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