projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Total file size handling fix
[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#include "mutex.h"
21
22#ifdef FIO_HAVE_SYSLET
23#include "syslet.h"
24#endif
25
26enum fio_ddir {
27 DDIR_READ = 0,
28 DDIR_WRITE,
29 DDIR_SYNC,
30};
31
32enum td_ddir {
33 TD_DDIR_READ = 1 << 0,
34 TD_DDIR_WRITE = 1 << 1,
35 TD_DDIR_RAND = 1 << 2,
36 TD_DDIR_RW = TD_DDIR_READ | TD_DDIR_WRITE,
37 TD_DDIR_RANDREAD = TD_DDIR_READ | TD_DDIR_RAND,
38 TD_DDIR_RANDWRITE = TD_DDIR_WRITE | TD_DDIR_RAND,
39 TD_DDIR_RANDRW = TD_DDIR_RW | TD_DDIR_RAND,
40};
41
42/*
43 * Use for maintaining statistics
44 */
45struct io_stat {
46 unsigned long max_val;
47 unsigned long min_val;
48 unsigned long samples;
49
50 double mean;
51 double S;
52};
53
54/*
55 * A single data sample
56 */
57struct io_sample {
58 unsigned long time;
59 unsigned long val;
60 enum fio_ddir ddir;
61};
62
63/*
64 * Dynamically growing data sample log
65 */
66struct io_log {
67 unsigned long nr_samples;
68 unsigned long max_samples;
69 struct io_sample *log;
70};
71
72/*
73 * When logging io actions, this matches a single sent io_u
74 */
75struct io_piece {
76 struct list_head list;
77 struct fio_file *file;
78 unsigned long long offset;
79 unsigned long len;
80 enum fio_ddir ddir;
81};
82
83#ifdef FIO_HAVE_SYSLET
84struct syslet_req {
85 struct syslet_uatom atom; /* the atom to submit */
86 struct syslet_uatom *head; /* head of the sequence */
87 long ret; /* syscall return value */
88};
89#endif
90
91enum {
92 IO_U_F_FREE = 1 << 0,
93 IO_U_F_FLIGHT = 1 << 1,
94};
95
96/*
97 * The io unit
98 */
99struct io_u {
100 union {
101#ifdef FIO_HAVE_LIBAIO
102 struct iocb iocb;
103#endif
104#ifdef FIO_HAVE_POSIXAIO
105 struct aiocb aiocb;
106#endif
107#ifdef FIO_HAVE_SGIO
108 struct sg_io_hdr hdr;
109#endif
110#ifdef FIO_HAVE_SYSLET
111 struct syslet_req req;
112#endif
113 };
114 struct timeval start_time;
115 struct timeval issue_time;
116
117 /*
118 * Allocated/set buffer and length
119 */
120 void *buf;
121 unsigned long buflen;
122 unsigned long long offset;
123
124 /*
125 * IO engine state, may be different from above when we get
126 * partial transfers / residual data counts
127 */
128 void *xfer_buf;
129 unsigned long xfer_buflen;
130
131 unsigned int resid;
132 unsigned int error;
133
134 enum fio_ddir ddir;
135
136 /*
137 * io engine private data
138 */
139 union {
140 unsigned int index;
141 unsigned int seen;
142 };
143
144 unsigned int flags;
145
146 struct fio_file *file;
147
148 struct list_head list;
149
150 /*
151 * Callback for io completion
152 */
153 int (*end_io)(struct io_u *);
154};
155
156/*
157 * io_ops->queue() return values
158 */
159enum {
160 FIO_Q_COMPLETED = 0, /* completed sync */
161 FIO_Q_QUEUED = 1, /* queued, will complete async */
162 FIO_Q_BUSY = 2, /* no more room, call ->commit() */
163};
164
165#define FIO_HDR_MAGIC 0xf00baaef
166
167enum {
168 VERIFY_NONE = 0, /* no verification */
169 VERIFY_MD5, /* md5 sum data blocks */
170 VERIFY_CRC32, /* crc32 sum data blocks */
171};
172
173/*
174 * A header structure associated with each checksummed data block
175 */
176struct verify_header {
177 unsigned int fio_magic;
178 unsigned int len;
179 unsigned int verify_type;
180 union {
181 char md5_digest[MD5_HASH_WORDS * 4];
