[PATCH] Fix busy-looping with aio engine and depth > 1
[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 {
ef9c5c40
JA
28 unsigned long long val;
29 unsigned long long val_sq;
ebac4655
JA
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 127 MEM_MMAP, /* use anonynomous mmap */
d0bdaf49 128 MEM_MMAPHUGE, /* memory mapped huge file */
e9c047a0
JA
129};
130
131/*
132 * The type of object we are working on
133 */
134enum fio_filetype {
135 FIO_TYPE_FILE = 1,
136 FIO_TYPE_BD,
137 FIO_TYPE_CHAR,
138};
139
2866c82d 140enum fio_ioengine_flags {
e9c047a0 141 FIO_SYNCIO = 1 << 0,
2866c82d
JA
142 FIO_CPUIO = 1 << 1,
143 FIO_MMAPIO = 1 << 2,
b2a15192 144 FIO_RAWIO = 1 << 3,
e9c047a0
JA
145};
146
53cdc686
JA
147struct fio_file {
148 /*
149 * A file may not be a file descriptor, let the io engine decide
150 */
151 union {
152 unsigned long file_data;
153 int fd;
154 };
155 char *file_name;
156 void *mmap;
157 unsigned long long file_size;
158 unsigned long long real_file_size;
159 unsigned long long file_offset;
160 unsigned long long last_pos;
02bcaa8c 161 unsigned long long last_completed_pos;
53cdc686
JA
162
163 unsigned long *file_map;
164 unsigned int num_maps;
9b031dc8
JA
165
166 unsigned int unlink;
53cdc686
JA
167};
168
e9c047a0
JA
169/*
170 * This describes a single thread/process executing a fio job.
171 */
ebac4655 172struct thread_data {
b4692828 173 char *name;
ef899b63 174 char *directory;
13f8e2d2 175 char *filename;
ebac4655
JA
176 char verror[80];
177 pthread_t thread;
178 int thread_number;
179 int groupid;
e9c047a0 180 enum fio_filetype filetype;
53cdc686
JA
181 struct fio_file *files;
182 unsigned int nr_files;
13f8e2d2 183 unsigned int nr_uniq_files;
53cdc686 184 unsigned int next_file;
ebac4655 185 int error;
ebac4655
JA
186 pid_t pid;
187 char *orig_buffer;
188 size_t orig_buffer_size;
189 volatile int terminate;
190 volatile int runstate;
e9c047a0 191 enum fio_ddir ddir;
3d60d1ed 192 unsigned int iomix;
ebac4655 193 unsigned int ioprio;
87dc1ab1 194 unsigned int last_was_sync;
e9c047a0 195
9158d2f7
JA
196 unsigned int sequential;
197 unsigned int odirect;
198 unsigned int invalidate_cache;
199 unsigned int create_serialize;
200 unsigned int create_fsync;
201 unsigned int end_fsync;
202 unsigned int sync_io;
203 unsigned int verify;
204 unsigned int use_thread;
205 unsigned int unlink;
206 unsigned int do_disk_util;
207 unsigned int override_sync;
208 unsigned int rand_repeatable;
209 unsigned int write_lat_log;
210 unsigned int write_bw_log;
bb8895e0 211 unsigned int norandommap;
690adba3 212 unsigned int bs_unaligned;
e9c047a0 213
a00735e6
JA
214 unsigned int bs[2];
215 unsigned int min_bs[2];
216 unsigned int max_bs[2];
56bb17f2 217 unsigned int hugepage_size;
a00735e6 218 unsigned int rw_min_bs;
ebac4655 219 unsigned int thinktime;
9c1f7434 220 unsigned int thinktime_blocks;
ebac4655
JA
221 unsigned int fsync_blocks;
222 unsigned int start_delay;
906c8d75 223 unsigned long timeout;
ebac4655 224 unsigned int overwrite;
ebac4655 225 unsigned int bw_avg_time;
ebac4655 226 unsigned int loops;
20dc95c4
JA
227 unsigned long long zone_size;
228 unsigned long long zone_skip;
e9c047a0 229 enum fio_memtype mem_type;
313cb206
JA
230 char *mmapfile;
231 int mmapfd;
ebac4655
JA
232 unsigned int stonewall;
233 unsigned int numjobs;
ebac4655
JA
234 unsigned int iodepth;
235 os_cpu_mask_t cpumask;
aea47d44 236 unsigned int iolog;
843a7413 237 unsigned int read_iolog;
a6ccc7be
JA
238 unsigned int rwmixcycle;
239 unsigned int rwmixread;
e1f36503 240 unsigned int rwmixwrite;
b6f4d880 241 unsigned int nice;
aea47d44 242
076efc7c
JA
243 char *read_iolog_file;
244 char *write_iolog_file;
843a7413
JA
245 void *iolog_buf;
246 FILE *iolog_f;
ebac4655 247
da86774e 248 char *sysfs_root;
da86774e
JA
249 char *ioscheduler;
250
6dfd46b9
JA
251 os_random_state_t bsrange_state;
252 os_random_state_t verify_state;
ebac4655
JA
253
254 int shm_id;
255
e9c047a0
JA
256 /*
257 * IO engine hooks, contains everything needed to submit an io_u
258 * to any of the available IO engines.
