[PATCH] libaio engine: fake work-around for sync issue
[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
cec6b55d
JA
78 void *xfer_buf;
79 unsigned int xfer_buflen;
80
ebac4655
JA
81 unsigned int resid;
82 unsigned int error;
83
1e97cce9 84 enum fio_ddir ddir;
ebac4655 85
dfd7bc2c
JA
86 /*
87 * io engine private data
88 */
89 union {
90 unsigned int index;
91 unsigned int seen;
92 };
93
53cdc686
JA
94 struct fio_file *file;
95
ebac4655
JA
96 struct list_head list;
97};
98
99#define FIO_HDR_MAGIC 0xf00baaef
100
101enum {
102 VERIFY_NONE = 0,
103 VERIFY_MD5,
104 VERIFY_CRC32,
105};
106
107struct verify_header {
108 unsigned int fio_magic;
109 unsigned int len;
110 unsigned int verify_type;
111 union {
112 char md5_digest[MD5_HASH_WORDS * 4];
113 unsigned long crc32;
114 };
115};
116
117struct group_run_stats {
9104f874
JA
118 unsigned long long max_run[2], min_run[2];
119 unsigned long long max_bw[2], min_bw[2];
e9b2a3fa 120 unsigned long long io_kb[2];
9104f874 121 unsigned long long agg[2];
ebac4655
JA
122};
123
e9c047a0
JA
124/*
125 * What type of allocation to use for io buffers
126 */
127enum fio_memtype {
128 MEM_MALLOC = 0, /* ordinary malloc */
129 MEM_SHM, /* use shared memory segments */
74b025b0 130 MEM_SHMHUGE, /* use shared memory segments with huge pages */
e9c047a0 131 MEM_MMAP, /* use anonynomous mmap */
d0bdaf49 132 MEM_MMAPHUGE, /* memory mapped huge file */
e9c047a0
JA
133};
134
135/*
136 * The type of object we are working on
137 */
138enum fio_filetype {
139 FIO_TYPE_FILE = 1,
140 FIO_TYPE_BD,
141 FIO_TYPE_CHAR,
142};
143
2866c82d 144enum fio_ioengine_flags {
e9c047a0 145 FIO_SYNCIO = 1 << 0,
2866c82d
JA
146 FIO_CPUIO = 1 << 1,
147 FIO_MMAPIO = 1 << 2,
b2a15192 148 FIO_RAWIO = 1 << 3,
ed92ac0c 149 FIO_NETIO = 1 << 4,
4d9345ae 150 FIO_NULLIO = 1 << 5,
e9c047a0
JA
151};
152
53cdc686
JA
153struct fio_file {
154 /*
155 * A file may not be a file descriptor, let the io engine decide
156 */
157 union {
158 unsigned long file_data;
159 int fd;
160 };
161 char *file_name;
162 void *mmap;
163 unsigned long long file_size;
164 unsigned long long real_file_size;
165 unsigned long long file_offset;
166 unsigned long long last_pos;
02bcaa8c 167 unsigned long long last_completed_pos;
53cdc686
JA
168
169 unsigned long *file_map;
170 unsigned int num_maps;
9b031dc8
JA
171
172 unsigned int unlink;
53cdc686
JA
173};
174
71619dc2
JA
175/*
176 * How many depth levels to log
177 */
178#define FIO_IO_U_MAP_NR 8
179
e9c047a0
JA
180/*
181 * This describes a single thread/process executing a fio job.