182 unsigned long crc32;
183 };
184};
185
186struct group_run_stats {
187 unsigned long long max_run[2], min_run[2];
188 unsigned long long max_bw[2], min_bw[2];
189 unsigned long long io_kb[2];
190 unsigned long long agg[2];
191};
192
193/*
194 * What type of allocation to use for io buffers
195 */
196enum fio_memtype {
197 MEM_MALLOC = 0, /* ordinary malloc */
198 MEM_SHM, /* use shared memory segments */
199 MEM_SHMHUGE, /* use shared memory segments with huge pages */
200 MEM_MMAP, /* use anonynomous mmap */
201 MEM_MMAPHUGE, /* memory mapped huge file */
202};
203
204/*
205 * The type of object we are working on
206 */
207enum fio_filetype {
208 FIO_TYPE_FILE = 1, /* plain file */
209 FIO_TYPE_BD, /* block device */
210 FIO_TYPE_CHAR, /* character device */
211};
212
213enum fio_ioengine_flags {
214 FIO_SYNCIO = 1 << 0, /* io engine has synchronous ->queue */
215 FIO_RAWIO = 1 << 1, /* some sort of direct/raw io */
216 FIO_DISKLESSIO = 1 << 2, /* no disk involved */
217 FIO_NOEXTEND = 1 << 3, /* engine can't extend file */
218 FIO_NODISKUTIL = 1 << 4, /* diskutil can't handle filename */
219};
220
221enum fio_file_flags {
222 FIO_FILE_OPEN = 1 << 0, /* file is open */
223 FIO_FILE_UNLINK = 1 << 1, /* unlink on close */
224 FIO_FILE_CLOSING = 1 << 2, /* file being closed */
225};
226
227/*
228 * Each thread_data structure has a number of files associated with it,
229 * this structure holds state information for a single file.
230 */
231struct fio_file {
232 enum fio_filetype filetype;
233
234 /*
235 * A file may not be a file descriptor, let the io engine decide
236 */
237 union {
238 unsigned long file_data;
239 int fd;
240 };
241 char *file_name;
242 void *mmap;
243 unsigned long long file_size;
244 unsigned long long real_file_size;
245 unsigned long long file_offset;
246 unsigned long long last_pos;
247 unsigned long long last_completed_pos;
248
249 /*
250 * block map for random io
251 */
252 unsigned long *file_map;
253 unsigned int num_maps;
254 unsigned int last_free_lookup;
255
256 int references;
257 enum fio_file_flags flags;
258};
259
260/*
261 * How many depth levels to log
262 */
263#define FIO_IO_U_MAP_NR 8
264#define FIO_IO_U_LAT_NR 12
265
266struct thread_stat {
267 char *name;
268 char *verror;
269 int error;
270 int groupid;
271 pid_t pid;
272 char *description;
273 int members;
274
275 struct io_log *slat_log;
276 struct io_log *clat_log;
277 struct io_log *bw_log;
278
279 /*
280 * bandwidth and latency stats
281 */
282 struct io_stat clat_stat[2]; /* completion latency */
283 struct io_stat slat_stat[2]; /* submission latency */
284 struct io_stat bw_stat[2]; /* bandwidth stats */
285
286 unsigned long long stat_io_bytes[2];
287 struct timeval stat_sample_time[2];
288
289 /*
290 * fio system usage accounting
291 */
292 struct rusage ru_start;
293 struct rusage ru_end;
294 unsigned long usr_time;
295 unsigned long sys_time;
296 unsigned long ctx;
297
298 /*
299 * IO depth and latency stats
300 */
301 unsigned int io_u_map[FIO_IO_U_MAP_NR];
302 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
303 unsigned long total_io_u[2];
304
305 unsigned long long io_bytes[2];
306 unsigned long runtime[2];
307 unsigned long total_run_time;
308};
309
310/*
311 * This describes a single thread/process executing a fio job.
