projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Excessive passing around of struct fio_file
[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
31enum td_ddir {
32 TD_DDIR_READ = 1 << 0,
33 TD_DDIR_WRITE = 1 << 1,
34 TD_DDIR_RAND = 1 << 2,
35 TD_DDIR_RW = TD_DDIR_READ | TD_DDIR_WRITE,
36 TD_DDIR_RANDREAD = TD_DDIR_READ | TD_DDIR_RAND,
37 TD_DDIR_RANDWRITE = TD_DDIR_WRITE | TD_DDIR_READ,
38 TD_DDIR_RANDRW = TD_DDIR_RW | TD_DDIR_RAND,
39};
40
41/*
42 * Use for maintaining statistics
43 */
44struct io_stat {
45 unsigned long max_val;
46 unsigned long min_val;
47 unsigned long samples;
48
49 double mean;
50 double S;
51};
52
53/*
54 * A single data sample
55 */
56struct io_sample {
57 unsigned long time;
58 unsigned long val;
59 enum fio_ddir ddir;
60};
61
62/*
63 * Dynamically growing data sample log
64 */
65struct io_log {
66 unsigned long nr_samples;
67 unsigned long max_samples;
68 struct io_sample *log;
69};
70
71/*
72 * When logging io actions, this matches a single sent io_u
73 */
74struct io_piece {
75 struct list_head list;
76 struct fio_file *file;
77 unsigned long long offset;
78 unsigned long len;
79 enum fio_ddir ddir;
80};
81
82#ifdef FIO_HAVE_SYSLET
83struct syslet_req {
84 struct syslet_uatom atom; /* the atom to submit */
85 struct syslet_uatom *head; /* head of the sequence */
86 long ret; /* syscall return value */
87};
88#endif
89
90enum {
91 IO_U_F_FREE = 1 << 0,
92 IO_U_F_FLIGHT = 1 << 1,
93};
94
95/*
96 * The io unit
97 */
98struct io_u {
99 union {
100#ifdef FIO_HAVE_LIBAIO
101 struct iocb iocb;
102#endif
103#ifdef FIO_HAVE_POSIXAIO
104 struct aiocb aiocb;
105#endif
106#ifdef FIO_HAVE_SGIO
107 struct sg_io_hdr hdr;
108#endif
109#ifdef FIO_HAVE_SYSLET
110 struct syslet_req req;
111#endif
112 };
113 struct timeval start_time;
114 struct timeval issue_time;
115
116 /*
117 * Allocated/set buffer and length
118 */
119 void *buf;
120 unsigned long buflen;
121 unsigned long long offset;
122
123 /*
124 * IO engine state, may be different from above when we get
125 * partial transfers / residual data counts
126 */
127 void *xfer_buf;
128 unsigned long xfer_buflen;
129
130 unsigned int resid;
131 unsigned int error;
132
133 enum fio_ddir ddir;
134
135 /*
136 * io engine private data
137 */
138 union {
139 unsigned int index;
140 unsigned int seen;
141 };
142
143 unsigned int flags;
144
145 struct fio_file *file;
146
147 struct list_head list;
148
149 /*
150 * Callback for io completion
151 */
152 int (*end_io)(struct io_u *);
153};
154
155/*
156 * io_ops->queue() return values
157 */
158enum {
159 FIO_Q_COMPLETED = 0, /* completed sync */
160 FIO_Q_QUEUED = 1, /* queued, will complete async */
161 FIO_Q_BUSY = 2, /* no more room, call ->commit() */
162};
163
164#define FIO_HDR_MAGIC 0xf00baaef
165
166enum {
167 VERIFY_NONE = 0, /* no verification */
168 VERIFY_MD5, /* md5 sum data blocks */
169 VERIFY_CRC32, /* crc32 sum data blocks */
170};
171
172/*
173 * A header structure associated with each checksummed data block
174 */
175struct verify_header {
176 unsigned int fio_magic;
177 unsigned int len;
178 unsigned int verify_type;
179 union {
180 char md5_digest[MD5_HASH_WORDS * 4];
181 unsigned long crc32;
182 };
183};
184
185struct group_run_stats {
186 unsigned long long max_run[2], min_run[2];
187 unsigned long long max_bw[2], min_bw[2];
188 unsigned long long io_kb[2];
189 unsigned long long agg[2];
190};
191
192/*
193 * What type of allocation to use for io buffers
194 */
195enum fio_memtype {
196 MEM_MALLOC = 0, /* ordinary malloc */
197 MEM_SHM, /* use shared memory segments */
198 MEM_SHMHUGE, /* use shared memory segments with huge pages */
199 MEM_MMAP, /* use anonynomous mmap */
200 MEM_MMAPHUGE, /* memory mapped huge file */
201};
202
203/*
204 * The type of object we are working on
205 */
206enum fio_filetype {
207 FIO_TYPE_FILE = 1, /* plain file */
208 FIO_TYPE_BD, /* block device */
209 FIO_TYPE_CHAR, /* character device */
210};
211
212enum fio_ioengine_flags {
213 FIO_SYNCIO = 1 << 0, /* io engine has synchronous ->queue */
214 FIO_CPUIO = 1 << 1, /* cpu burner, doesn't do real io */
215 FIO_MMAPIO = 1 << 2, /* uses memory mapped io */
216 FIO_RAWIO = 1 << 3, /* some sort of direct/raw io */
217 FIO_DISKLESSIO = 1 << 4, /* no disk involved */
218 FIO_SELFOPEN = 1 << 5, /* opens its own devices */
219 FIO_NULLIO = 1 << 6, /* no real data transfer (cpu/null) */
220};
221
222/*
223 * Each thread_data structure has a number of files associated with it,
224 * this structure holds state information for a single file.
