[PATCH] Split the io_u io handling out of fio.c
[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>
ebac4655
JA
13
14#include "list.h"
15#include "md5.h"
16#include "crc32.h"
17#include "arch.h"
18#include "os.h"
19
20struct io_stat {
21 unsigned long val;
22 unsigned long val_sq;
23 unsigned long max_val;
24 unsigned long min_val;
25 unsigned long samples;
26};
27
28struct io_sample {
29 unsigned long time;
30 unsigned long val;
31 unsigned int ddir;
32};
33
34struct io_log {
35 unsigned long nr_samples;
36 unsigned long max_samples;
37 struct io_sample *log;
38};
39
40struct io_piece {
41 struct list_head list;
53cdc686 42 struct fio_file *file;
ebac4655
JA
43 unsigned long long offset;
44 unsigned int len;
aea47d44 45 int ddir;
ebac4655
JA
46};
47
48/*
49 * The io unit
50 */
51struct io_u {
52 union {
53#ifdef FIO_HAVE_LIBAIO
54 struct iocb iocb;
55#endif
56#ifdef FIO_HAVE_POSIXAIO
57 struct aiocb aiocb;
58#endif
59#ifdef FIO_HAVE_SGIO
60 struct sg_io_hdr hdr;
61#endif
62 };
63 struct timeval start_time;
64 struct timeval issue_time;
65
66 char *buf;
67 unsigned int buflen;
68 unsigned long long offset;
b1ff3403 69 unsigned int index;
ebac4655
JA
70
71 unsigned int resid;
72 unsigned int error;
73
74 unsigned char seen;
75 unsigned char ddir;
76
53cdc686
JA
77 struct fio_file *file;
78
ebac4655
JA
79 struct list_head list;
80};
81
82#define FIO_HDR_MAGIC 0xf00baaef
83
84enum {
85 VERIFY_NONE = 0,
86 VERIFY_MD5,
87 VERIFY_CRC32,
88};
89
90struct verify_header {
91 unsigned int fio_magic;
92 unsigned int len;
93 unsigned int verify_type;
94 union {
95 char md5_digest[MD5_HASH_WORDS * 4];
96 unsigned long crc32;
97 };
98};
99
100struct group_run_stats {
9104f874
JA
101 unsigned long long max_run[2], min_run[2];
102 unsigned long long max_bw[2], min_bw[2];
e9b2a3fa 103 unsigned long long io_kb[2];
9104f874 104 unsigned long long agg[2];
ebac4655
JA
105};
106
e9c047a0
JA
107enum fio_ddir {
108 DDIR_READ = 0,
109 DDIR_WRITE,
110};
111
112/*
113 * What type of allocation to use for io buffers
114 */
115enum fio_memtype {
116 MEM_MALLOC = 0, /* ordinary malloc */
117 MEM_SHM, /* use shared memory segments */
118 MEM_MMAP, /* use anonynomous mmap */
119};
120
121/*
122 * The type of object we are working on
123 */
124enum fio_filetype {
125 FIO_TYPE_FILE = 1,
126 FIO_TYPE_BD,
127 FIO_TYPE_CHAR,
128};
129
2866c82d 130enum fio_ioengine_flags {
e9c047a0 131 FIO_SYNCIO = 1 << 0,
2866c82d
JA
132 FIO_CPUIO = 1 << 1,
133 FIO_MMAPIO = 1 << 2,
b2a15192 134 FIO_RAWIO = 1 << 3,
e9c047a0
JA
135};
136
53cdc686
JA
137struct fio_file {
138 /*
139 * A file may not be a file descriptor, let the io engine decide
140 */
141 union {
142 unsigned long file_data;
143 int fd;
144 };
145 char *file_name;
146 void *mmap;
147 unsigned long long file_size;
148 unsigned long long real_file_size;
149 unsigned long long file_offset;
150 unsigned long long last_pos;
151
152 unsigned long *file_map;
153 unsigned int num_maps;
b2a15192
JA
154
155 int fileno;
53cdc686
JA
156};
157
e9c047a0
JA
158/*
159 * This describes a single thread/process executing a fio job.
