Better handling of file creation vs existing files
[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 FIO_FILE_EXISTS = 1 << 3, /* no need to create */
226};
227
228/*
229 * Each thread_data structure has a number of files associated with it,
230 * this structure holds state information for a single file.
231 */
232struct fio_file {
233 enum fio_filetype filetype;
234
235 /*
236 * A file may not be a file descriptor, let the io engine decide
237 */
238 union {
239 unsigned long file_data;
240 int fd;
241 };
242 char *file_name;
243 void *mmap;
244 unsigned long long file_size;
245 unsigned long long real_file_size;
246 unsigned long long file_offset;
247 unsigned long long last_pos;
248 unsigned long long last_completed_pos;
249
250 /*
251 * block map for random io
252 */
253 unsigned long *file_map;
254 unsigned int num_maps;
255 unsigned int last_free_lookup;
256
257 int references;
258 enum fio_file_flags flags;
259};
260
261/*
262 * How many depth levels to log
263 */
264#define FIO_IO_U_MAP_NR 8
265#define FIO_IO_U_LAT_NR 12
266
267struct thread_stat {
268 char *name;
269 char *verror;
270 int error;
271 int groupid;
272 pid_t pid;
273 char *description;
274 int members;
275
276 struct io_log *slat_log;
277 struct io_log *clat_log;
278 struct io_log *bw_log;
279
280 /*
281 * bandwidth and latency stats
282 */
283 struct io_stat clat_stat[2]; /* completion latency */
284 struct io_stat slat_stat[2]; /* submission latency */
285 struct io_stat bw_stat[2]; /* bandwidth stats */
286
287 unsigned long long stat_io_bytes[2];
288 struct timeval stat_sample_time[2];
289
290 /*
291 * fio system usage accounting
292 */
293 struct rusage ru_start;
294 struct rusage ru_end;
295 unsigned long usr_time;
296 unsigned long sys_time;
297 unsigned long ctx;
298
299 /*
300 * IO depth and latency stats
301 */
302 unsigned int io_u_map[FIO_IO_U_MAP_NR];
303 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
304 unsigned long total_io_u[2];
305
306 unsigned long long io_bytes[2];
307 unsigned long runtime[2];
308 unsigned long total_run_time;
309};
310
311/*
312 * This describes a single thread/process executing a fio job.
313 */
314struct thread_data {
315 int pad;
316 char *description;
317 char *name;
318 char *directory;
319 char *filename;
320 char *ioengine;
321 char verror[128];
322 pthread_t thread;
323 int thread_number;
324 int groupid;
325 struct thread_stat ts;
326 struct fio_file *files;
327 unsigned int files_index;
328 unsigned int nr_files;
329 unsigned int nr_open_files;
330 unsigned int nr_normal_files;
331 union {
332 unsigned int next_file;
333 os_random_state_t next_file_state;
334 };
335 int error;
336 pid_t pid;
337 char *orig_buffer;
338 size_t orig_buffer_size;
339 volatile int terminate;
340 volatile int runstate;
341 enum td_ddir td_ddir;
342 unsigned int ioprio;
343 unsigned int last_was_sync;
344
345 unsigned int odirect;
346 unsigned int invalidate_cache;
347 unsigned int create_serialize;
348 unsigned int create_fsync;
349 unsigned int end_fsync;
350 unsigned int sync_io;
351 unsigned int verify;
352 unsigned int use_thread;
353 unsigned int unlink;
354 unsigned int do_disk_util;
355 unsigned int override_sync;
356 unsigned int rand_repeatable;
357 unsigned int write_lat_log;
358 unsigned int write_bw_log;
359 unsigned int norandommap;
360 unsigned int bs_unaligned;
361 unsigned int fsync_on_close;
362
363 unsigned int bs[2];
364 unsigned int min_bs[2];
365 unsigned int max_bs[2];
366 unsigned int hugepage_size;
367 unsigned int rw_min_bs;
368 unsigned int thinktime;
369 unsigned int thinktime_spin;
370 unsigned int thinktime_blocks;
371 unsigned int fsync_blocks;
372 unsigned int start_delay;
373 unsigned long timeout;
374 unsigned int overwrite;
375 unsigned int bw_avg_time;
376 unsigned int loops;
377 unsigned long long zone_size;
378 unsigned long long zone_skip;
379 enum fio_memtype mem_type;
380 char *mmapfile;
381 int mmapfd;
382 unsigned int stonewall;
383 unsigned int numjobs;
384 unsigned int iodepth;
385 unsigned int iodepth_low;
386 unsigned int iodepth_batch;
387 os_cpu_mask_t cpumask;
388 unsigned int iolog;
389 unsigned int read_iolog;
390 unsigned int rwmixcycle;
391 unsigned int rwmixread;
392 unsigned int rwmixwrite;
393 unsigned int nice;
394 unsigned int file_service_type;
395 unsigned int group_reporting;
396 unsigned int open_files;
397 char *opendir;
398
399 char *read_iolog_file;
400 char *write_iolog_file;
401 void *iolog_buf;
402 FILE *iolog_f;
403
404 char *sysfs_root;
405 char *ioscheduler;
406
407 os_random_state_t bsrange_state;
408 os_random_state_t verify_state;
409
410 int shm_id;
411
412 /*
413 * IO engine hooks, contains everything needed to submit an io_u
414 * to any of the available IO engines.
