projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
[PATCH] Let the syslet ioengine register check supportability
[fio.git] / fio.h
... / ...
CommitLineData
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>
9#include <errno.h>
10#include <stdlib.h>
11#include <stdio.h>
12#include <unistd.h>
13#include <string.h>
14
15#include "list.h"
16#include "md5.h"
17#include "crc32.h"
18#include "arch.h"
19#include "os.h"
20
21#include "syslet.h"
22
23enum fio_ddir {
24 DDIR_READ = 0,
25 DDIR_WRITE,
26 DDIR_SYNC,
27};
28
29/*
30 * Use for maintaining statistics
31 */
32struct io_stat {
33 unsigned long max_val;
34 unsigned long min_val;
35 unsigned long samples;
36
37 double mean;
38 double S;
39};
40
41/*
42 * A single data sample
43 */
44struct io_sample {
45 unsigned long time;
46 unsigned long val;
47 enum fio_ddir ddir;
48};
49
50/*
51 * Dynamically growing data sample log
52 */
53struct io_log {
54 unsigned long nr_samples;
55 unsigned long max_samples;
56 struct io_sample *log;
57};
58
59/*
60 * When logging io actions, this matches a single sent io_u
61 */
62struct io_piece {
63 struct list_head list;
64 struct fio_file *file;
65 unsigned long long offset;
66 unsigned long len;
67 enum fio_ddir ddir;
68};
69
70#ifdef FIO_HAVE_SYSLET
71struct syslet_req {
72 struct syslet_uatom atom;
73 unsigned long cmd;
74 long ret;
75};
76#endif
77
78/*
79 * The io unit
80 */
81struct io_u {
82 union {
83#ifdef FIO_HAVE_LIBAIO
84 struct iocb iocb;
85#endif
86#ifdef FIO_HAVE_POSIXAIO
87 struct aiocb aiocb;
88#endif
89#ifdef FIO_HAVE_SGIO
90 struct sg_io_hdr hdr;
91#endif
92#ifdef FIO_HAVE_SYSLET
93 struct syslet_req rw_atom;
94 struct syslet_req seek_atom;
95#endif
96 };
97 struct timeval start_time;
98 struct timeval issue_time;
99
100 /*
101 * Allocated/set buffer and length
102 */
103 void *buf;
104 unsigned long buflen;
105 unsigned long long offset;
106
107 /*
108 * IO engine state, may be different from above when we get
109 * partial transfers / residual data counts
110 */
111 void *xfer_buf;
112 unsigned long xfer_buflen;
113
114 unsigned int resid;
115 unsigned int error;
116
117 enum fio_ddir ddir;
118
119 /*
120 * io engine private data
121 */
122 union {
123 unsigned int index;
124 unsigned int seen;
125 };
126
127 struct fio_file *file;
128
129 struct list_head list;
130};
131
132#define FIO_HDR_MAGIC 0xf00baaef
133
134enum {
135 VERIFY_NONE = 0, /* no verification */
136 VERIFY_MD5, /* md5 sum data blocks */
137 VERIFY_CRC32, /* crc32 sum data blocks */
138};
139
140/*
141 * A header structure associated with each checksummed data block
142 */
143struct verify_header {
144 unsigned int fio_magic;
145 unsigned int len;
146 unsigned int verify_type;
147 union {
148 char md5_digest[MD5_HASH_WORDS * 4];
149 unsigned long crc32;
150 };
151};
152
153struct group_run_stats {
154 unsigned long long max_run[2], min_run[2];
155 unsigned long long max_bw[2], min_bw[2];
156 unsigned long long io_kb[2];
157 unsigned long long agg[2];
158};
159
160/*
161 * What type of allocation to use for io buffers
162 */
163enum fio_memtype {
164 MEM_MALLOC = 0, /* ordinary malloc */
165 MEM_SHM, /* use shared memory segments */
166 MEM_SHMHUGE, /* use shared memory segments with huge pages */
167 MEM_MMAP, /* use anonynomous mmap */
168 MEM_MMAPHUGE, /* memory mapped huge file */
169};
170
171/*
172 * The type of object we are working on
173 */
174enum fio_filetype {
175 FIO_TYPE_FILE = 1, /* plain file */
176 FIO_TYPE_BD, /* block device */
177 FIO_TYPE_CHAR, /* character device */
178};
179
180enum fio_ioengine_flags {
181 FIO_SYNCIO = 1 << 0, /* io engine has synchronous ->queue */
182 FIO_CPUIO = 1 << 1, /* cpu burner, doesn't do real io */
183 FIO_MMAPIO = 1 << 2, /* uses memory mapped io */
184 FIO_RAWIO = 1 << 3, /* some sort of direct/raw io */
185 FIO_NETIO = 1 << 4, /* networked io */
186 FIO_NULLIO = 1 << 5, /* no real data transfer (cpu/null) */
187};
188
189/*
190 * Each thread_data structure has a number of files associated with it,
191 * this structure holds state information for a single file.
