[PATCH] Network engine updates
[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
78 unsigned int resid;
79 unsigned int error;
80
1e97cce9 81 enum fio_ddir ddir;
ebac4655 82
dfd7bc2c
JA
83 /*
84 * io engine private data
85 */
86 union {
87 unsigned int index;
88 unsigned int seen;
89 };
90
53cdc686
JA
91 struct fio_file *file;
92
ebac4655
JA
93 struct list_head list;
94};
95
96#define FIO_HDR_MAGIC 0xf00baaef
97
98enum {
99 VERIFY_NONE = 0,
100 VERIFY_MD5,
101 VERIFY_CRC32,
102};
103
104struct verify_header {
105 unsigned int fio_magic;
106 unsigned int len;
107 unsigned int verify_type;
108 union {
109 char md5_digest[MD5_HASH_WORDS * 4];
110 unsigned long crc32;
111 };
112};
113
114struct group_run_stats {
9104f874
JA
115 unsigned long long max_run[2], min_run[2];
116 unsigned long long max_bw[2], min_bw[2];
e9b2a3fa 117 unsigned long long io_kb[2];
9104f874 118 unsigned long long agg[2];
ebac4655
JA
119};
120
e9c047a0
JA
121/*
122 * What type of allocation to use for io buffers
123 */
124enum fio_memtype {
125 MEM_MALLOC = 0, /* ordinary malloc */
126 MEM_SHM, /* use shared memory segments */
74b025b0 127 MEM_SHMHUGE, /* use shared memory segments with huge pages */
e9c047a0 128 MEM_MMAP, /* use anonynomous mmap */
d0bdaf49 129 MEM_MMAPHUGE, /* memory mapped huge file */
e9c047a0
JA
130};
131
132/*
133 * The type of object we are working on
134 */
135enum fio_filetype {
136 FIO_TYPE_FILE = 1,
137 FIO_TYPE_BD,
138 FIO_TYPE_CHAR,
139};
140
2866c82d 141enum fio_ioengine_flags {
e9c047a0 142 FIO_SYNCIO = 1 << 0,
2866c82d
JA
143 FIO_CPUIO = 1 << 1,
144 FIO_MMAPIO = 1 << 2,
b2a15192 145 FIO_RAWIO = 1 << 3,
ed92ac0c 146 FIO_NETIO = 1 << 4,
e9c047a0
JA
147};
148
53cdc686
JA
149struct fio_file {
150 /*
151 * A file may not be a file descriptor, let the io engine decide
152 */
153 union {
154 unsigned long file_data;
155 int fd;
156 };
157 char *file_name;
158 void *mmap;
159 unsigned long long file_size;
160 unsigned long long real_file_size;
161 unsigned long long file_offset;
162 unsigned long long last_pos;
02bcaa8c 163 unsigned long long last_completed_pos;
53cdc686
JA
164
165 unsigned long *file_map;
166 unsigned int num_maps;
9b031dc8
JA
167
168 unsigned int unlink;
53cdc686
JA
169};
170
71619dc2
JA
171/*
172 * How many depth levels to log
173 */
174#define FIO_IO_U_MAP_NR 8
175
e9c047a0
JA
176/*
177 * This describes a single thread/process executing a fio job.
