 unsigned long long this_io_bytes[2];
bbfd6b00 291 volatile int mutex;
ebac4655 292
e9c047a0
JA
293 /*
294 * State for random io, a bitmap of blocks done vs not done
295 */
6dfd46b9 296 os_random_state_t random_state;
ebac4655 297
b990b5c0
JA
298 /*
299 * CPU "io" cycle burner
300 */
301 unsigned int cpuload;
302 unsigned int cpucycle;
303
ebac4655
JA
304 /*
305 * bandwidth and latency stats
306 */
307 struct io_stat clat_stat[2]; /* completion latency */
308 struct io_stat slat_stat[2]; /* submission latency */
309 struct io_stat bw_stat[2]; /* bandwidth stats */
310
9104f874 311 unsigned long long stat_io_bytes[2];
ebac4655
JA
312 struct timeval stat_sample_time[2];
313
314 struct io_log *slat_log;
315 struct io_log *clat_log;
316 struct io_log *bw_log;
317
318 struct timeval start; /* start of this loop */
319 struct timeval epoch; /* time job was started */
69008999 320 struct timeval end_time;/* time job ended */
ebac4655 321
e9c047a0
JA
322 /*
323 * fio system usage accounting
324 */
ebac4655
JA
325 struct rusage ru_start;
326 struct rusage ru_end;
327 unsigned long usr_time;
328 unsigned long sys_time;
329 unsigned long ctx;
330
e9c047a0
JA
331 /*
332 * read/write mixed workload state
333 */
6dfd46b9 334 os_random_state_t rwmix_state;
a6ccc7be 335 struct timeval rwmix_switch;
e9c047a0 336 enum fio_ddir rwmix_ddir;
a6ccc7be 337
4e0ba8af
JA
338 /*
339 * Pre-run and post-run shell
340 */
341 char *exec_prerun;
342 char *exec_postrun;
343
e9c047a0
JA
344 /*
345 * IO historic logs
346 */
ebac4655 347 struct list_head io_hist_list;
aea47d44 348 struct list_head io_log_list;
ebac4655
JA
349};
350
b990b5c0 351#define __td_verror(td, err, msg) \
ebac4655
JA
352 do { \
353 int e = (err); \
354 (td)->error = e; \
b990b5c0 355 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, error=%s", __FILE__, __LINE__, (msg)); \
ebac4655
JA
356 } while (0)
357
b990b5c0
JA
358
359#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
360#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
361
ebac4655
JA
362extern int exitall_on_terminate;
363extern int thread_number;
364extern int shm_id;
365extern int groupid;
c6ae0a5b 366extern int terse_output;
eb8bbf48
JA
367extern FILE *f_out;
368extern FILE *f_err;
53cdc686 369extern int temp_stall_ts;
1e97cce9 370extern unsigned long long mlock_size;
ebac4655
JA
371
372extern struct thread_data *threads;
373
ebac4655
JA
374#define td_read(td) ((td)->ddir == DDIR_READ)
375#define td_write(td) ((td)->ddir == DDIR_WRITE)
3d60d1ed 376#define td_rw(td) ((td)->iomix != 0)
ebac4655
JA
377
378#define BLOCKS_PER_MAP (8 * sizeof(long))
a00735e6 379#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
53cdc686
JA
380#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
381#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
ebac4655
JA
382
383#define MAX_JOBS (1024)
384
87dc1ab1
JA
385static inline int should_fsync(struct thread_data *td)
386{
387 if (td->last_was_sync)
388 return 0;
389 if (td->odirect)
390 return 0;
391 if (td_write(td) || td_rw(td) || td->override_sync)
392 return 1;
393
394 return 0;
395}
396
ebac4655
JA
397struct disk_util_stat {
398 unsigned ios[2];
399 unsigned merges[2];
400 unsigned long long sectors[2];
401 unsigned ticks[2];
402 unsigned io_ticks;
403 unsigned time_in_queue;
404};
405
406struct disk_util {
