Speedup init_disk_util() when dealing with many jobs
[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_NETIO = 1 << 4, /* networked io */
218 FIO_NULLIO = 1 << 5, /* no real data transfer (cpu/null) */
219};
220
221/*
222 * Each thread_data structure has a number of files associated with it,
223 * this structure holds state information for a single file.
224 */
225struct fio_file {
226 /*
227 * A file may not be a file descriptor, let the io engine decide
228 */
229 union {
230 unsigned long file_data;
231 int fd;
232 };
233 char *file_name;
234 void *mmap;
235 unsigned long long file_size;
236 unsigned long long real_file_size;
237 unsigned long long file_offset;
238 unsigned long long last_pos;
239 unsigned long long last_completed_pos;
240
241 /*
242 * block map for random io
243 */
244 unsigned long *file_map;
245 unsigned int num_maps;
246 unsigned int last_free_lookup;
247
248 unsigned int unlink;
249};
250
251struct thread_stat {
252 struct io_log *slat_log;
253 struct io_log *clat_log;
254 struct io_log *bw_log;
255
256 /*
257 * bandwidth and latency stats
258 */
259 struct io_stat clat_stat[2]; /* completion latency */
260 struct io_stat slat_stat[2]; /* submission latency */
261 struct io_stat bw_stat[2]; /* bandwidth stats */
262
263 unsigned long long stat_io_bytes[2];
264 struct timeval stat_sample_time[2];
265
266 /*
267 * fio system usage accounting
268 */
269 struct rusage ru_start;
270 struct rusage ru_end;
271 unsigned long usr_time;
272 unsigned long sys_time;
273 unsigned long ctx;
274};
275
276/*
277 * How many depth levels to log
278 */
279#define FIO_IO_U_MAP_NR 8
280#define FIO_IO_U_LAT_NR 12
281
282/*
283 * This describes a single thread/process executing a fio job.
284 */
285struct thread_data {
286 int pad;
287 char *description;
288 char *name;
289 char *directory;
290 char *filename;
291 char verror[80];
292 pthread_t thread;
293 int thread_number;
294 int groupid;
295 struct thread_stat ts;
296 enum fio_filetype filetype;
297 struct fio_file *files;
298 unsigned int nr_files;
299 unsigned int nr_uniq_files;
300 union {
301 unsigned int next_file;
302 os_random_state_t next_file_state;
303 };
304 int error;
305 pid_t pid;
306 char *orig_buffer;
307 size_t orig_buffer_size;
308 volatile int terminate;
309 volatile int runstate;
310 enum td_ddir td_ddir;
311 unsigned int ioprio;
312 unsigned int last_was_sync;
313
314 unsigned int odirect;
315 unsigned int invalidate_cache;
316 unsigned int create_serialize;
317 unsigned int create_fsync;
318 unsigned int end_fsync;
319 unsigned int sync_io;
320 unsigned int verify;
321 unsigned int use_thread;
322 unsigned int unlink;
323 unsigned int do_disk_util;
324 unsigned int override_sync;
325 unsigned int rand_repeatable;
326 unsigned int write_lat_log;
327 unsigned int write_bw_log;
328 unsigned int norandommap;
329 unsigned int bs_unaligned;
330
331 unsigned int bs[2];
332 unsigned int min_bs[2];
333 unsigned int max_bs[2];
334 unsigned int hugepage_size;
335 unsigned int rw_min_bs;
336 unsigned int thinktime;
337 unsigned int thinktime_spin;
338 unsigned int thinktime_blocks;
339 unsigned int fsync_blocks;
340 unsigned int start_delay;
341 unsigned long timeout;
342 unsigned int overwrite;
343 unsigned int bw_avg_time;
344 unsigned int loops;
345 unsigned long long zone_size;
346 unsigned long long zone_skip;
347 enum fio_memtype mem_type;
348 char *mmapfile;
349 int mmapfd;
350 unsigned int stonewall;
351 unsigned int numjobs;
352 unsigned int iodepth;
353 unsigned int iodepth_low;
354 os_cpu_mask_t cpumask;
355 unsigned int iolog;
356 unsigned int read_iolog;
357 unsigned int rwmixcycle;
358 unsigned int rwmixread;
359 unsigned int rwmixwrite;
360 unsigned int nice;
361 unsigned int file_service_type;
362
363 char *read_iolog_file;
364 char *write_iolog_file;
365 void *iolog_buf;
366 FILE *iolog_f;
367
368 char *sysfs_root;
369 char *ioscheduler;
370
371 os_random_state_t bsrange_state;
372 os_random_state_t verify_state;
373
374 int shm_id;
375
376 /*
377 * IO engine hooks, contains everything needed to submit an io_u
378 * to any of the available IO engines.
