8 #include <sys/resource.h>
21 #ifdef FIO_HAVE_SYSLET
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,
42 * Use for maintaining statistics
45 unsigned long max_val;
46 unsigned long min_val;
47 unsigned long samples;
54 * A single data sample
63 * Dynamically growing data sample log
66 unsigned long nr_samples;
67 unsigned long max_samples;
68 struct io_sample *log;
72 * When logging io actions, this matches a single sent io_u
75 struct list_head list;
76 struct fio_file *file;
77 unsigned long long offset;
82 #ifdef FIO_HAVE_SYSLET
84 struct syslet_uatom atom; /* the atom to submit */
85 struct syslet_uatom *head; /* head of the sequence */
86 long ret; /* syscall return value */
92 IO_U_F_FLIGHT = 1 << 1,
100 #ifdef FIO_HAVE_LIBAIO
103 #ifdef FIO_HAVE_POSIXAIO
107 struct sg_io_hdr hdr;
109 #ifdef FIO_HAVE_SYSLET
110 struct syslet_req req;
113 struct timeval start_time;
114 struct timeval issue_time;
117 * Allocated/set buffer and length
120 unsigned long buflen;
121 unsigned long long offset;
124 * IO engine state, may be different from above when we get
125 * partial transfers / residual data counts
128 unsigned long xfer_buflen;
136 * io engine private data
145 struct fio_file *file;
147 struct list_head list;
150 * Callback for io completion
152 int (*end_io)(struct io_u *);
156 * io_ops->queue() return values
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() */
164 #define FIO_HDR_MAGIC 0xf00baaef
167 VERIFY_NONE = 0, /* no verification */
168 VERIFY_MD5, /* md5 sum data blocks */
169 VERIFY_CRC32, /* crc32 sum data blocks */
173 * A header structure associated with each checksummed data block
175 struct verify_header {
176 unsigned int fio_magic;
178 unsigned int verify_type;
180 char md5_digest[MD5_HASH_WORDS * 4];
185 struct 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];
193 * What type of allocation to use for io buffers
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 */
204 * The type of object we are working on
207 FIO_TYPE_FILE = 1, /* plain file */
208 FIO_TYPE_BD, /* block device */
209 FIO_TYPE_CHAR, /* character device */
212 enum 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) */
223 * Each thread_data structure has a number of files associated with it,
224 * this structure holds state information for a single file.
228 * A file may not be a file descriptor, let the io engine decide
231 unsigned long file_data;
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;
243 * block map for random io
245 unsigned long *file_map;
246 unsigned int num_maps;
247 unsigned int last_free_lookup;
253 struct io_log *slat_log;
254 struct io_log *clat_log;
255 struct io_log *bw_log;
258 * bandwidth and latency stats
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 */
264 unsigned long long stat_io_bytes[2];
265 struct timeval stat_sample_time[2];
268 * fio system usage accounting
270 struct rusage ru_start;
271 struct rusage ru_end;
272 unsigned long usr_time;
273 unsigned long sys_time;
278 * How many depth levels to log
280 #define FIO_IO_U_MAP_NR 8
281 #define FIO_IO_U_LAT_NR 12
284 * This describes a single thread/process executing a fio job.