178 */
ebac4655 179struct thread_data {
61697c37 180 char *description;
b4692828 181 char *name;
ef899b63 182 char *directory;
13f8e2d2 183 char *filename;
ebac4655
JA
184 char verror[80];
185 pthread_t thread;
186 int thread_number;
187 int groupid;
e9c047a0 188 enum fio_filetype filetype;
53cdc686
JA
189 struct fio_file *files;
190 unsigned int nr_files;
13f8e2d2 191 unsigned int nr_uniq_files;
53cdc686 192 unsigned int next_file;
ebac4655 193 int error;
ebac4655
JA
194 pid_t pid;
195 char *orig_buffer;
196 size_t orig_buffer_size;
197 volatile int terminate;
198 volatile int runstate;
e9c047a0 199 enum fio_ddir ddir;
3d60d1ed 200 unsigned int iomix;
ebac4655 201 unsigned int ioprio;
87dc1ab1 202 unsigned int last_was_sync;
e9c047a0 203
9158d2f7
JA
204 unsigned int sequential;
205 unsigned int odirect;
206 unsigned int invalidate_cache;
207 unsigned int create_serialize;
208 unsigned int create_fsync;
209 unsigned int end_fsync;
210 unsigned int sync_io;
211 unsigned int verify;
212 unsigned int use_thread;
213 unsigned int unlink;
214 unsigned int do_disk_util;
215 unsigned int override_sync;
216 unsigned int rand_repeatable;
217 unsigned int write_lat_log;
218 unsigned int write_bw_log;
bb8895e0 219 unsigned int norandommap;
690adba3 220 unsigned int bs_unaligned;
e9c047a0 221
a00735e6
JA
222 unsigned int bs[2];
223 unsigned int min_bs[2];
224 unsigned int max_bs[2];
56bb17f2 225 unsigned int hugepage_size;
a00735e6 226 unsigned int rw_min_bs;
ebac4655 227 unsigned int thinktime;
9c1f7434 228 unsigned int thinktime_blocks;
ebac4655
JA
229 unsigned int fsync_blocks;
230 unsigned int start_delay;
906c8d75 231 unsigned long timeout;
ebac4655 232 unsigned int overwrite;
ebac4655 233 unsigned int bw_avg_time;
ebac4655 234 unsigned int loops;
20dc95c4
JA
235 unsigned long long zone_size;
236 unsigned long long zone_skip;
e9c047a0 237 enum fio_memtype mem_type;
313cb206
JA
238 char *mmapfile;
239 int mmapfd;
ebac4655
JA
240 unsigned int stonewall;
241 unsigned int numjobs;
ebac4655
JA
242 unsigned int iodepth;
243 os_cpu_mask_t cpumask;
aea47d44 244 unsigned int iolog;
843a7413 245 unsigned int read_iolog;
a6ccc7be
JA
246 unsigned int rwmixcycle;
247 unsigned int rwmixread;
e1f36503 248 unsigned int rwmixwrite;
b6f4d880 249 unsigned int nice;
aea47d44 250
076efc7c
JA
251 char *read_iolog_file;
252 char *write_iolog_file;
843a7413
JA
253 void *iolog_buf;
254 FILE *iolog_f;
ebac4655 255
da86774e 256 char *sysfs_root;
da86774e
JA
257 char *ioscheduler;
258
6dfd46b9
JA
259 os_random_state_t bsrange_state;
260 os_random_state_t verify_state;
ebac4655
JA
261
262 int shm_id;
263
e9c047a0
JA
264 /*
265 * IO engine hooks, contains everything needed to submit an io_u
266 * to any of the available IO engines.
267 */
2866c82d 268 struct ioengine_ops *io_ops;
ebac4655 269
e9c047a0
JA
270 /*
271 * Current IO depth and list of free and busy io_u's.
