8 #include <sys/resource.h>
24 #ifdef FIO_HAVE_SYSLET
39 TD_DDIR_READ = 1 << 0,
40 TD_DDIR_WRITE = 1 << 1,
41 TD_DDIR_RAND = 1 << 2,
42 TD_DDIR_RW = TD_DDIR_READ | TD_DDIR_WRITE,
43 TD_DDIR_RANDREAD = TD_DDIR_READ | TD_DDIR_RAND,
44 TD_DDIR_RANDWRITE = TD_DDIR_WRITE | TD_DDIR_RAND,
45 TD_DDIR_RANDRW = TD_DDIR_RW | TD_DDIR_RAND,
49 * Use for maintaining statistics
52 unsigned long max_val;
53 unsigned long min_val;
54 unsigned long samples;
61 * A single data sample
70 * Dynamically growing data sample log
73 unsigned long nr_samples;
74 unsigned long max_samples;
75 struct io_sample *log;
79 * When logging io actions, this matches a single sent io_u
83 struct rb_node rb_node;
84 struct list_head list;
86 struct fio_file *file;
87 unsigned long long offset;
92 #ifdef FIO_HAVE_SYSLET
94 struct syslet_uatom atom; /* the atom to submit */
95 struct syslet_uatom *head; /* head of the sequence */
96 long ret; /* syscall return value */
101 IO_U_F_FREE = 1 << 0,
102 IO_U_F_FLIGHT = 1 << 1,
110 #ifdef FIO_HAVE_LIBAIO
113 #ifdef FIO_HAVE_POSIXAIO
117 struct sg_io_hdr hdr;
119 #ifdef FIO_HAVE_SYSLET
120 struct syslet_req req;
122 #ifdef FIO_HAVE_GUASI
126 struct timeval start_time;
127 struct timeval issue_time;
130 * Allocated/set buffer and length
133 unsigned long buflen;
134 unsigned long long offset;
137 * IO engine state, may be different from above when we get
138 * partial transfers / residual data counts
141 unsigned long xfer_buflen;
149 * io engine private data
158 struct fio_file *file;
160 struct list_head list;
163 * Callback for io completion
165 int (*end_io)(struct io_u *);
169 * io_ops->queue() return values
172 FIO_Q_COMPLETED = 0, /* completed sync */
173 FIO_Q_QUEUED = 1, /* queued, will complete async */
174 FIO_Q_BUSY = 2, /* no more room, call ->commit() */
177 #define FIO_HDR_MAGIC 0xf00baaef
180 VERIFY_NONE = 0, /* no verification */
181 VERIFY_MD5, /* md5 sum data blocks */
182 VERIFY_CRC32, /* crc32 sum data blocks */
186 * A header structure associated with each checksummed data block
188 struct verify_header {
189 unsigned int fio_magic;
191 unsigned int verify_type;
193 char md5_digest[MD5_HASH_WORDS * 4];
198 struct group_run_stats {
199 unsigned long long max_run[2], min_run[2];
200 unsigned long long max_bw[2], min_bw[2];
201 unsigned long long io_kb[2];
202 unsigned long long agg[2];
206 * What type of allocation to use for io buffers
209 MEM_MALLOC = 0, /* ordinary malloc */
210 MEM_SHM, /* use shared memory segments */
211 MEM_SHMHUGE, /* use shared memory segments with huge pages */
212 MEM_MMAP, /* use anonynomous mmap */
213 MEM_MMAPHUGE, /* memory mapped huge file */
217 * The type of object we are working on
220 FIO_TYPE_FILE = 1, /* plain file */
221 FIO_TYPE_BD, /* block device */
222 FIO_TYPE_CHAR, /* character device */
225 enum fio_ioengine_flags {
226 FIO_SYNCIO = 1 << 0, /* io engine has synchronous ->queue */
227 FIO_RAWIO = 1 << 1, /* some sort of direct/raw io */
228 FIO_DISKLESSIO = 1 << 2, /* no disk involved */
229 FIO_NOEXTEND = 1 << 3, /* engine can't extend file */
230 FIO_NODISKUTIL = 1 << 4, /* diskutil can't handle filename */
233 enum fio_file_flags {
234 FIO_FILE_OPEN = 1 << 0, /* file is open */
235 FIO_FILE_UNLINK = 1 << 1, /* unlink on close */
236 FIO_FILE_CLOSING = 1 << 2, /* file being closed */
237 FIO_FILE_EXISTS = 1 << 3, /* no need to create */
241 * Each thread_data structure has a number of files associated with it,
242 * this structure holds state information for a single file.