379 */
380 struct ioengine_ops *io_ops;
381
382 /*
383 * Current IO depth and list of free and busy io_u's.
384 */
385 unsigned int cur_depth;
386 unsigned int io_u_map[FIO_IO_U_MAP_NR];
387 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
388 unsigned long total_io_u;
389 struct list_head io_u_freelist;
390 struct list_head io_u_busylist;
391 struct list_head io_u_requeues;
392
393 /*
394 * Rate state
395 */
396 unsigned int rate;
397 unsigned int ratemin;
398 unsigned int ratecycle;
399 unsigned long rate_usec_cycle;
400 long rate_pending_usleep;
401 unsigned long rate_bytes;
402 struct timeval lastrate;
403
404 unsigned long runtime[2]; /* msec */
405 unsigned long long io_size;
406 unsigned long long total_file_size;
407 unsigned long long start_offset;
408 unsigned long long total_io_size;
409
410 unsigned long io_issues[2];
411 unsigned long long io_blocks[2];
412 unsigned long long io_bytes[2];
413 unsigned long long this_io_bytes[2];
414 unsigned long long zone_bytes;
415 volatile int mutex;
416
417 /*
418 * State for random io, a bitmap of blocks done vs not done
419 */
420 os_random_state_t random_state;
421
422 /*
423 * CPU "io" cycle burner
424 */
425 unsigned int cpuload;
426 unsigned int cpucycle;
427
428 struct timeval start; /* start of this loop */
429 struct timeval epoch; /* time job was started */
430 struct timeval end_time;/* time job ended */
431
432 /*
433 * read/write mixed workload state
434 */
435 os_random_state_t rwmix_state;
436 struct timeval rwmix_switch;
437 enum fio_ddir rwmix_ddir;
438
439 /*
440 * Pre-run and post-run shell
441 */
442 char *exec_prerun;
443 char *exec_postrun;
444
445 /*
446 * IO historic logs
447 */
448 struct list_head io_hist_list;
449 struct list_head io_log_list;
450
451 /*
452 * timeout handling
453 */
454 struct timeval timeout_end;
455 struct itimerval timer;
456};
457
458/*
459 * roundrobin available files, or choose one at random.
460 */
461enum {
462 FIO_FSERVICE_RANDOM = 1,
463 FIO_FSERVICE_RR = 2,
464};
465
466/*
467 * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
468 * the timer on each queue operation.