160 */
ebac4655 161struct thread_data {
e9c047a0 162 char name[32];
ef899b63 163 char *directory;
ebac4655
JA
164 char verror[80];
165 pthread_t thread;
166 int thread_number;
167 int groupid;
e9c047a0 168 enum fio_filetype filetype;
53cdc686
JA
169 struct fio_file *files;
170 unsigned int nr_files;
171 unsigned int next_file;
ebac4655 172 int error;
ebac4655
JA
173 pid_t pid;
174 char *orig_buffer;
175 size_t orig_buffer_size;
176 volatile int terminate;
177 volatile int runstate;
e9c047a0 178 enum fio_ddir ddir;
3d60d1ed 179 unsigned int iomix;
ebac4655 180 unsigned int ioprio;
e9c047a0
JA
181
182 unsigned char sequential;
183 unsigned char odirect;
e9c047a0
JA
184 unsigned char invalidate_cache;
185 unsigned char create_serialize;
186 unsigned char create_fsync;
187 unsigned char end_fsync;
188 unsigned char sync_io;
189 unsigned char verify;
190 unsigned char use_thread;
f6cbb269 191 unsigned char unlink;
e9c047a0
JA
192 unsigned char do_disk_util;
193 unsigned char override_sync;
9ebc27e1 194 unsigned char rand_repeatable;
ec94ec56
JA
195 unsigned char write_lat_log;
196 unsigned char write_bw_log;
e9c047a0 197
ebac4655
JA
198 unsigned int bs;
199 unsigned int min_bs;
200 unsigned int max_bs;
ebac4655
JA
201 unsigned int thinktime;
202 unsigned int fsync_blocks;
203 unsigned int start_delay;
906c8d75 204 unsigned long timeout;
ebac4655 205 unsigned int overwrite;
ebac4655 206 unsigned int bw_avg_time;
ebac4655 207 unsigned int loops;
20dc95c4
JA
208 unsigned long long zone_size;
209 unsigned long long zone_skip;
e9c047a0 210 enum fio_memtype mem_type;
ebac4655
JA
211 unsigned int stonewall;
212 unsigned int numjobs;
ebac4655
JA
213 unsigned int iodepth;
214 os_cpu_mask_t cpumask;
aea47d44 215 unsigned int iolog;
843a7413
JA
216 unsigned int read_iolog;
217 unsigned int write_iolog;
a6ccc7be
JA
218 unsigned int rwmixcycle;
219 unsigned int rwmixread;
b6f4d880 220 unsigned int nice;
aea47d44 221
ef899b63 222 char *iolog_file;
843a7413
JA
223 void *iolog_buf;
224 FILE *iolog_f;
ebac4655 225
da86774e 226 char *sysfs_root;
da86774e
JA
227 char *ioscheduler;
228
6dfd46b9
JA
229 os_random_state_t bsrange_state;
230 os_random_state_t verify_state;
ebac4655
JA
231
232 int shm_id;
233
e9c047a0
JA
234 /*
235 * IO engine hooks, contains everything needed to submit an io_u
236 * to any of the available IO engines.
237 */
2866c82d 238 struct ioengine_ops *io_ops;
ebac4655 239
e9c047a0
JA
240 /*
241 * Current IO depth and list of free and busy io_u's.
242 */
ebac4655
JA
243 unsigned int cur_depth;
244 struct list_head io_u_freelist;
245 struct list_head io_u_busylist;
246
e9c047a0
JA
247 /*
248 * Rate state
249 */
ebac4655
JA
250 unsigned int rate;
251 unsigned int ratemin;
252 unsigned int ratecycle;
253 unsigned long rate_usec_cycle;
254 long rate_pending_usleep;
255 unsigned long rate_bytes;
256 struct timeval lastrate;
257
258 unsigned long runtime[2]; /* msec */
259 unsigned long long io_size;
53cdc686
JA
260 unsigned long long total_file_size;
261 unsigned long long start_offset;
ebac4655
JA
262 unsigned long long total_io_size;
263
9104f874
JA
264 unsigned long long io_blocks[2];
265 unsigned long long io_bytes[2];
266 unsigned long long zone_bytes;
267 unsigned long long this_io_bytes[2];
bbfd6b00 268 volatile int mutex;
ebac4655 269
e9c047a0
JA
270 /*
271 * State for random io, a bitmap of blocks done vs not done
272 */
6dfd46b9 273 os_random_state_t random_state;
ebac4655 274
b990b5c0
JA
275 /*
276 * CPU "io" cycle burner
277 */
278 unsigned int cpuload;
279 unsigned int cpucycle;
280
ebac4655
JA
281 /*
282 * bandwidth and latency stats
283 */
284 struct io_stat clat_stat[2]; /* completion latency */
285 struct io_stat slat_stat[2]; /* submission latency */
286 struct io_stat bw_stat[2]; /* bandwidth stats */
287
9104f874 288 unsigned long long stat_io_bytes[2];
ebac4655
JA
289 struct timeval stat_sample_time[2];
290
291 struct io_log *slat_log;
292 struct io_log *clat_log;
293 struct io_log *bw_log;
294
295 struct timeval start; /* start of this loop */
296 struct timeval epoch; /* time job was started */
297
e9c047a0
JA
298 /*
299 * fio system usage accounting
300 */
ebac4655
JA
301 struct rusage ru_start;
302 struct rusage ru_end;
303 unsigned long usr_time;
304 unsigned long sys_time;
305 unsigned long ctx;
306
e9c047a0
JA
307 /*
308 * read/write mixed workload state
309 */
6dfd46b9 310 os_random_state_t rwmix_state;
a6ccc7be 311 struct timeval rwmix_switch;
e9c047a0 312 enum fio_ddir rwmix_ddir;
a6ccc7be 313
4e0ba8af
JA
314 /*