245 enum fio_filetype filetype;
248 * A file may not be a file descriptor, let the io engine decide
251 unsigned long file_data;
256 unsigned long long file_size;
257 unsigned long long real_file_size;
258 unsigned long long file_offset;
259 unsigned long long last_pos;
260 unsigned long long last_completed_pos;
263 * block map for random io
265 unsigned long *file_map;
266 unsigned int num_maps;
267 unsigned int last_free_lookup;
270 enum fio_file_flags flags;
274 * How many depth levels to log
276 #define FIO_IO_U_MAP_NR 8
277 #define FIO_IO_U_LAT_NR 12
288 struct io_log *slat_log;
289 struct io_log *clat_log;
290 struct io_log *bw_log;
293 * bandwidth and latency stats
295 struct io_stat clat_stat[2]; /* completion latency */
296 struct io_stat slat_stat[2]; /* submission latency */
297 struct io_stat bw_stat[2]; /* bandwidth stats */
299 unsigned long long stat_io_bytes[2];
300 struct timeval stat_sample_time[2];
303 * fio system usage accounting
305 struct rusage ru_start;
306 struct rusage ru_end;
307 unsigned long usr_time;
308 unsigned long sys_time;
312 * IO depth and latency stats
314 unsigned int io_u_map[FIO_IO_U_MAP_NR];
315 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
316 unsigned long total_io_u[2];
318 unsigned long long io_bytes[2];
319 unsigned long runtime[2];
320 unsigned long total_run_time;
323 struct thread_options {
331 enum td_ddir td_ddir;
332 unsigned int ddir_nr;
333 unsigned int iodepth;
334 unsigned int iodepth_low;
335 unsigned int iodepth_batch;
337 unsigned long long size;
338 unsigned long long file_size_low;
339 unsigned long long file_size_high;
340 unsigned long long start_offset;
343 unsigned int min_bs[2];
344 unsigned int max_bs[2];
346 unsigned int nr_files;
347 unsigned int open_files;
349 unsigned int odirect;
350 unsigned int invalidate_cache;
351 unsigned int create_serialize;
352 unsigned int create_fsync;
353 unsigned int end_fsync;
354 unsigned int sync_io;
356 unsigned int use_thread;
358 unsigned int do_disk_util;
359 unsigned int override_sync;
360 unsigned int rand_repeatable;
361 unsigned int write_lat_log;
362 unsigned int write_bw_log;
363 unsigned int norandommap;
364 unsigned int bs_unaligned;
365 unsigned int fsync_on_close;
367 unsigned int hugepage_size;
368 unsigned int rw_min_bs;
369 unsigned int thinktime;
370 unsigned int thinktime_spin;
371 unsigned int thinktime_blocks;
372 unsigned int fsync_blocks;
373 unsigned int start_delay;
374 unsigned long timeout;
375 unsigned int overwrite;
376 unsigned int bw_avg_time;
378 unsigned long long zone_size;
379 unsigned long long zone_skip;
380 enum fio_memtype mem_type;
382 unsigned int stonewall;
383 unsigned int new_group;
384 unsigned int numjobs;
385 os_cpu_mask_t cpumask;
387 unsigned int read_iolog;
388 unsigned int rwmixcycle;
389 unsigned int rwmix[2];
391 unsigned int file_service_type;
392 unsigned int group_reporting;
393 unsigned int fadvise_hint;
395 char *read_iolog_file;
396 char *write_iolog_file;
399 * Pre-run and post-run shell
405 unsigned int ratemin;
406 unsigned int ratecycle;
407 unsigned int rate_iops;
408 unsigned int rate_iops_min;
413 * CPU "io" cycle burner
415 unsigned int cpuload;
416 unsigned int cpucycle;
419 #define FIO_VERROR_SIZE 128
422 * This describes a single thread/process executing a fio job.
