8 #include <sys/resource.h>
19 #include "compiler/compiler.h"
23 #include "arch/arch.h"
36 #include "lib/getopt.h"
44 #ifdef FIO_HAVE_SOLARISAIO
45 #include <sys/asynch.h>
48 struct group_run_stats {
49 unsigned long long max_run[2], min_run[2];
50 unsigned long long max_bw[2], min_bw[2];
51 unsigned long long io_kb[2];
52 unsigned long long agg[2];
57 * What type of allocation to use for io buffers
60 MEM_MALLOC = 0, /* ordinary malloc */
61 MEM_SHM, /* use shared memory segments */
62 MEM_SHMHUGE, /* use shared memory segments with huge pages */
63 MEM_MMAP, /* use anonynomous mmap */
64 MEM_MMAPHUGE, /* memory mapped huge file */
68 * offset generator types
76 * How many depth levels to log
78 #define FIO_IO_U_MAP_NR 7
79 #define FIO_IO_U_LAT_U_NR 10
80 #define FIO_IO_U_LAT_M_NR 12
83 * Aggregate clat samples to report percentile(s) of them.
87 * FIO_IO_U_PLAT_BITS determines the maximum statistical error on the
88 * value of resulting percentiles. The error will be approximately
89 * 1/2^(FIO_IO_U_PLAT_BITS+1) of the value.
91 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the maximum
92 * range being tracked for latency samples. The maximum value tracked
93 * accurately will be 2^(GROUP_NR + PLAT_BITS -1) microseconds.
95 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the memory
96 * requirement of storing those aggregate counts. The memory used will
97 * be (FIO_IO_U_PLAT_GROUP_NR * 2^FIO_IO_U_PLAT_BITS) * sizeof(int)
100 * FIO_IO_U_PLAT_NR is the total number of buckets.
104 * Suppose the clat varies from 0 to 999 (usec), the straightforward
105 * method is to keep an array of (999 + 1) buckets, in which a counter
106 * keeps the count of samples which fall in the bucket, e.g.,
107 * {[0],[1],...,[999]}. However this consumes a huge amount of space,
108 * and can be avoided if an approximation is acceptable.
110 * One such method is to let the range of the bucket to be greater
111 * than one. This method has low accuracy when the value is small. For
112 * example, let the buckets be {[0,99],[100,199],...,[900,999]}, and
113 * the represented value of each bucket be the mean of the range. Then
114 * a value 0 has an round-off error of 49.5. To improve on this, we
115 * use buckets with non-uniform ranges, while bounding the error of
116 * each bucket within a ratio of the sample value. A simple example
117 * would be when error_bound = 0.005, buckets are {
118 * {[0],[1],...,[99]}, {[100,101],[102,103],...,[198,199]},..,
119 * {[900,909],[910,919]...} }. The total range is partitioned into
120 * groups with different ranges, then buckets with uniform ranges. An
121 * upper bound of the error is (range_of_bucket/2)/value_of_bucket
123 * For better efficiency, we implement this using base two. We group
124 * samples by their Most Significant Bit (MSB), extract the next M bit
125 * of them as an index within the group, and discard the rest of the
128 * E.g., assume a sample 'x' whose MSB is bit n (starting from bit 0),
129 * and use M bit for indexing
131 * | n | M bits | bit (n-M-1) ... bit 0 |
133 * Because x is at least 2^n, and bit 0 to bit (n-M-1) is at most
134 * (2^(n-M) - 1), discarding bit 0 to (n-M-1) makes the round-off
137 * 2^(n-M)-1 2^(n-M) 1
138 * e <= --------- <= ------- = ---
141 * Furthermore, we use "mean" of the range to represent the bucket,
142 * the error e can be lowered by half to 1 / 2^(M+1). By using M bits
143 * as the index, each group must contains 2^M buckets.
145 * E.g. Let M (FIO_IO_U_PLAT_BITS) be 6
146 * Error bound is 1/2^(6+1) = 0.0078125 (< 1%)
148 * Group MSB #discarded range of #buckets
150 * ----------------------------------------------------------------
155 * 4 9 3 [512,1023] 64
156 * ... ... ... [...,...] ...
