8 #include <sys/resource.h>
23 #ifdef FIO_HAVE_SYSLET
34 TD_DDIR_READ = 1 << 0,
35 TD_DDIR_WRITE = 1 << 1,
36 TD_DDIR_RAND = 1 << 2,
37 TD_DDIR_RW = TD_DDIR_READ | TD_DDIR_WRITE,
38 TD_DDIR_RANDREAD = TD_DDIR_READ | TD_DDIR_RAND,
39 TD_DDIR_RANDWRITE = TD_DDIR_WRITE | TD_DDIR_RAND,
40 TD_DDIR_RANDRW = TD_DDIR_RW | TD_DDIR_RAND,
44 * Use for maintaining statistics
47 unsigned long max_val;
48 unsigned long min_val;
49 unsigned long samples;
56 * A single data sample
65 * Dynamically growing data sample log
68 unsigned long nr_samples;
69 unsigned long max_samples;
70 struct io_sample *log;
74 * When logging io actions, this matches a single sent io_u
77 struct list_head list;
78 struct fio_file *file;
79 unsigned long long offset;
84 #ifdef FIO_HAVE_SYSLET
86 struct syslet_uatom atom; /* the atom to submit */
87 struct syslet_uatom *head; /* head of the sequence */
88 long ret; /* syscall return value */
94 IO_U_F_FLIGHT = 1 << 1,
102 #ifdef FIO_HAVE_LIBAIO
105 #ifdef FIO_HAVE_POSIXAIO
109 struct sg_io_hdr hdr;
111 #ifdef FIO_HAVE_SYSLET
112 struct syslet_req req;
115 struct timeval start_time;
116 struct timeval issue_time;
119 * Allocated/set buffer and length
122 unsigned long buflen;
123 unsigned long long offset;
126 * IO engine state, may be different from above when we get
127 * partial transfers / residual data counts
130 unsigned long xfer_buflen;
138 * io engine private data
147 struct fio_file *file;
149 struct list_head list;
152 * Callback for io completion
154 int (*end_io)(struct io_u *);
158 * io_ops->queue() return values
161 FIO_Q_COMPLETED = 0, /* completed sync */
162 FIO_Q_QUEUED = 1, /* queued, will complete async */
163 FIO_Q_BUSY = 2, /* no more room, call ->commit() */
166 #define FIO_HDR_MAGIC 0xf00baaef
169 VERIFY_NONE = 0, /* no verification */
170 VERIFY_MD5, /* md5 sum data blocks */
171 VERIFY_CRC32, /* crc32 sum data blocks */
175 * A header structure associated with each checksummed data block
177 struct verify_header {
178 unsigned int fio_magic;
180 unsigned int verify_type;
182 char md5_digest[MD5_HASH_WORDS * 4];
187 struct group_run_stats {
188 unsigned long long max_run[2], min_run[2];
189 unsigned long long max_bw[2], min_bw[2];
190 unsigned long long io_kb[2];
191 unsigned long long agg[2];
195 * What type of allocation to use for io buffers
198 MEM_MALLOC = 0, /* ordinary malloc */
199 MEM_SHM, /* use shared memory segments */
200 MEM_SHMHUGE, /* use shared memory segments with huge pages */
201 MEM_MMAP, /* use anonynomous mmap */
202 MEM_MMAPHUGE, /* memory mapped huge file */
206 * The type of object we are working on
209 FIO_TYPE_FILE = 1, /* plain file */
210 FIO_TYPE_BD, /* block device */
211 FIO_TYPE_CHAR, /* character device */
214 enum fio_ioengine_flags {
215 FIO_SYNCIO = 1 << 0, /* io engine has synchronous ->queue */
216 FIO_RAWIO = 1 << 1, /* some sort of direct/raw io */
217 FIO_DISKLESSIO = 1 << 2, /* no disk involved */
218 FIO_NOEXTEND = 1 << 3, /* engine can't extend file */
219 FIO_NODISKUTIL = 1 << 4, /* diskutil can't handle filename */
222 enum fio_file_flags {
223 FIO_FILE_OPEN = 1 << 0, /* file is open */
224 FIO_FILE_UNLINK = 1 << 1, /* unlink on close */
225 FIO_FILE_CLOSING = 1 << 2, /* file being closed */
226 FIO_FILE_EXISTS = 1 << 3, /* no need to create */
230 * Each thread_data structure has a number of files associated with it,
231 * this structure holds state information for a single file.