272 */
ebac4655 273 unsigned int cur_depth;
71619dc2
JA
274 unsigned int io_u_map[FIO_IO_U_MAP_NR];
275 unsigned long total_io_u;
ebac4655
JA
276 struct list_head io_u_freelist;
277 struct list_head io_u_busylist;
278
e9c047a0
JA
279 /*
280 * Rate state
281 */
ebac4655
JA
282 unsigned int rate;
283 unsigned int ratemin;
284 unsigned int ratecycle;
285 unsigned long rate_usec_cycle;
286 long rate_pending_usleep;
287 unsigned long rate_bytes;
288 struct timeval lastrate;
289
290 unsigned long runtime[2]; /* msec */
291 unsigned long long io_size;
53cdc686
JA
292 unsigned long long total_file_size;
293 unsigned long long start_offset;
ebac4655
JA
294 unsigned long long total_io_size;
295
9104f874
JA
296 unsigned long long io_blocks[2];
297 unsigned long long io_bytes[2];
298 unsigned long long zone_bytes;
299 unsigned long long this_io_bytes[2];
bbfd6b00 300 volatile int mutex;
ebac4655 301
e9c047a0
JA
302 /*
303 * State for random io, a bitmap of blocks done vs not done
304 */
6dfd46b9 305 os_random_state_t random_state;
ebac4655 306
b990b5c0
JA
307 /*
308 * CPU "io" cycle burner
309 */
310 unsigned int cpuload;
311 unsigned int cpucycle;
312
ebac4655
JA
313 /*
314 * bandwidth and latency stats
315 */
316 struct io_stat clat_stat[2]; /* completion latency */
317 struct io_stat slat_stat[2]; /* submission latency */
318 struct io_stat bw_stat[2]; /* bandwidth stats */
319
9104f874 320 unsigned long long stat_io_bytes[2];
ebac4655
JA
321 struct timeval stat_sample_time[2];
322
323 struct io_log *slat_log;
324 struct io_log *clat_log;
325 struct io_log *bw_log;
326
327 struct timeval start; /* start of this loop */
328 struct timeval epoch; /* time job was started */
69008999 329 struct timeval end_time;/* time job ended */
ebac4655 330
e9c047a0
JA
331 /*
332 * fio system usage accounting
333 */
ebac4655
JA
334 struct rusage ru_start;
335 struct rusage ru_end;
336 unsigned long usr_time;
337 unsigned long sys_time;
338 unsigned long ctx;
339
e9c047a0
JA
340 /*
341 * read/write mixed workload state
342 */
6dfd46b9 343 os_random_state_t rwmix_state;
a6ccc7be 344 struct timeval rwmix_switch;
e9c047a0 345 enum fio_ddir rwmix_ddir;
a6ccc7be 346
4e0ba8af
JA
347 /*
348 * Pre-run and post-run shell
349 */
350 char *exec_prerun;
351 char *exec_postrun;
352
e9c047a0
JA
353 /*
354 * IO historic logs
355 */
ebac4655 356 struct list_head io_hist_list;
aea47d44 357 struct list_head io_log_list;
ebac4655
JA
358};
359
b990b5c0 360#define __td_verror(td, err, msg) \
ebac4655
JA
361 do { \
362 int e = (err); \
363 (td)->error = e; \
b990b5c0 364 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, error=%s", __FILE__, __LINE__, (msg)); \
ebac4655
JA
365 } while (0)
366
b990b5c0
JA
367
368#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
369#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
370
ebac4655
JA
371extern int exitall_on_terminate;
372extern int thread_number;
373extern int shm_id;
374extern int groupid;
c6ae0a5b 375extern int terse_output;
eb8bbf48
JA
376extern FILE *f_out;
377extern FILE *f_err;
53cdc686 378extern int temp_stall_ts;
1e97cce9 379extern unsigned long long mlock_size;
ebac4655
JA
380
381extern struct thread_data *threads;
382
ebac4655
JA
383#define td_read(td) ((td)->ddir == DDIR_READ)
384#define td_write(td) ((td)->ddir == DDIR_WRITE)
3d60d1ed 385#define td_rw(td) ((td)->iomix != 0)
ebac4655
JA
386
387#define BLOCKS_PER_MAP (8 * sizeof(long))
a00735e6 388#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
53cdc686
JA
389#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
390#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
ebac4655
JA
391
392#define MAX_JOBS (1024)
393
87dc1ab1
JA
394static inline int should_fsync(struct thread_data *td)
395{
396 if (td->last_was_sync)
397 return 0;
398 if (td->odirect)
399 return 0;
