8 #include <sys/resource.h>
23 #ifdef FIO_HAVE_SYSLET
38 TD_DDIR_READ = 1 << 0,
39 TD_DDIR_WRITE = 1 << 1,
40 TD_DDIR_RAND = 1 << 2,
41 TD_DDIR_RW = TD_DDIR_READ | TD_DDIR_WRITE,
42 TD_DDIR_RANDREAD = TD_DDIR_READ | TD_DDIR_RAND,
43 TD_DDIR_RANDWRITE = TD_DDIR_WRITE | TD_DDIR_RAND,
44 TD_DDIR_RANDRW = TD_DDIR_RW | TD_DDIR_RAND,
48 * Use for maintaining statistics
51 unsigned long max_val;
52 unsigned long min_val;
53 unsigned long samples;
60 * A single data sample
69 * Dynamically growing data sample log
72 unsigned long nr_samples;
73 unsigned long max_samples;
74 struct io_sample *log;
78 * When logging io actions, this matches a single sent io_u
81 struct list_head list;
82 struct fio_file *file;
83 unsigned long long offset;
88 #ifdef FIO_HAVE_SYSLET
90 struct syslet_uatom atom; /* the atom to submit */
91 struct syslet_uatom *head; /* head of the sequence */
92 long ret; /* syscall return value */
98 IO_U_F_FLIGHT = 1 << 1,
106 #ifdef FIO_HAVE_LIBAIO
109 #ifdef FIO_HAVE_POSIXAIO
113 struct sg_io_hdr hdr;
115 #ifdef FIO_HAVE_SYSLET
116 struct syslet_req req;
118 #ifdef FIO_HAVE_GUASI
122 struct timeval start_time;
123 struct timeval issue_time;
126 * Allocated/set buffer and length
129 unsigned long buflen;
130 unsigned long long offset;
133 * IO engine state, may be different from above when we get
134 * partial transfers / residual data counts
137 unsigned long xfer_buflen;
145 * io engine private data
154 struct fio_file *file;
156 struct list_head list;
159 * Callback for io completion
161 int (*end_io)(struct io_u *);
165 * io_ops->queue() return values
168 FIO_Q_COMPLETED = 0, /* completed sync */
169 FIO_Q_QUEUED = 1, /* queued, will complete async */
170 FIO_Q_BUSY = 2, /* no more room, call ->commit() */
173 #define FIO_HDR_MAGIC 0xf00baaef
176 VERIFY_NONE = 0, /* no verification */
177 VERIFY_MD5, /* md5 sum data blocks */
178 VERIFY_CRC32, /* crc32 sum data blocks */
182 * A header structure associated with each checksummed data block
184 struct verify_header {
185 unsigned int fio_magic;
187 unsigned int verify_type;
189 char md5_digest[MD5_HASH_WORDS * 4];
194 struct group_run_stats {
195 unsigned long long max_run[2], min_run[2];
196 unsigned long long max_bw[2], min_bw[2];
197 unsigned long long io_kb[2];
198 unsigned long long agg[2];
202 * What type of allocation to use for io buffers
205 MEM_MALLOC = 0, /* ordinary malloc */
206 MEM_SHM, /* use shared memory segments */
207 MEM_SHMHUGE, /* use shared memory segments with huge pages */
208 MEM_MMAP, /* use anonynomous mmap */
209 MEM_MMAPHUGE, /* memory mapped huge file */
213 * The type of object we are working on
216 FIO_TYPE_FILE = 1, /* plain file */
217 FIO_TYPE_BD, /* block device */
218 FIO_TYPE_CHAR, /* character device */
221 enum fio_ioengine_flags {
222 FIO_SYNCIO = 1 << 0, /* io engine has synchronous ->queue */
223 FIO_RAWIO = 1 << 1, /* some sort of direct/raw io */
224 FIO_DISKLESSIO = 1 << 2, /* no disk involved */
225 FIO_NOEXTEND = 1 << 3, /* engine can't extend file */
226 FIO_NODISKUTIL = 1 << 4, /* diskutil can't handle filename */
229 enum fio_file_flags {
230 FIO_FILE_OPEN = 1 << 0, /* file is open */
231 FIO_FILE_UNLINK = 1 << 1, /* unlink on close */
232 FIO_FILE_CLOSING = 1 << 2, /* file being closed */
233 FIO_FILE_EXISTS = 1 << 3, /* no need to create */
237 * Each thread_data structure has a number of files associated with it,
238 * this structure holds state information for a single file.