312 */
313struct thread_data {
314 int pad;
315 char *description;
316 char *name;
317 char *directory;
318 char *filename;
319 char *ioengine;
320 char verror[128];
321 pthread_t thread;
322 int thread_number;
323 int groupid;
324 struct thread_stat ts;
325 struct fio_file *files;
326 unsigned int files_index;
327 unsigned int nr_files;
328 unsigned int nr_open_files;
329 unsigned int nr_normal_files;
330 union {
331 unsigned int next_file;
332 os_random_state_t next_file_state;
333 };
334 int error;
335 pid_t pid;
336 char *orig_buffer;
337 size_t orig_buffer_size;
338 volatile int terminate;
339 volatile int runstate;
340 enum td_ddir td_ddir;
341 unsigned int ioprio;
342 unsigned int last_was_sync;
343
344 unsigned int odirect;
345 unsigned int invalidate_cache;
346 unsigned int create_serialize;
347 unsigned int create_fsync;
348 unsigned int end_fsync;
349 unsigned int sync_io;
350 unsigned int verify;
351 unsigned int use_thread;
352 unsigned int unlink;
353 unsigned int do_disk_util;
354 unsigned int override_sync;
355 unsigned int rand_repeatable;
356 unsigned int write_lat_log;
357 unsigned int write_bw_log;
358 unsigned int norandommap;
359 unsigned int bs_unaligned;
360
361 unsigned int bs[2];
362 unsigned int min_bs[2];
363 unsigned int max_bs[2];
364 unsigned int hugepage_size;
365 unsigned int rw_min_bs;
366 unsigned int thinktime;
367 unsigned int thinktime_spin;
368 unsigned int thinktime_blocks;
369 unsigned int fsync_blocks;
370 unsigned int start_delay;
371 unsigned long timeout;
372 unsigned int overwrite;
373 unsigned int bw_avg_time;
374 unsigned int loops;
375 unsigned long long zone_size;
376 unsigned long long zone_skip;
377 enum fio_memtype mem_type;
378 char *mmapfile;
379 int mmapfd;
380 unsigned int stonewall;
381 unsigned int numjobs;
382 unsigned int iodepth;
383 unsigned int iodepth_low;
384 unsigned int iodepth_batch;
385 os_cpu_mask_t cpumask;
386 unsigned int iolog;
387 unsigned int read_iolog;
388 unsigned int rwmixcycle;
389 unsigned int rwmixread;
390 unsigned int rwmixwrite;
391 unsigned int nice;
392 unsigned int file_service_type;
393 unsigned int group_reporting;
394 unsigned int open_files;
395
396 char *read_iolog_file;
397 char *write_iolog_file;
398 void *iolog_buf;
399 FILE *iolog_f;
400
401 char *sysfs_root;
402 char *ioscheduler;
403
404 os_random_state_t bsrange_state;
405 os_random_state_t verify_state;
406
407 int shm_id;
408
409 /*
410 * IO engine hooks, contains everything needed to submit an io_u
411 * to any of the available IO engines.
412 */
413 struct ioengine_ops *io_ops;
414
415 /*
416 * Current IO depth and list of free and busy io_u's.
