projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Fix sync engine completion latency
[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#ifdef FIO_HAVE_SYSLET
22#include "syslet.h"
23#endif
24
25enum fio_ddir {
26 DDIR_READ = 0,
27 DDIR_WRITE,
28 DDIR_SYNC,
29};
30
31/*
32 * Use for maintaining statistics
33 */
34struct io_stat {
35 unsigned long max_val;
36 unsigned long min_val;
37 unsigned long samples;
38
39 double mean;
40 double S;
41};
42
43/*
44 * A single data sample
45 */
46struct io_sample {
47 unsigned long time;
48 unsigned long val;
49 enum fio_ddir ddir;
50};
51
52/*
53 * Dynamically growing data sample log
54 */
55struct io_log {
56 unsigned long nr_samples;
57 unsigned long max_samples;
58 struct io_sample *log;
59};
60
61/*
62 * When logging io actions, this matches a single sent io_u
63 */
64struct io_piece {
65 struct list_head list;
66 struct fio_file *file;
67 unsigned long long offset;
68 unsigned long len;
69 enum fio_ddir ddir;
70};
71
72#ifdef FIO_HAVE_SYSLET
73struct syslet_req {
74 struct syslet_uatom atom;
75 long ret;
76};
77#endif
78
79/*
80 * The io unit
81 */
82struct io_u {
83 union {
84#ifdef FIO_HAVE_LIBAIO
85 struct iocb iocb;
86#endif
87#ifdef FIO_HAVE_POSIXAIO
88 struct aiocb aiocb;
89#endif
90#ifdef FIO_HAVE_SGIO
91 struct sg_io_hdr hdr;
92#endif
93#ifdef FIO_HAVE_SYSLET
94 struct syslet_req req;
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/*
133 * io_ops->queue() return values
134 */
135enum {
136 FIO_Q_COMPLETED = 0, /* completed sync */
137 FIO_Q_QUEUED = 1, /* queued, will complete async */
138 FIO_Q_BUSY = 2, /* no more room, call ->commit() */
139};
140
141#define FIO_HDR_MAGIC 0xf00baaef
142
143enum {
144 VERIFY_NONE = 0, /* no verification */
145 VERIFY_MD5, /* md5 sum data blocks */
146 VERIFY_CRC32, /* crc32 sum data blocks */
147};
148
149/*
150 * A header structure associated with each checksummed data block
151 */
152struct verify_header {
153 unsigned int fio_magic;
154 unsigned int len;
155 unsigned int verify_type;
156 union {
157 char md5_digest[MD5_HASH_WORDS * 4];
158 unsigned long crc32;
159 };
160};
161
162struct group_run_stats {
163 unsigned long long max_run[2], min_run[2];
164 unsigned long long max_bw[2], min_bw[2];
165 unsigned long long io_kb[2];
166 unsigned long long agg[2];
167};
168
169/*
170 * What type of allocation to use for io buffers
171 */
172enum fio_memtype {
173 MEM_MALLOC = 0, /* ordinary malloc */
174 MEM_SHM, /* use shared memory segments */
175 MEM_SHMHUGE, /* use shared memory segments with huge pages */
176 MEM_MMAP, /* use anonynomous mmap */
177 MEM_MMAPHUGE, /* memory mapped huge file */
178};
179
180/*
181 * The type of object we are working on
182 */
183enum fio_filetype {
184 FIO_TYPE_FILE = 1, /* plain file */
185 FIO_TYPE_BD, /* block device */
186 FIO_TYPE_CHAR, /* character device */
187};
188
189enum fio_ioengine_flags {
190 FIO_SYNCIO = 1 << 0, /* io engine has synchronous ->queue */
191 FIO_CPUIO = 1 << 1, /* cpu burner, doesn't do real io */
192 FIO_MMAPIO = 1 << 2, /* uses memory mapped io */
193 FIO_RAWIO = 1 << 3, /* some sort of direct/raw io */
194 FIO_NETIO = 1 << 4, /* networked io */
195 FIO_NULLIO = 1 << 5, /* no real data transfer (cpu/null) */
196};
197
198/*
199 * Each thread_data structure has a number of files associated with it,
200 * this structure holds state information for a single file.