259 */
2866c82d 260 struct ioengine_ops *io_ops;
ebac4655 261
e9c047a0
JA
262 /*
263 * Current IO depth and list of free and busy io_u's.
264 */
ebac4655
JA
265 unsigned int cur_depth;
266 struct list_head io_u_freelist;
267 struct list_head io_u_busylist;
268
e9c047a0
JA
269 /*
270 * Rate state
271 */
ebac4655
JA
272 unsigned int rate;
273 unsigned int ratemin;
274 unsigned int ratecycle;
275 unsigned long rate_usec_cycle;
276 long rate_pending_usleep;
277 unsigned long rate_bytes;
278 struct timeval lastrate;
279
280 unsigned long runtime[2]; /* msec */
281 unsigned long long io_size;
53cdc686
JA
282 unsigned long long total_file_size;
283 unsigned long long start_offset;
ebac4655
JA
284 unsigned long long total_io_size;
285
9104f874
JA
286 unsigned long long io_blocks[2];
287 unsigned long long io_bytes[2];
288 unsigned long long zone_bytes;
289 unsigned long long this_io_bytes[2];
bbfd6b00 290 volatile int mutex;
ebac4655 291
e9c047a0
JA
292 /*
293 * State for random io, a bitmap of blocks done vs not done
294 */
6dfd46b9 295 os_random_state_t random_state;
ebac4655 296
b990b5c0
JA
297 /*
298 * CPU "io" cycle burner
299 */
300 unsigned int cpuload;
301 unsigned int cpucycle;
302
ebac4655
JA
303 /*
304 * bandwidth and latency stats
305 */
306 struct io_stat clat_stat[2]; /* completion latency */
307 struct io_stat slat_stat[2]; /* submission latency */
308 struct io_stat bw_stat[2]; /* bandwidth stats */
309
9104f874 310 unsigned long long stat_io_bytes[2];
ebac4655
JA
311 struct timeval stat_sample_time[2];
312
313 struct io_log *slat_log;
314 struct io_log *clat_log;
315 struct io_log *bw_log;
316
317 struct timeval start; /* start of this loop */
318 struct timeval epoch; /* time job was started */
69008999 319 struct timeval end_time;/* time job ended */
ebac4655 320
e9c047a0
JA
321 /*
322 * fio system usage accounting
323 */
ebac4655
JA
324 struct rusage ru_start;
325 struct rusage ru_end;
326 unsigned long usr_time;
327 unsigned long sys_time;
328 unsigned long ctx;
329
e9c047a0
JA
330 /*
331 * read/write mixed workload state
332 */
6dfd46b9 333 os_random_state_t rwmix_state;
a6ccc7be 334 struct timeval rwmix_switch;
e9c047a0 335 enum fio_ddir rwmix_ddir;
a6ccc7be 336
4e0ba8af
JA
337 /*
338 * Pre-run and post-run shell
339 */
340 char *exec_prerun;
341 char *exec_postrun;
342
e9c047a0
JA
343 /*
344 * IO historic logs
345 */
ebac4655 346 struct list_head io_hist_list;
aea47d44 347 struct list_head io_log_list;
ebac4655
JA
348};
349
b990b5c0 350#define __td_verror(td, err, msg) \
ebac4655
JA
351 do { \
352 int e = (err); \
353 (td)->error = e; \
b990b5c0 354 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, error=%s", __FILE__, __LINE__, (msg)); \
ebac4655
JA
355 } while (0)
356
b990b5c0
JA
357
358#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
359#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
360
ebac4655
JA
361extern int exitall_on_terminate;
362extern int thread_number;
363extern int shm_id;
364extern int groupid;
c6ae0a5b 365extern int terse_output;
eb8bbf48
JA
366extern FILE *f_out;
367extern FILE *f_err;
53cdc686 368extern int temp_stall_ts;
1e97cce9 369extern unsigned long long mlock_size;
ebac4655
JA
370
371extern struct thread_data *threads;
372
ebac4655
JA
373#define td_read(td) ((td)->ddir == DDIR_READ)
374#define td_write(td) ((td)->ddir == DDIR_WRITE)
3d60d1ed 375#define td_rw(td) ((td)->iomix != 0)
ebac4655
JA
376
377#define BLOCKS_PER_MAP (8 * sizeof(long))
a00735e6 378#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
53cdc686
JA
379#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
380#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
ebac4655
JA
381
382#define MAX_JOBS (1024)