417 */
418 unsigned int cur_depth;
419 struct list_head io_u_freelist;
420 struct list_head io_u_busylist;
421 struct list_head io_u_requeues;
422 unsigned int io_u_queued;
423
424 /*
425 * Rate state
426 */
427 unsigned int rate;
428 unsigned int ratemin;
429 unsigned int ratecycle;
430 unsigned long rate_usec_cycle;
431 long rate_pending_usleep;
432 unsigned long rate_bytes;
433 struct timeval lastrate;
434
435 unsigned long long io_size;
436 unsigned long long total_file_size;
437 unsigned long long start_offset;
438 unsigned long long total_io_size;
439
440 unsigned long io_issues[2];
441 unsigned long long io_blocks[2];
442 unsigned long long io_bytes[2];
443 unsigned long long this_io_bytes[2];
444 unsigned long long zone_bytes;
445 struct fio_sem *mutex;
446
447 /*
448 * State for random io, a bitmap of blocks done vs not done
449 */
450 os_random_state_t random_state;
451
452 /*
453 * CPU "io" cycle burner
454 */
455 unsigned int cpuload;
456 unsigned int cpucycle;
457
458 struct timeval start; /* start of this loop */
459 struct timeval epoch; /* time job was started */
460
461 /*
462 * read/write mixed workload state
463 */
464 os_random_state_t rwmix_state;
465 struct timeval rwmix_switch;
466 enum fio_ddir rwmix_ddir;
467
468 /*
469 * Pre-run and post-run shell
470 */
471 char *exec_prerun;
472 char *exec_postrun;
473
474 /*
475 * IO historic logs
476 */
477 struct list_head io_hist_list;
478 struct list_head io_log_list;
479
480 /*
481 * timeout handling
482 */
483 struct timeval timeout_end;
484 struct itimerval timer;
485
486 /*
487 * for fileservice, how often to switch to a new file
488 */
489 unsigned int file_service_nr;
490 unsigned int file_service_left;
491 struct fio_file *file_service_file;
492};
493
494/*
495 * roundrobin available files, or choose one at random.
496 */
497enum {
498 FIO_FSERVICE_RANDOM = 1,
499 FIO_FSERVICE_RR = 2,
500};
501
502/*
503 * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
504 * the timer on each queue operation.
505 */
506#define IO_U_TIMEOUT_INC 5
507#define IO_U_TIMEOUT 30
508
509#define __td_verror(td, err, msg, func) \
510 do { \
511 if ((td)->error) \
512 break; \
513 int e = (err); \
514 (td)->error = e; \
515 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
516 } while (0)
517
518
519#define td_verror(td, err, func) \
520 __td_verror((td), (err), strerror((err)), (func))
521#define td_vmsg(td, err, msg, func) \
522 __td_verror((td), (err), (msg), (func))
523
524extern int exitall_on_terminate;
525extern int thread_number;
526extern int nr_process, nr_thread;
527extern int shm_id;
528extern int groupid;
529extern int terse_output;
530extern FILE *f_out;
531extern FILE *f_err;
532extern int temp_stall_ts;
533extern unsigned long long mlock_size;
534
535extern struct thread_data *threads;
536
537#define td_read(td) ((td)->td_ddir & TD_DDIR_READ)
538#define td_write(td) ((td)->td_ddir & TD_DDIR_WRITE)
539#define td_rw(td) (((td)->td_ddir & TD_DDIR_RW) == TD_DDIR_RW)
540#define td_random(td) ((td)->td_ddir & TD_DDIR_RAND)
541
542#define BLOCKS_PER_MAP (8 * sizeof(long))
543#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
544#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
545#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
546
547#define MAX_JOBS (1024)
548
549static inline int should_fsync(struct thread_data *td)
550{
551 if (td->last_was_sync)
552 return 0;
553 if (td->odirect)
554 return 0;
555 if (td_write(td) || td_rw(td) || td->override_sync)
556 return 1;
557
558 return 0;
559}
560
561/*
562 * Disk utils as read in /sys/block/<dev>/stat
563 */
564struct disk_util_stat {
565 unsigned ios[2];
566 unsigned merges[2];
