[PATCH] Shrink io_u a little
[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;
69
70 unsigned int resid;
71 unsigned int error;
72
ebac4655
JA
73 unsigned char ddir;
74
dfd7bc2c
JA
75 /*
76 * io engine private data
77 */
78 union {
79 unsigned int index;
80 unsigned int seen;
81 };
82
53cdc686
JA
83 struct fio_file *file;
84
ebac4655
JA
85 struct list_head list;
86};
87
88#define FIO_HDR_MAGIC 0xf00baaef
89
90enum {
91 VERIFY_NONE = 0,
92 VERIFY_MD5,
93 VERIFY_CRC32,
94};
95
96struct verify_header {
97 unsigned int fio_magic;
98 unsigned int len;
99 unsigned int verify_type;
100 union {
101 char md5_digest[MD5_HASH_WORDS * 4];
102 unsigned long crc32;
103 };
104};
105
106struct group_run_stats {
9104f874
JA
107 unsigned long long max_run[2], min_run[2];
108 unsigned long long max_bw[2], min_bw[2];
e9b2a3fa 109 unsigned long long io_kb[2];
9104f874 110 unsigned long long agg[2];
ebac4655
JA
111};
112
e9c047a0
JA
113enum fio_ddir {
114 DDIR_READ = 0,
115 DDIR_WRITE,
116};
117
118/*
119 * What type of allocation to use for io buffers
120 */
121enum fio_memtype {
122 MEM_MALLOC = 0, /* ordinary malloc */
123 MEM_SHM, /* use shared memory segments */
124 MEM_MMAP, /* use anonynomous mmap */
125};
126
127/*
128 * The type of object we are working on
129 */
130enum fio_filetype {
131 FIO_TYPE_FILE = 1,
132 FIO_TYPE_BD,
133 FIO_TYPE_CHAR,
134};
135
2866c82d 136enum fio_ioengine_flags {
e9c047a0 137 FIO_SYNCIO = 1 << 0,
2866c82d
JA
138 FIO_CPUIO = 1 << 1,
139 FIO_MMAPIO = 1 << 2,
b2a15192 140 FIO_RAWIO = 1 << 3,
e9c047a0
JA
141};
142
53cdc686
JA
143struct fio_file {
144 /*
145 * A file may not be a file descriptor, let the io engine decide
146 */
147 union {
148 unsigned long file_data;
149 int fd;
150 };
151 char *file_name;
152 void *mmap;
153 unsigned long long file_size;
154 unsigned long long real_file_size;
155 unsigned long long file_offset;
156 unsigned long long last_pos;
157
158 unsigned long *file_map;
159 unsigned int num_maps;
160};
161
e9c047a0
JA
162/*
163 * This describes a single thread/process executing a fio job.
164 */
ebac4655 165struct thread_data {
e9c047a0 166 char name[32];
ef899b63 167 char *directory;
13f8e2d2 168 char *filename;
ebac4655
JA
169 char verror[80];
170 pthread_t thread;
171 int thread_number;
172 int groupid;
e9c047a0 173 enum fio_filetype filetype;
53cdc686
JA
174 struct fio_file *files;
175 unsigned int nr_files;
13f8e2d2 176 unsigned int nr_uniq_files;
53cdc686 177 unsigned int next_file;
ebac4655 178 int error;
ebac4655
JA
179 pid_t pid;
180 char *orig_buffer;
181 size_t orig_buffer_size;
182 volatile int terminate;
183 volatile int runstate;
e9c047a0 184 enum fio_ddir ddir;
3d60d1ed 185 unsigned int iomix;
ebac4655 186 unsigned int ioprio;
e9c047a0
JA
187
188 unsigned char sequential;
189 unsigned char odirect;
e9c047a0
JA
190 unsigned char invalidate_cache;
191 unsigned char create_serialize;
192 unsigned char create_fsync;
193 unsigned char end_fsync;
194 unsigned char sync_io;
195 unsigned char verify;
196 unsigned char use_thread;
f6cbb269 197 unsigned char unlink;
e9c047a0
JA
198 unsigned char do_disk_util;
199 unsigned char override_sync;
9ebc27e1 200 unsigned char rand_repeatable;
ec94ec56
JA
201 unsigned char write_lat_log;
202 unsigned char write_bw_log;
e9c047a0 203
ebac4655
JA
204 unsigned int bs;
205 unsigned int min_bs;
206 unsigned int max_bs;
ebac4655
JA
207 unsigned int thinktime;
208 unsigned int fsync_blocks;
209 unsigned int start_delay;
906c8d75 210 unsigned long timeout;
ebac4655 211 unsigned int overwrite;
ebac4655 212 unsigned int bw_avg_time;
ebac4655 213 unsigned int loops;
20dc95c4
JA
214 unsigned long long zone_size;
215 unsigned long long zone_skip;
e9c047a0 216 enum fio_memtype mem_type;
ebac4655