157 * 18 23 17 [8838608,+inf]** 64
159 * * Special cases: when n < (M-1) or when n == (M-1), in both cases,
160 * the value cannot be rounded off. Use all bits of the sample as
163 * ** If a sample's MSB is greater than 23, it will be counted as 23.
166 #define FIO_IO_U_PLAT_BITS 6
167 #define FIO_IO_U_PLAT_VAL (1 << FIO_IO_U_PLAT_BITS)
168 #define FIO_IO_U_PLAT_GROUP_NR 19
169 #define FIO_IO_U_PLAT_NR (FIO_IO_U_PLAT_GROUP_NR * FIO_IO_U_PLAT_VAL)
170 #define FIO_IO_U_LIST_MAX_LEN 20 /* The size of the default and user-specified
171 list of percentiles */
173 #define MAX_PATTERN_SIZE 512
184 struct io_log *slat_log;
185 struct io_log *clat_log;
186 struct io_log *lat_log;
187 struct io_log *bw_log;
190 * bandwidth and latency stats
192 struct io_stat clat_stat[2]; /* completion latency */
193 struct io_stat slat_stat[2]; /* submission latency */
194 struct io_stat lat_stat[2]; /* total latency */
195 struct io_stat bw_stat[2]; /* bandwidth stats */
197 unsigned long long stat_io_bytes[2];
198 struct timeval stat_sample_time[2];
201 * fio system usage accounting
203 struct rusage ru_start;
204 struct rusage ru_end;
205 unsigned long usr_time;
206 unsigned long sys_time;
208 unsigned long minf, majf;
211 * IO depth and latency stats
213 unsigned int clat_percentiles;
214 double* percentile_list;
216 unsigned int io_u_map[FIO_IO_U_MAP_NR];
217 unsigned int io_u_submit[FIO_IO_U_MAP_NR];
218 unsigned int io_u_complete[FIO_IO_U_MAP_NR];
219 unsigned int io_u_lat_u[FIO_IO_U_LAT_U_NR];
220 unsigned int io_u_lat_m[FIO_IO_U_LAT_M_NR];
221 unsigned int io_u_plat[2][FIO_IO_U_PLAT_NR];
222 unsigned long total_io_u[3];
223 unsigned long short_io_u[3];
224 unsigned long total_submit;
225 unsigned long total_complete;
227 unsigned long long io_bytes[2];
228 unsigned long long runtime[2];
229 unsigned long total_run_time;
232 * IO Error related stats
234 unsigned continue_on_error;
235 unsigned long total_err_count;
238 unsigned int kb_base;
246 struct thread_options {
254 enum td_ddir td_ddir;
256 unsigned int kb_base;
257 unsigned int ddir_seq_nr;
259 unsigned int iodepth;
260 unsigned int iodepth_low;
261 unsigned int iodepth_batch;
262 unsigned int iodepth_batch_complete;
264 unsigned long long size;
265 unsigned int size_percent;
266 unsigned int fill_device;
267 unsigned long long file_size_low;
268 unsigned long long file_size_high;
269 unsigned long long start_offset;
273 unsigned int min_bs[2];
274 unsigned int max_bs[2];
275 struct bssplit *bssplit[2];
276 unsigned int bssplit_nr[2];
278 unsigned int nr_files;
279 unsigned int open_files;
280 enum file_lock_mode file_lock_mode;
281 unsigned int lockfile_batch;
283 unsigned int odirect;
284 unsigned int invalidate_cache;
285 unsigned int create_serialize;
286 unsigned int create_fsync;
287 unsigned int create_on_open;
288 unsigned int end_fsync;
289 unsigned int pre_read;
290 unsigned int sync_io;
292 unsigned int do_verify;
293 unsigned int verifysort;
294 unsigned int verify_interval;
295 unsigned int verify_offset;
296 char verify_pattern[MAX_PATTERN_SIZE];
297 unsigned int verify_pattern_bytes;
298 unsigned int verify_fatal;
299 unsigned int verify_dump;
300 unsigned int verify_async;
301 unsigned long long verify_backlog;
302 unsigned int verify_batch;
303 unsigned int use_thread;
305 unsigned int do_disk_util;
306 unsigned int override_sync;
307 unsigned int rand_repeatable;
308 unsigned int use_os_rand;
309 unsigned int write_lat_log;
310 unsigned int write_bw_log;