182 */
ebac4655 183struct thread_data {
61697c37 184 char *description;
b4692828 185 char *name;
ef899b63 186 char *directory;
13f8e2d2 187 char *filename;
ebac4655
JA
188 char verror[80];
189 pthread_t thread;
190 int thread_number;
191 int groupid;
e9c047a0 192 enum fio_filetype filetype;
53cdc686
JA
193 struct fio_file *files;
194 unsigned int nr_files;
13f8e2d2 195 unsigned int nr_uniq_files;
53cdc686 196 unsigned int next_file;
ebac4655 197 int error;
ebac4655
JA
198 pid_t pid;
199 char *orig_buffer;
200 size_t orig_buffer_size;
201 volatile int terminate;
202 volatile int runstate;
e9c047a0 203 enum fio_ddir ddir;
3d60d1ed 204 unsigned int iomix;
ebac4655 205 unsigned int ioprio;
87dc1ab1 206 unsigned int last_was_sync;
e9c047a0 207
9158d2f7
JA
208 unsigned int sequential;
209 unsigned int odirect;
210 unsigned int invalidate_cache;
211 unsigned int create_serialize;
212 unsigned int create_fsync;
213 unsigned int end_fsync;
214 unsigned int sync_io;
215 unsigned int verify;
216 unsigned int use_thread;
217 unsigned int unlink;
218 unsigned int do_disk_util;
219 unsigned int override_sync;
220 unsigned int rand_repeatable;
221 unsigned int write_lat_log;
222 unsigned int write_bw_log;
bb8895e0 223 unsigned int norandommap;
690adba3 224 unsigned int bs_unaligned;
e9c047a0 225
a00735e6
JA
226 unsigned int bs[2];
227 unsigned int min_bs[2];
228 unsigned int max_bs[2];
56bb17f2 229 unsigned int hugepage_size;
a00735e6 230 unsigned int rw_min_bs;
ebac4655 231 unsigned int thinktime;
9c1f7434 232 unsigned int thinktime_blocks;
ebac4655
JA
233 unsigned int fsync_blocks;
234 unsigned int start_delay;
906c8d75 235 unsigned long timeout;
ebac4655 236 unsigned int overwrite;
ebac4655 237 unsigned int bw_avg_time;
ebac4655 238 unsigned int loops;
20dc95c4
JA
239 unsigned long long zone_size;
240 unsigned long long zone_skip;
e9c047a0 241 enum fio_memtype mem_type;
313cb206
JA
242 char *mmapfile;
243 int mmapfd;
ebac4655
JA
244 unsigned int stonewall;
245 unsigned int numjobs;
ebac4655
JA
246 unsigned int iodepth;
247 os_cpu_mask_t cpumask;
aea47d44 248 unsigned int iolog;
843a7413 249 unsigned int read_iolog;
a6ccc7be
JA
250 unsigned int rwmixcycle;
251 unsigned int rwmixread;
e1f36503 252 unsigned int rwmixwrite;
b6f4d880 253 unsigned int nice;
aea47d44 254
076efc7c
JA
255 char *read_iolog_file;
256 char *write_iolog_file;
843a7413
JA
257 void *iolog_buf;
258 FILE *iolog_f;
ebac4655 259
da86774e 260 char *sysfs_root;
da86774e
JA
261 char *ioscheduler;
262
6dfd46b9
JA
263 os_random_state_t bsrange_state;
264 os_random_state_t verify_state;
ebac4655
JA
265
266 int shm_id;
267
e9c047a0
JA
268 /*
269 * IO engine hooks, contains everything needed to submit an io_u
270 * to any of the available IO engines.
271 */
2866c82d 272 struct ioengine_ops *io_ops;
ebac4655 273
e9c047a0
JA
274 /*
275 * Current IO depth and list of free and busy io_u's.