567 unsigned long long sectors[2];
568 unsigned ticks[2];
569 unsigned io_ticks;
570 unsigned time_in_queue;
571};
572
573/*
574 * Per-device disk util management
575 */
576struct disk_util {
577 struct list_head list;
578
579 char *name;
580 char path[256];
581 dev_t dev;
582
583 struct disk_util_stat dus;
584 struct disk_util_stat last_dus;
585
586 unsigned long msec;
587 struct timeval time;
588};
589
590#define DISK_UTIL_MSEC (250)
591
592#ifndef min
593#define min(a, b) ((a) < (b) ? (a) : (b))
594#endif
595#ifndef max
596#define max(a, b) ((a) > (b) ? (a) : (b))
597#endif
598
599/*
600 * Log exports
601 */
602extern int __must_check read_iolog_get(struct thread_data *, struct io_u *);
603extern void write_iolog_put(struct thread_data *, struct io_u *);
604extern int __must_check init_iolog(struct thread_data *td);
605extern void log_io_piece(struct thread_data *, struct io_u *);
606extern void prune_io_piece_log(struct thread_data *);
607extern void write_iolog_close(struct thread_data *);
608
609/*
610 * Logging
611 */
612extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
613extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
614extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
615extern void show_run_stats(void);
616extern void init_disk_util(struct thread_data *);
617extern void update_rusage_stat(struct thread_data *);
618extern void update_io_ticks(void);
619extern void disk_util_timer_arm(void);
620extern void setup_log(struct io_log **);
621extern void finish_log(struct thread_data *, struct io_log *, const char *);
622extern void __finish_log(struct io_log *, const char *);
623extern int setup_rate(struct thread_data *);
624extern struct io_log *agg_io_log[2];
625extern int write_bw_log;
626extern void add_agg_sample(unsigned long, enum fio_ddir);
627
628/*
629 * Time functions
630 */
631extern unsigned long utime_since(struct timeval *, struct timeval *);
632extern unsigned long utime_since_now(struct timeval *);
633extern unsigned long mtime_since(struct timeval *, struct timeval *);
634extern unsigned long mtime_since_now(struct timeval *);
635extern unsigned long time_since_now(struct timeval *);
636extern unsigned long mtime_since_genesis(void);
637extern void __usec_sleep(unsigned int);
638extern void usec_sleep(struct thread_data *, unsigned long);
639extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
640extern void fill_start_time(struct timeval *);
641extern void fio_gettime(struct timeval *, void *);
642extern void set_genesis_time(void);
643
644/*
645 * Init functions
646 */
647extern int __must_check parse_options(int, char **);
648extern int __must_check init_random_state(struct thread_data *);
649
650/*
651 * File setup/shutdown
652 */
653extern void close_files(struct thread_data *);
654extern int __must_check setup_files(struct thread_data *);
655extern int __must_check open_files(struct thread_data *);
656extern int __must_check file_invalidate_cache(struct thread_data *, struct fio_file *);
657extern int __must_check generic_open_file(struct thread_data *, struct fio_file *);
658extern void generic_close_file(struct thread_data *, struct fio_file *);
659extern void add_file(struct thread_data *, const char *);
660extern void get_file(struct fio_file *);
661extern void put_file(struct thread_data *, struct fio_file *);
662
663/*
664 * ETA/status stuff
665 */
666extern void print_thread_status(void);
667extern void print_status_init(int);
668
669/*
670 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
671 * will never back again. It may cycle between running/verififying/fsyncing.
672 * Once the thread reaches TD_EXITED, it is just waiting for the core to
673 * reap it.