407 struct list_head list;
408
409 char *name;
410 char path[256];
411 dev_t dev;
412
413 struct disk_util_stat dus;
414 struct disk_util_stat last_dus;
415
416 unsigned long msec;
417 struct timeval time;
418};
419
420struct io_completion_data {
421 int nr; /* input */
422
423 int error; /* output */
424 unsigned long bytes_done[2]; /* output */
02bcaa8c 425 struct timeval time; /* output */
ebac4655
JA
426};
427
428#define DISK_UTIL_MSEC (250)
429
6a0106a0
JA
430#ifndef min
431#define min(a, b) ((a) < (b) ? (a) : (b))
432#endif
a00735e6
JA
433#ifndef max
434#define max(a, b) ((a) > (b) ? (a) : (b))
435#endif
6a0106a0 436
6796209a
JA
437/*
438 * Log exports
439 */
440extern int read_iolog_get(struct thread_data *, struct io_u *);
441extern void write_iolog_put(struct thread_data *, struct io_u *);
442extern int init_iolog(struct thread_data *td);
443extern void log_io_piece(struct thread_data *, struct io_u *);
444extern void prune_io_piece_log(struct thread_data *);
445extern void write_iolog_close(struct thread_data *);
446
447/*
448 * Logging
449 */
1e97cce9
JA
450extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
451extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
452extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
6796209a
JA
453extern void show_run_stats(void);
454extern void init_disk_util(struct thread_data *);
455extern void update_rusage_stat(struct thread_data *);
456extern void update_io_ticks(void);
457extern void disk_util_timer_arm(void);
8914a9d8
JA
458extern void setup_log(struct io_log **);
459extern void finish_log(struct thread_data *, struct io_log *, const char *);
460extern int setup_rate(struct thread_data *);
6796209a
JA
461
462/*
463 * Time functions
464 */
465extern unsigned long utime_since(struct timeval *, struct timeval *);
69008999 466extern unsigned long utime_since_now(struct timeval *);
6796209a
JA
467extern unsigned long mtime_since(struct timeval *, struct timeval *);
468extern unsigned long mtime_since_now(struct timeval *);
469extern unsigned long time_since_now(struct timeval *);
263e529f 470extern unsigned long mtime_since_genesis(void);
b990b5c0 471extern void __usec_sleep(unsigned int);
6796209a 472extern void usec_sleep(struct thread_data *, unsigned long);
a00735e6 473extern void rate_throttle(struct thread_data *, unsigned long, unsigned int, int);
6043c579 474extern void fill_start_time(struct timeval *);
02bcaa8c 475extern void fio_gettime(struct timeval *, void *);
6796209a 476
8914a9d8
JA
477/*
478 * Init functions
479 */
480extern int parse_options(int, char **);
481extern int init_random_state(struct thread_data *);
482
53cdc686
JA
483/*
484 * File setup/shutdown
485 */
486extern void close_files(struct thread_data *);
487extern int setup_files(struct thread_data *);
21972cde 488extern int open_files(struct thread_data *);
e5b401d4 489extern int file_invalidate_cache(struct thread_data *, struct fio_file *);
53cdc686 490
263e529f
JA
491/*
492 * ETA/status stuff
493 */
494extern void print_thread_status(void);
495extern void print_status_init(int);
496
497/*
498 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
499 * will never back again. It may cycle between running/verififying/fsyncing.
500 * Once the thread reaches TD_EXITED, it is just waiting for the core to
501 * reap it.
