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