201 */
202struct fio_file {
203 /*
204 * A file may not be a file descriptor, let the io engine decide
205 */
206 union {
207 unsigned long file_data;
208 int fd;
209 };
210 char *file_name;
211 void *mmap;
212 unsigned long long file_size;
213 unsigned long long real_file_size;
214 unsigned long long file_offset;
215 unsigned long long last_pos;
216 unsigned long long last_completed_pos;
217
218 /*
219 * block map for random io
220 */
221 unsigned long *file_map;
222 unsigned int num_maps;
223 unsigned int last_free_lookup;
224
225 unsigned int unlink;
226};
227
228/*
229 * How many depth levels to log
230 */
231#define FIO_IO_U_MAP_NR 8
232#define FIO_IO_U_LAT_NR 12
233
234/*
235 * This describes a single thread/process executing a fio job.
236 */
237struct thread_data {
238 char *description;
239 char *name;
240 char *directory;
241 char *filename;
242 char verror[80];
243 pthread_t thread;
244 int thread_number;
245 int groupid;
246 enum fio_filetype filetype;
247 struct fio_file *files;
248 unsigned int nr_files;
249 unsigned int nr_uniq_files;
250 unsigned int next_file;
251 int error;
252 pid_t pid;
253 char *orig_buffer;
254 size_t orig_buffer_size;
255 volatile int terminate;
256 volatile int runstate;
257 enum fio_ddir ddir;
258 unsigned int iomix;
259 unsigned int ioprio;
260 unsigned int last_was_sync;
261
262 unsigned int sequential;
263 unsigned int odirect;
264 unsigned int invalidate_cache;
265 unsigned int create_serialize;
266 unsigned int create_fsync;
267 unsigned int end_fsync;
268 unsigned int sync_io;
269 unsigned int verify;
270 unsigned int use_thread;
271 unsigned int unlink;
272 unsigned int do_disk_util;
273 unsigned int override_sync;
274 unsigned int rand_repeatable;
275 unsigned int write_lat_log;
276 unsigned int write_bw_log;
277 unsigned int norandommap;
278 unsigned int bs_unaligned;
279
280 unsigned int bs[2];
281 unsigned int min_bs[2];
282 unsigned int max_bs[2];
283 unsigned int hugepage_size;
284 unsigned int rw_min_bs;
285 unsigned int thinktime;
286 unsigned int thinktime_spin;
287 unsigned int thinktime_blocks;
288 unsigned int fsync_blocks;
289 unsigned int start_delay;
290 unsigned long timeout;
291 unsigned int overwrite;
292 unsigned int bw_avg_time;
293 unsigned int loops;
294 unsigned long long zone_size;
295 unsigned long long zone_skip;
296 enum fio_memtype mem_type;
297 char *mmapfile;
298 int mmapfd;
299 unsigned int stonewall;
300 unsigned int numjobs;
301 unsigned int iodepth;
302 os_cpu_mask_t cpumask;
303 unsigned int iolog;
304 unsigned int read_iolog;
305 unsigned int rwmixcycle;
306 unsigned int rwmixread;
307 unsigned int rwmixwrite;
308 unsigned int nice;
309
310 char *read_iolog_file;
311 char *write_iolog_file;
312 void *iolog_buf;
313 FILE *iolog_f;
314
315 char *sysfs_root;
316 char *ioscheduler;
317
318 os_random_state_t bsrange_state;
319 os_random_state_t verify_state;
320
321 int shm_id;
322
323 /*
324 * IO engine hooks, contains everything needed to submit an io_u
325 * to any of the available IO engines.