JA
217 unsigned int stonewall;
218 unsigned int numjobs;
ebac4655
JA
219 unsigned int iodepth;
220 os_cpu_mask_t cpumask;
aea47d44 221 unsigned int iolog;
843a7413
JA
222 unsigned int read_iolog;
223 unsigned int write_iolog;
a6ccc7be
JA
224 unsigned int rwmixcycle;
225 unsigned int rwmixread;
b6f4d880 226 unsigned int nice;
aea47d44 227
ef899b63 228 char *iolog_file;
843a7413
JA
229 void *iolog_buf;
230 FILE *iolog_f;
ebac4655 231
da86774e 232 char *sysfs_root;
da86774e
JA
233 char *ioscheduler;
234
6dfd46b9
JA
235 os_random_state_t bsrange_state;
236 os_random_state_t verify_state;
ebac4655
JA
237
238 int shm_id;
239
e9c047a0
JA
240 /*
241 * IO engine hooks, contains everything needed to submit an io_u
242 * to any of the available IO engines.
243 */
2866c82d 244 struct ioengine_ops *io_ops;
ebac4655 245
e9c047a0
JA
246 /*
247 * Current IO depth and list of free and busy io_u's.
248 */
ebac4655
JA
249 unsigned int cur_depth;
250 struct list_head io_u_freelist;
251 struct list_head io_u_busylist;
252
e9c047a0
JA
253 /*
254 * Rate state
255 */
ebac4655
JA
256 unsigned int rate;
257 unsigned int ratemin;
258 unsigned int ratecycle;
259 unsigned long rate_usec_cycle;
260 long rate_pending_usleep;
261 unsigned long rate_bytes;
262 struct timeval lastrate;
263
264 unsigned long runtime[2]; /* msec */
265 unsigned long long io_size;
53cdc686
JA
266 unsigned long long total_file_size;
267 unsigned long long start_offset;
ebac4655
JA
268 unsigned long long total_io_size;
269
9104f874
JA
270 unsigned long long io_blocks[2];
271 unsigned long long io_bytes[2];
272 unsigned long long zone_bytes;
273 unsigned long long this_io_bytes[2];
bbfd6b00 274 volatile int mutex;
ebac4655 275
e9c047a0
JA
276 /*
277 * State for random io, a bitmap of blocks done vs not done
278 */
6dfd46b9 279 os_random_state_t random_state;
ebac4655 280
b990b5c0
JA
281 /*
282 * CPU "io" cycle burner
283 */
284 unsigned int cpuload;
285 unsigned int cpucycle;
286
ebac4655
JA
287 /*
288 * bandwidth and latency stats
289 */
290 struct io_stat clat_stat[2]; /* completion latency */
291 struct io_stat slat_stat[2]; /* submission latency */
292 struct io_stat bw_stat[2]; /* bandwidth stats */
293
9104f874 294 unsigned long long stat_io_bytes[2];
ebac4655
JA
295 struct timeval stat_sample_time[2];
296
297 struct io_log *slat_log;
298 struct io_log *clat_log;
299 struct io_log *bw_log;
300
301 struct timeval start; /* start of this loop */
302 struct timeval epoch; /* time job was started */
303
e9c047a0
JA
304 /*
305 * fio system usage accounting
306 */
ebac4655
JA
307 struct rusage ru_start;
308 struct rusage ru_end;
309 unsigned long usr_time;
310 unsigned long sys_time;
311 unsigned long ctx;
312
e9c047a0
JA
313 /*
314 * read/write mixed workload state
315 */
6dfd46b9 316 os_random_state_t rwmix_state;
a6ccc7be 317 struct timeval rwmix_switch;
e9c047a0 318 enum fio_ddir rwmix_ddir;
a6ccc7be 319
4e0ba8af
JA
320 /*
321 * Pre-run and post-run shell
322 */
323 char *exec_prerun;
324 char *exec_postrun;
325
e9c047a0
JA
326 /*
327 * IO historic logs
328 */
ebac4655 329 struct list_head io_hist_list;
aea47d44 330 struct list_head io_log_list;
ebac4655
JA
331};
332
b990b5c0 333#define __td_verror(td, err, msg) \
ebac4655
JA
334 do { \
335 int e = (err); \
336 (td)->error = e; \
b990b5c0 337 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, error=%s", __FILE__, __LINE__, (msg)); \
ebac4655
JA
338 } while (0)
339
b990b5c0
JA
340
341#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
342#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
343
ebac4655 344extern int rate_quit;
ebac4655
JA
345extern int exitall_on_terminate;
346extern int thread_number;
347extern int shm_id;
348extern int groupid;