234 enum fio_filetype filetype;
237 * A file may not be a file descriptor, let the io engine decide
240 unsigned long file_data;
245 unsigned long long file_size;
246 unsigned long long real_file_size;
247 unsigned long long file_offset;
248 unsigned long long last_pos;
249 unsigned long long last_completed_pos;
252 * block map for random io
254 unsigned long *file_map;
255 unsigned int num_maps;
256 unsigned int last_free_lookup;
259 enum fio_file_flags flags;
263 * How many depth levels to log
265 #define FIO_IO_U_MAP_NR 8
266 #define FIO_IO_U_LAT_NR 12
277 struct io_log *slat_log;
278 struct io_log *clat_log;
279 struct io_log *bw_log;
282 * bandwidth and latency stats
284 struct io_stat clat_stat[2]; /* completion latency */
285 struct io_stat slat_stat[2]; /* submission latency */
286 struct io_stat bw_stat[2]; /* bandwidth stats */
288 unsigned long long stat_io_bytes[2];
289 struct timeval stat_sample_time[2];
292 * fio system usage accounting
294 struct rusage ru_start;
295 struct rusage ru_end;
296 unsigned long usr_time;
297 unsigned long sys_time;
301 * IO depth and latency stats
303 unsigned int io_u_map[FIO_IO_U_MAP_NR];
304 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
305 unsigned long total_io_u[2];
307 unsigned long long io_bytes[2];
308 unsigned long runtime[2];
309 unsigned long total_run_time;
312 struct thread_options {
320 enum td_ddir td_ddir;
321 unsigned int iodepth;
322 unsigned int iodepth_low;
323 unsigned int iodepth_batch;
325 unsigned long long size;
326 unsigned long long file_size_low;
327 unsigned long long file_size_high;
328 unsigned long long start_offset;
331 unsigned int min_bs[2];
332 unsigned int max_bs[2];
334 unsigned int nr_files;
335 unsigned int open_files;
337 unsigned int odirect;
338 unsigned int invalidate_cache;
339 unsigned int create_serialize;
340 unsigned int create_fsync;
341 unsigned int end_fsync;
342 unsigned int sync_io;
344 unsigned int use_thread;
346 unsigned int do_disk_util;
347 unsigned int override_sync;
348 unsigned int rand_repeatable;
349 unsigned int write_lat_log;
350 unsigned int write_bw_log;
351 unsigned int norandommap;
352 unsigned int bs_unaligned;
353 unsigned int fsync_on_close;
355 unsigned int hugepage_size;
356 unsigned int rw_min_bs;
357 unsigned int thinktime;
358 unsigned int thinktime_spin;
359 unsigned int thinktime_blocks;
360 unsigned int fsync_blocks;
361 unsigned int start_delay;
362 unsigned long timeout;
363 unsigned int overwrite;
364 unsigned int bw_avg_time;
366 unsigned long long zone_size;
367 unsigned long long zone_skip;
368 enum fio_memtype mem_type;
370 unsigned int stonewall;
371 unsigned int numjobs;
372 os_cpu_mask_t cpumask;
374 unsigned int read_iolog;
375 unsigned int rwmixcycle;
376 unsigned int rwmixread;
377 unsigned int rwmixwrite;
379 unsigned int file_service_type;
380 unsigned int group_reporting;
382 char *read_iolog_file;
383 char *write_iolog_file;
386 * Pre-run and post-run shell
392 unsigned int ratemin;
393 unsigned int ratecycle;
394 unsigned int rate_iops;
395 unsigned int rate_iops_min;
400 * CPU "io" cycle burner
402 unsigned int cpuload;
403 unsigned int cpucycle;
407 * This describes a single thread/process executing a fio job.
410 struct thread_options o;
415 struct thread_stat ts;
416 struct fio_file *files;
417 unsigned int files_index;
418 unsigned int nr_open_files;
419 unsigned int nr_normal_files;
421 unsigned int next_file;
422 os_random_state_t next_file_state;
427 size_t orig_buffer_size;
428 volatile int terminate;
429 volatile int runstate;
431 unsigned int last_was_sync;
441 os_random_state_t bsrange_state;
442 os_random_state_t verify_state;
447 * IO engine hooks, contains everything needed to submit an io_u
448 * to any of the available IO engines.
450 struct ioengine_ops *io_ops;
453 * Current IO depth and list of free and busy io_u's.