425 struct thread_options o;
426 char verror[FIO_VERROR_SIZE];
430 struct thread_stat ts;
431 struct fio_file *files;
432 unsigned int files_index;
433 unsigned int nr_open_files;
434 unsigned int nr_normal_files;
436 unsigned int next_file;
437 os_random_state_t next_file_state;
442 size_t orig_buffer_size;
443 volatile int terminate;
444 volatile int runstate;
446 unsigned int last_was_sync;
456 os_random_state_t bsrange_state;
457 os_random_state_t verify_state;
462 * IO engine hooks, contains everything needed to submit an io_u
463 * to any of the available IO engines.
465 struct ioengine_ops *io_ops;
468 * Current IO depth and list of free and busy io_u's.
470 unsigned int cur_depth;
471 struct list_head io_u_freelist;
472 struct list_head io_u_busylist;
473 struct list_head io_u_requeues;
474 unsigned int io_u_queued;
479 unsigned long rate_usec_cycle;
480 long rate_pending_usleep;
481 unsigned long rate_bytes;
482 unsigned long rate_blocks;
483 struct timeval lastrate;
485 unsigned long long io_size;
486 unsigned long long total_io_size;
488 unsigned long io_issues[2];
489 unsigned long long io_blocks[2];
490 unsigned long long io_bytes[2];
491 unsigned long long this_io_bytes[2];
492 unsigned long long zone_bytes;
493 struct fio_sem *mutex;
496 * State for random io, a bitmap of blocks done vs not done
498 os_random_state_t random_state;
500 struct timeval start; /* start of this loop */
501 struct timeval epoch; /* time job was started */
502 struct timeval rw_end[2];
503 unsigned int rw_end_set[2];
506 * read/write mixed workload state
508 os_random_state_t rwmix_state;
509 unsigned long long rwmix_bytes;
510 struct timeval rwmix_switch;
511 enum fio_ddir rwmix_ddir;
512 unsigned int ddir_nr;
517 struct rb_root io_hist_tree;
518 struct list_head io_log_list;
523 struct timeval timeout_end;
524 struct itimerval timer;
527 * for fileservice, how often to switch to a new file
529 unsigned int file_service_nr;
530 unsigned int file_service_left;
531 struct fio_file *file_service_file;
534 * For generating file sizes
536 os_random_state_t file_size_state;
540 * roundrobin available files, or choose one at random.
543 FIO_FSERVICE_RANDOM = 1,
548 * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
549 * the timer on each queue operation.