326 */
327 struct ioengine_ops *io_ops;
328
329 /*
330 * Current IO depth and list of free and busy io_u's.
331 */
332 unsigned int cur_depth;
333 unsigned int io_u_map[FIO_IO_U_MAP_NR];
334 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
335 unsigned long total_io_u;
336 struct list_head io_u_freelist;
337 struct list_head io_u_busylist;
338 struct list_head io_u_requeues;
339
340 /*
341 * Rate state
342 */
343 unsigned int rate;
344 unsigned int ratemin;
345 unsigned int ratecycle;
346 unsigned long rate_usec_cycle;
347 long rate_pending_usleep;
348 unsigned long rate_bytes;
349 struct timeval lastrate;
350
351 unsigned long runtime[2]; /* msec */
352 unsigned long long io_size;
353 unsigned long long total_file_size;
354 unsigned long long start_offset;
355 unsigned long long total_io_size;
356
357 unsigned long io_issues[2];
358 unsigned long long io_blocks[2];
359 unsigned long long io_bytes[2];
360 unsigned long long zone_bytes;
361 unsigned long long this_io_bytes[2];
362 volatile int mutex;
363
364 /*
365 * State for random io, a bitmap of blocks done vs not done
366 */
367 os_random_state_t random_state;
368
369 /*
370 * CPU "io" cycle burner
371 */
372 unsigned int cpuload;
373 unsigned int cpucycle;
374
375 /*
376 * bandwidth and latency stats
377 */
378 struct io_stat clat_stat[2]; /* completion latency */
379 struct io_stat slat_stat[2]; /* submission latency */
380 struct io_stat bw_stat[2]; /* bandwidth stats */
381
382 unsigned long long stat_io_bytes[2];
383 struct timeval stat_sample_time[2];
384
385 struct io_log *slat_log;
386 struct io_log *clat_log;
387 struct io_log *bw_log;
388
389 struct timeval start; /* start of this loop */
390 struct timeval epoch; /* time job was started */
391 struct timeval end_time;/* time job ended */
392
393 /*
394 * fio system usage accounting
395 */
396 struct rusage ru_start;
397 struct rusage ru_end;
398 unsigned long usr_time;
399 unsigned long sys_time;
400 unsigned long ctx;
401
402 /*
403 * read/write mixed workload state
404 */
405 os_random_state_t rwmix_state;
406 struct timeval rwmix_switch;
407 enum fio_ddir rwmix_ddir;
408
409 /*
410 * Pre-run and post-run shell
411 */
412 char *exec_prerun;
413 char *exec_postrun;
414
415 /*
416 * IO historic logs
417 */
418 struct list_head io_hist_list;
419 struct list_head io_log_list;
420};
421
422#define __td_verror(td, err, msg) \
423 do { \
424 int e = (err); \
425 (td)->error = e; \
426 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, error=%s", __FILE__, __LINE__, (msg)); \
427 } while (0)
428
429
430#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
431#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
432
433extern int exitall_on_terminate;
434extern int thread_number;
435extern int shm_id;
436extern int groupid;
437extern int terse_output;
438extern FILE *f_out;
439extern FILE *f_err;
440extern int temp_stall_ts;
441extern unsigned long long mlock_size;
442
443extern struct thread_data *threads;
444
445#define td_read(td) ((td)->ddir == DDIR_READ)
446#define td_write(td) ((td)->ddir == DDIR_WRITE)
447#define td_rw(td) ((td)->iomix != 0)
448
449#define BLOCKS_PER_MAP (8 * sizeof(long))
450#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
451#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
452#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
453
454#define MAX_JOBS (1024)
455
456static inline int should_fsync(struct thread_data *td)
457{
458 if (td->last_was_sync)
459 return 0;
460 if (td->odirect)
461 return 0;
462 if (td_write(td) || td_rw(td) || td->override_sync)