225 */
226struct fio_file {
227 /*
228 * A file may not be a file descriptor, let the io engine decide
229 */
230 union {
231 unsigned long file_data;
232 int fd;
233 };
234 char *file_name;
235 void *mmap;
236 unsigned long long file_size;
237 unsigned long long real_file_size;
238 unsigned long long file_offset;
239 unsigned long long last_pos;
240 unsigned long long last_completed_pos;
241
242 /*
243 * block map for random io
244 */
245 unsigned long *file_map;
246 unsigned int num_maps;
247 unsigned int last_free_lookup;
248
249 unsigned int unlink;
250};
251
252struct thread_stat {
253 struct io_log *slat_log;
254 struct io_log *clat_log;
255 struct io_log *bw_log;
256
257 /*
258 * bandwidth and latency stats
259 */
260 struct io_stat clat_stat[2]; /* completion latency */
261 struct io_stat slat_stat[2]; /* submission latency */
262 struct io_stat bw_stat[2]; /* bandwidth stats */
263
264 unsigned long long stat_io_bytes[2];
265 struct timeval stat_sample_time[2];
266
267 /*
268 * fio system usage accounting
269 */
270 struct rusage ru_start;
271 struct rusage ru_end;
272 unsigned long usr_time;
273 unsigned long sys_time;
274 unsigned long ctx;
275};
276
277/*
278 * How many depth levels to log
279 */
280#define FIO_IO_U_MAP_NR 8
281#define FIO_IO_U_LAT_NR 12
282
283/*
284 * This describes a single thread/process executing a fio job.
285 */
286struct thread_data {
287 int pad;
288 char *description;
289 char *name;
290 char *directory;
291 char *filename;
292 char verror[80];
293 pthread_t thread;
294 int thread_number;
295 int groupid;
296 struct thread_stat ts;
297 enum fio_filetype filetype;
298 struct fio_file *files;
299 unsigned int nr_files;
300 unsigned int nr_uniq_files;
301 union {
302 unsigned int next_file;
303 os_random_state_t next_file_state;
304 };
305 int error;
306 pid_t pid;
307 char *orig_buffer;
308 size_t orig_buffer_size;
309 volatile int terminate;
310 volatile int runstate;
311 enum td_ddir td_ddir;
312 unsigned int ioprio;
313 unsigned int last_was_sync;
314
315 unsigned int odirect;
316 unsigned int invalidate_cache;
317 unsigned int create_serialize;
318 unsigned int create_fsync;
319 unsigned int end_fsync;
320 unsigned int sync_io;
321 unsigned int verify;
322 unsigned int use_thread;
323 unsigned int unlink;
324 unsigned int do_disk_util;
325 unsigned int override_sync;
326 unsigned int rand_repeatable;
327 unsigned int write_lat_log;
328 unsigned int write_bw_log;
329 unsigned int norandommap;
330 unsigned int bs_unaligned;
331
332 unsigned int bs[2];
333 unsigned int min_bs[2];
334 unsigned int max_bs[2];
335 unsigned int hugepage_size;
336 unsigned int rw_min_bs;
337 unsigned int thinktime;
338 unsigned int thinktime_spin;
339 unsigned int thinktime_blocks;
340 unsigned int fsync_blocks;
341 unsigned int start_delay;
342 unsigned long timeout;
343 unsigned int overwrite;
344 unsigned int bw_avg_time;
345 unsigned int loops;
346 unsigned long long zone_size;