241 enum fio_filetype filetype;
244 * A file may not be a file descriptor, let the io engine decide
247 unsigned long file_data;
252 unsigned long long file_size;
253 unsigned long long real_file_size;
254 unsigned long long file_offset;
255 unsigned long long last_pos;
256 unsigned long long last_completed_pos;
259 * block map for random io
261 unsigned long *file_map;
262 unsigned int num_maps;
263 unsigned int last_free_lookup;
266 enum fio_file_flags flags;
270 * How many depth levels to log
272 #define FIO_IO_U_MAP_NR 8
273 #define FIO_IO_U_LAT_NR 12
284 struct io_log *slat_log;
285 struct io_log *clat_log;
286 struct io_log *bw_log;
289 * bandwidth and latency stats
291 struct io_stat clat_stat[2]; /* completion latency */
292 struct io_stat slat_stat[2]; /* submission latency */
293 struct io_stat bw_stat[2]; /* bandwidth stats */
295 unsigned long long stat_io_bytes[2];
296 struct timeval stat_sample_time[2];
299 * fio system usage accounting
301 struct rusage ru_start;
302 struct rusage ru_end;
303 unsigned long usr_time;
304 unsigned long sys_time;
308 * IO depth and latency stats
310 unsigned int io_u_map[FIO_IO_U_MAP_NR];
311 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
312 unsigned long total_io_u[2];
314 unsigned long long io_bytes[2];
315 unsigned long runtime[2];
316 unsigned long total_run_time;
319 struct thread_options {
327 enum td_ddir td_ddir;
328 unsigned int iodepth;
329 unsigned int iodepth_low;
330 unsigned int iodepth_batch;
332 unsigned long long size;
333 unsigned long long file_size_low;
334 unsigned long long file_size_high;
335 unsigned long long start_offset;
338 unsigned int min_bs[2];
339 unsigned int max_bs[2];
341 unsigned int nr_files;
342 unsigned int open_files;
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;
351 unsigned int use_thread;
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 unsigned int fsync_on_close;
362 unsigned int hugepage_size;
363 unsigned int rw_min_bs;
364 unsigned int thinktime;
365 unsigned int thinktime_spin;
366 unsigned int thinktime_blocks;
367 unsigned int fsync_blocks;
368 unsigned int start_delay;
369 unsigned long timeout;
370 unsigned int overwrite;
371 unsigned int bw_avg_time;
373 unsigned long long zone_size;
374 unsigned long long zone_skip;
375 enum fio_memtype mem_type;
377 unsigned int stonewall;
378 unsigned int new_group;
379 unsigned int numjobs;
380 os_cpu_mask_t cpumask;
382 unsigned int read_iolog;
383 unsigned int rwmixcycle;
384 unsigned int rwmix[2];
386 unsigned int file_service_type;
387 unsigned int group_reporting;
389 char *read_iolog_file;
390 char *write_iolog_file;
393 * Pre-run and post-run shell
399 unsigned int ratemin;
400 unsigned int ratecycle;
401 unsigned int rate_iops;
402 unsigned int rate_iops_min;
407 * CPU "io" cycle burner
409 unsigned int cpuload;
410 unsigned int cpucycle;
413 #define FIO_VERROR_SIZE 128
416 * This describes a single thread/process executing a fio job.
419 struct thread_options o;
420 char verror[FIO_VERROR_SIZE];
424 struct thread_stat ts;
425 struct fio_file *files;
426 unsigned int files_index;
427 unsigned int nr_open_files;
428 unsigned int nr_normal_files;
430 unsigned int next_file;
431 os_random_state_t next_file_state;
436 size_t orig_buffer_size;
437 volatile int terminate;
438 volatile int runstate;
440 unsigned int last_was_sync;
450 os_random_state_t bsrange_state;
451 os_random_state_t verify_state;
456 * IO engine hooks, contains everything needed to submit an io_u
457 * to any of the available IO engines.
459 struct ioengine_ops *io_ops;
462 * Current IO depth and list of free and busy io_u's.
464 unsigned int cur_depth;
465 struct list_head io_u_freelist;
466 struct list_head io_u_busylist;
467 struct list_head io_u_requeues;
468 unsigned int io_u_queued;
473 unsigned long rate_usec_cycle;
474 long rate_pending_usleep;
475 unsigned long rate_bytes;
476 unsigned long rate_blocks;
477 struct timeval lastrate;
479 unsigned long long io_size;
480 unsigned long long total_io_size;
482 unsigned long io_issues[2];
483 unsigned long long io_blocks[2];
484 unsigned long long io_bytes[2];
485 unsigned long long this_io_bytes[2];
486 unsigned long long zone_bytes;
487 struct fio_sem *mutex;
490 * State for random io, a bitmap of blocks done vs not done
492 os_random_state_t random_state;
494 struct timeval start; /* start of this loop */
495 struct timeval epoch; /* time job was started */
496 struct timeval rw_end[2];
497 unsigned int rw_end_set[2];
500 * read/write mixed workload state
502 os_random_state_t rwmix_state;
503 unsigned long long rwmix_bytes;
504 struct timeval rwmix_switch;
505 enum fio_ddir rwmix_ddir;
510 struct list_head io_hist_list;
511 struct list_head io_log_list;
516 struct timeval timeout_end;
517 struct itimerval timer;
520 * for fileservice, how often to switch to a new file
522 unsigned int file_service_nr;
523 unsigned int file_service_left;
524 struct fio_file *file_service_file;
527 * For generating file sizes
529 os_random_state_t file_size_state;
533 * roundrobin available files, or choose one at random.