463 return 1;
464
465 return 0;
466}
467
468/*
469 * Disk utils as read in /sys/block/<dev>/stat
470 */
471struct disk_util_stat {
472 unsigned ios[2];
473 unsigned merges[2];
474 unsigned long long sectors[2];
475 unsigned ticks[2];
476 unsigned io_ticks;
477 unsigned time_in_queue;
478};
479
480/*
481 * Per-device disk util management
482 */
483struct disk_util {
484 struct list_head list;
485
486 char *name;
487 char path[256];
488 dev_t dev;
489
490 struct disk_util_stat dus;
491 struct disk_util_stat last_dus;
492
493 unsigned long msec;
494 struct timeval time;
495};
496
497/*
498 * Callback for io completion
499 */
500typedef int (endio_handler)(struct io_u *);
501
502#define DISK_UTIL_MSEC (250)
503
504#ifndef min
505#define min(a, b) ((a) < (b) ? (a) : (b))
506#endif
507#ifndef max
508#define max(a, b) ((a) > (b) ? (a) : (b))
509#endif
510
511/*
512 * Log exports
513 */
514extern int read_iolog_get(struct thread_data *, struct io_u *);
515extern void write_iolog_put(struct thread_data *, struct io_u *);
516extern int init_iolog(struct thread_data *td);
517extern void log_io_piece(struct thread_data *, struct io_u *);
518extern void prune_io_piece_log(struct thread_data *);
519extern void write_iolog_close(struct thread_data *);
520
521/*
522 * Logging
523 */
524extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
525extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
526extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
527extern void show_run_stats(void);
528extern void init_disk_util(struct thread_data *);
529extern void update_rusage_stat(struct thread_data *);
530extern void update_io_ticks(void);
531extern void disk_util_timer_arm(void);
532extern void setup_log(struct io_log **);
533extern void finish_log(struct thread_data *, struct io_log *, const char *);
534extern void __finish_log(struct io_log *, const char *);
535extern int setup_rate(struct thread_data *);
536extern struct io_log *agg_io_log[2];
537extern int write_bw_log;
538extern void add_agg_sample(unsigned long, enum fio_ddir);
539
540/*
541 * Time functions
542 */
543extern unsigned long utime_since(struct timeval *, struct timeval *);
544extern unsigned long utime_since_now(struct timeval *);
545extern unsigned long mtime_since(struct timeval *, struct timeval *);
546extern unsigned long mtime_since_now(struct timeval *);
547extern unsigned long time_since_now(struct timeval *);
548extern unsigned long mtime_since_genesis(void);
549extern void __usec_sleep(unsigned int);
550extern void usec_sleep(struct thread_data *, unsigned long);
551extern void rate_throttle(struct thread_data *, unsigned long, unsigned int, int);
552extern void fill_start_time(struct timeval *);
553extern void fio_gettime(struct timeval *, void *);
554
555/*
556 * Init functions
557 */
558extern int parse_options(int, char **);
559extern int init_random_state(struct thread_data *);
560
561/*
562 * File setup/shutdown
563 */
564extern void close_files(struct thread_data *);
565extern int setup_files(struct thread_data *);
566extern int open_files(struct thread_data *);
567extern int file_invalidate_cache(struct thread_data *, struct fio_file *);
568
569/*
570 * ETA/status stuff
571 */
572extern void print_thread_status(void);
573extern void print_status_init(int);
574
575/*
576 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
577 * will never back again. It may cycle between running/verififying/fsyncing.
578 * Once the thread reaches TD_EXITED, it is just waiting for the core to
579 * reap it.