551 #define IO_U_TIMEOUT_INC 5
552 #define IO_U_TIMEOUT 30
554 #define __td_verror(td, err, msg, func) \
560 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
564 #define td_verror(td, err, func) \
565 __td_verror((td), (err), strerror((err)), (func))
566 #define td_vmsg(td, err, msg, func) \
567 __td_verror((td), (err), (msg), (func))
569 extern int exitall_on_terminate;
570 extern int thread_number;
571 extern int nr_process, nr_thread;
574 extern int terse_output;
577 extern int temp_stall_ts;
578 extern unsigned long long mlock_size;
579 extern unsigned long page_mask, page_size;
581 extern struct thread_data *threads;
583 #define td_read(td) ((td)->o.td_ddir & TD_DDIR_READ)
584 #define td_write(td) ((td)->o.td_ddir & TD_DDIR_WRITE)
585 #define td_rw(td) (((td)->o.td_ddir & TD_DDIR_RW) == TD_DDIR_RW)
586 #define td_random(td) ((td)->o.td_ddir & TD_DDIR_RAND)
588 #define BLOCKS_PER_MAP (8 * sizeof(long))
589 #define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->o.rw_min_bs))
590 #define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
591 #define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
593 #define MAX_JOBS (1024)
595 static inline int should_fsync(struct thread_data *td)
597 if (td->last_was_sync)
601 if (td_write(td) || td_rw(td) || td->o.override_sync)
608 * Disk utils as read in /sys/block/<dev>/stat
610 struct disk_util_stat {
613 unsigned long long sectors[2];
616 unsigned time_in_queue;
620 * Per-device disk util management
623 struct list_head list;
629 struct disk_util_stat dus;
630 struct disk_util_stat last_dus;
636 #define DISK_UTIL_MSEC (250)
639 #define min(a, b) ((a) < (b) ? (a) : (b))
642 #define max(a, b) ((a) > (b) ? (a) : (b))
648 extern int __must_check read_iolog_get(struct thread_data *, struct io_u *);
649 extern void write_iolog_put(struct thread_data *, struct io_u *);
650 extern int __must_check init_iolog(struct thread_data *td);
651 extern void log_io_piece(struct thread_data *, struct io_u *);
652 extern void prune_io_piece_log(struct thread_data *);
653 extern void write_iolog_close(struct thread_data *);
658 extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
659 extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
660 extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
661 extern void show_run_stats(void);
662 extern void init_disk_util(struct thread_data *);
663 extern void update_rusage_stat(struct thread_data *);
664 extern void update_io_ticks(void);
665 extern void disk_util_timer_arm(void);
666 extern void setup_log(struct io_log **);
667 extern void finish_log(struct thread_data *, struct io_log *, const char *);
668 extern void __finish_log(struct io_log *, const char *);
669 extern struct io_log *agg_io_log[2];
670 extern int write_bw_log;
671 extern void add_agg_sample(unsigned long, enum fio_ddir);
676 extern unsigned long utime_since(struct timeval *, struct timeval *);
677 extern unsigned long utime_since_now(struct timeval *);
678 extern unsigned long mtime_since(struct timeval *, struct timeval *);
679 extern unsigned long mtime_since_now(struct timeval *);
680 extern unsigned long time_since_now(struct timeval *);
681 extern unsigned long mtime_since_genesis(void);
682 extern void __usec_sleep(unsigned int);
683 extern void usec_sleep(struct thread_data *, unsigned long);
684 extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
685 extern void fill_start_time(struct timeval *);
686 extern void fio_gettime(struct timeval *, void *);
687 extern void set_genesis_time(void);
690 * Init/option functions
692 extern int __must_check parse_options(int, char **);
693 extern int fio_option_parse(struct thread_data *, const char *);
694 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
695 extern void fio_fill_default_options(struct thread_data *);
696 extern int fio_show_option_help(const char *);
697 extern void fio_options_dup_and_init(struct option *);
698 extern void options_mem_dupe(struct thread_data *);
699 extern void options_mem_free(struct thread_data *);
700 #define FIO_GETOPT_JOB 0x89988998
701 #define FIO_NR_OPTIONS 128
704 * File setup/shutdown
706 extern void close_files(struct thread_data *);
707 extern int __must_check setup_files(struct thread_data *);
708 extern int __must_check open_files(struct thread_data *);
709 extern int __must_check file_invalidate_cache(struct thread_data *, struct fio_file *);
710 extern int __must_check generic_open_file(struct thread_data *, struct fio_file *);
711 extern void generic_close_file(struct thread_data *, struct fio_file *);
712 extern void add_file(struct thread_data *, const char *);
713 extern void get_file(struct fio_file *);
714 extern void put_file(struct thread_data *, struct fio_file *);
715 extern int add_dir_files(struct thread_data *, const char *);
716 extern int init_random_map(struct thread_data *);
717 extern void dup_files(struct thread_data *, struct thread_data *);
722 extern void print_thread_status(void);
723 extern void print_status_init(int);
726 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
727 * will never back again. It may cycle between running/verififying/fsyncing.