347 unsigned long long zone_skip;
348 enum fio_memtype mem_type;
349 char *mmapfile;
350 int mmapfd;
351 unsigned int stonewall;
352 unsigned int numjobs;
353 unsigned int iodepth;
354 unsigned int iodepth_low;
355 unsigned int iodepth_batch;
356 os_cpu_mask_t cpumask;
357 unsigned int iolog;
358 unsigned int read_iolog;
359 unsigned int rwmixcycle;
360 unsigned int rwmixread;
361 unsigned int rwmixwrite;
362 unsigned int nice;
363 unsigned int file_service_type;
364
365 char *read_iolog_file;
366 char *write_iolog_file;
367 void *iolog_buf;
368 FILE *iolog_f;
369
370 char *sysfs_root;
371 char *ioscheduler;
372
373 os_random_state_t bsrange_state;
374 os_random_state_t verify_state;
375
376 int shm_id;
377
378 /*
379 * IO engine hooks, contains everything needed to submit an io_u
380 * to any of the available IO engines.
381 */
382 struct ioengine_ops *io_ops;
383
384 /*
385 * Current IO depth and list of free and busy io_u's.
386 */
387 unsigned int cur_depth;
388 unsigned int io_u_map[FIO_IO_U_MAP_NR];
389 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
390 unsigned long total_io_u;
391 struct list_head io_u_freelist;
392 struct list_head io_u_busylist;
393 struct list_head io_u_requeues;
394 unsigned int io_u_queued;
395
396 /*
397 * Rate state
398 */
399 unsigned int rate;
400 unsigned int ratemin;
401 unsigned int ratecycle;
402 unsigned long rate_usec_cycle;
403 long rate_pending_usleep;
404 unsigned long rate_bytes;
405 struct timeval lastrate;
406
407 unsigned long runtime[2]; /* msec */
408 unsigned long long io_size;
409 unsigned long long total_file_size;
410 unsigned long long start_offset;
411 unsigned long long total_io_size;
412
413 unsigned long io_issues[2];
414 unsigned long long io_blocks[2];
415 unsigned long long io_bytes[2];
416 unsigned long long this_io_bytes[2];
417 unsigned long long zone_bytes;
418 volatile int mutex;
419
420 /*
421 * State for random io, a bitmap of blocks done vs not done
422 */
423 os_random_state_t random_state;
424
425 /*
426 * CPU "io" cycle burner
427 */
428 unsigned int cpuload;
429 unsigned int cpucycle;
430
431 struct timeval start; /* start of this loop */
432 struct timeval epoch; /* time job was started */
433 struct timeval end_time;/* time job ended */
434
435 /*
436 * read/write mixed workload state
437 */
438 os_random_state_t rwmix_state;
439 struct timeval rwmix_switch;
440 enum fio_ddir rwmix_ddir;
441
442 /*
443 * Pre-run and post-run shell
444 */
445 char *exec_prerun;
446 char *exec_postrun;
447
448 /*
449 * IO historic logs
450 */
451 struct list_head io_hist_list;
452 struct list_head io_log_list;
453
454 /*
455 * timeout handling
456 */
457 struct timeval timeout_end;
458 struct itimerval timer;
459};
460
461/*
462 * roundrobin available files, or choose one at random.
463 */
464enum {
465 FIO_FSERVICE_RANDOM = 1,
466 FIO_FSERVICE_RR = 2,
467};
468
469/*
470 * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
471 * the timer on each queue operation.