580 */
581enum {
582 TD_NOT_CREATED = 0,
583 TD_CREATED,
584 TD_INITIALIZED,
585 TD_RUNNING,
586 TD_VERIFYING,
587 TD_FSYNCING,
588 TD_EXITED,
589 TD_REAPED,
590};
591
592/*
593 * Verify helpers
594 */
595extern void populate_verify_io_u(struct thread_data *, struct io_u *);
596extern int get_next_verify(struct thread_data *td, struct io_u *);
597extern int verify_io_u(struct io_u *);
598
599/*
600 * Memory helpers
601 */
602extern int fio_pin_memory(void);
603extern void fio_unpin_memory(void);
604extern int allocate_io_mem(struct thread_data *);
605extern void free_io_mem(struct thread_data *);
606
607/*
608 * io unit handling
609 */
610#define queue_full(td) list_empty(&(td)->io_u_freelist)
611extern struct io_u *__get_io_u(struct thread_data *);
612extern struct io_u *get_io_u(struct thread_data *);
613extern void put_io_u(struct thread_data *, struct io_u *);
614extern void requeue_io_u(struct thread_data *, struct io_u **);
615extern long io_u_sync_complete(struct thread_data *, struct io_u *, endio_handler *);
616extern long io_u_queued_complete(struct thread_data *, int, endio_handler *);
617extern void io_u_queued(struct thread_data *, struct io_u *);
618
619/*
620 * io engine entry points
621 */
622extern int td_io_init(struct thread_data *);
623extern int td_io_prep(struct thread_data *, struct io_u *);
624extern int td_io_queue(struct thread_data *, struct io_u *);
625extern int td_io_sync(struct thread_data *, struct fio_file *);
626extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
627extern int td_io_commit(struct thread_data *);
628
629/*
630 * This is a pretty crappy semaphore implementation, but with the use that fio
631 * has (just signalling start/go conditions), it doesn't have to be better.
632 * Naturally this would not work for any type of contended semaphore or
633 * for real locking.
634 */
635static inline void fio_sem_init(volatile int *sem, int val)
636{
637 *sem = val;
638}
639
640static inline void fio_sem_down(volatile int *sem)
641{
642 while (*sem == 0)
643 usleep(10000);
644
645 (*sem)--;
646}
647
648static inline void fio_sem_up(volatile int *sem)
649{
650 (*sem)++;
651}
652
653/*
654 * If logging output to a file, stderr should go to both stderr and f_err
655 */
656#define log_err(args...) do { \
657 fprintf(f_err, ##args); \
658 if (f_err != stderr) \
659 fprintf(stderr, ##args); \
660 } while (0)
661
662struct ioengine_ops {
663 struct list_head list;
664 char name[16];
665 int version;
666 int flags;
667 int (*setup)(struct thread_data *);
668 int (*init)(struct thread_data *);
669 int (*prep)(struct thread_data *, struct io_u *);
670 int (*queue)(struct thread_data *, struct io_u *);
671 int (*commit)(struct thread_data *);
672 int (*getevents)(struct thread_data *, int, int, struct timespec *);
673 struct io_u *(*event)(struct thread_data *, int);
674 int (*cancel)(struct thread_data *, struct io_u *);
675 void (*cleanup)(struct thread_data *);
676 void *data;
677 void *dlhandle;
678 unsigned long priv;
679};
680
681#define FIO_IOOPS_VERSION 5
682
683extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
684extern int register_ioengine(struct ioengine_ops *);
685extern void unregister_ioengine(struct ioengine_ops *);
686extern void close_ioengine(struct thread_data *);
687
688/*
689 * Mark unused variables passed to ops functions as unused, to silence gcc
690 */
691#define fio_unused __attribute((__unused__))
692#define fio_init __attribute__((constructor))
693#define fio_exit __attribute__((destructor))
694
695#define for_each_td(td, i) \
696 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
697#define for_each_file(td, f, i) \
698 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
699
700#define fio_assert(td, cond) do { \
701 if (!(cond)) { \
702 int *__foo = NULL; \
703 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
704 (td)->runstate = TD_EXITED; \
705 (td)->error = EFAULT; \
706 *__foo = 0; \
707 } \
708} while (0)
709
710#endif
fio - Flexible IO Tester