383
87dc1ab1
JA
384static inline int should_fsync(struct thread_data *td)
385{
386 if (td->last_was_sync)
387 return 0;
388 if (td->odirect)
389 return 0;
390 if (td_write(td) || td_rw(td) || td->override_sync)
391 return 1;
392
393 return 0;
394}
395
ebac4655
JA
396struct disk_util_stat {
397 unsigned ios[2];
398 unsigned merges[2];
399 unsigned long long sectors[2];
400 unsigned ticks[2];
401 unsigned io_ticks;
402 unsigned time_in_queue;
403};
404
405struct disk_util {
406 struct list_head list;
407
408 char *name;
409 char path[256];
410 dev_t dev;
411
412 struct disk_util_stat dus;
413 struct disk_util_stat last_dus;
414
415 unsigned long msec;
416 struct timeval time;
417};
418
419struct io_completion_data {
420 int nr; /* input */
421
422 int error; /* output */
423 unsigned long bytes_done[2]; /* output */
02bcaa8c 424 struct timeval time; /* output */
ebac4655
JA
425};
426
427#define DISK_UTIL_MSEC (250)
428
6a0106a0
JA
429#ifndef min
430#define min(a, b) ((a) < (b) ? (a) : (b))
431#endif
a00735e6
JA
432#ifndef max
433#define max(a, b) ((a) > (b) ? (a) : (b))
434#endif
6a0106a0 435
6796209a
JA
436/*
437 * Log exports
438 */
439extern int read_iolog_get(struct thread_data *, struct io_u *);
440extern void write_iolog_put(struct thread_data *, struct io_u *);
441extern int init_iolog(struct thread_data *td);
442extern void log_io_piece(struct thread_data *, struct io_u *);
443extern void prune_io_piece_log(struct thread_data *);
444extern void write_iolog_close(struct thread_data *);
445
446/*
447 * Logging
448 */
1e97cce9
JA
449extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
450extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
451extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
6796209a
JA
452extern void show_run_stats(void);
453extern void init_disk_util(struct thread_data *);
454extern void update_rusage_stat(struct thread_data *);
455extern void update_io_ticks(void);
456extern void disk_util_timer_arm(void);
8914a9d8
JA
457extern void setup_log(struct io_log **);
458extern void finish_log(struct thread_data *, struct io_log *, const char *);
459extern int setup_rate(struct thread_data *);
6796209a
JA
460
461/*
462 * Time functions
463 */
464extern unsigned long utime_since(struct timeval *, struct timeval *);
69008999 465extern unsigned long utime_since_now(struct timeval *);
6796209a
JA
466extern unsigned long mtime_since(struct timeval *, struct timeval *);
467extern unsigned long mtime_since_now(struct timeval *);
468extern unsigned long time_since_now(struct timeval *);
263e529f 469extern unsigned long mtime_since_genesis(void);
b990b5c0 470extern void __usec_sleep(unsigned int);
6796209a 471extern void usec_sleep(struct thread_data *, unsigned long);
a00735e6 472extern void rate_throttle(struct thread_data *, unsigned long, unsigned int, int);
6043c579 473extern void fill_start_time(struct timeval *);
02bcaa8c 474extern void fio_gettime(struct timeval *, void *);
6796209a 475
8914a9d8
JA
476/*
477 * Init functions
478 */
479extern int parse_options(int, char **);
480extern int init_random_state(struct thread_data *);
481
53cdc686
JA
482/*
483 * File setup/shutdown
484 */
485extern void close_files(struct thread_data *);
486extern int setup_files(struct thread_data *);
21972cde 487extern int open_files(struct thread_data *);
e5b401d4 488extern int file_invalidate_cache(struct thread_data *, struct fio_file *);
53cdc686 489
263e529f
JA
490/*
491 * ETA/status stuff
492 */
493extern void print_thread_status(void);
494extern void print_status_init(int);
495
496/*
497 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
498 * will never back again. It may cycle between running/verififying/fsyncing.