536 FIO_FSERVICE_RANDOM = 1,
541 * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
542 * the timer on each queue operation.
544 #define IO_U_TIMEOUT_INC 5
545 #define IO_U_TIMEOUT 30
547 #define __td_verror(td, err, msg, func) \
553 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
557 #define td_verror(td, err, func) \
558 __td_verror((td), (err), strerror((err)), (func))
559 #define td_vmsg(td, err, msg, func) \
560 __td_verror((td), (err), (msg), (func))
562 extern int exitall_on_terminate;
563 extern int thread_number;
564 extern int nr_process, nr_thread;
567 extern int terse_output;
570 extern int temp_stall_ts;
571 extern unsigned long long mlock_size;
572 extern unsigned long page_mask, page_size;
574 extern struct thread_data *threads;
576 #define td_read(td) ((td)->o.td_ddir & TD_DDIR_READ)
577 #define td_write(td) ((td)->o.td_ddir & TD_DDIR_WRITE)
578 #define td_rw(td) (((td)->o.td_ddir & TD_DDIR_RW) == TD_DDIR_RW)
579 #define td_random(td) ((td)->o.td_ddir & TD_DDIR_RAND)
581 #define BLOCKS_PER_MAP (8 * sizeof(long))
582 #define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->o.rw_min_bs))
583 #define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
584 #define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
586 #define MAX_JOBS (1024)
588 static inline int should_fsync(struct thread_data *td)
590 if (td->last_was_sync)
594 if (td_write(td) || td_rw(td) || td->o.override_sync)
601 * Disk utils as read in /sys/block/<dev>/stat
603 struct disk_util_stat {
606 unsigned long long sectors[2];
609 unsigned time_in_queue;
613 * Per-device disk util management
616 struct list_head list;
622 struct disk_util_stat dus;
623 struct disk_util_stat last_dus;
629 #define DISK_UTIL_MSEC (250)
632 #define min(a, b) ((a) < (b) ? (a) : (b))
635 #define max(a, b) ((a) > (b) ? (a) : (b))
641 extern int __must_check read_iolog_get(struct thread_data *, struct io_u *);
642 extern void write_iolog_put(struct thread_data *, struct io_u *);
643 extern int __must_check init_iolog(struct thread_data *td);
644 extern void log_io_piece(struct thread_data *, struct io_u *);
645 extern void prune_io_piece_log(struct thread_data *);
646 extern void write_iolog_close(struct thread_data *);
651 extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
652 extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
653 extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
654 extern void show_run_stats(void);
655 extern void init_disk_util(struct thread_data *);
656 extern void update_rusage_stat(struct thread_data *);
657 extern void update_io_ticks(void);
658 extern void disk_util_timer_arm(void);
659 extern void setup_log(struct io_log **);
660 extern void finish_log(struct thread_data *, struct io_log *, const char *);
661 extern void __finish_log(struct io_log *, const char *);
662 extern struct io_log *agg_io_log[2];
663 extern int write_bw_log;
664 extern void add_agg_sample(unsigned long, enum fio_ddir);
669 extern unsigned long utime_since(struct timeval *, struct timeval *);
670 extern unsigned long utime_since_now(struct timeval *);
671 extern unsigned long mtime_since(struct timeval *, struct timeval *);
672 extern unsigned long mtime_since_now(struct timeval *);
673 extern unsigned long time_since_now(struct timeval *);
674 extern unsigned long mtime_since_genesis(void);
675 extern void __usec_sleep(unsigned int);
676 extern void usec_sleep(struct thread_data *, unsigned long);
677 extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
678 extern void fill_start_time(struct timeval *);
679 extern void fio_gettime(struct timeval *, void *);
680 extern void set_genesis_time(void);
683 * Init/option functions
685 extern int __must_check parse_options(int, char **);
686 extern int fio_option_parse(struct thread_data *, const char *);
687 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
688 extern void fio_fill_default_options(struct thread_data *);
689 extern int fio_show_option_help(const char *);
690 extern void fio_options_dup_and_init(struct option *);
691 #define FIO_GETOPT_JOB 0x89988998
692 #define FIO_NR_OPTIONS 128
695 * File setup/shutdown
697 extern void close_files(struct thread_data *);
698 extern int __must_check setup_files(struct thread_data *);
699 extern int __must_check open_files(struct thread_data *);
700 extern int __must_check file_invalidate_cache(struct thread_data *, struct fio_file *);
701 extern int __must_check generic_open_file(struct thread_data *, struct fio_file *);
702 extern void generic_close_file(struct thread_data *, struct fio_file *);
703 extern void add_file(struct thread_data *, const char *);
704 extern void get_file(struct fio_file *);
705 extern void put_file(struct thread_data *, struct fio_file *);
706 extern int add_dir_files(struct thread_data *, const char *);
707 extern int init_random_map(struct thread_data *);
712 extern void print_thread_status(void);
713 extern void print_status_init(int);
716 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
717 * will never back again. It may cycle between running/verififying/fsyncing.