192 */
193struct fio_file {
194 /*
195 * A file may not be a file descriptor, let the io engine decide
196 */
197 union {
198 unsigned long file_data;
199 long fd;
200 };
201 char *file_name;
202 void *mmap;
203 unsigned long long file_size;
204 unsigned long long real_file_size;
205 unsigned long long file_offset;
206 unsigned long long last_pos;
207 unsigned long long last_completed_pos;
208
209 /*
210 * block map for random io
211 */
212 unsigned long *file_map;
213 unsigned int num_maps;
214 unsigned int last_free_lookup;
215
216 unsigned int unlink;
217};
218
219/*
220 * How many depth levels to log
221 */
222#define FIO_IO_U_MAP_NR 8
223
224/*
225 * This describes a single thread/process executing a fio job.
226 */
227struct thread_data {
228 char *description;
229 char *name;
230 char *directory;
231 char *filename;
232 char verror[80];
233 pthread_t thread;
234 int thread_number;
235 int groupid;
236 enum fio_filetype filetype;
237 struct fio_file *files;
238 unsigned int nr_files;
239 unsigned int nr_uniq_files;
240 unsigned int next_file;
241 int error;
242 pid_t pid;
243 char *orig_buffer;
244 size_t orig_buffer_size;
245 volatile int terminate;
246 volatile int runstate;
247 enum fio_ddir ddir;
248 unsigned int iomix;
249 unsigned int ioprio;
250 unsigned int last_was_sync;
251
252 unsigned int sequential;
253 unsigned int odirect;
254 unsigned int invalidate_cache;
255 unsigned int create_serialize;
256 unsigned int create_fsync;
257 unsigned int end_fsync;
258 unsigned int sync_io;
259 unsigned int verify;
260 unsigned int use_thread;
261 unsigned int unlink;
262 unsigned int do_disk_util;
263 unsigned int override_sync;
264 unsigned int rand_repeatable;
265 unsigned int write_lat_log;
266 unsigned int write_bw_log;
267 unsigned int norandommap;
268 unsigned int bs_unaligned;
269
270 unsigned int bs[2];
271 unsigned int min_bs[2];
272 unsigned int max_bs[2];
273 unsigned int hugepage_size;
274 unsigned int rw_min_bs;
275 unsigned int thinktime;
276 unsigned int thinktime_blocks;
277 unsigned int fsync_blocks;
278 unsigned int start_delay;
279 unsigned long timeout;
280 unsigned int overwrite;
281 unsigned int bw_avg_time;
282 unsigned int loops;
283 unsigned long long zone_size;
284 unsigned long long zone_skip;
285 enum fio_memtype mem_type;
286 char *mmapfile;
287 int mmapfd;
288 unsigned int stonewall;
289 unsigned int numjobs;
290 unsigned int iodepth;
291 os_cpu_mask_t cpumask;
292 unsigned int iolog;
293 unsigned int read_iolog;
294 unsigned int rwmixcycle;
295 unsigned int rwmixread;
296 unsigned int rwmixwrite;
297 unsigned int nice;
298
299 char *read_iolog_file;
300 char *write_iolog_file;
301 void *iolog_buf;
302 FILE *iolog_f;
303
304 char *sysfs_root;
305 char *ioscheduler;
306
307 os_random_state_t bsrange_state;
308 os_random_state_t verify_state;
309
310 int shm_id;
311
312 /*
313 * IO engine hooks, contains everything needed to submit an io_u
314 * to any of the available IO engines.