499 * Once the thread reaches TD_EXITED, it is just waiting for the core to
500 * reap it.
501 */
502enum {
503 TD_NOT_CREATED = 0,
504 TD_CREATED,
505 TD_INITIALIZED,
506 TD_RUNNING,
507 TD_VERIFYING,
508 TD_FSYNCING,
509 TD_EXITED,
510 TD_REAPED,
511};
512
e29d1b70
JA
513/*
514 * Verify helpers
515 */
516extern void populate_verify_io_u(struct thread_data *, struct io_u *);
517extern int get_next_verify(struct thread_data *td, struct io_u *);
a9619d44 518extern int do_io_u_verify(struct thread_data *, struct io_u **);
e29d1b70 519
2f9ade3c
JA
520/*
521 * Memory helpers
522 */
523extern int fio_pin_memory(void);
524extern void fio_unpin_memory(void);
525extern int allocate_io_mem(struct thread_data *);
526extern void free_io_mem(struct thread_data *);
527
10ba535a
JA
528/*
529 * io unit handling
530 */
531#define queue_full(td) list_empty(&(td)->io_u_freelist)
532extern struct io_u *__get_io_u(struct thread_data *);
533extern struct io_u *get_io_u(struct thread_data *, struct fio_file *);
534extern void put_io_u(struct thread_data *, struct io_u *);
535extern void ios_completed(struct thread_data *, struct io_completion_data *);
536extern void io_completed(struct thread_data *, struct io_u *, struct io_completion_data *);
537
538/*
539 * io engine entry points
540 */
8c16d840 541extern int td_io_init(struct thread_data *);
10ba535a
JA
542extern int td_io_prep(struct thread_data *, struct io_u *);
543extern int td_io_queue(struct thread_data *, struct io_u *);
544extern int td_io_sync(struct thread_data *, struct fio_file *);
545extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
546
bbfd6b00
JA
547/*
548 * This is a pretty crappy semaphore implementation, but with the use that fio
549 * has (just signalling start/go conditions), it doesn't have to be better.
550 * Naturally this would not work for any type of contended semaphore or
551 * for real locking.
552 */
1056eaad 553static inline void fio_sem_init(volatile int *sem, int val)
bbfd6b00
JA
554{
555 *sem = val;
556}
557
1056eaad 558static inline void fio_sem_down(volatile int *sem)
bbfd6b00
JA
559{
560 while (*sem == 0)
561 usleep(10000);
562
563 (*sem)--;
564}
565
1056eaad 566static inline void fio_sem_up(volatile int *sem)
bbfd6b00
JA
567{
568 (*sem)++;
569}
570
3b70d7e5
JA
571/*
572 * If logging output to a file, stderr should go to both stderr and f_err
573 */
574#define log_err(args...) do { \
575 fprintf(f_err, ##args); \
576 if (f_err != stderr) \
577 fprintf(stderr, ##args); \
578 } while (0)
579
2866c82d 580struct ioengine_ops {
5f350952 581 struct list_head list;
2866c82d
JA
582 char name[16];
583 int version;
584 int flags;
ea2877a4 585 int (*setup)(struct thread_data *);
2866c82d
JA
586 int (*init)(struct thread_data *);
587 int (*prep)(struct thread_data *, struct io_u *);
588 int (*queue)(struct thread_data *, struct io_u *);
589 int (*getevents)(struct thread_data *, int, int, struct timespec *);
590 struct io_u *(*event)(struct thread_data *, int);
591 int (*cancel)(struct thread_data *, struct io_u *);
592 void (*cleanup)(struct thread_data *);
2866c82d
JA
593 void *data;
594 void *dlhandle;
595};
596
87dc1ab1 597#define FIO_IOOPS_VERSION 3
2866c82d 598
b4692828 599extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
5f350952
JA
600extern int register_ioengine(struct ioengine_ops *);
601extern void unregister_ioengine(struct ioengine_ops *);
2866c82d
JA
602extern void close_ioengine(struct thread_data *);
603
604/*
605 * Mark unused variables passed to ops functions as unused, to silence gcc
606 */
607#define fio_unused __attribute((__unused__))
5f350952
JA
608#define fio_init __attribute__((constructor))
609#define fio_exit __attribute__((destructor))
2866c82d 610
34572e28
JA
611#define for_each_td(td, i) \
612 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
53cdc686 613#define for_each_file(td, f, i) \
34572e28 614 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
53cdc686 615
ebac4655 616#endif