315 * Pre-run and post-run shell
316 */
317 char *exec_prerun;
318 char *exec_postrun;
319
e9c047a0
JA
320 /*
321 * IO historic logs
322 */
ebac4655 323 struct list_head io_hist_list;
aea47d44 324 struct list_head io_log_list;
ebac4655
JA
325};
326
b990b5c0 327#define __td_verror(td, err, msg) \
ebac4655
JA
328 do { \
329 int e = (err); \
330 (td)->error = e; \
b990b5c0 331 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, error=%s", __FILE__, __LINE__, (msg)); \
ebac4655
JA
332 } while (0)
333
b990b5c0
JA
334
335#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
336#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
337
ebac4655 338extern int rate_quit;
ebac4655
JA
339extern int exitall_on_terminate;
340extern int thread_number;
341extern int shm_id;
342extern int groupid;
c6ae0a5b 343extern int terse_output;
eb8bbf48
JA
344extern FILE *f_out;
345extern FILE *f_err;
c1d5725e 346extern char *fio_inst_prefix;
53cdc686 347extern int temp_stall_ts;
ebac4655
JA
348
349extern struct thread_data *threads;
350
ebac4655
JA
351#define td_read(td) ((td)->ddir == DDIR_READ)
352#define td_write(td) ((td)->ddir == DDIR_WRITE)
3d60d1ed 353#define td_rw(td) ((td)->iomix != 0)
ebac4655
JA
354
355#define BLOCKS_PER_MAP (8 * sizeof(long))
53cdc686
JA
356#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->min_bs))
357#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
358#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
ebac4655
JA
359
360#define MAX_JOBS (1024)
361
362struct disk_util_stat {
363 unsigned ios[2];
364 unsigned merges[2];
365 unsigned long long sectors[2];
366 unsigned ticks[2];
367 unsigned io_ticks;
368 unsigned time_in_queue;
369};
370
371struct disk_util {
372 struct list_head list;
373
374 char *name;
375 char path[256];
376 dev_t dev;
377
378 struct disk_util_stat dus;
379 struct disk_util_stat last_dus;
380
381 unsigned long msec;
382 struct timeval time;
383};
384
385struct io_completion_data {
386 int nr; /* input */
387
388 int error; /* output */
389 unsigned long bytes_done[2]; /* output */
390};
391
392#define DISK_UTIL_MSEC (250)
393
6a0106a0
JA
394#ifndef min
395#define min(a, b) ((a) < (b) ? (a) : (b))
396#endif
397
6796209a
JA
398/*
399 * Log exports
400 */
401extern int read_iolog_get(struct thread_data *, struct io_u *);
402extern void write_iolog_put(struct thread_data *, struct io_u *);
403extern int init_iolog(struct thread_data *td);
404extern void log_io_piece(struct thread_data *, struct io_u *);
405extern void prune_io_piece_log(struct thread_data *);
406extern void write_iolog_close(struct thread_data *);
407
408/*
409 * Logging
410 */
411extern void add_clat_sample(struct thread_data *, int, unsigned long);
412extern void add_slat_sample(struct thread_data *, int, unsigned long);
413extern void add_bw_sample(struct thread_data *, int);
414extern void show_run_stats(void);
415extern void init_disk_util(struct thread_data *);
416extern void update_rusage_stat(struct thread_data *);
417extern void update_io_ticks(void);
418extern void disk_util_timer_arm(void);
8914a9d8
JA
419extern void setup_log(struct io_log **);
420extern void finish_log(struct thread_data *, struct io_log *, const char *);
421extern int setup_rate(struct thread_data *);
6796209a
JA
422
423/*
424 * Time functions
425 */
263e529f 426extern void time_init(void);
6796209a
JA
427extern unsigned long utime_since(struct timeval *, struct timeval *);
428extern unsigned long mtime_since(struct timeval *, struct timeval *);
429extern unsigned long mtime_since_now(struct timeval *);
430extern unsigned long time_since_now(struct timeval *);
263e529f 431extern unsigned long mtime_since_genesis(void);
b990b5c0 432extern void __usec_sleep(unsigned int);
6796209a
JA
433extern void usec_sleep(struct thread_data *, unsigned long);
434extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
435
8914a9d8
JA
436/*
437 * Init functions
438 */
439extern int parse_options(int, char **);
440extern int init_random_state(struct thread_data *);
441
53cdc686
JA
442/*
443 * File setup/shutdown
444 */
445extern void close_files(struct thread_data *);
446extern int setup_files(struct thread_data *);
e5b401d4 447extern int file_invalidate_cache(struct thread_data *, struct fio_file *);
53cdc686 448
263e529f
JA
449/*
450 * ETA/status stuff
451 */
452extern void print_thread_status(void);
453extern void print_status_init(int);
454
455/*
456 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
457 * will never back again. It may cycle between running/verififying/fsyncing.