455 unsigned int cur_depth;
456 struct list_head io_u_freelist;
457 struct list_head io_u_busylist;
458 struct list_head io_u_requeues;
459 unsigned int io_u_queued;
464 unsigned long rate_usec_cycle;
465 long rate_pending_usleep;
466 unsigned long rate_bytes;
467 unsigned long rate_blocks;
468 struct timeval lastrate;
470 unsigned long long io_size;
471 unsigned long long total_io_size;
473 unsigned long io_issues[2];
474 unsigned long long io_blocks[2];
475 unsigned long long io_bytes[2];
476 unsigned long long this_io_bytes[2];
477 unsigned long long zone_bytes;
478 struct fio_sem *mutex;
481 * State for random io, a bitmap of blocks done vs not done
483 os_random_state_t random_state;
485 struct timeval start; /* start of this loop */
486 struct timeval epoch; /* time job was started */
489 * read/write mixed workload state
491 os_random_state_t rwmix_state;
492 unsigned long long rwmix_bytes;
493 struct timeval rwmix_switch;
494 enum fio_ddir rwmix_ddir;
499 struct list_head io_hist_list;
500 struct list_head io_log_list;
505 struct timeval timeout_end;
506 struct itimerval timer;
509 * for fileservice, how often to switch to a new file
511 unsigned int file_service_nr;
512 unsigned int file_service_left;
513 struct fio_file *file_service_file;
516 * For generating file sizes
518 os_random_state_t file_size_state;
522 * roundrobin available files, or choose one at random.
525 FIO_FSERVICE_RANDOM = 1,
530 * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
531 * the timer on each queue operation.
533 #define IO_U_TIMEOUT_INC 5
534 #define IO_U_TIMEOUT 30
536 #define __td_verror(td, err, msg, func) \
542 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
546 #define td_verror(td, err, func) \
547 __td_verror((td), (err), strerror((err)), (func))
548 #define td_vmsg(td, err, msg, func) \
549 __td_verror((td), (err), (msg), (func))
551 extern int exitall_on_terminate;
552 extern int thread_number;
553 extern int nr_process, nr_thread;
556 extern int terse_output;
559 extern int temp_stall_ts;
560 extern unsigned long long mlock_size;
561 extern unsigned long page_mask, page_size;
563 extern struct thread_data *threads;
565 #define td_read(td) ((td)->o.td_ddir & TD_DDIR_READ)
566 #define td_write(td) ((td)->o.td_ddir & TD_DDIR_WRITE)
567 #define td_rw(td) (((td)->o.td_ddir & TD_DDIR_RW) == TD_DDIR_RW)
568 #define td_random(td) ((td)->o.td_ddir & TD_DDIR_RAND)
570 #define BLOCKS_PER_MAP (8 * sizeof(long))
571 #define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->o.rw_min_bs))
572 #define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
573 #define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
575 #define MAX_JOBS (1024)
577 static inline int should_fsync(struct thread_data *td)
579 if (td->last_was_sync)
583 if (td_write(td) || td_rw(td) || td->o.override_sync)
590 * Disk utils as read in /sys/block/<dev>/stat
592 struct disk_util_stat {
595 unsigned long long sectors[2];
598 unsigned time_in_queue;
602 * Per-device disk util management
605 struct list_head list;
611 struct disk_util_stat dus;
612 struct disk_util_stat last_dus;
618 #define DISK_UTIL_MSEC (250)
621 #define min(a, b) ((a) < (b) ? (a) : (b))
624 #define max(a, b) ((a) > (b) ? (a) : (b))
630 extern int __must_check read_iolog_get(struct thread_data *, struct io_u *);
631 extern void write_iolog_put(struct thread_data *, struct io_u *);
632 extern int __must_check init_iolog(struct thread_data *td);
633 extern void log_io_piece(struct thread_data *, struct io_u *);
634 extern void prune_io_piece_log(struct thread_data *);
635 extern void write_iolog_close(struct thread_data *);
640 extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
641 extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
642 extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
643 extern void show_run_stats(void);
644 extern void init_disk_util(struct thread_data *);
645 extern void update_rusage_stat(struct thread_data *);
646 extern void update_io_ticks(void);
647 extern void disk_util_timer_arm(void);
648 extern void setup_log(struct io_log **);
649 extern void finish_log(struct thread_data *, struct io_log *, const char *);
650 extern void __finish_log(struct io_log *, const char *);
651 extern struct io_log *agg_io_log[2];
652 extern int write_bw_log;
653 extern void add_agg_sample(unsigned long, enum fio_ddir);
658 extern unsigned long utime_since(struct timeval *, struct timeval *);
659 extern unsigned long utime_since_now(struct timeval *);
660 extern unsigned long mtime_since(struct timeval *, struct timeval *);
661 extern unsigned long mtime_since_now(struct timeval *);
662 extern unsigned long time_since_now(struct timeval *);
663 extern unsigned long mtime_since_genesis(void);
664 extern void __usec_sleep(unsigned int);
665 extern void usec_sleep(struct thread_data *, unsigned long);
666 extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
667 extern void fill_start_time(struct timeval *);
668 extern void fio_gettime(struct timeval *, void *);
669 extern void set_genesis_time(void);
672 * Init/option functions
674 extern int __must_check parse_options(int, char **);
675 extern int fio_option_parse(struct thread_data *, const char *);
676 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
677 extern void fio_fill_default_options(struct thread_data *);
678 extern int fio_show_option_help(const char *);
679 extern void fio_options_dup_and_init(struct option *);
680 #define FIO_GETOPT_JOB 0x89988998
681 #define FIO_NR_OPTIONS 128
684 * File setup/shutdown
686 extern void close_files(struct thread_data *);
687 extern int __must_check setup_files(struct thread_data *);
688 extern int __must_check open_files(struct thread_data *);
689 extern int __must_check file_invalidate_cache(struct thread_data *, struct fio_file *);
690 extern int __must_check generic_open_file(struct thread_data *, struct fio_file *);
691 extern void generic_close_file(struct thread_data *, struct fio_file *);
692 extern void add_file(struct thread_data *, const char *);
693 extern void get_file(struct fio_file *);
694 extern void put_file(struct thread_data *, struct fio_file *);
695 extern int add_dir_files(struct thread_data *, const char *);
696 extern int init_random_map(struct thread_data *);
701 extern void print_thread_status(void);
702 extern void print_status_init(int);
705 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
706 * will never back again. It may cycle between running/verififying/fsyncing.