502 */
503enum {
504 TD_NOT_CREATED = 0,
505 TD_CREATED,
506 TD_INITIALIZED,
507 TD_RUNNING,
508 TD_VERIFYING,
509 TD_FSYNCING,
510 TD_EXITED,
511 TD_REAPED,
512};
513
e29d1b70
JA
514/*
515 * Verify helpers
516 */
517extern void populate_verify_io_u(struct thread_data *, struct io_u *);
518extern int get_next_verify(struct thread_data *td, struct io_u *);
a9619d44 519extern int do_io_u_verify(struct thread_data *, struct io_u **);
e29d1b70 520
2f9ade3c
JA
521/*
522 * Memory helpers
523 */
524extern int fio_pin_memory(void);
525extern void fio_unpin_memory(void);
526extern int allocate_io_mem(struct thread_data *);
527extern void free_io_mem(struct thread_data *);
528
10ba535a
JA
529/*
530 * io unit handling
531 */
532#define queue_full(td) list_empty(&(td)->io_u_freelist)
533extern struct io_u *__get_io_u(struct thread_data *);
534extern struct io_u *get_io_u(struct thread_data *, struct fio_file *);
535extern void put_io_u(struct thread_data *, struct io_u *);
536extern void ios_completed(struct thread_data *, struct io_completion_data *);
537extern void io_completed(struct thread_data *, struct io_u *, struct io_completion_data *);
538
539/*
540 * io engine entry points
541 */
8c16d840 542extern int td_io_init(struct thread_data *);
10ba535a
JA
543extern int td_io_prep(struct thread_data *, struct io_u *);
544extern int td_io_queue(struct thread_data *, struct io_u *);
545extern int td_io_sync(struct thread_data *, struct fio_file *);
546extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
547
bbfd6b00
JA
548/*
549 * This is a pretty crappy semaphore implementation, but with the use that fio
550 * has (just signalling start/go conditions), it doesn't have to be better.
551 * Naturally this would not work for any type of contended semaphore or
552 * for real locking.
553 */
1056eaad 554static inline void fio_sem_init(volatile int *sem, int val)
bbfd6b00
JA
555{
556 *sem = val;
557}
558
1056eaad 559static inline void fio_sem_down(volatile int *sem)
bbfd6b00
JA
560{
561 while (*sem == 0)
562 usleep(10000);
563
564 (*sem)--;
565}
566
1056eaad 567static inline void fio_sem_up(volatile int *sem)
bbfd6b00
JA
568{
569 (*sem)++;
570}
571
3b70d7e5
JA
572/*
573 * If logging output to a file, stderr should go to both stderr and f_err
574 */
575#define log_err(args...) do { \
576 fprintf(f_err, ##args); \
577 if (f_err != stderr) \
578 fprintf(stderr, ##args); \
579 } while (0)
580
2866c82d 581struct ioengine_ops {
5f350952 582 struct list_head list;
2866c82d
JA
583 char name[16];
584 int version;
585 int flags;
ea2877a4 586 int (*setup)(struct thread_data *);
2866c82d
JA
587 int (*init)(struct thread_data *);
588 int (*prep)(struct thread_data *, struct io_u *);
589 int (*queue)(struct thread_data *, struct io_u *);
590 int (*getevents)(struct thread_data *, int, int, struct timespec *);
591 struct io_u *(*event)(struct thread_data *, int);
592 int (*cancel)(struct thread_data *, struct io_u *);
593 void (*cleanup)(struct thread_data *);
2866c82d
JA
594 void *data;
595 void *dlhandle;
596};
597
87dc1ab1 598#define FIO_IOOPS_VERSION 3
2866c82d 599
b4692828 600extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
5f350952
JA
601extern int register_ioengine(struct ioengine_ops *);
602extern void unregister_ioengine(struct ioengine_ops *);
2866c82d
JA
603extern void close_ioengine(struct thread_data *);
604
605/*
606 * Mark unused variables passed to ops functions as unused, to silence gcc
607 */
608#define fio_unused __attribute((__unused__))
5f350952
JA
609#define fio_init __attribute__((constructor))
610#define fio_exit __attribute__((destructor))
2866c82d 611
34572e28
JA
612#define for_each_td(td, i) \
613 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
53cdc686 614#define for_each_file(td, f, i) \
34572e28 615 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
53cdc686 616
ebac4655 617#endif