469 */
470#define IO_U_TIMEOUT_INC 5
471#define IO_U_TIMEOUT 30
472
473#define __td_verror(td, err, msg, func) \
474 do { \
475 if ((td)->error) \
476 break; \
477 int e = (err); \
478 (td)->error = e; \
479 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
480 } while (0)
481
482
483#define td_verror(td, err, func) \
484 __td_verror((td), (err), strerror((err)), (func))
485#define td_vmsg(td, err, msg, func) \
486 __td_verror((td), (err), (msg), (func))
487
488extern int exitall_on_terminate;
489extern int thread_number;
490extern int shm_id;
491extern int groupid;
492extern int terse_output;
493extern FILE *f_out;
494extern FILE *f_err;
495extern int temp_stall_ts;
496extern unsigned long long mlock_size;
497
498extern struct thread_data *threads;
499
500#define td_read(td) ((td)->td_ddir & TD_DDIR_READ)
501#define td_write(td) ((td)->td_ddir & TD_DDIR_WRITE)
502#define td_rw(td) (((td)->td_ddir & TD_DDIR_RW) == TD_DDIR_RW)
503#define td_random(td) ((td)->td_ddir & TD_DDIR_RAND)
504
505#define BLOCKS_PER_MAP (8 * sizeof(long))
506#define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
507#define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
508#define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
509
510#define MAX_JOBS (1024)
511
512static inline int should_fsync(struct thread_data *td)
513{
514 if (td->last_was_sync)
515 return 0;
516 if (td->odirect)
517 return 0;
518 if (td_write(td) || td_rw(td) || td->override_sync)
519 return 1;
520
521 return 0;
522}
523
524/*
525 * Disk utils as read in /sys/block/<dev>/stat
526 */
527struct disk_util_stat {
528 unsigned ios[2];
529 unsigned merges[2];
530 unsigned long long sectors[2];
531 unsigned ticks[2];
532 unsigned io_ticks;
533 unsigned time_in_queue;
534};
535
536/*
537 * Per-device disk util management
538 */
539struct disk_util {
540 struct list_head list;
541
542 char *name;
543 char path[256];
544 dev_t dev;
545
546 struct disk_util_stat dus;
547 struct disk_util_stat last_dus;
548
549 unsigned long msec;
550 struct timeval time;
551};
552
553#define DISK_UTIL_MSEC (250)
554
555#ifndef min
556#define min(a, b) ((a) < (b) ? (a) : (b))
557#endif
558#ifndef max
559#define max(a, b) ((a) > (b) ? (a) : (b))
560#endif
561
562/*
563 * Log exports
564 */
565extern int __must_check read_iolog_get(struct thread_data *, struct io_u *);
566extern void write_iolog_put(struct thread_data *, struct io_u *);
567extern int __must_check init_iolog(struct thread_data *td);
568extern void log_io_piece(struct thread_data *, struct io_u *);
569extern void prune_io_piece_log(struct thread_data *);
570extern void write_iolog_close(struct thread_data *);
571
572/*
573 * Logging
574 */
575extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
576extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
577extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
578extern void show_run_stats(void);
579extern void init_disk_util(struct thread_data *);
580extern void update_rusage_stat(struct thread_data *);
581extern void update_io_ticks(void);
582extern void disk_util_timer_arm(void);
583extern void setup_log(struct io_log **);
584extern void finish_log(struct thread_data *, struct io_log *, const char *);
585extern void __finish_log(struct io_log *, const char *);
586extern int setup_rate(struct thread_data *);
587extern struct io_log *agg_io_log[2];
588extern int write_bw_log;
589extern void add_agg_sample(unsigned long, enum fio_ddir);
590
591/*
592 * Time functions
593 */
594extern unsigned long utime_since(struct timeval *, struct timeval *);
595extern unsigned long utime_since_now(struct timeval *);
596extern unsigned long mtime_since(struct timeval *, struct timeval *);
597extern unsigned long mtime_since_now(struct timeval *);
598extern unsigned long time_since_now(struct timeval *);
599extern unsigned long mtime_since_genesis(void);
600extern void __usec_sleep(unsigned int);
601extern void usec_sleep(struct thread_data *, unsigned long);
602extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
603extern void fill_start_time(struct timeval *);
604extern void fio_gettime(struct timeval *, void *);
605extern void set_genesis_time(void);
606
607/*
608 * Init functions
609 */
610extern int __must_check parse_options(int, char **);
611extern int __must_check init_random_state(struct thread_data *);
612
613/*
614 * File setup/shutdown
615 */
616extern void close_files(struct thread_data *);
617extern int __must_check setup_files(struct thread_data *);
618extern int __must_check open_files(struct thread_data *);
619extern int __must_check file_invalidate_cache(struct thread_data *, struct fio_file *);
620
621/*
622 * ETA/status stuff
623 */
624extern void print_thread_status(void);
625extern void print_status_init(int);
626
627/*
628 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
629 * will never back again. It may cycle between running/verififying/fsyncing.
630 * Once the thread reaches TD_EXITED, it is just waiting for the core to
631 * reap it.