728 * Once the thread reaches TD_EXITED, it is just waiting for the core to
745 extern void populate_verify_io_u(struct thread_data *, struct io_u *);
746 extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
747 extern int __must_check verify_io_u(struct io_u *);
752 extern int __must_check fio_pin_memory(void);
753 extern void fio_unpin_memory(void);
754 extern int __must_check allocate_io_mem(struct thread_data *);
755 extern void free_io_mem(struct thread_data *);
760 #define queue_full(td) list_empty(&(td)->io_u_freelist)
761 extern struct io_u *__get_io_u(struct thread_data *);
762 extern struct io_u *get_io_u(struct thread_data *);
763 extern void put_io_u(struct thread_data *, struct io_u *);
764 extern void requeue_io_u(struct thread_data *, struct io_u **);
765 extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *);
766 extern long __must_check io_u_queued_complete(struct thread_data *, int);
767 extern void io_u_queued(struct thread_data *, struct io_u *);
768 extern void io_u_log_error(struct thread_data *, struct io_u *);
769 extern void io_u_init_timeout(void);
770 extern void io_u_set_timeout(struct thread_data *);
771 extern void io_u_mark_depth(struct thread_data *, struct io_u *);
774 * io engine entry points
776 extern int __must_check td_io_init(struct thread_data *);
777 extern int __must_check td_io_prep(struct thread_data *, struct io_u *);
778 extern int __must_check td_io_queue(struct thread_data *, struct io_u *);
779 extern int __must_check td_io_sync(struct thread_data *, struct fio_file *);
780 extern int __must_check td_io_getevents(struct thread_data *, int, int, struct timespec *);
781 extern int __must_check td_io_commit(struct thread_data *);
782 extern int __must_check td_io_open_file(struct thread_data *, struct fio_file *);
783 extern void td_io_close_file(struct thread_data *, struct fio_file *);
786 * If logging output to a file, stderr should go to both stderr and f_err
788 #define log_err(args...) do { \
789 fprintf(f_err, ##args); \
790 if (f_err != stderr) \
791 fprintf(stderr, ##args); \
794 #define log_info(args...) fprintf(f_out, ##args)
796 FILE *get_f_out(void);
797 FILE *get_f_err(void);
799 struct ioengine_ops {
800 struct list_head list;
804 int (*setup)(struct thread_data *);
805 int (*init)(struct thread_data *);
806 int (*prep)(struct thread_data *, struct io_u *);
807 int (*queue)(struct thread_data *, struct io_u *);
808 int (*commit)(struct thread_data *);
809 int (*getevents)(struct thread_data *, int, int, struct timespec *);
810 struct io_u *(*event)(struct thread_data *, int);
811 int (*cancel)(struct thread_data *, struct io_u *);
812 void (*cleanup)(struct thread_data *);
813 int (*open_file)(struct thread_data *, struct fio_file *);
814 void (*close_file)(struct thread_data *, struct fio_file *);
819 #define FIO_IOOPS_VERSION 6
821 extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
822 extern void register_ioengine(struct ioengine_ops *);
823 extern void unregister_ioengine(struct ioengine_ops *);
824 extern void close_ioengine(struct thread_data *);
827 * Mark unused variables passed to ops functions as unused, to silence gcc
829 #define fio_unused __attribute((__unused__))
830 #define fio_init __attribute__((constructor))
831 #define fio_exit __attribute__((destructor))
833 #define for_each_td(td, i) \
834 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
835 #define for_each_file(td, f, i) \
836 for ((i) = 0, (f) = &(td)->files[0]; (i) < (td)->o.nr_files; (i)++, (f)++)
838 #define fio_assert(td, cond) do { \
841 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
842 (td)->runstate = TD_EXITED; \
843 (td)->error = EFAULT; \