296 struct thread_stat ts;
297 enum fio_filetype filetype;
298 struct fio_file *files;
299 unsigned int nr_files;
300 unsigned int nr_open_files;
301 unsigned int nr_uniq_files;
303 unsigned int next_file;
304 os_random_state_t next_file_state;
309 size_t orig_buffer_size;
310 volatile int terminate;
311 volatile int runstate;
312 enum td_ddir td_ddir;
314 unsigned int last_was_sync;
316 unsigned int odirect;
317 unsigned int invalidate_cache;
318 unsigned int create_serialize;
319 unsigned int create_fsync;
320 unsigned int end_fsync;
321 unsigned int sync_io;
323 unsigned int use_thread;
325 unsigned int do_disk_util;
326 unsigned int override_sync;
327 unsigned int rand_repeatable;
328 unsigned int write_lat_log;
329 unsigned int write_bw_log;
330 unsigned int norandommap;
331 unsigned int bs_unaligned;
334 unsigned int min_bs[2];
335 unsigned int max_bs[2];
336 unsigned int hugepage_size;
337 unsigned int rw_min_bs;
338 unsigned int thinktime;
339 unsigned int thinktime_spin;
340 unsigned int thinktime_blocks;
341 unsigned int fsync_blocks;
342 unsigned int start_delay;
343 unsigned long timeout;
344 unsigned int overwrite;
345 unsigned int bw_avg_time;
347 unsigned long long zone_size;
348 unsigned long long zone_skip;
349 enum fio_memtype mem_type;
352 unsigned int stonewall;
353 unsigned int numjobs;
354 unsigned int iodepth;
355 unsigned int iodepth_low;
356 unsigned int iodepth_batch;
357 os_cpu_mask_t cpumask;
359 unsigned int read_iolog;
360 unsigned int rwmixcycle;
361 unsigned int rwmixread;
362 unsigned int rwmixwrite;
364 unsigned int file_service_type;
366 char *read_iolog_file;
367 char *write_iolog_file;
374 os_random_state_t bsrange_state;
375 os_random_state_t verify_state;
380 * IO engine hooks, contains everything needed to submit an io_u
381 * to any of the available IO engines.
383 struct ioengine_ops *io_ops;
386 * Current IO depth and list of free and busy io_u's.
388 unsigned int cur_depth;
389 unsigned int io_u_map[FIO_IO_U_MAP_NR];
390 unsigned int io_u_lat[FIO_IO_U_LAT_NR];
391 unsigned long total_io_u;
392 struct list_head io_u_freelist;
393 struct list_head io_u_busylist;
394 struct list_head io_u_requeues;
395 unsigned int io_u_queued;
401 unsigned int ratemin;
402 unsigned int ratecycle;
403 unsigned long rate_usec_cycle;
404 long rate_pending_usleep;
405 unsigned long rate_bytes;
406 struct timeval lastrate;
408 unsigned long runtime[2]; /* msec */
409 unsigned long long io_size;
410 unsigned long long total_file_size;
411 unsigned long long start_offset;
412 unsigned long long total_io_size;
414 unsigned long io_issues[2];
415 unsigned long long io_blocks[2];
416 unsigned long long io_bytes[2];
417 unsigned long long this_io_bytes[2];
418 unsigned long long zone_bytes;
422 * State for random io, a bitmap of blocks done vs not done
424 os_random_state_t random_state;
427 * CPU "io" cycle burner
429 unsigned int cpuload;
430 unsigned int cpucycle;
432 struct timeval start; /* start of this loop */
433 struct timeval epoch; /* time job was started */
434 struct timeval end_time;/* time job ended */
437 * read/write mixed workload state
439 os_random_state_t rwmix_state;
440 struct timeval rwmix_switch;
441 enum fio_ddir rwmix_ddir;
444 * Pre-run and post-run shell
452 struct list_head io_hist_list;
453 struct list_head io_log_list;
458 struct timeval timeout_end;
459 struct itimerval timer;
463 * roundrobin available files, or choose one at random.
466 FIO_FSERVICE_RANDOM = 1,
471 * 30 second per-io_u timeout, with 5 second intervals to avoid resetting
472 * the timer on each queue operation.