632 */
633enum {
634 TD_NOT_CREATED = 0,
635 TD_CREATED,
636 TD_INITIALIZED,
637 TD_RUNNING,
638 TD_VERIFYING,
639 TD_FSYNCING,
640 TD_EXITED,
641 TD_REAPED,
642};
643
644/*
645 * Verify helpers
646 */
647extern void populate_verify_io_u(struct thread_data *, struct io_u *);
648extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
649extern int __must_check verify_io_u(struct io_u *);
650
651/*
652 * Memory helpers
653 */
654extern int __must_check fio_pin_memory(void);
655extern void fio_unpin_memory(void);
656extern int __must_check allocate_io_mem(struct thread_data *);
657extern void free_io_mem(struct thread_data *);
658
659/*
660 * io unit handling
661 */
662#define queue_full(td) list_empty(&(td)->io_u_freelist)
663extern struct io_u *__get_io_u(struct thread_data *);
664extern struct io_u *get_io_u(struct thread_data *);
665extern void put_io_u(struct thread_data *, struct io_u *);
666extern void requeue_io_u(struct thread_data *, struct io_u **);
667extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *);
668extern long __must_check io_u_queued_complete(struct thread_data *, int);
669extern void io_u_queued(struct thread_data *, struct io_u *);
670extern void io_u_init_timeout(void);
671extern void io_u_set_timeout(struct thread_data *);
672
673/*
674 * io engine entry points
675 */
676extern int __must_check td_io_init(struct thread_data *);
677extern int __must_check td_io_prep(struct thread_data *, struct io_u *);
678extern int __must_check td_io_queue(struct thread_data *, struct io_u *);
679extern int __must_check td_io_sync(struct thread_data *, struct fio_file *);
680extern int __must_check td_io_getevents(struct thread_data *, int, int, struct timespec *);
681extern int __must_check td_io_commit(struct thread_data *);
682
683/*
684 * This is a pretty crappy semaphore implementation, but with the use that fio
685 * has (just signalling start/go conditions), it doesn't have to be better.
686 * Naturally this would not work for any type of contended semaphore or
687 * for real locking.
688 */
689static inline void fio_sem_init(volatile int *sem, int val)
690{
691 *sem = val;
692}
693
694static inline void fio_sem_down(volatile int *sem)
695{
696 while (*sem == 0)
697 usleep(10000);
698
699 (*sem)--;
700}
701
702static inline void fio_sem_up(volatile int *sem)
703{
704 (*sem)++;
705}
706
707/*
708 * If logging output to a file, stderr should go to both stderr and f_err
709 */
710#define log_err(args...) do { \
711 fprintf(f_err, ##args); \
712 if (f_err != stderr) \
713 fprintf(stderr, ##args); \
714 } while (0)
715
716struct ioengine_ops {
717 struct list_head list;
718 char name[16];
719 int version;
720 int flags;
721 int (*setup)(struct thread_data *);
722 int (*init)(struct thread_data *);
723 int (*prep)(struct thread_data *, struct io_u *);
724 int (*queue)(struct thread_data *, struct io_u *);
725 int (*commit)(struct thread_data *);
726 int (*getevents)(struct thread_data *, int, int, struct timespec *);
727 struct io_u *(*event)(struct thread_data *, int);
728 int (*cancel)(struct thread_data *, struct io_u *);
729 void (*cleanup)(struct thread_data *);
730 void *data;
731 void *dlhandle;
732 unsigned long priv;
733};
734
735#define FIO_IOOPS_VERSION 5
736
737extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
738extern void register_ioengine(struct ioengine_ops *);
739extern void unregister_ioengine(struct ioengine_ops *);
740extern void close_ioengine(struct thread_data *);
741
742/*
743 * Mark unused variables passed to ops functions as unused, to silence gcc
744 */
745#define fio_unused __attribute((__unused__))
746#define fio_init __attribute__((constructor))
747#define fio_exit __attribute__((destructor))
748
749#define for_each_td(td, i) \
750 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
751#define for_each_file(td, f, i) \
752 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
753
754#define fio_assert(td, cond) do { \
755 if (!(cond)) { \
756 int *__foo = NULL; \
757 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
758 (td)->runstate = TD_EXITED; \
759 (td)->error = EFAULT; \
760 *__foo = 0; \
761 } \
762} while (0)
763
764#endif