472 */
473#define IO_U_TIMEOUT_INC 5
474#define IO_U_TIMEOUT 30
475
476#define __td_verror(td, err, msg, func) \
477 do { \
478 if ((td)->error) \
479 break; \
480 int e = (err); \
481 (td)->error = e; \
482 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
483 } while (0)
484
485
486#define td_verror(td, err, func) \
487 __td_verror((td), (err), strerror((err)), (func))
488#define td_vmsg(td, err, msg, func) \
489 __td_verror((td), (err), (msg), (func))
490
491extern int exitall_on_terminate;
492extern int thread_number;
493extern int shm_id;
494extern int groupid;
495extern int terse_output;
496extern FILE *f_out;
497extern FILE *f_err;
498extern int temp_stall_ts;
499extern unsigned long long mlock_size;
500
501extern struct thread_data *threads;
502
503#define td_read(td) ((td)->td_ddir & TD_DDIR_READ)
504#define td_write(td) ((td)->td_ddir & TD_DDIR_WRITE)
505#define td_rw(td) (((td)->td_ddir & TD_DDIR_RW) == TD_DDIR_RW)
506#define td_random(td) ((td)->td_ddir & TD_DDIR_RAND)
507
508#define BLOCKS_PER_MAP (8 * sizeof(long))
509#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
510#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
511#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
512
513#define MAX_JOBS (1024)
514
515static inline int should_fsync(struct thread_data *td)
516{
517 if (td->last_was_sync)
518 return 0;
519 if (td->odirect)
520 return 0;
521 if (td_write(td) || td_rw(td) || td->override_sync)
522 return 1;
523
524 return 0;
525}
526
527/*
528 * Disk utils as read in /sys/block/<dev>/stat
529 */
530struct disk_util_stat {
531 unsigned ios[2];
532 unsigned merges[2];
533 unsigned long long sectors[2];
534 unsigned ticks[2];
535 unsigned io_ticks;
536 unsigned time_in_queue;
537};
538
539/*
540 * Per-device disk util management
541 */
542struct disk_util {
543 struct list_head list;
544
545 char *name;
546 char path[256];
547 dev_t dev;
548
549 struct disk_util_stat dus;
550 struct disk_util_stat last_dus;
551
552 unsigned long msec;
553 struct timeval time;
554};
555
556#define DISK_UTIL_MSEC (250)
557
558#ifndef min
559#define min(a, b) ((a) < (b) ? (a) : (b))
560#endif
561#ifndef max
562#define max(a, b) ((a) > (b) ? (a) : (b))
563#endif
564
565/*
566 * Log exports
567 */
568extern int __must_check read_iolog_get(struct thread_data *, struct io_u *);
569extern void write_iolog_put(struct thread_data *, struct io_u *);
570extern int __must_check init_iolog(struct thread_data *td);
571extern void log_io_piece(struct thread_data *, struct io_u *);
572extern void prune_io_piece_log(struct thread_data *);
573extern void write_iolog_close(struct thread_data *);
574
575/*
576 * Logging
577 */
578extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
579extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
580extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
581extern void show_run_stats(void);
582extern void init_disk_util(struct thread_data *);
583extern void update_rusage_stat(struct thread_data *);
584extern void update_io_ticks(void);
585extern void disk_util_timer_arm(void);
586extern void setup_log(struct io_log **);
587extern void finish_log(struct thread_data *, struct io_log *, const char *);
588extern void __finish_log(struct io_log *, const char *);
589extern int setup_rate(struct thread_data *);
590extern struct io_log *agg_io_log[2];
591extern int write_bw_log;
592extern void add_agg_sample(unsigned long, enum fio_ddir);
593
594/*
595 * Time functions
596 */
597extern unsigned long utime_since(struct timeval *, struct timeval *);
598extern unsigned long utime_since_now(struct timeval *);
599extern unsigned long mtime_since(struct timeval *, struct timeval *);
600extern unsigned long mtime_since_now(struct timeval *);
601extern unsigned long time_since_now(struct timeval *);
602extern unsigned long mtime_since_genesis(void);
603extern void __usec_sleep(unsigned int);
604extern void usec_sleep(struct thread_data *, unsigned long);
605extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
606extern void fill_start_time(struct timeval *);
607extern void fio_gettime(struct timeval *, void *);
608extern void set_genesis_time(void);
609
610/*
611 * Init functions
612 */
613extern int __must_check parse_options(int, char **);
614extern int __must_check init_random_state(struct thread_data *);
615
616/*
617 * File setup/shutdown
618 */
619extern void close_files(struct thread_data *);
620extern int __must_check setup_files(struct thread_data *);
621extern int __must_check open_files(struct thread_data *);
622extern int __must_check file_invalidate_cache(struct thread_data *, struct fio_file *);
623
624/*
625 * ETA/status stuff
626 */
627extern void print_thread_status(void);
628extern void print_status_init(int);
629
630/*
631 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
632 * will never back again. It may cycle between running/verififying/fsyncing.
633 * Once the thread reaches TD_EXITED, it is just waiting for the core to
634 * reap it.