276 */
ebac4655 277 unsigned int cur_depth;
71619dc2
JA
278 unsigned int io_u_map[FIO_IO_U_MAP_NR];
279 unsigned long total_io_u;
ebac4655
JA
280 struct list_head io_u_freelist;
281 struct list_head io_u_busylist;
282
e9c047a0
JA
283 /*
284 * Rate state
285 */
ebac4655
JA
286 unsigned int rate;
287 unsigned int ratemin;
288 unsigned int ratecycle;
289 unsigned long rate_usec_cycle;
290 long rate_pending_usleep;
291 unsigned long rate_bytes;
292 struct timeval lastrate;
293
294 unsigned long runtime[2]; /* msec */
295 unsigned long long io_size;
53cdc686
JA
296 unsigned long long total_file_size;
297 unsigned long long start_offset;
ebac4655
JA
298 unsigned long long total_io_size;
299
9104f874
JA
300 unsigned long long io_blocks[2];
301 unsigned long long io_bytes[2];
302 unsigned long long zone_bytes;
303 unsigned long long this_io_bytes[2];
bbfd6b00 304 volatile int mutex;
ebac4655 305
e9c047a0
JA
306 /*
307 * State for random io, a bitmap of blocks done vs not done
308 */
6dfd46b9 309 os_random_state_t random_state;
ebac4655 310
b990b5c0
JA
311 /*
312 * CPU "io" cycle burner
313 */
314 unsigned int cpuload;
315 unsigned int cpucycle;
316
ebac4655
JA
317 /*
318 * bandwidth and latency stats
319 */
320 struct io_stat clat_stat[2]; /* completion latency */
321 struct io_stat slat_stat[2]; /* submission latency */
322 struct io_stat bw_stat[2]; /* bandwidth stats */
323
9104f874 324 unsigned long long stat_io_bytes[2];
ebac4655
JA
325 struct timeval stat_sample_time[2];
326
327 struct io_log *slat_log;
328 struct io_log *clat_log;
329 struct io_log *bw_log;
330
331 struct timeval start; /* start of this loop */
332 struct timeval epoch; /* time job was started */
69008999 333 struct timeval end_time;/* time job ended */
ebac4655 334
e9c047a0
JA
335 /*
336 * fio system usage accounting
337 */
ebac4655
JA
338 struct rusage ru_start;
339 struct rusage ru_end;
340 unsigned long usr_time;
341 unsigned long sys_time;
342 unsigned long ctx;
343
e9c047a0
JA
344 /*
345 * read/write mixed workload state
346 */
6dfd46b9 347 os_random_state_t rwmix_state;
a6ccc7be 348 struct timeval rwmix_switch;
e9c047a0 349 enum fio_ddir rwmix_ddir;
a6ccc7be 350
4e0ba8af
JA
351 /*
352 * Pre-run and post-run shell
353 */
354 char *exec_prerun;
355 char *exec_postrun;
356
e9c047a0
JA
357 /*
358 * IO historic logs
359 */
ebac4655 360 struct list_head io_hist_list;
aea47d44 361 struct list_head io_log_list;
ebac4655
JA
362};
363
b990b5c0 364#define __td_verror(td, err, msg) \
ebac4655
JA
365 do { \
366 int e = (err); \
367 (td)->error = e; \
b990b5c0 368 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, error=%s", __FILE__, __LINE__, (msg)); \
ebac4655
JA
369 } while (0)
370
b990b5c0
JA
371
372#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
373#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
374
ebac4655
JA
375extern int exitall_on_terminate;
376extern int thread_number;
377extern int shm_id;
378extern int groupid;
c6ae0a5b 379extern int terse_output;
eb8bbf48
JA
380extern FILE *f_out;
381extern FILE *f_err;
53cdc686 382extern int temp_stall_ts;
1e97cce9 383extern unsigned long long mlock_size;
ebac4655
JA
384
385extern struct thread_data *threads;
386
ebac4655
JA
387#define td_read(td) ((td)->ddir == DDIR_READ)
388#define td_write(td) ((td)->ddir == DDIR_WRITE)
3d60d1ed 389#define td_rw(td) ((td)->iomix != 0)
ebac4655
JA
390
391#define BLOCKS_PER_MAP (8 * sizeof(long))
a00735e6 392#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
53cdc686
JA
393#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
394#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
ebac4655
JA
395
396#define MAX_JOBS (1024)
397
87dc1ab1
JA
398static inline int should_fsync(struct thread_data *td)
399{
400 if (td->last_was_sync)
401 return 0;
402 if (td->odirect)
403 return 0;