400 if (td_write(td) || td_rw(td) || td->override_sync)
401 return 1;
402
403 return 0;
404}
405
ebac4655
JA
406struct disk_util_stat {
407 unsigned ios[2];
408 unsigned merges[2];
409 unsigned long long sectors[2];
410 unsigned ticks[2];
411 unsigned io_ticks;
412 unsigned time_in_queue;
413};
414
415struct disk_util {
416 struct list_head list;
417
418 char *name;
419 char path[256];
420 dev_t dev;
421
422 struct disk_util_stat dus;
423 struct disk_util_stat last_dus;
424
425 unsigned long msec;
426 struct timeval time;
427};
428
429struct io_completion_data {
430 int nr; /* input */
431
432 int error; /* output */
433 unsigned long bytes_done[2]; /* output */
02bcaa8c 434 struct timeval time; /* output */
ebac4655
JA
435};
436
437#define DISK_UTIL_MSEC (250)
438
6a0106a0
JA
439#ifndef min
440#define min(a, b) ((a) < (b) ? (a) : (b))
441#endif
a00735e6
JA
442#ifndef max
443#define max(a, b) ((a) > (b) ? (a) : (b))
444#endif
6a0106a0 445
6796209a
JA
446/*
447 * Log exports
448 */
449extern int read_iolog_get(struct thread_data *, struct io_u *);
450extern void write_iolog_put(struct thread_data *, struct io_u *);
451extern int init_iolog(struct thread_data *td);
452extern void log_io_piece(struct thread_data *, struct io_u *);
453extern void prune_io_piece_log(struct thread_data *);
454extern void write_iolog_close(struct thread_data *);
455
456/*
457 * Logging
458 */
1e97cce9
JA
459extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
460extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
461extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
6796209a
JA
462extern void show_run_stats(void);
463extern void init_disk_util(struct thread_data *);
464extern void update_rusage_stat(struct thread_data *);
465extern void update_io_ticks(void);
466extern void disk_util_timer_arm(void);
8914a9d8
JA
467extern void setup_log(struct io_log **);
468extern void finish_log(struct thread_data *, struct io_log *, const char *);
bb3884d8 469extern void __finish_log(struct io_log *, const char *);
8914a9d8 470extern int setup_rate(struct thread_data *);
bb3884d8
JA
471extern struct io_log *agg_io_log[2];
472extern int write_bw_log;
473extern void add_agg_sample(unsigned long, enum fio_ddir);
6796209a
JA
474
475/*
476 * Time functions
477 */
478extern unsigned long utime_since(struct timeval *, struct timeval *);
69008999 479extern unsigned long utime_since_now(struct timeval *);
6796209a
JA
480extern unsigned long mtime_since(struct timeval *, struct timeval *);
481extern unsigned long mtime_since_now(struct timeval *);
482extern unsigned long time_since_now(struct timeval *);
263e529f 483extern unsigned long mtime_since_genesis(void);
b990b5c0 484extern void __usec_sleep(unsigned int);
6796209a 485extern void usec_sleep(struct thread_data *, unsigned long);
a00735e6 486extern void rate_throttle(struct thread_data *, unsigned long, unsigned int, int);
6043c579 487extern void fill_start_time(struct timeval *);
02bcaa8c 488extern void fio_gettime(struct timeval *, void *);
6796209a 489
8914a9d8
JA
490/*
491 * Init functions
492 */
493extern int parse_options(int, char **);
494extern int init_random_state(struct thread_data *);
495
53cdc686
JA
496/*
497 * File setup/shutdown
498 */
499extern void close_files(struct thread_data *);
500extern int setup_files(struct thread_data *);
21972cde 501extern int open_files(struct thread_data *);
e5b401d4 502extern int file_invalidate_cache(struct thread_data *, struct fio_file *);
53cdc686 503
263e529f
JA
504/*
505 * ETA/status stuff
506 */
507extern void print_thread_status(void);
508extern void print_status_init(int);
509
510/*
511 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
512 * will never back again. It may cycle between running/verififying/fsyncing.
513 * Once the thread reaches TD_EXITED, it is just waiting for the core to
514 * reap it.