311 unsigned int norandommap;
312 unsigned int softrandommap;
313 unsigned int bs_unaligned;
314 unsigned int fsync_on_close;
316 unsigned int hugepage_size;
317 unsigned int rw_min_bs;
318 unsigned int thinktime;
319 unsigned int thinktime_spin;
320 unsigned int thinktime_blocks;
321 unsigned int fsync_blocks;
322 unsigned int fdatasync_blocks;
323 unsigned int barrier_blocks;
324 unsigned long long start_delay;
325 unsigned long long timeout;
326 unsigned long long ramp_time;
327 unsigned int overwrite;
328 unsigned int bw_avg_time;
330 unsigned long long zone_size;
331 unsigned long long zone_skip;
332 enum fio_memtype mem_type;
333 unsigned int mem_align;
335 unsigned int stonewall;
336 unsigned int new_group;
337 unsigned int numjobs;
338 os_cpu_mask_t cpumask;
339 unsigned int cpumask_set;
340 os_cpu_mask_t verify_cpumask;
341 unsigned int verify_cpumask_set;
343 unsigned int rwmixcycle;
344 unsigned int rwmix[2];
346 unsigned int file_service_type;
347 unsigned int group_reporting;
348 unsigned int fadvise_hint;
349 enum fio_fallocate_mode fallocate_mode;
350 unsigned int zero_buffers;
351 unsigned int refill_buffers;
352 unsigned int scramble_buffers;
353 unsigned int time_based;
354 unsigned int disable_lat;
355 unsigned int disable_clat;
356 unsigned int disable_slat;
357 unsigned int disable_bw;
358 unsigned int gtod_reduce;
359 unsigned int gtod_cpu;
360 unsigned int gtod_offload;
361 enum fio_cs clocksource;
362 unsigned int no_stall;
363 unsigned int trim_percentage;
364 unsigned int trim_batch;
365 unsigned int trim_zero;
366 unsigned long long trim_backlog;
367 unsigned int clat_percentiles;
368 unsigned int overwrite_plist;
369 double percentile_list[FIO_IO_U_LIST_MAX_LEN];
371 char *read_iolog_file;
372 char *write_iolog_file;
375 char *replay_redirect;
378 * Pre-run and post-run shell
383 unsigned int rate[2];
384 unsigned int ratemin[2];
385 unsigned int ratecycle;
386 unsigned int rate_iops[2];
387 unsigned int rate_iops_min[2];
392 * CPU "io" cycle burner
394 unsigned int cpuload;
395 unsigned int cpucycle;
400 unsigned int continue_on_error;
403 * Benchmark profile type
408 * blkio cgroup support
411 unsigned int cgroup_weight;
412 unsigned int cgroup_nodelete;
417 unsigned int sync_file_range;
419 unsigned int userspace_libaio_reap;
422 #define FIO_VERROR_SIZE 128
425 * This describes a single thread/process executing a fio job.
428 struct thread_options o;
429 char verror[FIO_VERROR_SIZE];
433 struct thread_stat ts;
434 struct fio_file **files;
435 unsigned int files_size;
436 unsigned int files_index;
437 unsigned int nr_open_files;
438 unsigned int nr_done_files;
439 unsigned int nr_normal_files;
441 unsigned int next_file;
442 os_random_state_t next_file_state;
443 struct frand_state __next_file_state;
449 size_t orig_buffer_size;
450 volatile int terminate;
451 volatile int runstate;
453 unsigned int ioprio_set;
454 unsigned int last_was_sync;
455 enum fio_ddir last_ddir;
465 unsigned long rand_seeds[8];
468 os_random_state_t bsrange_state;
469 struct frand_state __bsrange_state;
472 os_random_state_t verify_state;
473 struct frand_state __verify_state;
476 os_random_state_t trim_state;
477 struct frand_state __trim_state;
480 struct frand_state buf_state;
482 unsigned int verify_batch;
483 unsigned int trim_batch;
488 * IO engine hooks, contains everything needed to submit an io_u
489 * to any of the available IO engines.
491 struct ioengine_ops *io_ops;
494 * Current IO depth and list of free and busy io_u's.