404 if (td_write(td) || td_rw(td) || td->override_sync)
405 return 1;
406
407 return 0;
408}
409
ebac4655
JA
410struct disk_util_stat {
411 unsigned ios[2];
412 unsigned merges[2];
413 unsigned long long sectors[2];
414 unsigned ticks[2];
415 unsigned io_ticks;
416 unsigned time_in_queue;
417};
418
419struct disk_util {
420 struct list_head list;
421
422 char *name;
423 char path[256];
424 dev_t dev;
425
426 struct disk_util_stat dus;
427 struct disk_util_stat last_dus;
428
429 unsigned long msec;
430 struct timeval time;
431};
432
433struct io_completion_data {
434 int nr; /* input */
435
436 int error; /* output */
437 unsigned long bytes_done[2]; /* output */
02bcaa8c 438 struct timeval time; /* output */
ebac4655
JA
439};
440
441#define DISK_UTIL_MSEC (250)
442
6a0106a0
JA
443#ifndef min
444#define min(a, b) ((a) < (b) ? (a) : (b))
445#endif
a00735e6
JA
446#ifndef max
447#define max(a, b) ((a) > (b) ? (a) : (b))
448#endif
6a0106a0 449
6796209a
JA
450/*
451 * Log exports
452 */
453extern int read_iolog_get(struct thread_data *, struct io_u *);
454extern void write_iolog_put(struct thread_data *, struct io_u *);
455extern int init_iolog(struct thread_data *td);
456extern void log_io_piece(struct thread_data *, struct io_u *);
457extern void prune_io_piece_log(struct thread_data *);
458extern void write_iolog_close(struct thread_data *);
459
460/*
461 * Logging
462 */
1e97cce9
JA
463extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
464extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
465extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
6796209a
JA
466extern void show_run_stats(void);
467extern void init_disk_util(struct thread_data *);
468extern void update_rusage_stat(struct thread_data *);
469extern void update_io_ticks(void);
470extern void disk_util_timer_arm(void);
8914a9d8
JA
471extern void setup_log(struct io_log **);
472extern void finish_log(struct thread_data *, struct io_log *, const char *);
bb3884d8 473extern void __finish_log(struct io_log *, const char *);
8914a9d8 474extern int setup_rate(struct thread_data *);
bb3884d8
JA
475extern struct io_log *agg_io_log[2];
476extern int write_bw_log;
477extern void add_agg_sample(unsigned long, enum fio_ddir);
6796209a
JA
478
479/*
480 * Time functions
481 */
482extern unsigned long utime_since(struct timeval *, struct timeval *);
69008999 483extern unsigned long utime_since_now(struct timeval *);
6796209a
JA
484extern unsigned long mtime_since(struct timeval *, struct timeval *);
485extern unsigned long mtime_since_now(struct timeval *);
486extern unsigned long time_since_now(struct timeval *);
263e529f 487extern unsigned long mtime_since_genesis(void);
b990b5c0 488extern void __usec_sleep(unsigned int);
6796209a 489extern void usec_sleep(struct thread_data *, unsigned long);
a00735e6 490extern void rate_throttle(struct thread_data *, unsigned long, unsigned int, int);
6043c579 491extern void fill_start_time(struct timeval *);
02bcaa8c 492extern void fio_gettime(struct timeval *, void *);
6796209a 493
8914a9d8
JA
494/*
495 * Init functions
496 */
497extern int parse_options(int, char **);
498extern int init_random_state(struct thread_data *);
499
53cdc686
JA
500/*
501 * File setup/shutdown
502 */
503extern void close_files(struct thread_data *);
504extern int setup_files(struct thread_data *);
21972cde 505extern int open_files(struct thread_data *);
e5b401d4 506extern int file_invalidate_cache(struct thread_data *, struct fio_file *);
53cdc686 507
263e529f
JA
508/*
509 * ETA/status stuff
510 */
511extern void print_thread_status(void);
512extern void print_status_init(int);
513
514/*
515 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
516 * will never back again. It may cycle between running/verififying/fsyncing.
517 * Once the thread reaches TD_EXITED, it is just waiting for the core to
518 * reap it.