c6ae0a5b 349extern int terse_output;
eb8bbf48
JA
350extern FILE *f_out;
351extern FILE *f_err;
c1d5725e 352extern char *fio_inst_prefix;
53cdc686 353extern int temp_stall_ts;
ebac4655
JA
354
355extern struct thread_data *threads;
356
ebac4655
JA
357#define td_read(td) ((td)->ddir == DDIR_READ)
358#define td_write(td) ((td)->ddir == DDIR_WRITE)
3d60d1ed 359#define td_rw(td) ((td)->iomix != 0)
ebac4655
JA
360
361#define BLOCKS_PER_MAP (8 * sizeof(long))
53cdc686
JA
362#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->min_bs))
363#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
364#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
ebac4655
JA
365
366#define MAX_JOBS (1024)
367
368struct disk_util_stat {
369 unsigned ios[2];
370 unsigned merges[2];
371 unsigned long long sectors[2];
372 unsigned ticks[2];
373 unsigned io_ticks;
374 unsigned time_in_queue;
375};
376
377struct disk_util {
378 struct list_head list;
379
380 char *name;
381 char path[256];
382 dev_t dev;
383
384 struct disk_util_stat dus;
385 struct disk_util_stat last_dus;
386
387 unsigned long msec;
388 struct timeval time;
389};
390
391struct io_completion_data {
392 int nr; /* input */
393
394 int error; /* output */
395 unsigned long bytes_done[2]; /* output */
396};
397
398#define DISK_UTIL_MSEC (250)
399
6a0106a0
JA
400#ifndef min
401#define min(a, b) ((a) < (b) ? (a) : (b))
402#endif
403
6796209a
JA
404/*
405 * Log exports
406 */
407extern int read_iolog_get(struct thread_data *, struct io_u *);
408extern void write_iolog_put(struct thread_data *, struct io_u *);
409extern int init_iolog(struct thread_data *td);
410extern void log_io_piece(struct thread_data *, struct io_u *);
411extern void prune_io_piece_log(struct thread_data *);
412extern void write_iolog_close(struct thread_data *);
413
414/*
415 * Logging
416 */
417extern void add_clat_sample(struct thread_data *, int, unsigned long);
418extern void add_slat_sample(struct thread_data *, int, unsigned long);
419extern void add_bw_sample(struct thread_data *, int);
420extern void show_run_stats(void);
421extern void init_disk_util(struct thread_data *);
422extern void update_rusage_stat(struct thread_data *);
423extern void update_io_ticks(void);
424extern void disk_util_timer_arm(void);
8914a9d8
JA
425extern void setup_log(struct io_log **);
426extern void finish_log(struct thread_data *, struct io_log *, const char *);
427extern int setup_rate(struct thread_data *);
6796209a
JA
428
429/*
430 * Time functions
431 */
263e529f 432extern void time_init(void);
6796209a
JA
433extern unsigned long utime_since(struct timeval *, struct timeval *);
434extern unsigned long mtime_since(struct timeval *, struct timeval *);
435extern unsigned long mtime_since_now(struct timeval *);
436extern unsigned long time_since_now(struct timeval *);
263e529f 437extern unsigned long mtime_since_genesis(void);
b990b5c0 438extern void __usec_sleep(unsigned int);
6796209a
JA
439extern void usec_sleep(struct thread_data *, unsigned long);
440extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
441
8914a9d8
JA
442/*
443 * Init functions
444 */
445extern int parse_options(int, char **);
446extern int init_random_state(struct thread_data *);
447
53cdc686
JA
448/*
449 * File setup/shutdown
450 */
451extern void close_files(struct thread_data *);
452extern int setup_files(struct thread_data *);
e5b401d4 453extern int file_invalidate_cache(struct thread_data *, struct fio_file *);
53cdc686 454
263e529f
JA
455/*
456 * ETA/status stuff
457 */
458extern void print_thread_status(void);
459extern void print_status_init(int);
460
461/*
462 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
463 * will never back again. It may cycle between running/verififying/fsyncing.
