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