519 */
520enum {
521 TD_NOT_CREATED = 0,
522 TD_CREATED,
523 TD_INITIALIZED,
524 TD_RUNNING,
525 TD_VERIFYING,
526 TD_FSYNCING,
527 TD_EXITED,
528 TD_REAPED,
529};
530
e29d1b70
JA
531/*
532 * Verify helpers
533 */
534extern void populate_verify_io_u(struct thread_data *, struct io_u *);
535extern int get_next_verify(struct thread_data *td, struct io_u *);
a9619d44 536extern int do_io_u_verify(struct thread_data *, struct io_u **);
e29d1b70 537
2f9ade3c
JA
538/*
539 * Memory helpers
540 */
541extern int fio_pin_memory(void);
542extern void fio_unpin_memory(void);
543extern int allocate_io_mem(struct thread_data *);
544extern void free_io_mem(struct thread_data *);
545
10ba535a
JA
546/*
547 * io unit handling
548 */
549#define queue_full(td) list_empty(&(td)->io_u_freelist)
550extern struct io_u *__get_io_u(struct thread_data *);
551extern struct io_u *get_io_u(struct thread_data *, struct fio_file *);
552extern void put_io_u(struct thread_data *, struct io_u *);
553extern void ios_completed(struct thread_data *, struct io_completion_data *);
554extern void io_completed(struct thread_data *, struct io_u *, struct io_completion_data *);
555
556/*
557 * io engine entry points
558 */
8c16d840 559extern int td_io_init(struct thread_data *);
10ba535a
JA
560extern int td_io_prep(struct thread_data *, struct io_u *);
561extern int td_io_queue(struct thread_data *, struct io_u *);
562extern int td_io_sync(struct thread_data *, struct fio_file *);
563extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
564
bbfd6b00
JA
565/*
566 * This is a pretty crappy semaphore implementation, but with the use that fio
567 * has (just signalling start/go conditions), it doesn't have to be better.
568 * Naturally this would not work for any type of contended semaphore or
569 * for real locking.
570 */
1056eaad 571static inline void fio_sem_init(volatile int *sem, int val)
bbfd6b00
JA
572{
573 *sem = val;
574}
575
1056eaad 576static inline void fio_sem_down(volatile int *sem)
bbfd6b00
JA
577{
578 while (*sem == 0)
579 usleep(10000);
580
581 (*sem)--;
582}
583
1056eaad 584static inline void fio_sem_up(volatile int *sem)
bbfd6b00
JA
585{
586 (*sem)++;
587}
588
3b70d7e5
JA
589/*
590 * If logging output to a file, stderr should go to both stderr and f_err
591 */
592#define log_err(args...) do { \
593 fprintf(f_err, ##args); \
594 if (f_err != stderr) \
595 fprintf(stderr, ##args); \
596 } while (0)
597
2866c82d 598struct ioengine_ops {
5f350952 599 struct list_head list;
2866c82d
JA
600 char name[16];
601 int version;
602 int flags;
ea2877a4 603 int (*setup)(struct thread_data *);
2866c82d
JA
604 int (*init)(struct thread_data *);
605 int (*prep)(struct thread_data *, struct io_u *);
606 int (*queue)(struct thread_data *, struct io_u *);
607 int (*getevents)(struct thread_data *, int, int, struct timespec *);
608 struct io_u *(*event)(struct thread_data *, int);
609 int (*cancel)(struct thread_data *, struct io_u *);
610 void (*cleanup)(struct thread_data *);
2866c82d
JA
611 void *data;
612 void *dlhandle;
613};
614
87dc1ab1 615#define FIO_IOOPS_VERSION 3
2866c82d 616
b4692828 617extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
5f350952
JA
618extern int register_ioengine(struct ioengine_ops *);
619extern void unregister_ioengine(struct ioengine_ops *);
2866c82d
JA
620extern void close_ioengine(struct thread_data *);
621
622/*
623 * Mark unused variables passed to ops functions as unused, to silence gcc
624 */
625#define fio_unused __attribute((__unused__))
5f350952
JA
626#define fio_init __attribute__((constructor))
627#define fio_exit __attribute__((destructor))
2866c82d 628
34572e28
JA
629#define for_each_td(td, i) \
630 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
53cdc686 631#define for_each_file(td, f, i) \
34572e28 632 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
53cdc686 633
ebac4655 634#endif