496 unsigned int cur_depth;
497 unsigned int io_u_queued;
498 struct flist_head io_u_freelist;
499 struct flist_head io_u_busylist;
500 struct flist_head io_u_requeues;
501 pthread_mutex_t io_u_lock;
502 pthread_cond_t free_cond;
505 * async verify offload
507 struct flist_head verify_list;
508 pthread_t *verify_threads;
509 unsigned int nr_verify_threads;
510 pthread_cond_t verify_cond;
511 int verify_thread_exit;
516 unsigned long rate_nsec_cycle[2];
517 long rate_pending_usleep[2];
518 unsigned long rate_bytes[2];
519 unsigned long rate_blocks[2];
520 struct timeval lastrate[2];
522 unsigned long long total_io_size;
523 unsigned long long fill_device_size;
525 unsigned long io_issues[2];
526 unsigned long long io_blocks[2];
527 unsigned long long io_bytes[2];
528 unsigned long long io_skip_bytes;
529 unsigned long long this_io_bytes[2];
530 unsigned long long zone_bytes;
531 struct fio_mutex *mutex;
534 * State for random io, a bitmap of blocks done vs not done
537 os_random_state_t random_state;
538 struct frand_state __random_state;
541 struct timeval start; /* start of this loop */
542 struct timeval epoch; /* time job was started */
543 struct timeval last_issue;
544 struct timeval tv_cache;
545 unsigned int tv_cache_nr;
546 unsigned int tv_cache_mask;
547 unsigned int ramp_time_over;
550 * read/write mixed workload state
553 os_random_state_t rwmix_state;
554 struct frand_state __rwmix_state;
556 unsigned long rwmix_issues;
557 enum fio_ddir rwmix_ddir;
558 unsigned int ddir_seq_nr;
561 * IO history logs for verification. We use a tree for sorting,
562 * if we are overwriting. Otherwise just use a fifo.
564 struct rb_root io_hist_tree;
565 struct flist_head io_hist_list;
566 unsigned long io_hist_len;
571 struct flist_head io_log_list;
574 * For tracking/handling discards
576 struct flist_head trim_list;
577 unsigned long trim_entries;
580 * for fileservice, how often to switch to a new file
582 unsigned int file_service_nr;
583 unsigned int file_service_left;
584 struct fio_file *file_service_file;
586 unsigned int sync_file_range_nr;
589 * For generating file sizes
592 os_random_state_t file_size_state;
593 struct frand_state __file_size_state;
599 unsigned int total_err_count;
603 * Can be overloaded by profiles
605 struct prof_io_ops prof_io_ops;
610 * when should interactive ETA output be generated
618 #define __td_verror(td, err, msg, func) \
624 if (!(td)->first_error) \
625 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
629 #define td_clear_error(td) \
631 #define td_verror(td, err, func) \
632 __td_verror((td), (err), strerror((err)), (func))
633 #define td_vmsg(td, err, msg, func) \
634 __td_verror((td), (err), (msg), (func))
636 extern int exitall_on_terminate;
637 extern int thread_number;
638 extern int nr_process, nr_thread;
641 extern int terse_output;
642 extern int temp_stall_ts;
643 extern unsigned long long mlock_size;
644 extern unsigned long page_mask, page_size;
645 extern int read_only;
646 extern int eta_print;
647 extern unsigned long done_secs;
648 extern char *job_section;
649 extern int fio_gtod_offload;
650 extern int fio_gtod_cpu;
651 extern enum fio_cs fio_clock_source;
652 extern int warnings_fatal;
653 extern int terse_version;
654 extern int is_backend;
655 extern int is_client;
657 extern struct thread_data *threads;
659 static inline void fio_ro_check(struct thread_data *td, struct io_u *io_u)
661 assert(!(io_u->ddir == DDIR_WRITE && !td_write(td)));
664 #define BLOCKS_PER_MAP (8 * sizeof(unsigned long))
665 #define TO_MAP_BLOCK(f, b) (b)
666 #define RAND_MAP_IDX(f, b) (TO_MAP_BLOCK(f, b) / BLOCKS_PER_MAP)
667 #define RAND_MAP_BIT(f, b) (TO_MAP_BLOCK(f, b) & (BLOCKS_PER_MAP - 1))
669 #define REAL_MAX_JOBS 2048
671 #define td_non_fatal_error(e) ((e) == EIO || (e) == EILSEQ)
673 static inline void update_error_count(struct thread_data *td, int err)