635 */
636enum {
637 TD_NOT_CREATED = 0,
638 TD_CREATED,
639 TD_INITIALIZED,
640 TD_RUNNING,
641 TD_VERIFYING,
642 TD_FSYNCING,
643 TD_EXITED,
644 TD_REAPED,
645};
646
647/*
648 * Verify helpers
649 */
650extern void populate_verify_io_u(struct thread_data *, struct io_u *);
651extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
652extern int __must_check verify_io_u(struct io_u *);
653
654/*
655 * Memory helpers
656 */
657extern int __must_check fio_pin_memory(void);
658extern void fio_unpin_memory(void);
659extern int __must_check allocate_io_mem(struct thread_data *);
660extern void free_io_mem(struct thread_data *);
661
662/*
663 * io unit handling
664 */
665#define queue_full(td) list_empty(&(td)->io_u_freelist)
666extern struct io_u *__get_io_u(struct thread_data *);
667extern struct io_u *get_io_u(struct thread_data *);
668extern void put_io_u(struct thread_data *, struct io_u *);
669extern void requeue_io_u(struct thread_data *, struct io_u **);
670extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *);
671extern long __must_check io_u_queued_complete(struct thread_data *, int);
672extern void io_u_queued(struct thread_data *, struct io_u *);
673extern void io_u_init_timeout(void);
674extern void io_u_set_timeout(struct thread_data *);
675
676/*
677 * io engine entry points
678 */
679extern int __must_check td_io_init(struct thread_data *);
680extern int __must_check td_io_prep(struct thread_data *, struct io_u *);
681extern int __must_check td_io_queue(struct thread_data *, struct io_u *);
682extern int __must_check td_io_sync(struct thread_data *, struct fio_file *);
683extern int __must_check td_io_getevents(struct thread_data *, int, int, struct timespec *);
684extern int __must_check td_io_commit(struct thread_data *);
685
686/*
687 * This is a pretty crappy semaphore implementation, but with the use that fio
688 * has (just signalling start/go conditions), it doesn't have to be better.
689 * Naturally this would not work for any type of contended semaphore or
690 * for real locking.
691 */
692static inline void fio_sem_init(volatile int *sem, int val)
693{
694 *sem = val;
695}
696
697static inline void fio_sem_down(volatile int *sem)
698{
699 while (*sem == 0)
700 usleep(10000);
701
702 (*sem)--;
703}
704
705static inline void fio_sem_up(volatile int *sem)
706{
707 (*sem)++;
708}
709
710/*
711 * If logging output to a file, stderr should go to both stderr and f_err
712 */
713#define log_err(args...) do { \
714 fprintf(f_err, ##args); \
715 if (f_err != stderr) \
716 fprintf(stderr, ##args); \
717 } while (0)
718
719FILE *get_f_out(void);
720FILE *get_f_err(void);
721
722struct ioengine_ops {
723 struct list_head list;
724 char name[16];
725 int version;
726 int flags;
727 int (*setup)(struct thread_data *);
728 int (*init)(struct thread_data *);
729 int (*prep)(struct thread_data *, struct io_u *);
730 int (*queue)(struct thread_data *, struct io_u *);
731 int (*commit)(struct thread_data *);
732 int (*getevents)(struct thread_data *, int, int, struct timespec *);
733 struct io_u *(*event)(struct thread_data *, int);
734 int (*cancel)(struct thread_data *, struct io_u *);
735 void (*cleanup)(struct thread_data *);
736 void *data;
737 void *dlhandle;
738 unsigned long priv;
739};
740
741#define FIO_IOOPS_VERSION 5
742
743extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
744extern void register_ioengine(struct ioengine_ops *);
745extern void unregister_ioengine(struct ioengine_ops *);
746extern void close_ioengine(struct thread_data *);
747
748/*
749 * Mark unused variables passed to ops functions as unused, to silence gcc
750 */
751#define fio_unused __attribute((__unused__))
752#define fio_init __attribute__((constructor))
753#define fio_exit __attribute__((destructor))
754
755#define for_each_td(td, i) \
756 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
757#define for_each_file(td, f, i) \
758 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
759
760#define fio_assert(td, cond) do { \
761 if (!(cond)) { \
762 int *__foo = NULL; \
763 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
764 (td)->runstate = TD_EXITED; \
765 (td)->error = EFAULT; \
766 *__foo = 0; \
767 } \
768} while (0)
769
770#endif
fio - Flexible IO Tester