415 */
416 struct ioengine_ops *io_ops;
417
418 /*
419 * Current IO depth and list of free and busy io_u's.
420 */
421 unsigned int cur_depth;
422 struct list_head io_u_freelist;
423 struct list_head io_u_busylist;
424 struct list_head io_u_requeues;
425 unsigned int io_u_queued;
426
427 /*
428 * Rate state
429 */
430 unsigned int rate;
431 unsigned int ratemin;
432 unsigned int ratecycle;
433 unsigned long rate_usec_cycle;
434 long rate_pending_usleep;
435 unsigned long rate_bytes;
436 struct timeval lastrate;
437
438 unsigned long long io_size;
439 unsigned long long total_file_size;
440 unsigned long long start_offset;
441 unsigned long long total_io_size;
442
443 unsigned long io_issues[2];
444 unsigned long long io_blocks[2];
445 unsigned long long io_bytes[2];
446 unsigned long long this_io_bytes[2];
447 unsigned long long zone_bytes;
448 struct fio_sem *mutex;
449
450 /*
451 * State for random io, a bitmap of blocks done vs not done
452 */
453 os_random_state_t random_state;
454
455 /*
456 * CPU "io" cycle burner
457 */
458 unsigned int cpuload;
459 unsigned int cpucycle;
460
461 struct timeval start; /* start of this loop */
462 struct timeval epoch; /* time job was started */
463
464 /*
465 * read/write mixed workload state
466 */
467 os_random_state_t rwmix_state;
468 struct timeval rwmix_switch;
469 enum fio_ddir rwmix_ddir;
470
471 /*
472 * Pre-run and post-run shell
473 */
474 char *exec_prerun;
475 char *exec_postrun;
476
477 /*
478 * IO historic logs
479 */
480 struct list_head io_hist_list;
481 struct list_head io_log_list;
482
483 /*
484 * timeout handling
485 */
486 struct timeval timeout_end;
487 struct itimerval timer;
488
489 /*
490 * for fileservice, how often to switch to a new file
491 */
492 unsigned int file_service_nr;
493 unsigned int file_service_left;
494 struct fio_file *file_service_file;
495};
496
497/*
498 * roundrobin available files, or choose one at random.
499 */
500enum {
501 FIO_FSERVICE_RANDOM = 1,
502 FIO_FSERVICE_RR = 2,
503};
504
505/*
506 * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
507 * the timer on each queue operation.