674 */
675enum {
676 TD_NOT_CREATED = 0,
677 TD_CREATED,
678 TD_INITIALIZED,
679 TD_RUNNING,
680 TD_VERIFYING,
681 TD_FSYNCING,
682 TD_EXITED,
683 TD_REAPED,
684};
685
686/*
687 * Verify helpers
688 */
689extern void populate_verify_io_u(struct thread_data *, struct io_u *);
690extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
691extern int __must_check verify_io_u(struct io_u *);
692
693/*
694 * Memory helpers
695 */
696extern int __must_check fio_pin_memory(void);
697extern void fio_unpin_memory(void);
698extern int __must_check allocate_io_mem(struct thread_data *);
699extern void free_io_mem(struct thread_data *);
700
701/*
702 * io unit handling
703 */
704#define queue_full(td) list_empty(&(td)->io_u_freelist)
705extern struct io_u *__get_io_u(struct thread_data *);
706extern struct io_u *get_io_u(struct thread_data *);
707extern void put_io_u(struct thread_data *, struct io_u *);
708extern void requeue_io_u(struct thread_data *, struct io_u **);
709extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *);
710extern long __must_check io_u_queued_complete(struct thread_data *, int);
711extern void io_u_queued(struct thread_data *, struct io_u *);
712extern void io_u_log_error(struct thread_data *, struct io_u *);
713extern void io_u_init_timeout(void);
714extern void io_u_set_timeout(struct thread_data *);
715extern void io_u_mark_depth(struct thread_data *, struct io_u *);
716
717/*
718 * io engine entry points
719 */
720extern int __must_check td_io_init(struct thread_data *);
721extern int __must_check td_io_prep(struct thread_data *, struct io_u *);
722extern int __must_check td_io_queue(struct thread_data *, struct io_u *);
723extern int __must_check td_io_sync(struct thread_data *, struct fio_file *);
724extern int __must_check td_io_getevents(struct thread_data *, int, int, struct timespec *);
725extern int __must_check td_io_commit(struct thread_data *);
726extern int __must_check td_io_open_file(struct thread_data *, struct fio_file *);
727extern void td_io_close_file(struct thread_data *, struct fio_file *);
728
729/*
730 * If logging output to a file, stderr should go to both stderr and f_err
731 */
732#define log_err(args...) do { \
733 fprintf(f_err, ##args); \
734 if (f_err != stderr) \
735 fprintf(stderr, ##args); \
736 } while (0)
737
738FILE *get_f_out(void);
739FILE *get_f_err(void);
740
741struct ioengine_ops {
742 struct list_head list;
743 char name[16];
744 int version;
745 int flags;
746 int (*setup)(struct thread_data *);
747 int (*init)(struct thread_data *);
748 int (*prep)(struct thread_data *, struct io_u *);
749 int (*queue)(struct thread_data *, struct io_u *);
750 int (*commit)(struct thread_data *);
751 int (*getevents)(struct thread_data *, int, int, struct timespec *);
752 struct io_u *(*event)(struct thread_data *, int);
753 int (*cancel)(struct thread_data *, struct io_u *);
754 void (*cleanup)(struct thread_data *);
755 int (*open_file)(struct thread_data *, struct fio_file *);
756 void (*close_file)(struct thread_data *, struct fio_file *);
757 void *data;
758 void *dlhandle;
759};
760
761#define FIO_IOOPS_VERSION 6
762
763extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
764extern void register_ioengine(struct ioengine_ops *);
765extern void unregister_ioengine(struct ioengine_ops *);
766extern void close_ioengine(struct thread_data *);
767
768/*
769 * Mark unused variables passed to ops functions as unused, to silence gcc
770 */
771#define fio_unused __attribute((__unused__))
772#define fio_init __attribute__((constructor))
773#define fio_exit __attribute__((destructor))
774
775#define for_each_td(td, i) \
776 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
777#define for_each_file(td, f, i) \
778 for ((i) = 0, (f) = &(td)->files[0]; (i) < (td)->nr_files; (i)++, (f)++)
779
780#define fio_assert(td, cond) do { \
781 if (!(cond)) { \
782 int *__foo = NULL; \
783 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
784 (td)->runstate = TD_EXITED; \
785 (td)->error = EFAULT; \
786 *__foo = 0; \
787 } \
788} while (0)
789
790#endif
fio - Flexible IO Tester