458 * Once the thread reaches TD_EXITED, it is just waiting for the core to
459 * reap it.
460 */
461enum {
462 TD_NOT_CREATED = 0,
463 TD_CREATED,
464 TD_INITIALIZED,
465 TD_RUNNING,
466 TD_VERIFYING,
467 TD_FSYNCING,
468 TD_EXITED,
469 TD_REAPED,
470};
471
e29d1b70
JA
472/*
473 * Verify helpers
474 */
475extern void populate_verify_io_u(struct thread_data *, struct io_u *);
476extern int get_next_verify(struct thread_data *td, struct io_u *);
a9619d44 477extern int do_io_u_verify(struct thread_data *, struct io_u **);
e29d1b70 478
2f9ade3c
JA
479/*
480 * Memory helpers
481 */
482extern int fio_pin_memory(void);
483extern void fio_unpin_memory(void);
484extern int allocate_io_mem(struct thread_data *);
485extern void free_io_mem(struct thread_data *);
486
10ba535a
JA
487/*
488 * io unit handling
489 */
490#define queue_full(td) list_empty(&(td)->io_u_freelist)
491extern struct io_u *__get_io_u(struct thread_data *);
492extern struct io_u *get_io_u(struct thread_data *, struct fio_file *);
493extern void put_io_u(struct thread_data *, struct io_u *);
494extern void ios_completed(struct thread_data *, struct io_completion_data *);
495extern void io_completed(struct thread_data *, struct io_u *, struct io_completion_data *);
496
497/*
498 * io engine entry points
499 */
500extern int td_io_prep(struct thread_data *, struct io_u *);
501extern int td_io_queue(struct thread_data *, struct io_u *);
502extern int td_io_sync(struct thread_data *, struct fio_file *);
503extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
504
bbfd6b00
JA
505/*
506 * This is a pretty crappy semaphore implementation, but with the use that fio
507 * has (just signalling start/go conditions), it doesn't have to be better.
508 * Naturally this would not work for any type of contended semaphore or
509 * for real locking.
510 */
1056eaad 511static inline void fio_sem_init(volatile int *sem, int val)
bbfd6b00
JA
512{
513 *sem = val;
514}
515
1056eaad 516static inline void fio_sem_down(volatile int *sem)
bbfd6b00
JA
517{
518 while (*sem == 0)
519 usleep(10000);
520
521 (*sem)--;
522}
523
1056eaad 524static inline void fio_sem_up(volatile int *sem)
bbfd6b00
JA
525{
526 (*sem)++;
527}
528
3b70d7e5
JA
529/*
530 * If logging output to a file, stderr should go to both stderr and f_err
531 */
532#define log_err(args...) do { \
533 fprintf(f_err, ##args); \
534 if (f_err != stderr) \
535 fprintf(stderr, ##args); \
536 } while (0)
537
2866c82d
JA
538struct ioengine_ops {
539 char name[16];
540 int version;
541 int flags;
ea2877a4 542 int (*setup)(struct thread_data *);
2866c82d
JA
543 int (*init)(struct thread_data *);
544 int (*prep)(struct thread_data *, struct io_u *);
545 int (*queue)(struct thread_data *, struct io_u *);
546 int (*getevents)(struct thread_data *, int, int, struct timespec *);
547 struct io_u *(*event)(struct thread_data *, int);
548 int (*cancel)(struct thread_data *, struct io_u *);
549 void (*cleanup)(struct thread_data *);
53cdc686 550 int (*sync)(struct thread_data *, struct fio_file *);
2866c82d
JA
551 void *data;
552 void *dlhandle;
553};
554
53cdc686 555#define FIO_IOOPS_VERSION 2
2866c82d
JA
556
557extern struct ioengine_ops *load_ioengine(struct thread_data *, char *);
558extern void close_ioengine(struct thread_data *);
559
560/*
561 * Mark unused variables passed to ops functions as unused, to silence gcc
562 */
563#define fio_unused __attribute((__unused__))
564
34572e28
JA
565#define for_each_td(td, i) \
566 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
53cdc686 567#define for_each_file(td, f, i) \
34572e28 568 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
53cdc686 569
ebac4655 570#endif