508 */
509#define IO_U_TIMEOUT_INC 5
510#define IO_U_TIMEOUT 30
511
512#define __td_verror(td, err, msg, func) \
513 do { \
514 if ((td)->error) \
515 break; \
516 int e = (err); \
517 (td)->error = e; \
518 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
519 } while (0)
520
521
522#define td_verror(td, err, func) \
523 __td_verror((td), (err), strerror((err)), (func))
524#define td_vmsg(td, err, msg, func) \
525 __td_verror((td), (err), (msg), (func))
526
527extern int exitall_on_terminate;
528extern int thread_number;
529extern int nr_process, nr_thread;
530extern int shm_id;
531extern int groupid;
532extern int terse_output;
533extern FILE *f_out;
534extern FILE *f_err;
535extern int temp_stall_ts;
536extern unsigned long long mlock_size;
537
538extern struct thread_data *threads;
539
540#define td_read(td) ((td)->td_ddir & TD_DDIR_READ)
541#define td_write(td) ((td)->td_ddir & TD_DDIR_WRITE)
542#define td_rw(td) (((td)->td_ddir & TD_DDIR_RW) == TD_DDIR_RW)
543#define td_random(td) ((td)->td_ddir & TD_DDIR_RAND)
544
545#define BLOCKS_PER_MAP (8 * sizeof(long))
546#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
547#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
548#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
549
550#define MAX_JOBS (1024)
551
552static inline int should_fsync(struct thread_data *td)
553{
554 if (td->last_was_sync)
555 return 0;
556 if (td->odirect)
557 return 0;
558 if (td_write(td) || td_rw(td) || td->override_sync)
559 return 1;
560
561 return 0;
562}
563
564/*
565 * Disk utils as read in /sys/block/<dev>/stat
566 */
567struct disk_util_stat {
568 unsigned ios[2];
569 unsigned merges[2];
570 unsigned long long sectors[2];
571 unsigned ticks[2];
572 unsigned io_ticks;
573 unsigned time_in_queue;
574};
575
576/*
577 * Per-device disk util management
578 */
579struct disk_util {
580 struct list_head list;
581
582 char *name;
583 char path[256];
584 dev_t dev;
585
586 struct disk_util_stat dus;
587 struct disk_util_stat last_dus;
588
589 unsigned long msec;
590 struct timeval time;
591};
592
593#define DISK_UTIL_MSEC (250)
594
595#ifndef min
596#define min(a, b) ((a) < (b) ? (a) : (b))
597#endif
598#ifndef max
599#define max(a, b) ((a) > (b) ? (a) : (b))
600#endif
601
602/*
603 * Log exports
604 */
605extern int __must_check read_iolog_get(struct thread_data *, struct io_u *);
606extern void write_iolog_put(struct thread_data *, struct io_u *);
607extern int __must_check init_iolog(struct thread_data *td);
608extern void log_io_piece(struct thread_data *, struct io_u *);
609extern void prune_io_piece_log(struct thread_data *);
610extern void write_iolog_close(struct thread_data *);
611
612/*
613 * Logging
614 */
615extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
616extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
617extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
618extern void show_run_stats(void);
619extern void init_disk_util(struct thread_data *);
620extern void update_rusage_stat(struct thread_data *);
621extern void update_io_ticks(void);
622extern void disk_util_timer_arm(void);
623extern void setup_log(struct io_log **);
624extern void finish_log(struct thread_data *, struct io_log *, const char *);
625extern void __finish_log(struct io_log *, const char *);
626extern int setup_rate(struct thread_data *);
627extern struct io_log *agg_io_log[2];
628extern int write_bw_log;
629extern void add_agg_sample(unsigned long, enum fio_ddir);
630
631/*
632 * Time functions
633 */
634extern unsigned long utime_since(struct timeval *, struct timeval *);
635extern unsigned long utime_since_now(struct timeval *);
636extern unsigned long mtime_since(struct timeval *, struct timeval *);
637extern unsigned long mtime_since_now(struct timeval *);
638extern unsigned long time_since_now(struct timeval *);
639extern unsigned long mtime_since_genesis(void);
640extern void __usec_sleep(unsigned int);
641extern void usec_sleep(struct thread_data *, unsigned long);
642extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
643extern void fill_start_time(struct timeval *);
644extern void fio_gettime(struct timeval *, void *);
645extern void set_genesis_time(void);
646
647/*
648 * Init functions
649 */
650extern int __must_check parse_options(int, char **);
651extern int __must_check init_random_state(struct thread_data *);
652
653/*
654 * File setup/shutdown
655 */
656extern void close_files(struct thread_data *);
657extern int __must_check setup_files(struct thread_data *);
658extern int __must_check open_files(struct thread_data *);
659extern int __must_check file_invalidate_cache(struct thread_data *, struct fio_file *);
660extern int __must_check generic_open_file(struct thread_data *, struct fio_file *);
661extern void generic_close_file(struct thread_data *, struct fio_file *);
662extern void add_file(struct thread_data *, const char *);
663extern void get_file(struct fio_file *);
664extern void put_file(struct thread_data *, struct fio_file *);
665extern int add_dir_files(struct thread_data *, const char *);
666
667/*
668 * ETA/status stuff
669 */
670extern void print_thread_status(void);
671extern void print_status_init(int);
672
673/*
674 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
675 * will never back again. It may cycle between running/verififying/fsyncing.