474 #define IO_U_TIMEOUT_INC 5
475 #define IO_U_TIMEOUT 30
477 #define __td_verror(td, err, msg, func) \
483 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
487 #define td_verror(td, err, func) \
488 __td_verror((td), (err), strerror((err)), (func))
489 #define td_vmsg(td, err, msg, func) \
490 __td_verror((td), (err), (msg), (func))
492 extern int exitall_on_terminate;
493 extern int thread_number;
496 extern int terse_output;
499 extern int temp_stall_ts;
500 extern unsigned long long mlock_size;
502 extern struct thread_data *threads;
504 #define td_read(td) ((td)->td_ddir & TD_DDIR_READ)
505 #define td_write(td) ((td)->td_ddir & TD_DDIR_WRITE)
506 #define td_rw(td) (((td)->td_ddir & TD_DDIR_RW) == TD_DDIR_RW)
507 #define td_random(td) ((td)->td_ddir & TD_DDIR_RAND)
509 #define BLOCKS_PER_MAP (8 * sizeof(long))
510 #define TO_MAP_BLOCK(td, f, b) ((b) - ((f)->file_offset / (td)->rw_min_bs))
511 #define RAND_MAP_IDX(td, f, b) (TO_MAP_BLOCK(td, f, b) / BLOCKS_PER_MAP)
512 #define RAND_MAP_BIT(td, f, b) (TO_MAP_BLOCK(td, f, b) & (BLOCKS_PER_MAP - 1))
514 #define MAX_JOBS (1024)
516 static inline int should_fsync(struct thread_data *td)
518 if (td->last_was_sync)
522 if (td_write(td) || td_rw(td) || td->override_sync)
529 * Disk utils as read in /sys/block/<dev>/stat
531 struct disk_util_stat {
534 unsigned long long sectors[2];
537 unsigned time_in_queue;
541 * Per-device disk util management
544 struct list_head list;
550 struct disk_util_stat dus;
551 struct disk_util_stat last_dus;
557 #define DISK_UTIL_MSEC (250)
560 #define min(a, b) ((a) < (b) ? (a) : (b))
563 #define max(a, b) ((a) > (b) ? (a) : (b))
569 extern int __must_check read_iolog_get(struct thread_data *, struct io_u *);
570 extern void write_iolog_put(struct thread_data *, struct io_u *);
571 extern int __must_check init_iolog(struct thread_data *td);
572 extern void log_io_piece(struct thread_data *, struct io_u *);
573 extern void prune_io_piece_log(struct thread_data *);
574 extern void write_iolog_close(struct thread_data *);
579 extern void add_clat_sample(struct thread_data *, enum fio_ddir, unsigned long);
580 extern void add_slat_sample(struct thread_data *, enum fio_ddir, unsigned long);
581 extern void add_bw_sample(struct thread_data *, enum fio_ddir, struct timeval *);
582 extern void show_run_stats(void);
583 extern void init_disk_util(struct thread_data *);
584 extern void update_rusage_stat(struct thread_data *);
585 extern void update_io_ticks(void);
586 extern void disk_util_timer_arm(void);
587 extern void setup_log(struct io_log **);
588 extern void finish_log(struct thread_data *, struct io_log *, const char *);
589 extern void __finish_log(struct io_log *, const char *);
590 extern int setup_rate(struct thread_data *);
591 extern struct io_log *agg_io_log[2];
592 extern int write_bw_log;
593 extern void add_agg_sample(unsigned long, enum fio_ddir);
598 extern unsigned long utime_since(struct timeval *, struct timeval *);
599 extern unsigned long utime_since_now(struct timeval *);
600 extern unsigned long mtime_since(struct timeval *, struct timeval *);
601 extern unsigned long mtime_since_now(struct timeval *);
602 extern unsigned long time_since_now(struct timeval *);
603 extern unsigned long mtime_since_genesis(void);
604 extern void __usec_sleep(unsigned int);
605 extern void usec_sleep(struct thread_data *, unsigned long);
606 extern void rate_throttle(struct thread_data *, unsigned long, unsigned int);
607 extern void fill_start_time(struct timeval *);
608 extern void fio_gettime(struct timeval *, void *);
609 extern void set_genesis_time(void);
614 extern int __must_check parse_options(int, char **);
615 extern int __must_check init_random_state(struct thread_data *);
618 * File setup/shutdown
620 extern void close_files(struct thread_data *);
621 extern int __must_check setup_files(struct thread_data *);
622 extern int __must_check open_files(struct thread_data *);
623 extern int open_file(struct thread_data *, struct fio_file *, int, int);
624 extern void close_file(struct thread_data *, struct fio_file *);
625 extern int __must_check file_invalidate_cache(struct thread_data *, struct fio_file *);
630 extern void print_thread_status(void);
631 extern void print_status_init(int);
634 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
635 * will never back again. It may cycle between running/verififying/fsyncing.