515 */
516enum {
517 TD_NOT_CREATED = 0,
518 TD_CREATED,
519 TD_INITIALIZED,
520 TD_RUNNING,
521 TD_VERIFYING,
522 TD_FSYNCING,
523 TD_EXITED,
524 TD_REAPED,
525};
526
e29d1b70
JA
527/*
528 * Verify helpers
529 */
530extern void populate_verify_io_u(struct thread_data *, struct io_u *);
531extern int get_next_verify(struct thread_data *td, struct io_u *);
a9619d44 532extern int do_io_u_verify(struct thread_data *, struct io_u **);
e29d1b70 533
2f9ade3c
JA
534/*
535 * Memory helpers
536 */
537extern int fio_pin_memory(void);
538extern void fio_unpin_memory(void);
539extern int allocate_io_mem(struct thread_data *);
540extern void free_io_mem(struct thread_data *);
541
10ba535a
JA
542/*
543 * io unit handling
544 */
545#define queue_full(td) list_empty(&(td)->io_u_freelist)
546extern struct io_u *__get_io_u(struct thread_data *);
547extern struct io_u *get_io_u(struct thread_data *, struct fio_file *);
548extern void put_io_u(struct thread_data *, struct io_u *);
549extern void ios_completed(struct thread_data *, struct io_completion_data *);
550extern void io_completed(struct thread_data *, struct io_u *, struct io_completion_data *);
551
552/*
553 * io engine entry points
554 */
8c16d840 555extern int td_io_init(struct thread_data *);
10ba535a
JA
556extern int td_io_prep(struct thread_data *, struct io_u *);
557extern int td_io_queue(struct thread_data *, struct io_u *);
558extern int td_io_sync(struct thread_data *, struct fio_file *);
559extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
560
bbfd6b00
JA
561/*
562 * This is a pretty crappy semaphore implementation, but with the use that fio
563 * has (just signalling start/go conditions), it doesn't have to be better.
564 * Naturally this would not work for any type of contended semaphore or
565 * for real locking.
566 */
1056eaad 567static inline void fio_sem_init(volatile int *sem, int val)
bbfd6b00
JA
568{
569 *sem = val;
570}
571
1056eaad 572static inline void fio_sem_down(volatile int *sem)
bbfd6b00
JA
573{
574 while (*sem == 0)
575 usleep(10000);
576
577 (*sem)--;
578}
579
1056eaad 580static inline void fio_sem_up(volatile int *sem)
bbfd6b00
JA
581{
582 (*sem)++;
583}
584
3b70d7e5
JA
585/*
586 * If logging output to a file, stderr should go to both stderr and f_err
587 */
588#define log_err(args...) do { \
589 fprintf(f_err, ##args); \
590 if (f_err != stderr) \
591 fprintf(stderr, ##args); \
592 } while (0)
593
2866c82d 594struct ioengine_ops {
5f350952 595 struct list_head list;
2866c82d
JA
596 char name[16];
597 int version;
598 int flags;
ea2877a4 599 int (*setup)(struct thread_data *);
2866c82d
JA
600 int (*init)(struct thread_data *);
601 int (*prep)(struct thread_data *, struct io_u *);
602 int (*queue)(struct thread_data *, struct io_u *);
603 int (*getevents)(struct thread_data *, int, int, struct timespec *);
604 struct io_u *(*event)(struct thread_data *, int);
605 int (*cancel)(struct thread_data *, struct io_u *);
606 void (*cleanup)(struct thread_data *);
2866c82d
JA
607 void *data;
608 void *dlhandle;
609};
610
87dc1ab1 611#define FIO_IOOPS_VERSION 3
2866c82d 612
b4692828 613extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
5f350952
JA
614extern int register_ioengine(struct ioengine_ops *);
615extern void unregister_ioengine(struct ioengine_ops *);
2866c82d
JA
616extern void close_ioengine(struct thread_data *);
617
618/*
619 * Mark unused variables passed to ops functions as unused, to silence gcc
620 */
621#define fio_unused __attribute((__unused__))
5f350952
JA
622#define fio_init __attribute__((constructor))
623#define fio_exit __attribute__((destructor))
2866c82d 624
34572e28
JA
625#define for_each_td(td, i) \
626 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
53cdc686 627#define for_each_file(td, f, i) \
34572e28 628 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
53cdc686 629
ebac4655 630#endif