676 * Once the thread reaches TD_EXITED, it is just waiting for the core to
677 * reap it.
678 */
679enum {
680 TD_NOT_CREATED = 0,
681 TD_CREATED,
682 TD_INITIALIZED,
683 TD_RUNNING,
684 TD_VERIFYING,
685 TD_FSYNCING,
686 TD_EXITED,
687 TD_REAPED,
688};
689
690/*
691 * Verify helpers
692 */
693extern void populate_verify_io_u(struct thread_data *, struct io_u *);
694extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
695extern int __must_check verify_io_u(struct io_u *);
696
697/*
698 * Memory helpers
699 */
700extern int __must_check fio_pin_memory(void);
701extern void fio_unpin_memory(void);
702extern int __must_check allocate_io_mem(struct thread_data *);
703extern void free_io_mem(struct thread_data *);
704
705/*
706 * io unit handling
707 */
708#define queue_full(td) list_empty(&(td)->io_u_freelist)
709extern struct io_u *__get_io_u(struct thread_data *);
710extern struct io_u *get_io_u(struct thread_data *);
711extern void put_io_u(struct thread_data *, struct io_u *);
712extern void requeue_io_u(struct thread_data *, struct io_u **);
713extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *);
714extern long __must_check io_u_queued_complete(struct thread_data *, int);
715extern void io_u_queued(struct thread_data *, struct io_u *);
716extern void io_u_log_error(struct thread_data *, struct io_u *);
717extern void io_u_init_timeout(void);
718extern void io_u_set_timeout(struct thread_data *);
719extern void io_u_mark_depth(struct thread_data *, struct io_u *);
720
721/*
722 * io engine entry points
723 */
724extern int __must_check td_io_init(struct thread_data *);
725extern int __must_check td_io_prep(struct thread_data *, struct io_u *);
726extern int __must_check td_io_queue(struct thread_data *, struct io_u *);
727extern int __must_check td_io_sync(struct thread_data *, struct fio_file *);
728extern int __must_check td_io_getevents(struct thread_data *, int, int, struct timespec *);
729extern int __must_check td_io_commit(struct thread_data *);
730extern int __must_check td_io_open_file(struct thread_data *, struct fio_file *);
731extern void td_io_close_file(struct thread_data *, struct fio_file *);
732
733/*
734 * If logging output to a file, stderr should go to both stderr and f_err
735 */
736#define log_err(args...) do { \
737 fprintf(f_err, ##args); \
738 if (f_err != stderr) \
739 fprintf(stderr, ##args); \
740 } while (0)
741
742FILE *get_f_out(void);
743FILE *get_f_err(void);
744
745struct ioengine_ops {
746 struct list_head list;
747 char name[16];
748 int version;
749 int flags;
750 int (*setup)(struct thread_data *);
751 int (*init)(struct thread_data *);
752 int (*prep)(struct thread_data *, struct io_u *);
753 int (*queue)(struct thread_data *, struct io_u *);
754 int (*commit)(struct thread_data *);
755 int (*getevents)(struct thread_data *, int, int, struct timespec *);
756 struct io_u *(*event)(struct thread_data *, int);
757 int (*cancel)(struct thread_data *, struct io_u *);
758 void (*cleanup)(struct thread_data *);
759 int (*open_file)(struct thread_data *, struct fio_file *);
760 void (*close_file)(struct thread_data *, struct fio_file *);
761 void *data;
762 void *dlhandle;
763};
764
765#define FIO_IOOPS_VERSION 6
766
767extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
768extern void register_ioengine(struct ioengine_ops *);
769extern void unregister_ioengine(struct ioengine_ops *);
770extern void close_ioengine(struct thread_data *);
771
772/*
773 * Mark unused variables passed to ops functions as unused, to silence gcc
774 */
775#define fio_unused __attribute((__unused__))
776#define fio_init __attribute__((constructor))
777#define fio_exit __attribute__((destructor))
778
779#define for_each_td(td, i) \
780 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
781#define for_each_file(td, f, i) \
782 for ((i) = 0, (f) = &(td)->files[0]; (i) < (td)->nr_files; (i)++, (f)++)
783
784#define fio_assert(td, cond) do { \
785 if (!(cond)) { \
786 int *__foo = NULL; \
787 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
788 (td)->runstate = TD_EXITED; \
789 (td)->error = EFAULT; \
790 *__foo = 0; \
791 } \
792} while (0)
793
794#endif