636 * Once the thread reaches TD_EXITED, it is just waiting for the core to
653 extern void populate_verify_io_u(struct thread_data *, struct io_u *);
654 extern int __must_check get_next_verify(struct thread_data *td, struct io_u *);
655 extern int __must_check verify_io_u(struct io_u *);
660 extern int __must_check fio_pin_memory(void);
661 extern void fio_unpin_memory(void);
662 extern int __must_check allocate_io_mem(struct thread_data *);
663 extern void free_io_mem(struct thread_data *);
668 #define queue_full(td) list_empty(&(td)->io_u_freelist)
669 extern struct io_u *__get_io_u(struct thread_data *);
670 extern struct io_u *get_io_u(struct thread_data *);
671 extern void put_io_u(struct thread_data *, struct io_u *);
672 extern void requeue_io_u(struct thread_data *, struct io_u **);
673 extern long __must_check io_u_sync_complete(struct thread_data *, struct io_u *);
674 extern long __must_check io_u_queued_complete(struct thread_data *, int);
675 extern void io_u_queued(struct thread_data *, struct io_u *);
676 extern void io_u_init_timeout(void);
677 extern void io_u_set_timeout(struct thread_data *);
680 * io engine entry points
682 extern int __must_check td_io_init(struct thread_data *);
683 extern int __must_check td_io_prep(struct thread_data *, struct io_u *);
684 extern int __must_check td_io_queue(struct thread_data *, struct io_u *);
685 extern int __must_check td_io_sync(struct thread_data *, struct fio_file *);
686 extern int __must_check td_io_getevents(struct thread_data *, int, int, struct timespec *);
687 extern int __must_check td_io_commit(struct thread_data *);
690 * This is a pretty crappy semaphore implementation, but with the use that fio
691 * has (just signalling start/go conditions), it doesn't have to be better.
692 * Naturally this would not work for any type of contended semaphore or
695 static inline void fio_sem_init(volatile int *sem, int val)
700 static inline void fio_sem_down(volatile int *sem)
708 static inline void fio_sem_up(volatile int *sem)
714 * If logging output to a file, stderr should go to both stderr and f_err
716 #define log_err(args...) do { \
717 fprintf(f_err, ##args); \
718 if (f_err != stderr) \
719 fprintf(stderr, ##args); \
722 FILE *get_f_out(void);
723 FILE *get_f_err(void);
725 struct ioengine_ops {
726 struct list_head list;
730 int (*setup)(struct thread_data *);
731 int (*init)(struct thread_data *);
732 int (*prep)(struct thread_data *, struct io_u *);
733 int (*queue)(struct thread_data *, struct io_u *);
734 int (*commit)(struct thread_data *);
735 int (*getevents)(struct thread_data *, int, int, struct timespec *);
736 struct io_u *(*event)(struct thread_data *, int);
737 int (*cancel)(struct thread_data *, struct io_u *);
738 void (*cleanup)(struct thread_data *);
744 #define FIO_IOOPS_VERSION 5
746 extern struct ioengine_ops *load_ioengine(struct thread_data *, const char *);
747 extern void register_ioengine(struct ioengine_ops *);
748 extern void unregister_ioengine(struct ioengine_ops *);
749 extern void close_ioengine(struct thread_data *);
752 * Mark unused variables passed to ops functions as unused, to silence gcc
754 #define fio_unused __attribute((__unused__))
755 #define fio_init __attribute__((constructor))
756 #define fio_exit __attribute__((destructor))
758 #define for_each_td(td, i) \
759 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
760 #define for_each_file(td, f, i) \
761 for ((i) = 0, (f) = &(td)->files[0]; (i) < (int) (td)->nr_files; (i)++, (f)++)
763 #define fio_assert(td, cond) do { \
766 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
767 (td)->runstate = TD_EXITED; \
768 (td)->error = EFAULT; \