707 * Once the thread reaches TD_EXITED, it is just waiting for the core to
724 extern void populate_verify_io_u(struct thread_data *, struct io_u *);
725 extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
726 extern int __must_check verify_io_u(struct io_u *);
731 extern int __must_check fio_pin_memory(void);
732 extern void fio_unpin_memory(void);
733 extern int __must_check allocate_io_mem(struct thread_data *);
734 extern void free_io_mem(struct thread_data *);
739 #define queue_full(td) list_empty(&(td)->io_u_freelist)
740 extern struct io_u *__get_io_u(struct thread_data *);
741 extern struct io_u *get_io_u(struct thread_data *);
742 extern void put_io_u(struct thread_data *, struct io_u *);
743 extern void requeue_io_u(struct thread_data *, struct io_u **);
744 extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *);
745 extern long __must_check io_u_queued_complete(struct thread_data *, int);
746 extern void io_u_queued(struct thread_data *, struct io_u *);
747 extern void io_u_log_error(struct thread_data *, struct io_u *);
748 extern void io_u_init_timeout(void);
749 extern void io_u_set_timeout(struct thread_data *);
750 extern void io_u_mark_depth(struct thread_data *, struct io_u *);
753 * io engine entry points
755 extern int __must_check td_io_init(struct thread_data *);
756 extern int __must_check td_io_prep(struct thread_data *, struct io_u *);
757 extern int __must_check td_io_queue(struct thread_data *, struct io_u *);
758 extern int __must_check td_io_sync(struct thread_data *, struct fio_file *);
759 extern int __must_check td_io_getevents(struct thread_data *, int, int, struct timespec *);
760 extern int __must_check td_io_commit(struct thread_data *);
761 extern int __must_check td_io_open_file(struct thread_data *, struct fio_file *);
762 extern void td_io_close_file(struct thread_data *, struct fio_file *);
765 * If logging output to a file, stderr should go to both stderr and f_err
767 #define log_err(args...) do { \
768 fprintf(f_err, ##args); \
769 if (f_err != stderr) \
770 fprintf(stderr, ##args); \
773 #define log_info(args...) fprintf(f_out, ##args)
775 FILE *get_f_out(void);
776 FILE *get_f_err(void);
778 struct ioengine_ops {
779 struct list_head list;
783 int (*setup)(struct thread_data *);
784 int (*init)(struct thread_data *);
785 int (*prep)(struct thread_data *, struct io_u *);
786 int (*queue)(struct thread_data *, struct io_u *);
787 int (*commit)(struct thread_data *);
788 int (*getevents)(struct thread_data *, int, int, struct timespec *);
789 struct io_u *(*event)(struct thread_data *, int);
790 int (*cancel)(struct thread_data *, struct io_u *);
791 void (*cleanup)(struct thread_data *);
792 int (*open_file)(struct thread_data *, struct fio_file *);
793 void (*close_file)(struct thread_data *, struct fio_file *);
798 #define FIO_IOOPS_VERSION 6
800 extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
801 extern void register_ioengine(struct ioengine_ops *);
802 extern void unregister_ioengine(struct ioengine_ops *);
803 extern void close_ioengine(struct thread_data *);
806 * Mark unused variables passed to ops functions as unused, to silence gcc
808 #define fio_unused __attribute((__unused__))
809 #define fio_init __attribute__((constructor))
810 #define fio_exit __attribute__((destructor))
812 #define for_each_td(td, i) \
813 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
814 #define for_each_file(td, f, i) \
815 for ((i) = 0, (f) = &(td)->files[0]; (i) < (td)->o.nr_files; (i)++, (f)++)
817 #define fio_assert(td, cond) do { \
820 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
821 (td)->runstate = TD_EXITED; \
822 (td)->error = EFAULT; \