675 td->total_err_count++;
676 if (td->total_err_count == 1)
677 td->first_error = err;
680 static inline int should_fsync(struct thread_data *td)
682 if (td->last_was_sync)
686 if (td_write(td) || td_rw(td) || td->o.override_sync)
693 * Init/option functions
695 extern int __must_check parse_options(int, char **);
696 extern int parse_jobs_ini(char *, int, int);
697 extern int exec_run(void);
698 extern void reset_fio_state(void);
699 extern int fio_options_parse(struct thread_data *, char **, int);
700 extern void fio_keywords_init(void);
701 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
702 extern void fio_fill_default_options(struct thread_data *);
703 extern int fio_show_option_help(const char *);
704 extern void fio_options_dup_and_init(struct option *);
705 extern void options_mem_dupe(struct thread_data *);
706 extern void options_mem_free(struct thread_data *);
707 extern void td_fill_rand_seeds(struct thread_data *);
708 extern void add_job_opts(const char **);
709 extern char *num2str(unsigned long, int, int, int);
711 #define FIO_GETOPT_JOB 0x89988998
712 #define FIO_NR_OPTIONS (FIO_MAX_OPTS + 128)
717 extern void print_thread_status(void);
718 extern void print_status_init(int);
721 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
722 * will never back again. It may cycle between running/verififying/fsyncing.
723 * Once the thread reaches TD_EXITED, it is just waiting for the core to
739 extern void td_set_runstate(struct thread_data *, int);
744 extern int __must_check fio_pin_memory(void);
745 extern void fio_unpin_memory(void);
746 extern int __must_check allocate_io_mem(struct thread_data *);
747 extern void free_io_mem(struct thread_data *);
750 * Reset stats after ramp time completes
752 extern void reset_all_stats(struct thread_data *);
757 #ifdef FIO_HAVE_BLKTRACE
758 extern int is_blktrace(const char *);
759 extern int load_blktrace(struct thread_data *, const char *);
763 * Mark unused variables passed to ops functions as unused, to silence gcc
765 #define fio_unused __attribute((__unused__))
766 #define fio_init __attribute__((constructor))
767 #define fio_exit __attribute__((destructor))
769 #define for_each_td(td, i) \
770 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
771 #define for_each_file(td, f, i) \
772 if ((td)->files_index) \
773 for ((i) = 0, (f) = (td)->files[0]; \
774 (i) < (td)->o.nr_files && ((f) = (td)->files[i]) != NULL; \
777 #define fio_assert(td, cond) do { \
780 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
781 td_set_runstate((td), TD_EXITED); \
782 (td)->error = EFAULT; \
787 static inline int fio_fill_issue_time(struct thread_data *td)
789 if (td->o.read_iolog_file ||
790 !td->o.disable_clat || !td->o.disable_slat || !td->o.disable_bw)
796 static inline int __should_check_rate(struct thread_data *td,
799 struct thread_options *o = &td->o;
802 * If some rate setting was given, we need to check it
804 if (o->rate[ddir] || o->ratemin[ddir] || o->rate_iops[ddir] ||
805 o->rate_iops_min[ddir])
811 static inline int should_check_rate(struct thread_data *td,
812 unsigned long *bytes_done)
817 ret |= __should_check_rate(td, 0);
819 ret |= __should_check_rate(td, 1);
824 static inline int is_power_of_2(unsigned int val)
826 return (val != 0 && ((val & (val - 1)) == 0));
830 * We currently only need to do locking if we have verifier threads
831 * accessing our internal structures too
833 static inline void td_io_u_lock(struct thread_data *td)
835 if (td->o.verify_async)
836 pthread_mutex_lock(&td->io_u_lock);
839 static inline void td_io_u_unlock(struct thread_data *td)
841 if (td->o.verify_async)
842 pthread_mutex_unlock(&td->io_u_lock);
845 static inline void td_io_u_free_notify(struct thread_data *td)
847 if (td->o.verify_async)
848 pthread_cond_signal(&td->free_cond);