315 */
316 struct ioengine_ops *io_ops;
317
318 /*
319 * Current IO depth and list of free and busy io_u's.
320 */
321 unsigned int cur_depth;
322 unsigned int io_u_map[FIO_IO_U_MAP_NR];
323 unsigned long total_io_u;
324 struct list_head io_u_freelist;
325 struct list_head io_u_busylist;
326
327 /*
328 * Rate state
329 */
330 unsigned int rate;
331 unsigned int ratemin;
332 unsigned int ratecycle;
333 unsigned long rate_usec_cycle;
334 long rate_pending_usleep;
335 unsigned long rate_bytes;
336 struct timeval lastrate;
337
338 unsigned long runtime[2]; /* msec */
339 unsigned long long io_size;
340 unsigned long long total_file_size;
341 unsigned long long start_offset;
342 unsigned long long total_io_size;
343
344 unsigned long long io_blocks[2];
345 unsigned long long io_bytes[2];
346 unsigned long long zone_bytes;
347 unsigned long long this_io_bytes[2];
348 volatile int mutex;
349
350 /*
351 * State for random io, a bitmap of blocks done vs not done
352 */
353 os_random_state_t random_state;
354
355 /*
356 * CPU "io" cycle burner
357 */
358 unsigned int cpuload;
359 unsigned int cpucycle;
360
361 /*
362 * bandwidth and latency stats
363 */
364 struct io_stat clat_stat[2]; /* completion latency */
365 struct io_stat slat_stat[2]; /* submission latency */
366 struct io_stat bw_stat[2]; /* bandwidth stats */
367
368 unsigned long long stat_io_bytes[2];
369 struct timeval stat_sample_time[2];
370
371 struct io_log *slat_log;
372 struct io_log *clat_log;
373 struct io_log *bw_log;
374
375 struct timeval start; /* start of this loop */
376 struct timeval epoch; /* time job was started */
377 struct timeval end_time;/* time job ended */
378
379 /*
380 * fio system usage accounting
381 */
382 struct rusage ru_start;
383 struct rusage ru_end;
384 unsigned long usr_time;
385 unsigned long sys_time;
386 unsigned long ctx;
387
388 /*
389 * read/write mixed workload state
390 */
391 os_random_state_t rwmix_state;
392 struct timeval rwmix_switch;
393 enum fio_ddir rwmix_ddir;
394
395 /*
396 * Pre-run and post-run shell
397 */
398 char *exec_prerun;
399 char *exec_postrun;
400
401 /*
402 * IO historic logs
403 */
404 struct list_head io_hist_list;
405 struct list_head io_log_list;
406};
407
408#define __td_verror(td, err, msg) \
409 do { \
410 int e = (err); \
411 (td)->error = e; \
412 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, error=%s", __FILE__, __LINE__, (msg)); \
413 } while (0)
414
415
416#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
417#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
418
419extern int exitall_on_terminate;
420extern int thread_number;
421extern int shm_id;
422extern int groupid;
423extern int terse_output;
424extern FILE *f_out;
425extern FILE *f_err;
426extern int temp_stall_ts;
427extern unsigned long long mlock_size;
428
429extern struct thread_data *threads;
430
431#define td_read(td) ((td)->ddir == DDIR_READ)
432#define td_write(td) ((td)->ddir == DDIR_WRITE)
433#define td_rw(td) ((td)->iomix != 0)
434
435#define BLOCKS_PER_MAP (8 * sizeof(long))
436#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
437#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
438#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
439
440#define MAX_JOBS (1024)
441
442static inline int should_fsync(struct thread_data *td)
443{
444 if (td->last_was_sync)
445 return 0;
446 if (td->odirect)
447 return 0;
448 if (td_write(td) || td_rw(td) || td->override_sync)
449 return 1;
450
451 return 0;
452}
453
454/*
455 * Disk utils as read in /sys/block/<dev>/stat
456 */