718 * Once the thread reaches TD_EXITED, it is just waiting for the core to
735 extern void populate_verify_io_u(struct thread_data *, struct io_u *);
736 extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
737 extern int __must_check verify_io_u(struct io_u *);
742 extern int __must_check fio_pin_memory(void);
743 extern void fio_unpin_memory(void);
744 extern int __must_check allocate_io_mem(struct thread_data *);
745 extern void free_io_mem(struct thread_data *);
750 #define queue_full(td) list_empty(&(td)->io_u_freelist)
751 extern struct io_u *__get_io_u(struct thread_data *);
752 extern struct io_u *get_io_u(struct thread_data *);
753 extern void put_io_u(struct thread_data *, struct io_u *);
754 extern void requeue_io_u(struct thread_data *, struct io_u **);
755 extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *);
756 extern long __must_check io_u_queued_complete(struct thread_data *, int);
757 extern void io_u_queued(struct thread_data *, struct io_u *);
758 extern void io_u_log_error(struct thread_data *, struct io_u *);
759 extern void io_u_init_timeout(void);
760 extern void io_u_set_timeout(struct thread_data *);
761 extern void io_u_mark_depth(struct thread_data *, struct io_u *);
764 * io engine entry points
766 extern int __must_check td_io_init(struct thread_data *);
767 extern int __must_check td_io_prep(struct thread_data *, struct io_u *);
768 extern int __must_check td_io_queue(struct thread_data *, struct io_u *);
769 extern int __must_check td_io_sync(struct thread_data *, struct fio_file *);
770 extern int __must_check td_io_getevents(struct thread_data *, int, int, struct timespec *);
771 extern int __must_check td_io_commit(struct thread_data *);
772 extern int __must_check td_io_open_file(struct thread_data *, struct fio_file *);
773 extern void td_io_close_file(struct thread_data *, struct fio_file *);
776 * If logging output to a file, stderr should go to both stderr and f_err
778 #define log_err(args...) do { \
779 fprintf(f_err, ##args); \
780 if (f_err != stderr) \
781 fprintf(stderr, ##args); \
784 #define log_info(args...) fprintf(f_out, ##args)
786 FILE *get_f_out(void);
787 FILE *get_f_err(void);
789 struct ioengine_ops {
790 struct list_head list;
794 int (*setup)(struct thread_data *);
795 int (*init)(struct thread_data *);
796 int (*prep)(struct thread_data *, struct io_u *);
797 int (*queue)(struct thread_data *, struct io_u *);
798 int (*commit)(struct thread_data *);
799 int (*getevents)(struct thread_data *, int, int, struct timespec *);
800 struct io_u *(*event)(struct thread_data *, int);
801 int (*cancel)(struct thread_data *, struct io_u *);
802 void (*cleanup)(struct thread_data *);
803 int (*open_file)(struct thread_data *, struct fio_file *);
804 void (*close_file)(struct thread_data *, struct fio_file *);
809 #define FIO_IOOPS_VERSION 6
811 extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
812 extern void register_ioengine(struct ioengine_ops *);
813 extern void unregister_ioengine(struct ioengine_ops *);
814 extern void close_ioengine(struct thread_data *);
817 * Mark unused variables passed to ops functions as unused, to silence gcc
819 #define fio_unused __attribute((__unused__))
820 #define fio_init __attribute__((constructor))
821 #define fio_exit __attribute__((destructor))
823 #define for_each_td(td, i) \
824 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
825 #define for_each_file(td, f, i) \
826 for ((i) = 0, (f) = &(td)->files[0]; (i) < (td)->o.nr_files; (i)++, (f)++)
828 #define fio_assert(td, cond) do { \
831 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
832 (td)->runstate = TD_EXITED; \
833 (td)->error = EFAULT; \