464 * Once the thread reaches TD_EXITED, it is just waiting for the core to
465 * reap it.
466 */
467enum {
468 TD_NOT_CREATED = 0,
469 TD_CREATED,
470 TD_INITIALIZED,
471 TD_RUNNING,
472 TD_VERIFYING,
473 TD_FSYNCING,
474 TD_EXITED,
475 TD_REAPED,
476};
477
e29d1b70
JA
478/*
479 * Verify helpers
480 */
481extern void populate_verify_io_u(struct thread_data *, struct io_u *);
482extern int get_next_verify(struct thread_data *td, struct io_u *);
a9619d44 483extern int do_io_u_verify(struct thread_data *, struct io_u **);
e29d1b70 484
2f9ade3c
JA
485/*
486 * Memory helpers
487 */
488extern int fio_pin_memory(void);
489extern void fio_unpin_memory(void);
490extern int allocate_io_mem(struct thread_data *);
491extern void free_io_mem(struct thread_data *);
492
10ba535a
JA
493/*
494 * io unit handling
495 */
496#define queue_full(td) list_empty(&(td)->io_u_freelist)
497extern struct io_u *__get_io_u(struct thread_data *);
498extern struct io_u *get_io_u(struct thread_data *, struct fio_file *);
499extern void put_io_u(struct thread_data *, struct io_u *);
500extern void ios_completed(struct thread_data *, struct io_completion_data *);
501extern void io_completed(struct thread_data *, struct io_u *, struct io_completion_data *);
502
503/*
504 * io engine entry points
505 */
8c16d840 506extern int td_io_init(struct thread_data *);
10ba535a
JA
507extern int td_io_prep(struct thread_data *, struct io_u *);
508extern int td_io_queue(struct thread_data *, struct io_u *);
509extern int td_io_sync(struct thread_data *, struct fio_file *);
510extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
511
bbfd6b00
JA
512/*
513 * This is a pretty crappy semaphore implementation, but with the use that fio
514 * has (just signalling start/go conditions), it doesn't have to be better.
515 * Naturally this would not work for any type of contended semaphore or
516 * for real locking.
517 */
1056eaad 518static inline void fio_sem_init(volatile int *sem, int val)
bbfd6b00
JA
519{
520 *sem = val;
521}
522
1056eaad 523static inline void fio_sem_down(volatile int *sem)
bbfd6b00
JA
524{
525 while (*sem == 0)
526 usleep(10000);
527
528 (*sem)--;
529}
530
1056eaad 531static inline void fio_sem_up(volatile int *sem)
bbfd6b00
JA
532{
533 (*sem)++;
534}
535
3b70d7e5
JA
536/*
537 * If logging output to a file, stderr should go to both stderr and f_err
538 */
539#define log_err(args...) do { \
540 fprintf(f_err, ##args); \
541 if (f_err != stderr) \
542 fprintf(stderr, ##args); \
543 } while (0)
544
2866c82d
JA
545struct ioengine_ops {
546 char name[16];
547 int version;
548 int flags;
ea2877a4 549 int (*setup)(struct thread_data *);
2866c82d
JA
550 int (*init)(struct thread_data *);
551 int (*prep)(struct thread_data *, struct io_u *);
552 int (*queue)(struct thread_data *, struct io_u *);
553 int (*getevents)(struct thread_data *, int, int, struct timespec *);
554 struct io_u *(*event)(struct thread_data *, int);
555 int (*cancel)(struct thread_data *, struct io_u *);
556 void (*cleanup)(struct thread_data *);
53cdc686 557 int (*sync)(struct thread_data *, struct fio_file *);
2866c82d
JA
558 void *data;
559 void *dlhandle;
560};
561
53cdc686 562#define FIO_IOOPS_VERSION 2
2866c82d
JA
563
564extern struct ioengine_ops *load_ioengine(struct thread_data *, char *);
565extern void close_ioengine(struct thread_data *);
566
567/*
568 * Mark unused variables passed to ops functions as unused, to silence gcc
569 */
570#define fio_unused __attribute((__unused__))
571
34572e28
JA
572#define for_each_td(td, i) \
573 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
53cdc686 574#define for_each_file(td, f, i) \
34572e28 575 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
53cdc686 576
ebac4655 577#endif