457struct disk_util_stat {
458 unsigned ios[2];
459 unsigned merges[2];
460 unsigned long long sectors[2];
461 unsigned ticks[2];
462 unsigned io_ticks;
463 unsigned time_in_queue;
464};
465
466/*
467 * Per-device disk util management
468 */
469struct disk_util {
470 struct list_head list;
471
472 char *name;
473 char path[256];
474 dev_t dev;
475
476 struct disk_util_stat dus;
477 struct disk_util_stat last_dus;
478
479 unsigned long msec;
480 struct timeval time;
481};
482
483/*
484 * Used for passing io_u completion data
485 */
486struct io_completion_data {
487 int nr; /* input */
488
489 int error; /* output */
490 unsigned long bytes_done[2]; /* output */
491 struct timeval time; /* output */
492};
493
494#define DISK_UTIL_MSEC (250)
495
496#ifndef min
497#define min(a, b) ((a) < (b) ? (a) : (b))
498#endif
499#ifndef max
500#define max(a, b) ((a) > (b) ? (a) : (b))
501#endif
502
503/*
504 * Log exports
505 */
506extern int read_iolog_get(struct thread_data *, struct io_u *);
507extern void write_iolog_put(struct thread_data *, struct io_u *);
508extern int init_iolog(struct thread_data *td);
509extern void log_io_piece(struct thread_data *, struct io_u *);
510extern void prune_io_piece_log(struct thread_data *);
511extern void write_iolog_close(struct thread_data *);
512
513/*
514 * Logging
515 */
516extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
517extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
518extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
519extern void show_run_stats(void);
520extern void init_disk_util(struct thread_data *);
521extern void update_rusage_stat(struct thread_data *);
522extern void update_io_ticks(void);
523extern void disk_util_timer_arm(void);
524extern void setup_log(struct io_log **);
525extern void finish_log(struct thread_data *, struct io_log *, const char *);
526extern void __finish_log(struct io_log *, const char *);
527extern int setup_rate(struct thread_data *);
528extern struct io_log *agg_io_log[2];
529extern int write_bw_log;
530extern void add_agg_sample(unsigned long, enum fio_ddir);
531
532/*
533 * Time functions
534 */
535extern unsigned long utime_since(struct timeval *, struct timeval *);
536extern unsigned long utime_since_now(struct timeval *);
537extern unsigned long mtime_since(struct timeval *, struct timeval *);
538extern unsigned long mtime_since_now(struct timeval *);
539extern unsigned long time_since_now(struct timeval *);
540extern unsigned long mtime_since_genesis(void);
541extern void __usec_sleep(unsigned int);
542extern void usec_sleep(struct thread_data *, unsigned long);
543extern void rate_throttle(struct thread_data *, unsigned long, unsigned int, int);
544extern void fill_start_time(struct timeval *);
545extern void fio_gettime(struct timeval *, void *);
546
547/*
548 * Init functions
549 */
550extern int parse_options(int, char **);
551extern int init_random_state(struct thread_data *);
552
553/*
554 * File setup/shutdown
555 */
556extern void close_files(struct thread_data *);
557extern int setup_files(struct thread_data *);
558extern int open_files(struct thread_data *);
559extern int file_invalidate_cache(struct thread_data *, struct fio_file *);
560
561/*
562 * ETA/status stuff
563 */
564extern void print_thread_status(void);
565extern void print_status_init(int);
566
567/*
568 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
569 * will never back again. It may cycle between running/verififying/fsyncing.
570 * Once the thread reaches TD_EXITED, it is just waiting for the core to
571 * reap it.
572 */
573enum {
574 TD_NOT_CREATED = 0,
575 TD_CREATED,
576 TD_INITIALIZED,
577 TD_RUNNING,
578 TD_VERIFYING,
579 TD_FSYNCING,
580 TD_EXITED,
581 TD_REAPED,
582};
583
584/*
585 * Verify helpers
586 */
587extern void populate_verify_io_u(struct thread_data *, struct io_u *);
588extern int get_next_verify(struct thread_data *td, struct io_u *);
589extern int do_io_u_verify(struct thread_data *, struct io_u **);
590
591/*
592 * Memory helpers
593 */
594extern int fio_pin_memory(void);
595extern void fio_unpin_memory(void);
596extern int allocate_io_mem(struct thread_data *);
597extern void free_io_mem(struct thread_data *);
598
599/*
600 * io unit handling
601 */
602#define queue_full(td) list_empty(&(td)->io_u_freelist)
603extern struct io_u *__get_io_u(struct thread_data *);
604extern struct io_u *get_io_u(struct thread_data *, struct fio_file *);
605extern void put_io_u(struct thread_data *, struct io_u *);
606extern void ios_completed(struct thread_data *, struct io_completion_data *);
607extern void io_completed(struct thread_data *, struct io_u *, struct io_completion_data *);
608
609/*
610 * io engine entry points
611 */
612extern int td_io_init(struct thread_data *);
613extern int td_io_prep(struct thread_data *, struct io_u *);
614extern int td_io_queue(struct thread_data *, struct io_u *);
615extern int td_io_sync(struct thread_data *, struct fio_file *);
616extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
617
618/*
619 * This is a pretty crappy semaphore implementation, but with the use that fio
620 * has (just signalling start/go conditions), it doesn't have to be better.
621 * Naturally this would not work for any type of contended semaphore or
622 * for real locking.
623 */
624static inline void fio_sem_init(volatile int *sem, int val)
625{
626 *sem = val;
627}
628
629static inline void fio_sem_down(volatile int *sem)
630{
631 while (*sem == 0)
632 usleep(10000);
633
634 (*sem)--;
635}
636
637static inline void fio_sem_up(volatile int *sem)
638{
639 (*sem)++;
640}
641
642/*
643 * If logging output to a file, stderr should go to both stderr and f_err
644 */
645#define log_err(args...) do { \
646 fprintf(f_err, ##args); \
647 if (f_err != stderr) \
648 fprintf(stderr, ##args); \
649 } while (0)
650
651struct ioengine_ops {
652 struct list_head list;
653 char name[16];
654 int version;
655 int flags;
656 int (*setup)(struct thread_data *);
657 int (*init)(struct thread_data *);
658 int (*prep)(struct thread_data *, struct io_u *);
659 int (*queue)(struct thread_data *, struct io_u *);
660 int (*getevents)(struct thread_data *, int, int, struct timespec *);
661 struct io_u *(*event)(struct thread_data *, int);
662 int (*cancel)(struct thread_data *, struct io_u *);
663 void (*cleanup)(struct thread_data *);
664 void *data;
665 void *dlhandle;
666};
667
668#define FIO_IOOPS_VERSION 3
669
670extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
671extern int register_ioengine(struct ioengine_ops *);
672extern void unregister_ioengine(struct ioengine_ops *);
673extern void close_ioengine(struct thread_data *);
674
675/*
676 * Mark unused variables passed to ops functions as unused, to silence gcc
677 */
678#define fio_unused __attribute((__unused__))
679#define fio_init __attribute__((constructor))
680#define fio_exit __attribute__((destructor))
681
682#define for_each_td(td, i) \
683 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
684#define for_each_file(td, f, i) \
685 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
686
687#define fio_assert(td, cond) do { \
688 if (!(cond)) { \
689 int *__foo = NULL; \
690 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
691 (td)->runstate = TD_EXITED; \
692 (td)->error = EFAULT; \
693 *__foo = 0; \
694 } \
695} while (0)
696
697#endif
fio - Flexible IO Tester